本文分享自华为云社区《
Ascend C 自定义算子 Kernel Launch调用入门
》，作者： jackwangcumt。
1 Kernel Launch概述
根据官方说明文档的介绍，Ascend C对外开放核函数的基础调用（Kernel Launch）方式，是为了简化Ascend C 自定义算子的开发流程，提供更易用的调试调优功能。当开发者完成算子核函数的开发和Tiling实现后，即可通过AscendCL运行时接口，完成算子的调用并实现自己的推理应用；同时提供简易的kernel开发工程，开发者仅需提供kernel侧实现，基于工程框架可以快速实现Kernel Launch。本文实验前提是完成了《Ascend C 自定义PRelu算子》博文的相关算子开发工程。网址为：https://bbs.huaweicloud.com/blogs/425244 。请注意：
2 Kernel Launch调用方式
ACLRT_LAUNCH_KERNEL调用方式对内核调用符方式进行了功能加强，核函数的调用是异步的，调用接口的使用方法如下：
ACLRT_LAUNCH_KERNEL(kernel_name)(blockDim, stream, argument list);
- kernel_name：算子核函数的名称。
- blockDim：规定了核函数将会在几个核上执行。每个执行该核函数的核会被分配一个逻辑ID，即block_idx，可以在核函数的实现中调用
GetBlockIdx
来获取block_idx。
- stream，类型为aclrtStream，stream用于维护一些异步操作的执行顺序，确保按照应用程序中的代码调用顺序在Device上执行。
- argument list：参数列表，与核函数的参数列表保持一致。
为帮助开发者快速的完成算子的Kernel Launch调试，官方提供了简易的算子工程，我们可以基于该算子工程中的样例代码和工程框架进行算子开发。算子工程支持的如下：
可以参考工程样例：https://gitee.com/ascend/samples/blob/master/operator/AddCustomSample/KernelLaunch/AddKernelInvocationTilingNeo ，其目录结构如下所示：
AddKernelInvocationNeo|-- cmake //CMake编译文件
|--scripts| ├── gen_data.py //输入数据和真值数据生成脚本文件
| ├── verify_result.py //验证输出数据和真值数据是否一致的验证脚本
|-- CMakeLists.txt //CMake编译配置文件
|-- add_custom.cpp //矢量算子kernel实现
|-- data_utils.h //数据读入写出函数
|-- main.cpp //主函数，调用算子的应用程序，含CPU域及NPU域调用
|-- run.sh //编译运行算子的脚本
基于该算子工程，开发者进行算子开发的步骤如下：
3 Kernel Launch实现
在PReluSample目录下新建一个目录KernelLaunch，用于存放Kernel Launch调用方式的工程代码，我这里参考官方的https://gitee.com/ascend/samples/tree/master/operator/LeakyReluCustomSample/KernelLaunch/
LeakyReluKernelInvocation样例工程，并修改了相关参数，p_relu_custom.cpp 代码如下所示：
#include "kernel_operator.h"
using namespaceAscendC;
constexpr int32_t BUFFER_NUM= 2;
constexpr int32_t TOTAL_LENGTH= 8 * 200 * 1024;
constexpr int32_t TILE_NUM= 32;
constexprfloat alpha = 0.002;classKernelPRelu {public:
__aicore__ inline KernelPRelu() {}
__aicore__ inlinevoid Init(GM_ADDR x, GM_ADDR y, uint32_t totalLength, uint32_t tileNum, floatalpha)
{
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__ inlinevoidProcess()
{//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__ inlinevoidCopyIn(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__ inlinevoidCompute(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__ inlinevoidCopyOut(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;floatalpha;
};extern "C" __global__ __aicore__ voidp_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 代码如下所示 :
/** 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__ voidp_relu_custom(GM_ADDR x, GM_ADDR y);#endifint32_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;
}
执行如下代码进行NPU上板调试和CPU调试：
#npu
bash run.sh Ascend310P1 npu_onboard
# cpu
bash run.sh Ascend310P1 cpu
点击关注，第一时间了解华为云新鲜技术~
