1,源码下载
GitHub - ROCm/HIPIFY: HIPIFY: Convert CUDA to Portable C++ CodeHIPIFY: Convert CUDA to Portable C++ Code. Contribute to ROCm/HIPIFY development by creating an account on GitHub.https://github.com/ROCm/HIPIFY.git
git clone --recursive https://github.com/ROCm/HIPIFY.git
sudo apt install clang-dev
2,编译并安装
2.1 通常方式
hipify-clang 文档:
https://github.com/ROCm/HIPIFY/blob/amd-staging/docs/hipify-clang.md
编译命令:
cmake -DCMAKE_INSTALL_PREFIX=../dist -DCMAKE_BUILD_TYPE=Release ..
make -j install
此时 hipify-clang 会被安装到 HIPIFY/dist/bin 中,
测试:
cd ../dist/bin
hipify --help
如果系统中存在多个llvm版本,在执行翻译命令时,比如hipify-clang ./vectorAdd.cu --cuda-path=/usr/local/cuda-12.1可能会发生错误,如下提示:
CommandLine Error: Option 'static-func-full-module-prefix' registered more than once!
LLVM ERROR: inconsistency in registered CommandLine option
这时需要使用自制的LLVM,如下2.2节所示。
2.2 自制LLVM的方式
2.2.1 下载llvm源码
wget https://github.com/llvm/llvm-project/archive/refs/tags/llvmorg-17.0.6.tar.gz
解压,tar zxf llvmorg.....
2.2.2 配置编译LLVM
cd llvmorg.....mkdir -p build ../dist/localcd buildcmake -G "Unix Makefiles" ../llvm \
-DLLVM_ENABLE_PROJECTS="clang;clang-tools-extra;compiler-rt" \
-DLLVM_BUILD_EXAMPLES=ON -DLLVM_TARGETS_TO_BUILD="host" \
-DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=ON \
-DLLVM_ENABLE_RUNTIMES=all -DLLVM_BUILD_LLVM_DYLIB=ON \
-DCMAKE_INSTALL_PREFIX=../../dist/local
make -j
make -j install
测试时,llvm 被install在如下文件夹:
/home/hipper/ex_dock_hipify/dist/local
ls /home/hipper/ex_dock_hipify/dist/local 如图:
2.2.3 配置编译HIPIFY
指定 LLVM 安装目录的配置方法:
-DCMAKE_PREFIX_PATH=/home/hipper/ex_dock_hipify/dist/local
cmake \-DCMAKE_BUILD_TYPE=Release \-DCMAKE_INSTALL_PREFIX=../dist \-DCMAKE_PREFIX_PATH=/home/hipper/ex_dock_hipify/dist/local \-DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-12.1 ..
make -j install
3. 示例
3.1翻译 .cu 文件到 .hip 文件
命令:
/home/hipper/ex_dock_hipify/HIPIFY/dist/bin/hipify-clang ./vectorAdd.cu --cuda-path=/usr/local/cuda-12.1
会在 ./ 目录中生成 vectoreAdd.cu.hip 的文件。
其中,hipify-clang 并不检查输入文件的扩展名,比如这里的.cu,它只检查文件内部的内容,将cuda生态的关键字有机地翻译成 hip生态的关键字,输出文件会在原文件名的基础上加上 .hip 后缀;
源代码分别如下。
使用 cuda samples中的vectoradd.cu为例,源码如下:
vectorAdd.cu
#include <stdio.h>
#include <cuda_runtime.h>__global__ void vectorAdd(const float *A, const float *B, float *C,int numElements) {int i = blockDim.x * blockIdx.x + threadIdx.x;if (i < numElements) {C[i] = A[i] + B[i] + 0.0f;}
}int main(void) {cudaError_t err = cudaSuccess;int numElements = 50000;size_t size = numElements * sizeof(float);printf("[Vector addition of %d elements]\n", numElements);float *h_A = (float *)malloc(size);float *h_B = (float *)malloc(size);float *h_C = (float *)malloc(size);if (h_A == NULL || h_B == NULL || h_C == NULL) {fprintf(stderr, "Failed to allocate host vectors!\n");exit(EXIT_FAILURE);}for (int i = 0; i < numElements; ++i) {h_A[i] = rand() / (float)RAND_MAX;h_B[i] = rand() / (float)RAND_MAX;}float *d_A = NULL;err = cudaMalloc((void **)&d_A, size);if (err != cudaSuccess) {fprintf(stderr, "Failed to allocate device vector A (error code %s)!\n",cudaGetErrorString(err));exit(EXIT_FAILURE);}float *d_B = NULL;err = cudaMalloc((void **)&d_B, size);if (err != cudaSuccess) {fprintf(stderr, "Failed to allocate device vector B (error code %s)!\n",cudaGetErrorString(err));exit(EXIT_FAILURE);}float *d_C = NULL;err = cudaMalloc((void **)&d_C, size);if (err != cudaSuccess) {fprintf(stderr, "Failed to allocate device vector C (error code %s)!\n",cudaGetErrorString(err));exit(EXIT_FAILURE);}printf("Copy input data from the host memory to the CUDA device\n");err = cudaMemcpy(d_A, h_A, size, cudaMemcpyHostToDevice);if (err != cudaSuccess) {fprintf(stderr,"Failed to copy vector A from host to device (error code %s)!\n",cudaGetErrorString(err));exit(EXIT_FAILURE);}err = cudaMemcpy(d_B, h_B, size, cudaMemcpyHostToDevice);if (err != cudaSuccess) {fprintf(stderr,"Failed to copy vector B from host to device (error code %s)!\n",cudaGetErrorString(err));exit(EXIT_FAILURE);}int threadsPerBlock = 256;int blocksPerGrid = (numElements + threadsPerBlock - 1) / threadsPerBlock;printf("CUDA kernel launch with %d blocks of %d threads\n", blocksPerGrid,threadsPerBlock);vectorAdd<<<blocksPerGrid, threadsPerBlock>>>(d_A, d_B, d_C, numElements);err = cudaGetLastError();if (err != cudaSuccess) {fprintf(stderr, "Failed to launch vectorAdd kernel (error code %s)!\n",cudaGetErrorString(err));exit(EXIT_FAILURE);}printf("Copy output data from the CUDA device to the host memory\n");err = cudaMemcpy(h_C, d_C, size, cudaMemcpyDeviceToHost);if (err != cudaSuccess) {fprintf(stderr,"Failed to copy vector C from device to host (error code %s)!\n",cudaGetErrorString(err));exit(EXIT_FAILURE);}for (int i = 0; i < numElements; ++i) {if (fabs(h_A[i] + h_B[i] - h_C[i]) > 1e-5) {fprintf(stderr, "Result verification failed at element %d!\n", i);exit(EXIT_FAILURE);}}printf("Test PASSED\n");err = cudaFree(d_A);if (err != cudaSuccess) {fprintf(stderr, "Failed to free device vector A (error code %s)!\n",cudaGetErrorString(err));exit(EXIT_FAILURE);}err = cudaFree(d_B);if (err != cudaSuccess) {fprintf(stderr, "Failed to free device vector B (error code %s)!\n",cudaGetErrorString(err));exit(EXIT_FAILURE);}err = cudaFree(d_C);if (err != cudaSuccess) {fprintf(stderr, "Failed to free device vector C (error code %s)!\n",cudaGetErrorString(err));exit(EXIT_FAILURE);}free(h_A);free(h_B);free(h_C);printf("Done\n");return 0;
}
生成的 vectorAdd.cu.hip :
#include <stdio.h>
#include <hip/hip_runtime.h>__global__ void vectorAdd(const float *A, const float *B, float *C,int numElements) {int i = blockDim.x * blockIdx.x + threadIdx.x;if (i < numElements) {C[i] = A[i] + B[i] + 0.0f;}
}int main(void) {hipError_t err = hipSuccess;int numElements = 50000;size_t size = numElements * sizeof(float);printf("[Vector addition of %d elements]\n", numElements);float *h_A = (float *)malloc(size);float *h_B = (float *)malloc(size);float *h_C = (float *)malloc(size);if (h_A == NULL || h_B == NULL || h_C == NULL) {fprintf(stderr, "Failed to allocate host vectors!\n");exit(EXIT_FAILURE);}for (int i = 0; i < numElements; ++i) {h_A[i] = rand() / (float)RAND_MAX;h_B[i] = rand() / (float)RAND_MAX;}float *d_A = NULL;err = hipMalloc((void **)&d_A, size);if (err != hipSuccess) {fprintf(stderr, "Failed to allocate device vector A (error code %s)!\n",hipGetErrorString(err));exit(EXIT_FAILURE);}float *d_B = NULL;err = hipMalloc((void **)&d_B, size);if (err != hipSuccess) {fprintf(stderr, "Failed to allocate device vector B (error code %s)!\n",hipGetErrorString(err));exit(EXIT_FAILURE);}float *d_C = NULL;err = hipMalloc((void **)&d_C, size);if (err != hipSuccess) {fprintf(stderr, "Failed to allocate device vector C (error code %s)!\n",hipGetErrorString(err));exit(EXIT_FAILURE);}printf("Copy input data from the host memory to the CUDA device\n");err = hipMemcpy(d_A, h_A, size, hipMemcpyHostToDevice);if (err != hipSuccess) {fprintf(stderr,"Failed to copy vector A from host to device (error code %s)!\n",hipGetErrorString(err));exit(EXIT_FAILURE);}err = hipMemcpy(d_B, h_B, size, hipMemcpyHostToDevice);if (err != hipSuccess) {fprintf(stderr,"Failed to copy vector B from host to device (error code %s)!\n",hipGetErrorString(err));exit(EXIT_FAILURE);}int threadsPerBlock = 256;int blocksPerGrid = (numElements + threadsPerBlock - 1) / threadsPerBlock;printf("CUDA kernel launch with %d blocks of %d threads\n", blocksPerGrid,threadsPerBlock);vectorAdd<<<blocksPerGrid, threadsPerBlock>>>(d_A, d_B, d_C, numElements);err = hipGetLastError();if (err != hipSuccess) {fprintf(stderr, "Failed to launch vectorAdd kernel (error code %s)!\n",hipGetErrorString(err));exit(EXIT_FAILURE);}printf("Copy output data from the CUDA device to the host memory\n");err = hipMemcpy(h_C, d_C, size, hipMemcpyDeviceToHost);if (err != hipSuccess) {fprintf(stderr,"Failed to copy vector C from device to host (error code %s)!\n",hipGetErrorString(err));exit(EXIT_FAILURE);}for (int i = 0; i < numElements; ++i) {if (fabs(h_A[i] + h_B[i] - h_C[i]) > 1e-5) {fprintf(stderr, "Result verification failed at element %d!\n", i);exit(EXIT_FAILURE);}}printf("Test PASSED\n");err = hipFree(d_A);if (err != hipSuccess) {fprintf(stderr, "Failed to free device vector A (error code %s)!\n",hipGetErrorString(err));exit(EXIT_FAILURE);}err = hipFree(d_B);if (err != hipSuccess) {fprintf(stderr, "Failed to free device vector B (error code %s)!\n",hipGetErrorString(err));exit(EXIT_FAILURE);}err = hipFree(d_C);if (err != hipSuccess) {fprintf(stderr, "Failed to free device vector C (error code %s)!\n",hipGetErrorString(err));exit(EXIT_FAILURE);}free(h_A);free(h_B);free(h_C);printf("Done\n");return 0;
}
3.2 编译运行 vectorAdd.cu.hip
编译:
$ /opt/rocm/bin/hipcc ./vectorAdd.cu.hip -o vectorAdd
运行效果如下图: