roctracer 的应用示例

Eloudy2024-09-15 9:11

1，不用 roctracer 的普通场景

mt.cpp

/* Copyright (c) 2018-2022 Advanced Micro Devices, Inc.

 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal
 in the Software without restriction, including without limitation the rights
 to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 copies of the Software, and to permit persons to whom the Software is
 furnished to do so, subject to the following conditions:

 The above copyright notice and this permission notice shall be included in
 all copies or substantial portions of the Software.

 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 THE SOFTWARE. */

#include <iostream>

// hip header file
#include <hip/hip_runtime.h>

#define HIP_CALL(call)                                                                             \
  do {                                                                                             \
    hipError_t err = call;                                                                         \
    if (err != hipSuccess) {                                                                       \
      fprintf(stderr, "%s\n", hipGetErrorString(err));                                             \
      abort();                                                                                     \
    }                                                                                              \
  } while (0)

#define WIDTH 1024


#define NUM (WIDTH * WIDTH)

#define THREADS_PER_BLOCK_X 4
#define THREADS_PER_BLOCK_Y 4
#define THREADS_PER_BLOCK_Z 1

// Device (Kernel) function, it must be void
__global__ void matrixTranspose(float* out, float* in, const int width) {
  int x = hipBlockDim_x * hipBlockIdx_x + hipThreadIdx_x;
  int y = hipBlockDim_y * hipBlockIdx_y + hipThreadIdx_y;

  out[y * width + x] = in[x * width + y];
}

// CPU implementation of matrix transpose
void matrixTransposeCPUReference(float* output, float* input, const unsigned int width) {
  for (unsigned int j = 0; j < width; j++) {
    for (unsigned int i = 0; i < width; i++) {
      output[i * width + j] = input[j * width + i];
    }
  }
}

int main() {
  float* Matrix;
  float* TransposeMatrix;
  float* cpuTransposeMatrix;

  float* gpuMatrix;
  float* gpuTransposeMatrix;

  hipDeviceProp_t devProp;
  HIP_CALL(hipGetDeviceProperties(&devProp, 0));

  std::cerr << "Device name " << devProp.name << std::endl;

  int i;
  int errors;

  Matrix = (float*)malloc(NUM * sizeof(float));
  TransposeMatrix = (float*)malloc(NUM * sizeof(float));
  cpuTransposeMatrix = (float*)malloc(NUM * sizeof(float));

  // initialize the input data
  for (i = 0; i < NUM; i++) {
    Matrix[i] = (float)i * 10.0f;
  }

  // allocate the memory on the device side
  HIP_CALL(hipMalloc((void**)&gpuMatrix, NUM * sizeof(float)));
  HIP_CALL(hipMalloc((void**)&gpuTransposeMatrix, NUM * sizeof(float)));

  uint32_t iterations = 100;
  while (iterations-- > 0) {
    std::cerr << "## Iteration (" << iterations << ") #################" << std::endl;

    // Memory transfer from host to device
    HIP_CALL(hipMemcpy(gpuMatrix, Matrix, NUM * sizeof(float), hipMemcpyHostToDevice));

    // Lauching kernel from host
    hipLaunchKernelGGL(
        matrixTranspose, dim3(WIDTH / THREADS_PER_BLOCK_X, WIDTH / THREADS_PER_BLOCK_Y),
        dim3(THREADS_PER_BLOCK_X, THREADS_PER_BLOCK_Y), 0, 0, gpuTransposeMatrix, gpuMatrix, WIDTH);


    HIP_CALL(
        hipMemcpy(TransposeMatrix, gpuTransposeMatrix, NUM * sizeof(float), hipMemcpyDeviceToHost));


    // CPU MatrixTranspose computation
    matrixTransposeCPUReference(cpuTransposeMatrix, Matrix, WIDTH);

    // verify the results
    errors = 0;
    double eps = 1.0E-6;
    for (i = 0; i < NUM; i++) {
      if (std::abs(TransposeMatrix[i] - cpuTransposeMatrix[i]) > eps) {
        errors++;
      }
    }
    if (errors != 0) {
      fprintf(stderr, "FAILED: %d errors\n", errors);
    } else {
      fprintf(stderr, "PASSED!\n");
    }
  }

  // free the resources on device side
  HIP_CALL(hipFree(gpuMatrix));
  HIP_CALL(hipFree(gpuTransposeMatrix));

  // free the resources on host side
  free(Matrix);
  free(TransposeMatrix);
  free(cpuTransposeMatrix);

  return errors;
}

编译：

$ hipcc mt.cpp -o mt

$ ./mt xxx

不会产生文件；

2，加入roctracer的源文件

MatrixTranspose.cpp：

cpp 复制代码 
/* Copyright (c) 2018-2022 Advanced Micro Devices, Inc.

 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal
 in the Software without restriction, including without limitation the rights
 to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 copies of the Software, and to permit persons to whom the Software is
 furnished to do so, subject to the following conditions:

 The above copyright notice and this permission notice shall be included in
 all copies or substantial portions of the Software.

 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 THE SOFTWARE. */

#include <iostream>

// hip header file
#include <hip/hip_runtime.h>
#include "roctracer_ext.h"
// roctx header file
#include <roctx.h>

#define HIP_CALL(call)                                                                             \
  do {                                                                                             \
    hipError_t err = call;                                                                         \
    if (err != hipSuccess) {                                                                       \
      fprintf(stderr, "%s\n", hipGetErrorString(err));                                             \
      abort();                                                                                     \
    }                                                                                              \
  } while (0)

#define WIDTH 1024


#define NUM (WIDTH * WIDTH)

#define THREADS_PER_BLOCK_X 4
#define THREADS_PER_BLOCK_Y 4
#define THREADS_PER_BLOCK_Z 1

// Device (Kernel) function, it must be void
__global__ void matrixTranspose(float* out, float* in, const int width) {
  int x = hipBlockDim_x * hipBlockIdx_x + hipThreadIdx_x;
  int y = hipBlockDim_y * hipBlockIdx_y + hipThreadIdx_y;

  out[y * width + x] = in[x * width + y];
}

// CPU implementation of matrix transpose
void matrixTransposeCPUReference(float* output, float* input, const unsigned int width) {
  for (unsigned int j = 0; j < width; j++) {
    for (unsigned int i = 0; i < width; i++) {
      output[i * width + j] = input[j * width + i];
    }
  }
}

int main() {
  float* Matrix;
  float* TransposeMatrix;
  float* cpuTransposeMatrix;

  float* gpuMatrix;
  float* gpuTransposeMatrix;

  hipDeviceProp_t devProp;
  HIP_CALL(hipGetDeviceProperties(&devProp, 0));

  std::cerr << "Device name " << devProp.name << std::endl;

  int i;
  int errors;

  Matrix = (float*)malloc(NUM * sizeof(float));
  TransposeMatrix = (float*)malloc(NUM * sizeof(float));
  cpuTransposeMatrix = (float*)malloc(NUM * sizeof(float));

  // initialize the input data
  for (i = 0; i < NUM; i++) {
    Matrix[i] = (float)i * 10.0f;
  }

  // allocate the memory on the device side
  HIP_CALL(hipMalloc((void**)&gpuMatrix, NUM * sizeof(float)));
  HIP_CALL(hipMalloc((void**)&gpuTransposeMatrix, NUM * sizeof(float)));

  uint32_t iterations = 100;
  while (iterations-- > 0) {
    std::cerr << "## Iteration (" << iterations << ") #################" << std::endl;

    // Memory transfer from host to device
    HIP_CALL(hipMemcpy(gpuMatrix, Matrix, NUM * sizeof(float), hipMemcpyHostToDevice));

    roctxMark("before hipLaunchKernel");
    int rangeId = roctxRangeStart("hipLaunchKernel range");
    roctxRangePush("hipLaunchKernel");
    // Lauching kernel from host
    hipLaunchKernelGGL(
        matrixTranspose, dim3(WIDTH / THREADS_PER_BLOCK_X, WIDTH / THREADS_PER_BLOCK_Y),
        dim3(THREADS_PER_BLOCK_X, THREADS_PER_BLOCK_Y), 0, 0, gpuTransposeMatrix, gpuMatrix, WIDTH);
    roctxMark("after hipLaunchKernel");

    // Memory transfer from device to host
    roctxRangePush("hipMemcpy");

    HIP_CALL(
        hipMemcpy(TransposeMatrix, gpuTransposeMatrix, NUM * sizeof(float), hipMemcpyDeviceToHost));

    roctxRangePop();  // for "hipMemcpy"
    roctxRangePop();  // for "hipLaunchKernel"
    roctxRangeStop(rangeId);

    // CPU MatrixTranspose computation
    matrixTransposeCPUReference(cpuTransposeMatrix, Matrix, WIDTH);

    // verify the results
    errors = 0;
    double eps = 1.0E-6;
    for (i = 0; i < NUM; i++) {
      if (std::abs(TransposeMatrix[i] - cpuTransposeMatrix[i]) > eps) {
        errors++;
      }
    }
    if (errors != 0) {
      fprintf(stderr, "FAILED: %d errors\n", errors);
    } else {
      fprintf(stderr, "PASSED!\n");
    }
  }

  // free the resources on device side
  HIP_CALL(hipFree(gpuMatrix));
  HIP_CALL(hipFree(gpuTransposeMatrix));

  // free the resources on host side
  free(Matrix);
  free(TransposeMatrix);
  free(cpuTransposeMatrix);

  return errors;
}

编译：

只使用hipcc无法直接编译这个源文件

需要指定include 目录和链接库：

cpp 复制代码 
$ hipcc ./MatrixTranspose.cpp  -I /opt/rocm/include/roctracer/ -lroctx64

运行：

./a.out

相关推荐
yyt_cdeyyds3 分钟前
FIFO和LRU算法实现操作系统中主存管理
算法
暮色_年华5 分钟前
Modern Effective C++item 9:优先考虑别名声明而非typedef
c++
重生之我是数学王子13 分钟前
QT基础 编码问题 定时器 事件 绘图事件 keyPressEvent QT5.12.3环境 C++实现
开发语言·c++·qt
alphaTao30 分钟前
LeetCode 每日一题 2024/11/18-2024/11/24
算法·leetcode
我们的五年37 分钟前
【Linux课程学习】：进程程序替换，execl，execv，execlp，execvp，execve，execle，execvpe函数
linux·c++·学习
kitesxian39 分钟前
Leetcode448. 找到所有数组中消失的数字(HOT100)+Leetcode139. 单词拆分(HOT100)
数据结构·算法·leetcode
做人不要太理性1 小时前
【C++】深入哈希表核心：从改造到封装，解锁 unordered_set 与 unordered_map 的终极奥义！
c++·哈希算法·散列表·unordered_map·unordered_set
程序员-King.1 小时前
2、桥接模式
c++·桥接模式
chnming19871 小时前
STL关联式容器之map
开发语言·c++
VertexGeek1 小时前
Rust学习（八）：异常处理和宏编程：
学习·算法·rust
热门推荐
01(欧拉)openEuler系统添加网卡文件配置流程、(欧拉)openEuler系统手动配置ipv6地址流程、(欧拉)openEuler系统网络管理说明02PyTorch机器学习实现液态神经网络03【HarmonyOS】HUAWEI DevEco Studio 下载地址汇总04玄机平台应急响应—webshell查杀05Coze扣子平台完整体验和实践（附国内和国际版对比）06Ubuntu 20.04使用Livox mid 360 测试 FAST_LIO07【目标跟踪】相机运动补偿08Unity中PICO实现 隔空取物 和 接触抓取物体09RAG 实践- Ollama+RagFlow 部署本地知识库10量化方法怎么选？如何评估量化后的大模型LLM？