slow OpenCL with AMD 21.30 driver

[andyrtr]

I've received my AMD RX 460 card (4GB) and played with getting OpenCL working using Arch Linux in DT 3.6.1. I've found that all closed AMD drivers up to 21.20 are a nice speedup and report it to be

[opencl_init] device 0: Baffin 
     GLOBAL_MEM_SIZE:          3493MB
     MAX_WORK_GROUP_SIZE:      256
     MAX_WORK_ITEM_DIMENSIONS: 3
     MAX_WORK_ITEM_SIZES:      [ 1024 1024 1024 ]
     DRIVER_VERSION:           3224.4
     DEVICE_VERSION:           OpenCL 1.2 AMD-APP (3224.4)

while the latest 21.30 version is a lot slower than cpu only

[opencl_init] device 0: Baffin 
     GLOBAL_MEM_SIZE:          4096MB
     MAX_WORK_GROUP_SIZE:      256
     MAX_WORK_ITEM_DIMENSIONS: 3
     MAX_WORK_ITEM_SIZES:      [ 1024 1024 1024 ]
     DRIVER_VERSION:           3302.5 (PAL,HSAIL)
     DEVICE_VERSION:           OpenCL 2.0 AMD-APP (3302.5)

I couldn't menage to fix it with all the opencl options in darktablerc. Could this be a darktable issue or driver version related where we can't do much or is there a way to fix it for 21.30 and future driver releases?

I'm attaching logs of
time darktable-cli bench.SRW test.jpg --core -d perf -d opencl > log_21.x0.txt

log_21.20.txt
log_21.30.txt

[johnny-bit]

If dt version haven't changed, but drivers did... And performance went down from drivers... Why filling issue report with darktable rather than driver vendor? I'm sure AMD has some way to do so.

[andyrtr]

I was in hope this can be "fixed" by tweaking DT opencl settings or DT opencl checks might need some additional improvement. Sure I'd prefer to use a stable and OS opencl driver some day (ROCm is not supported proably due to my PCIe v2 mainboard and there's no mesa opencl image support yet).

[johnny-bit]

You use drivers, you detect problem, you should report it to AMD.

Please report to AMD, link here and close the issue.

[andyrtr]

Done here (I hope they remove the spam marking there soon).