softmax_op_xpu.cc 5.0 KB
Newer Older
Z
zhupengyang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
/* Copyright (c) 2020 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. */

#ifdef PADDLE_WITH_XPU

#include "paddle/fluid/operators/softmax_op.h"
#include "paddle/fluid/framework/op_registry.h"

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
using DDim = framework::DDim;

template <typename DeviceContext, typename T>
class SoftmaxXPUKernel : public framework::OpKernel<T> {
25 26
  using XPUType = typename XPUTypeTrait<T>::Type;

Z
zhupengyang 已提交
27 28 29 30 31
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* x = context.Input<Tensor>("X");
    auto* out = context.Output<Tensor>("Out");
    const int rank = x->dims().size();
32
    int axis = CanonicalAxis(context.Attr<int>("axis"), rank);
Z
zhupengyang 已提交
33 34 35 36

    // allocate memory on device.
    out->mutable_data<T>(context.GetPlace());

37 38 39 40 41 42 43
    std::vector<int> x_dims;
    for (int i = 0; i < rank; i++) {
      x_dims.push_back(x->dims()[i]);
    }
    if (axis < 0) {
      axis += rank;
    }
Z
zhupengyang 已提交
44 45

    auto& dev_ctx = context.template device_context<DeviceContext>();
T
taixiurong 已提交
46 47

    int r = XPU_SUCCESS;
48
    auto version = platform::get_xpu_version(context.GetPlace().GetDeviceId());
W
Wilber 已提交
49
    if (version == pten::backends::xpu::XPUVersion::XPU1) {
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 75 76 77
      xpu::ctx_guard RAII_GUARD(dev_ctx.x_context());
      XPUType* clip_x_data_l3 = RAII_GUARD.alloc_l3_or_gm<XPUType>(x->numel());
      r = xpu::clip_v2(dev_ctx.x_context(),
                       reinterpret_cast<const XPUType*>(x->data<T>()),
                       clip_x_data_l3, x->numel(), static_cast<XPUType>(-1e20),
                       static_cast<XPUType>(1e20));
      PADDLE_ENFORCE_EQ(r, XPU_SUCCESS,
                        platform::errors::External(
                            "XPU API(clip_v2) return wrong value[%d %s]", r,
                            XPUAPIErrorMsg[r]));
      r = xpu::softmax<XPUType>(dev_ctx.x_context(), clip_x_data_l3,
                                reinterpret_cast<XPUType*>(out->data<T>()),
                                x_dims, axis);
      PADDLE_ENFORCE_EQ(
          r, XPU_SUCCESS,
          platform::errors::External("XPU API(softmax2d_forward) return wrong "
                                     "value[%d %s]",
                                     r, XPUAPIErrorMsg[r]));
    } else {
      r = xpu::softmax<XPUType>(
          dev_ctx.x_context(), reinterpret_cast<const XPUType*>(x->data<T>()),
          reinterpret_cast<XPUType*>(out->data<T>()), x_dims, axis);
      PADDLE_ENFORCE_EQ(
          r, XPU_SUCCESS,
          platform::errors::External("XPU API(softmax2d_forward) return wrong "
                                     "value[%d %s]",
                                     r, XPUAPIErrorMsg[r]));
    }
Z
zhupengyang 已提交
78 79 80 81 82
  }
};

template <typename DeviceContext, typename T>
class SoftmaxGradXPUKernel : public framework::OpKernel<T> {
83 84
  using XPUType = typename XPUTypeTrait<T>::Type;

Z
zhupengyang 已提交
85 86 87 88 89 90
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* out = context.Input<Tensor>("Out");
    auto* dout = context.Input<Tensor>(framework::GradVarName("Out"));
    auto* dx = context.Output<Tensor>(framework::GradVarName("X"));
    const int rank = dx->dims().size();
91
    int axis = CanonicalAxis(context.Attr<int>("axis"), rank);
Z
zhupengyang 已提交
92 93 94 95

    // allocate memory on device.
    dx->mutable_data<T>(context.GetPlace());

96 97 98 99 100 101 102
    std::vector<int> x_dims;
    for (int i = 0; i < rank; i++) {
      x_dims.push_back(dx->dims()[i]);
    }
    if (axis < 0) {
      axis += rank;
    }
Z
zhupengyang 已提交
103 104

    auto& dev_ctx = context.template device_context<DeviceContext>();
105 106 107 108
    int r = xpu::softmax_grad<XPUType>(
        dev_ctx.x_context(), reinterpret_cast<const XPUType*>(out->data<T>()),
        reinterpret_cast<const XPUType*>(dout->data<T>()),
        reinterpret_cast<XPUType*>(dx->data<T>()), x_dims, axis);
Z
zhupengyang 已提交
109 110 111
    PADDLE_ENFORCE_EQ(
        r, XPU_SUCCESS,
        platform::errors::External("XPU API(softmax2d_backward) return wrong "
112 113
                                   "value[%d %s]",
                                   r, XPUAPIErrorMsg[r]));
Z
zhupengyang 已提交
114 115 116 117 118 119 120 121 122
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;

REGISTER_OP_XPU_KERNEL(
123 124 125
    softmax, ops::SoftmaxXPUKernel<paddle::platform::XPUDeviceContext, float>,
    ops::SoftmaxXPUKernel<paddle::platform::XPUDeviceContext,
                          paddle::platform::float16>);
Z
zhupengyang 已提交
126 127
REGISTER_OP_XPU_KERNEL(
    softmax_grad,
128 129 130
    ops::SoftmaxGradXPUKernel<paddle::platform::XPUDeviceContext, float>,
    ops::SoftmaxGradXPUKernel<paddle::platform::XPUDeviceContext,
                              paddle::platform::float16>);
Z
zhupengyang 已提交
131 132

#endif  // PADDLE_WITH_XPU