[flang-commits] [flang] [OpenMP][libomptarget] Add map checks when running under unified shared memory (PR #69005)
via flang-commits
flang-commits at lists.llvm.org
Fri Oct 13 13:23:52 PDT 2023
================
@@ -289,13 +306,38 @@ TargetPointerResultTy DeviceTy::getTargetPointer(
// In addition to the mapping rules above, the close map modifier forces the
// mapping of the variable to the device.
if (Size) {
- DP("Return HstPtrBegin " DPxMOD " Size=%" PRId64 " for unified shared "
- "memory\n",
- DPxPTR((uintptr_t)HstPtrBegin), Size);
- LR.TPR.Flags.IsPresent = false;
+ LR.TPR.Flags.IsNewEntry = true;
+ assert(TgtPadding == 0 && "TgtPadding must always be zero in USM mode");
+ uintptr_t TgtPtrBegin = (uintptr_t)HstPtrBegin + TgtPadding;
+ LR.TPR.setEntry(
+ HDTTMap
+ ->emplace(new HostDataToTargetTy(
+ (uintptr_t)HstPtrBase, (uintptr_t)HstPtrBegin,
+ (uintptr_t)HstPtrBegin + Size, (uintptr_t)HstPtrBegin,
+ TgtPtrBegin, HasHoldModifier, HstPtrName))
+ .first->HDTT);
+ INFO(OMP_INFOTYPE_MAPPING_CHANGED, DeviceID,
+ "Creating new map entry ONLY with HstPtrBase=" DPxMOD
+ ", HstPtrBegin=" DPxMOD ", TgtAllocBegin=" DPxMOD
+ ", TgtPtrBegin=" DPxMOD
+ ", Size=%ld, DynRefCount=%s, HoldRefCount=%s, Name=%s\n",
+ DPxPTR(HstPtrBase), DPxPTR(HstPtrBegin), DPxPTR(HstPtrBegin),
+ DPxPTR(TgtPtrBegin), Size,
+ LR.TPR.getEntry()->dynRefCountToStr().c_str(),
+ LR.TPR.getEntry()->holdRefCountToStr().c_str(),
+ (HstPtrName) ? getNameFromMapping(HstPtrName).c_str() : "unknown");
LR.TPR.Flags.IsHostPointer = true;
+
+ // The following assert should catch any case in which the pointers
+ // do not match to understand if this case can ever happen.
+ assert((uintptr_t)HstPtrBegin == TgtPtrBegin &&
+ "Pointers must always match");
+
+ // If the above assert is ever hit the following should be changed to =
+ // TgtPtrBegin
LR.TPR.TargetPointer = HstPtrBegin;
}
+ LR.TPR.Flags.IsPresent = false;
----------------
carlobertolli wrote:
How do we implement omp_target_is_present()?
If we implement it by checking if the input pointer is contained within previously mapped memory, then we might want to change this to "true".
Here's a reminder of the definition of the API above:
> The omp_target_is_present routine tests whether a host pointer refers to storage that is
> mapped to a given device.
Under unified_shared_memory, mapping memory should mean the same as in non unified_shared_memory: not that memory is being allocated and copied on some physical storage device, but that the device data environment contains that memory, which in this case it does (and it is accessed via the host pointer).
I believe omp_target_is_present implementation has been discussed in the context of OpenMP acceleration subcommittee, so we may want to check on what TR11 and any "in fieri" document requires about this.
https://github.com/llvm/llvm-project/pull/69005
More information about the flang-commits
mailing list