wehub-resource-sync
97 lines
4.0 KiB
Plaintext
97 lines
4.0 KiB
Plaintext
/*
|
|
* SPDX-FileCopyrightText: Copyright (c) 1993-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
* SPDX-License-Identifier: Apache-2.0
|
|
*
|
|
* 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 "seqLen2SpatialKernel.h"
|
|
#include <MNN/MNNDefine.h>
|
|
|
|
template <typename T, int32_t C, int32_t TPB>
|
|
__global__ void SeqLen2SpatialKernel(T const* input, T const* biasInput, T const* residualInput, T* output)
|
|
{
|
|
int32_t baseOffset = blockIdx.x * C + threadIdx.x;
|
|
int32_t biasOffset = threadIdx.x;
|
|
#pragma unroll
|
|
for (int32_t i = 0; i < C / TPB; ++i)
|
|
{
|
|
output[baseOffset] = input[baseOffset] + biasInput[biasOffset] + residualInput[baseOffset];
|
|
baseOffset += TPB;
|
|
biasOffset += TPB;
|
|
}
|
|
}
|
|
|
|
template __global__ void SeqLen2SpatialKernel<float, 320, 320>(float const*, float const*, float const*, float*);
|
|
template __global__ void SeqLen2SpatialKernel<float, 640, 320>(float const*, float const*, float const*, float*);
|
|
template __global__ void SeqLen2SpatialKernel<float, 1280, 320>(float const*, float const*, float const*, float*);
|
|
template __global__ void SeqLen2SpatialKernel<half, 320, 320>(half const*, half const*, half const*, half*);
|
|
template __global__ void SeqLen2SpatialKernel<half, 640, 320>(half const*, half const*, half const*, half*);
|
|
template __global__ void SeqLen2SpatialKernel<half, 1280, 320>(half const*, half const*, half const*, half*);
|
|
|
|
int32_t launchSeqLen2SpatialKernel(void const* input0, void const* input1, void const* input2, void* output0, bool isHalf,
|
|
int32_t gridSize, int32_t C, cudaStream_t stream)
|
|
{
|
|
if (!isHalf)
|
|
{
|
|
auto const input = static_cast<float const*>(input0);
|
|
auto const biasInput = static_cast<float const*>(input1);
|
|
auto const residualInput = static_cast<float const*>(input2);
|
|
auto output = static_cast<float*>(output0);
|
|
constexpr int32_t TPB = 320; // thread per block
|
|
switch (C)
|
|
{
|
|
case 320:
|
|
(SeqLen2SpatialKernel<float, 320, TPB>) <<<gridSize, TPB, 0, stream>>>(
|
|
input, biasInput, residualInput, output);
|
|
break;
|
|
case 640:
|
|
(SeqLen2SpatialKernel<float, 640, TPB>) <<<gridSize, TPB, 0, stream>>>(
|
|
input, biasInput, residualInput, output);
|
|
break;
|
|
case 1280:
|
|
(SeqLen2SpatialKernel<float, 1280, TPB>) <<<gridSize, TPB, 0, stream>>>(
|
|
input, biasInput, residualInput, output);
|
|
break;
|
|
default: MNN_ERROR("SeqLen2Spatial Unsupported number of channels!\n");
|
|
}
|
|
checkKernelErrors;
|
|
}
|
|
else
|
|
{
|
|
auto const input = static_cast<half const*>(input0);
|
|
auto const biasInput = static_cast<half const*>(input1);
|
|
auto const residualInput = static_cast<half const*>(input2);
|
|
auto output = static_cast<half*>(output0);
|
|
constexpr int32_t TPB = 320; // thread per block
|
|
switch (C)
|
|
{
|
|
case 320:
|
|
(SeqLen2SpatialKernel<half, 320, TPB>) <<<gridSize, TPB, 0, stream>>>(
|
|
input, biasInput, residualInput, output);
|
|
break;
|
|
case 640:
|
|
(SeqLen2SpatialKernel<half, 640, TPB>) <<<gridSize, TPB, 0, stream>>>(
|
|
input, biasInput, residualInput, output);
|
|
break;
|
|
case 1280:
|
|
(SeqLen2SpatialKernel<half, 1280, TPB>) <<<gridSize, TPB, 0, stream>>>(
|
|
input, biasInput, residualInput, output);
|
|
break;
|
|
default: MNN_ERROR("SeqLen2Spatial Unsupported number of channels!\n");
|
|
}
|
|
checkKernelErrors;
|
|
}
|
|
return 0;
|
|
}
|