File size: 44,003 Bytes

838f737

/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under the BSD-style license found in the
 * LICENSE file in the root directory of this source tree.
 */

#pragma once

/**
 * Top level include file for FBGEMM.
 */
#include <cassert>
#include <memory>
#include "./ConvUtils.h" // @manual
#include "./FbgemmBuild.h" // @manual
#include "./FbgemmEmbedding.h" // @manual
#include "./FbgemmI8DepthwiseAvx2.h" // @manual
#include "./FbgemmI8DirectconvAvx2.h" // @manual
#include "./FbgemmI8Spmdm.h" // @manual
#include "./FloatConversion.h" // @manual
#include "./QuantUtilsAvx2.h" // @manual
#include "./Types.h" // @manual
#include "./Utils.h" // @manual

// Turning on this option will print out time breakdown of each stage (e.g.,
// input packing, the main GEMM kernel, each output processing pipeline).
// Please note that currently this option won't report accurate timing if
// multiple threads are used.
// #define FBGEMM_MEASURE_TIME_BREAKDOWN

#ifdef FBGEMM_MEASURE_TIME_BREAKDOWN
#include <chrono>
#include <iostream>
extern double packing_time;
extern double computing_time;
extern double kernel_time;
extern double postprocessing_time;
extern double run_time;
#endif

namespace fbgemm {

/**
 * @brief Templatized struct for packing parameters for A and B matrices.
 *
 * @tparam T input type
 * @tparam accT the type used for accumulation
 * @tparam instSet anyarch/avx2/avx512
 * @tparam int8Type an auxiliary template parameter to specialize for 8-bit
 *                  input types.
 */
template <
    typename T,
    typename accT,
    inst_set_t instSet,
    typename int8Type = void>
struct PackingTraits;

// type specialized implementation in an include file
#include "./PackingTraits-inl.h" // @manual

/**
 * @brief Base class for packing matrices for higher GEMM performance.
 *
 * Matrix is tiled into blockRows() * blockCols() blocks.
 * Each block is with size blockRowSize() * blockColSize().
 * This class is designed using CRTP
 * (https://en.wikipedia.org/wiki/Curiously_recurring_template_pattern)
 *
 * @tparam PT actual packing type, e.g., PackAWithRowOffset
 */
template <typename PT, typename inpType, typename accType = std::int32_t>
class PackMatrix {
 public:
  PackMatrix() = delete; // no default constructor
  PackMatrix(const PackMatrix&) = delete; // no copy
  PackMatrix& operator=(const PackMatrix&) = delete; // no copy
  PackMatrix(PackMatrix&&) = delete; // no move
  PackMatrix& operator=(PackMatrix&& rhs) noexcept = delete; // no move

  /**
   * @param rows total number of rows in the matrix
   *             (packed rows can be less than rows).
   * @param cols total number of columns in the matrix
   * @param pmat A buffer to contain the packed matrix.
   *             If nullptr, a buffer owned by PackMatrix will be allocated
   *             internally to contain the packed matrix.
   *             For non-constant matrices like activation matrices, the client
   *             code may want to pass a pre-allocated pmat to avoid the
   *             overhead of internal memory allocation everytime a PackMatrix
   *             is constructed. The client code can query how big patm should
   *             be with packedBufferSize function.
   * @param groups when groups > 1, we compute groups number of GEMMs each
   *               multiplies A.rows by A.cols/A.groups matrix with
   *               B.rows/B.groups by B.cols matrix (in conventional BLAS
   *               terminology, this is a batched GEMM but we use the name group
   *               to follow deep learning terminology). The result matrix has
   *               dimension A.rows by B.cols*B.groups .
   *               A.groups must be same as B.groups, A.groups must divide
   *               A.cols, and B.groups must divide B.rows and C.cols.
   */
  PackMatrix(
      std::int32_t rows,
      std::int32_t cols,
      inpType* pmat,
      int groups = 1,
      const BlockingFactors* params = nullptr);

  /**
   * @return true usually when the matrix is constant matrix (e.g., weight
   *         matrices) that can be prepacked
   */
  bool isPrePacked() const {
    return static_cast<const PT*>(this)->isPrePacked();
  }

  /**
   * @return true if this is the first input matrix in GEMM (i.e., A in C = A *
   *         B)
   */
  static bool isA() {
    return PT::isA();
  }

  /**
   * @brief The size of the buffer used for packing (The size is in number of
   *        elements).
   *
   * rows and cols are only used for fully packing, i.e., for B matrix.  The
   * client code can use this function to query how big the buffer used for
   * packing should be.
   */
  static int packedBufferSize(
      int rows = 0,
      int cols = 0,
      const BlockingFactors* params = nullptr);

  FBGEMM_PUSH_WARNING_AND_DISABLE("-Wpragmas")
  FBGEMM_PUSH_WARNING_AND_DISABLE("-Winfinite-recursion")
  /**
   * @return Pointer to a buffer containing row offset results. Some packing
   *         objects fuse row offset computation for later requantization step.
   */
  std::int32_t* getRowOffsetBuffer() const {
    return static_cast<const PT*>(this)->getRowOffsetBuffer();
  }
  /**
   * @brief When k loop is also tiled/blocked, this function is used to check if
   * have executed computations for the last k block so that we can perform
   *        post-GEMM operations.
   */
  bool isThisLastKBlock(int block_id) const {
    return static_cast<const PT*>(this)->isThisLastKBlock(block_id);
  }
  FBGEMM_POP_WARNING
  FBGEMM_POP_WARNING

  /**
   * @brief Actual packing of a block of the source matrix in pmat buffer.
   */
  void pack(const block_type_t& block) {
#if defined(FBGEMM_FBCODE) || !defined(__aarch64__)
    static_cast<PT*>(this)->pack(block);
#else
    throw std::runtime_error("PackMatrix::pack() not implemented for aarch64");
#endif // __aarch64__
  }

  std::int32_t numRows() const {
    return nrows_;
  }

  std::int32_t numCols() const {
    return ncols_;
  }

  /**
   * @return The number of rows in each block
   */
  std::int32_t blockRowSize() const {
    return brow_;
  }

  /**
   * @return The number of columns in each block
   */
  std::int32_t blockColSize() const {
    return bcol_;
  }

  /**
   * @return The number of blocks along rows
   */
  std::int32_t blockRows() const {
    return nbrow_;
  }

  /**
   * @return The number of blocks along columns
   */
  std::int32_t blockCols() const {
    return nbcol_;
  }

  /**
   * @return The number of the rows in the currently packed block of a matrix.
   *         For pre-packed (i.e., fully-packed), it's equal to the total number
   * of rows.
   */
  std::int32_t numPackedRows() const {
    return packedBlock_.row_size;
  }

  /**
   * @return The number of columns in the currently packed block of a matrix.
   *         For pre-packed (i.e., fully-packed), it's equal to the number of
   * columns.
   */
  std::int32_t numPackedCols() const {
    return packedBlock_.col_size;
  }

  /**
   * @return The first row of the block we're working on.
   */
  std::int32_t packedRowStart() const {
    return packedBlock_.row_start;
  }

  /**
   * @return The first column of the block we're working on.
   */
  std::int32_t packedColStart() const {
    return packedBlock_.col_start;
  }

  /**
   * @return The beginning of (rowBlockNum, colBlockNum)th block
   */
  inpType* getBuf(std::int32_t rowBlockNum = 0, std::int32_t colBlockNum = 0) {
    return buf_ + blockRowSize() * blockColSize() * rowBlockNum +
        blockRowSize() * blockColSize() * blockCols() * colBlockNum;
  }

  /**
   * @brief Print the packed block.
   */
  void printPackedMatrix(const std::string& name) {
    static_cast<PT*>(this)->printPackedMatrix(name);
  }

  /**
   * @return The number of rows in the last row block.
   */
  std::int32_t lastBrow() const {
    return last_brow_;
  }

  /**
   * @return The number of columns in the last column block.
   */
  std::int32_t lastBcol() const {
    return last_bcol_;
  }

  int numGroups() const {
    return G_;
  }

  /**
   * @return True if the last column block has fewer columns than the block
   *         size.
   */
  bool isThereColRemainder() const {
    return last_bcol_ != blockColSize();
  }

  virtual ~PackMatrix() {
    if (bufAllocatedHere_) {
      fbgemmAlignedFree(buf_);
    }
  }

 protected:
  /**
   * Set which block we're packing
   */
  void packedBlock(const block_type_t& block) {
    packedBlock_ = block;
    nbrow_ = (numPackedRows() + blockRowSize() - 1) / blockRowSize();
    nbcol_ = (numPackedCols() + blockColSize() - 1) / blockColSize();

    last_brow_ = ((numPackedRows() % blockRowSize()) == 0)
        ? blockRowSize()
        : (numPackedRows() % blockRowSize());
    last_bcol_ = ((numPackedCols() % blockColSize()) == 0)
        ? blockColSize()
        : (numPackedCols() % blockColSize());
  }

  inpType* buf_;
  std::int32_t brow_; ///< the number of rows in each block
  std::int32_t bcol_; ///< the number of columns in each block
  std::int32_t nbrow_; ///< the number of blocks along rows
  std::int32_t nbcol_; ///< the number of blocks along columns
  bool bufAllocatedHere_{false};
  const BlockingFactors*
      blocking_params; ///< MCB, KCB, NCB, MR, NR, NR_MIN, ROW_INTERLEAVE;

 private:
  std::int32_t nrows_, ncols_;
  int G_;
  block_type_t packedBlock_; ///< The block in the source matrix just packed
  std::int32_t last_brow_, last_bcol_;
};

/**
 * @brief Matrix packed for the first input matrix in GEMM (usually
 *        activation).  The source matrix is already quantized. Default
 * accumulation type is int32.
 */
template <typename T, typename accT = std::int32_t>
class FBGEMM_API PackAMatrix final
    : public PackMatrix<PackAMatrix<T, accT>, T, accT> {
 public:
  using This = PackAMatrix<T, accT>;
  using BaseType = PackMatrix<This, T, accT>;
  using inpType = T;
  using accType = accT;

  PackAMatrix() = delete; // no default constructor

  PackAMatrix(
      matrix_op_t trans,
      std::int32_t nRow,
      std::int32_t nCol,
      const inpType* smat,
      std::int32_t ld,
      inpType* pmat = nullptr,
      int groups = 1,
      const BlockingFactors* params = nullptr);

  /**
   * Activation matrices are not constant so cannot amortize the cost of
   * pre-packing.
   */
  bool isPrePacked() const {
    return false;
  }

  /**
   * @return True if this is used as A matrix.
   */
  static constexpr bool isA() {
    return true;
  }

  /**
   * @return A pointer to the row offset buffer. There is no row offset buffer
   *         calculations with this packing class, hence, it returns nullptr.
   */
  std::int32_t* getRowOffsetBuffer() const {
    return nullptr;
  }

  /**
   * @return Offset of the element in the packed matrix that was at (i, j) in
   *         the source matrix.
   */
  std::int32_t addr(std::int32_t i, std::int32_t j) const;

  /**
   * @brief Packs a block of source matrix into pmat buffer.
   */
  void pack(const block_type_t& block);

  /**
   * @brief Print the packed block.
   */
  void printPackedMatrix(const std::string& name);

 private:
  matrix_op_t trans_;
  const T* smat_;
  std::int32_t ld_;
  std::int32_t row_interleave_B_;
};

/**
 * @brief Matrix packed for the second input matrix in GEMM (usually weight).
 *        The source matrix is already quantized. Default accumulation
 *        type is int32.
 */
template <typename T, typename accT = std::int32_t>
class FBGEMM_API PackBMatrix final
    : public PackMatrix<PackBMatrix<T, accT>, T, accT> {
 public:
  using This = PackBMatrix<T, accT>;
  using BaseType = PackMatrix<This, T, accT>;
  using inpType = T;
  using accType = accT;

  PackBMatrix() = delete; // no default constructor

  /**
   * @param groups if > 1 and trans == NoTranspose, smat is nRow x nCol with
   *               groups are vertically concatenated: each group is
   *               (nRow / groups) x nCol .
   *               if > 1 and trans == Transpose, smat is (nCol * groups) x
   *               (nRow / groups) with groups are horizontally concatenated:
   *               each group is nCol x (nRow / groups) . Each group is
   *               transposed and vertically concatenated to match with the
   *               NoTranspose case.
   */
  PackBMatrix(
      matrix_op_t trans,
      std::int32_t nRow,
      std::int32_t nCol,
      const inpType* smat,
      std::int32_t ld,
      inpType* pmat = nullptr,
      int groups = 1,
      const BlockingFactors* params = nullptr);

  /**
   * Weight matrices are usually constant so worth pre-packing.
   */
  bool isPrePacked() const {
    return true;
  }

  /**
   * @return True if to be used as A matrix, False otherwise.
   */
  static constexpr bool isA() {
    return false;
  }

  /**
   * @brief When k loop is also tiled/blocked, this function is used to check if
   * have executed computations for the last k block so that we can perform
   *        post-GEMM operations.
   */
  bool isThisLastKBlock(int block_id) const {
    return (BaseType::blockRows() - 1) == block_id;
  }

  /**
   * @return Offset of the element in the packed matrix that was at (i, j) in
   *         the source matrix.
   */
  std::int32_t addr(std::int32_t i, std::int32_t j) const;

  /**
   * @brief Packs a block of source matrix into pmat buffer. The blocking
   *        parameters are needed to compute the buffer size of each group.
   *        It will use default blocking parameters if params is not provided.
   */
  void pack(const block_type_t& block, const BlockingFactors* params = nullptr);

  /**
   * @brief Print the packed block.
   */
  void printPackedMatrix(
      const std::string& name,
      const BlockingFactors* params = nullptr);

  /**
   * @return true if meta information like matrix shape is the same.
   */
  bool metaEquals(const PackBMatrix<T, accT>& that) const;
  /**
   * @return true if matrices are the same.
   */
  bool equals(const PackBMatrix<T, accT>& that) const;

  /**
   * @brief Unpack pmat buffer to the origin_buf (Used for the serialization to
   * recover weight matrix).
   */
  void unpack(T* origin_buf, const BlockingFactors* params = nullptr);

  ~PackBMatrix() override = default;

 private:
  matrix_op_t trans_;
  const T* smat_;
  std::int32_t ld_;
  std::int32_t row_interleave_;

  /**
   * @brief Internal function performing both pack & unpack
   */
  void pack_unpack_(
      const block_type_t& block,
      T* unpack_buf,
      T* pack_buf,
      bool ispack,
      const BlockingFactors* params = nullptr);
};

/**
 * @brief Matrix packed for direct group convolution.
 *        The source matrix is already quantized. Default accumulation
 *        type is int32.
 */
template <typename T, typename accT = std::int32_t, int SPATIAL_DIM = 2>
class FBGEMM_API PackWeightMatrixForGConv {
 public:
  using This = PackWeightMatrixForGConv<T, accT, SPATIAL_DIM>;
  using inpType = T;
  using accType = accT;

  PackWeightMatrixForGConv() = delete; // no default constructor
  PackWeightMatrixForGConv(const PackWeightMatrixForGConv&) = delete; // no copy
  PackWeightMatrixForGConv& operator=(const PackWeightMatrixForGConv&) =
      delete; // no copy

  PackWeightMatrixForGConv(PackWeightMatrixForGConv&&) = delete; // no move
  PackWeightMatrixForGConv& operator=(PackWeightMatrixForGConv&&) =
      delete; // no move

  /**
   * @param pmat if nullptr, a buffer is allocated and owned by this class.
   */
  PackWeightMatrixForGConv(
      matrix_op_t trans,
      const conv_param_t<SPATIAL_DIM>& conv_param,
      const inpType* sdata,
      inpType* pdata = nullptr);

  /**
   * Number of groups we work at a time to fill the full simd width
   * e.g., IC_PER_G = 4 and OC_PER_G = 4, we work on two groups at a time
   * to fill the avx2 width of 256 bits.
   */
  static int numOfGroupsTogether(const conv_param_t<SPATIAL_DIM>& conv_param);

  /**
   * @brief Packs a block of source matrix into pmat buffer.
   */
  void pack();

  /**
   * @brief Unpacks a pmat buffer into source matrix.
   */
  void unpack(T* origin_buf);

  /**
   * @brief Return packed data
   */
  inpType* getBuf() {
    return pdata_;
  }

  ~PackWeightMatrixForGConv() {
    if (bufAllocatedHere_) {
      fbgemmAlignedFree(pdata_);
    }
  }

 private:
  matrix_op_t trans_;
  const conv_param_t<SPATIAL_DIM> conv_param_;
  const T* sdata_;
  T* pdata_;
  bool bufAllocatedHere_{false};
  // Number of groups we work at a time to fill the full simd width
  int GTogether_;

  /**
   * @brief Internal function performing both pack & unpack
   */
  void pack_unpack_(const T* src, T* dst, bool ispack);

  /**
   * @brief Get the index of the unpacked data
   */
  int unpacked_index_(int t, int r, int s, int k, int g, int c, bool tr);

  /**
   * @brief Get the index of the packed data
   */
  int packed_index_(int t, int r, int s, int k, int g, int c);
};

/**
 * @brief A container class to keep packed weight tensor for convolution.
 *        The source tensor should already be quantized.
 *
 * @tparam SPATIAL_DIM is equal to 2 for 2D convolutions and 3 for 3D
 *                     convolutions. Default value is 2.
 * @tparam T is the datatype for source tensor. Default value is int8.
 * @tparam accT is the datatype to accumulate into. Default value is int32.
 */
template <
    int SPATIAL_DIM = 2,
    typename T = std::int8_t,
    typename accT = std::int32_t>
class FBGEMM_API PackWeightsForConv {
 public:
  using This = PackWeightsForConv<SPATIAL_DIM, T, accT>;
  using inpType = T;
  using accType = accT;

  PackWeightsForConv() = delete; // no default constructor

  PackWeightsForConv(
      const conv_param_t<SPATIAL_DIM>& conv_param,
      const inpType* sdata,
      const BlockingFactors* blocking_params = nullptr);

  std::shared_ptr<PackBMatrix<T, accT>> getPackedWForIm2col() {
    return W_im2col_packed_;
  }

#if defined(FBGEMM_FBCODE) || !defined(__aarch64__)
  std::shared_ptr<PackedDepthWiseConvMatrix> getPackedWForDepthwise() {
    return W_dw_packed_;
  }
#endif // __aarch64__

  std::shared_ptr<PackedDirectConvMatrix> getPackedWForDirectconv() {
    return W_dc_packed_;
  }

  std::shared_ptr<PackWeightMatrixForGConv<T, accT, SPATIAL_DIM>>
  getPackedWForGroupwise() {
    return W_gconv_packed_;
  }

  std::shared_ptr<PackBMatrix<T, accT>> getPackedWForPointwise() {
    return W_pointwise_packed_;
  }

  int inputChannels() {
    return conv_param_.IC;
  }

  int outputChannels() {
    return conv_param_.OC;
  }

  std::array<int, SPATIAL_DIM> kernelDims() {
    return conv_param_.K;
  }

  int groups() {
    return conv_param_.G;
  }

  /**
   * @brief Returns true if the packed weights would work for the given
   * convolution parameters, and false otherwise
   */
  bool isPackingCompliant(const conv_param_t<SPATIAL_DIM>& conv_p);

  /**
   * @brief Returns a string of mismatching parameters
   */
  std::string mismatchingParams(const conv_param_t<SPATIAL_DIM>& conv_p);

  /**
   * @brief Unpack packed matric into origin_buf (Used for the serialization to
   * recover weight matrix).
   */
  void unpack(T* origin_buf);

 private:
  const conv_param_t<SPATIAL_DIM> conv_param_;
  // Packed weights if we use im2col based convolution implementation
  std::shared_ptr<PackBMatrix<T, accT>> W_im2col_packed_;
#if defined(FBGEMM_FBCODE) || !defined(__aarch64__)
  // Packed weights if we use depthwise convolution implementation
  std::shared_ptr<PackedDepthWiseConvMatrix> W_dw_packed_;
#endif // __aarch64__
  // Packed weights if we use direct convolution implementation
  std::shared_ptr<PackedDirectConvMatrix> W_dc_packed_;
  // Packed weights if we use groupwise (small channels per group) convolution
  // implementation
  std::shared_ptr<PackWeightMatrixForGConv<T, accT, SPATIAL_DIM>>
      W_gconv_packed_;
  // Packed weights if we use direct gemm for pointwise convolution
  std::shared_ptr<PackBMatrix<T, accT>> W_pointwise_packed_;
};

/**
 * @brief Matrix packed for the first input matrix in GEMM (usually activation),
 *        and row offsets used for requantization is computed during packing.
 *        Im2col is fused with packing here. The source matrix is already
 * quantized.
 */
template <typename T, typename accT = std::int32_t, int SPATIAL_DIM = 2>
class FBGEMM_API PackAWithIm2Col
    : public PackMatrix<PackAWithIm2Col<T, accT, SPATIAL_DIM>, T, accT> {
 public:
  using This = PackAWithIm2Col<T, accT, SPATIAL_DIM>;
  using BaseType = PackMatrix<This, T, accT>;
  using inpType = T;
  using accType = accT;

  PackAWithIm2Col() = delete; // no default constructor
  /**
   * @param zero_pt the quantized value that maps to 0.0f floating-point number.
   * @param row_offset If nullptr, this constructor internally allocates a
   *                   buffer and owns it. Otherwise, this class doesn't own
   *                   the buffer. The buffer will be populated when pack
   *                   function is called.
   * @param b_symmetric if true we skip row offset computation
   */
  PackAWithIm2Col(
      const conv_param_t<SPATIAL_DIM>& conv_param,
      const T* sdata,
      inpType* pmat = nullptr,
      std::int32_t a_zero_pt = 0,
      std::int32_t* row_offset = nullptr,
      bool b_symmetric = false,
      const BlockingFactors* params = nullptr);

  PackAWithIm2Col(const PackAWithIm2Col&) = delete;
  PackAWithIm2Col(PackAWithIm2Col&&) = delete;
  PackAWithIm2Col& operator=(const PackAWithIm2Col&) = delete;
  PackAWithIm2Col& operator=(PackAWithIm2Col&&) = delete;

  /**
   * Activation matrices are not constant so cannot amortize the cost of
   * pre-packing.
   */
  bool isPrePacked() const {
    return false;
  }

  /**
   * @return True if this is used as A matrix.
   */
  static constexpr bool isA() {
    return true;
  }

  /**
   * @brief Packs a block of source matrix into pmat buffer.
   */
  void pack(const block_type_t& block);

  /**
   * @return A pointer to the row offset buffer.
   */
  std::int32_t* getRowOffsetBuffer() const {
    return row_offset_;
  }

  /**
   * @brief Print the packed block.
   */
  void printPackedMatrix(const std::string& name);

  /**
   * @return Size of row offset buffer in number of elements
   */
  static int rowOffsetBufferSize(const BlockingFactors* params = nullptr);

  ~PackAWithIm2Col() override {
    if (rowOffsetAllocatedHere) {
      fbgemmAlignedFree(row_offset_);
    }
  }

 private:
  const conv_param_t<SPATIAL_DIM> conv_p_;
  const T* sdata_;
  std::int32_t a_zero_pt_;
  std::int32_t* row_offset_{nullptr};
  bool rowOffsetAllocatedHere{false};
  std::int32_t row_interleave_B_;
};

/**
 * @brief Matrix packed for the first input matrix in GEMM (usually activation),
 *        and row offsets used for requantization is computed during packing.
 *        The source matrix is already quantized.
 */
template <typename T, typename accT = std::int32_t>
class FBGEMM_API PackAWithRowOffset final
    : public PackMatrix<PackAWithRowOffset<T, accT>, T, accT> {
 public:
  using This = PackAWithRowOffset<T, accT>;
  using BaseType = PackMatrix<This, T, accT>;
  using inpType = T;
  using accType = accT;

  PackAWithRowOffset() = delete; // no default constructor
  /**
   * @param row_offset If nullptr, this constructor internally allocates a
   *                   buffer and owns it. Otherwise, this class doesn't own
   *                   the buffer. The buffer will be populated when pack
   *                   function is called.
   */
  PackAWithRowOffset(
      matrix_op_t trans,
      std::uint32_t nRow,
      std::uint32_t nCol,
      const T* smat,
      std::uint32_t ld,
      inpType* pmat = nullptr,
      int groups = 1,
      std::int32_t* row_offset = nullptr,
      const BlockingFactors* params = nullptr);

  PackAWithRowOffset(const PackAWithRowOffset&) = delete;
  PackAWithRowOffset(PackAWithRowOffset&&) = delete;
  PackAWithRowOffset& operator=(const PackAWithRowOffset&) = delete;
  PackAWithRowOffset& operator=(PackAWithRowOffset&&) = delete;

  /**
   * Activation matrices are not constant so cannot amortize the cost of
   * pre-packing.
   */
  bool isPrePacked() const {
    return false;
  }

  /**
   * @return True if this is used as A matrix.
   */
  static constexpr bool isA() {
    return true;
  }

  /**
   * @return Offset of the element in the packed matrix that was at (i, j) in
   *         the source matrix
   */
  std::int32_t addr(std::int32_t i, std::int32_t j) const;

  /**
   * @brief Packs a block of source matrix into pmat buffer.
   */
  void pack(const block_type_t& block);

  /**
   * @return A pointer to the row offset buffer.
   */
  std::int32_t* getRowOffsetBuffer() const {
    return row_offset_;
  }

  /**
   * @brief Print the packed block.
   */
  void printPackedMatrix(const std::string& name);

  /**
   * @return size of row offset buffer in number of elements
   */
  static int rowOffsetBufferSize(const BlockingFactors* params = nullptr);

  ~PackAWithRowOffset() override {
    if (rowOffsetAllocatedHere) {
      fbgemmAlignedFree(row_offset_);
    }
  }

 private:
  matrix_op_t trans_;
  const T* smat_;
  std::uint32_t ld_;
  std::int32_t* row_offset_{nullptr};
  bool rowOffsetAllocatedHere{false};
  std::int32_t row_interleave_B_;
};

/**
 * @brief Matrix packed for the first input matrix in GEMM (usually activation),
 *        and row offsets used for requantization is computed during packing.
 *        The source matrix is in fp32 and quantized during packing.
 */
template <typename T, typename accT = std::int32_t>
class FBGEMM_API PackAWithQuantRowOffset final
    : public PackMatrix<PackAWithQuantRowOffset<T, accT>, T, accT> {
 public:
  using This = PackAWithQuantRowOffset<T, accT>;
  using BaseType = PackMatrix<This, T, accT>;
  using inpType = T;
  using accType = accT;

  PackAWithQuantRowOffset() = delete; // no default constructor
  /**
   * @param row_offset If nullptr, this constructor internally allocates a
   *                   buffer and owns it. Otherwise, this class doesn't own
   *                   the buffer. The buffer will be populated when pack
   *                   function is called.
   */
  PackAWithQuantRowOffset(
      matrix_op_t trans,
      std::int32_t nRow,
      std::int32_t nCol,
      const float* smat,
      std::int32_t ld,
      inpType* pmat = nullptr,
      float scale = 1.0f,
      std::int32_t zero_pt = 0,
      int groups = 1,
      std::int32_t* row_offset = nullptr,
      const BlockingFactors* params = nullptr);
  PackAWithQuantRowOffset(const PackAWithQuantRowOffset&) = delete;
  PackAWithQuantRowOffset(PackAWithQuantRowOffset&&) = delete;
  PackAWithQuantRowOffset& operator=(const PackAWithQuantRowOffset&) = delete;
  PackAWithQuantRowOffset& operator=(PackAWithQuantRowOffset&&) = delete;

  /**
   * Activation matrices are not constant so cannot amortize the cost of
   * pre-packing.
   */
  bool isPrePacked() const {
    return false;
  }

  /**
   * @return True if this is used as A matrix.
   */
  static constexpr bool isA() {
    return true;
  }

  /**
   * @return offset of the element in the packed matrix that was at (i, j) in
   *         the source matrix
   */
  std::int32_t addr(std::int32_t i, std::int32_t j) const;

  /**
   * @brief Packs a block of source matrix into pmat buffer.
   */
  void pack(const block_type_t& block);

  /**
   * @return A pointer to the row offset buffer.
   */
  std::int32_t* getRowOffsetBuffer() const {
    return row_offset_;
  }

  /**
   * @brief Print the packed block.
   */
  void printPackedMatrix(const std::string& name);

  /**
   * @return Size of row offset buffer in number of elements
   */
  static int rowOffsetBufferSize(const BlockingFactors* params = nullptr);

  ~PackAWithQuantRowOffset() override {
    if (rowOffsetAllocatedHere) {
      fbgemmAlignedFree(row_offset_);
    }
  }

 private:
  matrix_op_t trans_;
  const float* smat_;
  std::int32_t ld_;
  float scale_;
  std::int32_t zero_pt_;
  std::int32_t* row_offset_{nullptr};
  bool rowOffsetAllocatedHere{false};
  std::int32_t row_interleave_B_;
};

/*
 *
 * Post Processing of outputs
 *
 */

/**
 * @brief Does nothing. NoOp. Used as the last operation in the output
 *        processing pipeline.
 *
 */
template <typename outT = std::uint8_t, typename inT = std::uint8_t>
class FBGEMM_API DoNothing {
 public:
  using outType = outT;
  using inpType = inT;
  DoNothing() = default;
  template <inst_set_t instSet>
  int f(
      outType* /* unused */,
      inpType* /* unused */,
      const block_type_t& /* unused */,
      int /* unused */,
      int /* unused */) const {
    return 0;
  }
};

/**
 * @brief Copy data pointed by inp ptr to out ptr when
 *        inp ptr and out ptr are not the same.
 *        inp buffer: row and column start points: (0, 0)
 *        output buffer: row and column start points:
 *        (block.row_start, block.col_start)
 *
 * This is the output processing stage that should passed when there is no
 * requantization and output is required in the same format as internal buffer
 * used for accumulation.
 */
template <
    typename outT = std::int32_t,
    typename inT = std::int32_t,
    typename nextOPType = DoNothing<outT, outT>>
class FBGEMM_API memCopy {
 public:
  using outType = outT;
  using inpType = inT;
  explicit memCopy(nextOPType& nextop) : nextop_(nextop) {}
  template <inst_set_t instSet>
  inline int f(
      outType* out,
      inpType* inp,
      const block_type_t& block,
      int ld_out,
      int ld_in) const;

 private:
  nextOPType& nextop_;
};

/**
 * @brief Perform scaling on accumulated data.
 */
template <
    typename outT = std::int32_t,
    typename inT = std::int32_t,
    typename nextOPType = DoNothing<outT, outT>>
class ScaleOP {
 public:
  using outType = outT;
  using inpType = inT;
  explicit ScaleOP(inpType scalingFactor) : scalingFactor_(scalingFactor) {}

  template <inst_set_t instSet>
  inline int f(
      outType* out,
      inpType* inp,
      const block_type_t& block,
      int ld_out,
      int ld_in) const;

 private:
  inpType scalingFactor_;
};

/**
 * @brief Perform Relu on accumulated data.
 */
template <
    typename outT = std::int32_t,
    typename inT = std::int32_t,
    typename nextOPType = DoNothing<outT, outT>>
class ReluOutput {
 public:
  using outType = outT;
  using inpType = inT;
  explicit ReluOutput(inpType zero_pt) : zero_pt_(zero_pt) {}

  template <inst_set_t instSet>
  inline int f(
      outType* out,
      inpType* inp,
      const block_type_t& block,
      int ld_out,
      int ld_in) const;

 private:
  inpType zero_pt_;
};

/**
 * @brief Perform Dense-Matrix * Sparse-Matrix as a part the of output
 * processing pipeline.
 *
 * SPMDM (SParse Matrix times Dense Matrix) inplace on the 32-bit input buffer
 * (inp). After modifying the input buffer, pass it to the next op.
 * When groups > 1, each group is numRows() x (numCols()/groups) matrix.
 */
template <
    typename outT = std::int32_t,
    typename inT = std::int32_t,
    typename nextOPType = DoNothing<inT, inT>>
class FBGEMM_API DoSpmdmOnInpBuffer {
 public:
  using outType = outT;
  using inpType = inT;
  DoSpmdmOnInpBuffer(
      nextOPType& nextop,
      const std::uint8_t* A,
      int lda,
      const CompressedSparseColumn& B_csc,
      int groups = 1)
      : nextop_(nextop), A_(A), lda_(lda), B_csc_(B_csc), groups_(groups) {}

  template <inst_set_t instSet>
  inline int f(
      outT* out,
      inT* inp,
      const block_type_t& block,
      int ld_out,
      int ld_in) const;

 private:
  nextOPType& nextop_;
  const std::uint8_t* A_;
  const int lda_;
  const CompressedSparseColumn& B_csc_;
  const int groups_;
};

/**
 * @brief Perform Dense-Matrix * Sparse-Matrix as a part the of output
 * processing pipeline.
 *
 * SPMDM (SParse Matrix times Dense Matrix) inplace on the 32-bit input buffer
 * (inp). After modifying the input buffer, pass it to the next op.
 * When groups > 1, each group is numRows() x (numCols()/groups) matrix.
 */
template <
    typename outT = std::int32_t,
    typename inT = std::int32_t,
    typename nextOPType = DoNothing<inT, inT>>
class FBGEMM_API DoSConvOnInpBuffer {
 public:
  using outType = outT;
  using inpType = inT;
  DoSConvOnInpBuffer(
      nextOPType& nextop,
      const std::uint8_t* A,
      const conv_param_t<>& conv_p,
      std::int32_t A_zero_point,
      const CompressedSparseColumn& B_csc)
      : nextop_(nextop),
        A_(A),
        conv_p_(conv_p),
        A_zero_point_(A_zero_point),
        B_csc_(B_csc) {}

  template <inst_set_t instSet>
  inline int f(
      outT* out,
      inT* inp,
      const block_type_t& block,
      int ld_out,
      int ld_in) const;

 private:
  nextOPType& nextop_;
  const std::uint8_t* A_;
  const conv_param_t<> conv_p_;
  const std::int32_t A_zero_point_;
  const CompressedSparseColumn& B_csc_;
};

/**
 * @brief Requantize values in inp buffer and write to out buffer.
 *        pass the out buffer to next op for further processing.
 */
template <
    bool FUSE_RELU,
    QuantizationGranularity Q_GRAN = QuantizationGranularity::TENSOR,
    typename BIAS_TYPE = std::int32_t,
    typename outT = std::uint8_t,
    typename inT = std::int32_t,
    typename nextOPType = DoNothing<outT, outT>>
class FBGEMM_API ReQuantizeOutput {
 public:
  static constexpr int RELU_FUSED = FUSE_RELU;
  static constexpr QuantizationGranularity QGRANType = Q_GRAN;
  using BIAS_T = BIAS_TYPE;
  using outType = outT;
  using inpType = inT;
  /**
   * @param C_multiplier The length of this array is
   *                     1 when Q_GRAN == QuantizationGranularity::TENSOR,
   *                     groups when Q_GRAN == QuantizationGranularity::GROUP,
   *                     nCol if Q_GRAN == QuantizationGranularity::OUT_CHANNEL
   * @param Bq_zero_point The length of this array should be the same as
   *                      C_multiplier.
   * @param row_offsets Typically, this should've been computed by a
   *                    PackAMatrix and should be obtained by
   *                    PackMatrix::getRowOffsetBuffer().
   *                    If Bq_zero_point == 0 (symmetric quantization of B
   *                    matrix), we can pass nullptr.
   * @param col_offsets This should be pre-computed for example using
   *                    col_offsets_with_zero_pt_s8acc32_ref.
   *                    The length should be nCol.
   *                    See PackedRequantizeTest.cc for an example.
   *                    TODO: if Aq_zero_point == 0, allow passing nullptr.
   * @param bias can be nullptr otherwise the length should be nCol
   * @param act_times_w_scale activation_scale * weight_scale. This is only
   *                          used if bias is unquantized (i.e., float).
   */
  ReQuantizeOutput(
      nextOPType& nextop,
      const float* C_multiplier,
      std::int32_t C_zero_point,
      std::int32_t Aq_zero_point,
      const std::int32_t* Bq_zero_point,
      const std::int32_t* row_offsets,
      const std::int32_t* col_offsets,
      const BIAS_T* bias,
      std::uint32_t nCol,
      int groups = 1,
      const float* act_times_w_scale = nullptr)
      : nextop_(nextop),
        C_multiplier_(C_multiplier),
        C_zero_point_(C_zero_point),
        Aq_zero_point_(Aq_zero_point),
        Bq_zero_point_(Bq_zero_point),
        q_row_offsets_(row_offsets),
        q_col_offsets_(col_offsets),
        bias_(bias),
        ncols_(nCol),
        groups_(groups),
        act_times_w_scale_(act_times_w_scale) {}

  template <inst_set_t instSet>
  inline int f(
      outT* out,
      const inT* inp,
      const block_type_t& block,
      int ld_out,
      int ld_in) const;

  const float* getCMultiplier() const {
    return C_multiplier_;
  }
  std::int32_t getAZeroPoint() const {
    return Aq_zero_point_;
  }
  std::int32_t getCZeroPoint() const {
    return C_zero_point_;
  }
  const std::int32_t* getBZeroPoint() const {
    return Bq_zero_point_;
  }
  const std::int32_t* getRowOffsets() const {
    return q_row_offsets_;
  }
  const std::int32_t* getColOffsets() const {
    return q_col_offsets_;
  }
  const BIAS_T* getBias() const {
    return bias_;
  }
  std::uint32_t getNCols() const {
    return ncols_;
  }
  const float* getActWScale() const {
    return act_times_w_scale_;
  }

  void setRowOffsets(const std::int32_t* row_offsets) {
    q_row_offsets_ = row_offsets;
  }

 private:
  nextOPType& nextop_;
  const float* C_multiplier_;
  std::int32_t C_zero_point_;
  std::int32_t Aq_zero_point_;
  const std::int32_t* Bq_zero_point_;
  const std::int32_t* q_row_offsets_;
  const std::int32_t* q_col_offsets_;
  const BIAS_T* bias_;
  std::uint32_t ncols_;
  int groups_;
  const float* act_times_w_scale_;
};

/**
 * @brief Requantize to convert accumulated data to be used as float, i.e., the
 *        output would be used as float.
 */
template <
    bool FUSE_RELU,
    QuantizationGranularity Q_GRAN = QuantizationGranularity::TENSOR,
    typename outT = float,
    typename inT = std::int32_t,
    typename nextOPType = DoNothing<outT, outT>>
class FBGEMM_API ReQuantizeForFloat {
 public:
  using outType = outT;
  using inpType = inT;
  /**
   * @param Bq_scale The length of this array is
   *                 1 when Q_GRAN == QuantizationGranularity::TENSOR,
   *                 groups when Q_GRAN == QuantizationGranularity::GROUP,
   *                 nCol if Q_GRAN == QuantizationGranularity::OUT_CHANNEL
   * @param Bq_zero_point The length of this array should be the same as
   *                      Bq_scale.
   * @param row_offsets Typically, this should've been computed by a
   *                    PackAMatrix and should be obtained by
   *                    PackMatrix::getRowOffsetBuffer().
   *                    If Bq_zero_point == 0 (symmetric quantization of B
   *                    matrix), we can pass nullptr.
   * @param col_offsets This should be pre-computed for example using
   *                    col_offsets_with_zero_pt_s8acc32_ref.
   *                    The length should be nCol.
   *                    See PackedRequantizeTest.cc for an example.
   *                    TODO: if Aq_zero_point == 0, allow passing nullptr.
   * @param bias can be nullptr otherwise the length should be nCol
   */
  ReQuantizeForFloat(
      nextOPType& nextop,
      float Aq_scale,
      const float* Bq_scale,
      std::int32_t Aq_zero_point,
      const std::int32_t* Bq_zero_point,
      const std::int32_t* row_offsets,
      const std::int32_t* col_offsets,
      const float* bias,
      std::uint32_t nCol,
      int groups = 1)
      : nextop_(nextop),
        Aq_scale_(Aq_scale),
        Bq_scale_(Bq_scale),
        Aq_zero_point_(Aq_zero_point),
        Bq_zero_point_(Bq_zero_point),
        q_row_offsets_(row_offsets),
        q_col_offsets_(col_offsets),
        bias_(bias),
        ncols_(nCol),
        groups_(groups) {}

  template <inst_set_t instSet>
  inline int f(
      outT* out,
      inT* inp,
      const block_type_t& block,
      int ld_out,
      int ld_in) const;

 private:
  nextOPType& nextop_;
  float Aq_scale_;
  const float* Bq_scale_;
  std::int32_t Aq_zero_point_;
  const std::int32_t* Bq_zero_point_;
  const std::int32_t* q_row_offsets_;
  const std::int32_t* q_col_offsets_;
  const float* bias_;
  std::uint32_t ncols_;
  int groups_;
};

// type specialized implementation in an include file
#include "./OutputProcessing-inl.h" // @manual

/*
 *
 * ####### GEMM related functions #######
 *
 */

/**
 * Matrix B must be prepacked. For matrix A, packA.pack function is called to
 * pack it.
 *
 * @tparam packingAMatrix processing of A matrix while packing,
 *                        e.g., PackAWithQuantRowOffset
 *
 * @tparam packingBMatrix processing of B matrix while packing,
 *                        e.g.,  pre-multiply by alpha
 * @tparam cT data type of C matrix
 * @tparam processOutputType further processing of outputs, e.g., Relu
 */
template <
    typename packingAMatrix,
    typename packingBMatrix,
    typename cT,
    typename processOutputType>
FBGEMM_API void fbgemmPacked(
    PackMatrix<
        packingAMatrix,
        typename packingAMatrix::inpType,
        typename packingAMatrix::accType>& packA,
    PackMatrix<
        packingBMatrix,
        typename packingBMatrix::inpType,
        typename packingBMatrix::accType>& packB,
    cT* C,
    std::int32_t* C_buffer,
    std::uint32_t ldc,
    const processOutputType& outProcess,
    int thread_id,
    int num_threads,
    const BlockingFactors* blocking_params = nullptr);

/**
 * @brief Perform small-channels-per-group groupwise convolution
 *        Note: Currently threading is not supported. This function does
 *              nothing for thread_ids > 0, i.e., returns early.
 *
 * @param rowOffsetBuf nullptr if B uses symmetric quantization
 *        Note: Currently threading is not supported. This function does
 *              nothing for thread_ids > 0, i.e., returns early.
 */
template <
    typename packed_W,
    typename outType,
    bool FUSE_RELU,
    QuantizationGranularity Q_GRAN,
    int SPATIAL_DIM = 2,
    typename BIAS_TYPE = std::int32_t>
FBGEMM_API void fbgemmGroupwiseConv(
    const conv_param_t<SPATIAL_DIM>& conv_param,
    const std::uint8_t* activations,
    std::int32_t a_zero_point,
    std::int32_t* rowOffsetBuf,
    packed_W& packed_weights,
    outType* out,
    std::int32_t* outBuffer,
    const ReQuantizeOutput<FUSE_RELU, Q_GRAN, BIAS_TYPE>& outProcess,
    int thread_id,
    int num_threads);

template <
    int SPATIAL_DIM,
    QuantizationGranularity Q_GRAN,
    bool FUSE_RELU,
    typename BIAS_TYPE = std::int32_t>
FBGEMM_API void fbgemmDirectConv(
    const conv_param_t<SPATIAL_DIM>& conv_p,
    const uint8_t* Aint8,
    PackedDirectConvMatrix& Bint8_tr,
    uint8_t* C,
    int32_t* C_buffer,
    const ReQuantizeOutput<FUSE_RELU, Q_GRAN, BIAS_TYPE>& outProcess,
    const BIAS_TYPE* bias,
    int thread_id,
    int num_threads);

/**
 * @return Size of row offset buffer in number of elements needed for
 * fbgemmGroupwiseConv
 */
template <int SPATIAL_DIM = 2>
FBGEMM_API int rowOffsetBufferSizeGConv(
    const conv_param_t<SPATIAL_DIM>& conv_param);

/**
 * @brief Is this depthwise convolution optimized?
 */
template <int SPATIAL_DIM = 2, typename ACC_T = std::int32_t>
bool takeDepthWiseFastPath(const conv_param_t<SPATIAL_DIM>& conv_p);

/**
 * @brief Is this groupwise convolution supported?
 */
template <int SPATIAL_DIM>
FBGEMM_API bool fbgemmOptimizedGConv(const conv_param_t<SPATIAL_DIM>& conv_p);

/**
 * @brief Is this convolution a direct matrix-matrix multiplication, i.e., 1x1
 * (aka pointwise) with right paddings etc.?
 */
template <int SPATIAL_DIM>
FBGEMM_API bool takePointWiseFastPath(const conv_param_t<SPATIAL_DIM>& conv_p);

/**
 * @brief Are we running on a fbgemm supported cpu?
 */
FBGEMM_API bool fbgemmSupportedCPU();

/**
 * @brief Performs convolution using fastest path available.
 *
 * @tparam SPATIAL_DIM It's 2 for 2D convolutions and 3 for 3D convolutions.
 */
template <
    typename processOutputType,
    int SPATIAL_DIM = 2,
    typename ACC_T = std::int32_t>
FBGEMM_API int fbgemmConv(
    const conv_param_t<SPATIAL_DIM>& conv_p,
    const std::uint8_t* activations,
    PackWeightsForConv<SPATIAL_DIM, std::int8_t, ACC_T>& packed_weights,
    typename processOutputType::outType* out,
    std::int32_t* outBuffer,
    processOutputType& outProcess,
    int thread_id,
    int num_threads,
    const BlockingFactors* blocking_params = nullptr);

/**
 * @brief Returns which fast path to take
 *
 * @tparam SPATIAL_DIM It's 2 for 2D convolutions and 3 for 3D convolutions.
 *
 * @return optimized_conv_t::depthwise, optimized_conv_t::groupwise or
 *         optimized_conv_t::im2col
 *
 */
template <int SPATIAL_DIM = 2, typename ACC_T = std::int32_t>
FBGEMM_API optimized_conv_t
ConvFastPath(const conv_param_t<SPATIAL_DIM>& conv_p);
} // namespace fbgemm