OpenMP_VV icon indicating copy to clipboard operation
OpenMP_VV copied to clipboard

Avoid stack allocation for test_requires_unified_shared_memory_heap*

Open tob2 opened this issue 2 years ago • 6 comments

Some (pseuo)USM implementations allocate memory in a special memory section that permits either direct memory access (DMA) from the device or on-access migration. In that case, stack and static memory may or may not be accessible.

This pull request changes the test_requires_unified_shared_memory_heap{,_map,_is_device_ptr}.{c,F90} testcases such that only heap memory is used on the device. The existing test_unified_shared_memory*.* testcases handle the other variants.

@nolanbaker31 @spophale @jrreap @krishols @mjcarr458 – please review

This is a consistency fix – permitting a partial pass with pseudo USM. For full USM, kernel + hardware support is required to access all memory. I some (but not all) post-Pascal Nvidia cards permit to access all memory and, if I understood it correctly (big “if”), AMD cards since VEGA support accessing all memory in principle, but effectively, only gfx90a (MI 200 series) permits fast access and has proper.

tob2 avatar Oct 24 '22 11:10 tob2