Hello! In attach SSE support for libac3. Please test it and apply if it works, otherwise - simply forget about. Best regards! Nick
Hello Nick, Monday, July 09, 2001, 2:21:53 PM, you wrote: NK> Hello! NK> In attach SSE support for libac3. NK> Please test it and apply if it works, otherwise - simply forget about. Great, it works (I've tested with 2.0 and 5.1 48kHz AC3) and it gives speedup of approx 10-15% on PIII coppermine. Btw. grabbed pcm dump of output and it differs, but think it's due to different srfft algorithm I think, I think it's no prob cause it sounds ok (did diff of audiodump.wav of mplayer with --disable-sse and --enable-sse). NK> Best regards! Nick -- Best regards, Felix mailto:atmosfear@users.sourceforge.net _______________________________________________ Mplayer-dev-eng mailing list Mplayer-dev-eng@lists.sourceforge.net http://lists.sourceforge.net/lists/listinfo/mplayer-dev-eng
Hi, Felix Bьnemann! On 2001-07-10 10:29:35 you wrote:
NK> In attach SSE support for libac3. NK> Please test it and apply if it works, otherwise - simply forget about.
Great, it works (I've tested with 2.0 and 5.1 48kHz AC3) and it gives speedup of approx 10-15% on PIII coppermine.
Btw. grabbed pcm dump of output and it differs, but think it's due to different srfft algorithm I think, I think it's no prob cause it sounds ok (did diff of audiodump.wav of mplayer with --disable-sse and --enable-sse).
The same problem was with old mp3 stuff, IMHO. It's normal since P3 has minor different precision for FPU and SSE units (I still don't study this question in detail). But problem is not in algorithms since there is used same algorithm for all types of decoders (SSE, 3DNOW). Maybe you'll want port dct64_3dnow.s (dct64_k7.s even better) to SSE as dct64_sse.s? At least phases 1-5? They can be simply converted from 3dnow to sse, but phases 6-10 could be interleaved with 1-5 to get better performance. (I think that per 1 SSE opcodes need 2 FPU opcodes to load cpu optimally). Best regards! Nick (2001-07-11 00:27:43) _______________________________________________ Mplayer-dev-eng mailing list Mplayer-dev-eng@lists.sourceforge.net http://lists.sourceforge.net/lists/listinfo/mplayer-dev-eng
Hello Nick, Wednesday, July 11, 2001, 2:35:11 AM, you wrote: NK> Hi, Felix Bьnemann! NK> On 2001-07-10 10:29:35 you wrote:
NK> In attach SSE support for libac3. NK> Please test it and apply if it works, otherwise - simply forget about.
Great, it works (I've tested with 2.0 and 5.1 48kHz AC3) and it gives speedup of approx 10-15% on PIII coppermine.
Btw. grabbed pcm dump of output and it differs, but think it's due to different srfft algorithm I think, I think it's no prob cause it sounds ok (did diff of audiodump.wav of mplayer with --disable-sse and --enable-sse).
NK> The same problem was with old mp3 stuff, IMHO. It's normal since P3 has NK> minor different precision for FPU and SSE units (I still don't study this question NK> in detail). But problem is not in algorithms since there is used same algorithm NK> for all types of decoders (SSE, 3DNOW).
NK> Maybe you'll want port dct64_3dnow.s (dct64_k7.s even better) to SSE as NK> dct64_sse.s? At least phases 1-5? They can be simply converted from NK> 3dnow to sse, but phases 6-10 could be interleaved with 1-5 to get better NK> performance. (I think that per 1 SSE opcodes need 2 FPU opcodes to load NK> cpu optimally). of course I'd like to, question is if I manage, think I'll get the basic stuff, but some of the complex instructions sometimes confuse me. But I'll take a look at it. Btw. how are instructions with 3 parameters converted from intel to at&t syntax? NK> Best regards! Nick (2001-07-11 00:27:43) -- Best regards, Felix mailto:atmosfear@users.sourceforge.net _______________________________________________ Mplayer-dev-eng mailing list Mplayer-dev-eng@lists.sourceforge.net http://lists.sourceforge.net/lists/listinfo/mplayer-dev-eng
Hi, Felix Bьnemann! On 2001-07-10 23:36:28 you wrote:
NK> In attach SSE support for libac3. NK> Please test it and apply if it works, otherwise - simply forget about.
Great, it works (I've tested with 2.0 and 5.1 48kHz AC3) and it gives speedup of approx 10-15% on PIII coppermine.
Btw. grabbed pcm dump of output and it differs, but think it's due to different srfft algorithm I think, I think it's no prob cause it sounds ok (did diff of audiodump.wav of mplayer with --disable-sse and --enable-sse).
NK> The same problem was with old mp3 stuff, IMHO. It's normal since P3 has NK> minor different precision for FPU and SSE units (I still don't study this question NK> in detail). But problem is not in algorithms since there is used same algorithm NK> for all types of decoders (SSE, 3DNOW).
NK> Maybe you'll want port dct64_3dnow.s (dct64_k7.s even better) to SSE as NK> dct64_sse.s? At least phases 1-5? They can be simply converted from NK> 3dnow to sse, but phases 6-10 could be interleaved with 1-5 to get better NK> performance. (I think that per 1 SSE opcodes need 2 FPU opcodes to load NK> cpu optimally). of course I'd like to, question is if I manage, think I'll get the basic stuff, but some of the complex instructions sometimes confuse me. But I'll take a look at it.
Btw. how are instructions with 3 parameters converted from intel to at&t syntax?
In reverse order ;) If you mean shufps or pshuw then: Intel syntax: shufps mm0, mm2, 5 at&t: shufps $5, %mm2, %mm0 Best regards! Nick (2001-07-11 10:25:56) _______________________________________________ Mplayer-dev-eng mailing list Mplayer-dev-eng@lists.sourceforge.net http://lists.sourceforge.net/lists/listinfo/mplayer-dev-eng
participants (2)
-
Felix Bьnemann -
Nick Kurshev