本次主要討論下AscendC算子的開發(fā)流程,基于Kernel直調工程的算子開發(fā)。
1 AscendC算子開發(fā)的基本流程
使用Ascend C完成Add算子核函數(shù)開發(fā);
使用ICPU_RUN_KF CPU調測宏完成算子核函數(shù)CPU側運行驗證;
使用<<<>>>內核調用符完成算子核函數(shù)NPU側運行驗證。
在正式的開發(fā)之前,還需要先完成環(huán)境準備和算子分析工作,開發(fā)Ascend C算子的基本流程如下圖所示:

2 核函數(shù)開發(fā)
本次以add_custom.cpp作為參考用例。Gitee也有對應工程和完整代碼。
operator/AddCustomSample/KernelLaunch/AddKernelInvocationNeo · Ascend/samples - 碼云 - 開源中國 (gitee.com)
2.1 核函數(shù)定義
首先要根據(jù)核函數(shù)定義 核函數(shù)-編程模型-Ascend C算子開發(fā)-算子開發(fā)-開發(fā)指南-CANN社區(qū)版8.0.RC3.alpha003開發(fā)文檔-昇騰社區(qū) (hiascend.com) 的規(guī)則進行核函數(shù)的定義,并在核函數(shù)中調用算子類的Init和Process函數(shù)。
// 給CPU調用
extern "C" __global__ __aicore__ void add_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z)
{
KernelAdd op;
op.Init(x, y, z);
op.Process();
}
// 給NPU調用
#ifndef ASCENDC_CPU_DEBUG
void add_custom_do(uint32_t blockDim, void *stream, uint8_t *x, uint8_t *y, uint8_t *z)
{
add_custom<<<blockDim, nullptr, stream>>>(x, y, z);
}
#endif
2.2 算子類定義
根據(jù)矢量編程范式實現(xiàn)算子類,本樣例中定義KernelAdd算子類,其具體成員如下:
class KernelAdd {
public:
__aicore__ inline KernelAdd(){}
// 初始化函數(shù),完成內存初始化相關操作
__aicore__ inline void Init(GM_ADDR x, GM_ADDR y, GM_ADDR z){}
// 核心處理函數(shù),實現(xiàn)算子邏輯,調用私有成員函數(shù)CopyIn、Compute、CopyOut完成矢量算子的三級流水操作
__aicore__ inline void Process(){}
private:
// 搬入函數(shù),完成CopyIn階段的處理,被核心Process函數(shù)調用
__aicore__ inline void CopyIn(int32_t progress){}
// 計算函數(shù),完成Compute階段的處理,被核心Process函數(shù)調用
__aicore__ inline void Compute(int32_t progress){}
// 搬出函數(shù),完成CopyOut階段的處理,被核心Process函數(shù)調用
__aicore__ inline void CopyOut(int32_t progress){}
private:
AscendC::TPipe pipe; //Pipe內存管理對象
AscendC::TQue<AscendC::QuePosition::VECIN, BUFFER_NUM> inQueueX, inQueueY; //輸入數(shù)據(jù)Queue隊列管理對象,QuePosition為VECIN
AscendC::TQue<AscendC::QuePosition::VECOUT, BUFFER_NUM> outQueueZ; //輸出數(shù)據(jù)Queue隊列管理對象,QuePosition為VECOUT
AscendC::GlobalTensor<half> xGm; //管理輸入輸出Global Memory內存地址的對象,其中xGm, yGm為輸入,zGm為輸出
AscendC::GlobalTensor<half> yGm;
AscendC::GlobalTensor<half> zGm;
};
核函數(shù)調用關系圖

2.3 實現(xiàn)Init,CopyIn,Compute,CopyOut這個4個關鍵函數(shù)
Init函數(shù)初始化輸入資源
__aicore__ inline void Init(GM_ADDR x, GM_ADDR y, GM_ADDR z)
{
xGm.SetGlobalBuffer((__gm__ half *)x + BLOCK_LENGTH * AscendC::GetBlockIdx(), BLOCK_LENGTH);
yGm.SetGlobalBuffer((__gm__ half *)y + BLOCK_LENGTH * AscendC::GetBlockIdx(), BLOCK_LENGTH);
zGm.SetGlobalBuffer((__gm__ half *)z + BLOCK_LENGTH * AscendC::GetBlockIdx(), BLOCK_LENGTH);
pipe.InitBuffer(inQueueX, BUFFER_NUM, TILE_LENGTH * sizeof(half));
pipe.InitBuffer(inQueueY, BUFFER_NUM, TILE_LENGTH * sizeof(half));
pipe.InitBuffer(outQueueZ, BUFFER_NUM, TILE_LENGTH * sizeof(half));
}
Process函數(shù)中通過如下方式調用這三個:
__aicore__ inline void Process()
{
// loop count need to be doubled, due to double buffer
constexpr int32_t loopCount = TILE_NUM * BUFFER_NUM;
// tiling strategy, pipeline parallel
for (int32_t i = 0; i < loopCount; i++) {
CopyIn(i);
Compute(i);
CopyOut(i);
}
}
CopyIn函數(shù)中通過如下方式調用這三個:
1、使用DataCopy接口將GlobalTensor數(shù)據(jù)拷貝到LocalTensor。
2、使用EnQue將LocalTensor放入VecIn的Queue中。
__aicore__ inline void CopyIn(int32_t progress)
{
// alloc tensor from queue memory
AscendC::LocalTensor<half> xLocal = inQueueX.AllocTensor<half>();
AscendC::LocalTensor<half> yLocal = inQueueY.AllocTensor<half>();
// copy progress_th tile from global tensor to local tensor
AscendC::DataCopy(xLocal, xGm[progress * TILE_LENGTH], TILE_LENGTH);
AscendC::DataCopy(yLocal, yGm[progress * TILE_LENGTH], TILE_LENGTH);
// enque input tensors to VECIN queue
inQueueX.EnQue(xLocal);
inQueueY.EnQue(yLocal);
}
Compute函數(shù)實現(xiàn)。
1、使用DeQue從VecIn中取出LocalTensor。
2、使用Ascend C接口Add完成矢量計算。
3、使用EnQue將計算結果LocalTensor放入到VecOut的Queue中。
4、使用FreeTensor將釋放不再使用的LocalTensor。
__aicore__ inline void Compute(int32_t progress)
{
// deque input tensors from VECIN queue
AscendC::LocalTensor<half> xLocal = inQueueX.DeQue<half>();
AscendC::LocalTensor<half> yLocal = inQueueY.DeQue<half>();
AscendC::LocalTensor<half> zLocal = outQueueZ.AllocTensor<half>();
// call Add instr for computation
AscendC::Add(zLocal, xLocal, yLocal, TILE_LENGTH);
// enque the output tensor to VECOUT queue
outQueueZ.EnQue<half>(zLocal);
// free input tensors for reuse
inQueueX.FreeTensor(xLocal);
inQueueY.FreeTensor(yLocal);
}
CopyOut函數(shù)實現(xiàn)。
1、使用DeQue接口從VecOut的Queue中取出LocalTensor。
2、使用DataCopy接口將LocalTensor拷貝到GlobalTensor上。
3、使用FreeTensor將不再使用的LocalTensor進行回收。
__aicore__ inline void CopyOut(int32_t progress)
{
// deque output tensor from VECOUT queue
AscendC::LocalTensor<half> zLocal = outQueueZ.DeQue<half>();
// copy progress_th tile from local tensor to global tensor
AscendC::DataCopy(zGm[progress * TILE_LENGTH], zLocal, TILE_LENGTH);
// free output tensor for reuse
outQueueZ.FreeTensor(zLocal);
}
3 核函數(shù)的運行驗證
異構計算架構中,NPU(kernel側)與CPU(host側)是協(xié)同工作的,完成了kernel側核函數(shù)開發(fā)后,即可編寫host側的核函數(shù)調用程序,實現(xiàn)從host側的APP程序調用算子,執(zhí)行計算過程。
3.1 編寫CPU側調用程序

// 使用GmAlloc分配共享內存,并進行數(shù)據(jù)初始化
uint8_t* x = (uint8_t*)AscendC::GmAlloc(inputByteSize);
uint8_t* y = (uint8_t*)AscendC::GmAlloc(inputByteSize);
uint8_t* z = (uint8_t*)AscendC::GmAlloc(outputByteSize);
ReadFile("./input/input_x.bin", inputByteSize, x, inputByteSize);
ReadFile("./input/input_y.bin", inputByteSize, y, inputByteSize);
// 調用ICPU_RUN_KF調測宏,完成核函數(shù)CPU側的調用
AscendC::SetKernelMode(KernelMode::AIV_MODE);
ICPU_RUN_KF(add_custom, blockDim, x, y, z); // use this macro for cpu debug
// 輸出數(shù)據(jù)寫出
WriteFile("./output/output_z.bin", z, outputByteSize);
// 調用GmFree釋放申請的資源
AscendC::GmFree((void *)x);
AscendC::GmFree((void *)y);
AscendC::GmFree((void *)z);
3.2 編寫NPU側運行算子的調用程序

// AscendCL初始化
CHECK_ACL(aclInit(nullptr));
// 運行管理資源申請
int32_t deviceId = 0;
CHECK_ACL(aclrtSetDevice(deviceId));
aclrtStream stream = nullptr;
CHECK_ACL(aclrtCreateStream(&stream));
// 分配Host內存
uint8_t *xHost, *yHost, *zHost;
uint8_t *xDevice, *yDevice, *zDevice;
CHECK_ACL(aclrtMallocHost((void**)(&xHost), inputByteSize));
CHECK_ACL(aclrtMallocHost((void**)(&yHost), inputByteSize));
CHECK_ACL(aclrtMallocHost((void**)(&zHost), outputByteSize));
// 分配Device內存
CHECK_ACL(aclrtMalloc((void**)&xDevice, inputByteSize, ACL_MEM_MALLOC_HUGE_FIRST));
CHECK_ACL(aclrtMalloc((void**)&yDevice, inputByteSize, ACL_MEM_MALLOC_HUGE_FIRST));
CHECK_ACL(aclrtMalloc((void**)&zDevice, outputByteSize, ACL_MEM_MALLOC_HUGE_FIRST));
// Host內存初始化
ReadFile("./input/input_x.bin", inputByteSize, xHost, inputByteSize);
ReadFile("./input/input_y.bin", inputByteSize, yHost, inputByteSize);
CHECK_ACL(aclrtMemcpy(xDevice, inputByteSize, xHost, inputByteSize, ACL_MEMCPY_HOST_TO_DEVICE));
CHECK_ACL(aclrtMemcpy(yDevice, inputByteSize, yHost, inputByteSize, ACL_MEMCPY_HOST_TO_DEVICE));
// 用內核調用符<<<>>>調用核函數(shù)完成指定的運算,add_custom_do中封裝了<<<>>>調用
add_custom_do(blockDim, nullptr, stream, xDevice, yDevice, zDevice);
CHECK_ACL(aclrtSynchronizeStream(stream));
// 將Device上的運算結果拷貝回Host
CHECK_ACL(aclrtMemcpy(zHost, outputByteSize, zDevice, outputByteSize, ACL_MEMCPY_DEVICE_TO_HOST));
WriteFile("./output/output_z.bin", zHost, outputByteSize);
// 釋放申請的資源
CHECK_ACL(aclrtFree(xDevice));
CHECK_ACL(aclrtFree(yDevice));
CHECK_ACL(aclrtFree(zDevice));
CHECK_ACL(aclrtFreeHost(xHost));
CHECK_ACL(aclrtFreeHost(yHost));
CHECK_ACL(aclrtFreeHost(zHost));
// AscendCL去初始化
CHECK_ACL(aclrtDestroyStream(stream));
CHECK_ACL(aclrtResetDevice(deviceId));
CHECK_ACL(aclFinalize());
3.3 完整main.cpp
/**
* @file main.cpp
*
* Copyright (C) 2024. Huawei Technologies Co., Ltd. All rights reserved.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
*/
#include "data_utils.h"
#ifndef ASCENDC_CPU_DEBUG
#include "acl/acl.h"
extern void add_custom_do(uint32_t blockDim, void *stream, uint8_t *x, uint8_t *y, uint8_t *z);
#else
#include "tikicpulib.h"
extern "C" __global__ __aicore__ void add_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z);
#endif
int32_t main(int32_t argc, char *argv[])
{
uint32_t blockDim = 8;
size_t inputByteSize = 8 * 2048 * sizeof(uint16_t);
size_t outputByteSize = 8 * 2048 * sizeof(uint16_t);
#ifdef ASCENDC_CPU_DEBUG
uint8_t *x = (uint8_t *)AscendC::GmAlloc(inputByteSize);
uint8_t *y = (uint8_t *)AscendC::GmAlloc(inputByteSize);
uint8_t *z = (uint8_t *)AscendC::GmAlloc(outputByteSize);
ReadFile("./input/input_x.bin", inputByteSize, x, inputByteSize);
ReadFile("./input/input_y.bin", inputByteSize, y, inputByteSize);
AscendC::SetKernelMode(KernelMode::AIV_MODE);
ICPU_RUN_KF(add_custom, blockDim, x, y, z); // use this macro for cpu debug
WriteFile("./output/output_z.bin", z, outputByteSize);
AscendC::GmFree((void *)x);
AscendC::GmFree((void *)y);
AscendC::GmFree((void *)z);
#else
CHECK_ACL(aclInit(nullptr));
int32_t deviceId = 0;
CHECK_ACL(aclrtSetDevice(deviceId));
aclrtStream stream = nullptr;
CHECK_ACL(aclrtCreateStream(&stream));
uint8_t *xHost, *yHost, *zHost;
uint8_t *xDevice, *yDevice, *zDevice;
CHECK_ACL(aclrtMallocHost((void **)(&xHost), inputByteSize));
CHECK_ACL(aclrtMallocHost((void **)(&yHost), inputByteSize));
CHECK_ACL(aclrtMallocHost((void **)(&zHost), outputByteSize));
CHECK_ACL(aclrtMalloc((void **)&xDevice, inputByteSize, ACL_MEM_MALLOC_HUGE_FIRST));
CHECK_ACL(aclrtMalloc((void **)&yDevice, inputByteSize, ACL_MEM_MALLOC_HUGE_FIRST));
CHECK_ACL(aclrtMalloc((void **)&zDevice, outputByteSize, ACL_MEM_MALLOC_HUGE_FIRST));
ReadFile("./input/input_x.bin", inputByteSize, xHost, inputByteSize);
ReadFile("./input/input_y.bin", inputByteSize, yHost, inputByteSize);
CHECK_ACL(aclrtMemcpy(xDevice, inputByteSize, xHost, inputByteSize, ACL_MEMCPY_HOST_TO_DEVICE));
CHECK_ACL(aclrtMemcpy(yDevice, inputByteSize, yHost, inputByteSize, ACL_MEMCPY_HOST_TO_DEVICE));
add_custom_do(blockDim, stream, xDevice, yDevice, zDevice);
CHECK_ACL(aclrtSynchronizeStream(stream));
CHECK_ACL(aclrtMemcpy(zHost, outputByteSize, zDevice, outputByteSize, ACL_MEMCPY_DEVICE_TO_HOST));
WriteFile("./output/output_z.bin", zHost, outputByteSize);
CHECK_ACL(aclrtFree(xDevice));
CHECK_ACL(aclrtFree(yDevice));
CHECK_ACL(aclrtFree(zDevice));
CHECK_ACL(aclrtFreeHost(xHost));
CHECK_ACL(aclrtFreeHost(yHost));
CHECK_ACL(aclrtFreeHost(zHost));
CHECK_ACL(aclrtDestroyStream(stream));
CHECK_ACL(aclrtResetDevice(deviceId));
CHECK_ACL(aclFinalize());
#endif
return 0;
}