VP vs CUDA decoding performance

[admin]

Frames corrupted with macroblocking. It's not subtle at all. I seriously doubt that this bug is still present. Unfortunately, I cannot remember which stream invoked the issue so I can't test it.

[hydra3333]

Frames corrupted with macroblocking. It's not subtle at all. I seriously doubt that this bug is still present. Unfortunately, I cannot remember which stream invoked the issue so I can't test it.

OK thanks.

I just did a test with a 720x576 MPEG-2 clip with my brand new GTX 750 using AVSMeter (GPU/VPU monitoring enabled):
DECODE_MODES 0,1,0 (default, GPU):

Code: Select all

[Runtime info]
Frames processed:               29315 (0 - 29314)
FPS (min | max | average):      139.0 | 1638 | 1034
Memory usage (phys | virt):     172 | 233 MB
Thread count:                   4
CPU usage (average):            29%
GPU usage (average):            30%
Video engine load (average):    0%
GPU memory usage:               79 MB
Time (elapsed):                 00:00:28.353

DECODE_MODES 0,0,0 (VPU):

Code: Select all

[Runtime info]
Frames processed:               29315 (0 - 29314)
FPS (min | max | average):      100.8 | 2075 | 838.7
Memory usage (phys | virt):     162 | 222 MB
Thread count:                   4
CPU usage (average):            3%
GPU usage (average):            17%
Video engine load (average):    50%
GPU memory usage:               69 MB
Time (elapsed):                 00:00:34.953

VPU mode seems to have the advantage, it's only slightly slower but the CPU load is virtually zero.

On the assumption that my process is cpu bound (achieving circa 3fps with 576i mpeg2 -> avisynth cleanup script -> x264.exe) then "any" lower NV CPU usage translates to more CPU cycles available for avisynth/x265 ... I am guessing due to the low fps throughput, the impact will be less than negligible however every bit helps

[admin]

If our concern is investigating the performance of the source filter, it's highly desirable to do this kind of benchmarking with a minimal script for processing, i.e., just the source filter. I'd be interested to see hydra3333's results doing that. It looks strange that VPU showed a significantly higher maximum decoding rate but a slower average one. Of course I understand hydra3333 wanted to show the effect on CPU usage in his use case. Anyway, hydra3333's results appear to further justify the return to 0,0,0.

[hydra3333]

Crikey. That must be one hell of a cleanup script. Does anything of the original video survive?

not a lot. Crikey is about right, source is Aussie OTA broadcast of fast ball sports action with 36 players on a very large grassed field and the broadcaster almost destroys the video with low bandwidth and terrible blocking. I'm nearly sure I'm doing something wrong with the script however the result turns out to be slightly more watchable than the source (if my team wins).