split_kernel.cc 2.4 KB
Newer Older
C
chentianyu03 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// Copyright (c) 2022 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/phi/kernels/split_kernel.h"
C
chentianyu03 已提交
16 17

#include "paddle/fluid/operators/strided_memcpy.h"
18 19
#include "paddle/phi/common/float16.h"
#include "paddle/phi/core/kernel_registry.h"
C
chentianyu03 已提交
20

21
#include "paddle/phi/infermeta/unary.h"
L
Leo Chen 已提交
22
#include "paddle/phi/kernels/funcs/concat_and_split_functor.h"
23
namespace phi {
C
chentianyu03 已提交
24 25 26 27 28 29 30 31

template <typename T, typename Context>
void SplitKernel(const Context& dev_ctx,
                 const DenseTensor& x,
                 const ScalarArray& num_or_sections,
                 const Scalar& axis_scalar,
                 std::vector<DenseTensor*> outs) {
  // need to infershape output
32
  if (num_or_sections.FromTensor() || axis_scalar.FromTensor()) {
C
chentianyu03 已提交
33 34 35 36 37
    std::vector<MetaTensor> out_metas;
    for (size_t i = 0; i < outs.size(); ++i) {
      out_metas.push_back(outs[i]);
    }

38
    phi::SplitInferMeta(x, num_or_sections, axis_scalar, &out_metas, true);
C
chentianyu03 已提交
39 40 41 42 43 44 45 46

    for (size_t i = 0; i < out_metas.size(); ++i) {
      outs[i]->Resize(out_metas[i].dims());
    }
  }

  std::vector<const DenseTensor*> shape_refer;
  for (size_t j = 0; j < outs.size(); ++j) {
47
    dev_ctx.template Alloc<T>(outs[j]);
C
chentianyu03 已提交
48 49 50 51 52 53 54 55 56
    shape_refer.emplace_back(outs[j]);
  }

  int axis = axis_scalar.to<int>();
  // Sometimes direct copies will be faster, this maybe need deeply analysis.
  if (axis == 0 && outs.size() < 10) {
    paddle::operators::StridedMemcpyWithAxis0<T>(
        dev_ctx, x, shape_refer, &outs);
  } else {
L
Leo Chen 已提交
57 58
    phi::funcs::SplitFunctor<Context, T> functor;
    functor(dev_ctx, x, shape_refer, axis, &outs);
C
chentianyu03 已提交
59 60 61
  }
}

62
}  // namespace phi
C
chentianyu03 已提交
63

64
PD_REGISTER_KERNEL(split,
C
chentianyu03 已提交
65 66
                   CPU,
                   ALL_LAYOUT,
67
                   phi::SplitKernel,
C
chentianyu03 已提交
68 69 70 71 72
                   float,
                   double,
                   int64_t,
                   int,
                   bool,
73
                   phi::dtype::float16) {}