roctracer 的应用示例

Eloudy2024-09-15 9:11

1，不用 roctracer 的普通场景

mt.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>

#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
这儿有一堆花1 小时前
比特币：固若金汤的数字堡垒与它的四道防线
算法·区块链·哈希算法
客卿1231 小时前
力扣100-移动0
算法·leetcode·职场和发展
CM莫问4 小时前
＜论文＞（微软）WINA：用于加速大语言模型推理的权重感知神经元激活
人工智能·算法·语言模型·自然语言处理·大模型·推理加速
计信金边罗6 小时前
是否存在路径（FIFOBB算法）
算法·蓝桥杯·图论
MZWeiei6 小时前
KMP 算法中 next 数组的构建函数 get_next
算法·kmp
Fanxt_Ja8 小时前
【JVM】三色标记法原理
java·开发语言·jvm·算法
luofeiju8 小时前
行列式的性质
线性代数·算法·矩阵
緈福的街口8 小时前
【leetcode】347. 前k个高频元素
算法·leetcode·职场和发展
南郁8 小时前
007-nlohmann/json 项目应用-C++开源库108杰
c++·开源·json·nlohmann·现代c++·d2school·108杰
pen-ai9 小时前
【统计方法】基础分类器: logistic, knn, svm, lda
算法·机器学习·支持向量机
热门推荐
01基于STM32的智能电池管理系统02KGG转MP3工具|非KGM文件|解密音频03YOLOv8入门 | 重要性能衡量指标、训练结果评价及分析及影响mAP的因素【发论文关注的指标】04海康Visionmaster-常见问题排查方法-启动阶段05从零安装 LLaMA-Factory 微调 Qwen 大模型成功及所有的坑06【SpeedAI科研小助手】2分钟极速解决知网维普重复率、AIGC率过高，一键全文降！文件格式不变，公式都保留的！07VMware虚拟机安装Win7专业版保姆级教程（附镜像包）08DeepSeek各版本说明与优缺点分析09组基轨迹建模 GBTM的介绍与实现（Stata 或 R）10柬埔寨语翻译通App，一款真正实现高棉语翻译、语音识别和中柬双语无障碍交流的应用程序。