/**
* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
/* !
* \file erfc.asc
* \brief
*/
#include "acl/acl.h"
#include "data_utils.h"
#include "kernel_operator.h"
template <typename T, int32_t apiMode>
class KernelErfc {
public:
__aicore__ inline KernelErfc() {}
__aicore__ inline void Init(GM_ADDR srcGm, GM_ADDR dstGm, uint32_t srcSize, uint32_t calCount,
AscendC::TPipe* pipeIn)
{
pipe = pipeIn;
srcGlobal.SetGlobalBuffer(reinterpret_cast<__gm__ T*>(srcGm), srcSize);
dstGlobal.SetGlobalBuffer(reinterpret_cast<__gm__ T*>(dstGm), srcSize);
bufferSize = srcSize;
count = calCount;
alignDataSize = AscendC::DivCeil(srcSize * sizeof(T), 32) * 32;
alignDataNum = alignDataSize / sizeof(T);
pipe->InitBuffer(inQueueX, 1, alignDataSize);
pipe->InitBuffer(inQueueX2, 1, alignDataNum * sizeof(uint8_t));
pipe->InitBuffer(outQueue, 1, alignDataSize);
}
__aicore__ inline void Process()
{
AscendC::AscendCUtils::SetOverflow(1);
CopyIn();
Compute();
CopyOut();
AscendC::AscendCUtils::SetOverflow(0);
}
__aicore__ inline void CopyIn()
{
AscendC::LocalTensor<T> srcLocal = inQueueX.AllocTensor<T>();
AscendC::LocalTensor<uint8_t> src1Local = inQueueX2.AllocTensor<uint8_t>();
AscendC::DataCopy(srcLocal, srcGlobal, bufferSize);
inQueueX.EnQue(srcLocal);
inQueueX2.EnQue(src1Local);
}
__aicore__ inline void Compute()
{
AscendC::LocalTensor<T> dstLocal = outQueue.AllocTensor<T>();
AscendC::LocalTensor<T> srcLocal = inQueueX.DeQue<T>();
AscendC::LocalTensor<uint8_t> tmpBuf = inQueueX2.DeQue<uint8_t>();
T zero = 0;
AscendC::Duplicate(dstLocal, zero, alignDataNum);
if constexpr (apiMode == 0) {
AscendC::Erfc<T, false>(dstLocal, srcLocal);
} else if constexpr (apiMode == 1) {
AscendC::Erfc<T, false>(dstLocal, srcLocal, count);
} else if constexpr (apiMode == 2) {
AscendC::Erfc<T, false>(dstLocal, srcLocal, tmpBuf, count);
} else {
AscendC::Erfc<T, false>(dstLocal, srcLocal, tmpBuf);
}
outQueue.EnQue<T>(dstLocal);
inQueueX.FreeTensor(srcLocal);
inQueueX2.FreeTensor(tmpBuf);
}
__aicore__ inline void CopyOut()
{
AscendC::LocalTensor<T> dstLocal = outQueue.DeQue<T>();
AscendC::DataCopy(dstGlobal, dstLocal, alignDataNum);
outQueue.FreeTensor(dstLocal);
}
private:
AscendC::TPipe* pipe;
AscendC::TQue<AscendC::QuePosition::VECIN, 1> inQueueX;
AscendC::TQue<AscendC::QuePosition::VECIN, 1> inQueueX2;
AscendC::TQue<AscendC::QuePosition::VECOUT, 1> outQueue;
AscendC::GlobalTensor<T> srcGlobal;
AscendC::GlobalTensor<T> dstGlobal;
uint32_t bufferSize = 0;
uint32_t alignDataNum = 0;
uint32_t alignDataSize = 0;
uint32_t count = 0;
};
__global__ __aicore__ void erfc_custom(GM_ADDR srcGm, GM_ADDR dstGm)
{
KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_AIV_ONLY);
AscendC::TPipe pipe;
constexpr uint32_t srcSize = 4096;
constexpr uint32_t calCount = 4096;
constexpr uint32_t apiMode = 0;
KernelErfc<float, apiMode> op;
op.Init(srcGm, dstGm, srcSize, calCount, &pipe);
op.Process();
}
static bool CompareResult(const void* outputData, uint32_t outSize)
{
void* goldenData;
aclrtMallocHost((void**)(&goldenData), outSize);
size_t goldenSize = outSize;
bool ret = ReadFile("./output/golden.bin", goldenSize, goldenData, goldenSize);
if (ret) {
printf("ReadFile golden.bin success!\n");
} else {
printf("test failed!\n");
return false;
}
constexpr float EPS = 1e-4;
int64_t wrongNum = 0;
for (size_t i = 0; i < outSize / sizeof(float); i++) {
float a = (reinterpret_cast<const float*>(outputData))[i];
float b = (reinterpret_cast<const float*>(goldenData))[i];
float ae = std::abs(a - b);
float re = ae / std::abs(b);
if (ae > EPS && re > EPS) {
printf("CompareResult golden.bin failed output is %lf, golden is %lf\n", a, b);
wrongNum++;
}
}
aclrtFreeHost(goldenData);
if (wrongNum != 0) {
return false;
} else {
printf("CompareResult golden.bin success!\n");
return true;
}
}
int32_t main(int32_t argc, char* argv[])
{
size_t param1FileSize = 4096 * sizeof(float);
size_t param2FileSize = 4096 * sizeof(float);
uint32_t numBlocks = 1;
aclInit(nullptr);
aclrtContext context;
int32_t deviceId = 0;
aclrtSetDevice(deviceId);
aclrtCreateContext(&context, deviceId);
aclrtStream stream = nullptr;
aclrtCreateStream(&stream);
uint8_t* param1Host;
uint8_t* param1Device;
aclrtMallocHost((void**)(¶m1Host), param1FileSize);
aclrtMalloc((void**)¶m1Device, param1FileSize, ACL_MEM_MALLOC_HUGE_FIRST);
ReadFile("./input/input_x.bin", param1FileSize, param1Host, param1FileSize);
aclrtMemcpy(param1Device, param1FileSize, param1Host, param1FileSize, ACL_MEMCPY_HOST_TO_DEVICE);
uint8_t* param2Host;
uint8_t* param2Device;
aclrtMallocHost((void**)(¶m2Host), param2FileSize);
aclrtMalloc((void**)¶m2Device, param2FileSize, ACL_MEM_MALLOC_HUGE_FIRST);
erfc_custom<<<numBlocks, nullptr, stream>>>(param1Device, param2Device);
aclrtSynchronizeStream(stream);
aclrtFree(param1Device);
aclrtFreeHost(param1Host);
aclrtMemcpy(param2Host, param2FileSize, param2Device, param2FileSize, ACL_MEMCPY_DEVICE_TO_HOST);
WriteFile("./output/output.bin", param2Host, param2FileSize);
bool goldenResult = true;
goldenResult = CompareResult(param2Host, param2FileSize);
if (goldenResult) {
printf("test pass!\n");
} else {
printf("test failed!\n");
}
aclrtFree(param2Device);
aclrtFreeHost(param2Host);
aclrtDestroyStream(stream);
aclrtDestroyContext(context);
aclrtResetDevice(deviceId);
aclFinalize();
return 0;
}