conv_bn_kernel.cpp 2.6 KB
Newer Older
Z
zhangyang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
/* 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. */

#ifdef FUSION_CONVBN_OP

#include "operators/kernel/conv_bn_kernel.h"

namespace paddle_mobile {
namespace operators {

template <>
N
nhzlx 已提交
23
bool ConvBNKernel<FPGA, float>::Init(FusionConvBNParam<FPGA> *param) {
Z
zhangyang 已提交
24
  bool relu_enabled = false;
Z
zhangyang 已提交
25 26 27
  auto input = const_cast<Tensor *>(param->Input());
  auto filter = const_cast<Tensor *>(param->Filter());
  auto out = param->Output();
Z
zhangyang 已提交
28 29 30 31 32
  auto bn_mean_ptr = param->InputMean()->data<float>();
  auto bn_var_ptr = param->InputVariance()->data<float>();
  auto bn_scale_ptr = param->InputScale()->data<float>();
  auto bn_bias_ptr = param->InputBias()->data<float>();
  const float epsilon = param->Epsilon();
Z
zhangyang 已提交
33 34 35
  PADDLE_MOBILE_ENFORCE(out->dims()[1] == param->InputBias()->dims()[0],
                        "Output channel should be equal to bias number");
  const int channel = out->dims()[1];
36
  auto bs_ptr =
37
      (float *)fpga::fpga_malloc(2 * channel * sizeof(float));  // // NOLINT
Z
zhangyang 已提交
38 39
  auto new_scale = new Tensor();
  auto new_bias = new Tensor();
Z
zhangyang 已提交
40 41 42 43 44 45
  auto new_scale_ptr = new_scale->mutable_data<float>({channel});
  auto new_bias_ptr = new_bias->mutable_data<float>({channel});

  for (int i = 0; i < channel; i++) {
    new_scale_ptr[i] = bn_scale_ptr[i] /
                       static_cast<float>(pow((bn_var_ptr[i] + epsilon), 0.5));
Z
zhangyang 已提交
46
    new_bias_ptr[i] = bn_bias_ptr[i] + (0 - bn_mean_ptr[i]) * new_scale_ptr[i];
Z
zhangyang 已提交
47 48
    bs_ptr[i + channel] = new_scale_ptr[i];
    bs_ptr[i] = new_bias_ptr[i];
Z
zhangyang 已提交
49 50 51
  }
  param->SetNewScale(new_scale);
  param->SetNewBias(new_bias);
Z
zhangyang 已提交
52

qnqinan's avatar
qnqinan 已提交
53
  fpga::format_conv_data(filter, out, &bs_ptr, param->Groups());
Z
zhangyang 已提交
54

Z
zhangyang 已提交
55 56 57 58 59
  fpga::SplitConvArgs conv_arg = {0};
  fpga::fill_split_arg(&conv_arg, input, out, filter, relu_enabled,
                       param->Groups(), param->Strides()[0],
                       param->Strides()[1], param->Paddings()[0],
                       param->Paddings()[1], bs_ptr);
60
  param->SetFpgaArgs(conv_arg);
Z
zhangyang 已提交
61 62 63 64
  return true;
}

template <>
65
void ConvBNKernel<FPGA, float>::Compute(const FusionConvBNParam<FPGA> &param) {
Z
zhangyang 已提交
66 67 68 69 70 71 72
  fpga::ComputeFpgaConv(param.FpgaArgs());
}

}  // namespace operators
}  // namespace paddle_mobile

#endif