search

ROCm / MIOpen

AMD's Machine Intelligence Library
https://rocm.docs.amd.com/projects/MIOpen/en/latest/
Other
1.06k stars 220 forks source link

asm igemm wrw computation error with stride=2, padding=2, filter=3, h=w=1 #2867

Open carlushuang opened 5 months ago

carlushuang commented 5 months ago

SWDEV-451681 MIOPEN_DEBUG_CONV_DIRECT=0 MIOPEN_DEBUG_CONV_FFT=0 MIOPEN_DEBUG_CONV_WINOGRAD=0 MIOPEN_DEBUG_CONV_GEMM=0 MIOPEN_DEBUG_CONV_IMPLICIT_GEMM=1 /opt/rocm/bin/MIOpenDriver conv -n 2 -c 1 -H 1 -W 1 -k 1 -y 3 -x 3 -p 2 -q 2 -u 2 -v 2 -l 1 -j 1 -m conv -g 1 -F 4 -t 1 MIOpenDriver conv -n 2 -c 1 -H 1 -W 1 -k 1 -y 3 -x 3 -p 2 -q 2 -u 2 -v 2 -l 1 -j 1 -m conv -g 1 -F 4 -t 1

atamazov commented 5 months ago

@carlushuang It is unclear what is wrong from the description. Can you please attach the faulty log (at level 5 or 6)?

AFAICS from the title, the issue is specific to 1x1 image, 3x3 filter, 2x2 padding and stride. The workaround #2869 is narrower, it cuts configs where C=1 && K=1. Is it correct, i.e. the bug only appears when C=K=1?

Maybe you know, has anyone tested adjacent configs, e.g. 2x2 image, or 2x2 filter, or different C and K?