concat_compute.cc 3.4 KB
Newer Older
Y
Yan Chunwei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
// Copyright (c) 2019 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 "lite/kernels/arm/concat_compute.h"
#include <string>
#include <vector>
18
#include "lite/backends/arm/math/funcs.h"
Y
Yan Chunwei 已提交
19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
#include "lite/core/op_registry.h"
#include "lite/core/tensor.h"
#include "lite/core/type_system.h"

namespace paddle {
namespace lite {
namespace kernels {
namespace arm {

std::vector<size_t> stride_numel(const DDim& ddim) {
  std::vector<size_t> strides(ddim.size());
  strides[ddim.size() - 1] = ddim[ddim.size() - 1];
  for (int i = ddim.size() - 2; i >= 0; --i) {
    strides[i] = strides[i + 1] * ddim[i];
  }
  return strides;
}

37 38 39 40 41
template <typename T>
void ConcatFunc(const std::vector<lite::Tensor*> inputs,
                int axis,
                lite::Tensor* out) {
  // Sometimes direct copies will be faster, this maybe need deeply analysis.
Y
Yan Chunwei 已提交
42 43 44 45 46
  if (axis == 0 && inputs.size() < 10) {
    size_t output_offset = 0;
    for (auto* in : inputs) {
      auto in_stride = stride_numel(in->dims());
      auto out_stride = stride_numel(out->dims());
47 48
      void* dst = out->mutable_data<T>() + output_offset;
      const void* src = in->data<T>();
Y
Yan Chunwei 已提交
49 50
      // src and dst tensor should have the same dims size.
      CHECK(in_stride.size() == out_stride.size());
51
      std::memcpy(dst, src, sizeof(T) * in_stride[0]);
Y
Yan Chunwei 已提交
52 53 54 55 56 57 58
      output_offset += in_stride[0];
    }
  } else {
    std::vector<lite::Tensor*> inputs_concat(inputs.size());
    for (int j = 0; j < inputs.size(); ++j) {
      inputs_concat[j] = inputs[j];
    }
59 60 61 62 63 64 65 66 67 68 69 70 71 72 73
    lite::arm::math::concat_func<T>(inputs_concat, axis, out);
  }
}

void ConcatCompute::Run() {
  auto& param = Param<operators::ConcatParam>();
  std::vector<lite::Tensor*> inputs = param.x;
  CHECK_GE(inputs.size(), 1);
  auto* out = param.output;
  int axis = param.axis;
  auto* axis_tensor = param.axis_tensor;
  if (axis_tensor != nullptr) {
    auto* axis_tensor_data = axis_tensor->data<int>();
    axis = axis_tensor_data[0];
  }
74 75 76
  if (axis < 0) {
    axis += inputs[0]->dims().size();
  }
77 78 79 80 81 82 83 84 85 86 87 88 89 90 91

  switch (inputs.front()->precision()) {
    case PRECISION(kFloat):
      ConcatFunc<float>(inputs, axis, out);
      break;
    case PRECISION(kInt32):
      ConcatFunc<int32_t>(inputs, axis, out);
      break;
    case PRECISION(kInt64):
      ConcatFunc<int64_t>(inputs, axis, out);
      break;
    default:
      LOG(FATAL) << "Concat does not implement for the "
                 << "input type:"
                 << static_cast<int>(inputs.front()->precision());
Y
Yan Chunwei 已提交
92 93 94 95 96 97 98 99 100
  }
}

}  // namespace arm
}  // namespace kernels
}  // namespace lite
}  // namespace paddle

REGISTER_LITE_KERNEL(
101 102
    concat, kARM, kAny, kNCHW, paddle::lite::kernels::arm::ConcatCompute, def)
    .BindInput("X", {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kAny))})
103 104
    .BindInput("AxisTensor",
               {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kInt32))})
105
    .BindOutput("Out", {LiteType::GetTensorTy(TARGET(kARM), PRECISION(kAny))})
Y
Yan Chunwei 已提交
106
    .Finalize();