目前cuda最新的版本已经来到了13.0，今天在cuda13.0的环境下编译gpu-burn报错，查了下资料，cuCtxCreate函数升级为了最新的v4，需要传入4个参数，但是gpu-burn还没更新，所以需要修正下这个问题，我们先来看下编译报错的错误信息

g++ -O3 -Wno-unused-result -I/usr/local/cuda/include -std=c++11 -DIS_JETSON=false -c gpu_burn-drv.cpp
gpu_burn-drv.cpp: In constructor ‘GPU_Test<T>::GPU_Test(int, bool, bool, const char*)’:
gpu_burn-drv.cpp:113:31: error: too few arguments to function ‘CUresult cuCtxCreate_v4(CUctx_st**, CUctxCreateParams*, unsigned int, CUdevice)’
  113 |         checkError(cuCtxCreate(&d_ctx, 0, d_dev));
gpu_burn-drv.cpp:98:17: note: in definition of macro ‘checkError’
   98 |     _checkError(rCode, __FILE__, __LINE__, ##__VA_ARGS__)
      |                 ^~~~~
In file included from gpu_burn-drv.cpp:65:
/usr/local/cuda/include/cuda.h:90:45: note: declared here
   90 | #define cuCtxCreate                         cuCtxCreate_v4
      |                                             ^~~~~~~~~~~~~~
/usr/local/cuda/include/cuda.h:6275:18: note: in expansion of macro ‘cuCtxCreate’
 6275 | CUresult CUDAAPI cuCtxCreate(CUcontext *pctx, CUctxCreateParams *ctxCreateParams, unsigned int flags, CUdevice dev);
      |                  ^~~~~~~~~~~
make: *** [Makefile:45: gpu_burn-drv.o] Error 1

然后我们需要修改gpu_burn-drv.cpp文件，大概在113行

//checkError(cuCtxCreate(&d_ctx, 0, d_dev)); 注释此行
checkError(cuDevicePrimaryCtxRetain(&d_ctx, d_dev)); #添加此行

最后如下图所示

然后再进行编译，发现还有报错

g++ -O3 -Wno-unused-result -I/usr/local/cuda/include -std=c++11 -DIS_JETSON=false -c gpu_burn-drv.cpp
PATH="/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/snap/bin::." /usr/local/cuda/bin/nvcc -I/usr/local/cuda/include -arch=compute_50 -ptx compare.cu -o compare.ptx
nvcc fatal   : Unsupported gpu architecture 'compute_50'
make: *** [Makefile:48: compare.ptx] Error 1

因为我使用的5090，报错的arch=compute_50明显和我的算力不一致，需要传入计算能力的参数再进行编译

make COMPUTE=120

然后再进行编译就可以顺利编译了。