[x264-devel] x264 thread scalability (was unaligned accesses in IA64)

[Mauricio Alvarez]

On Tue, Feb 26, 2008 at 1:06 PM, Mauricio Alvarez <lokifo at gmail.com> wrote:
>  I'm testing X264 on an Itanium machine . I want to test the parallel
>  scalability of the threading model implemented in X264.

BTW, Just some numbers of my experiments with the multiprocessor machine:
   - icc compiler (-O2)
   - input sequence: 1088p25_blue_sky,  100 frames
   - x264 command:  x264 --qp=20 --bframes 2 --keyint 100 --no-b-adapt
--scenecut=-1 --b-bias=0 --ref 1  --fps 25 --frames 100
--threads=${PTHREAD_NUM_THREADS}  -o 1088p25_blue_sky.h264
1088p25_blue_sky.yuv 1920x1088
  - time results are average of 10 executions

processors - time [s] - speed-up
1	100.84	1
2	65.08	1.55
4	31.13	3.24
8	18.33	5.5
12	13.22	7.63
16	9.56	10.54
20	9.15	11.02
24	8.15	12.37
28	8.01	12.58
32	6.14	16.43
36	5.93	17.01
40	6.2	16.28
44	5.98	16.87
48	6.22	16.21
52	6.38	15.8
56	6.37	15.83

results in terms of parallel scalability are similar to those reported
in http://www.via.ecp.fr/via/ml/x264-devel/2007-02/msg00064-en.html
but in absolute time I obtaining less performance I think due to the
exceptions that the program is getting due to the unaligned accesses.

regards,

Mauricio A.