conv_kernel.cpp 4.7 KB
Newer Older
L
liuruilong 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 23
/* 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 CONV_OP

#include "operators/kernel/conv_kernel.h"

namespace paddle_mobile {
namespace operators {

template <>
bool ConvKernel<GPU_CL, float>::Init(ConvParam<GPU_CL> *param) {
L
liuruilong 已提交
24 25 26 27 28
  PADDLE_MOBILE_ENFORCE(
      param->Filter()->dims()[2] == param->Filter()->dims()[3] &&
          param->Paddings()[0] == param->Paddings()[1],
      "need equal");

L
liuruilong 已提交
29 30
  param->Filter()->InitCLImage(cl_helper_.CLContext());

L
liuruilong 已提交
31 32 33 34
  int offset = static_cast<int>(param->Filter()->dims()[2]) / 2 -
               static_cast<int>(param->Paddings()[1]);
  param->SetOffset(offset);

L
liuruilong 已提交
35 36 37 38 39 40
  DLOG << " init helper: " << &cl_helper_;
  DLOG << " conv kernel add kernel ~ ";
  DLOG << " width of one block: " << param->Filter()->WidthOfOneBlock();
  DLOG << " height of one block: " << param->Filter()->HeightOfOneBlock();
  DLOG << " filter dims: " << param->Filter()->dims();

L
liuruilong 已提交
41 42
  if (param->Filter()->WidthOfOneBlock() == 1 &&
      param->Filter()->HeightOfOneBlock() == 1) {
L
liuruilong 已提交
43
    DLOG << " here1 ";
L
liuruilong 已提交
44
    this->cl_helper_.AddKernel("conv_1x1", "conv_kernel.cl");
L
liuruilong 已提交
45

L
liuruilong 已提交
46
  } else if (param->Filter()->dims()[1] == 1) {
L
liuruilong 已提交
47
    DLOG << " here2 ";
L
liuruilong 已提交
48
    this->cl_helper_.AddKernel("depth_conv_3x3", "conv_kernel.cl");
L
liuruilong 已提交
49

L
liuruilong 已提交
50 51
  } else if (param->Filter()->WidthOfOneBlock() == 3 &&
             param->Filter()->HeightOfOneBlock() == 3) {
L
liuruilong 已提交
52
    DLOG << " here3 ";
L
liuruilong 已提交
53
    this->cl_helper_.AddKernel("conv_3x3", "conv_kernel.cl");
L
liuruilong 已提交
54

L
liuruilong 已提交
55 56 57
  } else {
    PADDLE_MOBILE_THROW_EXCEPTION(" not support ");
  }
L
liuruilong 已提交
58

L
liuruilong 已提交
59 60 61 62 63
  return true;
}

template <>
void ConvKernel<GPU_CL, float>::Compute(const ConvParam<GPU_CL> &param) {
L
liuruilong 已提交
64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80
  DLOG << " Compute helper: " << &cl_helper_;
  DLOG << " begin compute ";
  auto kernel = this->cl_helper_.KernelAt(0);
  DLOG << " get work size ";
  auto default_work_size = this->cl_helper_.DefaultWorkSize(*param.Output());
  DLOG << " end work size ";
  int c_block = default_work_size[0];
  int w = default_work_size[1];
  int nh = default_work_size[2];
  auto input = param.Input()->GetCLImage();

  DLOG << " get Input ";

  auto filter = param.Filter()->GetCLImage();

  DLOG << " get Filter ";

L
liuruilong 已提交
81
  auto output = param.Output()->GetCLImage();
L
liuruilong 已提交
82 83 84 85 86 87 88 89 90 91

  DLOG << " get Output ";

  int stride = param.Strides()[0];
  int offset = param.Offset();
  int input_c = param.Input()->CBlock();
  int dilation = param.Dilations()[0];
  int input_width = param.Input()->WidthOfOneBlock();
  int input_height = param.Input()->HeightOfOneBlock();

L
liuruilong 已提交
92 93 94
  int output_width = param.Output()->WidthOfOneBlock();
  int output_height = param.Output()->HeightOfOneBlock();

L
liuruilong 已提交
95 96 97 98
  cl_int status;

  DLOG << " begin set kernel arg ";

L
liuruilong 已提交
99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139
  status = clSetKernelArg(kernel, 0, sizeof(int), &c_block);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 1, sizeof(int), &w);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 2, sizeof(int), &nh);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 3, sizeof(cl_mem), &input);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 4, sizeof(cl_mem), &filter);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 5, sizeof(cl_mem), &output);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 6, sizeof(int), &stride);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 7, sizeof(int), &offset);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 8, sizeof(int), &input_c);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 9, sizeof(int), &dilation);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 10, sizeof(int), &input_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 11, sizeof(int), &input_height);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 12, sizeof(int), &output_width);
  CL_CHECK_ERRORS(status);

  status = clSetKernelArg(kernel, 13, sizeof(int), &output_height);
  CL_CHECK_ERRORS(status);
L
liuruilong 已提交
140 141 142 143 144 145 146 147

  DLOG << " end set kernel arg ";

  DLOG << " begin enqueue ";

  status =
      clEnqueueNDRangeKernel(this->cl_helper_.CLCommandQueue(), kernel, 3, NULL,
                             default_work_size.data(), NULL, 0, NULL, NULL);
L
liuruilong 已提交
148
  CL_CHECK_ERRORS(status);
L
liuruilong 已提交
149
  DLOG << " end enqueue ";
L
liuruilong 已提交
150 151 152 153 154 155 156 157
}

template class ConvKernel<GPU_CL, float>;

}  // namespace operators
}  // namespace paddle_mobile

#endif