affine_channel_op.cu

/* 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.
Indicesou 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 __NVCC__
#include "cub/cub.cuh"
#endif

#ifdef __HIPCC__
#include <hipcub/hipcub.hpp>
namespace cub = hipcub;
#endif

#include "paddle/fluid/framework/data_layout.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/platform/device/gpu/gpu_primitives.h"

namespace paddle {
namespace operators {

template <typename T, framework::DataLayout layout, bool HasBias>
__global__ void KeAffineChannelCUDA(const T* x, const T* scale, const T* bias,
                                    const int C, const int HxW, const int num,
                                    T* y) {
  int gid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  for (int i = gid; i < num; i += stride) {
    const int c = layout == framework::DataLayout::kNCHW ? i / HxW % C : i % C;
    if (HasBias) {
      y[i] = scale[c] * x[i] + bias[c];
    } else {
      y[i] = scale[c] * x[i];
    }
  }
}

template <typename DeviceContext, typename T>
class AffineChannelCUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* x = ctx.Input<framework::Tensor>("X");
    auto* scale = ctx.Input<framework::Tensor>("Scale");
    auto* bias = ctx.Input<framework::Tensor>("Bias");

    auto* y = ctx.Output<framework::Tensor>("Out");
    y->mutable_data<T>(ctx.GetPlace());

    const framework::DataLayout layout =
        framework::StringToDataLayout(ctx.Attr<std::string>("data_layout"));
    auto& dev_ctx = ctx.template device_context<DeviceContext>();

    auto dims = x->dims();
    const int num = x->numel();
    int N = dims[0];
    int C = layout == framework::DataLayout::kNCHW ? dims[1]
                                                   : dims[dims.size() - 1];
    int HxW = num / N / C;

    const T* x_d = x->data<T>();
    const T* scale_d = scale->data<T>();
    const T* bias_d = bias->data<T>();
    T* y_d = y->data<T>();

#ifdef PADDLE_WITH_HIP
    int block = 256;
#else
    int block = 1024;
#endif  // PADDLE_WITH_HIP
    int grid = (num + block - 1) / block;

    int max_threads = dev_ctx.GetMaxPhysicalThreadCount();
    grid = std::min(std::max(max_threads / block, 1), grid);
    if (layout == framework::DataLayout::kNCHW) {
      KeAffineChannelCUDA<T, framework::DataLayout::kNCHW, true>
          <<<grid, block, 0, dev_ctx.stream()>>>(x_d, scale_d, bias_d, C, HxW,
                                                 num, y_d);
    } else {
      KeAffineChannelCUDA<T, framework::DataLayout::kNHWC, true>
          <<<grid, block, 0, dev_ctx.stream()>>>(x_d, scale_d, bias_d, C, HxW,
                                                 num, y_d);
    }
  }
};

template <typename T, int BlockDim, framework::DataLayout layout>
__global__ void AffineChannelScaleBiasGradientCUDAKernel(
    const T* dy, const T* x, const int N, const int C, const int HxW, T* dscale,
    T* dbias) {
  const int outer_size = C;
  const int inner_size = N * HxW;
  typedef cub::BlockReduce<double, BlockDim> BlockReduce;
  __shared__ typename BlockReduce::TempStorage ds_storage;
  __shared__ typename BlockReduce::TempStorage db_storage;

  for (int i = blockIdx.x; i < outer_size; i += gridDim.x) {
    T ds_sum = 0;
    T db_sum = 0;
    for (int j = threadIdx.x; j < inner_size; j += blockDim.x) {
      const int index = layout == framework::DataLayout::kNCHW
                            ? (j / HxW * C + i) * HxW + j % HxW
                            : j * outer_size + i;
      ds_sum += dy[index] * x[index];
      db_sum += dy[index];
    }
    __syncthreads();
    auto ds_out =
        BlockReduce(ds_storage).Reduce(static_cast<double>(ds_sum), cub::Sum());
    auto db_out =
        BlockReduce(db_storage).Reduce(static_cast<double>(db_sum), cub::Sum());
    __syncthreads();
    if (threadIdx.x == 0) {
      dscale[i] = ds_out;
      dbias[i] = db_out;
    }
  }
}

template <typename DeviceContext, typename T>
class AffineChannelGradCUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* x = ctx.Input<framework::Tensor>("X");
    auto* scale = ctx.Input<framework::Tensor>("Scale");
    auto* bias = ctx.Input<framework::Tensor>("Bias");
    auto* dy = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));

    auto* dx = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
    auto* dscale =
        ctx.Output<framework::Tensor>(framework::GradVarName("Scale"));
    auto* dbias = ctx.Output<framework::Tensor>(framework::GradVarName("Bias"));

    const framework::DataLayout layout =
        framework::StringToDataLayout(ctx.Attr<std::string>("data_layout"));
    auto& dev_ctx = ctx.template device_context<DeviceContext>();

    auto dims = dy->dims();
    const int num = dy->numel();
    int N = dims[0];
    int C = layout == framework::DataLayout::kNCHW ? dims[1]
                                                   : dims[dims.size() - 1];
    int HxW = num / N / C;

    const T* dy_d = dy->data<T>();
    const T* s_d = scale->data<T>();

    T* dx_d = dx ? dx->mutable_data<T>(ctx.GetPlace()) : nullptr;
    T* ds_d = dscale ? dscale->mutable_data<T>(ctx.GetPlace()) : nullptr;
    T* db_d = dbias ? dbias->mutable_data<T>(ctx.GetPlace()) : nullptr;

#ifdef PADDLE_WITH_HIP
    const int block = 256;
#else
    const int block = 1024;
#endif  // PADDLE_WITH_HIP
    int max_threads = dev_ctx.GetMaxPhysicalThreadCount();
    const int max_blocks = std::max(max_threads / block, 1);
    int grid1 = (num + block - 1) / block;
    int grid2 = std::min(C, max_blocks);
    if (layout == framework::DataLayout::kNCHW) {
      if (dscale && dbias) {
        const T* x_d = x->data<T>();
        AffineChannelScaleBiasGradientCUDAKernel<T, block,
                                                 framework::DataLayout::kNCHW>
            <<<grid2, block, 0, dev_ctx.stream()>>>(dy_d, x_d, N, C, HxW, ds_d,
                                                    db_d);
      }
      if (dx) {
        KeAffineChannelCUDA<T, framework::DataLayout::kNCHW, false>
            <<<grid1, block, 0, dev_ctx.stream()>>>(dy_d, s_d, nullptr, C, HxW,
                                                    num, dx_d);
      }
    } else {
      if (dscale && dbias) {
        const T* x_d = x->data<T>();
        AffineChannelScaleBiasGradientCUDAKernel<T, block,
                                                 framework::DataLayout::kNHWC>
            <<<grid2, block, 0, dev_ctx.stream()>>>(dy_d, x_d, N, C, HxW, ds_d,
                                                    db_d);
      }

      if (dx) {
        KeAffineChannelCUDA<T, framework::DataLayout::kNHWC, false>
            <<<grid1, block, 0, dev_ctx.stream()>>>(dy_d, s_d, nullptr, C, HxW,
                                                    num, dx_d);
      }
    }
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
using CUDA = paddle::platform::CUDADeviceContext;

REGISTER_OP_CUDA_KERNEL(affine_channel,
                        ops::AffineChannelCUDAKernel<CUDA, float>,
                        ops::AffineChannelCUDAKernel<CUDA, double>);
REGISTER_OP_CUDA_KERNEL(affine_channel_grad,
                        ops::AffineChannelGradCUDAKernel<CUDA, float>,
                        ops::AffineChannelGradCUDAKernel<CUDA, double>);