elementwise_add_op_npu.cc 5.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
/* Copyright (c) 2021 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. */

#include <memory>
#include <string>

18
#include "paddle/fluid/framework/tensor_util.h"
19
#include "paddle/fluid/operators/elementwise/elementwise_add_op.h"
20
#include "paddle/fluid/operators/elementwise/elementwise_npu.h"
21 22 23 24
#include "paddle/fluid/operators/npu_op_runner.h"

namespace paddle {
namespace operators {
25
using Tensor = framework::Tensor;
26

27
template <typename T>
28 29 30
class ElementwiseAddNPUKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
31 32
    auto& dev_ctx =
        ctx.template device_context<paddle::platform::NPUDeviceContext>();
33 34 35 36 37
    auto* x = ctx.Input<framework::LoDTensor>("X");
    auto* y = ctx.Input<framework::LoDTensor>("Y");
    auto* out = ctx.Output<framework::LoDTensor>("Out");
    out->mutable_data<T>(ctx.GetPlace());

38 39 40 41 42 43 44
    int axis = ctx.Attr<int>("axis");

    bool direct_compute = false;
    auto x_dims = x->dims();
    auto y_dims = y->dims();
    axis = (axis == -1 ? std::abs(x_dims.size() - y_dims.size()) : axis);
    if (x_dims.size() >= y_dims.size()) {
45
      direct_compute = x_dims.size() == (y_dims.size() + axis);
46
    } else {
47
      direct_compute = y_dims.size() == (x_dims.size() + axis);
48 49 50
    }

    if (direct_compute) {
51 52
      const auto& runner = NpuOpRunner("Add", {*x, *y}, {*out}, {});
      runner.Run(dev_ctx.stream());
53
    } else {
54
      Tensor transformed_x, transformed_y;
55 56
      NpuElementWiseOpBroadcast<T>(dev_ctx, x, y, axis, &transformed_x,
                                   &transformed_y);
57 58 59
      const auto& runner =
          NpuOpRunner("Add", {transformed_x, transformed_y}, {*out}, {});
      runner.Run(dev_ctx.stream());
60
    }
61 62 63
  }
};

64 65 66 67
template <typename T>
class ElementwiseAddGradNPUKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
68 69 70 71 72 73 74 75 76 77 78
    auto& dev_ctx =
        ctx.template device_context<paddle::platform::NPUDeviceContext>();
    auto* x = ctx.Input<framework::Tensor>("X");
    auto* y = ctx.Input<framework::Tensor>("Y");
    auto* dout = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
    auto* dx = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
    auto* dy = ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
    int axis = ctx.Attr<int>("axis");

    axis = (axis == -1 ? std::abs(x->dims().size() - y->dims().size()) : axis);
    auto stream = dev_ctx.stream();
79 80
    if (dx) {
      dx->mutable_data<T>(ctx.GetPlace());
81 82 83 84 85 86 87 88 89 90 91 92 93 94
      if (dx->dims() != dout->dims()) {
        std::vector<int> dst_dims_vec;
        std::vector<int> reduce_axes;
        auto src_dims = dx->dims();
        auto dout_dims = dout->dims();

        int src_axis = (src_dims.size() < dout_dims.size() ? axis : 0);
        for (int ax = 0; ax < dout_dims.size(); ++ax) {
          if ((ax < src_axis || ax >= src_axis + src_dims.size()) ||
              (dout_dims[ax] > 1 && src_dims[ax - src_axis] == 1)) {
            reduce_axes.push_back(ax);
          } else {
            dst_dims_vec.push_back(dout_dims[ax]);
          }
95
        }
96 97 98 99 100 101 102 103
        if (!reduce_axes.empty()) {
          Tensor tmp;
          tmp.ShareDataWith(*dx);
          tmp.Resize(framework::make_ddim(dst_dims_vec));
          const auto& runner =
              NpuOpRunner("ReduceSumD", {*dout}, {tmp},
                          {{"axes", reduce_axes}, {"keep_dims", false}});
          runner.Run(stream);
104 105
        }
      } else {
106
        framework::TensorCopy(*dout, ctx.GetPlace(), dev_ctx, dx);
107 108 109
      }
    }
    if (dy) {
110 111 112 113 114 115 116 117 118 119 120 121 122 123 124
      dy->mutable_data<T>(ctx.GetPlace());
      if (dy->dims() != dout->dims()) {
        std::vector<int> dst_dims_vec;
        std::vector<int> reduce_axes;
        auto src_dims = dy->dims();
        auto dout_dims = dout->dims();

        int src_axis = (src_dims.size() < dout_dims.size() ? axis : 0);
        for (int ax = 0; ax < dout_dims.size(); ++ax) {
          if ((ax < src_axis || ax >= src_axis + src_dims.size()) ||
              (dout_dims[ax] > 1 && src_dims[ax - src_axis] == 1)) {
            reduce_axes.push_back(ax);
          } else {
            dst_dims_vec.push_back(dout_dims[ax]);
          }
125
        }
126 127 128 129 130 131 132 133
        if (!reduce_axes.empty()) {
          Tensor tmp;
          tmp.ShareDataWith(*dy);
          tmp.Resize(framework::make_ddim(dst_dims_vec));
          const auto& runner =
              NpuOpRunner("ReduceSumD", {*dout}, {tmp},
                          {{"axes", reduce_axes}, {"keep_dims", false}});
          runner.Run(stream);
134 135
        }
      } else {
136
        framework::TensorCopy(*dout, ctx.GetPlace(), dev_ctx, dy);
137 138 139 140 141
      }
    }
  }
};

142 143 144 145
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
146 147 148
namespace plat = paddle::platform;

REGISTER_OP_NPU_KERNEL(elementwise_add, ops::ElementwiseAddNPUKernel<float>,
149 150 151
#ifdef PADDLE_WITH_ASCEND_INT64
                       ops::ElementwiseAddNPUKernel<int64_t>,
#endif
152
                       ops::ElementwiseAddNPUKernel<plat::float16>);
153

154 155 156
REGISTER_OP_NPU_KERNEL(elementwise_add_grad,
                       ops::ElementwiseAddGradNPUKernel<float>,
                       ops::ElementwiseAddGradNPUKernel<plat::float16>);