top_k_op.cu 6.9 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
武毅 已提交
2

L
Luo Tao 已提交
3 4 5
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
武毅 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
武毅 已提交
8

L
Luo Tao 已提交
9 10 11 12 13
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. */
武毅 已提交
14

W
wawltor 已提交
15
#pragma once
16
#include <cstdio>
W
wawltor 已提交
17
#include <vector>
18
#ifdef __NVCC__
19
#include "cub/cub.cuh"
20 21 22 23
#endif
#ifdef __HIPCC__
#include <hipcub/hipcub.hpp>
#endif
Y
Yi Wang 已提交
24
#include "paddle/fluid/framework/op_registry.h"
W
wawltor 已提交
25
#include "paddle/fluid/operators/top_k_function_cuda.h"
26
#include "paddle/fluid/operators/top_k_op.h"
W
Wu Yi 已提交
27
#include "paddle/fluid/platform/float16.h"
28 29
// set cub base traits in order to handle float16

武毅 已提交
30 31 32 33 34
namespace paddle {
namespace operators {

using Tensor = framework::Tensor;

35 36 37 38 39 40 41 42 43 44 45 46
#define FIXED_BLOCK_DIM_BASE(dim, ...) \
  case (dim): {                        \
    constexpr auto kBlockDim = (dim);  \
    __VA_ARGS__;                       \
  } break

#define FIXED_BLOCK_DIM(...)                \
  FIXED_BLOCK_DIM_BASE(256, ##__VA_ARGS__); \
  FIXED_BLOCK_DIM_BASE(128, ##__VA_ARGS__); \
  FIXED_BLOCK_DIM_BASE(64, ##__VA_ARGS__);  \
  FIXED_BLOCK_DIM_BASE(32, ##__VA_ARGS__)

47
template <typename DeviceContext, typename T>
Y
Yu Yang 已提交
48
class TopkOpCUDAKernel : public framework::OpKernel<T> {
武毅 已提交
49 50
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
J
Jiawei Wang 已提交
51 52 53
    PADDLE_ENFORCE_EQ(
        platform::is_gpu_place(ctx.GetPlace()), true,
        platform::errors::InvalidArgument("It must use CUDAPlace."));
武毅 已提交
54 55 56
    auto* input = ctx.Input<Tensor>("X");
    auto* output = ctx.Output<Tensor>("Out");
    auto* indices = ctx.Output<Tensor>("Indices");
57
    int k = static_cast<int>(ctx.Attr<int>("k"));
武毅 已提交
58

W
whs 已提交
59 60 61 62 63 64 65 66 67 68 69
    auto* k_t = ctx.Input<Tensor>("K");
    if (k_t) {
      Tensor k_host;
      framework::TensorCopySync(*k_t, platform::CPUPlace(), &k_host);
      k = k_host.data<int>()[0];
      framework::DDim output_dims = output->dims();
      output_dims[output_dims.size() - 1] = k;
      output->Resize(output_dims);
      indices->Resize(output_dims);
    }

武毅 已提交
70 71 72 73
    const T* input_data = input->data<T>();
    T* output_data = output->mutable_data<T>(ctx.GetPlace());
    // FIXME(typhoonzero): data is always converted to type T?

Q
qingqing01 已提交
74
    framework::DDim inputdims = input->dims();
75
    const int64_t input_height = framework::product(
Q
qingqing01 已提交
76
        framework::slice_ddim(inputdims, 0, inputdims.size() - 1));
77
    const int64_t input_width = inputdims[inputdims.size() - 1];
78 79 80 81 82 83 84 85 86 87 88 89
    const auto& dev_ctx = ctx.cuda_device_context();
    if ((input_width <= 1024 || k >= 128 || k == input_width)) {
      if (SortTopk<T>(dev_ctx, input, input_width, input_height, k, output,
                      indices)) {
        // Successed, return.
        return;
      } else {
        LOG(INFO) << "TopKOP: Some errors happened when use cub sorting, use "
                     "default topk kernel.";
      }
    }
    int64_t* indices_data = indices->mutable_data<int64_t>(ctx.GetPlace());
武毅 已提交
90 91 92 93 94
    if (k > input_width) k = input_width;

    // NOTE: pass lds and dim same to input width.
    // NOTE: old matrix implementation of stride is different to eigen.
    // TODO(typhoonzero): refine this kernel.
95 96 97 98 99 100
    const int kMaxHeight = 2048;
    int gridx = input_height < kMaxHeight ? input_height : kMaxHeight;
    switch (GetDesiredBlockDim(input_width)) {
      FIXED_BLOCK_DIM(
          KeMatrixTopK<T, 5,
                       kBlockDim><<<gridx, kBlockDim, 0, dev_ctx.stream()>>>(
Q
qingqing01 已提交
101 102
              output_data, k, indices_data, input_data, input_width,
              input_width, static_cast<int>(k), gridx, input_height));
103
      default:
104 105
        PADDLE_THROW(platform::errors::Unavailable(
            "Calculation error occurred in TopK Operator's CUDA Kernel."));
106
    }
武毅 已提交
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
template <typename DeviceContext, typename T>
class TopkOpGradCUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    PADDLE_ENFORCE_EQ(
        platform::is_gpu_place(context.GetPlace()), true,
        platform::errors::InvalidArgument("It must use CUDAPlace."));
    auto* x = context.Input<Tensor>("X");
    auto* out_grad = context.Input<Tensor>(framework::GradVarName("Out"));
    auto* indices = context.Input<Tensor>("Indices");
    auto* x_grad = context.Output<Tensor>(framework::GradVarName("X"));

    T* x_grad_data = x_grad->mutable_data<T>(context.GetPlace());
    const T* out_grad_data = out_grad->data<T>();
    const int64_t* indices_data = indices->data<int64_t>();
    size_t k = indices->dims()[indices->dims().size() - 1];

    framework::DDim xdims = x->dims();
    const size_t row =
        framework::product(framework::slice_ddim(xdims, 0, xdims.size() - 1));
    const size_t col = xdims[xdims.size() - 1];
    const auto& dev_ctx = context.cuda_device_context();
    const int kMaxHeight = 2048;
    int gridx = row < kMaxHeight ? row : kMaxHeight;
    switch (GetDesiredBlockDim(col)) {
      FIXED_BLOCK_DIM(
          AssignGrad<T, 5,
                     kBlockDim><<<gridx, kBlockDim, 0, dev_ctx.stream()>>>(
              x_grad_data, indices_data, out_grad_data, row, col, k));
      default:
        PADDLE_THROW(
            platform::errors::Unavailable("Error occurs when Assign Grad."));
    }
  }
};
145 146 147
#undef FIXED_BLOCK_DIM_BASE
#undef FIXED_BLOCK_DIM

武毅 已提交
148 149
}  // namespace operators
}  // namespace paddle
W
Wu Yi 已提交
150
REGISTER_OP_CUDA_KERNEL(
151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174
    top_k,
    paddle::operators::TopkOpCUDAKernel<paddle::platform::CUDADeviceContext,
                                        float>,
    paddle::operators::TopkOpCUDAKernel<paddle::platform::CUDADeviceContext,
                                        double>,
    paddle::operators::TopkOpCUDAKernel<paddle::platform::CUDADeviceContext,
                                        int>,
    paddle::operators::TopkOpCUDAKernel<paddle::platform::CUDADeviceContext,
                                        int64_t>,
    paddle::operators::TopkOpCUDAKernel<paddle::platform::CUDADeviceContext,
                                        paddle::platform::float16>);

REGISTER_OP_CUDA_KERNEL(
    top_k_grad,
    paddle::operators::TopkOpGradCUDAKernel<paddle::platform::CUDADeviceContext,
                                            float>,
    paddle::operators::TopkOpGradCUDAKernel<paddle::platform::CUDADeviceContext,
                                            double>,
    paddle::operators::TopkOpGradCUDAKernel<paddle::platform::CUDADeviceContext,
                                            int>,
    paddle::operators::TopkOpGradCUDAKernel<paddle::platform::CUDADeviceContext,
                                            int64_t>,
    paddle::operators::TopkOpGradCUDAKernel<paddle::platform::CUDADeviceContext,
                                            paddle::platform::float16>);