https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128590
--- Comment #2 from Jerry DeLisle <jvdelisle at gcc dot gnu.org> --- Created attachment 66131 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=66131&action=edit Benchmark used to test Some preliminary results: Benchmark (RX 6600 XT, offload mandatory, 3 interleaved runs, spread under 5%): ┌─────────────────────────────────┬───────────────┬───────────────┐ │ │ stencil │ compute-bound │ ├─────────────────────────────────┼───────────────┼───────────────┤ │ serial │ 1.24 s │ 2.38 s │ ├─────────────────────────────────┼───────────────┼───────────────┤ │ host │ 1.33 s │ 0.158 s (15×) │ ├─────────────────────────────────┼───────────────┼───────────────┤ │ device │ 4.52 s │ 0.083 s (29×) │ ├─────────────────────────────────┼───────────────┼───────────────┤ │ device inside !$omp target data │ 0.48 s (2.6×) │ 0.035 s (68×) │ └─────────────────────────────────┴───────────────┴───────────────┘ The test case from PR83064 gives correct results.Initially I got a segfault because I was short on stack memory. I am still tweaking and testing this and I only have available my older GPU here.
