Closed jjennychen closed 1 year ago
There are also some other failed compilations for other benchmarks. Would adding them as a list under this issue or creating a separate issue for each error be more convenient/desirable? Thank you!!
The benchmark has a warp vote function:
...
tmp = dx*dx + (dy*dy + (dz*dz + epssqd)); // compute distance squared (plus softening)
if ((n < nbodiesd) || __all(tmp >= dq[depth])) {
...
For warp vote functions you''ll need patched clang and llvm-spirv to compile the application.
Would adding them as a list under this issue or creating a separate issue for each error be more convenient/desirable?
If the cases fail for other reasons than the warp functions, it might be better to put them into separate issues. If you are unsure, you may list them here.
I'd like to add a note. Running the "bh-hip" on an AMD GPU (e.g. wave-front size is 32) hangs. Not sure if this is reproducible. I don't think the CUDA program "bh-cuda" hangs on modern NVIDIA GPUs.
The issue was also reported here: https://github.com/oneapi-src/SYCLomatic/issues/791
Thank you for the note! Yeah the "bh-hip" hung when I ran it on Intel Gen9 too (with ChipStar/CHIP-SPV)... But it did compile successfully with the patched clang and llvm-spirv! Thank you again!!
I hope patched clang and llvm-spirv are installed by default in the chipStar install guide.
Yep, there's a mention in the beginning of the README.md. I suppose this can be now closed?
When running a benchmark from the benchmark set (https://github.com/zjin-lcf/HeCBench.git), I got the following error that caused the compilation to fail: (seems like some issues with the llvm-spirv...)
[To Reproduce the Issue]
git clone https://github.com/zjin-lcf/HeCBench.git
qsub -I -n 1 -q iris -t 360
cd HeCBench/bh-hip
make VERIFY=yes DEBUG=yes
After that make command, the error about llvm-spirv should show up.
(Extra note: to run the benchmark after compiling, do
make run
)