Insert scratch memory into scratch cache on full profile systems.

Scratch cache was not updated for IOMMUv2 systems previously.
This both negates the cache and causes segfault during scratch
release.

Change-Id: I71e81d6b642d65ca135868ff7225ea173529d458


[ROCm/ROCR-Runtime commit: 191664cd20]
This commit is contained in:
Sean Keely
2021-03-02 21:38:56 -06:00
förälder 519b7d4642
incheckning e2aca270b7
@@ -1136,15 +1136,13 @@ void GpuAgent::AcquireQueueScratch(ScratchInfo& scratch) {
assert(((!scratch.large) | use_reclaim) && "Large scratch used with reclaim disabled.");
if (scratch.queue_base != nullptr) {
if (profile_ == HSA_PROFILE_FULL) return;
if (profile_ == HSA_PROFILE_BASE) {
HSAuint64 alternate_va;
if (hsaKmtMapMemoryToGPU(scratch.queue_base, scratch.size, &alternate_va) ==
HSAKMT_STATUS_SUCCESS) {
if (scratch.large) scratch_used_large_ += scratch.size;
scratch_cache_.insert(scratch);
return;
}
HSAuint64 alternate_va;
if ((profile_ == HSA_PROFILE_FULL) ||
(hsaKmtMapMemoryToGPU(scratch.queue_base, scratch.size, &alternate_va) ==
HSAKMT_STATUS_SUCCESS)) {
if (scratch.large) scratch_used_large_ += scratch.size;
scratch_cache_.insert(scratch);
return;
}
}