[SYCL] Support UR_DEVICE_INFO_USM_CONTEXT_MEMCPY_SUPPORT_EXP on Cuda #19267
Conversation
|
It's tested by an existing |
jchlanda
force-pushed
the
jakub/device_globals_gpu
branch
from
5ded23f to
1561006
Compare
| void *pDst, | ||
| const void *pSrc, | ||
| size_t Size) { | ||
| UR_CHECK_ERROR(cuMemcpyHtoD((CUdeviceptr)pDst, pSrc, Size)); |
There was a problem hiding this comment.
Just curious - where in the API does it say that pSrc is a host pointer and pDst is a device pointer? I can't find much information about the extension.
There was a problem hiding this comment.
Seems you're right, the docs for the function declaration don't mention it explicitly. Are you worried that this call could potentially handle situation where both src and dest are device pointers?
There was a problem hiding this comment.
Yeah, or host->host or device->host. It's not clear to me what this API is meant to handle.
There was a problem hiding this comment.
I've discussed it offline, you are right, the extension does not give any guarantees about the location of the pointers. I'll put it back to draft to address this issue.
There was a problem hiding this comment.
We should be bale to trace the pointers with cuPointerGetAttribute and CU_POINTER_ATTRIBUTE_MEMORY_TYPE.
There was a problem hiding this comment.
I've moved to cuMemcpy that automatically handles the provenience of the src/dst pointers.
jchlanda
force-pushed
the
jakub/device_globals_gpu
branch
from
b07d3fc to
e84cf2b
Compare
jchlanda
force-pushed
the
jakub/device_globals_gpu
branch
from
e84cf2b to
8d70bec
Compare
jchlanda
force-pushed
the
jakub/device_globals_gpu
branch
from
8d70bec to
40707c8
Compare
jchlanda
force-pushed
the
jakub/device_globals_gpu
branch
from
40707c8 to
cba6baa
Compare
jchlanda
force-pushed
the
jakub/device_globals_gpu
branch
from
cba6baa to
5eee551
Compare
|
This should be ready to roll @intel/llvm-gatekeepers. Thank you. |
This is an optimisation that uses a direct memcopy from host to device.