wehub-resource-sync
203 lines
6.9 KiB
Plaintext
203 lines
6.9 KiB
Plaintext
// 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/temporal_shift_kernel.h"
|
|
|
|
#include <cstdint>
|
|
|
|
#include "paddle/common/enforce.h"
|
|
#include "paddle/common/layout.h"
|
|
#include "paddle/phi/backends/gpu/gpu_context.h"
|
|
#include "paddle/phi/core/kernel_registry.h"
|
|
|
|
namespace phi {
|
|
|
|
template <typename T, typename IndexT>
|
|
__global__ void KeTemporalShiftFwNCHW(const T* input,
|
|
T* output,
|
|
const IndexT ntchw,
|
|
const IndexT tchw,
|
|
const IndexT chw,
|
|
const IndexT hw,
|
|
const int t,
|
|
const IndexT c1,
|
|
const IndexT c2) {
|
|
IndexT tid = static_cast<IndexT>(blockIdx.x) * blockDim.x + threadIdx.x;
|
|
IndexT stride = static_cast<IndexT>(blockDim.x) * gridDim.x;
|
|
IndexT src_it = 0;
|
|
|
|
for (; tid < ntchw; tid += stride) {
|
|
IndexT it = (tid % tchw) / chw;
|
|
IndexT ic = (tid % chw) / hw;
|
|
|
|
if (ic < c1) {
|
|
src_it = it - 1;
|
|
} else if (ic < c2) {
|
|
src_it = it + 1;
|
|
} else {
|
|
src_it = it;
|
|
}
|
|
|
|
if (src_it < 0 || src_it >= t) {
|
|
output[tid] = 0;
|
|
} else {
|
|
output[tid] = input[tid + (src_it - it) * chw];
|
|
}
|
|
}
|
|
}
|
|
|
|
template <typename T, typename IndexT>
|
|
__global__ void KeTemporalShiftFwNHWC(const T* input,
|
|
T* output,
|
|
const IndexT nthwc,
|
|
const IndexT thwc,
|
|
const IndexT hwc,
|
|
const int t,
|
|
const IndexT c,
|
|
const IndexT c1,
|
|
const IndexT c2) {
|
|
IndexT tid = static_cast<IndexT>(blockIdx.x) * blockDim.x + threadIdx.x;
|
|
IndexT stride = static_cast<IndexT>(blockDim.x) * gridDim.x;
|
|
IndexT src_it = 0;
|
|
|
|
for (; tid < nthwc; tid += stride) {
|
|
IndexT it = (tid % thwc) / hwc;
|
|
IndexT ic = tid % c;
|
|
|
|
if (ic < c1) {
|
|
src_it = it - 1;
|
|
} else if (ic < c2) {
|
|
src_it = it + 1;
|
|
} else {
|
|
src_it = it;
|
|
}
|
|
|
|
if (src_it < 0 || src_it >= t) {
|
|
output[tid] = 0;
|
|
} else {
|
|
output[tid] = input[tid + (src_it - it) * hwc];
|
|
}
|
|
}
|
|
}
|
|
|
|
template <typename T, typename Context>
|
|
void TemporalShiftKernel(const Context& dev_ctx,
|
|
const DenseTensor& x,
|
|
int seg_num,
|
|
float shift_ratio,
|
|
const std::string& data_format_str,
|
|
DenseTensor* out) {
|
|
if (out && out->numel() == 0) {
|
|
dev_ctx.template Alloc<T>(out);
|
|
return;
|
|
}
|
|
auto* input = &x;
|
|
auto* output = out;
|
|
int t = seg_num;
|
|
const DataLayout data_layout = StringToDataLayout(data_format_str);
|
|
|
|
const int64_t nt = input->dims()[0];
|
|
const int64_t c =
|
|
(data_layout == DataLayout::NCHW ? input->dims()[1] : input->dims()[3]);
|
|
const int64_t h =
|
|
(data_layout == DataLayout::NCHW ? input->dims()[2] : input->dims()[1]);
|
|
const int64_t w =
|
|
(data_layout == DataLayout::NCHW ? input->dims()[3] : input->dims()[2]);
|
|
|
|
const int64_t hw = h * w;
|
|
const int64_t chw = c * hw;
|
|
const int64_t tchw = t * chw;
|
|
const int64_t ntchw = nt * chw;
|
|
|
|
const int64_t c1 = static_cast<int64_t>(c * shift_ratio);
|
|
const int64_t c2 = static_cast<int64_t>(c * 2 * shift_ratio);
|
|
|
|
DDim out_dims = (data_layout == DataLayout::NCHW ? make_ddim({nt, c, h, w})
|
|
: make_ddim({nt, h, w, c}));
|
|
const T* input_data = input->data<T>();
|
|
output->Resize(out_dims);
|
|
T* output_data = dev_ctx.template Alloc<T>(output);
|
|
|
|
int64_t pixelNum = nt * chw;
|
|
int64_t threads = 1024;
|
|
int64_t grid = (pixelNum + threads - 1) / threads;
|
|
int64_t blocks_per_sm = dev_ctx.GetMaxPhysicalThreadCount() / threads;
|
|
grid = std::min(dev_ctx.GetSMCount() * blocks_per_sm, grid);
|
|
PADDLE_ENFORCE_LE_UINT32_MAX(grid, "grid");
|
|
PADDLE_ENFORCE_LE_UINT32_MAX(threads, "threads");
|
|
const uint32_t grid_32 = static_cast<uint32_t>(grid);
|
|
const uint32_t threads_32 = static_cast<uint32_t>(threads);
|
|
|
|
if (data_layout == DataLayout::NCHW) {
|
|
if (x.numel() < std::numeric_limits<int32_t>::max()) {
|
|
PADDLE_ENFORCE_LE_INT_MAX(ntchw, "ntchw");
|
|
PADDLE_ENFORCE_LE_INT_MAX(tchw, "tchw");
|
|
PADDLE_ENFORCE_LE_INT_MAX(chw, "chw");
|
|
PADDLE_ENFORCE_LE_INT_MAX(hw, "hw");
|
|
PADDLE_ENFORCE_LE_INT_MAX(c1, "c1");
|
|
PADDLE_ENFORCE_LE_INT_MAX(c2, "c2");
|
|
KeTemporalShiftFwNCHW<T, int32_t>
|
|
<<<grid_32, threads_32, 0, dev_ctx.stream()>>>(
|
|
input_data,
|
|
output_data,
|
|
static_cast<int32_t>(ntchw),
|
|
static_cast<int32_t>(tchw),
|
|
static_cast<int32_t>(chw),
|
|
static_cast<int32_t>(hw),
|
|
t,
|
|
static_cast<int32_t>(c1),
|
|
static_cast<int32_t>(c2));
|
|
} else {
|
|
KeTemporalShiftFwNCHW<T, int64_t>
|
|
<<<grid_32, threads_32, 0, dev_ctx.stream()>>>(
|
|
input_data, output_data, ntchw, tchw, chw, hw, t, c1, c2);
|
|
}
|
|
} else {
|
|
if (x.numel() < std::numeric_limits<int32_t>::max()) {
|
|
PADDLE_ENFORCE_LE_INT_MAX(ntchw, "ntchw");
|
|
PADDLE_ENFORCE_LE_INT_MAX(tchw, "tchw");
|
|
PADDLE_ENFORCE_LE_INT_MAX(chw, "chw");
|
|
PADDLE_ENFORCE_LE_INT_MAX(c, "c");
|
|
PADDLE_ENFORCE_LE_INT_MAX(c1, "c1");
|
|
PADDLE_ENFORCE_LE_INT_MAX(c2, "c2");
|
|
KeTemporalShiftFwNHWC<T, int32_t>
|
|
<<<grid_32, threads_32, 0, dev_ctx.stream()>>>(
|
|
input_data,
|
|
output_data,
|
|
static_cast<int32_t>(ntchw),
|
|
static_cast<int32_t>(tchw),
|
|
static_cast<int32_t>(chw),
|
|
t,
|
|
static_cast<int32_t>(c),
|
|
static_cast<int32_t>(c1),
|
|
static_cast<int32_t>(c2));
|
|
} else {
|
|
KeTemporalShiftFwNHWC<T, int64_t>
|
|
<<<grid_32, threads_32, 0, dev_ctx.stream()>>>(
|
|
input_data, output_data, ntchw, tchw, chw, t, c, c1, c2);
|
|
}
|
|
}
|
|
}
|
|
|
|
} // namespace phi
|
|
|
|
PD_REGISTER_KERNEL(temporal_shift,
|
|
GPU,
|
|
ALL_LAYOUT,
|
|
phi::TemporalShiftKernel,
|
|
float,
|
|
double,
|
|
phi::float16,
|
|
phi::bfloat16) {}
|