Prioritized list of Regent features for HTR (PSAAP)

[elliottslaughter]

This is mostly a bookkeeping issue, to help us track the current set of prioritized Legion/Regent/Realm issues for HTR. My understanding is that these issues are currently (in order of priority):

Regent

• Separate Compilation for Regent #637
• Regent: CUDA compiler breaks when passing large structures as an argument #958
• Regent: Support function-call calling convention for Regent tasks #1073
• NVIDIA CUDA Ampere architecture support for Regent #1184
• Regent: Errors during executable linking on IBM Power 9 machine with LLVM 90 #1038
• Regent: Link error on PowerPC when using LLVM 9 #939
• Regent: Performance regression in CUDA codegen with LLVM 6 vs 3.8 #1041
• Regent: LLVM error when including .h files and with -fcuda 1 #873
• Regent: HIP support #1224
• building HTR with llvm > 6.0 #1264
• Regent: bug in returning structs on PowerPC #1276
• Calling Pygion tasks from Regent #1188
• regent_cuda.cu build failure on PPC #1511
• Regent: support compact sparse instances  #878
• Regent: support for local regions (DeferredBuffer) #995
• Regent: Add privileges for regions contained in an fspace #789
• Regent: improve constant folding of Terra macros #694
• Regent: missing debug symbols #1190
• Regent: potential bug in inliner #1189
• Regent: lifting variables to Futures across __local tasks boundaries #1354

Legion

• Realm/Legion: Interleaved Compact Instance Layouts #1057
• Legion: non informative error in case of impossible compact sparse instance covering #1388
• Deterministic error affecting HTR #1466
• Non-deterministic hangs while running HTR #1167
• Assertion sit->second == local_space failed #1473
• Assertion `!subscription_event.exists()' failed. #1484
• Legion: regression in padded constraint conflicts #1537
• non-deterministic test failures #1538
• Legion: mixing user-defined and Legion-generated layout constraint IDs #1570
• Legion: AllGatherCollective<false>::~AllGatherCollective() [INORDER = false]: Assertion `done_triggered' failed. #1590
• Deadlock at shutdown when collecting Legion_prof logs #1586
• HTR crash on multiple nodes #1573
• Non-deterministic segmentation fault #1632
• Hang at start up on multinode runs #1643
• [HTR] Segmentation faults at 16 nodes #1420
• HTR correctness failure at 64 nodes #1566
• remove_nested_valid_ref assertion failure #1672
• [HTR] segfault at 4 nodes #1677
• Legion: assertion mapping_tracker != __null #1441
• finder != operations.end() assertion failure #1686
• Legion: handling multiple layout constraint sets for one region requirement #1629
• Performance degradation when running HTR #1652
• Legion: warning ‘it.Legion::Domain::DomainPointIterator::is_valid’ may be used uninitialized #1714
• [HTR] Control replication violation #1689
• Legion: default mapper disregards field constraints for region requirements with reduction privileges #1703
• Legion: non deterministic assertion #1790
• Legion: assertion when destroying a deferred buffer #1816

Realm

• OPENMP dynamic scheduling #1154
• Realm: packet corruption with gasnet-ex layer #1141
• Realm: Bad data in HDF5 file #1450
• Realm: slow data transposes on GPUs #1494
• Realm: segmentation fault at startup #1523
• Regression in GPU executions #1550
• Realm: compile time error in shm.cc #1589
• Realm: Too many simultaneous local client threads #1612
• Reservation Acquire Poisoning #1610
• Realm: Failed to send message #1597
• [HTR] device-side assert triggered #1687
• Realm: Random failure in shutdown_check #1424
• Realm: HIP implementation of multi-affine accessor #1646
• Realm::UnfairMutex::unlock_slow(): Assertion #1535
• Realm: OOM error when running on recent versions of the runtime #1513
• Realm: possible off-by-one error in compute_covering of sparse compact instances #1387
• Deadlock when running with AMD GPUs #1688
• [BUG] HTR failure on lassen  #1755
• Realm: missed detection of HIP hijack #1557
• Realm: support for reductions to HDF5 instances #521
• Assertion failure (KD tree) or hang #1262
• Realm: MultiAffineAccessors are not thread safe #1364
• Realm: OpenMP compatible RectInDomainIterator #1164
• Legion/Realm: interface for CUDA constants #964
• Bounded Unfairness in Realm Locks #1292
• Realm: Option to activate all the loggers that match a pattern #1386
• Realm: slow data transposes on CPUs #1493
• Realm: Duplicate registration of function _ZN3cub11EmptyKernelIvEEvv #1583
• Random "poisoned deferred unlock skipped" #1631
• Realm: error when compiling with HIP with NVIDIA GPUs #1713
• Realm: Slow data transposes on GPU using HIP #1789

Build

• Error while linking on Tioga (AMD) with GasNET #1555
• HTR build failure in CI #1756

Tools

• Legion_prof: assertion failed: copy_inst_info._src.unwrap() == chan_src #1606
• Legion built without zlib generates uncompressed logs without warning #1738

[mariodirenzo]

I think that HTR would also benefit from the solution of #873 and #694.
They are not critical so they should have the lowest priority.

[streichler]

Pinging here too: #1154 has a proposed fix that I need somebody to test out

[mariodirenzo]

@elliottslaughter, could you please insert #1276 in the Regent section with high priority?

[mariodirenzo]

@elliottslaughter, could you add #1292 to the Realm section with low priority?

[mariodirenzo]

@elliottslaughter, could you add #1386, #1387, and #1388 to the Realm (first two) and Legion (third issue) sections?
All these issues have low priority.
Thanks

[mariodirenzo]

@elliottslaughter, could you add #1364 to the Realm issues with priority just below #521?

[mariodirenzo]

@elliottslaughter, could you please add #1420 to the legion section?

[mariodirenzo]

@elliottslaughter, sorry for asking to add issues so frequently. Could you add #1262 to the Realm section above 1364?

[mariodirenzo]

Hi @elliottslaughter, could please you add #1473 (high priority), #1441 (low priority) to the Legion section and #1424 (low priority) to the Realm section?