affine_channel.cc 3.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
// Copyright (c) 2018 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.

#include "paddle/fluid/inference/anakin/convert/affine_channel.h"
#include <algorithm>
#include <string>
#include <vector>
19
#include "paddle/fluid/inference/anakin/convert/helper.h"
20 21 22 23 24

namespace paddle {
namespace inference {
namespace anakin {

25 26
template <typename TargetT, ::anakin::Precision PrecisionT>
void AffineChannelOpConverter<TargetT, PrecisionT>::operator()(
27 28 29 30 31 32 33 34 35
    const framework::proto::OpDesc &op, const framework::BlockDesc &block_desc,
    const framework::Scope &scope, bool test_mode) {
  framework::OpDesc op_desc(op, nullptr);
  PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1);
  PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1);

  auto op_name = op_desc.Type() + ":" + op_desc.Output("Out").front();
  auto input_name = op_desc.Input("X").front();
  auto output_name = op_desc.Output("Out").front();
36
  this->engine_->AddOp(op_name, "AffineChannel", {input_name}, {output_name});
37 38 39 40

  // Copy the Scale to CPUPlace and get the pointer.
  auto *scale_v = scope.FindVar(op_desc.Input("Scale").front());
  PADDLE_ENFORCE_NOT_NULL(scale_v);
41
  auto weight1 = pblock_from_var<TargetT, PrecisionT>(*scale_v, this->engine_);
42
  this->engine_->AddOpAttr(op_name, "weight_1", *weight1);
43 44 45 46

  // Copy the Bias to CPUPlace and get the pointer.
  auto *bias_v = scope.FindVar(op_desc.Input("Bias").front());
  PADDLE_ENFORCE_NOT_NULL(bias_v);
47
  auto weight2 = pblock_from_var<TargetT, PrecisionT>(*bias_v, this->engine_);
48
  this->engine_->AddOpAttr(op_name, "weight_2", *weight2);
49 50 51 52 53 54
}

}  // namespace anakin
}  // namespace inference
}  // namespace paddle

55
#ifdef PADDLE_WITH_CUDA
56 57 58 59 60 61 62 63
using affine_channel_nv_fp32 =
    ::paddle::inference::anakin::AffineChannelOpConverter<
        ::anakin::saber::NV, ::anakin::Precision::FP32>;
using affine_channel_nv_int8 =
    ::paddle::inference::anakin::AffineChannelOpConverter<
        ::anakin::saber::NV, ::anakin::Precision::INT8>;
REGISTER_CUDA_ANAKIN_OP_CONVERTER(affine_channel, affine_channel_nv_fp32);
REGISTER_CUDA_INT8_ANAKIN_OP_CONVERTER(affine_channel, affine_channel_nv_int8);
64
#endif
65 66 67 68 69 70 71 72 73

using affine_channel_cpu_fp32 =
    ::paddle::inference::anakin::AffineChannelOpConverter<
        ::anakin::saber::X86, ::anakin::Precision::FP32>;
using affine_channel_cpu_int8 =
    ::paddle::inference::anakin::AffineChannelOpConverter<
        ::anakin::saber::X86, ::anakin::Precision::INT8>;
REGISTER_CPU_ANAKIN_OP_CONVERTER(affine_channel, affine_channel_cpu_fp32);
REGISTER_CPU_INT8_ANAKIN_OP_CONVERTER(affine_channel, affine_channel_cpu_int8);