manipulation.cu 7.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
//   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.

15
#include "paddle/pten/api/ext/dispatch.h"
16
#include "paddle/pten/infermeta/unary.h"
17 18
#include "paddle/pten/kernels/cuda/manipulation.h"
#include "paddle/pten/kernels/cuda/utils.h"
19
#include "paddle/pten/kernels/functions/cuda/cast_kernel_impl.h"
20
#include "paddle/pten/kernels/functions/general/manipulation.h"
21 22 23 24 25 26 27 28 29

namespace pten {

template <typename T>
void Flatten(const CUDAContext& dev_ctx,
             const DenseTensor& x,
             int start_axis,
             int stop_axis,
             DenseTensor* out) {
30
  auto out_dims = out->dims();
31
  pten::Copy(dev_ctx, x, false, out);
32
  out->Resize(out_dims);
33 34 35 36 37 38 39 40 41 42 43 44 45
}

// TODO(yuanrisheng): this kernel is for training and xshape is a Intermediate
// Output Tensor,
// is there a more flexible way to deal with this case?
template <typename T>
void FlattenWithXShape(const CUDAContext& dev_ctx,
                       const DenseTensor& x,
                       int start_axis,
                       int stop_axis,
                       DenseTensor* out,
                       DenseTensor* xshape) {
  Flatten<T>(dev_ctx, x, start_axis, stop_axis, out);
46 47 48 49 50
  general::SetXShape(x, xshape);
}

void ReshapeFromVectorVal(const CUDAContext& dev_ctx,
                          const DenseTensor& x,
51
                          const std::vector<int64_t>& shape,
52
                          DenseTensor* out) {
53
  auto out_meta = InferMetaFromVecValue(x.meta(), shape);
54 55 56 57 58 59
  if (&x == out) {
    out->Resize(out_meta.dims);
    return;
  }
  pten::Copy(dev_ctx, x, false, out);
  out->Resize(out_meta.dims);
60 61 62 63
}

void ReshapeFromVectorValWithXShape(const CUDAContext& dev_ctx,
                                    const DenseTensor& x,
64
                                    const std::vector<int64_t>& shape,
65 66 67
                                    DenseTensor* xshape,
                                    DenseTensor* out) {
  general::SetXShape(x, xshape);
68
  ReshapeFromVectorVal(dev_ctx, x, shape, out);
69 70 71 72 73 74 75
}

void ReshapeFromDT(const CUDAContext& dev_ctx,
                   const DenseTensor& x,
                   const DenseTensor& shape,
                   DenseTensor* out) {
  auto* shape_data = shape.data<int>();
76 77
  auto vector_shape =
      std::vector<int64_t>(shape_data, shape_data + shape.numel());
78 79
  ReshapeFromVectorVal(dev_ctx, x, vector_shape, out);
  out->ResetLoD(x.lod());
80 81 82 83 84 85 86 87
}

void ReshapeFromDTWithXShape(const CUDAContext& dev_ctx,
                             const DenseTensor& x,
                             const DenseTensor& shape,
                             DenseTensor* xshape,
                             DenseTensor* out) {
  general::SetXShape(x, xshape);
88
  ReshapeFromDT(dev_ctx, x, shape, out);
89 90 91 92 93 94
}

void ReshapeFromVectorDT(const CUDAContext& dev_ctx,
                         const DenseTensor& x,
                         const std::vector<DenseTensor>& shape,
                         DenseTensor* out) {
95
  std::vector<int64_t> vector_shape;
96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115
  for (auto& tensor : shape) {
    PADDLE_ENFORCE_EQ(
        tensor.dims(),
        paddle::framework::make_ddim({1}),
        paddle::platform::errors::InvalidArgument(
            "If the element type of 'shape' in ReshapeOp is Tensor, "
            "the element's shape must be [1]. But received the element's shape "
            "is [%s]",
            tensor.dims()));
    vector_shape.push_back(*tensor.data<int32_t>());
  }
  ReshapeFromVectorVal(dev_ctx, x, vector_shape, out);
}

void ReshapeFromVectorDTWithXShape(const CUDAContext& dev_ctx,
                                   const DenseTensor& x,
                                   const std::vector<DenseTensor>& shape,
                                   DenseTensor* xshape,
                                   DenseTensor* out) {
  general::SetXShape(x, xshape);
116 117 118 119 120 121 122 123 124 125
  ReshapeFromVectorDT(dev_ctx, x, shape, out);
}

template <typename T>
void Cast(const CUDAContext& dev_ctx,
          const DenseTensor& x,
          DataType out_dtype,
          DataType in_dtype,
          DenseTensor* out) {
  PD_VISIT_ALL_TYPES(out_dtype, "CastKernelImpl", ([&] {
126
                       detail::CastCUDAKernelImpl<T, data_t>(dev_ctx, x, out);
127
                     }));
128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159
}

}  // namespace pten

// TODO(chenweihang): replace by better impl
PT_REGISTER_MODULE(ManipulationCUDA);

using float16 = paddle::platform::float16;
// TODO(yuanrisheng): "flatten_contiguous_range" is compatible with old kernel
// architecture, kernel_name should be "flatten".
PT_REGISTER_KERNEL("flatten_contiguous_range",
                   CUDA,
                   ANY,
                   pten::Flatten,
                   float,
                   float16,
                   double,
                   uint8_t,
                   int8_t,
                   int,
                   int64_t) {}

PT_REGISTER_KERNEL("flatten_contiguous_range.mid",
                   CUDA,
                   ANY,
                   pten::FlattenWithXShape,
                   float,
                   double,
                   uint8_t,
                   int8_t,
                   int,
                   int64_t) {}
160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185

#define PTEN_REGISTER_CAST_CUDA_BASE_TYPE(op_name, ...) \
  PT_REGISTER_KERNEL("cast",                            \
                     CUDA,                              \
                     ANY,                               \
                     pten::Cast,                        \
                     float,                             \
                     double,                            \
                     int,                               \
                     int64_t,                           \
                     int16_t,                           \
                     bool,                              \
                     uint8_t,                           \
                     paddle::platform::float16,         \
                     paddle::platform::complex<float>,  \
                     paddle::platform::complex<double>, \
                     ##__VA_ARGS__) {                   \
    kernel->OutputAt(0).SetDataType(                    \
        paddle::experimental::DataType::UNDEFINED);     \
  }

#if !defined(PADDLE_WITH_HIP)
PTEN_REGISTER_CAST_CUDA_BASE_TYPE(cast, paddle::platform::bfloat16)
#else
PTEN_REGISTER_CAST_CUDA_BASE_TYPE(cast)
#endif
186 187 188 189 190 191 192 193 194 195

PT_REGISTER_KERNEL_WITH_NO_TYPE("reshape2",
                                CUDA,
                                ANY,
                                pten::ReshapeFromVectorVal) {}

PT_REGISTER_KERNEL_WITH_NO_TYPE("reshape2.mid",
                                CUDA,
                                ANY,
                                pten::ReshapeFromVectorValWithXShape) {}
新手
引导
客服 返回
顶部