17/09/2025 12:06, Thomas Monjalon: > 15/09/2025 16:41, Gregory Etelson: > > From: Bing Zhao <[email protected]> > > > > The API name and prototype have been changed in the new CUDA, this > > commit will ensure the compatibility with different versions. > [...] > > /* CUDA Driver functions loaded with cuGetProcAddress for versioning */ > > +#if defined(CUDA_VERSION) && (CUDA_VERSION >= 13000) > > +static PFN_cuGetErrorString_v6000 pfn_cuGetErrorString; > > +static PFN_cuGetErrorName_v6000 pfn_cuGetErrorName; > > +static PFN_cuPointerSetAttribute_v6000 pfn_cuPointerSetAttribute; > > +static PFN_cuDeviceGetAttribute_v2000 pfn_cuDeviceGetAttribute; > > +static PFN_cuDeviceGetByPCIBusId_v4010 pfn_cuDeviceGetByPCIBusId; > > +static PFN_cuDevicePrimaryCtxRetain_v7000 pfn_cuDevicePrimaryCtxRetain; > > +static PFN_cuDevicePrimaryCtxRelease_v11000 pfn_cuDevicePrimaryCtxRelease; > > +static PFN_cuDeviceTotalMem_v3020 pfn_cuDeviceTotalMem; > > +static PFN_cuDeviceGetName_v2000 pfn_cuDeviceGetName; > > +static PFN_cuCtxGetApiVersion_v3020 pfn_cuCtxGetApiVersion; > > +static PFN_cuCtxSetCurrent_v4000 pfn_cuCtxSetCurrent; > > +static PFN_cuCtxGetCurrent_v4000 pfn_cuCtxGetCurrent; > > +static PFN_cuCtxGetDevice_v2000 pfn_cuCtxGetDevice; > > +static PFN_cuCtxGetExecAffinity_v11040 pfn_cuCtxGetExecAffinity; > > +static PFN_cuMemAlloc_v3020 pfn_cuMemAlloc; > > +static PFN_cuMemFree_v3020 pfn_cuMemFree; > > +static PFN_cuMemHostRegister_v6050 pfn_cuMemHostRegister; > > +static PFN_cuMemHostUnregister_v4000 pfn_cuMemHostUnregister; > > +static PFN_cuMemHostGetDevicePointer_v3020 pfn_cuMemHostGetDevicePointer; > > +static PFN_cuFlushGPUDirectRDMAWrites_v11030 > > pfn_cuFlushGPUDirectRDMAWrites; > > +#else > > static PFN_cuGetErrorString pfn_cuGetErrorString; > > static PFN_cuGetErrorName pfn_cuGetErrorName; > > What is the logic here? > The old version keeps its original name, and new versions have a suffix? > Why not use the old version? > What will happen with next versions?
Looks like we cannot do better than this versioning when using cuGetProcAddress to avoid linking CUDA. Applied
