requantize_mkldnn_op.cc 5.8 KB
Newer Older
X
xiaolil1 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.

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. */

15
#include "dnnl.hpp"  // NOLINT
X
xiaolil1 已提交
16 17 18 19 20 21 22 23
#include "paddle/fluid/framework/data_layout_transform.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/operators/requantize_op.h"
#include "paddle/fluid/platform/mkldnn_helper.h"

namespace paddle {
namespace operators {

24 25
using dnnl::memory;
using dnnl::reorder;
X
xiaolil1 已提交
26
using platform::to_void_cast;
27
using Tensor = phi::DenseTensor;
X
xiaolil1 已提交
28

29 30 31 32 33 34 35 36
namespace {

inline uint8_t clip_to_uint8(float x) {
  return std::max(0L, std::min(255L, std::lround(x)));
}

}  // namespace

X
xiaolil1 已提交
37 38 39 40
template <typename T>
class ReQuantOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
41
    auto* input = ctx.Input<phi::DenseTensor>("Input");
X
xiaolil1 已提交
42
    auto scale_in = ctx.Attr<float>("Scale_in");
43
    auto shift_in = ctx.Attr<float>("Shift_in");
X
xiaolil1 已提交
44
    auto scale_out = ctx.Attr<float>("Scale_out");
45 46
    auto shift_out = ctx.Attr<float>("Shift_out");
    bool with_shift = shift_in != 0.0f || shift_out != 0.0f;
47
    auto* output = ctx.Output<phi::DenseTensor>("Output");
48

49
    PADDLE_ENFORCE_NE(
50 51
        scale_in,
        0.0f,
52 53
        platform::errors::InvalidArgument("Scale of input cannot be 0.0"));
    PADDLE_ENFORCE_NE(
54 55
        scale_out,
        0.0f,
56
        platform::errors::InvalidArgument("Scale of output cannot be 0.0"));
57 58
    if (shift_in != 0.0f) {
      PADDLE_ENFORCE_EQ(
59 60
          framework::TransToProtoVarType(input->dtype()),
          framework::proto::VarType::UINT8,
61 62 63 64
          platform::errors::Unimplemented("Requantize does not support nonzero "
                                          "shift for signed input."));
    }

X
xiaolil1 已提交
65 66 67 68
    auto& dev_ctx =
        ctx.template device_context<platform::MKLDNNDeviceContext>();
    const auto& engine = dev_ctx.GetEngine();

69
    auto src_tz = phi::vectorize(input->dims());
X
xiaolil1 已提交
70

71 72
    float reorder_scale = scale_out / scale_in;

73 74
    std::string key = platform::CreateKey(
        dev_ctx, src_tz, scale_in, scale_out, ctx.OutputName("Output"));
75
    key = platform::ExtendKeyWithThreadInfoIfNeeded(dev_ctx, key);
76 77 78
    const std::string key_prim = key + "@r";
    const std::string key_src_mem = key + "@s";
    const std::string key_dst_mem = key + "@d";
X
xiaolil1 已提交
79

80 81 82 83
    std::shared_ptr<dnnl::memory> src_memory;
    std::shared_ptr<dnnl::memory> dst_memory;
    std::shared_ptr<reorder> reorder_p;
    reorder_p = std::static_pointer_cast<reorder>(dev_ctx.GetBlob(key_prim));
X
xiaolil1 已提交
84

85
    const T* input_data = input->data<T>();
A
Adam 已提交
86

87
    if (reorder_p == nullptr) {
88
      auto dst_tz = phi::vectorize(output->dims());
89 90
      auto src_dt = framework::ToMKLDNNDataType(
          framework::TransToProtoVarType(input->dtype()));
91
      auto dst_dt = with_shift ? framework::MKLDNNDataType::u8 : src_dt;
92

93
      auto src_md = platform::MKLDNNMemDesc({src_tz}, src_dt, input->format());
94 95
      src_memory = std::make_shared<dnnl::memory>(
          src_md, engine, to_void_cast<T>(input_data));
96
      auto dst_md = platform::MKLDNNMemDesc({dst_tz}, dst_dt, input->format());
97 98 99 100 101

      dnnl::primitive_attr attri;
      int mask = 0;
      attri.set_output_scales(mask, {reorder_scale});
      if (with_shift) {
102
        dnnl::post_ops post_operations;
103 104 105 106 107 108 109 110 111 112 113 114 115
        post_operations.append_sum();
        attri.set_post_ops(post_operations);
        uint8_t* output_data = output->mutable_data<uint8_t>(ctx.GetPlace());
        uint8_t reorder_shift =
            clip_to_uint8(shift_out - reorder_scale * shift_in);
        std::memset(output_data, reorder_shift, output->numel());
        dst_memory = std::make_shared<dnnl::memory>(
            dst_md, engine, to_void_cast<uint8_t>(output_data));
      } else {
        T* output_data = output->mutable_data<T>(ctx.GetPlace());
        dst_memory = std::make_shared<dnnl::memory>(
            dst_md, engine, to_void_cast<T>(output_data));
      }
116 117 118 119 120 121 122 123 124 125 126 127 128 129 130

      auto reorder_pd =
          reorder::primitive_desc(*src_memory, *dst_memory, attri);
      reorder_p = std::make_shared<reorder>(reorder_pd);

      dev_ctx.SetBlob(key_prim, reorder_p);
      dev_ctx.SetBlob(key_src_mem, src_memory);
      dev_ctx.SetBlob(key_dst_mem, dst_memory);
    } else {
      src_memory =
          std::static_pointer_cast<dnnl::memory>(dev_ctx.GetBlob(key_src_mem));
      src_memory->set_data_handle(to_void_cast<T>(input_data));

      dst_memory =
          std::static_pointer_cast<dnnl::memory>(dev_ctx.GetBlob(key_dst_mem));
131 132 133 134 135 136 137 138 139 140 141
      if (with_shift) {
        uint8_t* output_data = output->mutable_data<uint8_t>(ctx.GetPlace());
        uint8_t reorder_shift =
            clip_to_uint8(shift_out - reorder_scale * shift_in);
        std::memset(output_data, reorder_shift, output->numel());
        dst_memory->set_data_handle(output_data);

      } else {
        T* output_data = output->mutable_data<T>(ctx.GetPlace());
        dst_memory->set_data_handle(output_data);
      }
142 143
    }

144
    auto& astream = platform::MKLDNNDeviceContext::tls().get_stream();
145 146 147

    reorder_p->execute(astream, *src_memory, *dst_memory);
    astream.wait();
X
xiaolil1 已提交
148

149
    output->set_mem_desc(dst_memory->get_desc());
X
xiaolil1 已提交
150 151 152 153 154 155 156 157
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;

158 159 160 161 162
REGISTER_OP_KERNEL(requantize,
                   MKLDNN,
                   ::paddle::platform::CPUPlace,
                   ops::ReQuantOpKernel<int8_t>,
                   ops::ReQuantOpKernel<uint8_t>,
163
                   ops::ReQuantOpKernel<paddle::platform::bfloat16>);