preda / gpuowl

GPU Mersenne primality test.
GNU General Public License v3.0
127 stars 35 forks source link

Performance regression on Ubuntu 22.0 with ROCm 5.4.3/5.4.5/5.5 and latest gpuOwl version, exponent 114710069 #268

Closed selroc closed 7 months ago

selroc commented 1 year ago

ROCm 5.4.3: GPU Radeon VII us/it 772

selroc commented 1 year ago

ROCm 5.4.5: GPU Radeon VII us/it 775

selroc commented 1 year ago

Getting worse: ROCm 5.5: GPU Radeon VII us/it 797

selroc commented 1 year ago

ROCm 5.5.1: GPU Radeon VII us/it 799

selroc commented 1 year ago

ROCm 5.6: GPU Radeon VII (with barrier CLK_LOCAL_MEM_FENCE) us/it 1233 (with barrier 0) us/it 1206

selroc commented 1 year ago

I'm stopping Mersenne search with gpuOwl until it is Re-optimized.

selroc commented 1 year ago

ROCm 5.6.1: Radeon VII Still 1206 us/it for the same exponent.

selroc commented 1 year ago

ROCm 5.7.0: Radeon VII Performance Back To Normal 786 us/it.

selroc commented 1 year ago

ROCm 5.7.0: Radeon VII Performance Back To Normal 783 us/it with barrier(0) instead of CLK_LOCAL_MEM_FENCE.