I got the kqtgmc , knnedi3 , working with avisynth+cuda .
Some quick tests on a gtx860 were consistently slower than avisynth+mt for KQTGMC, and KNNEDI3 (ie. avisynth+mt on CPU was about 1.2-1.4x faster using the normal CPU versions QTGMC and NNEDI3), maybe a faster card would put it over the top . I didn't test validity/quality results yet.
The same .dll was able to run the "normal" avisynth+mt script as is too (ie. with prefetch(x) ), with the same speed
Quote:
Originally Posted by tuanden0
Here's my test with GTX 1060 3GB:
Without CUDA: https://i.imgur.com/W2WFT9R.png
With CUDA: https://i.imgur.com/aE1SbfM.png
I tried to remove DumpfilterGraph and it still work but no FPS change.
If I remove Prefetch then FPS go down to 48 and time go up to 2 min.
I put "OnCUDA" on all filteres but it's not work, except "OnCPU" work with source filter and FPS not change.
|
Look at the readme (use google translate) . You need to wrap in OnCPU, OnCUDA
Did you notice no speed change ? What about GPU-z ? monitor GPU card usage, or try device_index=x to swap GPU# (take out KNLMeansCL for testing now, because that will use GPU too)
Im guessing not all filters are accelerated, or maybe your card is as fast as CPU processing ?