wehub-resource-sync
135 lines
4.3 KiB
C++
135 lines
4.3 KiB
C++
// 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.
|
|
|
|
#include "paddle/phi/kernels/concat_kernel.h"
|
|
|
|
#include "paddle/phi/backends/cpu/cpu_context.h"
|
|
#include "paddle/phi/common/scalar.h"
|
|
#include "paddle/phi/core/dense_tensor.h"
|
|
#include "paddle/phi/core/kernel_registry.h"
|
|
#include "paddle/phi/core/lod_utils.h"
|
|
#include "paddle/phi/kernels/funcs/concat_and_split_functor.h"
|
|
#include "paddle/phi/kernels/funcs/concat_funcs.h"
|
|
#include "paddle/phi/kernels/funcs/strided_memcpy.h"
|
|
|
|
namespace phi {
|
|
|
|
template <typename T, typename Context>
|
|
void ConcatKernel(const Context& dev_ctx,
|
|
const std::vector<const DenseTensor*>& x,
|
|
const Scalar& axis_scalar,
|
|
DenseTensor* out) {
|
|
int64_t axis = axis_scalar.to<int64_t>();
|
|
|
|
axis = funcs::ComputeAxis(axis, x[0]->dims().size());
|
|
|
|
std::vector<DDim> x_dims;
|
|
x_dims.reserve(x.size());
|
|
for (auto item : x) {
|
|
x_dims.push_back(item->dims());
|
|
}
|
|
|
|
DDim out_dims = funcs::ComputeAndCheckShape(true, x_dims, axis);
|
|
out->Resize(out_dims);
|
|
dev_ctx.template Alloc<T>(out);
|
|
if (out->numel() == 0) {
|
|
return;
|
|
}
|
|
// If axis is 0, the lod of the output is not the same as inputs.
|
|
if (axis == 0 && !x[0]->lod().empty()) {
|
|
size_t lod_size_0 = x[0]->lod().size();
|
|
size_t lod_size = lod_size_0;
|
|
for (size_t i = 1; i < x.size(); ++i) {
|
|
if (!x[i]->lod().empty()) {
|
|
PADDLE_ENFORCE_EQ(
|
|
x[i]->lod().size(),
|
|
lod_size_0,
|
|
common::errors::Unimplemented(
|
|
"The lod level of all input DenseTensors should be same. "
|
|
"Maybe different lod level of input DenseTensors can concat,"
|
|
"it is not supported currently. The lod level of %dth input "
|
|
"is %d and first input is %d.",
|
|
i,
|
|
x[i]->lod().size(),
|
|
lod_size_0));
|
|
} else {
|
|
lod_size = 0;
|
|
break;
|
|
}
|
|
}
|
|
if (lod_size) {
|
|
auto* out_lod = out->mutable_lod();
|
|
for (size_t i = 1; i < x.size(); ++i) {
|
|
auto in_lod = ConvertToLengthBasedLegacyLoD(x[i]->lod());
|
|
AppendLegacyLoD(out_lod, in_lod);
|
|
}
|
|
}
|
|
}
|
|
|
|
// Sometimes direct copies will be faster, this maybe need deeply analysis.
|
|
if (axis == 0 && x.size() < 10) {
|
|
size_t output_offset = 0;
|
|
for (const auto* in : x) {
|
|
if (in->numel() == 0UL) {
|
|
continue;
|
|
}
|
|
auto in_stride = common::stride_numel(in->dims());
|
|
auto out_stride = common::stride_numel(out->dims());
|
|
funcs::StridedNumelCopyWithAxis<T, Context>(
|
|
dev_ctx,
|
|
axis,
|
|
out->data<T>() + output_offset,
|
|
out_stride,
|
|
in->data<T>(),
|
|
in_stride,
|
|
in_stride[static_cast<int>(axis)]);
|
|
output_offset += in_stride[static_cast<int>(axis)];
|
|
}
|
|
} else {
|
|
// TODO(chenweihang): concat functor support vector<DenseTensor*> input
|
|
std::vector<DenseTensor> inputs;
|
|
inputs.reserve(x.size());
|
|
for (auto item : x) {
|
|
if (item->numel() > 0) {
|
|
inputs.emplace_back(*item);
|
|
} else {
|
|
continue;
|
|
}
|
|
}
|
|
funcs::ConcatFunctor<Context, T> functor;
|
|
functor(dev_ctx, inputs, axis, out);
|
|
}
|
|
}
|
|
|
|
} // namespace phi
|
|
|
|
PD_REGISTER_KERNEL(concat,
|
|
CPU,
|
|
ALL_LAYOUT,
|
|
phi::ConcatKernel,
|
|
float,
|
|
double,
|
|
bool,
|
|
int64_t,
|
|
int,
|
|
uint8_t,
|
|
int8_t,
|
|
int16_t,
|
|
phi::float16,
|
|
phi::bfloat16,
|
|
phi::float8_e4m3fn,
|
|
phi::float8_e5m2,
|
|
phi::complex64,
|
|
phi::complex128) {}
|