conv_transpose_compute.cc

// Copyright (c) 2019 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 "lite/kernels/arm/conv_transpose_compute.h"
#include <vector>
#include "lite/backends/arm/math/funcs.h"
#include "lite/core/op_registry.h"
#include "lite/core/type_system.h"

namespace paddle {
namespace lite {
namespace kernels {
namespace arm {

void Conv2DTransposeCompute::PrepareForRun() {
  auto& param = this->Param<param_t>();
  auto x_dims = param.x->dims();
  auto w_dims = param.filter->dims();
  auto o_dims = param.output->dims();
  int win = x_dims[3];  // nchw
  int hin = x_dims[2];
  int chin = x_dims[1];
  int num = x_dims[0];
  int wout = o_dims[3];
  int hout = o_dims[2];
  int chout = o_dims[1];
  int kw = w_dims[3];  // oihw
  int kh = w_dims[2];
  int group = param.groups;

  // deconv weights layout: chin * chout * kh * kw
  int m = chout * kw * kh / group;
  int n = hin * win;
  int k = chin / group;

  workspace_size_ = group * m * n * sizeof(float);

  auto& ctx = this->ctx_->template As<ARMContext>();
  lite::Tensor tmp_weights;
  lite::arm::math::prepackA(
      &tmp_weights, *(param.filter), 1.f, m, k, group, true, &ctx);
  param.filter->Resize(tmp_weights.dims());
  param.filter->CopyDataFrom(tmp_weights);
  param.filter->Resize(w_dims);
  is_first_epoch_ = false;
}

void Conv2DTransposeCompute::Run() {
  auto& ctx = this->ctx_->template As<ARMContext>();
  ctx.ExtendWorkspace(workspace_size_);
  auto& param = this->Param<param_t>();
  auto x_dims = param.x->dims();
  auto o_dims = param.output->dims();
  auto w_dims = param.filter->dims();
  int num = x_dims[0];
  int chin = x_dims[1];
  int hin = x_dims[2];
  int win = x_dims[3];
  int chout = o_dims[1];
  int hout = o_dims[2];
  int wout = o_dims[3];
  int kw = w_dims[3];  // oihw
  int kh = w_dims[2];
  int group = param.groups;
  bool fuse_relu = param.fuse_relu;
  bool flag_bias = (param.bias != nullptr);

  int m = chout * kw * kh / group;
  int n = hin * win;
  int k = chin / group;
  int group_size_in = win * hin * chin / group;
  int group_size_out = wout * hout * chout / group;
  int group_size_coldata = m * n;
  int hblock = lite::arm::math::get_hblock(&ctx);
  int m_roundup = hblock * ((m + hblock - 1) / hblock);
  int group_size_weights = ((m_roundup * k + 15) / 16) * 16;
  bool flag_1x1s1p1 = (kw == 1) && (kh == 1) && (param.strides[0] == 1) &&
                      (param.strides[1] == 1) && (param.paddings[0] == 0) &&
                      (param.paddings[1] == 0) && (param.dilations[0] == 1) &&
                      (param.dilations[1] == 1);
  ctx.ExtendWorkspace(sizeof(float) * group * m * n);

  auto din = param.x->data<float>();
  auto dout = param.output->mutable_data<float>();
  auto weights = param.filter->data<float>();
  for (int i = 0; i < num; i++) {
    const float* din_batch = din + i * chin * hin * win;
    float* dout_batch = dout + i * chout * hout * wout;
    float* col_data = static_cast<float*>(ctx.workspace_data<float>()) +
                      ctx.llc_size() / sizeof(float);
    if (flag_1x1s1p1) {
      col_data = dout_batch;
    }
    for (int g = 0; g < group; g++) {
      const float* din_group = din_batch + g * group_size_in;
      const float* weights_group = weights + g * group_size_weights;
      float* coldata_group = col_data + g * group_size_coldata;

      lite::arm::math::sgemm_prepack(false,
                                     m,
                                     n,
                                     k,
                                     weights_group,
                                     din_group,
                                     n,
                                     0.f,
                                     coldata_group,
                                     n,
                                     nullptr,
                                     false,
                                     fuse_relu && (!flag_bias),
                                     &ctx);
    }
    if (!flag_1x1s1p1) {
      lite::arm::math::col2im<float>(col_data,
                                     chout,
                                     hout,
                                     wout,
                                     kh,
                                     kw,
                                     param.paddings[0],
                                     param.paddings[1],
                                     param.strides[0],
                                     param.strides[1],
                                     param.dilations[0],
                                     param.dilations[1],
                                     dout_batch);
    }
    if (flag_bias) {
      lite::arm::math::fill_bias_relu<float>(
          dout_batch,
          static_cast<const float*>(param.bias->data<float>()),
          chout,
          wout * hout,
          flag_bias,
          fuse_relu);
    }
  }
}
}  // namespace arm
}  // namespace kernels
}  // namespace lite
}  // namespace paddle

REGISTER_LITE_KERNEL(conv2d_transpose,
                     kARM,
                     kFloat,
                     kNCHW,
                     paddle::lite::kernels::arm::Conv2DTransposeCompute,
                     def)
    .BindInput("Input", {LiteType::GetTensorTy(TARGET(kARM))})
    .BindInput("Bias", {LiteType::GetTensorTy(TARGET(kARM))})
    .BindInput("Filter", {LiteType::GetTensorTy(TARGET(kARM))})
    .BindOutput("Output", {LiteType::GetTensorTy(TARGET(kARM))})
    .Finalize();