溄

Hello, I’m HONG !

4 minute read

华为昇腾CANN算子开发实战：从Host到Device的向量减法实现

本文通过一个完整的向量减法（z = x - y）算子代码，深入讲解基于华为昇腾CANN的算子开发流程。从AscendCL的Host侧内存管理，到AscendC的核函数与流水线并行设计，带你一步步理解昇腾AI芯片的编程模型。

1. 背景与目标

在昇腾AI处理器的开发中，算子（Operator） 是计算的核心单元。CANN（Compute Architecture for Neural Networks）提供了AscendCL（Host侧）和AscendC（Device侧）两套编程接口。本文的目标是：

  • 实现一个向量减法算子 sub_custom，支持 z[i] = x[i] - y[i]。
  • 利用昇腾芯片的 AI Core 进行并行计算。
  • 采用 TPipe + TQue 实现流水线数据处理，提升吞吐。

2. 整体架构概览

代码结构上分为三个层级：

层级 文件/模块 作用
Host 主程序 main() 初始化环境，分配内存，调用核函数，验证结果
Host 调用层 kernel_sub() 封装AscendCL API，完成Host↔Device数据搬运
Device 核函数 sub_custom + KernelSub 实际在AI Core上执行的并行计算逻辑

整个流程如下：

Host (CPU) → aclrtMalloc → 搬运数据 → 核函数启动 → 同步 → 回拷结果 → Host
                            ↘  Device (AI Core) 执行 sub_custom

3. Host 侧：AscendCL 环境与内存管理

Host 侧的 kernel_sub 函数完成了标准的 “初始化 → 分配 → 搬运 → 启动 → 同步 → 回拷 → 释放” 闭环。

关键代码段：

aclInit(nullptr);
aclrtSetDevice(0);
aclrtCreateStream(&stream);

// 分配Device内存
aclrtMalloc(&xDevice, totalByteSize, ACL_MEM_MALLOC_HUGE_FIRST);
aclrtMalloc(&yDevice, totalByteSize, ACL_MEM_MALLOC_HUGE_FIRST);
aclrtMalloc(&zDevice, totalByteSize, ACL_MEM_MALLOC_HUGE_FIRST);

// 数据搬运 HOST → DEVICE
aclrtMemcpy(xDevice, totalByteSize, xHost, totalByteSize, ACL_MEMCPY_HOST_TO_DEVICE);
aclrtMemcpy(yDevice, totalByteSize, yHost, totalByteSize, ACL_MEMCPY_HOST_TO_DEVICE);

// 启动核函数
sub_custom<<<blockDim, nullptr, stream>>>(xDevice, yDevice, zDevice, tiling);
aclrtSynchronizeStream(stream);

// 结果回拷 DEVICE → HOST
aclrtMemcpy(zHost, totalByteSize, zDevice, totalByteSize, ACL_MEMCPY_DEVICE_TO_HOST);

💡 Tips：aclrtSynchronizeStream 是关键同步点，必须确保核函数完成后再回拷数据，否则结果错误。

4. Device 侧：AscendC 核函数与流水线设计

昇腾的核函数采用了 “多核并行 + 流水线分块” 的编程模型。

4.1 多核数据切分

this->blockLength = totalLength / AscendC::GetBlockNum();

每个核处理的数据范围由 totalLength / blockDim 确定，GetBlockIdx() 获取当前核的编号，实现无冲突的并行。

4.2 流水线并行（TPipe + TQue）

核心思路：将数据搬运（CopyIn）、计算（Compute）、结果搬出（CopyOut）重叠进行，避免计算单元等待数据。

  • TPipe：内存管理管道，负责管理 LocalMemory 中的缓冲区。
  • TQue：队列，承载数据流。我们使用两个 VECIN 队列（输入）和一个 VECOUT 队列（输出）。
AscendC::TPipe pipe;
AscendC::TQue<AscendC::TPosition::VECIN, BUFFER_NUM> inQueueX, inQueueY;
AscendC::TQue<AscendC::TPosition::VECOUT, BUFFER_NUM> outQueueZ;

三步循环流程：

  1. CopyIn：从 GlobalMemory 搬运数据到 LocalTensor 
    AscendC::LocalTensor<float> xLocal = inQueueX.AllocTensor<float>();
AscendC::DataCopy(xLocal, xGm[progress * tileLength], tileLength);
inQueueX.EnQue(xLocal);
  2. Compute：从队列取出数据，计算后放入输出队列 
    AscendC::LocalTensor<float> xLocal = inQueueX.DeQue<float>();
AscendC::LocalTensor<float> zLocal = outQueueZ.AllocTensor<float>();
AscendC::Sub(zLocal, xLocal, yLocal, tileLength);
outQueueZ.EnQue(zLocal);
inQueueX.FreeTensor(xLocal);
  3. CopyOut：将结果写回 GlobalMemory 
    AscendC::LocalTensor<float> zLocal = outQueueZ.DeQue<float>();
AscendC::DataCopy(zGm[progress * tileLength], zLocal, tileLength);
outQueueZ.FreeTensor(zLocal);

这种设计使得数据搬运和计算可以流水线并行，极大提升硬件利用率。

5. 详细代码逐行解析

5.1 自定义分块数据结构

struct SubCustomTilingData {
    uint32_t totalLength;
    uint32_t tileNum;
};

用于向核函数传递总长度和分块数量。

5.2 核函数入口

__global__ __aicore__ void sub_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z, SubCustomTilingData tiling) {
    KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_AIV_ONLY);
    KernelSub kernelSub;
    kernelSub.Init(x, y, z, tiling.totalLength, tiling.tileNum);
    kernelSub.Process();
}

KERNEL_TASK_TYPE_DEFAULT 指定任务类型，Init 完成初始化，Process 启动流水线。

5.3 验证与测试

std::vector<float> golden(totalLength, valueX - valueY);
return VerifyResult(output, golden);

构造 Golden 结果并对比，输出成功或失败信息。

6. 关键知识点总结

概念 说明
GM_ADDR Global Memory 地址类型，由 Host 传入
LocalTensor 位于 LocalMemory 的片上数据块，计算核心操作对象
TPipe 管理片上缓冲区（LocalMemory）的生命周期
TQue 队列，支持 AllocTensor → EnQue → DeQue → FreeTensor 模式
DataCopy 在 GlobalMemory 和 LocalMemory 之间搬运数据
AscendC::Sub 向量减法内置指令（SVE 向量化）

7. 完整代码


#include <cstdint>
#include <iostream>
#include <vector>
#include <algorithm>
#include <iterator>
#include "acl/acl.h"
#include "kernel_operator.h"

constexpr uint32_t BUFFER_NUM = 2; // tensor num for each queue

struct SubCustomTilingData
{
    uint32_t totalLength;
    uint32_t tileNum;
};

class KernelSub {
public:
    __aicore__ inline KernelSub(){}
    __aicore__ inline void Init(GM_ADDR x, GM_ADDR y, GM_ADDR z, uint32_t totalLength, uint32_t tileNum)
    {
        // 请补充……

        this->blockLength = totalLength / AscendC::GetBlockNum();
        this->tileNum = tileNum;
        this->tileLength = this->blockLength / this->tileNum / BUFFER_NUM;

        xGm.SetGlobalBuffer((__gm__ float *)x + this->blockLength *AscendC::GetBlockIdx(),this->blockLength);
        yGm.SetGlobalBuffer((__gm__ float *)y + this->blockLength *AscendC::GetBlockIdx(),this->blockLength);
        zGm.SetGlobalBuffer((__gm__ float *)z + this->blockLength *AscendC::GetBlockIdx(),this->blockLength);

        this->pipe.InitBuffer(this->inQueueX,BUFFER_NUM,this->tileLength * sizeof(float));
        this->pipe.InitBuffer(this->inQueueY,BUFFER_NUM,this->tileLength * sizeof(float));
        this->pipe.InitBuffer(this->outQueueZ,BUFFER_NUM,this->tileLength * sizeof(float));
    }
    __aicore__ inline void Process()
    {
        // 请补充……
        int32_t loopCount = this->tileNum * BUFFER_NUM;
        for(int32_t i = 0; i < loopCount; i++)
        {
            this->CopyIn(i);
            this->Compute(i);
            this->CopyOut(i);
        }
    }

private:
    __aicore__ inline void CopyIn(int32_t progress)
    {
        // 请补充……
        AscendC::LocalTensor<float> xLocal = this->inQueueX.AllocTensor<float>();
        AscendC::LocalTensor<float> yLocal = this->inQueueY.AllocTensor<float>();

        AscendC::DataCopy(xLocal,this->xGm[progress * this->tileLength],this->tileLength);
        AscendC::DataCopy(yLocal,this->yGm[progress * this->tileLength],this->tileLength);

        this->inQueueX.EnQue(xLocal);
        this->inQueueY.EnQue(yLocal);
    }
    __aicore__ inline void Compute(int32_t progress)
    {
        // 请补充……
        AscendC::LocalTensor<float> xLocal = this->inQueueX.DeQue<float>();
        AscendC::LocalTensor<float> yLocal = this->inQueueY.DeQue<float>();

        AscendC::LocalTensor<float> zLocal = this->outQueueZ.AllocTensor<float>();

        AscendC::Sub(zLocal,xLocal,yLocal,this->tileLength);

        this->outQueueZ.EnQue<float>(zLocal);

        this->inQueueX.FreeTensor(xLocal);
        this->inQueueY.FreeTensor(yLocal);
    }
    __aicore__ inline void CopyOut(int32_t progress)
    {
        // 请补充……
        AscendC::LocalTensor<float> zLocal = this->outQueueZ.DeQue<float>();

        AscendC::DataCopy(this->zGm[progress * this->tileLength],zLocal,this->tileLength);

        this->outQueueZ.FreeTensor(zLocal);
    }

private:
    // 请补充……
    AscendC::TPipe pipe; //TPipe内存管理对象
    AscendC::TQue<AscendC::TPosition::VECIN,BUFFER_NUM> inQueueX,inQueueY;
    AscendC::TQue<AscendC::TPosition::VECOUT,BUFFER_NUM> outQueueZ;
    AscendC::GlobalTensor<float> xGm;
    AscendC::GlobalTensor<float> yGm;
    AscendC::GlobalTensor<float> zGm;
    uint32_t blockLength;
    uint32_t tileNum;
    uint32_t tileLength;
};

__global__ __aicore__ void sub_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z, SubCustomTilingData tiling)
{
    // 请补充……
    KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_AIV_ONLY);
    KernelSub kernelSub;
    kernelSub.Init(x, y, z, tiling.totalLength, tiling.tileNum);
    kernelSub.Process();
}

std::vector<float> kernel_sub(std::vector<float> &x, std::vector<float> &y)
{
    // 请补充……
    constexpr uint32_t blockDim = 8;
    uint32_t totalLength = x.size();
    size_t totalByteSize = totalLength * sizeof(float);
    int32_t deviceId = 0;
    aclrtStream stream = nullptr;
    SubCustomTilingData tiling = {totalLength, 8};

    // HOST侧数据指针
    uint8_t *xHost = reinterpret_cast<uint8_t *>(x.data());
    uint8_t *yHost = reinterpret_cast<uint8_t *>(y.data());
    uint8_t *zHost = nullptr;

    // DEVICE侧数据指针
    uint8_t *xDevice = nullptr;
    uint8_t *yDevice = nullptr;
    uint8_t *zDevice = nullptr;

    aclInit(nullptr);

    aclrtSetDevice(deviceId);
    aclrtCreateStream(&stream); 


    aclrtMallocHost((void **)(&zHost),totalByteSize);
    // DEVICE侧数据指针
    aclrtMalloc((void **)(&xDevice),totalByteSize,ACL_MEM_MALLOC_HUGE_FIRST );
    aclrtMalloc((void **)(&yDevice),totalByteSize,ACL_MEM_MALLOC_HUGE_FIRST );
    aclrtMalloc((void **)(&zDevice),totalByteSize,ACL_MEM_MALLOC_HUGE_FIRST );


    aclrtMemcpy(xDevice, totalByteSize, xHost, totalByteSize, ACL_MEMCPY_HOST_TO_DEVICE);
    aclrtMemcpy(yDevice, totalByteSize, yHost, totalByteSize, ACL_MEMCPY_HOST_TO_DEVICE);


    sub_custom<<<blockDim,nullptr,stream>>>(xDevice,yDevice,zDevice,tiling);
    aclrtSynchronizeStream(stream);


    aclrtMemcpy(zHost, totalByteSize, zDevice, totalByteSize, ACL_MEMCPY_DEVICE_TO_HOST);

    std::vector<float> output((float *)zHost, (float *)(zHost + totalByteSize));



    aclrtFree((void *)xDevice);
    aclrtFree((void *)yDevice);
    aclrtFree((void *)zDevice);
    aclrtFreeHost((void *)zHost);
    aclrtDestroyStream(stream);
    aclrtResetDevice(deviceId);

    aclFinalize();

    return output;

}

uint32_t VerifyResult(std::vector<float> &output, std::vector<float> &golden)
{
    auto printTensor = [](std::vector<float> &tensor, const char *name) {
        constexpr size_t maxPrintSize = 200;
        std::cout << name << ": ";
        std::copy(tensor.begin(), tensor.begin() + std::min(tensor.size(), maxPrintSize),
            std::ostream_iterator<float>(std::cout, " "));
        if (tensor.size() > maxPrintSize) {
            std::cout << "...";
        }
        std::cout << std::endl;
    };
    printTensor(output, "Output");
    printTensor(golden, "Golden");
    if (std::equal(golden.begin(), golden.end(), output.begin())) {
        std::cout << "[Success] Case accuracy is verification passed." << std::endl;
        return 0;
    } else {
        std::cout << "[Failed] Case accuracy is verification failed!" << std::endl;
        return 1;
    }
    return 0;
}

int32_t main(int32_t argc, char *argv[])
{
    constexpr uint32_t totalLength = 8 * 2048;
    constexpr float valueX = 1.2f;
    constexpr float valueY = 2.3f;
    std::vector<float> x(totalLength, valueX);
    std::vector<float> y(totalLength, valueY);

    // 请补充……
    std::vector<float> output = kernel_sub(x, y);

    std::vector<float> golden(totalLength, valueX - valueY);
    return VerifyResult(output, golden);
}

参考文献：

  • 华为昇腾 CANN 官方文档 (v6.3.0+)
  • 《Ascend C 编程指南》
  • 《AscendCL 接口参考》

You May Also Enjoy