[OpenMP] complex reduction failure on AMD GPU

[ye-luo]

reproducer code https://github.com/ye-luo/openmp-target/blob/master/tests/complex/complex_reduction.cpp

$ clang++ -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906 -fopenmp-target-new-runtime -O3 complex_reduction.cpp -lm
$ ./a.out 
Testing float
hierarchical parallelism wrong reduction value check(0,0) correct value (4950,-4950)
Testing double
hierarchical parallelism wrong reduction value check(0,0) correct value (4950,-4950)

On a different machine with MI50, different OS. I got

[GPU Memory Error] Addr: 0x7f0613e00000 Reason: Page not present or supervisor privilege. 
Memory access fault by GPU node-2 (Agent handle: 0x1b883b0) on address 0x7f0613e00000. Reason: Page not present or supervisor privilege.

[jdoerfert]

I'm assuming/hoping this is fixed on mainline now.

[ye-luo]

Still got zero with 06f346c

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

[jdoerfert]

This worked for me now. I think the manual unroll (a619072) hides the error. I further believe the AMD backend pass transforming allocas to vectors is the problem, or better the handling of pointer vectors by the backend. (Fyi @arsenm)

[arsenm]

This worked for me now. I think the manual unroll (a619072) hides the error. I further believe the AMD backend pass transforming allocas to vectors is the problem, or better the handling of pointer vectors by the backend. (Fyi @arsenm)

I tried reverting this patch and I'm not seeing anything. Do you have the failing IR/ISA from when this was broken?

[arsenm]

This worked for me now. I think the manual unroll (a619072) hides the error. I further believe the AMD backend pass transforming allocas to vectors is the problem, or better the handling of pointer vectors by the backend. (Fyi @arsenm)

I tried reverting this patch and I'm not seeing anything. Do you have the failing IR/ISA from when this was broken?

Ping, do you have the before/after sample IR/ISA?

[ye-luo]

This issue is obsolete. The most recent breakage was reported as #58774

[arsenm]

This issue is obsolete. The most recent breakage was reported as #58774

My understanding is the backend failure was worked around with a loop unroll, indicating there is still something broken here