[OpenMP] offload hierachical parallelism gives wrong results. #156805
Description
Caused by
#146404
Test OvO. https://github.com/TApplencourt/OvO
run
CXX="/soft/compilers/llvm/main-patched/bin/clang++" CXXFLAGS="-fopenmp --offload-arch=sm_90 -O3" ./ovo.sh run test_src/cpp/hierarchical_parallelism
results of 87db8e9
./ovo.sh report --summary --tablefmt github
>> Overall result for test_result/2025-09-04_00-08_hopper01
| pass rate(%) | test(#) | success(#) | compilation error(#) | runtime error(#) | wrong value(#) | timeout(#) |
|----------------|-----------|--------------|------------------------|--------------------|------------------|--------------|
| 74.2% | 310 | 230 | 0 | 0 | 80 | 0 |
>> Summary
| language | category | name | pass rate(%) | test(#) | success(#) | compilation error(#) | runtime error(#) | wrong value(#) | timeout(#) |
|------------|--------------------------|------------------------------|----------------|-----------|--------------|------------------------|--------------------|------------------|--------------|
| cpp | hierarchical_parallelism | memcopy-complex_double | 64.4% | 45 | 29 | 0 | 0 | 16 | 0 |
| cpp | hierarchical_parallelism | memcopy-float | 64.4% | 45 | 29 | 0 | 0 | 16 | 0 |
| cpp | hierarchical_parallelism | atomic_add-float | 77.8% | 72 | 56 | 0 | 0 | 16 | 0 |
| cpp | hierarchical_parallelism | reduction_add-complex_double | 78.4% | 74 | 58 | 0 | 0 | 16 | 0 |
| cpp | hierarchical_parallelism | reduction_add-float | 78.4% | 74 | 58 | 0 | 0 | 16 | 0 |
before change 87db8e9
>> Overall result for test_result/2025-09-04_00-24_hopper01
| pass rate(%) | test(#) | success(#) | compilation error(#) | runtime error(#) | wrong value(#) | timeout(#) |
|----------------|-----------|--------------|------------------------|--------------------|------------------|--------------|
| 100.0% | 310 | 310 | 0 | 0 | 0 | 0 |
>> Summary
| language | category | name | pass rate(%) | test(#) | success(#) | compilation error(#) | runtime error(#) | wrong value(#) | timeout(#) |
|------------|--------------------------|------------------------------|----------------|-----------|--------------|------------------------|--------------------|------------------|--------------|
| cpp | hierarchical_parallelism | atomic_add-float | 100.0% | 72 | 72 | 0 | 0 | 0 | 0 |
| cpp | hierarchical_parallelism | memcopy-complex_double | 100.0% | 45 | 45 | 0 | 0 | 0 | 0 |
| cpp | hierarchical_parallelism | memcopy-float | 100.0% | 45 | 45 | 0 | 0 | 0 | 0 |
| cpp | hierarchical_parallelism | reduction_add-complex_double | 100.0% | 74 | 74 | 0 | 0 | 0 | 0 |
| cpp | hierarchical_parallelism | reduction_add-float | 100.0% | 74 | 74 | 0 | 0 | 0 | 0 |
LIBOMPTARGET_DEBUG=1 shows the old code launching kernels in generic-SPMD mode but new code in generic mode.