repo_id
stringlengths
21
96
file_path
stringlengths
31
155
content
stringlengths
1
92.9M
__index_level_0__
int64
0
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail/epsilon_neighborhood.cuh
/* * Copyright (c) 2020-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/linalg/contractions.cuh> #include <raft/util/device_utils.cuh> namespace cuvs { namespace spatial { namespace knn { namespace detail { template <typename DataT, typename IdxT, typename Policy, typename BaseClass = raft::linalg::Contractions_NT<DataT, IdxT, Policy>> struct EpsUnexpL2SqNeighborhood : public BaseClass { private: typedef Policy P; bool* adj; DataT eps; IdxT* vd; char* smem; // for final reductions DataT acc[P::AccRowsPerTh][P::AccColsPerTh]; public: DI EpsUnexpL2SqNeighborhood(bool* _adj, IdxT* _vd, const DataT* _x, const DataT* _y, IdxT _m, IdxT _n, IdxT _k, DataT _eps, char* _smem) : BaseClass(_x, _y, _m, _n, _k, _smem), adj(_adj), eps(_eps), vd(_vd), smem(_smem) { } DI void run() { prolog(); loop(); epilog(); } private: DI void prolog() { this->ldgXY(IdxT(blockIdx.x) * P::Mblk, IdxT(blockIdx.y) * P::Nblk, 0); #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { acc[i][j] = BaseClass::Zero; } } this->stsXY(); __syncthreads(); this->switch_write_buffer(); } DI void loop() { for (int kidx = P::Kblk; kidx < this->k; kidx += P::Kblk) { this->ldgXY(IdxT(blockIdx.x) * P::Mblk, IdxT(blockIdx.y) * P::Nblk, kidx); accumulate(); // on the previous k-block this->stsXY(); __syncthreads(); this->switch_write_buffer(); this->switch_read_buffer(); } accumulate(); // last iteration } DI void epilog() { IdxT startx = blockIdx.x * P::Mblk + this->accrowid; IdxT starty = blockIdx.y * P::Nblk + this->acccolid; auto lid = raft::laneId(); IdxT sums[P::AccColsPerTh]; #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { sums[j] = 0; } #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { auto xid = startx + i * P::AccThRows; #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { auto yid = starty + j * P::AccThCols; auto is_neigh = acc[i][j] <= eps; ///@todo: fix uncoalesced writes using shared mem if (xid < this->m && yid < this->n) { adj[xid * this->n + yid] = is_neigh; sums[j] += is_neigh; } } } // perform reduction of adjacency values to compute vertex degrees if (vd != nullptr) { updateVertexDegree(sums); } } DI void accumulate() { #pragma unroll for (int ki = 0; ki < P::Kblk; ki += P::Veclen) { this->ldsXY(ki); #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { #pragma unroll for (int v = 0; v < P::Veclen; ++v) { auto diff = this->regx[i][v] - this->regy[j][v]; acc[i][j] += diff * diff; } } } } } DI void updateVertexDegree(IdxT (&sums)[P::AccColsPerTh]) { __syncthreads(); // so that we can safely reuse smem int gid = threadIdx.x / P::AccThCols; int lid = threadIdx.x % P::AccThCols; auto cidx = IdxT(blockIdx.y) * P::Nblk + lid; IdxT totalSum = 0; // update the individual vertex degrees #pragma unroll for (int i = 0; i < P::AccColsPerTh; ++i) { sums[i] = batchedBlockReduce<IdxT, P::AccThCols>(sums[i], smem); auto cid = cidx + i * P::AccThCols; if (gid == 0 && cid < this->n) { atomicUpdate(cid, sums[i]); totalSum += sums[i]; } __syncthreads(); // for safe smem reuse } // update the total edge count totalSum = raft::blockReduce<IdxT>(totalSum, smem); if (threadIdx.x == 0) { atomicUpdate(this->n, totalSum); } } DI void atomicUpdate(IdxT addrId, IdxT val) { if (sizeof(IdxT) == 4) { raft::myAtomicAdd<unsigned>((unsigned*)(vd + addrId), val); } else if (sizeof(IdxT) == 8) { raft::myAtomicAdd<unsigned long long>((unsigned long long*)(vd + addrId), val); } } }; // struct EpsUnexpL2SqNeighborhood template <typename DataT, typename IdxT, typename Policy> __launch_bounds__(Policy::Nthreads, 2) RAFT_KERNEL epsUnexpL2SqNeighKernel( bool* adj, IdxT* vd, const DataT* x, const DataT* y, IdxT m, IdxT n, IdxT k, DataT eps) { extern __shared__ char smem[]; EpsUnexpL2SqNeighborhood<DataT, IdxT, Policy> obj(adj, vd, x, y, m, n, k, eps, smem); obj.run(); } template <typename DataT, typename IdxT, int VecLen> void epsUnexpL2SqNeighImpl(bool* adj, IdxT* vd, const DataT* x, const DataT* y, IdxT m, IdxT n, IdxT k, DataT eps, cudaStream_t stream) { typedef typename raft::linalg::Policy4x4<DataT, VecLen>::Policy Policy; dim3 grid(raft::ceildiv<int>(m, Policy::Mblk), raft::ceildiv<int>(n, Policy::Nblk)); dim3 blk(Policy::Nthreads); epsUnexpL2SqNeighKernel<DataT, IdxT, Policy> <<<grid, blk, Policy::SmemSize, stream>>>(adj, vd, x, y, m, n, k, eps); RAFT_CUDA_TRY(cudaGetLastError()); } /** * @brief Computes epsilon neighborhood for the L2-Squared distance metric * * @tparam DataT IO and math type * @tparam IdxT Index type * * @param[out] adj adjacency matrix [row-major] [on device] [dim = m x n] * @param[out] vd vertex degree array [on device] [len = m + 1] * `vd + m` stores the total number of edges in the adjacency * matrix. Pass a nullptr if you don't need this info. * @param[in] x first matrix [row-major] [on device] [dim = m x k] * @param[in] y second matrix [row-major] [on device] [dim = n x k] * @param[in] eps defines epsilon neighborhood radius (should be passed as * squared as we compute L2-squared distance in this method) * @param[in] fop device lambda to do any other custom functions * @param[in] stream cuda stream */ template <typename DataT, typename IdxT> void epsUnexpL2SqNeighborhood(bool* adj, IdxT* vd, const DataT* x, const DataT* y, IdxT m, IdxT n, IdxT k, DataT eps, cudaStream_t stream) { size_t bytes = sizeof(DataT) * k; if (16 % sizeof(DataT) == 0 && bytes % 16 == 0) { epsUnexpL2SqNeighImpl<DataT, IdxT, 16 / sizeof(DataT)>(adj, vd, x, y, m, n, k, eps, stream); } else if (8 % sizeof(DataT) == 0 && bytes % 8 == 0) { epsUnexpL2SqNeighImpl<DataT, IdxT, 8 / sizeof(DataT)>(adj, vd, x, y, m, n, k, eps, stream); } else { epsUnexpL2SqNeighImpl<DataT, IdxT, 1>(adj, vd, x, y, m, n, k, eps, stream); } } } // namespace detail } // namespace knn } // namespace spatial } // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail/ann_utils.cuh
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/distance/distance_types.hpp> #include <raft/core/logger.hpp> #include <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <raft/util/integer_utils.hpp> #include <rmm/cuda_stream_view.hpp> #include <rmm/device_scalar.hpp> #include <rmm/device_uvector.hpp> #include <memory> #include <optional> #include <cuda_fp16.hpp> namespace cuvs::spatial::knn::detail::utils { /** Whether pointers are accessible on the device or on the host. */ enum class pointer_residency { /** Some of the pointers are on the device, some on the host. */ mixed, /** All pointers accessible from both the device and the host. */ host_and_device, /** All pointers are host accessible. */ host_only, /** All poitners are device accessible. */ device_only }; template <typename... Types> struct pointer_residency_count {}; template <> struct pointer_residency_count<> { static inline auto run() -> std::tuple<int, int> { return std::make_tuple(0, 0); } }; template <typename Type, typename... Types> struct pointer_residency_count<Type, Types...> { static inline auto run(const Type* ptr, const Types*... ptrs) -> std::tuple<int, int> { auto [on_device, on_host] = pointer_residency_count<Types...>::run(ptrs...); cudaPointerAttributes attr; RAFT_CUDA_TRY(cudaPointerGetAttributes(&attr, ptr)); switch (attr.type) { case cudaMemoryTypeUnregistered: return std::make_tuple(on_device, on_host + 1); case cudaMemoryTypeHost: return std::make_tuple(on_device + int(attr.devicePointer == ptr), on_host + 1); case cudaMemoryTypeDevice: return std::make_tuple(on_device + 1, on_host); case cudaMemoryTypeManaged: return std::make_tuple(on_device + 1, on_host + 1); default: return std::make_tuple(on_device, on_host); } } }; /** Check if all argument pointers reside on the host or on the device. */ template <typename... Types> auto check_pointer_residency(const Types*... ptrs) -> pointer_residency { auto [on_device, on_host] = pointer_residency_count<Types...>::run(ptrs...); int n_args = sizeof...(Types); if (on_device == n_args && on_host == n_args) { return pointer_residency::host_and_device; } if (on_device == n_args) { return pointer_residency::device_only; } if (on_host == n_args) { return pointer_residency::host_only; } return pointer_residency::mixed; } /** RAII helper to access the host data from gpu when necessary. */ template <typename PtrT, typename Action> struct with_mapped_memory_t { with_mapped_memory_t(PtrT ptr, size_t size, Action action) : action_(action) { if (ptr == nullptr) { return; } switch (utils::check_pointer_residency(ptr)) { case utils::pointer_residency::device_only: case utils::pointer_residency::host_and_device: { dev_ptr_ = (void*)ptr; // NOLINT } break; default: { host_ptr_ = (void*)ptr; // NOLINT RAFT_CUDA_TRY(cudaHostRegister(host_ptr_, size, choose_flags(ptr))); RAFT_CUDA_TRY(cudaHostGetDevicePointer(&dev_ptr_, host_ptr_, 0)); } break; } } ~with_mapped_memory_t() { if (host_ptr_ != nullptr) { cudaHostUnregister(host_ptr_); } } auto operator()() { return action_((PtrT)dev_ptr_); } // NOLINT private: Action action_; void* host_ptr_ = nullptr; void* dev_ptr_ = nullptr; template <typename T> static auto choose_flags(const T*) -> unsigned int { int dev_id, readonly_supported; RAFT_CUDA_TRY(cudaGetDevice(&dev_id)); RAFT_CUDA_TRY(cudaDeviceGetAttribute( &readonly_supported, cudaDevAttrHostRegisterReadOnlySupported, dev_id)); if (readonly_supported) { return cudaHostRegisterMapped | cudaHostRegisterReadOnly; } else { return cudaHostRegisterMapped; } } template <typename T> static auto choose_flags(T*) -> unsigned int { return cudaHostRegisterMapped; } }; template <typename T> struct config {}; template <> struct config<double> { using value_t = double; static constexpr double kDivisor = 1.0; }; template <> struct config<float> { using value_t = float; static constexpr double kDivisor = 1.0; }; template <> struct config<half> { using value_t = half; static constexpr double kDivisor = 1.0; }; template <> struct config<uint8_t> { using value_t = uint32_t; static constexpr double kDivisor = 256.0; }; template <> struct config<int8_t> { using value_t = int32_t; static constexpr double kDivisor = 128.0; }; /** * @brief Converting values between the types taking into account scaling factors * for the integral types. * * @tparam T target type of the mapping. */ template <typename T> struct mapping { /** * @defgroup * @brief Cast and possibly scale a value of the source type `S` to the target type `T`. * * @tparam S source type * @param x source value * @{ */ template <typename S> HDI constexpr auto operator()(const S& x) const -> std::enable_if_t<std::is_same_v<S, T>, T> { return x; }; template <typename S> HDI constexpr auto operator()(const S& x) const -> std::enable_if_t<!std::is_same_v<S, T>, T> { constexpr double kMult = config<T>::kDivisor / config<S>::kDivisor; if constexpr (std::is_floating_point_v<S>) { return static_cast<T>(x * static_cast<S>(kMult)); } if constexpr (std::is_floating_point_v<T>) { return static_cast<T>(x) * static_cast<T>(kMult); } return static_cast<T>(static_cast<float>(x) * static_cast<float>(kMult)); }; /** @} */ }; /** * @brief Sets the first num bytes of the block of memory pointed by ptr to the specified value. * * @param[out] ptr host or device pointer * @param[in] value * @param[in] n_bytes */ template <typename T, typename IdxT> inline void memzero(T* ptr, IdxT n_elems, rmm::cuda_stream_view stream) { switch (check_pointer_residency(ptr)) { case pointer_residency::host_and_device: case pointer_residency::device_only: { RAFT_CUDA_TRY(cudaMemsetAsync(ptr, 0, n_elems * sizeof(T), stream)); } break; case pointer_residency::host_only: { stream.synchronize(); ::memset(ptr, 0, n_elems * sizeof(T)); } break; default: RAFT_FAIL("memset: unreachable code"); } } template <typename T, typename IdxT> RAFT_KERNEL outer_add_kernel(const T* a, IdxT len_a, const T* b, IdxT len_b, T* c) { IdxT gid = threadIdx.x + blockDim.x * static_cast<IdxT>(blockIdx.x); IdxT i = gid / len_b; IdxT j = gid % len_b; if (i >= len_a) return; c[gid] = (a == nullptr ? T(0) : a[i]) + (b == nullptr ? T(0) : b[j]); } template <typename T, typename IdxT> RAFT_KERNEL block_copy_kernel(const IdxT* in_offsets, const IdxT* out_offsets, IdxT n_blocks, const T* in_data, T* out_data, IdxT n_mult) { IdxT i = static_cast<IdxT>(blockDim.x) * static_cast<IdxT>(blockIdx.x) + threadIdx.x; // find the source offset using the binary search. uint32_t l = 0; uint32_t r = n_blocks; IdxT in_offset = 0; if (in_offsets[r] * n_mult <= i) return; while (l + 1 < r) { uint32_t c = (l + r) >> 1; IdxT o = in_offsets[c] * n_mult; if (o <= i) { l = c; in_offset = o; } else { r = c; } } // copy the data out_data[out_offsets[l] * n_mult - in_offset + i] = in_data[i]; } /** * Copy chunks of data from one array to another at given offsets. * * @tparam T element type * @tparam IdxT index type * * @param[in] in_offsets * @param[in] out_offsets * @param n_blocks size of the offset arrays minus one. * @param[in] in_data * @param[out] out_data * @param n_mult constant multiplier for offset values (such as e.g. `dim`) * @param stream */ template <typename T, typename IdxT> void block_copy(const IdxT* in_offsets, const IdxT* out_offsets, IdxT n_blocks, const T* in_data, T* out_data, IdxT n_mult, rmm::cuda_stream_view stream) { IdxT in_size; update_host(&in_size, in_offsets + n_blocks, 1, stream); stream.synchronize(); dim3 threads(128, 1, 1); dim3 blocks(raft::ceildiv<IdxT>(in_size * n_mult, threads.x), 1, 1); block_copy_kernel<<<blocks, threads, 0, stream>>>( in_offsets, out_offsets, n_blocks, in_data, out_data, n_mult); } /** * @brief Fill matrix `c` with all combinations of sums of vectors `a` and `b`. * * NB: device-only function * * @tparam T element type * @tparam IdxT index type * * @param[in] a device pointer to a vector [len_a] * @param len_a number of elements in `a` * @param[in] b device pointer to a vector [len_b] * @param len_b number of elements in `b` * @param[out] c row-major matrix [len_a, len_b] * @param stream */ template <typename T, typename IdxT> void outer_add(const T* a, IdxT len_a, const T* b, IdxT len_b, T* c, rmm::cuda_stream_view stream) { dim3 threads(128, 1, 1); dim3 blocks(raft::ceildiv<IdxT>(len_a * len_b, threads.x), 1, 1); outer_add_kernel<<<blocks, threads, 0, stream>>>(a, len_a, b, len_b, c); } template <typename T, typename S, typename IdxT, typename LabelT> RAFT_KERNEL copy_selected_kernel( IdxT n_rows, IdxT n_cols, const S* src, const LabelT* row_ids, IdxT ld_src, T* dst, IdxT ld_dst) { IdxT gid = threadIdx.x + blockDim.x * static_cast<IdxT>(blockIdx.x); IdxT j = gid % n_cols; IdxT i_dst = gid / n_cols; if (i_dst >= n_rows) return; auto i_src = static_cast<IdxT>(row_ids[i_dst]); dst[ld_dst * i_dst + j] = mapping<T>{}(src[ld_src * i_src + j]); } /** * @brief Copy selected rows of a matrix while mapping the data from the source to the target * type. * * @tparam T target type * @tparam S source type * @tparam IdxT index type * @tparam LabelT label type * * @param n_rows * @param n_cols * @param[in] src input matrix [..., ld_src] * @param[in] row_ids selection of rows to be copied [n_rows] * @param ld_src number of cols in the input (ld_src >= n_cols) * @param[out] dst output matrix [n_rows, ld_dst] * @param ld_dst number of cols in the output (ld_dst >= n_cols) * @param stream */ template <typename T, typename S, typename IdxT, typename LabelT> void copy_selected(IdxT n_rows, IdxT n_cols, const S* src, const LabelT* row_ids, IdxT ld_src, T* dst, IdxT ld_dst, rmm::cuda_stream_view stream) { switch (check_pointer_residency(src, dst, row_ids)) { case pointer_residency::host_and_device: case pointer_residency::device_only: { IdxT block_dim = 128; IdxT grid_dim = raft::ceildiv(n_rows * n_cols, block_dim); copy_selected_kernel<T, S> <<<grid_dim, block_dim, 0, stream>>>(n_rows, n_cols, src, row_ids, ld_src, dst, ld_dst); } break; case pointer_residency::host_only: { stream.synchronize(); for (IdxT i_dst = 0; i_dst < n_rows; i_dst++) { auto i_src = static_cast<IdxT>(row_ids[i_dst]); for (IdxT j = 0; j < n_cols; j++) { dst[ld_dst * i_dst + j] = mapping<T>{}(src[ld_src * i_src + j]); } } stream.synchronize(); } break; default: RAFT_FAIL("All pointers must reside on the same side, host or device."); } } /** * A batch input iterator over the data source. * Given an input pointer, it decides whether the current device has the access to the data and * gives it back to the user in batches. Three scenarios are possible: * * 1. if `source == nullptr`: then `batch.data() == nullptr` * 2. if `source` is accessible from the device, `batch.data()` points directly at the source at * the proper offsets on each iteration. * 3. if `source` is not accessible from the device, `batch.data()` points to an intermediate * buffer; the corresponding data is copied in the given `stream` on every iterator dereference * (i.e. batches can be skipped). Dereferencing the same batch two times in a row does not force * the copy. * * In all three scenarios, the number of iterations, batch offsets and sizes are the same. * * The iterator can be reused. If the number of iterations is one, at most one copy will ever be * invoked (i.e. small datasets are not reloaded multiple times). */ template <typename T> struct batch_load_iterator { using size_type = size_t; /** A single batch of data residing in device memory. */ struct batch { /** Logical width of a single row in a batch, in elements of type `T`. */ [[nodiscard]] auto row_width() const -> size_type { return row_width_; } /** Logical offset of the batch, in rows (`row_width()`) */ [[nodiscard]] auto offset() const -> size_type { return pos_.value_or(0) * batch_size_; } /** Logical size of the batch, in rows (`row_width()`) */ [[nodiscard]] auto size() const -> size_type { return batch_len_; } /** Logical size of the batch, in rows (`row_width()`) */ [[nodiscard]] auto data() const -> const T* { return const_cast<const T*>(dev_ptr_); } /** Whether this batch copies the data (i.e. the source is inaccessible from the device). */ [[nodiscard]] auto does_copy() const -> bool { return needs_copy_; } private: batch(const T* source, size_type n_rows, size_type row_width, size_type batch_size, rmm::cuda_stream_view stream, rmm::mr::device_memory_resource* mr) : stream_(stream), buf_(0, stream, mr), source_(source), dev_ptr_(nullptr), n_rows_(n_rows), row_width_(row_width), batch_size_(std::min(batch_size, n_rows)), pos_(std::nullopt), n_iters_(raft::div_rounding_up_safe(n_rows, batch_size)), needs_copy_(false) { if (source_ == nullptr) { return; } cudaPointerAttributes attr; RAFT_CUDA_TRY(cudaPointerGetAttributes(&attr, source_)); dev_ptr_ = reinterpret_cast<T*>(attr.devicePointer); if (dev_ptr_ == nullptr) { buf_.resize(row_width_ * batch_size_, stream); dev_ptr_ = buf_.data(); needs_copy_ = true; } } rmm::cuda_stream_view stream_; rmm::device_uvector<T> buf_; const T* source_; size_type n_rows_; size_type row_width_; size_type batch_size_; size_type n_iters_; bool needs_copy_; std::optional<size_type> pos_; size_type batch_len_; T* dev_ptr_; friend class batch_load_iterator<T>; /** * Changes the state of the batch to point at the `pos` index. * If necessary, copies the data from the source in the registered stream. */ void load(const size_type& pos) { // No-op if the data is already loaded, or it's the end of the input. if (pos == pos_ || pos >= n_iters_) { return; } pos_.emplace(pos); batch_len_ = std::min(batch_size_, n_rows_ - std::min(offset(), n_rows_)); if (source_ == nullptr) { return; } if (needs_copy_) { if (size() > 0) { RAFT_LOG_TRACE("batch_load_iterator::copy(offset = %zu, size = %zu, row_width = %zu)", size_t(offset()), size_t(size()), size_t(row_width())); copy(dev_ptr_, source_ + offset() * row_width(), size() * row_width(), stream_); } } else { dev_ptr_ = const_cast<T*>(source_) + offset() * row_width(); } } }; using value_type = batch; using reference = const value_type&; using pointer = const value_type*; /** * Create a batch iterator over the data `source`. * * For convenience, the data `source` is read in logical units of size `row_width`; batch sizes * and offsets are calculated in logical rows. Hence, can interpret the data as a contiguous * row-major matrix of size [n_rows, row_width], and the batches are the sub-matrices of size * [x<=batch_size, n_rows]. * * @param source the input data -- host, device, or nullptr. * @param n_rows the size of the input in logical rows. * @param row_width the size of the logical row in the elements of type `T`. * @param batch_size the desired size of the batch. * @param stream the ordering for the host->device copies, if applicable. * @param mr a custom memory resource for the intermediate buffer, if applicable. */ batch_load_iterator(const T* source, size_type n_rows, size_type row_width, size_type batch_size, rmm::cuda_stream_view stream, rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource()) : cur_batch_(new batch(source, n_rows, row_width, batch_size, stream, mr)), cur_pos_(0) { } /** * Whether this iterator copies the data on every iteration * (i.e. the source is inaccessible from the device). */ [[nodiscard]] auto does_copy() const -> bool { return cur_batch_->does_copy(); } /** Reset the iterator position to `begin()` */ void reset() { cur_pos_ = 0; } /** Reset the iterator position to `end()` */ void reset_to_end() { cur_pos_ = cur_batch_->n_iters_; } [[nodiscard]] auto begin() const -> const batch_load_iterator<T> { batch_load_iterator<T> x(*this); x.reset(); return x; } [[nodiscard]] auto end() const -> const batch_load_iterator<T> { batch_load_iterator<T> x(*this); x.reset_to_end(); return x; } [[nodiscard]] auto operator*() const -> reference { cur_batch_->load(cur_pos_); return *cur_batch_; } [[nodiscard]] auto operator->() const -> pointer { cur_batch_->load(cur_pos_); return cur_batch_.get(); } friend auto operator==(const batch_load_iterator<T>& x, const batch_load_iterator<T>& y) -> bool { return x.cur_batch_ == y.cur_batch_ && x.cur_pos_ == y.cur_pos_; }; friend auto operator!=(const batch_load_iterator<T>& x, const batch_load_iterator<T>& y) -> bool { return x.cur_batch_ != y.cur_batch_ || x.cur_pos_ != y.cur_pos_; }; auto operator++() -> batch_load_iterator<T>& { ++cur_pos_; return *this; } auto operator++(int) -> batch_load_iterator<T> { batch_load_iterator<T> x(*this); ++cur_pos_; return x; } auto operator--() -> batch_load_iterator<T>& { --cur_pos_; return *this; } auto operator--(int) -> batch_load_iterator<T> { batch_load_iterator<T> x(*this); --cur_pos_; return x; } private: std::shared_ptr<value_type> cur_batch_; size_type cur_pos_; }; } // namespace cuvs::spatial::knn::detail::utils
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail/ball_cover.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resource/thrust_policy.hpp> #include <raft/core/resources.hpp> #include "../ball_cover_types.hpp" #include "ball_cover/common.cuh" #include "ball_cover/registers.cuh" #include "haversine_distance.cuh" #include <cstdint> #include <limits.h> #include <raft/util/cuda_utils.cuh> #include <cuvs/neighbors/detail/faiss_select/key_value_block_select.cuh> #include <cuvs/neighbors/brute_force.cuh> #include <raft/matrix/copy.cuh> #include <raft/random/rng.cuh> #include <raft/sparse/convert/csr.cuh> #include <rmm/device_uvector.hpp> #include <rmm/exec_policy.hpp> #include <thrust/fill.h> #include <thrust/for_each.h> #include <thrust/functional.h> #include <thrust/iterator/counting_iterator.h> #include <thrust/iterator/zip_iterator.h> #include <thrust/reduce.h> #include <thrust/sequence.h> #include <thrust/sort.h> #include <thrust/tuple.h> namespace cuvs { namespace spatial { namespace knn { namespace detail { /** * Given a set of points in row-major order which are to be * used as a set of index points, uniformly samples a subset * of points to be used as landmarks. * @tparam value_idx * @tparam value_t * @param handle * @param index */ template <typename value_idx, typename value_t, typename value_int = std::uint32_t> void sample_landmarks(raft::resources const& handle, BallCoverIndex<value_idx, value_t, value_int>& index) { rmm::device_uvector<value_idx> R_1nn_cols2(index.n_landmarks, resource::get_cuda_stream(handle)); rmm::device_uvector<value_t> R_1nn_ones(index.m, resource::get_cuda_stream(handle)); rmm::device_uvector<value_idx> R_indices(index.n_landmarks, resource::get_cuda_stream(handle)); thrust::sequence(raft::resource::get_thrust_policy(handle), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_cols().data_handle() + index.m, (value_idx)0); thrust::fill(raft::resource::get_thrust_policy(handle), R_1nn_ones.data(), R_1nn_ones.data() + R_1nn_ones.size(), 1.0); thrust::fill(raft::resource::get_thrust_policy(handle), R_indices.data(), R_indices.data() + R_indices.size(), 0.0); /** * 1. Randomly sample sqrt(n) points from X */ raft::random::RngState rng_state(12345); raft::random::sampleWithoutReplacement(handle, rng_state, R_indices.data(), R_1nn_cols2.data(), index.get_R_1nn_cols().data_handle(), R_1nn_ones.data(), (value_idx)index.n_landmarks, (value_idx)index.m); // index.get_X() returns the wrong indextype (uint32_t where we need value_idx), so need to // create new device_matrix_view here auto x = index.get_X(); auto r = index.get_R(); raft::matrix::copy_rows<value_t, value_idx>( handle, make_device_matrix_view<const value_t, value_idx>(x.data_handle(), x.extent(0), x.extent(1)), make_device_matrix_view<value_t, value_idx>(r.data_handle(), r.extent(0), r.extent(1)), make_device_vector_view(R_1nn_cols2.data(), index.n_landmarks)); } /** * Constructs a 1-nn index mapping each landmark to their closest points. * @tparam value_idx * @tparam value_t * @param handle * @param R_knn_inds_ptr * @param R_knn_dists_ptr * @param k * @param index */ template <typename value_idx, typename value_t, typename value_int = std::uint32_t> void construct_landmark_1nn(raft::resources const& handle, const value_idx* R_knn_inds_ptr, const value_t* R_knn_dists_ptr, value_int k, BallCoverIndex<value_idx, value_t, value_int>& index) { rmm::device_uvector<value_idx> R_1nn_inds(index.m, resource::get_cuda_stream(handle)); thrust::fill(raft::resource::get_thrust_policy(handle), R_1nn_inds.data(), R_1nn_inds.data() + index.m, std::numeric_limits<value_idx>::max()); value_idx* R_1nn_inds_ptr = R_1nn_inds.data(); value_t* R_1nn_dists_ptr = index.get_R_1nn_dists().data_handle(); auto idxs = thrust::make_counting_iterator<value_idx>(0); thrust::for_each( resource::get_thrust_policy(handle), idxs, idxs + index.m, [=] __device__(value_idx i) { R_1nn_inds_ptr[i] = R_knn_inds_ptr[i * k]; R_1nn_dists_ptr[i] = R_knn_dists_ptr[i * k]; }); auto keys = thrust::make_zip_iterator( thrust::make_tuple(R_1nn_inds.data(), index.get_R_1nn_dists().data_handle())); // group neighborhoods for each reference landmark and sort each group by distance thrust::sort_by_key(raft::resource::get_thrust_policy(handle), keys, keys + index.m, index.get_R_1nn_cols().data_handle(), NNComp()); // convert to CSR for fast lookup raft::sparse::convert::sorted_coo_to_csr(R_1nn_inds.data(), index.m, index.get_R_indptr().data_handle(), index.n_landmarks + 1, resource::get_cuda_stream(handle)); } /** * Computes the k closest landmarks to a set of query points. * @tparam value_idx * @tparam value_t * @tparam value_int * @param handle * @param index * @param query_pts * @param n_query_pts * @param k * @param R_knn_inds * @param R_knn_dists */ template <typename value_idx, typename value_t, typename value_int = std::uint32_t> void k_closest_landmarks(raft::resources const& handle, const BallCoverIndex<value_idx, value_t, value_int>& index, const value_t* query_pts, value_int n_query_pts, value_int k, value_idx* R_knn_inds, value_t* R_knn_dists) { std::vector<raft::device_matrix_view<const value_t, value_int>> inputs = {index.get_R()}; cuvs::neighbors::brute_force::knn<value_idx, value_t, value_int>( handle, inputs, make_device_matrix_view(query_pts, n_query_pts, inputs[0].extent(1)), make_device_matrix_view(R_knn_inds, n_query_pts, k), make_device_matrix_view(R_knn_dists, n_query_pts, k), index.get_metric()); } /** * Uses the sorted data points in the 1-nn landmark index to compute * an array of radii for each landmark. * @tparam value_idx * @tparam value_t * @param handle * @param index */ template <typename value_idx, typename value_t, typename value_int = std::uint32_t> void compute_landmark_radii(raft::resources const& handle, BallCoverIndex<value_idx, value_t, value_int>& index) { auto entries = thrust::make_counting_iterator<value_idx>(0); const value_idx* R_indptr_ptr = index.get_R_indptr().data_handle(); const value_t* R_1nn_dists_ptr = index.get_R_1nn_dists().data_handle(); value_t* R_radius_ptr = index.get_R_radius().data_handle(); thrust::for_each(raft::resource::get_thrust_policy(handle), entries, entries + index.n_landmarks, [=] __device__(value_idx input) { value_idx last_row_idx = R_indptr_ptr[input + 1] - 1; R_radius_ptr[input] = R_1nn_dists_ptr[last_row_idx]; }); } /** * 4. Perform k-select over original KNN, using L_r to filter distances * * a. Map 1 row to each warp/block * b. Add closest k R points to heap * c. Iterate through batches of R, having each thread in the warp load a set * of distances y from R (only if d(q, r) < 3 * distance to closest r) and * marking the distance to be computed between x, y only * if knn[k].distance >= d(x_i, R_k) + d(R_k, y) */ template <typename value_idx, typename value_t, typename value_int = std::uint32_t, typename dist_func> void perform_rbc_query(raft::resources const& handle, const BallCoverIndex<value_idx, value_t, value_int>& index, const value_t* query, value_int n_query_pts, std::uint32_t k, const value_idx* R_knn_inds, const value_t* R_knn_dists, dist_func dfunc, value_idx* inds, value_t* dists, value_int* dists_counter, value_int* post_dists_counter, float weight = 1.0, bool perform_post_filtering = true) { // initialize output inds and dists thrust::fill(raft::resource::get_thrust_policy(handle), inds, inds + (k * n_query_pts), std::numeric_limits<value_idx>::max()); thrust::fill(raft::resource::get_thrust_policy(handle), dists, dists + (k * n_query_pts), std::numeric_limits<value_t>::max()); if (index.n == 2) { // Compute nearest k for each neighborhood in each closest R rbc_low_dim_pass_one<value_idx, value_t, value_int, 2>(handle, index, query, n_query_pts, k, R_knn_inds, R_knn_dists, dfunc, inds, dists, weight, dists_counter); if (perform_post_filtering) { rbc_low_dim_pass_two<value_idx, value_t, value_int, 2>(handle, index, query, n_query_pts, k, R_knn_inds, R_knn_dists, dfunc, inds, dists, weight, post_dists_counter); } } else if (index.n == 3) { // Compute nearest k for each neighborhood in each closest R rbc_low_dim_pass_one<value_idx, value_t, value_int, 3>(handle, index, query, n_query_pts, k, R_knn_inds, R_knn_dists, dfunc, inds, dists, weight, dists_counter); if (perform_post_filtering) { rbc_low_dim_pass_two<value_idx, value_t, value_int, 3>(handle, index, query, n_query_pts, k, R_knn_inds, R_knn_dists, dfunc, inds, dists, weight, post_dists_counter); } } } /** * Similar to a ball tree, the random ball cover algorithm * uses the triangle inequality to prune distance computations * in any metric space with a guarantee of sqrt(n) * c^{3/2} * where `c` is an expansion constant based on the distance * metric. * * This function variant performs an all nearest neighbors * query which is useful for algorithms that need to perform * A * A.T. */ template <typename value_idx = std::int64_t, typename value_t, typename value_int = std::uint32_t, typename distance_func> void rbc_build_index(raft::resources const& handle, BallCoverIndex<value_idx, value_t, value_int>& index, distance_func dfunc) { ASSERT(index.n <= 3, "only 2d and 3d vectors are supported in current implementation"); ASSERT(!index.is_index_trained(), "index cannot be previously trained"); rmm::device_uvector<value_idx> R_knn_inds(index.m, resource::get_cuda_stream(handle)); // Initialize the uvectors thrust::fill(raft::resource::get_thrust_policy(handle), R_knn_inds.begin(), R_knn_inds.end(), std::numeric_limits<value_idx>::max()); thrust::fill(raft::resource::get_thrust_policy(handle), index.get_R_closest_landmark_dists().data_handle(), index.get_R_closest_landmark_dists().data_handle() + index.m, std::numeric_limits<value_t>::max()); /** * 1. Randomly sample sqrt(n) points from X */ sample_landmarks<value_idx, value_t>(handle, index); /** * 2. Perform knn = bfknn(X, R, k) */ value_int k = 1; k_closest_landmarks(handle, index, index.get_X().data_handle(), index.m, k, R_knn_inds.data(), index.get_R_closest_landmark_dists().data_handle()); /** * 3. Create L_r = knn[:,0].T (CSR) * * Slice closest neighboring R * Secondary sort by (R_knn_inds, R_knn_dists) */ construct_landmark_1nn( handle, R_knn_inds.data(), index.get_R_closest_landmark_dists().data_handle(), k, index); /** * Compute radius of each R for filtering: p(q, r) <= p(q, q_r) + radius(r) * (need to take the */ compute_landmark_radii(handle, index); } /** * Performs an all neighbors knn query (e.g. index == query) */ template <typename value_idx = std::int64_t, typename value_t, typename value_int = std::uint32_t, typename distance_func> void rbc_all_knn_query(raft::resources const& handle, BallCoverIndex<value_idx, value_t, value_int>& index, value_int k, value_idx* inds, value_t* dists, distance_func dfunc, // approximate nn options bool perform_post_filtering = true, float weight = 1.0) { ASSERT(index.n <= 3, "only 2d and 3d vectors are supported in current implementation"); ASSERT(index.n_landmarks >= k, "number of landmark samples must be >= k"); ASSERT(!index.is_index_trained(), "index cannot be previously trained"); rmm::device_uvector<value_idx> R_knn_inds(k * index.m, raft::resource::get_cuda_stream(handle)); rmm::device_uvector<value_t> R_knn_dists(k * index.m, raft::resource::get_cuda_stream(handle)); // Initialize the uvectors thrust::fill(raft::resource::get_thrust_policy(handle), R_knn_inds.begin(), R_knn_inds.end(), std::numeric_limits<value_idx>::max()); thrust::fill(raft::resource::get_thrust_policy(handle), R_knn_dists.begin(), R_knn_dists.end(), std::numeric_limits<value_t>::max()); thrust::fill(raft::resource::get_thrust_policy(handle), inds, inds + (k * index.m), std::numeric_limits<value_idx>::max()); thrust::fill(raft::resource::get_thrust_policy(handle), dists, dists + (k * index.m), std::numeric_limits<value_t>::max()); // For debugging / verification. Remove before releasing rmm::device_uvector<value_int> dists_counter(index.m, raft::resource::get_cuda_stream(handle)); rmm::device_uvector<value_int> post_dists_counter(index.m, raft::resource::get_cuda_stream(handle)); sample_landmarks<value_idx, value_t>(handle, index); k_closest_landmarks( handle, index, index.get_X().data_handle(), index.m, k, R_knn_inds.data(), R_knn_dists.data()); construct_landmark_1nn(handle, R_knn_inds.data(), R_knn_dists.data(), k, index); compute_landmark_radii(handle, index); perform_rbc_query(handle, index, index.get_X().data_handle(), index.m, k, R_knn_inds.data(), R_knn_dists.data(), dfunc, inds, dists, dists_counter.data(), post_dists_counter.data(), weight, perform_post_filtering); } /** * Performs a knn query against an index. This assumes the index has * already been built. */ template <typename value_idx = std::int64_t, typename value_t, typename value_int = std::uint32_t, typename distance_func> void rbc_knn_query(raft::resources const& handle, const BallCoverIndex<value_idx, value_t, value_int>& index, value_int k, const value_t* query, value_int n_query_pts, value_idx* inds, value_t* dists, distance_func dfunc, // approximate nn options bool perform_post_filtering = true, float weight = 1.0) { ASSERT(index.n <= 3, "only 2d and 3d vectors are supported in current implementation"); ASSERT(index.n_landmarks >= k, "number of landmark samples must be >= k"); ASSERT(index.is_index_trained(), "index must be previously trained"); rmm::device_uvector<value_idx> R_knn_inds(k * n_query_pts, raft::resource::get_cuda_stream(handle)); rmm::device_uvector<value_t> R_knn_dists(k * n_query_pts, raft::resource::get_cuda_stream(handle)); // Initialize the uvectors thrust::fill(raft::resource::get_thrust_policy(handle), R_knn_inds.begin(), R_knn_inds.end(), std::numeric_limits<value_idx>::max()); thrust::fill(raft::resource::get_thrust_policy(handle), R_knn_dists.begin(), R_knn_dists.end(), std::numeric_limits<value_t>::max()); thrust::fill(raft::resource::get_thrust_policy(handle), inds, inds + (k * n_query_pts), std::numeric_limits<value_idx>::max()); thrust::fill(raft::resource::get_thrust_policy(handle), dists, dists + (k * n_query_pts), std::numeric_limits<value_t>::max()); k_closest_landmarks(handle, index, query, n_query_pts, k, R_knn_inds.data(), R_knn_dists.data()); // For debugging / verification. Remove before releasing rmm::device_uvector<value_int> dists_counter(index.m, raft::resource::get_cuda_stream(handle)); rmm::device_uvector<value_int> post_dists_counter(index.m, raft::resource::get_cuda_stream(handle)); thrust::fill(raft::resource::get_thrust_policy(handle), post_dists_counter.data(), post_dists_counter.data() + post_dists_counter.size(), 0); thrust::fill(raft::resource::get_thrust_policy(handle), dists_counter.data(), dists_counter.data() + dists_counter.size(), 0); perform_rbc_query(handle, index, query, n_query_pts, k, R_knn_inds.data(), R_knn_dists.data(), dfunc, inds, dists, dists_counter.data(), post_dists_counter.data(), weight, perform_post_filtering); } }; // namespace detail }; // namespace knn }; // namespace spatial }; // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail/ann_quantized.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "../ann_common.h" #include "../ivf_flat.cuh" #include <raft/core/resource/cuda_stream.hpp> #include "processing.cuh" #include <raft/core/operators.hpp> #include <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <cuvs/distance/distance.cuh> #include <cuvs/distance/distance_types.hpp> #include <cuvs/neighbors/ivf_pq.cuh> #include <raft/label/classlabels.cuh> #include <raft/core/device_mdspan.hpp> #include <rmm/cuda_stream_view.hpp> #include <thrust/iterator/transform_iterator.h> namespace cuvs::spatial::knn::detail { template <typename T = float, typename IntType = int> void approx_knn_build_index(raft::resources const& handle, knnIndex* index, knnIndexParam* params, cuvs::distance::DistanceType metric, float metricArg, T* index_array, IntType n, IntType D) { auto stream = resource::get_cuda_stream(handle); index->metric = metric; index->metricArg = metricArg; if (dynamic_cast<const IVFParam*>(params)) { index->nprobe = dynamic_cast<const IVFParam*>(params)->nprobe; } auto ivf_ft_pams = dynamic_cast<IVFFlatParam*>(params); auto ivf_pq_pams = dynamic_cast<IVFPQParam*>(params); if constexpr (std::is_same_v<T, float>) { index->metric_processor = create_processor<float>(metric, n, D, 0, false, stream); // For cosine/correlation distance, the metric processor translates distance // to inner product via pre/post processing - pass the translated metric to // ANN index if (metric == cuvs::distance::DistanceType::CosineExpanded || metric == cuvs::distance::DistanceType::CorrelationExpanded) { metric = index->metric = cuvs::distance::DistanceType::InnerProduct; } } if constexpr (std::is_same_v<T, float>) { index->metric_processor->preprocess(index_array); } if (ivf_ft_pams) { auto new_params = from_legacy_index_params(*ivf_ft_pams, metric, metricArg); index->ivf_flat<T, int64_t>() = std::make_unique<const ivf_flat::index<T, int64_t>>( ivf_flat::build(handle, new_params, index_array, int64_t(n), D)); } else if (ivf_pq_pams) { neighbors::ivf_pq::index_params params; params.metric = metric; params.metric_arg = metricArg; params.n_lists = ivf_pq_pams->nlist; params.pq_bits = ivf_pq_pams->n_bits; params.pq_dim = ivf_pq_pams->M; // TODO: handle ivf_pq_pams.usePrecomputedTables ? auto index_view = raft::make_device_matrix_view<const T, int64_t>(index_array, n, D); index->ivf_pq = std::make_unique<const neighbors::ivf_pq::index<int64_t>>( neighbors::ivf_pq::build(handle, params, index_view)); } else { RAFT_FAIL("Unrecognized index type."); } if constexpr (std::is_same_v<T, float>) { index->metric_processor->revert(index_array); } } template <typename T = float, typename IntType = int> void approx_knn_search(raft::resources const& handle, float* distances, int64_t* indices, knnIndex* index, IntType k, T* query_array, IntType n) { if constexpr (std::is_same_v<T, float>) { index->metric_processor->preprocess(query_array); index->metric_processor->set_num_queries(k); } // search if (index->ivf_flat<T, int64_t>()) { ivf_flat::search_params params; params.n_probes = index->nprobe; ivf_flat::search( handle, params, *(index->ivf_flat<T, int64_t>()), query_array, n, k, indices, distances); } else if (index->ivf_pq) { neighbors::ivf_pq::search_params params; params.n_probes = index->nprobe; auto query_view = raft::make_device_matrix_view<const T, uint32_t>(query_array, n, index->ivf_pq->dim()); auto indices_view = raft::make_device_matrix_view<int64_t, uint32_t>(indices, n, k); auto distances_view = raft::make_device_matrix_view<float, uint32_t>(distances, n, k); neighbors::ivf_pq::search( handle, params, *index->ivf_pq, query_view, indices_view, distances_view); } else { RAFT_FAIL("The model is not trained"); } // revert changes to the query if constexpr (std::is_same_v<T, float>) { index->metric_processor->revert(query_array); } // perform post-processing to show the real distances if (index->metric == cuvs::distance::DistanceType::L2SqrtExpanded || index->metric == cuvs::distance::DistanceType::L2SqrtUnexpanded || index->metric == cuvs::distance::DistanceType::LpUnexpanded) { /** * post-processing */ float p = 0.5; // standard l2 if (index->metric == cuvs::distance::DistanceType::LpUnexpanded) p = 1.0 / index->metricArg; raft::linalg::unaryOp<float>( distances, distances, n * k, raft::pow_const_op<float>(p), resource::get_cuda_stream(handle)); } if constexpr (std::is_same_v<T, float>) { index->metric_processor->postprocess(distances); } } } // namespace cuvs::spatial::knn::detail
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail/processing.cuh
/* * Copyright (c) 2020-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "processing.hpp" #include <cuvs/distance/distance_types.hpp> #include <raft/core/operators.hpp> #include <raft/linalg/matrix_vector_op.cuh> #include <raft/linalg/norm.cuh> #include <raft/linalg/unary_op.cuh> #include <raft/stats/mean.cuh> #include <raft/stats/mean_center.cuh> #include <rmm/device_uvector.hpp> namespace cuvs { namespace spatial { namespace knn { template <typename math_t> class CosineMetricProcessor : public MetricProcessor<math_t> { protected: int k_; bool row_major_; size_t n_rows_; size_t n_cols_; cudaStream_t stream_; rmm::device_uvector<math_t> colsums_; public: CosineMetricProcessor(size_t n_rows, size_t n_cols, int k, bool row_major, cudaStream_t stream) : stream_(stream), colsums_(n_rows, stream), n_cols_(n_cols), n_rows_(n_rows), row_major_(row_major), k_(k) { } void preprocess(math_t* data) { raft::linalg::rowNorm(colsums_.data(), data, n_cols_, n_rows_, raft::linalg::NormType::L2Norm, row_major_, stream_, raft::sqrt_op{}); raft::linalg::matrixVectorOp( data, data, colsums_.data(), n_cols_, n_rows_, row_major_, false, raft::div_op{}, stream_); } void revert(math_t* data) { raft::linalg::matrixVectorOp( data, data, colsums_.data(), n_cols_, n_rows_, row_major_, false, raft::mul_op{}, stream_); } void postprocess(math_t* data) { raft::linalg::unaryOp( data, data, k_ * n_rows_, [] __device__(math_t in) { return 1 - in; }, stream_); } void set_num_queries(int k) override { k_ = k; } ~CosineMetricProcessor() = default; }; template <typename math_t> class CorrelationMetricProcessor : public CosineMetricProcessor<math_t> { using cosine = CosineMetricProcessor<math_t>; public: CorrelationMetricProcessor( size_t n_rows, size_t n_cols, int k, bool row_major, cudaStream_t stream) : CosineMetricProcessor<math_t>(n_rows, n_cols, k, row_major, stream), means_(n_rows, stream) { } void preprocess(math_t* data) { math_t normalizer_const = 1.0 / (math_t)cosine::n_cols_; raft::linalg::reduce(means_.data(), data, cosine::n_cols_, cosine::n_rows_, (math_t)0.0, cosine::row_major_, true, cosine::stream_); raft::linalg::unaryOp(means_.data(), means_.data(), cosine::n_rows_, raft::mul_const_op<math_t>(normalizer_const), cosine::stream_); raft::stats::meanCenter(data, data, means_.data(), cosine::n_cols_, cosine::n_rows_, cosine::row_major_, false, cosine::stream_); CosineMetricProcessor<math_t>::preprocess(data); } void revert(math_t* data) { CosineMetricProcessor<math_t>::revert(data); raft::stats::meanAdd(data, data, means_.data(), cosine::n_cols_, cosine::n_rows_, cosine::row_major_, false, cosine::stream_); } void postprocess(math_t* data) { CosineMetricProcessor<math_t>::postprocess(data); } ~CorrelationMetricProcessor() = default; rmm::device_uvector<math_t> means_; }; template <typename math_t> class DefaultMetricProcessor : public MetricProcessor<math_t> { public: void preprocess(math_t* data) {} void revert(math_t* data) {} void postprocess(math_t* data) {} ~DefaultMetricProcessor() = default; }; template <typename math_t> inline std::unique_ptr<MetricProcessor<math_t>> create_processor( distance::DistanceType metric, int n, int D, int k, bool rowMajorQuery, cudaStream_t userStream) { MetricProcessor<math_t>* mp = nullptr; switch (metric) { case distance::DistanceType::CosineExpanded: mp = new CosineMetricProcessor<math_t>(n, D, k, rowMajorQuery, userStream); break; case distance::DistanceType::CorrelationExpanded: mp = new CorrelationMetricProcessor<math_t>(n, D, k, rowMajorQuery, userStream); break; default: mp = new DefaultMetricProcessor<math_t>(); } return std::unique_ptr<MetricProcessor<math_t>>(mp); } // Currently only being used by floats template class MetricProcessor<float>; template class CosineMetricProcessor<float>; template class CorrelationMetricProcessor<float>; template class DefaultMetricProcessor<float>; } // namespace knn } // namespace spatial } // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail/ball_cover/registers-ext.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "../../ball_cover_types.hpp" // BallCoverIndex #include "registers_types.cuh" // DistFunc #include <cstdint> // uint32_t #include <raft/util/raft_explicit.hpp> //RAFT_EXPLICIT #if defined(RAFT_EXPLICIT_INSTANTIATE_ONLY) namespace cuvs::spatial::knn::detail { template <typename value_idx, typename value_t, typename value_int = std::uint32_t, int dims = 2, typename dist_func> void rbc_low_dim_pass_one(raft::resources const& handle, const BallCoverIndex<value_idx, value_t, value_int>& index, const value_t* query, const value_int n_query_rows, value_int k, const value_idx* R_knn_inds, const value_t* R_knn_dists, dist_func& dfunc, value_idx* inds, value_t* dists, float weight, value_int* dists_counter) RAFT_EXPLICIT; template <typename value_idx, typename value_t, typename value_int = std::uint32_t, int dims = 2, typename dist_func> void rbc_low_dim_pass_two(raft::resources const& handle, const BallCoverIndex<value_idx, value_t, value_int>& index, const value_t* query, const value_int n_query_rows, value_int k, const value_idx* R_knn_inds, const value_t* R_knn_dists, dist_func& dfunc, value_idx* inds, value_t* dists, float weight, value_int* post_dists_counter) RAFT_EXPLICIT; }; // namespace cuvs::spatial::knn::detail #endif // RAFT_EXPLICIT_INSTANTIATE_ONLY #define instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_one( \ Mvalue_idx, Mvalue_t, Mvalue_int, Mdims, Mdist_func) \ extern template void \ cuvs::spatial::knn::detail::rbc_low_dim_pass_one<Mvalue_idx, Mvalue_t, Mvalue_int, Mdims>( \ raft::resources const& handle, \ const BallCoverIndex<Mvalue_idx, Mvalue_t, Mvalue_int>& index, \ const Mvalue_t* query, \ const Mvalue_int n_query_rows, \ Mvalue_int k, \ const Mvalue_idx* R_knn_inds, \ const Mvalue_t* R_knn_dists, \ Mdist_func<Mvalue_t, Mvalue_int>& dfunc, \ Mvalue_idx* inds, \ Mvalue_t* dists, \ float weight, \ Mvalue_int* dists_counter) #define instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_two( \ Mvalue_idx, Mvalue_t, Mvalue_int, Mdims, Mdist_func) \ extern template void \ cuvs::spatial::knn::detail::rbc_low_dim_pass_two<Mvalue_idx, Mvalue_t, Mvalue_int, Mdims>( \ raft::resources const& handle, \ const BallCoverIndex<Mvalue_idx, Mvalue_t, Mvalue_int>& index, \ const Mvalue_t* query, \ const Mvalue_int n_query_rows, \ Mvalue_int k, \ const Mvalue_idx* R_knn_inds, \ const Mvalue_t* R_knn_dists, \ Mdist_func<Mvalue_t, Mvalue_int>& dfunc, \ Mvalue_idx* inds, \ Mvalue_t* dists, \ float weight, \ Mvalue_int* dists_counter) instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_one( std::int64_t, float, std::uint32_t, 2, cuvs::spatial::knn::detail::HaversineFunc); instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_one( std::int64_t, float, std::uint32_t, 3, cuvs::spatial::knn::detail::HaversineFunc); instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_one( std::int64_t, float, std::uint32_t, 2, cuvs::spatial::knn::detail::EuclideanFunc); instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_one( std::int64_t, float, std::uint32_t, 3, cuvs::spatial::knn::detail::EuclideanFunc); instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_one( std::int64_t, float, std::uint32_t, 2, cuvs::spatial::knn::detail::DistFunc); instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_one( std::int64_t, float, std::uint32_t, 3, cuvs::spatial::knn::detail::DistFunc); instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_two( std::int64_t, float, std::uint32_t, 2, cuvs::spatial::knn::detail::HaversineFunc); instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_two( std::int64_t, float, std::uint32_t, 3, cuvs::spatial::knn::detail::HaversineFunc); instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_two( std::int64_t, float, std::uint32_t, 2, cuvs::spatial::knn::detail::EuclideanFunc); instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_two( std::int64_t, float, std::uint32_t, 3, cuvs::spatial::knn::detail::EuclideanFunc); instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_two( std::int64_t, float, std::uint32_t, 2, cuvs::spatial::knn::detail::DistFunc); instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_two( std::int64_t, float, std::uint32_t, 3, cuvs::spatial::knn::detail::DistFunc); #undef instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_two #undef instantiate_raft_spatial_knn_detail_rbc_low_dim_pass_one
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail/ball_cover/registers.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #ifndef RAFT_EXPLICIT_INSTANTIATE_ONLY #include "registers-inl.cuh" #endif #ifdef RAFT_COMPILED #include "registers-ext.cuh" #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail/ball_cover/registers-inl.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "common.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resource/thrust_policy.hpp> #include "../../ball_cover_types.hpp" #include "../haversine_distance.cuh" #include "registers_types.cuh" // DistFunc #include <cstdint> #include <limits.h> #include <cuvs/neighbors/detail/faiss_select/key_value_block_select.cuh> #include <raft/util/cuda_utils.cuh> #include <thrust/fill.h> namespace cuvs { namespace spatial { namespace knn { namespace detail { /** * To find exact neighbors, we perform a post-processing stage * that filters out those points which might have neighbors outside * of their k closest landmarks. This is usually a very small portion * of the total points. * @tparam value_idx * @tparam value_t * @tparam value_int * @tparam tpb * @param X * @param n_cols * @param R_knn_inds * @param R_knn_dists * @param R_radius * @param landmarks * @param n_landmarks * @param bitset_size * @param k * @param output * @param weight */ template <typename value_idx, typename value_t, typename value_int = std::uint32_t, int col_q = 2, int tpb = 32, typename distance_func> RAFT_KERNEL perform_post_filter_registers(const value_t* X, value_int n_cols, const value_idx* R_knn_inds, const value_t* R_knn_dists, const value_t* R_radius, const value_t* landmarks, int n_landmarks, value_int bitset_size, value_int k, distance_func dfunc, std::uint32_t* output, float weight = 1.0) { // allocate array of size n_landmarks / 32 ints extern __shared__ std::uint32_t shared_mem[]; // Start with all bits on for (value_int i = threadIdx.x; i < bitset_size; i += tpb) { shared_mem[i] = 0xffffffff; } __syncthreads(); // TODO: Would it be faster to use L1 for this? value_t local_x_ptr[col_q]; for (value_int j = 0; j < n_cols; ++j) { local_x_ptr[j] = X[n_cols * blockIdx.x + j]; } value_t closest_R_dist = R_knn_dists[blockIdx.x * k + (k - 1)]; // zero out bits for closest k landmarks for (value_int j = threadIdx.x; j < k; j += tpb) { _zero_bit(shared_mem, (std::uint32_t)R_knn_inds[blockIdx.x * k + j]); } __syncthreads(); // Discard any landmarks where p(q, r) > p(q, r_q) + radius(r) // That is, the distance between the current point and the current // landmark is > the distance between the current point and // its closest landmark + the radius of the current landmark. for (value_int l = threadIdx.x; l < n_landmarks; l += tpb) { // compute p(q, r) value_t dist = dfunc(local_x_ptr, landmarks + (n_cols * l), n_cols); if (dist > weight * (closest_R_dist + R_radius[l]) || dist > 3 * closest_R_dist) { _zero_bit(shared_mem, l); } } __syncthreads(); /** * Output bitset */ for (value_int l = threadIdx.x; l < bitset_size; l += tpb) { output[blockIdx.x * bitset_size + l] = shared_mem[l]; } } /** * @tparam value_idx * @tparam value_t * @tparam value_int * @tparam bitset_type * @tparam warp_q number of registers to use per warp * @tparam thread_q number of registers to use within each thread * @tparam tpb number of threads per block * @param X * @param n_cols * @param bitset * @param bitset_size * @param R_knn_dists * @param R_indptr * @param R_1nn_inds * @param R_1nn_dists * @param knn_inds * @param knn_dists * @param n_landmarks * @param k * @param dist_counter */ template <typename value_idx, typename value_t, typename value_int = std::uint32_t, typename bitset_type = std::uint32_t, typename dist_func, int warp_q = 32, int thread_q = 2, int tpb = 128, int col_q = 2> RAFT_KERNEL compute_final_dists_registers(const value_t* X_index, const value_t* X, const value_int n_cols, bitset_type* bitset, value_int bitset_size, const value_t* R_closest_landmark_dists, const value_idx* R_indptr, const value_idx* R_1nn_inds, const value_t* R_1nn_dists, value_idx* knn_inds, value_t* knn_dists, value_int n_landmarks, value_int k, dist_func dfunc, value_int* dist_counter) { static constexpr int kNumWarps = tpb / raft::WarpSize; __shared__ value_t shared_memK[kNumWarps * warp_q]; __shared__ KeyValuePair<value_t, value_idx> shared_memV[kNumWarps * warp_q]; const value_t* x_ptr = X + (n_cols * blockIdx.x); value_t local_x_ptr[col_q]; for (value_int j = 0; j < n_cols; ++j) { local_x_ptr[j] = x_ptr[j]; } using namespace cuvs::neighbors::detail::faiss_select; KeyValueBlockSelect<value_t, value_idx, false, Comparator<value_t>, warp_q, thread_q, tpb> heap( std::numeric_limits<value_t>::max(), std::numeric_limits<value_t>::max(), -1, shared_memK, shared_memV, k); const value_int n_k = raft::Pow2<WarpSize>::roundDown(k); value_int i = threadIdx.x; for (; i < n_k; i += tpb) { value_idx ind = knn_inds[blockIdx.x * k + i]; heap.add(knn_dists[blockIdx.x * k + i], R_closest_landmark_dists[ind], ind); } if (i < k) { value_idx ind = knn_inds[blockIdx.x * k + i]; heap.addThreadQ(knn_dists[blockIdx.x * k + i], R_closest_landmark_dists[ind], ind); } heap.checkThreadQ(); for (value_int cur_R_ind = 0; cur_R_ind < n_landmarks; ++cur_R_ind) { // if cur R overlaps cur point's closest R, it could be a // candidate if (_get_val(bitset + (blockIdx.x * bitset_size), cur_R_ind)) { value_idx R_start_offset = R_indptr[cur_R_ind]; value_idx R_stop_offset = R_indptr[cur_R_ind + 1]; value_idx R_size = R_stop_offset - R_start_offset; // Loop through R's neighborhood in parallel // Round R_size to the nearest warp threads so they can // all be computing in parallel. const value_int limit = raft::Pow2<WarpSize>::roundDown(R_size); i = threadIdx.x; for (; i < limit; i += tpb) { value_idx cur_candidate_ind = R_1nn_inds[R_start_offset + i]; value_t cur_candidate_dist = R_1nn_dists[R_start_offset + i]; value_t z = heap.warpKTopRDist == 0.00 ? 0.0 : (abs(heap.warpKTop - heap.warpKTopRDist) * abs(heap.warpKTopRDist - cur_candidate_dist) - heap.warpKTop * cur_candidate_dist) / heap.warpKTopRDist; z = isnan(z) || isinf(z) ? 0.0 : z; // If lower bound on distance could possibly be in // the closest k neighbors, compute it and add to k-select value_t dist = std::numeric_limits<value_t>::max(); if (z <= heap.warpKTop) { const value_t* y_ptr = X_index + (n_cols * cur_candidate_ind); value_t local_y_ptr[col_q]; for (value_int j = 0; j < n_cols; ++j) { local_y_ptr[j] = y_ptr[j]; } dist = dfunc(local_x_ptr, local_y_ptr, n_cols); } heap.add(dist, cur_candidate_dist, cur_candidate_ind); } // second round guarantees to be only a single warp. if (i < R_size) { value_idx cur_candidate_ind = R_1nn_inds[R_start_offset + i]; value_t cur_candidate_dist = R_1nn_dists[R_start_offset + i]; value_t z = heap.warpKTopRDist == 0.00 ? 0.0 : (abs(heap.warpKTop - heap.warpKTopRDist) * abs(heap.warpKTopRDist - cur_candidate_dist) - heap.warpKTop * cur_candidate_dist) / heap.warpKTopRDist; z = isnan(z) || isinf(z) ? 0.0 : z; // If lower bound on distance could possibly be in // the closest k neighbors, compute it and add to k-select value_t dist = std::numeric_limits<value_t>::max(); if (z <= heap.warpKTop) { const value_t* y_ptr = X_index + (n_cols * cur_candidate_ind); value_t local_y_ptr[col_q]; for (value_int j = 0; j < n_cols; ++j) { local_y_ptr[j] = y_ptr[j]; } dist = dfunc(local_x_ptr, local_y_ptr, n_cols); } heap.addThreadQ(dist, cur_candidate_dist, cur_candidate_ind); } heap.checkThreadQ(); } } heap.reduce(); for (value_int i = threadIdx.x; i < k; i += tpb) { knn_dists[blockIdx.x * k + i] = shared_memK[i]; knn_inds[blockIdx.x * k + i] = shared_memV[i].value; } } /** * Random ball cover kernel for n_dims == 2 * @tparam value_idx * @tparam value_t * @tparam warp_q * @tparam thread_q * @tparam tpb * @tparam value_idx * @tparam value_t * @param R_knn_inds * @param R_knn_dists * @param m * @param k * @param R_indptr * @param R_1nn_cols * @param R_1nn_dists */ template <typename value_idx = std::int64_t, typename value_t, int warp_q = 32, int thread_q = 2, int tpb = 128, int col_q = 2, typename value_int = std::uint32_t, typename distance_func> RAFT_KERNEL block_rbc_kernel_registers(const value_t* X_index, const value_t* X, value_int n_cols, // n_cols should be 2 or 3 dims const value_idx* R_knn_inds, const value_t* R_knn_dists, value_int m, value_int k, const value_idx* R_indptr, const value_idx* R_1nn_cols, const value_t* R_1nn_dists, value_idx* out_inds, value_t* out_dists, value_int* dist_counter, const value_t* R_radius, distance_func dfunc, float weight = 1.0) { static constexpr value_int kNumWarps = tpb / raft::WarpSize; __shared__ value_t shared_memK[kNumWarps * warp_q]; __shared__ KeyValuePair<value_t, value_idx> shared_memV[kNumWarps * warp_q]; // TODO: Separate kernels for different widths: // 1. Very small (between 3 and 32) just use registers for columns of "blockIdx.x" // 2. Can fit comfortably in shared memory (32 to a few thousand?) // 3. Load each time individually. const value_t* x_ptr = X + (n_cols * blockIdx.x); // Use registers only for 2d or 3d value_t local_x_ptr[col_q]; for (value_int i = 0; i < n_cols; ++i) { local_x_ptr[i] = x_ptr[i]; } // Each warp works on 1 R using namespace cuvs::neighbors::detail::faiss_select; KeyValueBlockSelect<value_t, value_idx, false, Comparator<value_t>, warp_q, thread_q, tpb> heap( std::numeric_limits<value_t>::max(), std::numeric_limits<value_t>::max(), -1, shared_memK, shared_memV, k); value_t min_R_dist = R_knn_dists[blockIdx.x * k + (k - 1)]; value_int n_dists_computed = 0; /** * First add distances for k closest neighbors of R * to the heap */ // Start iterating through elements of each set from closest R elements, // determining if the distance could even potentially be in the heap. for (value_int cur_k = 0; cur_k < k; ++cur_k) { // index and distance to current blockIdx.x's closest landmark value_t cur_R_dist = R_knn_dists[blockIdx.x * k + cur_k]; value_idx cur_R_ind = R_knn_inds[blockIdx.x * k + cur_k]; // Equation (2) in Cayton's paper- prune out R's which are > 3 * p(q, r_q) if (cur_R_dist > weight * (min_R_dist + R_radius[cur_R_ind])) continue; if (cur_R_dist > 3 * min_R_dist) return; // The whole warp should iterate through the elements in the current R value_idx R_start_offset = R_indptr[cur_R_ind]; value_idx R_stop_offset = R_indptr[cur_R_ind + 1]; value_idx R_size = R_stop_offset - R_start_offset; value_int limit = raft::Pow2<WarpSize>::roundDown(R_size); value_int i = threadIdx.x; for (; i < limit; i += tpb) { // Index and distance of current candidate's nearest landmark value_idx cur_candidate_ind = R_1nn_cols[R_start_offset + i]; value_t cur_candidate_dist = R_1nn_dists[R_start_offset + i]; // Take 2 landmarks l_1 and l_2 where l_1 is the furthest point in the heap // and l_2 is the current landmark R. s is the current data point and // t is the new candidate data point. We know that: // d(s, t) cannot possibly be any smaller than | d(s, l_1) - d(l_1, l_2) | * | d(l_1, l_2) - // d(l_2, t) | - d(s, l_1) * d(l_2, t) // Therefore, if d(s, t) >= d(s, l_1) from the computation above, we know that the distance to // the candidate point cannot possibly be in the nearest neighbors. However, if d(s, t) < d(s, // l_1) then we should compute the distance because it's possible it could be smaller. // value_t z = heap.warpKTopRDist == 0.00 ? 0.0 : (abs(heap.warpKTop - heap.warpKTopRDist) * abs(heap.warpKTopRDist - cur_candidate_dist) - heap.warpKTop * cur_candidate_dist) / heap.warpKTopRDist; z = isnan(z) || isinf(z) ? 0.0 : z; value_t dist = std::numeric_limits<value_t>::max(); if (z <= heap.warpKTop) { const value_t* y_ptr = X_index + (n_cols * cur_candidate_ind); value_t local_y_ptr[col_q]; for (value_int j = 0; j < n_cols; ++j) { local_y_ptr[j] = y_ptr[j]; } dist = dfunc(local_x_ptr, local_y_ptr, n_cols); ++n_dists_computed; } heap.add(dist, cur_candidate_dist, cur_candidate_ind); } if (i < R_size) { value_idx cur_candidate_ind = R_1nn_cols[R_start_offset + i]; value_t cur_candidate_dist = R_1nn_dists[R_start_offset + i]; value_t z = heap.warpKTopRDist == 0.0 ? 0.0 : (abs(heap.warpKTop - heap.warpKTopRDist) * abs(heap.warpKTopRDist - cur_candidate_dist) - heap.warpKTop * cur_candidate_dist) / heap.warpKTopRDist; z = isnan(z) || isinf(z) ? 0.0 : z; value_t dist = std::numeric_limits<value_t>::max(); if (z <= heap.warpKTop) { const value_t* y_ptr = X_index + (n_cols * cur_candidate_ind); value_t local_y_ptr[col_q]; for (value_int j = 0; j < n_cols; ++j) { local_y_ptr[j] = y_ptr[j]; } dist = dfunc(local_x_ptr, local_y_ptr, n_cols); ++n_dists_computed; } heap.addThreadQ(dist, cur_candidate_dist, cur_candidate_ind); } heap.checkThreadQ(); } heap.reduce(); for (int i = threadIdx.x; i < k; i += tpb) { out_dists[blockIdx.x * k + i] = shared_memK[i]; out_inds[blockIdx.x * k + i] = shared_memV[i].value; } } template <typename value_idx, typename value_t, typename value_int = std::uint32_t, int dims = 2, typename dist_func> void rbc_low_dim_pass_one(raft::resources const& handle, const BallCoverIndex<value_idx, value_t, value_int>& index, const value_t* query, const value_int n_query_rows, value_int k, const value_idx* R_knn_inds, const value_t* R_knn_dists, dist_func& dfunc, value_idx* inds, value_t* dists, float weight, value_int* dists_counter) { if (k <= 32) block_rbc_kernel_registers<value_idx, value_t, 32, 2, 128, dims, value_int> <<<n_query_rows, 128, 0, resource::get_cuda_stream(handle)>>>( index.get_X().data_handle(), query, index.n, R_knn_inds, R_knn_dists, index.m, k, index.get_R_indptr().data_handle(), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_dists().data_handle(), inds, dists, dists_counter, index.get_R_radius().data_handle(), dfunc, weight); else if (k <= 64) block_rbc_kernel_registers<value_idx, value_t, 64, 3, 128, 2, value_int> <<<n_query_rows, 128, 0, resource::get_cuda_stream(handle)>>>( index.get_X().data_handle(), query, index.n, R_knn_inds, R_knn_dists, index.m, k, index.get_R_indptr().data_handle(), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_dists().data_handle(), inds, dists, dists_counter, index.get_R_radius().data_handle(), dfunc, weight); else if (k <= 128) block_rbc_kernel_registers<value_idx, value_t, 128, 3, 128, dims, value_int> <<<n_query_rows, 128, 0, resource::get_cuda_stream(handle)>>>( index.get_X().data_handle(), query, index.n, R_knn_inds, R_knn_dists, index.m, k, index.get_R_indptr().data_handle(), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_dists().data_handle(), inds, dists, dists_counter, index.get_R_radius().data_handle(), dfunc, weight); else if (k <= 256) block_rbc_kernel_registers<value_idx, value_t, 256, 4, 128, dims, value_int> <<<n_query_rows, 128, 0, resource::get_cuda_stream(handle)>>>( index.get_X().data_handle(), query, index.n, R_knn_inds, R_knn_dists, index.m, k, index.get_R_indptr().data_handle(), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_dists().data_handle(), inds, dists, dists_counter, index.get_R_radius().data_handle(), dfunc, weight); else if (k <= 512) block_rbc_kernel_registers<value_idx, value_t, 512, 8, 64, dims, value_int> <<<n_query_rows, 64, 0, resource::get_cuda_stream(handle)>>>( index.get_X().data_handle(), query, index.n, R_knn_inds, R_knn_dists, index.m, k, index.get_R_indptr().data_handle(), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_dists().data_handle(), inds, dists, dists_counter, index.get_R_radius().data_handle(), dfunc, weight); else if (k <= 1024) block_rbc_kernel_registers<value_idx, value_t, 1024, 8, 64, dims, value_int> <<<n_query_rows, 64, 0, resource::get_cuda_stream(handle)>>>( index.get_X().data_handle(), query, index.n, R_knn_inds, R_knn_dists, index.m, k, index.get_R_indptr().data_handle(), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_dists().data_handle(), inds, dists, dists_counter, index.get_R_radius().data_handle(), dfunc, weight); } template <typename value_idx, typename value_t, typename value_int = std::uint32_t, int dims = 2, typename dist_func> void rbc_low_dim_pass_two(raft::resources const& handle, const BallCoverIndex<value_idx, value_t, value_int>& index, const value_t* query, const value_int n_query_rows, value_int k, const value_idx* R_knn_inds, const value_t* R_knn_dists, dist_func& dfunc, value_idx* inds, value_t* dists, float weight, value_int* post_dists_counter) { const value_int bitset_size = ceil(index.n_landmarks / 32.0); rmm::device_uvector<std::uint32_t> bitset(bitset_size * n_query_rows, resource::get_cuda_stream(handle)); thrust::fill( resource::get_thrust_policy(handle), bitset.data(), bitset.data() + bitset.size(), 0); perform_post_filter_registers<value_idx, value_t, value_int, dims, 128> <<<n_query_rows, 128, bitset_size * sizeof(std::uint32_t), resource::get_cuda_stream(handle)>>>( query, index.n, R_knn_inds, R_knn_dists, index.get_R_radius().data_handle(), index.get_R().data_handle(), index.n_landmarks, bitset_size, k, dfunc, bitset.data(), weight); if (k <= 32) compute_final_dists_registers<value_idx, value_t, value_int, std::uint32_t, dist_func, 32, 2, 128, dims> <<<n_query_rows, 128, 0, resource::get_cuda_stream(handle)>>>( index.get_X().data_handle(), query, index.n, bitset.data(), bitset_size, index.get_R_closest_landmark_dists().data_handle(), index.get_R_indptr().data_handle(), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_dists().data_handle(), inds, dists, index.n_landmarks, k, dfunc, post_dists_counter); else if (k <= 64) compute_final_dists_registers<value_idx, value_t, value_int, std::uint32_t, dist_func, 64, 3, 128, dims> <<<n_query_rows, 128, 0, resource::get_cuda_stream(handle)>>>( index.get_X().data_handle(), query, index.n, bitset.data(), bitset_size, index.get_R_closest_landmark_dists().data_handle(), index.get_R_indptr().data_handle(), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_dists().data_handle(), inds, dists, index.n_landmarks, k, dfunc, post_dists_counter); else if (k <= 128) compute_final_dists_registers<value_idx, value_t, value_int, std::uint32_t, dist_func, 128, 3, 128, dims> <<<n_query_rows, 128, 0, resource::get_cuda_stream(handle)>>>( index.get_X().data_handle(), query, index.n, bitset.data(), bitset_size, index.get_R_closest_landmark_dists().data_handle(), index.get_R_indptr().data_handle(), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_dists().data_handle(), inds, dists, index.n_landmarks, k, dfunc, post_dists_counter); else if (k <= 256) compute_final_dists_registers<value_idx, value_t, value_int, std::uint32_t, dist_func, 256, 4, 128, dims> <<<n_query_rows, 128, 0, resource::get_cuda_stream(handle)>>>( index.get_X().data_handle(), query, index.n, bitset.data(), bitset_size, index.get_R_closest_landmark_dists().data_handle(), index.get_R_indptr().data_handle(), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_dists().data_handle(), inds, dists, index.n_landmarks, k, dfunc, post_dists_counter); else if (k <= 512) compute_final_dists_registers<value_idx, value_t, value_int, std::uint32_t, dist_func, 512, 8, 64, dims><<<n_query_rows, 64, 0, resource::get_cuda_stream(handle)>>>( index.get_X().data_handle(), query, index.n, bitset.data(), bitset_size, index.get_R_closest_landmark_dists().data_handle(), index.get_R_indptr().data_handle(), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_dists().data_handle(), inds, dists, index.n_landmarks, k, dfunc, post_dists_counter); else if (k <= 1024) compute_final_dists_registers<value_idx, value_t, value_int, std::uint32_t, dist_func, 1024, 8, 64, dims><<<n_query_rows, 64, 0, resource::get_cuda_stream(handle)>>>( index.get_X().data_handle(), query, index.n, bitset.data(), bitset_size, index.get_R_closest_landmark_dists().data_handle(), index.get_R_indptr().data_handle(), index.get_R_1nn_cols().data_handle(), index.get_R_1nn_dists().data_handle(), inds, dists, index.n_landmarks, k, dfunc, post_dists_counter); } }; // namespace detail }; // namespace knn }; // namespace spatial }; // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail/ball_cover/common.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "../haversine_distance.cuh" #include "registers_types.cuh" #include <cstdint> #include <thrust/functional.h> #include <thrust/tuple.h> namespace cuvs { namespace spatial { namespace knn { namespace detail { struct NNComp { template <typename one, typename two> __host__ __device__ bool operator()(const one& t1, const two& t2) { // sort first by each sample's reference landmark, if (thrust::get<0>(t1) < thrust::get<0>(t2)) return true; if (thrust::get<0>(t1) > thrust::get<0>(t2)) return false; // then by closest neighbor, return thrust::get<1>(t1) < thrust::get<1>(t2); } }; /** * Zeros the bit at location h in a one-hot encoded 32-bit int array */ __device__ inline void _zero_bit(std::uint32_t* arr, std::uint32_t h) { int bit = h % 32; int idx = h / 32; std::uint32_t assumed; std::uint32_t old = arr[idx]; do { assumed = old; old = atomicCAS(arr + idx, assumed, assumed & ~(1 << bit)); } while (assumed != old); } /** * Returns whether or not bit at location h is nonzero in a one-hot * encoded 32-bit in array. */ __device__ inline bool _get_val(std::uint32_t* arr, std::uint32_t h) { int bit = h % 32; int idx = h / 32; return (arr[idx] & (1 << bit)) > 0; } }; // namespace detail }; // namespace knn }; // namespace spatial }; // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/detail/ball_cover/registers_types.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "../haversine_distance.cuh" // compute_haversine #include <cstdint> // uint32_t namespace cuvs { namespace spatial { namespace knn { namespace detail { template <typename value_t, typename value_int = std::uint32_t> struct DistFunc { virtual __device__ __host__ __forceinline__ value_t operator()(const value_t* a, const value_t* b, const value_int n_dims) { return -1; }; }; template <typename value_t, typename value_int = std::uint32_t> struct HaversineFunc : public DistFunc<value_t, value_int> { __device__ __host__ __forceinline__ value_t operator()(const value_t* a, const value_t* b, const value_int n_dims) override { return cuvs::spatial::knn::detail::compute_haversine(a[0], b[0], a[1], b[1]); } }; template <typename value_t, typename value_int = std::uint32_t> struct EuclideanFunc : public DistFunc<value_t, value_int> { __device__ __host__ __forceinline__ value_t operator()(const value_t* a, const value_t* b, const value_int n_dims) override { value_t sum_sq = 0; for (value_int i = 0; i < n_dims; ++i) { value_t diff = a[i] - b[i]; sum_sq += diff * diff; } return raft::sqrt(sum_sq); } }; }; // namespace detail }; // namespace knn }; // namespace spatial }; // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn
rapidsai_public_repos/cuvs/cpp/include/cuvs/spatial/knn/specializations/knn.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #pragma message( \ __FILE__ \ " is deprecated and will be removed." \ " Including specializations is not necessary any more." \ " For more information, see: https://docs.rapids.ai/api/raft/nightly/using_libraft.html")
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/cluster_solvers.cuh
/* * Copyright (c) 2019-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #ifndef __CLUSTER_SOLVERS_H #define __CLUSTER_SOLVERS_H #pragma once #include <cuvs/cluster/kmeans.cuh> #include <raft/core/resource/thrust_policy.hpp> #include <utility> // for std::pair namespace cuvs { namespace spectral { using namespace matrix; // aggregate of control params for Eigen Solver: // template <typename index_type_t, typename value_type_t, typename size_type_t = index_type_t> struct cluster_solver_config_t { size_type_t n_clusters; size_type_t maxIter; value_type_t tol; unsigned long long seed{123456}; }; template <typename index_type_t, typename value_type_t, typename size_type_t = index_type_t> struct kmeans_solver_t { explicit kmeans_solver_t( cluster_solver_config_t<index_type_t, value_type_t, size_type_t> const& config) : config_(config) { } std::pair<value_type_t, index_type_t> solve(raft::resources const& handle, size_type_t n_obs_vecs, size_type_t dim, value_type_t const* __restrict__ obs, index_type_t* __restrict__ codes) const { RAFT_EXPECTS(obs != nullptr, "Null obs buffer."); RAFT_EXPECTS(codes != nullptr, "Null codes buffer."); value_type_t residual{}; index_type_t iters{}; cuvs::cluster::KMeansParams km_params; km_params.n_clusters = config_.n_clusters; km_params.tol = config_.tol; km_params.max_iter = config_.maxIter; km_params.rng_state.seed = config_.seed; auto X = raft::make_device_matrix_view<const value_type_t>(obs, n_obs_vecs, dim); auto labels = raft::make_device_vector_view<index_type_t>(codes, n_obs_vecs); auto centroids = raft::make_device_matrix<value_type_t, index_type_t>(handle, config_.n_clusters, dim); auto weight = raft::make_device_vector<value_type_t, index_type_t>(handle, n_obs_vecs); thrust::fill(raft::resource::get_thrust_policy(handle), weight.data_handle(), weight.data_handle() + n_obs_vecs, 1); auto sw = std::make_optional((raft::device_vector_view<const value_type_t>)weight.view()); cuvs::cluster::kmeans_fit_predict<value_type_t, index_type_t>( handle, km_params, X, sw, centroids.view(), labels, raft::make_host_scalar_view(&residual), raft::make_host_scalar_view(&iters)); return std::make_pair(residual, iters); } auto const& get_config(void) const { return config_; } private: cluster_solver_config_t<index_type_t, value_type_t, size_type_t> config_; }; } // namespace spectral } // namespace cuvs #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/eigen_solvers.cuh
/* * Copyright (c) 2019-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #ifndef __EIGEN_SOLVERS_H #define __EIGEN_SOLVERS_H #pragma once #include <raft/sparse/solver/lanczos.cuh> #include <raft/spectral/matrix_wrappers.hpp> namespace cuvs { namespace spectral { // aggregate of control params for Eigen Solver: // template <typename index_type_t, typename value_type_t, typename size_type_t = index_type_t> struct eigen_solver_config_t { size_type_t n_eigVecs; size_type_t maxIter; size_type_t restartIter; value_type_t tol; bool reorthogonalize{false}; unsigned long long seed{ 1234567}; // CAVEAT: this default value is now common to all instances of using seed in // Lanczos; was not the case before: there were places where a default seed = 123456 // was used; this may trigger slightly different # solver iterations }; template <typename index_type_t, typename value_type_t, typename size_type_t = index_type_t> struct lanczos_solver_t { explicit lanczos_solver_t( eigen_solver_config_t<index_type_t, value_type_t, size_type_t> const& config) : config_(config) { } index_type_t solve_smallest_eigenvectors( raft::resources const& handle, raft::matrix::sparse_matrix_t<index_type_t, value_type_t> const& A, value_type_t* __restrict__ eigVals, value_type_t* __restrict__ eigVecs) const { RAFT_EXPECTS(eigVals != nullptr, "Null eigVals buffer."); RAFT_EXPECTS(eigVecs != nullptr, "Null eigVecs buffer."); index_type_t iters{}; sparse::solver::computeSmallestEigenvectors(handle, A, config_.n_eigVecs, config_.maxIter, config_.restartIter, config_.tol, config_.reorthogonalize, iters, eigVals, eigVecs, config_.seed); return iters; } index_type_t solve_largest_eigenvectors( raft::resources const& handle, raft::matrix::sparse_matrix_t<index_type_t, value_type_t> const& A, value_type_t* __restrict__ eigVals, value_type_t* __restrict__ eigVecs) const { RAFT_EXPECTS(eigVals != nullptr, "Null eigVals buffer."); RAFT_EXPECTS(eigVecs != nullptr, "Null eigVecs buffer."); index_type_t iters{}; sparse::solver::computeLargestEigenvectors(handle, A, config_.n_eigVecs, config_.maxIter, config_.restartIter, config_.tol, config_.reorthogonalize, iters, eigVals, eigVecs, config_.seed); return iters; } auto const& get_config(void) const { return config_; } private: eigen_solver_config_t<index_type_t, value_type_t, size_type_t> config_; }; } // namespace spectral } // namespace cuvs #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/matrix_wrappers.hpp
/* * Copyright (c) 2020-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/spectral/detail/matrix_wrappers.hpp> // ========================================================= // Useful macros // ========================================================= namespace cuvs { namespace spectral { namespace matrix { using size_type = int; // for now; TODO: move it in appropriate header // specifies type of algorithm used // for SpMv: // using detail::sparse_mv_alg_t; // Vector "view"-like aggregate for linear algebra purposes // using detail::vector_view_t; using detail::vector_t; using detail::sparse_matrix_t; using detail::laplacian_matrix_t; using detail::modularity_matrix_t; } // namespace matrix } // namespace spectral } // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/specializations.cuh
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #pragma message( \ __FILE__ \ " is deprecated and will be removed." \ " Including specializations is not necessary any more." \ " For more information, see: https://docs.rapids.ai/api/raft/nightly/using_libraft.html")
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/modularity_maximization.cuh
/* * Copyright (c) 2020-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #ifndef __MODULARITY_MAXIMIZATION_H #define __MODULARITY_MAXIMIZATION_H #pragma once #include <tuple> #include <raft/spectral/detail/modularity_maximization.hpp> namespace cuvs { namespace spectral { // ========================================================= // Spectral modularity_maximization // ========================================================= /** Compute partition for a weighted undirected graph. This * partition attempts to minimize the cost function: * Cost = \f$sum_i\f$ (Edges cut by ith partition)/(Vertices in ith partition) * * @param handle raft handle for managing expensive resources * @param csr_m Weighted graph in CSR format * @param eigen_solver Eigensolver implementation * @param cluster_solver Cluster solver implementation * @param clusters (Output, device memory, n entries) Partition * assignments. * @param eigVals Output eigenvalue array pointer on device * @param eigVecs Output eigenvector array pointer on device * @return statistics: number of eigensolver iterations, . */ template <typename vertex_t, typename weight_t, typename EigenSolver, typename ClusterSolver> std::tuple<vertex_t, weight_t, vertex_t> modularity_maximization( raft::resources const& handle, raft::matrix::sparse_matrix_t<vertex_t, weight_t> const& csr_m, EigenSolver const& eigen_solver, ClusterSolver const& cluster_solver, vertex_t* __restrict__ clusters, weight_t* eigVals, weight_t* eigVecs) { return raft::spectral::detail:: modularity_maximization<vertex_t, weight_t, EigenSolver, ClusterSolver>( handle, csr_m, eigen_solver, cluster_solver, clusters, eigVals, eigVecs); } //=================================================== // Analysis of graph partition // ========================================================= /// Compute modularity /** This function determines the modularity based on a graph and cluster assignments * @param handle raft handle for managing expensive resources * @param csr_m Weighted graph in CSR format * @param nClusters Number of clusters. * @param clusters (Input, device memory, n entries) Cluster assignments. * @param modularity On exit, modularity */ template <typename vertex_t, typename weight_t> void analyzeModularity(raft::resources const& handle, raft::matrix::sparse_matrix_t<vertex_t, weight_t> const& csr_m, vertex_t nClusters, vertex_t const* __restrict__ clusters, weight_t& modularity) { raft::spectral::detail::analyzeModularity<vertex_t, weight_t>( handle, csr_m, nClusters, clusters, modularity); } } // namespace spectral } // namespace cuvs #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/cluster_solvers_deprecated.cuh
/* * Copyright (c) 2019-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ /** * Note: This file is deprecated and will be removed in a future release * Please use include/cuvs/cluster/kmeans.cuh instead */ #ifndef __CLUSTER_SOLVERS_deprecated_H #define __CLUSTER_SOLVERS_deprecated_H #pragma once #include <cuvs/cluster/kmeans_deprecated.cuh> #include <utility> // for std::pair namespace cuvs { namespace spectral { using namespace matrix; // aggregate of control params for Eigen Solver: // template <typename index_type_t, typename value_type_t, typename size_type_t = index_type_t> struct cluster_solver_config_deprecated_t { size_type_t n_clusters; size_type_t maxIter; value_type_t tol; unsigned long long seed{123456}; }; template <typename index_type_t, typename value_type_t, typename size_type_t = index_type_t> struct kmeans_solver_deprecated_t { explicit kmeans_solver_deprecated_t( cluster_solver_config_deprecated_t<index_type_t, value_type_t, size_type_t> const& config) : config_(config) { } std::pair<value_type_t, index_type_t> solve(raft::resources const& handle, size_type_t n_obs_vecs, size_type_t dim, value_type_t const* __restrict__ obs, index_type_t* __restrict__ codes) const { RAFT_EXPECTS(obs != nullptr, "Null obs buffer."); RAFT_EXPECTS(codes != nullptr, "Null codes buffer."); value_type_t residual{}; index_type_t iters{}; cuvs::cluster::kmeans(handle, n_obs_vecs, dim, config_.n_clusters, config_.tol, config_.maxIter, obs, codes, residual, iters, config_.seed); return std::make_pair(residual, iters); } auto const& get_config(void) const { return config_; } private: cluster_solver_config_deprecated_t<index_type_t, value_type_t, size_type_t> config_; }; } // namespace spectral } // namespace cuvs #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/partition.cuh
/* * Copyright (c) 2019-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #ifndef __PARTITION_H #define __PARTITION_H #pragma once #include <tuple> #include <raft/spectral/detail/partition.hpp> namespace cuvs { namespace spectral { // ========================================================= // Spectral partitioner // ========================================================= /// Compute spectral graph partition /** Compute partition for a weighted undirected graph. This * partition attempts to minimize the cost function: * Cost = \f$sum_i\f$ (Edges cut by ith partition)/(Vertices in ith partition) * * @param handle raft handle for managing expensive resources * @param csr_m Weighted graph in CSR format * @param eigen_solver Eigensolver implementation * @param cluster_solver Cluster solver implementation * @param clusters (Output, device memory, n entries) Partition * assignments. * @param eigVals Output eigenvalue array pointer on device * @param eigVecs Output eigenvector array pointer on device * @return statistics: number of eigensolver iterations, . */ template <typename vertex_t, typename weight_t, typename EigenSolver, typename ClusterSolver> std::tuple<vertex_t, weight_t, vertex_t> partition( raft::resources const& handle, raft::matrix::sparse_matrix_t<vertex_t, weight_t> const& csr_m, EigenSolver const& eigen_solver, ClusterSolver const& cluster_solver, vertex_t* __restrict__ clusters, weight_t* eigVals, weight_t* eigVecs) { return raft::spectral::detail::partition<vertex_t, weight_t, EigenSolver, ClusterSolver>( handle, csr_m, eigen_solver, cluster_solver, clusters, eigVals, eigVecs); } // ========================================================= // Analysis of graph partition // ========================================================= /// Compute cost function for partition /** This function determines the edges cut by a partition and a cost * function: * Cost = \f$sum_i\f$ (Edges cut by ith partition)/(Vertices in ith partition) * Graph is assumed to be weighted and undirected. * * @param handle raft handle for managing expensive resources * @param csr_m Weighted graph in CSR format * @param nClusters Number of partitions. * @param clusters (Input, device memory, n entries) Partition * assignments. * @param edgeCut On exit, weight of edges cut by partition. * @param cost On exit, partition cost function. */ template <typename vertex_t, typename weight_t> void analyzePartition(raft::resources const& handle, raft::matrix::sparse_matrix_t<vertex_t, weight_t> const& csr_m, vertex_t nClusters, const vertex_t* __restrict__ clusters, weight_t& edgeCut, weight_t& cost) { raft::spectral::detail::analyzePartition<vertex_t, weight_t>( handle, csr_m, nClusters, clusters, edgeCut, cost); } } // namespace spectral } // namespace cuvs #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/detail/warn_dbg.hpp
/* * Copyright (c) 2020-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <stdexcept> #include <string> #include <raft/core/detail/macros.hpp> #ifdef DEBUG #define COUT() (std::cout) #define CERR() (std::cerr) // nope: // #define WARNING(message) \ do { \ std::stringstream ss; \ ss << "Warning (" << __FILE__ << ":" << __LINE__ << "): " << message; \ CERR() << ss.str() << std::endl; \ } while (0) #else // DEBUG #define WARNING(message) #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/detail/spectral_util.cuh
/* * Copyright (c) 2020-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/core/resource/cublas_handle.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resource/thrust_policy.hpp> #include <raft/core/resources.hpp> #include <raft/linalg/detail/cublas_wrappers.hpp> #include <raft/spectral/matrix_wrappers.hpp> #include <raft/util/cudart_utils.hpp> #include <thrust/device_ptr.h> #include <thrust/fill.h> #include <thrust/for_each.h> #include <thrust/functional.h> #include <thrust/iterator/constant_iterator.h> #include <thrust/iterator/zip_iterator.h> #include <thrust/reduce.h> #include <thrust/transform.h> #include <thrust/tuple.h> #include <algorithm> namespace cuvs { namespace spectral { template <typename index_type_t, typename value_type_t> RAFT_KERNEL scale_obs_kernel(index_type_t m, index_type_t n, value_type_t* obs) { index_type_t i, j, k, index, mm; value_type_t alpha, v, last; bool valid; // ASSUMPTION: kernel is launched with either 2, 4, 8, 16 or 32 threads in x-dimension // compute alpha mm = (((m + blockDim.x - 1) / blockDim.x) * blockDim.x); // m in multiple of blockDim.x alpha = 0.0; for (j = threadIdx.y + blockIdx.y * blockDim.y; j < n; j += blockDim.y * gridDim.y) { for (i = threadIdx.x; i < mm; i += blockDim.x) { // check if the thread is valid valid = i < m; // get the value of the last thread last = __shfl_sync(warp_full_mask(), alpha, blockDim.x - 1, blockDim.x); // if you are valid read the value from memory, otherwise set your value to 0 alpha = (valid) ? obs[i + j * m] : 0.0; alpha = alpha * alpha; // do prefix sum (of size warpSize=blockDim.x =< 32) for (k = 1; k < blockDim.x; k *= 2) { v = __shfl_up_sync(warp_full_mask(), alpha, k, blockDim.x); if (threadIdx.x >= k) alpha += v; } // shift by last alpha += last; } } // scale by alpha alpha = __shfl_sync(warp_full_mask(), alpha, blockDim.x - 1, blockDim.x); alpha = raft::sqrt(alpha); for (j = threadIdx.y + blockIdx.y * blockDim.y; j < n; j += blockDim.y * gridDim.y) { for (i = threadIdx.x; i < m; i += blockDim.x) { // blockDim.x=32 index = i + j * m; obs[index] = obs[index] / alpha; } } } template <typename index_type_t> index_type_t next_pow2(index_type_t n) { index_type_t v; // Reference: // http://graphics.stanford.edu/~seander/bithacks.html#RoundUpPowerOf2Float v = n - 1; v |= v >> 1; v |= v >> 2; v |= v >> 4; v |= v >> 8; v |= v >> 16; return v + 1; } template <typename index_type_t, typename value_type_t> cudaError_t scale_obs(index_type_t m, index_type_t n, value_type_t* obs) { index_type_t p2m; // find next power of 2 p2m = next_pow2<index_type_t>(m); // setup launch configuration unsigned int xsize = std::max(2, std::min(p2m, 32)); dim3 nthreads{xsize, 256 / xsize, 1}; dim3 nblocks{1, (n + nthreads.y - 1) / nthreads.y, 1}; // launch scaling kernel (scale each column of obs by its norm) scale_obs_kernel<index_type_t, value_type_t><<<nblocks, nthreads>>>(m, n, obs); return cudaSuccess; } template <typename vertex_t, typename edge_t, typename weight_t> void transform_eigen_matrix(raft::resources const& handle, edge_t n, vertex_t nEigVecs, weight_t* eigVecs) { auto stream = resource::get_cuda_stream(handle); auto cublas_h = resource::get_cublas_handle(handle); auto thrust_exec_policy = resource::get_thrust_policy(handle); const weight_t zero{0.0}; const weight_t one{1.0}; // Whiten eigenvector matrix for (auto i = 0; i < nEigVecs; ++i) { weight_t mean, std; mean = thrust::reduce(thrust_exec_policy, thrust::device_pointer_cast(eigVecs + IDX(0, i, n)), thrust::device_pointer_cast(eigVecs + IDX(0, i + 1, n))); RAFT_CHECK_CUDA(stream); mean /= n; thrust::transform(thrust_exec_policy, thrust::device_pointer_cast(eigVecs + IDX(0, i, n)), thrust::device_pointer_cast(eigVecs + IDX(0, i + 1, n)), thrust::make_constant_iterator(mean), thrust::device_pointer_cast(eigVecs + IDX(0, i, n)), thrust::minus<weight_t>()); RAFT_CHECK_CUDA(stream); // TODO: Call from public API when ready RAFT_CUBLAS_TRY( raft::linalg::detail::cublasnrm2(cublas_h, n, eigVecs + IDX(0, i, n), 1, &std, stream)); std /= std::sqrt(static_cast<weight_t>(n)); thrust::transform(thrust_exec_policy, thrust::device_pointer_cast(eigVecs + IDX(0, i, n)), thrust::device_pointer_cast(eigVecs + IDX(0, i + 1, n)), thrust::make_constant_iterator(std), thrust::device_pointer_cast(eigVecs + IDX(0, i, n)), thrust::divides<weight_t>()); RAFT_CHECK_CUDA(stream); } // Transpose eigenvector matrix // TODO: in-place transpose { raft::spectral::matrix::vector_t<weight_t> work(handle, nEigVecs * n); // TODO: Call from public API when ready RAFT_CUBLAS_TRY( raft::linalg::detail::cublassetpointermode(cublas_h, CUBLAS_POINTER_MODE_HOST, stream)); // TODO: Call from public API when ready RAFT_CUBLAS_TRY(raft::linalg::detail::cublasgeam(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N, nEigVecs, n, &one, eigVecs, n, &zero, (weight_t*)NULL, nEigVecs, work.raw(), nEigVecs, stream)); RAFT_CUDA_TRY(cudaMemcpyAsync( eigVecs, work.raw(), nEigVecs * n * sizeof(weight_t), cudaMemcpyDeviceToDevice, stream)); } } namespace { /// Functor to generate indicator vectors /** For use in Thrust transform */ template <typename index_type_t, typename value_type_t> struct equal_to_i_op { const index_type_t i; public: equal_to_i_op(index_type_t _i) : i(_i) {} template <typename Tuple_> __host__ __device__ void operator()(Tuple_ t) { thrust::get<1>(t) = (thrust::get<0>(t) == i) ? (value_type_t)1.0 : (value_type_t)0.0; } }; } // namespace // Construct indicator vector for ith partition // template <typename vertex_t, typename edge_t, typename weight_t> bool construct_indicator(raft::resources const& handle, edge_t index, edge_t n, weight_t& clustersize, weight_t& partStats, vertex_t const* __restrict__ clusters, raft::spectral::matrix::vector_t<weight_t>& part_i, raft::spectral::matrix::vector_t<weight_t>& Bx, raft::spectral::matrix::laplacian_matrix_t<vertex_t, weight_t> const& B) { auto stream = resource::get_cuda_stream(handle); auto cublas_h = resource::get_cublas_handle(handle); auto thrust_exec_policy = resource::get_thrust_policy(handle); thrust::for_each( thrust_exec_policy, thrust::make_zip_iterator(thrust::make_tuple(thrust::device_pointer_cast(clusters), thrust::device_pointer_cast(part_i.raw()))), thrust::make_zip_iterator(thrust::make_tuple(thrust::device_pointer_cast(clusters + n), thrust::device_pointer_cast(part_i.raw() + n))), equal_to_i_op<vertex_t, weight_t>(index)); RAFT_CHECK_CUDA(stream); // Compute size of ith partition // TODO: Call from public API when ready RAFT_CUBLAS_TRY(raft::linalg::detail::cublasdot( cublas_h, n, part_i.raw(), 1, part_i.raw(), 1, &clustersize, stream)); clustersize = round(clustersize); if (clustersize < 0.5) { return false; } // Compute part stats B.mv(1, part_i.raw(), 0, Bx.raw()); // TODO: Call from public API when ready RAFT_CUBLAS_TRY( raft::linalg::detail::cublasdot(cublas_h, n, Bx.raw(), 1, part_i.raw(), 1, &partStats, stream)); return true; } } // namespace spectral } // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/detail/lapack.hpp
/* * Copyright (c) 2020-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cusolverDn.h> #include <raft/core/error.hpp> #include <raft/linalg/detail/cublas_wrappers.hpp> #include <raft/linalg/detail/cusolver_wrappers.hpp> // for now; TODO: check if/where this `define` should be; // #define USE_LAPACK namespace cuvs { #define lapackCheckError(status) \ { \ if (status < 0) { \ std::stringstream ss; \ ss << "Lapack error: argument number " << -status << " had an illegal value."; \ throw exception(ss.str()); \ } else if (status > 0) \ RAFT_FAIL("Lapack error: internal error."); \ } extern "C" void sgeqrf_( int* m, int* n, float* a, int* lda, float* tau, float* work, int* lwork, int* info); extern "C" void dgeqrf_( int* m, int* n, double* a, int* lda, double* tau, double* work, int* lwork, int* info); extern "C" void sormqr_(char* side, char* trans, int* m, int* n, int* k, float* a, int* lda, const float* tau, float* c, int* ldc, float* work, int* lwork, int* info); extern "C" void dormqr_(char* side, char* trans, int* m, int* n, int* k, double* a, int* lda, const double* tau, double* c, int* ldc, double* work, int* lwork, int* info); extern "C" int dgeev_(char* jobvl, char* jobvr, int* n, double* a, int* lda, double* wr, double* wi, double* vl, int* ldvl, double* vr, int* ldvr, double* work, int* lwork, int* info); extern "C" int sgeev_(char* jobvl, char* jobvr, int* n, float* a, int* lda, float* wr, float* wi, float* vl, int* ldvl, float* vr, int* ldvr, float* work, int* lwork, int* info); extern "C" cusolverStatus_t cusolverDnSgemmHost(cublasOperation_t transa, cublasOperation_t transb, int m, int n, int k, const float* alpha, const float* A, int lda, const float* B, int ldb, const float* beta, float* C, int ldc); extern "C" cusolverStatus_t cusolverDnDgemmHost(cublasOperation_t transa, cublasOperation_t transb, int m, int n, int k, const double* alpha, const double* A, int lda, const double* B, int ldb, const double* beta, double* C, int ldc); extern "C" cusolverStatus_t cusolverDnSsterfHost(int n, float* d, float* e, int* info); extern "C" cusolverStatus_t cusolverDnDsterfHost(int n, double* d, double* e, int* info); extern "C" cusolverStatus_t cusolverDnSsteqrHost( const signed char* compz, int n, float* d, float* e, float* z, int ldz, float* work, int* info); extern "C" cusolverStatus_t cusolverDnDsteqrHost(const signed char* compz, int n, double* d, double* e, double* z, int ldz, double* work, int* info); template <typename T> class Lapack { private: Lapack(); ~Lapack(); public: static void check_lapack_enabled(); static void gemm(bool transa, bool transb, int m, int n, int k, T alpha, const T* A, int lda, const T* B, int ldb, T beta, T* C, int ldc); // special QR for lanczos static void sterf(int n, T* d, T* e); static void steqr(char compz, int n, T* d, T* e, T* z, int ldz, T* work); // QR // computes the QR factorization of a general matrix static void geqrf(int m, int n, T* a, int lda, T* tau, T* work, int* lwork); // Generates the real orthogonal matrix Q of the QR factorization formed by geqrf. // multiply C by implicit Q static void ormqr(bool right_side, bool transq, int m, int n, int k, T* a, int lda, T* tau, T* c, int ldc, T* work, int* lwork); static void geev(T* A, T* eigenvalues, int dim, int lda); static void geev(T* A, T* eigenvalues, T* eigenvectors, int dim, int lda, int ldvr); static void geev(T* A, T* eigenvalues_r, T* eigenvalues_i, T* eigenvectors_r, T* eigenvectors_i, int dim, int lda, int ldvr); private: static void lapack_gemm(const char transa, const char transb, int m, int n, int k, float alpha, const float* a, int lda, const float* b, int ldb, float beta, float* c, int ldc) { cublasOperation_t cublas_transa = (transa == 'N') ? CUBLAS_OP_N : CUBLAS_OP_T; cublasOperation_t cublas_transb = (transb == 'N') ? CUBLAS_OP_N : CUBLAS_OP_T; cusolverDnSgemmHost( cublas_transa, cublas_transb, m, n, k, &alpha, (float*)a, lda, (float*)b, ldb, &beta, c, ldc); } static void lapack_gemm(const signed char transa, const signed char transb, int m, int n, int k, double alpha, const double* a, int lda, const double* b, int ldb, double beta, double* c, int ldc) { cublasOperation_t cublas_transa = (transa == 'N') ? CUBLAS_OP_N : CUBLAS_OP_T; cublasOperation_t cublas_transb = (transb == 'N') ? CUBLAS_OP_N : CUBLAS_OP_T; cusolverDnDgemmHost(cublas_transa, cublas_transb, m, n, k, &alpha, (double*)a, lda, (double*)b, ldb, &beta, c, ldc); } static void lapack_sterf(int n, float* d, float* e, int* info) { cusolverDnSsterfHost(n, d, e, info); } static void lapack_sterf(int n, double* d, double* e, int* info) { cusolverDnDsterfHost(n, d, e, info); } static void lapack_steqr( const signed char compz, int n, float* d, float* e, float* z, int ldz, float* work, int* info) { cusolverDnSsteqrHost(&compz, n, d, e, z, ldz, work, info); } static void lapack_steqr(const signed char compz, int n, double* d, double* e, double* z, int ldz, double* work, int* info) { cusolverDnDsteqrHost(&compz, n, d, e, z, ldz, work, info); } static void lapack_geqrf( int m, int n, float* a, int lda, float* tau, float* work, int* lwork, int* info) { sgeqrf_(&m, &n, a, &lda, tau, work, lwork, info); } static void lapack_geqrf( int m, int n, double* a, int lda, double* tau, double* work, int* lwork, int* info) { dgeqrf_(&m, &n, a, &lda, tau, work, lwork, info); } static void lapack_ormqr(char side, char trans, int m, int n, int k, float* a, int lda, float* tau, float* c, int ldc, float* work, int* lwork, int* info) { sormqr_(&side, &trans, &m, &n, &k, a, &lda, tau, c, &ldc, work, lwork, info); } static void lapack_ormqr(char side, char trans, int m, int n, int k, double* a, int lda, double* tau, double* c, int ldc, double* work, int* lwork, int* info) { dormqr_(&side, &trans, &m, &n, &k, a, &lda, tau, c, &ldc, work, lwork, info); } static int lapack_geev_dispatch(char* jobvl, char* jobvr, int* n, double* a, int* lda, double* wr, double* wi, double* vl, int* ldvl, double* vr, int* ldvr, double* work, int* lwork, int* info) { return dgeev_(jobvl, jobvr, n, a, lda, wr, wi, vl, ldvl, vr, ldvr, work, lwork, info); } static int lapack_geev_dispatch(char* jobvl, char* jobvr, int* n, float* a, int* lda, float* wr, float* wi, float* vl, int* ldvl, float* vr, int* ldvr, float* work, int* lwork, int* info) { return sgeev_(jobvl, jobvr, n, a, lda, wr, wi, vl, ldvl, vr, ldvr, work, lwork, info); } // real eigenvalues static void lapack_geev(T* A, T* eigenvalues, int dim, int lda) { char job = 'N'; std::vector<T> WI(dim); int ldv = 1; T* vl = 0; int work_size = 6 * dim; std::vector<T> work(work_size); int info; lapack_geev_dispatch(&job, &job, &dim, A, &lda, eigenvalues, WI.data(), vl, &ldv, vl, &ldv, work.data(), &work_size, &info); lapackCheckError(info); } // real eigenpairs static void lapack_geev(T* A, T* eigenvalues, T* eigenvectors, int dim, int lda, int ldvr) { char jobvl = 'N'; char jobvr = 'V'; std::vector<T> WI(dim); int work_size = 6 * dim; T* vl = 0; int ldvl = 1; std::vector<T> work(work_size); int info; lapack_geev_dispatch(&jobvl, &jobvr, &dim, A, &lda, eigenvalues, WI.data(), vl, &ldvl, eigenvectors, &ldvr, work.data(), &work_size, &info); lapackCheckError(info); } // complex eigenpairs static void lapack_geev(T* A, T* eigenvalues_r, T* eigenvalues_i, T* eigenvectors_r, T* eigenvectors_i, int dim, int lda, int ldvr) { char jobvl = 'N'; char jobvr = 'V'; int work_size = 8 * dim; int ldvl = 1; std::vector<T> work(work_size); int info; lapack_geev_dispatch(&jobvl, &jobvr, &dim, A, &lda, eigenvalues_r, eigenvalues_i, 0, &ldvl, eigenvectors_r, &ldvr, work.data(), &work_size, &info); lapackCheckError(info); } }; template <typename T> void Lapack<T>::check_lapack_enabled() { #ifndef USE_LAPACK RAFT_FAIL("Error: LAPACK not enabled."); #endif } template <typename T> void Lapack<T>::gemm(bool transa, bool transb, int m, int n, int k, T alpha, const T* A, int lda, const T* B, int ldb, T beta, T* C, int ldc) { // check_lapack_enabled(); // #ifdef NVGRAPH_USE_LAPACK const char transA_char = transa ? 'T' : 'N'; const char transB_char = transb ? 'T' : 'N'; lapack_gemm(transA_char, transB_char, m, n, k, alpha, A, lda, B, ldb, beta, C, ldc); // #endif } template <typename T> void Lapack<T>::sterf(int n, T* d, T* e) { // check_lapack_enabled(); // #ifdef NVGRAPH_USE_LAPACK int info; lapack_sterf(n, d, e, &info); lapackCheckError(info); // #endif } template <typename T> void Lapack<T>::steqr(char compz, int n, T* d, T* e, T* z, int ldz, T* work) { // check_lapack_enabled(); // #ifdef NVGRAPH_USE_LAPACK int info; lapack_steqr(compz, n, d, e, z, ldz, work, &info); lapackCheckError(info); // #endif } template <typename T> void Lapack<T>::geqrf(int m, int n, T* a, int lda, T* tau, T* work, int* lwork) { check_lapack_enabled(); #ifdef USE_LAPACK int info; lapack_geqrf(m, n, a, lda, tau, work, lwork, &info); lapackCheckError(info); #endif } template <typename T> void Lapack<T>::ormqr(bool right_side, bool transq, int m, int n, int k, T* a, int lda, T* tau, T* c, int ldc, T* work, int* lwork) { check_lapack_enabled(); #ifdef USE_LAPACK char side = right_side ? 'R' : 'L'; char trans = transq ? 'T' : 'N'; int info; lapack_ormqr(side, trans, m, n, k, a, lda, tau, c, ldc, work, lwork, &info); lapackCheckError(info); #endif } // real eigenvalues template <typename T> void Lapack<T>::geev(T* A, T* eigenvalues, int dim, int lda) { check_lapack_enabled(); #ifdef USE_LAPACK lapack_geev(A, eigenvalues, dim, lda); #endif } // real eigenpairs template <typename T> void Lapack<T>::geev(T* A, T* eigenvalues, T* eigenvectors, int dim, int lda, int ldvr) { check_lapack_enabled(); #ifdef USE_LAPACK lapack_geev(A, eigenvalues, eigenvectors, dim, lda, ldvr); #endif } // complex eigenpairs template <typename T> void Lapack<T>::geev(T* A, T* eigenvalues_r, T* eigenvalues_i, T* eigenvectors_r, T* eigenvectors_i, int dim, int lda, int ldvr) { check_lapack_enabled(); #ifdef USE_LAPACK lapack_geev(A, eigenvalues_r, eigenvalues_i, eigenvectors_r, eigenvectors_i, dim, lda, ldvr); #endif } } // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/detail/partition.hpp
/* * Copyright (c) 2019-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <math.h> #include <raft/core/resource/cublas_handle.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <stdio.h> #include <cuda.h> #include <thrust/fill.h> #include <thrust/reduce.h> #include <thrust/transform.h> #include <tuple> #include <raft/linalg/detail/cublas_wrappers.hpp> #include <raft/spectral/cluster_solvers.cuh> #include <raft/spectral/detail/spectral_util.cuh> #include <raft/spectral/eigen_solvers.cuh> #include <raft/spectral/matrix_wrappers.hpp> namespace cuvs { namespace spectral { namespace detail { // ========================================================= // Spectral partitioner // ========================================================= /// Compute spectral graph partition /** Compute partition for a weighted undirected graph. This * partition attempts to minimize the cost function: * Cost = \sum_i (Edges cut by ith partition)/(Vertices in ith partition) * * @param G Weighted graph in CSR format * @param nClusters Number of partitions. * @param nEigVecs Number of eigenvectors to compute. * @param maxIter_lanczos Maximum number of Lanczos iterations. * @param restartIter_lanczos Maximum size of Lanczos system before * implicit restart. * @param tol_lanczos Convergence tolerance for Lanczos method. * @param maxIter_kmeans Maximum number of k-means iterations. * @param tol_kmeans Convergence tolerance for k-means algorithm. * @param clusters (Output, device memory, n entries) Partition * assignments. * @param iters_lanczos On exit, number of Lanczos iterations * performed. * @param iters_kmeans On exit, number of k-means iterations * performed. * @return statistics: number of eigensolver iterations, . */ template <typename vertex_t, typename weight_t, typename EigenSolver, typename ClusterSolver> std::tuple<vertex_t, weight_t, vertex_t> partition( raft::resources const& handle, spectral::matrix::sparse_matrix_t<vertex_t, weight_t> const& csr_m, EigenSolver const& eigen_solver, ClusterSolver const& cluster_solver, vertex_t* __restrict__ clusters, weight_t* eigVals, weight_t* eigVecs) { RAFT_EXPECTS(clusters != nullptr, "Null clusters buffer."); RAFT_EXPECTS(eigVals != nullptr, "Null eigVals buffer."); RAFT_EXPECTS(eigVecs != nullptr, "Null eigVecs buffer."); auto stream = resource::get_cuda_stream(handle); auto cublas_h = resource::get_cublas_handle(handle); std::tuple<vertex_t, weight_t, vertex_t> stats; //{iters_eig_solver,residual_cluster,iters_cluster_solver} // # iters eigen solver, // cluster solver residual, # iters cluster solver vertex_t n = csr_m.nrows_; // ------------------------------------------------------- // Spectral partitioner // ------------------------------------------------------- // Compute eigenvectors of Laplacian // Initialize Laplacian /// sparse_matrix_t<vertex_t, weight_t> A{handle, graph}; spectral::matrix::laplacian_matrix_t<vertex_t, weight_t> L{handle, csr_m}; auto eigen_config = eigen_solver.get_config(); auto nEigVecs = eigen_config.n_eigVecs; // Compute smallest eigenvalues and eigenvectors std::get<0>(stats) = eigen_solver.solve_smallest_eigenvectors(handle, L, eigVals, eigVecs); // Whiten eigenvector matrix transform_eigen_matrix(handle, n, nEigVecs, eigVecs); // Find partition clustering auto pair_cluster = cluster_solver.solve(handle, n, nEigVecs, eigVecs, clusters); std::get<1>(stats) = pair_cluster.first; std::get<2>(stats) = pair_cluster.second; return stats; } // ========================================================= // Analysis of graph partition // ========================================================= /// Compute cost function for partition /** This function determines the edges cut by a partition and a cost * function: * Cost = \sum_i (Edges cut by ith partition)/(Vertices in ith partition) * Graph is assumed to be weighted and undirected. * * @param G Weighted graph in CSR format * @param nClusters Number of partitions. * @param clusters (Input, device memory, n entries) Partition * assignments. * @param edgeCut On exit, weight of edges cut by partition. * @param cost On exit, partition cost function. * @return error flag. */ template <typename vertex_t, typename weight_t> void analyzePartition(raft::resources const& handle, spectral::matrix::sparse_matrix_t<vertex_t, weight_t> const& csr_m, vertex_t nClusters, const vertex_t* __restrict__ clusters, weight_t& edgeCut, weight_t& cost) { RAFT_EXPECTS(clusters != nullptr, "Null clusters buffer."); vertex_t i; vertex_t n = csr_m.nrows_; auto stream = resource::get_cuda_stream(handle); auto cublas_h = resource::get_cublas_handle(handle); weight_t partEdgesCut, clustersize; // Device memory spectral::matrix::vector_t<weight_t> part_i(handle, n); spectral::matrix::vector_t<weight_t> Lx(handle, n); // Initialize cuBLAS RAFT_CUBLAS_TRY( raft::linalg::detail::cublassetpointermode(cublas_h, CUBLAS_POINTER_MODE_HOST, stream)); // Initialize Laplacian /// sparse_matrix_t<vertex_t, weight_t> A{handle, graph}; spectral::matrix::laplacian_matrix_t<vertex_t, weight_t> L{handle, csr_m}; // Initialize output cost = 0; edgeCut = 0; // Iterate through partitions for (i = 0; i < nClusters; ++i) { // Construct indicator vector for ith partition if (!construct_indicator(handle, i, n, clustersize, partEdgesCut, clusters, part_i, Lx, L)) { WARNING("empty partition"); continue; } // Record results cost += partEdgesCut / clustersize; edgeCut += partEdgesCut / 2; } } } // namespace detail } // namespace spectral } // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/detail/matrix_wrappers.hpp
/* * Copyright (c) 2020-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/core/resource/cublas_handle.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resource/cusparse_handle.hpp> #include <raft/core/resource/thrust_policy.hpp> #include <raft/core/resources.hpp> #include <raft/linalg/detail/cublas_wrappers.hpp> #include <raft/sparse/detail/cusparse_wrappers.h> #include <raft/util/cudart_utils.hpp> #include <rmm/device_uvector.hpp> #include <thrust/execution_policy.h> #include <thrust/fill.h> #include <thrust/reduce.h> #include <thrust/system/cuda/execution_policy.h> #include <algorithm> // ========================================================= // Useful macros // ========================================================= // Get index of matrix entry #define IDX(i, j, lda) ((i) + (j) * (lda)) namespace cuvs { namespace spectral { namespace matrix { namespace detail { using size_type = int; // for now; TODO: move it in appropriate header // Apply diagonal matrix to vector: // template <typename IndexType_, typename ValueType_> RAFT_KERNEL diagmv(IndexType_ n, ValueType_ alpha, const ValueType_* __restrict__ D, const ValueType_* __restrict__ x, ValueType_* __restrict__ y) { IndexType_ i = threadIdx.x + blockIdx.x * blockDim.x; while (i < n) { y[i] += alpha * D[i] * x[i]; i += blockDim.x * gridDim.x; } } // specifies type of algorithm used // for SpMv: // enum struct sparse_mv_alg_t : int { SPARSE_MV_UNDEFINED = -1, SPARSE_MV_ALG_DEFAULT, // generic, for any sparse matrix SPARSE_MV_ALG1, // typical for CSR SPARSE_MV_ALG2 // may provide better performance for irregular sparse matrices }; // Vector "view"-like aggregate for linear algebra purposes // template <typename value_type> struct vector_view_t { value_type* buffer_; size_type size_; vector_view_t(value_type* buffer, size_type sz) : buffer_(buffer), size_(sz) {} vector_view_t(vector_view_t&& other) : buffer_(other.raw()), size_(other.size()) {} vector_view_t& operator=(vector_view_t&& other) { buffer_ = other.raw(); size_ = other.size(); } }; template <typename value_type> class vector_t { public: vector_t(raft::resources const& raft_handle, size_type sz) : buffer_(sz, resource::get_cuda_stream(raft_handle)), thrust_policy(raft::resource::get_thrust_policy(raft_handle)) { } size_type size(void) const { return buffer_.size(); } value_type* raw(void) { return buffer_.data(); } value_type const* raw(void) const { return buffer_.data(); } value_type nrm1() const { return thrust::reduce(thrust_policy, buffer_.data(), buffer_.data() + buffer_.size(), value_type{0}, [] __device__(auto left, auto right) { auto abs_left = left > 0 ? left : -left; auto abs_right = right > 0 ? right : -right; return abs_left + abs_right; }); } void fill(value_type value) { thrust::fill_n(thrust_policy, buffer_.data(), buffer_.size(), value); } private: using thrust_exec_policy_t = thrust::detail::execute_with_allocator<rmm::mr::thrust_allocator<char>, thrust::cuda_cub::execute_on_stream_base>; rmm::device_uvector<value_type> buffer_; const thrust_exec_policy_t thrust_policy; }; template <typename index_type, typename value_type> struct sparse_matrix_t { sparse_matrix_t(raft::resources const& raft_handle, index_type const* row_offsets, index_type const* col_indices, value_type const* values, index_type const nrows, index_type const ncols, index_type const nnz) : handle_(raft_handle), row_offsets_(row_offsets), col_indices_(col_indices), values_(values), nrows_(nrows), ncols_(ncols), nnz_(nnz) { } sparse_matrix_t(raft::resources const& raft_handle, index_type const* row_offsets, index_type const* col_indices, value_type const* values, index_type const nrows, index_type const nnz) : handle_(raft_handle), row_offsets_(row_offsets), col_indices_(col_indices), values_(values), nrows_(nrows), ncols_(nrows), nnz_(nnz) { } template <typename CSRView> sparse_matrix_t(raft::resources const& raft_handle, CSRView const& csr_view) : handle_(raft_handle), row_offsets_(csr_view.offsets), col_indices_(csr_view.indices), values_(csr_view.edge_data), nrows_(csr_view.number_of_vertices), ncols_(csr_view.number_of_vertices), nnz_(csr_view.number_of_edges) { } virtual ~sparse_matrix_t(void) = default; // virtual because used as base for following matrix types // y = alpha*A*x + beta*y //(Note: removed const-ness of x, because CUDA 11 SpMV // descriptor creation works with non-const, and const-casting // down is dangerous) // virtual void mv(value_type alpha, value_type* __restrict__ x, value_type beta, value_type* __restrict__ y, sparse_mv_alg_t alg = sparse_mv_alg_t::SPARSE_MV_ALG1, bool transpose = false, bool symmetric = false) const { using namespace sparse; RAFT_EXPECTS(x != nullptr, "Null x buffer."); RAFT_EXPECTS(y != nullptr, "Null y buffer."); auto cusparse_h = resource::get_cusparse_handle(handle_); auto stream = resource::get_cuda_stream(handle_); cusparseOperation_t trans = transpose ? CUSPARSE_OPERATION_TRANSPOSE : // transpose CUSPARSE_OPERATION_NON_TRANSPOSE; // non-transpose #if not defined CUDA_ENFORCE_LOWER and CUDA_VER_10_1_UP auto size_x = transpose ? nrows_ : ncols_; auto size_y = transpose ? ncols_ : nrows_; cusparseSpMVAlg_t spmv_alg = translate_algorithm(alg); // create descriptors: //(below casts are necessary, because // cusparseCreateCsr(...) takes non-const // void*; the casts should be harmless) // cusparseSpMatDescr_t matA; RAFT_CUSPARSE_TRY(raft::sparse::detail::cusparsecreatecsr(&matA, nrows_, ncols_, nnz_, const_cast<index_type*>(row_offsets_), const_cast<index_type*>(col_indices_), const_cast<value_type*>(values_))); cusparseDnVecDescr_t vecX; RAFT_CUSPARSE_TRY(raft::sparse::detail::cusparsecreatednvec(&vecX, size_x, x)); rmm::device_uvector<value_type> y_tmp(size_y, stream); raft::copy(y_tmp.data(), y, size_y, stream); cusparseDnVecDescr_t vecY; RAFT_CUSPARSE_TRY(raft::sparse::detail::cusparsecreatednvec(&vecY, size_y, y_tmp.data())); // get (scratch) external device buffer size: // size_t bufferSize; RAFT_CUSPARSE_TRY(raft::sparse::detail::cusparsespmv_buffersize( cusparse_h, trans, &alpha, matA, vecX, &beta, vecY, spmv_alg, &bufferSize, stream)); // allocate external buffer: // vector_t<value_type> external_buffer(handle_, bufferSize); // finally perform SpMV: // RAFT_CUSPARSE_TRY(raft::sparse::detail::cusparsespmv( cusparse_h, trans, &alpha, matA, vecX, &beta, vecY, spmv_alg, external_buffer.raw(), stream)); // FIXME: This is a workaround for a cusparse issue being encountered in CUDA 12 raft::copy(y, y_tmp.data(), size_y, stream); // free descriptors: //(TODO: maybe wrap them in a RAII struct?) // RAFT_CUSPARSE_TRY(cusparseDestroyDnVec(vecY)); RAFT_CUSPARSE_TRY(cusparseDestroyDnVec(vecX)); RAFT_CUSPARSE_TRY(cusparseDestroySpMat(matA)); #else RAFT_CUSPARSE_TRY( raft::sparse::detail::cusparsesetpointermode(cusparse_h, CUSPARSE_POINTER_MODE_HOST, stream)); cusparseMatDescr_t descr = 0; RAFT_CUSPARSE_TRY(cusparseCreateMatDescr(&descr)); if (symmetric) { RAFT_CUSPARSE_TRY(cusparseSetMatType(descr, CUSPARSE_MATRIX_TYPE_SYMMETRIC)); } else { RAFT_CUSPARSE_TRY(cusparseSetMatType(descr, CUSPARSE_MATRIX_TYPE_GENERAL)); } RAFT_CUSPARSE_TRY(cusparseSetMatIndexBase(descr, CUSPARSE_INDEX_BASE_ZERO)); RAFT_CUSPARSE_TRY(raft::sparse::detail::cusparsecsrmv(cusparse_h, trans, nrows_, ncols_, nnz_, &alpha, descr, values_, row_offsets_, col_indices_, x, &beta, y, stream)); RAFT_CUSPARSE_TRY(cusparseDestroyMatDescr(descr)); #endif } resources const& get_handle(void) const { return handle_; } #if not defined CUDA_ENFORCE_LOWER and CUDA_VER_10_1_UP cusparseSpMVAlg_t translate_algorithm(sparse_mv_alg_t alg) const { switch (alg) { case sparse_mv_alg_t::SPARSE_MV_ALG1: return CUSPARSE_SPMV_CSR_ALG1; case sparse_mv_alg_t::SPARSE_MV_ALG2: return CUSPARSE_SPMV_CSR_ALG2; default: return CUSPARSE_SPMV_ALG_DEFAULT; } } #endif // private: // maybe not, keep this ASAPBNS ("as simple as possible, but not simpler"); hence, // aggregate raft::resources const& handle_; index_type const* row_offsets_; index_type const* col_indices_; value_type const* values_; index_type const nrows_; index_type const ncols_; index_type const nnz_; }; template <typename index_type, typename value_type> struct laplacian_matrix_t : sparse_matrix_t<index_type, value_type> { laplacian_matrix_t(raft::resources const& raft_handle, index_type const* row_offsets, index_type const* col_indices, value_type const* values, index_type const nrows, index_type const nnz) : sparse_matrix_t<index_type, value_type>( raft_handle, row_offsets, col_indices, values, nrows, nnz), diagonal_(raft_handle, nrows) { vector_t<value_type> ones{raft_handle, nrows}; ones.fill(1.0); sparse_matrix_t<index_type, value_type>::mv(1, ones.raw(), 0, diagonal_.raw()); } laplacian_matrix_t(raft::resources const& raft_handle, sparse_matrix_t<index_type, value_type> const& csr_m) : sparse_matrix_t<index_type, value_type>(raft_handle, csr_m.row_offsets_, csr_m.col_indices_, csr_m.values_, csr_m.nrows_, csr_m.nnz_), diagonal_(raft_handle, csr_m.nrows_) { vector_t<value_type> ones{raft_handle, csr_m.nrows_}; ones.fill(1.0); sparse_matrix_t<index_type, value_type>::mv(1, ones.raw(), 0, diagonal_.raw()); } // y = alpha*A*x + beta*y // void mv(value_type alpha, value_type* __restrict__ x, value_type beta, value_type* __restrict__ y, sparse_mv_alg_t alg = sparse_mv_alg_t::SPARSE_MV_ALG1, bool transpose = false, bool symmetric = false) const override { constexpr int BLOCK_SIZE = 1024; auto n = sparse_matrix_t<index_type, value_type>::nrows_; auto handle = sparse_matrix_t<index_type, value_type>::get_handle(); auto cublas_h = resource::get_cublas_handle(handle); auto stream = resource::get_cuda_stream(handle); // scales y by beta: // if (beta == 0) { RAFT_CUDA_TRY(cudaMemsetAsync(y, 0, n * sizeof(value_type), stream)); } else if (beta != 1) { // TODO: Call from public API when ready RAFT_CUBLAS_TRY(raft::linalg::detail::cublasscal(cublas_h, n, &beta, y, 1, stream)); } // Apply diagonal matrix // dim3 gridDim{std::min<unsigned int>((n + BLOCK_SIZE - 1) / BLOCK_SIZE, 65535), 1, 1}; dim3 blockDim{BLOCK_SIZE, 1, 1}; diagmv<<<gridDim, blockDim, 0, stream>>>(n, alpha, diagonal_.raw(), x, y); RAFT_CHECK_CUDA(stream); // Apply adjacency matrix // sparse_matrix_t<index_type, value_type>::mv(-alpha, x, 1, y, alg, transpose, symmetric); } vector_t<value_type> diagonal_; }; template <typename index_type, typename value_type> struct modularity_matrix_t : laplacian_matrix_t<index_type, value_type> { modularity_matrix_t(raft::resources const& raft_handle, index_type const* row_offsets, index_type const* col_indices, value_type const* values, index_type const nrows, index_type const nnz) : laplacian_matrix_t<index_type, value_type>( raft_handle, row_offsets, col_indices, values, nrows, nnz) { edge_sum_ = laplacian_matrix_t<index_type, value_type>::diagonal_.nrm1(); } modularity_matrix_t(raft::resources const& raft_handle, sparse_matrix_t<index_type, value_type> const& csr_m) : laplacian_matrix_t<index_type, value_type>(raft_handle, csr_m) { edge_sum_ = laplacian_matrix_t<index_type, value_type>::diagonal_.nrm1(); } // y = alpha*A*x + beta*y // void mv(value_type alpha, value_type* __restrict__ x, value_type beta, value_type* __restrict__ y, sparse_mv_alg_t alg = sparse_mv_alg_t::SPARSE_MV_ALG1, bool transpose = false, bool symmetric = false) const override { auto n = sparse_matrix_t<index_type, value_type>::nrows_; auto handle = sparse_matrix_t<index_type, value_type>::get_handle(); auto cublas_h = resource::get_cublas_handle(handle); auto stream = resource::get_cuda_stream(handle); // y = A*x // sparse_matrix_t<index_type, value_type>::mv(alpha, x, 0, y, alg, transpose, symmetric); value_type dot_res; // gamma = d'*x // // Cublas::dot(this->n, D.raw(), 1, x, 1, &dot_res); // TODO: Call from public API when ready RAFT_CUBLAS_TRY( raft::linalg::detail::cublasdot(cublas_h, n, laplacian_matrix_t<index_type, value_type>::diagonal_.raw(), 1, x, 1, &dot_res, stream)); // y = y -(gamma/edge_sum)*d // value_type gamma_ = -dot_res / edge_sum_; // TODO: Call from public API when ready RAFT_CUBLAS_TRY( raft::linalg::detail::cublasaxpy(cublas_h, n, &gamma_, laplacian_matrix_t<index_type, value_type>::diagonal_.raw(), 1, y, 1, stream)); } value_type edge_sum_; }; } // namespace detail } // namespace matrix } // namespace spectral } // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral
rapidsai_public_repos/cuvs/cpp/include/cuvs/spectral/detail/modularity_maximization.hpp
/* * Copyright (c) 2020-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <math.h> #include <raft/core/resource/cublas_handle.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <stdio.h> #include <cuda.h> #include <thrust/fill.h> #include <thrust/reduce.h> #include <thrust/transform.h> #include <tuple> #include <raft/linalg/detail/cublas_wrappers.hpp> #include <raft/spectral/cluster_solvers.cuh> #include <raft/spectral/detail/spectral_util.cuh> #include <raft/spectral/eigen_solvers.cuh> #include <raft/spectral/matrix_wrappers.hpp> namespace cuvs { namespace spectral { namespace detail { // ========================================================= // Spectral modularity_maximization // ========================================================= /** Compute partition for a weighted undirected graph. This * partition attempts to minimize the cost function: * Cost = \sum_i (Edges cut by ith partition)/(Vertices in ith partition) * * @param G Weighted graph in CSR format * @param nClusters Number of partitions. * @param nEigVecs Number of eigenvectors to compute. * @param maxIter_lanczos Maximum number of Lanczos iterations. * @param restartIter_lanczos Maximum size of Lanczos system before * implicit restart. * @param tol_lanczos Convergence tolerance for Lanczos method. * @param maxIter_kmeans Maximum number of k-means iterations. * @param tol_kmeans Convergence tolerance for k-means algorithm. * @param clusters (Output, device memory, n entries) Cluster * assignments. * @param iters_lanczos On exit, number of Lanczos iterations * performed. * @param iters_kmeans On exit, number of k-means iterations * performed. * @return error flag. */ template <typename vertex_t, typename weight_t, typename EigenSolver, typename ClusterSolver> std::tuple<vertex_t, weight_t, vertex_t> modularity_maximization( raft::resources const& handle, raft::spectral::matrix::sparse_matrix_t<vertex_t, weight_t> const& csr_m, EigenSolver const& eigen_solver, ClusterSolver const& cluster_solver, vertex_t* __restrict__ clusters, weight_t* eigVals, weight_t* eigVecs) { RAFT_EXPECTS(clusters != nullptr, "Null clusters buffer."); RAFT_EXPECTS(eigVals != nullptr, "Null eigVals buffer."); RAFT_EXPECTS(eigVecs != nullptr, "Null eigVecs buffer."); auto stream = resource::get_cuda_stream(handle); auto cublas_h = resource::get_cublas_handle(handle); std::tuple<vertex_t, weight_t, vertex_t> stats; // # iters eigen solver, cluster solver residual, # iters cluster solver vertex_t n = csr_m.nrows_; // Compute eigenvectors of Modularity Matrix // Initialize Modularity Matrix raft::spectral::matrix::modularity_matrix_t<vertex_t, weight_t> B{handle, csr_m}; auto eigen_config = eigen_solver.get_config(); auto nEigVecs = eigen_config.n_eigVecs; // Compute eigenvectors corresponding to largest eigenvalues std::get<0>(stats) = eigen_solver.solve_largest_eigenvectors(handle, B, eigVals, eigVecs); // Whiten eigenvector matrix transform_eigen_matrix(handle, n, nEigVecs, eigVecs); // notice that at this point the matrix has already been transposed, so we are scaling // columns scale_obs(nEigVecs, n, eigVecs); RAFT_CHECK_CUDA(stream); // Find partition clustering auto pair_cluster = cluster_solver.solve(handle, n, nEigVecs, eigVecs, clusters); std::get<1>(stats) = pair_cluster.first; std::get<2>(stats) = pair_cluster.second; return stats; } //=================================================== // Analysis of graph partition // ========================================================= /// Compute modularity /** This function determines the modularity based on a graph and cluster assignments * @param G Weighted graph in CSR format * @param nClusters Number of clusters. * @param clusters (Input, device memory, n entries) Cluster assignments. * @param modularity On exit, modularity */ template <typename vertex_t, typename weight_t> void analyzeModularity(raft::resources const& handle, raft::spectral::matrix::sparse_matrix_t<vertex_t, weight_t> const& csr_m, vertex_t nClusters, vertex_t const* __restrict__ clusters, weight_t& modularity) { RAFT_EXPECTS(clusters != nullptr, "Null clusters buffer."); vertex_t i; vertex_t n = csr_m.nrows_; weight_t partModularity, clustersize; auto cublas_h = resource::get_cublas_handle(handle); auto stream = resource::get_cuda_stream(handle); // Device memory raft::spectral::matrix::vector_t<weight_t> part_i(handle, n); raft::spectral::matrix::vector_t<weight_t> Bx(handle, n); // Initialize cuBLAS RAFT_CUBLAS_TRY(linalg::detail::cublassetpointermode(cublas_h, CUBLAS_POINTER_MODE_HOST, stream)); // Initialize Modularity raft::spectral::matrix::modularity_matrix_t<vertex_t, weight_t> B{handle, csr_m}; // Initialize output modularity = 0; // Iterate through partitions for (i = 0; i < nClusters; ++i) { if (!construct_indicator(handle, i, n, clustersize, partModularity, clusters, part_i, Bx, B)) { WARNING("empty partition"); continue; } // Record results modularity += partModularity; } modularity = modularity / B.diagonal_.nrm1(); } } // namespace detail } // namespace spectral } // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/distance.cuh
/* * Copyright (c) 2018-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #ifndef RAFT_EXPLICIT_INSTANTIATE_ONLY #include "distance-inl.cuh" #endif #ifdef RAFT_COMPILED #include "distance-ext.cuh" #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/distance_types.hpp
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once namespace cuvs::distance { /** enum to tell how to compute distance */ enum DistanceType : unsigned short { /** evaluate as dist_ij = sum(x_ik^2) + sum(y_ij)^2 - 2*sum(x_ik * y_jk) */ L2Expanded = 0, /** same as above, but inside the epilogue, perform square root operation */ L2SqrtExpanded = 1, /** cosine distance */ CosineExpanded = 2, /** L1 distance */ L1 = 3, /** evaluate as dist_ij += (x_ik - y-jk)^2 */ L2Unexpanded = 4, /** same as above, but inside the epilogue, perform square root operation */ L2SqrtUnexpanded = 5, /** basic inner product **/ InnerProduct = 6, /** Chebyshev (Linf) distance **/ Linf = 7, /** Canberra distance **/ Canberra = 8, /** Generalized Minkowski distance **/ LpUnexpanded = 9, /** Correlation distance **/ CorrelationExpanded = 10, /** Jaccard distance **/ JaccardExpanded = 11, /** Hellinger distance **/ HellingerExpanded = 12, /** Haversine distance **/ Haversine = 13, /** Bray-Curtis distance **/ BrayCurtis = 14, /** Jensen-Shannon distance**/ JensenShannon = 15, /** Hamming distance **/ HammingUnexpanded = 16, /** KLDivergence **/ KLDivergence = 17, /** RusselRao **/ RusselRaoExpanded = 18, /** Dice-Sorensen distance **/ DiceExpanded = 19, /** Precomputed (special value) **/ Precomputed = 100 }; /** * Whether minimal distance corresponds to similar elements (using the given metric). */ inline bool is_min_close(DistanceType metric) { bool select_min; switch (metric) { case DistanceType::InnerProduct: // Similarity metrics have the opposite meaning, i.e. nearest neighbors are those with larger // similarity (See the same logic at cpp/include/raft/sparse/spatial/detail/knn.cuh:362 // {perform_k_selection}) select_min = false; break; default: select_min = true; } return select_min; } namespace kernels { enum KernelType { LINEAR, POLYNOMIAL, RBF, TANH }; /** * Parameters for kernel matrices. * The following kernels are implemented: * - LINEAR \f[ K(x_1,x_2) = <x_1,x_2>, \f] where \f$< , >\f$ is the dot product * - POLYNOMIAL \f[ K(x_1, x_2) = (\gamma <x_1,x_2> + \mathrm{coef0})^\mathrm{degree} \f] * - RBF \f[ K(x_1, x_2) = \exp(- \gamma |x_1-x_2|^2) \f] * - TANH \f[ K(x_1, x_2) = \tanh(\gamma <x_1,x_2> + \mathrm{coef0}) \f] */ struct KernelParams { // Kernel function parameters KernelType kernel; //!< Type of the kernel function int degree; //!< Degree of polynomial kernel (ignored by others) double gamma; //!< multiplier in the double coef0; //!< additive constant in poly and tanh kernels }; } // end namespace kernels }; // namespace cuvs::distance
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/distance-inl.cuh
/* * Copyright (c) 2018-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/distance/detail/distance.cuh> #include <cuvs/distance/distance_types.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resources.hpp> #include <rmm/device_uvector.hpp> #include <type_traits> #include <raft/core/device_mdspan.hpp> namespace cuvs { namespace distance { /** * @defgroup pairwise_distance pointer-based pairwise distance prims * @{ */ /** * @brief Evaluate pairwise distances with the user epilogue lamba allowed * @tparam DistanceType which distance to evaluate * @tparam DataT input argument type * @tparam AccT accumulation type * @tparam OutT output type * @tparam FinalLambda user-defined epilogue lamba * @tparam IdxT Index type * @param handle raft handle for managing expensive resources * @param x first set of points * @param y second set of points * @param dist output distance matrix * @param m number of points in x * @param n number of points in y * @param k dimensionality * @param workspace temporary workspace needed for computations * @param worksize number of bytes of the workspace * @param fin_op the final gemm epilogue lambda * @param isRowMajor whether the matrices are row-major or col-major * @param metric_arg metric argument (used for Minkowski distance) * * @note fin_op: This is a device lambda which is supposed to operate upon the * input which is AccT and returns the output in OutT. It's signature is * as follows: <pre>OutT fin_op(AccT in, int g_idx);</pre>. If one needs * any other parameters, feel free to pass them via closure. */ template <cuvs::distance::DistanceType DistT, typename DataT, typename AccT, typename OutT, typename FinalLambda, typename IdxT = int> void distance(raft::resources const& handle, const DataT* x, const DataT* y, OutT* dist, IdxT m, IdxT n, IdxT k, void* workspace, size_t worksize, FinalLambda fin_op, bool isRowMajor = true, DataT metric_arg = 2.0f) { detail::distance<DistT, DataT, AccT, OutT, FinalLambda, IdxT>( handle, x, y, dist, m, n, k, workspace, worksize, fin_op, isRowMajor, metric_arg); } /** * @brief Evaluate pairwise distances for the simple use case * @tparam DistanceType which distance to evaluate * @tparam DataT input argument type * @tparam AccT accumulation type * @tparam OutT output type * @tparam IdxT Index type * @param handle raft handle for managing expensive resources * @param x first set of points * @param y second set of points * @param dist output distance matrix * @param m number of points in x * @param n number of points in y * @param k dimensionality * @param workspace temporary workspace needed for computations * @param worksize number of bytes of the workspace * @param isRowMajor whether the matrices are row-major or col-major * @param metric_arg metric argument (used for Minkowski distance) */ template <cuvs::distance::DistanceType DistT, typename DataT, typename AccT, typename OutT, typename IdxT = int> void distance(raft::resources const& handle, const DataT* x, const DataT* y, OutT* dist, IdxT m, IdxT n, IdxT k, void* workspace, size_t worksize, bool isRowMajor = true, DataT metric_arg = 2.0f) { detail::distance<DistT, DataT, AccT, OutT, IdxT>( handle, x, y, dist, m, n, k, workspace, worksize, isRowMajor, metric_arg); } /** * @brief Return the exact workspace size to compute the distance * @tparam DistanceType which distance to evaluate * @tparam DataT input argument type * @tparam AccT accumulation type * @tparam OutT output type * @tparam IdxT Index type * @param x first set of points * @param y second set of points * @param m number of points in x * @param n number of points in y * @param k dimensionality * * @note If the specified DistT doesn't need the workspace at all, it * returns 0. */ template <cuvs::distance::DistanceType DistT, typename DataT, typename AccT, typename OutT, typename IdxT = int> size_t getWorkspaceSize(const DataT* x, const DataT* y, IdxT m, IdxT n, IdxT k) { return detail::getWorkspaceSize<DistT, DataT, AccT, OutT, IdxT>(x, y, m, n, k); } /** * @brief Return the exact workspace size to compute the distance * @tparam DistanceType which distance to evaluate * @tparam DataT input argument type * @tparam AccT accumulation type * @tparam OutT output type * @tparam IdxT Index type * @param x first set of points (size m*k) * @param y second set of points (size n*k) * @return number of bytes needed in workspace * * @note If the specified DistT doesn't need the workspace at all, it * returns 0. */ template <cuvs::distance::DistanceType DistT, typename DataT, typename AccT, typename OutT, typename IdxT = int, typename layout> size_t getWorkspaceSize(raft::device_matrix_view<DataT, IdxT, layout> const& x, raft::device_matrix_view<DataT, IdxT, layout> const& y) { RAFT_EXPECTS(x.extent(1) == y.extent(1), "Number of columns must be equal."); return getWorkspaceSize<DistT, DataT, AccT, OutT, IdxT>( x.data_handle(), y.data_handle(), x.extent(0), y.extent(0), x.extent(1)); } /** * @brief Evaluate pairwise distances for the simple use case * @tparam DistanceType which distance to evaluate * @tparam DataT input argument type * @tparam AccT accumulation type * @tparam OutT output type * @tparam IdxT Index type * @param handle raft handle for managing expensive resources * @param x first set of points * @param y second set of points * @param dist output distance matrix * @param m number of points in x * @param n number of points in y * @param k dimensionality * @param isRowMajor whether the matrices are row-major or col-major * @param metric_arg metric argument (used for Minkowski distance) */ template <cuvs::distance::DistanceType DistT, typename DataT, typename AccT, typename OutT, typename IdxT = int> void distance(raft::resources const& handle, const DataT* x, const DataT* y, OutT* dist, IdxT m, IdxT n, IdxT k, bool isRowMajor = true, DataT metric_arg = 2.0f) { auto stream = raft::resource::get_cuda_stream(handle); rmm::device_uvector<char> workspace(0, stream); auto worksize = getWorkspaceSize<DistT, DataT, AccT, OutT, IdxT>(x, y, m, n, k); workspace.resize(worksize, stream); detail::distance<DistT, DataT, AccT, OutT, IdxT>( handle, x, y, dist, m, n, k, workspace.data(), worksize, isRowMajor, metric_arg); } /** * @brief Convenience wrapper around 'distance' prim to convert runtime metric * into compile time for the purpose of dispatch * @tparam Type input/accumulation/output data-type * @tparam IdxT indexing type * @param handle raft handle for managing expensive resources * @param x first set of points * @param y second set of points * @param dist output distance matrix * @param m number of points in x * @param n number of points in y * @param k dimensionality * @param workspace temporary workspace buffer which can get resized as per the * needed workspace size * @param metric distance metric * @param isRowMajor whether the matrices are row-major or col-major * @param metric_arg metric argument (used for Minkowski distance) */ template <typename Type, typename IdxT = int> void pairwise_distance(raft::resources const& handle, const Type* x, const Type* y, Type* dist, IdxT m, IdxT n, IdxT k, rmm::device_uvector<char>& workspace, cuvs::distance::DistanceType metric, bool isRowMajor = true, Type metric_arg = 2.0f) { cudaStream_t stream = raft::resource::get_cuda_stream(handle); auto dispatch = [&](auto distance_type) { auto worksize = getWorkspaceSize<distance_type(), Type, Type, Type, IdxT>(x, y, m, n, k); workspace.resize(worksize, stream); detail::distance<distance_type(), Type, Type, Type, IdxT>( handle, x, y, dist, m, n, k, workspace.data(), worksize, isRowMajor, metric_arg); }; switch (metric) { case DistanceType::Canberra: dispatch(std::integral_constant<DistanceType, DistanceType::Canberra>{}); break; case DistanceType::CorrelationExpanded: dispatch(std::integral_constant<DistanceType, DistanceType::CorrelationExpanded>{}); break; case DistanceType::CosineExpanded: dispatch(std::integral_constant<DistanceType, DistanceType::CosineExpanded>{}); break; case DistanceType::HammingUnexpanded: dispatch(std::integral_constant<DistanceType, DistanceType::HammingUnexpanded>{}); break; case DistanceType::HellingerExpanded: dispatch(std::integral_constant<DistanceType, DistanceType::HellingerExpanded>{}); break; case cuvs::distance::DistanceType::InnerProduct: dispatch(std::integral_constant<DistanceType, DistanceType::InnerProduct>{}); break; case DistanceType::JensenShannon: dispatch(std::integral_constant<DistanceType, DistanceType::JensenShannon>{}); break; case DistanceType::KLDivergence: dispatch(std::integral_constant<DistanceType, DistanceType::KLDivergence>{}); break; case DistanceType::L1: dispatch(std::integral_constant<DistanceType, DistanceType::L1>{}); break; case DistanceType::L2Expanded: dispatch(std::integral_constant<DistanceType, DistanceType::L2Expanded>{}); break; case DistanceType::L2SqrtExpanded: dispatch(std::integral_constant<DistanceType, DistanceType::L2SqrtExpanded>{}); break; case DistanceType::L2SqrtUnexpanded: dispatch(std::integral_constant<DistanceType, DistanceType::L2SqrtUnexpanded>{}); break; case DistanceType::L2Unexpanded: dispatch(std::integral_constant<DistanceType, DistanceType::L2Unexpanded>{}); break; case DistanceType::Linf: dispatch(std::integral_constant<DistanceType, DistanceType::Linf>{}); break; case DistanceType::LpUnexpanded: dispatch(std::integral_constant<DistanceType, DistanceType::LpUnexpanded>{}); break; case DistanceType::RusselRaoExpanded: dispatch(std::integral_constant<DistanceType, DistanceType::RusselRaoExpanded>{}); break; default: THROW("Unknown or unsupported distance metric '%d'!", (int)metric); }; } /** * @brief Convenience wrapper around 'distance' prim to convert runtime metric * into compile time for the purpose of dispatch * @tparam Type input/accumulation/output data-type * @tparam IdxT indexing type * @param handle raft handle for managing expensive resources * @param x first set of points * @param y second set of points * @param dist output distance matrix * @param m number of points in x * @param n number of points in y * @param k dimensionality * @param metric distance metric * @param isRowMajor whether the matrices are row-major or col-major * @param metric_arg metric argument (used for Minkowski distance) */ template <typename Type, typename IdxT = int> void pairwise_distance(raft::resources const& handle, const Type* x, const Type* y, Type* dist, IdxT m, IdxT n, IdxT k, cuvs::distance::DistanceType metric, bool isRowMajor = true, Type metric_arg = 2.0f) { auto stream = raft::resource::get_cuda_stream(handle); rmm::device_uvector<char> workspace(0, stream); pairwise_distance<Type, IdxT>( handle, x, y, dist, m, n, k, workspace, metric, isRowMajor, metric_arg); } /** @} */ /** * \defgroup distance_mdspan Pairwise distance functions * @{ */ /** * @brief Evaluate pairwise distances for the simple use case. * * Note: Only contiguous row- or column-major layouts supported currently. * * Usage example: * @code{.cpp} * #include <raft/core/resources.hpp> * #include <raft/core/device_mdarray.hpp> * #include <raft/random/make_blobs.cuh> * #include <cuvs/distance/distance.cuh> * * raft::raft::resources handle; * int n_samples = 5000; * int n_features = 50; * * auto input = raft::make_device_matrix<float>(handle, n_samples, n_features); * auto labels = raft::make_device_vector<int>(handle, n_samples); * auto output = raft::make_device_matrix<float>(handle, n_samples, n_samples); * * raft::random::make_blobs(handle, input.view(), labels.view()); * auto metric = cuvs::distance::DistanceType::L2SqrtExpanded; * cuvs::distance::pairwise_distance(handle, input.view(), input.view(), output.view(), metric); * @endcode * * @tparam DistanceType which distance to evaluate * @tparam DataT input argument type * @tparam AccT accumulation type * @tparam OutT output type * @tparam IdxT Index type * @param handle raft handle for managing expensive resources * @param x first set of points (size n*k) * @param y second set of points (size m*k) * @param dist output distance matrix (size n*m) * @param metric_arg metric argument (used for Minkowski distance) */ template <cuvs::distance::DistanceType DistT, typename DataT, typename AccT, typename OutT, typename layout = raft::layout_c_contiguous, typename IdxT = int> void distance(raft::resources const& handle, raft::device_matrix_view<DataT, IdxT, layout> const x, raft::device_matrix_view<DataT, IdxT, layout> const y, raft::device_matrix_view<OutT, IdxT, layout> dist, DataT metric_arg = 2.0f) { RAFT_EXPECTS(x.extent(1) == y.extent(1), "Number of columns must be equal."); RAFT_EXPECTS(dist.extent(0) == x.extent(0), "Number of rows in output must be equal to " "number of rows in X"); RAFT_EXPECTS(dist.extent(1) == y.extent(0), "Number of columns in output must be equal to " "number of rows in Y"); RAFT_EXPECTS(x.is_exhaustive(), "Input x must be contiguous."); RAFT_EXPECTS(y.is_exhaustive(), "Input y must be contiguous."); constexpr auto is_rowmajor = std::is_same_v<layout, raft::layout_c_contiguous>; distance<DistT, DataT, AccT, OutT, IdxT>(handle, x.data_handle(), y.data_handle(), dist.data_handle(), x.extent(0), y.extent(0), x.extent(1), is_rowmajor, metric_arg); } /** * @brief Convenience wrapper around 'distance' prim to convert runtime metric * into compile time for the purpose of dispatch * @tparam Type input/accumulation/output data-type * @tparam IdxT indexing type * @param handle raft handle for managing expensive resources * @param x first matrix of points (size mxk) * @param y second matrix of points (size nxk) * @param dist output distance matrix (size mxn) * @param metric distance metric * @param metric_arg metric argument (used for Minkowski distance) */ template <typename Type, typename layout = raft::layout_c_contiguous, typename IdxT = int> void pairwise_distance(raft::resources const& handle, raft::device_matrix_view<Type, IdxT, layout> const x, raft::device_matrix_view<Type, IdxT, layout> const y, raft::device_matrix_view<Type, IdxT, layout> dist, cuvs::distance::DistanceType metric, Type metric_arg = 2.0f) { RAFT_EXPECTS(x.extent(1) == y.extent(1), "Number of columns must be equal."); RAFT_EXPECTS(dist.extent(0) == x.extent(0), "Number of rows in output must be equal to " "number of rows in X"); RAFT_EXPECTS(dist.extent(1) == y.extent(0), "Number of columns in output must be equal to " "number of rows in Y"); RAFT_EXPECTS(x.is_exhaustive(), "Input x must be contiguous."); RAFT_EXPECTS(y.is_exhaustive(), "Input y must be contiguous."); RAFT_EXPECTS(dist.is_exhaustive(), "Output must be contiguous."); constexpr auto rowmajor = std::is_same_v<layout, raft::layout_c_contiguous>; auto stream = raft::resource::get_cuda_stream(handle); rmm::device_uvector<char> workspace(0, stream); pairwise_distance(handle, x.data_handle(), y.data_handle(), dist.data_handle(), x.extent(0), y.extent(0), x.extent(1), metric, rowmajor, metric_arg); } /** @} */ }; // namespace distance }; // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/fused_l2_nn.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #ifndef RAFT_EXPLICIT_INSTANTIATE_ONLY #include "fused_l2_nn-inl.cuh" #endif #ifdef RAFT_COMPILED #include "fused_l2_nn-ext.cuh" #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/fused_l2_nn_helpers.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/distance/detail/fused_l2_nn.cuh> #include <raft/core/resource/cuda_stream.hpp> namespace cuvs::distance { /** * \defgroup fused_l2_nn Fused 1-nearest neighbors * @{ */ template <typename LabelT, typename DataT> using KVPMinReduce = detail::KVPMinReduceImpl<LabelT, DataT>; template <typename LabelT, typename DataT> using MinAndDistanceReduceOp = detail::MinAndDistanceReduceOpImpl<LabelT, DataT>; template <typename LabelT, typename DataT> using MinReduceOp = detail::MinReduceOpImpl<LabelT, DataT>; /** @} */ /** * Initialize array using init value from reduction op */ template <typename DataT, typename OutT, typename IdxT, typename ReduceOpT> void initialize(raft::resources const& handle, OutT* min, IdxT m, DataT maxVal, ReduceOpT redOp) { detail::initialize<DataT, OutT, IdxT, ReduceOpT>( min, m, maxVal, redOp, resource::get_cuda_stream(handle)); } } // namespace cuvs::distance
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/fused_l2_nn-ext.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cstdint> // int64_t #include <cuvs/distance/fused_l2_nn_helpers.cuh> // include initialize and reduce operations #include <raft/core/kvp.hpp> // raft::KeyValuePair #include <raft/core/resources.hpp> // raft::resources #include <raft/util/raft_explicit.hpp> // RAFT_EXPLICIT #ifdef RAFT_EXPLICIT_INSTANTIATE_ONLY namespace cuvs { namespace distance { template <typename DataT, typename OutT, typename IdxT> void fusedL2NNMinReduce(OutT* min, const DataT* x, const DataT* y, const DataT* xn, const DataT* yn, IdxT m, IdxT n, IdxT k, void* workspace, bool sqrt, bool initOutBuffer, cudaStream_t stream) RAFT_EXPLICIT; } // namespace distance } // namespace cuvs #endif // RAFT_EXPLICIT_INSTANTIATE_ONLY #define instantiate_raft_distance_fusedL2NNMinReduce(DataT, OutT, IdxT) \ extern template void cuvs::distance::fusedL2NNMinReduce<DataT, OutT, IdxT>(OutT * min, \ const DataT* x, \ const DataT* y, \ const DataT* xn, \ const DataT* yn, \ IdxT m, \ IdxT n, \ IdxT k, \ void* workspace, \ bool sqrt, \ bool initOutBuffer, \ cudaStream_t stream) instantiate_raft_distance_fusedL2NNMinReduce(double, double, int); instantiate_raft_distance_fusedL2NNMinReduce(double, double, int64_t); instantiate_raft_distance_fusedL2NNMinReduce(float, float, int); instantiate_raft_distance_fusedL2NNMinReduce(float, float, int64_t); // We can't have comma's in the macro expansion, so we use the COMMA macro: #define COMMA , instantiate_raft_distance_fusedL2NNMinReduce(double, raft::KeyValuePair<int COMMA double>, int); instantiate_raft_distance_fusedL2NNMinReduce(double, raft::KeyValuePair<int64_t COMMA double>, int64_t); instantiate_raft_distance_fusedL2NNMinReduce(float, raft::KeyValuePair<int COMMA float>, int); instantiate_raft_distance_fusedL2NNMinReduce(float, raft::KeyValuePair<int64_t COMMA float>, int64_t); #undef COMMA #undef instantiate_raft_distance_fusedL2NNMinReduce
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/masked_nn.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #ifndef __MASKED_L2_NN_H #define __MASKED_L2_NN_H #pragma once #include <cuvs/distance/detail/masked_nn.cuh> #include <cuvs/distance/fused_l2_nn.cuh> #include <limits> #include <raft/core/handle.hpp> #include <raft/util/cuda_utils.cuh> #include <stdint.h> namespace cuvs { namespace distance { /** * \defgroup masked_nn Masked 1-nearest neighbors * @{ */ /** * @brief Parameter struct for masked_l2_nn function * * @tparam ReduceOpT Type of reduction operator in the epilogue. * @tparam KVPReduceOpT Type of Reduction operation on key value pairs. * * Usage example: * @code{.cpp} * #include <cuvs/distance/masked_nn.cuh> * * using IdxT = int; * using DataT = float; * using RedOpT = cuvs::distance::MinAndDistanceReduceOp<IdxT, DataT>; * using PairRedOpT = cuvs::distance::KVPMinReduce<IdxT, DataT>; * using ParamT = cuvs::distance::masked_l2_nn_params<RedOpT, PairRedOpT>; * * bool init_out = true; * bool sqrt = false; * * ParamT masked_l2_params{RedOpT{}, PairRedOpT{}, sqrt, init_out}; * @endcode * * Prescribes how to reduce a distance to an intermediate type (`redOp`), and * how to reduce two intermediate types (`pairRedOp`). Typically, a distance is * mapped to an (index, value) pair and (index, value) pair with the lowest * value (distance) is selected. * * In addition, prescribes whether to compute the square root of the distance * (`sqrt`) and whether to initialize the output buffer (`initOutBuffer`). */ template <typename ReduceOpT, typename KVPReduceOpT> struct masked_l2_nn_params { /** Reduction operator in the epilogue */ ReduceOpT redOp; /** Reduction operation on key value pairs */ KVPReduceOpT pairRedOp; /** Whether the output `minDist` should contain L2-sqrt */ bool sqrt; /** Whether to initialize the output buffer before the main kernel launch */ bool initOutBuffer; }; /** * @brief Masked L2 distance and 1-nearest-neighbor computation in a single call. * * This function enables faster computation of nearest neighbors if the * computation of distances between certain point pairs can be skipped. * * We use an adjacency matrix that describes which distances to calculate. The * points in `y` are divided into groups, and the adjacency matrix indicates * whether to compute distances between points in `x` and groups in `y`. In other * words, if `adj[i,k]` is true then distance between point `x_i`, and points in * `group_k` will be calculated. * * **Performance considerations** * * The points in `x` are processed in tiles of `M` points (`M` is currently 64, * but may change in the future). As a result, the largest compute time * reduction occurs if all `M` points can skip a group. If only part of the `M` * points can skip a group, then at most a minor compute time reduction and a * modest energy use reduction can be expected. * * The points in `y` are also grouped into tiles of `N` points (`N` is currently * 64, but may change in the future). As a result, group sizes should be larger * than `N` to avoid wasting computational resources. If the group sizes are * evenly divisible by `N`, then the computation is most efficient, although for * larger group sizes this effect is minor. * * * **Comparison to SDDM** * * [SDDMM](https://ieeexplore.ieee.org/document/8638042) (sampled dense-dense * matrix multiplication) is a matrix-matrix multiplication where only part of * the output is computed. Compared to masked_l2_nn, there are a few differences: * * - The output of masked_l2_nn is a single vector (of nearest neighbors) and not * a sparse matrix. * * - The sampling in masked_l2_nn is expressed through intermediate "groups" rather than a CSR format. * * @tparam DataT data type * @tparam OutT output type to either store 1-NN indices and their minimum * distances or store only the min distances. Accordingly, one * has to pass an appropriate `ReduceOpT` * @tparam IdxT indexing arithmetic type * @tparam ReduceOpT A struct to perform the final needed reduction operation * and also to initialize the output array elements with the * appropriate initial value needed for reduction. * * @param handle RAFT handle for managing expensive resources * @param params Parameter struct specifying the reduction operations. * @param[in] x First matrix. Row major. Dim = `m x k`. * (on device). * @param[in] y Second matrix. Row major. Dim = `n x k`. * (on device). * @param[in] x_norm L2 squared norm of `x`. Length = `m`. (on device). * @param[in] y_norm L2 squared norm of `y`. Length = `n`. (on device) * @param[in] adj A boolean adjacency matrix indicating for each * row of `x` and each group in `y` whether to compute the * distance. Dim = `m x num_groups`. * @param[in] group_idxs An array containing the *end* indices of each group * in `y`. The value of group_idxs[j] indicates the * start of group j + 1, i.e., it is the inclusive * scan of the group lengths. The first group is * always assumed to start at index 0 and the last * group typically ends at index `n`. Length = * `num_groups`. * @param[out] out will contain the reduced output (Length = `m`) * (on device) */ template <typename DataT, typename OutT, typename IdxT, typename ReduceOpT, typename KVPReduceOpT> void masked_l2_nn(raft::resources const& handle, cuvs::distance::masked_l2_nn_params<ReduceOpT, KVPReduceOpT> params, raft::device_matrix_view<const DataT, IdxT, raft::layout_c_contiguous> x, raft::device_matrix_view<const DataT, IdxT, raft::layout_c_contiguous> y, raft::device_vector_view<const DataT, IdxT, raft::layout_c_contiguous> x_norm, raft::device_vector_view<const DataT, IdxT, raft::layout_c_contiguous> y_norm, raft::device_matrix_view<const bool, IdxT, raft::layout_c_contiguous> adj, raft::device_vector_view<const IdxT, IdxT, raft::layout_c_contiguous> group_idxs, raft::device_vector_view<OutT, IdxT, raft::layout_c_contiguous> out) { IdxT m = x.extent(0); IdxT n = y.extent(0); IdxT k = x.extent(1); IdxT num_groups = group_idxs.extent(0); // Match k dimension of x, y RAFT_EXPECTS(x.extent(1) == y.extent(1), "Dimension of vectors in x and y must be equal."); // Match x, x_norm and y, y_norm RAFT_EXPECTS(m == x_norm.extent(0), "Length of `x_norm` must match input `x`."); RAFT_EXPECTS(n == y_norm.extent(0), "Length of `y_norm` must match input `y` "); // Match adj to x and group_idxs RAFT_EXPECTS(m == adj.extent(0), "#rows in `adj` must match input `x`."); RAFT_EXPECTS(num_groups == adj.extent(1), "#cols in `adj` must match length of `group_idxs`."); // NOTE: We do not check if all indices in group_idxs actually points *inside* y. // If there is no work to be done, return immediately. if (m == 0 || n == 0 || k == 0 || num_groups == 0) { return; } detail::masked_l2_nn_impl<DataT, OutT, IdxT, ReduceOpT>(handle, out.data_handle(), x.data_handle(), y.data_handle(), x_norm.data_handle(), y_norm.data_handle(), adj.data_handle(), group_idxs.data_handle(), num_groups, m, n, k, params.redOp, params.pairRedOp, params.sqrt, params.initOutBuffer); } /** @} */ } // namespace distance } // namespace cuvs #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/distance-ext.cuh
/* * Copyright (c) 2018-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/distance/detail/kernels/rbf_fin_op.cuh> // rbf_fin_op #include <cuvs/distance/distance_types.hpp> // cuvs::distance::DistanceType #include <raft/core/device_mdspan.hpp> // raft::device_matrix_view #include <raft/core/operators.hpp> // raft::identity_op #include <raft/core/resources.hpp> // raft::resources #include <raft/util/raft_explicit.hpp> // RAFT_EXPLICIT #include <rmm/device_uvector.hpp> // rmm::device_uvector #ifdef RAFT_EXPLICIT_INSTANTIATE_ONLY namespace cuvs { namespace distance { template <cuvs::distance::DistanceType DistT, typename DataT, typename AccT, typename OutT, typename FinalLambda, typename IdxT = int> void distance(raft::resources const& handle, const DataT* x, const DataT* y, OutT* dist, IdxT m, IdxT n, IdxT k, void* workspace, size_t worksize, FinalLambda fin_op, bool isRowMajor = true, DataT metric_arg = 2.0f) RAFT_EXPLICIT; template <cuvs::distance::DistanceType DistT, typename DataT, typename AccT, typename OutT, typename IdxT = int> void distance(raft::resources const& handle, const DataT* x, const DataT* y, OutT* dist, IdxT m, IdxT n, IdxT k, void* workspace, size_t worksize, bool isRowMajor = true, DataT metric_arg = 2.0f) RAFT_EXPLICIT; template <cuvs::distance::DistanceType DistT, typename DataT, typename AccT, typename OutT, typename IdxT = int> size_t getWorkspaceSize(const DataT* x, const DataT* y, IdxT m, IdxT n, IdxT k) RAFT_EXPLICIT; template <cuvs::distance::DistanceType DistT, typename DataT, typename AccT, typename OutT, typename IdxT = int, typename layout> size_t getWorkspaceSize(raft::device_matrix_view<DataT, IdxT, layout> const& x, raft::device_matrix_view<DataT, IdxT, layout> const& y) RAFT_EXPLICIT; template <cuvs::distance::DistanceType DistT, typename DataT, typename AccT, typename OutT, typename IdxT = int> void distance(raft::resources const& handle, const DataT* x, const DataT* y, OutT* dist, IdxT m, IdxT n, IdxT k, bool isRowMajor = true, DataT metric_arg = 2.0f) RAFT_EXPLICIT; template <typename Type, typename IdxT = int> void pairwise_distance(raft::resources const& handle, const Type* x, const Type* y, Type* dist, IdxT m, IdxT n, IdxT k, rmm::device_uvector<char>& workspace, cuvs::distance::DistanceType metric, bool isRowMajor = true, Type metric_arg = 2.0f) RAFT_EXPLICIT; template <typename Type, typename IdxT = int> void pairwise_distance(raft::resources const& handle, const Type* x, const Type* y, Type* dist, IdxT m, IdxT n, IdxT k, cuvs::distance::DistanceType metric, bool isRowMajor = true, Type metric_arg = 2.0f) RAFT_EXPLICIT; template <cuvs::distance::DistanceType DistT, typename DataT, typename AccT, typename OutT, typename layout = raft::layout_c_contiguous, typename IdxT = int> void distance(raft::resources const& handle, raft::device_matrix_view<DataT, IdxT, layout> const x, raft::device_matrix_view<DataT, IdxT, layout> const y, raft::device_matrix_view<OutT, IdxT, layout> dist, DataT metric_arg = 2.0f) RAFT_EXPLICIT; template <typename Type, typename layout = raft::layout_c_contiguous, typename IdxT = int> void pairwise_distance(raft::resources const& handle, device_matrix_view<Type, IdxT, layout> const x, device_matrix_view<Type, IdxT, layout> const y, device_matrix_view<Type, IdxT, layout> dist, cuvs::distance::DistanceType metric, Type metric_arg = 2.0f) RAFT_EXPLICIT; }; // namespace distance }; // namespace cuvs #endif // RAFT_EXPLICIT_INSTANTIATE_ONLY /* * Hierarchy of instantiations: * * This file defines the extern template instantiations for the public API of * cuvs::distance. To improve compile times, the extern template instantiation * of the distance kernels is handled in * distance/detail/pairwise_matrix/dispatch-ext.cuh. * * After adding an instance here, make sure to also add the instance to * dispatch-ext.cuh and the corresponding .cu files. */ #define instantiate_raft_distance_distance(DT, DataT, AccT, OutT, FinalLambda, IdxT) \ extern template void cuvs::distance::distance<DT, DataT, AccT, OutT, FinalLambda, IdxT>( \ raft::resources const& handle, \ const DataT* x, \ const DataT* y, \ OutT* dist, \ IdxT m, \ IdxT n, \ IdxT k, \ void* workspace, \ size_t worksize, \ FinalLambda fin_op, \ bool isRowMajor, \ DataT metric_arg) // The following two instances are used in test/distance/gram.cu. Note the use // of int64_t for the index type. instantiate_raft_distance_distance(cuvs::distance::DistanceType::L2Unexpanded, float, float, float, cuvs::distance::kernels::detail::rbf_fin_op<float>, int64_t); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L2Unexpanded, double, double, double, cuvs::distance::kernels::detail::rbf_fin_op<double>, int64_t); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Canberra, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Canberra, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::CorrelationExpanded, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::CorrelationExpanded, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::CosineExpanded, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::CosineExpanded, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::HammingUnexpanded, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::HammingUnexpanded, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::HellingerExpanded, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::HellingerExpanded, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::InnerProduct, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::InnerProduct, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::JensenShannon, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::JensenShannon, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::KLDivergence, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::KLDivergence, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L1, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L1, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Expanded, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Expanded, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2SqrtExpanded, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2SqrtExpanded, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2SqrtUnexpanded, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2SqrtUnexpanded, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Unexpanded, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Unexpanded, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Linf, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Linf, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::LpUnexpanded, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::LpUnexpanded, double, double, double, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::RusselRaoExpanded, float, float, float, raft::identity_op, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::RusselRaoExpanded, double, double, double, raft::identity_op, int); #undef instantiate_raft_distance_distance // Same, but without raft::identity_op #define instantiate_raft_distance_distance(DT, DataT, AccT, OutT, IdxT) \ extern template void cuvs::distance::distance<DT, DataT, AccT, OutT, IdxT>( \ raft::resources const& handle, \ const DataT* x, \ const DataT* y, \ OutT* dist, \ IdxT m, \ IdxT n, \ IdxT k, \ void* workspace, \ size_t worksize, \ bool isRowMajor, \ DataT metric_arg) instantiate_raft_distance_distance( cuvs::distance::DistanceType::Canberra, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Canberra, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::CorrelationExpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::CorrelationExpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::CosineExpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::CosineExpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::HammingUnexpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::HammingUnexpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::HellingerExpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::HellingerExpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::InnerProduct, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::InnerProduct, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::JensenShannon, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::JensenShannon, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::KLDivergence, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::KLDivergence, double, double, double, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L1, float, float, float, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L1, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Expanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Expanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2SqrtExpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2SqrtExpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2SqrtUnexpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2SqrtUnexpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Unexpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Unexpanded, double, double, double, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::Linf, float, float, float, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::Linf, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::LpUnexpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::LpUnexpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::RusselRaoExpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::RusselRaoExpanded, double, double, double, int); #undef instantiate_raft_distance_distance // Same, but without workspace #define instantiate_raft_distance_distance(DT, DataT, AccT, OutT, IdxT) \ extern template void cuvs::distance::distance<DT, DataT, AccT, OutT, IdxT>( \ raft::resources const& handle, \ const DataT* x, \ const DataT* y, \ OutT* dist, \ IdxT m, \ IdxT n, \ IdxT k, \ bool isRowMajor, \ DataT metric_arg) instantiate_raft_distance_distance( cuvs::distance::DistanceType::Canberra, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Canberra, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::CorrelationExpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::CorrelationExpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::CosineExpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::CosineExpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::HammingUnexpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::HammingUnexpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::HellingerExpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::HellingerExpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::InnerProduct, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::InnerProduct, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::JensenShannon, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::JensenShannon, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::KLDivergence, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::KLDivergence, double, double, double, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L1, float, float, float, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L1, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Expanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Expanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2SqrtExpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2SqrtExpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2SqrtUnexpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2SqrtUnexpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Unexpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Unexpanded, double, double, double, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::Linf, float, float, float, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::Linf, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::LpUnexpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::LpUnexpanded, double, double, double, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::RusselRaoExpanded, float, float, float, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::RusselRaoExpanded, double, double, double, int); #undef instantiate_raft_distance_distance #define instantiate_raft_distance_getWorkspaceSize(DistT, DataT, AccT, OutT, IdxT) \ extern template size_t cuvs::distance::getWorkspaceSize<DistT, DataT, AccT, OutT, IdxT>( \ const DataT* x, const DataT* y, IdxT m, IdxT n, IdxT k) instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::Canberra, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::Canberra, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::CorrelationExpanded, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::CorrelationExpanded, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::CosineExpanded, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::CosineExpanded, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::HammingUnexpanded, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::HammingUnexpanded, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::HellingerExpanded, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::HellingerExpanded, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::InnerProduct, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::InnerProduct, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::JensenShannon, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::JensenShannon, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::KLDivergence, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::KLDivergence, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L1, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L1, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2Expanded, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2Expanded, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2SqrtExpanded, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2SqrtExpanded, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2SqrtUnexpanded, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2SqrtUnexpanded, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2Unexpanded, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2Unexpanded, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::Linf, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::Linf, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::LpUnexpanded, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::LpUnexpanded, double, double, double, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::RusselRaoExpanded, float, float, float, int); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::RusselRaoExpanded, double, double, double, int); #undef instantiate_raft_distance_getWorkspaceSize #define instantiate_raft_distance_getWorkspaceSize(DistT, DataT, AccT, OutT, IdxT, layout) \ extern template size_t cuvs::distance::getWorkspaceSize<DistT, DataT, AccT, OutT, IdxT, layout>( \ raft::device_matrix_view<DataT, IdxT, layout> const& x, \ raft::device_matrix_view<DataT, IdxT, layout> const& y) // We could consider not taking template parameters for this function. The // number of instantiations seems a bit excessive.. instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::Canberra, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::Canberra, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::Canberra, float, float, float, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::Canberra, double, double, double, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::CorrelationExpanded, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::CorrelationExpanded, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::CorrelationExpanded, float, float, float, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::CorrelationExpanded, double, double, double, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::CosineExpanded, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::CosineExpanded, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::CosineExpanded, float, float, float, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::CosineExpanded, double, double, double, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::HammingUnexpanded, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::HammingUnexpanded, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::HammingUnexpanded, float, float, float, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::HammingUnexpanded, double, double, double, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::HellingerExpanded, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::HellingerExpanded, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::HellingerExpanded, float, float, float, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::HellingerExpanded, double, double, double, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::InnerProduct, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::InnerProduct, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::InnerProduct, float, float, float, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::InnerProduct, double, double, double, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::JensenShannon, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::JensenShannon, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::JensenShannon, float, float, float, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::JensenShannon, double, double, double, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::KLDivergence, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::KLDivergence, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::KLDivergence, float, float, float, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::KLDivergence, double, double, double, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L1, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L1, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L1, float, float, float, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L1, double, double, double, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2Expanded, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2Expanded, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2Expanded, float, float, float, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2Expanded, double, double, double, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::L2SqrtExpanded, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::L2SqrtExpanded, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::L2SqrtExpanded, float, float, float, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::L2SqrtExpanded, double, double, double, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::L2SqrtUnexpanded, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::L2SqrtUnexpanded, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::L2SqrtUnexpanded, float, float, float, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::L2SqrtUnexpanded, double, double, double, int, raft::layout_f_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2Unexpanded, float, float, float, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize(cuvs::distance::DistanceType::L2Unexpanded, double, double, double, int, raft::layout_c_contiguous); instantiate_raft_distance_getWorkspaceSize( cuvs::distance::DistanceType::L2Unexpanded, float, float, float, int, raft::layout_f_contiguous); #undef instantiate_raft_distance_getWorkspaceSize #define instantiate_raft_distance_pairwise_distance(DataT, IdxT) \ extern template void cuvs::distance::pairwise_distance(raft::resources const& handle, \ const DataT* x, \ const DataT* y, \ DataT* dist, \ IdxT m, \ IdxT n, \ IdxT k, \ rmm::device_uvector<char>& workspace, \ cuvs::distance::DistanceType metric, \ bool isRowMajor, \ DataT metric_arg) instantiate_raft_distance_pairwise_distance(float, int); instantiate_raft_distance_pairwise_distance(double, int); #undef instantiate_raft_distance_pairwise_distance // Same, but without workspace #define instantiate_raft_distance_pairwise_distance(DataT, IdxT) \ extern template void cuvs::distance::pairwise_distance(raft::resources const& handle, \ const DataT* x, \ const DataT* y, \ DataT* dist, \ IdxT m, \ IdxT n, \ IdxT k, \ cuvs::distance::DistanceType metric, \ bool isRowMajor, \ DataT metric_arg) instantiate_raft_distance_pairwise_distance(float, int); instantiate_raft_distance_pairwise_distance(double, int); #undef instantiate_raft_distance_pairwise_distance // Version with mdspan #define instantiate_raft_distance_distance(DistT, DataT, AccT, OutT, layout, IdxT) \ extern template void cuvs::distance::distance<DistT, DataT, AccT, OutT, layout, IdxT>( \ raft::resources const& handle, \ raft::device_matrix_view<DataT, IdxT, layout> const x, \ raft::device_matrix_view<DataT, IdxT, layout> const y, \ raft::device_matrix_view<OutT, IdxT, layout> dist, \ DataT metric_arg) // Again, we might want to consider reigning in the number of instantiations... instantiate_raft_distance_distance( cuvs::distance::DistanceType::Canberra, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Canberra, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Canberra, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Canberra, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::CorrelationExpanded, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::CorrelationExpanded, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::CorrelationExpanded, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::CorrelationExpanded, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::CosineExpanded, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::CosineExpanded, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::CosineExpanded, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::CosineExpanded, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::HammingUnexpanded, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::HammingUnexpanded, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::HammingUnexpanded, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::HammingUnexpanded, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::HellingerExpanded, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::HellingerExpanded, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::HellingerExpanded, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::HellingerExpanded, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::InnerProduct, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::InnerProduct, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::InnerProduct, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::InnerProduct, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::JensenShannon, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::JensenShannon, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::JensenShannon, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::JensenShannon, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::KLDivergence, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::KLDivergence, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::KLDivergence, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::KLDivergence, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L1, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L1, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L1, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L1, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Expanded, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Expanded, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Expanded, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Expanded, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L2SqrtExpanded, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L2SqrtExpanded, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L2SqrtExpanded, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L2SqrtExpanded, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L2SqrtUnexpanded, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L2SqrtUnexpanded, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L2SqrtUnexpanded, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L2SqrtUnexpanded, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Unexpanded, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L2Unexpanded, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::L2Unexpanded, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::L2Unexpanded, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Linf, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Linf, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Linf, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::Linf, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::LpUnexpanded, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::LpUnexpanded, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance( cuvs::distance::DistanceType::LpUnexpanded, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::LpUnexpanded, double, double, double, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::RusselRaoExpanded, float, float, float, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::RusselRaoExpanded, double, double, double, raft::layout_c_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::RusselRaoExpanded, float, float, float, raft::layout_f_contiguous, int); instantiate_raft_distance_distance(cuvs::distance::DistanceType::RusselRaoExpanded, double, double, double, raft::layout_f_contiguous, int); #undef instantiate_raft_distance_distance #define instantiate_raft_distance_pairwise_distance(DataT, layout, IdxT) \ extern template void cuvs::distance::pairwise_distance( \ raft::resources const& handle, \ raft::device_matrix_view<DataT, IdxT, layout> const x, \ raft::device_matrix_view<DataT, IdxT, layout> const y, \ raft::device_matrix_view<DataT, IdxT, layout> dist, \ cuvs::distance::DistanceType metric, \ DataT metric_arg) instantiate_raft_distance_pairwise_distance(float, raft::layout_c_contiguous, int); instantiate_raft_distance_pairwise_distance(float, raft::layout_f_contiguous, int); instantiate_raft_distance_pairwise_distance(double, raft::layout_c_contiguous, int); instantiate_raft_distance_pairwise_distance(double, raft::layout_f_contiguous, int); #undef instantiate_raft_distance_pairwise_distance
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/fused_l2_nn-inl.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #ifndef __FUSED_L2_NN_H #define __FUSED_L2_NN_H #pragma once #include <cub/cub.cuh> #include <cuvs/distance/detail/fused_l2_nn.cuh> #include <cuvs/distance/fused_l2_nn_helpers.cuh> #include <limits> #include <raft/core/resources.hpp> #include <raft/linalg/contractions.cuh> #include <raft/util/cuda_utils.cuh> #include <stdint.h> #include <type_traits> namespace cuvs { namespace distance { /** * \ingroup fused_l2_nn * @{ */ /** * @brief Fused L2 distance and 1-nearest-neighbor computation in a single call. * * The benefits of such a call are 2-fold: 1) eliminate the need for an * intermediate buffer to store the output of gemm 2) reduce the memory read * traffic on this intermediate buffer, otherwise needed during the reduction * phase for 1-NN. * * @tparam DataT data type * @tparam OutT output type to either store 1-NN indices and their minimum * distances or store only the min distances. Accordingly, one * has to pass an appropriate `ReduceOpT` * @tparam IdxT indexing arithmetic type * @tparam ReduceOpT A struct to perform the final needed reduction operation * and also to initialize the output array elements with the * appropriate initial value needed for reduction. * * @param[out] min will contain the reduced output (Length = `m`) * (on device) * @param[in] x first matrix. Row major. Dim = `m x k`. * (on device). * @param[in] y second matrix. Row major. Dim = `n x k`. * (on device). * @param[in] xn L2 squared norm of `x`. Length = `m`. (on device). * @param[in] yn L2 squared norm of `y`. Length = `n`. (on device) * @param[in] m gemm m * @param[in] n gemm n * @param[in] k gemm k * @param[in] workspace temp workspace. Size = sizeof(int)*m. (on device) * @param[in] redOp reduction operator in the epilogue * @param[in] pairRedOp reduction operation on key value pairs * @param[in] sqrt Whether the output `minDist` should contain L2-sqrt * @param[in] initOutBuffer whether to initialize the output buffer before the * main kernel launch * @param[in] stream cuda stream */ template <typename DataT, typename OutT, typename IdxT, typename ReduceOpT, typename KVPReduceOpT> void fusedL2NN(OutT* min, const DataT* x, const DataT* y, const DataT* xn, const DataT* yn, IdxT m, IdxT n, IdxT k, void* workspace, ReduceOpT redOp, KVPReduceOpT pairRedOp, bool sqrt, bool initOutBuffer, cudaStream_t stream) { // When k is smaller than 32, the Policy4x4 results in redundant calculations // as it uses tiles that have k=32. Therefore, use a "skinny" policy instead // that uses tiles with a smaller value of k. bool is_skinny = k < 32; size_t bytes = sizeof(DataT) * k; auto px = reinterpret_cast<uintptr_t>(x); auto py = reinterpret_cast<uintptr_t>(y); if (16 % sizeof(DataT) == 0 && bytes % 16 == 0 && px % 16 == 0 && py % 16 == 0) { if (is_skinny) { detail::fusedL2NNImpl< DataT, OutT, IdxT, typename raft::linalg::Policy4x4Skinny<DataT, 16 / sizeof(DataT)>::Policy, ReduceOpT>( min, x, y, xn, yn, m, n, k, (int*)workspace, redOp, pairRedOp, sqrt, initOutBuffer, stream); } else { detail::fusedL2NNImpl<DataT, OutT, IdxT, typename raft::linalg::Policy4x4<DataT, 16 / sizeof(DataT)>::Policy, ReduceOpT>( min, x, y, xn, yn, m, n, k, (int*)workspace, redOp, pairRedOp, sqrt, initOutBuffer, stream); } } else if (8 % sizeof(DataT) == 0 && bytes % 8 == 0 && px % 8 == 0 && py % 8 == 0) { if (is_skinny) { detail::fusedL2NNImpl< DataT, OutT, IdxT, typename raft::linalg::Policy4x4Skinny<DataT, 8 / sizeof(DataT)>::Policy, ReduceOpT>( min, x, y, xn, yn, m, n, k, (int*)workspace, redOp, pairRedOp, sqrt, initOutBuffer, stream); } else { detail::fusedL2NNImpl<DataT, OutT, IdxT, typename raft::linalg::Policy4x4<DataT, 8 / sizeof(DataT)>::Policy, ReduceOpT>( min, x, y, xn, yn, m, n, k, (int*)workspace, redOp, pairRedOp, sqrt, initOutBuffer, stream); } } else { if (is_skinny) { detail::fusedL2NNImpl<DataT, OutT, IdxT, typename raft::linalg::Policy4x4Skinny<DataT, 1>::Policy, ReduceOpT>( min, x, y, xn, yn, m, n, k, (int*)workspace, redOp, pairRedOp, sqrt, initOutBuffer, stream); } else { detail::fusedL2NNImpl<DataT, OutT, IdxT, typename raft::linalg::Policy4x4<DataT, 1>::Policy, ReduceOpT>( min, x, y, xn, yn, m, n, k, (int*)workspace, redOp, pairRedOp, sqrt, initOutBuffer, stream); } } } /** * @brief Wrapper around fusedL2NN with minimum reduction operators. * * fusedL2NN cannot be compiled in the distance library due to the lambda * operators, so this wrapper covers the most common case (minimum). * This should be preferred to the more generic API when possible, in order to * reduce compilation times for users of the shared library. * * @tparam DataT data type * @tparam OutT output type to either store 1-NN indices and their minimum * distances (e.g. raft::KeyValuePair<int, float>) or store only the min * distances. * @tparam IdxT indexing arithmetic type * @param[out] min will contain the reduced output (Length = `m`) * (on device) * @param[in] x first matrix. Row major. Dim = `m x k`. * (on device). * @param[in] y second matrix. Row major. Dim = `n x k`. * (on device). * @param[in] xn L2 squared norm of `x`. Length = `m`. (on device). * @param[in] yn L2 squared norm of `y`. Length = `n`. (on device) * @param[in] m gemm m * @param[in] n gemm n * @param[in] k gemm k * @param[in] workspace temp workspace. Size = sizeof(int)*m. (on device) * @param[in] sqrt Whether the output `minDist` should contain L2-sqrt * @param[in] initOutBuffer whether to initialize the output buffer before the * main kernel launch * @param[in] stream cuda stream */ template <typename DataT, typename OutT, typename IdxT> void fusedL2NNMinReduce(OutT* min, const DataT* x, const DataT* y, const DataT* xn, const DataT* yn, IdxT m, IdxT n, IdxT k, void* workspace, bool sqrt, bool initOutBuffer, cudaStream_t stream) { MinAndDistanceReduceOp<IdxT, DataT> redOp; KVPMinReduce<IdxT, DataT> pairRedOp; fusedL2NN<DataT, OutT, IdxT>( min, x, y, xn, yn, m, n, k, workspace, redOp, pairRedOp, sqrt, initOutBuffer, stream); } /** @} */ } // namespace distance } // namespace cuvs #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/kernels.cuh
/* * Copyright (c) 2019-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/distance/detail/kernels/gram_matrix.cuh> #include <cuvs/distance/detail/kernels/kernel_factory.cuh> #include <raft/util/cuda_utils.cuh> #include <cuvs/distance/distance.cuh> #include <raft/linalg/gemm.cuh> namespace cuvs::distance::kernels { // TODO: Need to expose formal APIs for this that are more consistent w/ other APIs in RAFT using cuvs::distance::kernels::detail::GramMatrixBase; using cuvs::distance::kernels::detail::KernelFactory; }; // end namespace cuvs::distance::kernels
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance.cuh
/* * Copyright (c) 2018-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/distance/detail/distance_ops/all_ops.cuh> #include <cuvs/distance/detail/pairwise_matrix/dispatch.cuh> #include <cuvs/distance/detail/pairwise_matrix/dispatch_sm60.cuh> #include <cuvs/distance/detail/pairwise_matrix/dispatch_sm80.cuh> #include <cuvs/distance/distance_types.hpp> #include <raft/core/operators.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/linalg/gemm.cuh> #include <raft/linalg/norm.cuh> #include <raft/linalg/reduce.cuh> #include <raft/linalg/unary_op.cuh> #include <type_traits> namespace cuvs { namespace distance { namespace detail { /** * @brief: A tag type for overload resolution based on DistanceType * * It is not possible to partially specialize function templates on a single * parameter. Instead, it is often easier to use a combination of conventional * method overloading and a parameter with a specific tag type. The following * type is used to help method overloading based on the DistanceType enum. */ template <DistanceType d> using distance_tag = std::integral_constant<DistanceType, d>; /** * @brief Implement pairwise_matrix for specific distance * * There are multiple overloads for this function, one for each distance type. * They are implemented below. The documentation of this function serves as * documentation for all functions. The following overloads are defined: * * - DistanceType::Canberra: * - DistanceType::CorrelationExpanded: * - DistanceType::CosineExpanded: * - DistanceType::HammingUnexpanded: * - DistanceType::HellingerExpanded: * - DistanceType::JensenShannon: * - DistanceType::KLDivergence: * - DistanceType::L1: * - DistanceType::L2Expanded: * - DistanceType::L2SqrtExpanded: * - DistanceType::L2Unexpanded: * - DistanceType::L2SqrtUnexpanded: * - DistanceType::Linf: * - DistanceType::LpUnexpanded: * - DistanceType::RusselRaoExpanded: * * @tparam DataT Input data type * @tparam AccT Accumulation data type * @tparam OutT Output data type * @tparam FinOpT Type of final operation * @tparam IdxT Index type * * @param handle RAFT resources handle * @param distance_type A tag type to indicate which distance is calculated. * @param x First set of points * @param y Second set of points * @param out Output distance matrix * @param m Number of points in x * @param n Number of points in y * @param k Dimensionality of points in x, y * @param workspace Temporary workspace needed for computations * @param worksize Number of bytes of the workspace * @param is_row_major Whether the matrices are row-major or col-major * @param metric_arg The `p` argument for Lp. */ template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::Canberra> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT* workspace, // unused size_t worksize, // unused FinOpT fin_op, bool is_row_major, DataT metric_arg) // unused { ops::canberra_distance_op<DataT, AccT, IdxT> distance_op{}; const DataT* x_norm = nullptr; const DataT* y_norm = nullptr; cudaStream_t stream = raft::resource::get_cuda_stream(handle); pairwise_matrix_dispatch<decltype(distance_op), DataT, AccT, OutT, FinOpT, IdxT>( distance_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::CorrelationExpanded> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT* workspace, size_t worksize, FinOpT fin_op, bool is_row_major, DataT) // unused { ASSERT(!(worksize < 2 * (m + n) * sizeof(AccT)), "workspace size error"); ASSERT(workspace != nullptr, "workspace is null"); cudaStream_t stream = raft::resource::get_cuda_stream(handle); AccT* x_norm = workspace; AccT* y_norm = workspace; AccT* sq_x_norm = workspace; AccT* sq_y_norm = workspace; // TODO: Column major case looks to have lower accuracy for X == Y, // perhaps the use of stridedSummationKernel could be causing this, // need to investigate and fix. if (x == y && is_row_major) { raft::linalg::reduce(x_norm, x, k, std::max(m, n), (AccT)0, is_row_major, true, stream, false, raft::identity_op(), raft::add_op()); sq_x_norm += std::max(m, n); sq_y_norm = sq_x_norm; raft::linalg::rowNorm( sq_x_norm, x, k, std::max(m, n), raft::linalg::L2Norm, is_row_major, stream); } else { y_norm += m; raft::linalg::reduce(x_norm, x, k, m, (AccT)0, is_row_major, true, stream, false, raft::identity_op(), raft::add_op()); raft::linalg::reduce(y_norm, y, k, n, (AccT)0, is_row_major, true, stream, false, raft::identity_op(), raft::add_op()); sq_x_norm += (m + n); sq_y_norm = sq_x_norm + m; raft::linalg::rowNorm(sq_x_norm, x, k, m, raft::linalg::L2Norm, is_row_major, stream); raft::linalg::rowNorm(sq_y_norm, y, k, n, raft::linalg::L2Norm, is_row_major, stream); } using OpT = ops::correlation_distance_op<DataT, AccT, IdxT>; OpT corr_op(is_row_major, sq_x_norm, sq_y_norm, m, n, k); pairwise_matrix_dispatch<decltype(corr_op), DataT, AccT, OutT, FinOpT, IdxT>( corr_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::CosineExpanded> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT* workspace, size_t worksize, FinOpT fin_op, bool is_row_major, DataT) // unused { // raft distance support inputs as float/double and output as uint8_t/float/double. static_assert(!((sizeof(OutT) > 1) && (sizeof(AccT) != sizeof(OutT))), "OutT can be uint8_t, float, double," "if sizeof(OutT) > 1 then sizeof(AccT) == sizeof(OutT)."); ASSERT(!(worksize < (m + n) * sizeof(AccT)), "workspace size error"); ASSERT(workspace != nullptr, "workspace is null"); cudaStream_t stream = raft::resource::get_cuda_stream(handle); DataT* x_norm = workspace; DataT* y_norm = workspace; // TODO: Column major case looks to have lower accuracy for X == Y, // perhaps the use of stridedSummationKernel could be causing this, // need to investigate and fix. if (x == y && is_row_major) { raft::linalg::rowNorm( x_norm, x, k, std::max(m, n), raft::linalg::L2Norm, is_row_major, stream, raft::sqrt_op{}); } else { y_norm += m; raft::linalg::rowNorm( x_norm, x, k, m, raft::linalg::L2Norm, is_row_major, stream, raft::sqrt_op{}); raft::linalg::rowNorm( y_norm, y, k, n, raft::linalg::L2Norm, is_row_major, stream, raft::sqrt_op{}); } ops::cosine_distance_op<DataT, AccT, IdxT> distance_op{}; pairwise_matrix_dispatch<decltype(distance_op), DataT, AccT, OutT, FinOpT, IdxT>( distance_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::HammingUnexpanded> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT*, // workspace unused size_t, // worksize unused FinOpT fin_op, bool is_row_major, DataT) // metric_arg unused { ops::hamming_distance_op<DataT, AccT, IdxT> distance_op{k}; const DataT* x_norm = nullptr; const DataT* y_norm = nullptr; cudaStream_t stream = raft::resource::get_cuda_stream(handle); pairwise_matrix_dispatch<decltype(distance_op), DataT, AccT, OutT, FinOpT, IdxT>( distance_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::InnerProduct> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT*, // workspace unused size_t, // worksize unused FinOpT fin_op, bool is_row_major, DataT) // metric_arg unused { cudaStream_t stream = raft::resource::get_cuda_stream(handle); raft::linalg::gemm(handle, out, const_cast<DataT*>(x), const_cast<DataT*>(y), m, n, k, !is_row_major, !is_row_major, is_row_major, stream); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::HellingerExpanded> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT*, // workspace unused size_t, // worksize unused FinOpT fin_op, bool is_row_major, DataT) // metric_arg unused { cudaStream_t stream = raft::resource::get_cuda_stream(handle); // First sqrt x and y const auto raft_sqrt = raft::linalg::unaryOp<DataT, raft::sqrt_op, IdxT>; raft_sqrt((DataT*)x, x, m * k, raft::sqrt_op{}, stream); if (x != y) { raft_sqrt((DataT*)y, y, n * k, raft::sqrt_op{}, stream); } // Then calculate Hellinger distance ops::hellinger_distance_op<DataT, AccT, IdxT> distance_op{}; const DataT* x_norm = nullptr; const DataT* y_norm = nullptr; pairwise_matrix_dispatch<decltype(distance_op), DataT, AccT, OutT, FinOpT, IdxT>( distance_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); // Finally revert sqrt of x and y raft_sqrt((DataT*)x, x, m * k, raft::sqrt_op{}, stream); if (x != y) { raft_sqrt((DataT*)y, y, n * k, raft::sqrt_op{}, stream); } RAFT_CUDA_TRY(cudaGetLastError()); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::JensenShannon> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT*, // workspace unused size_t, // worksize unused FinOpT fin_op, bool is_row_major, DataT) // metric_arg unused { ops::jensen_shannon_distance_op<DataT, AccT, IdxT> distance_op{}; const DataT* x_norm = nullptr; const DataT* y_norm = nullptr; cudaStream_t stream = raft::resource::get_cuda_stream(handle); pairwise_matrix_dispatch<decltype(distance_op), DataT, AccT, OutT, FinOpT, IdxT>( distance_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::KLDivergence> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT*, // workspace unused size_t, // worksize unused FinOpT fin_op, bool is_row_major, DataT) // metric_arg unused { cudaStream_t stream = raft::resource::get_cuda_stream(handle); auto unaryOp_lambda = [] __device__(DataT input) { const bool x_zero = (input == 0); return (!x_zero) * raft::log(input + x_zero); }; auto unaryOp_lambda_reverse = [] __device__(DataT input) { // reverse previous log (x) back to x using (e ^ log(x)) const bool x_zero = (input == 0); return (!x_zero) * raft::exp(input); }; if (x != y) { raft::linalg::unaryOp<DataT, decltype(unaryOp_lambda), IdxT>( (DataT*)y, y, n * k, unaryOp_lambda, stream); } const DataT* x_norm = nullptr; const DataT* y_norm = nullptr; // This op takes some shortcuts when x equals y. So its behavior changes based // on this. ops::kl_divergence_op<DataT, AccT, IdxT> distance_op{is_row_major, x == y}; pairwise_matrix_dispatch<decltype(distance_op), DataT, AccT, OutT, FinOpT, IdxT>( distance_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); if (x != y) { // Now reverse previous log (x) back to x using (e ^ log(x)) raft::linalg::unaryOp<DataT, decltype(unaryOp_lambda_reverse), IdxT>( (DataT*)y, y, n * k, unaryOp_lambda_reverse, stream); } } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::L1> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT*, // workspace unused size_t, // worksize unused FinOpT fin_op, bool is_row_major, DataT) // metric_arg unused { ops::l1_distance_op<DataT, AccT, IdxT> distance_op{}; const DataT* x_norm = nullptr; const DataT* y_norm = nullptr; cudaStream_t stream = raft::resource::get_cuda_stream(handle); pairwise_matrix_dispatch<decltype(distance_op), DataT, AccT, OutT, FinOpT, IdxT>( distance_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl_l2_expanded( // NOTE: different name bool perform_sqrt, // dispatch on sqrt const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT* workspace, size_t worksize, FinOpT fin_op, cudaStream_t stream, bool is_row_major) { // raft distance support inputs as float/double and output as uint8_t/float/double. static_assert(!((sizeof(OutT) > 1) && (sizeof(AccT) != sizeof(OutT))), "OutT can be uint8_t, float, double," "if sizeof(OutT) > 1 then sizeof(AccT) == sizeof(OutT)."); ASSERT(!(worksize < (m + n) * sizeof(AccT)), "workspace size error"); ASSERT(workspace != nullptr, "workspace is null"); DataT* x_norm = workspace; DataT* y_norm = workspace; // TODO: Column major case looks to have lower accuracy for X == Y, // perhaps the use of stridedSummationKernel could be causing this, // need to investigate and fix. if ((x == y) && is_row_major) { raft::linalg::rowNorm(x_norm, x, k, std::max(m, n), raft::linalg::L2Norm, is_row_major, stream, raft::identity_op{}); } else { y_norm += m; raft::linalg::rowNorm( x_norm, x, k, m, raft::linalg::L2Norm, is_row_major, stream, raft::identity_op{}); raft::linalg::rowNorm( y_norm, y, k, n, raft::linalg::L2Norm, is_row_major, stream, raft::identity_op{}); } ops::l2_exp_distance_op<DataT, AccT, IdxT> distance_op{perform_sqrt}; pairwise_matrix_dispatch<decltype(distance_op), DataT, AccT, OutT, FinOpT, IdxT>( distance_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::L2Expanded> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT* workspace, size_t worksize, FinOpT fin_op, bool is_row_major, DataT) // metric_arg unused { bool perform_sqrt = false; cudaStream_t stream = raft::resource::get_cuda_stream(handle); distance_impl_l2_expanded( perform_sqrt, x, y, out, m, n, k, workspace, worksize, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::L2SqrtExpanded> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT* workspace, size_t worksize, FinOpT fin_op, bool is_row_major, DataT) // metric_arg unused { bool perform_sqrt = true; cudaStream_t stream = raft::resource::get_cuda_stream(handle); distance_impl_l2_expanded( perform_sqrt, x, y, out, m, n, k, workspace, worksize, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::L2Unexpanded> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT*, // workspace unused size_t, // worksize unused FinOpT fin_op, bool is_row_major, DataT) // metric_arg unused { bool perform_sqrt = false; ops::l2_unexp_distance_op<DataT, AccT, IdxT> l2_op(perform_sqrt); // The unexpanded L2 does not require the norms of a and b to be calculated. const DataT* x_norm = nullptr; const DataT* y_norm = nullptr; cudaStream_t stream = raft::resource::get_cuda_stream(handle); pairwise_matrix_dispatch<decltype(l2_op), DataT, AccT, OutT, FinOpT, IdxT>( l2_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::L2SqrtUnexpanded> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT*, // workspace unused size_t, // worksize unused FinOpT fin_op, bool is_row_major, DataT) // metric_arg unused { bool perform_sqrt = true; ops::l2_unexp_distance_op<DataT, AccT, IdxT> l2_op(perform_sqrt); // The unexpanded L2 does not require the norms of a and b to be calculated. const DataT* x_norm = nullptr; const DataT* y_norm = nullptr; cudaStream_t stream = raft::resource::get_cuda_stream(handle); pairwise_matrix_dispatch<decltype(l2_op), DataT, AccT, OutT, FinOpT, IdxT>( l2_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::Linf> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT*, // workspace unused size_t, // worksize unused FinOpT fin_op, bool is_row_major, DataT) // metric_arg unused { ops::l_inf_distance_op<DataT, AccT, IdxT> distance_op{}; const DataT* x_norm = nullptr; const DataT* y_norm = nullptr; cudaStream_t stream = raft::resource::get_cuda_stream(handle); pairwise_matrix_dispatch<decltype(distance_op), DataT, AccT, OutT, FinOpT, IdxT>( distance_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::LpUnexpanded> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT*, // workspace unused size_t, // worksize unused FinOpT fin_op, bool is_row_major, DataT metric_arg) { ops::lp_unexp_distance_op<DataT, AccT, IdxT> distance_op{metric_arg}; const DataT* x_norm = nullptr; const DataT* y_norm = nullptr; cudaStream_t stream = raft::resource::get_cuda_stream(handle); pairwise_matrix_dispatch<decltype(distance_op), DataT, AccT, OutT, FinOpT, IdxT>( distance_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); } template <typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void distance_impl(raft::resources const& handle, distance_tag<DistanceType::RusselRaoExpanded> distance_type, const DataT* x, const DataT* y, OutT* out, IdxT m, IdxT n, IdxT k, AccT*, // workspace unused size_t, // worksize unused FinOpT fin_op, bool is_row_major, DataT) // metric_arg unused { ops::russel_rao_distance_op<DataT, AccT, IdxT> distance_op{k}; const DataT* x_norm = nullptr; const DataT* y_norm = nullptr; cudaStream_t stream = raft::resource::get_cuda_stream(handle); pairwise_matrix_dispatch<decltype(distance_op), DataT, AccT, OutT, FinOpT, IdxT>( distance_op, m, n, k, x, y, x_norm, y_norm, out, fin_op, stream, is_row_major); } /** * @brief Evaluate pairwise distances with the user epilogue lamba allowed * @tparam DistanceType which distance to evaluate * @tparam InType input argument type * @tparam AccType accumulation type * @tparam OutType output type * @tparam FinalLambda user-defined epilogue lamba * @tparam Index_ Index type * * @param x first set of points * @param y second set of points * @param out output distance matrix * @param m number of points in x * @param n number of points in y * @param k dimensionality * @param workspace temporary workspace needed for computations * @param worksize number of bytes of the workspace * @param fin_op the final gemm epilogue lambda * @param stream cuda stream * @param isRowMajor whether the matrices are row-major or col-major * * @note fin_op: This is a device lambda which is supposed to operate upon the * input which is AccType and returns the output in OutType. It's signature is * as follows: <pre>OutType fin_op(AccType in, int g_idx);</pre>. If one needs * any other parameters, feel free to pass them via closure. */ template <cuvs::distance::DistanceType distanceType, typename InType, typename AccType, typename OutType, typename FinalLambda, typename Index_ = int> void distance(raft::resources const& handle, const InType* x, const InType* y, OutType* out, Index_ m, Index_ n, Index_ k, void* workspace, size_t worksize, FinalLambda fin_op, bool isRowMajor = true, InType metric_arg = 2.0f) { // raft distance support inputs as float/double and output as uint8_t/float/double. static_assert(!((sizeof(OutType) > 1) && (sizeof(AccType) != sizeof(OutType))), "OutType can be uint8_t, float, double," "if sizeof(OutType) > 1 then sizeof(AccType) == sizeof(OutType)."); distance_impl<InType, AccType, OutType, FinalLambda, Index_>( handle, distance_tag<distanceType>{}, x, y, out, m, n, k, reinterpret_cast<AccType*>(workspace), worksize, fin_op, isRowMajor, metric_arg); RAFT_CUDA_TRY(cudaPeekAtLastError()); } /** * @brief Evaluate pairwise distances for the simple use case * @tparam DistanceType which distance to evaluate * @tparam InType input argument type * @tparam AccType accumulation type * @tparam OutType output type * @tparam Index_ Index type * @param x first set of points * @param y second set of points * @param dist output distance matrix * @param m number of points in x * @param n number of points in y * @param k dimensionality * @param workspace temporary workspace needed for computations * @param worksize number of bytes of the workspace * @param stream cuda stream * @param isRowMajor whether the matrices are row-major or col-major */ template <cuvs::distance::DistanceType distanceType, typename InType, typename AccType, typename OutType, typename Index_ = int> void distance(raft::resources const& handle, const InType* x, const InType* y, OutType* out, Index_ m, Index_ n, Index_ k, void* workspace, size_t worksize, bool isRowMajor = true, InType metric_arg = 2.0f) { auto fin_op = raft::identity_op(); distance<distanceType, InType, AccType, OutType, decltype(fin_op), Index_>( handle, x, y, out, m, n, k, workspace, worksize, fin_op, isRowMajor, metric_arg); } /** * @brief Return the exact workspace size to compute the distance * @tparam DistanceType which distance to evaluate * @tparam InType input argument type * @tparam AccType accumulation type * @tparam OutType output type * @tparam Index_ Index type * @param x first set of points * @param y second set of points * @param m number of points in x * @param n number of points in y * @param k dimensionality * * @note If the specified distanceType doesn't need the workspace at all, it * returns 0. */ template <cuvs::distance::DistanceType distanceType, typename InType, typename AccType, typename OutType, typename Index_ = int> size_t getWorkspaceSize(const InType* x, const InType* y, Index_ m, Index_ n, Index_ k) { size_t worksize = 0; constexpr bool is_allocated = (distanceType <= cuvs::distance::DistanceType::CosineExpanded) || (distanceType == cuvs::distance::DistanceType::CorrelationExpanded); constexpr int numOfBuffers = (distanceType == cuvs::distance::DistanceType::CorrelationExpanded) ? 2 : 1; if (is_allocated) { // TODO : when X == Y allocate std::max(m, n) instead of m + n when column major input // accuracy issue is resolved until then we allocate as m + n. worksize += numOfBuffers * m * sizeof(AccType); worksize += numOfBuffers * n * sizeof(AccType); } return worksize; } }; // namespace detail }; // namespace distance }; // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_distance_cutlass_base.cuh
/* * Copyright (c) 2018-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #pragma GCC diagnostic push #pragma GCC diagnostic ignored "-Wstrict-aliasing" #pragma GCC diagnostic ignored "-Wtautological-compare" // We define CUTLASS_NAMESPACE in case // RAFT cmake is not used #ifndef CUTLASS_NAMESPACE #define cutlass raft_cutlass #endif #include <rmm/device_uvector.hpp> #include <type_traits> #include <cutlass/cutlass.h> #include <cutlass/gemm/device/gemm.h> #include <cutlass/gemm/device/gemm_universal_adapter.h> #include <cutlass/layout/matrix.h> #include <cutlass/layout/tensor.h> #include <cutlass/matrix_coord.h> #include <cutlass/tensor_view.h> #include <cuvs/distance/detail/distance_ops/cutlass.cuh> #include <raft/util/cutlass_utils.cuh> #include "./pairwise_distance_epilogue_elementwise.h" #include "./pairwise_distance_gemm.h" namespace cuvs { namespace distance { namespace detail { template <typename DataT, typename AccT, typename OutT, typename IdxT, int VecLen, typename FinalLambda, typename OpT, bool isRowMajor> std::enable_if_t<ops::has_cutlass_op<OpT>::value> cutlassDistanceKernel(const DataT* x, const DataT* y, const DataT* xn, const DataT* yn, IdxT m, IdxT n, IdxT k, IdxT lda, IdxT ldb, IdxT ldd, OutT* dOutput, FinalLambda fin_op, OpT distance_op, cudaStream_t stream) { static_assert(!(std::is_same<OutT, bool>::value), "OutType bool is not supported use uint8_t instead"); auto dist_op = distance_op.get_cutlass_op(); using DistanceFn = decltype(dist_op); using EpilogueOutputOp = cutlass::epilogue::thread::PairwiseDistanceEpilogueElementwise<DataT, // ElementC_ AccT, // ElementAccumulator_ DataT, // ElementCompute_ AccT, // ElementZ_ OutT, // ElementT_ 1, // Elements per access 1 DistanceFn, FinalLambda>; constexpr int batch_count = 1; constexpr auto mode = cutlass::gemm::GemmUniversalMode::kGemm; typename EpilogueOutputOp::Params epilog_op_param(dist_op, fin_op); const DataT *a, *b; IdxT gemm_lda, gemm_ldb; // Number of pipelines you want to use constexpr int NumStages = 3; // Alignment constexpr int Alignment = VecLen; // default initialize problem size with row major inputs auto problem_size = cutlass::gemm::GemmCoord(n, m, k); using cutlassDistKernel = typename cutlass::gemm::kernel::PairwiseDistanceGemm<DataT, Alignment, DataT, Alignment, AccT, AccT, EpilogueOutputOp, NumStages, // Number of pipeline stages isRowMajor>::GemmKernel; using cutlassDist = cutlass::gemm::device::GemmUniversalAdapter<cutlassDistKernel>; if constexpr (isRowMajor) { a = y; b = x; gemm_lda = ldb; gemm_ldb = lda; } else { problem_size = cutlass::gemm::GemmCoord(m, n, k); a = x; b = y; gemm_lda = lda; gemm_ldb = ldb; } typename cutlassDist::Arguments arguments{ mode, problem_size, batch_count, epilog_op_param, a, b, xn, // C matrix eq vector param, which here is A norm nullptr, // tensor_Z, (DataT*)yn, // this is broadcast vec, which is required to be non-const param dOutput, // Output distance matrix (int64_t)0, // batch stride A (int64_t)0, // batch stride B (int64_t)0, // batch stride Norm A (int64_t)0, (int64_t)0, // batch stride Norm B (int64_t)0, // batch stride Output gemm_lda, // stride A gemm_ldb, // stride B 1, // stride A norm 0, // this is no-op for Z 0, // This must be zero ldd // stride Output matrix }; // Using the arguments, query for extra workspace required for matrix multiplication computation size_t workspace_size = cutlassDist::get_workspace_size(arguments); // Allocate workspace memory rmm::device_uvector<uint8_t> workspace(workspace_size, stream); // Instantiate CUTLASS kernel depending on templates cutlassDist cutlassDist_op; // Check the problem size is supported or not RAFT_CUTLASS_TRY(cutlassDist_op.can_implement(arguments)); // Initialize CUTLASS kernel with arguments and workspace pointer RAFT_CUTLASS_TRY(cutlassDist_op.initialize(arguments, workspace.data(), stream)); // Launch initialized CUTLASS kernel RAFT_CUTLASS_TRY(cutlassDist_op(stream)); } }; // namespace detail }; // namespace distance }; // namespace cuvs #pragma GCC diagnostic pop
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_distance_gemm.h
/* * Copyright (c) 2018-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cutlass/cutlass.h> #include <cutlass/gemm/kernel/default_gemm_universal.h> #include <cutlass/gemm/kernel/gemm_with_fused_epilogue.h> #include <cutlass/layout/matrix.h> #include <cutlass/layout/tensor.h> #include "./pairwise_distance_epilogue.h" ///////////////////////////////////////////////////////////////////////////////////////////////// namespace cutlass { namespace gemm { namespace kernel { ///////////////////////////////////////////////////////////////////////////////////////////////// template < /// Element type for A matrix operand typename ElementA_, /// Layout type for A matrix operand int kAlignmentA, /// Element type for B matrix operand typename ElementB_, /// Layout type for B matrix operand int kAlignmentB, /// Element type for C and D matrix operands typename ElementC_, /// Element type for internal accumulation typename ElementAccumulator, /// Element type for final output // typename ElementOutT, /// Epilogue output operator - must satisfy concept of 'EpilogueWithBroadcastOp' typename EpilogueOutputOp, /// Number of stages used in the pipelined mainloop int Stages, /// data layout row/column major of inputs bool isRowMajor> struct PairwiseDistanceGemm { // This struct is specialized for fp32/3xTF32 /// Threadblock-level tile size (concept: GemmShape) using ThreadblockShape = cutlass::gemm::GemmShape<128, 128, 16>; // <- threadblock tile M = 128, N = 128, K = 16 /// Warp-level tile size (concept: GemmShape) // This code section describes tile size a warp will compute using WarpShape = cutlass::gemm::GemmShape<64, 64, 16>; // <- warp tile M = 64, N = 64, K = 16 /// Warp-level tile size (concept: GemmShape) // This code section describes the size of MMA op using InstructionShape = cutlass::gemm::GemmShape<16, 8, 4>; // <- MMA Op tile M = 16, N = 8, K = 4 /// Operation performed by GEMM using Operator = cutlass::arch::OpMultiplyAddFastF32; // This code section describes whether you want to use tensor cores or regular SIMT cores on GPU // SM using OperatorClass = cutlass::arch::OpClassTensorOp; // This code section describes CUDA SM architecture number using ArchTag = cutlass::arch::Sm80; // This code section describes how threadblocks are scheduled on GPU /// Threadblock-level swizzling operator using ThreadblockSwizzle = cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>; /// data layout for final output matrix. // we keep this same layout even for column major inputs using LayoutOutput = cutlass::layout::RowMajor; typedef typename std::conditional<isRowMajor, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::type NormXLayout; typedef typename std:: conditional<isRowMajor, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::type LayoutA_; typedef typename std:: conditional<isRowMajor, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::type LayoutB_; using GemmBase = typename DefaultGemmUniversal<ElementA_, LayoutA_, cutlass::ComplexTransform::kNone, kAlignmentA, ElementB_, LayoutB_, cutlass::ComplexTransform::kNone, kAlignmentB, ElementC_, LayoutOutput, ElementAccumulator, OperatorClass, ArchTag, ThreadblockShape, WarpShape, InstructionShape, EpilogueOutputOp, ThreadblockSwizzle, Stages, Operator>::GemmKernel; // Replace epilogue using Epilogue = typename cutlass::epilogue::threadblock::PairwiseDistanceEpilogue< typename GemmBase::Epilogue::Shape, typename GemmBase::Epilogue::WarpMmaOperator, GemmBase::Epilogue::kPartitionsK, ElementAccumulator, typename EpilogueOutputOp::ElementT, ElementAccumulator, EpilogueOutputOp, NormXLayout, GemmBase::Epilogue::kElementsPerAccess>::Epilogue; // Compose the GEMM kernel using GemmKernel = GemmWithFusedEpilogue<typename GemmBase::Mma, Epilogue, ThreadblockSwizzle>; }; template < /// Layout type for A matrix operand int kAlignmentA, /// Layout type for B matrix operand int kAlignmentB, /// Element type for C and D matrix operands typename ElementC_, /// Element type for internal accumulation typename ElementAccumulator, /// Epilogue output operator - must satisfy concept of 'EpilogueWithBroadcastOp' typename EpilogueOutputOp, /// Number of stages used in the pipelined mainloop int Stages, /// data layout row/column major of inputs bool isRowMajor> struct PairwiseDistanceGemm<double, kAlignmentA, double, kAlignmentB, ElementC_, ElementAccumulator, EpilogueOutputOp, Stages, isRowMajor> { // using Transform = cutlass::ComplexTransform::kNone; // Threadblock-level tile size (concept: GemmShape) using ThreadblockShape = cutlass::gemm::GemmShape<64, 64, 16>; // <- threadblock tile M = 64, N = 64, K = 16 /// Warp-level tile size (concept: GemmShape) // This code section describes tile size a warp will compute using WarpShape = cutlass::gemm::GemmShape<32, 32, 16>; // <- warp tile M = 32, N = 32, K = 16 /// Warp-level tile size (concept: GemmShape) // This code section describes the size of MMA op using InstructionShape = cutlass::gemm::GemmShape<8, 8, 4>; // Operation performed by GEMM using Operator = cutlass::arch::OpMultiplyAdd; // This code section describes whether you want to use tensor cores or regular SIMT cores on GPU // SM using OperatorClass = cutlass::arch::OpClassTensorOp; // This code section describes CUDA SM architecture number using ArchTag = cutlass::arch::Sm80; // This code section describes how threadblocks are scheduled on GPU /// Threadblock-level swizzling operator using ThreadblockSwizzle = cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>; /// data layout for final output matrix. // we keep this same layout even for column major inputs using LayoutOutput = cutlass::layout::RowMajor; typedef typename std::conditional<isRowMajor, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::type NormXLayout; typedef typename std:: conditional<isRowMajor, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::type LayoutA_; typedef typename std:: conditional<isRowMajor, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::type LayoutB_; using GemmBase = typename DefaultGemmUniversal<double, LayoutA_, cutlass::ComplexTransform::kNone, 1, double, LayoutB_, cutlass::ComplexTransform::kNone, 1, ElementC_, LayoutOutput, ElementAccumulator, OperatorClass, ArchTag, ThreadblockShape, WarpShape, InstructionShape, EpilogueOutputOp, ThreadblockSwizzle, Stages, Operator>::GemmKernel; // Replace epilogue using Epilogue = typename cutlass::epilogue::threadblock::PairwiseDistanceEpilogue< typename GemmBase::Epilogue::Shape, typename GemmBase::Epilogue::WarpMmaOperator, GemmBase::Epilogue::kPartitionsK, ElementC_, typename EpilogueOutputOp::ElementT, ElementC_, EpilogueOutputOp, NormXLayout, GemmBase::Epilogue::kElementsPerAccess>::Epilogue; // Compose the GEMM kernel using GemmKernel = GemmWithFusedEpilogue<typename GemmBase::Mma, Epilogue, ThreadblockSwizzle>; }; ///////////////////////////////////////////////////////////////////////////////////////////////// } // namespace kernel } // namespace gemm } // namespace cutlass
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/fused_l2_nn.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cstddef> // size_t #include <cuvs/distance/detail/distance_ops/l2_exp.cuh> // ops::l2_exp_distance_op #include <cuvs/distance/detail/fused_distance_nn/cutlass_base.cuh> #include <cuvs/distance/detail/pairwise_distance_base.cuh> // PairwiseDistances #include <limits> // std::numeric_limits #include <raft/core/kvp.hpp> // raft::KeyValuePair #include <raft/core/operators.hpp> // raft::identity_op #include <raft/linalg/contractions.cuh> // Policy #include <raft/util/arch.cuh> // raft::util::arch::SM_* #include <raft/util/cuda_utils.cuh> // raft::ceildiv, raft::shfl namespace cuvs { namespace distance { namespace detail { template <typename LabelT, typename DataT> struct KVPMinReduceImpl { typedef raft::KeyValuePair<LabelT, DataT> KVP; DI KVP operator()(LabelT rit, const KVP& a, const KVP& b) { return b.value < a.value ? b : a; } DI KVP operator()(const KVP& a, const KVP& b) { return b.value < a.value ? b : a; } }; // KVPMinReduce template <typename LabelT, typename DataT> struct MinAndDistanceReduceOpImpl { typedef typename raft::KeyValuePair<LabelT, DataT> KVP; DI void operator()(LabelT rid, KVP* out, const KVP& other) const { if (other.value < out->value) { out->key = other.key; out->value = other.value; } } DI void operator()(LabelT rid, DataT* out, const KVP& other) const { if (other.value < *out) { *out = other.value; } } DI void operator()(LabelT rid, DataT* out, const DataT& other) const { if (other < *out) { *out = other; } } DI void init(DataT* out, DataT maxVal) const { *out = maxVal; } DI void init(KVP* out, DataT maxVal) const { out->value = maxVal; } DI void init_key(DataT& out, LabelT idx) const { return; } DI void init_key(KVP& out, LabelT idx) const { out.key = idx; } DI DataT get_value(KVP& out) const { return out.value; ; } DI DataT get_value(DataT& out) const { return out; } }; template <typename LabelT, typename DataT> struct MinReduceOpImpl { typedef typename raft::KeyValuePair<LabelT, DataT> KVP; DI void operator()(LabelT rid, DataT* out, const KVP& other) { if (other.value < *out) { *out = other.value; } } DI void init(DataT* out, DataT maxVal) { *out = maxVal; } }; template <typename DataT, typename OutT, typename IdxT, typename ReduceOpT> RAFT_KERNEL initKernel(OutT* min, IdxT m, DataT maxVal, ReduceOpT redOp) { auto tid = IdxT(blockIdx.x) * blockDim.x + threadIdx.x; if (tid < m) { redOp.init(min + tid, maxVal); } } template <typename DataT, typename OutT, typename IdxT, typename ReduceOpT> void initialize(OutT* min, IdxT m, DataT maxVal, ReduceOpT redOp, cudaStream_t stream) { auto blks = raft::ceildiv(m, 256); initKernel<DataT, OutT, IdxT><<<blks, 256, 0, stream>>>(min, m, maxVal, redOp); } // TODO: specialize this function for MinAndDistanceReduceOp<int, float> // with atomicCAS of 64 bit which will eliminate mutex and raft::shfls template <typename P, typename OutT, typename IdxT, typename KVPair, typename ReduceOpT> DI void updateReducedVal( int* mutex, OutT* min, KVPair* val, ReduceOpT red_op, IdxT m, IdxT gridStrideY) { const auto lid = threadIdx.x % raft::WarpSize; const auto accrowid = threadIdx.x / P::AccThCols; // Update each output row in order within a warp. This will resolve hang // issues with pre-Volta architectures #pragma unroll for (int j = 0; j < (raft::WarpSize / P::AccThCols); j++) { if (lid == j * P::AccThCols) { #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { auto rid = gridStrideY + accrowid + i * P::AccThRows; if (rid < m) { auto value = val[i]; while (atomicCAS(mutex + rid, 0, 1) == 1) ; __threadfence(); red_op(rid, min + rid, value); __threadfence(); atomicCAS(mutex + rid, 1, 0); } } } } } template <typename DataT, typename OutT, typename IdxT, typename P, typename ReduceOpT, typename KVPReduceOpT, typename OpT, typename FinalLambda> __launch_bounds__(P::Nthreads, 2) RAFT_KERNEL fusedL2NNkernel(OutT* min, const DataT* x, const DataT* y, const DataT* xn, const DataT* yn, IdxT m, IdxT n, IdxT k, DataT maxVal, int* mutex, ReduceOpT redOp, KVPReduceOpT pairRedOp, OpT distance_op, FinalLambda fin_op) { // compile only if below non-ampere arch. #if __CUDA_ARCH__ < 800 extern __shared__ char smem[]; typedef KeyValuePair<IdxT, DataT> KVPair; KVPair val[P::AccRowsPerTh]; #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { val[i] = {0, maxVal}; } // epilogue operation lambda for final value calculation auto epilog_lambda = [n, pairRedOp, &val, maxVal] __device__( DataT acc[P::AccRowsPerTh][P::AccColsPerTh], DataT * regxn, DataT * regyn, IdxT gridStrideX, IdxT gridStrideY) { KVPReduceOpT pairRed_op(pairRedOp); // intra thread reduce const auto acccolid = threadIdx.x % P::AccThCols; const auto accrowid = threadIdx.x / P::AccThCols; #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { auto tmpkey = acccolid + j * P::AccThCols + gridStrideX; KVPair tmp = {tmpkey, acc[i][j]}; if (tmpkey < n) { val[i] = pairRed_op(accrowid + i * P::AccThRows + gridStrideY, tmp, val[i]); } } } }; auto rowEpilog_lambda = [m, mutex, min, pairRedOp, redOp, &val, maxVal] __device__(IdxT gridStrideY) { KVPReduceOpT pairRed_op(pairRedOp); ReduceOpT red_op(redOp); const auto accrowid = threadIdx.x / P::AccThCols; const auto lid = raft::laneId(); // reduce #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { #pragma unroll for (int j = P::AccThCols / 2; j > 0; j >>= 1) { // Actually, the srcLane (lid +j) should be (lid +j) % P:AccThCols, // but the raft::shfl op applies the modulo internally. auto tmpkey = raft::shfl(val[i].key, lid + j, P::AccThCols); auto tmpvalue = raft::shfl(val[i].value, lid + j, P::AccThCols); KVPair tmp = {tmpkey, tmpvalue}; val[i] = pairRed_op(accrowid + i * P::AccThRows + gridStrideY, tmp, val[i]); } } updateReducedVal<P, OutT, IdxT, KVPair, ReduceOpT>(mutex, min, val, red_op, m, gridStrideY); // reset the val array. #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { val[i] = {0, maxVal}; } }; IdxT lda = k, ldb = k, ldd = n; constexpr bool row_major = true; constexpr bool write_out = false; PairwiseDistances<DataT, DataT, // OutT (unused in PairwiseDistances) IdxT, P, decltype(distance_op), decltype(epilog_lambda), FinalLambda, decltype(rowEpilog_lambda), row_major, write_out> obj(x, y, m, n, k, lda, ldb, ldd, xn, yn, nullptr, // Output pointer smem, distance_op, epilog_lambda, fin_op, rowEpilog_lambda); obj.run(); #endif } // cg::reduce functor for FusedDistanceNN used in its cutlass version // to output the min distance value & key(loc id). // This is used in fused_distance_nn/predicated_tile_iterator_reduced_vec.h // store_with_byte_offset() passed to cg::reduce() & select_reduce. template <typename AccType, typename Index, typename OutType> struct kvp_cg_min_reduce_op { typedef typename raft::KeyValuePair<Index, AccType> KVP; __host__ __device__ kvp_cg_min_reduce_op() noexcept {}; using AccTypeT = AccType; using IndexT = Index; // functor signature. __host__ __device__ KVP operator()(KVP a, KVP b) const { return a.value < b.value ? a : b; } __host__ __device__ AccType operator()(AccType a, AccType b) const { return min(a, b); } __host__ __device__ bool isAmin(AccType a, AccType b) const { return a < b ? true : false; } }; template <typename DataT, typename OutT, typename IdxT, typename Policy, typename ReduceOpT, typename KVPReduceOpT> void fusedL2NNImpl(OutT* min, const DataT* x, const DataT* y, const DataT* xn, const DataT* yn, IdxT m, IdxT n, IdxT k, int* workspace, ReduceOpT redOp, KVPReduceOpT pairRedOp, bool sqrt, bool initOutBuffer, cudaStream_t stream) { // The kernel policy is determined by fusedL2NN. typedef Policy P; dim3 blk(P::Nthreads); auto nblks = raft::ceildiv<int>(m, P::Nthreads); constexpr auto maxVal = std::numeric_limits<DataT>::max(); typedef raft::KeyValuePair<IdxT, DataT> KVPair; RAFT_CUDA_TRY(cudaMemsetAsync(workspace, 0, sizeof(int) * m, stream)); if (initOutBuffer) { initKernel<DataT, OutT, IdxT, ReduceOpT> <<<nblks, P::Nthreads, 0, stream>>>(min, m, maxVal, redOp); RAFT_CUDA_TRY(cudaGetLastError()); } namespace arch = raft::util::arch; using AccT = DataT; ops::l2_exp_distance_op<DataT, AccT, IdxT> distance_op{sqrt}; raft::identity_op fin_op{}; auto kernel = fusedL2NNkernel<DataT, OutT, IdxT, P, ReduceOpT, KVPReduceOpT, decltype(distance_op), decltype(fin_op)>; // Get pointer to fp32 SIMT kernel to determine the best compute architecture // out of all for which the kernel was compiled for that matches closely // to the current device. Other methods to determine the architecture (that do not // require a pointer) can be error prone. See: // https://github.com/NVIDIA/cub/issues/545 void* kernel_ptr = reinterpret_cast<void*>(kernel); auto runtime_arch = arch::kernel_virtual_arch(kernel_ptr); auto cutlass_range = arch::SM_range(arch::SM_80(), arch::SM_future()); if (cutlass_range.contains(runtime_arch)) { // If device is SM_80 or later, use CUTLASS-based kernel. using L2Op = cuvs::distance::detail::ops::l2_exp_cutlass_op<DataT, DataT>; using kvp_cg_min_reduce_op_ = kvp_cg_min_reduce_op<DataT, IdxT, OutT>; kvp_cg_min_reduce_op_ cg_reduce_op; L2Op L2_dist_op(sqrt); IdxT lda, ldb, ldd; lda = k, ldb = k, ldd = n; cutlassFusedDistanceNN<DataT, DataT, OutT, IdxT, P::Veclen, kvp_cg_min_reduce_op_, L2Op, ReduceOpT, KVPReduceOpT>(x, y, xn, yn, m, n, k, lda, ldb, ldd, min, workspace, cg_reduce_op, L2_dist_op, redOp, pairRedOp, stream); } else { // If device less than SM_80, use fp32 SIMT kernel. constexpr size_t shmemSize = P::SmemSize + ((P::Mblk + P::Nblk) * sizeof(DataT)); dim3 grid = launchConfigGenerator<P>(m, n, shmemSize, kernel); kernel<<<grid, blk, shmemSize, stream>>>( min, x, y, xn, yn, m, n, k, maxVal, workspace, redOp, pairRedOp, distance_op, fin_op); RAFT_CUDA_TRY(cudaGetLastError()); } } } // namespace detail } // namespace distance } // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/masked_nn.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <limits> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resource/device_memory_resource.hpp> #include <stdint.h> #include <cuvs/distance/detail/compress_to_bits.cuh> #include <cuvs/distance/detail/fused_l2_nn.cuh> #include <cuvs/distance/detail/masked_distance_base.cuh> #include <raft/linalg/contractions.cuh> #include <raft/util/cuda_utils.cuh> #include <rmm/device_uvector.hpp> namespace cuvs { namespace distance { namespace detail { template <typename DataT, typename OutT, typename IdxT, typename P, typename ReduceOpT, typename KVPReduceOpT, typename CoreLambda, typename FinalLambda> __launch_bounds__(P::Nthreads, 2) RAFT_KERNEL masked_l2_nn_kernel(OutT* min, const DataT* x, const DataT* y, const DataT* xn, const DataT* yn, const uint64_t* adj, const IdxT* group_idxs, IdxT num_groups, IdxT m, IdxT n, IdxT k, bool sqrt, DataT maxVal, int* mutex, ReduceOpT redOp, KVPReduceOpT pairRedOp, CoreLambda core_op, FinalLambda fin_op) { extern __shared__ char smem[]; typedef raft::KeyValuePair<IdxT, DataT> KVPair; KVPair val[P::AccRowsPerTh]; #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { val[i] = {-1, maxVal}; } // epilogue operation lambda for final value calculation auto epilog_lambda = [pairRedOp, &val, maxVal, sqrt] __device__( DataT acc[P::AccRowsPerTh][P::AccColsPerTh], int thread_adj, DataT* regxn, DataT* regyn, IdxT tile_idx_n, IdxT tile_idx_m, IdxT tile_end_n) { KVPReduceOpT pairRed_op(pairRedOp); #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { acc[i][j] = regxn[i] + regyn[j] - (DataT)2.0 * acc[i][j]; } } if (sqrt) { #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { acc[i][j] = raft::sqrt(acc[i][j]); } } } // intra thread reduce const auto acccolid = threadIdx.x % P::AccThCols; const auto accrowid = threadIdx.x / P::AccThCols; #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { // thread_adj is a bitfield that contains a 1 at location i iff we must // compute row i of acc (the accumulator register tile). It is described in // more detail in the maskedDistances.run() method. const bool ignore = (thread_adj & (1 << i)) == 0; if (ignore) { continue; } #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { auto tmpkey = acccolid + j * P::AccThCols + tile_idx_n; if (tile_end_n <= tmpkey) { // Do not process beyond end of tile. continue; } KVPair tmp = {tmpkey, acc[i][j]}; if (tmpkey < tile_end_n) { val[i] = pairRed_op(accrowid + i * P::AccThRows + tile_idx_m, tmp, val[i]); } } } }; auto rowEpilog_lambda = [m, mutex, min, pairRedOp, redOp, &val, maxVal] __device__(IdxT tile_idx_m) { KVPReduceOpT pairRed_op(pairRedOp); ReduceOpT red_op(redOp); const auto accrowid = threadIdx.x / P::AccThCols; const auto lid = raft::laneId(); // reduce #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { #pragma unroll for (int j = P::AccThCols / 2; j > 0; j >>= 1) { auto tmpkey = raft::shfl(val[i].key, lid + j); auto tmpvalue = raft::shfl(val[i].value, lid + j); KVPair tmp = {tmpkey, tmpvalue}; val[i] = pairRed_op(accrowid + i * P::AccThRows + tile_idx_m, tmp, val[i]); } } updateReducedVal<P, OutT, IdxT, KVPair, ReduceOpT>(mutex, min, val, red_op, m, tile_idx_m); // reset the val array. #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { val[i] = {-1, maxVal}; } }; IdxT lda = k, ldb = k, ldd = n; MaskedDistances<true, DataT, DataT, IdxT, P, CoreLambda, decltype(epilog_lambda), FinalLambda, decltype(rowEpilog_lambda), true> obj(x, y, m, n, k, lda, ldb, ldd, xn, yn, adj, group_idxs, num_groups, smem, core_op, epilog_lambda, fin_op, rowEpilog_lambda); obj.run(); } /** * @brief Wrapper for masked_l2_nn_kernel * * Responsibilities: * - Allocate (and initialize) workspace memory for: * - mutexes used in nearest neighbor update step * - adjacency matrix bitfield * - Compress adjacency matrix to bitfield * - Initialize output buffer (conditional on `initOutBuffer`) * - Specify core and final operations for the L2 norm * - Determine optimal launch configuration for kernel. * - Launch kernel and check for errors. * * @tparam DataT Input data-type (for x and y matrices). * @tparam OutT Output data-type (for key-value pairs). * @tparam IdxT Index data-type. * @tparam ReduceOpT A struct to perform the final needed reduction * operation and also to initialize the output array * elements with the appropriate initial value needed for * reduction. * @tparam KVPReduceOpT Type of Reduction operation on key value pairs. * * @param handle RAFT handle for managing expensive resources * @param[out] out Will contain reduced output (nn key-value pairs) * @param[in] x First matrix. Row major. Dim = `m x k`. (on device) * @param[in] y Second matrix. Row major. Dim = `n x k`. (on device) * @param[in] xn L2 squared norm of `x`. Length = `m`. * @param[in] yn L2 squared norm of `y`. Length = `n`. * @param[in] adj A boolean adjacency matrix indicating for each * row of `x` and each group in `y` whether to compute the * distance. Dim = `m x num_groups`. * @param[in] group_idxs An array containing the *end* indices of each group * in `y`. The value of group_idxs[j] indicates the * start of group j + 1, i.e., it is the inclusive * scan of the group lengths. The first group is * always assumed to start at index 0 and the last * group typically ends at index `n`. Length = * `num_groups`. * @param[in] num_groups Length of `group_idxs`. * @param m Rows of `x`. * @param n Rows of `y`. * @param k Cols of `x` and `y`. * @param redOp Reduction operator in the epilogue * @param pairRedOp Reduction operation on key value pairs * @param sqrt Whether to compute the squared or actual (i.e. sqrt) L2 norm. * @param initOutBuffer Whether to initialize the output buffer * * */ template <typename DataT, typename OutT, typename IdxT, typename ReduceOpT, typename KVPReduceOpT> void masked_l2_nn_impl(raft::resources const& handle, OutT* out, const DataT* x, const DataT* y, const DataT* xn, const DataT* yn, const bool* adj, const IdxT* group_idxs, IdxT num_groups, IdxT m, IdxT n, IdxT k, ReduceOpT redOp, KVPReduceOpT pairRedOp, bool sqrt, bool initOutBuffer) { typedef typename linalg::Policy4x4<DataT, 1>::Policy P; static_assert(P::Mblk == 64, "masked_l2_nn_impl only supports a policy with 64 rows per block."); // Get stream and workspace memory resource rmm::mr::device_memory_resource* ws_mr = dynamic_cast<rmm::mr::device_memory_resource*>(raft::resource::get_workspace_resource(handle)); auto stream = resource::get_cuda_stream(handle); // Acquire temporary buffers and initialize to zero: // 1) Adjacency matrix bitfield // 2) Workspace for fused nearest neighbor operation size_t m_div_64 = raft::ceildiv(m, IdxT(64)); rmm::device_uvector<uint64_t> ws_adj64{m_div_64 * num_groups, stream, ws_mr}; rmm::device_uvector<int> ws_fused_nn{size_t(m), stream, ws_mr}; RAFT_CUDA_TRY(cudaMemsetAsync(ws_adj64.data(), 0, ws_adj64.size() * sizeof(uint64_t), stream)); RAFT_CUDA_TRY(cudaMemsetAsync(ws_fused_nn.data(), 0, ws_fused_nn.size() * sizeof(int), stream)); // Compress boolean adjacency matrix to bitfield. auto adj_view = raft::make_device_matrix_view<const bool, int>(adj, m, num_groups); auto adj64_view = raft::make_device_matrix_view<uint64_t, int>(ws_adj64.data(), m_div_64, num_groups); compress_to_bits(handle, adj_view, adj64_view); // Initialize output buffer with keyvalue pairs as determined by the reduction // operator (it will be called with maxVal). constexpr auto maxVal = std::numeric_limits<DataT>::max(); if (initOutBuffer) { dim3 grid(raft::ceildiv<int>(m, P::Nthreads)); dim3 block(P::Nthreads); initKernel<DataT, OutT, IdxT, ReduceOpT><<<grid, block, 0, stream>>>(out, m, maxVal, redOp); RAFT_CUDA_TRY(cudaGetLastError()); } // Accumulation operation lambda auto core_lambda = [] __device__(DataT & acc, DataT & x, DataT & y) { acc += x * y; }; auto fin_op = raft::identity_op{}; auto kernel = masked_l2_nn_kernel<DataT, OutT, IdxT, P, ReduceOpT, KVPReduceOpT, decltype(core_lambda), decltype(fin_op)>; constexpr size_t smemSize = P::SmemSize + ((P::Mblk + P::Nblk) * sizeof(DataT)); dim3 block(P::Nthreads); dim3 grid = launchConfigGenerator<P>(m, n, smemSize, kernel); kernel<<<grid, block, smemSize, stream>>>(out, x, y, xn, yn, ws_adj64.data(), group_idxs, num_groups, m, n, k, sqrt, maxVal, ws_fused_nn.data(), redOp, pairRedOp, core_lambda, fin_op); RAFT_CUDA_TRY(cudaGetLastError()); } } // namespace detail } // namespace distance } // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_distance_epilogue_elementwise.h
/* * Copyright (c) 2018-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ // /*! \file \brief Functor performing distance operations used by epilogues of pairwise distance * kernels. * This is adapted from LinearCombinationBiasElementwise from CUTLASS 2.9.0 * customized for applying elementwise distance formula on accumulated GEMM value * and applying user-defined final custom operation on the distance value. */ #pragma once #include <cutlass/array.h> #include <cutlass/cutlass.h> #include <cutlass/functional.h> #include <cutlass/numeric_conversion.h> #include <cutlass/numeric_types.h> #include <cutlass/epilogue/thread/activation.h> ///////////////////////////////////////////////////////////////////////////////////////////////// namespace cutlass { namespace epilogue { namespace thread { ///////////////////////////////////////////////////////////////////////////////////////////////// /// This base class is meant to define the concept required of the /// EpilogueWithBroadcast::OutputOp template <typename ElementC_, typename ElementAccumulator_, typename ElementCompute_, typename ElementZ_, typename ElementT_, int ElementsPerAccess, typename DistanceOp_, typename FinalOp_> class PairwiseDistanceEpilogueElementwise { public: using ElementOutput = ElementC_; using ElementC = ElementC_; using ElementAccumulator = ElementAccumulator_; using ElementCompute = ElementCompute_; using ElementZ = ElementZ_; using ElementT = ElementT_; static int const kElementsPerAccess = ElementsPerAccess; static int const kCount = kElementsPerAccess; using DistanceOp = DistanceOp_; using FinalOp = FinalOp_; using FragmentAccumulator = Array<ElementAccumulator, kElementsPerAccess>; using FragmentCompute = Array<ElementCompute, kElementsPerAccess>; using FragmentC = Array<ElementOutput, kElementsPerAccess>; using FragmentZ = Array<ElementZ, kElementsPerAccess>; using FragmentT = Array<ElementT, kElementsPerAccess>; using FragmentOutput = FragmentZ; static bool const kIsHeavy = false; // ElementwiseOp::kIsHeavy; /// If true, the 'Z' tensor is stored static bool const kStoreZ = false; // We don't store anything in Z, /// If true, the 'T' tensor is stored static bool const kStoreT = true; // this is our final output storage. /// Host-constructable parameters structure struct Params { FinalOp_ final_op_; DistanceOp_ dist_op_; // // Methods // CUTLASS_HOST_DEVICE Params(DistanceOp_ dist_op, FinalOp final_op) : final_op_(final_op), dist_op_(dist_op) {} CUTLASS_HOST_DEVICE Params() {} }; private: // // Data members // FinalOp_ final_op; DistanceOp_ elementwise_op; public: // // Methods // /// Constructor from Params CUTLASS_HOST_DEVICE PairwiseDistanceEpilogueElementwise(Params const& params) : final_op(params.final_op_), elementwise_op(params.dist_op_) { } /// Returns true if source is needed CUTLASS_HOST_DEVICE bool is_source_needed() const { // we use for making sure C matrix path is used for A mat norm. return true; } /// Functionally required for serial reduction in the epilogue CUTLASS_HOST_DEVICE void set_k_partition(int k_partition, int k_partition_count) {} /// Applies the operation when is_source_needed() is true CUTLASS_HOST_DEVICE void operator()(FragmentZ& frag_Z, FragmentT& frag_T, FragmentAccumulator const& AB, FragmentC const& frag_C, FragmentCompute const& V) const { FragmentCompute tmp_Accum = NumericArrayConverter<ElementCompute, ElementAccumulator, kElementsPerAccess>()(AB); FragmentCompute tmp_C = NumericArrayConverter<ElementCompute, ElementC, kElementsPerAccess>()(frag_C); FragmentCompute result_Z; FragmentCompute result_T; CUTLASS_PRAGMA_UNROLL for (int i = 0; i < kElementsPerAccess; ++i) { result_Z[i] = elementwise_op(tmp_C[i], V[i], tmp_Accum[i]); result_T[i] = final_op(result_Z[i], 0); } NumericArrayConverter<ElementT, ElementCompute, kElementsPerAccess> convert_t; frag_T = convert_t(result_T); } /// Applies the operation when is_source_needed() is false CUTLASS_HOST_DEVICE void operator()(FragmentZ& frag_Z, FragmentT& frag_T, FragmentAccumulator const& AB, FragmentCompute const& V) const { } }; ///////////////////////////////////////////////////////////////////////////////////////////////// } // namespace thread } // namespace epilogue } // namespace cutlass /////////////////////////////////////////////////////////////////////////////////////////////////
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/predicated_tile_iterator_normvec.h
/* * Copyright (c) 2018-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ /*! \file \brief Epilogue for threadblock scoped GEMMs using Tensor Ops. This file contains a customized version of PredicatedTileIterator from CUTLASS 2.9.0 (https://github.com/NVIDIA/cutlass/blob/v2.9.0/include/cutlass/epilogue/threadblock/predicated_tile_iterator.h#L75) Changes: - added `Layout_` template param - Only the row index is used to load the data in load_with_byte_offset(). This way the same normalization data is used across all columns in a row. */ #pragma once #include <cutlass/arch/arch.h> #include <cutlass/arch/memory.h> #include <cutlass/array.h> #include <cutlass/cutlass.h> #include <cutlass/epilogue/threadblock/output_tile_thread_map.h> #include <cutlass/epilogue/threadblock/predicated_tile_iterator_params.h> #include <cutlass/layout/matrix.h> #include <cutlass/layout/tensor.h> #include <cutlass/matrix_shape.h> #include <cutlass/numeric_types.h> #include <cutlass/tensor_ref.h> #include <cutlass/transform/pitch_linear_thread_map.h> //////////////////////////////////////////////////////////////////////////////// namespace cutlass { //////////////////////////////////////////////////////////////////////////////// namespace epilogue { namespace threadblock { //////////////////////////////////////////////////////////////////////////////// /// Tile iterator used to load and store output tile from global memory in epilogue. /// /// Satisfies: ReadableTileIterator | PredicatedTileIterator | ForwardTileIterator /// template <typename ThreadMap_, ///< Thread map (conept: OutputTileThreadMap) typename Element_, ///< Element data type typename Layout_, bool ScatterD = false, ///< Scatter D operand or not bool UseCUDAStore = false> class PredicatedTileIteratorNormVec { public: using ThreadMap = ThreadMap_; using Shape = typename ThreadMap::Shape; using Element = Element_; using Layout = Layout_; using TensorRef = TensorRef<Element, Layout>; using ConstTensorRef = typename TensorRef::ConstTensorRef; using Index = typename Layout::Index; using LongIndex = typename Layout::LongIndex; using TensorCoord = MatrixCoord; static int const kElementsPerAccess = ThreadMap::kElementsPerAccess; static int const kThreads = ThreadMap::kThreads; static int const kIterations = ThreadMap::Count::kTile; static_assert(ThreadMap::Iterations::kRow > 0, "ThreadMap::Iterations::kRow must be > 0"); static_assert(ThreadMap::Iterations::kGroup > 0, "ThreadMap::Iterations::kGroup must be > 0"); static_assert(ThreadMap::Iterations::kCluster > 0, "ThreadMap::Iterations::kCluster must be > 0"); static_assert(ThreadMap::Iterations::kColumn > 0, "ThreadMap::Iterations::kColumn must be > 0"); /// Fragment object using Fragment = Array<Element, ThreadMap::Iterations::kColumn * ThreadMap::Iterations::kRow * ThreadMap::Iterations::kGroup * ThreadMap::Iterations::kCluster * ThreadMap::kElementsPerAccess>; /// Memory access size using AccessType = AlignedArray<Element, ThreadMap::kElementsPerAccess>; // // Parameters struct // /// Uses a non-template class struct Params : PredicatedTileIteratorParams { using Base = PredicatedTileIteratorParams; CUTLASS_HOST_DEVICE Params() {} CUTLASS_HOST_DEVICE Params(Layout const& layout) : PredicatedTileIteratorParams( layout.stride(0) * int(sizeof(AccessType)) / kElementsPerAccess, make_OutputTileThreadMapDesc<ThreadMap>()) { } CUTLASS_HOST_DEVICE Params(Base const& base) : Base(base) {} }; /// Mask object struct Mask { static int const kCount = ThreadMap::Iterations::kColumn; /// Predicate state bool predicates[kCount]; // // Mask // CUTLASS_HOST_DEVICE Mask() { enable(); } ///< Efficiently disables all accesses guarded by mask CUTLASS_HOST_DEVICE void clear() { CUTLASS_PRAGMA_UNROLL for (int i = 0; i < kCount; ++i) { predicates[i] = false; } } ///< CUTLASS_HOST_DEVICE enables all accesses guarded by mask CUTLASS_DEVICE void enable() { CUTLASS_PRAGMA_UNROLL for (int i = 0; i < kCount; ++i) { predicates[i] = true; } } }; private: // // Data members // /// Parameters structure containing reference and precomputed state. PredicatedTileIteratorParams params_; /// Byte-level pointer uint8_t* byte_pointer_; /// Array of boolean values to contain steady-state predicates Mask mask_; /// Extent of the matrix tile in rows Index extent_row_; /// Extent of the matrix tile in rows Index extent_column_; /// A thread's starting row position (assuming steady-state predicates have been computed) Index thread_start_row_; /// A thread's starting column Index thread_start_column_; /// Internal state counter int state_[3]; /// Scatter indices int const* indices_; // // Static asserts about internal strides // static_assert(sizeof(extent_row_) == 4, "Expected 32b extents"); static_assert(sizeof(thread_start_row_) == 4, "Expected 32b extents"); static_assert(sizeof(PredicatedTileIteratorParams::stride) == 8, "Expected 64b strides"); private: // // Methods // public: // // Methods // /// Constructor CUTLASS_DEVICE PredicatedTileIteratorNormVec(PredicatedTileIteratorParams const& params, Element* pointer, TensorCoord extent, int thread_idx, TensorCoord threadblock_offset = TensorCoord(), int const* indices = nullptr) : params_(params), indices_(indices) { TensorCoord thread_offset = ThreadMap::initial_offset(thread_idx) + threadblock_offset; extent_row_ = extent.row(); extent_column_ = extent.column(); thread_start_row_ = thread_offset.row(); thread_start_column_ = thread_offset.column(); // Initialize predicates CUTLASS_PRAGMA_UNROLL for (int c = 0; c < ThreadMap::Iterations::kColumn; ++c) { mask_.predicates[c] = ((thread_offset.column() + ThreadMap::Delta::kColumn * c) < extent.column()); } // Null pointer performs no accesses if (!pointer) { mask_.clear(); } if (ScatterD && !indices) { mask_.clear(); } // Initialize pointer byte_pointer_ = reinterpret_cast<uint8_t*>(pointer) + LongIndex(thread_offset.row()) * LongIndex(params_.stride); if (ScatterD) { byte_pointer_ = reinterpret_cast<uint8_t*>(pointer) + LongIndex(thread_offset.column()) * sizeof(AccessType) / kElementsPerAccess; } // Initialize internal state counter state_[0] = state_[1] = state_[2] = 0; } /// Adds a pointer offset in units of Element CUTLASS_HOST_DEVICE void add_pointer_offset(LongIndex pointer_offset) { byte_pointer_ += pointer_offset * sizeof_bits<Element>::value / 8; } /// Loads a fragment from memory CUTLASS_DEVICE void load_with_byte_offset(Fragment& frag, int64_t byte_offset) const { uint8_t* byte_pointer = byte_pointer_; AccessType* frag_ptr = reinterpret_cast<AccessType*>(&frag); CUTLASS_PRAGMA_UNROLL for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) { CUTLASS_PRAGMA_UNROLL for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) { CUTLASS_PRAGMA_UNROLL for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) { int frag_row_idx = (row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster)); int row_offset = row * ThreadMap::Delta::kRow + group * ThreadMap::Delta::kGroup + cluster * ThreadMap::Delta::kCluster; bool row_guard = ((row_offset + thread_start_row_) < extent_row_); AccessType* memory_pointer = reinterpret_cast<AccessType*>(byte_pointer + byte_offset); if (ScatterD && row_guard) { assert(indices_); memory_pointer = reinterpret_cast<AccessType*>( byte_pointer + byte_offset + LongIndex(indices_[row_offset + thread_start_row_]) * LongIndex(params_.stride)); } CUTLASS_PRAGMA_UNROLL for (int column = 0; column < ThreadMap::Iterations::kColumn; ++column) { bool guard = row_guard && mask_.predicates[column]; if (column == 0) { cutlass::arch::global_load<AccessType, sizeof(AccessType)>( frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column], (void*)&memory_pointer[0], guard); } else { frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column] = frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn]; } } if (row + 1 < ThreadMap::Iterations::kRow) { if (!ScatterD) { byte_pointer += params_.increment_row; } } } if (group + 1 < ThreadMap::Iterations::kGroup) { byte_pointer += params_.increment_group; } } if (cluster + 1 < ThreadMap::Iterations::kCluster) { byte_pointer += params_.increment_cluster; } } } /// Loads a fragment from memory CUTLASS_DEVICE void load(Fragment& frag) const { load_with_byte_offset(frag, 0); } /// Stores a fragment to memory CUTLASS_DEVICE void store_with_byte_offset(Fragment const& frag, int64_t byte_offset) const { uint8_t* byte_pointer = byte_pointer_; AccessType const* frag_ptr = reinterpret_cast<AccessType const*>(&frag); CUTLASS_PRAGMA_UNROLL for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) { CUTLASS_PRAGMA_UNROLL for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) { CUTLASS_PRAGMA_UNROLL for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) { int frag_row_idx = (row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster)); int row_offset = row * ThreadMap::Delta::kRow + group * ThreadMap::Delta::kGroup + cluster * ThreadMap::Delta::kCluster; bool row_guard = ((row_offset + thread_start_row_) < extent_row_); AccessType* memory_pointer = reinterpret_cast<AccessType*>(byte_pointer + byte_offset); if (ScatterD && row_guard) { assert(indices_); memory_pointer = reinterpret_cast<AccessType*>( byte_pointer + byte_offset + LongIndex(indices_[row_offset + thread_start_row_]) * LongIndex(params_.stride)); } CUTLASS_PRAGMA_UNROLL for (int column = 0; column < ThreadMap::Iterations::kColumn; ++column) { bool guard = row_guard && mask_.predicates[column]; if (UseCUDAStore) { if (guard) { memory_pointer[column * ThreadMap::Delta::kColumn / kElementsPerAccess] = frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column]; } } else { cutlass::arch::global_store<AccessType, sizeof(AccessType)>( frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column], (void*)&memory_pointer[column * ThreadMap::Delta::kColumn / kElementsPerAccess], guard); } } if (row + 1 < ThreadMap::Iterations::kRow) { if (!ScatterD) { byte_pointer += params_.increment_row; } } } if (group + 1 < ThreadMap::Iterations::kGroup) { byte_pointer += params_.increment_group; } } if (cluster + 1 < ThreadMap::Iterations::kCluster) { byte_pointer += params_.increment_cluster; } } } /// Stores a fragment to memory CUTLASS_DEVICE void store(Fragment const& frag) const { store_with_byte_offset(frag, 0); } /// Loads a fragment from memory CUTLASS_DEVICE void downsample_load_with_byte_offset(Fragment& frag, int64_t byte_offset, int convolution_P, int convolution_Q, int add_P, int add_Q, int problem_N) const { uint8_t* byte_pointer = byte_pointer_; AccessType* frag_ptr = reinterpret_cast<AccessType*>(&frag); CUTLASS_PRAGMA_UNROLL for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) { CUTLASS_PRAGMA_UNROLL for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) { CUTLASS_PRAGMA_UNROLL for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) { int frag_row_idx = (row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster)); int row_offset = row * ThreadMap::Delta::kRow + group * ThreadMap::Delta::kGroup + cluster * ThreadMap::Delta::kCluster; bool row_guard = ((row_offset + thread_start_row_) < extent_row_); int output_row = row_offset + thread_start_row_; int output_N = output_row / (convolution_P * convolution_Q); int output_PQ = output_row % (convolution_P * convolution_Q); int output_P = output_PQ / convolution_Q; int output_Q = output_PQ % convolution_Q; int input_row = output_N * 2 * convolution_P * 2 * convolution_Q + (2 * output_P + add_P) * 2 * convolution_Q + 2 * output_Q + add_Q; int64_t byte_offset = (input_row - output_row) * problem_N * sizeof(float); AccessType* memory_pointer = reinterpret_cast<AccessType*>(byte_pointer + byte_offset); CUTLASS_PRAGMA_UNROLL for (int column = 0; column < ThreadMap::Iterations::kColumn; ++column) { bool guard = row_guard && mask_.predicates[column]; cutlass::arch::global_load<AccessType, sizeof(AccessType)>( frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column], (void*)&memory_pointer[column * ThreadMap::Delta::kColumn / kElementsPerAccess], guard); } if (row + 1 < ThreadMap::Iterations::kRow) { byte_pointer += params_.increment_row; } } if (group + 1 < ThreadMap::Iterations::kGroup) { byte_pointer += params_.increment_group; } } if (cluster + 1 < ThreadMap::Iterations::kCluster) { byte_pointer += params_.increment_cluster; } } } /// Loads a fragment from memory CUTLASS_DEVICE void upsample_load_with_byte_offset(Fragment& frag, int64_t byte_offset, int convolution_P, int convolution_Q, int add_P, int add_Q, int problem_N) const { uint8_t* byte_pointer = byte_pointer_; AccessType* frag_ptr = reinterpret_cast<AccessType*>(&frag); CUTLASS_PRAGMA_UNROLL for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) { CUTLASS_PRAGMA_UNROLL for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) { CUTLASS_PRAGMA_UNROLL for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) { int frag_row_idx = (row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster)); int row_offset = row * ThreadMap::Delta::kRow + group * ThreadMap::Delta::kGroup + cluster * ThreadMap::Delta::kCluster; bool row_guard = ((row_offset + thread_start_row_) < extent_row_); int output_row = row_offset + thread_start_row_; int output_N = output_row / (convolution_P * convolution_Q); int output_PQ = output_row % (convolution_P * convolution_Q); int output_P = output_PQ / convolution_Q; int output_Q = output_PQ % convolution_Q; int row_add_P = add_P; int row_add_Q = add_Q; if (output_P > convolution_P - 2) row_add_P = 0; if (output_Q > convolution_Q - 2) row_add_Q = 0; int input_row = output_N * (convolution_P / 2) * (convolution_Q / 2) + ((output_P + row_add_P) / 2) * (convolution_Q / 2) + (output_Q + row_add_Q) / 2; int64_t byte_offset = (input_row - output_row) * problem_N * sizeof(float); AccessType* memory_pointer = reinterpret_cast<AccessType*>(byte_pointer + byte_offset); CUTLASS_PRAGMA_UNROLL for (int column = 0; column < ThreadMap::Iterations::kColumn; ++column) { bool guard = row_guard && mask_.predicates[column]; cutlass::arch::global_load<AccessType, sizeof(AccessType)>( frag_ptr[frag_row_idx * ThreadMap::Iterations::kColumn + column], (void*)&memory_pointer[column * ThreadMap::Delta::kColumn / kElementsPerAccess], guard); } if (row + 1 < ThreadMap::Iterations::kRow) { byte_pointer += params_.increment_row; } } if (group + 1 < ThreadMap::Iterations::kGroup) { byte_pointer += params_.increment_group; } } if (cluster + 1 < ThreadMap::Iterations::kCluster) { byte_pointer += params_.increment_cluster; } } } CUTLASS_DEVICE MatrixCoord thread_start() const { return MatrixCoord(thread_start_row_, thread_start_column_); } /// Need to get the thread start row from the tile iterator CUTLASS_DEVICE int32_t thread_start_row() const { return thread_start_row_; } /// Need to get the thread start row from the tile iterator CUTLASS_DEVICE int32_t thread_start_column() const { return thread_start_column_; } /// Extent of the matrix in rows CUTLASS_DEVICE Index extent_row() const { return extent_row_; } /// Extent of the matrix in columns CUTLASS_DEVICE Index extent_column() const { return extent_column_; } /// Advances to the next position to load or store CUTLASS_HOST_DEVICE PredicatedTileIteratorNormVec& operator++() { ++state_[0]; if (!ScatterD) { byte_pointer_ += params_.advance_row; } thread_start_row_ += ThreadMap::Shape::kRow; if (state_[0] == ThreadMap::Count::kRow) { state_[0] = 0; ++state_[1]; byte_pointer_ += params_.advance_group; thread_start_row_ += (ThreadMap::Shape::kGroup - 1) * ThreadMap::Shape::kRow * ThreadMap::Count::kRow; if (state_[1] == ThreadMap::Count::kGroup) { state_[1] = 0; ++state_[2]; byte_pointer_ += params_.advance_cluster; thread_start_row_ += ThreadMap::Count::kGroup * ThreadMap::Shape::kGroup * ThreadMap::Count::kRow * ThreadMap::Shape::kRow; if (state_[2] == ThreadMap::Count::kCluster) { state_[2] = 0; byte_pointer_ += params_.advance_tile; } } } return *this; } ///< Efficiently disables all accesses guarded by mask CUTLASS_DEVICE void clear_mask() { mask_.clear(); } ///< Efficiently enables all accesses guarded by mask CUTLASS_DEVICE void enable_mask() { mask_.enable(); } ///< Sets the mask CUTLASS_DEVICE void get_mask(Mask& mask) const { mask = mask_; } ///< Sets the mask CUTLASS_DEVICE void set_mask(Mask const& mask) { mask_ = mask; } }; /////////////////////////////////////////////////////////////////////////////// } // namespace threadblock } // namespace epilogue } // namespace cutlass ////////////////////////////////////////////////////////////////////////////////
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_distance_epilogue.h
/* * Copyright (c) 2018-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ /*! \file \brief Epilogue for threadblock scoped GEMMs using Tensor Ops. This is adapted from DefaultEpilogueWithBroadcastTensorOp from CUTLASS 2.9.0 (https://github.com/NVIDIA/cutlass/blob/master/include/cutlass/epilogue/threadblock/default_epilogue_with_broadcast.h#L75) This epilogue allows us to load norm buffers using PredicatedTileIteratorNormVec and EpilogueWithBroadcast used for distances L2/cosine as well as applies user-define elementwise operation. -- A norm load is provided PredicatedTileIteratorNormVec -- B norm load is provided by EpilogueWithBroadcast -- elementwise operation is provided by OutputOp */ #pragma once #include <cutlass/array.h> #include <cutlass/cutlass.h> #include <cutlass/numeric_types.h> #include <cutlass/gemm/gemm.h> #include "./predicated_tile_iterator_normvec.h" #include <cutlass/epilogue/threadblock/default_epilogue_tensor_op.h> #include <cutlass/epilogue/threadblock/default_epilogue_volta_tensor_op.h> #include <cutlass/epilogue/threadblock/epilogue.h> #include <cutlass/epilogue/threadblock/epilogue_with_broadcast.h> //////////////////////////////////////////////////////////////////////////////// namespace cutlass { namespace epilogue { namespace threadblock { //////////////////////////////////////////////////////////////////////////////// /// Defines sensible defaults for epilogues for TensorOps. template <typename Shape, typename WarpMmaTensorOp, int PartitionsK, typename ElementOutput, typename ElementTensor, typename ElementVector, typename OutputOp, typename LayoutT, int ElementsPerAccess, bool ScatterD = false> struct PairwiseDistanceEpilogue { /// Use defaults related to the existing epilogue using Base = DefaultEpilogueTensorOp<Shape, WarpMmaTensorOp, PartitionsK, OutputOp, ElementsPerAccess>; // // Stores the result z = (y = GEMM(A, B, C), broadcast) // using OutputTileIterator = cutlass::epilogue::threadblock:: PredicatedTileIteratorNormVec<typename Base::OutputTileThreadMap, ElementOutput, LayoutT>; // // Additional tensor tile iterator - stores t = Elementwise(z) // using TensorTileIterator = cutlass::epilogue::threadblock::PredicatedTileIterator<typename Base::OutputTileThreadMap, ElementTensor>; /// Define the epilogue using Epilogue = EpilogueWithBroadcast<Shape, WarpMmaTensorOp, PartitionsK, OutputTileIterator, TensorTileIterator, ElementVector, typename Base::AccumulatorFragmentIterator, typename Base::WarpTileIterator, typename Base::SharedLoadIterator, OutputOp, typename Base::Padding, Base::kFragmentsPerIteration>; }; } // namespace threadblock } // namespace epilogue } // namespace cutlass ////////////////////////////////////////////////////////////////////////////////
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_distance_base.cuh
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/linalg/contractions.cuh> // raft::linalg::Contractions_NT #include <raft/util/cuda_dev_essentials.cuh> // ceildiv #include <raft/util/cuda_rt_essentials.hpp> // RAFT_CUDA_TRY #include <cstddef> // size_t namespace cuvs { namespace distance { namespace detail { /** * @brief Device class for L1, L2 and cosine distance metrics. * @tparam DataT input data-type (for A and B matrices) * @tparam AccT accumulation data-type * @tparam OutT output data-type (for C and D matrices) * @tparam IdxT index data-type * @tparam Policy struct which tunes the Contraction kernel * @tparam OpT A distance operation, e.g., cosine_distance_op. * @tparam EpilogueLambda applies an elementwise function to compute final values. Its signature is: template <typename AccT, typename DataT> void epilogue_lambda (AccT acc[][], DataT* regxn, DataT* regyn); * @tparam FinalLambda the final lambda called on final distance value * @param[in] x input matrix * @param[in] y input matrix * @param[in] m number of rows of A and C/D * @param[in] n number of columns of B and C/D * @param[in] k number of cols of A and rows of B * @param[in] lda leading dimension of A * @param[in] ldb leading dimension of B * @param[in] ldd leading dimension of C/D * @param[in] xn row norms of input matrix A. Required for expanded L2, cosine * @param[in] yn row norms of input matrix B. Required for expanded L2, cosine * @param[output] pD output matrix * @param[in] smem shared mem buffer for intermediate storage of A, B, xn & yn. * @param distance_op the distance operation, e.g. cosine_distance_op * @param epilog_op the epilog operation lambda * @param fin_op the final gemm epilogue lambda * @param rowEpilog_op epilog lambda that executes when a full row has been processed */ template <typename DataT, typename OutT, typename IdxT, typename Policy, typename OpT, typename EpilogueLambda, typename FinalLambda, typename rowEpilogueLambda, bool isRowMajor = true, bool writeOut = true, typename BaseClass = raft::linalg::Contractions_NT<DataT, IdxT, Policy, isRowMajor>> struct PairwiseDistances : public BaseClass { // Get accumulation type from distance_op using AccT = typename OpT::AccT; private: typedef Policy P; const DataT* xn; const DataT* yn; const DataT* const yBase; OutT* dOutput; char* smem; OpT distance_op; EpilogueLambda epilog_op; FinalLambda fin_op; rowEpilogueLambda rowEpilog_op; const IdxT grid_stride_m; const IdxT grid_stride_n; const IdxT grid_offset_m; const IdxT grid_offset_n; AccT acc[P::AccRowsPerTh][P::AccColsPerTh]; public: // Constructor DI PairwiseDistances(const DataT* _x, const DataT* _y, IdxT _m, IdxT _n, IdxT _k, IdxT _lda, IdxT _ldb, IdxT _ldd, const DataT* _xn, const DataT* _yn, OutT* _dOutput, char* _smem, OpT _distance_op, EpilogueLambda _epilog_op, FinalLambda _fin_op, rowEpilogueLambda _rowEpilog_op) : BaseClass(_x, _y, _m, _n, _k, _lda, _ldb, _ldd, _smem), xn(_xn), yn(_yn), yBase(_y), dOutput(_dOutput), smem(_smem), distance_op(_distance_op), epilog_op(_epilog_op), fin_op(_fin_op), rowEpilog_op(_rowEpilog_op), grid_stride_m(P::Mblk * gridDim.y), grid_stride_n(P::Nblk * gridDim.x), grid_offset_m(P::Mblk * blockIdx.y), grid_offset_n(P::Nblk * blockIdx.x) { } DI void run() { for (auto tile_idx_m = grid_offset_m; tile_idx_m < this->m; tile_idx_m += grid_stride_m) { this->ldgXY(tile_idx_m, grid_offset_n, 0); for (auto tile_idx_n = grid_offset_n; tile_idx_n < this->n; tile_idx_n += grid_stride_n) { // Prolog: reset_accumulator(); this->stsXY(); __syncthreads(); this->switch_write_buffer(); // Main loop: for (int kidx = P::Kblk; kidx < this->k; kidx += P::Kblk) { this->ldgXY(tile_idx_m, tile_idx_n, kidx); // Process all data in shared memory (previous k-block) and // accumulate in registers. accumulate(); this->stsXY(); __syncthreads(); this->switch_write_buffer(); this->switch_read_buffer(); } accumulate(); // last iteration // The pre-condition for the loop over tile_idx_n is that write_buffer // and read_buffer point to the same buffer. This flips read_buffer back // so that it satisfies the pre-condition of this loop. this->switch_read_buffer(); // Epilog: if (distance_op.use_norms) { DataT regxn[P::AccRowsPerTh], regyn[P::AccColsPerTh]; load_norms(tile_idx_m, tile_idx_n, regxn, regyn); // Overlap ldg with epilog computation ldgNextGridStride(tile_idx_m, tile_idx_n); // Calculate distance_op epilog. // Use .template to disambiguate (See: // https://en.cppreference.com/w/cpp/language/dependent_name) distance_op.template epilog<Policy>(acc, regxn, regyn, tile_idx_n, tile_idx_m); // And any possible additional epilogs epilog_op(acc, regxn, regyn, tile_idx_n, tile_idx_m); } else { // Overlap ldg with epilog computation ldgNextGridStride(tile_idx_m, tile_idx_n); // Calculate distance_op epilog. // Use .template to disambiguate (See: // https://en.cppreference.com/w/cpp/language/dependent_name) distance_op.template epilog<Policy>(acc, nullptr, nullptr, tile_idx_n, tile_idx_m); // And any possible additional epilogs epilog_op(acc, nullptr, nullptr, tile_idx_n, tile_idx_m); } if (writeOut) { store_output(tile_idx_m, tile_idx_n); } } rowEpilog_op(tile_idx_m); } } private: DI void ldgNextGridStride(IdxT tile_idx_m, IdxT tile_idx_n) { // Fetch next grid stride ldg if within range const auto next_tile_tile_idx_n = tile_idx_n + grid_stride_n; const auto next_tile_tile_idx_m = tile_idx_m + grid_stride_m; if ((next_tile_tile_idx_n) < this->n) { this->ldgXY(tile_idx_m, next_tile_tile_idx_n, 0); } else if ((next_tile_tile_idx_m) < this->m) { this->ldgXY(next_tile_tile_idx_m, grid_offset_n, 0); } } DI void reset_accumulator() { // Reset accumulator registers to zero. #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { acc[i][j] = BaseClass::Zero; } } } DI void accumulate_reg_tile(DataT (&reg_x)[P::AccRowsPerTh][P::Veclen], DataT (&reg_y)[P::AccColsPerTh][P::Veclen]) { #pragma unroll for (int v = 0; v < P::Veclen; ++v) { #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { distance_op.core(acc[i][j], reg_x[i][v], reg_y[j][v]); } } } } DI void accumulate() { // We have a separate raft::ldsXY and accumulate_reg_tile outside the loop body, // so that these separated calls can be interspersed with preceding and // following instructions, thereby hiding latency. this->ldsXY(0); // If expensive inner loop, do not unroll loop. constexpr int num_iterations = P::Kblk / P::Veclen - 1; constexpr int unroll_count = decltype(distance_op)::expensive_inner_loop ? 1 : num_iterations; #pragma unroll unroll_count for (int ki = P::Veclen; ki < P::Kblk; ki += P::Veclen) { accumulate_reg_tile(this->regx, this->regy); this->ldsXY(ki); } // Accumulate last loaded tile. accumulate_reg_tile(this->regx, this->regy); } DI void load_norms(IdxT tile_idx_m, IdxT tile_idx_n, DataT (&regxn)[P::AccRowsPerTh], DataT (&regyn)[P::AccColsPerTh]) { DataT* sxNorm = (DataT*)(&smem[P::SmemSize]); DataT* syNorm = (&sxNorm[P::Mblk]); // Load x & y norms required by this threadblock in shmem buffer if (tile_idx_n == blockIdx.x * P::Nblk) { for (int i = threadIdx.x; i < P::Mblk; i += P::Nthreads) { auto idx = tile_idx_m + i; sxNorm[i] = idx < this->m ? xn[idx] : 0; } } for (int i = threadIdx.x; i < P::Nblk; i += P::Nthreads) { auto idx = tile_idx_n + i; syNorm[i] = idx < this->n ? yn[idx] : 0; } __syncthreads(); #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { regxn[i] = sxNorm[i * P::AccThRows + (threadIdx.x / P::AccThCols)]; } #pragma unroll for (int i = 0; i < P::AccColsPerTh; ++i) { regyn[i] = syNorm[i * P::AccThCols + (threadIdx.x % P::AccThCols)]; } } DI void store_output(IdxT tile_idx_m, IdxT tile_idx_n) { IdxT starty = tile_idx_m + this->accrowid; IdxT startx = tile_idx_n + this->acccolid; #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { auto rowId = starty + i * P::AccThRows; #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { auto colId = startx + j * P::AccThCols; if (rowId < this->m && colId < this->n) { // Promote to 64 bit index for final write, as output array can be > 2^31 dOutput[std::size_t(rowId) * this->n + colId] = fin_op(acc[i][j], 0); } } } } }; // struct PairwiseDistances template <typename P, typename IdxT, typename T> dim3 launchConfigGenerator(IdxT m, IdxT n, std::size_t sMemSize, T func) { int devId; RAFT_CUDA_TRY(cudaGetDevice(&devId)); int numSMs; RAFT_CUDA_TRY(cudaDeviceGetAttribute(&numSMs, cudaDevAttrMultiProcessorCount, devId)); int numBlocksPerSm = 0; dim3 grid; RAFT_CUDA_TRY( cudaOccupancyMaxActiveBlocksPerMultiprocessor(&numBlocksPerSm, func, P::Nthreads, sMemSize)); std::size_t minGridSize = numSMs * numBlocksPerSm; std::size_t yChunks = raft::ceildiv<int>(m, P::Mblk); std::size_t xChunks = raft::ceildiv<int>(n, P::Nblk); grid.y = yChunks > minGridSize ? minGridSize : yChunks; grid.x = (minGridSize - grid.y) <= 0 ? 1 : xChunks; if (grid.x != 1) { std::size_t i = 1; while (grid.y * i < minGridSize) { i++; } grid.x = i >= xChunks ? xChunks : i; } return grid; } }; // namespace detail }; // namespace distance }; // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/compress_to_bits.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/core/handle.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/util/cuda_utils.cuh> #include <raft/util/device_atomics.cuh> namespace cuvs::distance::detail { /** * @brief Compress 2D boolean matrix to bitfield * * Utility kernel for masked_l2_nn. * * @tparam T * * @parameter[in] in An `m x n` boolean matrix. Row major. * @parameter[out] out An `(m / bits_per_elem) x n` matrix with elements of * type T, where T is of size `bits_per_elem` bits. * Note: the division (`/`) is a ceilDiv. */ template <typename T = uint64_t, typename = std::enable_if_t<std::is_integral<T>::value>> RAFT_KERNEL compress_to_bits_kernel( raft::device_matrix_view<const bool, int, raft::layout_c_contiguous> in, raft::device_matrix_view<T, int, raft::layout_c_contiguous> out) { constexpr int bits_per_element = 8 * sizeof(T); constexpr int tile_dim_m = bits_per_element; constexpr int nthreads = 128; constexpr int tile_dim_n = nthreads; // read 128 bools at once = 1 sector // Tile in shared memory is transposed __shared__ bool smem[tile_dim_n][tile_dim_m]; const int num_tiles_per_m = raft::ceildiv(in.extent(0), tile_dim_m); const int num_tiles_per_n = raft::ceildiv(in.extent(1), tile_dim_n); for (int lin_tile_idx = blockIdx.x; true; lin_tile_idx += gridDim.x) { const int tile_idx_n = tile_dim_n * (lin_tile_idx % num_tiles_per_n); const int tile_idx_m = tile_dim_m * (lin_tile_idx / num_tiles_per_n); if (in.extent(0) <= tile_idx_m) { break; } // Fill shared memory tile bool reg_buf[tile_dim_m]; #pragma unroll for (int i = 0; i < tile_dim_m; ++i) { const int in_m = tile_idx_m + i; const int in_n = tile_idx_n + threadIdx.x; bool in_bounds = in_m < in.extent(0) && in_n < in.extent(1); reg_buf[i] = in_bounds ? in(in_m, in_n) : false; smem[threadIdx.x][i] = reg_buf[i]; } __syncthreads(); // Drain memory tile into single output element out_elem. T out_elem{0}; #pragma unroll for (int j = 0; j < tile_dim_n; ++j) { if (smem[threadIdx.x][j]) { out_elem |= T(1) << j; } } __syncthreads(); // Write output. int out_m = tile_idx_m / bits_per_element; int out_n = tile_idx_n + threadIdx.x; if (out_m < out.extent(0) && out_n < out.extent(1)) { out(out_m, out_n) = out_elem; } } } /** * @brief Compress 2D boolean matrix to bitfield * * Utility kernel for masked_l2_nn. * * @tparam T * * @parameter[in] in An `m x n` boolean matrix. Row major. * @parameter[out] out An `(m / bits_per_elem) x n` matrix with elements of * type T, where T is of size `bits_per_elem` bits. * Note: the division (`/`) is a ceilDiv. */ template <typename T = uint64_t, typename = std::enable_if_t<std::is_integral<T>::value>> void compress_to_bits(raft::resources const& handle, raft::device_matrix_view<const bool, int, raft::layout_c_contiguous> in, raft::device_matrix_view<T, int, raft::layout_c_contiguous> out) { auto stream = resource::get_cuda_stream(handle); constexpr int bits_per_element = 8 * sizeof(T); RAFT_EXPECTS(raft::ceildiv(in.extent(0), bits_per_element) == out.extent(0), "Number of output rows must be ceildiv(input rows, bits_per_elem)"); RAFT_EXPECTS(in.extent(1) == out.extent(1), "Number of output columns must equal input columns."); const int num_SMs = raft::getMultiProcessorCount(); int blocks_per_sm = 0; constexpr int num_threads = 128; constexpr int dyn_smem_size = 0; RAFT_CUDA_TRY(cudaOccupancyMaxActiveBlocksPerMultiprocessor( &blocks_per_sm, compress_to_bits_kernel<T>, num_threads, dyn_smem_size)); dim3 grid(num_SMs * blocks_per_sm); dim3 block(128); compress_to_bits_kernel<<<grid, block, 0, stream>>>(in, out); RAFT_CUDA_TRY(cudaGetLastError()); } }; // namespace cuvs::distance::detail
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/masked_distance_base.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/distance/detail/pairwise_distance_base.cuh> #include <raft/linalg/contractions.cuh> #include <raft/util/cuda_utils.cuh> #include <cstddef> namespace cuvs { namespace distance { namespace detail { /** * @brief Device class for masked nearest neighbor computations. * * @tparam useNorms whether norms are needed * @tparam DataT input data-type (for x and y matrices) * @tparam AccT accumulation data-type * @tparam IdxT index data-type * @tparam Policy struct which tunes the Contraction kernel * @tparam CoreLambda tells how to accumulate an x and y into acc. its signature: template <typename AccT, typename DataT> void core_lambda(AccT& acc, const DataT& x, const DataT& y) * @tparam EpilogueLambda applies an elementwise function to compute final values. Its signature is: template <typename AccT, typename DataT> void epilogue_lambda (AccT acc[][], DataT* regxn, DataT* regyn); * @tparam FinalLambda the final lambda called on final distance value * @tparam rowEpilogueLambda epilog lambda that executes when a full row has * been processed. * * @param[in] x input matrix * @param[in] y input matrix * @param[in] m number of rows of x * @param[in] n number of columns of y * @param[in] k number of cols of x and y * @param[in] lda leading dimension of x * @param[in] ldb leading dimension of y * @param[in] ldd parameter to keep Contractions_NT happy.. * @param[in] xn row norms of input matrix A. Required for expanded L2, cosine * @param[in] yn row norms of input matrix B. Required for expanded L2, cosine * @param[in] adj An adjacency matrix encoded as a bitfield indicating for each * row of `x` and each group in `y` whether to compute the * distance. Dim = `(m / 64) x num_groups`. * @param[in] group_idxs An array containing the *end* indices of each group * in `y`. The value of group_idxs[j] indicates the * start of group j + 1, i.e., it is the inclusive * scan of the group lengths. The first group is * always assumed to start at index 0 and the last * group typically ends at index `n`. Length = * `num_groups`. * @param[in] num_groups The number of groups in group_idxs. * @param[in] smem shared mem buffer for intermediate storage of x, y, xn & yn. * @param core_op the core accumulation operation lambda * @param epilog_op the epilog operation lambda * @param fin_op the final gemm epilogue lambda * @param rowEpilog_op epilog lambda that executes when a full row has been processed. */ template <bool useNorms, typename DataT, typename AccT, typename IdxT, typename Policy, typename CoreLambda, typename EpilogueLambda, typename FinalLambda, typename rowEpilogueLambda, bool isRowMajor = true, typename BaseClass = raft::linalg::Contractions_NT<DataT, IdxT, Policy, isRowMajor>> struct MaskedDistances : public BaseClass { private: typedef Policy P; const DataT* xn; const DataT* yn; const DataT* const yBase; const uint64_t* adj; const IdxT* group_idxs; IdxT num_groups; char* smem; CoreLambda core_op; EpilogueLambda epilog_op; FinalLambda fin_op; rowEpilogueLambda rowEpilog_op; AccT acc[P::AccRowsPerTh][P::AccColsPerTh]; public: // Constructor DI MaskedDistances(const DataT* _x, const DataT* _y, IdxT _m, IdxT _n, IdxT _k, IdxT _lda, IdxT _ldb, IdxT _ldd, const DataT* _xn, const DataT* _yn, const uint64_t* _adj, const IdxT* _group_idxs, IdxT _num_groups, char* _smem, CoreLambda _core_op, EpilogueLambda _epilog_op, FinalLambda _fin_op, rowEpilogueLambda _rowEpilog_op) : BaseClass(_x, _y, _m, _n, _k, _lda, _ldb, _ldd, _smem), xn(_xn), yn(_yn), yBase(_y), adj(_adj), group_idxs(_group_idxs), num_groups(_num_groups), smem(_smem), core_op(_core_op), epilog_op(_epilog_op), fin_op(_fin_op), rowEpilog_op(_rowEpilog_op) { } DI void run() { const auto grid_stride_m = (P::Mblk * gridDim.y); const auto grid_offset_m = (P::Mblk * blockIdx.y); const auto grid_stride_g = gridDim.x; const auto grid_offset_g = blockIdx.x; for (auto tile_idx_m = grid_offset_m; tile_idx_m < this->m; tile_idx_m += grid_stride_m) { // Start loop over groups for (auto idx_g = grid_offset_g; idx_g < this->num_groups; idx_g += grid_stride_g) { const uint64_t block_adj = get_block_adjacency(adj, tile_idx_m, idx_g); // block_adj is a bitfield that contains a 1 if a row is adjacent to the // current group. All zero means we can skip this group. if (block_adj == 0) { continue; } // thread_adj is a bitfield that contains a 1 at location i iff we must // compute row i of acc (the accumulator register tile). That is, // for i = 0,.., AccRowsPerTh and j = 0,.., AccColsPerTh: // // ((1 << i) & thread_adj) > 0 <=> acc[i][j] must be computed. // // We precompute this information because it is used in various // locations to skip thread-local computations, specifically: // // 1. To skip computations if thread_adj == 0, i.e., none of the values // of `acc` have to be computed. // // 2. In epilog_op, to consider only values of `acc` to be reduced that // are not masked of. // // Note 1: Even when the computation can be skipped for a specific thread, // the thread still participates in synchronization operations. // // Note 2: In theory, it should be possible to skip computations for // specific rows of `acc`. In practice, however, this does not improve // performance. int thread_adj = compute_thread_adjacency(block_adj); auto tile_idx_n = idx_g == 0 ? 0 : group_idxs[idx_g - 1]; const auto group_end_n = group_idxs[idx_g]; for (; tile_idx_n < group_end_n; tile_idx_n += P::Nblk) { // We provide group_end_n to limit the number of unnecessary data // points that are loaded from y. this->ldgXY(tile_idx_m, tile_idx_n, 0, group_end_n); reset_accumulator(); this->stsXY(); __syncthreads(); this->switch_write_buffer(); for (int kidx = P::Kblk; kidx < this->k; kidx += P::Kblk) { this->ldgXY(tile_idx_m, tile_idx_n, kidx, group_end_n); // Process all data in shared memory (previous k-block) and // accumulate in registers. if (thread_adj != 0) { accumulate(); } this->stsXY(); __syncthreads(); this->switch_write_buffer(); this->switch_read_buffer(); } if (thread_adj != 0) { accumulate(); // last iteration } // The pre-condition for the loop over tile_idx_n is that write_buffer // and read_buffer point to the same buffer. This flips read_buffer // back so that it satisfies the pre-condition of this loop. this->switch_read_buffer(); if (useNorms) { DataT regxn[P::AccRowsPerTh], regyn[P::AccColsPerTh]; load_norms(tile_idx_m, tile_idx_n, group_end_n, regxn, regyn); if (thread_adj != 0) { epilog_op(acc, thread_adj, regxn, regyn, tile_idx_n, tile_idx_m, group_end_n); } } else { if (thread_adj != 0) { epilog_op(acc, thread_adj, nullptr, nullptr, tile_idx_n, tile_idx_m, group_end_n); } } } // tile_idx_n } // idx_g rowEpilog_op(tile_idx_m); } // tile_idx_m } private: DI uint64_t get_block_adjacency(const uint64_t* adj, IdxT tile_idx_m, IdxT idx_group) { // A single element of `adj` contains exactly enough bits to indicate which // rows in the current tile to skip and which to compute. static_assert(P::Mblk == 8 * sizeof(adj[0]), "masked_l2_nn only supports a policy with 64 rows per block."); IdxT block_flag_idx = tile_idx_m / P::Mblk; // Index into adj at row tile_idx_m / 64 and column idx_group. return adj[block_flag_idx * this->num_groups + idx_group]; } DI uint32_t compute_thread_adjacency(const uint64_t block_adj) { // thread_adj is a bitfield that contains a 1 at location i iff we must // compute row i of acc (the accumulator register tile). It is described in // more detail in the run() method. uint32_t thread_adj = 0; #pragma unroll for (int thread_row_idx = 0; thread_row_idx < P::AccRowsPerTh; ++thread_row_idx) { // Index `thread_row_idx` refers to a row of the current threads' register // tile `acc`, i.e., acc[i][:]. Index `block_row_idx` refers to the // corresponding row of the current block tile in shared memory. const int block_row_idx = this->accrowid + thread_row_idx * P::AccThRows; // block_row_is_adjacent is true if the current block_row_idx is adjacent // to the current group. const uint64_t block_mask = 1ull << block_row_idx; const bool block_row_is_adjacent = (block_adj & block_mask) != 0; if (block_row_is_adjacent) { // If block row is adjacent, write a 1 bit to thread_adj at location // `thread_row_idx`. const uint32_t thread_mask = 1 << thread_row_idx; thread_adj |= thread_mask; } } return thread_adj; } DI void reset_accumulator() { // Reset accumulator registers to zero. #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { acc[i][j] = BaseClass::Zero; } } } DI void accumulate() { #pragma unroll for (int ki = 0; ki < P::Kblk; ki += P::Veclen) { this->ldsXY(ki); #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < P::AccColsPerTh; ++j) { #pragma unroll for (int v = 0; v < P::Veclen; ++v) { core_op(acc[i][j], this->regx[i][v], this->regy[j][v]); } } } } } DI void load_norms(IdxT tile_idx_m, IdxT tile_idx_n, IdxT end_n, DataT (&regxn)[P::AccRowsPerTh], DataT (&regyn)[P::AccColsPerTh]) { DataT* sxNorm = (DataT*)(&smem[P::SmemSize]); DataT* syNorm = (&sxNorm[P::Mblk]); // Load x & y norms required by this threadblock in shmem buffer for (int i = threadIdx.x; i < P::Mblk; i += P::Nthreads) { auto idx = tile_idx_m + i; sxNorm[i] = idx < this->m ? xn[idx] : 0; } for (int i = threadIdx.x; i < P::Nblk; i += P::Nthreads) { auto idx = tile_idx_n + i; syNorm[i] = idx < end_n ? yn[idx] : 0; } __syncthreads(); #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { regxn[i] = sxNorm[i * P::AccThRows + (threadIdx.x / P::AccThCols)]; } #pragma unroll for (int i = 0; i < P::AccColsPerTh; ++i) { regyn[i] = syNorm[i * P::AccThCols + (threadIdx.x % P::AccThCols)]; } } }; // struct MaskedDistances }; // namespace detail }; // namespace distance }; // namespace cuvs
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/canberra.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/core/operators.hpp> // raft::abs #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { /** * @brief The canberra distance matrix calculation * * It computes the following equation: * * c_ij = sum_k |x_ik - y_kj| / ( |x_ik| + |y_kj| ) */ template <typename DataType, typename AccType, typename IdxType> struct canberra_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; // Load norms of input data static constexpr bool use_norms = false; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = true; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize; } DI void core(AccT& acc, DataT& x, DataT& y) const { const auto diff = raft::abs(x - y); const auto add = raft::abs(x) + raft::abs(y); // deal with potential for 0 in denominator by // forcing 0/1 instead acc += ((add != 0) * diff / (add + (add == 0))); }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { return; } }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/l2_exp.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/core/math.hpp> #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { /** * Reserve 1 digit of precision from each floating-point type * for round-off error tolerance. * @tparam DataT */ template <typename DataT> __device__ constexpr DataT get_clamp_precision() { switch (sizeof(DataT)) { case 2: return 1e-3; case 4: return 1e-6; case 8: return 1e-15; default: return 0; } } // Epilogue operator for CUTLASS based kernel template <typename DataT, typename AccT> struct l2_exp_cutlass_op { bool sqrt; __device__ l2_exp_cutlass_op() noexcept : sqrt(false) {} __device__ l2_exp_cutlass_op(bool isSqrt) noexcept : sqrt(isSqrt) {} inline __device__ AccT operator()(DataT aNorm, DataT bNorm, DataT accVal) const noexcept { AccT outVal = aNorm + bNorm - DataT(2.0) * accVal; /** * Self-neighboring points should have (aNorm == bNorm) == accVal and the dot product (accVal) * can sometimes have round-off errors, which will cause (aNorm == bNorm) ~ accVal instead. */ outVal = outVal * !((outVal * outVal < get_clamp_precision<DataT>()) * (aNorm == bNorm)); return sqrt ? raft::sqrt(outVal * (outVal > 0)) : outVal; } __device__ AccT operator()(DataT aData) const noexcept { return aData; } }; /** * @brief the expanded euclidean distance matrix calculation * * It computes the following equation: * * c_ij = - 2 sum_k x_ik * y_kj + ||x_i.||_2 + ||y_.j||_2 * */ template <typename DataType, typename AccType, typename IdxType> struct l2_exp_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; const bool sqrt; l2_exp_distance_op(bool sqrt_) noexcept : sqrt(sqrt_) {} // Load norms of input data static constexpr bool use_norms = true; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = false; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize + ((Policy::Mblk + Policy::Nblk) * sizeof(DataT)); } DI void core(AccT& acc, DataT& x, DataT& y) const { acc += x * y; }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { #pragma unroll for (int i = 0; i < Policy::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < Policy::AccColsPerTh; ++j) { DataT accVal = acc[i][j]; DataT val = regxn[i] + regyn[j] - (DataT)2.0 * accVal; /** * Self-neighboring points should have (aNorm == bNorm) == accVal and the dot product * (accVal) can sometimes have round-off errors, which will cause (aNorm == bNorm) ~ accVal * instead. */ acc[i][j] = val * (val > 0) * !((val * val < get_clamp_precision<DataT>()) * (regxn[i] == regyn[j])); } } if (sqrt) { #pragma unroll for (int i = 0; i < Policy::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < Policy::AccColsPerTh; ++j) { acc[i][j] = raft::sqrt(acc[i][j]); } } } } constexpr l2_exp_cutlass_op<DataT, AccT> get_cutlass_op() const { return l2_exp_cutlass_op<DataT, AccT>(sqrt); } }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/l1.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { /** * @brief the L1 distance matrix calculation * * It computes the following equation: * * c_ij = sum_k abs(x_ik - y_kj) */ template <typename DataType, typename AccType, typename IdxType> struct l1_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; // Do not load norms of data, the computation of L1 distance does not use them. static constexpr bool use_norms = false; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = false; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize; } DI void core(AccT& acc, DataT& x, DataT& y) const { acc += raft::abs(x - y); }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { return; }; }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/all_ops.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once // Defines a named requirement "has_cutlass_op" #include <cuvs/distance/detail/distance_ops/cutlass.cuh> // The distance operations: #include <cuvs/distance/detail/distance_ops/canberra.cuh> #include <cuvs/distance/detail/distance_ops/correlation.cuh> #include <cuvs/distance/detail/distance_ops/cosine.cuh> #include <cuvs/distance/detail/distance_ops/hamming.cuh> #include <cuvs/distance/detail/distance_ops/hellinger.cuh> #include <cuvs/distance/detail/distance_ops/jensen_shannon.cuh> #include <cuvs/distance/detail/distance_ops/kl_divergence.cuh> #include <cuvs/distance/detail/distance_ops/l1.cuh> #include <cuvs/distance/detail/distance_ops/l2_exp.cuh> #include <cuvs/distance/detail/distance_ops/l2_unexp.cuh> #include <cuvs/distance/detail/distance_ops/l_inf.cuh> #include <cuvs/distance/detail/distance_ops/lp_unexp.cuh> #include <cuvs/distance/detail/distance_ops/russel_rao.cuh>
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/template.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { // Describes the computation the template distance // // Fill in the TODO items. template <typename DataType, typename AccType, typename IdxType> struct template_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; TODO member; template_distance_op(TODO member_) noexcept : member(member_) {} // Load norms of input data static constexpr bool use_norms = TODO; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = false; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize + TODO; } DI void core(AccT& acc, DataT& x, DataT& y) const { TODO; }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { TODO; } // If exist, returns a cutlass op that performs the same operation. // See cosine and l2_exp distance ops for an example. constexpr l2_exp_cutlass_op<DataT, AccT> get_cutlass_op() const { TODO; } }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/hamming.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { /** * @brief the Hamming Unexpanded distance matrix calculation * It computes the following equation: * * c_ij = sum_k (x_ik != y_kj) / k */ template <typename DataType, typename AccType, typename IdxType> struct hamming_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; IdxT k; hamming_distance_op(IdxT k_) noexcept : k(k_) {} // Load norms of input data static constexpr bool use_norms = false; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = false; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize; } DI void core(AccT& acc, DataT& x, DataT& y) const { acc += (x != y); }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { const DataT one_over_k = DataT(1.0) / k; #pragma unroll for (int i = 0; i < Policy::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < Policy::AccColsPerTh; ++j) { acc[i][j] *= one_over_k; } } } }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/kl_divergence.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/core/operators.hpp> // raft::log #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { /** * @brief the KL Divergence distance matrix calculation * * It computes the following equation: * * c_ij = 0.5 * sum(x * log (x / y)); */ template <typename DataType, typename AccType, typename IdxType> struct kl_divergence_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; const bool is_row_major; const bool x_equal_y; kl_divergence_op(bool row_major_, bool x_equal_y_ = false) noexcept : is_row_major(row_major_), x_equal_y(x_equal_y_) { } // Load norms of input data static constexpr bool use_norms = false; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = true; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize; } DI void core(AccT& acc, DataT& x, DataT& y) const { // TODO: make sure that these branches get hoisted out of main loop.. Could // be quite expensive otherwise. if (x_equal_y) { if (is_row_major) { const bool x_zero = (x == 0); const bool y_zero = (y == 0); acc += x * (raft::log(x + x_zero) - (!y_zero) * raft::log(y + y_zero)); } else { const bool y_zero = (y == 0); const bool x_zero = (x == 0); acc += y * (raft::log(y + y_zero) - (!x_zero) * raft::log(x + x_zero)); } } else { if (is_row_major) { const bool x_zero = (x == 0); acc += x * (raft::log(x + x_zero) - y); } else { const bool y_zero = (y == 0); acc += y * (raft::log(y + y_zero) - x); } } }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { #pragma unroll for (int i = 0; i < Policy::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < Policy::AccColsPerTh; ++j) { acc[i][j] = (0.5f * acc[i][j]); } } } }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/lp_unexp.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/core/operators.hpp> // raft::pow, raft::abs #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { /** * @brief the unexpanded Lp (Minkowski) distance matrix calculation * * It computes the following equation: * * c_ij = (sum_k |x_ik - y_jk|^p)^(1/p) */ template <typename DataType, typename AccType, typename IdxType> struct lp_unexp_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; DataT p; lp_unexp_distance_op(DataT p_) noexcept : p(p_) {} // Load norms of input data static constexpr bool use_norms = false; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = true; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize; } DI void core(AccT& acc, DataT& x, DataT& y) const { const auto diff = raft::abs(x - y); acc += raft::pow(diff, p); }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { const auto one_over_p = 1.0f / p; #pragma unroll for (int i = 0; i < Policy::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < Policy::AccColsPerTh; ++j) { acc[i][j] = raft::pow(acc[i][j], one_over_p); } } } }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/cutlass.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <type_traits> // std::false_type #include <utility> // std::declval namespace cuvs::distance::detail::ops { // This file defines the named requirement "has_cutlass_op" that can be used to // determine if a distance operation has a CUTLASS op that can be used to pass // to CUTLASS. Examples of distance operations that satisfy this requirement are // cosine_distance_op and l2_exp_distance_op. // Primary template handles types that do not support CUTLASS. // This pattern is described in: // https://en.cppreference.com/w/cpp/types/void_t template <typename, typename = void> struct has_cutlass_op : std::false_type {}; // Specialization recognizes types that do support CUTLASS template <typename T> struct has_cutlass_op<T, std::void_t<decltype(std::declval<T>().get_cutlass_op())>> : std::true_type {}; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/correlation.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { /** @brief The correlation distance * * It computes the following equation: * * d(x, y) = ((x - mean(x)) ⋅ (y - mean(y))) * / * (|| x - mean(x) ||_2 || y - mean(y) ||_2) */ template <typename DataType, typename AccType, typename IdxType> struct correlation_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; const DataT* x2n; const DataT* y2n; IdxT m; IdxT n; IdxT k; correlation_distance_op( bool is_row_major, const DataT* x2n_, const DataT* y2n_, IdxT m_, IdxT n_, IdxT k_) noexcept : x2n(x2n_), y2n(y2n_), m(m_), n(n_), k(k_) { // The distance op is typically created before the row-major/col-major // swapping has been done. So we do it here. if (!is_row_major) { std::swap<const DataT*>(x2n, y2n); std::swap(m, n); } } // Load norms of input data static constexpr bool use_norms = true; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = false; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize + (2 * (Policy::Mblk + Policy::Nblk) * sizeof(DataT)); } DI void core(AccT& acc, DataT& x, DataT& y) const { acc += x * y; }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { // Note how we can sneakily get a pointer to shared memory here, to store // more data. If the implementation of PairwiseDistanceMatKernel ever // changes, this will be where we find the bugs. extern __shared__ char smem[]; DataT regx2n[Policy::AccRowsPerTh], regy2n[Policy::AccColsPerTh]; DataT* sx2Norm = (DataT*)(&smem[Policy::SmemSize + (Policy::Mblk + Policy::Nblk) * sizeof(DataT)]); DataT* sy2Norm = (&sx2Norm[Policy::Mblk]); // Load x & y norms required by this threadblock in shmem buffer if (gridStrideX == blockIdx.x * Policy::Nblk) { for (int i = threadIdx.x; i < Policy::Mblk; i += Policy::Nthreads) { auto idx = gridStrideY + i; sx2Norm[i] = idx < m ? x2n[idx] : 0; } } for (int i = threadIdx.x; i < Policy::Nblk; i += Policy::Nthreads) { auto idx = gridStrideX + i; sy2Norm[i] = idx < n ? y2n[idx] : 0; } __syncthreads(); #pragma unroll for (int i = 0; i < Policy::AccRowsPerTh; ++i) { regx2n[i] = sx2Norm[i * Policy::AccThRows + (threadIdx.x / Policy::AccThCols)]; } #pragma unroll for (int i = 0; i < Policy::AccColsPerTh; ++i) { regy2n[i] = sy2Norm[i * Policy::AccThCols + (threadIdx.x % Policy::AccThCols)]; } #pragma unroll for (int i = 0; i < Policy::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < Policy::AccColsPerTh; ++j) { auto numer = k * acc[i][j] - (regxn[i] * regyn[j]); auto Q_denom = k * regx2n[i] - (regxn[i] * regxn[i]); auto R_denom = k * regy2n[j] - (regyn[j] * regyn[j]); acc[i][j] = 1 - (numer / raft::sqrt(Q_denom * R_denom)); } } } }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/russel_rao.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { /** * @brief the Russell Rao distance matrix calculation * * It computes the following equation: * * c_ij = (k - (sum_k x_ik * y_kj)) / k */ template <typename DataType, typename AccType, typename IdxType> struct russel_rao_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; IdxT k; const float one_over_k; russel_rao_distance_op(IdxT k_) noexcept : k(k_), one_over_k(1.0f / k_) {} // Load norms of input data static constexpr bool use_norms = false; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = false; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize; } DI void core(AccT& acc, DataT& x, DataT& y) const { acc += x * y; }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { #pragma unroll for (int i = 0; i < Policy::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < Policy::AccColsPerTh; ++j) { acc[i][j] = (k - acc[i][j]) * one_over_k; } } } }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/hellinger.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { /** * @brief the Hellinger distance matrix calculation * * It computes the following equation: * * c_ij = sqrt(1 - sum_k sqrt(x_ik * y_kj)) * */ template <typename DataType, typename AccType, typename IdxType> struct hellinger_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; // Load norms of input data static constexpr bool use_norms = false; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = false; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize; } DI void core(AccT& acc, DataT& x, DataT& y) const { // This is sqrt(x) * sqrt(y). const auto product = x * y; acc += product; }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { #pragma unroll for (int i = 0; i < Policy::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < Policy::AccColsPerTh; ++j) { // Adjust to replace NaN in sqrt with 0 if input to sqrt is negative const auto finalVal = (1 - acc[i][j]); const auto rectifier = (!signbit(finalVal)); acc[i][j] = raft::sqrt(rectifier * finalVal); } } } }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/jensen_shannon.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/core/operators.hpp> // raft::log #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { // Describes the computation the jensen_shannon distance /** * @brief the Jensen Shannon distance matrix calculation * * It computes the following equation: * * c_ij = sqrt(0.5 * sum( -x_i * (log(0.5 * (x_i + y_i)) - log(x_i)) * + (-y_i * (log(0.5 * (x_i + y_i)) - log(y_i))))) */ template <typename DataType, typename AccType, typename IdxType> struct jensen_shannon_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; // Load norms of input data static constexpr bool use_norms = false; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = true; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize; } DI void core(AccT& acc, DataT& x, DataT& y) const { const DataT m = 0.5f * (x + y); const bool m_zero = (m == 0); const auto logM = (!m_zero) * raft::log(m + m_zero); const bool x_zero = (x == 0); const bool y_zero = (y == 0); acc += (-x * (logM - raft::log(x + x_zero))) + (-y * (logM - raft::log(y + y_zero))); }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { #pragma unroll for (int i = 0; i < Policy::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < Policy::AccColsPerTh; ++j) { acc[i][j] = raft::sqrt(0.5 * acc[i][j]); } } } }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/cosine.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { // Epilogue operator for CUTLASS based kernel template <typename DataT, typename AccT> struct cosine_cutlass_op { __device__ cosine_cutlass_op() noexcept {} __device__ AccT operator()(DataT& aNorm, const DataT& bNorm, DataT& accVal) const noexcept { return static_cast<AccT>(1.0) - static_cast<AccT>(accVal / (aNorm * bNorm)); } __device__ AccT operator()(DataT aData) const noexcept { return aData; } }; /** * @brief the expanded cosine distance matrix calculation * * It computes the following equation: * * d(x, y) = 1 - (x ⋅ y) / ( ||x||_2 ||y||_2) */ template <typename DataType, typename AccType, typename IdxType> struct cosine_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; // Load norms of input data static constexpr bool use_norms = true; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = false; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize + ((Policy::Mblk + Policy::Nblk) * sizeof(DataT)); } DI void core(AccT& acc, DataT& x, DataT& y) const { acc += x * y; }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { #pragma unroll for (int i = 0; i < Policy::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < Policy::AccColsPerTh; ++j) { acc[i][j] = 1.0 - (acc[i][j] / (regxn[i] * regyn[j])); } } } constexpr cosine_cutlass_op<DataT, AccT> get_cutlass_op() const { return cosine_cutlass_op<DataT, AccT>(); } }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/l_inf.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { /** * @brief the L_inf (Chebyshev) distance matrix calculation * * It computes the following equation: * * c_ij = max_k | x_ik - y_kj | */ template <typename DataType, typename AccType, typename IdxType> struct l_inf_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; // Load norms of input data static constexpr bool use_norms = false; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = false; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize; } DI void core(AccT& acc, DataT& x, DataT& y) const { const auto diff = raft::abs(x - y); acc = raft::max(acc, diff); }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { return; } }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/distance_ops/l2_unexp.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <raft/util/cuda_dev_essentials.cuh> // DI namespace cuvs::distance::detail::ops { /** * @brief the unexpanded euclidean distance matrix calculation * * It computes the following equation: * * c_ij = optional_sqrt ( sum_k (x_ik - y_kj)^2 ) */ template <typename DataType, typename AccType, typename IdxType> struct l2_unexp_distance_op { using DataT = DataType; using AccT = AccType; using IdxT = IdxType; bool sqrt; l2_unexp_distance_op(bool sqrt_) noexcept : sqrt(sqrt_) {} // Do not load norms of data, the computation of L1 distance does not use them. static constexpr bool use_norms = false; // Whether the core function requires so many instructions that it makes sense // to reduce loop unrolling, etc. We do this to keep compile times in check. static constexpr bool expensive_inner_loop = false; // Size of shared memory. This is normally decided by the kernel policy, but // some ops such as correlation_distance_op use more. template <typename Policy> static constexpr size_t shared_mem_size() { return Policy::SmemSize; } DI void core(AccT& acc, DataT& x, DataT& y) const { const auto diff = x - y; acc += diff * diff; }; template <typename Policy> DI void epilog(AccT acc[Policy::AccRowsPerTh][Policy::AccColsPerTh], DataT* regxn, DataT* regyn, IdxT gridStrideX, IdxT gridStrideY) const { if (sqrt) { #pragma unroll for (int i = 0; i < Policy::AccRowsPerTh; ++i) { #pragma unroll for (int j = 0; j < Policy::AccColsPerTh; ++j) { acc[i][j] = raft::sqrt(acc[i][j]); } } } }; }; } // namespace cuvs::distance::detail::ops
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/fused_distance_nn/epilogue_elementwise.cuh
/*************************************************************************************************** * Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved. * SPDX-License-Identifier: BSD-3-Clause * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, this * list of conditions and the following disclaimer. * * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * 3. Neither the name of the copyright holder nor the names of its * contributors may be used to endorse or promote products derived from * this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * **************************************************************************************************/ /* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ // /*! \file \brief Functor performing distance operations used by epilogues of pairwise distance * kernels. * This is adapted from LinearCombinationBiasElementwise from CUTLASS 2.9.0 * customized for applying elementwise distance formula on accumulated GEMM value * and applying user-defined operation which can convert distance values to key-value pair. * . */ #pragma once #include <cutlass/array.h> #include <cutlass/cutlass.h> #include <cutlass/functional.h> #include <cutlass/numeric_conversion.h> #include <cutlass/numeric_types.h> #include <cutlass/epilogue/thread/activation.h> ///////////////////////////////////////////////////////////////////////////////////////////////// namespace cutlass { namespace epilogue { namespace thread { ///////////////////////////////////////////////////////////////////////////////////////////////// /// This base class is meant to define the concept required of the /// EpilogueWithBroadcast::OutputOp template <typename ElementC_, typename ElementAccumulator_, typename ElementCompute_, typename ElementZ_, typename ElementT_, int ElementsPerAccess, typename DistanceOp_, typename CGReduceOp_, typename ReduceOpT_, typename KVPReduceOpT_> class FusedDistanceNNEpilogueElementwise { public: using ElementOutput = ElementC_; using ElementC = ElementC_; using ElementAccumulator = ElementAccumulator_; using ElementCompute = ElementCompute_; using ElementZ = ElementZ_; using ElementT = ElementT_; static int const kElementsPerAccess = ElementsPerAccess; static int const kCount = kElementsPerAccess; using DistanceOp = DistanceOp_; using CGReduceOp = CGReduceOp_; using FragmentAccumulator = Array<ElementAccumulator, kElementsPerAccess>; using FragmentCompute = Array<ElementCompute, kElementsPerAccess>; using FragmentC = Array<ElementOutput, kElementsPerAccess>; using FragmentZ = Array<ElementZ, kElementsPerAccess>; using OutValT = typename CGReduceOp::AccTypeT; using FragmentT = Array<OutValT, kElementsPerAccess>; using FragmentOutput = FragmentZ; static bool const kIsHeavy = true; // ElementwiseOp::kIsHeavy; /// If true, the 'Z' tensor is stored static bool const kStoreZ = false; // We don't store anything in Z, /// If true, the 'T' tensor is stored static bool const kStoreT = true; // this is our final output storage. /// Host-constructable parameters structure struct Params { CGReduceOp_ cg_reduce_op; DistanceOp_ dist_op_; KVPReduceOpT_ pair_redop_; ReduceOpT_ red_op_; int* mutexes_; using CGReduceT = CGReduceOp_; // // Methods // CUTLASS_HOST_DEVICE Params(DistanceOp_ dist_op, CGReduceOp cg_reduce_op, ReduceOpT_ red_op, KVPReduceOpT_ pair_redop, int* mutexes) : cg_reduce_op(cg_reduce_op), dist_op_(dist_op), pair_redop_(pair_redop), red_op_(red_op), mutexes_(mutexes) { } CUTLASS_HOST_DEVICE Params() {} }; private: // // Data members // DistanceOp_ elementwise_op; KVPReduceOpT_ pair_redop; public: ReduceOpT_ red_op; // // Methods // /// Constructor from Params CUTLASS_HOST_DEVICE FusedDistanceNNEpilogueElementwise(Params const& params) : elementwise_op(params.dist_op_), pair_redop(params.pair_redop_), red_op(params.red_op_) { } /// Returns true if source is needed CUTLASS_HOST_DEVICE bool is_source_needed() const { // we use for making sure C matrix is used for A mat norm. return true; } /// Functionally required for serial reduction in the epilogue CUTLASS_HOST_DEVICE void set_k_partition(int k_partition, int k_partition_count) {} /// Applies the operation when is_source_needed() is true CUTLASS_HOST_DEVICE void operator()(FragmentT& frag_T, FragmentAccumulator const& AB, FragmentC const& frag_C, FragmentCompute const& V) const { FragmentCompute tmp_Accum = NumericArrayConverter<ElementCompute, ElementAccumulator, kElementsPerAccess>()(AB); FragmentCompute tmp_C = NumericArrayConverter<ElementCompute, ElementC, kElementsPerAccess>()(frag_C); FragmentCompute result_Z; CUTLASS_PRAGMA_UNROLL for (int i = 0; i < kElementsPerAccess; ++i) { ElementCompute res_Z = elementwise_op(tmp_C[i], V[i], tmp_Accum[i]); frag_T[i] = res_Z; } } /// Applies the operation when is_source_needed() is false CUTLASS_HOST_DEVICE void operator()(FragmentZ& frag_Z, FragmentT& frag_T, FragmentAccumulator const& AB, FragmentCompute const& V) const { } }; ///////////////////////////////////////////////////////////////////////////////////////////////// } // namespace thread } // namespace epilogue } // namespace cutlass /////////////////////////////////////////////////////////////////////////////////////////////////
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/fused_distance_nn/custom_epilogue_with_broadcast.h
/*************************************************************************************************** * Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved. * SPDX-License-Identifier: BSD-3-Clause * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, this * list of conditions and the following disclaimer. * * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * 3. Neither the name of the copyright holder nor the names of its * contributors may be used to endorse or promote products derived from * this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * **************************************************************************************************/ /* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ /*! \file \brief Epilogue for threadblock scoped GEMMs using Tensor Ops. The epilogue rearranges the result of a matrix product through shared memory to match canonical tensor layouts in global memory. Epilogues support conversion and reduction operations. This file contains a customized version of EpilogueWithBroadcast from CUTLASS 2.9.1 (https://github.com/NVIDIA/cutlass/blob/v2.9.1/include/cutlass/epilogue/threadblock/epilogue_with_broadcast.h) Changes: - customized the compute_source_needed_() and apply_output_operator_() to suit the needs of per row reduction */ #pragma once #if defined(__CUDACC_RTC__) #include <cuda/std/cassert> #include <cuda/std/utility> #else #include <assert.h> #include <utility> #endif #include <cutlass/aligned_buffer.h> #include <cutlass/array.h> #include <cutlass/cutlass.h> #include <cutlass/fast_math.h> #include <cutlass/functional.h> #include <cutlass/layout/tensor.h> #include <cutlass/layout/vector.h> #include <cutlass/numeric_conversion.h> #include <cutlass/numeric_types.h> #include <cutlass/tensor_coord.h> #include <cutlass/gemm/gemm.h> #include <cutlass/transform/pitch_linear_thread_map.h> #include <cutlass/transform/threadblock/regular_tile_iterator.h> #include <cutlass/epilogue/threadblock/epilogue_base.h> #include <cutlass/epilogue/threadblock/predicated_tile_iterator.h> #include <cutlass/numeric_types.h> ///////////////////////////////////////////////////////////////////////////////////////////////// namespace cutlass { namespace epilogue { namespace threadblock { ///////////////////////////////////////////////////////////////////////////////////////////////// /// This base class is meant to define the concept required of the /// EpilogueWithBroadcast::OutputOp template <typename ElementC_, typename ElementAccumulator_, typename ElementCompute_, typename ElementZ_, typename ElementT_, int ElementsPerAccess, bool StoreZ = true, bool StoreT = true> struct EpilogueWithBroadcastOpBaseCustom { using ElementOutput = ElementC_; using ElementAccumulator = ElementAccumulator_; using ElementCompute = ElementCompute_; using ElementZ = ElementZ_; using ElementT = ElementT_; static int const kElementsPerAccess = ElementsPerAccess; using FragmentAccumulator = Array<ElementAccumulator, kElementsPerAccess>; using FragmentCompute = Array<ElementCompute, kElementsPerAccess>; using FragmentC = Array<ElementOutput, kElementsPerAccess>; using FragmentZ = Array<ElementZ, kElementsPerAccess>; using FragmentT = Array<ElementT, kElementsPerAccess>; /// If true, the 'Z' tensor is stored static bool const kStoreZ = StoreZ; /// If true, the 'T' tensor is stored static bool const kStoreT = StoreT; /// Parameters structure - required struct Params {}; // // Methods // /// Constructor from Params EpilogueWithBroadcastOpBaseCustom(Params const& params_) {} /// Determine if the source is needed. May return false if bool is_source_needed() const { return true; } CUTLASS_HOST_DEVICE void set_k_partition(int k_partition, int k_partition_count) {} /// Applies the operation when is_source_needed() is true CUTLASS_HOST_DEVICE void operator()(FragmentZ& frag_Z, FragmentT& frag_T, FragmentAccumulator const& AB, FragmentC const& frag_C, FragmentCompute const& V) const { } /// Applies the operation when is_source_needed() is false CUTLASS_HOST_DEVICE void operator()(FragmentZ& frag_Z, FragmentT& frag_T, FragmentAccumulator const& AB, FragmentCompute const& V) const { } }; //////////////////////////////////////////////////////////////////////////////// /// Epilogue operator with bias vector broadcast over columns. /// /// Computes the following: /// /// /// Z, T = OutputOp(AB, C, Broadcast) /// /// if (ElementwiseOp::kStoreZ) { /// store(converted_u); /// } /// /// if (ElementwiseOp::kStoreT) { /// store(v); /// } /// template < typename Shape_, ///< Shape of threadblock tile (concept: GemmShape) typename WarpMmaOperator_, ///< Warp-level MMA operator (concept: gemm::warp::MmaTensorOp) int PartitionsK, ///< Number of partitions of the K dimension typename OutputTileIterator_, ///< Tile iterator reading and writing output tensors (z) typename TensorTileIterator_, ///< Additional tile iterator for tensor-valued operands (t) typename ElementVector_, ///< Pointer to broadcast vector typename AccumulatorFragmentIterator_, ///< Fragment iterator selecting accumulators typename WarpTileIterator_, ///< Warp-scoped tile iterator writing accumulators to SMEM typename SharedLoadIterator_, ///< Threadblock-scoped tile iterator loading from SMEM typename OutputOp_, ///< Output operator - concept is EpilogueWithBroadcastOp typename Padding_, ///< Padding added to SMEM allocation to avoid bank conflicts (concept: ///< MatrixShape) int FragmentsPerPartition = 1, ///< Used to coarsten the epilogue granularity int IterationsUnroll = ///< Used to reduce binary size when epilogue op is large (!IsEpilogueFunctorHeavy<OutputOp_>::value)> class EpilogueWithBroadcastCustom : public EpilogueBase<Shape_, typename WarpMmaOperator_::Shape, PartitionsK, AccumulatorFragmentIterator_, WarpTileIterator_, Padding_, FragmentsPerPartition> { public: using Base = EpilogueBase<Shape_, typename WarpMmaOperator_::Shape, PartitionsK, AccumulatorFragmentIterator_, WarpTileIterator_, Padding_, FragmentsPerPartition>; using Shape = Shape_; using WarpMmaOperator = WarpMmaOperator_; static int const kPartitionsK = PartitionsK; using OutputTileIterator = OutputTileIterator_; using TensorTileIterator = TensorTileIterator_; using ElementVector = ElementVector_; using AccumulatorFragmentIterator = AccumulatorFragmentIterator_; using WarpTileIterator = WarpTileIterator_; using SharedLoadIterator = SharedLoadIterator_; using OutputOp = OutputOp_; using Padding = Padding_; using Layout = layout::RowMajor; using LongIndex = typename Layout::LongIndex; /// The complete warp-level accumulator tile using AccumulatorTile = typename Base::AccumulatorTile; /// Accumulator element using ElementAccumulator = typename WarpTileIterator::Element; /// Compute data type produced by the output op using ElementCompute = typename OutputOp::ElementCompute; /// Compute fragment using FragmentCompute = Array<ElementCompute, OutputTileIterator::Fragment::kElements>; /// Thread map used by output tile iterators using ThreadMap = typename OutputTileIterator::ThreadMap; /// Fragment object used to store the broadcast values using BroadcastFragment = Array<ElementCompute, ThreadMap::Iterations::kColumn * ThreadMap::kElementsPerAccess>; /// Output element using ElementOutput = typename OutputTileIterator::Element; /// Data type of additional tensor using ElementTensor = typename TensorTileIterator::Element; /// Output access size static int const kElementsPerAccess = OutputTileIterator::kElementsPerAccess; /// Tensor reference to destination tensor using TensorRef = typename OutputTileIterator::TensorRef; /// Tensor reference to sync tensor using SyncTensorRef = typename cutlass::TensorRef<int, cutlass::layout::PackedVectorLayout>; /// Const tensor reference to source tensor using ConstTensorRef = typename OutputTileIterator::ConstTensorRef; /// Array type used to output using OutputAccessType = Array<typename OutputTileIterator::Element, OutputTileIterator::kElementsPerAccess>; /// Array type used by output functor using AccumulatorAccessType = Array<typename WarpTileIterator::Element, OutputTileIterator::kElementsPerAccess>; /// Array type used by output functor using ComputeAccessType = Array<ElementCompute, OutputTileIterator::kElementsPerAccess>; /// Tensor access type using TensorAccessType = Array<ElementTensor, OutputTileIterator::kElementsPerAccess>; /// Number of warps using WarpCount = typename Base::WarpCount; /// Shared memory allocation from epilogue base class using BaseSharedStorage = typename Base::SharedStorage; static int constexpr kSmemTiles = Base::kFragmentsPerIteration > 1 ? Base::kFragmentsPerIteration : kPartitionsK; static int constexpr kSmemPointerOffset = Base::SharedStorage::StorageShape::kCount / kSmemTiles; /// Used for the broadcast struct BroadcastDetail { /// Number of threads per warp static int const kWarpSize = 32; static int const kElementsPerAccess = ThreadMap::kElementsPerAccess; /// Number of distinct scalar column indices handled by each thread static int const kColumnsPerThread = ThreadMap::Iterations::kColumn * ThreadMap::kElementsPerAccess; /// Number of distinct scalar row indices handled by each thread static int const kRowsPerThread = ThreadMap::Iterations::kCount / ThreadMap::Iterations::kColumn; /// Number of threads per threadblock static int const kThreadCount = kWarpSize * WarpCount::kCount; /// Number of distinct threads per row of output tile static int const kThreadsPerRow = (Shape::kN / kColumnsPerThread); /// Number of distinct threads which must be reduced during the final reduction phase within the /// threadblock. static int const kThreadRows = kThreadCount / kThreadsPerRow; /// I'm not sure what I meant here. static int const kThreadAccessesPerRow = const_max(1, (Shape::kN + kThreadCount - 1) / kThreadCount); /// Shape of the shared memory allocation for the epilogue using StorageShape = MatrixShape<kThreadRows, Shape::kN>; /// Debug printing CUTLASS_DEVICE static void print() { #if 0 printf("BroadcastDetail {\n"); printf( " kColumnsPerThread: %d\nkRowsPerThread: %d\n,kThreadCount: %d\nkThreadsPerRow: %d\n" "kThreadRows: %d\nThreadAccessesPerRow: %d\nStorageShape: %d x %d (count: %d)\n", kColumnsPerThread, kRowsPerThread, kThreadCount, kThreadsPerRow, kThreadRows, kThreadAccessesPerRow, StorageShape::kRow, StorageShape::kColumn, StorageShape::kCount ); printf("};\n"); #endif } }; /// Shared storage structure (shadows base) with additional SMEM buffer for reduction struct SharedStorage { union { BaseSharedStorage base; }; CUTLASS_HOST_DEVICE SharedStorage() {} }; public: static_assert(SharedLoadIterator::Fragment::kElements == TensorTileIterator::Fragment::kElements, "Mismatch between shared load iterator and output tile iterator."); static_assert(OutputTileIterator::kElementsPerAccess, "OutputTileIterator::kElementsPerAccess must not be zero."); static_assert(!(OutputTileIterator::Fragment::kElements % OutputTileIterator::kElementsPerAccess), "Divisibility"); private: /// Loads fragment from shared memory aligned with output tensor SharedLoadIterator shared_load_iterator_; /// Thread index within the threadblock int thread_idx_; public: /// Constructor CUTLASS_DEVICE EpilogueWithBroadcastCustom(SharedStorage& shared_storage, ///< Shared storage object int thread_idx, ///< ID of a thread within the threadblock int warp_idx, ///< ID of warp within threadblock int lane_idx ///< Id of thread within warp ) : Base(shared_storage.base, thread_idx, warp_idx, lane_idx), shared_load_iterator_(shared_storage.base.reference(), thread_idx), thread_idx_(thread_idx) { } /// Streams the result to global memory CUTLASS_DEVICE void operator()( OutputOp const& output_op, ///< Output operator ElementVector const* broadcast_ptr, ///< Broadcast vector AccumulatorTile const& accumulators, ///< Complete warp-level accumulator tile OutputTileIterator source_iterator, ///< Tile iterator for source accumulator matrix TensorTileIterator tensor_iterator, ///< Threadblock tile iterator for additional tensor operand MatrixCoord const& problem_size = ///< Problem size needed to guard against out-of-bounds accesses MatrixCoord(Shape::kM, Shape::kN), MatrixCoord const& threadblock_offset = ///< Threadblock's initial offset within the problem size space MatrixCoord()) { BroadcastFragment broadcast_fragment; load_broadcast_fragment_(broadcast_fragment, broadcast_ptr, problem_size, threadblock_offset); compute_source_needed_( output_op, broadcast_fragment, accumulators, source_iterator, tensor_iterator); } private: CUTLASS_DEVICE void load_broadcast_fragment_( BroadcastFragment& broadcast_fragment, ///< Fragment containing the accumulated partial reduction over columns ElementVector const* broadcast_ptr, ///< Broadcast vector MatrixCoord const& problem_size, ///< Problem size needed to guard against out-of-bounds accesses MatrixCoord const& threadblock_offset ///< Threadblock's initial offset within the problem size space ) { broadcast_fragment.clear(); // If no pointer is supplied, set with all zeros and avoid memory accesses if (!broadcast_ptr) { return; } int thread_initial_column = ThreadMap::initial_offset(thread_idx_).column(); int thread_column_idx = threadblock_offset.column() + thread_initial_column; broadcast_ptr += thread_initial_column; NumericArrayConverter<ElementCompute, ElementVector, BroadcastDetail::kElementsPerAccess> converter; using AccessType = AlignedArray<ElementVector, BroadcastDetail::kElementsPerAccess>; using ComputeFragmentType = Array<ElementCompute, BroadcastDetail::kElementsPerAccess>; ComputeFragmentType* frag_ptr = reinterpret_cast<ComputeFragmentType*>(&broadcast_fragment); CUTLASS_PRAGMA_UNROLL for (int j = 0; j < ThreadMap::Iterations::kColumn; ++j) { AccessType loaded; loaded.clear(); if (thread_column_idx < problem_size.column()) { loaded = *reinterpret_cast<AccessType const*>(broadcast_ptr); } ComputeFragmentType cvt = converter(loaded); frag_ptr[j] = cvt; thread_column_idx += ThreadMap::Delta::kColumn; broadcast_ptr += ThreadMap::Delta::kColumn; } } template <class Seq> struct acc2smem_source_not_needed; template <size_t... Seq> struct acc2smem_source_not_needed<cutlass::index_sequence<Seq...>> { template <int Advance> CUTLASS_DEVICE static void helper(AccumulatorFragmentIterator accum_fragment_iterator, WarpTileIterator& warp_tile_iterator) { CUTLASS_PRAGMA_UNROLL for (int i = 0; i < Advance; i++) { ++accum_fragment_iterator; } CUTLASS_PRAGMA_UNROLL for (int p = 0; p < Base::kFragmentsPerIteration; ++p) { typename AccumulatorFragmentIterator::Fragment accum_fragment; accum_fragment_iterator.load(accum_fragment); ++accum_fragment_iterator; warp_tile_iterator.store(accum_fragment); if (p < Base::kFragmentsPerIteration - 1) { warp_tile_iterator.add_pointer_offset(kSmemPointerOffset); } } if (Base::kFragmentsPerIteration > 1) { warp_tile_iterator.add_pointer_offset(kSmemPointerOffset * (1 - Base::kFragmentsPerIteration)); } } CUTLASS_DEVICE static void push(size_t pos, AccumulatorFragmentIterator const& iterator_begin, WarpTileIterator& warp_tile_iterator) { int dummy[] = { (pos == (Seq * Base::kFragmentsPerIteration)) && (helper<Seq * Base::kFragmentsPerIteration>(iterator_begin, warp_tile_iterator), 0)...}; CUTLASS_UNUSED(dummy[0]); } }; /// Streams the result to global memory CUTLASS_DEVICE void compute_source_not_needed_( OutputOp const& output_op, ///< Output operator BroadcastFragment const& broadcast_fragment, ///< Fragment containing the accumulated partial reduction over columns OutputTileIterator destination_iterator, ///< Tile iterator for destination AccumulatorTile const& accumulators, ///< Complete warp-level accumulator tile TensorTileIterator tensor_iterator ///< Threadblock tile iterator for additioanl tensor operand ) { } template <class Seq> struct acc2smem_source_needed; template <size_t... Seq> struct acc2smem_source_needed<cutlass::index_sequence<Seq...>> { template <int Advance> CUTLASS_DEVICE static void helper(AccumulatorFragmentIterator accum_fragment_iterator, WarpTileIterator& warp_tile_iterator) { CUTLASS_PRAGMA_UNROLL for (int i = 0; i < Advance; i++) { ++accum_fragment_iterator; } typename AccumulatorFragmentIterator::Fragment accum_fragment; accum_fragment_iterator.load(accum_fragment); warp_tile_iterator.store(accum_fragment); } CUTLASS_DEVICE static void push(size_t pos, AccumulatorFragmentIterator const& iterator_begin, WarpTileIterator& warp_tile_iterator) { int dummy[] = {(pos == Seq) && (helper<Seq>(iterator_begin, warp_tile_iterator), 0)...}; } }; /// Streams the result to global memory CUTLASS_DEVICE void compute_source_needed_( OutputOp const& output_op, ///< Output operator BroadcastFragment const& broadcast_fragment, ///< Fragment containing the accumulated partial reduction over columns AccumulatorTile const& accumulators, ///< Complete warp-level accumulator tile OutputTileIterator source_iterator, ///< Threadblock tile coordinate in GEMM (in units of threadblock tiles) TensorTileIterator tensor_iterator ///< Threadblock tile iterator for additioanl tensor operand ) { typename OutputTileIterator::Fragment source_fragment; source_fragment.clear(); // // Iterator over warp-level accumulator fragment // AccumulatorFragmentIterator accum_fragment_iterator(accumulators); // // Iterate over accumulator tile // #pragma unroll(IterationsUnroll ? OutputTileIterator::kIterations : 1) for (int iter = 0; iter < OutputTileIterator::kIterations; ++iter) { // // Convert and store fragment // //__syncthreads(); acc2smem_source_needed<cutlass::make_index_sequence<OutputTileIterator::kIterations>>::push( iter, accum_fragment_iterator, this->warp_tile_iterator_); __syncthreads(); // // Load fragments from shared memory // typename SharedLoadIterator::Fragment aligned_accum_fragment[kPartitionsK]; shared_load_iterator_.load(aligned_accum_fragment[0]); // // Apply output operation // typename TensorTileIterator::Fragment frag_T; // // Load the source // source_iterator.load(source_fragment); ++source_iterator; apply_output_operator_( frag_T, output_op, aligned_accum_fragment[0], source_fragment, broadcast_fragment); // // Conditionally store fragments // if (OutputOp::kStoreT) { tensor_iterator.store(frag_T); ++tensor_iterator; } } } /// Helper to invoke the output functor over each vector of output CUTLASS_DEVICE void apply_output_operator_(typename TensorTileIterator::Fragment& frag_T, OutputOp const& output_op, typename SharedLoadIterator::Fragment const& frag_AB, typename OutputTileIterator::Fragment const& frag_C, BroadcastFragment const& frag_Broadcast) { using AccessTypeT = Array<typename TensorTileIterator::OutValT, kElementsPerAccess>; using AccessTypeBroadcast = Array<ElementCompute, kElementsPerAccess>; AccessTypeT* frag_T_ptr = reinterpret_cast<AccessTypeT*>(&frag_T); AccumulatorAccessType const* frag_AB_ptr = reinterpret_cast<AccumulatorAccessType const*>(&frag_AB); OutputAccessType const* frag_C_ptr = reinterpret_cast<OutputAccessType const*>(&frag_C); AccessTypeBroadcast const* frag_Broadcast_ptr = reinterpret_cast<AccessTypeBroadcast const*>(&frag_Broadcast); int const kOutputOpIterations = TensorTileIterator::Fragment::kElements / TensorTileIterator::kElementsPerAccess; CUTLASS_PRAGMA_UNROLL for (int i = 0; i < kOutputOpIterations; ++i) { output_op(frag_T_ptr[i], frag_AB_ptr[i], frag_C_ptr[(i / ThreadMap::Iterations::kColumn)], frag_Broadcast_ptr[i % ThreadMap::Iterations::kColumn]); } } /// Helper to invoke the output functor over each vector of output CUTLASS_DEVICE void apply_output_operator_source_not_needed_( typename OutputTileIterator::Fragment& frag_Z, typename TensorTileIterator::Fragment& frag_T, OutputOp const& output_op, typename SharedLoadIterator::Fragment const& frag_AB, BroadcastFragment const& frag_Broadcast) { } }; //////////////////////////////////////////////////////////////////////////////// } // namespace threadblock } // namespace epilogue } // namespace cutlass ////////////////////////////////////////////////////////////////////////////////
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/fused_distance_nn/cutlass_base.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #pragma GCC diagnostic push #pragma GCC diagnostic ignored "-Wstrict-aliasing" #pragma GCC diagnostic ignored "-Wtautological-compare" // We define CUTLASS_NAMESPACE in case // RAFT cmake is not used #ifndef CUTLASS_NAMESPACE #define cutlass raft_cutlass #endif #include <cutlass/cutlass.h> #include <cutlass/gemm/device/gemm.h> #include <cutlass/gemm/device/gemm_grouped.h> #include <cutlass/gemm/device/gemm_universal_adapter.h> #include <rmm/device_uvector.hpp> #include <cutlass/layout/matrix.h> #include <cutlass/layout/tensor.h> #include <cutlass/matrix_coord.h> #include <cutlass/tensor_view.h> #include <cuvs/distance/detail/fused_distance_nn/epilogue_elementwise.cuh> // FusedDistanceNNEpilogueElementwise #include <cuvs/distance/detail/fused_distance_nn/gemm.h> // FusedDistanceNNGemm #include <raft/util/cudart_utils.hpp> // getMultiProcessorCount #include <raft/util/cutlass_utils.cuh> // RAFT_CUTLASS_TRY namespace cuvs { namespace distance { namespace detail { template <typename DataT, typename AccT, typename OutT, typename IdxT, int VecLen, typename CGReduceOpT, typename DistanceFn, typename ReduceOpT, typename KVPReduceOpT> void cutlassFusedDistanceNN(const DataT* x, const DataT* y, const DataT* xn, const DataT* yn, IdxT m, IdxT n, IdxT k, IdxT lda, IdxT ldb, IdxT ldd, OutT* dOutput, int* mutexes, CGReduceOpT cg_reduce_op, DistanceFn dist_op, ReduceOpT redOp, KVPReduceOpT pairRedOp, cudaStream_t stream) { using EpilogueOutputOp = cutlass::epilogue::thread::FusedDistanceNNEpilogueElementwise< DataT, // ElementC_ AccT, // ElementAccumulator_ DataT, // ElementCompute_ AccT, // ElementZ_ OutT, // ElementT_ // 128 / cutlass::sizeof_bits<DataT>::value, 1, // Elements per access 1 DistanceFn, CGReduceOpT, ReduceOpT, KVPReduceOpT>; constexpr int batch_count = 1; typename EpilogueOutputOp::Params epilog_op_param( dist_op, cg_reduce_op, redOp, pairRedOp, mutexes); // Number of pipelines you want to use constexpr int NumStages = 3; // Alignment constexpr int Alignment = VecLen; // default initialize problem size with row major inputs auto problem_size = cutlass::gemm::GemmCoord(m, n, k); constexpr bool isRowMajor = true; using fusedDistanceNNKernel = typename cutlass::gemm::kernel::FusedDistanceNNGemm<DataT, Alignment, DataT, Alignment, AccT, AccT, EpilogueOutputOp, NumStages, // Number of pipeline stages isRowMajor>::GemmKernel; using fusedDistanceNN = cutlass::gemm::device::GemmGrouped<fusedDistanceNNKernel>; int num_blocks_per_sm = fusedDistanceNN::maximum_active_blocks(); int num_sms = raft::getMultiProcessorCount(); int full_wave = num_blocks_per_sm * num_sms; constexpr int mmaShapeM = fusedDistanceNNKernel::Mma::Shape::kM; constexpr int mmaShapeN = fusedDistanceNNKernel::Mma::Shape::kN; int columnTiles = (problem_size.n() - 1 + mmaShapeN) / mmaShapeN; int rowTiles = (problem_size.m() - 1 + mmaShapeM) / mmaShapeM; int totalTiles = columnTiles * rowTiles; int thread_blocks = rowTiles < full_wave ? (totalTiles < full_wave ? totalTiles : full_wave) : rowTiles; typename fusedDistanceNN::Arguments arguments{ problem_size, batch_count, // num of problems. thread_blocks, epilog_op_param, x, y, xn, // C matrix eq vector param, which here is A norm (DataT*)yn, // this is broadcast vec, which is required to be non-const param dOutput, // Output distance matrix (int64_t)lda, // stride A (int64_t)ldb, // stride B (int64_t)1, // stride A norm (int64_t)ldd // stride Output matrix }; // Using the arguments, query for extra workspace required for matrix multiplication computation size_t workspace_size = fusedDistanceNN::get_workspace_size(arguments); // Allocate workspace memory rmm::device_uvector<uint8_t> workspace(workspace_size, stream); // Instantiate CUTLASS kernel depending on templates fusedDistanceNN fusedDistanceNN_op; // Check the problem size is supported or not RAFT_CUTLASS_TRY(fusedDistanceNN_op.can_implement(arguments)); // Initialize CUTLASS kernel with arguments and workspace pointer RAFT_CUTLASS_TRY(fusedDistanceNN_op.initialize(arguments, workspace.data(), stream)); // Launch initialized CUTLASS kernel RAFT_CUTLASS_TRY(fusedDistanceNN_op.run(stream)); } }; // namespace detail }; // namespace distance }; // namespace cuvs #pragma GCC diagnostic pop
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/fused_distance_nn/predicated_tile_iterator_reduced_vec.h
/*************************************************************************************************** * Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved. * SPDX-License-Identifier: BSD-3-Clause * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, this * list of conditions and the following disclaimer. * * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * 3. Neither the name of the copyright holder nor the names of its * contributors may be used to endorse or promote products derived from * this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * **************************************************************************************************/ /* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ /*! \file \brief Epilogue for threadblock scoped GEMMs using Tensor Ops. This file contains a customized version of PredicatedTileIterator from CUTLASS 2.9.0 (https://github.com/NVIDIA/cutlass/blob/v2.9.0/include/cutlass/epilogue/threadblock/predicated_tile_iterator.h#L75) Changes: - added `Layout_` template param - PredicatedTileIteratorParams() is customized to not stride by layout.stride(0). - makes use of `SharedStorage` to store reduced values across warps to gmem in coalesced manner. - customized the store_with_byte_offset() to perform reduction per row and write final value to gmem. - customized the Params() struct to take user inputs from epilogueOp params. */ #pragma once #include <cooperative_groups.h> #include <cooperative_groups/reduce.h> #include <cutlass/arch/arch.h> #include <cutlass/arch/memory.h> #include <cutlass/array.h> #include <cutlass/cutlass.h> #include <cutlass/epilogue/threadblock/output_tile_thread_map.h> #include <cutlass/epilogue/threadblock/predicated_tile_iterator_params.h> #include <cutlass/layout/matrix.h> #include <cutlass/layout/tensor.h> #include <cutlass/matrix_shape.h> #include <cutlass/numeric_types.h> #include <cutlass/tensor_ref.h> #include <cutlass/transform/pitch_linear_thread_map.h> namespace cg = cooperative_groups; //////////////////////////////////////////////////////////////////////////////// namespace cutlass { //////////////////////////////////////////////////////////////////////////////// namespace epilogue { namespace threadblock { //////////////////////////////////////////////////////////////////////////////// /// Tile iterator used to load and store output tile from global memory in epilogue. /// /// Satisfies: ReadableTileIterator | PredicatedTileIterator | ForwardTileIterator /// template <typename ThreadMap_, ///< Thread map (conept: OutputTileThreadMap) typename Element_, ///< Element data type typename Layout_, typename EpilogueOpParams_, bool ScatterD = false, ///< Scatter D operand or not bool UseCUDAStore = false> class PredicatedTileIteratorReducedVec { public: using ThreadMap = ThreadMap_; using Shape = typename ThreadMap::Shape; using Element = Element_; using Layout = Layout_; using TensorRef = TensorRef<Element, Layout>; using ConstTensorRef = typename TensorRef::ConstTensorRef; using Index = typename Layout::Index; using LongIndex = typename Layout::LongIndex; using TensorCoord = MatrixCoord; using EpilogueOpParams = EpilogueOpParams_; using OutIdxT = typename EpilogueOpParams::CGReduceT::IndexT; using OutValT = typename EpilogueOpParams::CGReduceT::AccTypeT; static int const kElementsPerAccess = ThreadMap::kElementsPerAccess; static int const kThreads = ThreadMap::kThreads; static int const kIterations = ThreadMap::Count::kTile; static_assert(ThreadMap::Iterations::kRow > 0, "ThreadMap::Iterations::kRow must be > 0"); static_assert(ThreadMap::Iterations::kGroup > 0, "ThreadMap::Iterations::kGroup must be > 0"); static_assert(ThreadMap::Iterations::kCluster > 0, "ThreadMap::Iterations::kCluster must be > 0"); static_assert(ThreadMap::Iterations::kColumn > 0, "ThreadMap::Iterations::kColumn must be > 0"); static_assert(!UseCUDAStore, "UseCUDAStore path is not supported"); static int const total_rows = ThreadMap::kWarpCount * ThreadMap::Iterations::kRow * ThreadMap::Iterations::kGroup * ThreadMap::Iterations::kCluster * ThreadMap::Count::kTile * ThreadMap::Delta::kRow; /// Fragment object using Fragment = Array<OutValT, ThreadMap::Iterations::kColumn * ThreadMap::Iterations::kRow * ThreadMap::Iterations::kGroup * ThreadMap::Iterations::kCluster * kElementsPerAccess>; // Memory access size using AccessType = AlignedArray<Element, kElementsPerAccess>; using AccessTypeValT = AlignedArray<OutValT, kElementsPerAccess>; // // Parameters struct // /// Uses a non-template class struct Params : PredicatedTileIteratorParams { using Base = PredicatedTileIteratorParams; EpilogueOpParams user_param; CUTLASS_HOST_DEVICE Params() {} CUTLASS_HOST_DEVICE Params(Layout const& layout) : PredicatedTileIteratorParams( layout.stride(0) * int(sizeof(AccessType)) / kElementsPerAccess, make_OutputTileThreadMapDesc<ThreadMap>()) { } CUTLASS_HOST_DEVICE Params(Layout const& layout, EpilogueOpParams const& user_param_) : PredicatedTileIteratorParams(int(sizeof(AccessType)) / kElementsPerAccess, make_OutputTileThreadMapDesc<ThreadMap>()), user_param(user_param_) { } CUTLASS_HOST_DEVICE Params(Base const& base) : Base(base) {} }; /// Mask object struct Mask { // static int const kCount = ThreadMap::Iterations::kColumn; static int const kCount = ThreadMap::Iterations::kColumn * kElementsPerAccess; /// Predicate state bool predicates[kCount]; // // Mask // CUTLASS_HOST_DEVICE Mask() { enable(); } ///< Efficiently disables all accesses guarded by mask CUTLASS_HOST_DEVICE void clear() { CUTLASS_PRAGMA_UNROLL for (int i = 0; i < kCount; ++i) { predicates[i] = false; } } ///< CUTLASS_HOST_DEVICE enables all accesses guarded by mask CUTLASS_DEVICE void enable() { CUTLASS_PRAGMA_UNROLL for (int i = 0; i < kCount; ++i) { predicates[i] = true; } } }; /// Shared storage allocation needed by the predicated tile // iterator for reduction. struct SharedStorage { // // Type definitions // using Shape = MatrixShape<total_rows, 1>; /// Shape of the shared memory allocation for the reduced values store using StorageShape = MatrixShape<Shape::kRow, Shape::kColumn>; // // Data members // // Methods // AlignedBuffer<Element, StorageShape::kCount> storage; CUTLASS_DEVICE Element* data() { return storage.data(); } SharedStorage() {} CUTLASS_DEVICE void initSmem(EpilogueOpParams const& user_params) { Element* shared_elem_arr = data(); constexpr auto maxVal = std::numeric_limits<OutValT>::max(); for (int row = threadIdx.x; row < total_rows; row += blockDim.x) { user_params.red_op_.init(&shared_elem_arr[row], maxVal); } } }; template <typename cg_reduce_op_t, typename cg_group_t, typename IdxT, typename ValT, typename OutT> struct select_reduce { /// Performs warp level reduction and stores a reduced output to memory CUTLASS_DEVICE select_reduce(OutT value, ValT prev_red_val, cg_reduce_op_t reduce_op, cg_group_t cg_warp_group, OutT& shmem_ptr) { if (cg_warp_group.any(reduce_op.isAmin(value, prev_red_val))) { OutT reduced_val = cg::reduce(cg_warp_group, value, reduce_op); if (cg_warp_group.thread_rank() == 0) { shmem_ptr = reduced_val; } } } }; template <typename cg_reduce_op_t, typename cg_group_t, typename IdxT> struct select_reduce<cg_reduce_op_t, cg_group_t, IdxT, float, raft::KeyValuePair<IdxT, float>> { using ValT = float; using Ty = raft::KeyValuePair<IdxT, ValT>; /// Performs warp level reduction of key value pair and stores a reduced output to memory CUTLASS_DEVICE select_reduce(Ty val_to_red, float prev_red_val, cg_reduce_op_t cg_reduce_op, cg_group_t cg_warp_group, Ty& shmem_ptr) { ValT val = val_to_red.value; if (cg_warp_group.any(cg_reduce_op.isAmin(val, prev_red_val))) { ValT reduced_val = cg::reduce(cg_warp_group, val, cg_reduce_op); bool pred = (reduced_val == val); auto subTile = cg::binary_partition(cg_warp_group, pred); if (pred) { if (subTile.thread_rank() == 0) { shmem_ptr = val_to_red; } } } } }; template <typename cg_reduce_op_t, typename cg_group_t, typename IdxT> struct select_reduce<cg_reduce_op_t, cg_group_t, IdxT, double, raft::KeyValuePair<IdxT, double>> { using ValT = double; using Ty = raft::KeyValuePair<IdxT, ValT>; /// Performs warp level reduction of key value pair and stores a reduced output to memory CUTLASS_DEVICE select_reduce(Ty val_to_red, double prev_red_val, cg_reduce_op_t cg_reduce_op, cg_group_t cg_warp_group, Ty& shmem_ptr) { ValT val = val_to_red.value; if (cg_warp_group.any(cg_reduce_op.isAmin(val, prev_red_val))) { ValT reduced_val = cg::reduce(cg_warp_group, val, cg_reduce_op); bool pred = (reduced_val == val); auto subTile = cg::binary_partition(cg_warp_group, pred); if (pred) { if (subTile.thread_rank() == 0) { shmem_ptr = val_to_red; } } } } }; private: // // Data members // /// Parameters structure containing reference and precomputed state. Params params_; /// Byte-level pointer uint8_t* byte_pointer_; /// Byte-level pointer first tile offset of this threadblock. uint8_t* first_tile_byte_pointer_; /// Array of boolean values to contain steady-state predicates Mask mask_; /// Extent of the matrix tile in rows Index extent_row_; /// Extent of the matrix tile in rows Index extent_column_; /// A thread's starting row position (assuming steady-state predicates have been computed) Index thread_start_row_; Index block_start_row_first_tile_; /// A thread's starting column Index thread_start_column_; /// Internal state counter int state_[3]; // mutable int shared_tile_id; /// Scatter indices int const* indices_; // // Static asserts about internal strides // static_assert(sizeof(extent_row_) == 4, "Expected 32b extents"); static_assert(sizeof(thread_start_row_) == 4, "Expected 32b extents"); static_assert(sizeof(Params::stride) == 8, "Expected 64b strides"); protected: SharedStorage& shared_storage_; const bool& do_gmem_reduction_; private: // // Methods // public: // // Methods // /// Constructor CUTLASS_DEVICE PredicatedTileIteratorReducedVec(SharedStorage& shared_storage, Params const& params, Element* pointer, TensorCoord extent, int thread_idx, const bool& do_gmem_reduction, TensorCoord threadblock_offset = TensorCoord(), int const* indices = nullptr) : params_(params), indices_(indices), shared_storage_(shared_storage), do_gmem_reduction_(do_gmem_reduction) { TensorCoord thread_offset = ThreadMap::initial_offset(thread_idx) + threadblock_offset; extent_row_ = extent.row(); extent_column_ = extent.column(); thread_start_row_ = thread_offset.row(); thread_start_column_ = thread_offset.column(); TensorCoord block_offset = ThreadMap::initial_offset(0) + threadblock_offset; block_start_row_first_tile_ = block_offset.row(); // Initialize predicates CUTLASS_PRAGMA_UNROLL for (int c = 0; c < ThreadMap::Iterations::kColumn * kElementsPerAccess; ++c) { int columnPerAccess = (c / kElementsPerAccess); int columnWithinPerAccess = c % kElementsPerAccess; mask_.predicates[c] = ((thread_offset.column() + ThreadMap::Delta::kColumn * columnPerAccess + columnWithinPerAccess) < extent.column()); } if (threadblock_offset.column() == 0) { EpilogueOpParams const& user_params = params_.user_param; shared_storage_.initSmem(user_params); } // Null pointer performs no accesses if (!pointer) { mask_.clear(); } if (ScatterD && !indices) { mask_.clear(); } // Initialize pointer first_tile_byte_pointer_ = reinterpret_cast<uint8_t*>(pointer) + LongIndex(block_offset.row()) * LongIndex(params_.stride); if (ScatterD) { byte_pointer_ = reinterpret_cast<uint8_t*>(pointer) + LongIndex(thread_offset.column()) * sizeof(AccessType) / kElementsPerAccess; } // Initialize internal state counter state_[0] = state_[1] = state_[2] = 0; } /// Destructor CUTLASS_DEVICE ~PredicatedTileIteratorReducedVec() { if (do_gmem_reduction_) { EpilogueOpParams const& user_params = params_.user_param; auto gmem_ptr = reinterpret_cast<Element*>(first_tile_byte_pointer_); Element* shared_elem_arr = shared_storage_.data(); const uint32_t mutex_id = (block_start_row_first_tile_ / total_rows); bool useGmemMutex = (gridDim.x != ((extent_row_ - 1 + total_rows) / total_rows)); // If this is not optimal grid size perform mutex based gmem reduce. if (useGmemMutex) { // single lock per block for multiple rows if (threadIdx.x == 0 && block_start_row_first_tile_ < extent_row_) { // acquire mutex lock. unsigned int ns = 8; while (atomicCAS(user_params.mutexes_ + mutex_id, 0, 1) == 1) { __nanosleep(ns); if (ns < 256) { ns *= 2; } } } } __syncthreads(); for (int row = threadIdx.x; row < total_rows; row += blockDim.x) { if (block_start_row_first_tile_ + row < extent_row_) { user_params.red_op_( block_start_row_first_tile_ + row, &gmem_ptr[row], shared_elem_arr[row]); } } if (useGmemMutex) { __threadfence(); __syncthreads(); if (threadIdx.x == 0 && block_start_row_first_tile_ < extent_row_) { // release mutex lock. atomicExch(user_params.mutexes_ + mutex_id, 0); } } } } /// Adds a pointer offset in units of Element CUTLASS_HOST_DEVICE void add_pointer_offset(LongIndex pointer_offset) { byte_pointer_ += pointer_offset * sizeof_bits<Element>::value / 8; } /// Performs reduction and Stores a reduced output to memory CUTLASS_DEVICE void store_with_byte_offset(Fragment& frag, int64_t byte_offset) const { AccessTypeValT* frag_ptr = reinterpret_cast<AccessTypeValT*>(&frag); cg::thread_block cta = cg::this_thread_block(); // tile_width 16 is required if kElementPerAccess > 1 constexpr int tile_width = (32 / ThreadMap::Delta::kColumn) ? 32 : 16; cg::thread_block_tile<tile_width> tile32 = cg::tiled_partition<tile_width>(cta); EpilogueOpParams const& user_params = params_.user_param; using cg_reduce_t = decltype(user_params.cg_reduce_op); using tile32_t = decltype(tile32); Element* shared_elem_arr = shared_storage_.data(); constexpr auto maxVal = std::numeric_limits<OutValT>::max(); CUTLASS_PRAGMA_UNROLL for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) { CUTLASS_PRAGMA_UNROLL for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) { CUTLASS_PRAGMA_UNROLL for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) { int frag_row_idx = (row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster)); int row_offset = row * ThreadMap::Delta::kRow + group * ThreadMap::Delta::kGroup + cluster * ThreadMap::Delta::kCluster; const OutIdxT row_id = row_offset + thread_start_row_; bool row_guard = (row_id < extent_row_); const int frag_idx = frag_row_idx * ThreadMap::Iterations::kColumn * kElementsPerAccess; Element red_val; user_params.red_op_.init(&red_val, maxVal); if (row_guard) { const int iter_row = (row_id % total_rows); const auto prev_red_val = user_params.red_op_.get_value(shared_elem_arr[iter_row]); CUTLASS_PRAGMA_UNROLL for (int column = 0; column < ThreadMap::Iterations::kColumn * kElementsPerAccess; ++column) { int columnPerAccess = column / kElementsPerAccess; int columnWithPerAccess = column % kElementsPerAccess; bool guard = mask_.predicates[column]; if (guard) { const OutIdxT key_id = thread_start_column_ + ThreadMap::Delta::kColumn * columnPerAccess + columnWithPerAccess; const int frag_col_idx = frag_idx + column; Element this_val; user_params.red_op_.init(&this_val, (*frag_ptr)[frag_col_idx]); user_params.red_op_.init_key(this_val, key_id); user_params.red_op_(row_id, &red_val, this_val); } } // select_reduce doesn't need to use `red_op_` as at the warp level we use cg_reduce_op, // this satisfies the requirement of mst/single linkage of checking colors buffer. select_reduce<cg_reduce_t, tile32_t, OutIdxT, OutValT, Element> red_obj( red_val, prev_red_val, user_params.cg_reduce_op, tile32, shared_elem_arr[iter_row]); } } } } } /// Stores a fragment to memory CUTLASS_DEVICE void store(Fragment& frag) const { store_with_byte_offset(frag, 0); } CUTLASS_DEVICE MatrixCoord thread_start() const { return MatrixCoord(thread_start_row_, thread_start_column_); } /// Need to get the thread start row from the tile iterator CUTLASS_DEVICE int32_t thread_start_row() const { return thread_start_row_; } /// Need to get the thread start row from the tile iterator CUTLASS_DEVICE int32_t thread_start_column() const { return thread_start_column_; } /// Extent of the matrix in rows CUTLASS_DEVICE Index extent_row() const { return extent_row_; } /// Extent of the matrix in columns CUTLASS_DEVICE Index extent_column() const { return extent_column_; } /// Advances to the next position to load or store CUTLASS_HOST_DEVICE PredicatedTileIteratorReducedVec& operator++() { ++state_[0]; if (!ScatterD) { byte_pointer_ += params_.advance_row; } thread_start_row_ += ThreadMap::Shape::kRow; if (state_[0] == ThreadMap::Count::kRow) { state_[0] = 0; ++state_[1]; byte_pointer_ += params_.advance_group; thread_start_row_ += (ThreadMap::Shape::kGroup - 1) * ThreadMap::Shape::kRow * ThreadMap::Count::kRow; if (state_[1] == ThreadMap::Count::kGroup) { state_[1] = 0; ++state_[2]; byte_pointer_ += params_.advance_cluster; thread_start_row_ += ThreadMap::Count::kGroup * ThreadMap::Shape::kGroup * ThreadMap::Count::kRow * ThreadMap::Shape::kRow; if (state_[2] == ThreadMap::Count::kCluster) { state_[2] = 0; byte_pointer_ += params_.advance_tile; } } } return *this; } ///< Efficiently disables all accesses guarded by mask CUTLASS_DEVICE void clear_mask() { mask_.clear(); } ///< Efficiently enables all accesses guarded by mask CUTLASS_DEVICE void enable_mask() { mask_.enable(); } ///< Sets the mask CUTLASS_DEVICE void get_mask(Mask& mask) const { mask = mask_; } ///< Sets the mask CUTLASS_DEVICE void set_mask(Mask const& mask) { mask_ = mask; } }; /////////////////////////////////////////////////////////////////////////////// } // namespace threadblock } // namespace epilogue } // namespace cutlass ////////////////////////////////////////////////////////////////////////////////
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/fused_distance_nn/epilogue.cuh
/*************************************************************************************************** * Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved. * SPDX-License-Identifier: BSD-3-Clause * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, this * list of conditions and the following disclaimer. * * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * 3. Neither the name of the copyright holder nor the names of its * contributors may be used to endorse or promote products derived from * this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * **************************************************************************************************/ /* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ /*! \file \brief Epilogue for threadblock scoped GEMMs using Tensor Ops. This is adapted from DefaultEpilogueWithBroadcastTensorOp from CUTLASS 2.9.0 (https://github.com/NVIDIA/cutlass/blob/master/include/cutlass/epilogue/threadblock/default_epilogue_with_broadcast.h#L75) This epilogue allows us to load norm buffers using PredicatedTileIteratorNormVec and EpilogueWithBroadcast used for distances L2/cosine as well as applies user-define elementwise operation. -- A norm load is provided PredicatedTileIteratorNormVec -- B norm load is provided by EpilogueWithBroadcast -- elementwise operation is provided by OutputOp */ #pragma once #include <cutlass/array.h> #include <cutlass/cutlass.h> #include <cutlass/numeric_types.h> #include <cutlass/gemm/gemm.h> #include <cutlass/epilogue/threadblock/default_epilogue_tensor_op.h> #include <cutlass/epilogue/threadblock/default_epilogue_volta_tensor_op.h> #include <cutlass/epilogue/threadblock/epilogue.h> #include <cuvs/distance/detail/fused_distance_nn/custom_epilogue_with_broadcast.h> #include <cuvs/distance/detail/fused_distance_nn/predicated_tile_iterator_normvec_smem.h> #include <cuvs/distance/detail/fused_distance_nn/predicated_tile_iterator_reduced_vec.h> //////////////////////////////////////////////////////////////////////////////// namespace cutlass { namespace epilogue { namespace threadblock { //////////////////////////////////////////////////////////////////////////////// /// Defines sensible defaults for epilogues for TensorOps. template <typename Shape, typename WarpMmaTensorOp, int PartitionsK, typename ElementOutput, typename ElementTensor, typename ElementVector, typename OutputOp, typename LayoutT, int ElementsPerAccess, bool ScatterD = false> struct FusedDistanceNNEpilogue { /// Use defaults related to the existing epilogue using Base = DefaultEpilogueTensorOp<Shape, WarpMmaTensorOp, PartitionsK, OutputOp, ElementsPerAccess>; // // Stores the result z = (y = GEMM(A, B, C), broadcast) // using RowNormTileIterator = cutlass::epilogue::threadblock:: PredicatedTileIteratorNormVecSmem<typename Base::OutputTileThreadMap, ElementOutput, LayoutT>; // // Additional tensor tile iterator - stores t = Elementwise(z) // using OutputTileIterator = cutlass::epilogue::threadblock::PredicatedTileIteratorReducedVec< typename Base::OutputTileThreadMap, ElementTensor, LayoutT, typename OutputOp::Params>; /// Define the epilogue using Epilogue = cutlass::epilogue::threadblock::EpilogueWithBroadcastCustom< Shape, WarpMmaTensorOp, PartitionsK, RowNormTileIterator, OutputTileIterator, ElementVector, typename Base::AccumulatorFragmentIterator, typename Base::WarpTileIterator, typename Base::SharedLoadIterator, OutputOp, typename Base::Padding, Base::kFragmentsPerIteration>; }; } // namespace threadblock } // namespace epilogue } // namespace cutlass ////////////////////////////////////////////////////////////////////////////////
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/fused_distance_nn/persistent_gemm.h
/*************************************************************************************************** * Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved. * SPDX-License-Identifier: BSD-3-Clause * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, this * list of conditions and the following disclaimer. * * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * 3. Neither the name of the copyright holder nor the names of its * contributors may be used to endorse or promote products derived from * this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * **************************************************************************************************/ /* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ /*! \file \brief Problem visitor for grouped GEMMs This file contains heavily customized version of GemmGrouped from CUTLASS 2.10.0 (https://github.com/NVIDIA/cutlass/blob/v2.10.0/include/cutlass/gemm/kernel/gemm_grouped.h) Changes: - adds support for only single problem size to be launched persistently where each threablock processes more than one tile of the same problem. */ #pragma once #include <cutlass/complex.h> #include <cutlass/cutlass.h> #include <cutlass/fast_math.h> #include <cutlass/gemm/gemm.h> #include <cutlass/matrix_coord.h> #include <cutlass/semaphore.h> #include <cutlass/gemm/kernel/gemm_grouped_problem_visitor.h> #include <cutlass/gemm/kernel/gemm_transpose_operands.h> #include <cutlass/layout/matrix.h> #include <cutlass/trace.h> ///////////////////////////////////////////////////////////////////////////////////////////////// namespace cutlass { namespace gemm { namespace kernel { ///////////////////////////////////////////////////////////////////////////////////////////////// template <typename Mma_, ///! Threadblock-scoped matrix multiply-accumulate typename Epilogue_, ///! Epilogue typename ThreadblockSwizzle_, ///! Threadblock swizzling function GroupScheduleMode GroupScheduleMode_, ///! Type of scheduling to perform bool Transposed = false> struct FusedDistanceNNPersistent { public: using Mma = Mma_; using Epilogue = Epilogue_; using EpilogueOutputOp = typename Epilogue::OutputOp; using ThreadblockSwizzle = ThreadblockSwizzle_; static GroupScheduleMode const kGroupScheduleMode = GroupScheduleMode_; static bool const kTransposed = Transposed; // Optional transpose using MapArguments = kernel::detail::MapArguments<typename Mma::IteratorA::Element, typename Mma::IteratorA::Layout, Mma::kTransformA, Mma::IteratorA::AccessType::kElements, typename Mma::IteratorB::Element, typename Mma::IteratorB::Layout, Mma::kTransformB, Mma::IteratorB::AccessType::kElements, typename Mma::LayoutC, kTransposed>; // Public-facing type definitions related to operand element type, layout, and complex conjugate // operation. Must interact with the 'kTransposed' notion. using ElementA = typename MapArguments::ElementA; using LayoutA = typename MapArguments::LayoutA; using ElementB = typename MapArguments::ElementB; using LayoutB = typename MapArguments::LayoutB; using ElementC = typename Epilogue::OutputTileIterator::Element; using LayoutC = typename MapArguments::LayoutC; static ComplexTransform const kTransformA = MapArguments::kTransformA; static ComplexTransform const kTransformB = MapArguments::kTransformB; // Type definitions about the mainloop. using Operator = typename Mma::Operator; using OperatorClass = typename Mma::Operator::OperatorClass; using ThreadblockShape = typename Mma::Shape; using WarpShape = typename Mma::Operator::Shape; using InstructionShape = typename Mma::Policy::Operator::InstructionShape; using ArchTag = typename Mma::ArchTag; static int const kStages = Mma::kStages; static int const kAlignmentA = MapArguments::kAlignmentA; static int const kAlignmentB = MapArguments::kAlignmentB; static int const kAlignmentC = Epilogue::OutputTileIterator::kElementsPerAccess; /// Warp count (concept: GemmShape) using WarpCount = typename Mma::WarpCount; static int const kThreadCount = 32 * WarpCount::kCount; using ProblemVisitor = GemmGroupedProblemVisitor<ThreadblockShape, kGroupScheduleMode, kThreadCount, kThreadCount, kTransposed>; // // Structures // struct temp_problem_visitor { int problem_count; CUTLASS_HOST_DEVICE temp_problem_visitor() : problem_count(0){}; CUTLASS_HOST_DEVICE temp_problem_visitor(int problem_count_) : problem_count(problem_count_){}; }; /// Argument structure struct Arguments { // // Data members // GemmCoord problem_sizes; temp_problem_visitor problem_visitor; int problem_count; int threadblock_count; typename EpilogueOutputOp::Params output_op; void const* ptr_A; void const* ptr_B; void const* ptr_C; void* ptr_Vector; void* ptr_Tensor; typename LayoutA::Stride::Index lda; typename LayoutB::Stride::Index ldb; typename LayoutC::Stride::Index ldc; typename LayoutC::Stride::Index ldt; // Only used by device-level operator GemmCoord* host_problem_sizes; // // Methods // /// Default ctor CUTLASS_HOST_DEVICE Arguments() : // problem_count(0), threadblock_count(0), ptr_A(nullptr), ptr_B(nullptr), ptr_C(nullptr), ptr_Vector(nullptr), ptr_Tensor(nullptr), lda(0), ldb(0), ldc(0), ldt(0), host_problem_sizes(nullptr) { } /// Ctor CUTLASS_HOST_DEVICE Arguments(GemmCoord problem_sizes, int problem_count, int threadblock_count, typename EpilogueOutputOp::Params output_op, void const* ptr_A, void const* ptr_B, void const* ptr_C, void* ptr_Vector, void* ptr_Tensor, typename LayoutA::Stride::Index lda, typename LayoutB::Stride::Index ldb, typename LayoutC::Stride::Index ldc, typename LayoutC::Stride::Index ldt, GemmCoord* host_problem_sizes = nullptr) : problem_sizes(problem_sizes), threadblock_count(threadblock_count), output_op(output_op), ptr_A(ptr_A), ptr_B(ptr_B), ptr_C(ptr_C), ptr_Vector(ptr_Vector), ptr_Tensor(ptr_Tensor), lda(lda), ldb(ldb), ldc(ldc), ldt(ldt), host_problem_sizes(host_problem_sizes) { problem_visitor.problem_count = problem_count; } }; // // Structure for precomputing values in host memory and passing to kernels // /// Parameters structure struct Params { // typename ProblemVisitor::Params problem_visitor; temp_problem_visitor problem_visitor; int threadblock_count; typename Mma::IteratorA::Params params_A; typename Mma::IteratorB::Params params_B; typename Epilogue::OutputTileIterator::Params params_C; typename Epilogue::TensorTileIterator::Params params_Tensor; typename EpilogueOutputOp::Params output_op; void* ptr_A; void* ptr_B; void* ptr_C; void* ptr_Vector; void* ptr_Tensor; GemmCoord problem_size; typename LayoutA::Stride::Index lda; typename LayoutB::Stride::Index ldb; typename LayoutC::Stride::Index ldc; typename LayoutC::Stride::Index ldt; // // Methods // CUTLASS_HOST_DEVICE Params() : params_A(0), params_B(0), params_C(0), ptr_A(nullptr), ptr_B(nullptr), ptr_C(nullptr), ptr_Vector(nullptr), ptr_Tensor(nullptr), lda(0), ldb(0), ldc(0), ldt(0) { } CUTLASS_HOST_DEVICE Params(Arguments const& args, void* workspace = nullptr, int tile_count = 0) : problem_size(args.problem_sizes), threadblock_count(args.threadblock_count), output_op(args.output_op), params_A(args.lda), params_B(args.ldb), params_C(args.ldc), // Here we pass additional user args via args.output_op // to the reduction output tile iterator params_Tensor(args.ldt, args.output_op), ptr_A(const_cast<void*>(args.ptr_A)), ptr_B(const_cast<void*>(args.ptr_B)), ptr_C(const_cast<void*>(args.ptr_C)), ptr_Vector(args.ptr_Vector), ptr_Tensor(args.ptr_Tensor), lda(args.lda), ldb(args.ldb), ldc(args.ldc), ldt(args.ldt) { problem_visitor.problem_count = args.problem_visitor.problem_count; } CUTLASS_HOST_DEVICE void update(Arguments const& args, void* workspace = nullptr, int tile_count = 0) { threadblock_count = args.threadblock_count; output_op = args.output_op; ptr_A = const_cast<void*>(args.ptr_A); ptr_B = const_cast<void*>(args.ptr_B); ptr_C = const_cast<void*>(args.ptr_C); ptr_Vector = args.ptr_Vector; ptr_Tensor = args.ptr_Tensor; lda = args.lda; ldb = args.ldb; ldc = args.ldc; ldt = args.ldt; problem_size = args.problem_sizes; } }; /// Shared memory storage structure struct SharedStorage { union { typename Mma::SharedStorage main_loop; typename Epilogue::SharedStorage epilogue; } kernel; typename Epilogue::TensorTileIterator::SharedStorage reduced_store; typename Epilogue::OutputTileIterator::SharedStorage rownorm_store; }; public: // // Methods // CUTLASS_DEVICE FusedDistanceNNPersistent() {} /// Determines whether kernel satisfies alignment static Status can_implement(cutlass::gemm::GemmCoord const& problem_size) { return Status::kSuccess; } static Status can_implement(Arguments const& args) { return Status::kSuccess; } static size_t get_extra_workspace_size(Arguments const& args, cutlass::gemm::GemmCoord const& grid_tiled_shape) { return 0; } CUTLASS_DEVICE static uint32_t tile_count(const cutlass::MatrixCoord& grid) { return grid.row() * grid.column(); } /// Get the grid shape CUTLASS_DEVICE static cutlass::MatrixCoord grid_shape(const cutlass::gemm::GemmCoord& problem) { return cutlass::MatrixCoord(((problem.m() - 1 + ThreadblockShape::kM) / ThreadblockShape::kM), ((problem.n() - 1 + ThreadblockShape::kN) / ThreadblockShape::kN)); } /// Executes one GEMM CUTLASS_DEVICE void operator()(Params const& params, SharedStorage& shared_storage) { #if __CUDA_ARCH__ >= 800 // // These types shadow the type-level definitions and support the ability to implement // a 'transposed' GEMM that computes the transposed problems. // using ElementA = typename Mma::IteratorA::Element; using LayoutA = typename Mma::IteratorA::Layout; using ElementB = typename Mma::IteratorB::Element; using LayoutB = typename Mma::IteratorB::Layout; using ElementC = typename Epilogue::OutputTileIterator::Element; using LayoutC = typename Epilogue::OutputTileIterator::Layout; const GemmCoord& problem_size = params.problem_size; const auto grid_shape_ = grid_shape(problem_size); const uint32_t problem_chunk = (tile_count(grid_shape_) - 1 + gridDim.x) / gridDim.x; const uint32_t problem_chunk_end = blockIdx.x * problem_chunk + problem_chunk; typename LayoutB::Index column = ((blockIdx.x * problem_chunk) % grid_shape_.column()) * Mma::Shape::kN; typename LayoutB::Index row = ((blockIdx.x * problem_chunk) / grid_shape_.column()) * Mma::Shape::kM; if (column) { shared_storage.reduced_store.initSmem(params.output_op); shared_storage.rownorm_store.initSmem(params.ptr_C, problem_size.m(), row, sizeof(ElementC)); } // Outer 'persistent' loop to iterate over tiles for (uint32_t tile_idx = blockIdx.x * problem_chunk; tile_idx < problem_chunk_end; tile_idx++) { const auto grid_shape_ = grid_shape(problem_size); cutlass::MatrixCoord threadblock_offset( int(tile_idx / grid_shape_.column()) * Mma::Shape::kM, int(tile_idx % grid_shape_.column()) * Mma::Shape::kN); const bool isNextTile = ((tile_idx + 1) < problem_chunk_end); const bool doesRowChange = ((threadblock_offset.column() + Mma::Shape::kN) >= problem_size.n()); const bool do_gmem_reduce = (doesRowChange || !isNextTile) ? true : false; ElementA* ptr_A = static_cast<ElementA*>(params.ptr_A); ElementB* ptr_B = static_cast<ElementB*>(params.ptr_B); // Compute initial location in logical coordinates cutlass::MatrixCoord tb_offset_A{threadblock_offset.row(), 0}; cutlass::MatrixCoord tb_offset_B{0, threadblock_offset.column()}; // Compute position within threadblock int thread_idx = threadIdx.x; // Construct iterators to A and B operands typename Mma::IteratorA iterator_A( params.params_A, ptr_A, {problem_size.m(), problem_size.k()}, thread_idx, tb_offset_A); typename Mma::IteratorB iterator_B( params.params_B, ptr_B, {problem_size.k(), problem_size.n()}, thread_idx, tb_offset_B); // Broadcast the warp_id computed by lane 0 to ensure dependent code // is compiled as warp-uniform. int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0); int lane_idx = threadIdx.x % 32; // // Matrix multiply phase // // Construct thread-scoped matrix multiply Mma mma(shared_storage.kernel.main_loop, thread_idx, warp_idx, lane_idx); typename Mma::FragmentC accumulators; accumulators.clear(); // Compute threadblock-scoped matrix multiply-add int gemm_k_iterations = (problem_size.k() + Mma::Shape::kK - 1) / Mma::Shape::kK; // Wait for all threads to finish their epilogue phases from the previous tile. //__syncthreads(); // Compute threadblock-scoped matrix multiply-add mma(gemm_k_iterations, accumulators, iterator_A, iterator_B, accumulators); // // Epilogue // EpilogueOutputOp output_op(params.output_op); ElementC* ptr_C = static_cast<ElementC*>(params.ptr_C); typename Epilogue::ElementTensor* ptr_Tensor = static_cast<typename Epilogue::ElementTensor*>(params.ptr_Tensor); // Define the reduction output pointer and move to the appropriate place typename Epilogue::ElementVector* ptr_Vector = static_cast<typename Epilogue::ElementVector*>(params.ptr_Vector); // Tile iterator loading from source tensor. typename Epilogue::OutputTileIterator iterator_rownorm(shared_storage.rownorm_store, params.params_C, ptr_C, problem_size.mn(), thread_idx, threadblock_offset); // Additional tensor to load from typename Epilogue::TensorTileIterator tensor_iterator(shared_storage.reduced_store, params.params_Tensor, // Only the final block outputs Tensor ptr_Tensor, problem_size.mn(), thread_idx, do_gmem_reduce, threadblock_offset); Epilogue epilogue(shared_storage.kernel.epilogue, thread_idx, warp_idx, lane_idx); // Execute the epilogue operator to update the destination tensor. // Move to appropriate location for this output tile if (ptr_Vector) { ptr_Vector += threadblock_offset.column(); } // Execute the epilogue operator to update the destination tensor. epilogue(output_op, ptr_Vector, // iterator_D, accumulators, iterator_rownorm, tensor_iterator, problem_size.mn(), threadblock_offset); } #endif } }; ///////////////////////////////////////////////////////////////////////////////////////////////// } // namespace kernel } // namespace gemm } // namespace cutlass /////////////////////////////////////////////////////////////////////////////////////////////////
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/fused_distance_nn/predicated_tile_iterator_normvec_smem.h
/*************************************************************************************************** * Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved. * SPDX-License-Identifier: BSD-3-Clause * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, this * list of conditions and the following disclaimer. * * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * 3. Neither the name of the copyright holder nor the names of its * contributors may be used to endorse or promote products derived from * this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * **************************************************************************************************/ /* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ /*! \file \brief Epilogue for threadblock scoped GEMMs using Tensor Ops. This file contains a customized version of PredicatedTileIterator from CUTLASS 2.9.0 (https://github.com/NVIDIA/cutlass/blob/v2.9.0/include/cutlass/epilogue/threadblock/predicated_tile_iterator.h#L75) Changes: - added `Layout_` template param - Only the row index is used to load the data in load_with_byte_offset(). This way the same normalization data is used across all columns in a row. */ #pragma once #include <cutlass/arch/arch.h> #include <cutlass/arch/memory.h> #include <cutlass/array.h> #include <cutlass/cutlass.h> #include <cutlass/epilogue/threadblock/output_tile_thread_map.h> #include <cutlass/epilogue/threadblock/predicated_tile_iterator_params.h> #include <cutlass/layout/matrix.h> #include <cutlass/layout/tensor.h> #include <cutlass/matrix_shape.h> #include <cutlass/numeric_types.h> #include <cutlass/tensor_ref.h> #include <cutlass/transform/pitch_linear_thread_map.h> #include <raft/util/device_loads_stores.cuh> //////////////////////////////////////////////////////////////////////////////// namespace cutlass { //////////////////////////////////////////////////////////////////////////////// namespace epilogue { namespace threadblock { //////////////////////////////////////////////////////////////////////////////// /// Tile iterator used to load and store output tile from global memory in epilogue. /// /// Satisfies: ReadableTileIterator | PredicatedTileIterator | ForwardTileIterator /// template <typename ThreadMap_, ///< Thread map (conept: OutputTileThreadMap) typename Element_, ///< Element data type typename Layout_, bool ScatterD = false, ///< Scatter D operand or not bool UseCUDAStore = false> class PredicatedTileIteratorNormVecSmem { public: using ThreadMap = ThreadMap_; using Shape = typename ThreadMap::Shape; using Element = Element_; using Layout = Layout_; using TensorRef = TensorRef<Element, Layout>; using ConstTensorRef = typename TensorRef::ConstTensorRef; using Index = typename Layout::Index; using LongIndex = typename Layout::LongIndex; using TensorCoord = MatrixCoord; static int const kElementsPerAccess = ThreadMap::kElementsPerAccess; static int const kThreads = ThreadMap::kThreads; static int const kIterations = ThreadMap::Count::kTile; static int const total_rows = ThreadMap::kWarpCount * ThreadMap::Iterations::kRow * ThreadMap::Iterations::kGroup * ThreadMap::Iterations::kCluster * ThreadMap::Count::kTile * ThreadMap::Delta::kRow; static_assert(ThreadMap::Iterations::kRow > 0, "ThreadMap::Iterations::kRow must be > 0"); static_assert(ThreadMap::Iterations::kGroup > 0, "ThreadMap::Iterations::kGroup must be > 0"); static_assert(ThreadMap::Iterations::kCluster > 0, "ThreadMap::Iterations::kCluster must be > 0"); static_assert(ThreadMap::Iterations::kColumn > 0, "ThreadMap::Iterations::kColumn must be > 0"); using Fragment = Array<Element, ThreadMap::Iterations::kRow * ThreadMap::Iterations::kGroup * ThreadMap::Iterations::kCluster * ThreadMap::kElementsPerAccess>; /// Memory access size using AccessType = AlignedArray<Element, ThreadMap::kElementsPerAccess>; // // Parameters struct // /// Uses a non-template class struct Params : PredicatedTileIteratorParams { using Base = PredicatedTileIteratorParams; CUTLASS_HOST_DEVICE Params() {} CUTLASS_HOST_DEVICE Params(Layout const& layout) : PredicatedTileIteratorParams( layout.stride(0) * int(sizeof(AccessType)) / kElementsPerAccess, make_OutputTileThreadMapDesc<ThreadMap>()) { } CUTLASS_HOST_DEVICE Params(Base const& base) : Base(base) {} }; /// Mask object struct Mask { static int const kCount = ThreadMap::Iterations::kColumn; /// Predicate state bool predicates[kCount]; // // Mask // CUTLASS_HOST_DEVICE Mask() { enable(); } ///< Efficiently disables all accesses guarded by mask CUTLASS_HOST_DEVICE void clear() { CUTLASS_PRAGMA_UNROLL for (int i = 0; i < kCount; ++i) { predicates[i] = false; } } ///< CUTLASS_HOST_DEVICE enables all accesses guarded by mask CUTLASS_DEVICE void enable() { CUTLASS_PRAGMA_UNROLL for (int i = 0; i < kCount; ++i) { predicates[i] = true; } } }; /// Shared storage allocation needed by the predicated tile // iterator for storing rowNorm chunk. struct SharedStorage { // // Type definitions // using Shape = MatrixShape<total_rows, 1>; /// Shape of the shared memory allocation using StorageShape = MatrixShape<Shape::kRow, Shape::kColumn>; // // Data members // // Methods // AlignedBuffer<Element, StorageShape::kCount> storage; CUTLASS_DEVICE Element* data() { return storage.data(); } SharedStorage() {} CUTLASS_DEVICE void initSmem(void* pointer, const Index& num_rows, const Index& tb_row_offset, const LongIndex& stride) { Element* shared_elem_arr = data(); uint8_t* first_tile_byte_pointer_ = reinterpret_cast<uint8_t*>(pointer) + LongIndex(tb_row_offset) * LongIndex(stride); const auto gmem_ptr = reinterpret_cast<Element*>(first_tile_byte_pointer_); for (int row = threadIdx.x; row < total_rows; row += blockDim.x) { bool guard = (tb_row_offset + row) < num_rows; cutlass::arch::cp_async<sizeof(Element)>(shared_elem_arr + row, gmem_ptr + row, guard); cutlass::arch::cp_async_wait<0>(); } } }; private: // // Data members // /// Parameters structure containing reference and precomputed state. PredicatedTileIteratorParams params_; /// Byte-level pointer uint8_t* byte_pointer_; /// Array of boolean values to contain steady-state predicates Mask mask_; /// Extent of the matrix tile in rows Index extent_row_; /// Extent of the matrix tile in rows Index extent_column_; /// A thread's starting row position (assuming steady-state predicates have been computed) Index thread_start_row_; /// A thread's starting column Index thread_start_column_; /// Internal state counter int state_[3]; /// Scatter indices int const* indices_; // // Static asserts about internal strides // static_assert(sizeof(extent_row_) == 4, "Expected 32b extents"); static_assert(sizeof(thread_start_row_) == 4, "Expected 32b extents"); static_assert(sizeof(PredicatedTileIteratorParams::stride) == 8, "Expected 64b strides"); private: // // Methods // protected: SharedStorage& shared_storage_; public: // // Methods // /// Constructor CUTLASS_DEVICE PredicatedTileIteratorNormVecSmem(SharedStorage& shared_storage, PredicatedTileIteratorParams const& params, Element* pointer, TensorCoord extent, int thread_idx, TensorCoord& threadblock_offset, int const* indices = nullptr) : params_(params), indices_(indices), shared_storage_(shared_storage) { TensorCoord thread_offset = ThreadMap::initial_offset(thread_idx) + threadblock_offset; extent_row_ = extent.row(); extent_column_ = extent.column(); thread_start_row_ = thread_offset.row(); thread_start_column_ = thread_offset.column(); // Initialize predicates CUTLASS_PRAGMA_UNROLL for (int c = 0; c < ThreadMap::Iterations::kColumn; ++c) { mask_.predicates[c] = ((thread_offset.column() + ThreadMap::Delta::kColumn * c) < extent.column()); } // Null pointer performs no accesses if (!pointer) { mask_.clear(); return; } if (ScatterD && !indices) { mask_.clear(); } // Initialize pointer byte_pointer_ = reinterpret_cast<uint8_t*>(pointer) + LongIndex(thread_offset.row()) * LongIndex(params_.stride); if (ScatterD) { byte_pointer_ = reinterpret_cast<uint8_t*>(pointer) + LongIndex(thread_offset.column()) * sizeof(AccessType) / kElementsPerAccess; } if (threadblock_offset.column() == 0) { shared_storage_.initSmem(pointer, extent_row_, threadblock_offset.row(), params_.stride); } // Initialize internal state counter state_[0] = state_[1] = state_[2] = 0; } /// Adds a pointer offset in units of Element CUTLASS_HOST_DEVICE void add_pointer_offset(LongIndex pointer_offset) { byte_pointer_ += pointer_offset * sizeof_bits<Element>::value / 8; } /// Loads a fragment from memory CUTLASS_DEVICE void load_with_byte_offset(Fragment& frag, int64_t byte_offset) const { AccessType* frag_ptr = reinterpret_cast<AccessType*>(&frag); Element* shared_elem_arr = shared_storage_.data(); CUTLASS_PRAGMA_UNROLL for (int cluster = 0; cluster < ThreadMap::Iterations::kCluster; ++cluster) { CUTLASS_PRAGMA_UNROLL for (int group = 0; group < ThreadMap::Iterations::kGroup; ++group) { CUTLASS_PRAGMA_UNROLL for (int row = 0; row < ThreadMap::Iterations::kRow; ++row) { int frag_row_idx = (row + ThreadMap::Iterations::kRow * (group + ThreadMap::Iterations::kGroup * cluster)); int row_offset = row * ThreadMap::Delta::kRow + group * ThreadMap::Delta::kGroup + cluster * ThreadMap::Delta::kCluster; int iter_row = ((row_offset + thread_start_row_) % total_rows); Element val = shared_elem_arr[iter_row]; CUTLASS_PRAGMA_UNROLL for (int i = 0; i < kElementsPerAccess; ++i) { (*frag_ptr)[frag_row_idx + i] = val; } } } } } /// Loads a fragment from memory CUTLASS_DEVICE void load(Fragment& frag) const { load_with_byte_offset(frag, 0); } CUTLASS_DEVICE MatrixCoord thread_start() const { return MatrixCoord(thread_start_row_, thread_start_column_); } /// Need to get the thread start row from the tile iterator CUTLASS_DEVICE int32_t thread_start_row() const { return thread_start_row_; } /// Need to get the thread start row from the tile iterator CUTLASS_DEVICE int32_t thread_start_column() const { return thread_start_column_; } /// Extent of the matrix in rows CUTLASS_DEVICE Index extent_row() const { return extent_row_; } /// Extent of the matrix in columns CUTLASS_DEVICE Index extent_column() const { return extent_column_; } /// Advances to the next position to load or store CUTLASS_HOST_DEVICE PredicatedTileIteratorNormVecSmem& operator++() { ++state_[0]; if (!ScatterD) { byte_pointer_ += params_.advance_row; } thread_start_row_ += ThreadMap::Shape::kRow; if (state_[0] == ThreadMap::Count::kRow) { state_[0] = 0; ++state_[1]; byte_pointer_ += params_.advance_group; thread_start_row_ += (ThreadMap::Shape::kGroup - 1) * ThreadMap::Shape::kRow * ThreadMap::Count::kRow; if (state_[1] == ThreadMap::Count::kGroup) { state_[1] = 0; ++state_[2]; byte_pointer_ += params_.advance_cluster; thread_start_row_ += ThreadMap::Count::kGroup * ThreadMap::Shape::kGroup * ThreadMap::Count::kRow * ThreadMap::Shape::kRow; if (state_[2] == ThreadMap::Count::kCluster) { state_[2] = 0; byte_pointer_ += params_.advance_tile; } } } return *this; } ///< Efficiently disables all accesses guarded by mask CUTLASS_DEVICE void clear_mask() { mask_.clear(); } ///< Efficiently enables all accesses guarded by mask CUTLASS_DEVICE void enable_mask() { mask_.enable(); } ///< Sets the mask CUTLASS_DEVICE void get_mask(Mask& mask) const { mask = mask_; } ///< Sets the mask CUTLASS_DEVICE void set_mask(Mask const& mask) { mask_ = mask; } }; /////////////////////////////////////////////////////////////////////////////// } // namespace threadblock } // namespace epilogue } // namespace cutlass ////////////////////////////////////////////////////////////////////////////////
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/fused_distance_nn/gemm.h
/*************************************************************************************************** * Copyright (c) 2017 - 2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved. * SPDX-License-Identifier: BSD-3-Clause * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, this * list of conditions and the following disclaimer. * * 2. Redistributions in binary form must reproduce the above copyright notice, * this list of conditions and the following disclaimer in the documentation * and/or other materials provided with the distribution. * * 3. Neither the name of the copyright holder nor the names of its * contributors may be used to endorse or promote products derived from * this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * **************************************************************************************************/ /* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cutlass/cutlass.h> #include <cutlass/gemm/kernel/default_gemm_universal.h> #include <cutlass/layout/matrix.h> #include <cutlass/layout/tensor.h> #include <cuvs/distance/detail/fused_distance_nn/epilogue.cuh> #include <cuvs/distance/detail/fused_distance_nn/persistent_gemm.h> ///////////////////////////////////////////////////////////////////////////////////////////////// namespace cutlass { namespace gemm { namespace kernel { ///////////////////////////////////////////////////////////////////////////////////////////////// /* * This configuration is used for float inputs with veclen(kAlignmentA/B) = 2 or 4, * ideal threadblock tile shape is 32x256x16 for such cases as there is no * registers spills for it. * */ template < /// Element type for A matrix operand typename ElementA_, /// Layout type for A matrix operand int kAlignmentA, /// Element type for B matrix operand typename ElementB_, /// Layout type for B matrix operand int kAlignmentB, /// Element type for C and D matrix operands typename ElementC_, /// Element type for internal accumulation typename ElementAccumulator, /// Epilogue output operator - must satisfy concept of 'EpilogueWithBroadcastOp' typename EpilogueOutputOp, /// Number of stages used in the pipelined mainloop int Stages, /// data layout row/column major of inputs bool isRowMajor> struct FusedDistanceNNGemm { // This struct is specialized for fp32/3xTF32 /// Threadblock-level tile size (concept: GemmShape) // <- threadblock tile M = 32, N = 256, K = 16 // this is more performant but note that for veclen = 1 // this shape has register spills using ThreadblockShape = cutlass::gemm::GemmShape<32, 256, 16>; // <- threadblock tile M = 32, N = 128, K = 16 // this shape has high occupancy but less perf // this is less performant but this shape has *no* register spills // for any veclens(1, 2, 4) // using ThreadblockShape = cutlass::gemm::GemmShape<32, 128, 16>; /// Warp-level tile size (concept: GemmShape) // This code section describes tile size a warp will compute // <- warp tile M = 64, N = 64, K = 16 // this is more performant for veclen 2,4. using WarpShape = cutlass::gemm::GemmShape<32, 64, 16>; // this shape has high occupancy but less perf used for 32x128x16 // using WarpShape = cutlass::gemm::GemmShape<32, 32, 16>; /// Warp-level tile size (concept: GemmShape) // This code section describes the size of MMA op // <- MMA Op tile M = 16, N = 8, K = 4 using InstructionShape = cutlass::gemm::GemmShape<16, 8, 4>; /// Operation performed by GEMM using Operator = cutlass::arch::OpMultiplyAddFastF32; // using Operator = cutlass::arch::OpMultiplyAdd; // this runs only 1xTF32 for float inputs // This code section describes whether you want to use tensor cores or regular SIMT cores on GPU // SM using OperatorClass = cutlass::arch::OpClassTensorOp; // This code section describes CUDA SM architecture number using ArchTag = cutlass::arch::Sm80; // This code section describes how threadblocks are scheduled on GPU /// Threadblock-level swizzling operator using ThreadblockSwizzle = cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>; /// data layout for final output matrix. // we keep this same layout even for column major inputs using LayoutOutput = cutlass::layout::RowMajor; typedef typename std::conditional<isRowMajor, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::type NormXLayout; typedef typename std:: conditional<isRowMajor, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::type LayoutA_; typedef typename std:: conditional<isRowMajor, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::type LayoutB_; using GemmBase = typename DefaultGemmUniversal<ElementA_, LayoutA_, cutlass::ComplexTransform::kNone, kAlignmentA, ElementB_, LayoutB_, cutlass::ComplexTransform::kNone, kAlignmentB, ElementC_, LayoutOutput, ElementAccumulator, OperatorClass, ArchTag, ThreadblockShape, WarpShape, InstructionShape, EpilogueOutputOp, ThreadblockSwizzle, Stages, Operator>::GemmKernel; // Replace epilogue using Epilogue = typename cutlass::epilogue::threadblock::FusedDistanceNNEpilogue< typename GemmBase::Epilogue::Shape, typename GemmBase::Epilogue::WarpMmaOperator, GemmBase::Epilogue::kPartitionsK, ElementAccumulator, typename EpilogueOutputOp::ElementT, ElementAccumulator, EpilogueOutputOp, NormXLayout, GemmBase::Epilogue::kElementsPerAccess>::Epilogue; // Compose the GEMM kernel using GemmKernel = FusedDistanceNNPersistent<typename GemmBase::Mma, Epilogue, ThreadblockSwizzle, GroupScheduleMode::kDeviceOnly>; }; /* * This configuration is used for float inputs with veclen(kAlignmentA/B) = 1, * ideal threadblock tile shape is 32x128x16 for such cases as there is no * registers spills for it. * */ template < /// Element type for C and D matrix operands typename ElementC_, /// Element type for internal accumulation typename ElementAccumulator, /// Epilogue output operator - must satisfy concept of 'EpilogueWithBroadcastOp' typename EpilogueOutputOp, /// Number of stages used in the pipelined mainloop int Stages, /// data layout row/column major of inputs bool isRowMajor> struct FusedDistanceNNGemm<float, /// Element type for A matrix operand 1, /// Layout type (veclen) for A matrix operand float, /// Element type for B matrix operand 1, /// Layout type (veclen) for B matrix operand ElementC_, ElementAccumulator, EpilogueOutputOp, Stages, isRowMajor> { // This struct is specialized for fp32/3xTF32 using ElementA_ = float; using ElementB_ = float; /// Threadblock-level tile size (concept: GemmShape) // <- threadblock tile M = 32, N = 128, K = 16 // this shape has high occupancy and no register spills for veclen = 1. using ThreadblockShape = cutlass::gemm::GemmShape<32, 128, 16>; /// Warp-level tile size (concept: GemmShape) // This code section describes tile size a warp will compute // <- warp tile M = 32, N = 32, K = 16 using WarpShape = cutlass::gemm::GemmShape<32, 32, 16>; /// Warp-level tile size (concept: GemmShape) // This code section describes the size of MMA op // <- MMA Op tile M = 16, N = 8, K = 4 using InstructionShape = cutlass::gemm::GemmShape<16, 8, 4>; /// Operation performed by GEMM using Operator = cutlass::arch::OpMultiplyAddFastF32; // using Operator = cutlass::arch::OpMultiplyAdd; // this runs only 1xTF32 for float inputs // This code section describes whether you want to use tensor cores or regular SIMT cores on GPU // SM using OperatorClass = cutlass::arch::OpClassTensorOp; // This code section describes CUDA SM architecture number using ArchTag = cutlass::arch::Sm80; // This code section describes how threadblocks are scheduled on GPU /// Threadblock-level swizzling operator using ThreadblockSwizzle = cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>; /// data layout for final output matrix. // we keep this same layout even for column major inputs using LayoutOutput = cutlass::layout::RowMajor; typedef typename std::conditional<isRowMajor, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::type NormXLayout; typedef typename std:: conditional<isRowMajor, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::type LayoutA_; typedef typename std:: conditional<isRowMajor, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::type LayoutB_; using GemmBase = typename DefaultGemmUniversal<ElementA_, LayoutA_, cutlass::ComplexTransform::kNone, 1, ElementB_, LayoutB_, cutlass::ComplexTransform::kNone, 1, ElementC_, LayoutOutput, ElementAccumulator, OperatorClass, ArchTag, ThreadblockShape, WarpShape, InstructionShape, EpilogueOutputOp, ThreadblockSwizzle, Stages, Operator>::GemmKernel; // Replace epilogue using Epilogue = typename cutlass::epilogue::threadblock::FusedDistanceNNEpilogue< typename GemmBase::Epilogue::Shape, typename GemmBase::Epilogue::WarpMmaOperator, GemmBase::Epilogue::kPartitionsK, ElementAccumulator, typename EpilogueOutputOp::ElementT, ElementAccumulator, EpilogueOutputOp, NormXLayout, GemmBase::Epilogue::kElementsPerAccess>::Epilogue; // Compose the GEMM kernel using GemmKernel = FusedDistanceNNPersistent<typename GemmBase::Mma, Epilogue, ThreadblockSwizzle, GroupScheduleMode::kDeviceOnly>; }; template < /// Layout type for A matrix operand int kAlignmentA, /// Layout type for B matrix operand int kAlignmentB, /// Element type for C and D matrix operands typename ElementC_, /// Element type for internal accumulation typename ElementAccumulator, /// Epilogue output operator - must satisfy concept of 'EpilogueWithBroadcastOp' typename EpilogueOutputOp, /// Number of stages used in the pipelined mainloop int Stages, /// data layout row/column major of inputs bool isRowMajor> struct FusedDistanceNNGemm<double, kAlignmentA, double, kAlignmentB, ElementC_, ElementAccumulator, EpilogueOutputOp, Stages, isRowMajor> { // Threadblock-level tile size (concept: GemmShape) // <- threadblock tile M = 64, N = 64, K = 16 using ThreadblockShape = cutlass::gemm::GemmShape<64, 64, 16>; // using ThreadblockShape = cutlass::gemm::GemmShape<16, 128, 16>; /// Warp-level tile size (concept: GemmShape) // This code section describes tile size a warp will compute // <- warp tile M = 32, N = 32, K = 16 using WarpShape = cutlass::gemm::GemmShape<32, 32, 16>; // using WarpShape = cutlass::gemm::GemmShape<16, 32, 16>; /// Warp-level tile size (concept: GemmShape) // This code section describes the size of MMA op using InstructionShape = cutlass::gemm::GemmShape<8, 8, 4>; // Operation performed by GEMM using Operator = cutlass::arch::OpMultiplyAdd; // This code section describes whether you want to use tensor cores or regular SIMT cores on GPU // SM using OperatorClass = cutlass::arch::OpClassTensorOp; // This code section describes CUDA SM architecture number using ArchTag = cutlass::arch::Sm80; // This code section describes how threadblocks are scheduled on GPU /// Threadblock-level swizzling operator using ThreadblockSwizzle = cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>; /// data layout for final output matrix. // we keep this same layout even for column major inputs using LayoutOutput = cutlass::layout::RowMajor; typedef typename std::conditional<isRowMajor, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::type NormXLayout; typedef typename std:: conditional<isRowMajor, cutlass::layout::RowMajor, cutlass::layout::ColumnMajor>::type LayoutA_; typedef typename std:: conditional<isRowMajor, cutlass::layout::ColumnMajor, cutlass::layout::RowMajor>::type LayoutB_; using GemmBase = typename DefaultGemmUniversal<double, LayoutA_, cutlass::ComplexTransform::kNone, 1, double, LayoutB_, cutlass::ComplexTransform::kNone, 1, ElementC_, LayoutOutput, ElementAccumulator, OperatorClass, ArchTag, ThreadblockShape, WarpShape, InstructionShape, EpilogueOutputOp, ThreadblockSwizzle, Stages, Operator>::GemmKernel; // Replace epilogue using Epilogue = typename cutlass::epilogue::threadblock::FusedDistanceNNEpilogue< typename GemmBase::Epilogue::Shape, typename GemmBase::Epilogue::WarpMmaOperator, GemmBase::Epilogue::kPartitionsK, ElementC_, typename EpilogueOutputOp::ElementT, ElementC_, EpilogueOutputOp, NormXLayout, GemmBase::Epilogue::kElementsPerAccess>::Epilogue; // Compose the GEMM kernel using GemmKernel = FusedDistanceNNPersistent<typename GemmBase::Mma, Epilogue, ThreadblockSwizzle, GroupScheduleMode::kDeviceOnly>; }; ///////////////////////////////////////////////////////////////////////////////////////////////// } // namespace kernel } // namespace gemm } // namespace cutlass
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_matrix/dispatch_sm60.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <algorithm> // std::min #include <cuvs/distance/detail/pairwise_matrix/dispatch_layout.cuh> // dispatch_layout #include <cuvs/distance/detail/pairwise_matrix/kernel_sm60.cuh> // pairwise_matrix_sm60_wrapper #include <raft/linalg/contractions.cuh> // raft::linalg::Policy4x4 namespace cuvs::distance::detail { template <typename OpT, typename IdxT, typename DataT, typename OutT, typename FinOpT, typename SM_compat_t> pairwise_matrix_sm60_wrapper<OpT, IdxT, DataT, OutT, FinOpT> pairwise_matrix_sm60_get_wrapper( OpT distance_op, pairwise_matrix_params<IdxT, DataT, OutT, FinOpT> params, SM_compat_t sm_compat_range) { int vec_len = determine_vec_len(params); // f takes compile-time constants row_major and vec_len aligned and returns // the corresponding kernel wrapper. The wrapper contains the launch // parameters of the kernel: a pointer to the kernel function, grid size, // block size, and shared memory size. auto f = [&](auto row_major, auto vec_len_aligned) { // row_major and vec_len are std::integral_constants of type bool and int // respectively. // To keep compile times in check, we only specialize on veclen > 1 when // the inner loop is relatively cheap (< 5 flops). constexpr int vec_len_op = distance_op.expensive_inner_loop ? 1 : vec_len_aligned(); // Prevent double, vec_len=4 combination (this is not supported) constexpr int vec_len = std::min(vec_len_op, static_cast<int>(16 / sizeof(DataT))); using RowPolicy = typename raft::linalg::Policy4x4<DataT, vec_len>::Policy; using ColPolicy = typename raft::linalg::Policy4x4<DataT, vec_len>::ColPolicy; using Policy = typename std::conditional<row_major(), RowPolicy, ColPolicy>::type; auto wrapper = make_pairwise_matrix_sm60_wrapper<Policy, row_major()>(distance_op, params, sm_compat_range); return wrapper; }; // Dispatch_layout calls f with appropriate compile time constants based on // the runtime values of params.is_row_major and vec_len. return dispatch_layout(params.is_row_major, vec_len, f); } template <typename OpT, typename IdxT, typename DataT, typename OutT, typename FinOpT, typename SM_compat_t> void pairwise_matrix_sm60_dispatch(OpT distance_op, pairwise_matrix_params<IdxT, DataT, OutT, FinOpT> params, SM_compat_t sm_compat_range, cudaStream_t stream) { auto wrapper = pairwise_matrix_sm60_get_wrapper(distance_op, params, sm_compat_range); wrapper.launch(distance_op, params, stream); } } // namespace cuvs::distance::detail
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_matrix/dispatch_layout.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <algorithm> // std::min #include <cstdint> // size_t #include <cuvs/distance/detail/pairwise_matrix/params.cuh> // pairwise_matrix_params #include <raft/core/error.hpp> // RAFT_EXPECTS #include <type_traits> // std::integral_constant namespace cuvs::distance::detail { /** * @brief: Computes minimal common alignment of the rows in a 2D array in bytes * * The 2D matrix `x` is assumed to be row-major. This function computes the * minimal alignment in bytes of the first elements of each row. * Output can be 16, 8, 4, 2, 1. * * @param x Base pointer of row-major input matrix * @param stride Stride in number of element between consecutive rows. */ template <typename DataT> size_t alignment_of_2d_array(const DataT* x, size_t stride) { auto base = reinterpret_cast<uintptr_t>(x); size_t stride_bytes = sizeof(DataT) * stride; for (int align = 16; align >= 0; align /= 2) { bool base_aligned = base % align == 0; bool stride_aligned = stride_bytes % align == 0; if (base_aligned && stride_aligned) { return align; } } return 1; } /** * @brief: Computes the vec_len parameter kernel policy parameter * * @param params Kernel parameters */ template <typename IdxT, typename DataT, typename OutT, typename FinOpT> int determine_vec_len(pairwise_matrix_params<IdxT, DataT, OutT, FinOpT> params) { size_t align_x = alignment_of_2d_array(params.x, params.ldx); size_t align_y = alignment_of_2d_array(params.y, params.ldy); size_t byte_alignment = min(align_x, align_y); // Since alignment is in bytes, it could be smaller than sizeof(DataT). // Handle this (unlikely) case here. RAFT_EXPECTS(sizeof(DataT) <= byte_alignment, "Input matrix must be aligned to size of elements."); // Compute number of elements that can be loaded in one instruction // without causing misalignent errors. int vec_len_aligned = (byte_alignment % sizeof(DataT) == 0) ? byte_alignment / sizeof(DataT) : 1; // In the future, pairwise_matrix might support `int8_t` input. In that case, // byte_alignment / sizeof(DataT) might exceed 4. We maximize at 4 here, to // prevent adding more cases in dispatch_layout below (which are expensive to // compile). vec_len_aligned = std::min(vec_len_aligned, 4); return vec_len_aligned; } template <int n> using vec_len_constant = std::integral_constant<int, n>; /** * @brief: Converts run-time arguments to compile-time arguments * * Converts run-time arguments row_major and vec_len to compile-time arguments * and dispatches a lambda f with these compile-time arguments. * * This is equivalent to copying and pasting the lambda function `f` in each of * the switch case statements. * * @tparam F Type of lambda f. * @param row_major Boolean indicating whether input arrays have row-major layout. * @param vec_len Integer value 1, 2, or 4 specifying the Veclen template parameter of * the KernelPolicy. * @param f Lambda that takes two std::integral_constant parameters representing * row_major and vec_len. */ template <typename F> auto dispatch_layout(bool row_major, int vec_len, F&& f) { if (row_major) { switch (vec_len) { case 4: return f(std::true_type(), vec_len_constant<4>()); case 2: return f(std::true_type(), vec_len_constant<2>()); default: return f(std::true_type(), vec_len_constant<1>()); } } else { switch (vec_len) { case 4: return f(std::false_type(), vec_len_constant<4>()); case 2: return f(std::false_type(), vec_len_constant<2>()); default: return f(std::false_type(), vec_len_constant<1>()); } } } }; // namespace cuvs::distance::detail
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_matrix/dispatch.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #ifndef RAFT_EXPLICIT_INSTANTIATE_ONLY #include "dispatch-inl.cuh" #endif #ifdef RAFT_COMPILED #include "dispatch-ext.cuh" #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_matrix/dispatch-ext.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/distance/detail/distance_ops/all_ops.cuh> // ops::* #include <cuvs/distance/detail/distance_ops/cutlass.cuh> // ops::has_cutlass_op #include <cuvs/distance/detail/kernels/rbf_fin_op.cuh> // rbf_fin_op #include <cuvs/distance/detail/pairwise_matrix/params.cuh> // pairwise_matrix_params #include <raft/core/operators.hpp> // raft::identity_op #include <raft/util/raft_explicit.hpp> // RAFT_EXPLICIT #ifdef RAFT_EXPLICIT_INSTANTIATE_ONLY namespace cuvs::distance::detail { template <typename OpT, typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void pairwise_matrix_dispatch(OpT distance_op, IdxT m, IdxT n, IdxT k, const DataT* x, const DataT* y, const DataT* x_norm, const DataT* y_norm, OutT* out, FinOpT fin_op, cudaStream_t stream, bool is_row_major) RAFT_EXPLICIT; }; // namespace cuvs::distance::detail #endif // RAFT_EXPLICIT_INSTANTIATE_ONLY #define instantiate_raft_distance_detail_pairwise_matrix_dispatch( \ OpT, DataT, AccT, OutT, FinOpT, IdxT) \ extern template void cuvs::distance::detail:: \ pairwise_matrix_dispatch<OpT<DataT, AccT, IdxT>, DataT, AccT, OutT, FinOpT, IdxT>( \ OpT<DataT, AccT, IdxT> distance_op, \ IdxT m, \ IdxT n, \ IdxT k, \ const DataT* x, \ const DataT* y, \ const DataT* x_norm, \ const DataT* y_norm, \ OutT* out, \ FinOpT fin_op, \ cudaStream_t stream, \ bool is_row_major) /* * Hierarchy of instantiations: * * This file defines extern template instantiations of the distance kernels. The * instantiation of the public API is handled in cuvs/distance/distance-ext.cuh. * * After adding an instance here, make sure to also add the instance there. */ // The following two instances are used in the RBF kernel object. Note the use of int64_t for the // index type. instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::l2_unexp_distance_op, float, float, float, cuvs::distance::kernels::detail::rbf_fin_op<float>, int64_t); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::l2_unexp_distance_op, double, double, double, cuvs::distance::kernels::detail::rbf_fin_op<double>, int64_t); // Rest of instances instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::canberra_distance_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::canberra_distance_op, double, double, double, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::correlation_distance_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::correlation_distance_op, double, double, double, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::cosine_distance_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::cosine_distance_op, double, double, double, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::hamming_distance_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::hamming_distance_op, double, double, double, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::hellinger_distance_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::hellinger_distance_op, double, double, double, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::jensen_shannon_distance_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::jensen_shannon_distance_op, double, double, double, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::kl_divergence_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::kl_divergence_op, double, double, double, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::l1_distance_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::l1_distance_op, double, double, double, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::l2_exp_distance_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::l2_exp_distance_op, double, double, double, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::l2_unexp_distance_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::l2_unexp_distance_op, double, double, double, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::l_inf_distance_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::l_inf_distance_op, double, double, double, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::lp_unexp_distance_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::lp_unexp_distance_op, double, double, double, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::russel_rao_distance_op, float, float, float, raft::identity_op, int); instantiate_raft_distance_detail_pairwise_matrix_dispatch( cuvs::distance::detail::ops::russel_rao_distance_op, double, double, double, raft::identity_op, int); #undef instantiate_raft_distance_detail_pairwise_matrix_dispatch
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_matrix/dispatch-inl.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once /* This file has two responsibilities: * * 1. Dispatch to the correct implementation of a kernel based on the * architecture of the device on which the kernel will be launched. For * instance, the cosine distance has a CUTLASS-based implementation that can * be used on SM80+ and the normal implementation that is used on older * architectures. * * 2. Provide concise function templates that can be instantiated in * src/distance/detail/pairwise_matrix/. Previously, * cuvs::distance::detail::distance was instantiated. The function * necessarily required a large set of include files, which slowed down the * build. The cuvs::distance::detail::pairwise_matrix_arch_dispatch functions * do not require as large an include files set, which speeds up the build. */ #include <cuvs/distance/detail/distance_ops/cutlass.cuh> // ops::has_cutlass_op #include <cuvs/distance/detail/pairwise_matrix/dispatch_sm60.cuh> // dispatch_sm60 #include <cuvs/distance/detail/pairwise_matrix/params.cuh> // pairwise_matrix_params #include <raft/util/arch.cuh> // raft::util::arch::SM_* // NOTE: to minimize compile times, we do not include dispatch_sm80.cuh. // Including dispatch_sm80.cuh can slow down compile times (due to CUTLASS). // Therefore, it is the including file's responsibility to include the correct // dispatch_smXX.cuh headers, as is done in cuvs/distance/detail/distance.cuh // and src/distance/detail/pairwise_matrix/dispatch_*.cu. namespace cuvs::distance::detail { // This forward-declaration ensures that we do not need to include // dispatch_sm80.cuh if we are not calling it in practice. This makes compiling // all the non-CUTLASS based distance instantiations faster. For CUTLASS-based // distances, dispatch_sm80.cuh has to be included by the file including this // file. template <typename OpT, typename IdxT, typename DataT, typename OutT, typename FinOpT, typename SM_compat_t> void pairwise_matrix_sm80_dispatch(OpT, pairwise_matrix_params<IdxT, DataT, OutT, FinOpT>, SM_compat_t, cudaStream_t); template <typename OpT, typename DataT, typename AccT, typename OutT, typename FinOpT, typename IdxT = int> void pairwise_matrix_dispatch(OpT distance_op, IdxT m, IdxT n, IdxT k, const DataT* x, const DataT* y, const DataT* x_norm, const DataT* y_norm, OutT* out, FinOpT fin_op, cudaStream_t stream, bool is_row_major) { // Create kernel parameter struct. Flip x and y if column major. IdxT ldx = is_row_major ? k : m; IdxT ldy = is_row_major ? k : n; IdxT ld_out = is_row_major ? n : m; pairwise_matrix_params<IdxT, DataT, OutT, FinOpT> params{ m, n, k, ldx, ldy, ld_out, x, y, x_norm, y_norm, out, fin_op, is_row_major}; if (!params.is_row_major) { params.flip_x_and_y(); } // Dispatch rule: // - execute CUTLASS-based kernel on SM_80 and above // - execute normal kernel below SM_80 namespace arch = raft::util::arch; constexpr bool cutlass_op_unavailable = !ops::has_cutlass_op<OpT>(); if constexpr (cutlass_op_unavailable) { // Always execute legacy kernels when no cutlass op is available auto any_range = arch::SM_range(arch::SM_min(), arch::SM_future()); pairwise_matrix_sm60_dispatch(distance_op, params, any_range, stream); } else { auto cutlass_range = arch::SM_range(arch::SM_80(), arch::SM_future()); auto legacy_range = arch::SM_range(arch::SM_min(), arch::SM_80()); // Get pointer to SM60 kernel to determine the best compute architecture // out of all for which the kernel was compiled for that matches closely // to the current device. Other methods to determine the architecture (that do not // require a pointer) can be error prone. See: // https://github.com/NVIDIA/cub/issues/545 auto sm60_wrapper = pairwise_matrix_sm60_get_wrapper(distance_op, params, legacy_range); void* kernel_ptr = reinterpret_cast<void*>(sm60_wrapper.kernel_ptr); auto runtime_arch = arch::kernel_virtual_arch(kernel_ptr); if (cutlass_range.contains(runtime_arch)) { // If device is SM_80 or later, use CUTLASS-based kernel. pairwise_matrix_sm80_dispatch(distance_op, params, cutlass_range, stream); } else { // Reuse kernel wrapper that we obtained above. This avoids performing the // dispatch twice. sm60_wrapper.launch(distance_op, params, stream); } } } }; // namespace cuvs::distance::detail
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_matrix/params.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once namespace cuvs::distance::detail { template <typename IdxT, typename DataT, typename OutT, typename FinOpT> struct pairwise_matrix_params { IdxT m; IdxT n; IdxT k; IdxT ldx; IdxT ldy; IdxT ld_out; const DataT* x; const DataT* y; const DataT* x_norm; const DataT* y_norm; OutT* out; FinOpT fin_op; bool is_row_major; /// @brief: Flips the x and y input and corresponding sizes void flip_x_and_y() { // Flip m, n; ldx, ldy; x, y; x_norm, y_norm. std::swap(m, n); std::swap(ldx, ldy); std::swap(x, y); std::swap(x_norm, y_norm); } }; } // namespace cuvs::distance::detail
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_matrix/dispatch_sm80.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <algorithm> // std::min #include <cuvs/distance/detail/pairwise_distance_cutlass_base.cuh> // cutlassDistanceKernel #include <cuvs/distance/detail/pairwise_matrix/dispatch_layout.cuh> // dispatch_layout namespace cuvs::distance::detail { template <typename OpT, typename IdxT, typename DataT, typename OutT, typename FinOpT, typename SM_compat_t> void pairwise_matrix_sm80_dispatch(OpT distance_op, pairwise_matrix_params<IdxT, DataT, OutT, FinOpT> params, SM_compat_t sm_compat_range, cudaStream_t stream) { int vec_len = determine_vec_len(params); // f takes compile-time constants row_major and vec_len aligned and runs the // corresponding cutlass launch code. auto f = [&](auto row_major, auto vec_len_aligned) { // row_major and vec_len are std::integral_constants of type bool and int // respectively. // Prevent double, vec_len=4 combination (this is not supported) constexpr int vec_len = std::min(vec_len_aligned(), static_cast<int>(16 / sizeof(DataT))); using AccT = typename OpT::AccT; cutlassDistanceKernel<DataT, AccT, OutT, IdxT, vec_len, FinOpT, OpT, row_major()>(params.x, params.y, params.x_norm, params.y_norm, params.m, params.n, params.k, params.ldx, params.ldy, params.ld_out, params.out, params.fin_op, distance_op, stream); }; // Dispatch_layout calls f with appropriate compile time constants based on // the runtime values of params.is_row_major and vec_len. dispatch_layout(params.is_row_major, vec_len, f); } }; // namespace cuvs::distance::detail
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/pairwise_matrix/kernel_sm60.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cassert> // assert #include <cuvs/distance/detail/pairwise_distance_base.cuh> // PairwiseDistances #include <cuvs/distance/detail/pairwise_matrix/params.cuh> // pairwise_matrix_params #include <raft/core/operators.hpp> // raft::void_op #include <raft/util/arch.cuh> // raft::util::arch::SM_compute_arch namespace cuvs::distance::detail { template <typename Policy, bool row_major, typename SM_compat_t, typename OpT, typename IdxT, typename DataT, typename OutT, typename FinOpT> __launch_bounds__(Policy::Nthreads, 2) RAFT_KERNEL pairwise_matrix_kernel(OpT distance_op, pairwise_matrix_params<IdxT, DataT, OutT, FinOpT> params) { // Early exit to minimize the size of the kernel when it is not supposed to be compiled. constexpr SM_compat_t sm_compat_range{}; if constexpr (!sm_compat_range.contains(raft::util::arch::SM_compute_arch())) { assert(false); return; } extern __shared__ char smem[]; // The epilog is already provided by distance_op. Do not provide additional // epilogs. auto epilog_op = raft::void_op(); // No support for row_epilog_op. auto row_epilog_op = raft::void_op(); // Always write output constexpr bool write_out = true; constexpr bool use_norms = distance_op.use_norms; PairwiseDistances<DataT, OutT, IdxT, Policy, decltype(distance_op), decltype(epilog_op), decltype(params.fin_op), decltype(row_epilog_op), row_major, write_out> obj(params.x, params.y, params.m, params.n, params.k, params.ldx, params.ldy, params.ld_out, params.x_norm, params.y_norm, params.out, smem, distance_op, epilog_op, params.fin_op, row_epilog_op); obj.run(); } // The type of a pointer to the pairwise matrix kernel. The following template // arguments are type-erased: // // - The kernel policy // - row_major // - SM_compat_t template <typename OpT, typename IdxT, typename DataT, typename OutT, typename FinOpT> using pairwise_matrix_kernel_t = void (*)(OpT, pairwise_matrix_params<IdxT, DataT, OutT, FinOpT>); // A wrapper for the pairwise matrix kernel launch. Includes kernel launch // parameters. template <typename OpT, typename IdxT, typename DataT, typename OutT, typename FinOpT> struct pairwise_matrix_sm60_wrapper { dim3 grid; dim3 block; int smem_size; pairwise_matrix_kernel_t<OpT, IdxT, DataT, OutT, FinOpT> kernel_ptr; void launch(OpT distance_op, pairwise_matrix_params<IdxT, DataT, OutT, FinOpT> params, cudaStream_t stream) { kernel_ptr<<<grid, block, smem_size, stream>>>(distance_op, params); RAFT_CUDA_TRY(cudaGetLastError()); } }; /** @brief: Create kernel launch wrapper for pairwise matrix kernel * * This can be used to type-erase the kernel execution policy, row_major, and SM * compatibility range. * * @tparam Policy: Kernel execution policy * @tparam row_major: Indicates whether input matrices are row major * @tparam OpT: Type of distance operation * @tparam IdxT: Index type * @tparam DataT: Data type * @tparam OutT: Output data type * @tparam FinOpT: Final operation type * @tparam SM_compat_t: Type of the SM architecture compatibility * * @param distance_op: Distance operation * @param params: Parameters * @param sm_compat_range: Which SM architectures to compile for. */ template <typename Policy, bool row_major, typename OpT, typename IdxT, typename DataT, typename OutT, typename FinOpT, typename SM_compat_t> pairwise_matrix_sm60_wrapper<OpT, IdxT, DataT, OutT, FinOpT> make_pairwise_matrix_sm60_wrapper( OpT distance_op, pairwise_matrix_params<IdxT, DataT, OutT, FinOpT> params, SM_compat_t sm_compat_range) { dim3 block(Policy::Nthreads); // Use ::template to disambiguate (See: // https://en.cppreference.com/w/cpp/language/dependent_name) int smem_size = OpT::template shared_mem_size<Policy>(); // Obtain function pointer to kernel auto kernel = pairwise_matrix_kernel<Policy, row_major, SM_compat_t, OpT, IdxT, DataT, OutT, FinOpT>; dim3 grid = launchConfigGenerator<Policy>(params.m, params.n, smem_size, kernel); return pairwise_matrix_sm60_wrapper<OpT, IdxT, DataT, OutT, FinOpT>{ grid, block, smem_size, kernel}; } }; // namespace cuvs::distance::detail
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/kernels/kernel_matrices.cuh
/* * Copyright (c) 2019-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "gram_matrix.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <cuvs/distance/detail/kernels/rbf_fin_op.cuh> #include <cuvs/distance/distance.cuh> #include <raft/linalg/gemm.cuh> #include <raft/sparse/linalg/norm.cuh> #include <raft/util/cuda_utils.cuh> namespace cuvs::distance::kernels::detail { /** Epiloge function for polynomial kernel without padding. * Calculates output = (gain*in + offset)^exponent * @param inout device vector in column major format, size [len] * @param len array length * @param exponent * @param gain * @param offset */ template <typename math_t, typename exp_t> RAFT_KERNEL polynomial_kernel_nopad( math_t* inout, size_t len, exp_t exponent, math_t gain, math_t offset) { for (size_t tid = threadIdx.x + blockIdx.x * blockDim.x; tid < len; tid += blockDim.x * gridDim.x) { inout[tid] = pow(gain * inout[tid] + offset, exponent); } } /** Epiloge function for polynomial kernel with padding. * Calculates output = (gain*input + offset)^exponent * @param inout device vector in column major format, size [ld * cols] * @param ld leading dimension of the inout buffer * @param rows number of rows (rows <= ld) * @param cols number of columns * @param exponent * @param gain * @param offset */ template <typename math_t, typename exp_t> RAFT_KERNEL polynomial_kernel( math_t* inout, int ld, int rows, int cols, exp_t exponent, math_t gain, math_t offset) { for (size_t tidy = threadIdx.y + blockIdx.y * blockDim.y; tidy < cols; tidy += blockDim.y * gridDim.y) for (size_t tidx = threadIdx.x + blockIdx.x * blockDim.x; tidx < rows; tidx += blockDim.x * gridDim.x) { inout[tidx + tidy * ld] = pow(gain * inout[tidx + tidy * ld] + offset, exponent); } } /** Epiloge function for tanh kernel without padding. * Calculates output = tanh(gain*input + offset) * @param inout device vector, size [len] * @param len length of the input vector * @param gain * @param offset */ template <typename math_t> RAFT_KERNEL tanh_kernel_nopad(math_t* inout, size_t len, math_t gain, math_t offset) { for (size_t tid = threadIdx.x + blockIdx.x * blockDim.x; tid < len; tid += blockDim.x * gridDim.x) { inout[tid] = tanh(gain * inout[tid] + offset); } } /** Epiloge function for tanh kernel without padding. * Calculates output = tanh(gain*input + offset) * @param inout device vector in column major format, size [ld * cols] * @param ld leading dimension of the inout buffer * @param rows number of rows (rows <= ld) * @param cols number of columns * @param gain * @param offset */ template <typename math_t> RAFT_KERNEL tanh_kernel(math_t* inout, int ld, int rows, int cols, math_t gain, math_t offset) { for (size_t tidy = threadIdx.y + blockIdx.y * blockDim.y; tidy < cols; tidy += blockDim.y * gridDim.y) for (size_t tidx = threadIdx.x + blockIdx.x * blockDim.x; tidx < rows; tidx += blockDim.x * gridDim.x) { inout[tidx + tidy * ld] = tanh(gain * inout[tidx + tidy * ld] + offset); } } /** Epiloge function for rbf kernel using expansion. * * Calculates output_ij = exp(-gain * (norm_x_i + norm_y_j - 2*input_ij)); * * Intended usage * - input is the product of two matrices X and Y input_ij = sum_k X_ik * Y_jk * - norm_x_i = l2_norm(x_i), where x_i is the i-th row of matrix X * - norm_y_j = l2_norm(y_j), where y_j is the j-th row of matrix Y * * @param inout device vector in column major format, size [ld * cols] * @param ld leading dimension of the inout buffer * @param rows number of rows (rows <= ld) * @param cols number of columns * @param norm_x l2-norm of X's rows * @param norm_y l2-norm of Y's rows * @param gain */ template <typename math_t> RAFT_KERNEL rbf_kernel_expanded( math_t* inout, int ld, int rows, int cols, math_t* norm_x, math_t* norm_y, math_t gain) { for (size_t tidy = threadIdx.y + blockIdx.y * blockDim.y; tidy < cols; tidy += blockDim.y * gridDim.y) { math_t norm_y_val = norm_y[tidy]; for (size_t tidx = threadIdx.x + blockIdx.x * blockDim.x; tidx < rows; tidx += blockDim.x * gridDim.x) { inout[tidx + tidy * ld] = exp(-1.0 * gain * (norm_x[tidx] + norm_y_val - inout[tidx + tidy * ld] * 2)); } } } namespace { std::tuple<dim3, dim3> generateLaunchConfig2dElementwiseOp(int n1, int n2) { dim3 block_shape = dim3(32, 4); const int num_blocks_x = raft::ceildiv(n1, 32); const int num_blocks_y = std::min(raft::ceildiv(n2, 32), (1 << 16) - 1); dim3 grid_shape = dim3(num_blocks_x, num_blocks_y); return std::make_tuple(grid_shape, block_shape); } } // namespace /** * Create a kernel matrix using polynomial kernel function. */ template <typename math_t, typename exp_t> class PolynomialKernel : public GramMatrixBase<math_t> { exp_t exponent; math_t gain; math_t offset; void applyKernel( math_t* inout, int ld, int rows, int cols, bool is_row_major, cudaStream_t stream) { const int n_minor = is_row_major ? cols : rows; if (ld == n_minor) { polynomial_kernel_nopad<<<raft::ceildiv<size_t>((size_t)rows * cols, 128), 128, 0, stream>>>( inout, rows * cols, exponent, gain, offset); } else { int n1 = is_row_major ? cols : rows; int n2 = is_row_major ? rows : cols; auto [grid_shape, block_shape] = generateLaunchConfig2dElementwiseOp(n1, n2); polynomial_kernel<<<grid_shape, block_shape, 0, stream>>>( inout, ld, n1, n2, exponent, gain, offset); } RAFT_CUDA_TRY(cudaPeekAtLastError()); } public: /** * Constructs a polynomial kernel object. * It evaluates the kernel matrix using the following formula: * K_ij = (gain*<x1_i, x2_k> + offset)^exponent * * @tparam math_t floating point type * @tparam exp_t type of exponent * @param exponent * @param gain * @param offset */ PolynomialKernel(exp_t exponent, math_t gain, math_t offset) : GramMatrixBase<math_t>(), exponent(exponent), gain(gain), offset(offset) { } [[deprecated]] PolynomialKernel(exp_t exponent, math_t gain, math_t offset, cublasHandle_t handle) : GramMatrixBase<math_t>(handle), exponent(exponent), gain(gain), offset(offset) { } /** Evaluate kernel matrix using polynomial kernel. * * output[i,k] = (gain*<x1_i, x2_k> + offset)^exponent, * where x1_i is the i-th vector from the x1 set, and x2_k is k-th vector * in the x2 set, and < , > denotes dot product. * * @param [in] handle raft handle * @param [in] x1 dense device matrix view, size [n1*n_cols] * @param [in] x2 dense device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 unused. * @param norm_x2 unused. */ void evaluate(raft::resources const& handle, dense_input_matrix_view_t<math_t> x1, dense_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1, math_t* norm_x2) { bool is_row_major = GramMatrixBase<math_t>::get_is_row_major(out); int ld_out = is_row_major ? out.stride(0) : out.stride(1); GramMatrixBase<math_t>::linear(handle, x1, x2, out); applyKernel(out.data_handle(), ld_out, out.extent(0), out.extent(1), is_row_major, resource::get_cuda_stream(handle)); } /** Evaluate kernel matrix using polynomial kernel. * * output[i,k] = (gain*<x1_i, x2_k> + offset)^exponent, * where x1_i is the i-th vector from the x1 set, and x2_k is k-th vector * in the x2 set, and < , > denotes dot product. * * @param [in] handle raft handle * @param [in] x1 csr device matrix view, size [n1*n_cols] * @param [in] x2 dense device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 unused. * @param norm_x2 unused. */ void evaluate(raft::resources const& handle, csr_input_matrix_view_t<math_t> x1, dense_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1, math_t* norm_x2) { bool is_row_major = GramMatrixBase<math_t>::get_is_row_major(out); int ld_out = is_row_major ? out.stride(0) : out.stride(1); GramMatrixBase<math_t>::linear(handle, x1, x2, out); applyKernel(out.data_handle(), ld_out, out.extent(0), out.extent(1), is_row_major, resource::get_cuda_stream(handle)); } /** Evaluate kernel matrix using polynomial kernel. * * output[i,k] = (gain*<x1_i, x2_k> + offset)^exponent, * where x1_i is the i-th vector from the x1 set, and x2_k is k-th vector * in the x2 set, and < , > denotes dot product. * * @param [in] handle raft handle * @param [in] x1 csr device matrix view, size [n1*n_cols] * @param [in] x2 csr device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 unused. * @param norm_x2 unused. */ void evaluate(raft::resources const& handle, csr_input_matrix_view_t<math_t> x1, csr_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1, math_t* norm_x2) { bool is_row_major = GramMatrixBase<math_t>::get_is_row_major(out); int ld_out = is_row_major ? out.stride(0) : out.stride(1); GramMatrixBase<math_t>::linear(handle, x1, x2, out); applyKernel(out.data_handle(), ld_out, out.extent(0), out.extent(1), is_row_major, resource::get_cuda_stream(handle)); } /** Evaluate the Gram matrix using the legacy interface. * * @param [in] x1 device array of vectors, size [n1*n_cols] * @param [in] n1 number vectors in x1 * @param [in] n_cols number of columns (features) in x1 and x2 * @param [in] x2 device array of vectors, size [n2*n_cols] * @param [in] n2 number vectors in x2 * @param [out] out device buffer to store the Gram matrix, size [n1*n2] * @param [in] is_row_major whether the input and output matrices are in row * major format * @param [in] stream cuda stream * @param ld1 leading dimension of x1 (usually it is n1) * @param ld2 leading dimension of x2 (usually it is n2) * @param ld_out leading dimension of out (usually it is n1) */ [[deprecated]] void evaluate(const math_t* x1, int n1, int n_cols, const math_t* x2, int n2, math_t* out, bool is_row_major, cudaStream_t stream, int ld1, int ld2, int ld_out) { ASSERT(GramMatrixBase<math_t>::legacy_interface, "Legacy interface can only be used with legacy ctor."); GramMatrixBase<math_t>::linear( x1, n1, n_cols, x2, n2, out, is_row_major, stream, ld1, ld2, ld_out); applyKernel(out, ld_out, n1, n2, is_row_major, stream); } }; /** * Create a kernel matrix using tanh kernel function. */ template <typename math_t> class TanhKernel : public GramMatrixBase<math_t> { math_t gain, offset; void applyKernel( math_t* inout, int ld, int rows, int cols, bool is_row_major, cudaStream_t stream) { const int n_minor = is_row_major ? cols : rows; if (ld == n_minor) { tanh_kernel_nopad<<<raft::ceildiv<size_t>((size_t)rows * cols, 128), 128, 0, stream>>>( inout, rows * cols, gain, offset); } else { int n1 = is_row_major ? cols : rows; int n2 = is_row_major ? rows : cols; auto [grid_shape, block_shape] = generateLaunchConfig2dElementwiseOp(n1, n2); tanh_kernel<<<grid_shape, block_shape, 0, stream>>>(inout, ld, n1, n2, gain, offset); } RAFT_CUDA_TRY(cudaPeekAtLastError()); } public: /** * Constructs a tanh kernel object. * It evaluates the kernel matrix using the following formula: * K_ij = tanh(gain*<x1_i, x2_k> + offset) * * @tparam math_t floating point type * @param gain * @param offset */ TanhKernel(math_t gain, math_t offset) : GramMatrixBase<math_t>(), gain(gain), offset(offset) {} [[deprecated]] TanhKernel(math_t gain, math_t offset, cublasHandle_t handle) : GramMatrixBase<math_t>(handle), gain(gain), offset(offset) { } /** Evaluate kernel matrix using tanh kernel. * * output_[i + k*n1] = (gain*<x1_i, x2_k> + offset)^exponent, * where x1_i is the i-th vector from the x1 set, and x2_k is k-th vector * in the x2 set, and < , > denotes dot product. * * @param [in] handle raft handle * @param [in] x1 dense device matrix view, size [n1*n_cols] * @param [in] x2 dense device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 unused. * @param norm_x2 unused. */ void evaluate(raft::resources const& handle, dense_input_matrix_view_t<math_t> x1, dense_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1, math_t* norm_x2) { bool is_row_major = GramMatrixBase<math_t>::get_is_row_major(out); int ld_out = is_row_major ? out.stride(0) : out.stride(1); GramMatrixBase<math_t>::linear(handle, x1, x2, out); applyKernel(out.data_handle(), ld_out, out.extent(0), out.extent(1), is_row_major, resource::get_cuda_stream(handle)); } /** Evaluate kernel matrix using tanh kernel. * * output_[i + k*n1] = (gain*<x1_i, x2_k> + offset)^exponent, * where x1_i is the i-th vector from the x1 set, and x2_k is k-th vector * in the x2 set, and < , > denotes dot product. * * @param [in] handle raft handle * @param [in] x1 csr device matrix view, size [n1*n_cols] * @param [in] x2 dense device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 unused. * @param norm_x2 unused. */ void evaluate(raft::resources const& handle, csr_input_matrix_view_t<math_t> x1, dense_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1, math_t* norm_x2) { bool is_row_major = GramMatrixBase<math_t>::get_is_row_major(out); int ld_out = is_row_major ? out.stride(0) : out.stride(1); GramMatrixBase<math_t>::linear(handle, x1, x2, out); applyKernel(out.data_handle(), ld_out, out.extent(0), out.extent(1), is_row_major, resource::get_cuda_stream(handle)); } /** Evaluate kernel matrix using tanh kernel. * * output_[i + k*n1] = (gain*<x1_i, x2_k> + offset)^exponent, * where x1_i is the i-th vector from the x1 set, and x2_k is k-th vector * in the x2 set, and < , > denotes dot product. * * @param [in] handle raft handle * @param [in] x1 csr device matrix view, size [n1*n_cols] * @param [in] x2 csr device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 unused. * @param norm_x2 unused. */ void evaluate(raft::resources const& handle, csr_input_matrix_view_t<math_t> x1, csr_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1, math_t* norm_x2) { bool is_row_major = GramMatrixBase<math_t>::get_is_row_major(out); int ld_out = is_row_major ? out.stride(0) : out.stride(1); GramMatrixBase<math_t>::linear(handle, x1, x2, out); applyKernel(out.data_handle(), ld_out, out.extent(0), out.extent(1), is_row_major, resource::get_cuda_stream(handle)); } /** Evaluate the Gram matrix using the legacy interface. * * @param [in] x1 device array of vectors, size [n1*n_cols] * @param [in] n1 number vectors in x1 * @param [in] n_cols number of columns (features) in x1 and x2 * @param [in] x2 device array of vectors, size [n2*n_cols] * @param [in] n2 number vectors in x2 * @param [out] out device buffer to store the Gram matrix, size [n1*n2] * @param [in] is_row_major whether the input and output matrices are in row * major format * @param [in] stream cuda stream * @param ld1 leading dimension of x1 (usually it is n1) * @param ld2 leading dimension of x2 (usually it is n2) * @param ld_out leading dimension of out (usually it is n1) */ [[deprecated]] void evaluate(const math_t* x1, int n1, int n_cols, const math_t* x2, int n2, math_t* out, bool is_row_major, cudaStream_t stream, int ld1, int ld2, int ld_out) { ASSERT(GramMatrixBase<math_t>::legacy_interface, "Legacy interface can only be used with legacy ctor."); GramMatrixBase<math_t>::linear( x1, n1, n_cols, x2, n2, out, is_row_major, stream, ld1, ld2, ld_out); applyKernel(out, ld_out, n1, n2, is_row_major, stream); } }; /** * Create a kernel matrix using RBF kernel function. */ template <typename math_t> class RBFKernel : public GramMatrixBase<math_t> { math_t gain; void applyKernel(math_t* inout, int ld, int rows, int cols, math_t* norm_x1, math_t* norm_x2, bool is_row_major, cudaStream_t stream) { int n1 = is_row_major ? cols : rows; int n2 = is_row_major ? rows : cols; math_t* norm_n1 = is_row_major ? norm_x2 : norm_x1; math_t* norm_n2 = is_row_major ? norm_x1 : norm_x2; auto [grid_shape, block_shape] = generateLaunchConfig2dElementwiseOp(n1, n2); rbf_kernel_expanded<<<grid_shape, block_shape, 0, stream>>>( inout, ld, n1, n2, norm_n1, norm_n2, gain); } public: /** * Constructs a RBF kernel object. * It evaluates the kernel matrix using the following formula: * K_ij = exp(-gain*|x1_i- x2_k|^2) * * @tparam math_t floating point type * @param gain */ RBFKernel(math_t gain) : GramMatrixBase<math_t>(), gain(gain) {} [[deprecated]] RBFKernel(math_t gain, cublasHandle_t handle) : GramMatrixBase<math_t>(handle), gain(gain) { } void matrixRowNormL2(raft::resources const& handle, dense_input_matrix_view_t<math_t> matrix, math_t* target) { bool is_row_major = GramMatrixBase<math_t>::get_is_row_major(matrix); int minor = is_row_major ? matrix.extent(1) : matrix.extent(0); int ld = is_row_major ? matrix.stride(0) : matrix.stride(1); ASSERT(ld == minor, "RBF Kernel lazy rowNorm compute does not support ld parameter"); raft::linalg::rowNorm(target, matrix.data_handle(), matrix.extent(1), matrix.extent(0), raft::linalg::NormType::L2Norm, is_row_major, resource::get_cuda_stream(handle)); } void matrixRowNormL2(raft::resources const& handle, csr_input_matrix_view_t<math_t> matrix, math_t* target) { auto matrix_structure = matrix.structure_view(); raft::sparse::linalg::rowNormCsr(handle, matrix_structure.get_indptr().data(), matrix.get_elements().data(), matrix_structure.get_nnz(), matrix_structure.get_n_rows(), target, raft::linalg::NormType::L2Norm); } /** Evaluate kernel matrix using RBF kernel. * * output_[i + k*n1] = exp(-gain*|x1_i - x2_k|^2), * where x1_i is the i-th vector from the x1 set, and x2_k is k-th vector * in the x2 set, and | | euclidean distance. * * @param [in] handle raft handle * @param [in] x1 dense device matrix view, size [n1*n_cols] * @param [in] x2 dense device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 optional L2-norm of x1's rows for computation within RBF. * @param norm_x2 optional L2-norm of x2's rows for computation within RBF. */ void evaluate(raft::resources const& handle, dense_input_matrix_view_t<math_t> x1, dense_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1, math_t* norm_x2) { cudaStream_t stream = resource::get_cuda_stream(handle); // lazy compute norms if not given rmm::device_uvector<math_t> tmp_norm_x1(0, stream); rmm::device_uvector<math_t> tmp_norm_x2(0, stream); if (norm_x1 == nullptr) { tmp_norm_x1.reserve(x1.extent(0), stream); norm_x1 = tmp_norm_x1.data(); matrixRowNormL2(handle, x1, norm_x1); } if (norm_x2 == nullptr) { tmp_norm_x2.reserve(x2.extent(0), stream); norm_x2 = tmp_norm_x2.data(); matrixRowNormL2(handle, x2, norm_x2); } // compute L2expanded bool is_row_major = GramMatrixBase<math_t>::get_is_row_major(out); int ld_out = is_row_major ? out.stride(0) : out.stride(1); GramMatrixBase<math_t>::linear(handle, x1, x2, out); applyKernel(out.data_handle(), ld_out, out.extent(0), out.extent(1), norm_x1, norm_x2, is_row_major, resource::get_cuda_stream(handle)); } /** Evaluate kernel matrix using RBF kernel. * * output_[i + k*n1] = exp(-gain*|x1_i - x2_k|^2), * where x1_i is the i-th vector from the x1 set, and x2_k is k-th vector * in the x2 set, and | | euclidean distance. * * @param [in] handle raft handle * @param [in] x1 csr device matrix view, size [n1*n_cols] * @param [in] x2 dense device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 optional L2-norm of x1's rows for computation within RBF. * @param norm_x2 optional L2-norm of x2's rows for computation within RBF. */ void evaluate(raft::resources const& handle, csr_input_matrix_view_t<math_t> x1, dense_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1, math_t* norm_x2) { cudaStream_t stream = resource::get_cuda_stream(handle); // lazy compute norms if not given rmm::device_uvector<math_t> tmp_norm_x1(0, stream); rmm::device_uvector<math_t> tmp_norm_x2(0, stream); if (norm_x1 == nullptr) { tmp_norm_x1.reserve(x1.structure_view().get_n_rows(), stream); norm_x1 = tmp_norm_x1.data(); matrixRowNormL2(handle, x1, norm_x1); } if (norm_x2 == nullptr) { tmp_norm_x2.reserve(x2.extent(0), stream); norm_x2 = tmp_norm_x2.data(); matrixRowNormL2(handle, x2, norm_x2); } // compute L2expanded bool is_row_major = GramMatrixBase<math_t>::get_is_row_major(out); int ld_out = is_row_major ? out.stride(0) : out.stride(1); GramMatrixBase<math_t>::linear(handle, x1, x2, out); applyKernel(out.data_handle(), ld_out, out.extent(0), out.extent(1), norm_x1, norm_x2, is_row_major, resource::get_cuda_stream(handle)); } /** Evaluate kernel matrix using RBF kernel. * * output_[i + k*n1] = exp(-gain*|x1_i - x2_k|^2), * where x1_i is the i-th vector from the x1 set, and x2_k is k-th vector * in the x2 set, and | | euclidean distance. * * @param [in] handle raft handle * @param [in] x1 csr device matrix view, size [n1*n_cols] * @param [in] x2 csr device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 optional L2-norm of x1's rows for computation within RBF. * @param norm_x2 optional L2-norm of x2's rows for computation within RBF. */ void evaluate(raft::resources const& handle, csr_input_matrix_view_t<math_t> x1, csr_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1, math_t* norm_x2) { cudaStream_t stream = resource::get_cuda_stream(handle); // lazy compute norms if not given rmm::device_uvector<math_t> tmp_norm_x1(0, stream); rmm::device_uvector<math_t> tmp_norm_x2(0, stream); if (norm_x1 == nullptr) { tmp_norm_x1.reserve(x1.structure_view().get_n_rows(), stream); norm_x1 = tmp_norm_x1.data(); matrixRowNormL2(handle, x1, norm_x1); } if (norm_x2 == nullptr) { tmp_norm_x2.reserve(x2.structure_view().get_n_rows(), stream); norm_x2 = tmp_norm_x2.data(); matrixRowNormL2(handle, x2, norm_x2); } // compute L2expanded bool is_row_major = GramMatrixBase<math_t>::get_is_row_major(out); int ld_out = is_row_major ? out.stride(0) : out.stride(1); GramMatrixBase<math_t>::linear(handle, x1, x2, out); applyKernel(out.data_handle(), ld_out, out.extent(0), out.extent(1), norm_x1, norm_x2, is_row_major, resource::get_cuda_stream(handle)); } /** Evaluate the Gram matrix using the legacy interface. * * @param [in] x1 device array of vectors, size [n1*n_cols] * @param [in] n1 number vectors in x1 * @param [in] n_cols number of columns (features) in x1 and x2 * @param [in] x2 device array of vectors, size [n2*n_cols] * @param [in] n2 number vectors in x2 * @param [out] out device buffer to store the Gram matrix, size [n1*n2] * @param [in] is_row_major whether the input and output matrices are in row * major format * @param [in] stream cuda stream * @param ld1 leading dimension of x1 (usually it is n1) * @param ld2 leading dimension of x2 (usually it is n2) * @param ld_out leading dimension of out (usually it is n1) */ [[deprecated]] void evaluate(const math_t* x1, int n1, int n_cols, const math_t* x2, int n2, math_t* out, bool is_row_major, cudaStream_t stream, int ld1, int ld2, int ld_out) { ASSERT(GramMatrixBase<math_t>::legacy_interface, "Legacy interface can only be used with legacy ctor."); int minor1 = is_row_major ? n_cols : n1; int minor2 = is_row_major ? n_cols : n2; int minor_out = is_row_major ? n2 : n1; ASSERT(ld1 == minor1, "RBF Kernel distance does not support ld1 parameter"); ASSERT(ld2 == minor2, "RBF Kernel distance does not support ld2 parameter"); ASSERT(ld_out == minor_out, "RBF Kernel distance does not support ld_out parameter"); math_t gain = this->gain; using index_t = int64_t; rbf_fin_op fin_op{gain}; raft::resources handle; resource::set_cuda_stream(handle, stream); cuvs::distance::distance<cuvs::distance::DistanceType::L2Unexpanded, math_t, math_t, math_t, decltype(fin_op), index_t>(handle, const_cast<math_t*>(x1), const_cast<math_t*>(x2), out, n1, n2, n_cols, NULL, 0, fin_op, is_row_major); } }; }; // end namespace cuvs::distance::kernels::detail
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/kernels/rbf_fin_op.cuh
/* * Copyright (c) 2019-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once /* * This file defines rbf_fin_op, which is used in GramMatrixBase. * * This struct has been moved to a separate file, so that it is cheap to include * in distance/distance-ext.cuh, where an instance of cuvs::distance::distance * with the rbf_fin_op is instantiated. * */ #include <raft/core/math.hpp> // raft::exp #include <raft/util/cuda_dev_essentials.cuh> // HD namespace cuvs::distance::kernels::detail { /** @brief: Final op for Gram matrix with RBF kernel. * * Calculates output = e^(-gain * in) * */ template <typename OutT> struct rbf_fin_op { OutT gain; explicit HD rbf_fin_op(OutT gain_) noexcept : gain(gain_) {} template <typename... Args> HDI OutT operator()(OutT d_val, Args... unused_args) { return raft::exp(-gain * d_val); } }; // struct rbf_fin_op } // namespace cuvs::distance::kernels::detail
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/kernels/kernel_factory.cuh
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "gram_matrix.cuh" #include "kernel_matrices.cuh" #include <cuvs/distance/distance_types.hpp> #include <raft/util/cudart_utils.hpp> namespace cuvs::distance::kernels::detail { template <typename math_t> class KernelFactory { public: static GramMatrixBase<math_t>* create(KernelParams params) { GramMatrixBase<math_t>* res; // KernelParams is not templated, we convert the parameters to math_t here: math_t coef0 = params.coef0; math_t gamma = params.gamma; switch (params.kernel) { case LINEAR: res = new GramMatrixBase<math_t>(); break; case POLYNOMIAL: res = new PolynomialKernel<math_t, int>(params.degree, gamma, coef0); break; case TANH: res = new TanhKernel<math_t>(gamma, coef0); break; case RBF: res = new RBFKernel<math_t>(gamma); break; default: throw raft::exception("Kernel not implemented"); } return res; } [[deprecated]] static GramMatrixBase<math_t>* create(KernelParams params, cublasHandle_t handle) { GramMatrixBase<math_t>* res; // KernelParams is not templated, we convert the parameters to math_t here: math_t coef0 = params.coef0; math_t gamma = params.gamma; switch (params.kernel) { case LINEAR: res = new GramMatrixBase<math_t>(handle); break; case POLYNOMIAL: res = new PolynomialKernel<math_t, int>(params.degree, gamma, coef0, handle); break; case TANH: res = new TanhKernel<math_t>(gamma, coef0, handle); break; case RBF: res = new RBFKernel<math_t>(gamma, handle); break; default: throw raft::exception("Kernel not implemented"); } return res; } }; }; // end namespace cuvs::distance::kernels::detail
0
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail
rapidsai_public_repos/cuvs/cpp/include/cuvs/distance/detail/kernels/gram_matrix.cuh
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/distance/distance.cuh> #include <cuvs/distance/distance_types.hpp> #include <raft/core/device_csr_matrix.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resources.hpp> // #include <raft/sparse/detail/cusparse_wrappers.h> #include <raft/sparse/distance/distance.cuh> #include <raft/sparse/linalg/spmm.cuh> #include <raft/linalg/detail/cublas_wrappers.hpp> #include <raft/linalg/gemm.cuh> namespace cuvs::distance::kernels::detail { template <typename math_t> using dense_input_matrix_view_t = raft::device_matrix_view<const math_t, int, layout_stride>; template <typename math_t> using dense_output_matrix_view_t = raft::device_matrix_view<math_t, int, layout_stride>; template <typename math_t> using csr_input_matrix_view_t = raft::device_csr_matrix_view<const math_t, int, int, int>; /** * Base class for general Gram matrices * A Gram matrix is the Hermitian matrix of inner probucts G_ik = <x_i, x_k> * Here, the inner product is evaluated for all elements from vectors sets X1, * and X2. * * To be more precise, on exit the output buffer will store: * - if is_row_major == true: out[j+k*n1] = <x1_j, x2_k>, * - if is_row_major == false: out[j*n2 + k] = <x1_j, x2_k>, * where x1_j is the j-th vector from the x1 set and x2_k is the k-th vector * from the x2 set. */ template <typename math_t> class GramMatrixBase { protected: cublasHandle_t cublas_handle; bool legacy_interface; public: GramMatrixBase() : legacy_interface(false){}; [[deprecated]] GramMatrixBase(cublasHandle_t cublas_handle) : cublas_handle(cublas_handle), legacy_interface(true){}; virtual ~GramMatrixBase(){}; /** Convenience function to evaluate the Gram matrix for two vector sets. * Vector sets are provided in Matrix format * * @param [in] handle raft handle * @param [in] x1 dense device matrix view, size [n1*n_cols] * @param [in] x2 dense device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 optional L2-norm of x1's rows for computation within RBF. * @param norm_x2 optional L2-norm of x2's rows for computation within RBF. */ void operator()(raft::resources const& handle, dense_input_matrix_view_t<math_t> x1, dense_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1 = nullptr, math_t* norm_x2 = nullptr) { evaluate(handle, x1, x2, out, norm_x1, norm_x2); } /** Convenience function to evaluate the Gram matrix for two vector sets. * Vector sets are provided in Matrix format * * @param [in] handle raft handle * @param [in] x1 csr device matrix view, size [n1*n_cols] * @param [in] x2 dense device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 optional L2-norm of x1's rows for computation within RBF. * @param norm_x2 optional L2-norm of x2's rows for computation within RBF. */ void operator()(raft::resources const& handle, csr_input_matrix_view_t<math_t> x1, dense_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1 = nullptr, math_t* norm_x2 = nullptr) { evaluate(handle, x1, x2, out, norm_x1, norm_x2); } /** Convenience function to evaluate the Gram matrix for two vector sets. * Vector sets are provided in Matrix format * * @param [in] handle raft handle * @param [in] x1 csr device matrix view, size [n1*n_cols] * @param [in] x2 csr device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 optional L2-norm of x1's rows for computation within RBF. * @param norm_x2 optional L2-norm of x2's rows for computation within RBF. */ void operator()(raft::resources const& handle, csr_input_matrix_view_t<math_t> x1, csr_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1 = nullptr, math_t* norm_x2 = nullptr) { evaluate(handle, x1, x2, out, norm_x1, norm_x2); } // unfortunately, 'evaluate' cannot be templatized as it needs to be virtual /** Evaluate the Gram matrix for two vector sets using simple dot product. * * @param [in] handle raft handle * @param [in] x1 dense device matrix view, size [n1*n_cols] * @param [in] x2 dense device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 unused. * @param norm_x2 unused. */ virtual void evaluate(raft::resources const& handle, dense_input_matrix_view_t<math_t> x1, dense_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1, math_t* norm_x2) { linear(handle, x1, x2, out); } /** Evaluate the Gram matrix for two vector sets using simple dot product. * * @param [in] handle raft handle * @param [in] x1 csr device matrix view, size [n1*n_cols] * @param [in] x2 dense device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 unused. * @param norm_x2 unused. */ virtual void evaluate(raft::resources const& handle, csr_input_matrix_view_t<math_t> x1, dense_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1, math_t* norm_x2) { linear(handle, x1, x2, out); } /** Evaluate the Gram matrix for two vector sets using simple dot product. * * @param [in] handle raft handle * @param [in] x1 csr device matrix view, size [n1*n_cols] * @param [in] x2 csr device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] * @param norm_x1 unused. * @param norm_x2 unused. */ virtual void evaluate(raft::resources const& handle, csr_input_matrix_view_t<math_t> x1, csr_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out, math_t* norm_x1, math_t* norm_x2) { linear(handle, x1, x2, out); } /** Evaluate the Gram matrix for two vector sets using simple dot product. * * @param [in] x1 device array of vectors, size [n1*n_cols] * @param [in] n1 number vectors in x1 * @param [in] n_cols number of columns (features) in x1 and x2 * @param [in] x2 device array of vectors, size [n2*n_cols] * @param [in] n2 number vectors in x2 * @param [out] out device buffer to store the Gram matrix, size [n1*n2] * @param [in] is_row_major whether the input and output matrices are in row * major format * @param [in] stream cuda stream * @param ld1 leading dimension of x1 (usually it is n1) * @param ld2 leading dimension of x2 (usually it is n2) * @param ld_out leading dimension of out (usually it is n1) */ [[deprecated]] virtual void evaluate(const math_t* x1, int n1, int n_cols, const math_t* x2, int n2, math_t* out, bool is_row_major, cudaStream_t stream, int ld1, int ld2, int ld_out) { linear(x1, n1, n_cols, x2, n2, out, is_row_major, stream, ld1, ld2, ld_out); } /** Convenience function to evaluate the Gram matrix for two vector sets. * * @param [in] x1 device array of vectors, size [n1*n_cols] * @param [in] n1 number vectors in x1 * @param [in] n_cols number of columns (features) in x1 and x2 * @param [in] x2 device array of vectors, size [n2*n_cols] * @param [in] n2 number vectors in x2 * @param [out] out device buffer to store the Gram matrix, size [n1*n2] * @param [in] is_row_major whether the input and output matrices are in row * major format * @param [in] stream cuda stream * @param ld1 leading dimension of x1 * @param ld2 leading dimension of x2 * @param ld_out leading dimension of out */ [[deprecated]] void operator()(const math_t* x1, int n1, int n_cols, const math_t* x2, int n2, math_t* out, bool is_row_major, cudaStream_t stream, int ld1 = 0, int ld2 = 0, int ld_out = 0) { ASSERT(legacy_interface, "Legacy interface can only be used with legacy ctor."); if (ld1 <= 0) { ld1 = is_row_major ? n_cols : n1; } if (ld2 <= 0) { ld2 = is_row_major ? n_cols : n2; } if (ld_out <= 0) { ld_out = is_row_major ? n2 : n1; } evaluate(x1, n1, n_cols, x2, n2, out, is_row_major, stream, ld1, ld2, ld_out); } protected: /** Calculates the Gram matrix using simple dot product between vector sets. * * out = x1 * x2 * * Can be used as a building block for more complex kernel functions. * * @param [in] x1 device array of vectors, size [n1*n_cols] * @param [in] n1 number vectors in x1 * @param [in] n_cols number of columns (features) in x1 and x2 * @param [in] x2 device array of vectors, size [n2*n_cols] * @param [in] n2 number vectors in x2 * @param [out] out device buffer to store the Gram matrix, size [n1*n2] * @param [in] is_row_major whether the input and output matrices are in row * major format * @param [in] stream cuda stream * @param ld1 leading dimension of x1 * @param ld2 leading dimension of x2 * @param ld_out leading dimension of out */ [[deprecated]] void linear(const math_t* x1, int n1, int n_cols, const math_t* x2, int n2, math_t* out, bool is_row_major, cudaStream_t stream, int ld1, int ld2, int ld_out) { math_t alpha = 1.0; math_t beta = 0.0; if (is_row_major) { // #TODO: Call from public API when ready RAFT_CUBLAS_TRY(raft::linalg::detail::cublasgemm(cublas_handle, CUBLAS_OP_T, CUBLAS_OP_N, n2, n1, n_cols, &alpha, x2, ld2, x1, ld1, &beta, out, ld_out, stream)); } else { // #TODO: Call from public API when ready RAFT_CUBLAS_TRY(raft::linalg::detail::cublasgemm(cublas_handle, CUBLAS_OP_N, CUBLAS_OP_T, n1, n2, n_cols, &alpha, x1, ld1, x2, ld2, &beta, out, ld_out, stream)); } } protected: bool get_is_row_major(dense_output_matrix_view_t<math_t> matrix) { return (matrix.stride(1) == 1); } bool get_is_row_major(dense_input_matrix_view_t<math_t> matrix) { return (matrix.stride(1) == 1); } bool get_is_col_major(dense_output_matrix_view_t<math_t> matrix) { return (matrix.stride(0) == 1); } bool get_is_col_major(dense_input_matrix_view_t<math_t> matrix) { return (matrix.stride(0) == 1); } /** Calculates the Gram matrix using simple dot product between vector sets. * * out = x1 * x2 * * Can be used as a building block for more complex kernel functions. * * @param [in] handle raft handle * @param [in] x1 dense device matrix view, size [n1*n_cols] * @param [in] x2 dense device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] */ void linear(raft::resources const& handle, dense_input_matrix_view_t<math_t> x1, dense_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out) { // check is_row_major consistency bool is_row_major = get_is_row_major(x1) && get_is_row_major(x2) && get_is_row_major(out); bool is_col_major = get_is_col_major(x1) && get_is_col_major(x2) && get_is_col_major(out); ASSERT(is_row_major || is_col_major, "GramMatrix leading dimensions for x1, x2 and out do not match"); // check dimensions int n1 = out.extent(0); int n2 = out.extent(1); int n_cols = x1.extent(1); ASSERT(x1.extent(0) == n1, "GramMatrix input matrix dimensions for x1 and out do not match"); ASSERT(x2.extent(0) == n2, "GramMatrix input matrix dimensions for x2 and out do not match"); ASSERT(x2.extent(1) == n_cols, "GramMatrix input matrix dimensions for x1 and x2 do not match"); // extract major stride int ld1 = is_row_major ? x1.stride(0) : x1.stride(1); int ld2 = is_row_major ? x2.stride(0) : x2.stride(1); int ld_out = is_row_major ? out.stride(0) : out.stride(1); math_t alpha = 1.0; math_t beta = 0.0; if (is_row_major) { // #TODO: Use mdspan-based API when stride-capable // https://github.com/rapidsai/raft/issues/875 raft::linalg::gemm(handle, true, false, n2, n1, n_cols, &alpha, x2.data_handle(), ld2, x1.data_handle(), ld1, &beta, out.data_handle(), ld_out, resource::get_cuda_stream(handle)); } else { // #TODO: Use mdspan-based API when stride-capable // https://github.com/rapidsai/raft/issues/875 raft::linalg::gemm(handle, false, true, n1, n2, n_cols, &alpha, x1.data_handle(), ld1, x2.data_handle(), ld2, &beta, out.data_handle(), ld_out, resource::get_cuda_stream(handle)); } } /** Calculates the Gram matrix using simple dot product between vector sets. * * out = x1 * x2 * * Can be used as a building block for more complex kernel functions. * * @param [in] handle raft handle * @param [in] x1 csr device matrix view, size [n1*n_cols] * @param [in] x2 dense device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] */ void linear(raft::resources const& handle, csr_input_matrix_view_t<math_t> x1, dense_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out) { // check is_row_major consistency bool is_row_major = get_is_row_major(x2) && get_is_row_major(out); bool is_col_major = get_is_col_major(x2) && get_is_col_major(out); ASSERT(is_row_major || is_col_major, "GramMatrix leading dimensions for x2 and out do not match"); // check dimensions auto x1_structure = x1.structure_view(); ASSERT(x1_structure.get_n_rows() == out.extent(0), "GramMatrix input matrix dimensions for x1 and out do not match"); ASSERT(x2.extent(0) == out.extent(1), "GramMatrix input matrix dimensions for x2 and out do not match"); ASSERT(x2.extent(1) == x1_structure.get_n_cols(), "GramMatrix input matrix dimensions for x1 and x2 do not match"); math_t alpha = 1.0; math_t beta = 0.0; raft::sparse::linalg::spmm(handle, false, true, &alpha, x1, x2, &beta, out); } /** Calculates the Gram matrix using simple dot product between vector sets. * * out = x1 * x2 * * Can be used as a building block for more complex kernel functions. * * @param [in] handle raft handle * @param [in] x1 csr device matrix view, size [n1*n_cols] * @param [in] x2 csr device matrix view, size [n2*n_cols] * @param [out] out dense device matrix view for the Gram matrix, size [n1*n2] */ void linear(raft::resources const& handle, csr_input_matrix_view_t<math_t> x1, csr_input_matrix_view_t<math_t> x2, dense_output_matrix_view_t<math_t> out) { // check layout consistency (w.r.t. strides a matrix might be both row & col major) bool is_row_major_nopad = get_is_row_major(out) && out.stride(0) == out.extent(1); bool is_col_major_nopad = get_is_col_major(out) && out.stride(1) == out.extent(0); ASSERT(is_row_major_nopad || is_col_major_nopad, "Sparse linear Kernel distance does not support ld_out parameter"); // switch a,b based on is_row_major if (is_col_major_nopad) { auto out_row_major = raft::make_device_matrix_view<math_t, int, raft::row_major>( out.data_handle(), out.extent(1), out.extent(0)); raft::sparse::distance::pairwise_distance( handle, x2, x1, out_row_major, cuvs::distance::DistanceType::InnerProduct, 0.0); } else { auto out_row_major = raft::make_device_matrix_view<math_t, int, raft::row_major>( out.data_handle(), out.extent(0), out.extent(1)); raft::sparse::distance::pairwise_distance( handle, x1, x2, out_row_major, cuvs::distance::DistanceType::InnerProduct, 0.0); } } }; }; // end namespace cuvs::distance::kernels::detail
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/ball_cover-ext.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cstdint> // uint32_t #include <cuvs/distance/distance_types.hpp> // cuvs::distance::DistanceType #include <cuvs/neighbors/ball_cover_types.hpp> // BallCoverIndex #include <raft/util/raft_explicit.hpp> // RAFT_EXPLICIT #ifdef RAFT_EXPLICIT_INSTANTIATE_ONLY namespace cuvs::neighbors::ball_cover { template <typename idx_t, typename value_t, typename int_t, typename matrix_idx_t> void build_index(raft::resources const& handle, BallCoverIndex<idx_t, value_t, int_t, matrix_idx_t>& index) RAFT_EXPLICIT; template <typename idx_t, typename value_t, typename int_t, typename matrix_idx_t> void all_knn_query(raft::resources const& handle, BallCoverIndex<idx_t, value_t, int_t, matrix_idx_t>& index, int_t k, idx_t* inds, value_t* dists, bool perform_post_filtering = true, float weight = 1.0) RAFT_EXPLICIT; template <typename idx_t, typename value_t, typename int_t, typename matrix_idx_t> void all_knn_query(raft::resources const& handle, BallCoverIndex<idx_t, value_t, int_t, matrix_idx_t>& index, raft::device_matrix_view<idx_t, matrix_idx_t, raft::row_major> inds, raft::device_matrix_view<value_t, matrix_idx_t, raft::row_major> dists, int_t k, bool perform_post_filtering = true, float weight = 1.0) RAFT_EXPLICIT; template <typename idx_t, typename value_t, typename int_t> void knn_query(raft::resources const& handle, const BallCoverIndex<idx_t, value_t, int_t>& index, int_t k, const value_t* query, int_t n_query_pts, idx_t* inds, value_t* dists, bool perform_post_filtering = true, float weight = 1.0) RAFT_EXPLICIT; template <typename idx_t, typename value_t, typename int_t, typename matrix_idx_t> void knn_query(raft::resources const& handle, const BallCoverIndex<idx_t, value_t, int_t, matrix_idx_t>& index, raft::device_matrix_view<const value_t, matrix_idx_t, raft::row_major> query, raft::device_matrix_view<idx_t, matrix_idx_t, raft::row_major> inds, raft::device_matrix_view<value_t, matrix_idx_t, raft::row_major> dists, int_t k, bool perform_post_filtering = true, float weight = 1.0) RAFT_EXPLICIT; } // namespace cuvs::neighbors::ball_cover #endif // RAFT_EXPLICIT_INSTANTIATE_ONLY #define instantiate_raft_neighbors_ball_cover(idx_t, value_t, int_t, matrix_idx_t) \ extern template void \ cuvs::neighbors::ball_cover::build_index<idx_t, value_t, int_t, matrix_idx_t>( \ raft::resources const& handle, \ cuvs::neighbors::ball_cover::BallCoverIndex<idx_t, value_t, int_t, matrix_idx_t>& index); \ \ extern template void \ cuvs::neighbors::ball_cover::all_knn_query<idx_t, value_t, int_t, matrix_idx_t>( \ raft::resources const& handle, \ cuvs::neighbors::ball_cover::BallCoverIndex<idx_t, value_t, int_t, matrix_idx_t>& index, \ int_t k, \ idx_t* inds, \ value_t* dists, \ bool perform_post_filtering, \ float weight); \ \ extern template void \ cuvs::neighbors::ball_cover::all_knn_query<idx_t, value_t, int_t, matrix_idx_t>( \ raft::resources const& handle, \ cuvs::neighbors::ball_cover::BallCoverIndex<idx_t, value_t, int_t, matrix_idx_t>& index, \ raft::device_matrix_view<idx_t, matrix_idx_t, raft::row_major> inds, \ raft::device_matrix_view<value_t, matrix_idx_t, raft::row_major> dists, \ int_t k, \ bool perform_post_filtering, \ float weight); \ \ extern template void cuvs::neighbors::ball_cover::knn_query<idx_t, value_t, int_t>( \ raft::resources const& handle, \ const cuvs::neighbors::ball_cover::BallCoverIndex<idx_t, value_t, int_t>& index, \ int_t k, \ const value_t* query, \ int_t n_query_pts, \ idx_t* inds, \ value_t* dists, \ bool perform_post_filtering, \ float weight); \ \ extern template void \ cuvs::neighbors::ball_cover::knn_query<idx_t, value_t, int_t, matrix_idx_t>( \ raft::resources const& handle, \ const cuvs::neighbors::ball_cover::BallCoverIndex<idx_t, value_t, int_t, matrix_idx_t>& index, \ raft::device_matrix_view<const value_t, matrix_idx_t, raft::row_major> query, \ raft::device_matrix_view<idx_t, matrix_idx_t, raft::row_major> inds, \ raft::device_matrix_view<value_t, matrix_idx_t, raft::row_major> dists, \ int_t k, \ bool perform_post_filtering, \ float weight); instantiate_raft_neighbors_ball_cover(int64_t, float, uint32_t, uint32_t); #undef instantiate_raft_neighbors_ball_cover
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/refine-inl.cuh
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/neighbors/detail/refine.cuh> #include <cuvs/spatial/knn/detail/ann_utils.cuh> #include <raft/core/device_mdspan.hpp> #include <raft/core/host_mdspan.hpp> #include <raft/core/resources.hpp> namespace cuvs::neighbors { /** * @defgroup ann_refine Approximate Nearest Neighbors Refinement * @{ */ /** * @brief Refine nearest neighbor search. * * Refinement is an operation that follows an approximate NN search. The approximate search has * already selected n_candidates neighbor candidates for each query. We narrow it down to k * neighbors. For each query, we calculate the exact distance between the query and its * n_candidates neighbor candidate, and select the k nearest ones. * * The k nearest neighbors and distances are returned. * * Example usage * @code{.cpp} * using namespace cuvs::neighbors; * // use default index parameters * ivf_pq::index_params index_params; * // create and fill the index from a [N, D] dataset * auto index = ivf_pq::build(handle, index_params, dataset, N, D); * // use default search parameters * ivf_pq::search_params search_params; * // search m = 4 * k nearest neighbours for each of the N queries * ivf_pq::search(handle, search_params, index, queries, N, 4 * k, neighbor_candidates, * out_dists_tmp); * // refine it to the k nearest one * refine(handle, dataset, queries, neighbor_candidates, out_indices, out_dists, * index.metric()); * @endcode * * * @param[in] handle the raft handle * @param[in] dataset device matrix that stores the dataset [n_rows, dims] * @param[in] queries device matrix of the queries [n_queris, dims] * @param[in] neighbor_candidates indices of candidate vectors [n_queries, n_candidates], where * n_candidates >= k * @param[out] indices device matrix that stores the refined indices [n_queries, k] * @param[out] distances device matrix that stores the refined distances [n_queries, k] * @param[in] metric distance metric to use. Euclidean (L2) is used by default */ template <typename idx_t, typename data_t, typename distance_t, typename matrix_idx> void refine(raft::resources const& handle, raft::device_matrix_view<const data_t, matrix_idx, raft::row_major> dataset, raft::device_matrix_view<const data_t, matrix_idx, raft::row_major> queries, raft::device_matrix_view<const idx_t, matrix_idx, raft::row_major> neighbor_candidates, raft::device_matrix_view<idx_t, matrix_idx, raft::row_major> indices, raft::device_matrix_view<distance_t, matrix_idx, raft::row_major> distances, distance::DistanceType metric = distance::DistanceType::L2Unexpanded) { detail::refine_device(handle, dataset, queries, neighbor_candidates, indices, distances, metric); } /** Same as above, but all input and out data is in host memory. * @param[in] handle the raft handle * @param[in] dataset host matrix that stores the dataset [n_rows, dims] * @param[in] queries host matrix of the queries [n_queris, dims] * @param[in] neighbor_candidates host matrix with indices of candidate vectors [n_queries, * n_candidates], where n_candidates >= k * @param[out] indices host matrix that stores the refined indices [n_queries, k] * @param[out] distances host matrix that stores the refined distances [n_queries, k] * @param[in] metric distance metric to use. Euclidean (L2) is used by default */ template <typename idx_t, typename data_t, typename distance_t, typename matrix_idx> void refine(raft::resources const& handle, raft::host_matrix_view<const data_t, matrix_idx, raft::row_major> dataset, raft::host_matrix_view<const data_t, matrix_idx, raft::row_major> queries, raft::host_matrix_view<const idx_t, matrix_idx, raft::row_major> neighbor_candidates, raft::host_matrix_view<idx_t, matrix_idx, raft::row_major> indices, raft::host_matrix_view<distance_t, matrix_idx, raft::row_major> distances, distance::DistanceType metric = distance::DistanceType::L2Unexpanded) { detail::refine_host(dataset, queries, neighbor_candidates, indices, distances, metric); } /** @} */ // end group ann_refine } // namespace cuvs::neighbors
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/ivf_pq-inl.cuh
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/neighbors/detail/ivf_pq_build.cuh> #include <cuvs/neighbors/detail/ivf_pq_search.cuh> #include <cuvs/neighbors/ivf_pq_serialize.cuh> #include <cuvs/neighbors/ivf_pq_types.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/resource/device_memory_resource.hpp> #include <raft/core/resources.hpp> #include <rmm/mr/device/device_memory_resource.hpp> #include <memory> // shared_ptr namespace cuvs::neighbors::ivf_pq { /** * @defgroup ivf_pq IVF PQ Algorithm * @{ */ /** * @brief Build the index from the dataset for efficient search. * * NB: Currently, the following distance metrics are supported: * - L2Expanded * - L2Unexpanded * - InnerProduct * * @tparam T data element type * @tparam IdxT type of the indices in the source dataset * * @param[in] handle * @param[in] params configure the index building * @param[in] dataset a device matrix view to a row-major matrix [n_rows, dim] * * @return the constructed ivf-pq index */ template <typename T, typename IdxT = uint32_t> index<IdxT> build(raft::resources const& handle, const index_params& params, raft::device_matrix_view<const T, IdxT, raft::row_major> dataset) { IdxT n_rows = dataset.extent(0); IdxT dim = dataset.extent(1); return detail::build(handle, params, dataset.data_handle(), n_rows, dim); } /** * @brief Extend the index with the new data. * * * @tparam T data element type * @tparam IdxT type of the indices in the source dataset * * @param[in] handle * @param[in] new_vectors a device matrix view to a row-major matrix [n_rows, idx.dim()] * @param[in] new_indices a device vector view to a vector of indices [n_rows]. * If the original index is empty (`idx.size() == 0`), you can pass `std::nullopt` * here to imply a continuous range `[0...n_rows)`. * @param[inout] idx */ template <typename T, typename IdxT> index<IdxT> extend(raft::resources const& handle, raft::device_matrix_view<const T, IdxT, raft::row_major> new_vectors, std::optional<raft::device_vector_view<const IdxT, IdxT>> new_indices, const index<IdxT>& idx) { ASSERT(new_vectors.extent(1) == idx.dim(), "new_vectors should have the same dimension as the index"); IdxT n_rows = new_vectors.extent(0); if (new_indices.has_value()) { ASSERT(n_rows == new_indices.value().extent(0), "new_vectors and new_indices have different number of rows"); } return detail::extend(handle, idx, new_vectors.data_handle(), new_indices.has_value() ? new_indices.value().data_handle() : nullptr, n_rows); } /** * @brief Extend the index with the new data. * * * @tparam T data element type * @tparam IdxT type of the indices in the source dataset * * @param[in] handle * @param[in] new_vectors a device matrix view to a row-major matrix [n_rows, idx.dim()] * @param[in] new_indices a device vector view to a vector of indices [n_rows]. * If the original index is empty (`idx.size() == 0`), you can pass `std::nullopt` * here to imply a continuous range `[0...n_rows)`. * @param[inout] idx */ template <typename T, typename IdxT> void extend(raft::resources const& handle, raft::device_matrix_view<const T, IdxT, raft::row_major> new_vectors, std::optional<raft::device_vector_view<const IdxT, IdxT>> new_indices, index<IdxT>* idx) { ASSERT(new_vectors.extent(1) == idx->dim(), "new_vectors should have the same dimension as the index"); IdxT n_rows = new_vectors.extent(0); if (new_indices.has_value()) { ASSERT(n_rows == new_indices.value().extent(0), "new_vectors and new_indices have different number of rows"); } *idx = detail::extend(handle, *idx, new_vectors.data_handle(), new_indices.has_value() ? new_indices.value().data_handle() : nullptr, n_rows); } /** * @brief Search ANN using the constructed index with the given filter. * * See the [ivf_pq::build](#ivf_pq::build) documentation for a usage example. * * Note, this function requires a temporary buffer to store intermediate results between cuda kernel * calls, which may lead to undesirable allocations and slowdown. To alleviate the problem, you can * pass a pool memory resource or a large enough pre-allocated memory resource to reduce or * eliminate entirely allocations happening within `search`. * The exact size of the temporary buffer depends on multiple factors and is an implementation * detail. However, you can safely specify a small initial size for the memory pool, so that only a * few allocations happen to grow it during the first invocations of the `search`. * * @tparam T data element type * @tparam IdxT type of the indices * @tparam IvfSampleFilterT Device filter function, with the signature * `(uint32_t query_ix, uint32 cluster_ix, uint32_t sample_ix) -> bool` or * `(uint32_t query_ix, uint32 sample_ix) -> bool` * * @param[in] handle * @param[in] params configure the search * @param[in] idx ivf-pq constructed index * @param[in] queries a device matrix view to a row-major matrix [n_queries, index->dim()] * @param[out] neighbors a device matrix view to the indices of the neighbors in the source dataset * [n_queries, k] * @param[out] distances a device matrix view to the distances to the selected neighbors [n_queries, * k] * @param[in] sample_filter a device filter function that greenlights samples for a given query. */ template <typename T, typename IdxT, typename IvfSampleFilterT> void search_with_filtering(raft::resources const& handle, const search_params& params, const index<IdxT>& idx, raft::device_matrix_view<const T, uint32_t, raft::row_major> queries, raft::device_matrix_view<IdxT, uint32_t, raft::row_major> neighbors, raft::device_matrix_view<float, uint32_t, raft::row_major> distances, IvfSampleFilterT sample_filter = IvfSampleFilterT{}) { RAFT_EXPECTS( queries.extent(0) == neighbors.extent(0) && queries.extent(0) == distances.extent(0), "Number of rows in output neighbors and distances matrices must equal the number of queries."); RAFT_EXPECTS(neighbors.extent(1) == distances.extent(1), "Number of columns in output neighbors and distances matrices must equal k"); RAFT_EXPECTS(queries.extent(1) == idx.dim(), "Number of query dimensions should equal number of dimensions in the index."); std::uint32_t k = neighbors.extent(1); detail::search(handle, params, idx, queries.data_handle(), queries.extent(0), k, neighbors.data_handle(), distances.data_handle(), sample_filter); } /** * @brief Search ANN using the constructed index. * * See the [ivf_pq::build](#ivf_pq::build) documentation for a usage example. * * Note, this function requires a temporary buffer to store intermediate results between cuda kernel * calls, which may lead to undesirable allocations and slowdown. To alleviate the problem, you can * pass a pool memory resource or a large enough pre-allocated memory resource to reduce or * eliminate entirely allocations happening within `search`. * The exact size of the temporary buffer depends on multiple factors and is an implementation * detail. However, you can safely specify a small initial size for the memory pool, so that only a * few allocations happen to grow it during the first invocations of the `search`. * * @tparam T data element type * @tparam IdxT type of the indices * * @param[in] handle * @param[in] params configure the search * @param[in] idx ivf-pq constructed index * @param[in] queries a device matrix view to a row-major matrix [n_queries, index->dim()] * @param[out] neighbors a device matrix view to the indices of the neighbors in the source dataset * [n_queries, k] * @param[out] distances a device matrix view to the distances to the selected neighbors [n_queries, * k] */ template <typename T, typename IdxT> void search(raft::resources const& handle, const search_params& params, const index<IdxT>& idx, raft::device_matrix_view<const T, uint32_t, raft::row_major> queries, raft::device_matrix_view<IdxT, uint32_t, raft::row_major> neighbors, raft::device_matrix_view<float, uint32_t, raft::row_major> distances) { search_with_filtering(handle, params, idx, queries, neighbors, distances, cuvs::neighbors::filtering::none_ivf_sample_filter{}); } /** @} */ // end group ivf_pq /** * @brief Build the index from the dataset for efficient search. * * NB: Currently, the following distance metrics are supported: * - L2Expanded * - L2Unexpanded * - InnerProduct * * Usage example: * @code{.cpp} * using namespace cuvs::neighbors; * // use default index parameters * ivf_pq::index_params index_params; * // create and fill the index from a [N, D] dataset * auto index = ivf_pq::build(handle, index_params, dataset, N, D); * // use default search parameters * ivf_pq::search_params search_params; * // search K nearest neighbours for each of the N queries * ivf_pq::search(handle, search_params, index, queries, N, K, out_inds, out_dists); * @endcode * * @tparam T data element type * @tparam IdxT type of the indices in the source dataset * * @param[in] handle * @param[in] params configure the index building * @param[in] dataset a device/host pointer to a row-major matrix [n_rows, dim] * @param[in] n_rows the number of samples * @param[in] dim the dimensionality of the data * * @return the constructed ivf-pq index */ template <typename T, typename IdxT = uint32_t> auto build(raft::resources const& handle, const index_params& params, const T* dataset, IdxT n_rows, uint32_t dim) -> index<IdxT> { return detail::build(handle, params, dataset, n_rows, dim); } /** * @brief Build a new index containing the data of the original plus new extra vectors. * * Implementation note: * The new data is clustered according to existing kmeans clusters, the cluster * centers are unchanged. * * Usage example: * @code{.cpp} * using namespace cuvs::neighbors; * ivf_pq::index_params index_params; * index_params.add_data_on_build = false; // don't populate index on build * index_params.kmeans_trainset_fraction = 1.0; // use whole dataset for kmeans training * // train the index from a [N, D] dataset * auto index_empty = ivf_pq::build(handle, index_params, dataset, N, D); * // fill the index with the data * auto index = ivf_pq::extend(handle, index_empty, dataset, nullptr, N); * @endcode * * @tparam T data element type * @tparam IdxT type of the indices in the source dataset * * @param[in] handle * @param[inout] idx original index * @param[in] new_vectors a device/host pointer to a row-major matrix [n_rows, idx.dim()] * @param[in] new_indices a device/host pointer to a vector of indices [n_rows]. * If the original index is empty (`idx.size() == 0`), you can pass `nullptr` * here to imply a continuous range `[0...n_rows)`. * @param[in] n_rows the number of samples * * @return the constructed extended ivf-pq index */ template <typename T, typename IdxT> auto extend(raft::resources const& handle, const index<IdxT>& idx, const T* new_vectors, const IdxT* new_indices, IdxT n_rows) -> index<IdxT> { return detail::extend(handle, idx, new_vectors, new_indices, n_rows); } /** * @brief Extend the index with the new data. * * * @tparam T data element type * @tparam IdxT type of the indices in the source dataset * * @param[in] handle * @param[inout] idx * @param[in] new_vectors a device/host pointer to a row-major matrix [n_rows, idx.dim()] * @param[in] new_indices a device/host pointer to a vector of indices [n_rows]. * If the original index is empty (`idx.size() == 0`), you can pass `nullptr` * here to imply a continuous range `[0...n_rows)`. * @param[in] n_rows the number of samples */ template <typename T, typename IdxT> void extend(raft::resources const& handle, index<IdxT>* idx, const T* new_vectors, const IdxT* new_indices, IdxT n_rows) { detail::extend(handle, idx, new_vectors, new_indices, n_rows); } /** * @brief Search ANN using the constructed index with the given filter. * * See the [ivf_pq::build](#ivf_pq::build) documentation for a usage example. * * Note, this function requires a temporary buffer to store intermediate results between cuda kernel * calls, which may lead to undesirable allocations and slowdown. To alleviate the problem, you can * pass a pool memory resource or a large enough pre-allocated memory resource to reduce or * eliminate entirely allocations happening within `search`: * @code{.cpp} * ... * // use default search parameters * ivf_pq::search_params search_params; * filtering::none_ivf_sample_filter filter; * // Use the same allocator across multiple searches to reduce the number of * // cuda memory allocations * ivf_pq::search_with_filtering( * handle, search_params, index, queries1, N1, K, out_inds1, out_dists1, filter); * ivf_pq::search_with_filtering( * handle, search_params, index, queries2, N2, K, out_inds2, out_dists2, filter); * ivf_pq::search_with_filtering( * handle, search_params, index, queries3, N3, K, out_inds3, out_dists3, nfilter); * ... * @endcode * The exact size of the temporary buffer depends on multiple factors and is an implementation * detail. However, you can safely specify a small initial size for the memory pool, so that only a * few allocations happen to grow it during the first invocations of the `search`. * * @tparam T data element type * @tparam IdxT type of the indices * @tparam IvfSampleFilterT Device filter function, with the signature * `(uint32_t query_ix, uint32 cluster_ix, uint32_t sample_ix) -> bool` or * `(uint32_t query_ix, uint32 sample_ix) -> bool` * * @param[in] handle * @param[in] params configure the search * @param[in] idx ivf-pq constructed index * @param[in] queries a device pointer to a row-major matrix [n_queries, index->dim()] * @param[in] n_queries the batch size * @param[in] k the number of neighbors to find for each query. * @param[out] neighbors a device pointer to the indices of the neighbors in the source dataset * [n_queries, k] * @param[out] distances a device pointer to the distances to the selected neighbors [n_queries, k] * @param[in] sample_filter a device filter function that greenlights samples for a given query */ template <typename T, typename IdxT, typename IvfSampleFilterT> void search_with_filtering(raft::resources const& handle, const search_params& params, const index<IdxT>& idx, const T* queries, uint32_t n_queries, uint32_t k, IdxT* neighbors, float* distances, IvfSampleFilterT sample_filter = IvfSampleFilterT{}) { detail::search(handle, params, idx, queries, n_queries, k, neighbors, distances, sample_filter); } /** * This function is deprecated and will be removed in a future. * Please drop the `mr` argument and use `raft::resource::set_workspace_resource` instead. */ template <typename T, typename IdxT, typename IvfSampleFilterT> [[deprecated( "Drop the `mr` argument and use `raft::resource::set_workspace_resource` instead")]] void search_with_filtering(raft::resources const& handle, const search_params& params, const index<IdxT>& idx, const T* queries, uint32_t n_queries, uint32_t k, IdxT* neighbors, float* distances, rmm::mr::device_memory_resource* mr, IvfSampleFilterT sample_filter = IvfSampleFilterT{}) { if (mr != nullptr) { // Shallow copy of the resource with the automatic lifespan: // change the workspace resource temporarily raft::resources res_local(handle); resource::set_workspace_resource( res_local, std::shared_ptr<rmm::mr::device_memory_resource>{mr, void_op{}}); return search_with_filtering( res_local, params, idx, queries, n_queries, k, neighbors, distances, sample_filter); } else { return search_with_filtering( handle, params, idx, queries, n_queries, k, neighbors, distances, sample_filter); } } /** * @brief Search ANN using the constructed index. * * See the [ivf_pq::build](#ivf_pq::build) documentation for a usage example. * * Note, this function requires a temporary buffer to store intermediate results between cuda kernel * calls, which may lead to undesirable allocations and slowdown. To alleviate the problem, you can * pass a pool memory resource or a large enough pre-allocated memory resource to reduce or * eliminate entirely allocations happening within `search`: * @code{.cpp} * ... * // Create a pooling memory resource with a pre-defined initial size. * rmm::mr::pool_memory_resource<rmm::mr::device_memory_resource> mr( * rmm::mr::get_current_device_resource(), 1024 * 1024); * // use default search parameters * ivf_pq::search_params search_params; * // Use the same allocator across multiple searches to reduce the number of * // cuda memory allocations * ivf_pq::search(handle, search_params, index, queries1, N1, K, out_inds1, out_dists1, &mr); * ivf_pq::search(handle, search_params, index, queries2, N2, K, out_inds2, out_dists2, &mr); * ivf_pq::search(handle, search_params, index, queries3, N3, K, out_inds3, out_dists3, &mr); * ... * @endcode * The exact size of the temporary buffer depends on multiple factors and is an implementation * detail. However, you can safely specify a small initial size for the memory pool, so that only a * few allocations happen to grow it during the first invocations of the `search`. * * @tparam T data element type * @tparam IdxT type of the indices * * @param[in] handle * @param[in] params configure the search * @param[in] idx ivf-pq constructed index * @param[in] queries a device pointer to a row-major matrix [n_queries, index->dim()] * @param[in] n_queries the batch size * @param[in] k the number of neighbors to find for each query. * @param[out] neighbors a device pointer to the indices of the neighbors in the source dataset * [n_queries, k] * @param[out] distances a device pointer to the distances to the selected neighbors [n_queries, k] */ template <typename T, typename IdxT> void search(raft::resources const& handle, const search_params& params, const index<IdxT>& idx, const T* queries, uint32_t n_queries, uint32_t k, IdxT* neighbors, float* distances) { return search_with_filtering(handle, params, idx, queries, n_queries, k, neighbors, distances, cuvs::neighbors::filtering::none_ivf_sample_filter{}); } /** * This function is deprecated and will be removed in a future. * Please drop the `mr` argument and use `raft::resource::set_workspace_resource` instead. */ template <typename T, typename IdxT> [[deprecated( "Drop the `mr` argument and use `raft::resource::set_workspace_resource` instead")]] void search(raft::resources const& handle, const search_params& params, const index<IdxT>& idx, const T* queries, uint32_t n_queries, uint32_t k, IdxT* neighbors, float* distances, rmm::mr::device_memory_resource* mr) { return search_with_filtering(handle, params, idx, queries, n_queries, k, neighbors, distances, mr, cuvs::neighbors::filtering::none_ivf_sample_filter{}); } } // namespace cuvs::neighbors::ivf_pq
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/ball_cover-inl.cuh
/* * Copyright (c) 2021-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #ifndef __BALL_COVER_H #define __BALL_COVER_H #pragma once #include <cstdint> #include <cuvs/distance/distance_types.hpp> #include <cuvs/neighbors/ball_cover_types.hpp> #include <cuvs/spatial/knn/detail/ball_cover.cuh> #include <cuvs/spatial/knn/detail/ball_cover/common.cuh> #include <thrust/transform.h> namespace cuvs::neighbors::ball_cover { /** * @defgroup random_ball_cover Random Ball Cover algorithm * @{ */ /** * Builds and populates a previously unbuilt BallCoverIndex * * Usage example: * @code{.cpp} * * #include <raft/core/resources.hpp> * #include <cuvs/neighbors/ball_cover.cuh> * #include <cuvs/distance/distance_types.hpp> * using namespace cuvs::neighbors; * * raft::resources handle; * ... * auto metric = cuvs::distance::DistanceType::L2Expanded; * BallCoverIndex index(handle, X, metric); * * ball_cover::build_index(handle, index); * @endcode * * @tparam idx_t knn index type * @tparam value_t knn value type * @tparam int_t integral type for knn params * @tparam matrix_idx_t matrix indexing type * @param[in] handle library resource management handle * @param[inout] index an empty (and not previous built) instance of BallCoverIndex */ template <typename idx_t, typename value_t, typename int_t, typename matrix_idx_t> void build_index(raft::resources const& handle, BallCoverIndex<idx_t, value_t, int_t, matrix_idx_t>& index) { ASSERT(index.n <= 3, "only 2d and 3d vectors are supported in current implementation"); if (index.metric == cuvs::distance::DistanceType::Haversine) { cuvs::spatial::knn::detail::rbc_build_index( handle, index, spatial::knn::detail::HaversineFunc<value_t, int_t>()); } else if (index.metric == cuvs::distance::DistanceType::L2SqrtExpanded || index.metric == cuvs::distance::DistanceType::L2SqrtUnexpanded) { cuvs::spatial::knn::detail::rbc_build_index( handle, index, spatial::knn::detail::EuclideanFunc<value_t, int_t>()); } else { RAFT_FAIL("Metric not support"); } index.set_index_trained(); } /** @} */ // end group random_ball_cover /** * Performs a faster exact knn in metric spaces using the triangle * inequality with a number of landmark points to reduce the * number of distance computations from O(n^2) to O(sqrt(n)). This * performs an all neighbors knn, which can reuse memory when * the index and query are the same array. This function will * build the index and assumes rbc_build_index() has not already * been called. * @tparam idx_t knn index type * @tparam value_t knn distance type * @tparam int_t type for integers, such as number of rows/cols * @param[in] handle raft handle for resource management * @param[inout] index ball cover index which has not yet been built * @param[in] k number of nearest neighbors to find * @param[in] perform_post_filtering if this is false, only the closest k landmarks * are considered (which will return approximate * results). * @param[out] inds output knn indices * @param[out] dists output knn distances * @param[in] weight a weight for overlap between the closest landmark and * the radius of other landmarks when pruning distances. * Setting this value below 1 can effectively turn off * computing distances against many other balls, enabling * approximate nearest neighbors. Recall can be adjusted * based on how many relevant balls are ignored. Note that * many datasets can still have great recall even by only * looking in the closest landmark. */ template <typename idx_t, typename value_t, typename int_t, typename matrix_idx_t> void all_knn_query(raft::resources const& handle, BallCoverIndex<idx_t, value_t, int_t, matrix_idx_t>& index, int_t k, idx_t* inds, value_t* dists, bool perform_post_filtering = true, float weight = 1.0) { ASSERT(index.n <= 3, "only 2d and 3d vectors are supported in current implementation"); if (index.metric == cuvs::distance::DistanceType::Haversine) { cuvs::spatial::knn::detail::rbc_all_knn_query( handle, index, k, inds, dists, spatial::knn::detail::HaversineFunc<value_t, int_t>(), perform_post_filtering, weight); } else if (index.metric == cuvs::distance::DistanceType::L2SqrtExpanded || index.metric == cuvs::distance::DistanceType::L2SqrtUnexpanded) { cuvs::spatial::knn::detail::rbc_all_knn_query( handle, index, k, inds, dists, spatial::knn::detail::EuclideanFunc<value_t, int_t>(), perform_post_filtering, weight); } else { RAFT_FAIL("Metric not supported"); } index.set_index_trained(); } /** * @ingroup random_ball_cover * @{ */ /** * Performs a faster exact knn in metric spaces using the triangle * inequality with a number of landmark points to reduce the * number of distance computations from O(n^2) to O(sqrt(n)). This * performs an all neighbors knn, which can reuse memory when * the index and query are the same array. This function will * build the index and assumes rbc_build_index() has not already * been called. * * Usage example: * @code{.cpp} * * #include <raft/core/resources.hpp> * #include <cuvs/neighbors/ball_cover.cuh> * #include <cuvs/distance/distance_types.hpp> * using namespace cuvs::neighbors; * * raft::resources handle; * ... * auto metric = cuvs::distance::DistanceType::L2Expanded; * * // Construct a ball cover index * BallCoverIndex index(handle, X, metric); * * // Perform all neighbors knn query * ball_cover::all_knn_query(handle, index, inds, dists, k); * @endcode * * @tparam idx_t knn index type * @tparam value_t knn distance type * @tparam int_t type for integers, such as number of rows/cols * @tparam matrix_idx_t matrix indexing type * * @param[in] handle raft handle for resource management * @param[in] index ball cover index which has not yet been built * @param[out] inds output knn indices * @param[out] dists output knn distances * @param[in] k number of nearest neighbors to find * @param[in] perform_post_filtering if this is false, only the closest k landmarks * are considered (which will return approximate * results). * @param[in] weight a weight for overlap between the closest landmark and * the radius of other landmarks when pruning distances. * Setting this value below 1 can effectively turn off * computing distances against many other balls, enabling * approximate nearest neighbors. Recall can be adjusted * based on how many relevant balls are ignored. Note that * many datasets can still have great recall even by only * looking in the closest landmark. */ template <typename idx_t, typename value_t, typename int_t, typename matrix_idx_t> void all_knn_query(raft::resources const& handle, BallCoverIndex<idx_t, value_t, int_t, matrix_idx_t>& index, raft::device_matrix_view<idx_t, matrix_idx_t, raft::row_major> inds, raft::device_matrix_view<value_t, matrix_idx_t, raft::row_major> dists, int_t k, bool perform_post_filtering = true, float weight = 1.0) { RAFT_EXPECTS(index.n <= 3, "only 2d and 3d vectors are supported in current implementation"); RAFT_EXPECTS(k <= index.m, "k must be less than or equal to the number of data points in the index"); RAFT_EXPECTS(inds.extent(1) == dists.extent(1) && dists.extent(1) == static_cast<matrix_idx_t>(k), "Number of columns in output indices and distances matrices must be equal to k"); RAFT_EXPECTS(inds.extent(0) == dists.extent(0) && dists.extent(0) == index.get_X().extent(0), "Number of rows in output indices and distances matrices must equal number of rows " "in index matrix."); all_knn_query( handle, index, k, inds.data_handle(), dists.data_handle(), perform_post_filtering, weight); } /** @} */ /** * Performs a faster exact knn in metric spaces using the triangle * inequality with a number of landmark points to reduce the * number of distance computations from O(n^2) to O(sqrt(n)). This * function does not build the index and assumes rbc_build_index() has * already been called. Use this function when the index and * query arrays are different, otherwise use rbc_all_knn_query(). * @tparam idx_t index type * @tparam value_t distances type * @tparam int_t integer type for size info * @param[in] handle raft handle for resource management * @param[inout] index ball cover index which has not yet been built * @param[in] k number of nearest neighbors to find * @param[in] query the * @param[in] perform_post_filtering if this is false, only the closest k landmarks * are considered (which will return approximate * results). * @param[out] inds output knn indices * @param[out] dists output knn distances * @param[in] weight a weight for overlap between the closest landmark and * the radius of other landmarks when pruning distances. * Setting this value below 1 can effectively turn off * computing distances against many other balls, enabling * approximate nearest neighbors. Recall can be adjusted * based on how many relevant balls are ignored. Note that * many datasets can still have great recall even by only * looking in the closest landmark. * @param[in] n_query_pts number of query points */ template <typename idx_t, typename value_t, typename int_t> void knn_query(raft::resources const& handle, const BallCoverIndex<idx_t, value_t, int_t>& index, int_t k, const value_t* query, int_t n_query_pts, idx_t* inds, value_t* dists, bool perform_post_filtering = true, float weight = 1.0) { ASSERT(index.n <= 3, "only 2d and 3d vectors are supported in current implementation"); if (index.metric == cuvs::distance::DistanceType::Haversine) { cuvs::spatial::knn::detail::rbc_knn_query(handle, index, k, query, n_query_pts, inds, dists, spatial::knn::detail::HaversineFunc<value_t, int_t>(), perform_post_filtering, weight); } else if (index.metric == cuvs::distance::DistanceType::L2SqrtExpanded || index.metric == cuvs::distance::DistanceType::L2SqrtUnexpanded) { cuvs::spatial::knn::detail::rbc_knn_query(handle, index, k, query, n_query_pts, inds, dists, spatial::knn::detail::EuclideanFunc<value_t, int_t>(), perform_post_filtering, weight); } else { RAFT_FAIL("Metric not supported"); } } /** * @ingroup random_ball_cover * @{ */ /** * Performs a faster exact knn in metric spaces using the triangle * inequality with a number of landmark points to reduce the * number of distance computations from O(n^2) to O(sqrt(n)). This * function does not build the index and assumes rbc_build_index() has * already been called. Use this function when the index and * query arrays are different, otherwise use rbc_all_knn_query(). * * Usage example: * @code{.cpp} * * #include <raft/core/resources.hpp> * #include <cuvs/neighbors/ball_cover.cuh> * #include <cuvs/distance/distance_types.hpp> * using namespace cuvs::neighbors; * * raft::resources handle; * ... * auto metric = cuvs::distance::DistanceType::L2Expanded; * * // Build a ball cover index * BallCoverIndex index(handle, X, metric); * ball_cover::build_index(handle, index); * * // Perform all neighbors knn query * ball_cover::knn_query(handle, index, inds, dists, k); * @endcode * * @tparam idx_t index type * @tparam value_t distances type * @tparam int_t integer type for size info * @tparam matrix_idx_t * @param[in] handle raft handle for resource management * @param[in] index ball cover index which has not yet been built * @param[in] query device matrix containing query data points * @param[out] inds output knn indices * @param[out] dists output knn distances * @param[in] k number of nearest neighbors to find * @param[in] perform_post_filtering if this is false, only the closest k landmarks * are considered (which will return approximate * results). * @param[in] weight a weight for overlap between the closest landmark and * the radius of other landmarks when pruning distances. * Setting this value below 1 can effectively turn off * computing distances against many other balls, enabling * approximate nearest neighbors. Recall can be adjusted * based on how many relevant balls are ignored. Note that * many datasets can still have great recall even by only * looking in the closest landmark. */ template <typename idx_t, typename value_t, typename int_t, typename matrix_idx_t> void knn_query(raft::resources const& handle, const BallCoverIndex<idx_t, value_t, int_t, matrix_idx_t>& index, raft::device_matrix_view<const value_t, matrix_idx_t, raft::row_major> query, raft::device_matrix_view<idx_t, matrix_idx_t, raft::row_major> inds, raft::device_matrix_view<value_t, matrix_idx_t, raft::row_major> dists, int_t k, bool perform_post_filtering = true, float weight = 1.0) { RAFT_EXPECTS(k <= index.m, "k must be less than or equal to the number of data points in the index"); RAFT_EXPECTS(inds.extent(1) == dists.extent(1) && dists.extent(1) == static_cast<idx_t>(k), "Number of columns in output indices and distances matrices must be equal to k"); RAFT_EXPECTS(inds.extent(0) == dists.extent(0) && dists.extent(0) == query.extent(0), "Number of rows in output indices and distances matrices must equal number of rows " "in search matrix."); RAFT_EXPECTS(query.extent(1) == index.get_X().extent(1), "Number of columns in query and index matrices must match."); knn_query(handle, index, k, query.data_handle(), query.extent(0), inds.data_handle(), dists.data_handle(), perform_post_filtering, weight); } /** @} */ // TODO: implement functions for: // 4. rbc_eps_neigh() - given a populated index, perform query against different query array // 5. rbc_all_eps_neigh() - populate a BallCoverIndex and query against training data } // namespace cuvs::neighbors::ball_cover #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/sample_filter_types.hpp
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cstddef> #include <cstdint> #include <raft/core/detail/macros.hpp> namespace cuvs::neighbors::filtering { /* A filter that filters nothing. This is the default behavior. */ struct none_ivf_sample_filter { inline _RAFT_HOST_DEVICE bool operator()( // query index const uint32_t query_ix, // the current inverted list index const uint32_t cluster_ix, // the index of the current sample inside the current inverted list const uint32_t sample_ix) const { return true; } }; /* A filter that filters nothing. This is the default behavior. */ struct none_cagra_sample_filter { inline _RAFT_HOST_DEVICE bool operator()( // query index const uint32_t query_ix, // the index of the current sample const uint32_t sample_ix) const { return true; } }; template <typename filter_t, typename = void> struct takes_three_args : std::false_type {}; template <typename filter_t> struct takes_three_args< filter_t, std::void_t<decltype(std::declval<filter_t>()(uint32_t{}, uint32_t{}, uint32_t{}))>> : std::true_type {}; /** * @brief Filter used to convert the cluster index and sample index * of an IVF search into a sample index. This can be used as an * intermediate filter. * * @tparam index_t Indexing type * @tparam filter_t */ template <typename index_t, typename filter_t> struct ivf_to_sample_filter { const index_t* const* inds_ptrs_; const filter_t next_filter_; ivf_to_sample_filter(const index_t* const* inds_ptrs, const filter_t next_filter) : inds_ptrs_{inds_ptrs}, next_filter_{next_filter} { } /** If the original filter takes three arguments, then don't modify the arguments. * If the original filter takes two arguments, then we are using `inds_ptr_` to obtain the sample * index. */ inline _RAFT_HOST_DEVICE bool operator()( // query index const uint32_t query_ix, // the current inverted list index const uint32_t cluster_ix, // the index of the current sample inside the current inverted list const uint32_t sample_ix) const { if constexpr (takes_three_args<filter_t>::value) { return next_filter_(query_ix, cluster_ix, sample_ix); } else { return next_filter_(query_ix, inds_ptrs_[cluster_ix][sample_ix]); } } }; /** * If the filtering depends on the index of a sample, then the following * filter template can be used: * * template <typename IdxT> * struct index_ivf_sample_filter { * using index_type = IdxT; * * const index_type* const* inds_ptr = nullptr; * * index_ivf_sample_filter() {} * index_ivf_sample_filter(const index_type* const* _inds_ptr) * : inds_ptr{_inds_ptr} {} * index_ivf_sample_filter(const index_ivf_sample_filter&) = default; * index_ivf_sample_filter(index_ivf_sample_filter&&) = default; * index_ivf_sample_filter& operator=(const index_ivf_sample_filter&) = default; * index_ivf_sample_filter& operator=(index_ivf_sample_filter&&) = default; * * inline _RAFT_HOST_DEVICE bool operator()( * const uint32_t query_ix, * const uint32_t cluster_ix, * const uint32_t sample_ix) const { * index_type database_idx = inds_ptr[cluster_ix][sample_ix]; * * // return true or false, depending on the database_idx * return true; * } * }; * * Initialize it as: * using filter_type = index_ivf_sample_filter<idx_t>; * filter_type filter(raft_ivfpq_index.inds_ptrs().data_handle()); * * Use it as: * cuvs::neighbors::ivf_pq::search_with_filtering<data_t, idx_t, filter_type>( * ...regular parameters here..., * filter * ); * * Another example would be the following filter that greenlights samples according * to a contiguous bit mask vector. * * template <typename IdxT> * struct bitmask_ivf_sample_filter { * using index_type = IdxT; * * const index_type* const* inds_ptr = nullptr; * const uint64_t* const bit_mask_ptr = nullptr; * const int64_t bit_mask_stride_64 = 0; * * bitmask_ivf_sample_filter() {} * bitmask_ivf_sample_filter( * const index_type* const* _inds_ptr, * const uint64_t* const _bit_mask_ptr, * const int64_t _bit_mask_stride_64) * : inds_ptr{_inds_ptr}, * bit_mask_ptr{_bit_mask_ptr}, * bit_mask_stride_64{_bit_mask_stride_64} {} * bitmask_ivf_sample_filter(const bitmask_ivf_sample_filter&) = default; * bitmask_ivf_sample_filter(bitmask_ivf_sample_filter&&) = default; * bitmask_ivf_sample_filter& operator=(const bitmask_ivf_sample_filter&) = default; * bitmask_ivf_sample_filter& operator=(bitmask_ivf_sample_filter&&) = default; * * inline _RAFT_HOST_DEVICE bool operator()( * const uint32_t query_ix, * const uint32_t cluster_ix, * const uint32_t sample_ix) const { * const index_type database_idx = inds_ptr[cluster_ix][sample_ix]; * const uint64_t bit_mask_element = * bit_mask_ptr[query_ix * bit_mask_stride_64 + database_idx / 64]; * const uint64_t masked_bool = * bit_mask_element & (1ULL << (uint64_t)(database_idx % 64)); * const bool is_bit_set = (masked_bool != 0); * * return is_bit_set; * } * }; */ } // namespace cuvs::neighbors::filtering
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/nn_descent.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "detail/nn_descent.cuh" #include <raft/core/device_mdspan.hpp> #include <raft/core/host_mdspan.hpp> namespace cuvs::neighbors::experimental::nn_descent { /** * @defgroup nn-descent CUDA gradient descent nearest neighbor * @{ */ /** * @brief Build nn-descent Index with dataset in device memory * * The following distance metrics are supported: * - L2 * * Usage example: * @code{.cpp} * using namespace cuvs::neighbors::experimental; * // use default index parameters * nn_descent::index_params index_params; * // create and fill the index from a [N, D] raft::device_matrix_view dataset * auto index = cagra::build(res, index_params, dataset); * // index.graph() provides a raft::host_matrix_view of an * // all-neighbors knn graph of dimensions [N, k] of the input * // dataset * @endcode * * @tparam T data-type of the input dataset * @tparam IdxT data-type for the output index * @param[in] res raft::resources is an object mangaging resources * @param[in] params an instance of nn_descent::index_params that are parameters * to run the nn-descent algorithm * @param[in] dataset raft::device_matrix_view input dataset expected to be located * in device memory * @return index<IdxT> index containing all-neighbors knn graph in host memory */ template <typename T, typename IdxT = uint32_t> index<IdxT> build(raft::resources const& res, index_params const& params, raft::device_matrix_view<const T, int64_t, raft::row_major> dataset) { return detail::build<T, IdxT>(res, params, dataset); } /** * @brief Build nn-descent Index with dataset in device memory * * The following distance metrics are supported: * - L2 * * Usage example: * @code{.cpp} * using namespace cuvs::neighbors::experimental; * // use default index parameters * nn_descent::index_params index_params; * // create and fill the index from a [N, D] raft::device_matrix_view dataset * auto knn_graph = raft::make_host_matrix<uint32_t, int64_t>(N, D); * auto index = nn_descent::index{res, knn_graph.view()}; * cagra::build(res, index_params, dataset, index); * // index.graph() provides a raft::host_matrix_view of an * // all-neighbors knn graph of dimensions [N, k] of the input * // dataset * @endcode * * @tparam T data-type of the input dataset * @tparam IdxT data-type for the output index * @param res raft::resources is an object mangaging resources * @param[in] params an instance of nn_descent::index_params that are parameters * to run the nn-descent algorithm * @param[in] dataset raft::device_matrix_view input dataset expected to be located * in device memory * @param[out] idx cuvs::neighbors::experimental::nn_descentindex containing all-neighbors knn graph * in host memory */ template <typename T, typename IdxT = uint32_t> void build(raft::resources const& res, index_params const& params, raft::device_matrix_view<const T, int64_t, raft::row_major> dataset, index<IdxT>& idx) { detail::build<T, IdxT>(res, params, dataset, idx); } /** * @brief Build nn-descent Index with dataset in host memory * * The following distance metrics are supported: * - L2 * * Usage example: * @code{.cpp} * using namespace cuvs::neighbors::experimental; * // use default index parameters * nn_descent::index_params index_params; * // create and fill the index from a [N, D] raft::host_matrix_view dataset * auto index = cagra::build(res, index_params, dataset); * // index.graph() provides a raft::host_matrix_view of an * // all-neighbors knn graph of dimensions [N, k] of the input * // dataset * @endcode * * @tparam T data-type of the input dataset * @tparam IdxT data-type for the output index * @param res raft::resources is an object mangaging resources * @param[in] params an instance of nn_descent::index_params that are parameters * to run the nn-descent algorithm * @param[in] dataset raft::host_matrix_view input dataset expected to be located * in host memory * @return index<IdxT> index containing all-neighbors knn graph in host memory */ template <typename T, typename IdxT = uint32_t> index<IdxT> build(raft::resources const& res, index_params const& params, raft::host_matrix_view<const T, int64_t, raft::row_major> dataset) { return detail::build<T, IdxT>(res, params, dataset); } /** * @brief Build nn-descent Index with dataset in host memory * * The following distance metrics are supported: * - L2 * * Usage example: * @code{.cpp} * using namespace cuvs::neighbors::experimental; * // use default index parameters * nn_descent::index_params index_params; * // create and fill the index from a [N, D] raft::host_matrix_view dataset * auto knn_graph = raft::make_host_matrix<uint32_t, int64_t>(N, D); * auto index = nn_descent::index{res, knn_graph.view()}; * cagra::build(res, index_params, dataset, index); * // index.graph() provides a raft::host_matrix_view of an * // all-neighbors knn graph of dimensions [N, k] of the input * // dataset * @endcode * * @tparam T data-type of the input dataset * @tparam IdxT data-type for the output index * @param[in] res raft::resources is an object mangaging resources * @param[in] params an instance of nn_descent::index_params that are parameters * to run the nn-descent algorithm * @param[in] dataset raft::host_matrix_view input dataset expected to be located * in host memory * @param[out] idx cuvs::neighbors::experimental::nn_descentindex containing all-neighbors knn graph * in host memory */ template <typename T, typename IdxT = uint32_t> void build(raft::resources const& res, index_params const& params, raft::host_matrix_view<const T, int64_t, raft::row_major> dataset, index<IdxT>& idx) { detail::build<T, IdxT>(res, params, dataset, idx); } /** @} */ // end group nn-descent } // namespace cuvs::neighbors::experimental::nn_descent
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/ivf_pq_types.hpp
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/neighbors/ann_types.hpp> #include <cuvs/neighbors/ivf_list_types.hpp> #include <cuvs/distance/distance_types.hpp> #include <raft/core/device_mdarray.hpp> #include <raft/core/error.hpp> #include <raft/core/host_mdarray.hpp> #include <raft/core/mdspan_types.hpp> #include <raft/core/resources.hpp> #include <raft/util/integer_utils.hpp> #include <thrust/fill.h> #include <memory> #include <type_traits> namespace cuvs::neighbors::ivf_pq { /** * @addtogroup ivf_pq * @{ */ /** A type for specifying how PQ codebooks are created. */ enum class codebook_gen { // NOLINT PER_SUBSPACE = 0, // NOLINT PER_CLUSTER = 1, // NOLINT }; struct index_params : ann::index_params { /** * The number of inverted lists (clusters) * * Hint: the number of vectors per cluster (`n_rows/n_lists`) should be approximately 1,000 to * 10,000. */ uint32_t n_lists = 1024; /** The number of iterations searching for kmeans centers (index building). */ uint32_t kmeans_n_iters = 20; /** The fraction of data to use during iterative kmeans building. */ double kmeans_trainset_fraction = 0.5; /** * The bit length of the vector element after compression by PQ. * * Possible values: [4, 5, 6, 7, 8]. * * Hint: the smaller the 'pq_bits', the smaller the index size and the better the search * performance, but the lower the recall. */ uint32_t pq_bits = 8; /** * The dimensionality of the vector after compression by PQ. When zero, an optimal value is * selected using a heuristic. * * NB: `pq_dim * pq_bits` must be a multiple of 8. * * Hint: a smaller 'pq_dim' results in a smaller index size and better search performance, but * lower recall. If 'pq_bits' is 8, 'pq_dim' can be set to any number, but multiple of 8 are * desirable for good performance. If 'pq_bits' is not 8, 'pq_dim' should be a multiple of 8. * For good performance, it is desirable that 'pq_dim' is a multiple of 32. Ideally, 'pq_dim' * should be also a divisor of the dataset dim. */ uint32_t pq_dim = 0; /** How PQ codebooks are created. */ codebook_gen codebook_kind = codebook_gen::PER_SUBSPACE; /** * Apply a random rotation matrix on the input data and queries even if `dim % pq_dim == 0`. * * Note: if `dim` is not multiple of `pq_dim`, a random rotation is always applied to the input * data and queries to transform the working space from `dim` to `rot_dim`, which may be slightly * larger than the original space and and is a multiple of `pq_dim` (`rot_dim % pq_dim == 0`). * However, this transform is not necessary when `dim` is multiple of `pq_dim` * (`dim == rot_dim`, hence no need in adding "extra" data columns / features). * * By default, if `dim == rot_dim`, the rotation transform is initialized with the identity * matrix. When `force_random_rotation == true`, a random orthogonal transform matrix is generated * regardless of the values of `dim` and `pq_dim`. */ bool force_random_rotation = false; /** * By default, the algorithm allocates more space than necessary for individual clusters * (`list_data`). This allows to amortize the cost of memory allocation and reduce the number of * data copies during repeated calls to `extend` (extending the database). * * The alternative is the conservative allocation behavior; when enabled, the algorithm always * allocates the minimum amount of memory required to store the given number of records. Set this * flag to `true` if you prefer to use as little GPU memory for the database as possible. */ bool conservative_memory_allocation = false; }; struct search_params : ann::search_params { /** The number of clusters to search. */ uint32_t n_probes = 20; /** * Data type of look up table to be created dynamically at search time. * * Possible values: [CUDA_R_32F, CUDA_R_16F, CUDA_R_8U] * * The use of low-precision types reduces the amount of shared memory required at search time, so * fast shared memory kernels can be used even for datasets with large dimansionality. Note that * the recall is slightly degraded when low-precision type is selected. */ cudaDataType_t lut_dtype = CUDA_R_32F; /** * Storage data type for distance/similarity computed at search time. * * Possible values: [CUDA_R_16F, CUDA_R_32F] * * If the performance limiter at search time is device memory access, selecting FP16 will improve * performance slightly. */ cudaDataType_t internal_distance_dtype = CUDA_R_32F; /** * Preferred fraction of SM's unified memory / L1 cache to be used as shared memory. * * Possible values: [0.0 - 1.0] as a fraction of the `sharedMemPerMultiprocessor`. * * One wants to increase the carveout to make sure a good GPU occupancy for the main search * kernel, but not to keep it too high to leave some memory to be used as L1 cache. Note, this * value is interpreted only as a hint. Moreover, a GPU usually allows only a fixed set of cache * configurations, so the provided value is rounded up to the nearest configuration. Refer to the * NVIDIA tuning guide for the target GPU architecture. * * Note, this is a low-level tuning parameter that can have drastic negative effects on the search * performance if tweaked incorrectly. */ double preferred_shmem_carveout = 1.0; }; static_assert(std::is_aggregate_v<index_params>); static_assert(std::is_aggregate_v<search_params>); /** Size of the interleaved group. */ constexpr static uint32_t kIndexGroupSize = 32; /** Stride of the interleaved group for vectorized loads. */ constexpr static uint32_t kIndexGroupVecLen = 16; /** * Default value returned by `search` when the `n_probes` is too small and top-k is too large. * One may encounter it if the combined size of probed clusters is smaller than the requested * number of results per query. */ template <typename IdxT> constexpr static IdxT kOutOfBoundsRecord = std::numeric_limits<IdxT>::max(); template <typename SizeT, typename IdxT> struct list_spec { using value_type = uint8_t; using index_type = IdxT; /** PQ-encoded data stored in the interleaved format: * * [ raft::ceildiv(list_size, kIndexGroupSize) * , raft::ceildiv(pq_dim, (kIndexGroupVecLen * 8u) / pq_bits) * , kIndexGroupSize * , kIndexGroupVecLen * ]. */ using list_extents = extents<SizeT, raft::dynamic_extent, raft::dynamic_extent, kIndexGroupSize, kIndexGroupVecLen>; SizeT align_max; SizeT align_min; uint32_t pq_bits; uint32_t pq_dim; constexpr list_spec(uint32_t pq_bits, uint32_t pq_dim, bool conservative_memory_allocation) : pq_bits(pq_bits), pq_dim(pq_dim), align_min(kIndexGroupSize), align_max(conservative_memory_allocation ? kIndexGroupSize : 1024) { } // Allow casting between different size-types (for safer size and offset calculations) template <typename OtherSizeT> constexpr explicit list_spec(const list_spec<OtherSizeT, IdxT>& other_spec) : pq_bits{other_spec.pq_bits}, pq_dim{other_spec.pq_dim}, align_min{other_spec.align_min}, align_max{other_spec.align_max} { } /** Determine the extents of an array enough to hold a given amount of data. */ constexpr auto make_list_extents(SizeT n_rows) const -> list_extents { // how many elems of pq_dim fit into one kIndexGroupVecLen-byte chunk auto pq_chunk = (kIndexGroupVecLen * 8u) / pq_bits; return make_extents<SizeT>(div_rounding_up_safe<SizeT>(n_rows, kIndexGroupSize), div_rounding_up_safe<SizeT>(pq_dim, pq_chunk), kIndexGroupSize, kIndexGroupVecLen); } }; template <typename IdxT, typename SizeT = uint32_t> using list_data = ivf::list<list_spec, SizeT, IdxT>; /** * @brief IVF-PQ index. * * In the IVF-PQ index, a database vector y is approximated with two level quantization: * * y = Q_1(y) + Q_2(y - Q_1(y)) * * The first level quantizer (Q_1), maps the vector y to the nearest cluster center. The number of * clusters is n_lists. * * The second quantizer encodes the residual, and it is defined as a product quantizer [1]. * * A product quantizer encodes a `dim` dimensional vector with a `pq_dim` dimensional vector. * First we split the input vector into `pq_dim` subvectors (denoted by u), where each u vector * contains `pq_len` distinct components of y * * y_1, y_2, ... y_{pq_len}, y_{pq_len+1}, ... y_{2*pq_len}, ... y_{dim-pq_len+1} ... y_{dim} * \___________________/ \____________________________/ \______________________/ * u_1 u_2 u_{pq_dim} * * Then each subvector encoded with a separate quantizer q_i, end the results are concatenated * * Q_2(y) = q_1(u_1),q_2(u_2),...,q_{pq_dim}(u_pq_dim}) * * Each quantizer q_i outputs a code with pq_bit bits. The second level quantizers are also defined * by k-means clustering in the corresponding sub-space: the reproduction values are the centroids, * and the set of reproduction values is the codebook. * * When the data dimensionality `dim` is not multiple of `pq_dim`, the feature space is transformed * using a random orthogonal matrix to have `rot_dim = pq_dim * pq_len` dimensions * (`rot_dim >= dim`). * * The second-level quantizers are trained either for each subspace or for each cluster: * (a) codebook_gen::PER_SUBSPACE: * creates `pq_dim` second-level quantizers - one for each slice of the data along features; * (b) codebook_gen::PER_CLUSTER: * creates `n_lists` second-level quantizers - one for each first-level cluster. * In either case, the centroids are again found using k-means clustering interpreting the data as * having pq_len dimensions. * * [1] Product quantization for nearest neighbor search Herve Jegou, Matthijs Douze, Cordelia Schmid * * @tparam IdxT type of the indices in the source dataset * */ template <typename IdxT> struct index : ann::index { static_assert(!raft::is_narrowing_v<uint32_t, IdxT>, "IdxT must be able to represent all values of uint32_t"); public: /** Total length of the index. */ [[nodiscard]] constexpr inline auto size() const noexcept -> IdxT { return accum_sorted_sizes_(n_lists()); } /** Dimensionality of the input data. */ [[nodiscard]] constexpr inline auto dim() const noexcept -> uint32_t { return dim_; } /** * Dimensionality of the cluster centers: * input data dim extended with vector norms and padded to 8 elems. */ [[nodiscard]] constexpr inline auto dim_ext() const noexcept -> uint32_t { return raft::round_up_safe(dim() + 1, 8u); } /** * Dimensionality of the data after transforming it for PQ processing * (rotated and augmented to be muplitple of `pq_dim`). */ [[nodiscard]] constexpr inline auto rot_dim() const noexcept -> uint32_t { return pq_len() * pq_dim(); } /** The bit length of an encoded vector element after compression by PQ. */ [[nodiscard]] constexpr inline auto pq_bits() const noexcept -> uint32_t { return pq_bits_; } /** The dimensionality of an encoded vector after compression by PQ. */ [[nodiscard]] constexpr inline auto pq_dim() const noexcept -> uint32_t { return pq_dim_; } /** Dimensionality of a subspaces, i.e. the number of vector components mapped to a subspace */ [[nodiscard]] constexpr inline auto pq_len() const noexcept -> uint32_t { return raft::div_rounding_up_unsafe(dim(), pq_dim()); } /** The number of vectors in a PQ codebook (`1 << pq_bits`). */ [[nodiscard]] constexpr inline auto pq_book_size() const noexcept -> uint32_t { return 1 << pq_bits(); } /** Distance metric used for clustering. */ [[nodiscard]] constexpr inline auto metric() const noexcept -> cuvs::distance::DistanceType { return metric_; } /** How PQ codebooks are created. */ [[nodiscard]] constexpr inline auto codebook_kind() const noexcept -> codebook_gen { return codebook_kind_; } /** Number of clusters/inverted lists (first level quantization). */ [[nodiscard]] constexpr inline auto n_lists() const noexcept -> uint32_t { return lists_.size(); } /** * Whether to use convervative memory allocation when extending the list (cluster) data * (see index_params.conservative_memory_allocation). */ [[nodiscard]] constexpr inline auto conservative_memory_allocation() const noexcept -> bool { return conservative_memory_allocation_; } // Don't allow copying the index for performance reasons (try avoiding copying data) index(const index&) = delete; index(index&&) = default; auto operator=(const index&) -> index& = delete; auto operator=(index&&) -> index& = default; ~index() = default; /** Construct an empty index. It needs to be trained and then populated. */ index(raft::resources const& handle, cuvs::distance::DistanceType metric, codebook_gen codebook_kind, uint32_t n_lists, uint32_t dim, uint32_t pq_bits = 8, uint32_t pq_dim = 0, bool conservative_memory_allocation = false) : ann::index(), metric_(metric), codebook_kind_(codebook_kind), dim_(dim), pq_bits_(pq_bits), pq_dim_(pq_dim == 0 ? calculate_pq_dim(dim) : pq_dim), conservative_memory_allocation_(conservative_memory_allocation), pq_centers_{make_device_mdarray<float>(handle, make_pq_centers_extents())}, lists_{n_lists}, rotation_matrix_{make_device_matrix<float, uint32_t>(handle, this->rot_dim(), this->dim())}, list_sizes_{make_device_vector<uint32_t, uint32_t>(handle, n_lists)}, centers_{make_device_matrix<float, uint32_t>(handle, n_lists, this->dim_ext())}, centers_rot_{make_device_matrix<float, uint32_t>(handle, n_lists, this->rot_dim())}, data_ptrs_{make_device_vector<uint8_t*, uint32_t>(handle, n_lists)}, inds_ptrs_{make_device_vector<IdxT*, uint32_t>(handle, n_lists)}, accum_sorted_sizes_{make_host_vector<IdxT, uint32_t>(n_lists + 1)} { check_consistency(); accum_sorted_sizes_(n_lists) = 0; } /** Construct an empty index. It needs to be trained and then populated. */ index(raft::resources const& handle, const index_params& params, uint32_t dim) : index(handle, params.metric, params.codebook_kind, params.n_lists, dim, params.pq_bits, params.pq_dim, params.conservative_memory_allocation) { } using pq_centers_extents = std::experimental:: extents<uint32_t, raft::dynamic_extent, raft::dynamic_extent, raft::dynamic_extent>; /** * PQ cluster centers * * - codebook_gen::PER_SUBSPACE: [pq_dim , pq_len, pq_book_size] * - codebook_gen::PER_CLUSTER: [n_lists, pq_len, pq_book_size] */ inline auto pq_centers() noexcept -> raft::device_mdspan<float, pq_centers_extents, raft::row_major> { return pq_centers_.view(); } [[nodiscard]] inline auto pq_centers() const noexcept -> raft::device_mdspan<const float, pq_centers_extents, raft::row_major> { return pq_centers_.view(); } /** Lists' data and indices. */ inline auto lists() noexcept -> std::vector<std::shared_ptr<list_data<IdxT>>>& { return lists_; } [[nodiscard]] inline auto lists() const noexcept -> const std::vector<std::shared_ptr<list_data<IdxT>>>& { return lists_; } /** Pointers to the inverted lists (clusters) data [n_lists]. */ inline auto data_ptrs() noexcept -> raft::device_vector_view<uint8_t*, uint32_t, raft::row_major> { return data_ptrs_.view(); } [[nodiscard]] inline auto data_ptrs() const noexcept -> raft::device_vector_view<const uint8_t* const, uint32_t, raft::row_major> { return make_mdspan<const uint8_t* const, uint32_t, raft::row_major, false, true>( data_ptrs_.data_handle(), data_ptrs_.extents()); } /** Pointers to the inverted lists (clusters) indices [n_lists]. */ inline auto inds_ptrs() noexcept -> raft::device_vector_view<IdxT*, uint32_t, raft::row_major> { return inds_ptrs_.view(); } [[nodiscard]] inline auto inds_ptrs() const noexcept -> raft::device_vector_view<const IdxT* const, uint32_t, raft::row_major> { return make_mdspan<const IdxT* const, uint32_t, raft::row_major, false, true>( inds_ptrs_.data_handle(), inds_ptrs_.extents()); } /** The transform matrix (original space -> rotated padded space) [rot_dim, dim] */ inline auto rotation_matrix() noexcept -> raft::device_matrix_view<float, uint32_t, raft::row_major> { return rotation_matrix_.view(); } [[nodiscard]] inline auto rotation_matrix() const noexcept -> raft::device_matrix_view<const float, uint32_t, raft::row_major> { return rotation_matrix_.view(); } /** * Accumulated list sizes, sorted in descending order [n_lists + 1]. * The last value contains the total length of the index. * The value at index zero is always zero. * * That is, the content of this span is as if the `list_sizes` was sorted and then accumulated. * * This span is used during search to estimate the maximum size of the workspace. */ inline auto accum_sorted_sizes() noexcept -> raft::host_vector_view<IdxT, uint32_t, raft::row_major> { return accum_sorted_sizes_.view(); } [[nodiscard]] inline auto accum_sorted_sizes() const noexcept -> raft::host_vector_view<const IdxT, uint32_t, raft::row_major> { return accum_sorted_sizes_.view(); } /** Sizes of the lists [n_lists]. */ inline auto list_sizes() noexcept -> raft::device_vector_view<uint32_t, uint32_t, raft::row_major> { return list_sizes_.view(); } [[nodiscard]] inline auto list_sizes() const noexcept -> raft::device_vector_view<const uint32_t, uint32_t, raft::row_major> { return list_sizes_.view(); } /** Cluster centers corresponding to the lists in the original space [n_lists, dim_ext] */ inline auto centers() noexcept -> raft::device_matrix_view<float, uint32_t, raft::row_major> { return centers_.view(); } [[nodiscard]] inline auto centers() const noexcept -> raft::device_matrix_view<const float, uint32_t, raft::row_major> { return centers_.view(); } /** Cluster centers corresponding to the lists in the rotated space [n_lists, rot_dim] */ inline auto centers_rot() noexcept -> raft::device_matrix_view<float, uint32_t, raft::row_major> { return centers_rot_.view(); } [[nodiscard]] inline auto centers_rot() const noexcept -> raft::device_matrix_view<const float, uint32_t, raft::row_major> { return centers_rot_.view(); } /** fetch size of a particular IVF list in bytes using the list extents. * Usage example: * @code{.cpp} * raft::resources res; * // use default index params * ivf_pq::index_params index_params; * // extend the IVF lists while building the index * index_params.add_data_on_build = true; * // create and fill the index from a [N, D] dataset * auto index = cuvs::neighbors::ivf_pq::build<int64_t>(res, index_params, dataset, N, D); * // Fetch the size of the fourth list * uint32_t size = index.get_list_size_in_bytes(3); * @endcode * * @param[in] label list ID */ inline auto get_list_size_in_bytes(uint32_t label) -> uint32_t { RAFT_EXPECTS(label < this->n_lists(), "Expected label to be less than number of lists in the index"); auto list_data = this->lists()[label]->data; return list_data.size(); } private: cuvs::distance::DistanceType metric_; codebook_gen codebook_kind_; uint32_t dim_; uint32_t pq_bits_; uint32_t pq_dim_; bool conservative_memory_allocation_; // Primary data members std::vector<std::shared_ptr<list_data<IdxT>>> lists_; raft::device_vector<uint32_t, uint32_t, raft::row_major> list_sizes_; raft::device_mdarray<float, pq_centers_extents, raft::row_major> pq_centers_; raft::device_matrix<float, uint32_t, raft::row_major> centers_; raft::device_matrix<float, uint32_t, raft::row_major> centers_rot_; raft::device_matrix<float, uint32_t, raft::row_major> rotation_matrix_; // Computed members for accelerating search. raft::device_vector<uint8_t*, uint32_t, raft::row_major> data_ptrs_; raft::device_vector<IdxT*, uint32_t, raft::row_major> inds_ptrs_; raft::host_vector<IdxT, uint32_t, raft::row_major> accum_sorted_sizes_; /** Throw an error if the index content is inconsistent. */ void check_consistency() { RAFT_EXPECTS(pq_bits() >= 4 && pq_bits() <= 8, "`pq_bits` must be within closed range [4,8], but got %u.", pq_bits()); RAFT_EXPECTS((pq_bits() * pq_dim()) % 8 == 0, "`pq_bits * pq_dim` must be a multiple of 8, but got %u * %u = %u.", pq_bits(), pq_dim(), pq_bits() * pq_dim()); } auto make_pq_centers_extents() -> pq_centers_extents { switch (codebook_kind()) { case codebook_gen::PER_SUBSPACE: return make_extents<uint32_t>(pq_dim(), pq_len(), pq_book_size()); case codebook_gen::PER_CLUSTER: return make_extents<uint32_t>(n_lists(), pq_len(), pq_book_size()); default: RAFT_FAIL("Unreachable code"); } } static inline auto calculate_pq_dim(uint32_t dim) -> uint32_t { // If the dimensionality is large enough, we can reduce it to improve performance if (dim >= 128) { dim /= 2; } // Round it down to 32 to improve performance. auto r = raft::round_down_safe<uint32_t>(dim, 32); if (r > 0) return r; // If the dimensionality is really low, round it to the closest power-of-two r = 1; while ((r << 1) <= dim) { r = r << 1; } return r; } }; /** @} */ } // namespace cuvs::neighbors::ivf_pq
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/cagra.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "detail/cagra/cagra_build.cuh" #include "detail/cagra/cagra_search.cuh" #include "detail/cagra/graph_core.cuh" #include <cuvs/neighbors/cagra_types.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/host_device_accessor.hpp> #include <raft/core/mdspan.hpp> #include <raft/core/resources.hpp> #include <rmm/cuda_stream_view.hpp> namespace cuvs::neighbors::cagra { /** * @defgroup cagra CUDA ANN Graph-based nearest neighbor search * @{ */ /** * @brief Build a kNN graph using IVF-PQ. * * The kNN graph is the first building block for CAGRA index. * * The output is a dense matrix that stores the neighbor indices for each point in the dataset. * Each point has the same number of neighbors. * * See [cagra::build](#cagra::build) for an alternative method. * * The following distance metrics are supported: * - L2Expanded * * Usage example: * @code{.cpp} * using namespace cuvs::neighbors; * // use default index parameters * ivf_pq::index_params build_params; * ivf_pq::search_params search_params * auto knn_graph = raft::make_host_matrix<IdxT, IdxT>(dataset.extent(0), 128); * // create knn graph * cagra::build_knn_graph(res, dataset, knn_graph.view(), 2, build_params, search_params); * auto optimized_gaph = raft::make_host_matrix<IdxT, IdxT>(dataset.extent(0), 64); * cagra::optimize(res, dataset, knn_graph.view(), optimized_graph.view()); * // Construct an index from dataset and optimized knn_graph * auto index = cagra::index<T, IdxT>(res, build_params.metric(), dataset, * optimized_graph.view()); * @endcode * * @tparam DataT data element type * @tparam IdxT type of the dataset vector indices * * @param[in] res raft resources * @param[in] dataset a matrix view (host or device) to a row-major matrix [n_rows, dim] * @param[out] knn_graph a host matrix view to store the output knn graph [n_rows, graph_degree] * @param[in] refine_rate (optional) refinement rate for ivf-pq search * @param[in] build_params (optional) ivf_pq index building parameters for knn graph * @param[in] search_params (optional) ivf_pq search parameters */ template <typename DataT, typename IdxT, typename accessor> void build_knn_graph( raft::resources const& res, raft::mdspan<const DataT, raft::matrix_extent<int64_t>, raft::row_major, accessor> dataset, raft::host_matrix_view<IdxT, int64_t, raft::row_major> knn_graph, std::optional<float> refine_rate = std::nullopt, std::optional<ivf_pq::index_params> build_params = std::nullopt, std::optional<ivf_pq::search_params> search_params = std::nullopt) { using internal_IdxT = typename std::make_unsigned<IdxT>::type; auto knn_graph_internal = raft::make_host_matrix_view<internal_IdxT, int64_t>( reinterpret_cast<internal_IdxT*>(knn_graph.data_handle()), knn_graph.extent(0), knn_graph.extent(1)); auto dataset_internal = raft::mdspan<const DataT, raft::matrix_extent<int64_t>, raft::row_major, accessor>( dataset.data_handle(), dataset.extent(0), dataset.extent(1)); cagra::detail::build_knn_graph( res, dataset_internal, knn_graph_internal, refine_rate, build_params, search_params); } /** * @brief Build a kNN graph using NN-descent. * * The kNN graph is the first building block for CAGRA index. * * The output is a dense matrix that stores the neighbor indices for each point in the dataset. * Each point has the same number of neighbors. * * See [cagra::build](#cagra::build) for an alternative method. * * The following distance metrics are supported: * - L2Expanded * * Usage example: * @code{.cpp} * using namespace cuvs::neighbors; * using namespace cuvs::neighbors::experimental; * // use default index parameters * nn_descent::index_params build_params; * build_params.graph_degree = 128; * auto knn_graph = raft::make_host_matrix<IdxT, IdxT>(dataset.extent(0), 128); * // create knn graph * cagra::build_knn_graph(res, dataset, knn_graph.view(), build_params); * auto optimized_gaph = raft::make_host_matrix<IdxT, int64_t>(dataset.extent(0), 64); * cagra::optimize(res, dataset, nn_descent_index.graph.view(), optimized_graph.view()); * // Construct an index from dataset and optimized knn_graph * auto index = cagra::index<T, IdxT>(res, build_params.metric(), dataset, * optimized_graph.view()); * @endcode * * @tparam DataT data element type * @tparam IdxT type of the dataset vector indices * @tparam accessor host or device accessor_type for the dataset * @param[in] res raft::resources is an object mangaging resources * @param[in] dataset input raft::host/device_matrix_view that can be located in * in host or device memory * @param[out] knn_graph a host matrix view to store the output knn graph [n_rows, graph_degree] * @param[in] build_params an instance of experimental::nn_descent::index_params that are parameters * to run the nn-descent algorithm */ template <typename DataT, typename IdxT = uint32_t, typename accessor = host_device_accessor<std::experimental::default_accessor<DataT>, memory_type::device>> void build_knn_graph( raft::resources const& res, raft::mdspan<const DataT, raft::matrix_extent<int64_t>, raft::row_major, accessor> dataset, raft::host_matrix_view<IdxT, int64_t, raft::row_major> knn_graph, experimental::nn_descent::index_params build_params) { detail::build_knn_graph<DataT, IdxT>(res, dataset, knn_graph, build_params); } /** * @brief Sort a KNN graph index. * Preprocessing step for `cagra::optimize`: If a KNN graph is not built using * `cagra::build_knn_graph`, then it is necessary to call this function before calling * `cagra::optimize`. If the graph is built by `cagra::build_knn_graph`, it is already sorted and * you do not need to call this function. * * Usage example: * @code{.cpp} * using namespace cuvs::neighbors; * cagra::index_params build_params; * auto knn_graph = raft::make_host_matrix<IdxT, IdxT>(dataset.extent(0), 128); * // build KNN graph not using `cagra::build_knn_graph` * // build(knn_graph, dataset, ...); * // sort graph index * sort_knn_graph(res, dataset.view(), knn_graph.view()); * // optimize graph * cagra::optimize(res, dataset, knn_graph.view(), optimized_graph.view()); * // Construct an index from dataset and optimized knn_graph * auto index = cagra::index<T, IdxT>(res, build_params.metric(), dataset, * optimized_graph.view()); * @endcode * * @tparam DataT type of the data in the source dataset * @tparam IdxT type of the dataset vector indices * * @param[in] res raft resources * @param[in] dataset a matrix view (host or device) to a row-major matrix [n_rows, dim] * @param[in,out] knn_graph a matrix view (host or device) of the input knn graph [n_rows, * knn_graph_degree] */ template <typename DataT, typename IdxT = uint32_t, typename d_accessor = host_device_accessor<std::experimental::default_accessor<DataT>, memory_type::device>, typename g_accessor = host_device_accessor<std::experimental::default_accessor<IdxT>, memory_type::host>> void sort_knn_graph( raft::resources const& res, raft::mdspan<const DataT, raft::matrix_extent<int64_t>, raft::row_major, d_accessor> dataset, raft::mdspan<IdxT, raft::matrix_extent<int64_t>, raft::row_major, g_accessor> knn_graph) { using internal_IdxT = typename std::make_unsigned<IdxT>::type; using g_accessor_internal = host_device_accessor<std::experimental::default_accessor<internal_IdxT>, g_accessor::mem_type>; auto knn_graph_internal = raft::mdspan<internal_IdxT, raft::matrix_extent<int64_t>, raft::row_major, g_accessor_internal>( reinterpret_cast<internal_IdxT*>(knn_graph.data_handle()), knn_graph.extent(0), knn_graph.extent(1)); auto dataset_internal = raft::mdspan<const DataT, raft::matrix_extent<int64_t>, raft::row_major, d_accessor>( dataset.data_handle(), dataset.extent(0), dataset.extent(1)); cagra::detail::graph::sort_knn_graph(res, dataset_internal, knn_graph_internal); } /** * @brief Prune a KNN graph. * * Decrease the number of neighbors for each node. * * See [cagra::build_knn_graph](#cagra::build_knn_graph) for usage example * * @tparam IdxT type of the indices in the source dataset * * @param[in] res raft resources * @param[in] knn_graph a matrix view (host or device) of the input knn graph [n_rows, * knn_graph_degree] * @param[out] new_graph a host matrix view of the optimized knn graph [n_rows, graph_degree] */ template <typename IdxT = uint32_t, typename g_accessor = host_device_accessor<std::experimental::default_accessor<IdxT>, memory_type::host>> void optimize( raft::resources const& res, raft::mdspan<IdxT, raft::matrix_extent<int64_t>, raft::row_major, g_accessor> knn_graph, raft::host_matrix_view<IdxT, int64_t, raft::row_major> new_graph) { detail::optimize(res, knn_graph, new_graph); } /** * @brief Build the index from the dataset for efficient search. * * The build consist of two steps: build an intermediate knn-graph, and optimize it to * create the final graph. The index_params struct controls the node degree of these * graphs. * * It is required that dataset and the optimized graph fit the GPU memory. * * To customize the parameters for knn-graph building and pruning, and to reuse the * intermediate results, you could build the index in two steps using * [cagra::build_knn_graph](#cagra::build_knn_graph) and [cagra::optimize](#cagra::optimize). * * The following distance metrics are supported: * - L2 * * Usage example: * @code{.cpp} * using namespace cuvs::neighbors; * // use default index parameters * cagra::index_params index_params; * // create and fill the index from a [N, D] dataset * auto index = cagra::build(res, index_params, dataset); * // use default search parameters * cagra::search_params search_params; * // search K nearest neighbours * auto neighbors = raft::make_device_matrix<uint32_t>(res, n_queries, k); * auto distances = raft::make_device_matrix<float>(res, n_queries, k); * cagra::search(res, search_params, index, queries, neighbors, distances); * @endcode * * @tparam T data element type * @tparam IdxT type of the indices in the source dataset * * @param[in] res * @param[in] params parameters for building the index * @param[in] dataset a matrix view (host or device) to a row-major matrix [n_rows, dim] * * @return the constructed cagra index */ template <typename T, typename IdxT = uint32_t, typename Accessor = host_device_accessor<std::experimental::default_accessor<T>, memory_type::host>> index<T, IdxT> build( raft::resources const& res, const index_params& params, raft::mdspan<const T, raft::matrix_extent<int64_t>, raft::row_major, Accessor> dataset) { return detail::build<T, IdxT, Accessor>(res, params, dataset); } /** * @brief Search ANN using the constructed index. * * See the [cagra::build](#cagra::build) documentation for a usage example. * * @tparam T data element type * @tparam IdxT type of the indices * * @param[in] res raft resources * @param[in] params configure the search * @param[in] idx cagra index * @param[in] queries a device matrix view to a row-major matrix [n_queries, index->dim()] * @param[out] neighbors a device matrix view to the indices of the neighbors in the source dataset * [n_queries, k] * @param[out] distances a device matrix view to the distances to the selected neighbors [n_queries, * k] */ template <typename T, typename IdxT> void search(raft::resources const& res, const search_params& params, const index<T, IdxT>& idx, raft::device_matrix_view<const T, int64_t, raft::row_major> queries, raft::device_matrix_view<IdxT, int64_t, raft::row_major> neighbors, raft::device_matrix_view<float, int64_t, raft::row_major> distances) { RAFT_EXPECTS( queries.extent(0) == neighbors.extent(0) && queries.extent(0) == distances.extent(0), "Number of rows in output neighbors and distances matrices must equal the number of queries."); RAFT_EXPECTS(neighbors.extent(1) == distances.extent(1), "Number of columns in output neighbors and distances matrices must equal k"); RAFT_EXPECTS(queries.extent(1) == idx.dim(), "Number of query dimensions should equal number of dimensions in the index."); using internal_IdxT = typename std::make_unsigned<IdxT>::type; auto queries_internal = raft::make_device_matrix_view<const T, int64_t, raft::row_major>( queries.data_handle(), queries.extent(0), queries.extent(1)); auto neighbors_internal = raft::make_device_matrix_view<internal_IdxT, int64_t, raft::row_major>( reinterpret_cast<internal_IdxT*>(neighbors.data_handle()), neighbors.extent(0), neighbors.extent(1)); auto distances_internal = raft::make_device_matrix_view<float, int64_t, raft::row_major>( distances.data_handle(), distances.extent(0), distances.extent(1)); cagra::detail::search_main<T, internal_IdxT, decltype(cuvs::neighbors::filtering::none_cagra_sample_filter()), IdxT>(res, params, idx, queries_internal, neighbors_internal, distances_internal, cuvs::neighbors::filtering::none_cagra_sample_filter()); } /** * @brief Search ANN using the constructed index with the given sample filter. * * Usage example: * @code{.cpp} * using namespace cuvs::neighbors; * // use default index parameters * cagra::index_params index_params; * // create and fill the index from a [N, D] dataset * auto index = cagra::build(res, index_params, dataset); * // use default search parameters * cagra::search_params search_params; * // create a bitset to filter the search * auto removed_indices = raft::make_device_vector<IdxT>(res, n_removed_indices); * raft::core::bitset<std::uint32_t, IdxT> removed_indices_bitset( * res, removed_indices.view(), dataset.extent(0)); * // search K nearest neighbours according to a bitset * auto neighbors = raft::make_device_matrix<uint32_t>(res, n_queries, k); * auto distances = raft::make_device_matrix<float>(res, n_queries, k); * cagra::search_with_filtering(res, search_params, index, queries, neighbors, distances, * filtering::bitset_filter(removed_indices_bitset.view())); * @endcode * * @tparam T data element type * @tparam IdxT type of the indices * @tparam CagraSampleFilterT Device filter function, with the signature * `(uint32_t query ix, uint32_t sample_ix) -> bool` * * @param[in] res raft resources * @param[in] params configure the search * @param[in] idx cagra index * @param[in] queries a device matrix view to a row-major matrix [n_queries, index->dim()] * @param[out] neighbors a device matrix view to the indices of the neighbors in the source dataset * [n_queries, k] * @param[out] distances a device matrix view to the distances to the selected neighbors [n_queries, * k] * @param[in] sample_filter a device filter function that greenlights samples for a given query */ template <typename T, typename IdxT, typename CagraSampleFilterT> void search_with_filtering(raft::resources const& res, const search_params& params, const index<T, IdxT>& idx, raft::device_matrix_view<const T, int64_t, raft::row_major> queries, raft::device_matrix_view<IdxT, int64_t, raft::row_major> neighbors, raft::device_matrix_view<float, int64_t, raft::row_major> distances, CagraSampleFilterT sample_filter = CagraSampleFilterT()) { RAFT_EXPECTS( queries.extent(0) == neighbors.extent(0) && queries.extent(0) == distances.extent(0), "Number of rows in output neighbors and distances matrices must equal the number of queries."); RAFT_EXPECTS(neighbors.extent(1) == distances.extent(1), "Number of columns in output neighbors and distances matrices must equal k"); RAFT_EXPECTS(queries.extent(1) == idx.dim(), "Number of query dimensions should equal number of dimensions in the index."); using internal_IdxT = typename std::make_unsigned<IdxT>::type; auto queries_internal = raft::make_device_matrix_view<const T, int64_t, raft::row_major>( queries.data_handle(), queries.extent(0), queries.extent(1)); auto neighbors_internal = raft::make_device_matrix_view<internal_IdxT, int64_t, raft::row_major>( reinterpret_cast<internal_IdxT*>(neighbors.data_handle()), neighbors.extent(0), neighbors.extent(1)); auto distances_internal = raft::make_device_matrix_view<float, int64_t, raft::row_major>( distances.data_handle(), distances.extent(0), distances.extent(1)); cagra::detail::search_main<T, internal_IdxT, CagraSampleFilterT, IdxT>( res, params, idx, queries_internal, neighbors_internal, distances_internal, sample_filter); } /** @} */ // end group cagra } // namespace cuvs::neighbors::cagra // TODO: Remove deprecated experimental namespace in 23.12 release namespace cuvs::neighbors::experimental::cagra { using cuvs::neighbors::cagra::build; using cuvs::neighbors::cagra::build_knn_graph; using cuvs::neighbors::cagra::optimize; using cuvs::neighbors::cagra::search; using cuvs::neighbors::cagra::sort_knn_graph; } // namespace cuvs::neighbors::experimental::cagra
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/ivf_list.hpp
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/neighbors/ivf_list_types.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resource/thrust_policy.hpp> #include <raft/core/device_mdarray.hpp> #include <raft/core/error.hpp> #include <raft/core/host_mdarray.hpp> #include <raft/core/mdspan.hpp> #include <raft/core/resources.hpp> #include <raft/core/serialize.hpp> #include <raft/util/integer_utils.hpp> #include <thrust/fill.h> #include <fstream> #include <memory> #include <type_traits> namespace cuvs::neighbors::ivf { /** The data for a single IVF list. */ template <template <typename, typename...> typename SpecT, typename SizeT, typename... SpecExtraArgs> list<SpecT, SizeT, SpecExtraArgs...>::list(raft::resources const& res, const spec_type& spec, size_type n_rows) : size{n_rows}, data{res}, indices{res} { auto capacity = raft::round_up_safe<SizeT>(n_rows, spec.align_max); if (n_rows < spec.align_max) { capacity = bound_by_power_of_two<SizeT>(std::max<SizeT>(n_rows, spec.align_min)); capacity = std::min<SizeT>(capacity, spec.align_max); } try { data = raft::make_device_mdarray<value_type>(res, spec.make_list_extents(capacity)); indices = raft::make_device_vector<index_type, SizeT>(res, capacity); } catch (std::bad_alloc& e) { RAFT_FAIL( "ivf::list: failed to allocate a big enough list to hold all data " "(requested size: %zu records, selected capacity: %zu records). " "Allocator exception: %s", size_t(size), size_t(capacity), e.what()); } // Fill the index buffer with a pre-defined marker for easier debugging thrust::fill_n(raft::resource::get_thrust_policy(res), indices.data_handle(), indices.size(), ivf::kInvalidRecord<index_type>); } /** * Resize a list by the given id, so that it can contain the given number of records; * copy the data if necessary. */ template <typename ListT> void resize_list(raft::resources const& res, std::shared_ptr<ListT>& orig_list, // NOLINT const typename ListT::spec_type& spec, typename ListT::size_type new_used_size, typename ListT::size_type old_used_size) { bool skip_resize = false; if (orig_list) { if (new_used_size <= orig_list->indices.extent(0)) { auto shared_list_size = old_used_size; if (new_used_size <= old_used_size || orig_list->size.compare_exchange_strong(shared_list_size, new_used_size)) { // We don't need to resize the list if: // 1. The list exists // 2. The new size fits in the list // 3. The list doesn't grow or no-one else has grown it yet skip_resize = true; } } } else { old_used_size = 0; } if (skip_resize) { return; } auto new_list = std::make_shared<ListT>(res, spec, new_used_size); if (old_used_size > 0) { auto copied_data_extents = spec.make_list_extents(old_used_size); auto copied_view = make_mdspan<typename ListT::value_type, typename ListT::size_type, raft::row_major, false, true>(new_list->data.data_handle(), copied_data_extents); copy(copied_view.data_handle(), orig_list->data.data_handle(), copied_view.size(), resource::get_cuda_stream(res)); copy(new_list->indices.data_handle(), orig_list->indices.data_handle(), old_used_size, resource::get_cuda_stream(res)); } // swap the shared pointer content with the new list new_list.swap(orig_list); } template <typename ListT> auto serialize_list(const raft::resources& handle, std::ostream& os, const ListT& ld, const typename ListT::spec_type& store_spec, std::optional<typename ListT::size_type> size_override = std::nullopt) -> enable_if_valid_list_t<ListT> { using size_type = typename ListT::size_type; auto size = size_override.value_or(ld.size.load()); serialize_scalar(handle, os, size); if (size == 0) { return; } auto data_extents = store_spec.make_list_extents(size); auto data_array = raft::make_host_mdarray<typename ListT::value_type, size_type, raft::row_major>(data_extents); auto inds_array = raft::make_host_mdarray<typename ListT::index_type, size_type, raft::row_major>( make_extents<size_type>(size)); copy(data_array.data_handle(), ld.data.data_handle(), data_array.size(), resource::get_cuda_stream(handle)); copy(inds_array.data_handle(), ld.indices.data_handle(), inds_array.size(), resource::get_cuda_stream(handle)); resource::sync_stream(handle); serialize_mdspan(handle, os, data_array.view()); serialize_mdspan(handle, os, inds_array.view()); } template <typename ListT> auto serialize_list(const raft::resources& handle, std::ostream& os, const std::shared_ptr<ListT>& ld, const typename ListT::spec_type& store_spec, std::optional<typename ListT::size_type> size_override = std::nullopt) -> enable_if_valid_list_t<ListT> { if (ld) { return serialize_list<ListT>(handle, os, *ld, store_spec, size_override); } else { return serialize_scalar(handle, os, typename ListT::size_type{0}); } } template <typename ListT> auto deserialize_list(const raft::resources& handle, std::istream& is, std::shared_ptr<ListT>& ld, const typename ListT::spec_type& store_spec, const typename ListT::spec_type& device_spec) -> enable_if_valid_list_t<ListT> { using size_type = typename ListT::size_type; auto size = deserialize_scalar<size_type>(handle, is); if (size == 0) { return ld.reset(); } std::make_shared<ListT>(handle, device_spec, size).swap(ld); auto data_extents = store_spec.make_list_extents(size); auto data_array = raft::make_host_mdarray<typename ListT::value_type, size_type, raft::row_major>(data_extents); auto inds_array = raft::make_host_mdarray<typename ListT::index_type, size_type, raft::row_major>( make_extents<size_type>(size)); deserialize_mdspan(handle, is, data_array.view()); deserialize_mdspan(handle, is, inds_array.view()); copy(ld->data.data_handle(), data_array.data_handle(), data_array.size(), resource::get_cuda_stream(handle)); // NB: copying exactly 'size' indices to leave the rest 'kInvalidRecord' intact. copy( ld->indices.data_handle(), inds_array.data_handle(), size, resource::get_cuda_stream(handle)); // Make sure the data is copied from host to device before the host arrays get out of the scope. resource::sync_stream(handle); } } // namespace cuvs::neighbors::ivf
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/brute_force-ext.cuh
/* * Copyright (c) 2020-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <optional> #include <cuvs/distance/distance_types.hpp> // cuvs::distance::DistanceType #include <cuvs/neighbors/brute_force_types.hpp> #include <raft/core/device_mdspan.hpp> // raft::device_matrix_view #include <raft/core/operators.hpp> // raft::identity_op #include <raft/core/resources.hpp> // raft::resources #include <raft/util/raft_explicit.hpp> // RAFT_EXPLICIT #ifdef RAFT_EXPLICIT_INSTANTIATE_ONLY namespace cuvs::neighbors::brute_force { template <typename value_t, typename idx_t> inline void knn_merge_parts( raft::resources const& handle, raft::device_matrix_view<const value_t, idx_t, raft::row_major> in_keys, raft::device_matrix_view<const idx_t, idx_t, raft::row_major> in_values, raft::device_matrix_view<value_t, idx_t, raft::row_major> out_keys, raft::device_matrix_view<idx_t, idx_t, raft::row_major> out_values, size_t n_samples, std::optional<raft::device_vector_view<idx_t, idx_t>> translations = std::nullopt) RAFT_EXPLICIT; template <typename T, typename Accessor> index<T> build( raft::resources const& res, raft::mdspan<const T, raft::matrix_extent<int64_t>, raft::row_major, Accessor> dataset, cuvs::distance::DistanceType metric = distance::DistanceType::L2Unexpanded, T metric_arg = 0.0) RAFT_EXPLICIT; template <typename T, typename IdxT> void search(raft::resources const& res, const index<T>& idx, raft::device_matrix_view<const T, int64_t, raft::row_major> queries, raft::device_matrix_view<IdxT, int64_t, raft::row_major> neighbors, raft::device_matrix_view<T, int64_t, raft::row_major> distances) RAFT_EXPLICIT; template <typename idx_t, typename value_t, typename matrix_idx, typename index_layout, typename search_layout, typename epilogue_op = raft::identity_op> void knn(raft::resources const& handle, std::vector<raft::device_matrix_view<const value_t, matrix_idx, index_layout>> index, raft::device_matrix_view<const value_t, matrix_idx, search_layout> search, raft::device_matrix_view<idx_t, matrix_idx, raft::row_major> indices, raft::device_matrix_view<value_t, matrix_idx, raft::row_major> distances, distance::DistanceType metric = distance::DistanceType::L2Unexpanded, std::optional<float> metric_arg = std::make_optional<float>(2.0f), std::optional<idx_t> global_id_offset = std::nullopt, epilogue_op distance_epilogue = raft::identity_op()) RAFT_EXPLICIT; template <typename value_t, typename idx_t, typename idx_layout, typename query_layout> void fused_l2_knn(raft::resources const& handle, raft::device_matrix_view<const value_t, idx_t, idx_layout> index, raft::device_matrix_view<const value_t, idx_t, query_layout> query, raft::device_matrix_view<idx_t, idx_t, raft::row_major> out_inds, raft::device_matrix_view<value_t, idx_t, raft::row_major> out_dists, cuvs::distance::DistanceType metric) RAFT_EXPLICIT; } // namespace cuvs::neighbors::brute_force #endif // RAFT_EXPLICIT_INSTANTIATE_ONLY // No extern template for cuvs::neighbors::brute_force::knn_merge_parts #define instantiate_raft_neighbors_brute_force_knn( \ idx_t, value_t, matrix_idx, index_layout, search_layout, epilogue_op) \ extern template void cuvs::neighbors::brute_force:: \ knn<idx_t, value_t, matrix_idx, index_layout, search_layout, epilogue_op>( \ raft::resources const& handle, \ std::vector<raft::device_matrix_view<const value_t, matrix_idx, index_layout>> index, \ raft::device_matrix_view<const value_t, matrix_idx, search_layout> search, \ raft::device_matrix_view<idx_t, matrix_idx, raft::row_major> indices, \ raft::device_matrix_view<value_t, matrix_idx, raft::row_major> distances, \ cuvs::distance::DistanceType metric, \ std::optional<float> metric_arg, \ std::optional<idx_t> global_id_offset, \ epilogue_op distance_epilogue); instantiate_raft_neighbors_brute_force_knn( int64_t, float, uint32_t, raft::row_major, raft::row_major, raft::identity_op); instantiate_raft_neighbors_brute_force_knn( int64_t, float, int64_t, raft::row_major, raft::row_major, raft::identity_op); instantiate_raft_neighbors_brute_force_knn( int, float, int, raft::row_major, raft::row_major, raft::identity_op); instantiate_raft_neighbors_brute_force_knn( uint32_t, float, uint32_t, raft::row_major, raft::row_major, raft::identity_op); #undef instantiate_raft_neighbors_brute_force_knn namespace cuvs::neighbors::brute_force { extern template void search<float, int>( raft::resources const& res, const cuvs::neighbors::brute_force::index<float>& idx, raft::device_matrix_view<const float, int64_t, raft::row_major> queries, raft::device_matrix_view<int, int64_t, raft::row_major> neighbors, raft::device_matrix_view<float, int64_t, raft::row_major> distances); extern template void search<float, int64_t>( raft::resources const& res, const cuvs::neighbors::brute_force::index<float>& idx, raft::device_matrix_view<const float, int64_t, raft::row_major> queries, raft::device_matrix_view<int64_t, int64_t, raft::row_major> neighbors, raft::device_matrix_view<float, int64_t, raft::row_major> distances); extern template cuvs::neighbors::brute_force::index<float> build<float>( raft::resources const& res, raft::device_matrix_view<const float, int64_t, raft::row_major> dataset, cuvs::distance::DistanceType metric, float metric_arg); } // namespace cuvs::neighbors::brute_force #define instantiate_raft_neighbors_brute_force_fused_l2_knn( \ value_t, idx_t, idx_layout, query_layout) \ extern template void cuvs::neighbors::brute_force::fused_l2_knn( \ raft::resources const& handle, \ raft::device_matrix_view<const value_t, idx_t, idx_layout> index, \ raft::device_matrix_view<const value_t, idx_t, query_layout> query, \ raft::device_matrix_view<idx_t, idx_t, raft::row_major> out_inds, \ raft::device_matrix_view<value_t, idx_t, raft::row_major> out_dists, \ cuvs::distance::DistanceType metric); instantiate_raft_neighbors_brute_force_fused_l2_knn(float, int64_t, raft::row_major, raft::row_major) #undef instantiate_raft_neighbors_brute_force_fused_l2_knn
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/refine-ext.cuh
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cstdint> // int64_t #include <cuvs/distance/distance_types.hpp> // cuvs::distance::DistanceType #include <raft/core/device_mdspan.hpp> // raft::device_matrix_view #include <raft/core/host_mdspan.hpp> // // raft::host_matrix_view #include <raft/core/resources.hpp> // raft::resources #include <raft/util/raft_explicit.hpp> // RAFT_EXPLICIT #ifdef RAFT_EXPLICIT_INSTANTIATE_ONLY namespace cuvs::neighbors { template <typename idx_t, typename data_t, typename distance_t, typename matrix_idx> void refine(raft::resources const& handle, raft::device_matrix_view<const data_t, matrix_idx, raft::row_major> dataset, raft::device_matrix_view<const data_t, matrix_idx, raft::row_major> queries, raft::device_matrix_view<const idx_t, matrix_idx, raft::row_major> neighbor_candidates, raft::device_matrix_view<idx_t, matrix_idx, raft::row_major> indices, raft::device_matrix_view<distance_t, matrix_idx, raft::row_major> distances, cuvs::distance::DistanceType metric = distance::DistanceType::L2Unexpanded) RAFT_EXPLICIT; template <typename idx_t, typename data_t, typename distance_t, typename matrix_idx> void refine(raft::resources const& handle, raft::host_matrix_view<const data_t, matrix_idx, raft::row_major> dataset, raft::host_matrix_view<const data_t, matrix_idx, raft::row_major> queries, raft::host_matrix_view<const idx_t, matrix_idx, raft::row_major> neighbor_candidates, raft::host_matrix_view<idx_t, matrix_idx, raft::row_major> indices, raft::host_matrix_view<distance_t, matrix_idx, raft::row_major> distances, cuvs::distance::DistanceType metric = distance::DistanceType::L2Unexpanded) RAFT_EXPLICIT; } // namespace cuvs::neighbors #endif // RAFT_EXPLICIT_INSTANTIATE_ONLY #define instantiate_raft_neighbors_refine(idx_t, data_t, distance_t, matrix_idx) \ extern template void cuvs::neighbors::refine<idx_t, data_t, distance_t, matrix_idx>( \ raft::resources const& handle, \ raft::device_matrix_view<const data_t, matrix_idx, raft::row_major> dataset, \ raft::device_matrix_view<const data_t, matrix_idx, raft::row_major> queries, \ raft::device_matrix_view<const idx_t, matrix_idx, raft::row_major> neighbor_candidates, \ raft::device_matrix_view<idx_t, matrix_idx, raft::row_major> indices, \ raft::device_matrix_view<distance_t, matrix_idx, raft::row_major> distances, \ cuvs::distance::DistanceType metric); \ \ extern template void cuvs::neighbors::refine<idx_t, data_t, distance_t, matrix_idx>( \ raft::resources const& handle, \ raft::host_matrix_view<const data_t, matrix_idx, raft::row_major> dataset, \ raft::host_matrix_view<const data_t, matrix_idx, raft::row_major> queries, \ raft::host_matrix_view<const idx_t, matrix_idx, raft::row_major> neighbor_candidates, \ raft::host_matrix_view<idx_t, matrix_idx, raft::row_major> indices, \ raft::host_matrix_view<distance_t, matrix_idx, raft::row_major> distances, \ cuvs::distance::DistanceType metric); instantiate_raft_neighbors_refine(int64_t, float, float, int64_t); instantiate_raft_neighbors_refine(int64_t, int8_t, float, int64_t); instantiate_raft_neighbors_refine(int64_t, uint8_t, float, int64_t); #undef instantiate_raft_neighbors_refine
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/ivf_flat_serialize.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "detail/ivf_flat_serialize.cuh" namespace cuvs::neighbors::ivf_flat { /** * \defgroup ivf_flat_serialize IVF-Flat Serialize * @{ */ /** * Write the index to an output stream * * Experimental, both the API and the serialization format are subject to change. * * @code{.cpp} * #include <raft/core/resources.hpp> * * raft::resources handle; * * // create an output stream * std::ostream os(std::cout.rdbuf()); * // create an index with `auto index = ivf_flat::build(...);` * raft::serialize(handle, os, index); * @endcode * * @tparam T data element type * @tparam IdxT type of the indices * * @param[in] handle the raft handle * @param[in] os output stream * @param[in] index IVF-Flat index * */ template <typename T, typename IdxT> void serialize(raft::resources const& handle, std::ostream& os, const index<T, IdxT>& index) { detail::serialize(handle, os, index); } /** * Save the index to file. * * Experimental, both the API and the serialization format are subject to change. * * @code{.cpp} * #include <raft/core/resources.hpp> * * raft::resources handle; * * // create a string with a filepath * std::string filename("/path/to/index"); * // create an index with `auto index = ivf_flat::build(...);` * raft::serialize(handle, filename, index); * @endcode * * @tparam T data element type * @tparam IdxT type of the indices * * @param[in] handle the raft handle * @param[in] filename the file name for saving the index * @param[in] index IVF-Flat index * */ template <typename T, typename IdxT> void serialize(raft::resources const& handle, const std::string& filename, const index<T, IdxT>& index) { detail::serialize(handle, filename, index); } /** * Load index from input stream * * Experimental, both the API and the serialization format are subject to change. * * @code{.cpp} * #include <raft/core/resources.hpp> * * raft::resources handle; * * // create an input stream * std::istream is(std::cin.rdbuf()); * using T = float; // data element type * using IdxT = int; // type of the index * auto index = raft::deserialize<T, IdxT>(handle, is); * @endcode * * @tparam T data element type * @tparam IdxT type of the indices * * @param[in] handle the raft handle * @param[in] is input stream * * @return cuvs::neighbors::ivf_flat::index<T, IdxT> */ template <typename T, typename IdxT> index<T, IdxT> deserialize(raft::resources const& handle, std::istream& is) { return detail::deserialize<T, IdxT>(handle, is); } /** * Load index from file. * * Experimental, both the API and the serialization format are subject to change. * * @code{.cpp} * #include <raft/core/resources.hpp> * * raft::resources handle; * * // create a string with a filepath * std::string filename("/path/to/index"); * using T = float; // data element type * using IdxT = int; // type of the index * auto index = raft::deserialize<T, IdxT>(handle, filename); * @endcode * * @tparam T data element type * @tparam IdxT type of the indices * * @param[in] handle the raft handle * @param[in] filename the name of the file that stores the index * * @return cuvs::neighbors::ivf_flat::index<T, IdxT> */ template <typename T, typename IdxT> index<T, IdxT> deserialize(raft::resources const& handle, const std::string& filename) { return detail::deserialize<T, IdxT>(handle, filename); } /**@}*/ } // namespace cuvs::neighbors::ivf_flat
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/ivf_flat.cuh
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #ifndef RAFT_EXPLICIT_INSTANTIATE_ONLY #include "ivf_flat-inl.cuh" #endif #ifdef RAFT_COMPILED #include "ivf_flat-ext.cuh" #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/ivf_pq_serialize.cuh
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "detail/ivf_pq_serialize.cuh" namespace cuvs::neighbors::ivf_pq { /** * \defgroup ivf_pq_serialize IVF-PQ Serialize * @{ */ /** * Write the index to an output stream * * Experimental, both the API and the serialization format are subject to change. * * @code{.cpp} * #include <raft/core/resources.hpp> * * raft::resources handle; * * // create an output stream * std::ostream os(std::cout.rdbuf()); * // create an index with `auto index = ivf_pq::build(...);` * raft::serialize(handle, os, index); * @endcode * * @tparam IdxT type of the index * * @param[in] handle the raft handle * @param[in] os output stream * @param[in] index IVF-PQ index * */ template <typename IdxT> void serialize(raft::resources const& handle, std::ostream& os, const index<IdxT>& index) { detail::serialize(handle, os, index); } /** * Save the index to file. * * Experimental, both the API and the serialization format are subject to change. * * @code{.cpp} * #include <raft/core/resources.hpp> * * raft::resources handle; * * // create a string with a filepath * std::string filename("/path/to/index"); * // create an index with `auto index = ivf_pq::build(...);` * raft::serialize(handle, filename, index); * @endcode * * @tparam IdxT type of the index * * @param[in] handle the raft handle * @param[in] filename the file name for saving the index * @param[in] index IVF-PQ index * */ template <typename IdxT> void serialize(raft::resources const& handle, const std::string& filename, const index<IdxT>& index) { detail::serialize(handle, filename, index); } /** * Load index from input stream * * Experimental, both the API and the serialization format are subject to change. * * @code{.cpp} * #include <raft/core/resources.hpp> * * raft::resources handle; * * // create an input stream * std::istream is(std::cin.rdbuf()); * using IdxT = int; // type of the index * auto index = raft::deserialize<IdxT>(handle, is); * @endcode * * @tparam IdxT type of the index * * @param[in] handle the raft handle * @param[in] is input stream * * @return cuvs::neighbors::ivf_pq::index<IdxT> */ template <typename IdxT> index<IdxT> deserialize(raft::resources const& handle, std::istream& is) { return detail::deserialize<IdxT>(handle, is); } /** * Load index from file. * * Experimental, both the API and the serialization format are subject to change. * * @code{.cpp} * #include <raft/core/resources.hpp> * * raft::resources handle; * * // create a string with a filepath * std::string filename("/path/to/index"); * using IdxT = int; // type of the index * auto index = raft::deserialize<IdxT>(handle, filename); * @endcode * * @tparam IdxT type of the index * * @param[in] handle the raft handle * @param[in] filename the name of the file that stores the index * * @return cuvs::neighbors::ivf_pq::index<IdxT> */ template <typename IdxT> index<IdxT> deserialize(raft::resources const& handle, const std::string& filename) { return detail::deserialize<IdxT>(handle, filename); } /**@}*/ } // namespace cuvs::neighbors::ivf_pq
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/epsilon_neighborhood.cuh
/* * Copyright (c) 2020-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #ifndef __EPSILON_NEIGH_H #define __EPSILON_NEIGH_H #pragma once #include <cuvs/spatial/knn/detail/epsilon_neighborhood.cuh> #include <raft/core/device_mdspan.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resources.hpp> namespace cuvs::neighbors::epsilon_neighborhood { /** * @brief Computes epsilon neighborhood for the L2-Squared distance metric * * @tparam value_t IO and math type * @tparam idx_t Index type * * @param[out] adj adjacency matrix [row-major] [on device] [dim = m x n] * @param[out] vd vertex degree array [on device] [len = m + 1] * `vd + m` stores the total number of edges in the adjacency * matrix. Pass a nullptr if you don't need this info. * @param[in] x first matrix [row-major] [on device] [dim = m x k] * @param[in] y second matrix [row-major] [on device] [dim = n x k] * @param[in] m number of rows in x * @param[in] n number of rows in y * @param[in] k number of columns in x and k * @param[in] eps defines epsilon neighborhood radius (should be passed as * squared as we compute L2-squared distance in this method) * @param[in] stream cuda stream */ template <typename value_t, typename idx_t> void epsUnexpL2SqNeighborhood(bool* adj, idx_t* vd, const value_t* x, const value_t* y, idx_t m, idx_t n, idx_t k, value_t eps, cudaStream_t stream) { spatial::knn::detail::epsUnexpL2SqNeighborhood<value_t, idx_t>( adj, vd, x, y, m, n, k, eps, stream); } /** * @defgroup epsilon_neighbors Epislon Neighborhood Operations * @{ */ /** * @brief Computes epsilon neighborhood for the L2-Squared distance metric and given ball size. * The epsilon neighbors is represented by a dense boolean adjacency matrix of size m * n and * an array of degrees for each vertex, which can be used as a compressed sparse row (CSR) * indptr array. * * @code{.cpp} * #include <cuvs/neighbors/epsilon_neighborhood.cuh> * #include <raft/core/resources.hpp> * #include <raft/core/device_mdarray.hpp> * using namespace cuvs::neighbors; * raft::raft::resources handle; * ... * auto adj = raft::make_device_matrix<bool>(handle, m * n); * auto vd = raft::make_device_vector<int>(handle, m+1); * epsilon_neighborhood::eps_neighbors_l2sq(handle, x, y, adj.view(), vd.view(), eps); * @endcode * * @tparam value_t IO and math type * @tparam idx_t Index type * @tparam matrix_idx_t matrix indexing type * * @param[in] handle raft handle to manage library resources * @param[in] x first matrix [row-major] [on device] [dim = m x k] * @param[in] y second matrix [row-major] [on device] [dim = n x k] * @param[out] adj adjacency matrix [row-major] [on device] [dim = m x n] * @param[out] vd vertex degree array [on device] [len = m + 1] * `vd + m` stores the total number of edges in the adjacency * matrix. Pass a nullptr if you don't need this info. * @param[in] eps defines epsilon neighborhood radius (should be passed as * squared as we compute L2-squared distance in this method) */ template <typename value_t, typename idx_t, typename matrix_idx_t> void eps_neighbors_l2sq(raft::resources const& handle, raft::device_matrix_view<const value_t, matrix_idx_t, raft::row_major> x, raft::device_matrix_view<const value_t, matrix_idx_t, raft::row_major> y, raft::device_matrix_view<bool, matrix_idx_t, raft::row_major> adj, raft::device_vector_view<idx_t, matrix_idx_t> vd, value_t eps) { epsUnexpL2SqNeighborhood<value_t, idx_t>(adj.data_handle(), vd.data_handle(), x.data_handle(), y.data_handle(), x.extent(0), y.extent(0), x.extent(1), eps, resource::get_cuda_stream(handle)); } /** @} */ // end group epsilon_neighbors } // namespace cuvs::neighbors::epsilon_neighborhood #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/ivf_pq-ext.cuh
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cstdint> // int64_t #include <cuvs/neighbors/ivf_pq_types.hpp> // cuvs::neighbors::ivf_pq::index #include <raft/core/device_mdspan.hpp> // raft::device_matrix_view #include <raft/core/resources.hpp> // raft::resources #include <raft/util/raft_explicit.hpp> // RAFT_EXPLICIT #include <rmm/mr/device/per_device_resource.hpp> // rmm::mr::device_memory_resource #ifdef RAFT_EXPLICIT_INSTANTIATE_ONLY namespace cuvs::neighbors::ivf_pq { template <typename T, typename IdxT = uint32_t> index<IdxT> build(raft::resources const& handle, const index_params& params, raft::device_matrix_view<const T, IdxT, raft::row_major> dataset) RAFT_EXPLICIT; template <typename T, typename IdxT> index<IdxT> extend( raft::resources const& handle, raft::device_matrix_view<const T, IdxT, raft::row_major> new_vectors, std::optional<raft::device_vector_view<const IdxT, IdxT, raft::row_major>> new_indices, const index<IdxT>& idx) RAFT_EXPLICIT; template <typename T, typename IdxT> void extend(raft::resources const& handle, raft::device_matrix_view<const T, IdxT, raft::row_major> new_vectors, std::optional<raft::device_vector_view<const IdxT, IdxT, raft::row_major>> new_indices, index<IdxT>* idx) RAFT_EXPLICIT; template <typename T, typename IdxT, typename IvfSampleFilterT> void search_with_filtering(raft::resources const& handle, const search_params& params, const index<IdxT>& idx, raft::device_matrix_view<const T, uint32_t, raft::row_major> queries, raft::device_matrix_view<IdxT, uint32_t, raft::row_major> neighbors, raft::device_matrix_view<float, uint32_t, raft::row_major> distances, IvfSampleFilterT sample_filter) RAFT_EXPLICIT; template <typename T, typename IdxT> void search(raft::resources const& handle, const search_params& params, const index<IdxT>& idx, raft::device_matrix_view<const T, uint32_t, raft::row_major> queries, raft::device_matrix_view<IdxT, uint32_t, raft::row_major> neighbors, raft::device_matrix_view<float, uint32_t, raft::row_major> distances) RAFT_EXPLICIT; template <typename T, typename IdxT = uint32_t> auto build(raft::resources const& handle, const index_params& params, const T* dataset, IdxT n_rows, uint32_t dim) -> index<IdxT> RAFT_EXPLICIT; template <typename T, typename IdxT> auto extend(raft::resources const& handle, const index<IdxT>& idx, const T* new_vectors, const IdxT* new_indices, IdxT n_rows) -> index<IdxT> RAFT_EXPLICIT; template <typename T, typename IdxT> void extend(raft::resources const& handle, index<IdxT>* idx, const T* new_vectors, const IdxT* new_indices, IdxT n_rows) RAFT_EXPLICIT; template <typename T, typename IdxT, typename IvfSampleFilterT> void search_with_filtering(raft::resources const& handle, const cuvs::neighbors::ivf_pq::search_params& params, const index<IdxT>& idx, const T* queries, uint32_t n_queries, uint32_t k, IdxT* neighbors, float* distances, IvfSampleFilterT sample_filter = IvfSampleFilterT{}) RAFT_EXPLICIT; template <typename T, typename IdxT> void search(raft::resources const& handle, const cuvs::neighbors::ivf_pq::search_params& params, const index<IdxT>& idx, const T* queries, uint32_t n_queries, uint32_t k, IdxT* neighbors, float* distances) RAFT_EXPLICIT; template <typename T, typename IdxT, typename IvfSampleFilterT> [[deprecated( "Drop the `mr` argument and use `raft::resource::set_workspace_resource` instead")]] void search_with_filtering(raft::resources const& handle, const cuvs::neighbors::ivf_pq::search_params& params, const index<IdxT>& idx, const T* queries, uint32_t n_queries, uint32_t k, IdxT* neighbors, float* distances, rmm::mr::device_memory_resource* mr, IvfSampleFilterT sample_filter = IvfSampleFilterT{}) RAFT_EXPLICIT; template <typename T, typename IdxT> [[deprecated( "Drop the `mr` argument and use `raft::resource::set_workspace_resource` instead")]] void search(raft::resources const& handle, const cuvs::neighbors::ivf_pq::search_params& params, const index<IdxT>& idx, const T* queries, uint32_t n_queries, uint32_t k, IdxT* neighbors, float* distances, rmm::mr::device_memory_resource* mr) RAFT_EXPLICIT; } // namespace cuvs::neighbors::ivf_pq #endif // RAFT_EXPLICIT_INSTANTIATE_ONLY #define instantiate_raft_neighbors_ivf_pq_build(T, IdxT) \ extern template cuvs::neighbors::ivf_pq::index<IdxT> cuvs::neighbors::ivf_pq::build<T, IdxT>( \ raft::resources const& handle, \ const cuvs::neighbors::ivf_pq::index_params& params, \ raft::device_matrix_view<const T, IdxT, raft::row_major> dataset); \ \ extern template auto cuvs::neighbors::ivf_pq::build( \ raft::resources const& handle, \ const cuvs::neighbors::ivf_pq::index_params& params, \ const T* dataset, \ IdxT n_rows, \ uint32_t dim) \ ->cuvs::neighbors::ivf_pq::index<IdxT>; instantiate_raft_neighbors_ivf_pq_build(float, int64_t); instantiate_raft_neighbors_ivf_pq_build(int8_t, int64_t); instantiate_raft_neighbors_ivf_pq_build(uint8_t, int64_t); #undef instantiate_raft_neighbors_ivf_pq_build #define instantiate_raft_neighbors_ivf_pq_extend(T, IdxT) \ extern template cuvs::neighbors::ivf_pq::index<IdxT> cuvs::neighbors::ivf_pq::extend<T, IdxT>( \ raft::resources const& handle, \ raft::device_matrix_view<const T, IdxT, raft::row_major> new_vectors, \ std::optional<raft::device_vector_view<const IdxT, IdxT, raft::row_major>> new_indices, \ const cuvs::neighbors::ivf_pq::index<IdxT>& idx); \ \ extern template void cuvs::neighbors::ivf_pq::extend<T, IdxT>( \ raft::resources const& handle, \ raft::device_matrix_view<const T, IdxT, raft::row_major> new_vectors, \ std::optional<raft::device_vector_view<const IdxT, IdxT, raft::row_major>> new_indices, \ cuvs::neighbors::ivf_pq::index<IdxT>* idx); \ \ extern template auto cuvs::neighbors::ivf_pq::extend<T, IdxT>( \ raft::resources const& handle, \ const cuvs::neighbors::ivf_pq::index<IdxT>& idx, \ const T* new_vectors, \ const IdxT* new_indices, \ IdxT n_rows) \ ->cuvs::neighbors::ivf_pq::index<IdxT>; \ \ extern template void cuvs::neighbors::ivf_pq::extend<T, IdxT>( \ raft::resources const& handle, \ cuvs::neighbors::ivf_pq::index<IdxT>* idx, \ const T* new_vectors, \ const IdxT* new_indices, \ IdxT n_rows); instantiate_raft_neighbors_ivf_pq_extend(float, int64_t); instantiate_raft_neighbors_ivf_pq_extend(int8_t, int64_t); instantiate_raft_neighbors_ivf_pq_extend(uint8_t, int64_t); #undef instantiate_raft_neighbors_ivf_pq_extend #define instantiate_raft_neighbors_ivf_pq_search(T, IdxT) \ extern template void cuvs::neighbors::ivf_pq::search<T, IdxT>( \ raft::resources const& handle, \ const cuvs::neighbors::ivf_pq::search_params& params, \ const cuvs::neighbors::ivf_pq::index<IdxT>& idx, \ raft::device_matrix_view<const T, uint32_t, raft::row_major> queries, \ raft::device_matrix_view<IdxT, uint32_t, raft::row_major> neighbors, \ raft::device_matrix_view<float, uint32_t, raft::row_major> distances); \ \ extern template void cuvs::neighbors::ivf_pq::search<T, IdxT>( \ raft::resources const& handle, \ const cuvs::neighbors::ivf_pq::search_params& params, \ const cuvs::neighbors::ivf_pq::index<IdxT>& idx, \ const T* queries, \ uint32_t n_queries, \ uint32_t k, \ IdxT* neighbors, \ float* distances, \ rmm::mr::device_memory_resource* mr); \ \ extern template void cuvs::neighbors::ivf_pq::search<T, IdxT>( \ raft::resources const& handle, \ const cuvs::neighbors::ivf_pq::search_params& params, \ const cuvs::neighbors::ivf_pq::index<IdxT>& idx, \ const T* queries, \ uint32_t n_queries, \ uint32_t k, \ IdxT* neighbors, \ float* distances) instantiate_raft_neighbors_ivf_pq_search(float, int64_t); instantiate_raft_neighbors_ivf_pq_search(int8_t, int64_t); instantiate_raft_neighbors_ivf_pq_search(uint8_t, int64_t); #undef instantiate_raft_neighbors_ivf_pq_search
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/ann_types.hpp
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include <cuvs/distance/distance_types.hpp> namespace cuvs::neighbors::ann { /** * @defgroup ann_types Approximate Nearest Neighbors Types * @{ */ /** The base for approximate KNN index structures. */ struct index {}; /** The base for KNN index parameters. */ struct index_params { /** Distance type. */ cuvs::distance::DistanceType metric = distance::DistanceType::L2Expanded; /** The argument used by some distance metrics. */ float metric_arg = 2.0f; /** * Whether to add the dataset content to the index, i.e.: * * - `true` means the index is filled with the dataset vectors and ready to search after calling * `build`. * - `false` means `build` only trains the underlying model (e.g. quantizer or clustering), but * the index is left empty; you'd need to call `extend` on the index afterwards to populate it. */ bool add_data_on_build = true; }; struct search_params {}; /** @} */ // end group ann_types }; // namespace cuvs::neighbors::ann
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/ivf_pq.cuh
/* * Copyright (c) 2022-2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #ifndef RAFT_EXPLICIT_INSTANTIATE_ONLY #include "ivf_pq-inl.cuh" #endif #ifdef RAFT_COMPILED #include "ivf_pq-ext.cuh" #endif
0
rapidsai_public_repos/cuvs/cpp/include/cuvs
rapidsai_public_repos/cuvs/cpp/include/cuvs/neighbors/cagra_types.hpp
/* * Copyright (c) 2023, NVIDIA CORPORATION. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #pragma once #include "ann_types.hpp" #include <raft/core/resource/cuda_stream.hpp> #include <cuvs/distance/distance_types.hpp> #include <cuvs/neighbors/detail/cagra/utils.hpp> #include <raft/core/device_mdarray.hpp> #include <raft/core/error.hpp> #include <raft/core/host_mdarray.hpp> #include <raft/core/mdspan_types.hpp> #include <raft/core/resources.hpp> #include <raft/util/integer_utils.hpp> #include <memory> #include <optional> #include <string> #include <thrust/fill.h> #include <type_traits> #include <raft/core/logger.hpp> namespace cuvs::neighbors::cagra { /** * @addtogroup cagra * @{ */ /** * @brief ANN algorithm used by CAGRA to build knn graph * */ enum class graph_build_algo { /* Use IVF-PQ to build all-neighbors knn graph */ IVF_PQ, /* Experimental, use NN-Descent to build all-neighbors knn graph */ NN_DESCENT }; struct index_params : ann::index_params { /** Degree of input graph for pruning. */ size_t intermediate_graph_degree = 128; /** Degree of output graph. */ size_t graph_degree = 64; /** ANN algorithm to build knn graph. */ graph_build_algo build_algo = graph_build_algo::IVF_PQ; /** Number of Iterations to run if building with NN_DESCENT */ size_t nn_descent_niter = 20; }; enum class search_algo { /** For large batch sizes. */ SINGLE_CTA, /** For small batch sizes. */ MULTI_CTA, MULTI_KERNEL, AUTO }; enum class hash_mode { HASH, SMALL, AUTO }; struct search_params : ann::search_params { /** Maximum number of queries to search at the same time (batch size). Auto select when 0.*/ size_t max_queries = 0; /** Number of intermediate search results retained during the search. * * This is the main knob to adjust trade off between accuracy and search speed. * Higher values improve the search accuracy. */ size_t itopk_size = 64; /** Upper limit of search iterations. Auto select when 0.*/ size_t max_iterations = 0; // In the following we list additional search parameters for fine tuning. // Reasonable default values are automatically chosen. /** Which search implementation to use. */ search_algo algo = search_algo::AUTO; /** Number of threads used to calculate a single distance. 4, 8, 16, or 32. */ size_t team_size = 0; /** Number of graph nodes to select as the starting point for the search in each iteration. aka * search width?*/ size_t search_width = 1; /** Lower limit of search iterations. */ size_t min_iterations = 0; /** Thread block size. 0, 64, 128, 256, 512, 1024. Auto selection when 0. */ size_t thread_block_size = 0; /** Hashmap type. Auto selection when AUTO. */ hash_mode hashmap_mode = hash_mode::AUTO; /** Lower limit of hashmap bit length. More than 8. */ size_t hashmap_min_bitlen = 0; /** Upper limit of hashmap fill rate. More than 0.1, less than 0.9.*/ float hashmap_max_fill_rate = 0.5; /** Number of iterations of initial random seed node selection. 1 or more. */ uint32_t num_random_samplings = 1; /** Bit mask used for initial random seed node selection. */ uint64_t rand_xor_mask = 0x128394; }; static_assert(std::is_aggregate_v<index_params>); static_assert(std::is_aggregate_v<search_params>); /** * @brief CAGRA index. * * The index stores the dataset and a kNN graph in device memory. * * @tparam T data element type * @tparam IdxT type of the vector indices (represent dataset.extent(0)) * */ template <typename T, typename IdxT> struct index : ann::index { static_assert(!raft::is_narrowing_v<uint32_t, IdxT>, "IdxT must be able to represent all values of uint32_t"); public: /** Distance metric used for clustering. */ [[nodiscard]] constexpr inline auto metric() const noexcept -> cuvs::distance::DistanceType { return metric_; } /** Total length of the index (number of vectors). */ [[nodiscard]] constexpr inline auto size() const noexcept -> IdxT { return dataset_view_.extent(0); } /** Dimensionality of the data. */ [[nodiscard]] constexpr inline auto dim() const noexcept -> uint32_t { return dataset_view_.extent(1); } /** Graph degree */ [[nodiscard]] constexpr inline auto graph_degree() const noexcept -> uint32_t { return graph_view_.extent(1); } /** Dataset [size, dim] */ [[nodiscard]] inline auto dataset() const noexcept -> raft::device_matrix_view<const T, int64_t, raft::layout_stride> { return dataset_view_; } /** neighborhood graph [size, graph-degree] */ [[nodiscard]] inline auto graph() const noexcept -> raft::device_matrix_view<const IdxT, int64_t, raft::row_major> { return graph_view_; } // Don't allow copying the index for performance reasons (try avoiding copying data) index(const index&) = delete; index(index&&) = default; auto operator=(const index&) -> index& = delete; auto operator=(index&&) -> index& = default; ~index() = default; /** Construct an empty index. */ index(raft::resources const& res, cuvs::distance::DistanceType metric = cuvs::distance::DistanceType::L2Expanded) : ann::index(), metric_(metric), dataset_(raft::make_device_matrix<T, int64_t>(res, 0, 0)), graph_(raft::make_device_matrix<IdxT, int64_t>(res, 0, 0)) { } /** Construct an index from dataset and knn_graph arrays * * If the dataset and graph is already in GPU memory, then the index is just a thin wrapper around * these that stores a non-owning a reference to the arrays. * * The constructor also accepts host arrays. In that case they are copied to the device, and the * device arrays will be owned by the index. * * In case the dasates rows are not 16 bytes aligned, then we create a padded copy in device * memory to ensure alignment for vectorized load. * * Usage examples: * * - Cagra index is normally created by the cagra::build * @code{.cpp} * using namespace cuvs::neighbors::experimental; * auto dataset = raft::make_host_matrix<float, int64_t>(n_rows, n_cols); * load_dataset(dataset.view()); * // use default index parameters * cagra::index_params index_params; * // create and fill the index from a [N, D] dataset * auto index = cagra::build(res, index_params, dataset); * // use default search parameters * cagra::search_params search_params; * // search K nearest neighbours * auto neighbors = raft::make_device_matrix<uint32_t, int64_t>(res, n_queries, k); * auto distances = raft::make_device_matrix<float, int64_t>(res, n_queries, k); * cagra::search(res, search_params, index, queries, neighbors, distances); * @endcode * In the above example, we have passed a host dataset to build. The returned index will own a * device copy of the dataset and the knn_graph. In contrast, if we pass the dataset as a * raft::device_mdspan to build, then it will only store a reference to it. * * - Constructing index using existing knn-graph * @code{.cpp} * using namespace cuvs::neighbors::experimental; * * auto dataset = raft::make_device_matrix<float, int64_t>(res, n_rows, n_cols); * auto knn_graph = raft::make_device_matrix<uint32_n, int64_t>(res, n_rows, graph_degree); * * // custom loading and graph creation * // load_dataset(dataset.view()); * // create_knn_graph(knn_graph.view()); * * // Wrap the existing device arrays into an index structure * cagra::index<T, IdxT> index(res, metric, raft::make_const_mdspan(dataset.view()), * raft::make_const_mdspan(knn_graph.view())); * * // Both knn_graph and dataset objects have to be in scope while the index is used because * // the index only stores a reference to these. * cagra::search(res, search_params, index, queries, neighbors, distances); * @endcode * */ template <typename data_accessor, typename graph_accessor> index(raft::resources const& res, cuvs::distance::DistanceType metric, raft::mdspan<const T, raft::matrix_extent<int64_t>, raft::row_major, data_accessor> dataset, raft::mdspan<const IdxT, raft::matrix_extent<int64_t>, raft::row_major, graph_accessor> knn_graph) : ann::index(), metric_(metric), dataset_(raft::make_device_matrix<T, int64_t>(res, 0, 0)), graph_(raft::make_device_matrix<IdxT, int64_t>(res, 0, 0)) { RAFT_EXPECTS(dataset.extent(0) == knn_graph.extent(0), "Dataset and knn_graph must have equal number of rows"); update_dataset(res, dataset); update_graph(res, knn_graph); raft::resource::sync_stream(res); } /** * Replace the dataset with a new dataset. * * If the new dataset rows are aligned on 16 bytes, then only a reference is stored to the * dataset. It is the caller's responsibility to ensure that dataset stays alive as long as the * index. */ void update_dataset(raft::resources const& res, raft::device_matrix_view<const T, int64_t, raft::row_major> dataset) { if (dataset.extent(1) * sizeof(T) % 16 != 0) { RAFT_LOG_DEBUG("Creating a padded copy of CAGRA dataset in device memory"); copy_padded(res, dataset); } else { dataset_view_ = raft::make_device_strided_matrix_view<const T, int64_t>( dataset.data_handle(), dataset.extent(0), dataset.extent(1), dataset.extent(1)); } } /** * Replace the dataset with a new dataset. * * We create a copy of the dataset on the device. The index manages the lifetime of this copy. */ void update_dataset(raft::resources const& res, raft::host_matrix_view<const T, int64_t, raft::row_major> dataset) { RAFT_LOG_DEBUG("Copying CAGRA dataset from host to device"); copy_padded(res, dataset); } /** * Replace the graph with a new graph. * * Since the new graph is a device array, we store a reference to that, and it is * the caller's responsibility to ensure that knn_graph stays alive as long as the index. */ void update_graph(raft::resources const& res, raft::device_matrix_view<const IdxT, int64_t, raft::row_major> knn_graph) { graph_view_ = knn_graph; } /** * Replace the graph with a new graph. * * We create a copy of the graph on the device. The index manages the lifetime of this copy. */ void update_graph(raft::resources const& res, raft::host_matrix_view<const IdxT, int64_t, raft::row_major> knn_graph) { RAFT_LOG_DEBUG("Copying CAGRA knn graph from host to device"); if ((graph_.extent(0) != knn_graph.extent(0)) || (graph_.extent(1) != knn_graph.extent(1))) { // clear existing memory before allocating to prevent OOM errors on large graphs if (graph_.size()) { graph_ = raft::make_device_matrix<IdxT, int64_t>(res, 0, 0); } graph_ = raft::make_device_matrix<IdxT, int64_t>(res, knn_graph.extent(0), knn_graph.extent(1)); } raft::copy(graph_.data_handle(), knn_graph.data_handle(), knn_graph.size(), raft::resource::get_cuda_stream(res)); graph_view_ = graph_.view(); } private: /** Create a device copy of the dataset, and pad it if necessary. */ template <typename data_accessor> void copy_padded( raft::resources const& res, raft::mdspan<const T, raft::matrix_extent<int64_t>, raft::row_major, data_accessor> dataset) { detail::copy_with_padding(res, dataset_, dataset); dataset_view_ = raft::make_device_strided_matrix_view<const T, int64_t>( dataset_.data_handle(), dataset_.extent(0), dataset.extent(1), dataset_.extent(1)); RAFT_LOG_DEBUG("CAGRA dataset strided matrix view %zux%zu, stride %zu", static_cast<size_t>(dataset_view_.extent(0)), static_cast<size_t>(dataset_view_.extent(1)), static_cast<size_t>(dataset_view_.stride(0))); } cuvs::distance::DistanceType metric_; raft::device_matrix<T, int64_t, raft::row_major> dataset_; raft::device_matrix<IdxT, int64_t, raft::row_major> graph_; raft::device_matrix_view<const T, int64_t, raft::layout_stride> dataset_view_; raft::device_matrix_view<const IdxT, int64_t, raft::row_major> graph_view_; }; /** @} */ } // namespace cuvs::neighbors::cagra // TODO: Remove deprecated experimental namespace in 23.12 release namespace cuvs::neighbors::experimental::cagra { using cuvs::neighbors::cagra::graph_build_algo; using cuvs::neighbors::cagra::hash_mode; using cuvs::neighbors::cagra::index; using cuvs::neighbors::cagra::index_params; using cuvs::neighbors::cagra::search_algo; using cuvs::neighbors::cagra::search_params; } // namespace cuvs::neighbors::experimental::cagra
0