adagrad_op.cu 4.4 KB
Newer Older
1 2
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

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
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
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. */
14 15 16

#define EIGEN_USE_GPU
#include "paddle/operators/adagrad_op.h"
Q
QI JUN 已提交
17
#include "paddle/operators/math/math_function.h"
18
#include "paddle/operators/math/selected_rows_functor.h"
Q
QI JUN 已提交
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
#include "paddle/platform/cuda_helper.h"

namespace paddle {
namespace operators {

namespace {

template <typename T, int block_size>
__global__ void MergeGradKernel(const T* grad, const int64_t* grad_rows,
                                T* grad_merge, const int64_t* grad_merge_rows,
                                size_t grad_merge_rows_size,
                                int64_t row_numel) {
  const int ty = blockIdx.y;
  int tid = threadIdx.x;
  __shared__ size_t grad_merge_idx;

  if (tid == 0) {
    for (size_t i = 0; i < grad_merge_rows_size; i++) {
      if (grad_rows[ty] == grad_merge_rows[i]) {
        grad_merge_idx = i;
      }
    }
  }

  __syncthreads();

  grad += ty * row_numel;
  grad_merge += grad_merge_idx * row_numel;
  for (int index = tid; index < row_numel; index += block_size) {
    paddle::platform::CudaAtomicAdd(grad_merge + index, grad[index]);
  }
}

template <typename T, int block_size>
__global__ void SparseAdagradFunctorKernel(const T* grad, const int64_t* rows,
                                           const T* learning_rate, T* param,
                                           T* moment, int64_t row_numel,
                                           T epsilon) {
  const int ty = blockIdx.y;
  int tid = threadIdx.x;

  grad += ty * row_numel;
  param += rows[ty] * row_numel;
  moment += rows[ty] * row_numel;

  for (int index = tid; index < row_numel; index += block_size) {
    // Since index in rows of SelectedRows can be duplicate, we have to use
    // Atomic Operation to avoid concurrent write error.
    paddle::platform::CudaAtomicAdd(param + index,
                                    -1.0 * learning_rate[0] * grad[index] /
                                        (sqrt(moment[index]) + epsilon));
  }
}
}  // namespace

template <typename T>
Q
QI JUN 已提交
75 76
struct SparseAdagradFunctor<platform::CUDADeviceContext, T> {
  void operator()(const platform::CUDADeviceContext& context,
Q
QI JUN 已提交
77 78 79 80 81
                  const framework::SelectedRows& grad,
                  const framework::Tensor& learning_rate, T epsilon,
                  framework::Tensor* moment, framework::Tensor* param) {
    // 1. g_m.rows = set(g.rows)
    auto grad_width = grad.value().dims()[1];
T
typhoonzero 已提交
82
    math::scatter::MergeAdd<platform::CUDADeviceContext, T> merge_func;
T
wip  
typhoonzero 已提交
83 84
    auto grad_merge = merge_func(context, grad);
    auto* grad_merge_data = grad_merge.mutable_value()->template data<T>();
D
dzhwinter 已提交
85
    framework::Vector<int64_t> merge_rows(grad_merge.rows());
Q
QI JUN 已提交
86
    // 2. m += g_m * g_m
T
typhoonzero 已提交
87
    math::scatter::Mul<platform::CUDADeviceContext, T> sqare_func;
T
wip  
typhoonzero 已提交
88
    auto grad_square = sqare_func(context, grad_merge, grad_merge);
Q
QI JUN 已提交
89

Q
QI JUN 已提交
90
    math::SelectedRowsAddToTensor<platform::CUDADeviceContext, T> functor;
T
wip  
typhoonzero 已提交
91
    functor(context, grad_square, moment);
Q
QI JUN 已提交
92 93 94 95 96 97

    // 3. update parameter
    auto* lr = learning_rate.data<T>();
    auto* param_data = param->data<T>();
    auto* moment_data = moment->data<T>();

T
typhoonzero 已提交
98 99
    const int block_size = 256;
    dim3 threads(block_size, 1);
Q
QI JUN 已提交
100 101 102 103
    dim3 grid2(1, merge_rows.size());
    SparseAdagradFunctorKernel<
        T, 256><<<grid2, threads, 0,
                  reinterpret_cast<const platform::CUDADeviceContext&>(context)
D
dzhwinter 已提交
104 105
                      .stream()>>>(grad_merge_data, merge_rows.cuda_data(), lr,
                                   param_data, moment_data, grad_width,
106
                                   epsilon);
Q
QI JUN 已提交
107 108 109
  }
};

Q
QI JUN 已提交
110 111
template struct SparseAdagradFunctor<platform::CUDADeviceContext, float>;
template struct SparseAdagradFunctor<platform::CUDADeviceContext, double>;
Q
QI JUN 已提交
112 113 114

}  // namespace operators
}  // namespace paddle
115 116

namespace ops = paddle::operators;
Q
QI JUN 已提交
117 118 119
REGISTER_OP_CUDA_KERNEL(
    adagrad, ops::AdagradOpKernel<paddle::platform::CUDADeviceContext, float>,
    ops::AdagradOpKernel<paddle::platform::CUDADeviceContext, double>);