mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
1436507960
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/73452 Added a fused Int8FC path using PackAWithQuantRowOffset like the INT8 dynamic path. There are two ways to enable it (1) set an positive "X_scale" value in the arg list of Int8FC op (2) send both "Qparam" (for output requantization, could be dummy values) and "in_Qparam" (for fused input quantization) Differential Revision: D34034681 fbshipit-source-id: f25ca8a2b783ea597389d31c110448d19610218e (cherry picked from commit 6fa10ba0e3be2d46298b439fba0fe9ae7e329f3a)
68 lines
2.1 KiB
C++
68 lines
2.1 KiB
C++
#pragma once
|
|
|
|
#include <fbgemm/Fbgemm.h>
|
|
#include "caffe2/operators/fully_connected_op.h"
|
|
#include "caffe2/quantization/server/dnnlowp_op.h"
|
|
|
|
namespace caffe2 {
|
|
|
|
template <typename T, bool ReluFused = false>
|
|
class FullyConnectedDNNLowPOp
|
|
: public DNNLowPOp<T, FullyConnectedOp<CPUContext>> {
|
|
public:
|
|
FullyConnectedDNNLowPOp(const OperatorDef& operator_def, Workspace* ws);
|
|
bool RunOnDevice() override;
|
|
|
|
USE_OPERATOR_FUNCTIONS(CPUContext);
|
|
USE_DNNLOWP_OPERATOR_BASE_FUNCTIONS(T, FullyConnectedOp<CPUContext>);
|
|
|
|
protected:
|
|
bool GetQuantizationParameters_(float X_scale_=-1.0, int X_zero_point_=0);
|
|
|
|
std::size_t axis_{1};
|
|
std::size_t axis_w_{1};
|
|
float X_scale_{-1.0};
|
|
int X_zero_point_{0};
|
|
vector<std::int64_t> Y_shape_cache_;
|
|
|
|
std::vector<dnnlowp::RequantizationParams> requantization_params_;
|
|
bool requantization_param_selected_{false};
|
|
|
|
// x86 only provides SIMD instructions that multiply a signed integer with an
|
|
// unsigned integer. We use signed for weights.
|
|
using T_signed = typename std::make_signed<T>::type;
|
|
|
|
// used in fast path for T == uint8_t
|
|
std::shared_ptr<fbgemm::PackBMatrix<std::int8_t>> Wq_packed_;
|
|
std::vector<std::uint8_t> X_pack_buf_;
|
|
|
|
std::vector<std::int32_t> Y_int32_;
|
|
std::vector<dnnlowp::TensorQuantizationParams> filter_qparams_;
|
|
std::vector<float> filter_scales_;
|
|
std::vector<std::int32_t> filter_zero_points_;
|
|
|
|
std::vector<float> requantization_multipliers_;
|
|
bool quantize_channelwise_;
|
|
|
|
// used in slow path for T != uint8_t
|
|
std::vector<T_signed> W_quantized_;
|
|
|
|
// pre-computed biases and offsets
|
|
std::shared_ptr<std::vector<std::int32_t>> b_quantized_;
|
|
const std::int32_t* b_quantized_data_{nullptr};
|
|
std::vector<std::int32_t> row_offsets_;
|
|
std::shared_ptr<std::vector<std::int32_t>> column_offsets_;
|
|
|
|
// Dequantized bias populated when input bias is quantized and
|
|
// dequantized_output_ == true
|
|
std::vector<float> b_dequantized_;
|
|
const float* b_dequantized_data_{nullptr};
|
|
|
|
bool is_weight_constant_{true};
|
|
|
|
float in_qparams0_scale_old_ = 0;
|
|
std::int32_t in_qparams0_zero_point_old_ = 0;
|
|
}; // class FullyConnectedDNNLowPOp
|
|
|
|
} // namespace caffe2
|