Add -fopenmp-assume-no-thread-state to the amdflang offload flags (~1.2x on AFAR 24.3)

[sbryngelson]

Summary

Adds -fopenmp-assume-no-thread-state to the amdflang (LLVMFlang) OpenMP-offload compile flags. This is the one part of -fopenmp-target-fast that MFC did not already pass: on AFAR 24.3, -fopenmp-target-fast expands to -O3 -fopenmp-assume-no-nested-parallelism -fopenmp-assume-no-thread-state, and the first two are already in the flags. #1450 dropped -fopenmp-target-fast because it was in the reproducer for #1449.

The flag promises that no kernel changes OpenMP ICVs (the runtime's internal settings, such as thread counts) inside a target region, which lets the compiler drop the per-kernel thread-state bookkeeping. MFC's only runtime-setting call is omp_set_default_device, made on the host before any kernel launches. The GPU macros emit no num_threads, omp_set_*, or nested parallelism.

Results

HPCFund MI210 (gfx90a), AFAR 24.3.0, release build (-O3), no case optimization, ./mfc.sh bench --mem 2, same node as the master baseline. Grind time (lower is better):

case	master	this PR	speedup
5eq_rk3_weno3_hll	2.702	2.259	1.20x
5eq_rk3_weno3_hllc	2.495	2.094	1.19x
5eq_rk3_weno3_lf	2.448	2.078	1.18x
hypo_hll	1.885	1.402	1.35x
ibm	5.449	4.466	1.22x
igr	2.618	2.217	1.18x
viscous_weno5_sgb_acoustic	4.024	3.442	1.17x

Geometric mean: 1.21x. Same-node run-to-run noise in these measurements is about 1-2%. The Frontier (AMD) Bench job on this PR gives an independent MI250X number.

Testing

207 tests on the same build: all viscous, hypoelastic, WENO7, IBM (where #1449's corruption appeared), and chemistry tests, plus samples of HLL, HLLC, and LF. 207/207 passed.

Only amdflang GPU builds are affected; other compilers do not see this flag.

This PR was prepared with Claude Code (AI-assisted).

Acknowledgement

• I confirm this PR meets the above expectations and reflects my own understanding and real-world context.

[Copilot]

Copilot review overview

🟢 Approval recommended

The isolated compiler-flag change matches MFC’s device runtime usage and has targeted correctness and performance validation.

Review effort: Balanced
Findings: None

What changed in this PR

Adds a verified LLVMFlang OpenMP-offload optimization that removes unnecessary device thread-state bookkeeping.

Changes:

• Adds -fopenmp-assume-no-thread-state for LLVMFlang GPU builds.
• Documents the assumption and measured performance benefit.

File	Description
cmake/​MFCTargets.cmake	Adds the optimization flag to LLVMFlang offload compilation.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 62.80%. Comparing base (c7feed2) to head (198c5c0).

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1940   +/-   ##
=======================================
  Coverage   62.80%   62.80%           
=======================================
  Files          86       86           
  Lines       22385    22385           
  Branches     3304     3304           
=======================================
  Hits        14060    14060           
  Misses       6073     6073           
  Partials     2252     2252           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.