Quote:
Originally Posted by littlepox
Hi all, need some help with KNLMeansCL on AMD RX Vega64:
denoised Y clip comes with broken data (mostly zero, very few times random output). This is a known issue ( https://github.com/Khanattila/KNLMeansCL/issues/26) and we will await the patch.
Meanwhile, we would like to know whether using RGB is a safe workaround or not(assume we don't want to touch UV planes) :
Code:
a = "your file with YUV-8bit format.m2ts"
src8 = core.lsmas.LWLibavSource(a,threads=1)
src16 = core.fmtc.bitdepth(src8, bits=16)
Y = core.std.ShufflePlanes(src16, 0, vs.GRAY)
nr16y = core.resize.Bicubic(Y, format=vs.RGB48,matrix_in_s="709")
nr16y = core.knlm.KNLMeansCL(nr16y, d=1, a=2, s=3, h=1.3, channels="RGB", device_type="GPU")
nr16y = core.resize.Bicubic(nr16y, format=vs.GRAY16,matrix_s="709")
nr16y.set_output()
It looks OK for a few samples but we would like to know whether it is indeed unaffected by the same issue. Thanks for help! |
Hi littlepox, thanks for your report.
The main problem is I do not have a Vega GPU. In the country where I live the cost is very high, over 800€!
I also tried AMD's CodeXL, but for now, I do not see something strange. Future investigations are required.
I do not know if the workaround you suggest works or not, there is a problem with the new AMD OpenCL compiler for Vega.
There are probably some optimizations in place that are not clear.