[PD-dev] first auto-vector test
Hans-Christoph Steiner
hans at eds.org
Mon Dec 18 05:44:23 CET 2006
I made two builds of Pd-0.39-2-extended-test6 on Mac OSX using gcc
4.0.1, with differing CFLAGS. I used the attached patch to measure
the CPU usage, I tried to keep the operating circumstances equal, but
this is a rough test. But from this, the results look quite promising:
These flags are used in the standard build:
-Os -funroll-loops -fomit-frame-pointer -mcpu=powerpc -mtune=7450 -
mpowerpc-gfxopt
The second is optimized for G4+:
-fast -fPIC -mcpu=7450 -mtune=7450 -ftree-vectorize -ftree-vectorizer-
verbose=3
| filename | G4+ | powerpc |
------------------------------------------------------------------------
feedback_delay_network.pd
[linear 32 100 105( 91 165
[linear 65 10 15( 228 237
[exponential 64 1000 10( 249 280
(fdn~ NOT vectorized; d_soundfile.c, d_ctl.c, m_sched.c vectorized)
ramping_up_and_down.pd 61 64
doc/3.audio.examples/G90.pitchshift.pd 95 97
My guess its the [readsf~] that is so much quicker in the feedback
patch.
.hc
-------------- next part --------------
A non-text attachment was scrubbed...
Name: feedback_delay_network.pd
Type: application/octet-stream
Size: 1540 bytes
Desc: not available
URL: <http://lists.puredata.info/pipermail/pd-dev/attachments/20061217/ac8ed794/attachment.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: load-meter.pd
Type: application/octet-stream
Size: 863 bytes
Desc: not available
URL: <http://lists.puredata.info/pipermail/pd-dev/attachments/20061217/ac8ed794/attachment-0001.obj>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: ramping_up_and_down.pd
Type: application/octet-stream
Size: 1018 bytes
Desc: not available
URL: <http://lists.puredata.info/pipermail/pd-dev/attachments/20061217/ac8ed794/attachment-0002.obj>
-------------- next part --------------
------------------------------------------------------------------------
Mistrust authority - promote decentralization. - the hacker ethic
More information about the Pd-dev
mailing list