HIP

安装

sudo pacman -S hip-runtime-amd
pushd /opt/rocm/hip/bin
wget https://raw.githubusercontent.com/ROCm-Developer-Tools/HIPIFY/amd-staging/bin/hipify-perl; sudo chmod +x hipify-perl
popd

echo "export HIP_PATH="/opt/rocm/hip\nexport PATH="$PATH:/opt/rocm/bin:$HIP_PATH/bin" > $HOME/.zshrc
source $HOME/.zshrc

使用

将CUDA代码转换成HIP代码

hipify-perl test.cu > test.cpp

编译HIP代码

hipcc test.cpp -o test

免费的线上cuda环境

  1. 登录colab，需要有google账号

  2. 更改runtime选项

Runtime > change runtime type > GPU

  1. 添加测试代码块

%%writefile test.cu会将该代码块后面的代码写入到test.cu文件中

%%writefile test.cu
#include <cstdio>
#include <cuda_runtime.h>

__device__ void Print() {
  printf("this is a GPU\n");
}

void kernel_host() {
  printf("this is a CPU\n");
}
__global__ void kernel() {
  Print();
}

int main() {
  kernel<<<1,1>>>();
  cudaDeviceSynchronize();
  kernel_host();
}
  1. 编译

!会运行shell命令，使用nvcc编译步骤3中生成的test.cu文件

!nvcc test.cu -o test
  1. 运行
!./test

output:

this is a GPU
this is a CPU