repo_id
stringlengths
21
96
file_path
stringlengths
31
155
content
stringlengths
1
92.9M
__index_level_0__
int64
0
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/stats/minmax.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 __MINMAX_H #define __MINMAX_H #pragma once #include <optional> #include <raft/core/device_mdspan.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/stats/detail/minmax.cuh> #include <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <limits> namespace raft { namespace stats { /** * @brief Computes min/max across every column of the input matrix, as well as * optionally allow to subsample based on the given row/col ID mapping vectors * * @tparam T the data type * @tparam TPB number of threads per block * @param data input data * @param rowids actual row ID mappings. It is of length nrows. If you want to * skip this index lookup entirely, pass nullptr * @param colids actual col ID mappings. It is of length ncols. If you want to * skip this index lookup entirely, pass nullptr * @param nrows number of rows of data to be worked upon. The actual rows of the * input "data" can be bigger than this! * @param ncols number of cols of data to be worked upon. The actual cols of the * input "data" can be bigger than this! * @param row_stride stride (in number of elements) between 2 adjacent columns * @param globalmin final col-wise global minimum (size = ncols) * @param globalmax final col-wise global maximum (size = ncols) * @param sampledcols output sampled data. Pass nullptr if you don't need this * @param stream cuda stream * @note This method makes the following assumptions: * 1. input and output matrices are assumed to be col-major * 2. ncols is small enough to fit the whole of min/max values across all cols * in shared memory */ template <typename T, int TPB = 512> void minmax(const T* data, const unsigned* rowids, const unsigned* colids, int nrows, int ncols, int row_stride, T* globalmin, T* globalmax, T* sampledcols, cudaStream_t stream) { detail::minmax<T, TPB>( data, rowids, colids, nrows, ncols, row_stride, globalmin, globalmax, sampledcols, stream); } /** * @defgroup stats_minmax Min/Max * @{ */ /** * @brief Computes min/max across every column of the input matrix, as well as * optionally allow to subsample based on the given row/col ID mapping vectors * * @tparam value_t Data type of input matrix element. * @tparam idx_t Index type of matrix extent. * @param[in] handle the raft handle * @param[in] data input data col-major of size [nrows, ncols], unless rowids or * colids length is smaller * @param[in] rowids optional row ID mappings of length nrows. If you want to * skip this index lookup entirely, pass std::nullopt * @param[in] colids optional col ID mappings of length ncols. If you want to * skip this index lookup entirely, pass std::nullopt * @param[out] globalmin final col-wise global minimum (size = ncols) * @param[out] globalmax final col-wise global maximum (size = ncols) * @param[out] sampledcols output sampled data. Pass std::nullopt if you don't need this * @note This method makes the following assumptions: * 1. input and output matrices are assumed to be col-major * 2. ncols is small enough to fit the whole of min/max values across all cols * in shared memory */ template <typename value_t, typename idx_t> void minmax(raft::resources const& handle, raft::device_matrix_view<const value_t, idx_t, raft::col_major> data, std::optional<raft::device_vector_view<const unsigned, idx_t>> rowids, std::optional<raft::device_vector_view<const unsigned, idx_t>> colids, raft::device_vector_view<value_t, idx_t> globalmin, raft::device_vector_view<value_t, idx_t> globalmax, std::optional<raft::device_vector_view<value_t, idx_t>> sampledcols) { const unsigned* rowids_ptr = nullptr; const unsigned* colids_ptr = nullptr; value_t* sampledcols_ptr = nullptr; auto nrows = data.extent(0); auto ncols = data.extent(1); auto row_stride = data.stride(1); if (rowids.has_value()) { rowids_ptr = rowids.value().data_handle(); RAFT_EXPECTS(rowids.value().extent(0) <= nrows, "Rowids size is greater than nrows"); nrows = rowids.value().extent(0); } if (colids.has_value()) { colids_ptr = colids.value().data_handle(); RAFT_EXPECTS(colids.value().extent(0) <= ncols, "Colids size is greater than ncols"); ncols = colids.value().extent(0); } if (sampledcols.has_value()) { sampledcols_ptr = sampledcols.value().data_handle(); } RAFT_EXPECTS(globalmin.extent(0) == ncols, "Size mismatch between globalmin and ncols"); RAFT_EXPECTS(globalmax.extent(0) == ncols, "Size mismatch between globalmax and ncols"); detail::minmax<value_t>(data.data_handle(), rowids_ptr, colids_ptr, nrows, ncols, row_stride, globalmin.data_handle(), globalmax.data_handle(), sampledcols_ptr, resource::get_cuda_stream(handle)); } /** @} */ // end group stats_minmax }; // namespace stats }; // namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/stats/meanvar.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. */ #ifndef __MEANVAR_H #define __MEANVAR_H #pragma once #include <raft/core/device_mdspan.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/stats/detail/meanvar.cuh> namespace raft::stats { /** * @brief Compute mean and variance for each column of a given matrix. * * The operation is performed in a single sweep. Consider using it when you need to compute * both mean and variance, or when you need to compute variance but don't have the mean. * It's almost twice faster than running `mean` and `vars` sequentially, because all three * kernels are memory-bound. * * @tparam Type the data type * @tparam IdxType Integer type used for addressing * @param [out] mean the output mean vector of size D * @param [out] var the output variance vector of size D * @param [in] data the input matrix of size [N, D] * @param [in] D number of columns of data * @param [in] N number of rows of data * @param [in] sample whether to evaluate sample variance or not. In other words, whether to * normalize the variance using N-1 or N, for true or false respectively. * @param [in] rowMajor whether the input data is row- or col-major, for true or false respectively. * @param [in] stream */ template <typename Type, typename IdxType = int> void meanvar(Type* mean, Type* var, const Type* data, IdxType D, IdxType N, bool sample, bool rowMajor, cudaStream_t stream) { detail::meanvar(mean, var, data, D, N, sample, rowMajor, stream); } /** * @defgroup stats_mean_var Mean and Variance * @{ */ /** * @brief Compute mean and variance for each column of a given matrix. * * The operation is performed in a single sweep. Consider using it when you need to compute * both mean and variance, or when you need to compute variance but don't have the mean. * It's almost twice faster than running `mean` and `vars` sequentially, because all three * kernels are memory-bound. * * @tparam value_t the data type * @tparam idx_t Integer type used for addressing * @tparam layout_t Layout type of the input matrix. * @param[in] handle the raft handle * @param[in] data the input matrix of size [N, D] * @param[out] mean the output mean vector of size D * @param[out] var the output variance vector of size D * @param[in] sample whether to evaluate sample variance or not. In other words, whether to * normalize the variance using N-1 or N, for true or false respectively. */ template <typename value_t, typename idx_t, typename layout_t> void meanvar(raft::resources const& handle, raft::device_matrix_view<const value_t, idx_t, layout_t> data, raft::device_vector_view<value_t, idx_t> mean, raft::device_vector_view<value_t, idx_t> var, bool sample) { static_assert( std::is_same_v<layout_t, raft::row_major> || std::is_same_v<layout_t, raft::col_major>, "Data layout not supported"); RAFT_EXPECTS(data.extent(1) == var.extent(0), "Size mismatch between data and var"); RAFT_EXPECTS(mean.size() == var.size(), "Size mismatch between mean and var"); RAFT_EXPECTS(mean.is_exhaustive(), "mean must be contiguous"); RAFT_EXPECTS(var.is_exhaustive(), "var must be contiguous"); RAFT_EXPECTS(data.is_exhaustive(), "data must be contiguous"); detail::meanvar(mean.data_handle(), var.data_handle(), data.data_handle(), data.extent(1), data.extent(0), sample, std::is_same_v<layout_t, raft::row_major>, resource::get_cuda_stream(handle)); } /** @} */ // end group stats_mean_var }; // namespace raft::stats #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/stats/stats_types.hpp
/* * Copyright (c) 2019-2022, 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/cudart_utils.hpp> namespace raft::stats { /** * @ingroup stats_histogram * @{ */ /** * @brief Types of support histogram implementations */ enum HistType { /** shared mem atomics but with bins to be 1b int's */ HistTypeSmemBits1 = 1, /** shared mem atomics but with bins to be 2b int's */ HistTypeSmemBits2 = 2, /** shared mem atomics but with bins to be 4b int's */ HistTypeSmemBits4 = 4, /** shared mem atomics but with bins to ba 1B int's */ HistTypeSmemBits8 = 8, /** shared mem atomics but with bins to be 2B int's */ HistTypeSmemBits16 = 16, /** use only global atomics */ HistTypeGmem, /** uses shared mem atomics to reduce global traffic */ HistTypeSmem, /** * uses shared mem atomics with match_any intrinsic to further reduce shared * memory traffic. This can only be enabled on Volta and later architectures. * If one tries to enable this for older arch's, it will fall back to * `HistTypeSmem`. * @note This is to be used only when the input dataset leads to a lot of * repetitions in a given warp, else, this algo can be much slower than * `HistTypeSmem`! */ HistTypeSmemMatchAny, /** builds a hashmap of active bins in shared mem */ HistTypeSmemHash, /** decide at runtime the best algo for the given inputs */ HistTypeAuto }; /** @} */ /** * @ingroup stats_information_criterion * @{ */ /** * @brief Supported types of information criteria */ enum IC_Type { AIC, AICc, BIC }; /** @} */ }; // end namespace raft::stats
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/mean.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/linalg/eltwise.cuh> #include <raft/util/cuda_utils.cuh> #include <cub/cub.cuh> namespace raft { namespace stats { namespace detail { ///@todo: ColsPerBlk has been tested only for 32! template <typename Type, typename IdxType, int TPB, int ColsPerBlk = 32> RAFT_KERNEL meanKernelRowMajor(Type* mu, const Type* data, IdxType D, IdxType N) { const int RowsPerBlkPerIter = TPB / ColsPerBlk; IdxType thisColId = threadIdx.x % ColsPerBlk; IdxType thisRowId = threadIdx.x / ColsPerBlk; IdxType colId = thisColId + ((IdxType)blockIdx.y * ColsPerBlk); IdxType rowId = thisRowId + ((IdxType)blockIdx.x * RowsPerBlkPerIter); Type thread_data = Type(0); const IdxType stride = RowsPerBlkPerIter * gridDim.x; for (IdxType i = rowId; i < N; i += stride) thread_data += (colId < D) ? data[i * D + colId] : Type(0); __shared__ Type smu[ColsPerBlk]; if (threadIdx.x < ColsPerBlk) smu[threadIdx.x] = Type(0); __syncthreads(); raft::myAtomicAdd(smu + thisColId, thread_data); __syncthreads(); if (threadIdx.x < ColsPerBlk) raft::myAtomicAdd(mu + colId, smu[thisColId]); } template <typename Type, typename IdxType, int TPB> RAFT_KERNEL meanKernelColMajor(Type* mu, const Type* data, IdxType D, IdxType N) { typedef cub::BlockReduce<Type, TPB> BlockReduce; __shared__ typename BlockReduce::TempStorage temp_storage; Type thread_data = Type(0); IdxType colStart = N * blockIdx.x; for (IdxType i = threadIdx.x; i < N; i += TPB) { IdxType idx = colStart + i; thread_data += data[idx]; } Type acc = BlockReduce(temp_storage).Sum(thread_data); if (threadIdx.x == 0) { mu[blockIdx.x] = acc / N; } } template <typename Type, typename IdxType = int> void mean( Type* mu, const Type* data, IdxType D, IdxType N, bool sample, bool rowMajor, cudaStream_t stream) { static const int TPB = 256; if (rowMajor) { static const int RowsPerThread = 4; static const int ColsPerBlk = 32; static const int RowsPerBlk = (TPB / ColsPerBlk) * RowsPerThread; dim3 grid(raft::ceildiv(N, (IdxType)RowsPerBlk), raft::ceildiv(D, (IdxType)ColsPerBlk)); RAFT_CUDA_TRY(cudaMemsetAsync(mu, 0, sizeof(Type) * D, stream)); meanKernelRowMajor<Type, IdxType, TPB, ColsPerBlk><<<grid, TPB, 0, stream>>>(mu, data, D, N); RAFT_CUDA_TRY(cudaPeekAtLastError()); Type ratio = Type(1) / (sample ? Type(N - 1) : Type(N)); raft::linalg::scalarMultiply(mu, mu, ratio, D, stream); } else { meanKernelColMajor<Type, IdxType, TPB><<<D, TPB, 0, stream>>>(mu, data, D, N); } RAFT_CUDA_TRY(cudaPeekAtLastError()); } } // namespace detail } // namespace stats } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/cov.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 <raft/core/resource/cublas_handle.hpp> #include <raft/linalg/gemm.cuh> #include <raft/stats/mean_center.cuh> namespace raft { namespace stats { namespace detail { /** * @brief Compute covariance of the input matrix * * Mean operation is assumed to be performed on a given column. * * @tparam Type the data type * @param covar the output covariance matrix * @param data the input matrix (this will get mean-centered at the end!) * @param mu mean vector of the input matrix * @param D number of columns of data * @param N number of rows of data * @param sample whether to evaluate sample covariance or not. In other words, * whether to normalize the output using N-1 or N, for true or false, * respectively * @param rowMajor whether the input data is row or col major * @param stable whether to run the slower-but-numerically-stable version or not * @param handle cublas handle * @param stream cuda stream * @note if stable=true, then the input data will be mean centered after this * function returns! */ template <typename Type> void cov(raft::resources const& handle, Type* covar, Type* data, const Type* mu, std::size_t D, std::size_t N, bool sample, bool rowMajor, bool stable, cudaStream_t stream) { if (stable) { cublasHandle_t cublas_h = resource::get_cublas_handle(handle); // since mean operation is assumed to be along a given column, broadcast // must be along rows! raft::stats::meanCenter(data, data, mu, D, N, rowMajor, true, stream); Type alpha = Type(1) / (sample ? Type(N - 1) : Type(N)); Type beta = Type(0); if (rowMajor) { // #TODO: Call from public API when ready RAFT_CUBLAS_TRY(raft::linalg::detail::cublasgemm(cublas_h, CUBLAS_OP_N, CUBLAS_OP_T, D, D, N, &alpha, data, D, data, D, &beta, covar, D, stream)); } else { raft::linalg::gemm( handle, data, N, D, data, covar, D, D, CUBLAS_OP_T, CUBLAS_OP_N, alpha, beta, stream); } } else { ///@todo: implement this using cutlass + customized epilogue! ASSERT(false, "cov: Implement stable=false case!"); } RAFT_CUDA_TRY(cudaPeekAtLastError()); } }; // end namespace detail }; // end namespace stats }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/contingencyMatrix.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 <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <thrust/device_ptr.h> #include <thrust/execution_policy.h> #include <thrust/extrema.h> #include <thrust/reduce.h> #include <cub/cub.cuh> #include <math.h> namespace raft { namespace stats { namespace detail { typedef enum { IMPL_NONE, SMEM_ATOMICS, GLOBAL_ATOMICS, SORT_AND_GATOMICS } ContingencyMatrixImplType; template <typename T, typename OutT = int> RAFT_KERNEL devConstructContingencyMatrix(const T* groundTruth, const T* predicted, int nSamples, OutT* outMat, int outIdxOffset, int outMatWidth) { int elementId = threadIdx.x + blockDim.x * blockIdx.x; if (elementId < nSamples) { T gt = groundTruth[elementId]; T pd = predicted[elementId]; auto outputIdx = (gt - outIdxOffset) * outMatWidth + pd - outIdxOffset; raft::myAtomicAdd(outMat + outputIdx, OutT(1)); } } template <typename T, typename OutT = int> void computeCMatWAtomics(const T* groundTruth, const T* predictedLabel, int nSamples, OutT* outMat, int outIdxOffset, int outDimN, cudaStream_t stream) { RAFT_CUDA_TRY( cudaFuncSetCacheConfig(devConstructContingencyMatrix<T, OutT>, cudaFuncCachePreferL1)); static const int block = 128; auto grid = raft::ceildiv(nSamples, block); devConstructContingencyMatrix<T, OutT><<<grid, block, 0, stream>>>( groundTruth, predictedLabel, nSamples, outMat, outIdxOffset, outDimN); RAFT_CUDA_TRY(cudaGetLastError()); } template <typename T, typename OutT = int> RAFT_KERNEL devConstructContingencyMatrixSmem(const T* groundTruth, const T* predicted, int nSamples, OutT* outMat, int outIdxOffset, int outMatWidth) { extern __shared__ char smem[]; auto* sMemMatrix = reinterpret_cast<OutT*>(smem); for (int smemIdx = threadIdx.x; smemIdx < outMatWidth * outMatWidth; smemIdx += blockDim.x) { sMemMatrix[smemIdx] = 0; } __syncthreads(); int elementId = threadIdx.x + blockDim.x * blockIdx.x; if (elementId < nSamples) { T gt = groundTruth[elementId]; T pd = predicted[elementId]; auto outputIdx = (gt - outIdxOffset) * outMatWidth + pd - outIdxOffset; raft::myAtomicAdd(sMemMatrix + outputIdx, OutT(1)); } __syncthreads(); for (int smemIdx = threadIdx.x; smemIdx < outMatWidth * outMatWidth; smemIdx += blockDim.x) { raft::myAtomicAdd(outMat + smemIdx, sMemMatrix[smemIdx]); } } template <typename T, typename OutT = int> void computeCMatWSmemAtomics(const T* groundTruth, const T* predictedLabel, int nSamples, OutT* outMat, int outIdxOffset, int outDimN, cudaStream_t stream) { static const int block = 128; auto grid = raft::ceildiv(nSamples, block); size_t smemSizePerBlock = outDimN * outDimN * sizeof(OutT); devConstructContingencyMatrixSmem<T, OutT><<<grid, block, smemSizePerBlock, stream>>>( groundTruth, predictedLabel, nSamples, outMat, outIdxOffset, outDimN); RAFT_CUDA_TRY(cudaGetLastError()); } template <typename T, typename OutT = int> void contingencyMatrixWSort(const T* groundTruth, const T* predictedLabel, int nSamples, OutT* outMat, T minLabel, T maxLabel, void* workspace, size_t workspaceSize, cudaStream_t stream) { T* outKeys = reinterpret_cast<T*>(workspace); auto alignedBufferSz = raft::alignTo<size_t>(nSamples * sizeof(T), 256); T* outValue = reinterpret_cast<T*>((size_t)workspace + alignedBufferSz); void* pWorkspaceCub = reinterpret_cast<void*>((size_t)workspace + 2 * alignedBufferSz); auto bitsToSort = log2<int>(maxLabel); if (!raft::isPo2(maxLabel)) ++bitsToSort; // we dont really need perfect sorting, should get by with some sort of // binning-reordering operation ///@todo: future work - explore "efficient" custom binning kernels vs cub sort RAFT_CUDA_TRY(cub::DeviceRadixSort::SortPairs(pWorkspaceCub, workspaceSize, groundTruth, outKeys, predictedLabel, outValue, nSamples, 0, bitsToSort, stream)); auto outDimM_N = int(maxLabel - minLabel + 1); computeCMatWAtomics<T, OutT>(outKeys, outValue, nSamples, outMat, minLabel, outDimM_N, stream); } template <typename OutT = int> ContingencyMatrixImplType getImplVersion(OutT outDimN) { int currDevice = 0; int l2CacheSize = 0; // no way to query this from CUDA APIs, value for CC 7.0, 3.0 int maxBlocksResidentPerSM = 16; RAFT_CUDA_TRY(cudaGetDevice(&currDevice)); RAFT_CUDA_TRY(cudaDeviceGetAttribute(&l2CacheSize, cudaDevAttrL2CacheSize, currDevice)); auto maxSmemPerBlock = raft::getSharedMemPerBlock(); ContingencyMatrixImplType implVersion = IMPL_NONE; // keeping 8 block per SM to get good utilization // can go higher but reduced L1 size degrades perf OutT upperLimitSmemAtomics = std::floor(std::sqrt(maxSmemPerBlock / (sizeof(OutT) * (maxBlocksResidentPerSM / 2)))); OutT upperLimitL2Atomics = std::floor(std::sqrt(l2CacheSize / sizeof(OutT))); if (outDimN <= upperLimitSmemAtomics) implVersion = SMEM_ATOMICS; else if (outDimN <= upperLimitL2Atomics) implVersion = GLOBAL_ATOMICS; else implVersion = SORT_AND_GATOMICS; return implVersion; } /** * @brief use this to allocate output matrix size * size of matrix = (maxLabel - minLabel + 1)^2 * sizeof(int) * @param groundTruth: device 1-d array for ground truth (num of rows) * @param nSamples: number of elements in input array * @param stream: cuda stream for execution * @param minLabel: [out] calculated min value in input array * @param maxLabel: [out] calculated max value in input array */ template <typename T> void getInputClassCardinality( const T* groundTruth, const int nSamples, cudaStream_t stream, T& minLabel, T& maxLabel) { thrust::device_ptr<const T> dTrueLabel = thrust::device_pointer_cast(groundTruth); auto min_max = thrust::minmax_element(thrust::cuda::par.on(stream), dTrueLabel, dTrueLabel + nSamples); minLabel = *min_max.first; maxLabel = *min_max.second; } /** * @brief Calculate workspace size for running contingency matrix calculations * @tparam T label type * @tparam OutT output matrix type * @param nSamples: number of elements in input array * @param groundTruth: device 1-d array for ground truth (num of rows) * @param stream: cuda stream for execution * @param minLabel: Optional, min value in input array * @param maxLabel: Optional, max value in input array */ template <typename T, typename OutT = int> size_t getContingencyMatrixWorkspaceSize(int nSamples, const T* groundTruth, cudaStream_t stream, T minLabel = std::numeric_limits<T>::max(), T maxLabel = std::numeric_limits<T>::max()) { size_t workspaceSize = 0; // below is a redundant computation - can be avoided if (minLabel == std::numeric_limits<T>::max() || maxLabel == std::numeric_limits<T>::max()) { getInputClassCardinality<T>(groundTruth, nSamples, stream, minLabel, maxLabel); } auto outDimN = OutT(maxLabel - minLabel + 1); ContingencyMatrixImplType implVersion = getImplVersion<OutT>(outDimN); if (implVersion == SORT_AND_GATOMICS) { void* pWorkspaceCub{}; size_t tmpStorageBytes = 0; // no-op pointers to get workspace size T* pTmpUnused{}; RAFT_CUDA_TRY(cub::DeviceRadixSort::SortPairs( pWorkspaceCub, tmpStorageBytes, pTmpUnused, pTmpUnused, pTmpUnused, pTmpUnused, nSamples)); auto tmpStagingMemorySize = raft::alignTo<size_t>(nSamples * sizeof(T), 256); tmpStagingMemorySize *= 2; workspaceSize = tmpStagingMemorySize + tmpStorageBytes; } return workspaceSize; } /** * @brief construct contingency matrix given input ground truth and prediction * labels. Users should call function getInputClassCardinality to find * and allocate memory for output. Similarly workspace requirements * should be checked using function getContingencyMatrixWorkspaceSize * @tparam T label type * @tparam OutT output matrix type * @param groundTruth: device 1-d array for ground truth (num of rows) * @param predictedLabel: device 1-d array for prediction (num of columns) * @param nSamples: number of elements in input array * @param outMat: output buffer for contingecy matrix * @param stream: cuda stream for execution * @param workspace: Optional, workspace memory allocation * @param workspaceSize: Optional, size of workspace memory * @param minLabel: Optional, min value in input ground truth array * @param maxLabel: Optional, max value in input ground truth array */ template <typename T, typename OutT = int> void contingencyMatrix(const T* groundTruth, const T* predictedLabel, int nSamples, OutT* outMat, cudaStream_t stream, void* workspace = nullptr, size_t workspaceSize = 0, T minLabel = std::numeric_limits<T>::max(), T maxLabel = std::numeric_limits<T>::max()) { // assumptions: // output is not at par with scikit learn - output will be square matrix // always with numRows = numColumns = numOfClassesInTrueLabel // it is also assumed that true labels are monotically increasing // if for some reason groundTruth completely skips some labels // eg: {0,1,2,5} instead of {0,1,2,3}. // Output matrix will still have empty rows for label value {3,4} // Users can use "make_monotonic" to convert their discontinuous input label // range to a monotonically increasing one // // this also serves as way to measure co-occurrence/joint counts for NLP tasks which // can be used to then compute pointwise mutual information and mutual information if (minLabel == std::numeric_limits<T>::max() || maxLabel == std::numeric_limits<T>::max()) { getInputClassCardinality<T>(groundTruth, nSamples, stream, minLabel, maxLabel); } auto outDimM_N = OutT(maxLabel - minLabel + 1); RAFT_CUDA_TRY(cudaMemsetAsync(outMat, 0, sizeof(OutT) * outDimM_N * outDimM_N, stream)); ContingencyMatrixImplType implVersion = getImplVersion<OutT>(outDimM_N); switch (implVersion) { case SMEM_ATOMICS: // smem atomics and then single global mem atomics only works // when all label count can fit in smem for a block // helps when GLOBAL_ATOMICS performance blocked by atomic update // serialization -when very less labels ~10 labels computeCMatWSmemAtomics<T, OutT>( groundTruth, predictedLabel, nSamples, outMat, minLabel, outDimM_N, stream); break; case GLOBAL_ATOMICS: // launch kernel - global atomic ops per (groundTruth,predictedValue) pair computeCMatWAtomics<T, OutT>( groundTruth, predictedLabel, nSamples, outMat, minLabel, outDimM_N, stream); break; // more L2 thrashing if atomic OPs land in completely different mem // segment - when more labels case SORT_AND_GATOMICS: contingencyMatrixWSort<T, OutT>(groundTruth, predictedLabel, nSamples, outMat, minLabel, maxLabel, workspace, workspaceSize, stream); break; case IMPL_NONE: break; } } }; // namespace detail }; // namespace stats }; // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/trustworthiness_score.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. */ #include <raft/core/resource/cuda_stream.hpp> #include <raft/distance/distance.cuh> #include <raft/matrix/col_wise_sort.cuh> #include <raft/spatial/knn/knn.cuh> #include <rmm/device_scalar.hpp> #include <rmm/device_uvector.hpp> #define N_THREADS 512 namespace raft { namespace stats { namespace detail { /** * @brief Build the lookup table * @param[out] lookup_table: Lookup table giving nearest neighbor order * of pairwise distance calculations given sample index * @param[in] X_ind: Sorted indexes of pairwise distance calculations of X * @param n: Number of samples * @param work: Number of elements to consider */ RAFT_KERNEL build_lookup_table(int* lookup_table, const int* X_ind, int n, int work) { int i = blockIdx.x * blockDim.x + threadIdx.x; if (i >= work) return; int sample_idx = i / n; int nn_idx = i % n; int idx = X_ind[i]; lookup_table[(sample_idx * n) + idx] = nn_idx; } /** * @brief Compute a the rank of trustworthiness score * @param[out] rank: Resulting rank * @param[out] lookup_table: Lookup table giving nearest neighbor order * of pairwise distance calculations given sample index * @param[in] emb_ind: Indexes of KNN on embeddings * @param n: Number of samples * @param n_neighbors: Number of neighbors considered by trustworthiness score * @param work: Batch to consider (to do it at once use n * n_neighbors) */ template <typename knn_index_t> RAFT_KERNEL compute_rank(double* rank, const int* lookup_table, const knn_index_t* emb_ind, int n, int n_neighbors, int work) { int i = blockIdx.x * blockDim.x + threadIdx.x; if (i >= work) return; int sample_idx = i / n_neighbors; knn_index_t emb_nn_ind = emb_ind[i]; int r = lookup_table[(sample_idx * n) + emb_nn_ind]; int tmp = r - n_neighbors + 1; if (tmp > 0) raft::myAtomicAdd<double>(rank, tmp); } /** * @brief Compute a kNN and returns the indices of the nearest neighbors * @param h Raft handle * @param[in] input Input matrix containing the dataset * @param n Number of samples * @param d Number of features * @param n_neighbors number of neighbors * @param[out] indices KNN indexes * @param[out] distances KNN distances */ template <raft::distance::DistanceType distance_type, typename math_t> void run_knn(const raft::resources& h, math_t* input, int n, int d, int n_neighbors, int64_t* indices, math_t* distances) { std::vector<math_t*> ptrs(1); std::vector<int> sizes(1); ptrs[0] = input; sizes[0] = n; raft::spatial::knn::brute_force_knn<int64_t, float, int>(h, ptrs, sizes, d, input, n, indices, distances, n_neighbors, true, true, nullptr, distance_type); } /** * @brief Compute the trustworthiness score * @param h Raft handle * @param X[in]: Data in original dimension * @param X_embedded[in]: Data in target dimension (embedding) * @param n: Number of samples * @param m: Number of features in high/original dimension * @param d: Number of features in low/embedded dimension * @param n_neighbors Number of neighbors considered by trustworthiness score * @param batchSize Batch size * @return Trustworthiness score */ template <typename math_t, raft::distance::DistanceType distance_type> double trustworthiness_score(const raft::resources& h, const math_t* X, math_t* X_embedded, int n, int m, int d, int n_neighbors, int batchSize = 512) { cudaStream_t stream = resource::get_cuda_stream(h); const int KNN_ALLOC = n * (n_neighbors + 1); rmm::device_uvector<int64_t> emb_ind(KNN_ALLOC, stream); rmm::device_uvector<math_t> emb_dist(KNN_ALLOC, stream); run_knn<distance_type>(h, X_embedded, n, d, n_neighbors + 1, emb_ind.data(), emb_dist.data()); const int PAIRWISE_ALLOC = batchSize * n; rmm::device_uvector<int> X_ind(PAIRWISE_ALLOC, stream); rmm::device_uvector<math_t> X_dist(PAIRWISE_ALLOC, stream); rmm::device_uvector<int> lookup_table(PAIRWISE_ALLOC, stream); double t = 0.0; rmm::device_scalar<double> t_dbuf(stream); int toDo = n; while (toDo > 0) { int curBatchSize = min(toDo, batchSize); // Takes at most batchSize vectors at a time raft::distance::pairwise_distance( h, &X[(n - toDo) * m], X, X_dist.data(), curBatchSize, n, m, distance_type); size_t colSortWorkspaceSize = 0; bool bAllocWorkspace = false; raft::matrix::sort_cols_per_row(X_dist.data(), X_ind.data(), curBatchSize, n, bAllocWorkspace, nullptr, colSortWorkspaceSize, stream); if (bAllocWorkspace) { rmm::device_uvector<char> sortColsWorkspace(colSortWorkspaceSize, stream); raft::matrix::sort_cols_per_row(X_dist.data(), X_ind.data(), curBatchSize, n, bAllocWorkspace, sortColsWorkspace.data(), colSortWorkspaceSize, stream); } int work = curBatchSize * n; int n_blocks = raft::ceildiv(work, N_THREADS); build_lookup_table<<<n_blocks, N_THREADS, 0, stream>>>( lookup_table.data(), X_ind.data(), n, work); RAFT_CUDA_TRY(cudaMemsetAsync(t_dbuf.data(), 0, sizeof(double), stream)); work = curBatchSize * (n_neighbors + 1); n_blocks = raft::ceildiv(work, N_THREADS); compute_rank<<<n_blocks, N_THREADS, 0, stream>>>( t_dbuf.data(), lookup_table.data(), &emb_ind.data()[(n - toDo) * (n_neighbors + 1)], n, n_neighbors + 1, work); RAFT_CUDA_TRY(cudaPeekAtLastError()); t += t_dbuf.value(stream); toDo -= curBatchSize; } t = 1.0 - ((2.0 / ((n * n_neighbors) * ((2.0 * n) - (3.0 * n_neighbors) - 1.0))) * t); return t; } } // namespace detail } // namespace stats } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/v_measure.cuh
/* * Copyright (c) 2019-2022, 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 v_measure.cuh */ #include <raft/stats/homogeneity_score.cuh> namespace raft { namespace stats { namespace detail { /** * @brief Function to calculate the v-measure between two clusters * * @param truthClusterArray: the array of truth classes of type T * @param predClusterArray: the array of predicted classes of type T * @param size: the size of the data points of type int * @param lowerLabelRange: the lower bound of the range of labels * @param upperLabelRange: the upper bound of the range of labels * @param stream: the cudaStream object * @param beta: v_measure parameter */ template <typename T> double v_measure(const T* truthClusterArray, const T* predClusterArray, int size, T lowerLabelRange, T upperLabelRange, cudaStream_t stream, double beta = 1.0) { double computedHomogeity, computedCompleteness, computedVMeasure; computedHomogeity = raft::stats::homogeneity_score( truthClusterArray, predClusterArray, size, lowerLabelRange, upperLabelRange, stream); computedCompleteness = raft::stats::homogeneity_score( predClusterArray, truthClusterArray, size, lowerLabelRange, upperLabelRange, stream); if (computedCompleteness + computedHomogeity == 0.0) computedVMeasure = 0.0; else computedVMeasure = ((1 + beta) * computedHomogeity * computedCompleteness / (beta * computedHomogeity + computedCompleteness)); return computedVMeasure; } }; // end namespace detail }; // end namespace stats }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/rand_index.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. */ /** * @file rand_index.cuh * @todo TODO(Ganesh Venkataramana): * <pre> * The below rand_index calculation implementation is a Brute force one that uses (nElements*nElements) threads (2 dimensional grids and blocks) * For small datasets, this will suffice; but for larger ones, work done by the threads increase dramatically. * A more mathematically intensive implementation that uses half the above threads can be done, which will prove to be more efficient for larger datasets * the idea is as follows: * instead of 2D block and grid configuration with a total of (nElements*nElements) threads (where each (i,j) through these threads represent an ordered pair selection of 2 data points), a 1D block and grid configuration with a total of (nElements*(nElements))/2 threads (each thread index represents an element part of the set of unordered pairwise selections from the dataset (nChoose2)) * In this setup, one has to generate a one-to-one mapping between this 1D thread index (for each kernel) and the unordered pair of chosen datapoints. * More specifically, thread0-> {dataPoint1, dataPoint0}, thread1-> {dataPoint2, dataPoint0}, thread2-> {dataPoint2, dataPoint1} ... thread((nElements*(nElements))/2 - 1)-> {dataPoint(nElements-1),dataPoint(nElements-2)} * say , * threadNum: thread index | threadNum = threadIdx.x + BlockIdx.x*BlockDim.x, * i : index of dataPoint i * j : index of dataPoint j * then the mapping is as follows: * i = ceil((-1 + sqrt(1 + 8*(1 + threadNum)))/2) = floor((1 + sqrt(1 + 8*threadNum))/2) * j = threadNum - i(i-1)/2 * after obtaining the the pair of datapoints, calculation of rand index is the same as done in this implementation * Caveat: since the kernel implementation involves use of emulated sqrt() operations: * the number of instructions executed per kernel is ~40-50 times * as the O(nElements*nElements) increase beyond the floating point limit, floating point inaccuracies occur, and hence the above floor(...) != ceil(...) * </pre> */ #pragma once #include <cub/cub.cuh> #include <math.h> #include <raft/core/interruptible.hpp> #include <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <rmm/device_uvector.hpp> namespace raft { namespace stats { namespace detail { /** * @brief kernel to calculate the values of a and b * @param firstClusterArray: the array of classes of type T * @param secondClusterArray: the array of classes of type T * @param size: the size of the data points * @param a: number of pairs of points that both the clusters have classified the same * @param b: number of pairs of points that both the clusters have classified differently */ template <typename T, int BLOCK_DIM_X, int BLOCK_DIM_Y> RAFT_KERNEL computeTheNumerator( const T* firstClusterArray, const T* secondClusterArray, uint64_t size, uint64_t* a, uint64_t* b) { // calculating the indices of pairs of datapoints compared by the current thread uint64_t j = threadIdx.x + blockIdx.x * blockDim.x; uint64_t i = threadIdx.y + blockIdx.y * blockDim.y; // thread-local variables to count a and b uint64_t myA = 0, myB = 0; if (i < size && j < size && j < i) { // checking if the pair have been classified the same by both the clusters if (firstClusterArray[i] == firstClusterArray[j] && secondClusterArray[i] == secondClusterArray[j]) { ++myA; } // checking if the pair have been classified differently by both the clusters else if (firstClusterArray[i] != firstClusterArray[j] && secondClusterArray[i] != secondClusterArray[j]) { ++myB; } } // specialize blockReduce for a 2D block of 1024 threads of type uint64_t typedef cub::BlockReduce<uint64_t, BLOCK_DIM_X, cub::BLOCK_REDUCE_WARP_REDUCTIONS, BLOCK_DIM_Y> BlockReduce; // Allocate shared memory for blockReduce __shared__ typename BlockReduce::TempStorage temp_storage; // summing up thread-local counts specific to a block myA = BlockReduce(temp_storage).Sum(myA); __syncthreads(); myB = BlockReduce(temp_storage).Sum(myB); __syncthreads(); // executed once per block if (threadIdx.x == 0 && threadIdx.y == 0) { raft::myAtomicAdd<unsigned long long int>((unsigned long long int*)a, myA); raft::myAtomicAdd<unsigned long long int>((unsigned long long int*)b, myB); } } /** * @brief Function to calculate RandIndex * <a href="https://en.wikipedia.org/wiki/Rand_index">more info on rand index</a> * @param firstClusterArray: the array of classes of type T * @param secondClusterArray: the array of classes of type T * @param size: the size of the data points of type uint64_t * @param stream: the cudaStream object */ template <typename T> double compute_rand_index(const T* firstClusterArray, const T* secondClusterArray, uint64_t size, cudaStream_t stream) { // rand index for size less than 2 is not defined ASSERT(size >= 2, "Rand Index for size less than 2 not defined!"); // allocating and initializing memory for a and b in the GPU rmm::device_uvector<uint64_t> arr_buf(2, stream); RAFT_CUDA_TRY(cudaMemsetAsync(arr_buf.data(), 0, 2 * sizeof(uint64_t), stream)); // kernel configuration static const int BLOCK_DIM_Y = 16, BLOCK_DIM_X = 16; dim3 numThreadsPerBlock(BLOCK_DIM_X, BLOCK_DIM_Y); dim3 numBlocks(raft::ceildiv<int>(size, numThreadsPerBlock.x), raft::ceildiv<int>(size, numThreadsPerBlock.y)); // calling the kernel computeTheNumerator<T, BLOCK_DIM_X, BLOCK_DIM_Y><<<numBlocks, numThreadsPerBlock, 0, stream>>>( firstClusterArray, secondClusterArray, size, arr_buf.data(), arr_buf.data() + 1); // synchronizing and updating the calculated values of a and b from device to host uint64_t ab_host[2] = {0}; raft::update_host(ab_host, arr_buf.data(), 2, stream); raft::interruptible::synchronize(stream); // error handling RAFT_CUDA_TRY(cudaGetLastError()); // denominator uint64_t nChooseTwo = size * (size - 1) / 2; // calculating the rand_index return (double)(((double)(ab_host[0] + ab_host[1])) / (double)nChooseTwo); } }; // end namespace detail }; // end namespace stats }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/adjusted_rand_index.cuh
/* * Copyright (c) 2019-2022, 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 adjusted_rand_index.cuh * @brief The adjusted Rand index is the corrected-for-chance version of the Rand index. * Such a correction for chance establishes a baseline by using the expected similarity * of all pair-wise comparisons between clusterings specified by a random model. */ #pragma once #include "contingencyMatrix.cuh" #include <cub/cub.cuh> #include <math.h> #include <raft/linalg/map_then_reduce.cuh> #include <raft/linalg/reduce.cuh> #include <raft/stats/histogram.cuh> #include <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <rmm/device_scalar.hpp> #include <rmm/device_uvector.hpp> #include <thrust/device_ptr.h> #include <thrust/execution_policy.h> #include <thrust/extrema.h> namespace raft { namespace stats { namespace detail { /** * @brief Lambda to calculate the number of unordered pairs in a given input * * @tparam Type: Data type of the input * @param in: the input to the functional mapping * @param i: the indexing(not used in this case) */ template <typename Type> struct nCTwo { HDI Type operator()(Type in, int i = 0) { return in % 2 ? ((in - 1) >> 1) * in : (in >> 1) * (in - 1); } }; template <typename DataT, typename IdxT> struct Binner { Binner(DataT minL) : minLabel(minL) {} DI int operator()(DataT val, IdxT row, IdxT col) { return int(val - minLabel); } private: DataT minLabel; }; // struct Binner /** * @brief Function to count the number of unique elements in the input array * * @tparam T data-type for input arrays * * @param[in] arr input array [on device] [len = size] * @param[in] size the size of the input array * @param[out] minLabel the lower bound of the range of labels * @param[out] maxLabel the upper bound of the range of labels * @param[in] stream cuda stream * * @return the number of unique elements in the array */ template <typename T> int countUnique(const T* arr, int size, T& minLabel, T& maxLabel, cudaStream_t stream) { auto ptr = thrust::device_pointer_cast(arr); auto minmax = thrust::minmax_element(thrust::cuda::par.on(stream), ptr, ptr + size); minLabel = *minmax.first; maxLabel = *minmax.second; auto totalLabels = int(maxLabel - minLabel + 1); rmm::device_uvector<int> labelCounts(totalLabels, stream); rmm::device_scalar<int> nUniq(stream); raft::stats::histogram<T, int>( raft::stats::HistTypeAuto, labelCounts.data(), totalLabels, arr, size, 1, stream, [minLabel] __device__(T val, int row, int col) { return int(val - minLabel); }); raft::linalg::mapThenSumReduce<int>( nUniq.data(), totalLabels, [] __device__(const T& val) { return val != 0; }, stream, labelCounts.data()); auto numUniques = nUniq.value(stream); return numUniques; } /** * @brief Function to calculate Adjusted RandIndex as described * <a href="https://en.wikipedia.org/wiki/Rand_index">here</a> * @tparam T data-type for input label arrays * @tparam MathT integral data-type used for computing n-choose-r * @param firstClusterArray: the array of classes * @param secondClusterArray: the array of classes * @param size: the size of the data points of type int * @param stream: the cudaStream object */ template <typename T, typename MathT = int> double compute_adjusted_rand_index(const T* firstClusterArray, const T* secondClusterArray, int size, cudaStream_t stream) { ASSERT(size >= 2, "Rand Index for size less than 2 not defined!"); T minFirst, maxFirst, minSecond, maxSecond; auto nUniqFirst = countUnique(firstClusterArray, size, minFirst, maxFirst, stream); auto nUniqSecond = countUnique(secondClusterArray, size, minSecond, maxSecond, stream); auto lowerLabelRange = std::min(minFirst, minSecond); auto upperLabelRange = std::max(maxFirst, maxSecond); auto nClasses = upperLabelRange - lowerLabelRange + 1; // degenerate case of single cluster or clusters each with just one element if (nUniqFirst == nUniqSecond) { if (nUniqFirst == 1 || nUniqFirst == size) return 1.0; } auto nUniqClasses = MathT(nClasses); rmm::device_uvector<MathT> dContingencyMatrix(nUniqClasses * nUniqClasses, stream); RAFT_CUDA_TRY(cudaMemsetAsync( dContingencyMatrix.data(), 0, nUniqClasses * nUniqClasses * sizeof(MathT), stream)); auto workspaceSz = getContingencyMatrixWorkspaceSize<T, MathT>( size, firstClusterArray, stream, lowerLabelRange, upperLabelRange); rmm::device_uvector<char> workspaceBuff(workspaceSz, stream); contingencyMatrix<T, MathT>(firstClusterArray, secondClusterArray, size, dContingencyMatrix.data(), stream, workspaceBuff.data(), workspaceSz, lowerLabelRange, upperLabelRange); rmm::device_uvector<MathT> a(nUniqClasses, stream); rmm::device_uvector<MathT> b(nUniqClasses, stream); rmm::device_scalar<MathT> d_aCTwoSum(stream); rmm::device_scalar<MathT> d_bCTwoSum(stream); rmm::device_scalar<MathT> d_nChooseTwoSum(stream); MathT h_aCTwoSum, h_bCTwoSum, h_nChooseTwoSum; RAFT_CUDA_TRY(cudaMemsetAsync(a.data(), 0, nUniqClasses * sizeof(MathT), stream)); RAFT_CUDA_TRY(cudaMemsetAsync(b.data(), 0, nUniqClasses * sizeof(MathT), stream)); RAFT_CUDA_TRY(cudaMemsetAsync(d_aCTwoSum.data(), 0, sizeof(MathT), stream)); RAFT_CUDA_TRY(cudaMemsetAsync(d_bCTwoSum.data(), 0, sizeof(MathT), stream)); RAFT_CUDA_TRY(cudaMemsetAsync(d_nChooseTwoSum.data(), 0, sizeof(MathT), stream)); // calculating the sum of NijC2 raft::linalg::mapThenSumReduce<MathT, nCTwo<MathT>>(d_nChooseTwoSum.data(), nUniqClasses * nUniqClasses, nCTwo<MathT>(), stream, dContingencyMatrix.data(), dContingencyMatrix.data()); // calculating the row-wise sums raft::linalg::reduce<MathT, MathT>( a.data(), dContingencyMatrix.data(), nUniqClasses, nUniqClasses, 0, true, true, stream); // calculating the column-wise sums raft::linalg::reduce<MathT, MathT>( b.data(), dContingencyMatrix.data(), nUniqClasses, nUniqClasses, 0, true, false, stream); // calculating the sum of number of unordered pairs for every element in a raft::linalg::mapThenSumReduce<MathT, nCTwo<MathT>>( d_aCTwoSum.data(), nUniqClasses, nCTwo<MathT>(), stream, a.data(), a.data()); // calculating the sum of number of unordered pairs for every element of b raft::linalg::mapThenSumReduce<MathT, nCTwo<MathT>>( d_bCTwoSum.data(), nUniqClasses, nCTwo<MathT>(), stream, b.data(), b.data()); // updating in the host memory raft::update_host(&h_nChooseTwoSum, d_nChooseTwoSum.data(), 1, stream); raft::update_host(&h_aCTwoSum, d_aCTwoSum.data(), 1, stream); raft::update_host(&h_bCTwoSum, d_bCTwoSum.data(), 1, stream); // calculating the ARI auto nChooseTwo = double(size) * double(size - 1) / 2.0; auto expectedIndex = double(h_aCTwoSum) * double(h_bCTwoSum) / double(nChooseTwo); auto maxIndex = (double(h_bCTwoSum) + double(h_aCTwoSum)) / 2.0; auto index = double(h_nChooseTwoSum); if (maxIndex - expectedIndex) return (index - expectedIndex) / (maxIndex - expectedIndex); else return 0; } }; // end namespace detail }; // end namespace stats }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/kl_divergence.cuh
/* * Copyright (c) 2019-2022, 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 kl_divergence.cuh * @brief The KL divergence tells us how well the probability distribution Q AKA candidatePDF * approximates the probability distribution P AKA modelPDF. */ #pragma once #include <math.h> #include <raft/linalg/map_then_reduce.cuh> #include <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <rmm/device_scalar.hpp> namespace raft { namespace stats { namespace detail { /** * @brief the KL Diverence mapping function * * @tparam Type: Data type of the input * @param modelPDF: the model probability density function of type DataT * @param candidatePDF: the candidate probability density function of type DataT */ template <typename Type> struct KLDOp { HDI Type operator()(Type modelPDF, Type candidatePDF) { if (modelPDF == 0.0) return 0; else return modelPDF * (log(modelPDF) - log(candidatePDF)); } }; /** * @brief Function to calculate KL Divergence * <a href="https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence">more info on KL * Divergence</a> * * @tparam DataT: Data type of the input array * @param modelPDF: the model array of probability density functions of type DataT * @param candidatePDF: the candidate array of probability density functions of type DataT * @param size: the size of the data points of type int * @param stream: the cudaStream object */ template <typename DataT> DataT kl_divergence(const DataT* modelPDF, const DataT* candidatePDF, int size, cudaStream_t stream) { rmm::device_scalar<DataT> d_KLDVal(stream); RAFT_CUDA_TRY(cudaMemsetAsync(d_KLDVal.data(), 0, sizeof(DataT), stream)); raft::linalg::mapThenSumReduce<DataT, KLDOp<DataT>, size_t, 256, const DataT*>( d_KLDVal.data(), (size_t)size, KLDOp<DataT>(), stream, modelPDF, candidatePDF); DataT h_KLDVal; raft::update_host(&h_KLDVal, d_KLDVal.data(), 1, stream); raft::interruptible::synchronize(stream); return h_KLDVal; } }; // end namespace detail }; // end namespace stats }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/histogram.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 <raft/stats/stats_types.hpp> #include <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <raft/util/seive.hpp> #include <raft/util/vectorized.cuh> #include <stdint.h> // This file is a shameless amalgamation of independent works done by // Lars Nyland and Andy Adinets ///@todo: add cub's histogram as another option namespace raft { namespace stats { namespace detail { /** Default mapper which just returns the value of the data itself */ template <typename DataT, typename IdxT> struct IdentityBinner { DI int operator()(DataT val, IdxT row, IdxT col) { return int(val); } }; static const int ThreadsPerBlock = 256; template <typename IdxT, int VecLen> dim3 computeGridDim(IdxT nrows, IdxT ncols, const void* kernel) { int occupancy; RAFT_CUDA_TRY( cudaOccupancyMaxActiveBlocksPerMultiprocessor(&occupancy, kernel, ThreadsPerBlock, 0)); const auto maxBlks = occupancy * raft::getMultiProcessorCount(); int nblksx = raft::ceildiv<int>(VecLen ? nrows / VecLen : nrows, ThreadsPerBlock); // for cases when there aren't a lot of blocks for computing one histogram nblksx = std::min(nblksx, maxBlks); return dim3(nblksx, ncols); } template <typename DataT, typename BinnerOp, typename IdxT, int VecLen, typename CoreOp> DI void histCoreOp(const DataT* data, IdxT nrows, IdxT nbins, BinnerOp binner, CoreOp op, IdxT col) { IdxT offset = col * nrows; auto bdim = IdxT(blockDim.x); IdxT tid = threadIdx.x + bdim * blockIdx.x; tid *= VecLen; IdxT stride = bdim * gridDim.x * VecLen; int nCeil = raft::alignTo<int>(nrows, stride); typedef raft::TxN_t<DataT, VecLen> VecType; VecType a; for (auto i = tid; i < nCeil; i += stride) { if (i < nrows) { a.load(data, offset + i); } #pragma unroll for (int j = 0; j < VecLen; ++j) { int binId = binner(a.val.data[j], i + j, col); op(binId, i + j, col); } } } template <typename DataT, typename BinnerOp, typename IdxT, int VecLen> RAFT_KERNEL gmemHistKernel(int* bins, const DataT* data, IdxT nrows, IdxT nbins, BinnerOp binner) { auto op = [=] __device__(int binId, IdxT row, IdxT col) { if (row >= nrows) return; auto binOffset = col * nbins; #if __CUDA_ARCH__ < 700 raft::myAtomicAdd(bins + binOffset + binId, 1); #else auto amask = __activemask(); auto mask = __match_any_sync(amask, binId); auto leader = __ffs(mask) - 1; if (raft::laneId() == leader) { raft::myAtomicAdd(bins + binOffset + binId, __popc(mask)); } #endif // __CUDA_ARCH__ }; histCoreOp<DataT, BinnerOp, IdxT, VecLen>(data, nrows, nbins, binner, op, blockIdx.y); } template <typename DataT, typename BinnerOp, typename IdxT, int VecLen> void gmemHist(int* bins, IdxT nbins, const DataT* data, IdxT nrows, IdxT ncols, BinnerOp binner, cudaStream_t stream) { auto blks = computeGridDim<IdxT, VecLen>( nrows, ncols, (const void*)gmemHistKernel<DataT, BinnerOp, IdxT, VecLen>); gmemHistKernel<DataT, BinnerOp, IdxT, VecLen> <<<blks, ThreadsPerBlock, 0, stream>>>(bins, data, nrows, nbins, binner); } template <typename DataT, typename BinnerOp, typename IdxT, int VecLen, bool UseMatchAny> RAFT_KERNEL smemHistKernel(int* bins, const DataT* data, IdxT nrows, IdxT nbins, BinnerOp binner) { extern __shared__ unsigned sbins[]; for (auto i = threadIdx.x; i < nbins; i += blockDim.x) { sbins[i] = 0; } __syncthreads(); auto op = [=] __device__(int binId, IdxT row, IdxT col) { if (row >= nrows) return; #if __CUDA_ARCH__ < 700 raft::myAtomicAdd<unsigned int>(sbins + binId, 1); #else if (UseMatchAny) { auto amask = __activemask(); auto mask = __match_any_sync(amask, binId); auto leader = __ffs(mask) - 1; if (raft::laneId() == leader) { raft::myAtomicAdd<unsigned int>(sbins + binId, __popc(mask)); } } else { raft::myAtomicAdd<unsigned int>(sbins + binId, 1); } #endif // __CUDA_ARCH__ }; IdxT col = blockIdx.y; histCoreOp<DataT, BinnerOp, IdxT, VecLen>(data, nrows, nbins, binner, op, col); __syncthreads(); auto binOffset = col * nbins; for (auto i = threadIdx.x; i < nbins; i += blockDim.x) { auto val = sbins[i]; if (val > 0) { raft::myAtomicAdd<unsigned int>((unsigned int*)bins + binOffset + i, val); } } } template <typename DataT, typename BinnerOp, typename IdxT, int VecLen, bool UseMatchAny> void smemHist(int* bins, IdxT nbins, const DataT* data, IdxT nrows, IdxT ncols, BinnerOp binner, cudaStream_t stream) { auto blks = computeGridDim<IdxT, VecLen>( nrows, ncols, (const void*)smemHistKernel<DataT, BinnerOp, IdxT, VecLen, UseMatchAny>); size_t smemSize = nbins * sizeof(unsigned); smemHistKernel<DataT, BinnerOp, IdxT, VecLen, UseMatchAny> <<<blks, ThreadsPerBlock, smemSize, stream>>>(bins, data, nrows, nbins, binner); } template <unsigned _BIN_BITS> struct BitsInfo { static unsigned const BIN_BITS = _BIN_BITS; static unsigned const WORD_BITS = sizeof(unsigned) * 8; static unsigned const WORD_BINS = WORD_BITS / BIN_BITS; static unsigned const BIN_MASK = (1 << BIN_BITS) - 1; }; template <unsigned BIN_BITS> DI void incrementBin(unsigned* sbins, int* bins, int nbins, int binId) { typedef BitsInfo<BIN_BITS> Bits; auto iword = binId / Bits::WORD_BINS; auto ibin = binId % Bits::WORD_BINS; auto sh = ibin * Bits::BIN_BITS; auto old_word = atomicAdd(sbins + iword, unsigned(1 << sh)); auto new_word = old_word + unsigned(1 << sh); if ((new_word >> sh & Bits::BIN_MASK) != 0) return; // overflow raft::myAtomicAdd<unsigned int>((unsigned int*)bins + binId, Bits::BIN_MASK + 1); for (int dbin = 1; ibin + dbin < Bits::WORD_BINS && binId + dbin < nbins; ++dbin) { auto sh1 = (ibin + dbin) * Bits::BIN_BITS; if ((new_word >> sh1 & Bits::BIN_MASK) == 0) { // overflow raft::myAtomicAdd<unsigned int>((unsigned int*)bins + binId + dbin, Bits::BIN_MASK); } else { // correction raft::myAtomicAdd(bins + binId + dbin, -1); break; } } } template <> DI void incrementBin<1>(unsigned* sbins, int* bins, int nbins, int binId) { typedef BitsInfo<1> Bits; auto iword = binId / Bits::WORD_BITS; auto sh = binId % Bits::WORD_BITS; auto old_word = atomicXor(sbins + iword, unsigned(1 << sh)); if ((old_word >> sh & 1) != 0) raft::myAtomicAdd(bins + binId, 2); } template <typename DataT, typename BinnerOp, typename IdxT, int BIN_BITS, int VecLen> RAFT_KERNEL smemBitsHistKernel( int* bins, const DataT* data, IdxT nrows, IdxT nbins, BinnerOp binner) { extern __shared__ unsigned sbins[]; typedef BitsInfo<BIN_BITS> Bits; auto nwords = raft::ceildiv<int>(nbins, Bits::WORD_BINS); for (auto j = threadIdx.x; j < nwords; j += blockDim.x) { sbins[j] = 0; } __syncthreads(); IdxT col = blockIdx.y; IdxT binOffset = col * nbins; auto op = [=] __device__(int binId, IdxT row, IdxT col) { if (row >= nrows) return; incrementBin<Bits::BIN_BITS>(sbins, bins + binOffset, (int)nbins, binId); }; histCoreOp<DataT, BinnerOp, IdxT, VecLen>(data, nrows, nbins, binner, op, col); __syncthreads(); for (auto j = threadIdx.x; j < (int)nbins; j += blockDim.x) { auto shift = j % Bits::WORD_BINS * Bits::BIN_BITS; int count = sbins[j / Bits::WORD_BINS] >> shift & Bits::BIN_MASK; if (count > 0) raft::myAtomicAdd(bins + binOffset + j, count); } } template <typename DataT, typename BinnerOp, typename IdxT, int BIN_BITS, int VecLen> void smemBitsHist(int* bins, IdxT nbins, const DataT* data, IdxT nrows, IdxT ncols, BinnerOp binner, cudaStream_t stream) { typedef BitsInfo<BIN_BITS> Bits; auto blks = computeGridDim<IdxT, VecLen>( nrows, ncols, (const void*)smemBitsHistKernel<DataT, BinnerOp, IdxT, Bits::BIN_BITS, VecLen>); size_t smemSize = raft::ceildiv<size_t>(nbins, Bits::WORD_BITS / Bits::BIN_BITS) * sizeof(int); smemBitsHistKernel<DataT, BinnerOp, IdxT, Bits::BIN_BITS, VecLen> <<<blks, ThreadsPerBlock, smemSize, stream>>>(bins, data, nrows, nbins, binner); } #define INVALID_KEY -1 DI void clearHashTable(int2* ht, int hashSize) { for (auto i = threadIdx.x; i < hashSize; i += blockDim.x) { ht[i] = {INVALID_KEY, 0}; } } DI int findEntry(int2* ht, int hashSize, int binId, int threshold) { int idx = binId % hashSize; int t; int count = 0; while ((t = atomicCAS(&(ht[idx].x), INVALID_KEY, binId)) != INVALID_KEY && t != binId) { ++count; if (count >= threshold) { idx = INVALID_KEY; break; } ++idx; if (idx >= hashSize) { idx = 0; } } return idx; } DI void flushHashTable(int2* ht, int hashSize, int* bins, int nbins, int col) { int binOffset = col * nbins; for (auto i = threadIdx.x; i < hashSize; i += blockDim.x) { if (ht[i].x != INVALID_KEY && ht[i].y > 0) { raft::myAtomicAdd(bins + binOffset + ht[i].x, ht[i].y); } ht[i] = {INVALID_KEY, 0}; } } #undef INVALID_KEY ///@todo: honor VecLen template param template <typename DataT, typename BinnerOp, typename IdxT, int VecLen> RAFT_KERNEL smemHashHistKernel(int* bins, const DataT* data, IdxT nrows, IdxT nbins, BinnerOp binner, int hashSize, int threshold) { extern __shared__ int2 ht[]; int* needFlush = (int*)&(ht[hashSize]); if (threadIdx.x == 0) { needFlush[0] = 0; } clearHashTable(ht, hashSize); __syncthreads(); auto op = [=] __device__(int binId, IdxT row, IdxT col) { bool iNeedFlush = false; if (row < nrows) { int hidx = findEntry(ht, hashSize, binId, threshold); if (hidx >= 0) { raft::myAtomicAdd(&(ht[hidx].y), 1); } else { needFlush[0] = 1; iNeedFlush = true; } } __syncthreads(); if (needFlush[0]) { flushHashTable(ht, hashSize, bins, nbins, col); __syncthreads(); if (threadIdx.x == 0) { needFlush[0] = 0; } __syncthreads(); } if (iNeedFlush) { int hidx = findEntry(ht, hashSize, binId, threshold); // all threads are bound to get one valid entry as all threads in this // block will make forward progress due to the __syncthreads call in the // subsequent iteration raft::myAtomicAdd(&(ht[hidx].y), 1); } }; IdxT col = blockIdx.y; histCoreOp<DataT, BinnerOp, IdxT, VecLen>(data, nrows, nbins, binner, op, col); __syncthreads(); flushHashTable(ht, hashSize, bins, nbins, col); } inline int computeHashTableSize() { // we shouldn't have this much of shared memory available anytime soon! static const unsigned maxBinsEverPossible = 256 * 1024; static raft::common::Seive primes(maxBinsEverPossible); unsigned smem = raft::getSharedMemPerBlock(); // divide-by-2 because hash table entry stores 2 elements: idx and count auto binsPossible = smem / sizeof(unsigned) / 2; for (; binsPossible > 1; --binsPossible) { if (primes.isPrime(binsPossible)) return (int)binsPossible; } return 1; // should not happen! } template <typename DataT, typename BinnerOp, typename IdxT, int VecLen> void smemHashHist(int* bins, IdxT nbins, const DataT* data, IdxT nrows, IdxT ncols, BinnerOp binner, cudaStream_t stream) { static const int flushThreshold = 10; auto blks = computeGridDim<IdxT, 1>( nrows, ncols, (const void*)smemHashHistKernel<DataT, BinnerOp, IdxT, 1>); int hashSize = computeHashTableSize(); size_t smemSize = hashSize * sizeof(int2) + sizeof(int); smemHashHistKernel<DataT, BinnerOp, IdxT, 1><<<blks, ThreadsPerBlock, smemSize, stream>>>( bins, data, nrows, nbins, binner, hashSize, flushThreshold); } template <typename DataT, typename BinnerOp, typename IdxT, int VecLen> void histogramVecLen(HistType type, int* bins, IdxT nbins, const DataT* data, IdxT nrows, IdxT ncols, cudaStream_t stream, BinnerOp binner) { RAFT_CUDA_TRY(cudaMemsetAsync(bins, 0, ncols * nbins * sizeof(int), stream)); switch (type) { case HistTypeGmem: gmemHist<DataT, BinnerOp, IdxT, VecLen>(bins, nbins, data, nrows, ncols, binner, stream); break; case HistTypeSmem: smemHist<DataT, BinnerOp, IdxT, VecLen, false>( bins, nbins, data, nrows, ncols, binner, stream); break; case HistTypeSmemMatchAny: smemHist<DataT, BinnerOp, IdxT, VecLen, true>( bins, nbins, data, nrows, ncols, binner, stream); break; case HistTypeSmemBits16: smemBitsHist<DataT, BinnerOp, IdxT, 16, VecLen>( bins, nbins, data, nrows, ncols, binner, stream); break; case HistTypeSmemBits8: smemBitsHist<DataT, BinnerOp, IdxT, 8, VecLen>( bins, nbins, data, nrows, ncols, binner, stream); break; case HistTypeSmemBits4: smemBitsHist<DataT, BinnerOp, IdxT, 4, VecLen>( bins, nbins, data, nrows, ncols, binner, stream); break; case HistTypeSmemBits2: smemBitsHist<DataT, BinnerOp, IdxT, 2, VecLen>( bins, nbins, data, nrows, ncols, binner, stream); break; case HistTypeSmemBits1: smemBitsHist<DataT, BinnerOp, IdxT, 1, VecLen>( bins, nbins, data, nrows, ncols, binner, stream); break; case HistTypeSmemHash: smemHashHist<DataT, BinnerOp, IdxT, VecLen>(bins, nbins, data, nrows, ncols, binner, stream); break; default: ASSERT(false, "histogram: Invalid type passed '%d'!", type); }; RAFT_CUDA_TRY(cudaGetLastError()); } template <typename DataT, typename BinnerOp, typename IdxT> void histogramImpl(HistType type, int* bins, IdxT nbins, const DataT* data, IdxT nrows, IdxT ncols, cudaStream_t stream, BinnerOp binner) { size_t bytes = nrows * sizeof(DataT); if (nrows <= 0) return; if (16 % sizeof(DataT) == 0 && bytes % 16 == 0) { histogramVecLen<DataT, BinnerOp, IdxT, 16 / sizeof(DataT)>( type, bins, nbins, data, nrows, ncols, stream, binner); } else if (8 % sizeof(DataT) == 0 && bytes % 8 == 0) { histogramVecLen<DataT, BinnerOp, IdxT, 8 / sizeof(DataT)>( type, bins, nbins, data, nrows, ncols, stream, binner); } else if (4 % sizeof(DataT) == 0 && bytes % 4 == 0) { histogramVecLen<DataT, BinnerOp, IdxT, 4 / sizeof(DataT)>( type, bins, nbins, data, nrows, ncols, stream, binner); } else if (2 % sizeof(DataT) == 0 && bytes % 2 == 0) { histogramVecLen<DataT, BinnerOp, IdxT, 2 / sizeof(DataT)>( type, bins, nbins, data, nrows, ncols, stream, binner); } else { histogramVecLen<DataT, BinnerOp, IdxT, 1>( type, bins, nbins, data, nrows, ncols, stream, binner); } } template <typename IdxT> HistType selectBestHistAlgo(IdxT nbins) { size_t smem = raft::getSharedMemPerBlock(); size_t requiredSize = nbins * sizeof(unsigned); if (requiredSize <= smem) { return HistTypeSmem; } for (int bits = 16; bits >= 1; bits >>= 1) { auto nBytesForBins = raft::ceildiv<size_t>(bits * nbins, 8); requiredSize = raft::alignTo<size_t>(nBytesForBins, sizeof(unsigned)); if (requiredSize <= smem) { return static_cast<HistType>(bits); } } return HistTypeGmem; } /** * @brief Perform histogram on the input data. It chooses the right load size * based on the input data vector length. It also supports large-bin cases * using a specialized smem-based hashing technique. * @tparam DataT input data type * @tparam IdxT data type used to compute indices * @tparam BinnerOp takes the input data and computes its bin index * @param type histogram implementation type to choose * @param bins the output bins (length = ncols * nbins) * @param nbins number of bins * @param data input data (length = ncols * nrows) * @param nrows data array length in each column (or batch) * @param ncols number of columns (or batch size) * @param stream cuda stream * @param binner the operation that computes the bin index of the input data * * @note signature of BinnerOp is `int func(DataT, IdxT);` */ template <typename DataT, typename IdxT = int, typename BinnerOp = IdentityBinner<DataT, IdxT>> void histogram(HistType type, int* bins, IdxT nbins, const DataT* data, IdxT nrows, IdxT ncols, cudaStream_t stream, BinnerOp binner = IdentityBinner<DataT, IdxT>()) { HistType computedType = type; if (type == HistTypeAuto) { computedType = selectBestHistAlgo(nbins); } histogramImpl<DataT, BinnerOp, IdxT>( computedType, bins, nbins, data, nrows, ncols, stream, binner); } }; // end namespace detail }; // end namespace stats }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/weighted_mean.cuh
/* * Copyright (c) 2019-2022, 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/reduce.cuh> #include <raft/stats/sum.cuh> #include <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> namespace raft { namespace stats { namespace detail { /** * @brief Compute the row-wise weighted mean of the input matrix with a * vector of weights * * @tparam Type the data type * @tparam IdxType Integer type used to for addressing * @param mu the output mean vector * @param data the input matrix * @param weights weight of size D if along_row is true, else of size N * @param D number of columns of data * @param N number of rows of data * @param row_major data input matrix is row-major or not * @param along_rows whether to reduce along rows or columns * @param stream cuda stream to launch work on */ template <typename Type, typename IdxType = int> void weightedMean(Type* mu, const Type* data, const Type* weights, IdxType D, IdxType N, bool row_major, bool along_rows, cudaStream_t stream) { // sum the weights & copy back to CPU auto weight_size = along_rows ? D : N; Type WS = 0; raft::stats::sum(mu, weights, (IdxType)1, weight_size, false, stream); raft::update_host(&WS, mu, 1, stream); raft::linalg::reduce( mu, data, D, N, (Type)0, row_major, along_rows, stream, false, [weights] __device__(Type v, IdxType i) { return v * weights[i]; }, raft::add_op{}, raft::div_const_op<Type>(WS)); } }; // end namespace detail }; // end namespace stats }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/homogeneity_score.cuh
/* * Copyright (c) 2019-2022, 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 homogeneity_score.cuh * * @brief A clustering result satisfies homogeneity if all of its clusters * contain only data points which are members of a single class. */ #pragma once #include <raft/stats/entropy.cuh> #include <raft/stats/mutual_info_score.cuh> namespace raft { namespace stats { namespace detail { /** * @brief Function to calculate the homogeneity score between two clusters * <a href="https://en.wikipedia.org/wiki/Homogeneity_(statistics)">more info on mutual * information</a> * @param truthClusterArray: the array of truth classes of type T * @param predClusterArray: the array of predicted classes of type T * @param size: the size of the data points of type int * @param lowerLabelRange: the lower bound of the range of labels * @param upperLabelRange: the upper bound of the range of labels * @param stream: the cudaStream object */ template <typename T> double homogeneity_score(const T* truthClusterArray, const T* predClusterArray, int size, T lowerLabelRange, T upperLabelRange, cudaStream_t stream) { if (size == 0) return 1.0; double computedMI, computedEntropy; computedMI = raft::stats::mutual_info_score( truthClusterArray, predClusterArray, size, lowerLabelRange, upperLabelRange, stream); computedEntropy = raft::stats::entropy(truthClusterArray, size, lowerLabelRange, upperLabelRange, stream); double homogeneity; if (computedEntropy) { homogeneity = computedMI / computedEntropy; } else homogeneity = 1.0; return homogeneity; } }; // end namespace detail }; // end namespace stats }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/stddev.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/linalg/binary_op.cuh> #include <raft/util/cuda_utils.cuh> #include <cub/cub.cuh> namespace raft { namespace stats { namespace detail { ///@todo: ColPerBlk has been tested only for 32! template <typename Type, typename IdxType, int TPB, int ColsPerBlk = 32> RAFT_KERNEL stddevKernelRowMajor(Type* std, const Type* data, IdxType D, IdxType N) { const int RowsPerBlkPerIter = TPB / ColsPerBlk; IdxType thisColId = threadIdx.x % ColsPerBlk; IdxType thisRowId = threadIdx.x / ColsPerBlk; IdxType colId = thisColId + ((IdxType)blockIdx.y * ColsPerBlk); IdxType rowId = thisRowId + ((IdxType)blockIdx.x * RowsPerBlkPerIter); Type thread_data = Type(0); const IdxType stride = RowsPerBlkPerIter * gridDim.x; for (IdxType i = rowId; i < N; i += stride) { Type val = (colId < D) ? data[i * D + colId] : Type(0); thread_data += val * val; } __shared__ Type sstd[ColsPerBlk]; if (threadIdx.x < ColsPerBlk) sstd[threadIdx.x] = Type(0); __syncthreads(); raft::myAtomicAdd(sstd + thisColId, thread_data); __syncthreads(); if (threadIdx.x < ColsPerBlk) raft::myAtomicAdd(std + colId, sstd[thisColId]); } template <typename Type, typename IdxType, int TPB> RAFT_KERNEL stddevKernelColMajor(Type* std, const Type* data, const Type* mu, IdxType D, IdxType N) { typedef cub::BlockReduce<Type, TPB> BlockReduce; __shared__ typename BlockReduce::TempStorage temp_storage; Type thread_data = Type(0); IdxType colStart = N * blockIdx.x; Type m = mu[blockIdx.x]; for (IdxType i = threadIdx.x; i < N; i += TPB) { IdxType idx = colStart + i; Type diff = data[idx] - m; thread_data += diff * diff; } Type acc = BlockReduce(temp_storage).Sum(thread_data); if (threadIdx.x == 0) { std[blockIdx.x] = raft::sqrt(acc / N); } } template <typename Type, typename IdxType, int TPB> RAFT_KERNEL varsKernelColMajor(Type* var, const Type* data, const Type* mu, IdxType D, IdxType N) { typedef cub::BlockReduce<Type, TPB> BlockReduce; __shared__ typename BlockReduce::TempStorage temp_storage; Type thread_data = Type(0); IdxType colStart = N * blockIdx.x; Type m = mu[blockIdx.x]; for (IdxType i = threadIdx.x; i < N; i += TPB) { IdxType idx = colStart + i; Type diff = data[idx] - m; thread_data += diff * diff; } Type acc = BlockReduce(temp_storage).Sum(thread_data); if (threadIdx.x == 0) { var[blockIdx.x] = acc / N; } } /** * @brief Compute stddev of the input matrix * * Stddev operation is assumed to be performed on a given column. * * @tparam Type the data type * @tparam IdxType Integer type used to for addressing * @param std the output stddev vector * @param data the input matrix * @param mu the mean vector * @param D number of columns of data * @param N number of rows of data * @param sample whether to evaluate sample stddev or not. In other words, * whether * to normalize the output using N-1 or N, for true or false, respectively * @param rowMajor whether the input data is row or col major * @param stream cuda stream where to launch work */ template <typename Type, typename IdxType = int> void stddev(Type* std, const Type* data, const Type* mu, IdxType D, IdxType N, bool sample, bool rowMajor, cudaStream_t stream) { static const int TPB = 256; if (rowMajor) { static const int RowsPerThread = 4; static const int ColsPerBlk = 32; static const int RowsPerBlk = (TPB / ColsPerBlk) * RowsPerThread; dim3 grid(raft::ceildiv(N, (IdxType)RowsPerBlk), raft::ceildiv(D, (IdxType)ColsPerBlk)); RAFT_CUDA_TRY(cudaMemset(std, 0, sizeof(Type) * D)); stddevKernelRowMajor<Type, IdxType, TPB, ColsPerBlk><<<grid, TPB, 0, stream>>>(std, data, D, N); Type ratio = Type(1) / (sample ? Type(N - 1) : Type(N)); raft::linalg::binaryOp( std, std, mu, D, [ratio] __device__(Type a, Type b) { return raft::sqrt(a * ratio - b * b); }, stream); } else { stddevKernelColMajor<Type, IdxType, TPB><<<D, TPB, 0, stream>>>(std, data, mu, D, N); } RAFT_CUDA_TRY(cudaPeekAtLastError()); } /** * @brief Compute variance of the input matrix * * Variance operation is assumed to be performed on a given column. * * @tparam Type the data type * @tparam IdxType Integer type used to for addressing * @param var the output stddev vector * @param data the input matrix * @param mu the mean vector * @param D number of columns of data * @param N number of rows of data * @param sample whether to evaluate sample stddev or not. In other words, * whether * to normalize the output using N-1 or N, for true or false, respectively * @param rowMajor whether the input data is row or col major * @param stream cuda stream where to launch work */ template <typename Type, typename IdxType = int> void vars(Type* var, const Type* data, const Type* mu, IdxType D, IdxType N, bool sample, bool rowMajor, cudaStream_t stream) { static const int TPB = 256; if (rowMajor) { static const int RowsPerThread = 4; static const int ColsPerBlk = 32; static const int RowsPerBlk = (TPB / ColsPerBlk) * RowsPerThread; dim3 grid(raft::ceildiv(N, (IdxType)RowsPerBlk), raft::ceildiv(D, (IdxType)ColsPerBlk)); RAFT_CUDA_TRY(cudaMemset(var, 0, sizeof(Type) * D)); stddevKernelRowMajor<Type, IdxType, TPB, ColsPerBlk><<<grid, TPB, 0, stream>>>(var, data, D, N); Type ratio = Type(1) / (sample ? Type(N - 1) : Type(N)); raft::linalg::binaryOp( var, var, mu, D, [ratio] __device__(Type a, Type b) { return a * ratio - b * b; }, stream); } else { varsKernelColMajor<Type, IdxType, TPB><<<D, TPB, 0, stream>>>(var, data, mu, D, N); } RAFT_CUDA_TRY(cudaPeekAtLastError()); } } // namespace detail } // namespace stats } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/mean_center.cuh
/* * Copyright (c) 2022, 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/matrix_vector_op.cuh> #include <raft/util/cuda_utils.cuh> #include <raft/util/vectorized.cuh> namespace raft { namespace stats { namespace detail { /** * @brief Center the input matrix wrt its mean * @tparam Type the data type * @tparam IdxType Integer type used to for addressing * @tparam TPB threads per block of the cuda kernel launched * @param out the output mean-centered matrix * @param data input matrix * @param mu the mean vector * @param D number of columns of data * @param N number of rows of data * @param rowMajor whether input is row or col major * @param bcastAlongRows whether to broadcast vector along rows or columns * @param stream cuda stream where to launch work */ template <typename Type, typename IdxType = int, int TPB = 256> void meanCenter(Type* out, const Type* data, const Type* mu, IdxType D, IdxType N, bool rowMajor, bool bcastAlongRows, cudaStream_t stream) { raft::linalg::matrixVectorOp( out, data, mu, D, N, rowMajor, bcastAlongRows, raft::sub_op{}, stream); } /** * @brief Add the input matrix wrt its mean * @tparam Type the data type * @tparam IdxType Integer type used to for addressing * @tparam TPB threads per block of the cuda kernel launched * @param out the output mean-added matrix * @param data input matrix * @param mu the mean vector * @param D number of columns of data * @param N number of rows of data * @param rowMajor whether input is row or col major * @param bcastAlongRows whether to broadcast vector along rows or columns * @param stream cuda stream where to launch work */ template <typename Type, typename IdxType = int, int TPB = 256> void meanAdd(Type* out, const Type* data, const Type* mu, IdxType D, IdxType N, bool rowMajor, bool bcastAlongRows, cudaStream_t stream) { raft::linalg::matrixVectorOp( out, data, mu, D, N, rowMajor, bcastAlongRows, raft::add_op{}, stream); } }; // end namespace detail }; // end namespace stats }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/entropy.cuh
/* * Copyright (c) 2019-2022, 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 entropy.cuh * @brief Calculates the entropy for a labeling in nats.(ie, uses natural logarithm for the * calculations) */ #pragma once #include <cub/cub.cuh> #include <math.h> #include <raft/linalg/divide.cuh> #include <raft/linalg/map_then_reduce.cuh> #include <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <rmm/device_scalar.hpp> #include <rmm/device_uvector.hpp> namespace raft { namespace stats { namespace detail { /** * @brief Lambda to calculate the entropy of a sample given its probability value * * @param p: the input to the functional mapping * @param q: dummy param */ struct entropyOp { HDI double operator()(double p, double q) { if (p) return -1 * (p) * (log(p)); else return 0.0; } }; /** * @brief function to calculate the bincounts of number of samples in every label * * @tparam LabelT: type of the labels * @param labels: the pointer to the array containing labels for every data sample * @param binCountArray: pointer to the 1D array that contains the count of samples per cluster * @param nRows: number of data samples * @param lowerLabelRange * @param upperLabelRange * @param workspace: device buffer containing workspace memory * @param stream: the cuda stream where to launch this kernel */ template <typename LabelT> void countLabels(const LabelT* labels, double* binCountArray, int nRows, LabelT lowerLabelRange, LabelT upperLabelRange, rmm::device_uvector<char>& workspace, cudaStream_t stream) { int num_levels = upperLabelRange - lowerLabelRange + 2; LabelT lower_level = lowerLabelRange; LabelT upper_level = upperLabelRange + 1; size_t temp_storage_bytes = 0; RAFT_CUDA_TRY(cub::DeviceHistogram::HistogramEven(nullptr, temp_storage_bytes, labels, binCountArray, num_levels, lower_level, upper_level, nRows, stream)); workspace.resize(temp_storage_bytes, stream); RAFT_CUDA_TRY(cub::DeviceHistogram::HistogramEven(workspace.data(), temp_storage_bytes, labels, binCountArray, num_levels, lower_level, upper_level, nRows, stream)); } /** * @brief Function to calculate entropy * <a href="https://en.wikipedia.org/wiki/Entropy_(information_theory)">more info on entropy</a> * * @param clusterArray: the array of classes of type T * @param size: the size of the data points of type int * @param lowerLabelRange: the lower bound of the range of labels * @param upperLabelRange: the upper bound of the range of labels * @param stream: the cudaStream object * @return the entropy score */ template <typename T> double entropy(const T* clusterArray, const int size, const T lowerLabelRange, const T upperLabelRange, cudaStream_t stream) { if (!size) return 1.0; T numUniqueClasses = upperLabelRange - lowerLabelRange + 1; // declaring, allocating and initializing memory for bincount array and entropy values rmm::device_uvector<double> prob(numUniqueClasses, stream); RAFT_CUDA_TRY(cudaMemsetAsync(prob.data(), 0, numUniqueClasses * sizeof(double), stream)); rmm::device_scalar<double> d_entropy(stream); RAFT_CUDA_TRY(cudaMemsetAsync(d_entropy.data(), 0, sizeof(double), stream)); // workspace allocation rmm::device_uvector<char> workspace(1, stream); // calculating the bincounts and populating the prob array countLabels(clusterArray, prob.data(), size, lowerLabelRange, upperLabelRange, workspace, stream); // scalar dividing by size raft::linalg::divideScalar<double>( prob.data(), prob.data(), (double)size, numUniqueClasses, stream); // calculating the aggregate entropy raft::linalg::mapThenSumReduce<double, entropyOp>( d_entropy.data(), numUniqueClasses, entropyOp(), stream, prob.data(), prob.data()); // updating in the host memory double h_entropy; raft::update_host(&h_entropy, d_entropy.data(), 1, stream); raft::interruptible::synchronize(stream); return h_entropy; } }; // end namespace detail }; // end namespace stats }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/neighborhood_recall.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 <cstddef> #include <raft/core/device_mdspan.hpp> #include <raft/core/error.hpp> #include <raft/core/host_mdspan.hpp> #include <raft/core/math.hpp> #include <raft/core/mdspan_types.hpp> #include <raft/core/operators.hpp> #include <raft/core/resources.hpp> #include <cub/cub.cuh> #include <cuda/atomic> #include <optional> namespace raft::stats::detail { template <typename IndicesValueType, typename DistanceValueType, typename IndexType, typename ScalarType> RAFT_KERNEL neighborhood_recall( raft::device_matrix_view<const IndicesValueType, IndexType, raft::row_major> indices, raft::device_matrix_view<const IndicesValueType, IndexType, raft::row_major> ref_indices, std::optional<raft::device_matrix_view<const DistanceValueType, IndexType, raft::row_major>> distances, std::optional<raft::device_matrix_view<const DistanceValueType, IndexType, raft::row_major>> ref_distances, raft::device_scalar_view<ScalarType> recall_score, DistanceValueType const eps) { auto constexpr kThreadsPerBlock = 32; IndexType const row_idx = blockIdx.x; auto const lane_idx = threadIdx.x % kThreadsPerBlock; // Each warp stores a recall score computed across the columns per row IndexType thread_recall_score = 0; for (IndexType col_idx = lane_idx; col_idx < indices.extent(1); col_idx += kThreadsPerBlock) { for (IndexType ref_col_idx = 0; ref_col_idx < ref_indices.extent(1); ref_col_idx++) { if (indices(row_idx, col_idx) == ref_indices(row_idx, ref_col_idx)) { thread_recall_score += 1; break; } else if (distances.has_value()) { auto dist = distances.value()(row_idx, col_idx); auto ref_dist = ref_distances.value()(row_idx, ref_col_idx); DistanceValueType diff = raft::abs(dist - ref_dist); DistanceValueType m = std::max(raft::abs(dist), raft::abs(ref_dist)); DistanceValueType ratio = diff > eps ? diff / m : diff; if (ratio <= eps) { thread_recall_score += 1; break; } } } } // Reduce across a warp for row score typedef cub::BlockReduce<IndexType, kThreadsPerBlock> BlockReduce; __shared__ typename BlockReduce::TempStorage temp_storage; ScalarType row_recall_score = BlockReduce(temp_storage).Sum(thread_recall_score); // Reduce across all rows for global score if (lane_idx == 0) { cuda::atomic_ref<ScalarType, cuda::thread_scope_device> device_recall_score{ *recall_score.data_handle()}; std::size_t const total_count = indices.extent(0) * indices.extent(1); device_recall_score.fetch_add(row_recall_score / total_count); } } template <typename IndicesValueType, typename DistanceValueType, typename IndexType, typename ScalarType> void neighborhood_recall( raft::resources const& res, raft::device_matrix_view<const IndicesValueType, IndexType, raft::row_major> indices, raft::device_matrix_view<const IndicesValueType, IndexType, raft::row_major> ref_indices, std::optional<raft::device_matrix_view<const DistanceValueType, IndexType, raft::row_major>> distances, std::optional<raft::device_matrix_view<const DistanceValueType, IndexType, raft::row_major>> ref_distances, raft::device_scalar_view<ScalarType> recall_score, DistanceValueType const eps) { // One warp per row, launch a warp-width block per-row kernel auto constexpr kThreadsPerBlock = 32; auto const num_blocks = indices.extent(0); neighborhood_recall<<<num_blocks, kThreadsPerBlock>>>( indices, ref_indices, distances, ref_distances, recall_score, eps); } } // end namespace raft::stats::detail
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/silhouette_score.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 <algorithm> #include <cub/cub.cuh> #include <iostream> #include <math.h> #include <numeric> #include <raft/core/operators.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/distance/distance.cuh> #include <raft/distance/distance_types.hpp> #include <raft/linalg/add.cuh> #include <raft/linalg/eltwise.cuh> #include <raft/linalg/map_then_reduce.cuh> #include <raft/linalg/matrix_vector_op.cuh> #include <raft/linalg/reduce.cuh> #include <raft/linalg/reduce_cols_by_key.cuh> #include <raft/util/cuda_utils.cuh> #include <rmm/device_scalar.hpp> namespace raft { namespace stats { namespace detail { /** * @brief kernel that calculates the average intra-cluster distance for every sample data point and * updates the cluster distance to max value * @tparam DataT: type of the data samples * @tparam LabelT: type of the labels * @param sampleToClusterSumOfDistances: the pointer to the 2D array that contains the sum of * distances from every sample to every cluster (nRows x nLabels) * @param binCountArray: pointer to the 1D array that contains the count of samples per cluster (1 x * nLabels) * @param d_aArray: the pointer to the array of average intra-cluster distances for every sample in * device memory (1 x nRows) * @param labels: the pointer to the array containing labels for every data sample (1 x nRows) * @param nRows: number of data samples * @param nLabels: number of Labels * @param MAX_VAL: DataT specific upper limit */ template <typename DataT, typename LabelT> RAFT_KERNEL populateAKernel(DataT* sampleToClusterSumOfDistances, DataT* binCountArray, DataT* d_aArray, const LabelT* labels, int nRows, int nLabels, const DataT MAX_VAL) { // getting the current index int sampleIndex = threadIdx.x + blockIdx.x * blockDim.x; if (sampleIndex >= nRows) return; // sampleDistanceVector is an array that stores that particular row of the distanceMatrix DataT* sampleToClusterSumOfDistancesVector = &sampleToClusterSumOfDistances[sampleIndex * nLabels]; LabelT sampleCluster = labels[sampleIndex]; int sampleClusterIndex = (int)sampleCluster; if (binCountArray[sampleClusterIndex] - 1 <= 0) { d_aArray[sampleIndex] = -1; return; } else { d_aArray[sampleIndex] = (sampleToClusterSumOfDistancesVector[sampleClusterIndex]) / (binCountArray[sampleClusterIndex] - 1); // modifying the sampleDistanceVector to give sample average distance sampleToClusterSumOfDistancesVector[sampleClusterIndex] = MAX_VAL; } } /** * @brief function to calculate the bincounts of number of samples in every label * @tparam DataT: type of the data samples * @tparam LabelT: type of the labels * @param labels: the pointer to the array containing labels for every data sample (1 x nRows) * @param binCountArray: pointer to the 1D array that contains the count of samples per cluster (1 x * nLabels) * @param nRows: number of data samples * @param nUniqueLabels: number of Labels * @param workspace: device buffer containing workspace memory * @param stream: the cuda stream where to launch this kernel */ template <typename DataT, typename LabelT> void countLabels(const LabelT* labels, DataT* binCountArray, int nRows, int nUniqueLabels, rmm::device_uvector<char>& workspace, cudaStream_t stream) { int num_levels = nUniqueLabels + 1; LabelT lower_level = 0; LabelT upper_level = nUniqueLabels; size_t temp_storage_bytes = 0; rmm::device_uvector<int> countArray(nUniqueLabels, stream); RAFT_CUDA_TRY(cub::DeviceHistogram::HistogramEven(nullptr, temp_storage_bytes, labels, binCountArray, num_levels, lower_level, upper_level, nRows, stream)); workspace.resize(temp_storage_bytes, stream); RAFT_CUDA_TRY(cub::DeviceHistogram::HistogramEven(workspace.data(), temp_storage_bytes, labels, binCountArray, num_levels, lower_level, upper_level, nRows, stream)); } /** * @brief structure that defines the division Lambda for elementwise op */ template <typename DataT> struct DivOp { HDI DataT operator()(DataT a, int b, int c) { if (b == 0) return ULLONG_MAX; else return a / b; } }; /** * @brief structure that defines the elementwise operation to calculate silhouette score using * params 'a' and 'b' */ template <typename DataT> struct SilOp { HDI DataT operator()(DataT a, DataT b) { if (a == 0 && b == 0 || a == b) return 0; else if (a == -1) return 0; else if (a > b) return (b - a) / a; else return (b - a) / b; } }; /** * @brief main function that returns the average silhouette score for a given set of data and its * clusterings * @tparam DataT: type of the data samples * @tparam LabelT: type of the labels * @param X_in: pointer to the input Data samples array (nRows x nCols) * @param nRows: number of data samples * @param nCols: number of features * @param labels: the pointer to the array containing labels for every data sample (1 x nRows) * @param nLabels: number of Labels * @param silhouette_scorePerSample: pointer to the array that is optionally taken in as input and * is populated with the silhouette score for every sample (1 x nRows) * @param stream: the cuda stream where to launch this kernel * @param metric: the numerical value that maps to the type of distance metric to be used in the * calculations */ template <typename DataT, typename LabelT> DataT silhouette_score( raft::resources const& handle, const DataT* X_in, int nRows, int nCols, const LabelT* labels, int nLabels, DataT* silhouette_scorePerSample, cudaStream_t stream, raft::distance::DistanceType metric = raft::distance::DistanceType::L2Unexpanded) { ASSERT(nLabels >= 2 && nLabels <= (nRows - 1), "silhouette Score not defined for the given number of labels!"); // compute the distance matrix rmm::device_uvector<DataT> distanceMatrix(nRows * nRows, stream); rmm::device_uvector<char> workspace(1, stream); raft::distance::pairwise_distance( handle, X_in, X_in, distanceMatrix.data(), nRows, nRows, nCols, metric); // deciding on the array of silhouette scores for each dataPoint rmm::device_uvector<DataT> silhouette_scoreSamples(0, stream); DataT* perSampleSilScore = nullptr; if (silhouette_scorePerSample == nullptr) { silhouette_scoreSamples.resize(nRows, stream); perSampleSilScore = silhouette_scoreSamples.data(); } else { perSampleSilScore = silhouette_scorePerSample; } RAFT_CUDA_TRY(cudaMemsetAsync(perSampleSilScore, 0, nRows * sizeof(DataT), stream)); // getting the sample count per cluster rmm::device_uvector<DataT> binCountArray(nLabels, stream); RAFT_CUDA_TRY(cudaMemsetAsync(binCountArray.data(), 0, nLabels * sizeof(DataT), stream)); countLabels(labels, binCountArray.data(), nRows, nLabels, workspace, stream); // calculating the sample-cluster-distance-sum-array rmm::device_uvector<DataT> sampleToClusterSumOfDistances(nRows * nLabels, stream); RAFT_CUDA_TRY(cudaMemsetAsync( sampleToClusterSumOfDistances.data(), 0, nRows * nLabels * sizeof(DataT), stream)); raft::linalg::reduce_cols_by_key(distanceMatrix.data(), labels, sampleToClusterSumOfDistances.data(), nRows, nRows, nLabels, stream); // creating the a array and b array rmm::device_uvector<DataT> d_aArray(nRows, stream); rmm::device_uvector<DataT> d_bArray(nRows, stream); RAFT_CUDA_TRY(cudaMemsetAsync(d_aArray.data(), 0, nRows * sizeof(DataT), stream)); RAFT_CUDA_TRY(cudaMemsetAsync(d_bArray.data(), 0, nRows * sizeof(DataT), stream)); // kernel that populates the d_aArray // kernel configuration dim3 numThreadsPerBlock(32, 1, 1); dim3 numBlocks(raft::ceildiv<int>(nRows, numThreadsPerBlock.x), 1, 1); // calling the kernel populateAKernel<<<numBlocks, numThreadsPerBlock, 0, stream>>>( sampleToClusterSumOfDistances.data(), binCountArray.data(), d_aArray.data(), labels, nRows, nLabels, std::numeric_limits<DataT>::max()); // elementwise dividing by bincounts rmm::device_uvector<DataT> averageDistanceBetweenSampleAndCluster(nRows * nLabels, stream); RAFT_CUDA_TRY(cudaMemsetAsync( averageDistanceBetweenSampleAndCluster.data(), 0, nRows * nLabels * sizeof(DataT), stream)); raft::linalg::matrixVectorOp(averageDistanceBetweenSampleAndCluster.data(), sampleToClusterSumOfDistances.data(), binCountArray.data(), binCountArray.data(), nLabels, nRows, true, true, DivOp<DataT>(), stream); // calculating row-wise minimum raft::linalg::reduce<DataT, DataT, int, raft::identity_op, raft::min_op>( d_bArray.data(), averageDistanceBetweenSampleAndCluster.data(), nLabels, nRows, std::numeric_limits<DataT>::max(), true, true, stream, false, raft::identity_op{}, raft::min_op{}); // calculating the silhouette score per sample using the d_aArray and d_bArray raft::linalg::binaryOp<DataT, SilOp<DataT>>( perSampleSilScore, d_aArray.data(), d_bArray.data(), nRows, SilOp<DataT>(), stream); // calculating the sum of all the silhouette score rmm::device_scalar<DataT> d_avgSilhouetteScore(stream); RAFT_CUDA_TRY(cudaMemsetAsync(d_avgSilhouetteScore.data(), 0, sizeof(DataT), stream)); raft::linalg::mapThenSumReduce<double, raft::identity_op>(d_avgSilhouetteScore.data(), nRows, raft::identity_op(), stream, perSampleSilScore, perSampleSilScore); DataT avgSilhouetteScore = d_avgSilhouetteScore.value(stream); resource::sync_stream(handle, stream); avgSilhouetteScore /= nRows; return avgSilhouetteScore; } }; // namespace detail }; // namespace stats }; // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/dispersion.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 <cub/cub.cuh> #include <memory> #include <raft/core/interruptible.hpp> #include <raft/linalg/eltwise.cuh> #include <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <rmm/device_uvector.hpp> namespace raft { namespace stats { namespace detail { ///@todo: ColsPerBlk has been tested only for 32! template <typename DataT, typename IdxT, int TPB, int ColsPerBlk = 32> RAFT_KERNEL weightedMeanKernel(DataT* mu, const DataT* data, const IdxT* counts, IdxT D, IdxT N) { constexpr int RowsPerBlkPerIter = TPB / ColsPerBlk; IdxT thisColId = threadIdx.x % ColsPerBlk; IdxT thisRowId = threadIdx.x / ColsPerBlk; IdxT colId = thisColId + ((IdxT)blockIdx.y * ColsPerBlk); IdxT rowId = thisRowId + ((IdxT)blockIdx.x * RowsPerBlkPerIter); DataT thread_data = DataT(0); const IdxT stride = RowsPerBlkPerIter * gridDim.x; __shared__ DataT smu[ColsPerBlk]; if (threadIdx.x < ColsPerBlk) smu[threadIdx.x] = DataT(0); for (IdxT i = rowId; i < N; i += stride) { thread_data += (colId < D) ? data[i * D + colId] * (DataT)counts[i] : DataT(0); } __syncthreads(); raft::myAtomicAdd(smu + thisColId, thread_data); __syncthreads(); if (threadIdx.x < ColsPerBlk && colId < D) raft::myAtomicAdd(mu + colId, smu[thisColId]); } template <typename DataT, typename IdxT, int TPB> RAFT_KERNEL dispersionKernel(DataT* result, const DataT* clusters, const IdxT* clusterSizes, const DataT* mu, IdxT dim, IdxT nClusters) { IdxT tid = threadIdx.x + blockIdx.x * blockDim.x; IdxT len = dim * nClusters; IdxT stride = blockDim.x * gridDim.x; DataT sum = DataT(0); for (; tid < len; tid += stride) { IdxT col = tid % dim; IdxT row = tid / dim; DataT diff = clusters[tid] - mu[col]; sum += diff * diff * DataT(clusterSizes[row]); } typedef cub::BlockReduce<DataT, TPB> BlockReduce; __shared__ typename BlockReduce::TempStorage temp_storage; __syncthreads(); auto acc = BlockReduce(temp_storage).Sum(sum); __syncthreads(); if (threadIdx.x == 0) raft::myAtomicAdd(result, acc); } /** * @brief Compute cluster dispersion metric. This is very useful for * automatically finding the 'k' (in kmeans) that improves this metric. * @tparam DataT data type * @tparam IdxT index type * @tparam TPB threads block for kernels launched * @param centroids the cluster centroids. This is assumed to be row-major * and of dimension (nClusters x dim) * @param clusterSizes number of points in the dataset which belong to each * cluster. This is of length nClusters * @param globalCentroid compute the global weighted centroid of all cluster * centroids. This is of length dim. Pass a nullptr if this is not needed * @param nClusters number of clusters * @param nPoints number of points in the dataset * @param dim dataset dimensionality * @param stream cuda stream * @return the cluster dispersion value */ template <typename DataT, typename IdxT = int, int TPB = 256> DataT dispersion(const DataT* centroids, const IdxT* clusterSizes, DataT* globalCentroid, IdxT nClusters, IdxT nPoints, IdxT dim, cudaStream_t stream) { static const int RowsPerThread = 4; static const int ColsPerBlk = 32; static const int RowsPerBlk = (TPB / ColsPerBlk) * RowsPerThread; dim3 grid(raft::ceildiv(nPoints, (IdxT)RowsPerBlk), raft::ceildiv(dim, (IdxT)ColsPerBlk)); rmm::device_uvector<DataT> mean(0, stream); rmm::device_uvector<DataT> result(1, stream); DataT* mu = globalCentroid; if (globalCentroid == nullptr) { mean.resize(dim, stream); mu = mean.data(); } RAFT_CUDA_TRY(cudaMemsetAsync(mu, 0, sizeof(DataT) * dim, stream)); RAFT_CUDA_TRY(cudaMemsetAsync(result.data(), 0, sizeof(DataT), stream)); weightedMeanKernel<DataT, IdxT, TPB, ColsPerBlk> <<<grid, TPB, 0, stream>>>(mu, centroids, clusterSizes, dim, nClusters); RAFT_CUDA_TRY(cudaGetLastError()); DataT ratio = DataT(1) / DataT(nPoints); raft::linalg::scalarMultiply(mu, mu, ratio, dim, stream); // finally, compute the dispersion constexpr int ItemsPerThread = 4; int nblks = raft::ceildiv<int>(dim * nClusters, TPB * ItemsPerThread); dispersionKernel<DataT, IdxT, TPB> <<<nblks, TPB, 0, stream>>>(result.data(), centroids, clusterSizes, mu, dim, nClusters); RAFT_CUDA_TRY(cudaGetLastError()); DataT h_result; raft::update_host(&h_result, result.data(), 1, stream); raft::interruptible::synchronize(stream); return sqrt(h_result); } } // end namespace detail } // end namespace stats } // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/sum.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/linalg/eltwise.cuh> #include <raft/util/cuda_utils.cuh> #include <cub/cub.cuh> namespace raft { namespace stats { namespace detail { ///@todo: ColsPerBlk has been tested only for 32! template <typename Type, typename IdxType, int TPB, int ColsPerBlk = 32> RAFT_KERNEL sumKernelRowMajor(Type* mu, const Type* data, IdxType D, IdxType N) { const int RowsPerBlkPerIter = TPB / ColsPerBlk; IdxType thisColId = threadIdx.x % ColsPerBlk; IdxType thisRowId = threadIdx.x / ColsPerBlk; IdxType colId = thisColId + ((IdxType)blockIdx.y * ColsPerBlk); IdxType rowId = thisRowId + ((IdxType)blockIdx.x * RowsPerBlkPerIter); Type thread_data = Type(0); const IdxType stride = RowsPerBlkPerIter * gridDim.x; for (IdxType i = rowId; i < N; i += stride) thread_data += (colId < D) ? data[i * D + colId] : Type(0); __shared__ Type smu[ColsPerBlk]; if (threadIdx.x < ColsPerBlk) smu[threadIdx.x] = Type(0); __syncthreads(); raft::myAtomicAdd(smu + thisColId, thread_data); __syncthreads(); if (threadIdx.x < ColsPerBlk) raft::myAtomicAdd(mu + colId, smu[thisColId]); } template <typename Type, typename IdxType, int TPB> RAFT_KERNEL sumKernelColMajor(Type* mu, const Type* data, IdxType D, IdxType N) { typedef cub::BlockReduce<Type, TPB> BlockReduce; __shared__ typename BlockReduce::TempStorage temp_storage; Type thread_data = Type(0); IdxType colStart = N * blockIdx.x; for (IdxType i = threadIdx.x; i < N; i += TPB) { IdxType idx = colStart + i; thread_data += data[idx]; } Type acc = BlockReduce(temp_storage).Sum(thread_data); if (threadIdx.x == 0) { mu[blockIdx.x] = acc; } } template <typename Type, typename IdxType = int> void sum(Type* output, const Type* input, IdxType D, IdxType N, bool rowMajor, cudaStream_t stream) { static const int TPB = 256; if (rowMajor) { static const int RowsPerThread = 4; static const int ColsPerBlk = 32; static const int RowsPerBlk = (TPB / ColsPerBlk) * RowsPerThread; dim3 grid(raft::ceildiv(N, (IdxType)RowsPerBlk), raft::ceildiv(D, (IdxType)ColsPerBlk)); RAFT_CUDA_TRY(cudaMemset(output, 0, sizeof(Type) * D)); sumKernelRowMajor<Type, IdxType, TPB, ColsPerBlk> <<<grid, TPB, 0, stream>>>(output, input, D, N); } else { sumKernelColMajor<Type, IdxType, TPB><<<D, TPB, 0, stream>>>(output, input, D, N); } RAFT_CUDA_TRY(cudaPeekAtLastError()); } } // namespace detail } // namespace stats } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/mutual_info_score.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. */ /** * @file mutual_info_score.cuh * @brief The Mutual Information is a measure of the similarity between two labels of * the same data.This metric is independent of the absolute values of the labels: * a permutation of the class or cluster label values won't change the * score value in any way. * This metric is furthermore symmetric.This can be useful to * measure the agreement of two independent label assignments strategies * on the same dataset when the real ground truth is not known. */ #pragma once #include <cub/cub.cuh> #include <math.h> #include <raft/core/interruptible.hpp> #include <raft/linalg/reduce.cuh> #include <raft/stats/contingency_matrix.cuh> #include <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <rmm/device_scalar.hpp> #include <rmm/device_uvector.hpp> namespace raft { namespace stats { namespace detail { /** * @brief kernel to calculate the mutual info score * @param dContingencyMatrix: the contingency matrix corresponding to the two clusters * @param a: the row wise sum of the contingency matrix, which is also the bin counts of first * cluster array * @param b: the column wise sum of the contingency matrix, which is also the bin counts of second * cluster array * @param numUniqueClasses: number of unique classes * @param size: the size of array a and b (size of the contingency matrix is (size x size)) * @param d_MI: pointer to the device memory that stores the aggregate mutual information */ template <typename T, int BLOCK_DIM_X, int BLOCK_DIM_Y> RAFT_KERNEL mutual_info_kernel(const int* dContingencyMatrix, const int* a, const int* b, int numUniqueClasses, int size, double* d_MI) { // calculating the indices of pairs of datapoints compared by the current thread int j = threadIdx.x + blockIdx.x * blockDim.x; int i = threadIdx.y + blockIdx.y * blockDim.y; // thread-local variable to count the mutual info double localMI = 0.0; if (i < numUniqueClasses && j < numUniqueClasses && a[i] * b[j] != 0 && dContingencyMatrix[i * numUniqueClasses + j] != 0) { localMI += (double(dContingencyMatrix[i * numUniqueClasses + j])) * (log(double(size) * double(dContingencyMatrix[i * numUniqueClasses + j])) - log(double(a[i] * b[j]))); } // specialize blockReduce for a 2D block of 1024 threads of type uint64_t typedef cub::BlockReduce<double, BLOCK_DIM_X, cub::BLOCK_REDUCE_WARP_REDUCTIONS, BLOCK_DIM_Y> BlockReduce; // Allocate shared memory for blockReduce __shared__ typename BlockReduce::TempStorage temp_storage; // summing up thread-local counts specific to a block localMI = BlockReduce(temp_storage).Sum(localMI); __syncthreads(); // executed once per block if (threadIdx.x == 0 && threadIdx.y == 0) { raft::myAtomicAdd(d_MI, localMI); } } /** * @brief Function to calculate the mutual information between two clusters * <a href="https://en.wikipedia.org/wiki/Mutual_information">more info on mutual information</a> * @param firstClusterArray: the array of classes of type T * @param secondClusterArray: the array of classes of type T * @param size: the size of the data points of type int * @param lowerLabelRange: the lower bound of the range of labels * @param upperLabelRange: the upper bound of the range of labels * @param stream: the cudaStream object */ template <typename T> double mutual_info_score(const T* firstClusterArray, const T* secondClusterArray, int size, T lowerLabelRange, T upperLabelRange, cudaStream_t stream) { int numUniqueClasses = upperLabelRange - lowerLabelRange + 1; // declaring, allocating and initializing memory for the contingency marix rmm::device_uvector<int> dContingencyMatrix(numUniqueClasses * numUniqueClasses, stream); RAFT_CUDA_TRY(cudaMemsetAsync( dContingencyMatrix.data(), 0, numUniqueClasses * numUniqueClasses * sizeof(int), stream)); // workspace allocation size_t workspaceSz = raft::stats::getContingencyMatrixWorkspaceSize( size, firstClusterArray, stream, lowerLabelRange, upperLabelRange); rmm::device_uvector<char> pWorkspace(workspaceSz, stream); // calculating the contingency matrix raft::stats::contingencyMatrix(firstClusterArray, secondClusterArray, (int)size, (int*)dContingencyMatrix.data(), stream, (void*)pWorkspace.data(), workspaceSz, lowerLabelRange, upperLabelRange); // creating device buffers for all the parameters involved in ARI calculation // device variables rmm::device_uvector<int> a(numUniqueClasses, stream); rmm::device_uvector<int> b(numUniqueClasses, stream); rmm::device_scalar<double> d_MI(stream); // host variables double h_MI; // initializing device memory RAFT_CUDA_TRY(cudaMemsetAsync(a.data(), 0, numUniqueClasses * sizeof(int), stream)); RAFT_CUDA_TRY(cudaMemsetAsync(b.data(), 0, numUniqueClasses * sizeof(int), stream)); RAFT_CUDA_TRY(cudaMemsetAsync(d_MI.data(), 0, sizeof(double), stream)); // calculating the row-wise sums raft::linalg::reduce<int, int, int>( a.data(), dContingencyMatrix.data(), numUniqueClasses, numUniqueClasses, 0, true, true, stream); // calculating the column-wise sums raft::linalg::reduce<int, int, int>(b.data(), dContingencyMatrix.data(), numUniqueClasses, numUniqueClasses, 0, true, false, stream); // kernel configuration static const int BLOCK_DIM_Y = 16, BLOCK_DIM_X = 16; dim3 numThreadsPerBlock(BLOCK_DIM_X, BLOCK_DIM_Y); dim3 numBlocks(raft::ceildiv<int>(numUniqueClasses, numThreadsPerBlock.x), raft::ceildiv<int>(numUniqueClasses, numThreadsPerBlock.y)); // calling the kernel mutual_info_kernel<T, BLOCK_DIM_X, BLOCK_DIM_Y><<<numBlocks, numThreadsPerBlock, 0, stream>>>( dContingencyMatrix.data(), a.data(), b.data(), numUniqueClasses, size, d_MI.data()); // updating in the host memory h_MI = d_MI.value(stream); raft::interruptible::synchronize(stream); return h_MI / size; } }; // end namespace detail }; // end namespace stats }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/minmax.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 <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <limits> namespace raft { namespace stats { namespace detail { // TODO: replace with `std::bitcast` once we adopt C++20 or libcu++ adds it template <class To, class From> constexpr To bit_cast(const From& from) noexcept { To to{}; static_assert(sizeof(To) == sizeof(From)); memcpy(&to, &from, sizeof(To)); return to; } template <typename T> struct encode_traits {}; template <> struct encode_traits<float> { using E = int; }; template <> struct encode_traits<double> { using E = long long; }; HDI int encode(float val) { int i = detail::bit_cast<int>(val); return i >= 0 ? i : (1 << 31) | ~i; } HDI long long encode(double val) { std::int64_t i = detail::bit_cast<std::int64_t>(val); return i >= 0 ? i : (1ULL << 63) | ~i; } HDI float decode(int val) { if (val < 0) val = (1 << 31) | ~val; return detail::bit_cast<float>(val); } HDI double decode(long long val) { if (val < 0) val = (1ULL << 63) | ~val; return detail::bit_cast<double>(val); } template <typename T, typename E> DI T atomicMaxBits(T* address, T val) { E old = atomicMax((E*)address, encode(val)); return decode(old); } template <typename T, typename E> DI T atomicMinBits(T* address, T val) { E old = atomicMin((E*)address, encode(val)); return decode(old); } template <typename T, typename E> RAFT_KERNEL decodeKernel(T* globalmin, T* globalmax, int ncols) { int tid = threadIdx.x + blockIdx.x * blockDim.x; if (tid < ncols) { globalmin[tid] = decode(*(E*)&globalmin[tid]); globalmax[tid] = decode(*(E*)&globalmax[tid]); } } ///@todo: implement a proper "fill" kernel template <typename T, typename E> RAFT_KERNEL minmaxInitKernel(int ncols, T* globalmin, T* globalmax, T init_val) { int tid = threadIdx.x + blockIdx.x * blockDim.x; if (tid >= ncols) return; *(E*)&globalmin[tid] = encode(init_val); *(E*)&globalmax[tid] = encode(-init_val); } template <typename T, typename E> RAFT_KERNEL minmaxKernel(const T* data, const unsigned int* rowids, const unsigned int* colids, int nrows, int ncols, int row_stride, T* g_min, T* g_max, T* sampledcols, T init_min_val, int batch_ncols, int num_batches) { int tid = threadIdx.x + blockIdx.x * blockDim.x; extern __shared__ char shmem[]; T* s_min = (T*)shmem; T* s_max = (T*)(shmem + sizeof(T) * batch_ncols); int last_batch_ncols = ncols % batch_ncols; if (last_batch_ncols == 0) { last_batch_ncols = batch_ncols; } int orig_batch_ncols = batch_ncols; for (int batch_id = 0; batch_id < num_batches; batch_id++) { if (batch_id == num_batches - 1) { batch_ncols = last_batch_ncols; } for (int i = threadIdx.x; i < batch_ncols; i += blockDim.x) { *(E*)&s_min[i] = encode(init_min_val); *(E*)&s_max[i] = encode(-init_min_val); } __syncthreads(); for (int i = tid; i < nrows * batch_ncols; i += blockDim.x * gridDim.x) { int col = (batch_id * orig_batch_ncols) + (i / nrows); int row = i % nrows; if (colids != nullptr) { col = colids[col]; } if (rowids != nullptr) { row = rowids[row]; } int index = row + col * row_stride; T coldata = data[index]; if (!isnan(coldata)) { // Min max values are saved in shared memory and global memory as per the shuffled colids. atomicMinBits<T, E>(&s_min[(int)(i / nrows)], coldata); atomicMaxBits<T, E>(&s_max[(int)(i / nrows)], coldata); } if (sampledcols != nullptr) { sampledcols[batch_id * orig_batch_ncols + i] = coldata; } } __syncthreads(); // finally, perform global mem atomics for (int j = threadIdx.x; j < batch_ncols; j += blockDim.x) { atomicMinBits<T, E>(&g_min[batch_id * orig_batch_ncols + j], decode(*(E*)&s_min[j])); atomicMaxBits<T, E>(&g_max[batch_id * orig_batch_ncols + j], decode(*(E*)&s_max[j])); } __syncthreads(); } } /** * @brief Computes min/max across every column of the input matrix, as well as * optionally allow to subsample based on the given row/col ID mapping vectors * * @tparam T the data type * @tparam TPB number of threads per block * @param data input data * @param rowids actual row ID mappings. It is of length nrows. If you want to * skip this index lookup entirely, pass nullptr * @param colids actual col ID mappings. It is of length ncols. If you want to * skip this index lookup entirely, pass nullptr * @param nrows number of rows of data to be worked upon. The actual rows of the * input "data" can be bigger than this! * @param ncols number of cols of data to be worked upon. The actual cols of the * input "data" can be bigger than this! * @param row_stride stride (in number of elements) between 2 adjacent columns * @param globalmin final col-wise global minimum (size = ncols) * @param globalmax final col-wise global maximum (size = ncols) * @param sampledcols output sampled data. Pass nullptr if you don't need this * @param stream cuda stream * @note This method makes the following assumptions: * 1. input and output matrices are assumed to be col-major * 2. ncols is small enough to fit the whole of min/max values across all cols * in shared memory */ template <typename T, int TPB = 512> void minmax(const T* data, const unsigned* rowids, const unsigned* colids, int nrows, int ncols, int row_stride, T* globalmin, T* globalmax, T* sampledcols, cudaStream_t stream) { using E = typename encode_traits<T>::E; int nblks = raft::ceildiv(ncols, TPB); T init_val = std::numeric_limits<T>::max(); minmaxInitKernel<T, E><<<nblks, TPB, 0, stream>>>(ncols, globalmin, globalmax, init_val); RAFT_CUDA_TRY(cudaPeekAtLastError()); nblks = raft::ceildiv(nrows * ncols, TPB); nblks = min(nblks, 65536); size_t smemSize = sizeof(T) * 2 * ncols; // Compute the batch_ncols, in [1, ncols] range, that meet the available // shared memory constraints. auto smemPerBlk = raft::getSharedMemPerBlock(); int batch_ncols = min(ncols, (int)(smemPerBlk / (sizeof(T) * 2))); int num_batches = raft::ceildiv(ncols, batch_ncols); smemSize = sizeof(T) * 2 * batch_ncols; minmaxKernel<T, E><<<nblks, TPB, smemSize, stream>>>(data, rowids, colids, nrows, ncols, row_stride, globalmin, globalmax, sampledcols, init_val, batch_ncols, num_batches); RAFT_CUDA_TRY(cudaPeekAtLastError()); decodeKernel<T, E><<<nblks, TPB, 0, stream>>>(globalmin, globalmax, ncols); RAFT_CUDA_TRY(cudaPeekAtLastError()); } }; // end namespace detail }; // end namespace stats }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/scores.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 <memory> #include <raft/distance/distance.cuh> #include <raft/linalg/eltwise.cuh> #include <raft/linalg/power.cuh> #include <raft/linalg/subtract.cuh> #include <raft/spatial/knn/knn.cuh> #include <raft/stats/mean.cuh> #include <raft/util/cudart_utils.hpp> #include <rmm/device_scalar.hpp> #include <rmm/device_uvector.hpp> #include <thrust/count.h> #include <thrust/device_ptr.h> #include <thrust/execution_policy.h> #include <thrust/reduce.h> #define N_THREADS 512 namespace raft { namespace stats { namespace detail { /** * Calculates the "Coefficient of Determination" (R-Squared) score * normalizing the sum of squared errors by the total sum of squares. * * This score indicates the proportionate amount of variation in an * expected response variable is explained by the independent variables * in a linear regression model. The larger the R-squared value, the * more variability is explained by the linear regression model. * * @param y: Array of ground-truth response variables * @param y_hat: Array of predicted response variables * @param n: Number of elements in y and y_hat * @param stream: cuda stream * @return: The R-squared value. */ template <typename math_t> math_t r2_score(math_t* y, math_t* y_hat, int n, cudaStream_t stream) { rmm::device_scalar<math_t> y_bar(stream); raft::stats::mean(y_bar.data(), y, 1, n, false, false, stream); RAFT_CUDA_TRY(cudaPeekAtLastError()); rmm::device_uvector<math_t> sse_arr(n, stream); raft::linalg::eltwiseSub(sse_arr.data(), y, y_hat, n, stream); raft::linalg::powerScalar(sse_arr.data(), sse_arr.data(), math_t(2.0), n, stream); RAFT_CUDA_TRY(cudaPeekAtLastError()); rmm::device_uvector<math_t> ssto_arr(n, stream); raft::linalg::subtractDevScalar(ssto_arr.data(), y, y_bar.data(), n, stream); raft::linalg::powerScalar(ssto_arr.data(), ssto_arr.data(), math_t(2.0), n, stream); RAFT_CUDA_TRY(cudaPeekAtLastError()); thrust::device_ptr<math_t> d_sse = thrust::device_pointer_cast(sse_arr.data()); thrust::device_ptr<math_t> d_ssto = thrust::device_pointer_cast(ssto_arr.data()); math_t sse = thrust::reduce(thrust::cuda::par.on(stream), d_sse, d_sse + n); math_t ssto = thrust::reduce(thrust::cuda::par.on(stream), d_ssto, d_ssto + n); return 1.0 - sse / ssto; } /** * @brief Compute accuracy of predictions. Useful for classification. * @tparam math_t: data type for predictions (e.g., int for classification) * @param[in] predictions: array of predictions (GPU pointer). * @param[in] ref_predictions: array of reference (ground-truth) predictions (GPU pointer). * @param[in] n: number of elements in each of predictions, ref_predictions. * @param[in] stream: cuda stream. * @return: Accuracy score in [0, 1]; higher is better. */ template <typename math_t> float accuracy_score(const math_t* predictions, const math_t* ref_predictions, int n, cudaStream_t stream) { unsigned long long correctly_predicted = 0ULL; rmm::device_uvector<math_t> diffs_array(n, stream); // TODO could write a kernel instead raft::linalg::eltwiseSub(diffs_array.data(), predictions, ref_predictions, n, stream); RAFT_CUDA_TRY(cudaGetLastError()); correctly_predicted = thrust::count(thrust::cuda::par.on(stream), diffs_array.data(), diffs_array.data() + n, 0); float accuracy = correctly_predicted * 1.0f / n; return accuracy; } template <typename T> RAFT_KERNEL reg_metrics_kernel( const T* predictions, const T* ref_predictions, int n, double* abs_diffs, double* tmp_sums) { int tid = threadIdx.x + blockIdx.x * blockDim.x; __shared__ double shmem[2]; // {abs_difference_sum, squared difference sum} for (int i = threadIdx.x; i < 2; i += blockDim.x) { shmem[i] = 0; } __syncthreads(); for (int i = tid; i < n; i += blockDim.x * gridDim.x) { double diff = predictions[i] - ref_predictions[i]; double abs_diff = abs(diff); raft::myAtomicAdd(&shmem[0], abs_diff); raft::myAtomicAdd(&shmem[1], diff * diff); // update absolute difference in global memory for subsequent abs. median computation abs_diffs[i] = abs_diff; } __syncthreads(); // Update tmp_sum w/ total abs_difference_sum and squared difference sum. for (int i = threadIdx.x; i < 2; i += blockDim.x) { raft::myAtomicAdd(&tmp_sums[i], shmem[i]); } } /** * @brief Compute regression metrics mean absolute error, mean squared error, median absolute error * @tparam T: data type for predictions (e.g., float or double for regression). * @param[in] predictions: array of predictions (GPU pointer). * @param[in] ref_predictions: array of reference (ground-truth) predictions (GPU pointer). * @param[in] n: number of elements in each of predictions, ref_predictions. Should be > 0. * @param[in] stream: cuda stream. * @param[out] mean_abs_error: Mean Absolute Error. Sum over n of (|predictions[i] - * ref_predictions[i]|) / n. * @param[out] mean_squared_error: Mean Squared Error. Sum over n of ((predictions[i] - * ref_predictions[i])^2) / n. * @param[out] median_abs_error: Median Absolute Error. Median of |predictions[i] - * ref_predictions[i]| for i in [0, n). */ template <typename T> void regression_metrics(const T* predictions, const T* ref_predictions, int n, cudaStream_t stream, double& mean_abs_error, double& mean_squared_error, double& median_abs_error) { std::vector<double> mean_errors(2); std::vector<double> h_sorted_abs_diffs(n); int thread_cnt = 256; int block_cnt = raft::ceildiv(n, thread_cnt); int array_size = n * sizeof(double); rmm::device_uvector<double> abs_diffs_array(array_size, stream); rmm::device_uvector<double> sorted_abs_diffs(array_size, stream); rmm::device_uvector<double> tmp_sums(2 * sizeof(double), stream); RAFT_CUDA_TRY(cudaMemsetAsync(tmp_sums.data(), 0, 2 * sizeof(double), stream)); reg_metrics_kernel<T><<<block_cnt, thread_cnt, 0, stream>>>( predictions, ref_predictions, n, abs_diffs_array.data(), tmp_sums.data()); RAFT_CUDA_TRY(cudaGetLastError()); raft::update_host(&mean_errors[0], tmp_sums.data(), 2, stream); raft::interruptible::synchronize(stream); mean_abs_error = mean_errors[0] / n; mean_squared_error = mean_errors[1] / n; // Compute median error. Sort diffs_array and pick median value char* temp_storage = nullptr; size_t temp_storage_bytes; RAFT_CUDA_TRY(cub::DeviceRadixSort::SortKeys((void*)temp_storage, temp_storage_bytes, abs_diffs_array.data(), sorted_abs_diffs.data(), n, 0, 8 * sizeof(double), stream)); rmm::device_uvector<char> temp_storage_v(temp_storage_bytes, stream); temp_storage = temp_storage_v.data(); RAFT_CUDA_TRY(cub::DeviceRadixSort::SortKeys((void*)temp_storage, temp_storage_bytes, abs_diffs_array.data(), sorted_abs_diffs.data(), n, 0, 8 * sizeof(double), stream)); raft::update_host(h_sorted_abs_diffs.data(), sorted_abs_diffs.data(), n, stream); raft::interruptible::synchronize(stream); int middle = n / 2; if (n % 2 == 1) { median_abs_error = h_sorted_abs_diffs[middle]; } else { median_abs_error = (h_sorted_abs_diffs[middle] + h_sorted_abs_diffs[middle - 1]) / 2; } } } // namespace detail } // namespace stats } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/meanvar.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/reduce.cuh> #include <raft/util/cuda_utils.cuh> namespace raft::stats::detail { template <typename T> class mean_var { private: T w; T m; T s; public: /** Monoidal neutral. */ HDI mean_var() : w(0.0), m(0.0), s(0.0) {} /** Lift a single value. */ HDI explicit mean_var(T x) : w(1.0), m(x), s(0.0) {} /** * Monoidal binary op: combine means and vars of two sets. * (associative and commutative) */ friend HDI auto operator+(mean_var<T> a, mean_var<T> const& b) -> mean_var<T> { a += b; return a; } /** * Combine means and vars of two sets. * * Similar to: * https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Parallel_algorithm */ HDI auto operator+=(mean_var<T> const& b) & -> mean_var<T>& { mean_var<T>& a(*this); T cw = a.w + b.w; if (cw == 0) return a; T aw_frac = a.w / cw; T bw_frac = b.w / cw; a.w = cw; T d = a.m - b.m; a.s += b.s + cw * (d * aw_frac) * (d * bw_frac); a.m = a.m * aw_frac + b.m * bw_frac; return a; } /** Get the computed mean. */ HDI auto mean() const -> T { return m; } /** * @brief Get the computed variance. * * @param [in] sample whether to produce sample variance (divide by `N - 1` instead of `N`). * @return variance */ HDI auto var(bool sample) const -> T { return s / max(T(1.0), sample ? w - T(1.0) : w); } HDI void load(volatile mean_var<T>* address) { this->m = address->m; this->s = address->s; this->w = address->w; } HDI void store(volatile mean_var<T>* address) { address->m = this->m; address->s = this->s; address->w = this->w; } }; /* NB: current implementation here is not optimal, especially the rowmajor version; leaving this for further work (perhaps, as a more generic "linewiseReduce"). Vectorized loads/stores could speed things up a lot. */ /** * meanvar kernel - row-major version * * Assumptions: * * 1. blockDim.x == WarpSize * 2. Dimension X goes along columns (D) * 3. Dimension Y goes along rows (N) * * * @tparam T element type * @tparam I indexing type * @tparam BlockSize must be equal to blockDim.x * blockDim.y * blockDim.z * @param data input data * @param mvs meanvars -- output * @param locks guards for updating meanvars * @param len total length of input data (N * D) * @param D number of columns in the input data. */ template <typename T, typename I, int BlockSize> RAFT_KERNEL __launch_bounds__(BlockSize) meanvar_kernel_rowmajor(const T* data, volatile mean_var<T>* mvs, int* locks, I len, I D) { // read the data const I col = threadIdx.x + blockDim.x * blockIdx.x; mean_var<T> thread_data; if (col < D) { const I step = D * blockDim.y * gridDim.y; for (I i = col + D * (threadIdx.y + blockDim.y * blockIdx.y); i < len; i += step) { thread_data += mean_var<T>(data[i]); } } // aggregate within block if (blockDim.y > 1) { __shared__ uint8_t shm_bytes[BlockSize * sizeof(mean_var<T>)]; auto shm = (mean_var<T>*)shm_bytes; int tid = threadIdx.x + threadIdx.y * blockDim.x; shm[tid] = thread_data; for (int bs = BlockSize >> 1; bs >= blockDim.x; bs = bs >> 1) { __syncthreads(); if (tid < bs) { shm[tid] += shm[tid + bs]; } } thread_data = shm[tid]; } // aggregate across blocks if (threadIdx.y == 0) { int* lock = locks + blockIdx.x; if (threadIdx.x == 0 && col < D) { while (atomicCAS(lock, 0, 1) == 1) { __threadfence(); } } __syncthreads(); if (col < D) { __threadfence(); mean_var<T> global_data; global_data.load(mvs + col); global_data += thread_data; global_data.store(mvs + col); __threadfence(); } __syncthreads(); if (threadIdx.x == 0 && col < D) { __stwt(lock, 0); } } } template <typename T, typename I, int BlockSize> RAFT_KERNEL __launch_bounds__(BlockSize) meanvar_kernel_colmajor(T* mean, T* var, const T* data, I D, I N, bool sample) { using BlockReduce = cub::BlockReduce<mean_var<T>, BlockSize>; __shared__ typename BlockReduce::TempStorage shm; const T* block_data = data + N * blockIdx.x; mean_var<T> thread_data; for (I i = threadIdx.x; i < N; i += BlockSize) { thread_data += mean_var<T>(block_data[i]); } mean_var<T> acc = BlockReduce(shm).Sum(thread_data); if (threadIdx.x == 0) { mean[blockIdx.x] = acc.mean(); var[blockIdx.x] = acc.var(sample); } } template <typename T, typename I> RAFT_KERNEL meanvar_kernel_fill(T* mean, T* var, const mean_var<T>* aggr, I D, bool sample) { I i = threadIdx.x + blockDim.x * blockIdx.x; if (i >= D) return; auto x = aggr[i]; mean[i] = x.mean(); var[i] = x.var(sample); } template <typename T, typename I = int, int BlockSize = 256> void meanvar( T* mean, T* var, const T* data, I D, I N, bool sample, bool rowMajor, cudaStream_t stream) { if (rowMajor) { static_assert(BlockSize >= WarpSize, "Block size must be not smaller than the warp size."); const dim3 bs(WarpSize, BlockSize / WarpSize, 1); dim3 gs(raft::ceildiv<decltype(bs.x)>(D, bs.x), raft::ceildiv<decltype(bs.y)>(N, bs.y), 1); // Don't create more blocks than necessary to occupy the GPU int occupancy; RAFT_CUDA_TRY(cudaOccupancyMaxActiveBlocksPerMultiprocessor( &occupancy, meanvar_kernel_rowmajor<T, I, BlockSize>, BlockSize, 0)); gs.y = std::min(gs.y, raft::ceildiv<decltype(gs.y)>(occupancy * getMultiProcessorCount(), gs.x)); // Global memory: one mean_var<T> for each column // one lock per all blocks working on the same set of columns rmm::device_buffer buf(sizeof(mean_var<T>) * D + sizeof(int) * gs.x, stream); RAFT_CUDA_TRY(cudaMemsetAsync(buf.data(), 0, buf.size(), stream)); mean_var<T>* mvs = static_cast<mean_var<T>*>(buf.data()); int* locks = static_cast<int*>(static_cast<void*>(mvs + D)); const uint64_t len = uint64_t(D) * uint64_t(N); ASSERT(len <= uint64_t(std::numeric_limits<I>::max()), "N * D does not fit the indexing type"); meanvar_kernel_rowmajor<T, I, BlockSize><<<gs, bs, 0, stream>>>(data, mvs, locks, len, D); meanvar_kernel_fill<T, I> <<<raft::ceildiv<I>(D, BlockSize), BlockSize, 0, stream>>>(mean, var, mvs, D, sample); } else { meanvar_kernel_colmajor<T, I, BlockSize> <<<D, BlockSize, 0, stream>>>(mean, var, data, D, N, sample); } RAFT_CHECK_CUDA(stream); } }; // namespace raft::stats::detail
0
rapidsai_public_repos/raft/cpp/include/raft/stats/detail
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/batched/information_criterion.cuh
/* * Copyright (c) 2019-2022, 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/unary_op.cuh> #include <raft/stats/stats_types.hpp> #include <cmath> namespace raft { namespace stats { namespace batched { namespace detail { /** * Compute the given type of information criterion * * @note: it is safe to do the computation in-place (i.e give same pointer * as input and output) * * @param[out] d_ic Information criterion to be returned for each * series (device) * @param[in] d_loglikelihood Log-likelihood for each series (device) * @param[in] ic_type Type of criterion to compute. See IC_Type * @param[in] n_params Number of parameters in the model * @param[in] batch_size Number of series in the batch * @param[in] n_samples Number of samples in each series * @param[in] stream CUDA stream */ template <typename ScalarT, typename IdxT> void information_criterion(ScalarT* d_ic, const ScalarT* d_loglikelihood, IC_Type ic_type, IdxT n_params, IdxT batch_size, IdxT n_samples, cudaStream_t stream) { ScalarT ic_base{}; ScalarT N = static_cast<ScalarT>(n_params); ScalarT T = static_cast<ScalarT>(n_samples); switch (ic_type) { case AIC: ic_base = (ScalarT)2.0 * N; break; case AICc: ic_base = (ScalarT)2.0 * (N + (N * (N + (ScalarT)1.0)) / (T - N - (ScalarT)1.0)); break; case BIC: ic_base = std::log(T) * N; break; } /* Compute information criterion from log-likelihood and base term */ raft::linalg::unaryOp( d_ic, d_loglikelihood, batch_size, [=] __device__(ScalarT loglike) { return ic_base - (ScalarT)2.0 * loglike; }, stream); } } // namespace detail } // namespace batched } // namespace stats } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/stats/detail
rapidsai_public_repos/raft/cpp/include/raft/stats/detail/batched/silhouette_score.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 "../silhouette_score.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resource/cuda_stream_pool.hpp> #include <raft/core/resource/thrust_policy.hpp> #include <raft/util/cuda_utils.cuh> #include <raft/util/device_atomics.cuh> #include <rmm/device_uvector.hpp> #include <rmm/exec_policy.hpp> #include <thrust/device_vector.h> #include <thrust/fill.h> #include <thrust/reduce.h> namespace raft { namespace stats { namespace batched { namespace detail { /** * This kernel initializes matrix b (n_rows * n_labels) * For each label that the corresponding row is not a part of is initialized as 0 * If the corresponding row is the only sample in its label, again 0 * Only if the there are > 1 samples in the label, row is initialized to max */ template <typename value_t, typename value_idx, typename label_idx> RAFT_KERNEL fill_b_kernel(value_t* b, const label_idx* y, value_idx n_rows, label_idx n_labels, const value_idx* cluster_counts) { value_idx idx = threadIdx.x + blockIdx.x * blockDim.x; label_idx idy = threadIdx.y + blockIdx.y * blockDim.y; if (idx >= n_rows || idy >= n_labels) { return; } auto row_cluster = y[idx]; auto col_cluster_count = cluster_counts[idy]; // b for own cluster should be max value // so that it does not interfere with min operator // b is also max if col cluster count is 0 // however, b is 0 if self cluster count is 1 if (row_cluster == idy || col_cluster_count == 0) { if (cluster_counts[row_cluster] == 1) { b[idx * n_labels + idy] = 0; } else { b[idx * n_labels + idy] = std::numeric_limits<value_t>::max(); } } else { b[idx * n_labels + idy] = 0; } } /** * This kernel does an elementwise sweep of chunked pairwise distance matrix * By knowing the offsets of the chunked pairwise distance matrix in the * global pairwise distance matrix, we are able to calculate * intermediate values of a and b for the rows and columns present in the * current chunked pairwise distance matrix. */ template <typename value_t, typename value_idx, typename label_idx> RAFT_KERNEL compute_chunked_a_b_kernel(value_t* a, value_t* b, value_idx row_offset, value_idx col_offset, const label_idx* y, label_idx n_labels, const value_idx* cluster_counts, const value_t* distances, value_idx dist_rows, value_idx dist_cols) { value_idx row_id = threadIdx.x + blockIdx.x * blockDim.x; value_idx col_id = threadIdx.y + blockIdx.y * blockDim.y; // these are global offsets of current element // in the full pairwise distance matrix value_idx pw_row_id = row_id + row_offset; value_idx pw_col_id = col_id + col_offset; if (row_id >= dist_rows || col_id >= dist_cols || pw_row_id == pw_col_id) { return; } auto row_cluster = y[pw_row_id]; if (cluster_counts[row_cluster] == 1) { return; } auto col_cluster = y[pw_col_id]; auto col_cluster_counts = cluster_counts[col_cluster]; if (col_cluster == row_cluster) { atomicAdd(&a[pw_row_id], distances[row_id * dist_cols + col_id] / (col_cluster_counts - 1)); } else { atomicAdd(&b[pw_row_id * n_labels + col_cluster], distances[row_id * dist_cols + col_id] / col_cluster_counts); } } template <typename value_idx, typename label_idx> rmm::device_uvector<value_idx> get_cluster_counts(raft::resources const& handle, const label_idx* y, value_idx& n_rows, label_idx& n_labels) { auto stream = resource::get_cuda_stream(handle); rmm::device_uvector<value_idx> cluster_counts(n_labels, stream); rmm::device_uvector<char> workspace(1, stream); raft::stats::detail::countLabels(y, cluster_counts.data(), n_rows, n_labels, workspace, stream); return cluster_counts; } template <typename value_t, typename value_idx> rmm::device_uvector<value_t> get_pairwise_distance(raft::resources const& handle, const value_t* left_begin, const value_t* right_begin, value_idx& n_left_rows, value_idx& n_right_rows, value_idx& n_cols, raft::distance::DistanceType metric, cudaStream_t stream) { rmm::device_uvector<value_t> distances(n_left_rows * n_right_rows, stream); raft::distance::pairwise_distance( handle, left_begin, right_begin, distances.data(), n_left_rows, n_right_rows, n_cols, metric); return distances; } template <typename value_t, typename value_idx, typename label_idx> void compute_chunked_a_b(raft::resources const& handle, value_t* a, value_t* b, value_idx& row_offset, value_idx& col_offset, const label_idx* y, label_idx& n_labels, const value_idx* cluster_counts, const value_t* distances, value_idx& dist_rows, value_idx& dist_cols, cudaStream_t stream) { dim3 block_size(std::min(dist_rows, 32), std::min(dist_cols, 32)); dim3 grid_size(raft::ceildiv(dist_rows, (value_idx)block_size.x), raft::ceildiv(dist_cols, (value_idx)block_size.y)); detail::compute_chunked_a_b_kernel<<<grid_size, block_size, 0, stream>>>( a, b, row_offset, col_offset, y, n_labels, cluster_counts, distances, dist_rows, dist_cols); } template <typename value_t, typename value_idx, typename label_idx> value_t silhouette_score( raft::resources const& handle, const value_t* X, value_idx n_rows, value_idx n_cols, const label_idx* y, label_idx n_labels, value_t* scores, value_idx chunk, raft::distance::DistanceType metric = raft::distance::DistanceType::L2Unexpanded) { ASSERT(n_labels >= 2 && n_labels <= (n_rows - 1), "silhouette Score not defined for the given number of labels!"); rmm::device_uvector<value_idx> cluster_counts = get_cluster_counts(handle, y, n_rows, n_labels); auto stream = resource::get_cuda_stream(handle); auto policy = resource::get_thrust_policy(handle); auto b_size = n_rows * n_labels; value_t *a_ptr, *b_ptr; rmm::device_uvector<value_t> a(0, stream); rmm::device_uvector<value_t> b(b_size, stream); b_ptr = b.data(); // since a and silhouette score per sample are same size, reusing if (scores == nullptr || scores == NULL) { a.resize(n_rows, stream); a_ptr = a.data(); } else { a_ptr = scores; } thrust::fill(policy, a_ptr, a_ptr + n_rows, 0); dim3 block_size(std::min(n_rows, 32), std::min(n_labels, 32)); dim3 grid_size(raft::ceildiv(n_rows, (value_idx)block_size.x), raft::ceildiv(n_labels, (label_idx)block_size.y)); detail::fill_b_kernel<<<grid_size, block_size, 0, stream>>>( b_ptr, y, n_rows, n_labels, cluster_counts.data()); resource::wait_stream_pool_on_stream(handle); auto n_iters = 0; for (value_idx i = 0; i < n_rows; i += chunk) { for (value_idx j = 0; j < n_rows; j += chunk) { ++n_iters; auto chunk_stream = resource::get_next_usable_stream(handle, i + chunk * j); const auto* left_begin = X + (i * n_cols); const auto* right_begin = X + (j * n_cols); auto n_left_rows = (i + chunk) < n_rows ? chunk : (n_rows - i); auto n_right_rows = (j + chunk) < n_rows ? chunk : (n_rows - j); rmm::device_uvector<value_t> distances = get_pairwise_distance( handle, left_begin, right_begin, n_left_rows, n_right_rows, n_cols, metric, chunk_stream); compute_chunked_a_b(handle, a_ptr, b_ptr, i, j, y, n_labels, cluster_counts.data(), distances.data(), n_left_rows, n_right_rows, chunk_stream); } } resource::sync_stream_pool(handle); // calculating row-wise minimum in b // this prim only supports int indices for now raft::linalg::reduce<value_t, value_t, value_idx, raft::identity_op, raft::min_op>( b_ptr, b_ptr, n_labels, n_rows, std::numeric_limits<value_t>::max(), true, true, stream, false, raft::identity_op(), raft::min_op()); // calculating the silhouette score per sample raft::linalg::binaryOp<value_t, raft::stats::detail::SilOp<value_t>, value_t, value_idx>( a_ptr, a_ptr, b_ptr, n_rows, raft::stats::detail::SilOp<value_t>(), stream); return thrust::reduce(policy, a_ptr, a_ptr + n_rows, value_t(0)) / n_rows; } } // namespace detail } // namespace batched } // namespace stats } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/map_reduce.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. */ #ifndef __MAP_REDUCE_H #define __MAP_REDUCE_H #pragma once #include "detail/map_then_reduce.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> namespace raft::linalg { /** * @brief CUDA version of map and then generic reduction operation * @tparam Type data-type upon which the math operation will be performed * @tparam MapOp the device-lambda performing the actual map operation * @tparam ReduceLambda the device-lambda performing the actual reduction * @tparam TPB threads-per-block in the final kernel launched * @tparam Args additional parameters * @param out the output reduced value (assumed to be a device pointer) * @param len number of elements in the input array * @param neutral The neutral element of the reduction operation. For example: * 0 for sum, 1 for multiply, +Inf for Min, -Inf for Max * @param map the device-lambda * @param op the reduction device lambda * @param stream cuda-stream where to launch this kernel * @param in the input array * @param args additional input arrays */ template <typename InType, typename MapOp, typename ReduceLambda, typename IdxType = std::uint32_t, int TPB = 256, typename OutType = InType, typename... Args> void mapReduce(OutType* out, size_t len, OutType neutral, MapOp map, ReduceLambda op, cudaStream_t stream, const InType* in, Args... args) { detail::mapThenReduceImpl<InType, OutType, IdxType, MapOp, ReduceLambda, TPB, Args...>( out, len, neutral, map, op, stream, in, args...); } /** * @defgroup map_reduce Map-Reduce ops * @{ */ /** * @brief CUDA version of map and then generic reduction operation * @tparam InValueType the data-type of the input * @tparam MapOp the device-lambda performing the actual map operation * @tparam ReduceLambda the device-lambda performing the actual reduction * @tparam IndexType the index type * @tparam OutValueType the data-type of the output * @tparam ScalarIdxType index type of scalar * @tparam Args additional parameters * @param[in] handle raft::resources * @param[in] in the input of type raft::device_vector_view * @param[in] neutral The neutral element of the reduction operation. For example: * 0 for sum, 1 for multiply, +Inf for Min, -Inf for Max * @param[out] out the output reduced value assumed to be a raft::device_scalar_view * @param[in] map the fused device-lambda * @param[in] op the fused reduction device lambda * @param[in] args additional input arrays */ template <typename InValueType, typename MapOp, typename ReduceLambda, typename IndexType, typename OutValueType, typename ScalarIdxType, typename... Args> void map_reduce(raft::resources const& handle, raft::device_vector_view<const InValueType, IndexType> in, raft::device_scalar_view<OutValueType, ScalarIdxType> out, OutValueType neutral, MapOp map, ReduceLambda op, Args... args) { mapReduce<InValueType, MapOp, ReduceLambda, IndexType, 256, OutValueType, Args...>( out.data_handle(), in.extent(0), neutral, map, op, resource::get_cuda_stream(handle), in.data_handle(), args...); } /** @} */ // end of map_reduce } // end namespace raft::linalg #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/ternary_op.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. */ #ifndef __TERNARY_OP_H #define __TERNARY_OP_H #pragma once #include <raft/core/device_mdspan.hpp> #include <raft/core/resources.hpp> #include <raft/linalg/map.cuh> namespace raft { namespace linalg { /** * @brief perform element-wise ternary operation on the input arrays * @tparam math_t data-type upon which the math operation will be performed * @tparam Lambda the device-lambda performing the actual operation * @tparam IdxType Integer type used to for addressing * @tparam TPB threads-per-block in the final kernel launched * @param out the output array * @param in1 the first input array * @param in2 the second input array * @param in3 the third input array * @param len number of elements in the input array * @param op the device-lambda * @param stream cuda stream where to launch work */ template <typename math_t, typename Lambda, typename out_t, typename IdxType = int, int TPB = 256> void ternaryOp(out_t* out, const math_t* in1, const math_t* in2, const math_t* in3, IdxType len, Lambda op, cudaStream_t stream) { return detail::map<false>(stream, out, len, op, in1, in2, in3); } /** * @defgroup ternary_op Element-Wise Ternary Operation * @{ */ /** * @brief perform element-wise ternary operation on the input arrays * @tparam InType Input Type raft::device_mdspan * @tparam Lambda the device-lambda performing the actual operation * @tparam OutType Output Type raft::device_mdspan * @param[in] handle raft::resources * @param[in] in1 First input * @param[in] in2 Second input * @param[in] in3 Third input * @param[out] out Output * @param[in] op the device-lambda * @note Lambda must be a functor with the following signature: * `OutType func(const InType& val1, const InType& val2, const InType& val3);` */ template <typename InType, typename Lambda, typename OutType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void ternary_op( raft::resources const& handle, InType in1, InType in2, InType in3, OutType out, Lambda op) { return map(handle, in1, in2, in3, out, op); } /** @} */ // end of group ternary_op }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/unary_op.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. */ #ifndef __UNARY_OP_H #define __UNARY_OP_H #pragma once #include <raft/core/device_mdspan.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resources.hpp> #include <raft/linalg/map.cuh> namespace raft { namespace linalg { /** * @brief perform element-wise unary operation in the input array * @tparam InType input data-type * @tparam Lambda Device lambda performing the actual operation, with the signature * `OutType func(const InType& val);` * @tparam OutType output data-type * @tparam IdxType Integer type used to for addressing * @tparam TPB threads-per-block in the final kernel launched * @param[out] out Output array [on device], dim = [len] * @param[in] in Input array [on device], dim = [len] * @param[in] len Number of elements in the input array * @param[in] op Device lambda * @param[in] stream cuda stream where to launch work */ template <typename InType, typename Lambda, typename IdxType = int, typename OutType = InType, int TPB = 256> void unaryOp(OutType* out, const InType* in, IdxType len, Lambda op, cudaStream_t stream) { return detail::map<false>(stream, out, len, op, in); } /** * @brief Perform an element-wise unary operation into the output array * * Compared to `unaryOp()`, this method does not do any reads from any inputs * * @tparam OutType output data-type * @tparam Lambda Device lambda performing the actual operation, with the signature * `void func(OutType* outLocationOffset, IdxType idx);` * where outLocationOffset will be out + idx. * @tparam IdxType Integer type used to for addressing * @tparam TPB threads-per-block in the final kernel launched * * @param[out] out Output array [on device], dim = [len] * @param[in] len Number of elements in the input array * @param[in] op Device lambda * @param[in] stream cuda stream where to launch work */ template <typename OutType, typename Lambda, typename IdxType = int, int TPB = 256> void writeOnlyUnaryOp(OutType* out, IdxType len, Lambda op, cudaStream_t stream) { return detail::map<true>(stream, out, len, [op] __device__(IdxType offset) { OutType r; op(&r, offset); return r; }); } /** * @defgroup unary_op Element-Wise Unary Operations * @{ */ /** * @brief Perform an element-wise unary operation into the output array * @tparam InType Input Type raft::device_mdspan * @tparam Lambda Device lambda performing the actual operation, with the signature * `out_value_t func(const in_value_t& val);` * @tparam OutType Output Type raft::device_mdspan * @param[in] handle The raft handle * @param[in] in Input * @param[out] out Output * @param[in] op Device lambda */ template <typename InType, typename Lambda, typename OutType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void unary_op(raft::resources const& handle, InType in, OutType out, Lambda op) { return map(handle, in, out, op); } /** * @brief Perform an element-wise unary operation on the input index into the output array * * @note This operation is deprecated. Please use map_offset in `raft/linalg/map.cuh` instead. * * @tparam OutType Output Type raft::device_mdspan * @tparam Lambda Device lambda performing the actual operation, with the signature * `void func(out_value_t* out_location, index_t idx);` * @param[in] handle The raft handle * @param[out] out Output * @param[in] op Device lambda */ template <typename OutType, typename Lambda, typename = raft::enable_if_output_device_mdspan<OutType>> void write_only_unary_op(const raft::resources& handle, OutType out, Lambda op) { return writeOnlyUnaryOp(out.data_handle(), out.size(), op, resource::get_cuda_stream(handle)); } /** @} */ // end of group unary_op }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/lanczos.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. */ /** * This file is deprecated and will be removed in release 22.06. * Please use the cuh version instead. */ /** * DISCLAIMER: this file is deprecated: use lanczos.cuh instead */ #pragma once #pragma message(__FILE__ \ " is deprecated and will be removed in a future release." \ " Please use the sparse solvers version instead.") #include <raft/sparse/solver/lanczos.cuh> namespace raft::linalg { using raft::sparse::solver::computeLargestEigenvectors; using raft::sparse::solver::computeSmallestEigenvectors; } // namespace raft::linalg
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/axpy.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. */ #ifndef __AXPY_H #define __AXPY_H #pragma once #include "detail/axpy.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/host_mdspan.hpp> namespace raft::linalg { /** * @brief the wrapper of cublas axpy function * It computes the following equation: y = alpha * x + y * * @tparam T the element type * @tparam DevicePointerMode whether pointers alpha, beta point to device memory * @param [in] handle raft handle * @param [in] n number of elements in x and y * @param [in] alpha host or device scalar * @param [in] x vector of length n * @param [in] incx stride between consecutive elements of x * @param [inout] y vector of length n * @param [in] incy stride between consecutive elements of y * @param [in] stream */ template <typename T, bool DevicePointerMode = false> void axpy(raft::resources const& handle, const int n, const T* alpha, const T* x, const int incx, T* y, const int incy, cudaStream_t stream) { detail::axpy<T, DevicePointerMode>(handle, n, alpha, x, incx, y, incy, stream); } /** * @defgroup axpy axpy routine * @{ */ /** * @brief axpy function * It computes the following equation: y = alpha * x + y * * @param [in] handle raft::resources * @param [in] alpha raft::device_scalar_view * @param [in] x Input vector * @param [inout] y Output vector */ template <typename ElementType, typename IndexType, typename InLayoutPolicy, typename OutLayoutPolicy, typename ScalarIdxType> void axpy(raft::resources const& handle, raft::device_scalar_view<const ElementType, ScalarIdxType> alpha, raft::device_vector_view<const ElementType, IndexType, InLayoutPolicy> x, raft::device_vector_view<ElementType, IndexType, OutLayoutPolicy> y) { RAFT_EXPECTS(y.size() == x.size(), "Size mismatch between Output and Input"); axpy<ElementType, true>(handle, y.size(), alpha.data_handle(), x.data_handle(), x.stride(0), y.data_handle(), y.stride(0), resource::get_cuda_stream(handle)); } /** * @brief axpy function * It computes the following equation: y = alpha * x + y * @param [in] handle raft::resources * @param [in] alpha raft::device_scalar_view * @param [in] x Input vector * @param [inout] y Output vector */ template <typename ElementType, typename IndexType, typename InLayoutPolicy, typename OutLayoutPolicy, typename ScalarIdxType> void axpy(raft::resources const& handle, raft::host_scalar_view<const ElementType, ScalarIdxType> alpha, raft::device_vector_view<const ElementType, IndexType, InLayoutPolicy> x, raft::device_vector_view<ElementType, IndexType, OutLayoutPolicy> y) { RAFT_EXPECTS(y.size() == x.size(), "Size mismatch between Output and Input"); axpy<ElementType, false>(handle, y.size(), alpha.data_handle(), x.data_handle(), x.stride(0), y.data_handle(), y.stride(0), resource::get_cuda_stream(handle)); } /** @} */ // end of group axpy } // namespace raft::linalg #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/subtract.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. */ #ifndef __SUBTRACT_H #define __SUBTRACT_H #pragma once #include "detail/subtract.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/host_mdspan.hpp> #include <raft/util/input_validation.hpp> namespace raft { namespace linalg { /** * @brief Elementwise scalar subtraction operation on the input buffer * * @tparam InT input data-type. Also the data-type upon which the math ops * will be performed * @tparam OutT output data-type * @tparam IdxType Integer type used to for addressing * * @param out the output buffer * @param in the input buffer * @param scalar the scalar used in the operations * @param len number of elements in the input buffer * @param stream cuda stream where to launch work */ template <typename InT, typename OutT = InT, typename IdxType = int> void subtractScalar(OutT* out, const InT* in, InT scalar, IdxType len, cudaStream_t stream) { detail::subtractScalar(out, in, scalar, len, stream); } /** * @brief Elementwise subtraction operation on the input buffers * @tparam InT input data-type. Also the data-type upon which the math ops * will be performed * @tparam OutT output data-type * @tparam IdxType Integer type used to for addressing * * @param out the output buffer * @param in1 the first input buffer * @param in2 the second input buffer * @param len number of elements in the input buffers * @param stream cuda stream where to launch work */ template <typename InT, typename OutT = InT, typename IdxType = int> void subtract(OutT* out, const InT* in1, const InT* in2, IdxType len, cudaStream_t stream) { detail::subtract(out, in1, in2, len, stream); } /** Subtract single value pointed by singleScalarDev parameter in device memory from inDev[i] and * write result to outDev[i] * @tparam math_t data-type upon which the math operation will be performed * @tparam IdxType Integer type used to for addressing * @param outDev the output buffer * @param inDev the input buffer * @param singleScalarDev pointer to the scalar located in device memory * @param len number of elements in the input and output buffer * @param stream cuda stream * @remark block size has not been tuned */ template <typename math_t, typename IdxType = int, int TPB = 256> void subtractDevScalar(math_t* outDev, const math_t* inDev, const math_t* singleScalarDev, IdxType len, cudaStream_t stream) { detail::subtractDevScalar(outDev, inDev, singleScalarDev, len, stream); } /** * @defgroup sub Subtraction Arithmetic * @{ */ /** * @brief Elementwise subtraction operation on the input buffers * @tparam InType Input Type raft::device_mdspan * @tparam OutType Output Type raft::device_mdspan * @param handle raft::resources * @param[in] in1 First Input * @param[in] in2 Second Input * @param[out] out Output */ template <typename InType, typename OutType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void subtract(raft::resources const& handle, InType in1, InType in2, OutType out) { using in_value_t = typename InType::value_type; using out_value_t = typename OutType::value_type; RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in1), "Input 1 must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in2), "Input 2 must be contiguous"); RAFT_EXPECTS(out.size() == in1.size() && in1.size() == in2.size(), "Size mismatch between Output and Inputs"); if (out.size() <= std::numeric_limits<std::uint32_t>::max()) { subtract<in_value_t, out_value_t, std::uint32_t>(out.data_handle(), in1.data_handle(), in2.data_handle(), static_cast<std::uint32_t>(out.size()), resource::get_cuda_stream(handle)); } else { subtract<in_value_t, out_value_t, std::uint64_t>(out.data_handle(), in1.data_handle(), in2.data_handle(), static_cast<std::uint64_t>(out.size()), resource::get_cuda_stream(handle)); } } /** * @brief Elementwise subtraction of device scalar to input * @tparam InType Input Type raft::device_mdspan * @tparam OutType Output Type raft::device_mdspan * @tparam ScalarIdxType Index Type of scalar * @param[in] handle raft::resources * @param[in] in Input * @param[out] out Output * @param[in] scalar raft::device_scalar_view */ template <typename InType, typename OutType, typename ScalarIdxType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void subtract_scalar( raft::resources const& handle, InType in, OutType out, raft::device_scalar_view<const typename InType::element_type, ScalarIdxType> scalar) { using in_value_t = typename InType::value_type; using out_value_t = typename OutType::value_type; RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in), "Input must be contiguous"); RAFT_EXPECTS(out.size() == in.size(), "Size mismatch between Output and Input"); if (out.size() <= std::numeric_limits<std::uint32_t>::max()) { subtractDevScalar<in_value_t, out_value_t, std::uint32_t>( out.data_handle(), in.data_handle(), scalar.data_handle(), static_cast<std::uint32_t>(out.size()), resource::get_cuda_stream(handle)); } else { subtractDevScalar<in_value_t, out_value_t, std::uint64_t>( out.data_handle(), in.data_handle(), scalar.data_handle(), static_cast<std::uint64_t>(out.size()), resource::get_cuda_stream(handle)); } } /** * @brief Elementwise subtraction of host scalar to input * @tparam InType Input Type raft::device_mdspan * @tparam OutType Output Type raft::device_mdspan * @tparam ScalarIdxType Index Type of scalar * @param[in] handle raft::resources * @param[in] in Input * @param[out] out Output * @param[in] scalar raft::host_scalar_view */ template <typename InType, typename OutType, typename ScalarIdxType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void subtract_scalar( raft::resources const& handle, InType in, OutType out, raft::host_scalar_view<const typename InType::element_type, ScalarIdxType> scalar) { using in_value_t = typename InType::value_type; using out_value_t = typename OutType::value_type; RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in), "Input must be contiguous"); RAFT_EXPECTS(out.size() == in.size(), "Size mismatch between Output and Input"); if (out.size() <= std::numeric_limits<std::uint32_t>::max()) { subtractScalar<in_value_t, out_value_t, std::uint32_t>(out.data_handle(), in.data_handle(), *scalar.data_handle(), static_cast<std::uint32_t>(out.size()), resource::get_cuda_stream(handle)); } else { subtractScalar<in_value_t, out_value_t, std::uint64_t>(out.data_handle(), in.data_handle(), *scalar.data_handle(), static_cast<std::uint64_t>(out.size()), resource::get_cuda_stream(handle)); } } /** @} */ // end of group subtract }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/reduce.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. */ #ifndef __REDUCE_H #define __REDUCE_H #pragma once #include "detail/reduce.cuh" #include "linalg_types.hpp" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/operators.hpp> #include <raft/util/input_validation.hpp> namespace raft { namespace linalg { /** * @brief Compute reduction of the input matrix along the requested dimension * * @tparam InType the data type of the input * @tparam OutType the data type of the output (as well as the data type for * which reduction is performed) * @tparam IdxType data type of the indices of the array * @tparam MainLambda Unary lambda applied while acculumation (eg: L1 or L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*MainLambda)(InType, IdxType);</pre> * @tparam ReduceLambda Binary lambda applied for reduction (eg: addition(+) for L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*ReduceLambda)(OutType);</pre> * @tparam FinalLambda the final lambda applied before STG (eg: Sqrt for L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*FinalLambda)(OutType);</pre> * @param dots the output reduction vector * @param data the input matrix * @param D number of columns * @param N number of rows * @param init initial value to use for the reduction * @param rowMajor input matrix is row-major or not * @param alongRows whether to reduce along rows or columns * @param stream cuda stream where to launch work * @param inplace reduction result added inplace or overwrites old values? * @param main_op elementwise operation to apply before reduction * @param reduce_op binary reduction operation * @param final_op elementwise operation to apply before storing results */ template <typename InType, typename OutType = InType, typename IdxType = int, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void reduce(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, bool rowMajor, bool alongRows, cudaStream_t stream, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { detail::reduce<InType, OutType, IdxType>( dots, data, D, N, init, rowMajor, alongRows, stream, inplace, main_op, reduce_op, final_op); } /** * @defgroup reduction Reduction Along Requested Dimension * @{ */ /** * @brief Compute reduction of the input matrix along the requested dimension * This API computes a reduction of a matrix whose underlying storage * is either row-major or column-major, while allowing the choose the * dimension for reduction. Depending upon the dimension chosen for * reduction, the memory accesses may be coalesced or strided. * * @tparam InElementType the input data-type of underlying raft::matrix_view * @tparam LayoutPolicy The layout of Input/Output (row or col major) * @tparam OutElementType the output data-type of underlying raft::matrix_view and reduction * @tparam IndexType Integer type used to for addressing * @tparam MainLambda Unary lambda applied while acculumation (eg: L1 or L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*MainLambda)(InType, IdxType);</pre> * @tparam ReduceLambda Binary lambda applied for reduction (eg: addition(+) for L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*ReduceLambda)(OutType);</pre> * @tparam FinalLambda the final lambda applied before STG (eg: Sqrt for L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*FinalLambda)(OutType);</pre> * @param[in] handle raft::resources * @param[in] data Input of type raft::device_matrix_view * @param[out] dots Output of type raft::device_matrix_view * @param[in] init initial value to use for the reduction * @param[in] apply whether to reduce along rows or along columns (using raft::linalg::Apply) * @param[in] main_op fused elementwise operation to apply before reduction * @param[in] reduce_op fused binary reduction operation * @param[in] final_op fused elementwise operation to apply before storing results * @param[in] inplace reduction result added inplace or overwrites old values? */ template <typename InElementType, typename LayoutPolicy, typename OutElementType = InElementType, typename IdxType = std::uint32_t, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void reduce(raft::resources const& handle, raft::device_matrix_view<const InElementType, IdxType, LayoutPolicy> data, raft::device_vector_view<OutElementType, IdxType> dots, OutElementType init, Apply apply, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { RAFT_EXPECTS(raft::is_row_or_column_major(data), "Input must be contiguous"); auto constexpr row_major = std::is_same_v<typename decltype(data)::layout_type, raft::row_major>; bool along_rows = apply == Apply::ALONG_ROWS; if (along_rows) { RAFT_EXPECTS(static_cast<IdxType>(dots.size()) == data.extent(1), "Output should be equal to number of columns in Input"); } else { RAFT_EXPECTS(static_cast<IdxType>(dots.size()) == data.extent(0), "Output should be equal to number of rows in Input"); } reduce(dots.data_handle(), data.data_handle(), data.extent(1), data.extent(0), init, row_major, along_rows, resource::get_cuda_stream(handle), inplace, main_op, reduce_op, final_op); } /** @} */ // end of group reduction }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/eig.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. */ #ifndef __EIG_H #define __EIG_H #pragma once #include "detail/eig.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> namespace raft { namespace linalg { /** * @brief eig decomp with divide and conquer method for the column-major * symmetric matrices * @param handle raft handle * @param in the input buffer (symmetric matrix that has real eig values and * vectors. * @param n_rows: number of rows of the input * @param n_cols: number of cols of the input * @param eig_vectors: eigenvectors * @param eig_vals: eigen values * @param stream cuda stream */ template <typename math_t> void eigDC(raft::resources const& handle, const math_t* in, std::size_t n_rows, std::size_t n_cols, math_t* eig_vectors, math_t* eig_vals, cudaStream_t stream) { detail::eigDC(handle, in, n_rows, n_cols, eig_vectors, eig_vals, stream); } using detail::COPY_INPUT; using detail::EigVecMemUsage; using detail::OVERWRITE_INPUT; /** * @brief eig sel decomp with divide and conquer method for the column-major * symmetric matrices * @param handle raft handle * @param in the input buffer (symmetric matrix that has real eig values and * vectors. * @param n_rows: number of rows of the input * @param n_cols: number of cols of the input * @param n_eig_vals: number of eigenvectors to be generated * @param eig_vectors: eigenvectors * @param eig_vals: eigen values * @param memUsage: the memory selection for eig vector output * @param stream cuda stream */ template <typename math_t> void eigSelDC(raft::resources const& handle, math_t* in, std::size_t n_rows, std::size_t n_cols, std::size_t n_eig_vals, math_t* eig_vectors, math_t* eig_vals, EigVecMemUsage memUsage, cudaStream_t stream) { detail::eigSelDC(handle, in, n_rows, n_cols, n_eig_vals, eig_vectors, eig_vals, memUsage, stream); } /** * @brief overloaded function for eig decomp with Jacobi method for the * column-major symmetric matrices (in parameter) * @param handle: raft handle * @param in: input matrix * @param n_rows: number of rows of the input * @param n_cols: number of cols of the input * @param eig_vectors: eigenvectors * @param eig_vals: eigen values * @param stream: stream on which this function will be run * @param tol: error tolerance for the jacobi method. Algorithm stops when the * error is below tol * @param sweeps: number of sweeps in the Jacobi algorithm. The more the better * accuracy. */ template <typename math_t> void eigJacobi(raft::resources const& handle, const math_t* in, std::size_t n_rows, std::size_t n_cols, math_t* eig_vectors, math_t* eig_vals, cudaStream_t stream, math_t tol = 1.e-7, int sweeps = 15) { detail::eigJacobi(handle, in, n_rows, n_cols, eig_vectors, eig_vals, stream, tol, sweeps); } /** * @defgroup eig Eigen Decomposition Methods * @{ */ /** * @brief eig decomp with divide and conquer method for the column-major * symmetric matrices * @tparam ValueType the data-type of input and output * @tparam IntegerType Integer used for addressing * @param handle raft::resources * @param[in] in input raft::device_matrix_view (symmetric matrix that has real eig values and * vectors) * @param[out] eig_vectors: eigenvectors output of type raft::device_matrix_view * @param[out] eig_vals: eigen values output of type raft::device_vector_view */ template <typename ValueType, typename IndexType> void eig_dc(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> in, raft::device_matrix_view<ValueType, IndexType, raft::col_major> eig_vectors, raft::device_vector_view<ValueType, IndexType> eig_vals) { RAFT_EXPECTS(in.size() == eig_vectors.size(), "Size mismatch between Input and Eigen Vectors"); RAFT_EXPECTS(eig_vals.size() == in.extent(1), "Size mismatch between Input and Eigen Values"); eigDC(handle, in.data_handle(), in.extent(0), in.extent(1), eig_vectors.data_handle(), eig_vals.data_handle(), resource::get_cuda_stream(handle)); } /** * @brief eig decomp to select top-n eigen values with divide and conquer method * for the column-major symmetric matrices * @tparam ValueType the data-type of input and output * @tparam IntegerType Integer used for addressing * @param[in] handle raft::resources * @param[in] in input raft::device_matrix_view (symmetric matrix that has real eig values and * vectors) * @param[out] eig_vectors: eigenvectors output of type raft::device_matrix_view * @param[out] eig_vals: eigen values output of type raft::device_vector_view * @param[in] n_eig_vals: number of eigenvectors to be generated * @param[in] memUsage: the memory selection for eig vector output */ template <typename ValueType, typename IndexType> void eig_dc_selective(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> in, raft::device_matrix_view<ValueType, IndexType, raft::col_major> eig_vectors, raft::device_vector_view<ValueType, IndexType> eig_vals, std::size_t n_eig_vals, EigVecMemUsage memUsage) { RAFT_EXPECTS(eig_vectors.size() == n_eig_vals * in.extent(0), "Size mismatch between Input and Eigen Vectors"); RAFT_EXPECTS(eig_vals.size() == n_eig_vals, "Size mismatch between Input and Eigen Values"); raft::linalg::eigSelDC(handle, const_cast<ValueType*>(in.data_handle()), in.extent(0), in.extent(1), n_eig_vals, eig_vectors.data_handle(), eig_vals.data_handle(), memUsage, resource::get_cuda_stream(handle)); } /** * @brief overloaded function for eig decomp with Jacobi method for the * column-major symmetric matrices (in parameter) * @tparam ValueType the data-type of input and output * @tparam IntegerType Integer used for addressing * @param handle raft::resources * @param[in] in input raft::device_matrix_view (symmetric matrix that has real eig values and * vectors) * @param[out] eig_vectors: eigenvectors output of type raft::device_matrix_view * @param[out] eig_vals: eigen values output of type raft::device_vector_view * @param[in] tol: error tolerance for the jacobi method. Algorithm stops when the Frobenius norm of the absolute error is below tol * @param[in] sweeps: number of sweeps in the Jacobi algorithm. The more the better * accuracy. */ template <typename ValueType, typename IndexType> void eig_jacobi(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> in, raft::device_matrix_view<ValueType, IndexType, raft::col_major> eig_vectors, raft::device_vector_view<ValueType, IndexType> eig_vals, ValueType tol = 1.e-7, int sweeps = 15) { RAFT_EXPECTS(in.size() == eig_vectors.size(), "Size mismatch between Input and Eigen Vectors"); RAFT_EXPECTS(eig_vals.size() == in.extent(1), "Size mismatch between Input and Eigen Values"); eigJacobi(handle, in.data_handle(), in.extent(0), in.extent(1), eig_vectors.data_handle(), eig_vals.data_handle(), resource::get_cuda_stream(handle), tol, sweeps); } /** @} */ // end of eig }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/reduce_cols_by_key.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 __REDUCE_COLS_BY_KEY #define __REDUCE_COLS_BY_KEY #pragma once #include "detail/reduce_cols_by_key.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/resources.hpp> namespace raft { namespace linalg { /** * @brief Computes the sum-reduction of matrix columns for each given key * @tparam T the input data type (as well as the output reduced matrix) * @tparam KeyType data type of the keys * @tparam IdxType indexing arithmetic type * @param data the input data (dim = nrows x ncols). This is assumed to be in * row-major layout * @param keys keys array (len = ncols). It is assumed that each key in this * array is between [0, nkeys). In case this is not true, the caller is expected * to have called make_monotonic primitive to prepare such a contiguous and * monotonically increasing keys array. * @param out the output reduced matrix along columns (dim = nrows x nkeys). * This will be assumed to be in row-major layout * @param nrows number of rows in the input data * @param ncols number of columns in the input data * @param nkeys number of unique keys in the keys array * @param stream cuda stream to launch the kernel onto * @param reset_sums Whether to reset the output sums to zero before reducing */ template <typename T, typename KeyIteratorT, typename IdxType = int> void reduce_cols_by_key(const T* data, const KeyIteratorT keys, T* out, IdxType nrows, IdxType ncols, IdxType nkeys, cudaStream_t stream, bool reset_sums = true) { detail::reduce_cols_by_key(data, keys, out, nrows, ncols, nkeys, stream, reset_sums); } /** * @defgroup reduce_cols_by_key Reduce Across Columns by Key * @{ */ /** * @brief Computes the sum-reduction of matrix columns for each given key * TODO: Support generic reduction lambdas https://github.com/rapidsai/raft/issues/860 * @tparam ElementType the input data type (as well as the output reduced matrix) * @tparam KeyType data type of the keys * @tparam IndexType indexing arithmetic type * @param[in] handle raft::resources * @param[in] data the input data (dim = nrows x ncols). This is assumed to be in * row-major layout of type raft::device_matrix_view * @param[in] keys keys raft::device_vector_view (len = ncols). It is assumed that each key in this * array is between [0, nkeys). In case this is not true, the caller is expected * to have called make_monotonic primitive to prepare such a contiguous and * monotonically increasing keys array. * @param[out] out the output reduced raft::device_matrix_view along columns (dim = nrows x nkeys). * This will be assumed to be in row-major layout * @param[in] nkeys Number of unique keys in the keys array. By default, inferred from the number of * columns of out * @param[in] reset_sums Whether to reset the output sums to zero before reducing */ template <typename ElementType, typename KeyType = ElementType, typename IndexType = std::uint32_t> void reduce_cols_by_key( raft::resources const& handle, raft::device_matrix_view<const ElementType, IndexType, raft::row_major> data, raft::device_vector_view<const KeyType, IndexType> keys, raft::device_matrix_view<ElementType, IndexType, raft::row_major> out, IndexType nkeys = 0, bool reset_sums = true) { if (nkeys > 0) { RAFT_EXPECTS(out.extent(1) == nkeys, "Output doesn't have nkeys columns"); } else { nkeys = out.extent(1); } RAFT_EXPECTS(out.extent(0) == data.extent(0), "Output doesn't have the same number of rows as input"); RAFT_EXPECTS(keys.extent(0) == data.extent(1), "Keys is not of size ncols"); reduce_cols_by_key(data.data_handle(), keys.data_handle(), out.data_handle(), data.extent(0), data.extent(1), nkeys, resource::get_cuda_stream(handle), reset_sums); } /** @} */ // end of group reduce_cols_by_key }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/eltwise.cuh
/* * Copyright (c) 2022, 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 __ELTWISE_H #define __ELTWISE_H #pragma once #include "detail/eltwise.cuh" namespace raft { namespace linalg { /** * @defgroup ScalarOps Scalar operations on the input buffer * @tparam InType data-type upon which the math operation will be performed * @tparam IdxType Integer type used to for addressing * @param out the output buffer * @param in the input buffer * @param scalar the scalar used in the operations * @param len number of elements in the input buffer * @param stream cuda stream where to launch work * @{ */ template <typename InType, typename IdxType, typename OutType = InType> void scalarAdd(OutType* out, const InType* in, InType scalar, IdxType len, cudaStream_t stream) { detail::scalarAdd(out, in, scalar, len, stream); } template <typename InType, typename IdxType, typename OutType = InType> void scalarMultiply(OutType* out, const InType* in, InType scalar, IdxType len, cudaStream_t stream) { detail::scalarMultiply(out, in, scalar, len, stream); } /** @} */ /** * @defgroup BinaryOps Element-wise binary operations on the input buffers * @tparam InType data-type upon which the math operation will be performed * @tparam IdxType Integer type used to for addressing * @param out the output buffer * @param in1 the first input buffer * @param in2 the second input buffer * @param len number of elements in the input buffers * @param stream cuda stream where to launch work * @{ */ template <typename InType, typename IdxType, typename OutType = InType> void eltwiseAdd( OutType* out, const InType* in1, const InType* in2, IdxType len, cudaStream_t stream) { detail::eltwiseAdd(out, in1, in2, len, stream); } template <typename InType, typename IdxType, typename OutType = InType> void eltwiseSub( OutType* out, const InType* in1, const InType* in2, IdxType len, cudaStream_t stream) { detail::eltwiseSub(out, in1, in2, len, stream); } template <typename InType, typename IdxType, typename OutType = InType> void eltwiseMultiply( OutType* out, const InType* in1, const InType* in2, IdxType len, cudaStream_t stream) { detail::eltwiseMultiply(out, in1, in2, len, stream); } template <typename InType, typename IdxType, typename OutType = InType> void eltwiseDivide( OutType* out, const InType* in1, const InType* in2, IdxType len, cudaStream_t stream) { detail::eltwiseDivide(out, in1, in2, len, stream); } template <typename InType, typename IdxType, typename OutType = InType> void eltwiseDivideCheckZero( OutType* out, const InType* in1, const InType* in2, IdxType len, cudaStream_t stream) { detail::eltwiseDivideCheckZero(out, in1, in2, len, stream); } /** @} */ }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/matrix_vector_op.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. */ #ifndef __MATRIX_VECTOR_OP_H #define __MATRIX_VECTOR_OP_H #pragma once #include "detail/matrix_vector_op.cuh" #include "linalg_types.hpp" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/resources.hpp> #include <raft/util/input_validation.hpp> namespace raft { namespace linalg { /** * @brief Operations for all the columns or rows with a given vector. * Caution : Threads process multiple elements to speed up processing. These * are loaded in a single read thanks to type promotion. Faster processing * would thus only be enabled when addresses are optimally aligned for it. * Note : the function will also check that the size of the window of accesses * is a multiple of the number of elements processed by a thread in order to * enable faster processing * @tparam MatT the matrix type * @tparam Lambda a device function which represents a binary operator * @tparam VecT the input vector type * @tparam IdxType Integer type used to for addressing * @param out the output matrix (passing out = matrix makes it in-place) * @param matrix the input matrix * @param vec the vector * @param D number of columns of matrix * @param N number of rows of matrix * @param rowMajor whether input is row or col major * @param bcastAlongRows whether the broadcast of vector needs to happen along * the rows of the matrix or columns * @param op the mathematical operation * @param stream cuda stream where to launch work */ template <typename MatT, typename Lambda, typename VecT, typename IdxType = int> void matrixVectorOp(MatT* out, const MatT* matrix, const VecT* vec, IdxType D, IdxType N, bool rowMajor, bool bcastAlongRows, Lambda op, cudaStream_t stream) { detail::matrixVectorOp(out, matrix, vec, D, N, rowMajor, bcastAlongRows, op, stream); } /** * @brief Operations for all the columns or rows with the given vectors. * Caution : Threads process multiple elements to speed up processing. These * are loaded in a single read thanks to type promotion. Faster processing * would thus only be enabled when addresses are optimally aligned for it. * Note : the function will also check that the size of the window of accesses * is a multiple of the number of elements processed by a thread in order to * enable faster processing * @tparam MatT the matrix type * @tparam Lambda a device function which represents a binary operator * @tparam Vec1T the first input vector type * @tparam Vec2T the second input vector type * @tparam IdxType Integer type used to for addressing * @param out the output matrix (passing out = matrix makes it in-place) * @param matrix the input matrix * @param vec1 the first vector * @param vec2 the second vector * @param D number of columns of matrix * @param N number of rows of matrix * @param rowMajor whether input is row or col major * @param bcastAlongRows whether the broadcast of vector needs to happen along * the rows of the matrix or columns * @param op the mathematical operation * @param stream cuda stream where to launch work */ template <typename MatT, typename Lambda, typename Vec1T, typename Vec2T, typename IdxType = int> void matrixVectorOp(MatT* out, const MatT* matrix, const Vec1T* vec1, const Vec2T* vec2, IdxType D, IdxType N, bool rowMajor, bool bcastAlongRows, Lambda op, cudaStream_t stream) { detail::matrixVectorOp(out, matrix, vec1, vec2, D, N, rowMajor, bcastAlongRows, op, stream); } /** * @defgroup matrix_vector_op Matrix Vector Operations * @{ */ /** * @brief Operations for all the columns or rows with a given vector. * Caution : Threads process multiple elements to speed up processing. These * are loaded in a single read thanks to type promotion. Faster processing * would thus only be enabled when addresses are optimally aligned for it. * Note : the function will also check that the size of the window of accesses * is a multiple of the number of elements processed by a thread in order to * enable faster processing * @tparam MatValueType the data-type of the input matrix * @tparam VecValueType the data-type of the input vector * @tparam LayoutPolicy the layout of input and output (raft::row_major or raft::col_major) * @tparam Lambda a device function which represents a binary operator * @tparam IndexType Integer used for addressing * @param[in] handle raft::resources * @param[in] matrix input raft::matrix_view * @param[in] vec vector raft::vector_view * @param[out] out output raft::matrix_view * @param[in] apply whether the broadcast of vector needs to happen along * the rows of the matrix or columns using enum class raft::linalg::Apply * @param[in] op the mathematical operation */ template <typename MatValueType, typename VecValueType, typename LayoutPolicy, typename Lambda, typename IndexType> void matrix_vector_op(raft::resources const& handle, raft::device_matrix_view<const MatValueType, IndexType, LayoutPolicy> matrix, raft::device_vector_view<const VecValueType, IndexType> vec, raft::device_matrix_view<MatValueType, IndexType, LayoutPolicy> out, Apply apply, Lambda op) { RAFT_EXPECTS(raft::is_row_or_column_major(matrix), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(out), "Input must be contiguous"); RAFT_EXPECTS(out.size() == matrix.size(), "Size mismatch between Output and Input"); auto constexpr rowMajor = std::is_same_v<typename decltype(out)::layout_type, raft::row_major>; auto bcastAlongRows = apply == Apply::ALONG_ROWS; if (bcastAlongRows) { RAFT_EXPECTS(out.extent(1) == static_cast<IndexType>(vec.size()), "Size mismatch between matrix and vector"); } else { RAFT_EXPECTS(out.extent(0) == static_cast<IndexType>(vec.size()), "Size mismatch between matrix and vector"); } matrixVectorOp(out.data_handle(), matrix.data_handle(), vec.data_handle(), out.extent(1), out.extent(0), rowMajor, bcastAlongRows, op, resource::get_cuda_stream(handle)); } /** * @brief Operations for all the columns or rows with the given vectors. * Caution : Threads process multiple elements to speed up processing. These * are loaded in a single read thanks to type promotion. Faster processing * would thus only be enabled when addresses are optimally aligned for it. * Note : the function will also check that the size of the window of accesses * is a multiple of the number of elements processed by a thread in order to * enable faster processing * @tparam MatValueType the data-type of the input and output matrices * @tparam Vec1ValueType the data-type of the first input vector * @tparam Vec2ValueType the data-type of the second input vector * @tparam LayoutPolicy the layout of input and output (raft::row_major or raft::col_major) * @tparam Lambda a device function which represents a binary operator * @tparam IndexType Integer used for addressing * @param handle raft::resources * @param matrix input raft::matrix_view * @param vec1 the first vector raft::vector_view * @param vec2 the second vector raft::vector_view * @param out output raft::matrix_view * @param apply whether the broadcast of vector needs to happen along * the rows of the matrix or columns using enum class raft::linalg::Apply * @param op the mathematical operation */ template <typename MatValueType, typename Vec1ValueType, typename Vec2ValueType, typename LayoutPolicy, typename Lambda, typename IndexType> void matrix_vector_op(raft::resources const& handle, raft::device_matrix_view<const MatValueType, IndexType, LayoutPolicy> matrix, raft::device_vector_view<const Vec1ValueType, IndexType> vec1, raft::device_vector_view<const Vec2ValueType, IndexType> vec2, raft::device_matrix_view<MatValueType, IndexType, LayoutPolicy> out, Apply apply, Lambda op) { RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(matrix), "Input must be contiguous"); RAFT_EXPECTS(out.size() == matrix.size(), "Size mismatch between Output and Input"); auto constexpr rowMajor = std::is_same_v<typename decltype(out)::layout_type, raft::row_major>; auto bcastAlongRows = apply == Apply::ALONG_ROWS; if (bcastAlongRows) { RAFT_EXPECTS(out.extent(1) == static_cast<IndexType>(vec1.size()), "Size mismatch between matrix and vector"); RAFT_EXPECTS(out.extent(1) == static_cast<IndexType>(vec2.size()), "Size mismatch between matrix and vector"); } else { RAFT_EXPECTS(out.extent(0) == static_cast<IndexType>(vec1.size()), "Size mismatch between matrix and vector"); RAFT_EXPECTS(out.extent(0) == static_cast<IndexType>(vec2.size()), "Size mismatch between matrix and vector"); } matrixVectorOp(out.data_handle(), matrix.data_handle(), vec1.data_handle(), vec2.data_handle(), out.extent(1), out.extent(0), rowMajor, bcastAlongRows, op, resource::get_cuda_stream(handle)); } /** @} */ // end of group matrix_vector_op }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/cublas_macros.h
/* * Copyright (c) 2022, 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. */ /** * This file is deprecated and will be removed in release 22.06. * Please use core/cublas_macros.hpp instead. */ #pragma once #include <raft/core/cublas_macros.hpp>
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/matrix_vector.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/core/device_mdspan.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/linalg/linalg_types.hpp> #include <raft/matrix/detail/math.cuh> #include <raft/util/input_validation.hpp> namespace raft::linalg { /** * @defgroup matrix_vector Matrix-Vector Operations * @{ */ /** * @brief multiply each row or column of matrix with vector, skipping zeros in vector * @param [in] handle: raft handle for managing library resources * @param[inout] data: input matrix, results are in-place * @param[in] vec: input vector * @param[in] apply whether the broadcast of vector needs to happen along * the rows of the matrix or columns using enum class raft::linalg::Apply */ template <typename math_t, typename idx_t, typename layout_t> void binary_mult_skip_zero(raft::resources const& handle, raft::device_matrix_view<math_t, idx_t, layout_t> data, raft::device_vector_view<const math_t, idx_t> vec, Apply apply) { bool row_major = raft::is_row_major(data); auto bcast_along_rows = apply == Apply::ALONG_ROWS; idx_t vec_size = bcast_along_rows ? data.extent(1) : data.extent(0); RAFT_EXPECTS( vec.extent(0) == vec_size, "If `bcast_along_rows==true`, vector size must equal number of columns in the matrix." "If `bcast_along_rows==false`, vector size must equal number of rows in the matrix."); matrix::detail::matrixVectorBinaryMultSkipZero(data.data_handle(), vec.data_handle(), data.extent(0), data.extent(1), row_major, bcast_along_rows, resource::get_cuda_stream(handle)); } /** * @brief divide each row or column of matrix with vector * @param[in] handle: raft handle for managing library resources * @param[inout] data: input matrix, results are in-place * @param[in] vec: input vector * @param[in] apply whether the broadcast of vector needs to happen along * the rows of the matrix or columns using enum class raft::linalg::Apply */ template <typename math_t, typename idx_t, typename layout_t> void binary_div(raft::resources const& handle, raft::device_matrix_view<math_t, idx_t, layout_t> data, raft::device_vector_view<const math_t, idx_t> vec, Apply apply) { bool row_major = raft::is_row_major(data); auto bcast_along_rows = apply == Apply::ALONG_ROWS; idx_t vec_size = bcast_along_rows ? data.extent(1) : data.extent(0); RAFT_EXPECTS( vec.extent(0) == vec_size, "If `bcast_along_rows==true`, vector size must equal number of columns in the matrix." "If `bcast_along_rows==false`, vector size must equal number of rows in the matrix."); matrix::detail::matrixVectorBinaryDiv(data.data_handle(), vec.data_handle(), data.extent(0), data.extent(1), row_major, bcast_along_rows, resource::get_cuda_stream(handle)); } /** * @brief divide each row or column of matrix with vector, skipping zeros in vector * @param[in] handle: raft handle for managing library resources * @param[inout] data: input matrix, results are in-place * @param[in] vec: input vector * @param[in] apply whether the broadcast of vector needs to happen along * the rows of the matrix or columns using enum class raft::linalg::Apply * @param[in] return_zero: result is zero if true and vector value is below threshold, original * value if false */ template <typename math_t, typename idx_t, typename layout_t> void binary_div_skip_zero(raft::resources const& handle, raft::device_matrix_view<math_t, idx_t, layout_t> data, raft::device_vector_view<const math_t, idx_t> vec, Apply apply, bool return_zero = false) { bool row_major = raft::is_row_major(data); auto bcast_along_rows = apply == Apply::ALONG_ROWS; idx_t vec_size = bcast_along_rows ? data.extent(1) : data.extent(0); RAFT_EXPECTS( vec.extent(0) == vec_size, "If `bcast_along_rows==true`, vector size must equal number of columns in the matrix." "If `bcast_along_rows==false`, vector size must equal number of rows in the matrix."); matrix::detail::matrixVectorBinaryDivSkipZero(data.data_handle(), vec.data_handle(), data.extent(0), data.extent(1), row_major, bcast_along_rows, resource::get_cuda_stream(handle), return_zero); } /** * @brief add each row or column of matrix with vector * @param[in] handle: raft handle for managing library resources * @param[inout] data: input matrix, results are in-place * @param[in] vec: input vector * @param[in] apply whether the broadcast of vector needs to happen along * the rows of the matrix or columns using enum class raft::linalg::Apply */ template <typename math_t, typename idx_t, typename layout_t> void binary_add(raft::resources const& handle, raft::device_matrix_view<math_t, idx_t, layout_t> data, raft::device_vector_view<const math_t, idx_t> vec, Apply apply) { bool row_major = raft::is_row_major(data); auto bcast_along_rows = apply == Apply::ALONG_ROWS; idx_t vec_size = bcast_along_rows ? data.extent(1) : data.extent(0); RAFT_EXPECTS( vec.extent(0) == vec_size, "If `bcast_along_rows==true`, vector size must equal number of columns in the matrix." "If `bcast_along_rows==false`, vector size must equal number of rows in the matrix."); matrix::detail::matrixVectorBinaryAdd(data.data_handle(), vec.data_handle(), data.extent(0), data.extent(1), row_major, bcast_along_rows, resource::get_cuda_stream(handle)); } /** * @brief subtract each row or column of matrix with vector * @param[in] handle: raft handle for managing library resources * @param[inout] data: input matrix, results are in-place * @param[in] vec: input vector * @param[in] apply whether the broadcast of vector needs to happen along * the rows of the matrix or columns using enum class raft::linalg::Apply */ template <typename math_t, typename idx_t, typename layout_t> void binary_sub(raft::resources const& handle, raft::device_matrix_view<math_t, idx_t, layout_t> data, raft::device_vector_view<const math_t, idx_t> vec, Apply apply) { bool row_major = raft::is_row_major(data); auto bcast_along_rows = apply == Apply::ALONG_ROWS; idx_t vec_size = bcast_along_rows ? data.extent(1) : data.extent(0); RAFT_EXPECTS( vec.extent(0) == vec_size, "If `bcast_along_rows==true`, vector size must equal number of columns in the matrix." "If `bcast_along_rows==false`, vector size must equal number of rows in the matrix."); matrix::detail::matrixVectorBinarySub(data.data_handle(), vec.data_handle(), data.extent(0), data.extent(1), row_major, bcast_along_rows, resource::get_cuda_stream(handle)); } /** @} */ // end of matrix_vector } // namespace raft::linalg
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/cusolver_macros.h
/* * Copyright (c) 2022, 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. */ /** * This file is deprecated and will be removed in release 22.06. * Please use core/cusolver_macros.hpp instead. */ #pragma once #include <raft/core/cusolver_macros.hpp>
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/map.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. */ #ifndef __MAP_H #define __MAP_H #pragma once #include "detail/map.cuh" #include <raft/core/device_mdspan.hpp> #include <raft/core/resources.hpp> namespace raft::linalg { /** * @brief CUDA version of map * * Note: This call is deprecated, please use `map` from the same file. * * @tparam InType data-type upon which the math operation will be performed * @tparam MapOp the device-lambda performing the actual operation * @tparam TPB threads-per-block in the final kernel launched * @tparam Args additional parameters * @tparam OutType data-type in which the result will be stored * @param out the output of the map operation (assumed to be a device pointer) * @param len number of elements in the input array * @param map the device-lambda * @param stream cuda-stream where to launch this kernel * @param in the input array * @param args additional input arrays */ template <typename InType, typename MapOp, typename IdxType = std::uint32_t, int TPB = 256, typename OutType = InType, typename... Args> [[deprecated("Use function `map` from the same file")]] void map_k( OutType* out, IdxType len, MapOp map, cudaStream_t stream, const InType* in, Args... args) { return detail::map<false>(stream, out, len, map, in, args...); } /** * @defgroup map Mapping ops * @{ */ /** * @brief Map a function over zero or more input mdspans of the same size. * * The algorithm applied on `k` inputs can be described in a following pseudo-code: * @code * for (auto i: [0 ... out.size()]) { * out[i] = f(in_0[i], in_1[i], ..., in_k[i]) * } * @endcode * * _Performance note_: when possible, this function loads the argument arrays and stores the output * array using vectorized cuda load/store instructions. The size of the vectorization depends on the * size of the largest input/output element type and on the alignment of all pointers. * * Usage example: * @code{.cpp} * #include <raft/core/device_mdarray.hpp> * #include <raft/core/resources.hpp> * #include <raft/core/operators.hpp> * #include <raft/linalg/map.cuh> * * auto input = raft::make_device_vector<int>(res, n); * ... fill input .. * auto squares = raft::make_device_vector<int>(res, n); * raft::linalg::map_offset(res, squares.view(), raft::sq_op{}, input.view()); * @endcode * * @tparam OutType data-type of the result (device_mdspan) * @tparam Func the device-lambda performing the actual operation * @tparam InTypes data-types of the inputs (device_mdspan) * * @param[in] res raft::resources * @param[out] out the output of the map operation (device_mdspan) * @param[in] f device lambda * (InTypes::value_type xs...) -> OutType::value_type * @param[in] ins the inputs (each of the same size as the output) (device_mdspan) */ template <typename OutType, typename Func, typename... InTypes, typename = raft::enable_if_output_device_mdspan<OutType>, typename = raft::enable_if_input_device_mdspan<InTypes...>> void map(const raft::resources& res, OutType out, Func f, InTypes... ins) { return detail::map<false>(res, out, f, ins...); } /** * @brief Map a function over one mdspan. * * @tparam InType1 data-type of the input (device_mdspan) * @tparam OutType data-type of the result (device_mdspan) * @tparam Func the device-lambda performing the actual operation * * @param[in] res raft::resources * @param[in] in1 the input (the same size as the output) (device_mdspan) * @param[out] out the output of the map operation (device_mdspan) * @param[in] f device lambda * (InType1::value_type x) -> OutType::value_type */ template <typename InType1, typename OutType, typename Func, typename = raft::enable_if_output_device_mdspan<OutType>, typename = raft::enable_if_input_device_mdspan<InType1>> void map(const raft::resources& res, InType1 in1, OutType out, Func f) { return detail::map<false>(res, out, f, in1); } /** * @brief Map a function over two mdspans. * * @tparam InType1 data-type of the input (device_mdspan) * @tparam InType2 data-type of the input (device_mdspan) * @tparam OutType data-type of the result (device_mdspan) * @tparam Func the device-lambda performing the actual operation * * @param[in] res raft::resources * @param[in] in1 the input (the same size as the output) (device_mdspan) * @param[in] in2 the input (the same size as the output) (device_mdspan) * @param[out] out the output of the map operation (device_mdspan) * @param[in] f device lambda * (InType1::value_type x1, InType2::value_type x2) -> OutType::value_type */ template <typename InType1, typename InType2, typename OutType, typename Func, typename = raft::enable_if_output_device_mdspan<OutType>, typename = raft::enable_if_input_device_mdspan<InType1, InType2>> void map(const raft::resources& res, InType1 in1, InType2 in2, OutType out, Func f) { return detail::map<false>(res, out, f, in1, in2); } /** * @brief Map a function over three mdspans. * * @tparam InType1 data-type of the input 1 (device_mdspan) * @tparam InType2 data-type of the input 2 (device_mdspan) * @tparam InType3 data-type of the input 3 (device_mdspan) * @tparam OutType data-type of the result (device_mdspan) * @tparam Func the device-lambda performing the actual operation * * @param[in] res raft::resources * @param[in] in1 the input 1 (the same size as the output) (device_mdspan) * @param[in] in2 the input 2 (the same size as the output) (device_mdspan) * @param[in] in3 the input 3 (the same size as the output) (device_mdspan) * @param[out] out the output of the map operation (device_mdspan) * @param[in] f device lambda * (InType1::value_type x1, InType2::value_type x2, InType3::value_type x3) -> OutType::value_type */ template <typename InType1, typename InType2, typename InType3, typename OutType, typename Func, typename = raft::enable_if_output_device_mdspan<OutType>, typename = raft::enable_if_input_device_mdspan<InType1, InType2, InType3>> void map(const raft::resources& res, InType1 in1, InType2 in2, InType3 in3, OutType out, Func f) { return detail::map<false>(res, out, f, in1, in2, in3); } /** * @brief Map a function over zero-based flat index (element offset) and zero or more inputs. * * The algorithm applied on `k` inputs can be described in a following pseudo-code: * @code * for (auto i: [0 ... out.size()]) { * out[i] = f(i, in_0[i], in_1[i], ..., in_k[i]) * } * @endcode * * _Performance note_: when possible, this function loads the argument arrays and stores the output * array using vectorized cuda load/store instructions. The size of the vectorization depends on the * size of the largest input/output element type and on the alignment of all pointers. * * Usage example: * @code{.cpp} * #include <raft/core/device_mdarray.hpp> * #include <raft/core/resources.hpp> * #include <raft/core/operators.hpp> * #include <raft/linalg/map.cuh> * * auto squares = raft::make_device_vector<int>(handle, n); * raft::linalg::map_offset(res, squares.view(), raft::sq_op{}); * @endcode * * @tparam OutType data-type of the result (device_mdspan) * @tparam Func the device-lambda performing the actual operation * @tparam InTypes data-types of the inputs (device_mdspan) * * @param[in] res raft::resources * @param[out] out the output of the map operation (device_mdspan) * @param[in] f device lambda * (auto offset, InTypes::value_type xs...) -> OutType::value_type * @param[in] ins the inputs (each of the same size as the output) (device_mdspan) */ template <typename OutType, typename Func, typename... InTypes, typename = raft::enable_if_output_device_mdspan<OutType>, typename = raft::enable_if_input_device_mdspan<InTypes...>> void map_offset(const raft::resources& res, OutType out, Func f, InTypes... ins) { return detail::map<true>(res, out, f, ins...); } /** * @brief Map a function over zero-based flat index (element offset) and one mdspan. * * @tparam InType1 data-type of the input (device_mdspan) * @tparam OutType data-type of the result (device_mdspan) * @tparam Func the device-lambda performing the actual operation * * @param[in] res raft::resources * @param[in] in1 the input (the same size as the output) (device_mdspan) * @param[out] out the output of the map operation (device_mdspan) * @param[in] f device lambda * (auto offset, InType1::value_type x) -> OutType::value_type */ template <typename InType1, typename OutType, typename Func, typename = raft::enable_if_output_device_mdspan<OutType>, typename = raft::enable_if_input_device_mdspan<InType1>> void map_offset(const raft::resources& res, InType1 in1, OutType out, Func f) { return detail::map<true>(res, out, f, in1); } /** * @brief Map a function over zero-based flat index (element offset) and two mdspans. * * @tparam InType1 data-type of the input (device_mdspan) * @tparam InType2 data-type of the input (device_mdspan) * @tparam OutType data-type of the result (device_mdspan) * @tparam Func the device-lambda performing the actual operation * * @param[in] res raft::resources * @param[in] in1 the input (the same size as the output) (device_mdspan) * @param[in] in2 the input (the same size as the output) (device_mdspan) * @param[out] out the output of the map operation (device_mdspan) * @param[in] f device lambda * (auto offset, InType1::value_type x1, InType2::value_type x2) -> OutType::value_type */ template <typename InType1, typename InType2, typename OutType, typename Func, typename = raft::enable_if_output_device_mdspan<OutType>, typename = raft::enable_if_input_device_mdspan<InType1, InType2>> void map_offset(const raft::resources& res, InType1 in1, InType2 in2, OutType out, Func f) { return detail::map<true>(res, out, f, in1, in2); } /** * @brief Map a function over zero-based flat index (element offset) and three mdspans. * * @tparam InType1 data-type of the input 1 (device_mdspan) * @tparam InType2 data-type of the input 2 (device_mdspan) * @tparam InType3 data-type of the input 3 (device_mdspan) * @tparam OutType data-type of the result (device_mdspan) * @tparam Func the device-lambda performing the actual operation * * @param[in] res raft::resources * @param[in] in1 the input 1 (the same size as the output) (device_mdspan) * @param[in] in2 the input 2 (the same size as the output) (device_mdspan) * @param[in] in3 the input 3 (the same size as the output) (device_mdspan) * @param[out] out the output of the map operation (device_mdspan) * @param[in] f device lambda * (auto offset, InType1::value_type x1, InType2::value_type x2, InType3::value_type x3) * -> OutType::value_type */ template <typename InType1, typename InType2, typename InType3, typename OutType, typename Func, typename = raft::enable_if_output_device_mdspan<OutType>, typename = raft::enable_if_input_device_mdspan<InType1, InType2, InType3>> void map_offset( const raft::resources& res, InType1 in1, InType2 in2, InType3 in3, OutType out, Func f) { return detail::map<true>(res, out, f, in1, in2, in3); } /** @} */ // end of map } // namespace raft::linalg #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/strided_reduction.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. */ #ifndef __STRIDED_REDUCTION_H #define __STRIDED_REDUCTION_H #pragma once #include "detail/strided_reduction.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/operators.hpp> #include <raft/core/resources.hpp> #include <type_traits> namespace raft { namespace linalg { /** * @brief Compute reduction of the input matrix along the strided dimension * * @tparam InType the data type of the input * @tparam OutType the data type of the output (as well as the data type for * which reduction is performed) * @tparam IdxType data type of the indices of the array * @tparam MainLambda Unary lambda applied while acculumation (eg: L1 or L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*MainLambda)(InType, IdxType);</pre> * @tparam ReduceLambda Binary lambda applied for reduction (eg: addition(+) for L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*ReduceLambda)(OutType);</pre> * @tparam FinalLambda the final lambda applied before STG (eg: Sqrt for L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*FinalLambda)(OutType);</pre> * @param dots the output reduction vector * @param data the input matrix * @param D leading dimension of data * @param N second dimension data * @param init initial value to use for the reduction * @param main_op elementwise operation to apply before reduction * @param reduce_op binary reduction operation * @param final_op elementwise operation to apply before storing results * @param inplace reduction result added inplace or overwrites old values? * @param stream cuda stream where to launch work */ template <typename InType, typename OutType = InType, typename IdxType = int, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void stridedReduction(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, cudaStream_t stream, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { // Only compile for types supported by myAtomicReduce, but don't make the compilation fail in // other cases, because coalescedReduction supports arbitrary types. if constexpr (std::is_same_v<OutType, float> || std::is_same_v<OutType, double> || std::is_same_v<OutType, int> || std::is_same_v<OutType, long long> || std::is_same_v<OutType, unsigned long long>) { detail::stridedReduction<InType, OutType, IdxType>( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } else { THROW("Unsupported type for stridedReduction: %s", typeid(OutType).name()); } } /** * @defgroup strided_reduction Strided Memory Access Reductions * For reducing along rows for row-major and along columns for column-major * @{ */ /** * @brief Compute reduction of the input matrix along the strided dimension * This API is to be used when the desired reduction is NOT along the dimension * of the memory layout. For example, a row-major matrix will be reduced * along the rows whereas a column-major matrix will be reduced along * the columns. * * @tparam InValueType the input data-type of underlying raft::matrix_view * @tparam LayoutPolicy The layout of Input/Output (row or col major) * @tparam OutValueType the output data-type of underlying raft::matrix_view and reduction * @tparam IndexType Integer type used to for addressing * @tparam MainLambda Unary lambda applied while acculumation (eg: L1 or L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*MainLambda)(InType, IdxType);</pre> * @tparam ReduceLambda Binary lambda applied for reduction (eg: addition(+) for L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*ReduceLambda)(OutType);</pre> * @tparam FinalLambda the final lambda applied before STG (eg: Sqrt for L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*FinalLambda)(OutType);</pre> * @param[in] handle raft::resources * @param[in] data Input of type raft::device_matrix_view * @param[out] dots Output of type raft::device_matrix_view * @param[in] init initial value to use for the reduction * @param[in] main_op fused elementwise operation to apply before reduction * @param[in] reduce_op fused binary reduction operation * @param[in] final_op fused elementwise operation to apply before storing results * @param[in] inplace reduction result added inplace or overwrites old values? */ template <typename InValueType, typename LayoutPolicy, typename OutValueType, typename IndexType, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void strided_reduction(raft::resources const& handle, raft::device_matrix_view<const InValueType, IndexType, LayoutPolicy> data, raft::device_vector_view<OutValueType, IndexType> dots, OutValueType init, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { if constexpr (std::is_same_v<LayoutPolicy, raft::row_major>) { RAFT_EXPECTS(static_cast<IndexType>(dots.size()) == data.extent(1), "Output should be equal to number of columns in Input"); stridedReduction(dots.data_handle(), data.data_handle(), data.extent(1), data.extent(0), init, resource::get_cuda_stream(handle), inplace, main_op, reduce_op, final_op); } else if constexpr (std::is_same_v<LayoutPolicy, raft::col_major>) { RAFT_EXPECTS(static_cast<IndexType>(dots.size()) == data.extent(0), "Output should be equal to number of rows in Input"); stridedReduction(dots.data_handle(), data.data_handle(), data.extent(0), data.extent(1), init, resource::get_cuda_stream(handle), inplace, main_op, reduce_op, final_op); } } /** @} */ // end of group strided_reduction }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/gemv.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. */ #ifndef __GEMV_H #define __GEMV_H #pragma once #include "detail/gemv.hpp" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdarray.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/host_mdarray.hpp> #include <raft/core/host_mdspan.hpp> #include <raft/util/input_validation.hpp> namespace raft { namespace linalg { /** * @brief the wrapper of cublas gemv function * It computes the following equation: y = alpha .* op(A) * x + beta .* y * * @tparam math_t the element type * @tparam DevicePointerMode whether pointers alpha, beta point to device memory * @param [in] handle raft handle * @param [in] trans_a cublas transpose op for A * @param [in] m number of rows of A * @param [in] n number of columns of A * @param [in] alpha host or device scalar * @param [in] A column-major matrix of size [m, n] * @param [in] lda leading dimension of A * @param [in] x vector of length n if trans_a else m * @param [in] incx stride between consecutive elements of x * @param [in] beta host or device scalar * @param [inout] y vector of length m if trans_a else n * @param [in] incy stride between consecutive elements of y * @param [in] stream */ template <typename math_t, bool DevicePointerMode = false> void gemv(raft::resources const& handle, const bool trans_a, const int m, const int n, const math_t* alpha, const math_t* A, const int lda, const math_t* x, const int incx, const math_t* beta, math_t* y, const int incy, cudaStream_t stream) { detail::gemv<math_t, DevicePointerMode>( handle, trans_a, m, n, alpha, A, lda, x, incx, beta, y, incy, stream); } template <typename math_t> void gemv(raft::resources const& handle, const math_t* A, const int n_rows, const int n_cols, const math_t* x, const int incx, math_t* y, const int incy, const bool trans_a, const math_t alpha, const math_t beta, cudaStream_t stream) { detail::gemv(handle, A, n_rows, n_cols, x, incx, y, incy, trans_a, alpha, beta, stream); } /** * y = alpha * op(A) * x + beta * y * * where * * @param handle raft handle * @param A is a column-major matrix of size n_rows_a * n_cols_a. * op(A) is either the transpose operation (trans_a == true) or identity. * @param n_rows_a number of rows in A * @param n_cols_a number of cols in A * @param x is a vector of size `trans_a ? n_rows_a : n_cols_a`. * @param y is a vector of size `trans_a ? n_cols_a : n_rows_a`. * @param trans_a whether to take transpose of a * @param alpha is a scalar scale of Ax. * @param beta is a scalar scale of y. * @param stream stream on which this function is run */ template <typename math_t> void gemv(raft::resources const& handle, const math_t* A, const int n_rows_a, const int n_cols_a, const math_t* x, math_t* y, const bool trans_a, const math_t alpha, const math_t beta, cudaStream_t stream) { detail::gemv(handle, A, n_rows_a, n_cols_a, x, y, trans_a, alpha, beta, stream); } /** * y = op(A) * x * * where * * @param handle raft handle * @param A is a column-major matrix of size n_rows_a * n_cols_a. * op(A) is either the transpose operation (trans_a == true) or identity. * @param n_rows_a number of rows in A * @param n_cols_a number of cols in A * @param x is a vector of size `trans_a ? n_rows_a : n_cols_a`. * @param y is a vector of size `trans_a ? n_cols_a : n_rows_a`. * @param trans_a whether to take transpose of a * @param stream stream on which this function is run */ template <typename math_t> void gemv(raft::resources const& handle, const math_t* A, const int n_rows_a, const int n_cols_a, const math_t* x, math_t* y, const bool trans_a, cudaStream_t stream) { detail::gemv(handle, A, n_rows_a, n_cols_a, x, y, trans_a, stream); } /** * y = alpha * op(A) * x + beta * y * * where * @param handle raft handle * @param A is a column-major matrix of size n_rows_a * n_cols_a. * op(A) is either the transpose operation (trans_a == true) or identity. * @param n_rows_a number of rows in A * @param n_cols_a number of cols in A * @param lda is the leading dimension of A (number of rows); lda must be not smaller than n_rows_a. * set it when you need to use only the first n_rows_a rows of the matrix A, which has * (perhaps, due to padding) lda rows. * @param x is a vector of size `trans_a ? n_rows_a : n_cols_a`. * @param y is a vector of size `trans_a ? n_cols_a : n_rows_a`. * @param trans_a whether to take transpose of a * @param alpha is a scalar scale of Ax. * @param beta is a scalar scale of y. * @param stream stream on which this function is run */ template <typename math_t> void gemv(raft::resources const& handle, const math_t* A, const int n_rows_a, const int n_cols_a, const int lda, const math_t* x, math_t* y, const bool trans_a, const math_t alpha, const math_t beta, cudaStream_t stream) { detail::gemv(handle, A, n_rows_a, n_cols_a, lda, x, y, trans_a, alpha, beta, stream); } /** * y = op(A) * x * * where * @param handle raft handle * @param A is a column-major matrix of size n_rows_a * n_cols_a. * op(A) is either the transpose operation (trans_a == true) or identity. * @param n_rows_a number of rows in A * @param n_cols_a number of cols in A * @param lda is the leading dimension of A (number of rows); lda must be not smaller than n_rows_a. * set it when you need to use only the first n_rows_a rows of the matrix A, which has * (perhaps, due to padding) lda rows. * @param x is a vector of size `trans_a ? n_rows_a : n_cols_a`. * @param y is a vector of size `trans_a ? n_cols_a : n_rows_a`. * @param trans_a whether to take transpose of a * @param stream stream on which this function is run * */ template <typename math_t> void gemv(raft::resources const& handle, const math_t* A, const int n_rows_a, const int n_cols_a, const int lda, const math_t* x, math_t* y, const bool trans_a, cudaStream_t stream) { detail::gemv(handle, A, n_rows_a, n_cols_a, lda, x, y, trans_a, stream); } /** * @defgroup gemv Matrix-Vector Multiplication * @{ */ /** * @brief GEMV function designed for raft::col_major layout for A * It computes y = alpha * op(A) * x + beta * y, where length of y is number * of rows in A while length of x is number of columns in A * If layout for A is provided as raft::row_major, then a transpose of A * is used in the computation, where length of y is number of columns in A * while length of x is number of rows in A * If alpha is not provided, it is assumed to be 1.0 * If beta is not provided, it is assumed to be 0.0 * @tparam ValueType Data type of input/output matrices (float/double) * @tparam IndexType Type of index * @tparam LayoutPolicyX layout of X * @tparam LayoutPolicyY layout of Y * @tparam LayoutPolicyZ layout of Z * @param[in] handle raft handle * @param[in] A input raft::device_matrix_view of size (M, N) * @param[in] x input raft::device_matrix_view of size (N, 1) if A is raft::col_major, else (M, 1) * @param[out] y output raft::device_matrix_view of size (M, 1) if A is raft::col_major, else (N, 1) * @param[in] alpha optional raft::host_scalar_view or raft::device_scalar_view, default 1.0 * @param[in] beta optional raft::host_scalar_view or raft::device_scalar_view, default 0.0 */ template <typename ValueType, typename IndexType, typename LayoutPolicy, typename ScalarIdxType = std::uint32_t, typename ScalarViewType = raft::host_scalar_view<ValueType, ScalarIdxType>, typename = std::enable_if_t<std::disjunction_v< std::is_same<ScalarViewType, raft::host_scalar_view<ValueType, ScalarIdxType>>, std::is_same<ScalarViewType, raft::device_scalar_view<ValueType, ScalarIdxType>>>>> void gemv(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, LayoutPolicy> A, raft::device_vector_view<const ValueType, IndexType> x, raft::device_vector_view<ValueType, IndexType> y, std::optional<ScalarViewType> alpha = std::nullopt, std::optional<ScalarViewType> beta = std::nullopt) { RAFT_EXPECTS(raft::is_row_or_column_major(A), "A is not contiguous"); constexpr auto is_A_col_major = std::is_same_v<typename decltype(A)::layout_type, raft::col_major>; if (is_A_col_major) { RAFT_EXPECTS(x.extent(0) == A.extent(1), "Number of columns of A and length of x should be equal"); RAFT_EXPECTS(y.extent(0) == A.extent(0), "Number of rows of A and length of y should be equal"); } else { RAFT_EXPECTS(x.extent(0) == A.extent(0), "Number of rows of A and length of x should be equal"); RAFT_EXPECTS(y.extent(0) == A.extent(1), "Number of columns of A and length of y should be equal"); } constexpr auto device_mode = std::is_same_v<ScalarViewType, raft::device_scalar_view<ValueType, ScalarIdxType>>; ValueType alpha_value = 1; ValueType beta_value = 0; auto alpha_device = raft::make_device_scalar(handle, alpha_value); auto beta_device = raft::make_device_scalar(handle, beta_value); auto alpha_host = raft::make_host_scalar(alpha_value); auto beta_host = raft::make_host_scalar(beta_value); if constexpr (device_mode) { if (!alpha) { alpha = alpha_device.view(); } if (!beta) { beta = beta_device.view(); } } else { if (!alpha) { alpha = alpha_host.view(); } if (!beta) { beta = beta_host.view(); } } gemv<ValueType, device_mode>(handle, !is_A_col_major, A.extent(0), A.extent(1), alpha.value().data_handle(), A.data_handle(), A.extent(0), x.data_handle(), 1, beta.value().data_handle(), y.data_handle(), 1, resource::get_cuda_stream(handle)); } /** @} */ // end of gemv }; // namespace linalg }; // namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/norm_types.hpp
/* * Copyright (c) 2022, 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 raft { namespace linalg { /** Enum to tell how to compute a norm */ enum NormType : unsigned short { /** L0 (actually not a norm): sum((x_i != 0 ? 1 : 0)) */ L0PseudoNorm = 0, /** L1 norm or Manhattan: sum(abs(x_i)) */ L1Norm = 1, /** L2 norm or Euclidean: sqrt(sum(x_i^2)). Note that in some prims the square root is optional, in which case it can be specified using a boolean or a functor final_op */ L2Norm = 2, /** Linf norm or Chebyshev: max(abs(x_i)) */ LinfNorm }; } // namespace linalg } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/multiply.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. */ #ifndef __MULTIPLY_H #define __MULTIPLY_H #pragma once #include "detail/multiply.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/host_mdspan.hpp> #include <raft/util/input_validation.hpp> namespace raft { namespace linalg { /** * @defgroup ScalarOps Scalar operations on the input buffer * @tparam out_t data-type upon which the math operation will be performed * @tparam in_t input data-type * @tparam IdxType Integer type used to for addressing * @param out the output buffer * @param in the input buffer * @param scalar the scalar used in the operations * @param len number of elements in the input buffer * @param stream cuda stream where to launch work * @{ */ template <typename in_t, typename out_t = in_t, typename IdxType = int> void multiplyScalar(out_t* out, const in_t* in, in_t scalar, IdxType len, cudaStream_t stream) { detail::multiplyScalar(out, in, scalar, len, stream); } /** @} */ /** * @defgroup multiply Multiplication Arithmetic * @{ */ /** * @brief Element-wise multiplication of host scalar * @tparam InType Input Type raft::device_mdspan * @tparam OutType Output Type raft::device_mdspan * @tparam ScalarIdxType Index Type of scalar * @param[in] handle raft::resources * @param[in] in the input buffer * @param[out] out the output buffer * @param[in] scalar the scalar used in the operations * @{ */ template <typename InType, typename OutType, typename ScalarIdxType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void multiply_scalar( raft::resources const& handle, InType in, OutType out, raft::host_scalar_view<const typename InType::value_type, ScalarIdxType> scalar) { using in_value_t = typename InType::value_type; using out_value_t = typename OutType::value_type; RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in), "Input must be contiguous"); RAFT_EXPECTS(out.size() == in.size(), "Size mismatch between Output and Input"); if (out.size() <= std::numeric_limits<std::uint32_t>::max()) { multiplyScalar<in_value_t, out_value_t, std::uint32_t>(out.data_handle(), in.data_handle(), *scalar.data_handle(), static_cast<std::uint32_t>(out.size()), resource::get_cuda_stream(handle)); } else { multiplyScalar<in_value_t, out_value_t, std::uint64_t>(out.data_handle(), in.data_handle(), *scalar.data_handle(), static_cast<std::uint64_t>(out.size()), resource::get_cuda_stream(handle)); } } /** @} */ // end of group multiply }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/svd.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. */ #ifndef __SVD_H #define __SVD_H #pragma once #include "detail/svd.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <optional> namespace raft { namespace linalg { /** * @brief singular value decomposition (SVD) on the column major float type * input matrix using QR method * @param handle: raft handle * @param in: input matrix * @param n_rows: number rows of input matrix * @param n_cols: number columns of input matrix * @param sing_vals: singular values of input matrix * @param left_sing_vecs: left singular values of input matrix * @param right_sing_vecs: right singular values of input matrix * @param trans_right: transpose right vectors or not * @param gen_left_vec: generate left eig vector. Not activated. * @param gen_right_vec: generate right eig vector. Not activated. * @param stream cuda stream */ template <typename T> void svdQR(raft::resources const& handle, T* in, int n_rows, int n_cols, T* sing_vals, T* left_sing_vecs, T* right_sing_vecs, bool trans_right, bool gen_left_vec, bool gen_right_vec, cudaStream_t stream) { detail::svdQR(handle, in, n_rows, n_cols, sing_vals, left_sing_vecs, right_sing_vecs, trans_right, gen_left_vec, gen_right_vec, stream); } template <typename math_t, typename idx_t> void svdEig(raft::resources const& handle, math_t* in, idx_t n_rows, idx_t n_cols, math_t* S, math_t* U, math_t* V, bool gen_left_vec, cudaStream_t stream) { detail::svdEig(handle, in, n_rows, n_cols, S, U, V, gen_left_vec, stream); } /** * @brief on the column major input matrix using Jacobi method * @param handle: raft handle * @param in: input matrix * @param n_rows: number rows of input matrix * @param n_cols: number columns of input matrix * @param sing_vals: singular values of input matrix * @param left_sing_vecs: left singular vectors of input matrix * @param right_sing_vecs: right singular vectors of input matrix * @param gen_left_vec: generate left eig vector. Not activated. * @param gen_right_vec: generate right eig vector. Not activated. * @param tol: error tolerance for the jacobi method. Algorithm stops when the * error is below tol * @param max_sweeps: number of sweeps in the Jacobi algorithm. The more the better * accuracy. * @param stream cuda stream */ template <typename math_t> void svdJacobi(raft::resources const& handle, math_t* in, int n_rows, int n_cols, math_t* sing_vals, math_t* left_sing_vecs, math_t* right_sing_vecs, bool gen_left_vec, bool gen_right_vec, math_t tol, int max_sweeps, cudaStream_t stream) { detail::svdJacobi(handle, in, n_rows, n_cols, sing_vals, left_sing_vecs, right_sing_vecs, gen_left_vec, gen_right_vec, tol, max_sweeps, stream); } /** * @brief reconstruct a matrix use left and right singular vectors and * singular values * @param handle: raft handle * @param U: left singular vectors of size n_rows x k * @param S: square matrix with singular values on its diagonal, k x k * @param V: right singular vectors of size n_cols x k * @param out: reconstructed matrix to be returned * @param n_rows: number rows of output matrix * @param n_cols: number columns of output matrix * @param k: number of singular values * @param stream cuda stream */ template <typename math_t> void svdReconstruction(raft::resources const& handle, math_t* U, math_t* S, math_t* V, math_t* out, int n_rows, int n_cols, int k, cudaStream_t stream) { detail::svdReconstruction(handle, U, S, V, out, n_rows, n_cols, k, stream); } /** * @brief reconstruct a matrix use left and right singular vectors and * singular values * @param handle: raft handle * @param A_d: input matrix * @param U: left singular vectors of size n_rows x k * @param S_vec: singular values as a vector * @param V: right singular vectors of size n_cols x k * @param n_rows: number rows of output matrix * @param n_cols: number columns of output matrix * @param k: number of singular values to be computed, 1.0 for normal SVD * @param tol: tolerance for the evaluation * @param stream cuda stream */ template <typename math_t> bool evaluateSVDByL2Norm(raft::resources const& handle, math_t* A_d, math_t* U, math_t* S_vec, math_t* V, int n_rows, int n_cols, int k, math_t tol, cudaStream_t stream) { return detail::evaluateSVDByL2Norm(handle, A_d, U, S_vec, V, n_rows, n_cols, k, tol, stream); } /** * @defgroup svd Singular Value Decomposition * @{ */ /** * @brief singular value decomposition (SVD) on a column major * matrix using QR decomposition * @tparam ValueType value type of parameters * @tparam IndexType index type of parameters * @param[in] handle raft::resources * @param[in] in input raft::device_matrix_view with layout raft::col_major of shape (M, N) * @param[out] sing_vals singular values raft::device_vector_view of shape (K) * @param[out] U std::optional left singular values of raft::device_matrix_view with layout * raft::col_major and dimensions (m, n) * @param[out] V std::optional right singular values of raft::device_matrix_view with * layout raft::col_major and dimensions (n, n) */ template <typename ValueType, typename IndexType> void svd_qr( raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> in, raft::device_vector_view<ValueType, IndexType> sing_vals, std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::nullopt, std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> V = std::nullopt) { ValueType* left_sing_vecs_ptr = nullptr; ValueType* right_sing_vecs_ptr = nullptr; if (U) { RAFT_EXPECTS(in.extent(0) == U.value().extent(0) && in.extent(1) == U.value().extent(1), "U should have dimensions m * n"); left_sing_vecs_ptr = U.value().data_handle(); } if (V) { RAFT_EXPECTS(in.extent(1) == V.value().extent(0) && in.extent(1) == V.value().extent(1), "V should have dimensions n * n"); right_sing_vecs_ptr = V.value().data_handle(); } svdQR(handle, const_cast<ValueType*>(in.data_handle()), in.extent(0), in.extent(1), sing_vals.data_handle(), left_sing_vecs_ptr, right_sing_vecs_ptr, false, U.has_value(), V.has_value(), resource::get_cuda_stream(handle)); } /** * @brief Overload of `svd_qr` to help the * compiler find the above overload, in case users pass in * `std::nullopt` for one or both of the optional arguments. * * Please see above for documentation of `svd_qr`. */ template <typename ValueType, typename IndexType, typename UType, typename VType> void svd_qr(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> in, raft::device_vector_view<ValueType, IndexType> sing_vals, UType&& U_in = std::nullopt, VType&& V_in = std::nullopt) { std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::forward<UType>(U_in); std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> V = std::forward<VType>(V_in); svd_qr(handle, in, sing_vals, U, V); } /** * @brief singular value decomposition (SVD) on a column major * matrix using QR decomposition. Right singular vector matrix is transposed before returning * @tparam ValueType value type of parameters * @tparam IndexType index type of parameters * @param[in] handle raft::resources * @param[in] in input raft::device_matrix_view with layout raft::col_major of shape (M, N) * @param[out] sing_vals singular values raft::device_vector_view of shape (K) * @param[out] U std::optional left singular values of raft::device_matrix_view with layout * raft::col_major and dimensions (m, n) * @param[out] V std::optional right singular values of raft::device_matrix_view with * layout raft::col_major and dimensions (n, n) */ template <typename ValueType, typename IndexType> void svd_qr_transpose_right_vec( raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> in, raft::device_vector_view<ValueType, IndexType> sing_vals, std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::nullopt, std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> V = std::nullopt) { ValueType* left_sing_vecs_ptr = nullptr; ValueType* right_sing_vecs_ptr = nullptr; if (U) { RAFT_EXPECTS(in.extent(0) == U.value().extent(0) && in.extent(1) == U.value().extent(1), "U should have dimensions m * n"); left_sing_vecs_ptr = U.value().data_handle(); } if (V) { RAFT_EXPECTS(in.extent(1) == V.value().extent(0) && in.extent(1) == V.value().extent(1), "V should have dimensions n * n"); right_sing_vecs_ptr = V.value().data_handle(); } svdQR(handle, const_cast<ValueType*>(in.data_handle()), in.extent(0), in.extent(1), sing_vals.data_handle(), left_sing_vecs_ptr, right_sing_vecs_ptr, true, U.has_value(), V.has_value(), resource::get_cuda_stream(handle)); } /** * @brief Overload of `svd_qr_transpose_right_vec` to help the * compiler find the above overload, in case users pass in * `std::nullopt` for one or both of the optional arguments. * * Please see above for documentation of `svd_qr_transpose_right_vec`. */ template <typename ValueType, typename IndexType, typename UType, typename VType> void svd_qr_transpose_right_vec( raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> in, raft::device_vector_view<ValueType, IndexType> sing_vals, UType&& U_in = std::nullopt, VType&& V_in = std::nullopt) { std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::forward<UType>(U_in); std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> V = std::forward<VType>(V_in); svd_qr_transpose_right_vec(handle, in, sing_vals, U, V); } /** * @brief singular value decomposition (SVD) on a column major * matrix using Eigen decomposition. A square symmetric covariance matrix is constructed for the SVD * @param[in] handle raft::resources * @param[in] in input raft::device_matrix_view with layout raft::col_major of shape (M, N) * @param[out] S singular values raft::device_vector_view of shape (K) * @param[out] V right singular values of raft::device_matrix_view with layout * raft::col_major and dimensions (n, n) * @param[out] U optional left singular values of raft::device_matrix_view with layout * raft::col_major and dimensions (m, n) */ template <typename ValueType, typename IndexType> void svd_eig( raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> in, raft::device_vector_view<ValueType, IndexType> S, raft::device_matrix_view<ValueType, IndexType, raft::col_major> V, std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::nullopt) { ValueType* left_sing_vecs_ptr = nullptr; if (U) { RAFT_EXPECTS(in.extent(0) == U.value().extent(0) && in.extent(1) == U.value().extent(1), "U should have dimensions m * n"); left_sing_vecs_ptr = U.value().data_handle(); } RAFT_EXPECTS(in.extent(1) == V.extent(0) && in.extent(1) == V.extent(1), "V should have dimensions n * n"); svdEig(handle, const_cast<ValueType*>(in.data_handle()), in.extent(0), in.extent(1), S.data_handle(), left_sing_vecs_ptr, V.data_handle(), U.has_value(), resource::get_cuda_stream(handle)); } template <typename ValueType, typename IndexType, typename UType> void svd_eig(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> in, raft::device_vector_view<ValueType, IndexType> S, raft::device_matrix_view<ValueType, IndexType, raft::col_major> V, UType&& U = std::nullopt) { std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U_optional = std::forward<UType>(U); svd_eig(handle, in, S, V, U_optional); } /** * @brief reconstruct a matrix use left and right singular vectors and * singular values * @param[in] handle raft::resources * @param[in] U left singular values of raft::device_matrix_view with layout * raft::col_major and dimensions (m, k) * @param[in] S square matrix with singular values on its diagonal of shape (k, k) * @param[in] V right singular values of raft::device_matrix_view with layout * raft::col_major and dimensions (k, n) * @param[out] out output raft::device_matrix_view with layout raft::col_major of shape (m, n) */ template <typename ValueType, typename IndexType> void svd_reconstruction(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> U, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> S, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> V, raft::device_matrix_view<ValueType, IndexType, raft::col_major> out) { RAFT_EXPECTS(S.extent(0) == S.extent(1), "S should be a square matrix"); RAFT_EXPECTS(S.extent(0) == U.extent(1), "Number of rows of S should be equal to number of columns in U"); RAFT_EXPECTS(S.extent(1) == V.extent(0), "Number of columns of S should be equal to number of rows in V"); RAFT_EXPECTS(out.extent(0) == U.extent(0) && out.extent(1) == V.extent(1), "Number of rows should be equal in out and U and number of columns should be equal " "in out and V"); svdReconstruction(handle, const_cast<ValueType*>(U.data_handle()), const_cast<ValueType*>(S.data_handle()), const_cast<ValueType*>(V.data_handle()), out.data_handle(), out.extent(0), out.extent(1), S.extent(0), resource::get_cuda_stream(handle)); } /** @} */ // end of group svd }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/map_then_reduce.cuh
/* * Copyright (c) 2022, 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 __MAP_THEN_REDUCE_H #define __MAP_THEN_REDUCE_H #pragma once #include "detail/map_then_reduce.cuh" namespace raft { namespace linalg { /** * @brief CUDA version of map and then sum reduction operation * @tparam Type data-type upon which the math operation will be performed * @tparam MapOp the device-lambda performing the actual operation * @tparam TPB threads-per-block in the final kernel launched * @tparam Args additional parameters * @param out the output sum-reduced value (assumed to be a device pointer) * @param len number of elements in the input array * @param map the device-lambda * @param stream cuda-stream where to launch this kernel * @param in the input array * @param args additional input arrays */ template <typename InType, typename MapOp, typename IdxType = std::uint32_t, int TPB = 256, typename... Args, typename OutType = InType> void mapThenSumReduce( OutType* out, IdxType len, MapOp map, cudaStream_t stream, const InType* in, Args... args) { detail::mapThenReduceImpl<InType, OutType, IdxType, MapOp, detail::sum_tag, TPB, Args...>( out, len, (OutType)0, map, detail::sum_tag(), stream, in, args...); } /** * @brief CUDA version of map and then generic reduction operation * @tparam Type data-type upon which the math operation will be performed * @tparam MapOp the device-lambda performing the actual map operation * @tparam ReduceLambda the device-lambda performing the actual reduction * @tparam TPB threads-per-block in the final kernel launched * @tparam Args additional parameters * @param out the output reduced value (assumed to be a device pointer) * @param len number of elements in the input array * @param neutral The neutral element of the reduction operation. For example: * 0 for sum, 1 for multiply, +Inf for Min, -Inf for Max * @param map the device-lambda * @param op the reduction device lambda * @param stream cuda-stream where to launch this kernel * @param in the input array * @param args additional input arrays */ template <typename InType, typename MapOp, typename ReduceLambda, typename IdxType = std::uint32_t, int TPB = 256, typename OutType = InType, typename... Args> [[deprecated("Use function `mapReduce` from `raft/linalg/map_reduce.cuh")]] void mapThenReduce( OutType* out, size_t len, OutType neutral, MapOp map, ReduceLambda op, cudaStream_t stream, const InType* in, Args... args) { detail::mapThenReduceImpl<InType, OutType, IdxType, MapOp, ReduceLambda, TPB, Args...>( out, len, neutral, map, op, stream, in, args...); } }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/contractions.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. */ #ifndef __CONTRACTIONS_H #define __CONTRACTIONS_H #pragma once #include "detail/contractions.cuh" namespace raft { namespace linalg { /** * @brief This is the central enum that should be used to configure the perf * landscape of the Contraction kernel. * * Main goal of this Policy struct is to provide sufficient knobs to tune the * perf of Contraction kernel, as and when we see matrices of different shapes. * * @tparam DataT the IO and math datatype * @tparam _veclen number of k-elements loaded by each thread for every LDG call * it makes. This should be configured based on the input 'k' * value and the input data type. For eg: if DataT = float and * k is multiples of 4, then setting this to 4 gives the best * LDG pattern. Possible values are {1, 2, 4}. * @tparam _kblk number of k-elements operated upon per main-loop iteration. * Therefore total number of main-loop iterations will be * `ceil(k/_kblk)`. This must be multiples of `_veclen`. Do note * that bigger this value, the greater shared mem requirement. * @tparam _rpt Defines the number of rows that a given thread accumulates on. * This directly results in increased register pressure. This * also is used to compute the number of m-elements worked upon * by each thread block. * @tparam _cpt Defines the number of cols that a given thread accumulates on. * This directly results in increased register pressure. This * also is used to compute the number of n-elements worked upon * by each thread block. * @tparam _tr Number of threads working on the same output column. This is * used to compute the number of m-elements worked upon by each * thread block. This also determines the number of threads per * thread block * @tparam _tc Number of threads working on the same output row. This is * used to compute the number of m-elements worked upon by each * thread block. This also determines the number of threads per * thread block */ template <typename DataT, int _veclen, int _kblk, int _rpt, int _cpt, int _tr, int _tc> struct KernelPolicy { enum { /** number of elements along K worked upon per main loop iteration */ Kblk = _kblk, /** number of elements loaded per LDG */ Veclen = _veclen, /** number of rows a thread works on for accumulation */ AccRowsPerTh = _rpt, /** number of cols a thread works on for accumulation */ AccColsPerTh = _cpt, /** number of threads working the same output col */ AccThRows = _tr, /** number of threads working the same output row */ AccThCols = _tc, /** total threads per block */ Nthreads = AccThRows * AccThCols, /** output tile size along rows */ Mblk = AccRowsPerTh * AccThRows, /** output tile size along cols */ Nblk = AccColsPerTh * AccThCols, /** number of threads loading a single row */ LdgThRow = Kblk / Veclen, /** number of LDGs issued by a single thread for X */ LdgPerThX = Mblk * LdgThRow / Nthreads, /** number of LDGs issued by a single thread for Y */ LdgPerThY = Nblk * LdgThRow / Nthreads, /** number of rows of X covered per LDG */ LdgRowsX = Mblk / LdgPerThX, /** number of rows of Y covered per LDG */ LdgRowsY = Nblk / LdgPerThY, /** stride for accessing X/Y data in shared mem */ SmemStride = Kblk + Veclen, /** size of one page for storing X data */ SmemPageX = SmemStride * Mblk, /** size of one page for storing Y data */ SmemPageY = SmemStride * Nblk, /** size of one smem page */ SmemPage = SmemPageX + SmemPageY, /** size (in B) for smem needed */ SmemSize = 2 * SmemPage * sizeof(DataT), }; // enum }; // struct KernelPolicy template <typename DataT, int _veclen, int _kblk, int _rpt, int _cpt, int _tr, int _tc> struct ColKernelPolicy { enum { /** number of elements along K worked upon per main loop iteration */ Kblk = _kblk, /** number of elements loaded per LDG */ Veclen = _veclen, /** number of rows a thread works on for accumulation */ AccRowsPerTh = _rpt, /** number of cols a thread works on for accumulation */ AccColsPerTh = _cpt, /** number of threads working the same output col */ AccThRows = _tr, /** number of threads working the same output row */ AccThCols = _tc, /** total threads per block */ Nthreads = AccThRows * AccThCols, /** output tile size along rows */ Mblk = AccRowsPerTh * AccThRows, /** output tile size along cols */ Nblk = AccColsPerTh * AccThCols, /** number of threads loading a single col */ LdgThRow = Mblk / Veclen, /** number of LDGs issued by a single thread for X */ LdgPerThX = Kblk * LdgThRow / Nthreads, /** number of LDGs issued by a single thread for Y */ LdgPerThY = Kblk * LdgThRow / Nthreads, /** number of rows of X covered per LDG */ LdgRowsX = Kblk / LdgPerThX, /** number of rows of Y covered per LDG */ LdgRowsY = Kblk / LdgPerThY, /** stride for accessing X/Y data in shared mem */ SmemStride = Mblk + Veclen, /** size of one page for storing X data */ SmemPageX = SmemStride * Kblk, /** size of one page for storing Y data */ SmemPageY = SmemStride * Kblk, /** size of one smem page */ SmemPage = SmemPageX + SmemPageY, /** size (in B) for smem needed */ SmemSize = 2 * SmemPage * sizeof(DataT), }; // colMajor enum static_assert(Mblk == Nblk, "Mblk should be equal to Nblk"); }; /** * @defgroup Policy4x4 16 elements per thread Policy with k-block = 32 * @{ */ template <typename DataT, int _veclen> struct Policy4x4 {}; template <int _veclen> struct Policy4x4<float, _veclen> { typedef KernelPolicy<float, _veclen, 32, 4, 4, 16, 16> Policy; typedef ColKernelPolicy<float, _veclen, 32, 4, 4, 16, 16> ColPolicy; }; template <int _veclen> struct Policy4x4<double, _veclen> { typedef KernelPolicy<double, _veclen, 16, 4, 4, 16, 16> Policy; typedef ColKernelPolicy<double, _veclen, 16, 4, 4, 16, 16> ColPolicy; }; /** @} */ /** * A smaller k-block (8 instead of 32) with fewer threads per block (8x8 instead * of 16x16), which is faster for raft::distance::fusedL2NN on skinny matrices, * i.e., matrices with a small k dimension. * */ template <typename DataT, int _veclen> struct Policy4x4Skinny {}; template <int _veclen> struct Policy4x4Skinny<float, _veclen> { typedef KernelPolicy<float, _veclen, 8, 4, 4, 8, 8> Policy; typedef ColKernelPolicy<float, _veclen, 8, 4, 4, 8, 8> ColPolicy; }; template <int _veclen> struct Policy4x4Skinny<double, _veclen> { typedef KernelPolicy<double, _veclen, 8, 4, 4, 8, 8> Policy; typedef ColKernelPolicy<double, _veclen, 8, 4, 4, 8, 8> ColPolicy; }; /** * @defgroup Policy2x8 16 elements per thread Policy with k-block = 16 * @{ */ template <typename DataT, int _veclen = 1> struct Policy2x8 {}; template <int _veclen> struct Policy2x8<float, _veclen> { typedef KernelPolicy<float, _veclen, 16, 2, 8, 8, 32> Policy; }; template <int _veclen> struct Policy2x8<double, _veclen> { // this is not used just for keeping compiler happy. typedef KernelPolicy<double, _veclen, 32, 1, 2, 8, 32> Policy; }; /** @} */ /** * @brief Base class for gemm-like NT contractions * * This class does not provide any arithmetic operations, but only provides the * memory-related operations of loading the `x` and `y` matrix blocks from the * global memory into shared memory and then from shared into registers. Thus, * this class acts as a basic building block for further composing gemm-like NT * contractions on input matrices which are row-major (and so does the output) * * @tparam DataT IO and math data type * @tparam IdxT indexing type * @tparam Policy policy used to customize memory access behavior. * See documentation for `KernelPolicy` to know more. */ using detail::Contractions_NT; } // namespace linalg } // namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/binary_op.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. */ #ifndef __BINARY_OP_H #define __BINARY_OP_H #pragma once #include <raft/core/device_mdspan.hpp> #include <raft/core/resources.hpp> #include <raft/linalg/map.cuh> namespace raft { namespace linalg { /** * @brief perform element-wise binary operation on the input arrays * @tparam InType input data-type * @tparam Lambda the device-lambda performing the actual operation * @tparam OutType output data-type * @tparam IdxType Integer type used to for addressing * @tparam TPB threads-per-block in the final kernel launched * @param out the output array * @param in1 the first input array * @param in2 the second input array * @param len number of elements in the input array * @param op the device-lambda * @param stream cuda stream where to launch work * @note Lambda must be a functor with the following signature: * `OutType func(const InType& val1, const InType& val2);` */ template <typename InType, typename Lambda, typename OutType = InType, typename IdxType = int, int TPB = 256> void binaryOp( OutType* out, const InType* in1, const InType* in2, IdxType len, Lambda op, cudaStream_t stream) { return detail::map<false>(stream, out, len, op, in1, in2); } /** * @defgroup binary_op Element-Wise Binary Operation * @{ */ /** * @brief perform element-wise binary operation on the input arrays * @tparam InType Input Type raft::device_mdspan * @tparam Lambda the device-lambda performing the actual operation * @tparam OutType Output Type raft::device_mdspan * @param[in] handle raft::resources * @param[in] in1 First input * @param[in] in2 Second input * @param[out] out Output * @param[in] op the device-lambda * @note Lambda must be a functor with the following signature: * `OutType func(const InType& val1, const InType& val2);` */ template <typename InType, typename Lambda, typename OutType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void binary_op(raft::resources const& handle, InType in1, InType in2, OutType out, Lambda op) { return map(handle, in1, in2, out, op); } /** @} */ // end of group binary_op }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/sqrt.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. */ #ifndef __SQRT_H #define __SQRT_H #pragma once #include <raft/core/device_mdspan.hpp> #include <raft/core/operators.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/linalg/unary_op.cuh> namespace raft { namespace linalg { /** * @defgroup ScalarOps Scalar operations on the input buffer * @tparam math_t data-type upon which the math operation will be performed * @tparam IdxType Integer type used to for addressing * @param out the output buffer * @param in the input buffer * @param len number of elements in the input buffer * @param stream cuda stream where to launch work * @{ */ template <typename in_t, typename out_t = in_t, typename IdxType = int> void sqrt(out_t* out, const in_t* in, IdxType len, cudaStream_t stream) { raft::linalg::unaryOp(out, in, len, raft::sqrt_op{}, stream); } /** @} */ /** * @defgroup sqrt Sqrt Arithmetic * @{ */ /** * @brief Elementwise sqrt operation * @tparam InType Input Type raft::device_mdspan * @tparam OutType Output Type raft::device_mdspan * @param[in] handle raft::resources * @param[in] in Input * @param[out] out Output */ template <typename InType, typename OutType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void sqrt(raft::resources const& handle, InType in, OutType out) { using in_value_t = typename InType::value_type; using out_value_t = typename OutType::value_type; RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in), "Input 1 must be contiguous"); RAFT_EXPECTS(out.size() == in.size(), "Size mismatch between Output and Inputs"); if (out.size() <= std::numeric_limits<std::uint32_t>::max()) { sqrt<in_value_t, out_value_t, std::uint32_t>(out.data_handle(), in.data_handle(), static_cast<std::uint32_t>(out.size()), resource::get_cuda_stream(handle)); } else { sqrt<in_value_t, out_value_t, std::uint64_t>(out.data_handle(), in.data_handle(), static_cast<std::uint64_t>(out.size()), resource::get_cuda_stream(handle)); } } /** @} */ // end of group add }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/mean_squared_error.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. */ #ifndef __MSE_H #define __MSE_H #pragma once #include "detail/mean_squared_error.cuh" #include <raft/core/device_mdspan.hpp> #include <raft/core/resource/cuda_stream.hpp> namespace raft { namespace linalg { /** * @brief CUDA version mean squared error function mean((A-B)**2) * @tparam math_t data-type upon which the math operation will be performed * @tparam TPB threads-per-block * @param out the output mean squared error value (assumed to be a device pointer) * @param A input array (assumed to be a device pointer) * @param B input array (assumed to be a device pointer) * @param len number of elements in the input arrays * @param weight weight to apply to every term in the mean squared error calculation * @param stream cuda-stream where to launch this kernel */ template <typename in_t, typename out_t, typename idx_t = size_t> void meanSquaredError( out_t* out, const in_t* A, const in_t* B, idx_t len, in_t weight, cudaStream_t stream) { detail::meanSquaredError(out, A, B, len, weight, stream); } /** * @defgroup mean_squared_error Mean Squared Error * @{ */ /** * @brief CUDA version mean squared error function mean((A-B)**2) * @tparam InValueType Input data-type * @tparam IndexType Input/Output index type * @tparam OutValueType Output data-type * @tparam TPB threads-per-block * @param[in] handle raft::resources * @param[in] A input raft::device_vector_view * @param[in] B input raft::device_vector_view * @param[out] out the output mean squared error value of type raft::device_scalar_view * @param[in] weight weight to apply to every term in the mean squared error calculation */ template <typename InValueType, typename IndexType, typename OutValueType> void mean_squared_error(raft::resources const& handle, raft::device_vector_view<const InValueType, IndexType> A, raft::device_vector_view<const InValueType, IndexType> B, raft::device_scalar_view<OutValueType, IndexType> out, OutValueType weight) { RAFT_EXPECTS(A.size() == B.size(), "Size mismatch between inputs"); meanSquaredError(out.data_handle(), A.data_handle(), B.data_handle(), A.extent(0), weight, resource::get_cuda_stream(handle)); } /** @} */ // end of group mean_squared_error }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/dot.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. */ #ifndef __DOT_H #define __DOT_H #pragma once #include <raft/core/resource/cublas_handle.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/linalg/detail/cublas_wrappers.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/host_mdspan.hpp> #include <raft/core/resources.hpp> namespace raft::linalg { /** * @defgroup dot BLAS dot routine * @{ */ /** * @brief Computes the dot product of two vectors. * @param[in] handle raft::resources * @param[in] x First input vector * @param[in] y Second input vector * @param[out] out The output dot product between the x and y vectors. */ template <typename ElementType, typename IndexType, typename ScalarIndexType, typename LayoutPolicy1, typename LayoutPolicy2> void dot(raft::resources const& handle, raft::device_vector_view<const ElementType, IndexType, LayoutPolicy1> x, raft::device_vector_view<const ElementType, IndexType, LayoutPolicy2> y, raft::device_scalar_view<ElementType, ScalarIndexType> out) { RAFT_EXPECTS(x.size() == y.size(), "Size mismatch between x and y input vectors in raft::linalg::dot"); RAFT_CUBLAS_TRY(detail::cublasdot(resource::get_cublas_handle(handle), x.size(), x.data_handle(), x.stride(0), y.data_handle(), y.stride(0), out.data_handle(), resource::get_cuda_stream(handle))); } /** * @brief Computes the dot product of two vectors. * @param[in] handle raft::resources * @param[in] x First input vector * @param[in] y Second input vector * @param[out] out The output dot product between the x and y vectors. */ template <typename ElementType, typename IndexType, typename ScalarIndexType, typename LayoutPolicy1, typename LayoutPolicy2> void dot(raft::resources const& handle, raft::device_vector_view<const ElementType, IndexType, LayoutPolicy1> x, raft::device_vector_view<const ElementType, IndexType, LayoutPolicy2> y, raft::host_scalar_view<ElementType, ScalarIndexType> out) { RAFT_EXPECTS(x.size() == y.size(), "Size mismatch between x and y input vectors in raft::linalg::dot"); RAFT_CUBLAS_TRY(detail::cublasdot(resource::get_cublas_handle(handle), x.size(), x.data_handle(), x.stride(0), y.data_handle(), y.stride(0), out.data_handle(), resource::get_cuda_stream(handle))); } /** @} */ // end of group dot } // namespace raft::linalg #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/init.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. */ #ifndef __INIT_H #define __INIT_H #pragma once #include <raft/linalg/map.cuh> #include <raft/util/cudart_utils.hpp> namespace raft::linalg { /** * @brief Like Python range. * * Fills the output as out[i] = i. * * \param [out] out device array, size [end-start] * \param [in] start of the range * \param [in] end of range (exclusive) * \param [in] stream cuda stream */ template <typename T> void range(T* out, int start, int end, cudaStream_t stream) { return detail::map<true>( stream, out, end - start, compose_op{cast_op<T>{}, add_const_op<int>{start}}); } /** * @brief Like Python range. * * Fills the output as out[i] = i. * * \param [out] out device array, size [n] * \param [in] n length of the array * \param [in] stream cuda stream */ template <typename T, int TPB = 256> void range(T* out, int n, cudaStream_t stream) { return detail::map<true>(stream, out, n, cast_op<T>{}); } /** * @brief Zeros the output. * * \param [out] out device array, size [n] * \param [in] n length of the array * \param [in] stream cuda stream */ template <typename T> void zero(T* out, int n, cudaStream_t stream) { RAFT_CUDA_TRY(cudaMemsetAsync(static_cast<void*>(out), 0, n * sizeof(T), stream)); } } // namespace raft::linalg #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/rsvd.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. */ #ifndef __RSVD_H #define __RSVD_H #pragma once #include "detail/rsvd.cuh" #include <raft/core/device_mdspan.hpp> #include <raft/core/resource/cuda_stream.hpp> namespace raft { namespace linalg { /** * @brief randomized singular value decomposition (RSVD) on the column major * float type input matrix (Jacobi-based), by specifying no. of PCs and * upsamples directly * @param handle: raft handle * @param M: input matrix * @param n_rows: number rows of input matrix * @param n_cols: number columns of input matrix * @param S_vec: singular values of input matrix * @param U: left singular values of input matrix * @param V: right singular values of input matrix * @param k: no. of singular values to be computed * @param p: no. of upsamples * @param use_bbt: whether use eigen decomposition in computation or not * @param gen_left_vec: left vector needs to be generated or not? * @param gen_right_vec: right vector needs to be generated or not? * @param use_jacobi: whether to jacobi solver for decomposition * @param tol: tolerance for Jacobi-based solvers * @param max_sweeps: maximum number of sweeps for Jacobi-based solvers * @param stream cuda stream */ template <typename math_t> void rsvdFixedRank(raft::resources const& handle, math_t* M, int n_rows, int n_cols, math_t* S_vec, math_t* U, math_t* V, int k, int p, bool use_bbt, bool gen_left_vec, bool gen_right_vec, bool use_jacobi, math_t tol, int max_sweeps, cudaStream_t stream) { detail::rsvdFixedRank(handle, M, n_rows, n_cols, S_vec, U, V, k, p, use_bbt, gen_left_vec, gen_right_vec, use_jacobi, tol, max_sweeps, stream); } /** * @brief randomized singular value decomposition (RSVD) on the column major * float type input matrix (Jacobi-based), by specifying the PC and upsampling * ratio * @param handle: raft handle * @param M: input matrix * @param n_rows: number rows of input matrix * @param n_cols: number columns of input matrix * @param S_vec: singular values of input matrix * @param U: left singular values of input matrix * @param V: right singular values of input matrix * @param PC_perc: percentage of singular values to be computed * @param UpS_perc: upsampling percentage * @param use_bbt: whether use eigen decomposition in computation or not * @param gen_left_vec: left vector needs to be generated or not? * @param gen_right_vec: right vector needs to be generated or not? * @param use_jacobi: whether to jacobi solver for decomposition * @param tol: tolerance for Jacobi-based solvers * @param max_sweeps: maximum number of sweeps for Jacobi-based solvers * @param stream cuda stream */ template <typename math_t> void rsvdPerc(raft::resources const& handle, math_t* M, int n_rows, int n_cols, math_t* S_vec, math_t* U, math_t* V, math_t PC_perc, math_t UpS_perc, bool use_bbt, bool gen_left_vec, bool gen_right_vec, bool use_jacobi, math_t tol, int max_sweeps, cudaStream_t stream) { detail::rsvdPerc(handle, M, n_rows, n_cols, S_vec, U, V, PC_perc, UpS_perc, use_bbt, gen_left_vec, gen_right_vec, use_jacobi, tol, max_sweeps, stream); } /** * @defgroup rsvd Randomized Singular Value Decomposition * @{ */ /** * @brief randomized singular value decomposition (RSVD) on a column major * rectangular matrix using QR decomposition, by specifying no. of PCs and * upsamples directly * @tparam ValueType value type of parameters * @tparam IndexType index type of parameters * @tparam UType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * U_in * @tparam VType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * V_in * @param[in] handle raft::resources * @param[in] M input raft::device_matrix_view with layout raft::col_major of shape (M, N) * @param[out] S_vec singular values raft::device_vector_view of shape (K) * @param[in] p no. of upsamples * @param[out] U_in std::optional left singular values of raft::device_matrix_view with layout * raft::col_major * @param[out] V_in std::optional right singular values of raft::device_matrix_view with layout * raft::col_major */ template <typename ValueType, typename IndexType, typename UType, typename VType> void rsvd_fixed_rank(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> M, raft::device_vector_view<ValueType, IndexType> S_vec, IndexType p, UType&& U_in, VType&& V_in) { std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::forward<UType>(U_in); std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> V = std::forward<VType>(V_in); ValueType* U_ptr = nullptr; ValueType* V_ptr = nullptr; if (U) { RAFT_EXPECTS(M.extent(0) == U.value().extent(0), "Number of rows in M should be equal to U"); RAFT_EXPECTS(S_vec.extent(0) == U.value().extent(1), "Number of columns in U should be equal to length of S"); U_ptr = U.value().data_handle(); } if (V) { RAFT_EXPECTS(M.extent(1) == V.value().extent(1), "Number of columns in M should be equal to V"); RAFT_EXPECTS(S_vec.extent(0) == V.value().extent(0), "Number of rows in V should be equal to length of S"); V_ptr = V.value().data_handle(); } rsvdFixedRank(handle, const_cast<ValueType*>(M.data_handle()), M.extent(0), M.extent(1), S_vec.data_handle(), U_ptr, V_ptr, S_vec.extent(0), p, false, U.has_value(), V.has_value(), false, static_cast<ValueType>(0), 0, resource::get_cuda_stream(handle)); } /** * @brief Overload of `rsvd_fixed_rank` to help the * compiler find the above overload, in case users pass in * `std::nullopt` for one or both of the optional arguments. * * Please see above for documentation of `rsvd_fixed_rank`. */ template <typename... Args, typename = std::enable_if_t<sizeof...(Args) == 4>> void rsvd_fixed_rank(Args... args) { rsvd_fixed_rank(std::forward<Args>(args)..., std::nullopt, std::nullopt); } /** * @brief randomized singular value decomposition (RSVD) on a column major * rectangular matrix using symmetric Eigen decomposition, by specifying no. of PCs and * upsamples directly. The rectangular input matrix is made square and symmetric using B @ B^T * @tparam ValueType value type of parameters * @tparam IndexType index type of parameters * @tparam UType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * U_in * @tparam VType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * V_in * @param[in] handle raft::resources * @param[in] M input raft::device_matrix_view with layout raft::col_major of shape (M, N) * @param[out] S_vec singular values raft::device_vector_view of shape (K) * @param[in] p no. of upsamples * @param[out] U_in std::optional left singular values of raft::device_matrix_view with layout * raft::col_major * @param[out] V_in std::optional right singular values of raft::device_matrix_view with layout * raft::col_major */ template <typename ValueType, typename IndexType, typename UType, typename VType> void rsvd_fixed_rank_symmetric( raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> M, raft::device_vector_view<ValueType, IndexType> S_vec, IndexType p, UType&& U_in, VType&& V_in) { std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::forward<UType>(U_in); std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> V = std::forward<VType>(V_in); ValueType* U_ptr = nullptr; ValueType* V_ptr = nullptr; if (U) { U_ptr = U.value().data_handle(); RAFT_EXPECTS(M.extent(0) == U.value().extent(0), "Number of rows in M should be equal to U"); RAFT_EXPECTS(S_vec.extent(0) == U.value().extent(1), "Number of columns in U should be equal to length of S"); } if (V) { V_ptr = V.value().data_handle(); RAFT_EXPECTS(M.extent(1) == V.value().extent(1), "Number of columns in M should be equal to V"); RAFT_EXPECTS(S_vec.extent(0) == V.value().extent(0), "Number of rows in V should be equal to length of S"); } rsvdFixedRank(handle, const_cast<ValueType*>(M.data_handle()), M.extent(0), M.extent(1), S_vec.data_handle(), U_ptr, V_ptr, S_vec.extent(0), p, true, U.has_value(), V.has_value(), false, static_cast<ValueType>(0), 0, resource::get_cuda_stream(handle)); } /** * @brief Overload of `rsvd_fixed_rank_symmetric` to help the * compiler find the above overload, in case users pass in * `std::nullopt` for one or both of the optional arguments. * * Please see above for documentation of `rsvd_fixed_rank_symmetric`. */ template <typename... Args, typename = std::enable_if_t<sizeof...(Args) == 4>> void rsvd_fixed_rank_symmetric(Args... args) { rsvd_fixed_rank_symmetric(std::forward<Args>(args)..., std::nullopt, std::nullopt); } /** * @brief randomized singular value decomposition (RSVD) on a column major * rectangular matrix using Jacobi method, by specifying no. of PCs and * upsamples directly * @tparam ValueType value type of parameters * @tparam IndexType index type of parameters * @tparam UType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * U_in * @tparam VType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * V_in * @param[in] handle raft::resources * @param[in] M input raft::device_matrix_view with layout raft::col_major of shape (M, N) * @param[out] S_vec singular values raft::device_vector_view of shape (K) * @param[in] p no. of upsamples * @param[in] tol tolerance for Jacobi-based solvers * @param[in] max_sweeps maximum number of sweeps for Jacobi-based solvers * @param[out] U_in std::optional left singular values of raft::device_matrix_view with layout * raft::col_major * @param[out] V_in std::optional right singular values of raft::device_matrix_view with layout * raft::col_major */ template <typename ValueType, typename IndexType, typename UType, typename VType> void rsvd_fixed_rank_jacobi(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> M, raft::device_vector_view<ValueType, IndexType> S_vec, IndexType p, ValueType tol, int max_sweeps, UType&& U_in, VType&& V_in) { std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::forward<UType>(U_in); std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> V = std::forward<VType>(V_in); ValueType* U_ptr = nullptr; ValueType* V_ptr = nullptr; if (U) { U_ptr = U.value().data_handle(); RAFT_EXPECTS(M.extent(0) == U.value().extent(0), "Number of rows in M should be equal to U"); RAFT_EXPECTS(S_vec.extent(0) == U.value().extent(1), "Number of columns in U should be equal to length of S"); } if (V) { V_ptr = V.value().data_handle(); RAFT_EXPECTS(M.extent(1) == V.value().extent(1), "Number of columns in M should be equal to V"); RAFT_EXPECTS(S_vec.extent(0) == V.value().extent(0), "Number of rows in V should be equal to length of S"); } rsvdFixedRank(handle, const_cast<ValueType*>(M.data_handle()), M.extent(0), M.extent(1), S_vec.data_handle(), U_ptr, V_ptr, S_vec.extent(0), p, false, U.has_value(), V.has_value(), true, tol, max_sweeps, resource::get_cuda_stream(handle)); } /** * @brief Overload of `rsvd_fixed_rank_jacobi` to help the * compiler find the above overload, in case users pass in * `std::nullopt` for one or both of the optional arguments. * * Please see above for documentation of `rsvd_fixed_rank_jacobi`. */ template <typename... Args, typename = std::enable_if_t<sizeof...(Args) == 6>> void rsvd_fixed_rank_jacobi(Args... args) { rsvd_fixed_rank_jacobi(std::forward<Args>(args)..., std::nullopt, std::nullopt); } /** * @brief randomized singular value decomposition (RSVD) on a column major * rectangular matrix using Jacobi method, by specifying no. of PCs and * upsamples directly. The rectangular input matrix is made square and symmetric using B @ B^T * @tparam ValueType value type of parameters * @tparam IndexType index type of parameters * @tparam UType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * U_in * @tparam VType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * V_in * @param[in] handle raft::resources * @param[in] M input raft::device_matrix_view with layout raft::col_major of shape (M, N) * @param[out] S_vec singular values raft::device_vector_view of shape (K) * @param[in] p no. of upsamples * @param[in] tol tolerance for Jacobi-based solvers * @param[in] max_sweeps maximum number of sweeps for Jacobi-based solvers * @param[out] U_in std::optional left singular values of raft::device_matrix_view with layout * raft::col_major * @param[out] V_in std::optional right singular values of raft::device_matrix_view with layout * raft::col_major */ template <typename ValueType, typename IndexType, typename UType, typename VType> void rsvd_fixed_rank_symmetric_jacobi( raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> M, raft::device_vector_view<ValueType, IndexType> S_vec, IndexType p, ValueType tol, int max_sweeps, UType&& U_in, VType&& V_in) { std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::forward<UType>(U_in); std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> V = std::forward<VType>(V_in); ValueType* U_ptr = nullptr; ValueType* V_ptr = nullptr; if (U) { U_ptr = U.value().data_handle(); RAFT_EXPECTS(M.extent(0) == U.value().extent(0), "Number of rows in M should be equal to U"); RAFT_EXPECTS(S_vec.extent(0) == U.value().extent(1), "Number of columns in U should be equal to length of S"); } if (V) { V_ptr = V.value().data_handle(); RAFT_EXPECTS(M.extent(1) == V.value().extent(1), "Number of columns in M should be equal to V"); RAFT_EXPECTS(S_vec.extent(0) == V.value().extent(0), "Number of rows in V should be equal to length of S"); } rsvdFixedRank(handle, const_cast<ValueType*>(M.data_handle()), M.extent(0), M.extent(1), S_vec.data_handle(), U_ptr, V_ptr, S_vec.extent(0), p, true, U.has_value(), V.has_value(), true, tol, max_sweeps, resource::get_cuda_stream(handle)); } /** * @brief Overload of `rsvd_fixed_rank_symmetric_jacobi` to help the * compiler find the above overload, in case users pass in * `std::nullopt` for one or both of the optional arguments. * * Please see above for documentation of `rsvd_fixed_rank_symmetric_jacobi`. */ template <typename... Args, typename = std::enable_if_t<sizeof...(Args) == 6>> void rsvd_fixed_rank_symmetric_jacobi(Args... args) { rsvd_fixed_rank_symmetric_jacobi(std::forward<Args>(args)..., std::nullopt, std::nullopt); } /** * @brief randomized singular value decomposition (RSVD) on a column major * rectangular matrix using QR decomposition, by specifying the PC and upsampling * ratio * @tparam ValueType value type of parameters * @tparam IndexType index type of parameters * @tparam UType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * U_in * @tparam VType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * V_in * @param[in] handle raft::resources * @param[in] M input raft::device_matrix_view with layout raft::col_major of shape (M, N) * @param[out] S_vec singular values raft::device_vector_view of shape (K) * @param[in] PC_perc percentage of singular values to be computed * @param[in] UpS_perc upsampling percentage * @param[out] U_in std::optional left singular values of raft::device_matrix_view with layout * raft::col_major * @param[out] V_in std::optional right singular values of raft::device_matrix_view with layout * raft::col_major */ template <typename ValueType, typename IndexType, typename UType, typename VType> void rsvd_perc(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> M, raft::device_vector_view<ValueType, IndexType> S_vec, ValueType PC_perc, ValueType UpS_perc, UType&& U_in, VType&& V_in) { std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::forward<UType>(U_in); std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> V = std::forward<VType>(V_in); ValueType* U_ptr = nullptr; ValueType* V_ptr = nullptr; if (U) { U_ptr = U.value().data_handle(); RAFT_EXPECTS(M.extent(0) == U.value().extent(0), "Number of rows in M should be equal to U"); RAFT_EXPECTS(S_vec.extent(0) == U.value().extent(1), "Number of columns in U should be equal to length of S"); } if (V) { V_ptr = V.value().data_handle(); RAFT_EXPECTS(M.extent(1) == V.value().extent(1), "Number of columns in M should be equal to V"); RAFT_EXPECTS(S_vec.extent(0) == V.value().extent(0), "Number of rows in V should be equal to length of S"); } rsvdPerc(handle, const_cast<ValueType*>(M.data_handle()), M.extent(0), M.extent(1), S_vec.data_handle(), U_ptr, V_ptr, PC_perc, UpS_perc, false, U.has_value(), V.has_value(), false, static_cast<ValueType>(0), 0, resource::get_cuda_stream(handle)); } /** * @brief Overload of `rsvd_perc` to help the * compiler find the above overload, in case users pass in * `std::nullopt` for one or both of the optional arguments. * * Please see above for documentation of `rsvd_perc`. */ template <typename... Args, typename = std::enable_if_t<sizeof...(Args) == 5>> void rsvd_perc(Args... args) { rsvd_perc(std::forward<Args>(args)..., std::nullopt, std::nullopt); } /** * @brief randomized singular value decomposition (RSVD) on a column major * rectangular matrix using symmetric Eigen decomposition, by specifying the PC and upsampling * ratio. The rectangular input matrix is made square and symmetric using B @ B^T * @tparam ValueType value type of parameters * @tparam IndexType index type of parameters * @tparam UType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * U_in * @tparam VType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * V_in * @param[in] handle raft::resources * @param[in] M input raft::device_matrix_view with layout raft::col_major of shape (M, N) * @param[out] S_vec singular values raft::device_vector_view of shape (K) * @param[in] PC_perc percentage of singular values to be computed * @param[in] UpS_perc upsampling percentage * @param[out] U_in std::optional left singular values of raft::device_matrix_view with layout * raft::col_major * @param[out] V_in std::optional right singular values of raft::device_matrix_view with layout * raft::col_major */ template <typename ValueType, typename IndexType, typename UType, typename VType> void rsvd_perc_symmetric(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> M, raft::device_vector_view<ValueType, IndexType> S_vec, ValueType PC_perc, ValueType UpS_perc, UType&& U_in, VType&& V_in) { std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::forward<UType>(U_in); std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> V = std::forward<VType>(V_in); ValueType* U_ptr = nullptr; ValueType* V_ptr = nullptr; if (U) { U_ptr = U.value().data_handle(); RAFT_EXPECTS(M.extent(0) == U.value().extent(0), "Number of rows in M should be equal to U"); RAFT_EXPECTS(S_vec.extent(0) == U.value().extent(1), "Number of columns in U should be equal to length of S"); } if (V) { V_ptr = V.value().data_handle(); RAFT_EXPECTS(M.extent(1) == V.value().extent(1), "Number of columns in M should be equal to V"); RAFT_EXPECTS(S_vec.extent(0) == V.value().extent(0), "Number of rows in V should be equal to length of S"); } rsvdPerc(handle, const_cast<ValueType*>(M.data_handle()), M.extent(0), M.extent(1), S_vec.data_handle(), U_ptr, V_ptr, PC_perc, UpS_perc, true, U.has_value(), V.has_value(), false, static_cast<ValueType>(0), 0, resource::get_cuda_stream(handle)); } /** * @brief Overload of `rsvd_perc_symmetric` to help the * compiler find the above overload, in case users pass in * `std::nullopt` for one or both of the optional arguments. * * Please see above for documentation of `rsvd_perc_symmetric`. */ template <typename... Args, typename = std::enable_if_t<sizeof...(Args) == 5>> void rsvd_perc_symmetric(Args... args) { rsvd_perc_symmetric(std::forward<Args>(args)..., std::nullopt, std::nullopt); } /** * @brief randomized singular value decomposition (RSVD) on a column major * rectangular matrix using Jacobi method, by specifying the PC and upsampling * ratio * @tparam ValueType value type of parameters * @tparam IndexType index type of parameters * @tparam UType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * U_in * @tparam VType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * V_in * @param[in] handle raft::resources * @param[in] M input raft::device_matrix_view with layout raft::col_major of shape (M, N) * @param[out] S_vec singular values raft::device_vector_view of shape (K) * @param[in] PC_perc percentage of singular values to be computed * @param[in] UpS_perc upsampling percentage * @param[in] tol tolerance for Jacobi-based solvers * @param[in] max_sweeps maximum number of sweeps for Jacobi-based solvers * @param[out] U_in std::optional left singular values of raft::device_matrix_view with layout * raft::col_major * @param[out] V_in std::optional right singular values of raft::device_matrix_view with layout * raft::col_major */ template <typename ValueType, typename IndexType, typename UType, typename VType> void rsvd_perc_jacobi(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> M, raft::device_vector_view<ValueType, IndexType> S_vec, ValueType PC_perc, ValueType UpS_perc, ValueType tol, int max_sweeps, UType&& U_in, VType&& V_in) { std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::forward<UType>(U_in); std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> V = std::forward<VType>(V_in); ValueType* U_ptr = nullptr; ValueType* V_ptr = nullptr; if (U) { U_ptr = U.value().data_handle(); RAFT_EXPECTS(M.extent(0) == U.value().extent(0), "Number of rows in M should be equal to U"); RAFT_EXPECTS(S_vec.extent(0) == U.value().extent(1), "Number of columns in U should be equal to length of S"); } if (V) { V_ptr = V.value().data_handle(); RAFT_EXPECTS(M.extent(1) == V.value().extent(1), "Number of columns in M should be equal to V"); RAFT_EXPECTS(S_vec.extent(0) == V.value().extent(0), "Number of rows in V should be equal to length of S"); } rsvdPerc(handle, const_cast<ValueType*>(M.data_handle()), M.extent(0), M.extent(1), S_vec.data_handle(), U_ptr, V_ptr, PC_perc, UpS_perc, false, U.has_value(), V.has_value(), true, tol, max_sweeps, resource::get_cuda_stream(handle)); } /** * @brief Overload of `rsvd_perc_jacobi` to help the * compiler find the above overload, in case users pass in * `std::nullopt` for one or both of the optional arguments. * * Please see above for documentation of `rsvd_perc_jacobi`. */ template <typename... Args, typename = std::enable_if_t<sizeof...(Args) == 7>> void rsvd_perc_jacobi(Args... args) { rsvd_perc_jacobi(std::forward<Args>(args)..., std::nullopt, std::nullopt); } /** * @brief randomized singular value decomposition (RSVD) on a column major * rectangular matrix using Jacobi method, by specifying the PC and upsampling * ratio. The rectangular input matrix is made square and symmetric using B @ B^T * @tparam ValueType value type of parameters * @tparam IndexType index type of parameters * @tparam UType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * U_in * @tparam VType std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> @c * V_in * @param[in] handle raft::resources * @param[in] M input raft::device_matrix_view with layout raft::col_major of shape (M, N) * @param[out] S_vec singular values raft::device_vector_view of shape (K) * @param[in] PC_perc percentage of singular values to be computed * @param[in] UpS_perc upsampling percentage * @param[in] tol tolerance for Jacobi-based solvers * @param[in] max_sweeps maximum number of sweeps for Jacobi-based solvers * @param[out] U_in std::optional left singular values of raft::device_matrix_view with layout * raft::col_major * @param[out] V_in std::optional right singular values of raft::device_matrix_view with layout * raft::col_major */ template <typename ValueType, typename IndexType, typename UType, typename VType> void rsvd_perc_symmetric_jacobi( raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> M, raft::device_vector_view<ValueType, IndexType> S_vec, ValueType PC_perc, ValueType UpS_perc, ValueType tol, int max_sweeps, UType&& U_in, VType&& V_in) { std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> U = std::forward<UType>(U_in); std::optional<raft::device_matrix_view<ValueType, IndexType, raft::col_major>> V = std::forward<VType>(V_in); ValueType* U_ptr = nullptr; ValueType* V_ptr = nullptr; if (U) { U_ptr = U.value().data_handle(); RAFT_EXPECTS(M.extent(0) == U.value().extent(0), "Number of rows in M should be equal to U"); RAFT_EXPECTS(S_vec.extent(0) == U.value().extent(1), "Number of columns in U should be equal to length of S"); } if (V) { V_ptr = V.value().data_handle(); RAFT_EXPECTS(M.extent(1) == V.value().extent(1), "Number of columns in M should be equal to V"); RAFT_EXPECTS(S_vec.extent(0) == V.value().extent(0), "Number of rows in V should be equal to length of S"); } rsvdPerc(handle, const_cast<ValueType*>(M.data_handle()), M.extent(0), M.extent(1), S_vec.data_handle(), U_ptr, V_ptr, PC_perc, UpS_perc, true, U.has_value(), V.has_value(), true, tol, max_sweeps, resource::get_cuda_stream(handle)); } /** * @brief Overload of `rsvd_perc_symmetric_jacobi` to help the * compiler find the above overload, in case users pass in * `std::nullopt` for one or both of the optional arguments. * * Please see above for documentation of `rsvd_perc_symmetric_jacobi`. */ template <typename... Args, typename = std::enable_if_t<sizeof...(Args) == 7>> void rsvd_perc_symmetric_jacobi(Args... args) { rsvd_perc_symmetric_jacobi(std::forward<Args>(args)..., std::nullopt, std::nullopt); } /** * @brief randomized singular value decomposition (RSVD) using cusolver * @tparam math_t the data type * @tparam idx_t index type * @param[in] handle: raft handle * @param[in] in: input matrix in col-major format. * Warning: the content of this matrix is modified by the cuSOLVER routines. * [dim = n_rows * n_cols] * @param[out] S: array of singular values of input matrix. The rank k must be less than * min(m,n). [dim = k] * @param[out] U: optional left singular values of input matrix. Use std::nullopt to not * generate it. [dim = n_rows * k] * @param[out] V: optional right singular values of input matrix. Use std::nullopt to not * generate it. [dim = k * n_cols] * @param[in] p: Oversampling. The size of the subspace will be (k + p). (k+p) is less than * min(m,n). (Recommended to be at least 2*k) * @param[in] niters: Number of iteration of power method. (2 is recommended) */ template <typename math_t, typename idx_t> void randomized_svd(const raft::resources& handle, raft::device_matrix_view<const math_t, idx_t, raft::col_major> in, raft::device_vector_view<math_t, idx_t> S, std::optional<raft::device_matrix_view<math_t, idx_t, raft::col_major>> U, std::optional<raft::device_matrix_view<math_t, idx_t, raft::col_major>> V, std::size_t p, std::size_t niters) { auto k = S.extent(0); math_t* left_sing_vecs_ptr = nullptr; math_t* right_sing_vecs_ptr = nullptr; auto gen_U = U.has_value(); auto gen_V = V.has_value(); if (gen_U) { RAFT_EXPECTS(in.extent(0) == U.value().extent(0) && k == U.value().extent(1), "U should have dimensions n_rows * k"); left_sing_vecs_ptr = U.value().data_handle(); } if (gen_V) { RAFT_EXPECTS(k == V.value().extent(0) && in.extent(1) == V.value().extent(1), "V should have dimensions k * n_cols"); right_sing_vecs_ptr = V.value().data_handle(); } detail::randomized_svd(handle, in.data_handle(), in.extent(0), in.extent(1), k, p, niters, S.data_handle(), left_sing_vecs_ptr, right_sing_vecs_ptr, gen_U, gen_V); } /** * @brief Overload of `randomized_svd` to help the * compiler find the above overload, in case users pass in * `std::nullopt` for the optional arguments. * * Please see above for documentation of `randomized_svd`. */ template <typename math_t, typename idx_t, typename opt_u_vec_t, typename opt_v_vec_t> void randomized_svd(const raft::resources& handle, raft::device_matrix_view<const math_t, idx_t, raft::col_major> in, raft::device_vector_view<math_t, idx_t> S, opt_u_vec_t&& U, opt_v_vec_t&& V, std::size_t p, std::size_t niters) { std::optional<raft::device_matrix_view<math_t, idx_t, raft::col_major>> opt_u = std::forward<opt_u_vec_t>(U); std::optional<raft::device_matrix_view<math_t, idx_t, raft::col_major>> opt_v = std::forward<opt_v_vec_t>(V); randomized_svd(handle, in, S, opt_u, opt_v, p, niters); } /** @} */ // end of group rsvd }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/coalesced_reduction.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. */ #ifndef __COALESCED_REDUCTION_H #define __COALESCED_REDUCTION_H #pragma once #include "detail/coalesced_reduction.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/operators.hpp> #include <raft/core/resources.hpp> namespace raft { namespace linalg { /** * @brief Compute reduction of the input matrix along the leading dimension * * @tparam InType the data type of the input * @tparam OutType the data type of the output (as well as the data type for * which reduction is performed) * @tparam IdxType data type of the indices of the array * @tparam MainLambda Unary lambda applied while acculumation (eg: L1 or L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*MainLambda)(InType, IdxType);</pre> * @tparam ReduceLambda Binary lambda applied for reduction (eg: addition(+) for L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*ReduceLambda)(OutType);</pre> * @tparam FinalLambda the final lambda applied before STG (eg: Sqrt for L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*FinalLambda)(OutType);</pre> * @param dots the output reduction vector * @param data the input matrix * @param D leading dimension of data * @param N second dimension data * @param init initial value to use for the reduction * @param main_op elementwise operation to apply before reduction * @param reduce_op binary reduction operation * @param final_op elementwise operation to apply before storing results * @param inplace reduction result added inplace or overwrites old values? * @param stream cuda stream where to launch work */ template <typename InType, typename OutType = InType, typename IdxType = int, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void coalescedReduction(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, cudaStream_t stream, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { detail::coalescedReduction<InType, OutType, IdxType>( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } /** * @defgroup coalesced_reduction Coalesced Memory Access Reductions * For reducing along rows for col-major and along columns for row-major * @{ */ /** * @brief Compute reduction of the input matrix along the leading dimension * This API is to be used when the desired reduction is along the dimension * of the memory layout. For example, a row-major matrix will be reduced * along the columns whereas a column-major matrix will be reduced along * the rows. * * @tparam InValueType the input data-type of underlying raft::matrix_view * @tparam LayoutPolicy The layout of Input/Output (row or col major) * @tparam OutValueType the output data-type of underlying raft::matrix_view and reduction * @tparam IndexType Integer type used to for addressing * @tparam MainLambda Unary lambda applied while acculumation (eg: L1 or L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*MainLambda)(InType, IdxType);</pre> * @tparam ReduceLambda Binary lambda applied for reduction (eg: addition(+) for L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*ReduceLambda)(OutType);</pre> * @tparam FinalLambda the final lambda applied before STG (eg: Sqrt for L2 norm) * It must be a 'callable' supporting the following input and output: * <pre>OutType (*FinalLambda)(OutType);</pre> * @param handle raft::resources * @param[in] data Input of type raft::device_matrix_view * @param[out] dots Output of type raft::device_matrix_view * @param[in] init initial value to use for the reduction * @param[in] inplace reduction result added inplace or overwrites old values? * @param[in] main_op fused elementwise operation to apply before reduction * @param[in] reduce_op fused binary reduction operation * @param[in] final_op fused elementwise operation to apply before storing results */ template <typename InValueType, typename LayoutPolicy, typename OutValueType, typename IdxType, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void coalesced_reduction(raft::resources const& handle, raft::device_matrix_view<const InValueType, IdxType, LayoutPolicy> data, raft::device_vector_view<OutValueType, IdxType> dots, OutValueType init, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { if constexpr (std::is_same_v<LayoutPolicy, raft::row_major>) { RAFT_EXPECTS(static_cast<IdxType>(dots.size()) == data.extent(0), "Output should be equal to number of rows in Input"); coalescedReduction(dots.data_handle(), data.data_handle(), data.extent(1), data.extent(0), init, resource::get_cuda_stream(handle), inplace, main_op, reduce_op, final_op); } else if constexpr (std::is_same_v<LayoutPolicy, raft::col_major>) { RAFT_EXPECTS(static_cast<IdxType>(dots.size()) == data.extent(1), "Output should be equal to number of columns in Input"); coalescedReduction(dots.data_handle(), data.data_handle(), data.extent(0), data.extent(1), init, resource::get_cuda_stream(handle), inplace, main_op, reduce_op, final_op); } } /** @} */ // end of group coalesced_reduction }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/norm.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. */ #ifndef __NORM_H #define __NORM_H #pragma once #include "detail/norm.cuh" #include "linalg_types.hpp" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/operators.hpp> #include <raft/linalg/norm_types.hpp> #include <raft/util/input_validation.hpp> namespace raft { namespace linalg { /** * @brief Compute row-wise norm of the input matrix and perform fin_op lambda * * Row-wise norm is useful while computing pairwise distance matrix, for * example. * This is used in many clustering algos like knn, kmeans, dbscan, etc... * * @tparam Type the data type * @tparam Lambda device final lambda * @tparam IdxType Integer type used to for addressing * @param dots the output vector of row-wise dot products * @param data the input matrix * @param D number of columns of data * @param N number of rows of data * @param type the type of norm to be applied * @param rowMajor whether the input is row-major or not * @param stream cuda stream where to launch work * @param fin_op the final lambda op */ template <typename Type, typename IdxType = int, typename Lambda = raft::identity_op> void rowNorm(Type* dots, const Type* data, IdxType D, IdxType N, NormType type, bool rowMajor, cudaStream_t stream, Lambda fin_op = raft::identity_op()) { detail::rowNormCaller(dots, data, D, N, type, rowMajor, stream, fin_op); } /** * @brief Compute column-wise norm of the input matrix and perform fin_op * @tparam Type the data type * @tparam Lambda device final lambda * @tparam IdxType Integer type used to for addressing * @param dots the output vector of column-wise dot products * @param data the input matrix * @param D number of columns of data * @param N number of rows of data * @param type the type of norm to be applied * @param rowMajor whether the input is row-major or not * @param stream cuda stream where to launch work * @param fin_op the final lambda op */ template <typename Type, typename IdxType = int, typename Lambda = raft::identity_op> void colNorm(Type* dots, const Type* data, IdxType D, IdxType N, NormType type, bool rowMajor, cudaStream_t stream, Lambda fin_op = raft::identity_op()) { detail::colNormCaller(dots, data, D, N, type, rowMajor, stream, fin_op); } /** * @defgroup norm Row- or Col-norm computation * @{ */ /** * @brief Compute norm of the input matrix and perform fin_op * @tparam ElementType Input/Output data type * @tparam LayoutPolicy the layout of input (raft::row_major or raft::col_major) * @tparam IdxType Integer type used to for addressing * @tparam Lambda device final lambda * @param[in] handle raft::resources * @param[in] in the input raft::device_matrix_view * @param[out] out the output raft::device_vector_view * @param[in] type the type of norm to be applied * @param[in] apply Whether to apply the norm along rows (raft::linalg::Apply::ALONG_ROWS) or along columns (raft::linalg::Apply::ALONG_COLUMNS) * @param[in] fin_op the final lambda op */ template <typename ElementType, typename LayoutPolicy, typename IndexType, typename Lambda = raft::identity_op> void norm(raft::resources const& handle, raft::device_matrix_view<const ElementType, IndexType, LayoutPolicy> in, raft::device_vector_view<ElementType, IndexType> out, NormType type, Apply apply, Lambda fin_op = raft::identity_op()) { RAFT_EXPECTS(raft::is_row_or_column_major(in), "Input must be contiguous"); auto constexpr row_major = std::is_same_v<LayoutPolicy, raft::row_major>; auto along_rows = apply == Apply::ALONG_ROWS; if (along_rows) { RAFT_EXPECTS(static_cast<IndexType>(out.size()) == in.extent(0), "Output should be equal to number of rows in Input"); rowNorm(out.data_handle(), in.data_handle(), in.extent(1), in.extent(0), type, row_major, resource::get_cuda_stream(handle), fin_op); } else { RAFT_EXPECTS(static_cast<IndexType>(out.size()) == in.extent(1), "Output should be equal to number of columns in Input"); colNorm(out.data_handle(), in.data_handle(), in.extent(1), in.extent(0), type, row_major, resource::get_cuda_stream(handle), fin_op); } } /** @} */ }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/gemm.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. */ #ifndef __GEMM_H #define __GEMM_H #pragma once #include "detail/gemm.hpp" #include <raft/core/device_mdarray.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/host_mdarray.hpp> #include <raft/core/host_mdspan.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resources.hpp> #include <raft/util/input_validation.hpp> namespace raft { namespace linalg { /** * @brief the wrapper of cublas gemm function * It computes the following equation: C = alpha .* opA(A) * opB(B) + beta .* C * * @tparam math_t the element type * @tparam DevicePointerMode whether pointers alpha, beta point to device memory * @param [in] handle raft handle * @param [in] trans_a cublas transpose op for A * @param [in] trans_b cublas transpose op for B * @param [in] m number of rows of C * @param [in] n number of columns of C * @param [in] k number of rows of opB(B) / number of columns of opA(A) * @param [in] alpha host or device scalar * @param [in] A such a matrix that the shape of column-major opA(A) is [m, k] * @param [in] lda leading dimension of A * @param [in] B such a matrix that the shape of column-major opA(B) is [k, n] * @param [in] ldb leading dimension of B * @param [in] beta host or device scalar * @param [inout] C column-major matrix of size [m, n] * @param [in] ldc leading dimension of C * @param [in] stream */ template <typename math_t, bool DevicePointerMode = false> void gemm(raft::resources const& handle, const bool trans_a, const bool trans_b, const int m, const int n, const int k, const math_t* alpha, const math_t* A, const int lda, const math_t* B, const int ldb, const math_t* beta, math_t* C, const int ldc, cudaStream_t stream) { detail::gemm<math_t, DevicePointerMode>( handle, trans_a, trans_b, m, n, k, alpha, A, lda, B, ldb, beta, C, ldc, stream); } /** * @brief the wrapper of cublas gemm function * It computes the following equation: D = alpha . opA(A) * opB(B) + beta . C * @tparam math_t the type of input/output matrices * @param handle raft handle * @param a input matrix * @param n_rows_a number of rows of A * @param n_cols_a number of columns of A * @param b input matrix * @param c output matrix * @param n_rows_c number of rows of C * @param n_cols_c number of columns of C * @param trans_a cublas transpose op for A * @param trans_b cublas transpose op for B * @param alpha scalar * @param beta scalar * @param stream cuda stream */ template <typename math_t> void gemm(raft::resources const& handle, const math_t* a, int n_rows_a, int n_cols_a, const math_t* b, math_t* c, int n_rows_c, int n_cols_c, cublasOperation_t trans_a, cublasOperation_t trans_b, math_t alpha, math_t beta, cudaStream_t stream) { detail::gemm( handle, a, n_rows_a, n_cols_a, b, c, n_rows_c, n_cols_c, trans_a, trans_b, alpha, beta, stream); } /** * @brief the wrapper of cublas gemm function * It computes the following equation: D = alpha . opA(A) * opB(B) + beta . C * @tparam math_t the type of input/output matrices * @param handle raft handle * @param a input matrix * @param n_rows_a number of rows of A * @param n_cols_a number of columns of A * @param b input matrix * @param c output matrix * @param n_rows_c number of rows of C * @param n_cols_c number of columns of C * @param trans_a cublas transpose op for A * @param trans_b cublas transpose op for B * @param stream cuda stream */ template <typename math_t> void gemm(raft::resources const& handle, const math_t* a, int n_rows_a, int n_cols_a, const math_t* b, math_t* c, int n_rows_c, int n_cols_c, cublasOperation_t trans_a, cublasOperation_t trans_b, cudaStream_t stream) { detail::gemm(handle, a, n_rows_a, n_cols_a, b, c, n_rows_c, n_cols_c, trans_a, trans_b, stream); } /** * @brief A wrapper for CUBLS GEMM function designed for handling all possible * combinations of operand layouts. * It computes the following equation: Z = alpha . X * Y + beta . Z * @tparam T Data type of input/output matrices (float/double) * @param handle raft handle * @param z output matrix of size M rows x N columns * @param x input matrix of size M rows x K columns * @param y input matrix of size K rows x N columns * @param _M number of rows of X and Z * @param _N number of columns of Y and columns of Z * @param _K number of columns of X and rows of Y * @param isZColMajor Storage layout of Z. true = col major, false = row major * @param isXColMajor Storage layout of X. true = col major, false = row major * @param isYColMajor Storage layout of Y. true = col major, false = row major * @param stream cuda stream * @param alpha scalar * @param beta scalar */ template <typename T> void gemm(raft::resources const& handle, T* z, T* x, T* y, int _M, int _N, int _K, bool isZColMajor, bool isXColMajor, bool isYColMajor, cudaStream_t stream, T alpha = T(1.0), T beta = T(0.0)) { detail::gemm( handle, z, x, y, _M, _N, _K, isZColMajor, isXColMajor, isYColMajor, stream, &alpha, &beta); } /** * @defgroup gemm Matrix-Matrix Multiplication * @{ */ /** * @brief GEMM function designed for handling all possible * combinations of operand layouts (raft::row_major or raft::col_major) * with scalars alpha and beta on the host or device * It computes the following equation: Z = alpha . X * Y + beta . Z * If alpha is not provided, it is assumed to be 1.0 * If beta is not provided, it is assumed to be 0.0 * @tparam ValueType Data type of input/output matrices (float/double) * @tparam IndexType Type of index * @tparam LayoutPolicyX layout of X * @tparam LayoutPolicyY layout of Y * @tparam LayoutPolicyZ layout of Z * @param[in] handle raft handle * @param[in] x input raft::device_matrix_view of size M rows x K columns * @param[in] y input raft::device_matrix_view of size K rows x N columns * @param[out] z output raft::device_matrix_view of size M rows x N columns * @param[in] alpha optional raft::host_scalar_view or raft::device_scalar_view, default 1.0 * @param[in] beta optional raft::host_scalar_view or raft::device_scalar_view, default 0.0 */ template <typename ValueType, typename IndexType, typename LayoutPolicyX, typename LayoutPolicyY, typename LayoutPolicyZ, typename ScalarIdxType = std::uint32_t, typename ScalarViewType = raft::host_scalar_view<ValueType, ScalarIdxType>, typename = std::enable_if_t<std::disjunction_v< std::is_same<ScalarViewType, raft::host_scalar_view<ValueType, ScalarIdxType>>, std::is_same<ScalarViewType, raft::device_scalar_view<ValueType, ScalarIdxType>>>>> void gemm(raft::resources const& handle, raft::device_matrix_view<ValueType, IndexType, LayoutPolicyX> x, raft::device_matrix_view<ValueType, IndexType, LayoutPolicyY> y, raft::device_matrix_view<ValueType, IndexType, LayoutPolicyZ> z, std::optional<ScalarViewType> alpha = std::nullopt, std::optional<ScalarViewType> beta = std::nullopt) { RAFT_EXPECTS(raft::is_row_or_column_major(x), "X is not contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(y), "Y is not contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(z), "Z is not contiguous"); RAFT_EXPECTS(x.extent(0) == z.extent(0), "Number of rows of X and Z should be equal"); RAFT_EXPECTS(y.extent(1) == z.extent(1), "Number of columns of Y and Z should be equal"); RAFT_EXPECTS(x.extent(1) == y.extent(0), "Number of columns of X and rows of Y should be equal"); constexpr auto is_x_col_major = std::is_same_v<typename decltype(x)::layout_type, raft::col_major>; constexpr auto is_y_col_major = std::is_same_v<typename decltype(y)::layout_type, raft::col_major>; constexpr auto is_z_col_major = std::is_same_v<typename decltype(z)::layout_type, raft::col_major>; constexpr auto device_mode = std::is_same_v<ScalarViewType, raft::device_scalar_view<ValueType, ScalarIdxType>>; ValueType alpha_value = 1; ValueType beta_value = 0; auto alpha_device = raft::make_device_scalar(handle, alpha_value); auto beta_device = raft::make_device_scalar(handle, beta_value); auto alpha_host = raft::make_host_scalar(alpha_value); auto beta_host = raft::make_host_scalar(beta_value); if constexpr (device_mode) { if (!alpha) { alpha = alpha_device.view(); } if (!beta) { beta = beta_device.view(); } } else { if (!alpha) { alpha = alpha_host.view(); } if (!beta) { beta = beta_host.view(); } } detail::gemm<ValueType, device_mode>(handle, z.data_handle(), x.data_handle(), y.data_handle(), x.extent(0), y.extent(1), x.extent(1), is_z_col_major, is_x_col_major, is_y_col_major, resource::get_cuda_stream(handle), alpha.value().data_handle(), beta.value().data_handle()); } /** @} */ // end of gemm } // end namespace linalg } // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/linalg_types.hpp
/* * Copyright (c) 2022, 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 raft::linalg { /** * @brief Enum for reduction/broadcast where an operation is to be performed along * a matrix's rows or columns * */ enum class Apply { ALONG_ROWS, ALONG_COLUMNS }; /** * @brief Enum for reduction/broadcast where an operation is to be performed along * a matrix's rows or columns * */ enum class FillMode { UPPER, LOWER }; } // end namespace raft::linalg
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/cholesky_r1_update.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. */ #ifndef __CHOLESKY_R1_UPDATE_H #define __CHOLESKY_R1_UPDATE_H #pragma once #include "detail/cholesky_r1_update.cuh" #include <raft/core/resource/cublas_handle.hpp> namespace raft { namespace linalg { /** * @brief Rank 1 update of Cholesky decomposition. * NOTE: The new mdspan-based API will not be provided for this function. * * This method is useful if an algorithm iteratively builds up matrix A, and * the Cholesky decomposition of A is required at each step. * * On entry, L is the Cholesky decomposition of matrix A, where both A and L * have size n-1 x n-1. We are interested in the Cholesky decomposition of a new * matrix A', which we get by adding a row and column to A. In Python notation: * - A'[0:n-1, 0:n-1] = A; * - A'[:,n-1] = A[n-1,:] = A_new * * On entry, the new column A_new, is stored as the n-th column of L if uplo == * CUBLAS_FILL_MODE_UPPER, else A_new is stored as the n-th row of L. * * On exit L contains the Cholesky decomposition of A'. In practice the elements * of A_new are overwritten with new row/column of the L matrix. * * The uplo parameter is used to select the matrix layout. * If (uplo != CUBLAS_FILL_MODE_UPPER) then the input arg L stores the * lower triangular matrix L, so that A = L * L.T. Otherwise the input arg L * stores an upper triangular matrix U: A = U.T * U. * * On exit L will be updated to store the Cholesky decomposition of A'. * * If the matrix is not positive definite, or very ill conditioned then the new * diagonal element of L would be NaN. In such a case an exception is thrown. * The eps argument can be used to override this behavior: if eps >= 0 then * the diagonal element is replaced by eps in case the diagonal is NaN or * smaller than eps. Note: for an iterative solver it is probably better to * stop early in case of error, rather than relying on the eps parameter. * * Examples: * * - Lower triangular factorization: * @code{.cpp} * // Initialize arrays * int ld_L = n_rows; * rmm::device_uvector<math_t> L(ld_L * n_rows, stream); * raft::linalg::choleskyRank1Update(handle, L, n_rows, ld_L, nullptr, * &n_bytes, CUBLAS_FILL_MODE_LOWER, * stream); * rmm::device_uvector<char> workspace(n_bytes, stream); * * for (n=1; n<=n_rows; rank++) { * // Calculate a new row/column of matrix A into A_new * // ... * // Copy new row to L[rank-1,:] * RAFT_CUBLAS_TRY(cublasCopy(resource::get_cublas_handle(handle), n - 1, A_new, 1, * L + n - 1, ld_L, stream)); * // Update Cholesky factorization * raft::linalg::choleskyRank1Update( * handle, L, rank, ld_L, workspace, &n_bytes, CUBLAS_FILL_MODE_LOWER, * stream); * } * Now L stores the Cholesky decomposition of A: A = L * L.T * @endcode * * - Upper triangular factorization: * @code{.cpp} * // Initialize arrays * int ld_U = n_rows; * rmm::device_uvector<math_t> U(ld_U * n_rows, stream); * raft::linalg::choleskyRank1Update(handle, L, n_rows, ld_U, nullptr, * &n_bytes, CUBLAS_FILL_MODE_UPPER, * stream); * rmm::device_uvector<char> workspace(stream, n_bytes, stream); * * for (n=1; n<=n_rows; n++) { * // Calculate a new row/column of matrix A into array A_new * // ... * // Copy new row to U[:,n-1] (column major layout) * raft::copy(U + ld_U * (n-1), A_new, n-1, stream); * // * // Update Cholesky factorization * raft::linalg::choleskyRank1Update( * handle, U, n, ld_U, workspace, &n_bytes, CUBLAS_FILL_MODE_UPPER, * stream); * } * // Now U stores the Cholesky decomposition of A: A = U.T * U * @endcode * * @param handle RAFT handle (used to retrieve cuBLAS handles). * @param L device array for to store the triangular matrix L, and the new * column of A in column major format, size [n*n] * @param n number of elements in the new row. * @param ld stride of columns in L * @param workspace device pointer to workspace shall be nullptr ar an array * of size [n_bytes]. * @param n_bytes size of workspace is returned here if workspace==nullptr. * @param stream CUDA stream * @param uplo indicates whether L is stored as an upper or lower triangular * matrix (CUBLAS_FILL_MODE_UPPER or CUBLAS_FILL_MODE_LOWER) * @param eps numerical parameter that can act as a regularizer for ill * conditioned systems. Negative values mean no regularizaton. */ template <typename math_t> void choleskyRank1Update(raft::resources const& handle, math_t* L, int n, int ld, void* workspace, int* n_bytes, cublasFillMode_t uplo, cudaStream_t stream, math_t eps = -1) { detail::choleskyRank1Update(handle, L, n, ld, workspace, n_bytes, uplo, stream, eps); } }; // namespace linalg }; // namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/normalize.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 "detail/normalize.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/operators.hpp> #include <raft/linalg/norm_types.hpp> namespace raft { namespace linalg { /** * @defgroup normalize Row- or Col-norm computation * @{ */ /** * @brief Divide rows by their norm defined by main_op, reduce_op and fin_op * * @tparam ElementType Input/Output data type * @tparam IndexType Integer type used to for addressing * @tparam MainLambda Type of main_op * @tparam ReduceLambda Type of reduce_op * @tparam FinalLambda Type of fin_op * @param[in] handle raft::resources * @param[in] in the input raft::device_matrix_view * @param[out] out the output raft::device_matrix_view * @param[in] init Initialization value, i.e identity element for the reduction operation * @param[in] main_op Operation to apply to the elements before reducing them (e.g square for L2) * @param[in] reduce_op Operation to reduce a pair of elements (e.g sum for L2) * @param[in] fin_op Operation to apply once to the reduction result to finalize the norm * computation (e.g sqrt for L2) * @param[in] eps If the norm is below eps, the row is considered zero and no division is applied */ template <typename ElementType, typename IndexType, typename MainLambda, typename ReduceLambda, typename FinalLambda> void row_normalize(raft::resources const& handle, raft::device_matrix_view<const ElementType, IndexType, row_major> in, raft::device_matrix_view<ElementType, IndexType, row_major> out, ElementType init, MainLambda main_op, ReduceLambda reduce_op, FinalLambda fin_op, ElementType eps = ElementType(1e-8)) { RAFT_EXPECTS(raft::is_row_or_column_major(in), "Input must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(in.extent(0) == out.extent(0), "The number of rows of the input and output should be equal"); RAFT_EXPECTS(in.extent(1) == out.extent(1), "The number of columns of the input and output should be equal"); detail::coalesced_normalize(out.data_handle(), in.data_handle(), in.extent(1), in.extent(0), init, resource::get_cuda_stream(handle), main_op, reduce_op, fin_op, eps); } /** * @brief Divide rows by their norm. * * @tparam ElementType Input/Output data type * @tparam IndexType Integer type used to for addressing * @param[in] handle raft::resources * @param[in] in the input raft::device_matrix_view * @param[out] out the output raft::device_matrix_view * @param[in] norm_type the type of norm to be applied * @param[in] eps If the norm is below eps, the row is considered zero and no division is applied */ template <typename ElementType, typename IndexType> void row_normalize(raft::resources const& handle, raft::device_matrix_view<const ElementType, IndexType, row_major> in, raft::device_matrix_view<ElementType, IndexType, row_major> out, NormType norm_type, ElementType eps = ElementType(1e-8)) { switch (norm_type) { case L1Norm: row_normalize( handle, in, out, ElementType(0), raft::abs_op(), raft::add_op(), raft::identity_op(), eps); break; case L2Norm: row_normalize( handle, in, out, ElementType(0), raft::sq_op(), raft::add_op(), raft::sqrt_op(), eps); break; case LinfNorm: row_normalize( handle, in, out, ElementType(0), raft::abs_op(), raft::max_op(), raft::identity_op(), eps); break; default: THROW("Unsupported norm type: %d", norm_type); } } /** @} */ } // namespace linalg } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/qr.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. */ #ifndef __QR_H #define __QR_H #pragma once #include "detail/qr.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resources.hpp> namespace raft { namespace linalg { /** * @brief compute QR decomp and return only Q matrix * @param handle: raft handle * @param M: input matrix * @param Q: Q matrix to be returned (on GPU) * @param n_rows: number rows of input matrix * @param n_cols: number columns of input matrix * @param stream cuda stream */ template <typename math_t> void qrGetQ(raft::resources const& handle, const math_t* M, math_t* Q, int n_rows, int n_cols, cudaStream_t stream) { detail::qrGetQ(handle, M, Q, n_rows, n_cols, stream); } /** * @brief compute QR decomp and return both Q and R matrices * @param handle: raft handle * @param M: input matrix * @param Q: Q matrix to be returned (on GPU) * @param R: R matrix to be returned (on GPU) * @param n_rows: number rows of input matrix * @param n_cols: number columns of input matrix * @param stream cuda stream */ template <typename math_t> void qrGetQR(raft::resources const& handle, math_t* M, math_t* Q, math_t* R, int n_rows, int n_cols, cudaStream_t stream) { detail::qrGetQR(handle, M, Q, R, n_rows, n_cols, stream); } /** * @defgroup qr QR Decomposition * @{ */ /** * @brief Compute the QR decomposition of matrix M and return only the Q matrix. * @param[in] handle raft::resources * @param[in] M Input raft::device_matrix_view * @param[out] Q Output raft::device_matrix_view */ template <typename ElementType, typename IndexType> void qr_get_q(raft::resources const& handle, raft::device_matrix_view<const ElementType, IndexType, raft::col_major> M, raft::device_matrix_view<ElementType, IndexType, raft::col_major> Q) { RAFT_EXPECTS(Q.size() == M.size(), "Size mismatch between Output and Input"); qrGetQ(handle, M.data_handle(), Q.data_handle(), M.extent(0), M.extent(1), resource::get_cuda_stream(handle)); } /** * @brief Compute the QR decomposition of matrix M and return both the Q and R matrices. * @param[in] handle raft::resources * @param[in] M Input raft::device_matrix_view * @param[in] Q Output raft::device_matrix_view * @param[out] R Output raft::device_matrix_view */ template <typename ElementType, typename IndexType> void qr_get_qr(raft::resources const& handle, raft::device_matrix_view<const ElementType, IndexType, raft::col_major> M, raft::device_matrix_view<ElementType, IndexType, raft::col_major> Q, raft::device_matrix_view<ElementType, IndexType, raft::col_major> R) { RAFT_EXPECTS(Q.size() == M.size(), "Size mismatch between Output and Input"); qrGetQR(handle, M.data_handle(), Q.data_handle(), R.data_handle(), M.extent(0), M.extent(1), resource::get_cuda_stream(handle)); } /** @} */ }; // namespace linalg }; // namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/reduce_rows_by_key.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 __REDUCE_ROWS_BY_KEY #define __REDUCE_ROWS_BY_KEY #pragma once #include "detail/reduce_rows_by_key.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/resources.hpp> namespace raft { namespace linalg { /** Small helper function to convert from int->char and char->int Transform ncols*nrows read of int in 2*nrows reads of int + ncols*rows reads of chars **/ template <typename IteratorT1, typename IteratorT2> void convert_array(IteratorT1 dst, IteratorT2 src, int n, cudaStream_t st) { detail::convert_array(dst, src, n, st); } /** * @brief Computes the weighted reduction of matrix rows for each given key * * @tparam DataIteratorT Random-access iterator type, for reading input matrix * (may be a simple pointer type) * @tparam KeysIteratorT Random-access iterator type, for reading input keys * (may be a simple pointer type) * @tparam SumsT Type of the output sums * @tparam IdxT Index type * * @param[in] d_A Input data array (lda x nrows) * @param[in] lda Real row size for input data, d_A * @param[in] d_keys Keys for each row (1 x nrows) * @param[in] d_weights Weights for each observation in d_A (1 x nrows) * @param[out] d_keys_char Scratch memory for conversion of keys to char * @param[in] nrows Number of rows in d_A and d_keys * @param[in] ncols Number of data columns in d_A * @param[in] nkeys Number of unique keys in d_keys * @param[out] d_sums Row sums by key (ncols x d_keys) * @param[in] stream CUDA stream * @param[in] reset_sums Whether to reset the output sums to zero before reducing */ template <typename DataIteratorT, typename KeysIteratorT, typename WeightT, typename SumsT, typename IdxT> void reduce_rows_by_key(const DataIteratorT d_A, IdxT lda, const KeysIteratorT d_keys, const WeightT* d_weights, char* d_keys_char, IdxT nrows, IdxT ncols, IdxT nkeys, SumsT* d_sums, cudaStream_t stream, bool reset_sums = true) { detail::reduce_rows_by_key( d_A, lda, d_keys, d_weights, d_keys_char, nrows, ncols, nkeys, d_sums, stream, reset_sums); } /** * @brief Computes the reduction of matrix rows for each given key * @tparam DataIteratorT Random-access iterator type, for reading input matrix (may be a simple * pointer type) * @tparam KeysIteratorT Random-access iterator type, for reading input keys (may be a simple * pointer type) * @tparam SumsT Type of the output sums * @tparam IdxT Index type * @param[in] d_A Input data array (lda x nrows) * @param[in] lda Real row size for input data, d_A * @param[in] d_keys Keys for each row (1 x nrows) * @param d_keys_char Scratch memory for conversion of keys to char * @param[in] nrows Number of rows in d_A and d_keys * @param[in] ncols Number of data columns in d_A * @param[in] nkeys Number of unique keys in d_keys * @param[out] d_sums Row sums by key (ncols x d_keys) * @param[in] stream CUDA stream * @param[in] reset_sums Whether to reset the output sums to zero before reducing */ template <typename DataIteratorT, typename KeysIteratorT, typename SumsT, typename IdxT> void reduce_rows_by_key(const DataIteratorT d_A, IdxT lda, const KeysIteratorT d_keys, char* d_keys_char, IdxT nrows, IdxT ncols, IdxT nkeys, SumsT* d_sums, cudaStream_t stream, bool reset_sums = true) { typedef typename std::iterator_traits<DataIteratorT>::value_type DataType; reduce_rows_by_key(d_A, lda, d_keys, static_cast<DataType*>(nullptr), d_keys_char, nrows, ncols, nkeys, d_sums, stream, reset_sums); } /** * @defgroup reduce_rows_by_key Reduce Across Rows by Key * @{ */ /** * @brief Computes the weighted sum-reduction of matrix rows for each given key * TODO: Support generic reduction lambdas https://github.com/rapidsai/raft/issues/860 * @tparam ElementType data-type of input and output * @tparam KeyType data-type of keys * @tparam WeightType data-type of weights * @tparam IndexType index type * @param[in] handle raft::resources * @param[in] d_A Input raft::device_mdspan (ncols * nrows) * @param[in] d_keys Keys for each row raft::device_vector_view (1 x nrows) * @param[out] d_sums Row sums by key raft::device_matrix_view (ncols x d_keys) * @param[in] n_unique_keys Number of unique keys in d_keys * @param[out] d_keys_char Scratch memory for conversion of keys to char, raft::device_vector_view * @param[in] d_weights Weights for each observation in d_A raft::device_vector_view optional (1 * x nrows) * @param[in] reset_sums Whether to reset the output sums to zero before reducing */ template <typename ElementType, typename KeyType, typename WeightType, typename IndexType> void reduce_rows_by_key( raft::resources const& handle, raft::device_matrix_view<const ElementType, IndexType, raft::row_major> d_A, raft::device_vector_view<const KeyType, IndexType> d_keys, raft::device_matrix_view<ElementType, IndexType, raft::row_major> d_sums, IndexType n_unique_keys, raft::device_vector_view<char, IndexType> d_keys_char, std::optional<raft::device_vector_view<const WeightType, IndexType>> d_weights = std::nullopt, bool reset_sums = true) { RAFT_EXPECTS(d_A.extent(0) == d_A.extent(0) && d_sums.extent(1) == n_unique_keys, "Output is not of size ncols * n_unique_keys"); RAFT_EXPECTS(d_keys.extent(0) == d_A.extent(1), "Keys is not of size nrows"); if (d_weights) { RAFT_EXPECTS(d_weights.value().extent(0) == d_A.extent(1), "Weights is not of size nrows"); reduce_rows_by_key(d_A.data_handle(), d_A.extent(0), d_keys.data_handle(), d_weights.value().data_handle(), d_keys_char.data_handle(), d_A.extent(1), d_A.extent(0), n_unique_keys, d_sums.data_handle(), resource::get_cuda_stream(handle), reset_sums); } else { reduce_rows_by_key(d_A.data_handle(), d_A.extent(0), d_keys.data_handle(), d_keys_char.data_handle(), d_A.extent(1), d_A.extent(0), n_unique_keys, d_sums.data_handle(), resource::get_cuda_stream(handle), reset_sums); } } /** @} */ // end of group reduce_rows_by_key }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/power.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. */ #ifndef __POWER_H #define __POWER_H #pragma once #include <raft/core/host_mdspan.hpp> #include <raft/core/operators.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/linalg/binary_op.cuh> #include <raft/linalg/unary_op.cuh> #include <raft/util/input_validation.hpp> namespace raft { namespace linalg { /** * @defgroup ScalarOps Scalar operations on the input buffer * @tparam in_t Input data-type * @tparam out_t Output data-type * @param out the output buffer * @param in the input buffer * @param scalar the scalar used in the operations * @param len number of elements in the input buffer * @param stream cuda stream where to launch work * @{ */ template <typename in_t, typename out_t = in_t, typename IdxType = int> void powerScalar(out_t* out, const in_t* in, const in_t scalar, IdxType len, cudaStream_t stream) { raft::linalg::unaryOp(out, in, len, raft::pow_const_op<in_t>(scalar), stream); } /** @} */ /** * @defgroup BinaryOps Element-wise binary operations on the input buffers * @tparam in_t Input data-type * @tparam out_t Output data-type * @tparam IdxType Integer type used to for addressing * @param out the output buffer * @param in1 the first input buffer * @param in2 the second input buffer * @param len number of elements in the input buffers * @param stream cuda stream where to launch work * @{ */ template <typename in_t, typename out_t = in_t, typename IdxType = int> void power(out_t* out, const in_t* in1, const in_t* in2, IdxType len, cudaStream_t stream) { raft::linalg::binaryOp(out, in1, in2, len, raft::pow_op(), stream); } /** @} */ /** * @defgroup power Power Arithmetic * @{ */ /** * @brief Elementwise power operation on the input buffers * @tparam InType Input Type raft::device_mdspan * @tparam OutType Output Type raft::device_mdspan * @param[in] handle raft::resources * @param[in] in1 First Input * @param[in] in2 Second Input * @param[out] out Output */ template <typename InType, typename OutType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void power(raft::resources const& handle, InType in1, InType in2, OutType out) { using in_value_t = typename InType::value_type; using out_value_t = typename OutType::value_type; RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in1), "Input 1 must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in2), "Input 2 must be contiguous"); RAFT_EXPECTS(out.size() == in1.size() && in1.size() == in2.size(), "Size mismatch between Output and Inputs"); if (out.size() <= std::numeric_limits<std::uint32_t>::max()) { power<in_value_t, out_value_t, std::uint32_t>(out.data_handle(), in1.data_handle(), in2.data_handle(), static_cast<std::uint32_t>(out.size()), resource::get_cuda_stream(handle)); } else { power<in_value_t, out_value_t, std::uint64_t>(out.data_handle(), in1.data_handle(), in2.data_handle(), static_cast<std::uint64_t>(out.size()), resource::get_cuda_stream(handle)); } } /** * @brief Elementwise power of host scalar to input * @tparam InType Input Type raft::device_mdspan * @tparam OutType Output Type raft::device_mdspan * @tparam ScalarIdxType Index Type of scalar * @param[in] handle raft::resources * @param[in] in Input * @param[out] out Output * @param[in] scalar raft::host_scalar_view */ template <typename InType, typename OutType, typename ScalarIdxType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void power_scalar( raft::resources const& handle, InType in, OutType out, const raft::host_scalar_view<const typename InType::value_type, ScalarIdxType> scalar) { using in_value_t = typename InType::value_type; using out_value_t = typename OutType::value_type; RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in), "Input must be contiguous"); RAFT_EXPECTS(out.size() == in.size(), "Size mismatch between Output and Input"); if (out.size() <= std::numeric_limits<std::uint32_t>::max()) { powerScalar<in_value_t, out_value_t, std::uint32_t>(out.data_handle(), in.data_handle(), *scalar.data_handle(), static_cast<std::uint32_t>(out.size()), resource::get_cuda_stream(handle)); } else { powerScalar<in_value_t, out_value_t, std::uint64_t>(out.data_handle(), in.data_handle(), *scalar.data_handle(), static_cast<std::uint64_t>(out.size()), resource::get_cuda_stream(handle)); } } /** @} */ // end of group add }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/divide.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. */ #ifndef __DIVIDE_H #define __DIVIDE_H #pragma once #include "detail/divide.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/util/cuda_utils.cuh> #include <raft/util/input_validation.hpp> namespace raft { namespace linalg { /** * @defgroup ScalarOps Scalar operations on the input buffer * @tparam OutT output data-type upon which the math operation will be performed * @tparam InT input data-type upon which the math operation will be performed * @tparam IdxType Integer type used to for addressing * @param out the output buffer * @param in the input buffer * @param scalar the scalar used in the operations * @param len number of elements in the input buffer * @param stream cuda stream where to launch work * @{ */ template <typename InT, typename OutT = InT, typename IdxType = int> void divideScalar(OutT* out, const InT* in, InT scalar, IdxType len, cudaStream_t stream) { detail::divideScalar(out, in, scalar, len, stream); } /** @} */ /** * @defgroup divide Division Arithmetic * @{ */ /** * @brief Elementwise division of input by host scalar * @tparam InType Input Type raft::device_mdspan * @tparam OutType Output Type raft::device_mdspan * @tparam ScalarIdxType Index Type of scalar * @param[in] handle raft::resources * @param[in] in Input * @param[in] scalar raft::host_scalar_view * @param[out] out Output */ template <typename InType, typename OutType, typename ScalarIdxType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void divide_scalar(raft::resources const& handle, InType in, OutType out, raft::host_scalar_view<const typename InType::value_type, ScalarIdxType> scalar) { using in_value_t = typename InType::value_type; using out_value_t = typename OutType::value_type; RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in), "Input must be contiguous"); RAFT_EXPECTS(out.size() == in.size(), "Size mismatch between Output and Input"); if (out.size() <= std::numeric_limits<std::uint32_t>::max()) { divideScalar<in_value_t, out_value_t, std::uint32_t>(out.data_handle(), in.data_handle(), *scalar.data_handle(), static_cast<std::uint32_t>(out.size()), resource::get_cuda_stream(handle)); } else { divideScalar<in_value_t, out_value_t, std::uint64_t>(out.data_handle(), in.data_handle(), *scalar.data_handle(), static_cast<std::uint64_t>(out.size()), resource::get_cuda_stream(handle)); } } /** @} */ // end of group add }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/transpose.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. */ #ifndef __TRANSPOSE_H #define __TRANSPOSE_H #pragma once #include "detail/transpose.cuh" #include <raft/core/device_mdarray.hpp> #include <raft/core/resources.hpp> namespace raft { namespace linalg { /** * @brief transpose on the column major input matrix using Jacobi method * @param handle: raft handle * @param in: input matrix * @param out: output. Transposed input matrix * @param n_rows: number rows of input matrix * @param n_cols: number columns of input matrix * @param stream: cuda stream */ template <typename math_t> void transpose(raft::resources const& handle, math_t* in, math_t* out, int n_rows, int n_cols, cudaStream_t stream) { detail::transpose(handle, in, out, n_rows, n_cols, stream); } /** * @brief transpose on the column major input matrix using Jacobi method * @param inout: input and output matrix * @param n: number of rows and columns of input matrix * @param stream: cuda stream */ template <typename math_t> void transpose(math_t* inout, int n, cudaStream_t stream) { detail::transpose(inout, n, stream); } /** * @defgroup transpose Matrix transpose * @{ */ /** * @brief Transpose a matrix. The output has same layout policy as the input. * * @tparam T Data type of input matrix element. * @tparam IndexType Index type of matrix extent. * @tparam LayoutPolicy Layout type of the input matrix. When layout is strided, it can * be a submatrix of a larger matrix. Arbitrary stride is not supported. * @tparam AccessorPolicy Accessor for the input and output, must be valid accessor on * device. * * @param[in] handle raft handle for managing expensive cuda resources. * @param[in] in Input matrix. * @param[out] out Output matrix, storage is pre-allocated by caller. */ template <typename T, typename IndexType, typename LayoutPolicy, typename AccessorPolicy> auto transpose(raft::resources const& handle, raft::mdspan<T, raft::matrix_extent<IndexType>, LayoutPolicy, AccessorPolicy> in, raft::mdspan<T, raft::matrix_extent<IndexType>, LayoutPolicy, AccessorPolicy> out) -> std::enable_if_t<std::is_floating_point_v<T>, void> { RAFT_EXPECTS(out.extent(0) == in.extent(1), "Invalid shape for transpose."); RAFT_EXPECTS(out.extent(1) == in.extent(0), "Invalid shape for transpose."); if constexpr (std::is_same_v<typename decltype(in)::layout_type, layout_c_contiguous>) { detail::transpose_row_major_impl(handle, in, out); } else if (std::is_same_v<typename decltype(in)::layout_type, layout_f_contiguous>) { detail::transpose_col_major_impl(handle, in, out); } else { RAFT_EXPECTS(in.stride(0) == 1 || in.stride(1) == 1, "Unsupported matrix layout."); if (in.stride(1) == 1) { // row-major submatrix detail::transpose_row_major_impl(handle, in, out); } else { // col-major submatrix detail::transpose_col_major_impl(handle, in, out); } } } /** @} */ // end of group transpose }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/add.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. */ #ifndef __ADD_H #define __ADD_H #pragma once #include "detail/add.cuh" #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/device_mdspan.hpp> #include <raft/core/host_mdspan.hpp> #include <raft/util/input_validation.hpp> namespace raft { namespace linalg { /** * @ingroup arithmetic * @brief Elementwise scalar add operation on the input buffer * * @tparam InT input data-type. Also the data-type upon which the math ops * will be performed * @tparam OutT output data-type * @tparam IdxType Integer type used to for addressing * * @param out the output buffer * @param in the input buffer * @param scalar the scalar used in the operations * @param len number of elements in the input buffer * @param stream cuda stream where to launch work */ template <typename InT, typename OutT = InT, typename IdxType = int> void addScalar(OutT* out, const InT* in, const InT scalar, IdxType len, cudaStream_t stream) { detail::addScalar(out, in, scalar, len, stream); } /** * @brief Elementwise add operation on the input buffers * @tparam InT input data-type. Also the data-type upon which the math ops * will be performed * @tparam OutT output data-type * @tparam IdxType Integer type used to for addressing * * @param out the output buffer * @param in1 the first input buffer * @param in2 the second input buffer * @param len number of elements in the input buffers * @param stream cuda stream where to launch work */ template <typename InT, typename OutT = InT, typename IdxType = int> void add(OutT* out, const InT* in1, const InT* in2, IdxType len, cudaStream_t stream) { detail::add(out, in1, in2, len, stream); } /** Subtract single value pointed by singleScalarDev parameter in device memory from inDev[i] and * write result to outDev[i] * @tparam InT input data-type. Also the data-type upon which the math ops * will be performed * @tparam OutT output data-type * @tparam IdxType Integer type used to for addressing * @param outDev the output buffer * @param inDev the input buffer * @param singleScalarDev pointer to the scalar located in device memory * @param len number of elements in the input and output buffer * @param stream cuda stream */ template <typename InT, typename OutT = InT, typename IdxType = int> void addDevScalar( OutT* outDev, const InT* inDev, const InT* singleScalarDev, IdxType len, cudaStream_t stream) { detail::addDevScalar(outDev, inDev, singleScalarDev, len, stream); } /** * @defgroup add_dense Addition Arithmetic * @{ */ /** * @brief Elementwise add operation * @tparam InType Input Type raft::device_mdspan * @tparam OutType Output Type raft::device_mdspan * @param[in] handle raft::resources * @param[in] in1 First Input * @param[in] in2 Second Input * @param[out] out Output */ template <typename InType, typename OutType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void add(raft::resources const& handle, InType in1, InType in2, OutType out) { using in_value_t = typename InType::value_type; using out_value_t = typename OutType::value_type; RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in1), "Input 1 must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in2), "Input 2 must be contiguous"); RAFT_EXPECTS(out.size() == in1.size() && in1.size() == in2.size(), "Size mismatch between Output and Inputs"); if (out.size() <= std::numeric_limits<std::uint32_t>::max()) { add<in_value_t, out_value_t, std::uint32_t>(out.data_handle(), in1.data_handle(), in2.data_handle(), static_cast<std::uint32_t>(out.size()), resource::get_cuda_stream(handle)); } else { add<in_value_t, out_value_t, std::uint64_t>(out.data_handle(), in1.data_handle(), in2.data_handle(), static_cast<std::uint64_t>(out.size()), resource::get_cuda_stream(handle)); } } /** * @brief Elementwise addition of device scalar to input * @tparam InType Input Type raft::device_mdspan * @tparam OutType Output Type raft::device_mdspan * @tparam ScalarIdxType Index Type of scalar * @param[in] handle raft::resources * @param[in] in Input * @param[in] scalar raft::device_scalar_view * @param[in] out Output */ template <typename InType, typename OutType, typename ScalarIdxType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void add_scalar(raft::resources const& handle, InType in, OutType out, raft::device_scalar_view<const typename InType::value_type, ScalarIdxType> scalar) { using in_value_t = typename InType::value_type; using out_value_t = typename OutType::value_type; RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in), "Input must be contiguous"); RAFT_EXPECTS(out.size() == in.size(), "Size mismatch between Output and Input"); if (out.size() <= std::numeric_limits<std::uint32_t>::max()) { addDevScalar<in_value_t, out_value_t, std::uint32_t>(out.data_handle(), in.data_handle(), scalar.data_handle(), static_cast<std::uint32_t>(out.size()), resource::get_cuda_stream(handle)); } else { addDevScalar<in_value_t, out_value_t, std::uint64_t>(out.data_handle(), in.data_handle(), scalar.data_handle(), static_cast<std::uint64_t>(out.size()), resource::get_cuda_stream(handle)); } } /** * @brief Elementwise addition of host scalar to input * @tparam InType Input Type raft::device_mdspan * @tparam OutType Output Type raft::device_mdspan * @tparam ScalarIdxType Index Type of scalar * @param[in] handle raft::resources * @param[in] in Input * @param[in] scalar raft::host_scalar_view * @param[in] out Output */ template <typename InType, typename OutType, typename ScalarIdxType, typename = raft::enable_if_input_device_mdspan<InType>, typename = raft::enable_if_output_device_mdspan<OutType>> void add_scalar(raft::resources const& handle, const InType in, OutType out, raft::host_scalar_view<const typename InType::value_type, ScalarIdxType> scalar) { using in_value_t = typename InType::value_type; using out_value_t = typename OutType::value_type; RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); RAFT_EXPECTS(raft::is_row_or_column_major(in), "Input must be contiguous"); RAFT_EXPECTS(out.size() == in.size(), "Size mismatch between Output and Input"); if (out.size() <= std::numeric_limits<std::uint32_t>::max()) { addScalar<in_value_t, out_value_t, std::uint32_t>(out.data_handle(), in.data_handle(), *scalar.data_handle(), static_cast<std::uint32_t>(out.size()), resource::get_cuda_stream(handle)); } else { addScalar<in_value_t, out_value_t, std::uint64_t>(out.data_handle(), in.data_handle(), *scalar.data_handle(), static_cast<std::uint64_t>(out.size()), resource::get_cuda_stream(handle)); } } /** @} */ // end of group add }; // end namespace linalg }; // end namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft
rapidsai_public_repos/raft/cpp/include/raft/linalg/lstsq.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. */ #ifndef __LSTSQ_H #define __LSTSQ_H #pragma once #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resources.hpp> #include <raft/linalg/detail/lstsq.cuh> namespace raft { namespace linalg { /** Solves the linear ordinary least squares problem `Aw = b` * Via SVD decomposition of `A = U S Vt` using default cuSOLVER routine. * * @param[in] handle raft handle * @param[inout] A input feature matrix. * Warning: the content of this matrix is modified by the cuSOLVER routines. * @param[in] n_rows number of rows in A * @param[in] n_cols number of columns in A * @param[inout] b input target vector. * Warning: the content of this vector is modified by the cuSOLVER routines. * @param[out] w output coefficient vector * @param[in] stream cuda stream for ordering operations */ template <typename math_t> void lstsqSvdQR(raft::resources const& handle, math_t* A, const int n_rows, const int n_cols, const math_t* b, math_t* w, cudaStream_t stream) { detail::lstsqSvdQR(handle, A, n_rows, n_cols, b, w, stream); } /** Solves the linear ordinary least squares problem `Aw = b` * Via SVD decomposition of `A = U S V^T` using Jacobi iterations (cuSOLVER). * * @param[in] handle raft handle * @param[inout] A input feature matrix. * Warning: the content of this matrix is modified by the cuSOLVER routines. * @param[in] n_rows number of rows in A * @param[in] n_cols number of columns in A * @param[inout] b input target vector. * Warning: the content of this vector is modified by the cuSOLVER routines. * @param[out] w output coefficient vector * @param[in] stream cuda stream for ordering operations */ template <typename math_t> void lstsqSvdJacobi(raft::resources const& handle, math_t* A, const int n_rows, const int n_cols, const math_t* b, math_t* w, cudaStream_t stream) { detail::lstsqSvdJacobi(handle, A, n_rows, n_cols, b, w, stream); } /** Solves the linear ordinary least squares problem `Aw = b` * via eigenvalue decomposition of `A^T * A` (covariance matrix for dataset A). * (`w = (A^T A)^-1 A^T b`) */ template <typename math_t> void lstsqEig(raft::resources const& handle, const math_t* A, const int n_rows, const int n_cols, const math_t* b, math_t* w, cudaStream_t stream) { detail::lstsqEig(handle, A, n_rows, n_cols, b, w, stream); } /** Solves the linear ordinary least squares problem `Aw = b` * via QR decomposition of `A = QR`. * (triangular system of equations `Rw = Q^T b`) * * @param[in] handle raft handle * @param[inout] A input feature matrix. * Warning: the content of this matrix is modified by the cuSOLVER routines. * @param[in] n_rows number of rows in A * @param[in] n_cols number of columns in A * @param[inout] b input target vector. * Warning: the content of this vector is modified by the cuSOLVER routines. * @param[out] w output coefficient vector * @param[in] stream cuda stream for ordering operations */ template <typename math_t> void lstsqQR(raft::resources const& handle, math_t* A, const int n_rows, const int n_cols, math_t* b, math_t* w, cudaStream_t stream) { detail::lstsqQR(handle, A, n_rows, n_cols, b, w, stream); } /** * @defgroup lstsq Least Squares Methods * @{ */ /** * @brief Solves the linear ordinary least squares problem `Aw = b` * Via SVD decomposition of `A = U S Vt`. * * @tparam ValueType the data-type of input/output * @param[in] handle raft::resources * @param[inout] A input raft::device_matrix_view * Warning: the content of this matrix is modified. * @param[inout] b input target raft::device_vector_view * Warning: the content of this vector is modified. * @param[out] w output coefficient raft::device_vector_view */ template <typename ValueType, typename IndexType> void lstsq_svd_qr(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> A, raft::device_vector_view<const ValueType, IndexType> b, raft::device_vector_view<ValueType, IndexType> w) { RAFT_EXPECTS(A.extent(1) == w.size(), "Size mismatch between A and w"); RAFT_EXPECTS(A.extent(0) == b.size(), "Size mismatch between A and b"); lstsqSvdQR(handle, const_cast<ValueType*>(A.data_handle()), A.extent(0), A.extent(1), const_cast<ValueType*>(b.data_handle()), w.data_handle(), resource::get_cuda_stream(handle)); } /** * @brief Solves the linear ordinary least squares problem `Aw = b` * Via SVD decomposition of `A = U S V^T` using Jacobi iterations. * * @tparam ValueType the data-type of input/output * @param[in] handle raft::resources * @param[inout] A input raft::device_matrix_view * Warning: the content of this matrix is modified. * @param[inout] b input target raft::device_vector_view * Warning: the content of this vector is modified. * @param[out] w output coefficient raft::device_vector_view */ template <typename ValueType, typename IndexType> void lstsq_svd_jacobi(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> A, raft::device_vector_view<const ValueType, IndexType> b, raft::device_vector_view<ValueType, IndexType> w) { RAFT_EXPECTS(A.extent(1) == w.size(), "Size mismatch between A and w"); RAFT_EXPECTS(A.extent(0) == b.size(), "Size mismatch between A and b"); lstsqSvdJacobi(handle, const_cast<ValueType*>(A.data_handle()), A.extent(0), A.extent(1), const_cast<ValueType*>(b.data_handle()), w.data_handle(), resource::get_cuda_stream(handle)); } /** * @brief Solves the linear ordinary least squares problem `Aw = b` * via eigenvalue decomposition of `A^T * A` (covariance matrix for dataset A). * (`w = (A^T A)^-1 A^T b`) * * @tparam ValueType the data-type of input/output * @param[in] handle raft::resources * @param[inout] A input raft::device_matrix_view * Warning: the content of this matrix is modified by the cuSOLVER routines. * @param[inout] b input target raft::device_vector_view * Warning: the content of this vector is modified by the cuSOLVER routines. * @param[out] w output coefficient raft::device_vector_view */ template <typename ValueType, typename IndexType> void lstsq_eig(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> A, raft::device_vector_view<const ValueType, IndexType> b, raft::device_vector_view<ValueType, IndexType> w) { RAFT_EXPECTS(A.extent(1) == w.size(), "Size mismatch between A and w"); RAFT_EXPECTS(A.extent(0) == b.size(), "Size mismatch between A and b"); lstsqEig(handle, const_cast<ValueType*>(A.data_handle()), A.extent(0), A.extent(1), const_cast<ValueType*>(b.data_handle()), w.data_handle(), resource::get_cuda_stream(handle)); } /** * @brief Solves the linear ordinary least squares problem `Aw = b` * via QR decomposition of `A = QR`. * (triangular system of equations `Rw = Q^T b`) * * @tparam ValueType the data-type of input/output * @param[in] handle raft::resources * @param[inout] A input raft::device_matrix_view * Warning: the content of this matrix is modified. * @param[inout] b input target raft::device_vector_view * Warning: the content of this vector is modified. * @param[out] w output coefficient raft::device_vector_view */ template <typename ValueType, typename IndexType> void lstsq_qr(raft::resources const& handle, raft::device_matrix_view<const ValueType, IndexType, raft::col_major> A, raft::device_vector_view<const ValueType, IndexType> b, raft::device_vector_view<ValueType, IndexType> w) { RAFT_EXPECTS(A.extent(1) == w.size(), "Size mismatch between A and w"); RAFT_EXPECTS(A.extent(0) == b.size(), "Size mismatch between A and b"); lstsqQR(handle, const_cast<ValueType*>(A.data_handle()), A.extent(0), A.extent(1), const_cast<ValueType*>(b.data_handle()), w.data_handle(), resource::get_cuda_stream(handle)); } /** @} */ // end of lstsq }; // namespace linalg }; // namespace raft #endif
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/lanczos.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 // for cmath: #define _USE_MATH_DEFINES #include <cmath> #include <raft/core/resource/cublas_handle.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <vector> #include <cuda.h> #include <curand.h> #include "cublas_wrappers.hpp" #include <raft/core/resources.hpp> #include <raft/spectral/detail/lapack.hpp> #include <raft/spectral/detail/warn_dbg.hpp> #include <raft/spectral/matrix_wrappers.hpp> #include <raft/util/cudart_utils.hpp> namespace raft { namespace linalg { namespace detail { // curandGeneratorNormalX inline curandStatus_t curandGenerateNormalX( curandGenerator_t generator, float* outputPtr, size_t n, float mean, float stddev) { return curandGenerateNormal(generator, outputPtr, n, mean, stddev); } inline curandStatus_t curandGenerateNormalX( curandGenerator_t generator, double* outputPtr, size_t n, double mean, double stddev) { return curandGenerateNormalDouble(generator, outputPtr, n, mean, stddev); } // ========================================================= // Helper functions // ========================================================= /** * @brief Perform Lanczos iteration * Lanczos iteration is performed on a shifted matrix A+shift*I. * @tparam index_type_t the type of data used for indexing. * @tparam value_type_t the type of data used for weights, distances. * @param handle the raft handle. * @param A Matrix. * @param iter Pointer to current Lanczos iteration. On exit, the * variable is set equal to the final Lanczos iteration. * @param maxIter Maximum Lanczos iteration. This function will * perform a maximum of maxIter-*iter iterations. * @param shift Matrix shift. * @param tol Convergence tolerance. Lanczos iteration will * terminate when the residual norm (i.e. entry in beta_host) is * less than tol. * @param reorthogonalize Whether to reorthogonalize Lanczos * vectors. * @param alpha_host (Output, host memory, maxIter entries) * Diagonal entries of Lanczos system. * @param beta_host (Output, host memory, maxIter entries) * Off-diagonal entries of Lanczos system. * @param lanczosVecs_dev (Input/output, device memory, * n*(maxIter+1) entries) Lanczos vectors. Vectors are stored as * columns of a column-major matrix with dimensions * n x (maxIter+1). * @param work_dev (Output, device memory, maxIter entries) * Workspace. Not needed if full reorthogonalization is disabled. * @return Zero if successful. Otherwise non-zero. */ template <typename index_type_t, typename value_type_t> int performLanczosIteration(raft::resources const& handle, spectral::matrix::sparse_matrix_t<index_type_t, value_type_t> const* A, index_type_t* iter, index_type_t maxIter, value_type_t shift, value_type_t tol, bool reorthogonalize, value_type_t* __restrict__ alpha_host, value_type_t* __restrict__ beta_host, value_type_t* __restrict__ lanczosVecs_dev, value_type_t* __restrict__ work_dev) { // ------------------------------------------------------- // Variable declaration // ------------------------------------------------------- // Useful variables constexpr value_type_t one = 1; constexpr value_type_t negOne = -1; constexpr value_type_t zero = 0; value_type_t alpha; auto cublas_h = resource::get_cublas_handle(handle); auto stream = resource::get_cuda_stream(handle); RAFT_EXPECTS(A != nullptr, "Null matrix pointer."); index_type_t n = A->nrows_; // ------------------------------------------------------- // Compute second Lanczos vector // ------------------------------------------------------- if (*iter <= 0) { *iter = 1; // Apply matrix if (shift != 0) RAFT_CUDA_TRY(cudaMemcpyAsync(lanczosVecs_dev + n, lanczosVecs_dev, n * sizeof(value_type_t), cudaMemcpyDeviceToDevice, stream)); A->mv(1, lanczosVecs_dev, shift, lanczosVecs_dev + n); // Orthogonalize Lanczos vector RAFT_CUBLAS_TRY(cublasdot( cublas_h, n, lanczosVecs_dev, 1, lanczosVecs_dev + IDX(0, 1, n), 1, alpha_host, stream)); alpha = -alpha_host[0]; RAFT_CUBLAS_TRY(cublasaxpy( cublas_h, n, &alpha, lanczosVecs_dev, 1, lanczosVecs_dev + IDX(0, 1, n), 1, stream)); RAFT_CUBLAS_TRY(cublasnrm2(cublas_h, n, lanczosVecs_dev + IDX(0, 1, n), 1, beta_host, stream)); // Check if Lanczos has converged if (beta_host[0] <= tol) return 0; // Normalize Lanczos vector alpha = 1 / beta_host[0]; RAFT_CUBLAS_TRY(cublasscal(cublas_h, n, &alpha, lanczosVecs_dev + IDX(0, 1, n), 1, stream)); } // ------------------------------------------------------- // Compute remaining Lanczos vectors // ------------------------------------------------------- while (*iter < maxIter) { ++(*iter); // Apply matrix if (shift != 0) RAFT_CUDA_TRY(cudaMemcpyAsync(lanczosVecs_dev + (*iter) * n, lanczosVecs_dev + (*iter - 1) * n, n * sizeof(value_type_t), cudaMemcpyDeviceToDevice, stream)); A->mv(1, lanczosVecs_dev + IDX(0, *iter - 1, n), shift, lanczosVecs_dev + IDX(0, *iter, n)); // Full reorthogonalization // "Twice is enough" algorithm per Kahan and Parlett if (reorthogonalize) { RAFT_CUBLAS_TRY(cublasgemv(cublas_h, CUBLAS_OP_T, n, *iter, &one, lanczosVecs_dev, n, lanczosVecs_dev + IDX(0, *iter, n), 1, &zero, work_dev, 1, stream)); RAFT_CUBLAS_TRY(cublasgemv(cublas_h, CUBLAS_OP_N, n, *iter, &negOne, lanczosVecs_dev, n, work_dev, 1, &one, lanczosVecs_dev + IDX(0, *iter, n), 1, stream)); RAFT_CUDA_TRY(cudaMemcpyAsync(alpha_host + (*iter - 1), work_dev + (*iter - 1), sizeof(value_type_t), cudaMemcpyDeviceToHost, stream)); RAFT_CUBLAS_TRY(cublasgemv(cublas_h, CUBLAS_OP_T, n, *iter, &one, lanczosVecs_dev, n, lanczosVecs_dev + IDX(0, *iter, n), 1, &zero, work_dev, 1, stream)); RAFT_CUBLAS_TRY(cublasgemv(cublas_h, CUBLAS_OP_N, n, *iter, &negOne, lanczosVecs_dev, n, work_dev, 1, &one, lanczosVecs_dev + IDX(0, *iter, n), 1, stream)); } // Orthogonalization with 3-term recurrence relation else { RAFT_CUBLAS_TRY(cublasdot(cublas_h, n, lanczosVecs_dev + IDX(0, *iter - 1, n), 1, lanczosVecs_dev + IDX(0, *iter, n), 1, alpha_host + (*iter - 1), stream)); auto alpha = -alpha_host[*iter - 1]; RAFT_CUBLAS_TRY(cublasaxpy(cublas_h, n, &alpha, lanczosVecs_dev + IDX(0, *iter - 1, n), 1, lanczosVecs_dev + IDX(0, *iter, n), 1, stream)); alpha = -beta_host[*iter - 2]; RAFT_CUBLAS_TRY(cublasaxpy(cublas_h, n, &alpha, lanczosVecs_dev + IDX(0, *iter - 2, n), 1, lanczosVecs_dev + IDX(0, *iter, n), 1, stream)); } // Compute residual RAFT_CUBLAS_TRY(cublasnrm2( cublas_h, n, lanczosVecs_dev + IDX(0, *iter, n), 1, beta_host + *iter - 1, stream)); // Check if Lanczos has converged if (beta_host[*iter - 1] <= tol) break; // Normalize Lanczos vector alpha = 1 / beta_host[*iter - 1]; RAFT_CUBLAS_TRY(cublasscal(cublas_h, n, &alpha, lanczosVecs_dev + IDX(0, *iter, n), 1, stream)); } resource::sync_stream(handle, stream); return 0; } /** * @brief Find Householder transform for 3-dimensional system * Given an input vector v=[x,y,z]', this function finds a * Householder transform P such that P*v is a multiple of * e_1=[1,0,0]'. The input vector v is overwritten with the * Householder vector such that P=I-2*v*v'. * @tparam index_type_t the type of data used for indexing. * @tparam value_type_t the type of data used for weights, distances. * @param v (Input/output, host memory, 3 entries) Input * 3-dimensional vector. On exit, the vector is set to the * Householder vector. * @param Pv (Output, host memory, 1 entry) First entry of P*v * (here v is the input vector). Either equal to ||v||_2 or * -||v||_2. * @param P (Output, host memory, 9 entries) Householder transform * matrix. Matrix dimensions are 3 x 3. */ template <typename index_type_t, typename value_type_t> static void findHouseholder3(value_type_t* v, value_type_t* Pv, value_type_t* P) { // Compute norm of vector *Pv = std::sqrt(v[0] * v[0] + v[1] * v[1] + v[2] * v[2]); // Choose whether to reflect to e_1 or -e_1 // This choice avoids catastrophic cancellation if (v[0] >= 0) *Pv = -(*Pv); v[0] -= *Pv; // Normalize Householder vector value_type_t normHouseholder = std::sqrt(v[0] * v[0] + v[1] * v[1] + v[2] * v[2]); if (normHouseholder != 0) { v[0] /= normHouseholder; v[1] /= normHouseholder; v[2] /= normHouseholder; } else { v[0] = 0; v[1] = 0; v[2] = 0; } // Construct Householder matrix index_type_t i, j; for (j = 0; j < 3; ++j) for (i = 0; i < 3; ++i) P[IDX(i, j, 3)] = -2 * v[i] * v[j]; for (i = 0; i < 3; ++i) P[IDX(i, i, 3)] += 1; } /** * @brief Apply 3-dimensional Householder transform to 4 x 4 matrix * The Householder transform is pre-applied to the top three rows * of the matrix and post-applied to the left three columns. The * 4 x 4 matrix is intended to contain the bulge that is produced * in the Francis QR algorithm. * @tparam index_type_t the type of data used for indexing. * @tparam value_type_t the type of data used for weights, distances. * @param v (Input, host memory, 3 entries) Householder vector. * @param A (Input/output, host memory, 16 entries) 4 x 4 matrix. */ template <typename index_type_t, typename value_type_t> static void applyHouseholder3(const value_type_t* v, value_type_t* A) { // Loop indices index_type_t i, j; // Dot product between Householder vector and matrix row/column value_type_t vDotA; // Pre-apply Householder transform for (j = 0; j < 4; ++j) { vDotA = 0; for (i = 0; i < 3; ++i) vDotA += v[i] * A[IDX(i, j, 4)]; for (i = 0; i < 3; ++i) A[IDX(i, j, 4)] -= 2 * v[i] * vDotA; } // Post-apply Householder transform for (i = 0; i < 4; ++i) { vDotA = 0; for (j = 0; j < 3; ++j) vDotA += A[IDX(i, j, 4)] * v[j]; for (j = 0; j < 3; ++j) A[IDX(i, j, 4)] -= 2 * vDotA * v[j]; } } /** * @brief Perform one step of Francis QR algorithm * Equivalent to two steps of the classical QR algorithm on a * tridiagonal matrix. * @tparam index_type_t the type of data used for indexing. * @tparam value_type_t the type of data used for weights, distances. * @param n Matrix dimension. * @param shift1 QR algorithm shift. * @param shift2 QR algorithm shift. * @param alpha (Input/output, host memory, n entries) Diagonal * entries of tridiagonal matrix. * @param beta (Input/output, host memory, n-1 entries) * Off-diagonal entries of tridiagonal matrix. * @param V (Input/output, host memory, n*n entries) Orthonormal * transforms from previous steps of QR algorithm. Matrix * dimensions are n x n. On exit, the orthonormal transform from * this Francis QR step is post-applied to the matrix. * @param work (Output, host memory, 3*n entries) Workspace. * @return Zero if successful. Otherwise non-zero. */ template <typename index_type_t, typename value_type_t> static int francisQRIteration(index_type_t n, value_type_t shift1, value_type_t shift2, value_type_t* alpha, value_type_t* beta, value_type_t* V, value_type_t* work) { // ------------------------------------------------------- // Variable declaration // ------------------------------------------------------- // Temporary storage of 4x4 bulge and Householder vector value_type_t bulge[16]; // Householder vector value_type_t householder[3]; // Householder matrix value_type_t householderMatrix[3 * 3]; // Shifts are roots of the polynomial p(x)=x^2+b*x+c value_type_t b = -shift1 - shift2; value_type_t c = shift1 * shift2; // Loop indices index_type_t i, j, pos; // Temporary variable value_type_t temp; // ------------------------------------------------------- // Implementation // ------------------------------------------------------- // Compute initial Householder transform householder[0] = alpha[0] * alpha[0] + beta[0] * beta[0] + b * alpha[0] + c; householder[1] = beta[0] * (alpha[0] + alpha[1] + b); householder[2] = beta[0] * beta[1]; findHouseholder3<index_type_t, value_type_t>(householder, &temp, householderMatrix); // Apply initial Householder transform to create bulge memset(bulge, 0, 16 * sizeof(value_type_t)); for (i = 0; i < 4; ++i) bulge[IDX(i, i, 4)] = alpha[i]; for (i = 0; i < 3; ++i) { bulge[IDX(i + 1, i, 4)] = beta[i]; bulge[IDX(i, i + 1, 4)] = beta[i]; } applyHouseholder3<index_type_t, value_type_t>(householder, bulge); Lapack<value_type_t>::gemm(false, false, n, 3, 3, 1, V, n, householderMatrix, 3, 0, work, n); memcpy(V, work, 3 * n * sizeof(value_type_t)); // Chase bulge to bottom-right of matrix with Householder transforms for (pos = 0; pos < n - 4; ++pos) { // Move to next position alpha[pos] = bulge[IDX(0, 0, 4)]; householder[0] = bulge[IDX(1, 0, 4)]; householder[1] = bulge[IDX(2, 0, 4)]; householder[2] = bulge[IDX(3, 0, 4)]; for (j = 0; j < 3; ++j) for (i = 0; i < 3; ++i) bulge[IDX(i, j, 4)] = bulge[IDX(i + 1, j + 1, 4)]; bulge[IDX(3, 0, 4)] = 0; bulge[IDX(3, 1, 4)] = 0; bulge[IDX(3, 2, 4)] = beta[pos + 3]; bulge[IDX(0, 3, 4)] = 0; bulge[IDX(1, 3, 4)] = 0; bulge[IDX(2, 3, 4)] = beta[pos + 3]; bulge[IDX(3, 3, 4)] = alpha[pos + 4]; // Apply Householder transform findHouseholder3<index_type_t, value_type_t>(householder, beta + pos, householderMatrix); applyHouseholder3<index_type_t, value_type_t>(householder, bulge); Lapack<value_type_t>::gemm( false, false, n, 3, 3, 1, V + IDX(0, pos + 1, n), n, householderMatrix, 3, 0, work, n); memcpy(V + IDX(0, pos + 1, n), work, 3 * n * sizeof(value_type_t)); } // Apply penultimate Householder transform // Values in the last row and column are zero alpha[n - 4] = bulge[IDX(0, 0, 4)]; householder[0] = bulge[IDX(1, 0, 4)]; householder[1] = bulge[IDX(2, 0, 4)]; householder[2] = bulge[IDX(3, 0, 4)]; for (j = 0; j < 3; ++j) for (i = 0; i < 3; ++i) bulge[IDX(i, j, 4)] = bulge[IDX(i + 1, j + 1, 4)]; bulge[IDX(3, 0, 4)] = 0; bulge[IDX(3, 1, 4)] = 0; bulge[IDX(3, 2, 4)] = 0; bulge[IDX(0, 3, 4)] = 0; bulge[IDX(1, 3, 4)] = 0; bulge[IDX(2, 3, 4)] = 0; bulge[IDX(3, 3, 4)] = 0; findHouseholder3<index_type_t, value_type_t>(householder, beta + n - 4, householderMatrix); applyHouseholder3<index_type_t, value_type_t>(householder, bulge); Lapack<value_type_t>::gemm( false, false, n, 3, 3, 1, V + IDX(0, n - 3, n), n, householderMatrix, 3, 0, work, n); memcpy(V + IDX(0, n - 3, n), work, 3 * n * sizeof(value_type_t)); // Apply final Householder transform // Values in the last two rows and columns are zero alpha[n - 3] = bulge[IDX(0, 0, 4)]; householder[0] = bulge[IDX(1, 0, 4)]; householder[1] = bulge[IDX(2, 0, 4)]; householder[2] = 0; for (j = 0; j < 3; ++j) for (i = 0; i < 3; ++i) bulge[IDX(i, j, 4)] = bulge[IDX(i + 1, j + 1, 4)]; findHouseholder3<index_type_t, value_type_t>(householder, beta + n - 3, householderMatrix); applyHouseholder3<index_type_t, value_type_t>(householder, bulge); Lapack<value_type_t>::gemm( false, false, n, 2, 2, 1, V + IDX(0, n - 2, n), n, householderMatrix, 3, 0, work, n); memcpy(V + IDX(0, n - 2, n), work, 2 * n * sizeof(value_type_t)); // Bulge has been eliminated alpha[n - 2] = bulge[IDX(0, 0, 4)]; alpha[n - 1] = bulge[IDX(1, 1, 4)]; beta[n - 2] = bulge[IDX(1, 0, 4)]; return 0; } /** * @brief Perform implicit restart of Lanczos algorithm * Shifts are Chebyshev nodes of unwanted region of matrix spectrum. * @tparam index_type_t the type of data used for indexing. * @tparam value_type_t the type of data used for weights, distances. * @param handle the raft handle. * @param n Matrix dimension. * @param iter Current Lanczos iteration. * @param iter_new Lanczos iteration after restart. * @param shiftUpper Pointer (host memory) to upper bound for unwanted * region. Value is ignored if less than *shiftLower. If a * stronger upper bound has been found, the value is updated on * exit. * @param shiftLower Pointer (host memory) to lower bound for unwanted * region. Value is ignored if greater than *shiftUpper. If a * stronger lower bound has been found, the value is updated on * exit. * @param alpha_host (Input/output, host memory, iter entries) * Diagonal entries of Lanczos system. * @param beta_host (Input/output, host memory, iter entries) * Off-diagonal entries of Lanczos system. * @param V_host (Output, host memory, iter*iter entries) * Orthonormal transform used to obtain restarted system. Matrix * dimensions are iter x iter. * @param work_host (Output, host memory, 4*iter entries) * Workspace. * @param lanczosVecs_dev (Input/output, device memory, n*(iter+1) * entries) Lanczos vectors. Vectors are stored as columns of a * column-major matrix with dimensions n x (iter+1). * @param work_dev (Output, device memory, (n+iter)*iter entries) * Workspace. * @param smallest_eig specifies whether smallest (true) or largest * (false) eigenvalues are to be calculated. * @return error flag. */ template <typename index_type_t, typename value_type_t> static int lanczosRestart(raft::resources const& handle, index_type_t n, index_type_t iter, index_type_t iter_new, value_type_t* shiftUpper, value_type_t* shiftLower, value_type_t* __restrict__ alpha_host, value_type_t* __restrict__ beta_host, value_type_t* __restrict__ V_host, value_type_t* __restrict__ work_host, value_type_t* __restrict__ lanczosVecs_dev, value_type_t* __restrict__ work_dev, bool smallest_eig) { // ------------------------------------------------------- // Variable declaration // ------------------------------------------------------- // Useful constants constexpr value_type_t zero = 0; constexpr value_type_t one = 1; auto cublas_h = resource::get_cublas_handle(handle); auto stream = resource::get_cuda_stream(handle); // Loop index index_type_t i; // Number of implicit restart steps // Assumed to be even since each call to Francis algorithm is // equivalent to two calls of QR algorithm index_type_t restartSteps = iter - iter_new; // Ritz values from Lanczos method value_type_t* ritzVals_host = work_host + 3 * iter; // Shifts for implicit restart value_type_t* shifts_host; // Orthonormal matrix for similarity transform value_type_t* V_dev = work_dev + n * iter; // ------------------------------------------------------- // Implementation // ------------------------------------------------------- // Compute Ritz values memcpy(ritzVals_host, alpha_host, iter * sizeof(value_type_t)); memcpy(work_host, beta_host, (iter - 1) * sizeof(value_type_t)); Lapack<value_type_t>::sterf(iter, ritzVals_host, work_host); // Debug: Print largest eigenvalues // for (int i = iter-iter_new; i < iter; ++i) // std::cout <<*(ritzVals_host+i)<< " "; // std::cout <<std::endl; // Initialize similarity transform with identity matrix memset(V_host, 0, iter * iter * sizeof(value_type_t)); for (i = 0; i < iter; ++i) V_host[IDX(i, i, iter)] = 1; // Determine interval to suppress eigenvalues if (smallest_eig) { if (*shiftLower > *shiftUpper) { *shiftUpper = ritzVals_host[iter - 1]; *shiftLower = ritzVals_host[iter_new]; } else { *shiftUpper = std::max(*shiftUpper, ritzVals_host[iter - 1]); *shiftLower = std::min(*shiftLower, ritzVals_host[iter_new]); } } else { if (*shiftLower > *shiftUpper) { *shiftUpper = ritzVals_host[iter - iter_new - 1]; *shiftLower = ritzVals_host[0]; } else { *shiftUpper = std::max(*shiftUpper, ritzVals_host[iter - iter_new - 1]); *shiftLower = std::min(*shiftLower, ritzVals_host[0]); } } // Calculate Chebyshev nodes as shifts shifts_host = ritzVals_host; for (i = 0; i < restartSteps; ++i) { shifts_host[i] = cos((i + 0.5) * static_cast<value_type_t>(M_PI) / restartSteps); shifts_host[i] *= 0.5 * ((*shiftUpper) - (*shiftLower)); shifts_host[i] += 0.5 * ((*shiftUpper) + (*shiftLower)); } // Apply Francis QR algorithm to implicitly restart Lanczos for (i = 0; i < restartSteps; i += 2) if (francisQRIteration( iter, shifts_host[i], shifts_host[i + 1], alpha_host, beta_host, V_host, work_host)) WARNING("error in implicitly shifted QR algorithm"); // Obtain new residual RAFT_CUDA_TRY(cudaMemcpyAsync( V_dev, V_host, iter * iter * sizeof(value_type_t), cudaMemcpyHostToDevice, stream)); beta_host[iter - 1] = beta_host[iter - 1] * V_host[IDX(iter - 1, iter_new - 1, iter)]; RAFT_CUBLAS_TRY(cublasgemv(cublas_h, CUBLAS_OP_N, n, iter, beta_host + iter_new - 1, lanczosVecs_dev, n, V_dev + IDX(0, iter_new, iter), 1, beta_host + iter - 1, lanczosVecs_dev + IDX(0, iter, n), 1, stream)); // Obtain new Lanczos vectors RAFT_CUBLAS_TRY(cublasgemm(cublas_h, CUBLAS_OP_N, CUBLAS_OP_N, n, iter_new, iter, &one, lanczosVecs_dev, n, V_dev, iter, &zero, work_dev, n, stream)); RAFT_CUDA_TRY(cudaMemcpyAsync(lanczosVecs_dev, work_dev, n * iter_new * sizeof(value_type_t), cudaMemcpyDeviceToDevice, stream)); // Normalize residual to obtain new Lanczos vector RAFT_CUDA_TRY(cudaMemcpyAsync(lanczosVecs_dev + IDX(0, iter_new, n), lanczosVecs_dev + IDX(0, iter, n), n * sizeof(value_type_t), cudaMemcpyDeviceToDevice, stream)); RAFT_CUBLAS_TRY(cublasnrm2( cublas_h, n, lanczosVecs_dev + IDX(0, iter_new, n), 1, beta_host + iter_new - 1, stream)); auto h_beta = 1 / beta_host[iter_new - 1]; RAFT_CUBLAS_TRY( cublasscal(cublas_h, n, &h_beta, lanczosVecs_dev + IDX(0, iter_new, n), 1, stream)); return 0; } /** * @brief Compute smallest eigenvectors of symmetric matrix * Computes eigenvalues and eigenvectors that are least * positive. If matrix is positive definite or positive * semidefinite, the computed eigenvalues are smallest in * magnitude. * The largest eigenvalue is estimated by performing several * Lanczos iterations. An implicitly restarted Lanczos method is * then applied to A+s*I, where s is negative the largest * eigenvalue. * @tparam index_type_t the type of data used for indexing. * @tparam value_type_t the type of data used for weights, distances. * @param handle the raft handle. * @param A Matrix. * @param nEigVecs Number of eigenvectors to compute. * @param maxIter Maximum number of Lanczos steps. Does not include * Lanczos steps used to estimate largest eigenvalue. * @param restartIter Maximum size of Lanczos system before * performing an implicit restart. Should be at least 4. * @param tol Convergence tolerance. Lanczos iteration will * terminate when the residual norm is less than tol*theta, where * theta is an estimate for the smallest unwanted eigenvalue * (i.e. the (nEigVecs+1)th smallest eigenvalue). * @param reorthogonalize Whether to reorthogonalize Lanczos * vectors. * @param effIter On exit, pointer to final size of Lanczos system. * @param totalIter On exit, pointer to total number of Lanczos * iterations performed. Does not include Lanczos steps used to * estimate largest eigenvalue. * @param shift On exit, pointer to matrix shift (estimate for * largest eigenvalue). * @param alpha_host (Output, host memory, restartIter entries) * Diagonal entries of Lanczos system. * @param beta_host (Output, host memory, restartIter entries) * Off-diagonal entries of Lanczos system. * @param lanczosVecs_dev (Output, device memory, n*(restartIter+1) * entries) Lanczos vectors. Vectors are stored as columns of a * column-major matrix with dimensions n x (restartIter+1). * @param work_dev (Output, device memory, * (n+restartIter)*restartIter entries) Workspace. * @param eigVals_dev (Output, device memory, nEigVecs entries) * Largest eigenvalues of matrix. * @param eigVecs_dev (Output, device memory, n*nEigVecs entries) * Eigenvectors corresponding to smallest eigenvalues of * matrix. Vectors are stored as columns of a column-major matrix * with dimensions n x nEigVecs. * @param seed random seed. * @return error flag. */ template <typename index_type_t, typename value_type_t> int computeSmallestEigenvectors( raft::resources const& handle, spectral::matrix::sparse_matrix_t<index_type_t, value_type_t> const* A, index_type_t nEigVecs, index_type_t maxIter, index_type_t restartIter, value_type_t tol, bool reorthogonalize, index_type_t* effIter, index_type_t* totalIter, value_type_t* shift, value_type_t* __restrict__ alpha_host, value_type_t* __restrict__ beta_host, value_type_t* __restrict__ lanczosVecs_dev, value_type_t* __restrict__ work_dev, value_type_t* __restrict__ eigVals_dev, value_type_t* __restrict__ eigVecs_dev, unsigned long long seed) { // Useful constants constexpr value_type_t one = 1; constexpr value_type_t zero = 0; // Matrix dimension index_type_t n = A->nrows_; // Shift for implicit restart value_type_t shiftUpper; value_type_t shiftLower; // Lanczos iteration counters index_type_t maxIter_curr = restartIter; // Maximum size of Lanczos system // Status flags int status; // Loop index index_type_t i; // Host memory value_type_t* Z_host; // Eigenvectors in Lanczos basis value_type_t* work_host; // Workspace // ------------------------------------------------------- // Check that parameters are valid // ------------------------------------------------------- RAFT_EXPECTS(nEigVecs > 0 && nEigVecs <= n, "Invalid number of eigenvectors."); RAFT_EXPECTS(restartIter > 0, "Invalid restartIter."); RAFT_EXPECTS(tol > 0, "Invalid tolerance."); RAFT_EXPECTS(maxIter >= nEigVecs, "Invalid maxIter."); RAFT_EXPECTS(restartIter >= nEigVecs, "Invalid restartIter."); auto cublas_h = resource::get_cublas_handle(handle); auto stream = resource::get_cuda_stream(handle); // ------------------------------------------------------- // Variable initialization // ------------------------------------------------------- // Total number of Lanczos iterations *totalIter = 0; // Allocate host memory std::vector<value_type_t> Z_host_v(restartIter * restartIter); std::vector<value_type_t> work_host_v(4 * restartIter); Z_host = Z_host_v.data(); work_host = work_host_v.data(); // Initialize cuBLAS RAFT_CUBLAS_TRY(cublassetpointermode(cublas_h, CUBLAS_POINTER_MODE_HOST, stream)); // ------------------------------------------------------- // Compute largest eigenvalue to determine shift // ------------------------------------------------------- // Random number generator curandGenerator_t randGen; // Initialize random number generator curandCreateGenerator(&randGen, CURAND_RNG_PSEUDO_PHILOX4_32_10); curandSetPseudoRandomGeneratorSeed(randGen, seed); // Initialize initial Lanczos vector curandGenerateNormalX(randGen, lanczosVecs_dev, n + n % 2, zero, one); value_type_t normQ1; RAFT_CUBLAS_TRY(cublasnrm2(cublas_h, n, lanczosVecs_dev, 1, &normQ1, stream)); auto h_val = 1 / normQ1; RAFT_CUBLAS_TRY(cublasscal(cublas_h, n, &h_val, lanczosVecs_dev, 1, stream)); // Obtain tridiagonal matrix with Lanczos *effIter = 0; *shift = 0; status = performLanczosIteration<index_type_t, value_type_t>(handle, A, effIter, maxIter_curr, *shift, 0.0, reorthogonalize, alpha_host, beta_host, lanczosVecs_dev, work_dev); if (status) WARNING("error in Lanczos iteration"); // Determine largest eigenvalue Lapack<value_type_t>::sterf(*effIter, alpha_host, beta_host); *shift = -alpha_host[*effIter - 1]; // ------------------------------------------------------- // Compute eigenvectors of shifted matrix // ------------------------------------------------------- // Obtain tridiagonal matrix with Lanczos *effIter = 0; status = performLanczosIteration<index_type_t, value_type_t>(handle, A, effIter, maxIter_curr, *shift, 0, reorthogonalize, alpha_host, beta_host, lanczosVecs_dev, work_dev); if (status) WARNING("error in Lanczos iteration"); *totalIter += *effIter; // Apply Lanczos method until convergence shiftLower = 1; shiftUpper = -1; while (*totalIter < maxIter && beta_host[*effIter - 1] > tol * shiftLower) { // Determine number of restart steps // Number of steps must be even due to Francis algorithm index_type_t iter_new = nEigVecs + 1; if (restartIter - (maxIter - *totalIter) > nEigVecs + 1) iter_new = restartIter - (maxIter - *totalIter); if ((restartIter - iter_new) % 2) iter_new -= 1; if (iter_new == *effIter) break; // Implicit restart of Lanczos method status = lanczosRestart<index_type_t, value_type_t>(handle, n, *effIter, iter_new, &shiftUpper, &shiftLower, alpha_host, beta_host, Z_host, work_host, lanczosVecs_dev, work_dev, true); if (status) WARNING("error in Lanczos implicit restart"); *effIter = iter_new; // Check for convergence if (beta_host[*effIter - 1] <= tol * fabs(shiftLower)) break; // Proceed with Lanczos method status = performLanczosIteration<index_type_t, value_type_t>(handle, A, effIter, maxIter_curr, *shift, tol * fabs(shiftLower), reorthogonalize, alpha_host, beta_host, lanczosVecs_dev, work_dev); if (status) WARNING("error in Lanczos iteration"); *totalIter += *effIter - iter_new; } // Warning if Lanczos has failed to converge if (beta_host[*effIter - 1] > tol * fabs(shiftLower)) { WARNING("implicitly restarted Lanczos failed to converge"); } // Solve tridiagonal system memcpy(work_host + 2 * (*effIter), alpha_host, (*effIter) * sizeof(value_type_t)); memcpy(work_host + 3 * (*effIter), beta_host, (*effIter - 1) * sizeof(value_type_t)); Lapack<value_type_t>::steqr('I', *effIter, work_host + 2 * (*effIter), work_host + 3 * (*effIter), Z_host, *effIter, work_host); // Obtain desired eigenvalues by applying shift for (i = 0; i < *effIter; ++i) work_host[i + 2 * (*effIter)] -= *shift; for (i = *effIter; i < nEigVecs; ++i) work_host[i + 2 * (*effIter)] = 0; // Copy results to device memory RAFT_CUDA_TRY(cudaMemcpyAsync(eigVals_dev, work_host + 2 * (*effIter), nEigVecs * sizeof(value_type_t), cudaMemcpyHostToDevice, stream)); RAFT_CUDA_TRY(cudaMemcpyAsync(work_dev, Z_host, (*effIter) * nEigVecs * sizeof(value_type_t), cudaMemcpyHostToDevice, stream)); RAFT_CHECK_CUDA(stream); // Convert eigenvectors from Lanczos basis to standard basis RAFT_CUBLAS_TRY(cublasgemm(cublas_h, CUBLAS_OP_N, CUBLAS_OP_N, n, nEigVecs, *effIter, &one, lanczosVecs_dev, n, work_dev, *effIter, &zero, eigVecs_dev, n, stream)); // Clean up and exit curandDestroyGenerator(randGen); return 0; } template <typename index_type_t, typename value_type_t> int computeSmallestEigenvectors( raft::resources const& handle, spectral::matrix::sparse_matrix_t<index_type_t, value_type_t> const& A, index_type_t nEigVecs, index_type_t maxIter, index_type_t restartIter, value_type_t tol, bool reorthogonalize, index_type_t& iter, value_type_t* __restrict__ eigVals_dev, value_type_t* __restrict__ eigVecs_dev, unsigned long long seed = 1234567) { // Matrix dimension index_type_t n = A.nrows_; // Check that parameters are valid RAFT_EXPECTS(nEigVecs > 0 && nEigVecs <= n, "Invalid number of eigenvectors."); RAFT_EXPECTS(restartIter > 0, "Invalid restartIter."); RAFT_EXPECTS(tol > 0, "Invalid tolerance."); RAFT_EXPECTS(maxIter >= nEigVecs, "Invalid maxIter."); RAFT_EXPECTS(restartIter >= nEigVecs, "Invalid restartIter."); // Allocate memory std::vector<value_type_t> alpha_host_v(restartIter); std::vector<value_type_t> beta_host_v(restartIter); value_type_t* alpha_host = alpha_host_v.data(); value_type_t* beta_host = beta_host_v.data(); spectral::matrix::vector_t<value_type_t> lanczosVecs_dev(handle, n * (restartIter + 1)); spectral::matrix::vector_t<value_type_t> work_dev(handle, (n + restartIter) * restartIter); // Perform Lanczos method index_type_t effIter; value_type_t shift; int status = computeSmallestEigenvectors(handle, &A, nEigVecs, maxIter, restartIter, tol, reorthogonalize, &effIter, &iter, &shift, alpha_host, beta_host, lanczosVecs_dev.raw(), work_dev.raw(), eigVals_dev, eigVecs_dev, seed); // Clean up and return return status; } /** * @brief Compute largest eigenvectors of symmetric matrix * Computes eigenvalues and eigenvectors that are least * positive. If matrix is positive definite or positive * semidefinite, the computed eigenvalues are largest in * magnitude. * The largest eigenvalue is estimated by performing several * Lanczos iterations. An implicitly restarted Lanczos method is * then applied. * @tparam index_type_t the type of data used for indexing. * @tparam value_type_t the type of data used for weights, distances. * @param handle the raft handle. * @param A Matrix. * @param nEigVecs Number of eigenvectors to compute. * @param maxIter Maximum number of Lanczos steps. * @param restartIter Maximum size of Lanczos system before * performing an implicit restart. Should be at least 4. * @param tol Convergence tolerance. Lanczos iteration will * terminate when the residual norm is less than tol*theta, where * theta is an estimate for the largest unwanted eigenvalue * (i.e. the (nEigVecs+1)th largest eigenvalue). * @param reorthogonalize Whether to reorthogonalize Lanczos * vectors. * @param effIter On exit, pointer to final size of Lanczos system. * @param totalIter On exit, pointer to total number of Lanczos * iterations performed. * @param alpha_host (Output, host memory, restartIter entries) * Diagonal entries of Lanczos system. * @param beta_host (Output, host memory, restartIter entries) * Off-diagonal entries of Lanczos system. * @param lanczosVecs_dev (Output, device memory, n*(restartIter+1) * entries) Lanczos vectors. Vectors are stored as columns of a * column-major matrix with dimensions n x (restartIter+1). * @param work_dev (Output, device memory, * (n+restartIter)*restartIter entries) Workspace. * @param eigVals_dev (Output, device memory, nEigVecs entries) * Largest eigenvalues of matrix. * @param eigVecs_dev (Output, device memory, n*nEigVecs entries) * Eigenvectors corresponding to largest eigenvalues of * matrix. Vectors are stored as columns of a column-major matrix * with dimensions n x nEigVecs. * @param seed random seed. * @return error flag. */ template <typename index_type_t, typename value_type_t> int computeLargestEigenvectors( raft::resources const& handle, spectral::matrix::sparse_matrix_t<index_type_t, value_type_t> const* A, index_type_t nEigVecs, index_type_t maxIter, index_type_t restartIter, value_type_t tol, bool reorthogonalize, index_type_t* effIter, index_type_t* totalIter, value_type_t* __restrict__ alpha_host, value_type_t* __restrict__ beta_host, value_type_t* __restrict__ lanczosVecs_dev, value_type_t* __restrict__ work_dev, value_type_t* __restrict__ eigVals_dev, value_type_t* __restrict__ eigVecs_dev, unsigned long long seed) { // Useful constants constexpr value_type_t one = 1; constexpr value_type_t zero = 0; // Matrix dimension index_type_t n = A->nrows_; // Lanczos iteration counters index_type_t maxIter_curr = restartIter; // Maximum size of Lanczos system // Status flags int status; // Loop index index_type_t i; // Host memory value_type_t* Z_host; // Eigenvectors in Lanczos basis value_type_t* work_host; // Workspace // ------------------------------------------------------- // Check that LAPACK is enabled // ------------------------------------------------------- // Lapack<value_type_t>::check_lapack_enabled(); // ------------------------------------------------------- // Check that parameters are valid // ------------------------------------------------------- RAFT_EXPECTS(nEigVecs > 0 && nEigVecs <= n, "Invalid number of eigenvectors."); RAFT_EXPECTS(restartIter > 0, "Invalid restartIter."); RAFT_EXPECTS(tol > 0, "Invalid tolerance."); RAFT_EXPECTS(maxIter >= nEigVecs, "Invalid maxIter."); RAFT_EXPECTS(restartIter >= nEigVecs, "Invalid restartIter."); auto cublas_h = resource::get_cublas_handle(handle); auto stream = resource::get_cuda_stream(handle); // ------------------------------------------------------- // Variable initialization // ------------------------------------------------------- // Total number of Lanczos iterations *totalIter = 0; // Allocate host memory std::vector<value_type_t> Z_host_v(restartIter * restartIter); std::vector<value_type_t> work_host_v(4 * restartIter); Z_host = Z_host_v.data(); work_host = work_host_v.data(); // Initialize cuBLAS RAFT_CUBLAS_TRY(cublassetpointermode(cublas_h, CUBLAS_POINTER_MODE_HOST, stream)); // ------------------------------------------------------- // Compute largest eigenvalue // ------------------------------------------------------- // Random number generator curandGenerator_t randGen; // Initialize random number generator curandCreateGenerator(&randGen, CURAND_RNG_PSEUDO_PHILOX4_32_10); curandSetPseudoRandomGeneratorSeed(randGen, seed); // Initialize initial Lanczos vector curandGenerateNormalX(randGen, lanczosVecs_dev, n + n % 2, zero, one); value_type_t normQ1; RAFT_CUBLAS_TRY(cublasnrm2(cublas_h, n, lanczosVecs_dev, 1, &normQ1, stream)); auto h_val = 1 / normQ1; RAFT_CUBLAS_TRY(cublasscal(cublas_h, n, &h_val, lanczosVecs_dev, 1, stream)); // Obtain tridiagonal matrix with Lanczos *effIter = 0; value_type_t shift_val = 0.0; value_type_t* shift = &shift_val; status = performLanczosIteration<index_type_t, value_type_t>(handle, A, effIter, maxIter_curr, *shift, 0, reorthogonalize, alpha_host, beta_host, lanczosVecs_dev, work_dev); if (status) WARNING("error in Lanczos iteration"); *totalIter += *effIter; // Apply Lanczos method until convergence value_type_t shiftLower = 1; value_type_t shiftUpper = -1; while (*totalIter < maxIter && beta_host[*effIter - 1] > tol * shiftLower) { // Determine number of restart steps // Number of steps must be even due to Francis algorithm index_type_t iter_new = nEigVecs + 1; if (restartIter - (maxIter - *totalIter) > nEigVecs + 1) iter_new = restartIter - (maxIter - *totalIter); if ((restartIter - iter_new) % 2) iter_new -= 1; if (iter_new == *effIter) break; // Implicit restart of Lanczos method status = lanczosRestart<index_type_t, value_type_t>(handle, n, *effIter, iter_new, &shiftUpper, &shiftLower, alpha_host, beta_host, Z_host, work_host, lanczosVecs_dev, work_dev, false); if (status) WARNING("error in Lanczos implicit restart"); *effIter = iter_new; // Check for convergence if (beta_host[*effIter - 1] <= tol * fabs(shiftLower)) break; // Proceed with Lanczos method status = performLanczosIteration<index_type_t, value_type_t>(handle, A, effIter, maxIter_curr, *shift, tol * fabs(shiftLower), reorthogonalize, alpha_host, beta_host, lanczosVecs_dev, work_dev); if (status) WARNING("error in Lanczos iteration"); *totalIter += *effIter - iter_new; } // Warning if Lanczos has failed to converge if (beta_host[*effIter - 1] > tol * fabs(shiftLower)) { WARNING("implicitly restarted Lanczos failed to converge"); } for (int i = 0; i < restartIter; ++i) { for (int j = 0; j < restartIter; ++j) Z_host[i * restartIter + j] = 0; } // Solve tridiagonal system memcpy(work_host + 2 * (*effIter), alpha_host, (*effIter) * sizeof(value_type_t)); memcpy(work_host + 3 * (*effIter), beta_host, (*effIter - 1) * sizeof(value_type_t)); Lapack<value_type_t>::steqr('I', *effIter, work_host + 2 * (*effIter), work_host + 3 * (*effIter), Z_host, *effIter, work_host); // note: We need to pick the top nEigVecs eigenvalues // but effItter can be larger than nEigVecs // hence we add an offset for that case, because we want to access top nEigVecs eigenpairs in the // matrix of size effIter. remember the array is sorted, so it is not needed for smallest // eigenvalues case because the first ones are the smallest ones index_type_t top_eigenparis_idx_offset = *effIter - nEigVecs; // Debug : print nEigVecs largest eigenvalues // for (int i = top_eigenparis_idx_offset; i < *effIter; ++i) // std::cout <<*(work_host+(2*(*effIter)+i))<< " "; // std::cout <<std::endl; // Debug : print nEigVecs largest eigenvectors // for (int i = top_eigenparis_idx_offset; i < *effIter; ++i) //{ // for (int j = 0; j < *effIter; ++j) // std::cout <<Z_host[i*(*effIter)+j]<< " "; // std::cout <<std::endl; //} // Obtain desired eigenvalues by applying shift for (i = 0; i < *effIter; ++i) work_host[i + 2 * (*effIter)] -= *shift; for (i = 0; i < top_eigenparis_idx_offset; ++i) work_host[i + 2 * (*effIter)] = 0; // Copy results to device memory // skip smallest eigenvalue if needed RAFT_CUDA_TRY(cudaMemcpyAsync(eigVals_dev, work_host + 2 * (*effIter) + top_eigenparis_idx_offset, nEigVecs * sizeof(value_type_t), cudaMemcpyHostToDevice, stream)); // skip smallest eigenvector if needed RAFT_CUDA_TRY(cudaMemcpyAsync(work_dev, Z_host + (top_eigenparis_idx_offset * (*effIter)), (*effIter) * nEigVecs * sizeof(value_type_t), cudaMemcpyHostToDevice, stream)); RAFT_CHECK_CUDA(stream); // Convert eigenvectors from Lanczos basis to standard basis RAFT_CUBLAS_TRY(cublasgemm(cublas_h, CUBLAS_OP_N, CUBLAS_OP_N, n, nEigVecs, *effIter, &one, lanczosVecs_dev, n, work_dev, *effIter, &zero, eigVecs_dev, n, stream)); // Clean up and exit curandDestroyGenerator(randGen); return 0; } template <typename index_type_t, typename value_type_t> int computeLargestEigenvectors( raft::resources const& handle, spectral::matrix::sparse_matrix_t<index_type_t, value_type_t> const& A, index_type_t nEigVecs, index_type_t maxIter, index_type_t restartIter, value_type_t tol, bool reorthogonalize, index_type_t& iter, value_type_t* __restrict__ eigVals_dev, value_type_t* __restrict__ eigVecs_dev, unsigned long long seed = 123456) { // Matrix dimension index_type_t n = A.nrows_; // Check that parameters are valid RAFT_EXPECTS(nEigVecs > 0 && nEigVecs <= n, "Invalid number of eigenvectors."); RAFT_EXPECTS(restartIter > 0, "Invalid restartIter."); RAFT_EXPECTS(tol > 0, "Invalid tolerance."); RAFT_EXPECTS(maxIter >= nEigVecs, "Invalid maxIter."); RAFT_EXPECTS(restartIter >= nEigVecs, "Invalid restartIter."); // Allocate memory std::vector<value_type_t> alpha_host_v(restartIter); std::vector<value_type_t> beta_host_v(restartIter); value_type_t* alpha_host = alpha_host_v.data(); value_type_t* beta_host = beta_host_v.data(); spectral::matrix::vector_t<value_type_t> lanczosVecs_dev(handle, n * (restartIter + 1)); spectral::matrix::vector_t<value_type_t> work_dev(handle, (n + restartIter) * restartIter); // Perform Lanczos method index_type_t effIter; int status = computeLargestEigenvectors(handle, &A, nEigVecs, maxIter, restartIter, tol, reorthogonalize, &effIter, &iter, alpha_host, beta_host, lanczosVecs_dev.raw(), work_dev.raw(), eigVals_dev, eigVecs_dev, seed); // Clean up and return return status; } } // namespace detail } // namespace linalg } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/axpy.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 <cublas_v2.h> #include <raft/core/resource/cublas_handle.hpp> #include "cublas_wrappers.hpp" #include <raft/core/resources.hpp> namespace raft::linalg::detail { template <typename T, bool DevicePointerMode = false> void axpy(raft::resources const& handle, const int n, const T* alpha, const T* x, const int incx, T* y, const int incy, cudaStream_t stream) { auto cublas_h = resource::get_cublas_handle(handle); cublas_device_pointer_mode<DevicePointerMode> pmode(cublas_h); RAFT_CUBLAS_TRY(cublasaxpy(cublas_h, n, alpha, x, incx, y, incy, stream)); } } // namespace raft::linalg::detail
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/subtract.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/core/operators.hpp> #include <raft/linalg/binary_op.cuh> #include <raft/linalg/unary_op.cuh> #include <raft/util/cuda_utils.cuh> namespace raft { namespace linalg { namespace detail { template <typename InT, typename OutT = InT, typename IdxType = int> void subtractScalar(OutT* out, const InT* in, InT scalar, IdxType len, cudaStream_t stream) { raft::linalg::unaryOp(out, in, len, raft::sub_const_op<InT>(scalar), stream); } template <typename InT, typename OutT = InT, typename IdxType = int> void subtract(OutT* out, const InT* in1, const InT* in2, IdxType len, cudaStream_t stream) { raft::linalg::binaryOp(out, in1, in2, len, raft::sub_op(), stream); } template <class math_t, typename IdxType> RAFT_KERNEL subtract_dev_scalar_kernel(math_t* outDev, const math_t* inDev, const math_t* singleScalarDev, IdxType len) { // TODO: kernel do not use shared memory in current implementation int i = ((IdxType)blockIdx.x * (IdxType)blockDim.x) + threadIdx.x; if (i < len) { outDev[i] = inDev[i] - *singleScalarDev; } } template <typename math_t, typename IdxType = int, int TPB = 256> void subtractDevScalar(math_t* outDev, const math_t* inDev, const math_t* singleScalarDev, IdxType len, cudaStream_t stream) { // Just for the note - there is no way to express such operation with cuBLAS in effective way // https://stackoverflow.com/questions/14051064/add-scalar-to-vector-in-blas-cublas-cuda const IdxType nblks = raft::ceildiv(len, (IdxType)TPB); subtract_dev_scalar_kernel<math_t> <<<nblks, TPB, 0, stream>>>(outDev, inDev, singleScalarDev, len); RAFT_CUDA_TRY(cudaPeekAtLastError()); } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/reduce.cuh
/* * Copyright (c) 2022, 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> #include <raft/linalg/coalesced_reduction.cuh> #include <raft/linalg/strided_reduction.cuh> namespace raft { namespace linalg { namespace detail { template <typename InType, typename OutType = InType, typename IdxType = int, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void reduce(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, bool rowMajor, bool alongRows, cudaStream_t stream, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { if (rowMajor && alongRows) { raft::linalg::coalescedReduction<InType, OutType, IdxType>( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } else if (rowMajor && !alongRows) { raft::linalg::stridedReduction<InType, OutType, IdxType>( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } else if (!rowMajor && alongRows) { raft::linalg::stridedReduction<InType, OutType, IdxType>( dots, data, N, D, init, stream, inplace, main_op, reduce_op, final_op); } else { raft::linalg::coalescedReduction<InType, OutType, IdxType>( dots, data, N, D, init, stream, inplace, main_op, reduce_op, final_op); } } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/eig.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 "cusolver_wrappers.hpp" #include <cuda_runtime_api.h> #include <raft/core/resource/cusolver_dn_handle.hpp> #include <raft/core/resources.hpp> #include <raft/matrix/copy.cuh> #include <raft/util/cudart_utils.hpp> #include <rmm/device_scalar.hpp> #include <rmm/device_uvector.hpp> namespace raft { namespace linalg { namespace detail { template <typename math_t> void eigDC_legacy(raft::resources const& handle, const math_t* in, std::size_t n_rows, std::size_t n_cols, math_t* eig_vectors, math_t* eig_vals, cudaStream_t stream) { cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); int lwork; RAFT_CUSOLVER_TRY(cusolverDnsyevd_bufferSize(cusolverH, CUSOLVER_EIG_MODE_VECTOR, CUBLAS_FILL_MODE_UPPER, n_rows, in, n_cols, eig_vals, &lwork)); rmm::device_uvector<math_t> d_work(lwork, stream); rmm::device_scalar<int> d_dev_info(stream); raft::matrix::copy(handle, make_device_matrix_view<const math_t>(in, n_rows, n_cols), make_device_matrix_view<math_t>(eig_vectors, n_rows, n_cols)); RAFT_CUSOLVER_TRY(cusolverDnsyevd(cusolverH, CUSOLVER_EIG_MODE_VECTOR, CUBLAS_FILL_MODE_UPPER, n_rows, eig_vectors, n_cols, eig_vals, d_work.data(), lwork, d_dev_info.data(), stream)); RAFT_CUDA_TRY(cudaGetLastError()); auto dev_info = d_dev_info.value(stream); ASSERT(dev_info == 0, "eig.cuh: eigensolver couldn't converge to a solution. " "This usually occurs when some of the features do not vary enough."); } template <typename math_t> void eigDC(raft::resources const& handle, const math_t* in, std::size_t n_rows, std::size_t n_cols, math_t* eig_vectors, math_t* eig_vals, cudaStream_t stream) { #if CUDART_VERSION < 11010 eigDC_legacy(handle, in, n_rows, n_cols, eig_vectors, eig_vals, stream); #else cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); cusolverDnParams_t dn_params = nullptr; RAFT_CUSOLVER_TRY(cusolverDnCreateParams(&dn_params)); size_t workspaceDevice = 0; size_t workspaceHost = 0; RAFT_CUSOLVER_TRY(cusolverDnxsyevd_bufferSize(cusolverH, dn_params, CUSOLVER_EIG_MODE_VECTOR, CUBLAS_FILL_MODE_UPPER, static_cast<int64_t>(n_rows), eig_vectors, static_cast<int64_t>(n_cols), eig_vals, &workspaceDevice, &workspaceHost, stream)); rmm::device_uvector<math_t> d_work(workspaceDevice / sizeof(math_t), stream); rmm::device_scalar<int> d_dev_info(stream); std::vector<math_t> h_work(workspaceHost / sizeof(math_t)); raft::matrix::copy(handle, make_device_matrix_view<const math_t>(in, n_rows, n_cols), make_device_matrix_view<math_t>(eig_vectors, n_rows, n_cols)); RAFT_CUSOLVER_TRY(cusolverDnxsyevd(cusolverH, dn_params, CUSOLVER_EIG_MODE_VECTOR, CUBLAS_FILL_MODE_UPPER, static_cast<int64_t>(n_rows), eig_vectors, static_cast<int64_t>(n_cols), eig_vals, d_work.data(), workspaceDevice, h_work.data(), workspaceHost, d_dev_info.data(), stream)); RAFT_CUDA_TRY(cudaGetLastError()); RAFT_CUSOLVER_TRY(cusolverDnDestroyParams(dn_params)); int dev_info = d_dev_info.value(stream); ASSERT(dev_info == 0, "eig.cuh: eigensolver couldn't converge to a solution. " "This usually occurs when some of the features do not vary enough."); #endif } enum EigVecMemUsage { OVERWRITE_INPUT, COPY_INPUT }; template <typename math_t> void eigSelDC(raft::resources const& handle, math_t* in, std::size_t n_rows, std::size_t n_cols, std::size_t n_eig_vals, math_t* eig_vectors, math_t* eig_vals, EigVecMemUsage memUsage, cudaStream_t stream) { cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); int lwork; int h_meig; RAFT_CUSOLVER_TRY(cusolverDnsyevdx_bufferSize(cusolverH, CUSOLVER_EIG_MODE_VECTOR, CUSOLVER_EIG_RANGE_I, CUBLAS_FILL_MODE_UPPER, static_cast<int64_t>(n_rows), in, static_cast<int64_t>(n_cols), math_t(0.0), math_t(0.0), static_cast<int64_t>(n_cols - n_eig_vals + 1), static_cast<int64_t>(n_cols), &h_meig, eig_vals, &lwork)); rmm::device_uvector<math_t> d_work(lwork, stream); rmm::device_scalar<int> d_dev_info(stream); rmm::device_uvector<math_t> d_eig_vectors(0, stream); if (memUsage == OVERWRITE_INPUT) { RAFT_CUSOLVER_TRY(cusolverDnsyevdx(cusolverH, CUSOLVER_EIG_MODE_VECTOR, CUSOLVER_EIG_RANGE_I, CUBLAS_FILL_MODE_UPPER, static_cast<int64_t>(n_rows), in, static_cast<int64_t>(n_cols), math_t(0.0), math_t(0.0), static_cast<int64_t>(n_cols - n_eig_vals + 1), static_cast<int64_t>(n_cols), &h_meig, eig_vals, d_work.data(), lwork, d_dev_info.data(), stream)); } else if (memUsage == COPY_INPUT) { d_eig_vectors.resize(n_rows * n_cols, stream); raft::matrix::copy(handle, make_device_matrix_view<const math_t>(in, n_rows, n_cols), make_device_matrix_view(eig_vectors, n_rows, n_cols)); RAFT_CUSOLVER_TRY(cusolverDnsyevdx(cusolverH, CUSOLVER_EIG_MODE_VECTOR, CUSOLVER_EIG_RANGE_I, CUBLAS_FILL_MODE_UPPER, static_cast<int64_t>(n_rows), eig_vectors, static_cast<int64_t>(n_cols), math_t(0.0), math_t(0.0), static_cast<int64_t>(n_cols - n_eig_vals + 1), static_cast<int64_t>(n_cols), &h_meig, eig_vals, d_work.data(), lwork, d_dev_info.data(), stream)); } RAFT_CUDA_TRY(cudaGetLastError()); int dev_info = d_dev_info.value(stream); ASSERT(dev_info == 0, "eig.cuh: eigensolver couldn't converge to a solution. " "This usually occurs when some of the features do not vary enough."); if (memUsage == OVERWRITE_INPUT) { raft::matrix::trunc_zero_origin( handle, make_device_matrix_view<const math_t, size_t, col_major>(in, n_rows, n_eig_vals), make_device_matrix_view<math_t, size_t, col_major>(eig_vectors, n_rows, n_eig_vals)); } else if (memUsage == COPY_INPUT) { raft::matrix::trunc_zero_origin( handle, make_device_matrix_view<const math_t, size_t, col_major>( d_eig_vectors.data(), n_rows, n_eig_vals), make_device_matrix_view<math_t, size_t, col_major>(eig_vectors, n_rows, n_eig_vals)); } } template <typename math_t> void eigJacobi(raft::resources const& handle, const math_t* in, std::size_t n_rows, std::size_t n_cols, math_t* eig_vectors, math_t* eig_vals, cudaStream_t stream, math_t tol = 1.e-7, int sweeps = 15) { cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); syevjInfo_t syevj_params = nullptr; RAFT_CUSOLVER_TRY(cusolverDnCreateSyevjInfo(&syevj_params)); RAFT_CUSOLVER_TRY(cusolverDnXsyevjSetTolerance(syevj_params, tol)); RAFT_CUSOLVER_TRY(cusolverDnXsyevjSetMaxSweeps(syevj_params, sweeps)); int lwork; RAFT_CUSOLVER_TRY(cusolverDnsyevj_bufferSize(cusolverH, CUSOLVER_EIG_MODE_VECTOR, CUBLAS_FILL_MODE_UPPER, static_cast<int64_t>(n_rows), eig_vectors, static_cast<int64_t>(n_cols), eig_vals, &lwork, syevj_params)); rmm::device_uvector<math_t> d_work(lwork, stream); rmm::device_scalar<int> dev_info(stream); raft::matrix::copy(handle, make_device_matrix_view<const math_t>(in, n_rows, n_cols), make_device_matrix_view(eig_vectors, n_rows, n_cols)); RAFT_CUSOLVER_TRY(cusolverDnsyevj(cusolverH, CUSOLVER_EIG_MODE_VECTOR, CUBLAS_FILL_MODE_UPPER, static_cast<int64_t>(n_rows), eig_vectors, static_cast<int64_t>(n_cols), eig_vals, d_work.data(), lwork, dev_info.data(), syevj_params, stream)); int executed_sweeps; RAFT_CUSOLVER_TRY(cusolverDnXsyevjGetSweeps(cusolverH, syevj_params, &executed_sweeps)); RAFT_CUDA_TRY(cudaGetLastError()); RAFT_CUSOLVER_TRY(cusolverDnDestroySyevjInfo(syevj_params)); } } // namespace detail } // namespace linalg } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/reduce_cols_by_key.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 <cub/cub.cuh> #include <limits> #include <raft/util/cuda_utils.cuh> #include <stdlib.h> namespace raft { namespace linalg { namespace detail { ///@todo: support col-major ///@todo: specialize this to support shared-mem based atomics template <typename T, typename KeyIteratorT, typename IdxType> RAFT_KERNEL reduce_cols_by_key_direct_kernel( const T* data, const KeyIteratorT keys, T* out, IdxType nrows, IdxType ncols, IdxType nkeys) { typedef typename std::iterator_traits<KeyIteratorT>::value_type KeyType; IdxType idx = static_cast<IdxType>(blockIdx.x) * blockDim.x + threadIdx.x; if (idx >= (nrows * ncols)) return; ///@todo: yikes! use fast-int-div IdxType colId = idx % ncols; IdxType rowId = idx / ncols; KeyType key = keys[colId]; raft::myAtomicAdd(out + rowId * nkeys + key, data[idx]); } template <typename T, typename KeyIteratorT, typename IdxType> RAFT_KERNEL reduce_cols_by_key_cached_kernel( const T* data, const KeyIteratorT keys, T* out, IdxType nrows, IdxType ncols, IdxType nkeys) { typedef typename std::iterator_traits<KeyIteratorT>::value_type KeyType; extern __shared__ char smem[]; T* out_cache = reinterpret_cast<T*>(smem); // Initialize the shared memory accumulators to 0. for (IdxType idx = threadIdx.x; idx < nrows * nkeys; idx += blockDim.x) { out_cache[idx] = T{0}; } __syncthreads(); // Accumulate in shared memory for (IdxType idx = static_cast<IdxType>(blockIdx.x) * blockDim.x + threadIdx.x; idx < nrows * ncols; idx += blockDim.x * static_cast<IdxType>(gridDim.x)) { IdxType colId = idx % ncols; IdxType rowId = idx / ncols; KeyType key = keys[colId]; raft::myAtomicAdd(out_cache + rowId * nkeys + key, data[idx]); } // Add the shared-memory accumulators to the global results. __syncthreads(); for (IdxType idx = threadIdx.x; idx < nrows * nkeys; idx += blockDim.x) { T val = out_cache[idx]; if (val != T{0}) { raft::myAtomicAdd(out + idx, val); } } } /** * @brief Computes the sum-reduction of matrix columns for each given key * @tparam T the input data type (as well as the output reduced matrix) * @tparam KeyType data type of the keys * @tparam IdxType indexing arithmetic type * @param data the input data (dim = nrows x ncols). This is assumed to be in * row-major layout * @param keys keys array (len = ncols). It is assumed that each key in this * array is between [0, nkeys). In case this is not true, the caller is expected * to have called make_monotonic primitive to prepare such a contiguous and * monotonically increasing keys array. * @param out the output reduced matrix along columns (dim = nrows x nkeys). * This will be assumed to be in row-major layout * @param nrows number of rows in the input data * @param ncols number of columns in the input data * @param nkeys number of unique keys in the keys array * @param stream cuda stream to launch the kernel onto * @param reset_sums Whether to reset the output sums to zero before reducing */ template <typename T, typename KeyIteratorT, typename IdxType = int> void reduce_cols_by_key(const T* data, const KeyIteratorT keys, T* out, IdxType nrows, IdxType ncols, IdxType nkeys, cudaStream_t stream, bool reset_sums) { typedef typename std::iterator_traits<KeyIteratorT>::value_type KeyType; RAFT_EXPECTS(static_cast<size_t>(nrows) * static_cast<size_t>(ncols) <= static_cast<size_t>(std::numeric_limits<IdxType>::max()), "Index type too small to represent indices in the input array."); RAFT_EXPECTS(static_cast<size_t>(nrows) * static_cast<size_t>(nkeys) <= static_cast<size_t>(std::numeric_limits<IdxType>::max()), "Index type too small to represent indices in the output array."); // Memset the output to zero to use atomics-based reduction. if (reset_sums) { RAFT_CUDA_TRY(cudaMemsetAsync(out, 0, sizeof(T) * nrows * nkeys, stream)); } // The cached version is used when the cache fits in shared memory and the number of input // elements is above a threshold (the cached version is slightly slower for small input arrays, // and orders of magnitude faster for large input arrays). size_t cache_size = static_cast<size_t>(nrows * nkeys) * sizeof(T); if (cache_size <= 49152ull && nrows * ncols >= IdxType{8192}) { constexpr int TPB = 256; int n_sm = raft::getMultiProcessorCount(); int target_nblks = 4 * n_sm; int max_nblks = raft::ceildiv<IdxType>(nrows * ncols, TPB); int nblks = std::min(target_nblks, max_nblks); reduce_cols_by_key_cached_kernel<<<nblks, TPB, cache_size, stream>>>( data, keys, out, nrows, ncols, nkeys); } else { constexpr int TPB = 256; int nblks = raft::ceildiv<IdxType>(nrows * ncols, TPB); reduce_cols_by_key_direct_kernel<<<nblks, TPB, 0, stream>>>( data, keys, out, nrows, ncols, nkeys); } RAFT_CUDA_TRY(cudaPeekAtLastError()); } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/eltwise.cuh
/* * Copyright (c) 2022, 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> #include <raft/linalg/binary_op.cuh> #include <raft/linalg/unary_op.cuh> namespace raft { namespace linalg { namespace detail { template <typename InType, typename IdxType, typename OutType = InType> void scalarAdd(OutType* out, const InType* in, InType scalar, IdxType len, cudaStream_t stream) { raft::linalg::unaryOp(out, in, len, raft::add_const_op<InType>(scalar), stream); } template <typename InType, typename IdxType, typename OutType = InType> void scalarMultiply(OutType* out, const InType* in, InType scalar, IdxType len, cudaStream_t stream) { raft::linalg::unaryOp(out, in, len, raft::mul_const_op<InType>(scalar), stream); } template <typename InType, typename IdxType, typename OutType = InType> void eltwiseAdd( OutType* out, const InType* in1, const InType* in2, IdxType len, cudaStream_t stream) { raft::linalg::binaryOp(out, in1, in2, len, raft::add_op(), stream); } template <typename InType, typename IdxType, typename OutType = InType> void eltwiseSub( OutType* out, const InType* in1, const InType* in2, IdxType len, cudaStream_t stream) { raft::linalg::binaryOp(out, in1, in2, len, raft::sub_op(), stream); } template <typename InType, typename IdxType, typename OutType = InType> void eltwiseMultiply( OutType* out, const InType* in1, const InType* in2, IdxType len, cudaStream_t stream) { raft::linalg::binaryOp(out, in1, in2, len, raft::mul_op(), stream); } template <typename InType, typename IdxType, typename OutType = InType> void eltwiseDivide( OutType* out, const InType* in1, const InType* in2, IdxType len, cudaStream_t stream) { raft::linalg::binaryOp(out, in1, in2, len, raft::div_op(), stream); } template <typename InType, typename IdxType, typename OutType = InType> void eltwiseDivideCheckZero( OutType* out, const InType* in1, const InType* in2, IdxType len, cudaStream_t stream) { raft::linalg::binaryOp(out, in1, in2, len, raft::div_checkzero_op(), stream); } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/matrix_vector_op.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/core/resource/cuda_stream.hpp> #include <raft/matrix/linewise_op.cuh> namespace raft { namespace linalg { namespace detail { template <typename MatT, typename Lambda, typename VecT, typename IdxType = int, int TPB = 256> void matrixVectorOp(MatT* out, const MatT* matrix, const VecT* vec, IdxType D, IdxType N, bool rowMajor, bool bcastAlongRows, Lambda op, cudaStream_t stream) { raft::resources handle; resource::set_cuda_stream(handle, stream); bool along_lines = rowMajor == bcastAlongRows; if (rowMajor) { matrix::linewise_op<MatT, IdxType, row_major, Lambda>( handle, make_device_matrix_view<const MatT, IdxType, row_major>(matrix, N, D), make_device_matrix_view<MatT, IdxType, row_major>(out, N, D), along_lines, op, make_device_vector_view<const VecT, IdxType>(vec, bcastAlongRows ? N : D)); } else { matrix::linewise_op<MatT, IdxType, col_major, Lambda>( handle, make_device_matrix_view<const MatT, IdxType, col_major>(matrix, N, D), make_device_matrix_view<MatT, IdxType, col_major>(out, N, D), along_lines, op, make_device_vector_view<const VecT, IdxType>(vec, bcastAlongRows ? N : D)); } } template <typename MatT, typename Lambda, typename Vec1T, typename Vec2T, typename IdxType = int, int TPB = 256> void matrixVectorOp(MatT* out, const MatT* matrix, const Vec1T* vec1, const Vec2T* vec2, IdxType D, IdxType N, bool rowMajor, bool bcastAlongRows, Lambda op, cudaStream_t stream) { raft::resources handle; resource::set_cuda_stream(handle, stream); bool along_lines = rowMajor == bcastAlongRows; if (rowMajor) { matrix::linewise_op<MatT, IdxType, row_major, Lambda>( handle, make_device_matrix_view<const MatT, IdxType, row_major>(matrix, N, D), make_device_matrix_view<MatT, IdxType, row_major>(out, N, D), along_lines, op, make_device_vector_view<const Vec1T, IdxType>(vec1, bcastAlongRows ? N : D), make_device_vector_view<const Vec2T, IdxType>(vec2, bcastAlongRows ? N : D)); } else { matrix::linewise_op<MatT, IdxType, col_major, Lambda>( handle, make_device_matrix_view<const MatT, IdxType, col_major>(matrix, N, D), make_device_matrix_view<MatT, IdxType, col_major>(out, N, D), along_lines, op, make_device_vector_view<const Vec1T, IdxType>(vec1, bcastAlongRows ? N : D), make_device_vector_view<const Vec2T, IdxType>(vec2, bcastAlongRows ? N : D)); } } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/map.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/core/device_mdspan.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resources.hpp> #include <raft/util/cuda_utils.cuh> #include <raft/util/input_validation.hpp> #include <raft/util/integer_utils.hpp> #include <raft/util/pow2_utils.cuh> #include <raft/util/vectorized.cuh> #include <rmm/cuda_stream_view.hpp> #include <thrust/tuple.h> namespace raft::linalg::detail { template <bool PassOffset, typename OutT, typename IdxT, typename Func, typename... InTs> __device__ __forceinline__ auto map_apply(Func f, const IdxT& offset, const InTs&... ins) -> OutT { if constexpr (PassOffset) { return f(offset, ins...); } else { return f(ins...); } } template <int R, bool PassOffset, typename OutT, typename IdxT, typename Func, typename... InTs, size_t... Is> __device__ __forceinline__ void map_kernel_mainloop( OutT* out_ptr, IdxT offset, IdxT len, Func f, const InTs*... in_ptrs, std::index_sequence<Is...>) { TxN_t<OutT, R> wide; thrust::tuple<TxN_t<InTs, R>...> wide_args; if (offset + R <= len) { (thrust::get<Is>(wide_args).load(in_ptrs, offset), ...); #pragma unroll for (int j = 0; j < R; ++j) { wide.val.data[j] = map_apply<PassOffset, OutT, IdxT, Func, InTs...>( f, offset + j, thrust::get<Is>(wide_args).val.data[j]...); } wide.store(out_ptr, offset); } } template <int R, bool PassOffset, typename OutT, typename IdxT, typename Func, typename... InTs> RAFT_KERNEL map_kernel(OutT* out_ptr, IdxT len, Func f, const InTs*... in_ptrs) { const IdxT tid = blockIdx.x * blockDim.x + threadIdx.x; if constexpr (R <= 1) { if (tid < len) { out_ptr[tid] = map_apply<PassOffset, OutT, IdxT, Func, InTs...>(f, tid, in_ptrs[tid]...); } } else { using align_bytes = Pow2<sizeof(OutT) * size_t(R)>; using align_elems = Pow2<R>; // how many elements to skip in order to do aligned vectorized store const IdxT skip_cnt_left = std::min<IdxT>(IdxT(align_bytes::roundUp(out_ptr) - out_ptr), len); // The main loop: process all aligned data map_kernel_mainloop<R, PassOffset, OutT, IdxT, Func, InTs...>( out_ptr, tid * R + skip_cnt_left, len, f, in_ptrs..., std::index_sequence_for<InTs...>()); static_assert(WarpSize >= R); // Processes the skipped elements on the left if (tid < skip_cnt_left) { out_ptr[tid] = map_apply<PassOffset, OutT, IdxT, Func, InTs...>(f, tid, in_ptrs[tid]...); } // Processes the skipped elements on the right const IdxT skip_cnt_right = align_elems::mod(len - skip_cnt_left); const IdxT remain_i = len - skip_cnt_right + tid; if (remain_i < len) { out_ptr[remain_i] = map_apply<PassOffset, OutT, IdxT, Func, InTs...>(f, remain_i, in_ptrs[remain_i]...); } } } template <int R, bool PassOffset, typename OutT, typename IdxT, typename Func, typename... InTs> void map_call(rmm::cuda_stream_view stream, OutT* out_ptr, IdxT len, Func f, const InTs*... in_ptrs) { const IdxT len_vectorized = raft::div_rounding_up_safe<IdxT>(len, R); const int threads = std::max<int>(WarpSize, std::min<IdxT>(raft::bound_by_power_of_two<IdxT>(len_vectorized), 256)); const IdxT blocks = raft::div_rounding_up_unsafe<IdxT>(len_vectorized, threads); map_kernel<R, PassOffset><<<blocks, threads, 0, stream>>>(out_ptr, len, f, in_ptrs...); } constexpr int kCoalescedVectorSize = 16; constexpr int kSmallInputThreshold = 1024; struct ratio_selector { int ratio; int align; constexpr inline ratio_selector(int r, int a) : ratio(r), align(a) {} template <typename T> constexpr static auto ignoring_alignment() -> ratio_selector { constexpr bool T_evenly_fits_in_cache_line = (kCoalescedVectorSize % sizeof(T)) == 0; if constexpr (T_evenly_fits_in_cache_line) { return ratio_selector{size_t(kCoalescedVectorSize / sizeof(T)), 0}; } else { return ratio_selector{1, 0}; } } template <typename T> explicit ratio_selector(const T* ptr) { constexpr auto s = ignoring_alignment<T>(); // NOLINT if constexpr (s.ratio == 1) { align = 0; } else { align = int(Pow2<sizeof(T) * s.ratio>::roundUp(ptr) - ptr); } ratio = int(s.ratio); } }; constexpr inline auto operator*(const ratio_selector& a, const ratio_selector& b) -> ratio_selector { auto ratio = std::min<int>(a.ratio, b.ratio); while ((a.align % ratio) != (b.align % ratio)) { ratio >>= 1; } return ratio_selector{ratio, a.align % ratio}; } template <int R, bool PassOffset, typename OutT, typename IdxT, typename Func, typename... InTs> void map_call_rt( int r, rmm::cuda_stream_view stream, OutT* out_ptr, IdxT len, Func f, const InTs*... in_ptrs) { if (r >= R) { return map_call<R, PassOffset>(stream, out_ptr, len, f, in_ptrs...); } if constexpr (R > 1) { return map_call_rt<(R >> 1), PassOffset>(r, stream, out_ptr, len, f, in_ptrs...); } } template <bool PassOffset, typename OutT, typename IdxT, typename Func, typename... InTs> void map(rmm::cuda_stream_view stream, OutT* out_ptr, IdxT len, Func f, const InTs*... in_ptrs) { // don't vectorize on small inputs if (len <= kSmallInputThreshold) { return map_call<1, PassOffset>(stream, out_ptr, len, f, in_ptrs...); } constexpr int kRatio = (ratio_selector::ignoring_alignment<OutT>() * ... * ratio_selector::ignoring_alignment<InTs>()) .ratio; static_assert(kRatio > 0, "Unexpected zero vector size."); const int ratio = (ratio_selector(out_ptr) * ... * ratio_selector(in_ptrs)).ratio; return map_call_rt<kRatio, PassOffset>(ratio, stream, out_ptr, len, f, in_ptrs...); } template <typename OutType, typename InType, typename = raft::enable_if_output_device_mdspan<OutType>, typename = raft::enable_if_input_device_mdspan<InType>> void map_check_shape(OutType out, InType in) { RAFT_EXPECTS(raft::is_row_or_column_major(in) && out.size() == in.size(), "All inputs must be contiguous and have the same size as the output"); } /** * @brief Map a function over a zero or more inputs and optionally a 0-based flat index * (element offset). * * _Performance note_: when possible, this function loads the argument arrays and stores the output * array using vectorized cuda load/store instructions. The size of the vectorization depends on the * size of the largest input/output element type and on the alignment of all pointers. * * @tparam PassOffset whether to pass an offset as a first argument to Func * @tparam OutType data-type of the result (device_mdspan) * @tparam Func the device-lambda performing the actual operation * @tparam InTypes data-types of the inputs (device_mdspan) * * @param[in] res raft::resources * @param[out] out the output of the map operation (device_mdspan) * @param[in] f device lambda of type * ([auto offset], InTypes::value_type xs...) -> OutType::value_type * @param[in] ins the inputs (each of the same size as the output) (device_mdspan) */ template <bool PassOffset, typename OutType, typename Func, typename... InTypes, typename = raft::enable_if_output_device_mdspan<OutType>, typename = raft::enable_if_input_device_mdspan<InTypes...>> void map(const raft::resources& res, OutType out, Func f, InTypes... ins) { RAFT_EXPECTS(raft::is_row_or_column_major(out), "Output must be contiguous"); (map_check_shape(out, ins), ...); if (out.size() <= std::numeric_limits<std::uint32_t>::max()) { map<PassOffset, typename OutType::value_type, std::uint32_t, Func, typename InTypes::value_type...>(resource::get_cuda_stream(res), out.data_handle(), uint32_t(out.size()), f, ins.data_handle()...); } else { map<PassOffset, typename OutType::value_type, std::uint64_t, Func, typename InTypes::value_type...>(resource::get_cuda_stream(res), out.data_handle(), uint64_t(out.size()), f, ins.data_handle()...); } } } // namespace raft::linalg::detail
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/strided_reduction.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 <cub/cub.cuh> #include <raft/core/operators.hpp> #include <raft/linalg/unary_op.cuh> #include <raft/util/cuda_utils.cuh> #include <type_traits> namespace raft { namespace linalg { namespace detail { // Kernel to perform reductions along the strided dimension // of the matrix, i.e. reduce along columns for row major or reduce along rows // for column major layout template <typename Type, typename MainLambda> RAFT_KERNEL stridedSummationKernel( Type* dots, const Type* data, int D, int N, Type init, MainLambda main_op) { // Thread reduction Type thread_data = Type(init); int colStart = blockIdx.x * blockDim.x + threadIdx.x; if (colStart < D) { int rowStart = blockIdx.y * blockDim.y + threadIdx.y; int stride = blockDim.y * gridDim.y; for (int j = rowStart; j < N; j += stride) { int idx = colStart + j * D; thread_data += main_op(data[idx], j); } } // Block reduction extern __shared__ char tmp[]; // One element per thread in block Type* temp = (Type*)tmp; // Cast to desired type int myidx = threadIdx.x + blockDim.x * threadIdx.y; temp[myidx] = thread_data; __syncthreads(); for (int j = blockDim.y / 2; j > 0; j /= 2) { if (threadIdx.y < j) temp[myidx] += temp[myidx + j * blockDim.x]; __syncthreads(); } // Grid reduction if ((colStart < D) && (threadIdx.y == 0)) raft::myAtomicAdd(dots + colStart, temp[myidx]); } // Kernel to perform reductions along the strided dimension // of the matrix, i.e. reduce along columns for row major or reduce along rows // for column major layout template <typename InType, typename OutType, typename IdxType, typename MainLambda, typename ReduceLambda> RAFT_KERNEL stridedReductionKernel(OutType* dots, const InType* data, int D, int N, OutType init, MainLambda main_op, ReduceLambda reduce_op) { // Thread reduction OutType thread_data = init; IdxType colStart = blockIdx.x * blockDim.x + threadIdx.x; if (colStart < D) { IdxType rowStart = blockIdx.y * blockDim.y + threadIdx.y; IdxType stride = blockDim.y * gridDim.y; for (IdxType j = rowStart; j < N; j += stride) { IdxType idx = colStart + j * D; thread_data = reduce_op(thread_data, main_op(data[idx], j)); } } // Block reduction extern __shared__ char tmp[]; // One element per thread in block auto* temp = (OutType*)tmp; // Cast to desired type IdxType myidx = threadIdx.x + ((IdxType)blockDim.x * (IdxType)threadIdx.y); temp[myidx] = thread_data; __syncthreads(); for (int j = blockDim.y / 2; j > 0; j /= 2) { if (threadIdx.y < j) temp[myidx] = reduce_op(temp[myidx], temp[myidx + j * blockDim.x]); __syncthreads(); } // Grid reduction if ((colStart < D) && (threadIdx.y == 0)) raft::myAtomicReduce(dots + colStart, temp[myidx], reduce_op); } template <typename InType, typename OutType = InType, typename IdxType = int, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void stridedReduction(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, cudaStream_t stream, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { ///@todo: this extra should go away once we have eliminated the need /// for atomics in stridedKernel (redesign for this is already underway) if (!inplace) raft::linalg::unaryOp(dots, dots, D, raft::const_op(init), stream); // Arbitrary numbers for now, probably need to tune const dim3 thrds(32, 16); IdxType elemsPerThread = raft::ceildiv(N, (IdxType)thrds.y); elemsPerThread = (elemsPerThread > 8) ? 8 : elemsPerThread; const dim3 nblks(raft::ceildiv(D, (IdxType)thrds.x), raft::ceildiv(N, (IdxType)thrds.y * elemsPerThread)); const size_t shmemSize = sizeof(OutType) * thrds.x * thrds.y; ///@todo: this complication should go away once we have eliminated the need /// for atomics in stridedKernel (redesign for this is already underway) if constexpr (std::is_same<ReduceLambda, raft::add_op>::value && std::is_same<InType, OutType>::value) stridedSummationKernel<InType> <<<nblks, thrds, shmemSize, stream>>>(dots, data, D, N, init, main_op); else stridedReductionKernel<InType, OutType, IdxType> <<<nblks, thrds, shmemSize, stream>>>(dots, data, D, N, init, main_op, reduce_op); ///@todo: this complication should go away once we have eliminated the need /// for atomics in stridedKernel (redesign for this is already underway) // Perform final op on output data if (!std::is_same<FinalLambda, raft::identity_op>::value) raft::linalg::unaryOp(dots, dots, D, final_op, stream); } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/cusolver_wrappers.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 <cusolverDn.h> #include <cusolverSp.h> #include <raft/core/cusolver_macros.hpp> #include <raft/util/cudart_utils.hpp> #include <type_traits> namespace raft { namespace linalg { namespace detail { /** * @defgroup Getrf cusolver getrf operations * @{ */ template <typename T> cusolverStatus_t cusolverDngetrf(cusolverDnHandle_t handle, int m, // NOLINT int n, T* A, int lda, T* Workspace, int* devIpiv, int* devInfo, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDngetrf(cusolverDnHandle_t handle, // NOLINT int m, int n, float* A, int lda, float* Workspace, int* devIpiv, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnSgetrf(handle, m, n, A, lda, Workspace, devIpiv, devInfo); } template <> inline cusolverStatus_t cusolverDngetrf(cusolverDnHandle_t handle, // NOLINT int m, int n, double* A, int lda, double* Workspace, int* devIpiv, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnDgetrf(handle, m, n, A, lda, Workspace, devIpiv, devInfo); } template <typename T> cusolverStatus_t cusolverDngetrf_bufferSize( // NOLINT cusolverDnHandle_t handle, int m, int n, T* A, int lda, int* Lwork); template <> inline cusolverStatus_t cusolverDngetrf_bufferSize( // NOLINT cusolverDnHandle_t handle, int m, int n, float* A, int lda, int* Lwork) { return cusolverDnSgetrf_bufferSize(handle, m, n, A, lda, Lwork); } template <> inline cusolverStatus_t cusolverDngetrf_bufferSize( // NOLINT cusolverDnHandle_t handle, int m, int n, double* A, int lda, int* Lwork) { return cusolverDnDgetrf_bufferSize(handle, m, n, A, lda, Lwork); } /** * @defgroup Getrs cusolver getrs operations * @{ */ template <typename T> cusolverStatus_t cusolverDngetrs(cusolverDnHandle_t handle, // NOLINT cublasOperation_t trans, int n, int nrhs, const T* A, int lda, const int* devIpiv, T* B, int ldb, int* devInfo, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDngetrs(cusolverDnHandle_t handle, // NOLINT cublasOperation_t trans, int n, int nrhs, const float* A, int lda, const int* devIpiv, float* B, int ldb, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnSgetrs(handle, trans, n, nrhs, A, lda, devIpiv, B, ldb, devInfo); } template <> inline cusolverStatus_t cusolverDngetrs(cusolverDnHandle_t handle, // NOLINT cublasOperation_t trans, int n, int nrhs, const double* A, int lda, const int* devIpiv, double* B, int ldb, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnDgetrs(handle, trans, n, nrhs, A, lda, devIpiv, B, ldb, devInfo); } /** @} */ /** * @defgroup syevd cusolver syevd operations * @{ */ template <typename T> cusolverStatus_t cusolverDnsyevd_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, const T* A, int lda, const T* W, int* lwork); template <> inline cusolverStatus_t cusolverDnsyevd_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, const float* A, int lda, const float* W, int* lwork) { return cusolverDnSsyevd_bufferSize(handle, jobz, uplo, n, A, lda, W, lwork); } template <> inline cusolverStatus_t cusolverDnsyevd_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, const double* A, int lda, const double* W, int* lwork) { return cusolverDnDsyevd_bufferSize(handle, jobz, uplo, n, A, lda, W, lwork); } /** @} */ /** * @defgroup syevj cusolver syevj operations * @{ */ template <typename T> cusolverStatus_t cusolverDnsyevj(cusolverDnHandle_t handle, // NOLINT cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, T* A, int lda, T* W, T* work, int lwork, int* info, syevjInfo_t params, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDnsyevj( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, float* A, int lda, float* W, float* work, int lwork, int* info, syevjInfo_t params, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnSsyevj(handle, jobz, uplo, n, A, lda, W, work, lwork, info, params); } template <> inline cusolverStatus_t cusolverDnsyevj( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, double* A, int lda, double* W, double* work, int lwork, int* info, syevjInfo_t params, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnDsyevj(handle, jobz, uplo, n, A, lda, W, work, lwork, info, params); } template <typename T> cusolverStatus_t cusolverDnsyevj_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, const T* A, int lda, const T* W, int* lwork, syevjInfo_t params); template <> inline cusolverStatus_t cusolverDnsyevj_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, const float* A, int lda, const float* W, int* lwork, syevjInfo_t params) { return cusolverDnSsyevj_bufferSize(handle, jobz, uplo, n, A, lda, W, lwork, params); } template <> inline cusolverStatus_t cusolverDnsyevj_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, const double* A, int lda, const double* W, int* lwork, syevjInfo_t params) { return cusolverDnDsyevj_bufferSize(handle, jobz, uplo, n, A, lda, W, lwork, params); } /** @} */ /** * @defgroup syevd cusolver syevd operations * @{ */ template <typename T> cusolverStatus_t cusolverDnsyevd(cusolverDnHandle_t handle, // NOLINT cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, T* A, int lda, T* W, T* work, int lwork, int* devInfo, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDnsyevd(cusolverDnHandle_t handle, // NOLINT cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, float* A, int lda, float* W, float* work, int lwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnSsyevd(handle, jobz, uplo, n, A, lda, W, work, lwork, devInfo); } template <> inline cusolverStatus_t cusolverDnsyevd(cusolverDnHandle_t handle, // NOLINT cusolverEigMode_t jobz, cublasFillMode_t uplo, int n, double* A, int lda, double* W, double* work, int lwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnDsyevd(handle, jobz, uplo, n, A, lda, W, work, lwork, devInfo); } /** @} */ /** * @defgroup syevdx cusolver syevdx operations * @{ */ template <typename T> cusolverStatus_t cusolverDnsyevdx_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cusolverEigRange_t range, cublasFillMode_t uplo, int n, const T* A, int lda, T vl, T vu, int il, int iu, int* h_meig, const T* W, int* lwork); template <> inline cusolverStatus_t cusolverDnsyevdx_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cusolverEigRange_t range, cublasFillMode_t uplo, int n, const float* A, int lda, float vl, float vu, int il, int iu, int* h_meig, const float* W, int* lwork) { return cusolverDnSsyevdx_bufferSize( handle, jobz, range, uplo, n, A, lda, vl, vu, il, iu, h_meig, W, lwork); } template <> inline cusolverStatus_t cusolverDnsyevdx_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cusolverEigRange_t range, cublasFillMode_t uplo, int n, const double* A, int lda, double vl, double vu, int il, int iu, int* h_meig, const double* W, int* lwork) { return cusolverDnDsyevdx_bufferSize( handle, jobz, range, uplo, n, A, lda, vl, vu, il, iu, h_meig, W, lwork); } template <typename T> cusolverStatus_t cusolverDnsyevdx( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cusolverEigRange_t range, cublasFillMode_t uplo, int n, T* A, int lda, T vl, T vu, int il, int iu, int* h_meig, T* W, T* work, int lwork, int* devInfo, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDnsyevdx( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cusolverEigRange_t range, cublasFillMode_t uplo, int n, float* A, int lda, float vl, float vu, int il, int iu, int* h_meig, float* W, float* work, int lwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnSsyevdx( handle, jobz, range, uplo, n, A, lda, vl, vu, il, iu, h_meig, W, work, lwork, devInfo); } template <> inline cusolverStatus_t cusolverDnsyevdx( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, cusolverEigRange_t range, cublasFillMode_t uplo, int n, double* A, int lda, double vl, double vu, int il, int iu, int* h_meig, double* W, double* work, int lwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnDsyevdx( handle, jobz, range, uplo, n, A, lda, vl, vu, il, iu, h_meig, W, work, lwork, devInfo); } /** @} */ /** * @defgroup svd cusolver svd operations * @{ */ template <typename T> cusolverStatus_t cusolverDngesvd_bufferSize( // NOLINT cusolverDnHandle_t handle, int m, int n, int* lwork) { if (std::is_same<std::decay_t<T>, float>::value) { return cusolverDnSgesvd_bufferSize(handle, m, n, lwork); } else { return cusolverDnDgesvd_bufferSize(handle, m, n, lwork); } } template <typename T> cusolverStatus_t cusolverDngesvd( // NOLINT cusolverDnHandle_t handle, signed char jobu, signed char jobvt, int m, int n, T* A, int lda, T* S, T* U, int ldu, T* VT, int ldvt, T* work, int lwork, T* rwork, int* devInfo, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDngesvd( // NOLINT cusolverDnHandle_t handle, signed char jobu, signed char jobvt, int m, int n, float* A, int lda, float* S, float* U, int ldu, float* VT, int ldvt, float* work, int lwork, float* rwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnSgesvd( handle, jobu, jobvt, m, n, A, lda, S, U, ldu, VT, ldvt, work, lwork, rwork, devInfo); } template <> inline cusolverStatus_t cusolverDngesvd( // NOLINT cusolverDnHandle_t handle, signed char jobu, signed char jobvt, int m, int n, double* A, int lda, double* S, double* U, int ldu, double* VT, int ldvt, double* work, int lwork, double* rwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnDgesvd( handle, jobu, jobvt, m, n, A, lda, S, U, ldu, VT, ldvt, work, lwork, rwork, devInfo); } template <typename T> inline cusolverStatus_t CUSOLVERAPI cusolverDngesvdj_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, int econ, int m, int n, const T* A, int lda, const T* S, const T* U, int ldu, const T* V, int ldv, int* lwork, gesvdjInfo_t params); template <> inline cusolverStatus_t CUSOLVERAPI cusolverDngesvdj_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, int econ, int m, int n, const float* A, int lda, const float* S, const float* U, int ldu, const float* V, int ldv, int* lwork, gesvdjInfo_t params) { return cusolverDnSgesvdj_bufferSize( handle, jobz, econ, m, n, A, lda, S, U, ldu, V, ldv, lwork, params); } template <> inline cusolverStatus_t CUSOLVERAPI cusolverDngesvdj_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, int econ, int m, int n, const double* A, int lda, const double* S, const double* U, int ldu, const double* V, int ldv, int* lwork, gesvdjInfo_t params) { return cusolverDnDgesvdj_bufferSize( handle, jobz, econ, m, n, A, lda, S, U, ldu, V, ldv, lwork, params); } template <typename T> inline cusolverStatus_t CUSOLVERAPI cusolverDngesvdj( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, int econ, int m, int n, T* A, int lda, T* S, T* U, int ldu, T* V, int ldv, T* work, int lwork, int* info, gesvdjInfo_t params, cudaStream_t stream); template <> inline cusolverStatus_t CUSOLVERAPI cusolverDngesvdj( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, int econ, int m, int n, float* A, int lda, float* S, float* U, int ldu, float* V, int ldv, float* work, int lwork, int* info, gesvdjInfo_t params, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnSgesvdj( handle, jobz, econ, m, n, A, lda, S, U, ldu, V, ldv, work, lwork, info, params); } template <> inline cusolverStatus_t CUSOLVERAPI cusolverDngesvdj( // NOLINT cusolverDnHandle_t handle, cusolverEigMode_t jobz, int econ, int m, int n, double* A, int lda, double* S, double* U, int ldu, double* V, int ldv, double* work, int lwork, int* info, gesvdjInfo_t params, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnDgesvdj( handle, jobz, econ, m, n, A, lda, S, U, ldu, V, ldv, work, lwork, info, params); } #if CUDART_VERSION >= 11010 template <typename T> cusolverStatus_t cusolverDnxgesvdr_bufferSize( // NOLINT cusolverDnHandle_t handle, signed char jobu, signed char jobv, int64_t m, int64_t n, int64_t k, int64_t p, int64_t niters, const T* a, int64_t lda, const T* Srand, const T* Urand, int64_t ldUrand, const T* Vrand, int64_t ldVrand, size_t* workspaceInBytesOnDevice, size_t* workspaceInBytesOnHost, cudaStream_t stream) { RAFT_EXPECTS(std::is_floating_point_v<T>, "Unsupported data type"); cudaDataType dataType = std::is_same_v<T, float> ? CUDA_R_32F : CUDA_R_64F; RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); cusolverDnParams_t dn_params = nullptr; RAFT_CUSOLVER_TRY(cusolverDnCreateParams(&dn_params)); auto result = cusolverDnXgesvdr_bufferSize(handle, dn_params, jobu, jobv, m, n, k, p, niters, dataType, a, lda, dataType, Srand, dataType, Urand, ldUrand, dataType, Vrand, ldVrand, dataType, workspaceInBytesOnDevice, workspaceInBytesOnHost); RAFT_CUSOLVER_TRY(cusolverDnDestroyParams(dn_params)); return result; } template <typename T> cusolverStatus_t cusolverDnxgesvdr( // NOLINT cusolverDnHandle_t handle, signed char jobu, signed char jobv, int64_t m, int64_t n, int64_t k, int64_t p, int64_t niters, T* a, int64_t lda, T* Srand, T* Urand, int64_t ldUrand, T* Vrand, int64_t ldVrand, void* bufferOnDevice, size_t workspaceInBytesOnDevice, void* bufferOnHost, size_t workspaceInBytesOnHost, int* d_info, cudaStream_t stream) { cudaDataType dataType = std::is_same_v<T, float> ? CUDA_R_32F : CUDA_R_64F; RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); cusolverDnParams_t dn_params = nullptr; RAFT_CUSOLVER_TRY(cusolverDnCreateParams(&dn_params)); auto result = cusolverDnXgesvdr(handle, dn_params, jobu, jobv, m, n, k, p, niters, dataType, a, lda, dataType, Srand, dataType, Urand, ldUrand, dataType, Vrand, ldVrand, dataType, bufferOnDevice, workspaceInBytesOnDevice, bufferOnHost, workspaceInBytesOnHost, d_info); RAFT_CUSOLVER_TRY(cusolverDnDestroyParams(dn_params)); return result; } #endif // CUDART_VERSION >= 11010 /** @} */ /** * @defgroup potrf cusolver potrf operations * @{ */ template <typename T> cusolverStatus_t cusolverDnpotrf_bufferSize( // NOLINT cusolverDnHandle_t handle, cublasFillMode_t uplo, int n, T* A, int lda, int* Lwork); template <> inline cusolverStatus_t cusolverDnpotrf_bufferSize( // NOLINT cusolverDnHandle_t handle, cublasFillMode_t uplo, int n, float* A, int lda, int* Lwork) { return cusolverDnSpotrf_bufferSize(handle, uplo, n, A, lda, Lwork); } template <> inline cusolverStatus_t cusolverDnpotrf_bufferSize( // NOLINT cusolverDnHandle_t handle, cublasFillMode_t uplo, int n, double* A, int lda, int* Lwork) { return cusolverDnDpotrf_bufferSize(handle, uplo, n, A, lda, Lwork); } template <typename T> inline cusolverStatus_t cusolverDnpotrf(cusolverDnHandle_t handle, // NOLINT cublasFillMode_t uplo, int n, T* A, int lda, T* Workspace, int Lwork, int* devInfo, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDnpotrf(cusolverDnHandle_t handle, // NOLINT cublasFillMode_t uplo, int n, float* A, int lda, float* Workspace, int Lwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnSpotrf(handle, uplo, n, A, lda, Workspace, Lwork, devInfo); } template <> inline cusolverStatus_t cusolverDnpotrf(cusolverDnHandle_t handle, // NOLINT cublasFillMode_t uplo, int n, double* A, int lda, double* Workspace, int Lwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnDpotrf(handle, uplo, n, A, lda, Workspace, Lwork, devInfo); } /** @} */ /** * @defgroup potrs cusolver potrs operations * @{ */ template <typename T> cusolverStatus_t cusolverDnpotrs(cusolverDnHandle_t handle, // NOLINT cublasFillMode_t uplo, int n, int nrhs, const T* A, int lda, T* B, int ldb, int* devInfo, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDnpotrs(cusolverDnHandle_t handle, // NOLINT cublasFillMode_t uplo, int n, int nrhs, const float* A, int lda, float* B, int ldb, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnSpotrs(handle, uplo, n, nrhs, A, lda, B, ldb, devInfo); } template <> inline cusolverStatus_t cusolverDnpotrs(cusolverDnHandle_t handle, // NOLINT cublasFillMode_t uplo, int n, int nrhs, const double* A, int lda, double* B, int ldb, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnDpotrs(handle, uplo, n, nrhs, A, lda, B, ldb, devInfo); } /** @} */ /** * @defgroup geqrf cusolver geqrf operations * @{ */ template <typename T> cusolverStatus_t cusolverDngeqrf(cusolverDnHandle_t handle, int m, // NOLINT int n, T* A, int lda, T* TAU, T* Workspace, int Lwork, int* devInfo, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDngeqrf(cusolverDnHandle_t handle, // NOLINT int m, int n, float* A, int lda, float* TAU, float* Workspace, int Lwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnSgeqrf(handle, m, n, A, lda, TAU, Workspace, Lwork, devInfo); } template <> inline cusolverStatus_t cusolverDngeqrf(cusolverDnHandle_t handle, // NOLINT int m, int n, double* A, int lda, double* TAU, double* Workspace, int Lwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnDgeqrf(handle, m, n, A, lda, TAU, Workspace, Lwork, devInfo); } template <typename T> cusolverStatus_t cusolverDngeqrf_bufferSize( // NOLINT cusolverDnHandle_t handle, int m, int n, T* A, int lda, int* Lwork); template <> inline cusolverStatus_t cusolverDngeqrf_bufferSize( // NOLINT cusolverDnHandle_t handle, int m, int n, float* A, int lda, int* Lwork) { return cusolverDnSgeqrf_bufferSize(handle, m, n, A, lda, Lwork); } template <> inline cusolverStatus_t cusolverDngeqrf_bufferSize( // NOLINT cusolverDnHandle_t handle, int m, int n, double* A, int lda, int* Lwork) { return cusolverDnDgeqrf_bufferSize(handle, m, n, A, lda, Lwork); } /** @} */ /** * @defgroup orgqr cusolver orgqr operations * @{ */ template <typename T> cusolverStatus_t cusolverDnorgqr( // NOLINT cusolverDnHandle_t handle, int m, int n, int k, T* A, int lda, const T* tau, T* work, int lwork, int* devInfo, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDnorgqr( // NOLINT cusolverDnHandle_t handle, int m, int n, int k, float* A, int lda, const float* tau, float* work, int lwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnSorgqr(handle, m, n, k, A, lda, tau, work, lwork, devInfo); } template <> inline cusolverStatus_t cusolverDnorgqr( // NOLINT cusolverDnHandle_t handle, int m, int n, int k, double* A, int lda, const double* tau, double* work, int lwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnDorgqr(handle, m, n, k, A, lda, tau, work, lwork, devInfo); } template <typename T> cusolverStatus_t cusolverDnorgqr_bufferSize( // NOLINT cusolverDnHandle_t handle, int m, int n, int k, const T* A, int lda, const T* TAU, int* lwork); template <> inline cusolverStatus_t cusolverDnorgqr_bufferSize( // NOLINT cusolverDnHandle_t handle, int m, int n, int k, const float* A, int lda, const float* TAU, int* lwork) { return cusolverDnSorgqr_bufferSize(handle, m, n, k, A, lda, TAU, lwork); } template <> inline cusolverStatus_t cusolverDnorgqr_bufferSize( // NOLINT cusolverDnHandle_t handle, int m, int n, int k, const double* A, int lda, const double* TAU, int* lwork) { return cusolverDnDorgqr_bufferSize(handle, m, n, k, A, lda, TAU, lwork); } /** @} */ /** * @defgroup ormqr cusolver ormqr operations * @{ */ template <typename T> cusolverStatus_t cusolverDnormqr(cusolverDnHandle_t handle, // NOLINT cublasSideMode_t side, cublasOperation_t trans, int m, int n, int k, const T* A, int lda, const T* tau, T* C, int ldc, T* work, int lwork, int* devInfo, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDnormqr( // NOLINT cusolverDnHandle_t handle, cublasSideMode_t side, cublasOperation_t trans, int m, int n, int k, const float* A, int lda, const float* tau, float* C, int ldc, float* work, int lwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnSormqr(handle, side, trans, m, n, k, A, lda, tau, C, ldc, work, lwork, devInfo); } template <> inline cusolverStatus_t cusolverDnormqr( // NOLINT cusolverDnHandle_t handle, cublasSideMode_t side, cublasOperation_t trans, int m, int n, int k, const double* A, int lda, const double* tau, double* C, int ldc, double* work, int lwork, int* devInfo, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnDormqr(handle, side, trans, m, n, k, A, lda, tau, C, ldc, work, lwork, devInfo); } template <typename T> cusolverStatus_t cusolverDnormqr_bufferSize( // NOLINT cusolverDnHandle_t handle, cublasSideMode_t side, cublasOperation_t trans, int m, int n, int k, const T* A, int lda, const T* tau, const T* C, int ldc, int* lwork); template <> inline cusolverStatus_t cusolverDnormqr_bufferSize( // NOLINT cusolverDnHandle_t handle, cublasSideMode_t side, cublasOperation_t trans, int m, int n, int k, const float* A, int lda, const float* tau, const float* C, int ldc, int* lwork) { return cusolverDnSormqr_bufferSize(handle, side, trans, m, n, k, A, lda, tau, C, ldc, lwork); } template <> inline cusolverStatus_t cusolverDnormqr_bufferSize( // NOLINT cusolverDnHandle_t handle, cublasSideMode_t side, cublasOperation_t trans, int m, int n, int k, const double* A, int lda, const double* tau, const double* C, int ldc, int* lwork) { return cusolverDnDormqr_bufferSize(handle, side, trans, m, n, k, A, lda, tau, C, ldc, lwork); } /** @} */ /** * @defgroup csrqrBatched cusolver batched * @{ */ template <typename T> cusolverStatus_t cusolverSpcsrqrBufferInfoBatched( // NOLINT cusolverSpHandle_t handle, int m, int n, int nnzA, const cusparseMatDescr_t descrA, const T* csrValA, const int* csrRowPtrA, const int* csrColIndA, int batchSize, csrqrInfo_t info, size_t* internalDataInBytes, size_t* workspaceInBytes); template <> inline cusolverStatus_t cusolverSpcsrqrBufferInfoBatched( // NOLINT cusolverSpHandle_t handle, int m, int n, int nnzA, const cusparseMatDescr_t descrA, const float* csrValA, const int* csrRowPtrA, const int* csrColIndA, int batchSize, csrqrInfo_t info, size_t* internalDataInBytes, size_t* workspaceInBytes) { return cusolverSpScsrqrBufferInfoBatched(handle, m, n, nnzA, descrA, csrValA, csrRowPtrA, csrColIndA, batchSize, info, internalDataInBytes, workspaceInBytes); } template <> inline cusolverStatus_t cusolverSpcsrqrBufferInfoBatched( // NOLINT cusolverSpHandle_t handle, int m, int n, int nnzA, const cusparseMatDescr_t descrA, const double* csrValA, const int* csrRowPtrA, const int* csrColIndA, int batchSize, csrqrInfo_t info, size_t* internalDataInBytes, size_t* workspaceInBytes) { return cusolverSpDcsrqrBufferInfoBatched(handle, m, n, nnzA, descrA, csrValA, csrRowPtrA, csrColIndA, batchSize, info, internalDataInBytes, workspaceInBytes); } template <typename T> cusolverStatus_t cusolverSpcsrqrsvBatched( // NOLINT cusolverSpHandle_t handle, int m, int n, int nnzA, const cusparseMatDescr_t descrA, const T* csrValA, const int* csrRowPtrA, const int* csrColIndA, const T* b, T* x, int batchSize, csrqrInfo_t info, void* pBuffer, cudaStream_t stream); template <> inline cusolverStatus_t cusolverSpcsrqrsvBatched( // NOLINT cusolverSpHandle_t handle, int m, int n, int nnzA, const cusparseMatDescr_t descrA, const float* csrValA, const int* csrRowPtrA, const int* csrColIndA, const float* b, float* x, int batchSize, csrqrInfo_t info, void* pBuffer, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverSpSetStream(handle, stream)); return cusolverSpScsrqrsvBatched( handle, m, n, nnzA, descrA, csrValA, csrRowPtrA, csrColIndA, b, x, batchSize, info, pBuffer); } template <> inline cusolverStatus_t cusolverSpcsrqrsvBatched( // NOLINT cusolverSpHandle_t handle, int m, int n, int nnzA, const cusparseMatDescr_t descrA, const double* csrValA, const int* csrRowPtrA, const int* csrColIndA, const double* b, double* x, int batchSize, csrqrInfo_t info, void* pBuffer, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverSpSetStream(handle, stream)); return cusolverSpDcsrqrsvBatched( handle, m, n, nnzA, descrA, csrValA, csrRowPtrA, csrColIndA, b, x, batchSize, info, pBuffer); } /** @} */ #if CUDART_VERSION >= 11010 /** * @defgroup DnXsyevd cusolver DnXsyevd operations * @{ */ template <typename T> cusolverStatus_t cusolverDnxsyevd_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverDnParams_t params, cusolverEigMode_t jobz, cublasFillMode_t uplo, int64_t n, const T* A, int64_t lda, const T* W, size_t* workspaceInBytesOnDevice, size_t* workspaceInBytesOnHost, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDnxsyevd_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverDnParams_t params, cusolverEigMode_t jobz, cublasFillMode_t uplo, int64_t n, const float* A, int64_t lda, const float* W, size_t* workspaceInBytesOnDevice, size_t* workspaceInBytesOnHost, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnXsyevd_bufferSize(handle, params, jobz, uplo, n, CUDA_R_32F, A, lda, CUDA_R_32F, W, CUDA_R_32F, workspaceInBytesOnDevice, workspaceInBytesOnHost); } template <> inline cusolverStatus_t cusolverDnxsyevd_bufferSize( // NOLINT cusolverDnHandle_t handle, cusolverDnParams_t params, cusolverEigMode_t jobz, cublasFillMode_t uplo, int64_t n, const double* A, int64_t lda, const double* W, size_t* workspaceInBytesOnDevice, size_t* workspaceInBytesOnHost, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnXsyevd_bufferSize(handle, params, jobz, uplo, n, CUDA_R_64F, A, lda, CUDA_R_64F, W, CUDA_R_64F, workspaceInBytesOnDevice, workspaceInBytesOnHost); } template <typename T> cusolverStatus_t cusolverDnxsyevd( // NOLINT cusolverDnHandle_t handle, cusolverDnParams_t params, cusolverEigMode_t jobz, cublasFillMode_t uplo, int64_t n, T* A, int64_t lda, T* W, T* bufferOnDevice, size_t workspaceInBytesOnDevice, T* bufferOnHost, size_t workspaceInBytesOnHost, int* info, cudaStream_t stream); template <> inline cusolverStatus_t cusolverDnxsyevd( // NOLINT cusolverDnHandle_t handle, cusolverDnParams_t params, cusolverEigMode_t jobz, cublasFillMode_t uplo, int64_t n, float* A, int64_t lda, float* W, float* bufferOnDevice, size_t workspaceInBytesOnDevice, float* bufferOnHost, size_t workspaceInBytesOnHost, int* info, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnXsyevd(handle, params, jobz, uplo, n, CUDA_R_32F, A, lda, CUDA_R_32F, W, CUDA_R_32F, bufferOnDevice, workspaceInBytesOnDevice, bufferOnHost, workspaceInBytesOnHost, info); } template <> inline cusolverStatus_t cusolverDnxsyevd( // NOLINT cusolverDnHandle_t handle, cusolverDnParams_t params, cusolverEigMode_t jobz, cublasFillMode_t uplo, int64_t n, double* A, int64_t lda, double* W, double* bufferOnDevice, size_t workspaceInBytesOnDevice, double* bufferOnHost, size_t workspaceInBytesOnHost, int* info, cudaStream_t stream) { RAFT_CUSOLVER_TRY(cusolverDnSetStream(handle, stream)); return cusolverDnXsyevd(handle, params, jobz, uplo, n, CUDA_R_64F, A, lda, CUDA_R_64F, W, CUDA_R_64F, bufferOnDevice, workspaceInBytesOnDevice, bufferOnHost, workspaceInBytesOnHost, info); } /** @} */ #endif } // namespace detail } // namespace linalg } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/multiply.cuh
/* * Copyright (c) 2022, 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> #include <raft/linalg/unary_op.cuh> namespace raft { namespace linalg { namespace detail { template <typename math_t, typename IdxType = int> void multiplyScalar( math_t* out, const math_t* in, const math_t scalar, IdxType len, cudaStream_t stream) { raft::linalg::unaryOp(out, in, len, raft::mul_const_op<math_t>{scalar}, stream); } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/cublas_wrappers.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 <cublas_v2.h> #include <raft/core/cublas_macros.hpp> #include <raft/core/error.hpp> #include <cstdint> #include <cublas_v2.h> namespace raft { namespace linalg { namespace detail { /** * Assuming the default CUBLAS_POINTER_MODE_HOST, change it to host or device mode * temporary for the lifetime of this object. */ template <bool DevicePointerMode = false> class cublas_device_pointer_mode { public: explicit cublas_device_pointer_mode(cublasHandle_t handle) : handle_(handle) { if constexpr (DevicePointerMode) { RAFT_CUBLAS_TRY(cublasSetPointerMode(handle_, CUBLAS_POINTER_MODE_DEVICE)); } } auto operator=(const cublas_device_pointer_mode&) -> cublas_device_pointer_mode& = delete; auto operator=(cublas_device_pointer_mode&&) -> cublas_device_pointer_mode& = delete; static auto operator new(std::size_t) -> void* = delete; static auto operator new[](std::size_t) -> void* = delete; ~cublas_device_pointer_mode() { if constexpr (DevicePointerMode) { RAFT_CUBLAS_TRY_NO_THROW(cublasSetPointerMode(handle_, CUBLAS_POINTER_MODE_HOST)); } } private: cublasHandle_t handle_ = nullptr; }; /** * @defgroup Axpy cublas ax+y operations * @{ */ template <typename T> cublasStatus_t cublasaxpy(cublasHandle_t handle, int n, const T* alpha, const T* x, int incx, T* y, int incy, cudaStream_t stream); template <> inline cublasStatus_t cublasaxpy(cublasHandle_t handle, int n, const float* alpha, const float* x, int incx, float* y, int incy, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSaxpy(handle, n, alpha, x, incx, y, incy); } template <> inline cublasStatus_t cublasaxpy(cublasHandle_t handle, int n, const double* alpha, const double* x, int incx, double* y, int incy, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDaxpy(handle, n, alpha, x, incx, y, incy); } /** @} */ /** * @defgroup cublas swap operations * @{ */ template <typename T> cublasStatus_t cublasSwap( cublasHandle_t handle, int n, T* x, int incx, T* y, int incy, cudaStream_t stream); template <> inline cublasStatus_t cublasSwap( cublasHandle_t handle, int n, float* x, int incx, float* y, int incy, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSswap(handle, n, x, incx, y, incy); } template <> inline cublasStatus_t cublasSwap( cublasHandle_t handle, int n, double* x, int incx, double* y, int incy, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDswap(handle, n, x, incx, y, incy); } /** @} */ /** * @defgroup cublas copy operations * @{ */ template <typename T> cublasStatus_t cublasCopy( cublasHandle_t handle, int n, const T* x, int incx, T* y, int incy, cudaStream_t stream); template <> inline cublasStatus_t cublasCopy( cublasHandle_t handle, int n, const float* x, int incx, float* y, int incy, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasScopy(handle, n, x, incx, y, incy); } template <> inline cublasStatus_t cublasCopy( cublasHandle_t handle, int n, const double* x, int incx, double* y, int incy, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDcopy(handle, n, x, incx, y, incy); } /** @} */ /** * @defgroup gemv cublas gemv calls * @{ */ template <typename T> cublasStatus_t cublasgemv(cublasHandle_t handle, cublasOperation_t transA, int m, int n, const T* alfa, const T* A, int lda, const T* x, int incx, const T* beta, T* y, int incy, cudaStream_t stream); template <> inline cublasStatus_t cublasgemv(cublasHandle_t handle, cublasOperation_t transA, int m, int n, const float* alfa, const float* A, int lda, const float* x, int incx, const float* beta, float* y, int incy, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSgemv(handle, transA, m, n, alfa, A, lda, x, incx, beta, y, incy); } template <> inline cublasStatus_t cublasgemv(cublasHandle_t handle, cublasOperation_t transA, int m, int n, const double* alfa, const double* A, int lda, const double* x, int incx, const double* beta, double* y, int incy, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDgemv(handle, transA, m, n, alfa, A, lda, x, incx, beta, y, incy); } /** @} */ /** * @defgroup ger cublas a(x*y.T) + A calls * @{ */ template <typename T> cublasStatus_t cublasger(cublasHandle_t handle, int m, int n, const T* alpha, const T* x, int incx, const T* y, int incy, T* A, int lda, cudaStream_t stream); template <> inline cublasStatus_t cublasger(cublasHandle_t handle, int m, int n, const float* alpha, const float* x, int incx, const float* y, int incy, float* A, int lda, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSger(handle, m, n, alpha, x, incx, y, incy, A, lda); } template <> inline cublasStatus_t cublasger(cublasHandle_t handle, int m, int n, const double* alpha, const double* x, int incx, const double* y, int incy, double* A, int lda, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDger(handle, m, n, alpha, x, incx, y, incy, A, lda); } /** @} */ /** * @defgroup gemm cublas gemm calls * @{ */ template <typename T> cublasStatus_t cublasgemm(cublasHandle_t handle, cublasOperation_t transA, cublasOperation_t transB, int m, int n, int k, const T* alfa, const T* A, int lda, const T* B, int ldb, const T* beta, T* C, int ldc, cudaStream_t stream); template <> inline cublasStatus_t cublasgemm(cublasHandle_t handle, cublasOperation_t transA, cublasOperation_t transB, int m, int n, int k, const float* alfa, const float* A, int lda, const float* B, int ldb, const float* beta, float* C, int ldc, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSgemm(handle, transA, transB, m, n, k, alfa, A, lda, B, ldb, beta, C, ldc); } template <> inline cublasStatus_t cublasgemm(cublasHandle_t handle, cublasOperation_t transA, cublasOperation_t transB, int m, int n, int k, const double* alfa, const double* A, int lda, const double* B, int ldb, const double* beta, double* C, int ldc, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDgemm(handle, transA, transB, m, n, k, alfa, A, lda, B, ldb, beta, C, ldc); } /** @} */ /** * @defgroup gemmbatched cublas gemmbatched calls * @{ */ template <typename T> cublasStatus_t cublasgemmBatched(cublasHandle_t handle, // NOLINT cublasOperation_t transa, cublasOperation_t transb, int m, int n, int k, const T* alpha, const T* const Aarray[], // NOLINT int lda, const T* const Barray[], // NOLINT int ldb, const T* beta, T* Carray[], // NOLINT int ldc, int batchCount, cudaStream_t stream); template <> inline cublasStatus_t cublasgemmBatched( // NOLINT cublasHandle_t handle, cublasOperation_t transa, cublasOperation_t transb, int m, int n, int k, const float* alpha, const float* const Aarray[], // NOLINT int lda, const float* const Barray[], // NOLINT int ldb, const float* beta, float* Carray[], // NOLINT int ldc, int batchCount, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSgemmBatched(handle, transa, transb, m, n, k, alpha, Aarray, lda, Barray, ldb, beta, Carray, ldc, batchCount); } template <> inline cublasStatus_t cublasgemmBatched( // NOLINT cublasHandle_t handle, cublasOperation_t transa, cublasOperation_t transb, int m, int n, int k, const double* alpha, const double* const Aarray[], // NOLINT int lda, const double* const Barray[], // NOLINT int ldb, const double* beta, double* Carray[], // NOLINT int ldc, int batchCount, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDgemmBatched(handle, transa, transb, m, n, k, alpha, Aarray, lda, Barray, ldb, beta, Carray, ldc, batchCount); } /** @} */ /** * @defgroup gemmbatched cublas gemmbatched calls * @{ */ template <typename T> cublasStatus_t cublasgemmStridedBatched( // NOLINT cublasHandle_t handle, cublasOperation_t transa, cublasOperation_t transb, int m, int n, int k, const T* alpha, const T* const Aarray, int lda, int64_t strideA, const T* const Barray, int ldb, int64_t strideB, const T* beta, T* Carray, int ldc, int64_t strideC, int batchCount, cudaStream_t stream); template <> inline cublasStatus_t cublasgemmStridedBatched( // NOLINT cublasHandle_t handle, cublasOperation_t transa, cublasOperation_t transb, int m, int n, int k, const float* alpha, const float* const Aarray, int lda, int64_t strideA, const float* const Barray, int ldb, int64_t strideB, const float* beta, float* Carray, int ldc, int64_t strideC, int batchCount, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSgemmStridedBatched(handle, transa, transb, m, n, k, alpha, Aarray, lda, strideA, Barray, ldb, strideB, beta, Carray, ldc, strideC, batchCount); } template <> inline cublasStatus_t cublasgemmStridedBatched( // NOLINT cublasHandle_t handle, cublasOperation_t transa, cublasOperation_t transb, int m, int n, int k, const double* alpha, const double* const Aarray, int lda, int64_t strideA, const double* const Barray, int ldb, int64_t strideB, const double* beta, double* Carray, int ldc, int64_t strideC, int batchCount, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDgemmStridedBatched(handle, transa, transb, m, n, k, alpha, Aarray, lda, strideA, Barray, ldb, strideB, beta, Carray, ldc, strideC, batchCount); } /** @} */ /** * @defgroup solverbatched cublas getrf/gettribatched calls * @{ */ template <typename T> cublasStatus_t cublasgetrfBatched(cublasHandle_t handle, int n, // NOLINT T* const A[], // NOLINT int lda, int* P, int* info, int batchSize, cudaStream_t stream); template <> inline cublasStatus_t cublasgetrfBatched(cublasHandle_t handle, // NOLINT int n, float* const A[], // NOLINT int lda, int* P, int* info, int batchSize, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSgetrfBatched(handle, n, A, lda, P, info, batchSize); } template <> inline cublasStatus_t cublasgetrfBatched(cublasHandle_t handle, // NOLINT int n, double* const A[], // NOLINT int lda, int* P, int* info, int batchSize, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDgetrfBatched(handle, n, A, lda, P, info, batchSize); } template <typename T> cublasStatus_t cublasgetriBatched(cublasHandle_t handle, int n, // NOLINT const T* const A[], // NOLINT int lda, const int* P, T* const C[], // NOLINT int ldc, int* info, int batchSize, cudaStream_t stream); template <> inline cublasStatus_t cublasgetriBatched( // NOLINT cublasHandle_t handle, int n, const float* const A[], // NOLINT int lda, const int* P, float* const C[], // NOLINT int ldc, int* info, int batchSize, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSgetriBatched(handle, n, A, lda, P, C, ldc, info, batchSize); } template <> inline cublasStatus_t cublasgetriBatched( // NOLINT cublasHandle_t handle, int n, const double* const A[], // NOLINT int lda, const int* P, double* const C[], // NOLINT int ldc, int* info, int batchSize, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDgetriBatched(handle, n, A, lda, P, C, ldc, info, batchSize); } /** @} */ /** * @defgroup gelsbatched cublas gelsbatched calls * @{ */ template <typename T> inline cublasStatus_t cublasgelsBatched(cublasHandle_t handle, // NOLINT cublasOperation_t trans, int m, int n, int nrhs, T* Aarray[], // NOLINT int lda, T* Carray[], // NOLINT int ldc, int* info, int* devInfoArray, int batchSize, cudaStream_t stream); template <> inline cublasStatus_t cublasgelsBatched(cublasHandle_t handle, // NOLINT cublasOperation_t trans, int m, int n, int nrhs, float* Aarray[], // NOLINT int lda, float* Carray[], // NOLINT int ldc, int* info, int* devInfoArray, int batchSize, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSgelsBatched( handle, trans, m, n, nrhs, Aarray, lda, Carray, ldc, info, devInfoArray, batchSize); } template <> inline cublasStatus_t cublasgelsBatched(cublasHandle_t handle, // NOLINT cublasOperation_t trans, int m, int n, int nrhs, double* Aarray[], // NOLINT int lda, double* Carray[], // NOLINT int ldc, int* info, int* devInfoArray, int batchSize, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDgelsBatched( handle, trans, m, n, nrhs, Aarray, lda, Carray, ldc, info, devInfoArray, batchSize); } /** @} */ /** * @defgroup geam cublas geam calls * @{ */ template <typename T> cublasStatus_t cublasgeam(cublasHandle_t handle, cublasOperation_t transA, cublasOperation_t transB, int m, int n, const T* alfa, const T* A, int lda, const T* beta, const T* B, int ldb, T* C, int ldc, cudaStream_t stream); template <> inline cublasStatus_t cublasgeam(cublasHandle_t handle, cublasOperation_t transA, cublasOperation_t transB, int m, int n, const float* alfa, const float* A, int lda, const float* beta, const float* B, int ldb, float* C, int ldc, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSgeam(handle, transA, transB, m, n, alfa, A, lda, beta, B, ldb, C, ldc); } template <> inline cublasStatus_t cublasgeam(cublasHandle_t handle, cublasOperation_t transA, cublasOperation_t transB, int m, int n, const double* alfa, const double* A, int lda, const double* beta, const double* B, int ldb, double* C, int ldc, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDgeam(handle, transA, transB, m, n, alfa, A, lda, beta, B, ldb, C, ldc); } /** @} */ /** * @defgroup symm cublas symm calls * @{ */ template <typename T> cublasStatus_t cublassymm(cublasHandle_t handle, cublasSideMode_t side, cublasFillMode_t uplo, int m, int n, const T* alpha, const T* A, int lda, const T* B, int ldb, const T* beta, T* C, int ldc, cudaStream_t stream); template <> inline cublasStatus_t cublassymm(cublasHandle_t handle, cublasSideMode_t side, cublasFillMode_t uplo, int m, int n, const float* alpha, const float* A, int lda, const float* B, int ldb, const float* beta, float* C, int ldc, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSsymm(handle, side, uplo, m, n, alpha, A, lda, B, ldb, beta, C, ldc); } template <> inline cublasStatus_t cublassymm(cublasHandle_t handle, cublasSideMode_t side, cublasFillMode_t uplo, int m, int n, const double* alpha, const double* A, int lda, const double* B, int ldb, const double* beta, double* C, int ldc, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDsymm(handle, side, uplo, m, n, alpha, A, lda, B, ldb, beta, C, ldc); } /** @} */ /** * @defgroup syrk cublas syrk calls * @{ */ template <typename T> cublasStatus_t cublassyrk(cublasHandle_t handle, cublasFillMode_t uplo, cublasOperation_t trans, int n, int k, const T* alpha, const T* A, int lda, const T* beta, T* C, int ldc, cudaStream_t stream); template <> inline cublasStatus_t cublassyrk(cublasHandle_t handle, cublasFillMode_t uplo, cublasOperation_t trans, int n, int k, const float* alpha, const float* A, int lda, const float* beta, float* C, int ldc, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSsyrk(handle, uplo, trans, n, k, alpha, A, lda, beta, C, ldc); } template <> inline cublasStatus_t cublassyrk(cublasHandle_t handle, cublasFillMode_t uplo, cublasOperation_t trans, int n, int k, const double* alpha, const double* A, int lda, const double* beta, double* C, int ldc, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDsyrk(handle, uplo, trans, n, k, alpha, A, lda, beta, C, ldc); } /** @} */ /** * @defgroup nrm2 cublas nrm2 calls * @{ */ template <typename T> cublasStatus_t cublasnrm2( cublasHandle_t handle, int n, const T* x, int incx, T* result, cudaStream_t stream); template <> inline cublasStatus_t cublasnrm2( cublasHandle_t handle, int n, const float* x, int incx, float* result, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSnrm2(handle, n, x, incx, result); } template <> inline cublasStatus_t cublasnrm2( cublasHandle_t handle, int n, const double* x, int incx, double* result, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDnrm2(handle, n, x, incx, result); } /** @} */ template <typename T> cublasStatus_t cublastrsm(cublasHandle_t handle, cublasSideMode_t side, cublasFillMode_t uplo, cublasOperation_t trans, cublasDiagType_t diag, int m, int n, const T* alpha, const T* A, int lda, T* B, int ldb, cudaStream_t stream); template <> inline cublasStatus_t cublastrsm(cublasHandle_t handle, cublasSideMode_t side, cublasFillMode_t uplo, cublasOperation_t trans, cublasDiagType_t diag, int m, int n, const float* alpha, const float* A, int lda, float* B, int ldb, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasStrsm(handle, side, uplo, trans, diag, m, n, alpha, A, lda, B, ldb); } template <> inline cublasStatus_t cublastrsm(cublasHandle_t handle, cublasSideMode_t side, cublasFillMode_t uplo, cublasOperation_t trans, cublasDiagType_t diag, int m, int n, const double* alpha, const double* A, int lda, double* B, int ldb, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDtrsm(handle, side, uplo, trans, diag, m, n, alpha, A, lda, B, ldb); } /** * @defgroup dot cublas dot calls * @{ */ template <typename T> cublasStatus_t cublasdot(cublasHandle_t handle, int n, const T* x, int incx, const T* y, int incy, T* result, cudaStream_t stream); template <> inline cublasStatus_t cublasdot(cublasHandle_t handle, int n, const float* x, int incx, const float* y, int incy, float* result, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDotEx( handle, n, x, CUDA_R_32F, incx, y, CUDA_R_32F, incy, result, CUDA_R_32F, CUDA_R_32F); } template <> inline cublasStatus_t cublasdot(cublasHandle_t handle, int n, const double* x, int incx, const double* y, int incy, double* result, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDotEx( handle, n, x, CUDA_R_64F, incx, y, CUDA_R_64F, incy, result, CUDA_R_64F, CUDA_R_64F); } /** @} */ /** * @defgroup setpointermode cublas set pointer mode method * @{ */ // no T dependency... // template <typename T> // cublasStatus_t cublassetpointermode( // NOLINT // cublasHandle_t handle, // cublasPointerMode_t mode, // cudaStream_t stream); // template<> inline cublasStatus_t cublassetpointermode(cublasHandle_t handle, cublasPointerMode_t mode, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSetPointerMode(handle, mode); } /** @} */ /** * @defgroup scal cublas dot calls * @{ */ template <typename T> cublasStatus_t cublasscal( cublasHandle_t handle, int n, const T* alpha, T* x, int incx, cudaStream_t stream); template <> inline cublasStatus_t cublasscal( cublasHandle_t handle, int n, const float* alpha, float* x, int incx, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasSscal(handle, n, alpha, x, incx); } template <> inline cublasStatus_t cublasscal( cublasHandle_t handle, int n, const double* alpha, double* x, int incx, cudaStream_t stream) { RAFT_CUBLAS_TRY(cublasSetStream(handle, stream)); return cublasDscal(handle, n, alpha, x, incx); } /** @} */ } // namespace detail } // namespace linalg } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/gemv.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 <cublas_v2.h> #include <raft/core/resource/cublas_handle.hpp> #include "cublas_wrappers.hpp" #include <raft/core/resources.hpp> namespace raft { namespace linalg { namespace detail { template <typename math_t, bool DevicePointerMode = false> void gemv(raft::resources const& handle, const bool trans_a, const int m, const int n, const math_t* alpha, const math_t* A, const int lda, const math_t* x, const int incx, const math_t* beta, math_t* y, const int incy, cudaStream_t stream) { cublasHandle_t cublas_h = resource::get_cublas_handle(handle); detail::cublas_device_pointer_mode<DevicePointerMode> pmode(cublas_h); RAFT_CUBLAS_TRY(detail::cublasgemv(cublas_h, trans_a ? CUBLAS_OP_T : CUBLAS_OP_N, m, n, alpha, A, lda, x, incx, beta, y, incy, stream)); } template <typename math_t> void gemv(raft::resources const& handle, const math_t* A, const int n_rows, const int n_cols, const math_t* x, const int incx, math_t* y, const int incy, const bool trans_a, const math_t alpha, const math_t beta, cudaStream_t stream) { gemv(handle, trans_a, n_rows, n_cols, &alpha, A, n_rows, x, incx, &beta, y, incy, stream); } template <typename math_t> void gemv(raft::resources const& handle, const math_t* A, const int n_rows_a, const int n_cols_a, const math_t* x, math_t* y, const bool trans_a, const math_t alpha, const math_t beta, cudaStream_t stream) { gemv(handle, A, n_rows_a, n_cols_a, x, 1, y, 1, trans_a, alpha, beta, stream); } template <typename math_t> void gemv(raft::resources const& handle, const math_t* A, const int n_rows_a, const int n_cols_a, const math_t* x, math_t* y, const bool trans_a, cudaStream_t stream) { math_t alpha = math_t(1); math_t beta = math_t(0); gemv(handle, A, n_rows_a, n_cols_a, x, 1, y, 1, trans_a, alpha, beta, stream); } template <typename math_t> void gemv(raft::resources const& handle, const math_t* A, const int n_rows_a, const int n_cols_a, const int lda, const math_t* x, math_t* y, const bool trans_a, const math_t alpha, const math_t beta, cudaStream_t stream) { cublasHandle_t cublas_h = resource::get_cublas_handle(handle); cublasOperation_t op_a = trans_a ? CUBLAS_OP_T : CUBLAS_OP_N; RAFT_CUBLAS_TRY( cublasgemv(cublas_h, op_a, n_rows_a, n_cols_a, &alpha, A, lda, x, 1, &beta, y, 1, stream)); } template <typename math_t> void gemv(raft::resources const& handle, const math_t* A, const int n_rows_a, const int n_cols_a, const int lda, const math_t* x, math_t* y, const bool trans_a, cudaStream_t stream) { math_t alpha = math_t(1); math_t beta = math_t(0); gemv(handle, A, n_rows_a, n_cols_a, lda, x, y, trans_a, alpha, beta, stream); } }; // namespace detail }; // namespace linalg }; // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/svd.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 "cublas_wrappers.hpp" #include "cusolver_wrappers.hpp" #include <raft/core/resource/cublas_handle.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resource/cusolver_dn_handle.hpp> #include <raft/linalg/eig.cuh> #include <raft/linalg/gemm.cuh> #include <raft/linalg/transpose.cuh> #include <raft/common/nvtx.hpp> #include <raft/core/resources.hpp> #include <raft/matrix/diagonal.cuh> #include <raft/matrix/math.cuh> #include <raft/matrix/norm.cuh> #include <raft/matrix/reverse.cuh> #include <raft/util/cuda_utils.cuh> #include <raft/util/cudart_utils.hpp> #include <rmm/device_scalar.hpp> #include <rmm/device_uvector.hpp> namespace raft { namespace linalg { namespace detail { template <typename T> void svdQR(raft::resources const& handle, T* in, int n_rows, int n_cols, T* sing_vals, T* left_sing_vecs, T* right_sing_vecs, bool trans_right, bool gen_left_vec, bool gen_right_vec, cudaStream_t stream) { common::nvtx::range<common::nvtx::domain::raft> fun_scope( "raft::linalg::svdQR(%d, %d)", n_rows, n_cols); cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); cublasHandle_t cublasH = resource::get_cublas_handle(handle); const int m = n_rows; const int n = n_cols; rmm::device_scalar<int> devInfo(stream); T* d_rwork = nullptr; int lwork = 0; RAFT_CUSOLVER_TRY(cusolverDngesvd_bufferSize<T>(cusolverH, n_rows, n_cols, &lwork)); rmm::device_uvector<T> d_work(lwork, stream); char jobu = 'S'; char jobvt = 'A'; if (!gen_left_vec) { jobu = 'N'; } if (!gen_right_vec) { jobvt = 'N'; } RAFT_CUSOLVER_TRY(cusolverDngesvd(cusolverH, jobu, jobvt, m, n, in, m, sing_vals, left_sing_vecs, m, right_sing_vecs, n, d_work.data(), lwork, d_rwork, devInfo.data(), stream)); // Transpose the right singular vector back if (trans_right && right_sing_vecs != nullptr) raft::linalg::transpose(right_sing_vecs, n_cols, stream); RAFT_CUDA_TRY(cudaGetLastError()); int dev_info; raft::update_host(&dev_info, devInfo.data(), 1, stream); resource::sync_stream(handle, stream); ASSERT(dev_info == 0, "svd.cuh: svd couldn't converge to a solution. " "This usually occurs when some of the features do not vary enough."); } template <typename math_t, typename idx_t> void svdEig(raft::resources const& handle, math_t* in, idx_t n_rows, idx_t n_cols, math_t* S, math_t* U, math_t* V, bool gen_left_vec, cudaStream_t stream) { common::nvtx::range<common::nvtx::domain::raft> fun_scope( "raft::linalg::svdEig(%d, %d)", n_rows, n_cols); cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); cublasHandle_t cublasH = resource::get_cublas_handle(handle); auto len = n_cols * n_cols; rmm::device_uvector<math_t> in_cross_mult(len, stream); math_t alpha = math_t(1); math_t beta = math_t(0); raft::linalg::gemm(handle, in, n_rows, n_cols, in, in_cross_mult.data(), n_cols, n_cols, CUBLAS_OP_T, CUBLAS_OP_N, alpha, beta, stream); raft::linalg::eigDC(handle, in_cross_mult.data(), n_cols, n_cols, V, S, stream); raft::matrix::col_reverse(handle, make_device_matrix_view<math_t, idx_t, col_major>(V, n_cols, n_cols)); raft::matrix::row_reverse(handle, make_device_matrix_view<math_t, idx_t, col_major>(S, n_cols, idx_t(1))); raft::matrix::seqRoot(S, S, alpha, n_cols, stream, true); if (gen_left_vec) { raft::linalg::gemm(handle, in, n_rows, n_cols, V, U, n_rows, n_cols, CUBLAS_OP_N, CUBLAS_OP_N, alpha, beta, stream); raft::matrix::matrixVectorBinaryDivSkipZero(U, S, n_rows, n_cols, false, true, stream); } } template <typename math_t> void svdJacobi(raft::resources const& handle, math_t* in, int n_rows, int n_cols, math_t* sing_vals, math_t* left_sing_vecs, math_t* right_sing_vecs, bool gen_left_vec, bool gen_right_vec, math_t tol, int max_sweeps, cudaStream_t stream) { common::nvtx::range<common::nvtx::domain::raft> fun_scope( "raft::linalg::svdJacobi(%d, %d)", n_rows, n_cols); cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); gesvdjInfo_t gesvdj_params = NULL; RAFT_CUSOLVER_TRY(cusolverDnCreateGesvdjInfo(&gesvdj_params)); RAFT_CUSOLVER_TRY(cusolverDnXgesvdjSetTolerance(gesvdj_params, tol)); RAFT_CUSOLVER_TRY(cusolverDnXgesvdjSetMaxSweeps(gesvdj_params, max_sweeps)); int m = n_rows; int n = n_cols; rmm::device_scalar<int> devInfo(stream); int lwork = 0; int econ = 1; RAFT_CUSOLVER_TRY(cusolverDngesvdj_bufferSize(cusolverH, CUSOLVER_EIG_MODE_VECTOR, econ, m, n, in, m, sing_vals, left_sing_vecs, m, right_sing_vecs, n, &lwork, gesvdj_params)); rmm::device_uvector<math_t> d_work(lwork, stream); RAFT_CUSOLVER_TRY(cusolverDngesvdj(cusolverH, CUSOLVER_EIG_MODE_VECTOR, econ, m, n, in, m, sing_vals, left_sing_vecs, m, right_sing_vecs, n, d_work.data(), lwork, devInfo.data(), gesvdj_params, stream)); RAFT_CUSOLVER_TRY(cusolverDnDestroyGesvdjInfo(gesvdj_params)); } template <typename math_t> void svdReconstruction(raft::resources const& handle, math_t* U, math_t* S, math_t* V, math_t* out, int n_rows, int n_cols, int k, cudaStream_t stream) { const math_t alpha = 1.0, beta = 0.0; rmm::device_uvector<math_t> SVT(k * n_cols, stream); raft::linalg::gemm( handle, S, k, k, V, SVT.data(), k, n_cols, CUBLAS_OP_N, CUBLAS_OP_T, alpha, beta, stream); raft::linalg::gemm(handle, U, n_rows, k, SVT.data(), out, n_rows, n_cols, CUBLAS_OP_N, CUBLAS_OP_N, alpha, beta, stream); } template <typename math_t> bool evaluateSVDByL2Norm(raft::resources const& handle, math_t* A_d, math_t* U, math_t* S_vec, math_t* V, int n_rows, int n_cols, int k, math_t tol, cudaStream_t stream) { cublasHandle_t cublasH = resource::get_cublas_handle(handle); int m = n_rows, n = n_cols; // form product matrix rmm::device_uvector<math_t> P_d(m * n, stream); rmm::device_uvector<math_t> S_mat(k * k, stream); RAFT_CUDA_TRY(cudaMemsetAsync(P_d.data(), 0, sizeof(math_t) * m * n, stream)); RAFT_CUDA_TRY(cudaMemsetAsync(S_mat.data(), 0, sizeof(math_t) * k * k, stream)); raft::matrix::set_diagonal(handle, make_device_vector_view<const math_t>(S_vec, k), make_device_matrix_view<math_t>(S_mat.data(), k, k)); svdReconstruction(handle, U, S_mat.data(), V, P_d.data(), m, n, k, stream); // get norms of each math_t normA = raft::matrix::l2_norm(handle, make_device_matrix_view<const math_t>(A_d, m, n)); math_t normU = raft::matrix::l2_norm(handle, make_device_matrix_view<const math_t>(U, m, k)); math_t normS = raft::matrix::l2_norm(handle, make_device_matrix_view<const math_t>(S_mat.data(), k, k)); math_t normV = raft::matrix::l2_norm(handle, make_device_matrix_view<const math_t>(V, n, k)); math_t normP = raft::matrix::l2_norm(handle, make_device_matrix_view<const math_t>(P_d.data(), m, n)); // calculate percent error const math_t alpha = 1.0, beta = -1.0; rmm::device_uvector<math_t> A_minus_P(m * n, stream); RAFT_CUDA_TRY(cudaMemsetAsync(A_minus_P.data(), 0, sizeof(math_t) * m * n, stream)); RAFT_CUBLAS_TRY(cublasgeam(cublasH, CUBLAS_OP_N, CUBLAS_OP_N, m, n, &alpha, A_d, m, &beta, P_d.data(), m, A_minus_P.data(), m, stream)); math_t norm_A_minus_P = raft::matrix::l2_norm(handle, make_device_matrix_view<const math_t>(A_minus_P.data(), m, n)); math_t percent_error = 100.0 * norm_A_minus_P / normA; return (percent_error / 100.0 < tol); } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/map_then_reduce.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 <cub/cub.cuh> #include <raft/core/resources.hpp> #include <raft/util/cuda_utils.cuh> #include <raft/util/vectorized.cuh> namespace raft { namespace linalg { namespace detail { struct sum_tag {}; template <typename InType, typename OutType, int TPB> __device__ void reduce(OutType* out, const InType acc, sum_tag) { typedef cub::BlockReduce<InType, TPB> BlockReduce; __shared__ typename BlockReduce::TempStorage temp_storage; OutType tmp = BlockReduce(temp_storage).Sum(acc); if (threadIdx.x == 0) { raft::myAtomicAdd(out, tmp); } } template <typename InType, typename OutType, int TPB, typename ReduceLambda> __device__ void reduce(OutType* out, const InType acc, ReduceLambda op) { typedef cub::BlockReduce<InType, TPB> BlockReduce; __shared__ typename BlockReduce::TempStorage temp_storage; OutType tmp = BlockReduce(temp_storage).Reduce(acc, op); if (threadIdx.x == 0) { raft::myAtomicReduce(out, tmp, op); } } template <typename InType, typename OutType, typename IdxType, typename MapOp, typename ReduceLambda, int TPB, typename... Args> RAFT_KERNEL mapThenReduceKernel(OutType* out, IdxType len, OutType neutral, MapOp map, ReduceLambda op, const InType* in, Args... args) { OutType acc = neutral; auto idx = (threadIdx.x + (blockIdx.x * blockDim.x)); if (idx < len) { acc = map(in[idx], args[idx]...); } __syncthreads(); reduce<InType, OutType, TPB>(out, acc, op); } template <typename InType, typename OutType, typename IdxType, typename MapOp, typename ReduceLambda, int TPB, typename... Args> void mapThenReduceImpl(OutType* out, IdxType len, OutType neutral, MapOp map, ReduceLambda op, cudaStream_t stream, const InType* in, Args... args) { raft::update_device(out, &neutral, 1, stream); const int nblks = raft::ceildiv(len, IdxType(TPB)); mapThenReduceKernel<InType, OutType, IdxType, MapOp, ReduceLambda, TPB, Args...> <<<nblks, TPB, 0, stream>>>(out, len, neutral, map, op, in, args...); RAFT_CUDA_TRY(cudaPeekAtLastError()); } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/contractions.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/util/device_loads_stores.cuh> namespace raft { namespace linalg { namespace detail { template <typename DataT, typename IdxT, typename Policy, bool isRowMajor = true> struct Contractions_NT { protected: typedef Policy P; /** number of rows in X */ IdxT m; /** number of rows in Y */ IdxT n; /** number of columns in X and Y */ IdxT k; /** leading dimension in X */ IdxT lda; /** leading dimension in Y */ IdxT ldb; /** leading dimension in Output D */ IdxT ldd; /** global memory pointer to X matrix */ const DataT* x_base; /** global memory pointer to Y matrix */ const DataT* y_base; /** current thread's smem row id */ int srowid; /** current thread's smem column id */ int scolid; /** current thread's accumulation row id */ int accrowid; /** current thread's accumulation column id */ int acccolid; /** base smem pointer for X data storage */ DataT* sx; /** base smem pointer for Y data storage */ DataT* sy; /** index pointing the correct smem page for writing after `ldgXY()` */ int pageWr; /** index pointing the correct smem page for reading during `ldsXY()` */ int pageRd; /** block of X data loaded from smem after `ldsXY()` */ DataT regx[P::AccRowsPerTh][P::Veclen]; /** block of Y data loaded from smem after `ldsXY()` */ DataT regy[P::AccColsPerTh][P::Veclen]; /** block of X data loaded from global mem after `ldgXY()` */ DataT ldgDataX[P::LdgPerThX][P::Veclen]; /** block of Y data loaded from global mem after `ldgXY()` */ DataT ldgDataY[P::LdgPerThY][P::Veclen]; static constexpr DataT Zero = (DataT)0; public: /** * @brief Ctor * @param[in] _x X matrix. [on device] [dim = _m x _k] [row-major] * @param[in] _y Y matrix. [on device] [dim = _n x _k] [row-major] * @param[in] _m number of rows of X * @param[in] _n number of rows of Y * @param[in] _k number of cols of X and Y * @param[in] _smem shared memory region used during computations */ DI Contractions_NT(const DataT* _x, const DataT* _y, IdxT _m, IdxT _n, IdxT _k, char* _smem) : m(_m), n(_n), k(_k), lda(_k), ldb(_k), x_base(_x), y_base(_y), srowid(threadIdx.x / P::LdgThRow), scolid((threadIdx.x % P::LdgThRow) * P::Veclen), accrowid(threadIdx.x / P::AccThCols), acccolid(threadIdx.x % P::AccThCols), sx((DataT*)_smem), sy(&(sx[P::SmemPageX])), pageWr(0), pageRd(0) { } /** * @brief Ctor * @param[in] _x X matrix. [on device] [dim = _m x _k] [row-major] * @param[in] _y Y matrix. [on device] [dim = _n x _k] [row-major] * @param[in] _m number of rows of X * @param[in] _n number of rows of Y * @param[in] _k number of cols of X and Y * @param[in] _smem shared memory region used during computations */ DI Contractions_NT(const DataT* _x, const DataT* _y, IdxT _m, IdxT _n, IdxT _k, IdxT _lda, IdxT _ldb, IdxT _ldd, char* _smem) : m(_m), n(_n), k(_k), lda(_lda), ldb(_ldb), ldd(_ldd), x_base(_x), y_base(_y), srowid(threadIdx.x / P::LdgThRow), scolid((threadIdx.x % P::LdgThRow) * P::Veclen), accrowid(threadIdx.x / P::AccThCols), acccolid(threadIdx.x % P::AccThCols), sx((DataT*)_smem), sy(&(sx[P::SmemPageX])), pageWr(0), pageRd(0) { } protected: /** * @brief Load current block of X/Y from global memory to registers * @param[in] kidx current start index of k to be loaded */ DI void ldgXY(IdxT tile_idx_m, IdxT tile_idx_n, IdxT kidx) { ldgX(tile_idx_m, kidx); ldgY(tile_idx_n, kidx); } DI void ldgXY(IdxT tile_idx_m, IdxT tile_idx_n, IdxT kidx, IdxT tile_end_n) { ldgX(tile_idx_m, kidx); ldgY(tile_idx_n, kidx, tile_end_n); } /** * @brief Store current block of X/Y from registers to smem * @param[in] kidx current start index of k to be loaded */ DI void stsXY() { stsX(sx + pageWr * P::SmemPage); stsY(sy + pageWr * P::SmemPage); } /** * @brief Load X and Y block from shared memory to registers * @param[in] kidx k value from the current k-block to be loaded from smem */ DI void ldsXY(int kidx) { ldsX(kidx, sx + pageRd * P::SmemPage); ldsY(kidx, sy + pageRd * P::SmemPage); } DI void switch_read_buffer() { this->pageRd ^= 1; } DI void switch_write_buffer() { this->pageWr ^= 1; } private: DI void ldgX(IdxT tile_idx_m, IdxT kidx) { IdxT xrowid = isRowMajor ? tile_idx_m + srowid : tile_idx_m; auto x = isRowMajor ? x_base + xrowid * lda : x_base + xrowid + srowid * lda; if (isRowMajor) { auto numRows = m; auto koffset = kidx + scolid; #pragma unroll for (int i = 0; i < P::LdgPerThX; ++i) { if (koffset < lda && (xrowid + i * P::LdgRowsX) < numRows) { ldg(ldgDataX[i], x + i * P::LdgRowsX * lda + koffset); } else { #pragma unroll for (int j = 0; j < P::Veclen; ++j) { ldgDataX[i][j] = Zero; } } } } else { const auto numRows = k; auto koffset = scolid; #pragma unroll for (int i = 0; i < P::LdgPerThX; ++i) { if ((koffset + xrowid) < lda && (srowid + kidx + i * P::LdgRowsX) < numRows) { ldg(ldgDataX[i], x + (kidx + i * P::LdgRowsX) * lda + koffset); } else { #pragma unroll for (int j = 0; j < P::Veclen; ++j) { ldgDataX[i][j] = Zero; } } } } } DI void ldgY(IdxT tile_idx_n, IdxT kidx) { ldgY(tile_idx_n, kidx, n); } DI void ldgY(IdxT tile_idx_n, IdxT kidx, IdxT end_n) { IdxT yrowid = isRowMajor ? tile_idx_n + srowid : tile_idx_n; auto y = isRowMajor ? y_base + yrowid * ldb : y_base + yrowid + srowid * ldb; if (isRowMajor) { auto numRows = end_n; auto koffset = kidx + scolid; #pragma unroll for (int i = 0; i < P::LdgPerThY; ++i) { if (koffset < ldb && (yrowid + i * P::LdgRowsY) < numRows) { ldg(ldgDataY[i], y + i * P::LdgRowsY * ldb + koffset); } else { #pragma unroll for (int j = 0; j < P::Veclen; ++j) { ldgDataY[i][j] = Zero; } } } } else { auto numRows = k; auto koffset = scolid; #pragma unroll for (int i = 0; i < P::LdgPerThY; ++i) { if ((koffset + yrowid) < end_n && (srowid + kidx + i * P::LdgRowsY) < numRows) { ldg(ldgDataY[i], y + (kidx + i * P::LdgRowsY) * ldb + koffset); } else { #pragma unroll for (int j = 0; j < P::Veclen; ++j) { ldgDataY[i][j] = Zero; } } } } } DI void stsX(DataT* smem) { auto* saddr = smem + srowid * P::SmemStride + scolid; #pragma unroll for (int i = 0; i < P::LdgPerThX; ++i) { sts(saddr + i * P::LdgRowsX * P::SmemStride, ldgDataX[i]); } } DI void stsY(DataT* smem) { auto* saddr = smem + srowid * P::SmemStride + scolid; #pragma unroll for (int i = 0; i < P::LdgPerThY; ++i) { sts(saddr + i * P::LdgRowsY * P::SmemStride, ldgDataY[i]); } } DI void ldsX(int kidx, DataT* smem) { if (isRowMajor) { auto* saddr = smem + accrowid * P::SmemStride + kidx; #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { lds(regx[i], saddr + i * P::AccThRows * P::SmemStride); } } else { auto* saddr = smem + accrowid + kidx * P::SmemStride; #pragma unroll for (int i = 0; i < P::AccRowsPerTh; ++i) { #pragma unroll for (int v = 0; v < P::Veclen; ++v) { regx[i][v] = saddr[i * P::AccThRows + v * P::SmemStride]; } } } } DI void ldsY(int kidx, DataT* smem) { if (isRowMajor) { auto* saddr = smem + acccolid * P::SmemStride + kidx; #pragma unroll for (int i = 0; i < P::AccColsPerTh; ++i) { lds(regy[i], saddr + i * P::AccThCols * P::SmemStride); } } else { auto* saddr = smem + acccolid + kidx * P::SmemStride; #pragma unroll for (int i = 0; i < P::AccColsPerTh; ++i) { #pragma unroll for (int v = 0; v < P::Veclen; ++v) { regy[i][v] = saddr[i * P::AccThCols + v * P::SmemStride]; } } } } }; // struct Contractions_NT } // namespace detail } // namespace linalg } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/mean_squared_error.cuh
/* * Copyright (c) 2022, 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/map_then_reduce.cuh> namespace raft { namespace linalg { namespace detail { template <typename math_t, int TPB = 256> void meanSquaredError( math_t* out, const math_t* A, const math_t* B, size_t len, math_t weight, cudaStream_t stream) { auto sq_diff = [len, weight] __device__(const math_t a, const math_t b) { math_t diff = a - b; return diff * diff * weight / len; }; raft::linalg::mapThenSumReduce<math_t, decltype(sq_diff)>(out, len, sq_diff, stream, A, B); } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/rsvd.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 <raft/core/resource/cublas_handle.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resource/cusolver_dn_handle.hpp> #include <raft/linalg/eig.cuh> #include <raft/linalg/gemm.cuh> #include <raft/linalg/qr.cuh> #include <raft/linalg/svd.cuh> #include <raft/linalg/transpose.cuh> #include <raft/matrix/diagonal.cuh> #include <raft/matrix/math.cuh> #include <raft/matrix/reverse.cuh> #include <raft/matrix/slice.cuh> #include <raft/matrix/triangular.cuh> #include <raft/random/rng.cuh> #include <raft/util/cuda_utils.cuh> #include <algorithm> namespace raft { namespace linalg { namespace detail { template <typename math_t> void randomized_svd(const raft::resources& handle, const math_t* in, std::size_t n_rows, std::size_t n_cols, std::size_t k, std::size_t p, std::size_t niters, math_t* S, math_t* U, math_t* V, bool gen_U, bool gen_V) { common::nvtx::range<common::nvtx::domain::raft> fun_scope( "raft::linalg::randomized_svd(%d, %d, %d)", n_rows, n_cols, k); RAFT_EXPECTS(k < std::min(n_rows, n_cols), "k must be < min(n_rows, n_cols)"); RAFT_EXPECTS((k + p) < std::min(n_rows, n_cols), "k + p must be < min(n_rows, n_cols)"); RAFT_EXPECTS(!gen_U || (U != nullptr), "computation of U vector requested but found nullptr"); RAFT_EXPECTS(!gen_V || (V != nullptr), "computation of V vector requested but found nullptr"); #if CUDART_VERSION < 11050 RAFT_EXPECTS(gen_U && gen_V, "not computing U or V is not supported in CUDA version < 11.5"); #endif cudaStream_t stream = resource::get_cuda_stream(handle); cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); char jobu = gen_U ? 'S' : 'N'; char jobv = gen_V ? 'S' : 'N'; auto lda = n_rows; auto ldu = n_rows; auto ldv = n_cols; auto* in_ptr = const_cast<math_t*>(in); size_t workspaceDevice = 0; size_t workspaceHost = 0; RAFT_CUSOLVER_TRY(cusolverDnxgesvdr_bufferSize(cusolverH, jobu, jobv, n_rows, n_cols, k, p, niters, in_ptr, lda, S, U, ldu, V, ldv, &workspaceDevice, &workspaceHost, stream)); auto d_workspace = raft::make_device_vector<char>(handle, workspaceDevice); auto h_workspace = raft::make_host_vector<char>(workspaceHost); auto devInfo = raft::make_device_scalar<int>(handle, 0); RAFT_CUSOLVER_TRY(cusolverDnxgesvdr(cusolverH, jobu, jobv, n_rows, n_cols, k, p, niters, in_ptr, lda, S, U, ldu, V, ldv, d_workspace.data_handle(), workspaceDevice, h_workspace.data_handle(), workspaceHost, devInfo.data_handle(), stream)); RAFT_CUDA_TRY(cudaGetLastError()); int dev_info; raft::update_host(&dev_info, devInfo.data_handle(), 1, stream); resource::sync_stream(handle); ASSERT(dev_info == 0, "rsvd.cuh: Invalid parameter encountered."); } /** * @brief randomized singular value decomposition (RSVD) on the column major * float type input matrix (Jacobi-based), by specifying no. of PCs and * upsamples directly * @param handle: raft handle * @param M: input matrix * @param n_rows: number rows of input matrix * @param n_cols: number columns of input matrix * @param S_vec: singular values of input matrix * @param U: left singular values of input matrix * @param V: right singular values of input matrix * @param k: no. of singular values to be computed * @param p: no. of upsamples * @param use_bbt: whether use eigen decomposition in computation or not * @param gen_left_vec: left vector needs to be generated or not? * @param gen_right_vec: right vector needs to be generated or not? * @param use_jacobi: whether to jacobi solver for decomposition * @param tol: tolerance for Jacobi-based solvers * @param max_sweeps: maximum number of sweeps for Jacobi-based solvers * @param stream cuda stream */ template <typename math_t> void rsvdFixedRank(raft::resources const& handle, math_t* M, int n_rows, int n_cols, math_t* S_vec, math_t* U, math_t* V, int k, int p, bool use_bbt, bool gen_left_vec, bool gen_right_vec, bool use_jacobi, math_t tol, int max_sweeps, cudaStream_t stream) { cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); cublasHandle_t cublasH = resource::get_cublas_handle(handle); // All the notations are following Algorithm 4 & 5 in S. Voronin's paper: // https://arxiv.org/abs/1502.05366 int m = n_rows, n = n_cols; int l = k + p; // Total number of singular values to be computed before truncation int q = 2; // Number of power sampling counts int s = 1; // Frequency controller for QR decomposition during power sampling // scheme. s = 1: 2 QR per iteration; s = 2: 1 QR per iteration; s // > 2: less frequent QR const math_t alpha = 1.0, beta = 0.0; // Build temporary U, S, V matrices rmm::device_uvector<math_t> S_vec_tmp(l, stream); RAFT_CUDA_TRY(cudaMemsetAsync(S_vec_tmp.data(), 0, sizeof(math_t) * l, stream)); // build random matrix rmm::device_uvector<math_t> RN(n * l, stream); raft::random::RngState state{484}; raft::random::normal(handle, state, RN.data(), n * l, math_t(0.0), alpha); // multiply to get matrix of random samples Y rmm::device_uvector<math_t> Y(m * l, stream); raft::linalg::gemm( handle, M, m, n, RN.data(), Y.data(), m, l, CUBLAS_OP_N, CUBLAS_OP_N, alpha, beta, stream); // now build up (M M^T)^q R rmm::device_uvector<math_t> Z(n * l, stream); rmm::device_uvector<math_t> Yorth(m * l, stream); rmm::device_uvector<math_t> Zorth(n * l, stream); RAFT_CUDA_TRY(cudaMemsetAsync(Z.data(), 0, sizeof(math_t) * n * l, stream)); RAFT_CUDA_TRY(cudaMemsetAsync(Yorth.data(), 0, sizeof(math_t) * m * l, stream)); RAFT_CUDA_TRY(cudaMemsetAsync(Zorth.data(), 0, sizeof(math_t) * n * l, stream)); // power sampling scheme for (int j = 1; j < q; j++) { if ((2 * j - 2) % s == 0) { raft::linalg::qrGetQ(handle, Y.data(), Yorth.data(), m, l, stream); raft::linalg::gemm(handle, M, m, n, Yorth.data(), Z.data(), n, l, CUBLAS_OP_T, CUBLAS_OP_N, alpha, beta, stream); } else { raft::linalg::gemm( handle, M, m, n, Y.data(), Z.data(), n, l, CUBLAS_OP_T, CUBLAS_OP_N, alpha, beta, stream); } if ((2 * j - 1) % s == 0) { raft::linalg::qrGetQ(handle, Z.data(), Zorth.data(), n, l, stream); raft::linalg::gemm(handle, M, m, n, Zorth.data(), Y.data(), m, l, CUBLAS_OP_N, CUBLAS_OP_N, alpha, beta, stream); } else { raft::linalg::gemm( handle, M, m, n, Z.data(), Y.data(), m, l, CUBLAS_OP_N, CUBLAS_OP_N, alpha, beta, stream); } } // orthogonalize on exit from loop to get Q rmm::device_uvector<math_t> Q(m * l, stream); RAFT_CUDA_TRY(cudaMemsetAsync(Q.data(), 0, sizeof(math_t) * m * l, stream)); raft::linalg::qrGetQ(handle, Y.data(), Q.data(), m, l, stream); // either QR of B^T method, or eigendecompose BB^T method if (!use_bbt) { // form Bt = Mt*Q : nxm * mxl = nxl rmm::device_uvector<math_t> Bt(n * l, stream); RAFT_CUDA_TRY(cudaMemsetAsync(Bt.data(), 0, sizeof(math_t) * n * l, stream)); raft::linalg::gemm( handle, M, m, n, Q.data(), Bt.data(), n, l, CUBLAS_OP_T, CUBLAS_OP_N, alpha, beta, stream); // compute QR factorization of Bt // M is mxn ; Q is mxn ; R is min(m,n) x min(m,n) */ rmm::device_uvector<math_t> Qhat(n * l, stream); RAFT_CUDA_TRY(cudaMemsetAsync(Qhat.data(), 0, sizeof(math_t) * n * l, stream)); rmm::device_uvector<math_t> Rhat(l * l, stream); RAFT_CUDA_TRY(cudaMemsetAsync(Rhat.data(), 0, sizeof(math_t) * l * l, stream)); raft::linalg::qrGetQR(handle, Bt.data(), Qhat.data(), Rhat.data(), n, l, stream); // compute SVD of Rhat (lxl) rmm::device_uvector<math_t> Uhat(l * l, stream); RAFT_CUDA_TRY(cudaMemsetAsync(Uhat.data(), 0, sizeof(math_t) * l * l, stream)); rmm::device_uvector<math_t> Vhat(l * l, stream); RAFT_CUDA_TRY(cudaMemsetAsync(Vhat.data(), 0, sizeof(math_t) * l * l, stream)); if (use_jacobi) raft::linalg::svdJacobi(handle, Rhat.data(), l, l, S_vec_tmp.data(), Uhat.data(), Vhat.data(), true, true, tol, max_sweeps, stream); else raft::linalg::svdQR(handle, Rhat.data(), l, l, S_vec_tmp.data(), Uhat.data(), Vhat.data(), true, true, true, stream); // First k elements of S_vec raft::matrix::slice( handle, make_device_matrix_view<const math_t, int, col_major>(S_vec_tmp.data(), 1, l), make_device_matrix_view<math_t, int, col_major>(S_vec, 1, k), raft::matrix::slice_coordinates(0, 0, 1, k)); // Merge step 14 & 15 by calculating U = Q*Vhat[:,1:k] mxl * lxk = mxk if (gen_left_vec) { raft::linalg::gemm(handle, Q.data(), m, l, Vhat.data(), U, m, k /*used to be l and needs slicing*/, CUBLAS_OP_N, CUBLAS_OP_N, alpha, beta, stream); } // Merge step 14 & 15 by calculating V = Qhat*Uhat[:,1:k] nxl * lxk = nxk if (gen_right_vec) { raft::linalg::gemm(handle, Qhat.data(), n, l, Uhat.data(), V, n, k /*used to be l and needs slicing*/, CUBLAS_OP_N, CUBLAS_OP_N, alpha, beta, stream); } } else { // build the matrix B B^T = Q^T M M^T Q column by column // Bt = M^T Q ; nxm * mxk = nxk rmm::device_uvector<math_t> B(n * l, stream); raft::linalg::gemm( handle, Q.data(), m, l, M, B.data(), l, n, CUBLAS_OP_T, CUBLAS_OP_N, alpha, beta, stream); rmm::device_uvector<math_t> BBt(l * l, stream); raft::linalg::gemm(handle, B.data(), l, n, B.data(), BBt.data(), l, l, CUBLAS_OP_N, CUBLAS_OP_T, alpha, beta, stream); // compute eigendecomposition of BBt rmm::device_uvector<math_t> Uhat(l * l, stream); RAFT_CUDA_TRY(cudaMemsetAsync(Uhat.data(), 0, sizeof(math_t) * l * l, stream)); rmm::device_uvector<math_t> Uhat_dup(l * l, stream); RAFT_CUDA_TRY(cudaMemsetAsync(Uhat_dup.data(), 0, sizeof(math_t) * l * l, stream)); raft::matrix::upper_triangular( handle, make_device_matrix_view<const math_t, int, col_major>(BBt.data(), l, l), make_device_matrix_view<math_t, int, col_major>(Uhat_dup.data(), l, l)); if (use_jacobi) raft::linalg::eigJacobi( handle, Uhat_dup.data(), l, l, Uhat.data(), S_vec_tmp.data(), stream, tol, max_sweeps); else raft::linalg::eigDC(handle, Uhat_dup.data(), l, l, Uhat.data(), S_vec_tmp.data(), stream); raft::matrix::seqRoot(S_vec_tmp.data(), l, stream); auto S_vec_view = make_device_matrix_view<math_t, int, col_major>(S_vec, 1, k); raft::matrix::slice( handle, raft::make_device_matrix_view<const math_t, int, col_major>(S_vec_tmp.data(), 1, l), S_vec_view, raft::matrix::slice_coordinates(0, p, 1, l)); // Last k elements of S_vec raft::matrix::col_reverse(handle, S_vec_view); // Merge step 14 & 15 by calculating U = Q*Uhat[:,(p+1):l] mxl * lxk = mxk if (gen_left_vec) { raft::linalg::gemm(handle, Q.data(), m, l, Uhat.data() + p * l, U, m, k, CUBLAS_OP_N, CUBLAS_OP_N, alpha, beta, stream); raft::matrix::col_reverse(handle, make_device_matrix_view<math_t, int, col_major>(U, m, k)); } // Merge step 14 & 15 by calculating V = B^T Uhat[:,(p+1):l] * // Sigma^{-1}[(p+1):l, (p+1):l] nxl * lxk * kxk = nxk if (gen_right_vec) { rmm::device_uvector<math_t> Sinv(k * k, stream); RAFT_CUDA_TRY(cudaMemsetAsync(Sinv.data(), 0, sizeof(math_t) * k * k, stream)); rmm::device_uvector<math_t> UhatSinv(l * k, stream); RAFT_CUDA_TRY(cudaMemsetAsync(UhatSinv.data(), 0, sizeof(math_t) * l * k, stream)); raft::matrix::reciprocal(S_vec_tmp.data(), l, stream); raft::matrix::set_diagonal(handle, make_device_vector_view<const math_t>(S_vec_tmp.data() + p, k), make_device_matrix_view<math_t>(Sinv.data(), k, k)); raft::linalg::gemm(handle, Uhat.data() + p * l, l, k, Sinv.data(), UhatSinv.data(), l, k, CUBLAS_OP_N, CUBLAS_OP_N, alpha, beta, stream); raft::linalg::gemm(handle, B.data(), l, n, UhatSinv.data(), V, n, k, CUBLAS_OP_T, CUBLAS_OP_N, alpha, beta, stream); raft::matrix::col_reverse(handle, make_device_matrix_view<math_t, int, col_major>(V, n, k)); } } } /** * @brief randomized singular value decomposition (RSVD) on the column major * float type input matrix (Jacobi-based), by specifying the PC and upsampling * ratio * @param handle: raft handle * @param M: input matrix * @param n_rows: number rows of input matrix * @param n_cols: number columns of input matrix * @param S_vec: singular values of input matrix * @param U: left singular values of input matrix * @param V: right singular values of input matrix * @param PC_perc: percentage of singular values to be computed * @param UpS_perc: upsampling percentage * @param use_bbt: whether use eigen decomposition in computation or not * @param gen_left_vec: left vector needs to be generated or not? * @param gen_right_vec: right vector needs to be generated or not? * @param use_jacobi: whether to jacobi solver for decomposition * @param tol: tolerance for Jacobi-based solvers * @param max_sweeps: maximum number of sweeps for Jacobi-based solvers * @param stream cuda stream */ template <typename math_t> void rsvdPerc(raft::resources const& handle, math_t* M, int n_rows, int n_cols, math_t* S_vec, math_t* U, math_t* V, math_t PC_perc, math_t UpS_perc, bool use_bbt, bool gen_left_vec, bool gen_right_vec, bool use_jacobi, math_t tol, int max_sweeps, cudaStream_t stream) { int k = std::max((int)(std::min(n_rows, n_cols) * PC_perc), 1); // Number of singular values to be computed int p = std::max((int)(std::min(n_rows, n_cols) * UpS_perc), 1); // Upsamples rsvdFixedRank(handle, M, n_rows, n_cols, S_vec, U, V, k, p, use_bbt, gen_left_vec, gen_right_vec, use_jacobi, tol, max_sweeps, stream); } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/coalesced_reduction.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 // Always include inline definitions of coalesced reduction, because we do not // force explicit instantion. #include "coalesced_reduction-inl.cuh" // Do include the extern template instantiations when possible. #ifdef RAFT_COMPILED #include "coalesced_reduction-ext.cuh" #endif
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/norm.cuh
/* * Copyright (c) 2022, 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> #include <raft/linalg/norm_types.hpp> #include <raft/linalg/reduce.cuh> namespace raft { namespace linalg { namespace detail { template <typename Type, typename IdxType, typename Lambda> void rowNormCaller(Type* dots, const Type* data, IdxType D, IdxType N, NormType type, bool rowMajor, cudaStream_t stream, Lambda fin_op) { switch (type) { case L1Norm: raft::linalg::reduce<Type, Type, IdxType>(dots, data, D, N, (Type)0, rowMajor, true, stream, false, raft::abs_op(), raft::add_op(), fin_op); break; case L2Norm: raft::linalg::reduce<Type, Type, IdxType>(dots, data, D, N, (Type)0, rowMajor, true, stream, false, raft::sq_op(), raft::add_op(), fin_op); break; case LinfNorm: raft::linalg::reduce<Type, Type, IdxType>(dots, data, D, N, (Type)0, rowMajor, true, stream, false, raft::abs_op(), raft::max_op(), fin_op); break; default: THROW("Unsupported norm type: %d", type); }; } template <typename Type, typename IdxType, typename Lambda> void colNormCaller(Type* dots, const Type* data, IdxType D, IdxType N, NormType type, bool rowMajor, cudaStream_t stream, Lambda fin_op) { switch (type) { case L1Norm: raft::linalg::reduce<Type, Type, IdxType>(dots, data, D, N, (Type)0, rowMajor, false, stream, false, raft::abs_op(), raft::add_op(), fin_op); break; case L2Norm: raft::linalg::reduce<Type, Type, IdxType>(dots, data, D, N, (Type)0, rowMajor, false, stream, false, raft::sq_op(), raft::add_op(), fin_op); break; case LinfNorm: raft::linalg::reduce<Type, Type, IdxType>(dots, data, D, N, (Type)0, rowMajor, false, stream, false, raft::abs_op(), raft::max_op(), fin_op); break; default: THROW("Unsupported norm type: %d", type); }; } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/cholesky_r1_update.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 "cublas_wrappers.hpp" #include "cusolver_wrappers.hpp" #include <raft/core/resource/cublas_handle.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resources.hpp> #include <raft/linalg/binary_op.cuh> namespace raft { namespace linalg { namespace detail { template <typename math_t> void choleskyRank1Update(raft::resources const& handle, math_t* L, int n, int ld, void* workspace, int* n_bytes, cublasFillMode_t uplo, cudaStream_t stream, math_t eps = -1) { // The matrix A' is defined as: // A' = [[A_11, A_12] // [A_21, A_22]] // where: // - A_11 = A, matrix of size (n-1)x(n-1) // - A_21[j] = A_12.T[j] = A_new[j] j=0..n-2, vector with (n-1) elements // - A_22 = A_new[n-1] scalar. // // Instead of calculating the Cholelsky decomposition of A' from scratch, // we just update L with the new row. The new Cholesky decomposition will be // calculated as: // L' = [[L_11, 0] // [L_12, L_22]] // where L_11 is the Cholesky decomposition of A (size [n-1 x n-1]), and // L_12 and L_22 are the new quantities that we need to calculate. // We need a workspace in device memory to store a scalar. Additionally, in // CUBLAS_FILL_MODE_LOWER we need space for n-1 floats. const int align = 256; int offset = (uplo == CUBLAS_FILL_MODE_LOWER) ? raft::alignTo<int>(sizeof(math_t) * (n - 1), align) : 0; if (workspace == nullptr) { *n_bytes = offset + 1 * sizeof(math_t); return; } math_t* s = reinterpret_cast<math_t*>(((char*)workspace) + offset); math_t* L_22 = L + (n - 1) * ld + n - 1; math_t* A_new = nullptr; math_t* A_row = nullptr; if (uplo == CUBLAS_FILL_MODE_UPPER) { // A_new is stored as the n-1 th column of L A_new = L + (n - 1) * ld; } else { // If the input is lower triangular, then the new elements of A are stored // as the n-th row of L. Since the matrix is column major, this is non // contiguous. We copy elements from A_row to a contiguous workspace A_new. A_row = L + n - 1; A_new = reinterpret_cast<math_t*>(workspace); RAFT_CUBLAS_TRY( cublasCopy(resource::get_cublas_handle(handle), n - 1, A_row, ld, A_new, 1, stream)); } cublasOperation_t op = (uplo == CUBLAS_FILL_MODE_UPPER) ? CUBLAS_OP_T : CUBLAS_OP_N; if (n > 1) { // Calculate L_12 = x by solving equation L_11 x = A_12 math_t alpha = 1; RAFT_CUBLAS_TRY(cublastrsm(resource::get_cublas_handle(handle), CUBLAS_SIDE_LEFT, uplo, op, CUBLAS_DIAG_NON_UNIT, n - 1, 1, &alpha, L, ld, A_new, n - 1, stream)); // A_new now stores L_12, we calculate s = L_12 * L_12 RAFT_CUBLAS_TRY( cublasdot(resource::get_cublas_handle(handle), n - 1, A_new, 1, A_new, 1, s, stream)); if (uplo == CUBLAS_FILL_MODE_LOWER) { // Copy back the L_12 elements as the n-th row of L RAFT_CUBLAS_TRY( cublasCopy(resource::get_cublas_handle(handle), n - 1, A_new, 1, A_row, ld, stream)); } } else { // n == 1 case RAFT_CUDA_TRY(cudaMemsetAsync(s, 0, sizeof(math_t), stream)); } // L_22 = sqrt(A_22 - L_12 * L_12) math_t s_host; math_t L_22_host; raft::update_host(&s_host, s, 1, stream); raft::update_host(&L_22_host, L_22, 1, stream); // L_22 stores A_22 resource::sync_stream(handle, stream); L_22_host = std::sqrt(L_22_host - s_host); // Check for numeric error with sqrt. If the matrix is not positive definite or // the system is very ill conditioned then the A_22 - L_12 * L_12 can be // negative, which would result L_22 = NaN. A small positive eps parameter // can be used to prevent this. if (eps >= 0 && (std::isnan(L_22_host) || L_22_host < eps)) { L_22_host = eps; } ASSERT(!std::isnan(L_22_host), "Error during Cholesky rank one update"); raft::update_device(L_22, &L_22_host, 1, stream); } } // namespace detail } // namespace linalg } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/normalize.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/util/cuda_utils.cuh> namespace raft { namespace linalg { namespace detail { template <int warpSize, int rpb> struct NormalizeThinPolicy { static constexpr int LogicalWarpSize = warpSize; static constexpr int RowsPerBlock = rpb; static constexpr int ThreadsPerBlock = LogicalWarpSize * RowsPerBlock; }; template <typename Policy, typename Type, typename IdxType, typename MainLambda, typename ReduceLambda, typename FinalLambda> RAFT_KERNEL __launch_bounds__(Policy::ThreadsPerBlock) coalesced_normalize_thin_kernel(Type* out, const Type* in, IdxType D, IdxType N, Type init, MainLambda main_op, ReduceLambda reduce_op, FinalLambda fin_op, Type eps) { IdxType i = threadIdx.y + (Policy::RowsPerBlock * static_cast<IdxType>(blockIdx.x)); if (i >= N) return; Type acc = init; for (IdxType j = threadIdx.x; j < D; j += Policy::LogicalWarpSize) { Type val = in[j + D * i]; acc = reduce_op(acc, main_op(val, j)); } acc = raft::logicalWarpReduce<Policy::LogicalWarpSize>(acc, reduce_op); acc = fin_op(acc); if (acc <= eps) return; for (IdxType j = threadIdx.x; j < D; j += Policy::LogicalWarpSize) { out[j + D * i] = in[j + D * i] / acc; } } template <typename Policy, typename Type, typename IdxType, typename MainLambda, typename ReduceLambda, typename FinalLambda> inline void coalesced_normalize_thin(Type* out, const Type* in, IdxType D, IdxType N, Type init, cudaStream_t stream, MainLambda main_op, ReduceLambda reduce_op, FinalLambda fin_op, Type eps) { dim3 grid(ceildiv(N, (IdxType)Policy::RowsPerBlock), 1, 1); dim3 block(Policy::LogicalWarpSize, Policy::RowsPerBlock, 1); coalesced_normalize_thin_kernel<Policy> <<<grid, block, 0, stream>>>(out, in, D, N, init, main_op, reduce_op, fin_op, eps); RAFT_CUDA_TRY(cudaPeekAtLastError()); } template <int TPB, typename Type, typename IdxType, typename MainLambda, typename ReduceLambda, typename FinalLambda> RAFT_KERNEL __launch_bounds__(TPB) coalesced_normalize_medium_kernel(Type* out, const Type* in, IdxType D, IdxType N, Type init, MainLambda main_op, ReduceLambda reduce_op, FinalLambda fin_op, Type eps) { typedef cub::BlockReduce<Type, TPB, cub::BLOCK_REDUCE_RAKING> BlockReduce; __shared__ typename BlockReduce::TempStorage temp_storage; __shared__ Type bcast_acc; Type thread_data = init; IdxType rowStart = blockIdx.x * D; for (IdxType i = threadIdx.x; i < D; i += TPB) { IdxType idx = rowStart + i; thread_data = reduce_op(thread_data, main_op(in[idx], i)); } Type acc = BlockReduce(temp_storage).Reduce(thread_data, reduce_op); if (threadIdx.x == 0) { bcast_acc = fin_op(acc); } __syncthreads(); if (bcast_acc <= eps) return; for (IdxType i = threadIdx.x; i < D; i += TPB) { IdxType idx = rowStart + i; out[idx] = in[idx] / bcast_acc; } } template <int TPB, typename Type, typename IdxType, typename MainLambda, typename ReduceLambda, typename FinalLambda> inline void coalesced_normalize_medium(Type* out, const Type* in, IdxType D, IdxType N, Type init, cudaStream_t stream, MainLambda main_op, ReduceLambda reduce_op, FinalLambda fin_op, Type eps) { coalesced_normalize_medium_kernel<TPB> <<<N, TPB, 0, stream>>>(out, in, D, N, init, main_op, reduce_op, fin_op, eps); RAFT_CUDA_TRY(cudaPeekAtLastError()); } template <typename Type, typename IdxType, typename MainLambda, typename ReduceLambda, typename FinalLambda> void coalesced_normalize(Type* out, const Type* in, IdxType D, IdxType N, Type init, cudaStream_t stream, MainLambda main_op, ReduceLambda reduce_op, FinalLambda fin_op, Type eps) { const IdxType numSMs = raft::getMultiProcessorCount(); if (D <= IdxType(256) || (D <= IdxType(512) && N >= 4 * numSMs)) { if (D <= IdxType(2)) { coalesced_normalize_thin<NormalizeThinPolicy<2, 64>>( out, in, D, N, init, stream, main_op, reduce_op, fin_op, eps); } else if (D <= IdxType(4)) { coalesced_normalize_thin<NormalizeThinPolicy<4, 32>>( out, in, D, N, init, stream, main_op, reduce_op, fin_op, eps); } else if (D <= IdxType(8)) { coalesced_normalize_thin<NormalizeThinPolicy<8, 16>>( out, in, D, N, init, stream, main_op, reduce_op, fin_op, eps); } else if (D <= IdxType(16)) { coalesced_normalize_thin<NormalizeThinPolicy<16, 8>>( out, in, D, N, init, stream, main_op, reduce_op, fin_op, eps); } else { coalesced_normalize_thin<NormalizeThinPolicy<32, 4>>( out, in, D, N, init, stream, main_op, reduce_op, fin_op, eps); } } else { coalesced_normalize_medium<256>(out, in, D, N, init, stream, main_op, reduce_op, fin_op, eps); } } } // namespace detail } // namespace linalg } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/qr.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 "cublas_wrappers.hpp" #include "cusolver_wrappers.hpp" #include <raft/core/resource/cusolver_dn_handle.hpp> #include <raft/core/resources.hpp> #include <raft/matrix/triangular.cuh> #include <rmm/device_scalar.hpp> #include <rmm/device_uvector.hpp> #include <algorithm> namespace raft { namespace linalg { namespace detail { /** * @brief Calculate the QR decomposition and get matrix Q in place of the input. * * Subject to the algorithm constraint `n_rows >= n_cols`. * * @param handle * @param[inout] Q device pointer to input matrix and the output matrix Q, * both column-major and of size [n_rows, n_cols]. * @param n_rows * @param n_cols * @param stream */ template <typename math_t> void qrGetQ_inplace( raft::resources const& handle, math_t* Q, int n_rows, int n_cols, cudaStream_t stream) { RAFT_EXPECTS(n_rows >= n_cols, "QR decomposition expects n_rows >= n_cols."); cusolverDnHandle_t cusolver = resource::get_cusolver_dn_handle(handle); rmm::device_uvector<math_t> tau(n_cols, stream); RAFT_CUDA_TRY(cudaMemsetAsync(tau.data(), 0, sizeof(math_t) * n_cols, stream)); rmm::device_scalar<int> dev_info(stream); int ws_size; RAFT_CUSOLVER_TRY(cusolverDngeqrf_bufferSize(cusolver, n_rows, n_cols, Q, n_rows, &ws_size)); rmm::device_uvector<math_t> workspace(ws_size, stream); RAFT_CUSOLVER_TRY(cusolverDngeqrf(cusolver, n_rows, n_cols, Q, n_rows, tau.data(), workspace.data(), ws_size, dev_info.data(), stream)); RAFT_CUSOLVER_TRY( cusolverDnorgqr_bufferSize(cusolver, n_rows, n_cols, n_cols, Q, n_rows, tau.data(), &ws_size)); workspace.resize(ws_size, stream); RAFT_CUSOLVER_TRY(cusolverDnorgqr(cusolver, n_rows, n_cols, n_cols, Q, n_rows, tau.data(), workspace.data(), ws_size, dev_info.data(), stream)); } template <typename math_t> void qrGetQ(raft::resources const& handle, const math_t* M, math_t* Q, int n_rows, int n_cols, cudaStream_t stream) { raft::copy(Q, M, n_rows * n_cols, stream); qrGetQ_inplace(handle, Q, n_rows, n_cols, stream); } template <typename math_t> void qrGetQR(raft::resources const& handle, math_t* M, math_t* Q, math_t* R, int n_rows, int n_cols, cudaStream_t stream) { cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); int m = n_rows, n = n_cols; rmm::device_uvector<math_t> R_full(m * n, stream); rmm::device_uvector<math_t> tau(std::min(m, n), stream); RAFT_CUDA_TRY(cudaMemsetAsync(tau.data(), 0, sizeof(math_t) * std::min(m, n), stream)); int R_full_nrows = m, R_full_ncols = n; RAFT_CUDA_TRY( cudaMemcpyAsync(R_full.data(), M, sizeof(math_t) * m * n, cudaMemcpyDeviceToDevice, stream)); int Lwork; rmm::device_scalar<int> devInfo(stream); RAFT_CUSOLVER_TRY(cusolverDngeqrf_bufferSize( cusolverH, R_full_nrows, R_full_ncols, R_full.data(), R_full_nrows, &Lwork)); rmm::device_uvector<math_t> workspace(Lwork, stream); RAFT_CUSOLVER_TRY(cusolverDngeqrf(cusolverH, R_full_nrows, R_full_ncols, R_full.data(), R_full_nrows, tau.data(), workspace.data(), Lwork, devInfo.data(), stream)); raft::matrix::upper_triangular<math_t, int>( handle, make_device_matrix_view<const math_t, int, col_major>(R_full.data(), m, n), make_device_matrix_view<math_t, int, col_major>(R, std::min(m, n), std::min(m, n))); RAFT_CUDA_TRY( cudaMemcpyAsync(Q, R_full.data(), sizeof(math_t) * m * n, cudaMemcpyDeviceToDevice, stream)); int Q_nrows = m, Q_ncols = n; RAFT_CUSOLVER_TRY(cusolverDnorgqr_bufferSize( cusolverH, Q_nrows, Q_ncols, std::min(Q_ncols, Q_nrows), Q, Q_nrows, tau.data(), &Lwork)); workspace.resize(Lwork, stream); RAFT_CUSOLVER_TRY(cusolverDnorgqr(cusolverH, Q_nrows, Q_ncols, std::min(Q_ncols, Q_nrows), Q, Q_nrows, tau.data(), workspace.data(), Lwork, devInfo.data(), stream)); } }; // namespace detail }; // namespace linalg }; // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/reduce_rows_by_key.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 <raft/util/cuda_utils.cuh> #include <cub/cub.cuh> #include <limits> #define MAX_BLOCKS 65535u namespace raft { namespace linalg { namespace detail { // // Small helper function to convert from int->char and char->int // Transform ncols*nrows read of int in 2*nrows reads of int + ncols*rows reads of chars // template <typename IteratorT1, typename IteratorT2> RAFT_KERNEL convert_array_kernel(IteratorT1 dst, IteratorT2 src, int n) { for (int idx = blockDim.x * blockIdx.x + threadIdx.x; idx < n; idx += gridDim.x * blockDim.x) { dst[idx] = src[idx]; } } // // Small helper function to convert from int->char and char->int // Transform ncols*nrows read of int in 2*nrows reads of int + ncols*rows reads of chars // template <typename IteratorT1, typename IteratorT2> void convert_array(IteratorT1 dst, IteratorT2 src, int n, cudaStream_t st) { dim3 grid, block; block.x = 256; grid.x = raft::ceildiv(n, (int)block.x); grid.x = std::min(grid.x, MAX_BLOCKS); convert_array_kernel<<<grid, block, 0, st>>>(dst, src, n); } template <typename T> struct quad { T x, y, z, w; }; // // Functor for reduce by key, small k // template <typename T> struct quadSum { __host__ __device__ __forceinline__ quad<T> operator()(const quad<T>& a, const quad<T>& b) const { // wasting a double4.. quad<T> c; c.x = a.x + b.x; c.y = a.y + b.y; c.z = a.z + b.z; c.w = a.w + b.w; return c; } }; // // Reduce by keys // We need to sum each dimension by labels // The labels are not adjacent // // // Reduce by keys - for keys <= 4 // #define SUM_ROWS_SMALL_K_DIMX 256 #define SUM_ROWS_BY_KEY_SMALL_K_MAX_K 4 template <typename DataIteratorT, typename WeightT, typename SumsT, typename IdxT> __launch_bounds__(SUM_ROWS_SMALL_K_DIMX, 4) RAFT_KERNEL sum_rows_by_key_small_nkeys_kernel(const DataIteratorT d_A, IdxT lda, const char* d_keys, const WeightT* d_weights, IdxT nrows, IdxT ncols, IdxT nkeys, SumsT* d_sums) { typedef typename std::iterator_traits<DataIteratorT>::value_type DataType; typedef cub::BlockReduce<quad<SumsT>, SUM_ROWS_SMALL_K_DIMX> BlockReduce; __shared__ typename BlockReduce::TempStorage temp_storage; for (IdxT idim = static_cast<IdxT>(blockIdx.y); idim < ncols; idim += gridDim.y) { if (idim != static_cast<IdxT>(blockIdx.y)) __syncthreads(); // we're reusing temp_storage // threadIdx.x stores partial sum for current dim and key=threadIdx.x in this reg quad<SumsT> thread_sums; thread_sums.x = 0.0; thread_sums.y = 0.0; thread_sums.z = 0.0; thread_sums.w = 0.0; // May use vectorized load - not necessary for doubles for (IdxT block_offset_irow = blockIdx.x * blockDim.x; block_offset_irow < nrows; // we will syncthreads() inside the loop, no CTA divergence block_offset_irow += blockDim.x * gridDim.x) { IdxT irow = block_offset_irow + threadIdx.x; DataType val = (irow < nrows) ? d_A[irow * lda + idim] : 0.0; if (d_weights && irow < nrows) { val = val * d_weights[irow]; } // we are not reusing the keys - after profiling // d_keys is mainly loaded from L2, and this kernel is DRAM BW bounded // (experimentation gave a 10% speed up - not worth the many code lines added) IdxT row_key = (irow < nrows) ? d_keys[irow] : std::numeric_limits<IdxT>::max(); thread_sums.x += (row_key == 0) ? static_cast<SumsT>(val) : 0.0; thread_sums.y += (row_key == 1) ? static_cast<SumsT>(val) : 0.0; thread_sums.z += (row_key == 2) ? static_cast<SumsT>(val) : 0.0; thread_sums.w += (row_key == 3) ? static_cast<SumsT>(val) : 0.0; } // End of column // Saving local sums back to global mem // Strided access // Reducing by key thread_sums = BlockReduce(temp_storage).Reduce(thread_sums, quadSum<SumsT>()); if (threadIdx.x < 32) { // We only need 4 thread_sums = cub::ShuffleIndex<32>(thread_sums, 0, 0xffffffff); if (static_cast<IdxT>(threadIdx.x) < nkeys) { if (threadIdx.x == 0) raft::myAtomicAdd(&d_sums[threadIdx.x * ncols + idim], thread_sums.x); if (threadIdx.x == 1) raft::myAtomicAdd(&d_sums[threadIdx.x * ncols + idim], thread_sums.y); if (threadIdx.x == 2) raft::myAtomicAdd(&d_sums[threadIdx.x * ncols + idim], thread_sums.z); if (threadIdx.x == 3) raft::myAtomicAdd(&d_sums[threadIdx.x * ncols + idim], thread_sums.w); } } } } template <typename DataIteratorT, typename WeightT, typename SumsT, typename IdxT> void sum_rows_by_key_small_nkeys(const DataIteratorT d_A, IdxT lda, const char* d_keys, const WeightT* d_weights, IdxT nrows, IdxT ncols, IdxT nkeys, SumsT* d_sums, cudaStream_t st) { dim3 grid, block; block.x = SUM_ROWS_SMALL_K_DIMX; block.y = 1; // Necessary grid.x = raft::ceildiv(nrows, (IdxT)block.x); grid.x = std::min(grid.x, 32u); grid.y = ncols; grid.y = std::min(grid.y, MAX_BLOCKS); sum_rows_by_key_small_nkeys_kernel<<<grid, block, 0, st>>>( d_A, lda, d_keys, d_weights, nrows, ncols, nkeys, d_sums); } // // Reduce by keys - large number of keys // Computing a "weighted histogram" with local histograms in smem // Keeping it simple - not optimized // #define SUM_ROWS_BY_KEY_LARGE_K_MAX_K 1024 template <typename DataIteratorT, typename KeysIteratorT, typename WeightT, typename SumsT, typename IdxT> RAFT_KERNEL sum_rows_by_key_large_nkeys_kernel_colmajor(const DataIteratorT d_A, IdxT lda, KeysIteratorT d_keys, const WeightT* d_weights, IdxT nrows, IdxT ncols, int key_offset, IdxT nkeys, SumsT* d_sums) { typedef typename std::iterator_traits<KeysIteratorT>::value_type KeyType; typedef typename std::iterator_traits<DataIteratorT>::value_type DataType; __shared__ SumsT local_sums[SUM_ROWS_BY_KEY_LARGE_K_MAX_K]; for (IdxT local_key = threadIdx.x; local_key < nkeys; local_key += blockDim.x) local_sums[local_key] = 0.0; for (IdxT idim = blockIdx.y; idim < ncols; idim += gridDim.y) { __syncthreads(); // local_sums // At this point local_sums if full of zeros for (IdxT irow = blockIdx.x * blockDim.x + threadIdx.x; irow < nrows; irow += blockDim.x * gridDim.x) { // Branch div in this loop - not an issue with current code DataType val = d_A[idim * lda + irow]; if (d_weights) val = val * d_weights[irow]; IdxT local_key = d_keys[irow] - key_offset; // We could load next val here raft::myAtomicAdd(&local_sums[local_key], static_cast<SumsT>(val)); } __syncthreads(); // local_sums for (IdxT local_key = threadIdx.x; local_key < nkeys; local_key += blockDim.x) { SumsT local_sum = local_sums[local_key]; if (local_sum != 0.0) { KeyType global_key = key_offset + local_key; raft::myAtomicAdd(&d_sums[global_key * ncols + idim], local_sum); local_sums[local_key] = 0.0; } } } } template <typename DataIteratorT, typename KeysIteratorT, typename SumsT, typename IdxT> void sum_rows_by_key_large_nkeys_colmajor(const DataIteratorT d_A, IdxT lda, KeysIteratorT d_keys, IdxT nrows, IdxT ncols, int key_offset, IdxT nkeys, SumsT* d_sums, cudaStream_t st) { dim3 grid, block; block.x = SUM_ROWS_SMALL_K_DIMX; block.y = 1; // Necessary grid.x = raft::ceildiv(nrows, (IdxT)block.x); grid.x = std::min(grid.x, 32u); grid.y = ncols; grid.y = std::min(grid.y, MAX_BLOCKS); sum_rows_by_key_large_nkeys_kernel_colmajor<<<grid, block, 0, st>>>( d_A, lda, d_keys, nrows, ncols, key_offset, nkeys, d_sums); } template <typename DataIteratorT, typename KeysIteratorT, typename WeightT, typename SumsT, typename IdxT> RAFT_KERNEL sum_rows_by_key_large_nkeys_kernel_rowmajor(const DataIteratorT d_A, IdxT lda, const WeightT* d_weights, KeysIteratorT d_keys, IdxT nrows, IdxT ncols, SumsT* d_sums) { IdxT gid = threadIdx.x + (blockDim.x * static_cast<IdxT>(blockIdx.x)); IdxT j = gid % ncols; IdxT i = gid / ncols; if (i >= nrows) return; IdxT l = static_cast<IdxT>(d_keys[i]); SumsT val = d_A[j + lda * i]; if (d_weights != nullptr) val *= d_weights[i]; raft::myAtomicAdd(&d_sums[j + ncols * l], val); } template <typename DataIteratorT, typename KeysIteratorT, typename WeightT, typename SumsT, typename IdxT> void sum_rows_by_key_large_nkeys_rowmajor(const DataIteratorT d_A, IdxT lda, const KeysIteratorT d_keys, const WeightT* d_weights, IdxT nrows, IdxT ncols, SumsT* d_sums, cudaStream_t st) { uint32_t block_dim = 128; auto grid_dim = static_cast<uint32_t>(ceildiv<IdxT>(nrows * ncols, (IdxT)block_dim)); sum_rows_by_key_large_nkeys_kernel_rowmajor<<<grid_dim, block_dim, 0, st>>>( d_A, lda, d_weights, d_keys, nrows, ncols, d_sums); } /** * @brief Computes the weighted reduction of matrix rows for each given key * * @tparam DataIteratorT Random-access iterator type, for reading input matrix * (may be a simple pointer type) * @tparam KeysIteratorT Random-access iterator type, for reading input keys * (may be a simple pointer type) * @tparam SumsT Type of the output sums * @tparam IdxT Index type * * @param[in] d_A Input data array (lda x nrows) * @param[in] lda Real row size for input data, d_A * @param[in] d_keys Keys for each row (1 x nrows) * @param[in] d_weights Weights for each observation in d_A (1 x nrows) * @param[out] d_keys_char Scratch memory for conversion of keys to char * @param[in] nrows Number of rows in d_A and d_keys * @param[in] ncols Number of data columns in d_A * @param[in] nkeys Number of unique keys in d_keys * @param[out] d_sums Row sums by key (ncols x d_keys) * @param[in] stream CUDA stream * @param[in] reset_sums Whether to reset the output sums to zero before reducing */ template <typename DataIteratorT, typename KeysIteratorT, typename WeightT, typename SumsT, typename IdxT> void reduce_rows_by_key(const DataIteratorT d_A, IdxT lda, KeysIteratorT d_keys, const WeightT* d_weights, char* d_keys_char, IdxT nrows, IdxT ncols, IdxT nkeys, SumsT* d_sums, cudaStream_t stream, bool reset_sums) { typedef typename std::iterator_traits<KeysIteratorT>::value_type KeyType; // Following kernel needs memset if (reset_sums) { cudaMemsetAsync(d_sums, 0, ncols * nkeys * sizeof(SumsT), stream); } if (d_keys_char != nullptr && nkeys <= SUM_ROWS_BY_KEY_SMALL_K_MAX_K) { // sum_rows_by_key_small_k is BW bounded. d_keys is loaded ncols time - avoiding wasting BW // with doubles we have ~20% speed up - with floats we can hope something around 2x // Converting d_keys to char convert_array(d_keys_char, d_keys, nrows, stream); sum_rows_by_key_small_nkeys( d_A, lda, d_keys_char, d_weights, nrows, ncols, nkeys, d_sums, stream); } else { sum_rows_by_key_large_nkeys_rowmajor(d_A, lda, d_keys, d_weights, nrows, ncols, d_sums, stream); } } /** * @brief Computes the reduction of matrix rows for each given key * @tparam DataIteratorT Random-access iterator type, for reading input matrix (may be a simple * pointer type) * @tparam KeysIteratorT Random-access iterator type, for reading input keys (may be a simple * pointer type) * @tparam SumsT Type of the output sums * @tparam IdxT Index type * @param[in] d_A Input data array (lda x nrows) * @param[in] lda Real row size for input data, d_A * @param[in] d_keys Keys for each row (1 x nrows) * @param d_keys_char Scratch memory for conversion of keys to char * @param[in] nrows Number of rows in d_A and d_keys * @param[in] ncols Number of data columns in d_A * @param[in] nkeys Number of unique keys in d_keys * @param[out] d_sums Row sums by key (ncols x d_keys) * @param[in] stream CUDA stream */ template <typename DataIteratorT, typename KeysIteratorT, typename SumsT, typename IdxT> void reduce_rows_by_key(const DataIteratorT d_A, IdxT lda, KeysIteratorT d_keys, char* d_keys_char, IdxT nrows, IdxT ncols, IdxT nkeys, SumsT* d_sums, cudaStream_t stream, bool reset_sums) { typedef typename std::iterator_traits<DataIteratorT>::value_type DataType; reduce_rows_by_key(d_A, lda, d_keys, static_cast<DataType*>(nullptr), d_keys_char, nrows, ncols, nkeys, d_sums, stream, reset_sums); } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/divide.cuh
/* * Copyright (c) 2022, 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/host_mdspan.hpp> #include <raft/core/operators.hpp> #include <raft/linalg/unary_op.cuh> namespace raft { namespace linalg { namespace detail { template <typename InT, typename OutT = InT, typename IdxType = int> void divideScalar(OutT* out, const InT* in, InT scalar, IdxType len, cudaStream_t stream) { raft::linalg::unaryOp(out, in, len, raft::div_const_op<InT>(scalar), stream); } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/transpose.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 "cublas_wrappers.hpp" #include <raft/core/device_mdspan.hpp> #include <raft/core/resource/cublas_handle.hpp> #include <raft/core/resource/cuda_stream.hpp> #include <raft/core/resources.hpp> #include <rmm/exec_policy.hpp> #include <thrust/for_each.h> #include <thrust/iterator/counting_iterator.h> namespace raft { namespace linalg { namespace detail { template <typename math_t> void transpose(raft::resources const& handle, math_t* in, math_t* out, int n_rows, int n_cols, cudaStream_t stream) { cublasHandle_t cublas_h = resource::get_cublas_handle(handle); RAFT_CUBLAS_TRY(cublasSetStream(cublas_h, stream)); int out_n_rows = n_cols; int out_n_cols = n_rows; const math_t alpha = 1.0; const math_t beta = 0.0; RAFT_CUBLAS_TRY(cublasgeam(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N, out_n_rows, out_n_cols, &alpha, in, n_rows, &beta, out, out_n_rows, out, out_n_rows, stream)); } template <typename math_t> void transpose(math_t* inout, int n, cudaStream_t stream) { auto m = n; auto size = n * n; auto d_inout = inout; auto counting = thrust::make_counting_iterator<int>(0); thrust::for_each(rmm::exec_policy(stream), counting, counting + size, [=] __device__(int idx) { int s_row = idx % m; int s_col = idx / m; int d_row = s_col; int d_col = s_row; if (s_row < s_col) { auto temp = d_inout[d_col * m + d_row]; d_inout[d_col * m + d_row] = d_inout[s_col * m + s_row]; d_inout[s_col * m + s_row] = temp; } }); } template <typename T, typename IndexType, typename LayoutPolicy, typename AccessorPolicy> void transpose_row_major_impl( raft::resources const& handle, raft::mdspan<T, raft::matrix_extent<IndexType>, LayoutPolicy, AccessorPolicy> in, raft::mdspan<T, raft::matrix_extent<IndexType>, LayoutPolicy, AccessorPolicy> out) { auto out_n_rows = in.extent(1); auto out_n_cols = in.extent(0); T constexpr kOne = 1; T constexpr kZero = 0; CUBLAS_TRY(cublasgeam(resource::get_cublas_handle(handle), CUBLAS_OP_T, CUBLAS_OP_N, out_n_cols, out_n_rows, &kOne, in.data_handle(), in.stride(0), &kZero, static_cast<T*>(nullptr), out.stride(0), out.data_handle(), out.stride(0), resource::get_cuda_stream(handle))); } template <typename T, typename IndexType, typename LayoutPolicy, typename AccessorPolicy> void transpose_col_major_impl( raft::resources const& handle, raft::mdspan<T, raft::matrix_extent<IndexType>, LayoutPolicy, AccessorPolicy> in, raft::mdspan<T, raft::matrix_extent<IndexType>, LayoutPolicy, AccessorPolicy> out) { auto out_n_rows = in.extent(1); auto out_n_cols = in.extent(0); T constexpr kOne = 1; T constexpr kZero = 0; CUBLAS_TRY(cublasgeam(resource::get_cublas_handle(handle), CUBLAS_OP_T, CUBLAS_OP_N, out_n_rows, out_n_cols, &kOne, in.data_handle(), in.stride(1), &kZero, static_cast<T*>(nullptr), out.stride(1), out.data_handle(), out.stride(1), resource::get_cuda_stream(handle))); } }; // end namespace detail }; // end namespace linalg }; // end namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/add.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/core/operators.hpp> #include <raft/linalg/binary_op.cuh> #include <raft/linalg/unary_op.cuh> #include <raft/util/cuda_utils.cuh> namespace raft { namespace linalg { namespace detail { template <typename InT, typename OutT = InT, typename IdxType = int> void addScalar(OutT* out, const InT* in, InT scalar, IdxType len, cudaStream_t stream) { raft::linalg::unaryOp(out, in, len, raft::add_const_op<InT>(scalar), stream); } template <typename InT, typename OutT = InT, typename IdxType = int> void add(OutT* out, const InT* in1, const InT* in2, IdxType len, cudaStream_t stream) { raft::linalg::binaryOp(out, in1, in2, len, raft::add_op(), stream); } template <class InT, typename IdxType, typename OutT = InT> RAFT_KERNEL add_dev_scalar_kernel(OutT* outDev, const InT* inDev, const InT* singleScalarDev, IdxType len) { IdxType i = ((IdxType)blockIdx.x * (IdxType)blockDim.x) + threadIdx.x; if (i < len) { outDev[i] = inDev[i] + *singleScalarDev; } } template <typename InT, typename OutT = InT, typename IdxType = int> void addDevScalar( OutT* outDev, const InT* inDev, const InT* singleScalarDev, IdxType len, cudaStream_t stream) { // TODO: block dimension has not been tuned dim3 block(256); dim3 grid(raft::ceildiv(len, (IdxType)block.x)); add_dev_scalar_kernel<<<grid, block, 0, stream>>>(outDev, inDev, singleScalarDev, len); RAFT_CUDA_TRY(cudaPeekAtLastError()); } } // namespace detail } // namespace linalg } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/coalesced_reduction-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 <cub/cub.cuh> #include <raft/core/nvtx.hpp> #include <raft/core/operators.hpp> #include <raft/util/cuda_utils.cuh> #include <rmm/device_uvector.hpp> namespace raft { namespace linalg { namespace detail { template <int warpSize, int rpb> struct ReductionThinPolicy { static constexpr int LogicalWarpSize = warpSize; static constexpr int RowsPerBlock = rpb; static constexpr int ThreadsPerBlock = LogicalWarpSize * RowsPerBlock; }; template <typename Policy, typename InType, typename OutType, typename IdxType, typename MainLambda, typename ReduceLambda, typename FinalLambda> RAFT_KERNEL __launch_bounds__(Policy::ThreadsPerBlock) coalescedReductionThinKernel(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, MainLambda main_op, ReduceLambda reduce_op, FinalLambda final_op, bool inplace = false) { IdxType i = threadIdx.y + (Policy::RowsPerBlock * static_cast<IdxType>(blockIdx.x)); if (i >= N) return; OutType acc = init; for (IdxType j = threadIdx.x; j < D; j += Policy::LogicalWarpSize) { acc = reduce_op(acc, main_op(data[j + (D * i)], j)); } acc = raft::logicalWarpReduce<Policy::LogicalWarpSize>(acc, reduce_op); if (threadIdx.x == 0) { if (inplace) { dots[i] = final_op(reduce_op(dots[i], acc)); } else { dots[i] = final_op(acc); } } } template <typename Policy, typename InType, typename OutType = InType, typename IdxType = int, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void coalescedReductionThin(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, cudaStream_t stream, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { common::nvtx::range<common::nvtx::domain::raft> fun_scope( "coalescedReductionThin<%d,%d>", Policy::LogicalWarpSize, Policy::RowsPerBlock); dim3 threads(Policy::LogicalWarpSize, Policy::RowsPerBlock, 1); dim3 blocks(ceildiv<IdxType>(N, Policy::RowsPerBlock), 1, 1); coalescedReductionThinKernel<Policy> <<<blocks, threads, 0, stream>>>(dots, data, D, N, init, main_op, reduce_op, final_op, inplace); RAFT_CUDA_TRY(cudaPeekAtLastError()); } template <typename InType, typename OutType = InType, typename IdxType = int, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void coalescedReductionThinDispatcher(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, cudaStream_t stream, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { if (D <= IdxType(2)) { coalescedReductionThin<ReductionThinPolicy<2, 64>>( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } else if (D <= IdxType(4)) { coalescedReductionThin<ReductionThinPolicy<4, 32>>( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } else if (D <= IdxType(8)) { coalescedReductionThin<ReductionThinPolicy<8, 16>>( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } else if (D <= IdxType(16)) { coalescedReductionThin<ReductionThinPolicy<16, 8>>( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } else { coalescedReductionThin<ReductionThinPolicy<32, 4>>( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } } template <int TPB, typename InType, typename OutType, typename IdxType, typename MainLambda, typename ReduceLambda, typename FinalLambda> RAFT_KERNEL __launch_bounds__(TPB) coalescedReductionMediumKernel(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, MainLambda main_op, ReduceLambda reduce_op, FinalLambda final_op, bool inplace = false) { typedef cub::BlockReduce<OutType, TPB, cub::BLOCK_REDUCE_RAKING> BlockReduce; __shared__ typename BlockReduce::TempStorage temp_storage; OutType thread_data = init; IdxType rowStart = blockIdx.x * D; for (IdxType i = threadIdx.x; i < D; i += TPB) { IdxType idx = rowStart + i; thread_data = reduce_op(thread_data, main_op(data[idx], i)); } OutType acc = BlockReduce(temp_storage).Reduce(thread_data, reduce_op); if (threadIdx.x == 0) { if (inplace) { dots[blockIdx.x] = final_op(reduce_op(dots[blockIdx.x], acc)); } else { dots[blockIdx.x] = final_op(acc); } } } template <int TPB, typename InType, typename OutType = InType, typename IdxType = int, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void coalescedReductionMedium(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, cudaStream_t stream, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { common::nvtx::range<common::nvtx::domain::raft> fun_scope("coalescedReductionMedium<%d>", TPB); coalescedReductionMediumKernel<TPB> <<<N, TPB, 0, stream>>>(dots, data, D, N, init, main_op, reduce_op, final_op, inplace); RAFT_CUDA_TRY(cudaPeekAtLastError()); } template <typename InType, typename OutType = InType, typename IdxType = int, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void coalescedReductionMediumDispatcher(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, cudaStream_t stream, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { // Note: for now, this kernel is only used when D > 256. If this changes in the future, use // smaller block sizes when relevant. coalescedReductionMedium<256>( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } template <int tpb, int bpr> struct ReductionThickPolicy { static constexpr int ThreadsPerBlock = tpb; static constexpr int BlocksPerRow = bpr; static constexpr int BlockStride = tpb * bpr; }; template <typename Policy, typename InType, typename OutType, typename IdxType, typename MainLambda, typename ReduceLambda> RAFT_KERNEL __launch_bounds__(Policy::ThreadsPerBlock) coalescedReductionThickKernel(OutType* buffer, const InType* data, IdxType D, IdxType N, OutType init, MainLambda main_op, ReduceLambda reduce_op) { typedef cub::BlockReduce<OutType, Policy::ThreadsPerBlock, cub::BLOCK_REDUCE_RAKING> BlockReduce; __shared__ typename BlockReduce::TempStorage temp_storage; OutType thread_data = init; IdxType rowStart = blockIdx.x * D; for (IdxType i = blockIdx.y * Policy::ThreadsPerBlock + threadIdx.x; i < D; i += Policy::BlockStride) { IdxType idx = rowStart + i; thread_data = reduce_op(thread_data, main_op(data[idx], i)); } OutType acc = BlockReduce(temp_storage).Reduce(thread_data, reduce_op); if (threadIdx.x == 0) { buffer[Policy::BlocksPerRow * blockIdx.x + blockIdx.y] = acc; } } template <typename ThickPolicy, typename ThinPolicy, typename InType, typename OutType = InType, typename IdxType = int, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void coalescedReductionThick(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, cudaStream_t stream, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { common::nvtx::range<common::nvtx::domain::raft> fun_scope( "coalescedReductionThick<%d,%d>", ThickPolicy::ThreadsPerBlock, ThickPolicy::BlocksPerRow); dim3 threads(ThickPolicy::ThreadsPerBlock, 1, 1); dim3 blocks(N, ThickPolicy::BlocksPerRow, 1); rmm::device_uvector<OutType> buffer(N * ThickPolicy::BlocksPerRow, stream); /* We apply a two-step reduction: * 1. coalescedReductionThickKernel reduces the [N x D] input data to [N x BlocksPerRow]. It * applies the main_op but not the final op. * 2. coalescedReductionThinKernel reduces [N x BlocksPerRow] to [N x 1]. It doesn't apply any * main_op but applies final_op. If in-place, the existing and new values are reduced. */ coalescedReductionThickKernel<ThickPolicy> <<<blocks, threads, 0, stream>>>(buffer.data(), data, D, N, init, main_op, reduce_op); RAFT_CUDA_TRY(cudaPeekAtLastError()); coalescedReductionThin<ThinPolicy>(dots, buffer.data(), static_cast<IdxType>(ThickPolicy::BlocksPerRow), N, init, stream, inplace, raft::identity_op(), reduce_op, final_op); } template <typename InType, typename OutType = InType, typename IdxType = int, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void coalescedReductionThickDispatcher(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, cudaStream_t stream, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { // Note: multiple elements per thread to take advantage of the sequential reduction and loop // unrolling if (D < IdxType(32768)) { coalescedReductionThick<ReductionThickPolicy<256, 32>, ReductionThinPolicy<32, 4>>( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } else { coalescedReductionThick<ReductionThickPolicy<256, 64>, ReductionThinPolicy<32, 4>>( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } } // Primitive to perform reductions along the coalesced dimension of the matrix, i.e. reduce along // rows for row major or reduce along columns for column major layout. Can do an inplace reduction // adding to original values of dots if requested. template <typename InType, typename OutType = InType, typename IdxType = int, typename MainLambda = raft::identity_op, typename ReduceLambda = raft::add_op, typename FinalLambda = raft::identity_op> void coalescedReduction(OutType* dots, const InType* data, IdxType D, IdxType N, OutType init, cudaStream_t stream, bool inplace = false, MainLambda main_op = raft::identity_op(), ReduceLambda reduce_op = raft::add_op(), FinalLambda final_op = raft::identity_op()) { /* The primitive selects one of three implementations based on heuristics: * - Thin: very efficient when D is small and/or N is large * - Thick: used when N is very small and D very large * - Medium: used when N is too small to fill the GPU with the thin kernel */ const IdxType numSMs = raft::getMultiProcessorCount(); if (D <= IdxType(256) || N >= IdxType(4) * numSMs) { coalescedReductionThinDispatcher( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } else if (N < numSMs && D >= IdxType(16384)) { coalescedReductionThickDispatcher( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } else { coalescedReductionMediumDispatcher( dots, data, D, N, init, stream, inplace, main_op, reduce_op, final_op); } } } // namespace detail } // namespace linalg } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/gemm.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 <cublas_v2.h> #include "cublas_wrappers.hpp" #include <raft/core/resource/cublas_handle.hpp> #include <raft/core/resources.hpp> namespace raft { namespace linalg { namespace detail { /** * @brief the wrapper of cublas gemm function * It computes the following equation: C = alpha .* opA(A) * opB(B) + beta .* C * * @tparam math_t the element type * @tparam DevicePointerMode whether pointers alpha, beta point to device memory * @param [in] handle raft handle * @param [in] trans_a cublas transpose op for A * @param [in] trans_b cublas transpose op for B * @param [in] m number of rows of C * @param [in] n number of columns of C * @param [in] k number of rows of opB(B) / number of columns of opA(A) * @param [in] alpha host or device scalar * @param [in] A such a matrix that the shape of column-major opA(A) is [m, k] * @param [in] lda leading dimension of A * @param [in] B such a matrix that the shape of column-major opA(B) is [k, n] * @param [in] ldb leading dimension of B * @param [in] beta host or device scalar * @param [inout] C column-major matrix of size [m, n] * @param [in] ldc leading dimension of C * @param [in] stream */ template <typename math_t, bool DevicePointerMode = false> void gemm(raft::resources const& handle, const bool trans_a, const bool trans_b, const int m, const int n, const int k, const math_t* alpha, const math_t* A, const int lda, const math_t* B, const int ldb, const math_t* beta, math_t* C, const int ldc, cudaStream_t stream) { auto cublas_h = raft::resource::get_cublas_handle(handle); cublas_device_pointer_mode<DevicePointerMode> pmode(cublas_h); RAFT_CUBLAS_TRY(cublasgemm(cublas_h, trans_a ? CUBLAS_OP_T : CUBLAS_OP_N, trans_b ? CUBLAS_OP_T : CUBLAS_OP_N, m, n, k, alpha, A, lda, B, ldb, beta, C, ldc, stream)); } /** * @brief the wrapper of cublas gemm function * It computes the following equation: D = alpha . opA(A) * opB(B) + beta . C * @tparam math_t the type of input/output matrices * @param handle raft handle * @param a input matrix * @param n_rows_a number of rows of A * @param n_cols_a number of columns of A * @param b input matrix * @param c output matrix * @param n_rows_c number of rows of C * @param n_cols_c number of columns of C * @param trans_a cublas transpose op for A * @param trans_b cublas transpose op for B * @param alpha scalar * @param beta scalar * @param stream cuda stream */ template <typename math_t> void gemm(raft::resources const& handle, const math_t* a, int n_rows_a, int n_cols_a, const math_t* b, math_t* c, int n_rows_c, int n_cols_c, cublasOperation_t trans_a, cublasOperation_t trans_b, math_t alpha, math_t beta, cudaStream_t stream) { auto cublas_h = raft::resource::get_cublas_handle(handle); int m = n_rows_c; int n = n_cols_c; int k = trans_a == CUBLAS_OP_T ? n_rows_a : n_cols_a; int lda = trans_a == CUBLAS_OP_T ? k : m; int ldb = trans_b == CUBLAS_OP_T ? n : k; int ldc = m; RAFT_CUBLAS_TRY( cublasgemm(cublas_h, trans_a, trans_b, m, n, k, &alpha, a, lda, b, ldb, &beta, c, ldc, stream)); } template <typename math_t> void gemm(raft::resources const& handle, const math_t* a, int n_rows_a, int n_cols_a, const math_t* b, math_t* c, int n_rows_c, int n_cols_c, cublasOperation_t trans_a, cublasOperation_t trans_b, cudaStream_t stream) { math_t alpha = math_t(1); math_t beta = math_t(0); gemm( handle, a, n_rows_a, n_cols_a, b, c, n_rows_c, n_cols_c, trans_a, trans_b, alpha, beta, stream); } template <typename T, bool DevicePointerMode = false> void gemm(raft::resources const& handle, T* z, T* x, T* y, int _M, int _N, int _K, bool isZColMajor, bool isXColMajor, bool isYColMajor, cudaStream_t stream, T* alpha, T* beta) { auto cublas_h = raft::resource::get_cublas_handle(handle); cublas_device_pointer_mode<DevicePointerMode> pmode(cublas_h); cublasOperation_t trans_a, trans_b; T *a, *b, *c; int lda, ldb, ldc; int M, N, K; // This function performs c = a * b. Based on the required output layout, // either a = x, b = y or a = y, b = x. In either case c = z. if (isZColMajor == true) { // Result c is required in column major layout. Thus we perform, // z = x * y // Using BLAS call c = a * b. Therefore a = x, b = y and c = z a = x; // If x is in row major layout, cublas needs to transpose x first, // therefore trans_x needs to be CUBLAS_OP_T. If x is in column major // layout, trans_b needs to be CUBLAS_OP_N. trans_a = isXColMajor == true ? CUBLAS_OP_N : CUBLAS_OP_T; // Set leading dimension appropriately lda = isXColMajor == true ? _M : _K; b = y; // If y is in row major layout, cublas needs to transpose y first, // therefore trans_x needs to be CUBLAS_OP_T. If x is in column major // layout, trans_b needs to be CUBLAS_OP_N. trans_b = isYColMajor == true ? CUBLAS_OP_N : CUBLAS_OP_T; ldb = isYColMajor == true ? _K : _N; c = z; ldc = _M; M = _M; N = _N; K = _K; } else { // Result c is required in row major layout Thus we pick // a = y, b = x and c = a * b = y * x // cublas produces output matrix only in column major layout. To get output // matrix on row major layout, we need to produce transpose of output // in column major layout. Therefore we perform, // tr(z) = tr(y) * tr(x) // we model this using cublas call for c = a * b // therefore a = tr(y), b = tr(x) and c = tr(z) a = y; // If y is in row major layout, it can be/ interpreted as tr(y) on column // major layout. Therefore we can pass trans_a as CUBLAS_OP_N. If y is in // column major layout, cublas needs to transpose y first, therefore // trans_a needs to be CUBLAS_OP_T trans_a = isYColMajor == true ? CUBLAS_OP_T : CUBLAS_OP_N; // Set leading dimension appropriately lda = isYColMajor == true ? _K : _N; b = x; // If x is in row major layout, it can be interpreted as tr(x) on column // major layout. Therefore we can pass trans_b as CUBLAS_OP_N. If x is in // column major layout, cublas needs to trasponse x first, therefore // trans_b needs to be CUBLAS_OP_T trans_b = isXColMajor == true ? CUBLAS_OP_T : CUBLAS_OP_N; // Set leading dimension appropriately ldb = isXColMajor == true ? _M : _K; c = z; ldc = _N; M = _N; N = _M; K = _K; } // Actual cuBLAS call RAFT_CUBLAS_TRY( cublasgemm(cublas_h, trans_a, trans_b, M, N, K, alpha, a, lda, b, ldb, beta, c, ldc, stream)); } } // namespace detail } // namespace linalg } // namespace raft
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/coalesced_reduction-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 <raft/core/operators.hpp> // The explicit instantiation of raft::linalg::detail::coalescedReduction is not // forced because there would be too many instances. Instead, we cover the most // common instantiations with extern template instantiations below. #define instantiate_raft_linalg_detail_coalescedReduction( \ InType, OutType, IdxType, MainLambda, ReduceLambda, FinalLambda) \ extern template void raft::linalg::detail::coalescedReduction(OutType* dots, \ const InType* data, \ IdxType D, \ IdxType N, \ OutType init, \ cudaStream_t stream, \ bool inplace, \ MainLambda main_op, \ ReduceLambda reduce_op, \ FinalLambda final_op) instantiate_raft_linalg_detail_coalescedReduction( double, double, int, raft::identity_op, raft::min_op, raft::identity_op); instantiate_raft_linalg_detail_coalescedReduction( double, double, int, raft::sq_op, raft::add_op, raft::identity_op); instantiate_raft_linalg_detail_coalescedReduction( double, double, int, raft::sq_op, raft::add_op, raft::sqrt_op); instantiate_raft_linalg_detail_coalescedReduction( double, double, int, raft::abs_op, raft::add_op, raft::identity_op); instantiate_raft_linalg_detail_coalescedReduction( double, double, int, raft::abs_op, raft::max_op, raft::identity_op); instantiate_raft_linalg_detail_coalescedReduction( float, float, size_t, raft::abs_op, raft::add_op, raft::sqrt_op); instantiate_raft_linalg_detail_coalescedReduction( float, float, int, raft::abs_op, raft::add_op, raft::identity_op); instantiate_raft_linalg_detail_coalescedReduction( float, float, int, raft::identity_op, raft::add_op, raft::identity_op); instantiate_raft_linalg_detail_coalescedReduction( float, float, int, raft::identity_op, raft::min_op, raft::identity_op); instantiate_raft_linalg_detail_coalescedReduction( float, float, int, raft::sq_op, raft::add_op, raft::identity_op); instantiate_raft_linalg_detail_coalescedReduction( float, float, int, raft::sq_op, raft::add_op, raft::sqrt_op); instantiate_raft_linalg_detail_coalescedReduction( float, float, long, raft::sq_op, raft::add_op, raft::identity_op); instantiate_raft_linalg_detail_coalescedReduction( float, float, size_t, raft::identity_op, raft::add_op, raft::identity_op); instantiate_raft_linalg_detail_coalescedReduction( float, float, size_t, raft::sq_op, raft::add_op, raft::identity_op); instantiate_raft_linalg_detail_coalescedReduction( float, float, size_t, raft::abs_op, raft::max_op, raft::sqrt_op); instantiate_raft_linalg_detail_coalescedReduction( float, float, size_t, raft::sq_op, raft::add_op, raft::sqrt_op); instantiate_raft_linalg_detail_coalescedReduction( float, float, unsigned int, raft::sq_op, raft::add_op, raft::identity_op); #undef instantiate_raft_linalg_detail_coalescedReduction
0
rapidsai_public_repos/raft/cpp/include/raft/linalg
rapidsai_public_repos/raft/cpp/include/raft/linalg/detail/lstsq.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 <common/nvtx.hpp> #include <raft/common/nvtx.hpp> #include <raft/core/resource/cublas_handle.hpp> #include <raft/core/resource/cuda_stream_pool.hpp> #include <raft/core/resource/cusolver_dn_handle.hpp> #include <raft/linalg/detail/cublas_wrappers.hpp> #include <raft/linalg/detail/cusolver_wrappers.hpp> #include <raft/linalg/eig.cuh> #include <raft/linalg/eltwise.cuh> #include <raft/linalg/gemm.cuh> #include <raft/linalg/gemv.cuh> #include <raft/linalg/qr.cuh> #include <raft/linalg/svd.cuh> #include <raft/linalg/transpose.cuh> #include <raft/matrix/math.cuh> #include <raft/util/cudart_utils.hpp> #include <rmm/cuda_stream_view.hpp> #include <rmm/device_scalar.hpp> #include <rmm/device_uvector.hpp> namespace raft { namespace linalg { namespace detail { namespace { /** Operate a CUDA event if we're in the concurrent mode; no-op otherwise. */ struct DeviceEvent { private: cudaEvent_t e; public: DeviceEvent(bool concurrent) { if (concurrent) RAFT_CUDA_TRY(cudaEventCreateWithFlags(&e, cudaEventDisableTiming)); else e = nullptr; } ~DeviceEvent() { if (e != nullptr) RAFT_CUDA_TRY_NO_THROW(cudaEventDestroy(e)); } void record(cudaStream_t stream) { if (e != nullptr) RAFT_CUDA_TRY(cudaEventRecord(e, stream)); } void wait_by(cudaStream_t stream) { if (e != nullptr) RAFT_CUDA_TRY(cudaStreamWaitEvent(stream, e, 0u)); } DeviceEvent& operator=(const DeviceEvent& other) = delete; }; /** * @brief Tells if the viewed CUDA stream is implicitly synchronized with the given stream. * * This can happen e.g. * if the two views point to the same stream * or sometimes when one of them is the legacy default stream. */ bool are_implicitly_synchronized(rmm::cuda_stream_view a, rmm::cuda_stream_view b) { // any stream is "synchronized" with itself if (a.value() == b.value()) return true; // legacy + blocking streams unsigned int flags = 0; if (a.is_default()) { RAFT_CUDA_TRY(cudaStreamGetFlags(b.value(), &flags)); if ((flags & cudaStreamNonBlocking) == 0) return true; } if (b.is_default()) { RAFT_CUDA_TRY(cudaStreamGetFlags(a.value(), &flags)); if ((flags & cudaStreamNonBlocking) == 0) return true; } return false; } template <typename math_t> struct DivideByNonZero { constexpr static const math_t eps = math_t(1e-10); __device__ math_t operator()(const math_t a, const math_t b) const { return raft::abs<math_t>(b) >= eps ? a / b : a; } }; } // namespace /** Solves the linear ordinary least squares problem `Aw = b` * Via SVD decomposition of `A = U S Vt` using default cuSOLVER routine. * * @param A - input feature matrix; it's marked [in/out] in the used cuSOLVER routines, * so it's not guaranteed to stay unmodified. */ template <typename math_t> void lstsqSvdQR(raft::resources const& handle, math_t* A, const int n_rows, const int n_cols, const math_t* b, math_t* w, cudaStream_t stream) { const int minmn = min(n_rows, n_cols); cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); int cusolverWorkSetSize = 0; // #TODO: Call from public API when ready RAFT_CUSOLVER_TRY(raft::linalg::detail::cusolverDngesvd_bufferSize<math_t>( cusolverH, n_rows, n_cols, &cusolverWorkSetSize)); rmm::device_uvector<math_t> workset(cusolverWorkSetSize // cuSolver + n_rows * minmn // U + n_cols * n_cols // V + minmn // S + minmn // U^T * b + 1 // devInfo , stream); math_t* cusolverWorkSet = workset.data(); math_t* U = cusolverWorkSet + cusolverWorkSetSize; math_t* Vt = U + n_rows * minmn; math_t* S = Vt + n_cols * n_cols; math_t* Ub = S + minmn; int* devInfo = reinterpret_cast<int*>(Ub + minmn); // #TODO: Call from public API when ready RAFT_CUSOLVER_TRY(raft::linalg::detail::cusolverDngesvd<math_t>(cusolverH, 'S', 'S', n_rows, n_cols, A, n_rows, S, U, n_rows, Vt, n_cols, cusolverWorkSet, cusolverWorkSetSize, nullptr, devInfo, stream)); raft::linalg::gemv(handle, U, n_rows, minmn, b, Ub, true, stream); raft::linalg::binaryOp(Ub, Ub, S, minmn, DivideByNonZero<math_t>(), stream); raft::linalg::gemv(handle, Vt, minmn, n_cols, n_cols, Ub, w, true, stream); } /** Solves the linear ordinary least squares problem `Aw = b` * Via SVD decomposition of `A = U S V^T` using Jacobi iterations (cuSOLVER). * * @param A - input feature matrix; it's marked [in/out] in the used cuSOLVER routines, * so it's not guaranteed to stay unmodified. */ template <typename math_t> void lstsqSvdJacobi(raft::resources const& handle, math_t* A, const int n_rows, const int n_cols, const math_t* b, math_t* w, cudaStream_t stream) { const int minmn = min(n_rows, n_cols); gesvdjInfo_t gesvdj_params; RAFT_CUSOLVER_TRY(cusolverDnCreateGesvdjInfo(&gesvdj_params)); int cusolverWorkSetSize = 0; cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); // #TODO: Call from public API when ready RAFT_CUSOLVER_TRY( raft::linalg::detail::cusolverDngesvdj_bufferSize<math_t>(cusolverH, CUSOLVER_EIG_MODE_VECTOR, 1, n_rows, n_cols, A, n_rows, nullptr, nullptr, n_rows, nullptr, n_cols, &cusolverWorkSetSize, gesvdj_params)); rmm::device_uvector<math_t> workset(cusolverWorkSetSize // cuSolver + n_rows * minmn // U + n_cols * minmn // V + minmn // S + minmn // U^T * b + 1 // devInfo , stream); math_t* cusolverWorkSet = workset.data(); math_t* U = cusolverWorkSet + cusolverWorkSetSize; math_t* V = U + n_rows * minmn; math_t* S = V + n_cols * minmn; math_t* Ub = S + minmn; int* devInfo = reinterpret_cast<int*>(Ub + minmn); // #TODO: Call from public API when ready RAFT_CUSOLVER_TRY(raft::linalg::detail::cusolverDngesvdj<math_t>(cusolverH, CUSOLVER_EIG_MODE_VECTOR, 1, n_rows, n_cols, A, n_rows, S, U, n_rows, V, n_cols, cusolverWorkSet, cusolverWorkSetSize, devInfo, gesvdj_params, stream)); raft::linalg::gemv(handle, U, n_rows, minmn, b, Ub, true, stream); raft::linalg::binaryOp(Ub, Ub, S, minmn, DivideByNonZero<math_t>(), stream); raft::linalg::gemv(handle, V, n_cols, minmn, Ub, w, false, stream); } /** Solves the linear ordinary least squares problem `Aw = b` * via eigenvalue decomposition of `A^T * A` (covariance matrix for dataset A). * (`w = (A^T A)^-1 A^T b`) */ template <typename math_t> void lstsqEig(raft::resources const& handle, const math_t* A, const int n_rows, const int n_cols, const math_t* b, math_t* w, cudaStream_t stream) { rmm::cuda_stream_view mainStream = rmm::cuda_stream_view(stream); rmm::cuda_stream_view multAbStream = resource::get_next_usable_stream(handle); bool concurrent; // Check if the two streams can run concurrently. This is needed because a legacy default stream // would synchronize with other blocking streams. To avoid synchronization in such case, we try to // use an additional stream from the pool. if (!are_implicitly_synchronized(mainStream, multAbStream)) { concurrent = true; } else if (resource::get_stream_pool_size(handle) > 1) { mainStream = resource::get_next_usable_stream(handle); concurrent = true; } else { multAbStream = mainStream; concurrent = false; } rmm::device_uvector<math_t> workset(n_cols * n_cols * 3 + n_cols * 2, mainStream); // the event is created only if the given raft handle is capable of running // at least two CUDA streams without implicit synchronization. DeviceEvent worksetDone(concurrent); worksetDone.record(mainStream); math_t* Q = workset.data(); math_t* QS = Q + n_cols * n_cols; math_t* covA = QS + n_cols * n_cols; math_t* S = covA + n_cols * n_cols; math_t* Ab = S + n_cols; // covA <- A* A math_t alpha = math_t(1); math_t beta = math_t(0); raft::linalg::gemm(handle, A, n_rows, n_cols, A, covA, n_cols, n_cols, CUBLAS_OP_T, CUBLAS_OP_N, alpha, beta, mainStream); // Ab <- A* b worksetDone.wait_by(multAbStream); raft::linalg::gemv(handle, A, n_rows, n_cols, b, Ab, true, multAbStream); DeviceEvent multAbDone(concurrent); multAbDone.record(multAbStream); // Q S Q* <- covA raft::common::nvtx::push_range("raft::linalg::eigDC"); raft::linalg::eigDC(handle, covA, n_cols, n_cols, Q, S, mainStream); raft::common::nvtx::pop_range(); // QS <- Q invS raft::linalg::matrixVectorOp( QS, Q, S, n_cols, n_cols, false, true, DivideByNonZero<math_t>(), mainStream); // covA <- QS Q* == Q invS Q* == inv(A* A) raft::linalg::gemm(handle, QS, n_cols, n_cols, Q, covA, n_cols, n_cols, CUBLAS_OP_N, CUBLAS_OP_T, alpha, beta, mainStream); multAbDone.wait_by(mainStream); // w <- covA Ab == Q invS Q* A b == inv(A* A) A b raft::linalg::gemv(handle, covA, n_cols, n_cols, Ab, w, false, mainStream); // This event is created only if we use two worker streams, and `stream` is not the legacy stream, // and `mainStream` is not a non-blocking stream. In fact, with the current logic these conditions // are impossible together, but it still makes sense to put this construct here to emphasize that // `stream` must wait till the work here is done (for future refactorings). DeviceEvent mainDone(!are_implicitly_synchronized(mainStream, stream)); mainDone.record(mainStream); mainDone.wait_by(stream); } /** Solves the linear ordinary least squares problem `Aw = b` * via QR decomposition of `A = QR`. * (triangular system of equations `Rw = Q^T b`) * * @param A[in/out] - input feature matrix. * Warning: the content of this matrix is modified by the cuSOLVER routines. * @param b[in/out] - input target vector. * Warning: the content of this vector is modified by the cuSOLVER routines. */ template <typename math_t> void lstsqQR(raft::resources const& handle, math_t* A, const int n_rows, const int n_cols, math_t* b, math_t* w, cudaStream_t stream) { cublasHandle_t cublasH = resource::get_cublas_handle(handle); cusolverDnHandle_t cusolverH = resource::get_cusolver_dn_handle(handle); int m = n_rows; int n = n_cols; int info = 0; rmm::device_uvector<math_t> d_tau(n, stream); rmm::device_scalar<int> d_info(stream); const cublasSideMode_t side = CUBLAS_SIDE_LEFT; const cublasOperation_t trans = CUBLAS_OP_T; int lwork_geqrf = 0; int lwork_ormqr = 0; int lwork = 0; const int lda = m; const int ldb = m; // #TODO: Call from public API when ready RAFT_CUSOLVER_TRY( raft::linalg::detail::cusolverDngeqrf_bufferSize(cusolverH, m, n, A, lda, &lwork_geqrf)); // #TODO: Call from public API when ready RAFT_CUSOLVER_TRY(raft::linalg::detail::cusolverDnormqr_bufferSize(cusolverH, side, trans, m, 1, n, A, lda, d_tau.data(), b, // C, lda, // ldc, &lwork_ormqr)); lwork = (lwork_geqrf > lwork_ormqr) ? lwork_geqrf : lwork_ormqr; rmm::device_uvector<math_t> d_work(lwork, stream); // #TODO: Call from public API when ready RAFT_CUSOLVER_TRY(raft::linalg::detail::cusolverDngeqrf( cusolverH, m, n, A, lda, d_tau.data(), d_work.data(), lwork, d_info.data(), stream)); RAFT_CUDA_TRY(cudaMemcpyAsync(&info, d_info.data(), sizeof(int), cudaMemcpyDeviceToHost, stream)); RAFT_CUDA_TRY(cudaStreamSynchronize(stream)); ASSERT(0 == info, "lstsq.h: QR wasn't successful"); // #TODO: Call from public API when ready RAFT_CUSOLVER_TRY(raft::linalg::detail::cusolverDnormqr(cusolverH, side, trans, m, 1, n, A, lda, d_tau.data(), b, ldb, d_work.data(), lwork, d_info.data(), stream)); RAFT_CUDA_TRY(cudaMemcpyAsync(&info, d_info.data(), sizeof(int), cudaMemcpyDeviceToHost, stream)); RAFT_CUDA_TRY(cudaStreamSynchronize(stream)); ASSERT(0 == info, "lstsq.h: QR wasn't successful"); const math_t one = 1; // #TODO: Call from public API when ready RAFT_CUBLAS_TRY(raft::linalg::detail::cublastrsm(cublasH, side, CUBLAS_FILL_MODE_UPPER, CUBLAS_OP_N, CUBLAS_DIAG_NON_UNIT, n, 1, &one, A, lda, b, ldb, stream)); RAFT_CUDA_TRY(cudaMemcpyAsync(w, b, sizeof(math_t) * n, cudaMemcpyDeviceToDevice, stream)); } }; // namespace detail }; // namespace linalg }; // namespace raft
0