cast_op.cc 6.1 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Y
Yu Yang 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
Y
Yu Yang 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
Y
Yu Yang 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. */
Y
Yu Yang 已提交
14

Y
Yi Wang 已提交
15
#include "paddle/fluid/operators/cast_op.h"
T
tensor-tang 已提交
16
#include <memory>
Y
Yi Wang 已提交
17
#include "paddle/fluid/framework/op_registry.h"
K
Kexin Zhao 已提交
18
#include "paddle/fluid/platform/float16.h"
19 20 21
#ifdef PADDLE_WITH_MLU
#include "paddle/fluid/operators/mlu/mlu_baseop.h"
#endif
Y
Yu Yang 已提交
22 23 24 25 26 27

namespace paddle {
namespace operators {

class CastOpProtoMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
28
  void Make() override {
29 30
    AddInput("X", "The input tensor of cast op");
    AddOutput("Out", "The output tensor of cast op");
F
fengjiayi 已提交
31 32
    AddAttr<int>("out_dtype", "output data type");
    AddAttr<int>("in_dtype", "input data type");
33 34 35
    AddAttr<bool>("use_mkldnn",
                  "(bool, default false) Only used in mkldnn kernel")
        .SetDefault(false);
36 37 38 39
    AddComment(R"DOC(
Cast Operator.

This Operator casts the input tensor to another data type and
T
tensor-tang 已提交
40 41
returns the Output Tensor. It's meaningless if the output dtype equals
the input dtype, but it's fine if you do so.
42 43

)DOC");
Y
Yu Yang 已提交
44 45 46
  }
};

H
hong 已提交
47 48
template <typename T>
class CastOpGradMaker : public framework::SingleGradOpMaker<T> {
Y
Yu Yang 已提交
49
 public:
H
hong 已提交
50
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
Y
Yu Yang 已提交
51 52

 protected:
53
  void Apply(GradOpPtr<T> grad) const override {
Y
Yu Yang 已提交
54
    grad->SetType("cast");
H
hong 已提交
55 56 57 58
    grad->SetInput("X", this->OutputGrad("Out"));
    grad->SetOutput("Out", this->InputGrad("X"));
    grad->SetAttr("out_dtype", this->GetAttr("in_dtype"));
    grad->SetAttr("in_dtype", this->GetAttr("out_dtype"));
59
    grad->SetAttr("use_mkldnn", this->GetAttr("use_mkldnn"));
Y
Yu Yang 已提交
60 61 62
  }
};

Q
QI JUN 已提交
63 64 65 66 67
class CastOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

 protected:
Z
Zeng Jinle 已提交
68
  void InferShape(framework::InferShapeContext *context) const override {
69 70
    OP_INOUT_CHECK(context->HasInput("X"), "Input", "X", "cast");
    OP_INOUT_CHECK(context->HasOutput("Out"), "Output", "Out", "cast");
Z
Zeng Jinle 已提交
71 72 73 74
    context->SetOutputDim("Out", context->GetInputDim("X"));
    context->ShareLoD("X", "Out");
  }

Q
QI JUN 已提交
75 76 77
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext &ctx) const override {
    // CastOp kernel's device type is decided by input tensor place
78 79 80 81 82 83 84 85 86
    auto *tensor = ctx.Input<framework::LoDTensor>("X");
    PADDLE_ENFORCE_EQ(tensor->IsInitialized(), true,
                      platform::errors::PreconditionNotMet(
                          "The tensor of Input(X) is not initialized."));
    auto &tensor_place = tensor->place();
    // NOTE: cuda pinned tensor need to copy its data to target place
    if (platform::is_cuda_pinned_place(tensor_place)) {
      return framework::OpKernelType(tensor->type(), ctx.device_context());
    }
87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107

#ifdef PADDLE_WITH_MKLDNN
    int in_dtype = ctx.Attr<int>("in_dtype");
    int out_dtype = ctx.Attr<int>("out_dtype");

    auto MKLDNNSupportsCast = [&]() -> bool {
      int dtype_fp32 = static_cast<int>(framework::proto::VarType::FP32);
      int dtype_bf16 = static_cast<int>(framework::proto::VarType::BF16);

      if ((in_dtype != dtype_fp32 && in_dtype != dtype_bf16) ||
          (out_dtype != dtype_fp32 && out_dtype != dtype_bf16))
        return false;

      return true;
    };

    if (this->CanMKLDNNBeUsed(ctx, tensor->type()) && MKLDNNSupportsCast()) {
      return framework::OpKernelType(tensor->type(), ctx.GetPlace(),
                                     framework::DataLayout::kMKLDNN,
                                     framework::LibraryType::kMKLDNN);
    }
108 109 110 111 112 113 114 115 116 117 118 119 120
#endif
#ifdef PADDLE_WITH_MLU
    auto src_type = static_cast<VT::Type>(ctx.Attr<int>("in_dtype"));
    auto dst_type = static_cast<VT::Type>(ctx.Attr<int>("out_dtype"));
    if (src_type == dst_type || MLUSupportsCast(src_type, dst_type)) {
      return framework::OpKernelType(tensor->type(), tensor_place);
    } else {
      VLOG(3) << "MLU not support cast type: "
              << framework::DataTypeToString(src_type)
              << " to type: " << framework::DataTypeToString(dst_type)
              << ", fallbacking to CPU one!";
      return framework::OpKernelType(tensor->type(), platform::CPUPlace());
    }
121
#endif
122
    return framework::OpKernelType(tensor->type(), tensor_place);
Q
QI JUN 已提交
123
  }
124 125 126 127 128

  framework::KernelSignature GetExpectedPtenKernelArgs(
      const framework::ExecutionContext &ctx) const override {
    return framework::KernelSignature("cast", {"X"}, {"out_dtype"}, {"Out"});
  }
Q
QI JUN 已提交
129 130
};

Y
Yu Yang 已提交
131 132 133 134
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
Q
QI JUN 已提交
135
using CPU = paddle::platform::CPUDeviceContext;
136 137 138 139 140 141 142

// cast use pten kernel, so no need to REGISTER_OP_CPU_KERNEL here.
REGISTER_OPERATOR(cast, ops::CastOp,
                  ops::CastOpGradMaker<paddle::framework::OpDesc>,
                  ops::CastOpGradMaker<paddle::imperative::OpBase>,
                  ops::CastOpProtoMaker);

143 144 145
// [ why register transfer_dtype_op alias with cast_op? ]
// In case of InterpreterCore, if we reuse cast_op, we cannot distinguish
// which cast_op is inserted by new executor when we do profiling.
146 147 148 149 150 151 152 153 154 155 156 157 158 159
REGISTER_OPERATOR(transfer_dtype, ops::CastOp,
                  ops::CastOpGradMaker<paddle::framework::OpDesc>,
                  ops::CastOpGradMaker<paddle::imperative::OpBase>,
                  ops::CastOpProtoMaker);
REGISTER_OP_CPU_KERNEL(
    transfer_dtype, ops::CastOpKernel<CPU, float>,
    ops::CastOpKernel<CPU, double>, ops::CastOpKernel<CPU, int>,
    ops::CastOpKernel<CPU, int64_t>, ops::CastOpKernel<CPU, int>,
    ops::CastOpKernel<CPU, int16_t>, ops::CastOpKernel<CPU, bool>,
    ops::CastOpKernel<CPU, uint8_t>,
    ops::CastOpKernel<CPU, paddle::platform::float16>,
    ops::CastOpKernel<CPU, paddle::platform::bfloat16>,
    ops::CastOpKernel<CPU, paddle::platform::complex<float>>,
    ops::CastOpKernel<CPU, paddle::platform::complex<double>>);