wehub-resource-sync
87 lines
4.0 KiB
Plaintext
87 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 "splitGeLUKernel.h"
|
|
#include <MNN/MNNDefine.h>
|
|
|
|
template <typename T, int32_t tHHS, int32_t tTPB>
|
|
__global__ void splitGeLUKernel(T const* input0, T const* input1, T* output, float const fDivRecip, float const fAdd, float const fMul)
|
|
{
|
|
MNN_ASSERT(input0 != nullptr);
|
|
MNN_ASSERT(output != nullptr);
|
|
|
|
int32_t indexInput = blockIdx.x * tHHS * 2 + threadIdx.x;
|
|
int32_t indexInput1 = threadIdx.x;
|
|
int32_t indexOutput = blockIdx.x * tHHS + threadIdx.x;
|
|
float valueL, valueR;
|
|
#pragma unroll
|
|
for (int32_t i = 0; i < tHHS / tTPB; ++i)
|
|
{
|
|
if(input1 == nullptr) {
|
|
valueL = static_cast<float>(input0[indexInput]);
|
|
valueR = static_cast<float>(input0[indexInput + tHHS]);
|
|
} else {
|
|
valueL = static_cast<float>(input0[indexInput]) + static_cast<float>(input1[indexInput1]);
|
|
valueR = static_cast<float>(input0[indexInput + tHHS]) + static_cast<float>(input1[indexInput1 + tHHS]);
|
|
indexInput1 += tTPB;
|
|
}
|
|
float tmp = valueR;
|
|
tmp *= fDivRecip;
|
|
tmp = erff(tmp);
|
|
tmp += fAdd;
|
|
tmp *= valueR;
|
|
tmp *= fMul;
|
|
tmp *= valueL;
|
|
output[indexOutput] = static_cast<T>(tmp);
|
|
indexInput += tTPB;
|
|
indexOutput += tTPB;
|
|
}
|
|
return;
|
|
}
|
|
|
|
template <typename T>
|
|
int32_t launchSplitGeLUKernel(int32_t gridSize, int32_t nHalfHiddenSize, T const* input0, T const* input1, T* output,
|
|
float const fDiv, float const fAdd, float const fMul, cudaStream_t stream)
|
|
{
|
|
MNN_ASSERT(fDiv != 0.F);
|
|
|
|
auto const fDivRecip = 1.F / fDiv;
|
|
constexpr int32_t kTPB = 256; // thread per block
|
|
switch (nHalfHiddenSize)
|
|
{
|
|
case 1280: (splitGeLUKernel<T, 1280, kTPB>) <<<gridSize, kTPB, 0, stream>>>(input0, (T const*)input1, output, fDivRecip, fAdd, fMul); break;
|
|
case 2560: (splitGeLUKernel<T, 2560, kTPB>) <<<gridSize, kTPB, 0, stream>>>(input0, (T const*)input1, output, fDivRecip, fAdd, fMul); break;
|
|
case 5120: (splitGeLUKernel<T, 5120, kTPB>) <<<gridSize, kTPB, 0, stream>>>(input0, (T const*)input1, output, fDivRecip, fAdd, fMul); break;
|
|
}
|
|
|
|
checkKernelErrors;
|
|
return 0;
|
|
}
|
|
|
|
template __global__ void splitGeLUKernel<float, 1280, 256>(float const*, float const*, float*, float const, float const, float const);
|
|
template __global__ void splitGeLUKernel<float, 2560, 256>(float const*, float const*, float*, float const, float const, float const);
|
|
template __global__ void splitGeLUKernel<float, 5120, 256>(float const*, float const*, float*, float const, float const, float const);
|
|
template __global__ void splitGeLUKernel<half, 1280, 256>(half const*, half const*, half*, float const, float const, float const);
|
|
template __global__ void splitGeLUKernel<half, 2560, 256>(half const*, half const*, half*, float const, float const, float const);
|
|
template __global__ void splitGeLUKernel<half, 5120, 256>(half const*, half const*, half*, float const, float const, float const);
|
|
|
|
template int32_t launchSplitGeLUKernel<float>(int32_t gridSize, int32_t nHalfHiddenSize,
|
|
float const* input0, float const* input1, float* output, float const fDiv, float const fAdd, float const fMul, cudaStream_t stream);
|
|
|
|
template int32_t launchSplitGeLUKernel<half>(int32_t gridSize, int32_t nHalfHiddenSize,
|
|
half const* input0, half const* input1, half* output, float const fDiv, float const fAdd, float const fMul, cudaStream_t stream);
|