Program compiled with `hipcc -O0` from ROCm 5.3.2 do not run correctly

[fxzjshm]

Using vectoradd_hip.cpp from HIP examples, the executable produced by hipcc from ROCm 5.3.2 with -O0 flag cannot run correctly.

Log:

➜  test /opt/rocm-5.2.3/bin/hipcc -O0 vectoradd_hip.cpp -o vectoradd_hip_O0_5.2.3
➜  test /opt/rocm-5.2.3/bin/hipcc -O1 vectoradd_hip.cpp -o vectoradd_hip_O1_5.2.3
➜  test /opt/rocm-5.3.2/bin/hipcc -O0 vectoradd_hip.cpp -o vectoradd_hip_O0_5.3.2
➜  test /opt/rocm-5.3.2/bin/hipcc -O1 vectoradd_hip.cpp -o vectoradd_hip_O1_5.3.2
➜  test ./vectoradd_hip_O0_5.2.3
 System minor 0
 System major 9
 agent prop name AMD Radeon VII
hip Device prop succeeded 
PASSED!                                                                    /0.4s
➜  test ./vectoradd_hip_O1_5.2.3
 System minor 0
 System major 9
 agent prop name AMD Radeon VII
hip Device prop succeeded 
PASSED!                                                                    /0.4s
➜  test ./vectoradd_hip_O0_5.3.2
 System minor 0
 System major 9
 agent prop name AMD Radeon VII
hip Device prop succeeded 
FAILED: 1048576 errors                                                     /0.5s
➜  test ./vectoradd_hip_O1_5.3.2
 System minor 0
 System major 9
 agent prop name AMD Radeon VII
hip Device prop succeeded 
PASSED!                                                                    /0.5s

Seems a clang-15 + ROCm 5.3.2 problem.

Related discussion: illuhad/hipSYCL#869

[zjin-lcf]

I could reproduce the error with ROCm 5.4 on a MI-series GPU.

[abhimeda]

@fxzjshm Hi, is your issue resolved on the latest HIP? If so can we close this ticket?

[fxzjshm]

Not reproduced with ROCm 5.7.0 + gfx1100 & ROCm 6.0.0 + gfx1035, will test gfx906 once have access again.

[fxzjshm]

Still reproduced with ROCm 6.0.0 & 6.0.2 + gfx906.

[ppanchad-amd]

@fxzjshm Internal ticket has been created to resolve this issue. Thanks!

[nartmada]

@fxzjshm, which GPU are you currently to repro the issue? MI50, Radeon PRO VII, or Radeon VII? Thanks.

[fxzjshm]

@nartmada It's Radeon VII, as shown above in the log. Is this bug reproducible on your test environment?

[harkgill-amd]

Hi @fxzjshm, I was unable to reproduce your issue compiling the vectoradd_hip.cpp program with the -O0 flag on gfx906. Please see the output below.

rocm@rocm:~/vecadd$ hipcc -O0 vectoradd_hip.cpp -o vectoradd
#Warnings
rocm@rocm:~/vecadd$ ./vectoradd
 System minor 0
 System major 9
 agent prop name AMD Radeon VII
hip Device prop succeeded
PASSED!

This was tested on the latest ROCm 6.1.2 release. I will close this ticket for now, if you encounter this issue again, feel free to re-open this issue. Thanks!