gsoc2010-fftw-neon-misc:gsoc2010-fftw-neon-misc.git
8 years agoadjusted ffmpeg_fft/test/01-powers-of-two.c
Christopher Friedt [Mon, 12 Jul 2010 19:19:19 +0000 (15:19 -0400)]
adjusted ffmpeg_fft/test/01-powers-of-two.c

8 years agomodified exported symbols and ffmpeg_fft.h to use the correct ffmpeg fft api as defin...
Christopher Friedt [Mon, 12 Jul 2010 03:28:48 +0000 (23:28 -0400)]
modified exported symbols and ffmpeg_fft.h to use the correct ffmpeg fft api as defined in av_fft.h

8 years agoupdated README to reflect occasional DT_TEXTREL warning
Christopher Friedt [Sun, 11 Jul 2010 04:59:14 +0000 (00:59 -0400)]
updated README to reflect occasional DT_TEXTREL warning

8 years agofixed several random bugs, shared library things, Makefile things. powers of two...
Christopher Friedt [Sun, 11 Jul 2010 04:47:30 +0000 (00:47 -0400)]
fixed several random bugs, shared library things, Makefile things. powers of two work (at least for 2**2).

8 years agocompiles cleanly. fft and rdft contexts should really be opaque. need to finish power...
Christopher Friedt [Sun, 11 Jul 2010 03:27:36 +0000 (23:27 -0400)]
compiles cleanly. fft and rdft contexts should really be opaque. need to finish powers-of-two test.

8 years agoadded install target to Makefile. removed unused libavutil/avstring.*
Christopher Friedt [Sun, 11 Jul 2010 02:15:22 +0000 (22:15 -0400)]
added install target to Makefile. removed unused libavutil/avstring.*

8 years agoremoved log.*, pixfmt.h, error.*
Christopher Friedt [Sun, 11 Jul 2010 01:57:43 +0000 (21:57 -0400)]
removed log.*, pixfmt.h, error.*

8 years agoadded tests, ffmpeg_fft.h still needs a few fixumps
Christopher Friedt [Sun, 11 Jul 2010 01:40:21 +0000 (21:40 -0400)]
added tests, ffmpeg_fft.h still needs a few fixumps

8 years agohid all unnecessary symbols, so that only the necessities are exported in ffmpeg_fft...
Christopher Friedt [Sun, 11 Jul 2010 00:40:42 +0000 (20:40 -0400)]
hid all unnecessary symbols, so that only the necessities are exported in ffmpeg_fft.so. created ffmpeg_fft.h . still need to create a compatible interface for benchfft / libbench

8 years agoadjusted todos
Christopher Friedt [Sat, 10 Jul 2010 15:04:44 +0000 (11:04 -0400)]
adjusted todos

8 years agoadded beginnings of ffmpeg_fft library, for benchfft / libbench comparisons
Christopher Friedt [Sat, 10 Jul 2010 14:37:55 +0000 (10:37 -0400)]
added beginnings of ffmpeg_fft library, for benchfft / libbench comparisons

8 years agoModified benchfft-3.1 so that it only compares two different fftw3 implementations...
Christopher Friedt [Mon, 5 Jul 2010 05:14:02 +0000 (01:14 -0400)]
Modified benchfft-3.1 so that it only compares two different fftw3 implementations. More are possible, but each must have its own libfftw3f.

8 years agoinitial commit of benchfft-3.1
Christopher Friedt [Mon, 5 Jul 2010 04:58:23 +0000 (00:58 -0400)]
initial commit of benchfft-3.1

8 years agomodified print statements slightly
Christopher Friedt [Sat, 26 Jun 2010 02:00:12 +0000 (22:00 -0400)]
modified print statements slightly

8 years agofixed bugs with demo2/neon2.S, which uses d-regs. however, the speedup using d-regs...
Christopher Friedt [Sat, 26 Jun 2010 01:38:27 +0000 (21:38 -0400)]
fixed bugs with demo2/neon2.S, which uses d-regs. however, the speedup using d-regs was only apparent for relatiely short vectors. It might be poor instruction scheduling... will investigate later.

8 years agomodified demo2/timer.{c,h} to use the cycle counter. Requires patch from Mans. Now...
Christopher Friedt [Fri, 25 Jun 2010 23:31:05 +0000 (19:31 -0400)]
modified demo2/timer.{c,h} to use the cycle counter. Requires patch from Mans. Now counts cycles instead of using CLOCK_MONOTONIC (can be switched back using #define USE_CYCLE_COUNTER). Speedup measurement is unchanged, so the two methods are interchangeable to some degree. Modified demo2/neon1.S to provide a speedup of 11x: twice as fast as before! The trick was ensuring that vmla was not followed by vmul. Started working on demo2/neon2.S which will be using d-regs as opposed to q-regs. D-reg usage will give better granularity for tighter pipeline scheduling. Have one bug for chunk sizes >= 3 that I need to fix, but early tests for sizes 1, 2, are showing speedups of 18x. Getting instruction scheduling right is absolutely necessary before going deeper into fftw internals.

8 years agoadded code from Mans that uses cycle counters rather than CLOCK_MONOTONIC. Note:...
Christopher Friedt [Thu, 24 Jun 2010 12:37:21 +0000 (08:37 -0400)]
added code from Mans that uses cycle counters rather than CLOCK_MONOTONIC. Note: requires USER_PMON patch from git.mansr.com/?p=linux-omap;a=commitdiff;h=5170038fa

8 years agorefined main.c, fixed .L3 in neon1.S
Christopher Friedt [Tue, 22 Jun 2010 20:02:07 +0000 (16:02 -0400)]
refined main.c, fixed .L3 in neon1.S

8 years agorebased repository after accidentally committing incorrectly... i blame my current...
Christopher Friedt [Tue, 22 Jun 2010 14:12:05 +0000 (10:12 -0400)]
rebased repository after accidentally committing incorrectly... i blame my current fever and the drugs