/**
* 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 quant.asc
* \brief
*/
#include "acl/acl.h"
#include "data_utils.h"
#include "kernel_operator.h"
#include "tiling/tiling_api.h"
namespace optiling {
BEGIN_TILING_DATA_DEF(QuantCustomTilingData)
TILING_DATA_FIELD_DEF(uint32_t, dataLength);
TILING_DATA_FIELD_DEF(uint32_t, sharedTmpBufferSize);
END_TILING_DATA_DEF;
REGISTER_TILING_DATA_CLASS(QuantCustom, QuantCustomTilingData)
} // namespace optiling
void ComputeTiling(const uint32_t dataLength, optiling::QuantCustomTilingData& tiling)
{
std::vector<int64_t> shapeVec = {dataLength};
ge::Shape srcShape(shapeVec);
uint32_t typeSize = sizeof(float);
uint32_t maxTmpSize;
uint32_t minTmpSize;
AscendC::GetAscendQuantMaxMinTmpSize(srcShape, typeSize, maxTmpSize, minTmpSize);
uint32_t localWorkspaceSize = minTmpSize;
tiling.set_dataLength(dataLength);
tiling.set_sharedTmpBufferSize(localWorkspaceSize);
}
uint8_t* GetTilingBuf(optiling::QuantCustomTilingData* tilingData)
{
uint32_t tilingSize = sizeof(optiling::QuantCustomTilingData);
uint8_t* buf = (uint8_t*)malloc(tilingSize);
tilingData->SaveToBuffer(buf, tilingSize);
return buf;
}
uint8_t* GenerateTiling(uint32_t dataLength)
{
optiling::QuantCustomTilingData tiling;
ComputeTiling(dataLength, tiling);
return GetTilingBuf(&tiling);
}
namespace MyCustomKernel {
struct VecTiling {
uint32_t dataLength;
uint32_t sharedTmpBufferSize;
};
template <typename srcType>
class KernelQuant {
public:
__aicore__ inline KernelQuant() {}
__aicore__ inline void Init(GM_ADDR srcGm, GM_ADDR dstGm, uint32_t inputSize, AscendC::TPipe* pipeIn)
{
pipe = pipeIn;
dataSize = inputSize;
srcGlobal.SetGlobalBuffer(reinterpret_cast<__gm__ srcType*>(srcGm), dataSize);
dstGlobal.SetGlobalBuffer(reinterpret_cast<__gm__ int8_t*>(dstGm), dataSize);
pipe->InitBuffer(inQueueX, 1, dataSize * sizeof(srcType));
pipe->InitBuffer(outQueue, 1, dataSize * sizeof(int8_t));
}
__aicore__ inline void Process()
{
CopyIn();
Compute();
CopyOut();
}
private:
__aicore__ inline void CopyIn()
{
AscendC::LocalTensor<srcType> srcLocal = inQueueX.AllocTensor<srcType>();
AscendC::DataCopy(srcLocal, srcGlobal, dataSize);
inQueueX.EnQue(srcLocal);
}
__aicore__ inline void Compute()
{
const float scale = 2.0;
const float offset = 0.9;
AscendC::LocalTensor<int8_t> dstLocal = outQueue.AllocTensor<int8_t>();
AscendC::LocalTensor<srcType> srcLocal = inQueueX.DeQue<srcType>();
AscendC::AscendQuant<srcType>(dstLocal, srcLocal, scale, offset, dataSize);
outQueue.EnQue<int8_t>(dstLocal);
inQueueX.FreeTensor(srcLocal);
}
__aicore__ inline void CopyOut()
{
AscendC::LocalTensor<int8_t> dstLocal = outQueue.DeQue<int8_t>();
AscendC::DataCopy(dstGlobal, dstLocal, dataSize);
outQueue.FreeTensor(dstLocal);
}
private:
AscendC::GlobalTensor<srcType> srcGlobal;
AscendC::GlobalTensor<int8_t> dstGlobal;
AscendC::TPipe* pipe;
AscendC::TQue<AscendC::TPosition::VECIN, 1> inQueueX;
AscendC::TQue<AscendC::TPosition::VECOUT, 1> outQueue;
uint32_t dataSize = 0;
};
} // namespace MyCustomKernel
__aicore__ inline void CopyTiling(MyCustomKernel::VecTiling* tiling, GM_ADDR tilingGM)
{
uint32_t* ptr = reinterpret_cast<uint32_t*>(tiling);
auto tiling32 = reinterpret_cast<__gm__ uint32_t*>(tilingGM);
for (uint32_t i = 0; i < sizeof(MyCustomKernel::VecTiling) / sizeof(uint32_t); i++, ptr++) {
*ptr = *(tiling32 + i);
}
return;
}
extern "C" __global__ __aicore__ void quant_custom(GM_ADDR srcGm, GM_ADDR dstGm, GM_ADDR workspace, GM_ADDR tiling)
{
KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_AIV_ONLY);
AscendC::TPipe pipe;
MyCustomKernel::KernelQuant<float> op;
MyCustomKernel::VecTiling tilingData;
CopyTiling(&tilingData, tiling);
op.Init(srcGm, dstGm, tilingData.dataLength, &pipe);
op.Process();
}
constexpr uint32_t NUM_BLOCKS = 1;
constexpr uint32_t DATALENGTH = 1024;
constexpr uint32_t TILINGDATA_SIZE = 2;
constexpr uint32_t WORKSPACE_SIZE = 16 * 1024 * 1024;
static bool CompareResult(const void* outputData, int64_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");
aclrtFreeHost(goldenData);
return false;
}
constexpr float EPS = 1e-5;
int64_t wrongNum = 0;
for (int i = 0; i < outSize / sizeof(int8_t); i++) {
float a = (reinterpret_cast<const int8_t*>(outputData))[i];
float b = (reinterpret_cast<const int8_t*>(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[%d] is %lf, golden is %lf\n", i, 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[])
{
uint32_t numBlocks = NUM_BLOCKS;
size_t dataLength = DATALENGTH;
size_t inpSize = DATALENGTH * sizeof(float);
size_t outSize = DATALENGTH * sizeof(int8_t);
size_t tilingFileSize = TILINGDATA_SIZE * sizeof(uint32_t);
size_t workspaceSize = WORKSPACE_SIZE;
uint8_t* tilingBuf = GenerateTiling(dataLength);
aclInit(nullptr);
aclrtContext context;
int32_t deviceId = 0;
aclrtSetDevice(deviceId);
aclrtCreateContext(&context, deviceId);
aclrtStream stream = nullptr;
aclrtCreateStream(&stream);
uint8_t *inputHost, *outputHost, *workspaceHost, *tilingHost;
uint8_t *inputDevice, *outputDevice, *workspaceDevice, *tilingDevice;
aclrtMallocHost((void**)(&inputHost), inpSize);
aclrtMallocHost((void**)(&outputHost), outSize);
aclrtMallocHost((void**)(&workspaceHost), workspaceSize);
aclrtMallocHost((void**)(&tilingHost), tilingFileSize);
aclrtMalloc((void**)&inputDevice, inpSize, ACL_MEM_MALLOC_HUGE_FIRST);
aclrtMalloc((void**)&outputDevice, outSize, ACL_MEM_MALLOC_HUGE_FIRST);
aclrtMalloc((void**)&workspaceDevice, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
aclrtMalloc((void**)&tilingDevice, tilingFileSize, ACL_MEM_MALLOC_HUGE_FIRST);
ReadFile("./input/input.bin", inpSize, inputHost, inpSize);
aclrtMemcpy(tilingDevice, tilingFileSize, tilingBuf, tilingFileSize, ACL_MEMCPY_HOST_TO_DEVICE);
aclrtMemcpy(inputDevice, inpSize, inputHost, inpSize, ACL_MEMCPY_HOST_TO_DEVICE);
quant_custom<<<numBlocks, nullptr, stream>>>(inputDevice, outputDevice, workspaceDevice, tilingDevice);
aclrtSynchronizeStream(stream);
aclrtMemcpy(outputHost, outSize, outputDevice, outSize, ACL_MEMCPY_DEVICE_TO_HOST);
WriteFile("./output/output.bin", outputHost, outSize);
bool goldenResult = CompareResult(outputHost, outSize);
if (goldenResult) {
printf("test pass!\n");
} else {
printf("test failed!\n");
}
aclrtFree(inputDevice);
aclrtFree(outputDevice);
aclrtFree(tilingDevice);
aclrtFreeHost(inputHost);
aclrtFreeHost(outputHost);
aclrtFreeHost(tilingHost);
aclrtFree(workspaceDevice);
aclrtFreeHost(workspaceHost);
aclrtDestroyStream(stream);
aclrtDestroyContext(context);
aclrtResetDevice(deviceId);
aclFinalize();
free(tilingBuf);
return 0;
}