elementwise_add_op.cu 9.1 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
G
gongweibao 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
G
gongweibao 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
G
gongweibao 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. */
W
Wu Yi 已提交
14
#include "paddle/fluid/operators/elementwise/elementwise_add_op.h"
15
#include "paddle/fluid/operators/elementwise/elementwise_op_broadcast.cu.h"
16 17
#include "paddle/fluid/operators/reduce_ops/reduce_functor_op.h"
#include "paddle/fluid/operators/reduce_ops/reduce_op.cu.h"
18
#include "paddle/fluid/platform/complex.h"
K
Kexin Zhao 已提交
19
#include "paddle/fluid/platform/float16.h"
G
gongweibao 已提交
20 21

namespace ops = paddle::operators;
K
Kexin Zhao 已提交
22
namespace plat = paddle::platform;
G
gongweibao 已提交
23

24 25 26
namespace paddle {
namespace operators {

27 28 29 30 31 32
/*
   input: an array;
   return: the result of the math functor
   1. For Unary Op, the length of input array is 1,
      e.g. Relu: return args[0] > 0 ? args[0] : 0;
   2. For Binary Op, the length of input array is 2,
33
      e.g. Add: return args[0] expr args[1];
34
*/
35
template <typename T>
36
struct CudaAddFunctor {
37
  inline HOSTDEVICE T operator()(const T* args) const {
38 39
    return args[0] + args[1];
  }
40 41
};

42
template <typename T>
43 44 45 46
class ElementwiseAddKernel<platform::CUDADeviceContext, T>
    : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
47 48
    std::vector<const framework::Tensor*> ins;
    std::vector<framework::Tensor*> outs;
49 50 51 52
    const auto& cuda_ctx =
        ctx.template device_context<platform::CUDADeviceContext>();

    int axis = PackTensorsIntoVector<T>(ctx, &ins, &outs);
53
    LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
54
        cuda_ctx, ins, &outs, axis, CudaAddFunctor<T>());
55 56 57
  }
};

58
template <typename T>
59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74
static __global__ void SimpleElemwiseAddGradCUDAKernel(
    const T* __restrict__ dout, int size, int vec_size, T* dx, T* dy) {
  int tid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = gridDim.x * blockDim.x;
  int loop = size / vec_size;
  int remainder = size % vec_size;
  const float4* dout_vec = reinterpret_cast<const float4*>(dout);
  float4* dx_vec = reinterpret_cast<float4*>(dx);
  float4* dy_vec = reinterpret_cast<float4*>(dy);
  float4 tmp_loop;

  for (int i = tid; i < loop; i += stride) {
    tmp_loop = dout_vec[i];
    dx_vec[i] = tmp_loop;
    dy_vec[i] = tmp_loop;
  }
75

76 77 78 79 80 81 82 83 84
  if (tid == loop && remainder != 0) {
    T tmp_rem;
    while (remainder) {
      int idx = size - remainder;
      remainder--;
      tmp_rem = dout[idx];
      dx[idx] = tmp_rem;
      dy[idx] = tmp_rem;
    }
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
template <typename DeviceContext, typename T>
typename std::enable_if<
    std::is_same<DeviceContext, platform::CUDADeviceContext>::value>::type
default_elementwise_add_grad(const framework::ExecutionContext& ctx,
                             const framework::Tensor* x,
                             const framework::Tensor* y,
                             const framework::Tensor* out,
                             const framework::Tensor* dout,
                             framework::Tensor* dx, framework::Tensor* dy) {
  int axis = ctx.Attr<int>("axis");
  auto* dout_data = dout->data<T>();

  // dx
  if (dx != nullptr) {
    auto* dx_data = dx->mutable_data<T>(ctx.GetPlace());
    if (dx->dims() == dout->dims()) {
      if (dx_data != dout_data) {
        framework::TensorCopy(
            *dout, ctx.GetPlace(),
            ctx.template device_context<platform::DeviceContext>(), dx);
      }
    } else {
      // For inplace strategy, dx will be stored in addr of dout, which makes
      // the result of dy wrong.
      if (dx->IsSharedBufferWith(*dout)) {
        dx->clear();
        dx->mutable_data<T>(x->dims(), ctx.GetPlace());
      }
      std::vector<int> reduce_dims = GetReduceDim(x->dims(), out->dims(), axis);
      gpuStream_t stream = ctx.cuda_device_context().stream();
      TensorReduceFunctorImpl<T, T, CustomSum>(*dout, dx, reduce_dims, stream);
    }
  }
  // dy
  if (dy != nullptr) {
    auto* dy_data = dy->mutable_data<T>(ctx.GetPlace());
    if (dy->dims() == dout->dims()) {
      if (dy_data != dout_data) {
        framework::TensorCopy(
            *dout, ctx.GetPlace(),
            ctx.template device_context<platform::DeviceContext>(), dy);
      }
    } else {
      std::vector<int> reduce_dims = GetReduceDim(y->dims(), out->dims(), axis);
      gpuStream_t stream = ctx.cuda_device_context().stream();
      TensorReduceFunctorImpl<T, T, CustomSum>(*dout, dy, reduce_dims, stream);
    }
  }
}

138 139 140
template <typename DeviceContext, typename T>
typename std::enable_if<
    std::is_same<DeviceContext, plat::CUDADeviceContext>::value>::type
141 142 143 144 145
elementwise_add_grad(const framework::ExecutionContext& ctx,
                     const framework::Tensor* x, const framework::Tensor* y,
                     const framework::Tensor* out,
                     const framework::Tensor* dout, framework::Tensor* dx,
                     framework::Tensor* dy) {
146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178
  auto* dx_data = dx->mutable_data<T>(ctx.GetPlace());
  auto* dy_data = dy->mutable_data<T>(ctx.GetPlace());
  auto* dout_data = dout->data<T>();
  if (dx_data == dout_data && dy_data != dout_data) {
    VLOG(4) << "Special case when dx_data is the same as dout_data, "
               "only need copy dout to dy";
    framework::TensorCopy(
        *dout, ctx.GetPlace(),
        ctx.template device_context<platform::DeviceContext>(), dy);
  } else if (dx_data != dout_data && dy_data == dout_data) {
    VLOG(4) << "Special case when dy_data is the same as dout_data, "
               "only need copy dout to dx";
    framework::TensorCopy(
        *dout, ctx.GetPlace(),
        ctx.template device_context<platform::DeviceContext>(), dx);
  } else if (dx_data != dout_data && dy_data != dout_data) {
    auto size = x->numel();
    int vec_size = max(static_cast<int>(sizeof(float4) / sizeof(T)), 1);
    dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
    dim3 grid_size =
        dim3(((size + vec_size - 1) / vec_size + PADDLE_CUDA_THREAD_SIZE - 1) /
                 PADDLE_CUDA_THREAD_SIZE,
             1);
    SimpleElemwiseAddGradCUDAKernel<
        T><<<grid_size, block_size, 0,
             ctx.template device_context<plat::CUDADeviceContext>().stream()>>>(
        dout->data<T>(), size, vec_size, dx->mutable_data<T>(ctx.GetPlace()),
        dy->mutable_data<T>(ctx.GetPlace()));
  } else {
    VLOG(4) << "Special case when dy_data is the same as dout_data, "
               "and dx_data is the same as dout_data, do not need "
               "any operator";
  }
179 180 181 182
}

}  // namespace operators
}  // namespace paddle
Q
QI JUN 已提交
183
REGISTER_OP_CUDA_KERNEL(
K
Kexin Zhao 已提交
184 185 186
    elementwise_add, ops::ElementwiseAddKernel<plat::CUDADeviceContext, float>,
    ops::ElementwiseAddKernel<plat::CUDADeviceContext, double>,
    ops::ElementwiseAddKernel<plat::CUDADeviceContext, int>,
K
Kexin Zhao 已提交
187
    ops::ElementwiseAddKernel<plat::CUDADeviceContext, int64_t>,
188
    ops::ElementwiseAddKernel<plat::CUDADeviceContext, plat::float16>,
189 190
    ops::ElementwiseAddKernel<plat::CUDADeviceContext, plat::complex<float>>,
    ops::ElementwiseAddKernel<plat::CUDADeviceContext, plat::complex<double>>);
Q
QI JUN 已提交
191
REGISTER_OP_CUDA_KERNEL(
G
gongweibao 已提交
192
    elementwise_add_grad,
K
Kexin Zhao 已提交
193 194 195
    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, float>,
    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, double>,
    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, int>,
C
chengduo 已提交
196
    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, int64_t>,
197
    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, plat::float16>,
198 199 200 201
    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext,
                                  plat::complex<float>>,
    ops::ElementwiseAddGradKernel<plat::CUDADeviceContext,
                                  plat::complex<double>>);
202 203 204 205 206
REGISTER_OP_CUDA_KERNEL(
    elementwise_add_grad_grad,
    ops::ElementwiseAddDoubleGradKernel<plat::CUDADeviceContext, float>,
    ops::ElementwiseAddDoubleGradKernel<plat::CUDADeviceContext, double>,
    ops::ElementwiseAddDoubleGradKernel<plat::CUDADeviceContext, int>,
207
    ops::ElementwiseAddDoubleGradKernel<plat::CUDADeviceContext, int64_t>,
208
    ops::ElementwiseAddDoubleGradKernel<plat::CUDADeviceContext, plat::float16>,
209
    ops::ElementwiseAddDoubleGradKernel<plat::CUDADeviceContext,
210
                                        plat::complex<float>>,
211
    ops::ElementwiseAddDoubleGradKernel<plat::CUDADeviceContext,
212
                                        plat::complex<double>>);
213 214 215 216 217 218

REGISTER_OP_CUDA_KERNEL(
    grad_add, ops::ElementwiseAddKernel<plat::CUDADeviceContext, float>,
    ops::ElementwiseAddKernel<plat::CUDADeviceContext, double>,
    ops::ElementwiseAddKernel<plat::CUDADeviceContext, int>,
    ops::ElementwiseAddKernel<plat::CUDADeviceContext, int64_t>,
219
    ops::ElementwiseAddKernel<plat::CUDADeviceContext, plat::float16>,
220 221
    ops::ElementwiseAddKernel<plat::CUDADeviceContext, plat::complex<float>>,
    ops::ElementwiseAddKernel<plat::CUDADeviceContext, plat::complex<double>>);