roctracer 的应用示例

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
爱写Bug的浮游1 小时前
C++移动构造真的比拷贝快吗
c++·性能优化
Aaa西红柿鸡蛋2 小时前
网络协议分析实验:基于 UDP 广播模型的 C/S 通信设计与实现
c++
青少儿编程课堂2 小时前
蓝桥杯·全国青少年信息素养大赛·电子学会Python/C++考级:扫描线与线段树求矩形面积并解析
c++·python·算法·bfs·信息学竞赛
电科_银尘2 小时前
【大模型算法与工程 09】SGD、Momentum与AdamW:优化器如何选择
人工智能·算法
货拉拉技术4 小时前
视觉底座:货拉拉OCR的进阶之路
算法
__underline__4 小时前
冷门Python知识点005篇:list.sort( )的排序稳定性,是用0字节额外内存实现的
python·算法
风落青芜6 小时前
数据结构:堆的概念与实现
c语言·开发语言·数据结构·算法
Pniubi11 小时前
01背包类问题-力扣416.分割等和子集
算法·leetcode·职场和发展
eggwyw21 小时前
C++ union 的一种妙用法
开发语言·c++