pool_grad_kernel.cu 5.9 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
#include "paddle/phi/kernels/sparse/pool_grad_kernel.h"
16

17 18 19 20
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/backends/gpu/gpu_info.h"
#include "paddle/phi/backends/gpu/gpu_launch_config.h"
#include "paddle/phi/core/kernel_registry.h"
21
#include "paddle/phi/core/tensor_utils.h"
22
#include "paddle/phi/core/visit_type.h"
23
#include "paddle/phi/kernels/empty_kernel.h"
24 25 26 27 28 29 30
#include "paddle/phi/kernels/funcs/math_function.h"
#include "paddle/phi/kernels/funcs/pooling.h"
#include "paddle/phi/kernels/funcs/sparse/convolution.h"

namespace phi {
namespace sparse {

31
template <typename T, typename IntT = int>
32 33 34
__global__ void MaxPoolGradCudaKernel(const T* in_features_ptr,
                                      const T* out_features_ptr,
                                      const T* out_grad_ptr,
35
                                      const IntT* rulebook_ptr,
36 37 38 39 40 41 42 43
                                      const int n,
                                      const int rulebook_len,
                                      const int channels,
                                      T* x_grad_ptr) {
  phi::funcs::MaxPoolGrad<T> grad_functor;
  CUDA_KERNEL_LOOP_TYPE(i, n * channels, int64_t) {
    int real_i = i / channels;
    int c = i - real_i * channels;
44 45
    IntT in_i = rulebook_ptr[real_i];
    IntT out_i = rulebook_ptr[real_i + rulebook_len];
46 47 48 49 50 51 52 53
    grad_functor.compute(in_features_ptr[in_i * channels + c],
                         out_features_ptr[out_i * channels + c],
                         out_grad_ptr[out_i * channels + c],
                         1,
                         &x_grad_ptr[in_i * channels + c]);
  }
}

54
template <typename T, typename IntT = int>
55 56 57 58 59 60 61
void MaxPoolCooGradGPUKernel(const GPUContext& dev_ctx,
                             const SparseCooTensor& x,
                             const DenseTensor& rulebook,
                             const SparseCooTensor& out,
                             const SparseCooTensor& out_grad,
                             const std::vector<int>& kernel_sizes,
                             SparseCooTensor* x_grad) {
62 63 64
  int kernel_size = kernel_sizes[0] * kernel_sizes[1] * kernel_sizes[2];
  const int in_channels = x.dims()[4];
  int rulebook_len = rulebook.dims()[1];
65 66
  const IntT* rulebook_ptr = rulebook.data<IntT>();
  std::vector<IntT> offsets(kernel_size + 1), counter(kernel_size, 0),
Z
zhangkaihuo 已提交
67
      h_counter(rulebook_len, 0);
68 69
  phi::backends::gpu::GpuMemcpyAsync(&h_counter[0],
                                     rulebook_ptr,
70
                                     rulebook_len * sizeof(IntT),
71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
#ifdef PADDLE_WITH_HIP
                                     hipMemcpyDeviceToHost,
#else
                                     cudaMemcpyDeviceToHost,
#endif

                                     dev_ctx.stream());
  dev_ctx.Wait();
  for (int i = 0; i < rulebook_len; i++) {
    counter[h_counter[i]] += 1;
  }
  phi::funcs::sparse::PrefixSum(&counter[0], &offsets[0], kernel_size);

  const T* in_features_ptr = x.non_zero_elements().data<T>();
  const T* out_features_ptr = out.non_zero_elements().data<T>();
86 87 88 89 90 91 92 93 94 95 96 97 98 99
  const T* out_grad_ptr = out_grad.non_zero_elements().data<T>();
  // TODO(zhangkaihuo): call phi::sparse::EmptyLike
  DenseTensor x_grad_indices =
      phi::EmptyLike<IntT>(dev_ctx, x.non_zero_indices());
  DenseTensor x_grad_values = phi::EmptyLike<T>(dev_ctx, x.non_zero_elements());
  x_grad->SetMember(x_grad_indices, x_grad_values, x.dims(), true);
  T* x_grad_ptr = x_grad_values.data<T>();
  phi::funcs::SetConstant<GPUContext, T> set_zero;
  set_zero(dev_ctx, &x_grad_values, static_cast<T>(0.0f));
  phi::Copy<GPUContext>(dev_ctx,
                        x.non_zero_indices(),
                        dev_ctx.GetPlace(),
                        false,
                        &x_grad_indices);
100 101 102 103 104 105 106 107

  for (int i = 0; i < kernel_size; i++) {
    if (counter[i] <= 0) {
      continue;
    }

    auto config = phi::backends::gpu::GetGpuLaunchConfig1D(
        dev_ctx, counter[i] * in_channels, 1);
108 109 110 111 112 113 114 115 116 117 118 119
    MaxPoolGradCudaKernel<T, IntT>
        <<<config.block_per_grid.x,
           config.thread_per_block.x,
           0,
           dev_ctx.stream()>>>(in_features_ptr,
                               out_features_ptr,
                               out_grad_ptr,
                               rulebook_ptr + offsets[i] + rulebook_len,
                               counter[i],
                               rulebook_len,
                               in_channels,
                               x_grad_ptr);
120 121 122
  }
}

123
template <typename T, typename Context>
124 125 126 127 128 129 130
void MaxPoolCooGradKernel(const Context& dev_ctx,
                          const SparseCooTensor& x,
                          const DenseTensor& rulebook,
                          const SparseCooTensor& out,
                          const SparseCooTensor& out_grad,
                          const std::vector<int>& kernel_sizes,
                          SparseCooTensor* x_grad) {
131
  PD_VISIT_INTEGRAL_TYPES(
132 133
      x.non_zero_indices().dtype(), "MaxPoolCooGradGPUKernel", ([&] {
        MaxPoolCooGradGPUKernel<T, data_t>(
134 135 136 137
            dev_ctx, x, rulebook, out, out_grad, kernel_sizes, x_grad);
      }));
}

138 139 140
}  // namespace sparse
}  // namespace phi

141
PD_REGISTER_KERNEL(maxpool_coo_grad,
142 143
                   GPU,
                   ALL_LAYOUT,
144
                   phi::sparse::MaxPoolCooGradKernel,
145 146 147 148
                   float,
                   double) {
  kernel->InputAt(0).SetDataLayout(phi::DataLayout::SPARSE_COO);
}
反馈
建议
客服 返回
顶部