dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2024-11-20 12:28:26 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity
2,701 Commits 34 Branches 64 Tags 14 MiB
a53c336919
Commit Graph

513 Commits

Author SHA1 Message Date
James Zern
a53c336919 lossless_neon: add VP8LTransformColorInverse 
based on SSE2, only ~11% faster

Change-Id: I45434639d81e153f01f77c1f5d2da510b542170e
 2015-08-04 23:22:36 -07:00
James Zern
99131e7f8c Merge changes I9fb25a89,Ibc648e9e 
* changes:
  lossless_neon: remove predictors 5-13
  ll_enc_neon: enable VP8LSubtractGreenFromBlueAndRed
 2015-08-04 02:24:15 +00:00
Pascal Massimino
c455676680 simplify the main loop for downscaling 
(part of bug #254 investigation)

no speed change observed.

Change-Id: Ie21b33171def367f37643fef6a0bd378e49468c7
 2015-08-03 16:57:35 +02:00
James Zern
2a010f992a lossless_neon: remove predictors 5-13 
operating on single uint32's isn't helped by NEON.
this improves aarch64 performance by ~4%

Change-Id: I9fb25a8962de7b80e893e756ee7c76393cfd40c7
 2015-07-28 19:44:58 -07:00
James Zern
ca221bbc48 ll_enc_neon: enable VP8LSubtractGreenFromBlueAndRed 
this moves the function outside the WEBP_USE_INTRINSICS check.
there's no alternative version and it's ~54% faster at the
function level and mildly faster overall

Change-Id: Ibc648e9ee35021d48901e05aa596aa01067796a2
 2015-07-28 19:44:45 -07:00
Jyrki Alakuijala
85b44d8a69 lossless: encoding, don't compute unnecessary histo 
share the computation between different modes

3-5 % speedup for lossless alpha
1 % for lossy alpha

no change in compression density

Change-Id: I5e31413b3efcd4319121587da8320ac4f14550b2
 2015-07-07 20:24:26 -07:00
Pascal Massimino
0ae2c2e4b2 SSE2/SSE41: optimize SSE_16xN loops 
After several trials at re-organizing the main loop and accumulation scheme,
this is apparently the faster variant.

removed the SSE41 version, which is no longer faster now.
For some reason, the AVX variant seems to benefit most for the change.

Change-Id: Ib11ee18dbb69596cee1a3a289af8e2b4253de7b5
 2015-07-02 20:55:04 +02:00
James Zern
39216e59d9 cosmetics: fix indent after 32462a07 
Change-Id: If9a5d91c25e981bc4cd81adb476244e63fc7c3c8
 2015-07-01 23:49:20 -07:00
James Zern
559e54ca60 Merge "SSE2: slightly faster FTransformWHT" 2015-07-02 06:36:33 +00:00
Pascal Massimino
8ef9a63b45 SSE2: slightly faster FTransformWHT 
goes from 0.3% to 0.1% overall CPU time, but...

Change-Id: I4c9a92b1e1d6b58ed57c6b890366f1dbeaf84f84
 2015-07-01 23:03:17 -07:00
James Zern
f27f773576 lossless_neon: enable VP8LAddGreenToBlueAndRed 
this moves the function outside the WEBP_USE_INTRINSICS check.
there's no alternative version and it's ~70% faster at the
function level and 1-2% faster overall

Change-Id: I59fb4918ec86b1ac3a47cbd5d05ce62f007461cb
 2015-07-01 22:50:54 -07:00
Pascal Massimino
36e9c4bc50 SSE2: minor cosmetrics on in-loop filter code 
Change-Id: Ic0e6502081d7063bb2841df74e05c450d708aaf2
 2015-06-28 11:59:22 +02:00
James Zern
4741fac42e dsp/lossless_*sse2: remove some unnecessary inlines 
TransformColor / TransformColorInverse are the top-level function
pointer calls

Change-Id: Ieabdb4005ff3e4f9bb3ebcb140ccb6bef5d28f8b
 2015-06-25 21:02:01 -07:00
Pascal Massimino
1819965e0a fix warning ("left shift of negative value") using a cast 
Change-Id: Ie99e8ff87924a1d15e2c5d83bd9adf07dab04e94
 2015-06-24 23:46:09 -07:00
Pascal Massimino
7017001462 SSE2: speed-up some lossless-encoding functions 
optimized: CollectColorRedTransforms, CollectColorBlueTransforms, SubtractGreenFromBlueAndRed

overall effect is sub-1% speed-up, though.

Change-Id: I9cb49af5c56e4c03db417929b0a2cf575d60a5c6
 2015-06-24 20:09:13 -07:00
Pascal Massimino
abcb012841 Merge "SSE2: slightly faster (~5%) AddGreenToBlueAndRed()" 2015-06-24 09:37:46 +00:00
Pascal Massimino
2df5bd30a6 Merge "Speedup to HuffmanCostCombinedCount" 2015-06-24 07:42:26 +00:00
Pascal Massimino
9e356d6b25 SSE2: slightly faster (~5%) AddGreenToBlueAndRed() 
Change-Id: Ie147010b66544c4e959f26966ad588394302d418
 2015-06-24 09:36:44 +02:00
Pascal Massimino
fc6c75a2a2 SSE2: 53% faster TransformColor[Inverse] 
Changed the code (again) to process 4 pixels at a time. Loop is more
involved, but overall it's faster.

Removed the SSE4.1 implementation which is now slower than SSE2.

Change-Id: I7734e371033ad8929ace7f7e1373ba930d9bb5f1
 2015-06-23 14:52:01 -07:00
Pascal Massimino
49073da6d6 SSE2: 46% speed-up of TransformColor[Inverse] 
Change-Id: If3bf26dc8ed32a7c03cb438e5d5fc996e2e96b5e
 2015-06-23 20:09:04 +02:00
Pascal Massimino
32462a072c Speedup to HuffmanCostCombinedCount 
~3% speedup for lossless encoding
Improves compression ratio by ~0.03%

Change-Id: Ic6d05fb0b1099b5ca56689b92b1c6515d54a5d6b
 2015-06-23 16:41:03 +02:00
Pascal Massimino
f3d687e3fa SSE4.1 implementation of some lossless encoding functions 
New implementations: SubtractGreenFromBlueAndRed and TransformColor

around 1-2% faster lossless encoding.

Change-Id: I1668e36fdc316ba55b3b798b91b4a3e36ce62861
 2015-06-23 08:46:57 +02:00
Pascal Massimino
bfc300c7ff SSE4.1 implementation of some alpha-processing functions 
DispatchAlpha* functions are hard to speed up, compared to SSE2.
ExtractAlpha sees a ~15% speed-up though.

Change-Id: I8715c2defecbc832f469eed7e6ffd012146b52de
 2015-06-19 14:17:39 -07:00
Pascal Massimino
7f9c98f21d Merge "sse2 in-loop: simplify SignedShift8b() a bit" 2015-06-12 07:37:32 +00:00
James Zern
ef314a5d6c dec_sse2/GetNotHEV: micro optimization 
trade 2 subtractions + logical or for 1 max + 1 subtraction

Change-Id: I7d1f25f7cda2a89bc8247f3d3d5417f6b0e3d96c
 2015-06-11 22:46:24 -07:00
Pascal Massimino
a729cff987 sse2 in-loop: simplify SignedShift8b() a bit 
Change-Id: Ida3e096bb41451194d03dc7a97753a222ff0135c
 2015-06-11 15:26:31 -07:00
Pascal Massimino
422ec9fb62 simplify Load8x4() a bit 
Change-Id: I68cf09c432f48e34bbe1d47dd091417cfd40cf4e
 2015-06-10 12:35:50 -07:00
James Zern
8df238ec8a Merge "remove some duplicate FlipSign()" 2015-06-06 05:25:04 +00:00
Pascal Massimino
751506c484 remove some duplicate FlipSign() 
ApplyFilter2NoFlip is the new variant of ApplyFilter2 without the sign-flip

Change-Id: I2af54bd1499118c8321183e42251d265ba76219c
 2015-06-05 17:20:29 +02:00
James Zern
65ef5afc27 Merge "lossless: 0.13% compression density gain" 2015-06-03 03:02:09 +00:00
Jyrki Alakuijala
2beef2f245 lossless: 0.13% compression density gain 
over a 1000 image corpus

Single photograph benchmark:
Before:
Q=20: 2.560 MP/s
Q=40: 2.593 MP/s
Q=60: 1.795 MP/s
Q=80: 1.603 MP/s
Q=99: 1.122 MP/s

After:
Q=20: 3.334 MP/s
Q=40: 2.464 MP/s
Q=60: 2.009 MP/s
Q=80: 1.871 MP/s
Q=99: 1.163 MP/s

This CL allows for some further improvements that would not be possible
otherwise.

Change-Id: I61ba154beca2266cb96469281cf96e84a4412586
 2015-06-02 17:27:36 -07:00
Pascal Massimino
3033f24c26 lossless: 0.06 % compression density improvement 
Change-Id: Ib662e6aec53b40d6bc736d3ecfd6475bb005c790
 2015-06-02 14:51:51 +02:00
James Zern
64960da9e1 dec_neon: add VE8uv / VE16 
VE8uv/VE16: ~25%/~33% faster over 20M pixels

Change-Id: Ifac1114091527a05ed10edfcc43852edff012d14
 2015-05-30 13:40:00 -07:00
James Zern
14dbd87bed dec_neon: add HE8uv / HE16 
HE8uv/HE16: ~91%/~83% faster over 20M pixels

Change-Id: Ib0a776f7c193593ea0993e92cfa6e6be000fb810
 2015-05-30 13:39:24 -07:00
skal
ac76801159 introduce FTransform2 to perform two transforms at a time. 
FTransform goes from ~12.0% to 11.5% total CPU time.

Change-Id: Ibcb23155324f4fd8b235563f80668531c781f624
 2015-05-18 21:06:15 -07:00
James Zern
aa6065aedd dec_neon: use vld1_dup(mem) rather than vdup(mem[0]) 
should result in slightly less general purpose register use

Change-Id: I6069f49541392e56c8db2c28c8d1fdf88c1a1726
 2015-05-16 11:24:32 -07:00
Pascal Massimino
8b63ac78e0 Merge "dec_neon: add TM16" 2015-05-16 10:56:07 +00:00
Pascal Massimino
f51be09e1f Merge "dec_neon/TrueMotion: simply left border load" 2015-05-16 10:54:05 +00:00
James Zern
dc48196bd9 dec_neon: add TM16 
over 20M pixels ~78% faster

Change-Id: I420d5d590f275f19e08f86df1d1caa6b82fffbde
 2015-05-15 12:50:11 -07:00
James Zern
ea95b305ca dec_neon/TrueMotion: simply left border load 
use vld1_dup_u8() rather than a separate ld+dup after the values were
zero extended; mildly faster at the function level

Change-Id: I1b3666a6aeb465722a1214dbc6d71c27689a7f89
 2015-05-15 12:48:13 -07:00
Pascal Massimino
f262d6120e speed-up SetResidualSSE2 
(was unnecessarily complicated)

Before:
VP8SetResidualCoeffs: checksum = 1127918   elapsed = 475 ms.

Change-Id: Ia54bef86c45f9f474622ff16e594bf1da4f67ebd
After:
VP8SetResidualCoeffs: checksum = 1127918   elapsed = 404 ms.
 2015-05-14 21:24:24 -07:00
James Zern
bf46d0acff fix mips2 build target 
tested with mips1 and mips2; this should cover 3/4 as well.
fixes an ftbfs reported on the debian issue tracker:
https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=785000

Change-Id: I2458487c92bd638589fdfec5adb4f22102a5960c
 2015-05-13 10:36:22 -07:00
James Zern
929a0fdccd enc_sse2/TTransform: simplify abs calculation 
max(b, 0 - b) works as well as (b ^ sign) - b

Change-Id: Iad923236fd70db85ff58a64d3c8e25e4f42a525d
 2015-05-08 19:50:29 -07:00
James Zern
17dbd05819 enc_sse2/CollectHistogram: simplify abs calculation 
max(out, 0 - out) works as well as (out ^ sign) - out

Change-Id: Id820ab9b296512cb0d56c8026b986bf98e3d3909
 2015-05-08 19:49:08 -07:00
James Zern
a6c1593645 dec_neon: add DC16 intra predictors 
improvement over 20M pixels:
DC16: ~77%
DC16NoTop: ~78%
DC16NoLeft: ~83%
DC16NoTopLeft: ~83%

Change-Id: I4c4ee16a8fa0eb466eee45dfa6f6bbce5ce64b99
 2015-05-08 00:12:48 -07:00
James Zern
f274a96ce9 dsp/enc_sse2: add luma4 intra predictors 
VP8EncPredLuma4 improvement over ~20M pixels: ~39%

Change-Id: I9cd841250771276d2d1bef3991215a56e83f7f20
 2015-05-05 23:51:19 -07:00
James Zern
040b11bdf6 dsp/enc_sse2: add chroma intra predictors 
VP8EncPredChroma8 improvements over ~20M pixels
left/top: ~67%
left-only: ~52%
top-only: ~57%
none: ~61%

based on dec_sse2 versions with minor changes to benefit from the linear
storage of the left boundary

Change-Id: Iee7e387fb2570b4eb5af5bfd123e9c2e9ea49c76
 2015-05-05 23:51:14 -07:00
James Zern
aee021bbb1 dsp/enc_sse2: add luma16 intra predictors 
VP8EncPredLuma16 improvements over ~20M pixels
left/top: ~75%
left-only: ~47%
top-only: ~59%
none: ~63%

based on dec_sse2 versions with minor changes to benefit from the linear
storage of the left boundary

Change-Id: I7548be7214fa85c38fd11d30f5b8b271f437657d
 2015-05-05 23:51:07 -07:00
James Zern
4c9af02326 dec_neon: add DC8uvNoTopLeft 
~93% faster

Change-Id: Icf0fd5f85ac53c306a1b69d84275023e5b24a602
 2015-05-01 20:03:57 -07:00
Pascal Massimino
9287761d95 Merge "GetResidualCostSSE2: simplify abs calculation" 2015-04-30 06:30:58 +00:00