AMD Zambezi news, info, fans !

[undone]

Some interesting talk about x264 optimizing on bulldozer

http://www.planet3dnow.de/vbulletin/...&postcount=562
http://www.planet3dnow.de/vbulletin/...&postcount=585


Looks like fma4 & XOP bring greater help than AVX on bulldozer. Seems we would get a revolutionay change since MMX.

2011-09-16 23:42:16 < Dark_Shikari> Oh YI, we know now why AVX is useless on bulldozer
2011-09-16 23:42:20 < Dark_Shikari> *FYI
2011-09-16 23:42:22 < Dark_Shikari> Move elimination
2011-09-16 23:42:29 < Dark_Shikari> Their OOE engine eliminates moves and resolves them before ALU stage
2011-09-16 23:42:34 < Dark_Shikari> So moves are free, so AVX doesn't help
2011-09-16 23:42:39 < Dark_Shikari> Except reducing code size ofc

2011-09-23 18:56:03 < Dark_Shikari> Okay, so I have a massive series of bulldozer profiles ready
2011-09-23 18:56:13 < Dark_Shikari> It has instruction-based sampling and all sorts of awesome stuff
2011-09-23 18:56:43 < JEEB> AMD? Awesome stuff? This sounds like something that doesn't happen very often
2011-09-23 18:57:21 < Gramner> any NDA?
2011-09-23 18:59:53 < Dark_Shikari> Technically yeah
2011-09-23 19:00:08 < Dark_Shikari> Though a lot of the stuff isn't bulldozer-specific, its performance counters are just awesome
2011-09-23 19:00:32 < Dark_Shikari> Unsurprisingly, our load/store queue is full in pixel_avg functions.
2011-09-23 19:01:25 < Dark_Shikari> Er, load queue.
2011-09-23 19:01:36 < Dark_Shikari> Our store queue, on the other hand, fills in plane_copy, mc_copy...
2011-09-23 19:01:38 < Dark_Shikari> slicetype_mb_cost?
2011-09-23 19:02:12 < Dark_Shikari> cache_load and cache_save, guess that's obvious
2011-09-23 19:02:33 < Dark_Shikari> analyse_init, naturally
2011-09-23 19:02:50 < Dark_Shikari> Okay, time for INEFFECTIVE_SW_PREFETCHES
2011-09-23 19:03:05 < Dark_Shikari> Oh, this is awesome. It tells you when a prefetch is useless, i.e. the data was already in L1 cache
2011-09-23 19:03:12 < Dark_Shikari> Almost all of the "useless prefetches", pengvado, are in hpel_filter
2011-09-23 19:03:21 < Dark_Shikari> The rest are in cache_load
2011-09-23 19:03:23 < Dark_Shikari> Guess that's expected.
2011-09-23 19:04:02 < Dark_Shikari> Next: DECODER_EMPTY.
2011-09-23 19:04:17 < Dark_Shikari> I... think this is where the instruction decoder... hmm. Is this where the decoder is too fast, or too slow?
2011-09-23 19:04:43 < Dark_Shikari> Okay, it's where the decoder is too slow (there's nothing to dispatch)

(...)
2011-09-23 21:47:40 < Dark_Shikari> Thank you performance counters, I think I just made CABAC RD way faster
2011-09-23 21:48:37 < LordRPI> nice
2011-09-23 21:49:22 < Dark_Shikari> 50% of the branch mispredictions in cabac were on one line of code
2011-09-23 21:49:26 < Dark_Shikari> a restructure of the function, kabam

2011-09-27 00:55:51 < Dark_Shikari> pengvado: oh oops, vpermilps and pd are 5-operand (!!!!!)
2011-09-27 00:55:57 < Dark_Shikari> dst,src1,src2,selector,imm8
2011-09-27 00:56:25 < Dark_Shikari> I mean seriously wtf
2011-09-27 01:04:02 < Dark_Shikari> Also, they apparently dropped 3DNOW

2011-09-28 01:33:41 < Dark_Shikari> AVX mbtree propagate is slower than sse2
2011-09-28 01:33:49 < Dark_Shikari> FMA only barely manages to get it fast again.
2011-09-28 01:33:49 < kemuri-_9> lol
2011-09-28 01:33:52 < Sean_McG> hahah
2011-09-28 01:33:59 < Dark_Shikari> SSE2: 342 cycles
2011-09-28 01:34:00 < Dark_Shikari> AVX: 374
2011-09-28 01:34:05 < Dark_Shikari> FMA4: 340
2011-09-28 01:34:18 < kemuri-_9> lol
2011-09-28 01:34:26 < Dark_Shikari> I guess this makes sense given that it only has 128-bit execution units
2011-09-28 01:34:34 < Dark_Shikari> and the INT16_TO_FLOAT code is obnoxiously slow because avx sucks
2011-09-28 01:34:41 < Dark_Shikari> i.e. avx has no way of doing int16_t -> float fast
2011-09-28 01:35:18 < Dark_Shikari> Hmm. I wonder if FMA4 supports sse registers?
2011-09-28 01:35:37 < Dark_Shikari> Oh. It *does*...
2011-09-28 01:35:38 < Dark_Shikari> Let me try that.
2011-09-28 01:37:45 * codestr0m ears perk up
2011-09-28 01:49:29 < Dark_Shikari> FMA4: 314 cycles. Much better
2011-09-28 01:49:46 < codestr0m> Dark_Shikari: what was the change?
2011-09-28 02:01:21 < Dark_Shikari> using the sse instead of avx version
2011-09-28 02:01:26 < Dark_Shikari> as the basis for xop

2011-10-01 02:09:51 < Dark_Shikari> xop will make this a lot easier, but I'm trying to do ssse3 first

2011-10-04 04:46:38 < Dark_Shikari> C, with mode analysis shortcuts: 253 cycles
2011-10-04 04:46:45 < Dark_Shikari> My crappy, badly optimized XOP asm: 93 cycles
2011-10-04 04:46:56 < Dark_Shikari> This is kinda awesome
2011-10-04 04:49:35 < Dark_Shikari> Oh, and old without shortcuts: 379 cycles
2011-10-04 04:49:45 < Dark_Shikari> My asm is 4 times faster than the existing... wait where have we seen this before? XD
2011-10-04 04:49:57 < Dark_Shikari> It's just like SAD_4x4_x9 all over again!
2011-10-04 04:50:10 < JEEB>
2011-10-04 04:50:18 < JEEB> that sounds pretty awesome
2011-10-04 04:50:21 < Dark_Shikari> Except this time I'm still wondering how best to do it without vpperm
2011-10-04 04:50:33 < Dark_Shikari> Thanks AMD, for bringing back the best instruction ever after 15+ years of hiatus.