本文分享自華爲雲社區《Ascend C 自定義算子 Kernel Launch調用入門》,作者: jackwangcumt。
1 Kernel Launch概述
根據官方說明文檔的介紹,Ascend C對外開放核函數的基礎調用(Kernel Launch)方式,是爲了簡化Ascend C 自定義算子的開發流程,提供更易用的調試調優功能。當開發者完成算子核函數的開發和Tiling實現後,即可通過AscendCL運行時接口,完成算子的調用並實現自己的推理應用;同時提供簡易的kernel開發工程,開發者僅需提供kernel側實現,基於工程框架可以快速實現Kernel Launch。本文實驗前提是完成了《Ascend C 自定義PRelu算子》博文的相關算子開發工程。網址爲:https://bbs.huaweicloud.com/blogs/425244 。請注意:
- 8.0.RC1.alpha002 當前版本,Kernel Launch開放式編程爲試用特性,不支持應用於商用產品中。
- 8.0.RC1.alpha002 當前版本暫不支持獲取用戶workspace特性。
2 Kernel Launch調用方式
ACLRT_LAUNCH_KERNEL調用方式對內核調用符方式進行了功能加強,核函數的調用是異步的,調用接口的使用方法如下:
ACLRT_LAUNCH_KERNEL(kernel_name)(blockDim, stream, argument list);
- kernel_name:算子核函數的名稱。
- blockDim:規定了核函數將會在幾個核上執行。每個執行該核函數的核會被分配一個邏輯ID,即block_idx,可以在覈函數的實現中調用GetBlockIdx來獲取block_idx。
- stream,類型爲aclrtStream,stream用於維護一些異步操作的執行順序,確保按照應用程序中的代碼調用順序在Device上執行。
- argument list:參數列表,與核函數的參數列表保持一致。
爲幫助開發者快速的完成算子的Kernel Launch調試,官方提供了簡易的算子工程,我們可以基於該算子工程中的樣例代碼和工程框架進行算子開發。算子工程支持的如下:
- 該工程支持調試功能,如PRINTF功能、DumpTensor。
- 工程編譯生成的應用程序,可通過msprof命令行方式採集和解析性能數據。
可以參考工程樣例:https://gitee.com/ascend/samples/blob/master/operator/AddCustomSample/KernelLaunch/AddKernelInvocationTilingNeo ,其目錄結構如下所示:
AddKernelInvocationNeo |-- cmake // CMake編譯文件 |-- scripts | ├── gen_data.py // 輸入數據和真值數據生成腳本文件 | ├── verify_result.py // 驗證輸出數據和真值數據是否一致的驗證腳本 |-- CMakeLists.txt // CMake編譯配置文件 |-- add_custom.cpp // 矢量算子kernel實現 |-- data_utils.h // 數據讀入寫出函數 |-- main.cpp // 主函數,調用算子的應用程序,含CPU域及NPU域調用 |-- run.sh // 編譯運行算子的腳本
基於該算子工程,開發者進行算子開發的步驟如下:
- 完成算子kernel側實現。
- 編寫算子調用應用程序main.cpp。
-
編寫CMake編譯配置文件CMakeLists.txt。
- 根據實際需要修改輸入數據和真值數據生成腳本文件gen_data.py和驗證輸出數據和真值數據是否一致的驗證腳本verify_result.py。
- 根據實際需要修改編譯運行算子的腳本run.sh並執行該腳本,完成算子的編譯運行和結果驗證。
3 Kernel Launch實現
在PReluSample目錄下新建一個目錄KernelLaunch,用於存放Kernel Launch調用方式的工程代碼,我這裏參考官方的https://gitee.com/ascend/samples/tree/master/operator/LeakyReluCustomSample/KernelLaunch/
LeakyReluKernelInvocation樣例工程,並修改了相關參數,p_relu_custom.cpp 代碼如下所示:
#include "kernel_operator.h" using namespace AscendC; constexpr int32_t BUFFER_NUM = 2; constexpr int32_t TOTAL_LENGTH = 8 * 200 * 1024; constexpr int32_t TILE_NUM = 32; constexpr float alpha = 0.002; class KernelPRelu { public: __aicore__ inline KernelPRelu() {} __aicore__ inline void Init(GM_ADDR x, GM_ADDR y, uint32_t totalLength, uint32_t tileNum, float alpha) { PRINTF("[npu debug] >>> GetBlockNum() %d", GetBlockNum()); ASSERT(GetBlockNum() != 0 && "block dim can not be zero!"); this->blockLength = totalLength / GetBlockNum(); this->tileNum = tileNum; this->alpha = static_cast<float>(alpha); ASSERT(tileNum != 0 && "tile num can not be zero!"); this->tileLength = this->blockLength / tileNum / BUFFER_NUM; // get start index for current core, core parallel xGm.SetGlobalBuffer((__gm__ float*)x + this->blockLength * GetBlockIdx(), this->blockLength); yGm.SetGlobalBuffer((__gm__ float*)y + this->blockLength * GetBlockIdx(), this->blockLength); // pipe alloc memory to queue, the unit is Bytes pipe.InitBuffer(inQueueX, BUFFER_NUM, this->tileLength * sizeof(float)); pipe.InitBuffer(outQueueY, BUFFER_NUM, this->tileLength * sizeof(float)); pipe.InitBuffer(tmpBuffer1, this->tileLength * sizeof(float)); //pipe.InitBuffer(tmpBuffer2, this->tileLength * sizeof(float)); } __aicore__ inline void Process() { // loop count need to be doubled, due to double buffer int32_t loopCount = this->tileNum * BUFFER_NUM; // tiling strategy, pipeline parallel for (int32_t i = 0; i < loopCount; i++) { CopyIn(i); Compute(i); CopyOut(i); } } private: __aicore__ inline void CopyIn(int32_t progress) { // alloc tensor from queue memory LocalTensor<float> xLocal = inQueueX.AllocTensor<float>(); // copy progress_th tile from global tensor to local tensor DataCopy(xLocal, xGm[progress * this->tileLength], this->tileLength); // enque input tensors to VECIN queue inQueueX.EnQue(xLocal); } __aicore__ inline void Compute(int32_t progress) { // deque input tensors from VECIN queue LocalTensor<float> xLocal = inQueueX.DeQue<float>(); LocalTensor<float> yLocal = outQueueY.AllocTensor<float>(); LocalTensor<float> tmpTensor1 = tmpBuffer1.Get<float>(); float inputVal = 0.0; Maxs(tmpTensor1, xLocal, inputVal, this->tileLength); // x >= 0 --> x // x < 0 Mins(xLocal, xLocal, inputVal, this->tileLength); Muls(xLocal, xLocal, this->alpha, this->tileLength); Add(yLocal, xLocal, tmpTensor1, this->tileLength); outQueueY.EnQue<float>(yLocal); // free input tensors for reuse inQueueX.FreeTensor(xLocal); } __aicore__ inline void CopyOut(int32_t progress) { // deque output tensor from VECOUT queue LocalTensor<float> yLocal = outQueueY.DeQue<float>(); // copy progress_th tile from local tensor to global tensor DataCopy(yGm[progress * this->tileLength], yLocal, this->tileLength); // free output tensor for reuse outQueueY.FreeTensor(yLocal); } private: TPipe pipe; TBuf<QuePosition::VECCALC> tmpBuffer1; //TBuf<QuePosition::VECCALC> tmpBuffer1, tmpBuffer2; // create queues for input, in this case depth is equal to buffer num TQue<QuePosition::VECIN, BUFFER_NUM> inQueueX; // create queue for output, in this case depth is equal to buffer num TQue<QuePosition::VECOUT, BUFFER_NUM> outQueueY; GlobalTensor<float> xGm, yGm; uint32_t blockLength; uint32_t tileNum; uint32_t tileLength; float alpha; }; extern "C" __global__ __aicore__ void p_relu_custom(GM_ADDR x, GM_ADDR y) { //GET_TILING_DATA(tiling_data, tiling); // TODO: user kernel impl KernelPRelu op; op.Init(x, y, TOTAL_LENGTH, TILE_NUM, alpha); op.Process(); } #ifndef __CCE_KT_TEST__ // call of kernel function void p_relu_custom_do(uint32_t blockDim, void* l2ctrl, void* stream, uint8_t* x, uint8_t* y) { p_relu_custom<<<blockDim, l2ctrl, stream>>>(x, y); } #endif
main.cpp 代碼如下所示 :
/* * Copyright (c) Huawei Technologies Co., Ltd. 2022-2023. All rights reserved. * This file constains code of cpu debug and npu code.We read data from bin file * and write result to file. */ #include "data_utils.h" #ifndef __CCE_KT_TEST__ #include "acl/acl.h" extern void p_relu_custom_do(uint32_t coreDim, void* l2ctrl, void* stream, uint8_t* x, uint8_t* y); #else #include "tikicpulib.h" extern "C" __global__ __aicore__ void p_relu_custom(GM_ADDR x, GM_ADDR y); #endif int32_t main(int32_t argc, char* argv[]) { uint32_t blockDim = 8; size_t inputByteSize = 8 * 200 * 1024 * sizeof(float); size_t outputByteSize = 8 * 200 * 1024 * sizeof(float); #ifdef __CCE_KT_TEST__ // CPU uint8_t* x = (uint8_t*)AscendC::GmAlloc(inputByteSize); uint8_t* y = (uint8_t*)AscendC::GmAlloc(outputByteSize); printf("[cpu debug]>>> inputByteSize: %d\n", inputByteSize); ReadFile("./input/input_x.bin", inputByteSize, x, inputByteSize); AscendC::SetKernelMode(KernelMode::AIV_MODE); ICPU_RUN_KF(p_relu_custom, blockDim, x, y); // use this macro for cpu debug WriteFile("./output/output_y.bin", y, outputByteSize); AscendC::GmFree((void *)x); AscendC::GmFree((void *)y); #else // NPU //CHECK_ACL(aclInit(nullptr)); CHECK_ACL(aclInit("./acl.json")); aclrtContext context; int32_t deviceId = 0; CHECK_ACL(aclrtSetDevice(deviceId)); CHECK_ACL(aclrtCreateContext(&context, deviceId)); aclrtStream stream = nullptr; CHECK_ACL(aclrtCreateStream(&stream)); uint8_t *xHost, *yHost; uint8_t *xDevice, *yDevice; CHECK_ACL(aclrtMallocHost((void**)(&xHost), inputByteSize)); CHECK_ACL(aclrtMallocHost((void**)(&yHost), outputByteSize)); CHECK_ACL(aclrtMalloc((void**)&xDevice, inputByteSize, ACL_MEM_MALLOC_HUGE_FIRST)); CHECK_ACL(aclrtMalloc((void**)&yDevice, outputByteSize, ACL_MEM_MALLOC_HUGE_FIRST)); ReadFile("./input/input_x.bin", inputByteSize, xHost, inputByteSize); CHECK_ACL(aclrtMemcpy(xDevice, inputByteSize, xHost, inputByteSize, ACL_MEMCPY_HOST_TO_DEVICE)); p_relu_custom_do(blockDim, nullptr, stream, xDevice, yDevice); CHECK_ACL(aclrtSynchronizeStream(stream)); CHECK_ACL(aclrtMemcpy(yHost, outputByteSize, yDevice, outputByteSize, ACL_MEMCPY_DEVICE_TO_HOST)); WriteFile("./output/output_y.bin", yHost, outputByteSize); CHECK_ACL(aclrtFree(xDevice)); CHECK_ACL(aclrtFree(yDevice)); CHECK_ACL(aclrtFreeHost(xHost)); CHECK_ACL(aclrtFreeHost(yHost)); CHECK_ACL(aclrtDestroyStream(stream)); CHECK_ACL(aclrtDestroyContext(context)); CHECK_ACL(aclrtResetDevice(deviceId)); CHECK_ACL(aclFinalize()); #endif return 0; }
執行如下代碼進行NPU上板調試和CPU調試:
#npu bash run.sh Ascend310P1 npu_onboard # cpu bash run.sh Ascend310P1 cpu