sgd_op.cu 4.2 KB
Newer Older
L
liaogang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

   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. */

Q
qijun 已提交
15
#define EIGEN_USE_GPU
Q
Qiao Longfei 已提交
16
#include "paddle/operators/sgd_op.h"
Q
qijun 已提交
17 18 19 20 21 22
#include "paddle/platform/cuda_helper.h"

namespace paddle {
namespace operators {

namespace {
C
chengduoZH 已提交
23 24 25 26 27 28 29 30 31 32 33 34 35

template <typename T>
__global__ void SGDKernel(const T* g, const T* p, const T* learning_rate,
                          const int num, T* p_out) {
  T lr = learning_rate[0];
  int grid_size = blockDim.x * gridDim.x;
  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; i += grid_size) {
    T g_data = g[i];
    T p_data = p[i];
    p_out[i] = p_data - lr * g_data;
  }
}

Q
QI JUN 已提交
36
template <typename T, int block_size>
Q
qijun 已提交
37 38 39
__global__ void SparseSGDFunctorKernel(const T* selected_rows,
                                       const int64_t* rows,
                                       const T* learning_rate, T* tensor_out,
Q
QI JUN 已提交
40
                                       int64_t row_numel) {
Q
qijun 已提交
41 42 43 44 45 46 47 48 49
  const int ty = blockIdx.y;
  int tid = threadIdx.x;

  selected_rows += ty * row_numel;
  tensor_out += 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.
Q
qijun 已提交
50 51
    paddle::platform::CudaAtomicAdd(
        tensor_out + index, -1.0 * learning_rate[0] * selected_rows[index]);
Q
qijun 已提交
52 53 54 55 56
  }
}
}  // namespace

template <typename T>
C
chengduoZH 已提交
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
class SGDOpCUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* param = ctx.Input<framework::Tensor>("Param");
    auto* param_out = ctx.Output<framework::Tensor>("ParamOut");
    auto* learning_rate = ctx.Input<framework::Tensor>("LearningRate");

    auto* grad_var = ctx.InputVar("Grad");
    // Actually, all tensors are LoDTensor except SelectedRows.
    if (grad_var->IsType<framework::LoDTensor>()) {
      param_out->mutable_data<T>(ctx.GetPlace());
      auto* grad = ctx.Input<framework::Tensor>("Grad");
      auto* grad_data = grad->data<T>();
      auto* param_data = param->data<T>();
      auto* param_out_data = param_out->data<T>();

      int block = 512;
      int grid = (param->numel() + block - 1) / block;

      SGDKernel<T><<<grid, block, 0, ctx.cuda_device_context().stream()>>>(
          grad_data, param_data, learning_rate->data<T>(), param->numel(),
          param_out_data);

    } else if (grad_var->IsType<framework::SelectedRows>()) {
      // TODO(qijun): In Sparse SGD operator, in-place update is enforced.
      // This manual optimization brings difficulty to track data dependency.
      // It's better to find a more elegant solution.
      PADDLE_ENFORCE_EQ(param, param_out);
      auto* grad = ctx.Input<framework::SelectedRows>("Grad");

      auto in_height = grad->height();
      auto out_dims = param_out->dims();
      PADDLE_ENFORCE_EQ(in_height, out_dims[0]);

      auto& in_value = grad->value();
      auto& in_rows = grad->rows();

      int64_t in_row_numel = in_value.numel() / in_rows.size();
      PADDLE_ENFORCE_EQ(in_row_numel, param_out->numel() / in_height);

      auto* in_data = in_value.data<T>();
      auto* out_data = param_out->data<T>();

      const int block_size = 256;
      dim3 threads(block_size, 1);
      dim3 grid(1, in_rows.size());
      SparseSGDFunctorKernel<
          T, 256><<<grid, threads, 0, ctx.cuda_device_context().stream()>>>(
          in_data, in_rows.data(), learning_rate->data<T>(), out_data,
          in_row_numel);

    } else {
      PADDLE_THROW("Unsupported Variable Type of Grad");
    }
Q
qijun 已提交
111 112 113 114
  }
};
}  // namespace operators
}  // namespace paddle
Q
Qiao Longfei 已提交
115

D
dongzhihong 已提交
116
namespace ops = paddle::operators;
C
chengduoZH 已提交
117 118
REGISTER_OP_CUDA_KERNEL(sgd, ops::SGDOpCUDAKernel<float>,
                        ops::SGDOpCUDAKernel<double>);