search

Suvarna1488 / webm

Automatically exported from code.google.com/p/webm
0 stars 0 forks source link

vp9 cpu speed -6 is slower than cpu speed -5 on arm #908

Open GoogleCodeExporter opened 9 years ago

GoogleCodeExporter commented 9 years ago 
I have only tested with a 640x480 file on Arm.

My parameters:
vpxenc --codec=vp9 --profile=0 --kf-max-dist=90000 --static-thresh=0 
--end-usage=cbr --min-q=2 --max-q=56 --undershoot-pct=50 --overshoot-pct=50 
--buf-sz=1000 --buf-initial-sz=500 --buf-optimal-sz=600 --max-intra-rate=300 
--resize-allowed=0 --passes=1 --rt --lag-in-frames=0 --noise-sensitivity=0 
--error-resilient=1 --drop-frame=0 -D --target-bitrate=600 --cpu-used=-5 -w 640 
-h 480 -o <out> <in>

vpxenc --codec=vp9 --profile=0 --kf-max-dist=90000 --static-thresh=0 
--end-usage=cbr --min-q=2 --max-q=56 --undershoot-pct=50 --overshoot-pct=50 
--buf-sz=1000 --buf-initial-sz=500 --buf-optimal-sz=600 --max-intra-rate=300 
--resize-allowed=0 --passes=1 --rt --lag-in-frames=0 --noise-sensitivity=0 
--error-resilient=1 --drop-frame=0 -D --target-bitrate=600 --cpu-used=-6 -w 640 
-h 480 -o <out> <in>

-5:
Pass 1/1 frame  471/471  1171701B   19901b/f  597045b/s   48349 ms (9.74 fps)

-6:
Pass 1/1 frame  471/471  1171044B   19890b/f  596710b/s   53030 ms (8.88 fps)

Original issue reported on code.google.com by fgalli...@google.com on 7 Jan 2015 at 6:42

GoogleCodeExporter commented 9 years ago 
which test sequence?

Original comment by jingn...@google.com on 7 Jan 2015 at 6:45

GoogleCodeExporter commented 9 years ago 
Is it hitting more unoptimized code?

Original comment by johannko...@google.com on 7 Jan 2015 at 6:58

GoogleCodeExporter commented 9 years ago 
I was testing niklas_640_480_30.yuv and tacomanarrows_640_480_30.

Could be it is hitting more unoptimized code.

On the Nexus 5 we can see that the 640x480 -6 is slower than -5 as well. But 
for niklas 720p -6 is faster than -5.

Original comment by fgalli...@google.com on 7 Jan 2015 at 7:47

GoogleCodeExporter commented 9 years ago 
Is this function optimized in ARM?

vp9_fdct8x8_quant

Original comment by jingn...@google.com on 7 Jan 2015 at 8:03

GoogleCodeExporter commented 9 years ago 
No. I changed the title of the bug.

Thanks for looking into this.

Original comment by fgalli...@google.com on 7 Jan 2015 at 8:12

GoogleCodeExporter commented 9 years ago

Original comment by fgalli...@google.com on 16 Jan 2015 at 6:51

GoogleCodeExporter commented 9 years ago 
Actually after profiling it is because we do not have Neon optimized functions 
for:
vp9_sad64x64x4d_c
vp9_sad16x16x4d_c
vp9_sad32x32x4d_c
vp9_fdct8x8_quant

Original comment by fgalli...@google.com on 24 Jan 2015 at 11:09