[Openmp-commits] [openmp] [OpenMP][libomp] Fix dist barrier arrival synchronization (PR #213845)

[Yuta Mukai via Openmp-commits]

ytmukai wrote:

We measured the impact of this patch on Grace (AArch64 Neoverse V2). The benchmark is BARRIER from the EPCC OpenMP MicroBenchmark Suite.

The elapsed time (in microseconds) for each thread count is as follows:

| Algorithm | Patch | 2 | 4 | 8 | 16 | 32 | 64 |
| --- | --- | ---:| ---:| ---:| ---:| ---:| ---:|
| dist | Not applied | 0.55 | 0.54 | 0.77 | 1.21 | 1.54 | 2.64 |
| dist | Applied | 0.67 | 0.69 | 0.79 | 1.36 | 1.84 | 2.72 |
| dist+infinite* | Not applied | 0.48 | 0.66 | 0.55 | 0.73 | 0.76 | 0.85 |
| dist+infinite | Applied | 0.44 | 0.69 | 0.70 | 0.72 | 0.76 | 0.86 |
| default (hyper) | - | 0.60 | 0.82 | 1.54 | 1.57 | 1.95 | 2.26 |

\*: `KMP_BLOCKTIME=infinite`

The patch does not cause any significant performance degradation.
In addition, compared to the default algorithm, `dist+infinite` is the fastest, so we believe it is valuable for this algorithm to work correctly on non-x86 targets as well.


https://github.com/llvm/llvm-project/pull/213845