To work around mempools being unavailable, we need
- Check
CU_DEVICE_ATTRIBUTE_TCC_DRIVER and branch out a separate treatment
- For our purpose, it might also be sufficient to just check
CU_DEVICE_ATTRIBUTE_MEMORY_POOLS_SUPPORTED
- Implement a default allocator for TCC
- Implement a concrete
MemoryResource instance that just wraps cudaMalloc/cudaFree
- Document this divergence from other platforms as a known issue
To work around mempools being unavailable, we need
CU_DEVICE_ATTRIBUTE_TCC_DRIVERand branch out a separate treatmentCU_DEVICE_ATTRIBUTE_MEMORY_POOLS_SUPPORTEDMemoryResourceinstance that just wrapscudaMalloc/cudaFree