temporal_shift_grad_kernel.cu 5.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// Copyright (c) 2022 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.

15 16
#include "paddle/phi/kernels/temporal_shift_grad_kernel.h"

17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/common/layout.h"
#include "paddle/phi/core/kernel_registry.h"

namespace phi {

template <typename T>
__global__ void KeTemporalShiftBwNCHW(const T* output_grad,
                                      T* input_grad,
                                      const int ntchw,
                                      const int tchw,
                                      const int chw,
                                      const int hw,
                                      const int t,
                                      const int c1,
                                      const int c2) {
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  int src_it = 0;

  for (; tid < ntchw; tid += stride) {
    int it = (tid % tchw) / chw;
    int ic = (tid % chw) / hw;

    if (ic < c1) {
      src_it = it + 1;
    } else if (ic < c2) {
      src_it = it - 1;
    } else {
      src_it = it;
    }

    if (src_it >= 0 && src_it < t) {
      input_grad[tid] = output_grad[tid + (src_it - it) * chw];
    } else {
      input_grad[tid] = 0;
    }
  }
}

template <typename T>
__global__ void KeTemporalShiftBwNHWC(const T* output_grad,
                                      T* input_grad,
                                      const int nthwc,
                                      const int thwc,
                                      const int hwc,
                                      const int t,
                                      const int c,
                                      const int c1,
                                      const int c2) {
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  int src_it = 0;

  for (; tid < nthwc; tid += stride) {
    int it = (tid % thwc) / hwc;
    int ic = tid % c;

    if (ic < c1) {
      src_it = it + 1;
    } else if (ic < c2) {
      src_it = it - 1;
    } else {
      src_it = it;
    }

    if (src_it >= 0 && src_it < t) {
      input_grad[tid] = output_grad[tid + (src_it - it) * hwc];
    } else {
      input_grad[tid] = 0;
    }
  }
}

template <typename T, typename Context>
void TemporalShiftGradKernel(const Context& dev_ctx,
                             const DenseTensor& out_grad,
                             int seg_num,
                             float shift_ratio,
                             const std::string& data_format_str,
                             DenseTensor* x_grad) {
  auto* input_grad = x_grad;
  auto* output_grad = &out_grad;
  int t = seg_num;
  const DataLayout data_layout =
      paddle::framework::StringToDataLayout(data_format_str);

  const int nt = output_grad->dims()[0];
  const int c = (data_layout == DataLayout::kNCHW ? output_grad->dims()[1]
                                                  : output_grad->dims()[3]);
  const int h = (data_layout == DataLayout::kNCHW ? output_grad->dims()[2]
                                                  : output_grad->dims()[1]);
  const int w = (data_layout == DataLayout::kNCHW ? output_grad->dims()[3]
                                                  : output_grad->dims()[2]);

  const int hw = h * w;
  const int chw = c * hw;
  const int tchw = t * chw;
  const int ntchw = nt * chw;

  const int c1 = static_cast<int>(c * shift_ratio);
  const int c2 = static_cast<int>(c * 2 * shift_ratio);

  DDim in_grad_dims =
      (data_layout == DataLayout::kNCHW ? phi::make_ddim({nt, c, h, w})
                                        : phi::make_ddim({nt, h, w, c}));
  const T* output_grad_data = output_grad->data<T>();
  T* input_grad_data =
      input_grad->mutable_data<T>(in_grad_dims, dev_ctx.GetPlace());

  int pixelNum = nt * chw;
  int threads = 1024;
  int grid = (pixelNum + threads - 1) / threads;
  int blocks_per_sm = dev_ctx.GetMaxPhysicalThreadCount() / threads;
  grid = std::min(dev_ctx.GetSMCount() * blocks_per_sm, grid);

  if (data_layout == DataLayout::kNCHW) {
    KeTemporalShiftBwNCHW<T><<<grid, threads, 0, dev_ctx.stream()>>>(
        output_grad_data, input_grad_data, ntchw, tchw, chw, hw, t, c1, c2);
  } else {
    KeTemporalShiftBwNHWC<T><<<grid, threads, 0, dev_ctx.stream()>>>(
        output_grad_data, input_grad_data, ntchw, tchw, chw, t, c, c1, c2);
  }
}

}  // namespace phi

PD_REGISTER_KERNEL(temporal_shift_grad,
                   GPU,
                   ALL_LAYOUT,
                   phi::TemporalShiftGradKernel,
                   float,
                   double,
                   phi::dtype::float16) {}