Commit graph

692 commits

Author SHA1 Message Date
Pauli Oikkonen e382339182 Implement fast (butterfly) 32x32 DCT in AVX2 2019-10-28 16:19:42 +02:00
Pauli Oikkonen b5962dadac Tidy indentation in AVX2 16x16 iDCT 2019-10-28 16:19:42 +02:00
Pauli Oikkonen 36a8f89025 Fine-tune 16x16 AVX2 iDCT 2019-10-28 16:19:42 +02:00
Pauli Oikkonen ca9409de2b Implement 16x16 DCT as butterfly algorithm in AVX2 2019-10-28 16:19:42 +02:00
Pauli Oikkonen 7c69a26717 Use aligned loads and stores for AVX2 DCT 2019-10-28 16:19:42 +02:00
Pauli Oikkonen 8e9c65dca6 Align DCT matrices and temp transform buffers 2019-10-28 16:19:42 +02:00
Pauli Oikkonen 148a150522 Align DCT source and dest blocks to cache line 2019-10-28 16:19:42 +02:00
Pauli Oikkonen 8e60bbf6a6 Slightly tune 16x16 forward DCT
Use an array of __m256i's to store temporary value, essentially letting
the compiler enforce alignment and use aligned loads and stores.
2019-10-28 16:19:42 +02:00
Pauli Oikkonen c0cc0e8a75 Optimize 16x16 multiply by only slicing right mat once 2019-10-28 16:19:42 +02:00
Pauli Oikkonen e463d27f22 Implement streamlined generic 16x16 matrix multiply
It can't be this fast for real, can it?
2019-10-28 16:19:42 +02:00
Pauli Oikkonen beb85ce9d6 Reorder parameters for 8x8 matrix multiplies 2019-10-28 16:19:42 +02:00
Pauli Oikkonen 292af62256 Implement tailored 16x16 forward DCT 2019-10-28 16:19:42 +02:00
Pauli Oikkonen 30ce461d98 Redo 4x4 matrix multiplication 2019-10-28 16:19:42 +02:00
Pauli Oikkonen 07970ea82f Streamline by-the-book 8x8 matrix multiplication
Also chop up the forward transform into two tailored multiply functions
2019-10-28 16:19:42 +02:00
Pauli Oikkonen 7ec7ab3361 Implement a tailored AVX2 8x8 DCT 2019-10-28 16:19:42 +02:00
pkubaj 1d7fcf4227
Fix build on powerpc64 with LLVM 2019-09-12 15:05:00 +02:00
Pauli Oikkonen 99597b828a Work around the ancient Win32 calling convention hassle
See if this'll work now
2019-09-06 13:14:42 +03:00
Pauli Oikkonen c5ca18950c Revert "Revert to 6924d90052 due to broken visual studio build"
This reverts commit 1dd0619bd7.
2019-09-05 18:21:55 +03:00
Pauli Oikkonen 55529decd5 Implement _mm256_insert_epi32 and extract pseudo-ops
Visual Studio headers apparently lack these guys
2019-09-05 18:20:52 +03:00
Ari Lemmetti 557bcbc6aa Make luma or chroma only inter "recon" or predict possible 2019-09-02 17:15:28 +03:00
RLamm 60be6d411c Intra filtering fixed at least for luma. All intra modes output valid luma (hashes match), but chroma is still broken. 2019-08-30 16:14:00 +03:00
Marko Viitanen cb0d7c340a Use the new PDPC filtering in angular intra 2019-08-23 14:44:41 +03:00
Marko Viitanen 5bebb18943 Change intra filtering according to VTM6 2019-08-23 08:56:35 +03:00
Marko Viitanen a16efe6b52 Merge remote-tracking branch 'remotes/github_kvazaar/master'
# Conflicts:
#	build/kvazaar_VS2013.sln
#	build/kvazaar_VS2015.sln
#	build/kvazaar_VS2017.sln
#	build/kvazaar_cli/kvazaar_cli.vcxproj
#	build/kvazaar_lib/kvazaar_lib.vcxproj
#	build/kvazaar_tests/kvazaar_tests.vcxproj
#	src/encode_coding_tree.c
#	src/encode_coding_tree.h
#	src/encoder_state-bitstream.c
#	src/inter.c
#	src/strategies/avx2/quant-avx2.c
2019-08-22 15:12:01 +03:00
Ari Lemmetti 1dd0619bd7 Revert to 6924d90052 due to broken visual studio build 2019-08-08 15:15:34 +03:00
Pauli Oikkonen 2852baa673 Separate sign3_diff_epu8 from calc_eo_cat
Just to keep things simple, clear and obvious
2019-08-07 16:35:24 +03:00
Pauli Oikkonen a858e7dd4b Combine duplicate code into inline functions 2019-08-07 16:35:24 +03:00
Pauli Oikkonen de0e97f711 Take 8/16/24b loads and stores into separate functions 2019-08-07 16:35:24 +03:00
Pauli Oikkonen 10979f58fe Tidy up code 2019-08-07 16:35:24 +03:00
Pauli Oikkonen 9cc11976c0 Combine the delta accumulation from edge and band ddistortion into shared func
This won't reduce object size, but there'll be less duplicate code
2019-08-07 16:35:24 +03:00
Pauli Oikkonen 55d877bd66 Vectorize sao_edge_ddistortion 2019-08-07 16:35:24 +03:00
Pauli Oikkonen aef0f301d3 Fix function signatures
Mark anything intended as read-only to be const, and fix alignment
2019-08-07 16:35:24 +03:00
Pauli Oikkonen 997fd369b3 Redo calc_sao_edge_dir_avx2
Do it wider, 32 pixels at once!
2019-08-07 16:35:24 +03:00
Pauli Oikkonen db1e475e02 Use i32 instead of i8 for x/y offsets
Doesn't matter too much, because this number isn't used in SIMD
computation, only as a memory reference offset.
2019-08-07 16:35:24 +03:00
Pauli Oikkonen 12de466ef5 Reimplement non-band SAO color reconstruction in AVX2
Streamline things to work on 32 pixels at once instead of 8
2019-08-07 16:35:24 +03:00
Pauli Oikkonen e8bff99329 Redo the SAO_TYPE_BAND subsection of AVX2 SAO color reconstruction
Vectorize it all, hope this helps with perf
2019-08-07 16:35:24 +03:00
Pauli Oikkonen 7b5dffa855 Implement calc_sao_offset_array in AVX2
To be efficient, the AVX2 color reconstruction algorithm will need
offsets in byte, not dword, arrays. This is completely specific to 8-bit
pixels and the function signature is fundamentally distinct from the
generic algorithm, so it's better to not strategize SAO offset array
calculation.
2019-08-07 16:35:24 +03:00
Pauli Oikkonen 08881f5e9b (TEMP) (TODO) (whatever) Avoid compiler warnings
I want the CI to not crash on its -Wall -Werror, but instead to actually
build the thing and report me about actual memory errors etc
2019-08-07 16:35:24 +03:00
Pauli Oikkonen c18adc5ee0 Redo sao_band_ddistortion_avx2
Avoid branching and do the entire thing on 32 pixels at once in YMMs.
Also make the sao_bands function parameter const.
2019-08-07 16:35:24 +03:00
Pauli Oikkonen 1bb9a079a8 Fix indentation 2019-08-07 16:35:24 +03:00
Reima Hyvönen 7bc959c7c5 3 sao functions are now working 2019-08-07 16:35:24 +03:00
Reima Hyvönen 0e0f2d3490 made to clear sum vector after it has been set to memory 2019-08-07 16:35:24 +03:00
Reima Hyvönen f146de7acb removed some variables to prevent memory losses 2019-08-07 16:35:24 +03:00
Reima Hyvönen 247c3a7a71 conversed gined to unsigned int 2019-08-07 16:35:24 +03:00
Reima Hyvönen ac5c216974 Some more memory error preventing to sao_edge_ddistortion_avx2 2019-08-07 16:35:24 +03:00
Reima Hyvönen 3fb1cbca35 more editing sao_edge_ddistortion_avx2 2019-08-07 16:35:24 +03:00
Reima Hyvönen afbb6fb960 some more modifications to sao_edge_ddistortion_avx2 to prevent memory failures 2019-08-07 16:35:24 +03:00
Reima Hyvönen 3496a57f7a Edited sao_edge_ddistortion_avx2 to avoid memory overflow 2019-08-07 16:35:24 +03:00
Reima Hyvönen 267ba1d6ce Modified sao_band_ddistortion_avx2 2019-08-07 16:35:24 +03:00
Reima Hyvönen e70663b245 added some sub commands to avoid memory read errors 2019-08-07 16:35:24 +03:00
Reima Hyvönen 59dfb4570c Converted some loads to load int8_t instead ints 2019-08-07 16:35:24 +03:00
Reima Hyvönen 8b253209a8 Found false address load from calc_sao_edge_dir. Should now work like generic 2019-08-07 16:35:24 +03:00
Reima Hyvönen 50e0a47b7a Took away __restrict 2019-08-07 16:35:24 +03:00
Reima Hyvönen 8a39eb674e Removed c-variable from calc_sao_edge_dir_avx2 2019-08-07 16:35:24 +03:00
Reima Hyvönen bc0a36830d Clerified some 6 pixel loads 2019-08-07 16:35:24 +03:00
Reima Hyvönen 1a8b211e05 Added break to line 170 2019-08-07 16:35:24 +03:00
Reima Hyvönen d05e750ebe Added some switches to prevent segmentation fault from reading 2019-08-07 16:35:24 +03:00
Reima Hyvönen 203580047d Defined some AVX functions 2019-08-07 16:35:24 +03:00
Reima Hyvönen c884c738b1 Updated some commands to match the standard 2019-08-07 16:35:24 +03:00
Reima Hyvönen b412ed2f59 Removed some setr and used loads calc_sao_edge_dir_avx2 2019-08-07 16:35:24 +03:00
Reima Hyvönen c6cc063534 converted some hadd operations at calc_sao_edge_dir_avx2 to cast and extract 2019-08-07 16:35:24 +03:00
Reima Hyvönen 47ac109b10 optimated some sao_reconstruct_color_avx2 when sao->type == SAO_TYPE_BAND 2019-08-07 16:35:24 +03:00
Reima Hyvönen 96dc60a1ed first working optimation 2019-08-07 16:35:24 +03:00
Reima Hyvönen c148aff9fb Some optimation done to function sao_reconstruct_color_avx2 2019-08-07 16:35:24 +03:00
Reima Hyvönen bf16ba6cc4 Remade sao_edge_ddistortion_avx2 and calc_sao_edge_dir_avx2 2019-08-07 16:35:24 +03:00
Reima Hyvönen 79dc39a676 Some editing for sao_edge_ddistortion_avx2 2019-08-07 16:35:24 +03:00
Reima Hyvönen 06ee52924e some reconst done to calc_sao_edge_dir_avx2 2019-08-07 16:35:24 +03:00
Reima Hyvönen 5fbc65d823 reconst optimation doesn't work yet 2019-08-07 16:35:24 +03:00
Reima Hyvönen d29f834a69 Remove useless function 2019-08-07 16:35:24 +03:00
Reima Hyvönen a232a12160 calc_sao_edge_dir_avx2 updated 2019-08-07 16:35:24 +03:00
Reima Hyvönen b1febc02a5 sao_edge_ddistortion_avx2 now working proberly 2019-08-07 16:35:24 +03:00
Reima Hyvönen cd6092a1ec Still too much bits, looking for where they appear 2019-08-07 16:35:24 +03:00
Reima Hyvönen 7853be8eeb Incomple optimation 2019-08-07 16:35:24 +03:00
Marko Viitanen dfa5621024 Intrapred cleanup 2019-07-16 14:23:10 +03:00
Pauli Oikkonen 8d48bee180 Tidy fast coeff cost code 2019-07-09 18:01:54 +03:00
Pauli Oikkonen 201a43b08e Clean up the RD-estimation code 2019-07-09 18:01:54 +03:00
Pauli Oikkonen b111df5073 Create preliminary version of improved cost estimator 2019-07-09 18:01:54 +03:00
Marko Viitanen 10d850e98a Use index_offset in intra angular and change the offset to width+1 2019-07-08 14:23:19 +03:00
Marko Viitanen 3d1fa2a9cf Fixing angular intra prediction reference pixels 2019-07-08 14:00:02 +03:00
Marko Viitanen 0656c54cab Fix some problems with reference pixels in angular intra prediction kvz_angular_pred_generic() 2019-07-05 15:54:51 +03:00
Marko Viitanen 89ca2d4ba1 Use correct type for modedisp2sampledisp array 2019-07-05 14:12:10 +03:00
Marko Viitanen c6217e236f Enable 4-tap filtering for the intra angular 2019-07-04 16:26:10 +03:00
Marko Viitanen cda6d951c0 Change DCT arrays back to 8-bit -> some frames are now correct 2019-07-04 15:59:10 +03:00
Marko Viitanen 8280bd3217 Add channel info to angular_pred and fix the displacement tables.
Also includes 4-tap intra filtering code commented out
2019-07-04 09:35:47 +03:00
Pauli Oikkonen 081d16fc33 Fix intrinsics that may be missing on some systems
Create a header to collect all the workarounds for missing intrinsics
in one place
2019-05-23 19:59:40 +03:00
Marko Viitanen 30a8a7b97c WIP fixing the last significant xy coding 2019-05-07 15:01:02 +03:00
Pauli Oikkonen 87a9208db8 Eliminate cvtsi64_si128 intrinsic
Apparently it'll cause Win32 builds to break because it emits the movq
instruction or something..
2019-04-17 16:30:40 +03:00
Pauli Oikkonen 7175d20bb2 Still include stdint.h for non-vector builds 2019-04-15 19:36:01 +03:00
Pauli Oikkonen 1315c7e2b0 Do not compile any vector code for non-SSE4/AVX2 builds 2019-04-15 19:10:48 +03:00
Pauli Oikkonen f5f70e7bc5 Merge branch 'sad-optimization' 2019-04-15 19:02:01 +03:00
Pauli Oikkonen 6d43759604 Create a border-respecting 32-wide AVX hor_sad 2019-03-07 18:01:22 +02:00
Pauli Oikkonen f218cecb38 Remove offending hor_sad_avx2_w32 function
Consider possibly creating a non-offending AVX2 version instead, the
way hor_sad_sse41_w32 works. Or maybe there's more essential work to
do.
2019-03-05 22:51:41 +02:00
Pauli Oikkonen df2e6c54fd 4-unroll hor_sad_sse41_arbitrary
This may not increase perf though because it's so rarely used
function, so keeping icache footprint may be more essential...
2019-03-05 22:45:23 +02:00
Pauli Oikkonen 448eacba7b Avoid overreading block borders in hor_sad_sse41_arbitrary 2019-03-05 22:34:50 +02:00
Pauli Oikkonen 41f51c08c4 Avoid overrunning buffer in hor_sad_sse41_w32 2019-03-01 15:37:38 +02:00
Pauli Oikkonen bcd9879359 Include quant coeff range check in non-scaling list execution path too 2019-02-27 17:26:44 +02:00
Pauli Oikkonen 24e6363f64 Remove the kvz_quant_avx2 wrapper function 2019-02-27 16:32:58 +02:00
Pauli Oikkonen 748820f3c5 Eliminate unnecessary loading of coeffs if scaling lists are off 2019-02-27 16:26:35 +02:00
Pauli Oikkonen 5994350f40 Allow quant_flat_avx2 to be used with scaling lists on 2019-02-27 16:25:59 +02:00
Pauli Oikkonen 9b0e079262 Use SSE instructions for 64-bit SADs instead of MMX
VC++ seems to choke on MMX instructions
2019-02-18 20:13:33 +02:00
Pauli Oikkonen d8b8923028 Add LGPL notices to reg_sad headers 2019-02-18 17:52:47 +02:00
Pauli Oikkonen 770db825b9 Create hor_sad_w8 and w4 epol mask the way w16 works 2019-02-06 19:34:26 +02:00
Pauli Oikkonen aa19bcac8a Avoid branching in creating shuffle mask in hor_sad_w16 2019-02-06 18:58:46 +02:00
Pauli Oikkonen 2d05ca8520 Remove width from constant-width hor_sad func params
They should kinda know it already
2019-02-04 20:41:40 +02:00
Pauli Oikkonen 57db234d95 Move 32-wide SSE4.1 hor_sad to picture-sse41.c
It's not used by picture-avx2.c that also includes the header, so
it should not be in the header
2019-02-04 20:41:40 +02:00
Pauli Oikkonen dd7d989a39 Implement 32-wide hor_sad on AVX2 2019-02-04 20:41:40 +02:00
Pauli Oikkonen ff70c8a5ec Utilize horizontal SAD functions for SSE4.1 as well 2019-02-04 20:41:40 +02:00
Pauli Oikkonen f5ff4db01f 4-wide hor_sad border agnostic 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 35e7f9a700 Fix hor_sad w8 to work with both borders 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 836783dd6e Use hor_sad_w32 for both left and right borders 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 69687c8d24 Modify hor_sad_sse41_w16 to work over left and right borders 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 1e0eb1af30 Add generic strategy for hor_sad'ing an non-split width block 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 686fb2c957 Unroll arbitrary-width SSE4.1 hor_sad by 4 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 768203a2de First version of arbitrary-width SSE4.1 hor_sad 2019-02-04 20:41:40 +02:00
Pauli Oikkonen ccf683b9b6 Start work on left and right border aware hor_sad
Comes with 4, 8, 16 and 32 pixel wide implementations now, at some point
investigate if this can start to thrash icache
2019-02-04 20:41:40 +02:00
Pauli Oikkonen c36482a11a Fix bug in 24-wide SAD
*facepalm*
2019-02-04 20:41:40 +02:00
Pauli Oikkonen f781dc31f0 Create strategy for ver_sad
Easy to vectorize
2019-02-04 20:41:40 +02:00
Pauli Oikkonen 9db0a1bcda Create get_optimized_sad func for SSE4.1 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 91380729b1 Add generic get_optimized_sad implementation
NOTE: To force generic SAD implementation on devices supporting
vectorized variants, you now have to override both get_optimized_sad
and reg_sad to generic (only overriding get_optimized_sad on AVX2
hardware would just run all SAD blocks through reg_sad_avx2). Let's
see if there's a more sensible way to do it, but it's not trivial.
2019-02-04 20:41:40 +02:00
Pauli Oikkonen 45f36645a6 Move choosing of tailored SAD function higher up the calling chain 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 91cb0fbd45 Create strategy for directly obtaining pointer to constant-width SAD function 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 94035be342 Unify unrolling naming conventions 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 517a4338f6 Unroll SSE SAD for 8-wide blocks to process 4 lines at once 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 0f665b28f6 Unroll arbitrary width SSE4.1 SAD by 4 2019-02-04 20:41:40 +02:00
Pauli Oikkonen cbca3347b5 Unroll 64-wide AVX2 SAD by 2 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 84cf771dea Unroll 32 and 16 wide SAD vector implementations by 4 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 5df5c5f8a4 Cast all pointers to const types in vector SAD funcs
Also tidy up the pointer arithmetic
2019-02-04 20:41:40 +02:00
Pauli Oikkonen a711ce3df5 Inline fixed width vectorized SAD functions 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 6504145cce Remove 16-pixel wide AVX2 SAD implementation
At least on Skylake, it's noticeably slower than the very simple
version using SSE4.1
2019-02-04 20:41:40 +02:00
Pauli Oikkonen 4cb371184b Add SSE4.1 strategy for 24px wide SAD and an AVX2 strategy for 16 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 796568d9cc Add SSE4.1 strategies for SAD on widths 4 and 12 and AVX2 strategies for 32 and 64 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 4d45d828fa Use constant-width SSE4.1 SAD funcs for AVX2 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 2eaa7bc9d2 Move SSE4.1 SAD functions to separate header 2019-02-04 20:41:40 +02:00
Pauli Oikkonen d2db0086e1 Create constant width SAD versions for 8 and 16 pixels 2019-02-04 20:41:40 +02:00
Pauli Oikkonen a13fc51003 Include a blank AVX2 strategy registration function even in non-AVX2 builds 2019-02-04 19:52:24 +02:00
Pauli Oikkonen d55414db66 Only build AVX2 coeff encoding when supported
..whoops
2019-02-04 19:34:30 +02:00
Pauli Oikkonen 3fe2f29456 Merge branch 'encode-coeffs-avx2' 2019-02-04 18:52:31 +02:00
Pauli Oikkonen 722b738888 Fix more naming issues 2019-02-04 16:05:43 +02:00
Pauli Oikkonen e26d98fb75 Rename a couple variables and add crucial comments 2019-02-04 15:57:07 +02:00
Pauli Oikkonen f186455619 Move encode_last_significant_xy out of strategy modules
It's the exact same in both AVX2 and generic, and does not seem to
be worth even trying to vectorize
2019-02-04 14:55:41 +02:00
Pauli Oikkonen 3f7340c932 Fine-tune pack_16x16b_to_16x2b
Avoid mm_set1 operation when it's possible to create the constant with
one bit-shift operation from another instead. Thanks Intel for
3-operand instruction encoding!
2019-02-04 14:44:47 +02:00
Pauli Oikkonen 314f5b0e1f Rename 16x2b cmpgt function, comment it better, optimize it slightly
Eliminate an unnecessary bit masking to make it even more messy
2019-02-04 14:44:32 +02:00
Pauli Oikkonen d8ff6a6459 Fix _andn_u32 to work on old Visual Studio 2019-02-01 15:34:42 +02:00
Pauli Oikkonen 26e1b2c783 Use (u)int32_t instead of (unsigned) int in reg_sad_sse41 2019-01-10 14:37:04 +02:00
Pauli Oikkonen 3a1f2eb752 Prefer SSE4.1 implementation of SAD over AVX2
It seems that the 128-bit wide version consistently outperforms the
256-bit one
2019-01-10 13:48:55 +02:00
Pauli Oikkonen 9b24d81c6a Use SSE instead of AVX for small widths
Highly dubious if this will help performance at all
2019-01-07 20:12:13 +02:00
Pauli Oikkonen b2176bf72a Optimize SSE4.1 version of SAD
Make it use the same vblend trick as AVX2. Interestingly, on my test
setup this seems to be faster than the same code using 256-bit AVX
vectors.
2019-01-07 19:40:57 +02:00
Pauli Oikkonen 887d7700a8 Modify AVX2 SAD to mask data by byte granularity in AVX registers
Avoids using any SAD calculations narrower than 256 bits, and
simplifies the code. Also improves execution speed
2019-01-07 18:53:15 +02:00
Pauli Oikkonen 7585f79a71 AVX2-ize SAD calculation
Performance is no better than SSE though
2019-01-07 16:26:24 +02:00
Pauli Oikkonen ab3dc58df6 Copy SAD SSE4.1 impl to AVX2 2019-01-03 18:31:57 +02:00
Pauli Oikkonen 45ac6e6d03 Tidy pack_16x16b_to_16x2b comments 2019-01-03 16:37:05 +02:00
Pauli Oikkonen 016eb014ad Move packing 16x16b -> 16x2b into separate function 2018-12-20 10:51:44 +02:00
Ari Lemmetti b234897e8a Fix smp and amp blocks in fme and revert previous change.
Filter 8x8 (sub)blocks even with 8x4, 4x8, 16x4, 4x16 etc.
Calculate SATD on the 8x4, ... part
2018-12-19 21:30:53 +02:00
Pauli Oikkonen 9aaa6f260d Fixes to enable portability 2018-12-18 20:42:09 +02:00
Pauli Oikkonen 2fdbbe9730 Move CG reordering code from quant-avx2 to shared header 2018-12-18 19:42:18 +02:00
Pauli Oikkonen d02207306d Create a header file for shared AVX2 code 2018-12-18 19:41:09 +02:00
Pauli Oikkonen 361bf0c7db Precompute >=2 coeff encoding loop with 2-bit arithmetic
Who needs 16x16b vectors when you can do practically the same with
16x2b pseudovectors in 32-bit general purpose registers!
2018-12-18 19:41:09 +02:00
Pauli Oikkonen f66cb23d5b Optimize greater1 encoding loop
Calculating the c1 variable need not be a serial operation!
2018-12-18 19:41:09 +02:00
Pauli Oikkonen 8c8b791c35 Vectorize kvz_context_get_sig_ctx_inc 2018-12-18 19:41:09 +02:00
Pauli Oikkonen 033261eb74 Eliminate two branches using bit magic 2018-12-18 19:41:09 +02:00
Pauli Oikkonen c4434e8d04 Scan CG's in forward order to simplify finding last significant 2018-12-18 19:41:09 +02:00
Pauli Oikkonen efd097f5a5 Vectorize the coeff group loop to some extent 2018-12-18 19:41:09 +02:00
Pauli Oikkonen a01362e638 use the efficient method of reordering raster->scan 2018-12-18 19:41:09 +02:00
Pauli Oikkonen 50a888e789 Use the efficient method to find first and last nz coeffs in block 2018-12-18 19:41:09 +02:00
Pauli Oikkonen 7e9203f566 Scan coeff groups in scan order to help find last significant one 2018-12-18 19:41:09 +02:00
Pauli Oikkonen 9a5a6fdbc7 Simplify two ifs in encode_coeff_nxn-avx2 2018-12-18 19:41:09 +02:00
Pauli Oikkonen 37a2a8bac8 See if loop can be optimized by rearranging 2018-12-18 19:41:09 +02:00
Pauli Oikkonen 584f2f74b6 Vectorize significant coeff group scanning loop 2018-12-18 19:41:09 +02:00
Pauli Oikkonen 1bfed73221 Add AVX2 strategy for encode_coding_tree 2018-12-18 19:41:09 +02:00
Pauli Oikkonen c3a6f3112a Add generic strategy group for encode_coding_tree 2018-12-18 19:41:09 +02:00
Sergei Trofimovich 68a70e45a1 x86 asm: mark stack as non-executable
Gentoo's `scanelf` QA tool detects writable/executable stack
of assembly-writtent files as:

```
$ scanelf -qRa  .
 0644 LE !WX --- ---     ./src/strategies/x86_asm/.libs/picture-x86-asm-sad.o
 0644 LE !WX --- ---     ./src/strategies/x86_asm/.libs/picture-x86-asm-satd.o
 0644 LE !WX --- ---     ./src/strategies/x86_asm/picture-x86-asm-sad.o
 0644 LE !WX --- ---     ./src/strategies/x86_asm/picture-x86-asm-satd.o
```

Normally C compiler emits non-executable stack marking (or GNU assembler
via `-Wa,--noexecstack`).

The change adds non-executable stack marking for yasm-based assmbly files.

https://wiki.gentoo.org/wiki/Hardened/GNU_stack_quickstart has more details.

Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org>
2018-12-16 11:31:56 +00:00
Reima Hyvönen 1fcc5c6a8d Merge branch 'bipred_recon' 2018-12-11 09:59:35 +02:00
Reima Hyvönen e4a10880f3 Added case 12 to bipred_recon no mov 2018-12-11 09:52:17 +02:00
Marko Viitanen a4f3968e52 Fix Visual Studio errors by initializing some variables used in AVX2 signhiding 2018-12-11 09:33:26 +02:00
Pauli Oikkonen c465578048 Add a descriptive comment to coefficient reordering 2018-12-03 15:36:32 +02:00
Pauli Oikkonen f78bf2ebcb Optimize q_coefs usage for indexed fetch 2018-12-03 15:36:32 +02:00
Pauli Oikkonen d9591f1b49 Eliminate midway buffering of reordered coefs
TODO: For some mysterious reason seems slightly slower than the
buffered one
2018-12-03 15:36:32 +02:00
Pauli Oikkonen 7fe454c51f Optimize get_cheapest_alternative() 2018-12-03 15:36:32 +02:00
Pauli Oikkonen 6bbd3e5a44 Optimize rearrange_512 function 2018-12-03 15:36:32 +02:00
Pauli Oikkonen cb8209d1b3 Vectorize transform coefficient reordering loop 2018-12-03 15:36:32 +02:00
Pauli Oikkonen 7cf4c7ae5f Rename "reduce" functions to hsum
That's what the functions fundamendally do anyway
2018-12-03 15:36:32 +02:00
Pauli Oikkonen 316cd8a846 Fix ALIGNED keyword and grow alignment to 64B 2018-12-03 15:36:32 +02:00
Pauli Oikkonen 1befc69a4c Implement sign bit hiding in AVX2 2018-12-03 15:36:32 +02:00
Reima Hyvönen f8696b54a4 Updated bipred_recon_avx2 in avx2/picture-avx2.c. Now it detects blocks that can be not equal to 8 (ie. width = 12) 2018-11-20 17:09:19 +02:00
Reima Hyvönen 710ba288db Chroma has some problems 2018-11-15 16:42:48 +02:00
Ari Lemmetti a832206bb6 Replace 32-bit incompatible instrinsics 2018-11-12 18:54:33 +02:00
Ari Lemmetti 5c774c4105 Rewrite most of FME and interpolation filters
Changes had to break a lot of stuff and were just squashed into this horrible code dump
2018-11-08 20:21:16 +02:00
Reima Hyvönen 7406c33a42 Some more cleaning 2018-10-26 12:25:18 +03:00
Reima Hyvönen 4c71546b2e Cleaned some coding 2018-10-26 12:19:44 +03:00
Reima Hyvönen 4fe3909e48 Switched luma to use 32bits size ints intstead of 16bit size 2018-10-24 18:24:46 +03:00
Marko Viitanen 465bc2cfee [EMT] make functions static and prefix arrays with kvz_g 2018-10-18 10:54:33 +03:00
Marko Viitanen 169febd1c4 [EMT] Simplify DCT8, DCT5, DST1 and DST7 definitions 2018-10-17 12:17:54 +03:00
Marko Viitanen e015d7eb2b Fix compiler warnings 2018-10-17 10:43:11 +03:00
Marko Viitanen ad310c77d3 Added EMT transforms to the strategies 2018-10-17 08:56:49 +03:00
Reima Hyvönen 381e786e10 Trying to find the bug in luma 2018-10-11 18:08:41 +03:00
Reima Hyvönen 2f5f81bac3 removed the non-optimated bipred function 2018-10-09 11:19:23 +03:00
Reima Hyvönen 212a8e68fa Modified to avoid memory overflow, still some bug inside luma 2018-10-02 20:23:32 +03:00
Marko Viitanen 389aeebe07 Added 2x2 transform functions 2018-09-13 14:51:07 +03:00
Marko Viitanen 445c059b4a Fix transforms for VTM 2.0, generated new transform matrices and added a shift by 2 for forward and inverse 2018-09-13 14:39:49 +03:00
Marko Viitanen 382917bcd3 New table for choosing angular intra filtered references and a small bugfix on the end condition of angular intra 2018-09-13 09:35:55 +03:00
Marko Viitanen d4ed0ee3ad Fixed some array offsets in intra angular prediction 2018-09-12 08:53:17 +03:00
Sami Ahovainio 787264f568 Fixed dst indexing in kvz_angular_pred_generic 2018-08-31 10:36:28 +03:00
Sami Ahovainio d2291fea83 Intra mode scaling moved from angular prediction to kvz_intra_predict. pdpc implemented in kvz_intra_predict. 2018-08-31 10:01:28 +03:00
Sami Ahovainio 54ebadfc43 Clarifying comments and changes towards WAIP 2018-08-29 16:00:08 +03:00
Reima Hyvönen 896034b7cf Some renamed functions back 2018-08-28 15:31:10 +03:00
Reima Hyvönen e8b5e6db4c Did some merging 2018-08-28 15:26:27 +03:00
Reima Hyvönen 7de5c74434 Updated bipred_recon to work faster 2018-08-28 15:12:31 +03:00
Reima Hyvönen 47b357cca2 Comment one test 2018-08-27 18:52:14 +03:00
Reima Hyvönen 2ca99a44e8 Updated shuffle operation to be in right order 2018-08-27 18:16:38 +03:00
Sami Ahovainio 42741a2c40 Some changes for PCM and Intra towards VTM 2.0 compatibility. 2018-08-27 09:18:15 +03:00
Marko Viitanen 4f7da86285 Commented out sign hiding code, which is not used in VVC 2018-08-17 09:38:11 +03:00
Marko Viitanen c9cbdd5dc3 Added couple of ToDo comments for large CTU support 2018-08-17 09:37:14 +03:00
Marko Viitanen daf041406f Disable DST 2018-08-16 16:05:32 +03:00
Reima Hyvönen 508b218a12 some modifications made to prevent reading too much 2018-08-14 10:50:39 +03:00
Reima Hyvönen 1d935ee888 some useless stuff removed 2018-08-13 16:47:11 +03:00
Reima Hyvönen ce3ac4c05e some modifications to no_mov 2018-08-13 16:41:02 +03:00
Reima Hyvönen 15a613ae94 test if no_mov breaks testing 2018-08-13 16:02:56 +03:00
Reima Hyvönen 97a2049e58 removed pointer declaration out from switch 2018-08-10 16:42:26 +03:00
Reima Hyvönen aa94bcedbc Stream is now pointer 2018-08-10 16:38:49 +03:00
Reima Hyvönen fa5b227ece 256 to 32 doesn't work, made them by hand 2018-08-10 16:01:20 +03:00
Reima Hyvönen 408dedbcc8 removed _mm256_extract_epi8 and replaced with _mm_stream 2018-08-10 15:53:26 +03:00
Reima Hyvönen 31c35091c6 _mm256_cvtsi256_si32 removed 2018-08-10 10:06:40 +03:00
Reima Hyvönen 99dc43074f _mm256_cvtsi256_si32 breaks system, too much bits. back to extract 2018-08-10 09:59:33 +03:00
Reima Hyvönen 4f1f80b2cb Transformed convert from 256 to cast 256 -> 128 and then convert from 128 2018-08-09 15:35:54 +03:00
Reima Hyvönen 4957555eb3 Removed leftover from 939 2018-08-09 15:25:03 +03:00
Reima Hyvönen 28b165c971 Clearified some sections, added _MM_SHUFFLE macro 2018-08-09 15:23:01 +03:00
Reima Hyvönen dd04df8667 testing if error in both avx2 functions 2018-08-03 11:49:00 +03:00
Reima Hyvönen ed50d71fde Switched some variables to different location, altered inter_recon_bipred_avx2 function 2018-08-02 16:08:59 +03:00
Reima Hyvönen f5739a0028 Renaming and removing useless prints 2018-08-02 14:47:17 +03:00
Reima Hyvönen bc09f59bb6 Edited some definitions 2018-08-02 11:54:53 +03:00
Reima Hyvönen a4bf77f208 Tested some extract functions 2018-07-12 09:29:32 +03:00
Reima Hyvönen c05033a893 Even more useless vectors removed 2018-07-11 15:09:14 +03:00
Reima Hyvönen 884cb77238 Removed some not used vectors 2018-07-11 15:06:11 +03:00
Reima Hyvönen 792689a5ff Removed for-loops, added extract instead 2018-07-11 14:56:41 +03:00
Reima Hyvönen f9c7f6ee66 Added some break-operations for avx2 optimation 2018-07-11 14:15:38 +03:00
Reima Hyvönen cc064da143 some more optimation for bipred 2018-07-11 11:27:54 +03:00
Reima Hyvönen a22cf03ddb Updated to have no movement function to avx2 strategies 2018-07-10 16:07:15 +03:00
Reima Hyvönen ea83ae45f0 Toimiva ratkaisu 2018-07-03 11:18:51 +03:00
Reima Hyvönen 17babfffa4 25.6 working optimation, ~50% faster than original 2018-06-25 17:06:16 +03:00
Reima Hyvönen 9fed29f950 optimation for inter_recon_bipred 2018-04-18 15:25:44 +03:00
Arttu Ylä-Outinen 0a69e6d18f Fix selection of transform function for 4x4 blocks
DST function was returned for inter luma transform blocks of size 4x4
even though they must use DCT. Fixed by checking the prediction mode of
the block in addition to whether it is chroma or luma.
2018-01-18 10:36:25 +02:00
Arttu Ylä-Outinen 9694bd2fae Fix build on 32-bit systems
Function coeff_abs_sum_avx2 that was added in e950c9b was outside the
AVX2 #if directive.
2017-07-28 09:19:29 +03:00
Arttu Ylä-Outinen e950c9b101 Add AVX2 implementation for coefficient sum 2017-07-28 07:39:36 +03:00
Arttu Ylä-Outinen d50ae6990c Add sum of absolute coefficients to strategies 2017-07-28 07:39:15 +03:00
Arttu Ylä-Outinen fdb3480b54 Enable strategies for SAO reconstruction
Re-enables strategies for SAO reconstruction. They were disabled in
commit ec9ff42.
2017-07-11 10:35:18 +03:00
Arttu Ylä-Outinen 333dba3884 Add static to SAO strategies 2017-07-11 10:02:01 +03:00
Arttu Ylä-Outinen 563bc26e71 Fix out-of-bounds read in AVX2 SAO
AVX2 version of SAO loaded offsets with a 256 bit read even though there
are only five 32 bit integers.
2017-07-06 13:04:52 +03:00
Arttu Ylä-Outinen 2c66e0bbd2 Fix warnings about invalid reads in AVX2 ipol
AVX2 filter functions read pixels in chunks of 8 or 16 bytes. At the end
of the block, the read goes out of the bounds of the pixels array. The
extra pixels do not affect the result.

Fixes valgrind complaining about the invalid reads by allocating 5 extra
pixels in kvz_get_extended_block_avx2
2017-06-22 09:37:55 +03:00
Arttu Ylä-Outinen 95775a1645 Change coefficient storage order
Changes coefficient storage order to a zig-zag order. Reduces
unnecessary copying of coefficients to temporary arrays.
2017-05-12 16:46:57 +03:00
Arttu Ylä-Outinen 51786eda67 Drop redundant fields in encoder_control_t
Some of the fields in encoder_control_t were simply copies of the
corresponding fields in kvz_config. This commit drops the copied fields
in favor of using the fields in encoder_control_t.cfg directly.
2017-02-09 14:05:28 +09:00
Arttu Ylä-Outinen e78a8dfcf5 Copy the kvz_config passed to encoder_open
The kvz_config struct is created by the user but kvazaar keeps a pointer
to it. It is easy to break things by modifying the configuration outside
kvazaar. In addition, kvazaar modifies the struct even though it is has
a const modifier.

This commit changes the field cfg in encoder_control_t to be a copy of
the kvz_config struct instead of a pointer, removing modifications to
the const struct and allowing users to do whatever they want with it
after opening the encoder.
2017-02-09 13:23:54 +09:00
Arttu Ylä-Outinen 640ff94ecd Use separate lambda and QP for each LCU
Adds fields lambda, lambda_sqrt and qp to encoder_state_t. Drops field
cur_lambda_cost_sqrt from encoder_state_config_frame_t and renames
cur_lambda_cost to lambda.
2017-01-09 01:24:23 +09:00
Ari Lemmetti 70a52f0e48 10-bit: add missing bit depth adjustment to ssd 2016-11-17 19:28:04 +02:00
Ari Lemmetti 29153ed503 Remove unused variable 2016-10-21 17:28:42 +03:00
Ari Lemmetti 778e46dfd8 Add AVX2 version of SSD 2016-10-21 15:07:53 +03:00
Ari Lemmetti 6f5d7c9e06 Move SSD to strategies 2016-10-21 15:07:23 +03:00
Ari Lemmetti 89b941eab4 Fix typo 2016-10-21 15:07:02 +03:00
Ari Koivula cbfa824d1a Merge branch 'simd' 2016-09-27 20:49:45 +03:00
Ari Koivula 14a7bcba25 Use a faster function for clipped inter SAD
Use the vectorized general SSE41 inter SAD in AVX reg_sad for shapes
for which we don't have AVX versions yet.

Also improves speed of --smp and --amp a lot. Got a 1.25x speedup for:
--preset=ultrafast -q 27 --gop=lp-g4d3r3t1 --me-early-termination=on --rd=1 --pu-depth-inter=1-3 --smp --amp

* Suite speed_tests:
-PASS inter_sad: 0.898M x reg_sad(64x63):x86_asm_avx (1000 ticks, 1.000 sec)
+PASS inter_sad: 2.503M x reg_sad(64x63):x86_asm_avx (1000 ticks, 1.000 sec)
-PASS inter_sad: 115.054M x reg_sad(1x1):x86_asm_avx (1000 ticks, 1.000 sec)
+PASS inter_sad: 133.577M x reg_sad(1x1):x86_asm_avx (1000 ticks, 1.000 sec)
2016-09-27 20:48:30 +03:00
Eemeli Kallio f41e428e5f Removed kvz_skip_unnecessary_rdoq and reworked --rdoq-skip to skip 4x4 blocks when it is on. 2016-09-09 10:26:07 +03:00
Ari Koivula 02cd17b427 Add faster AVX inter SAD for 32x32 and 64x64
Add implementations for these functions that process the image line by
line instead of using the 16x16 function to process block by block.

The 32x32 is around 30% faster, and 64x64 is around 15% faster,
on Haswell.

PASS inter_sad: 28.744M x reg_sad(32x32):x86_asm_avx (1014 ticks, 1.014 sec)
PASS inter_sad: 7.882M x reg_sad(64x64):x86_asm_avx (1014 ticks, 1.014 sec)
to
PASS inter_sad: 37.828M x reg_sad(32x32):x86_asm_avx (1014 ticks, 1.014 sec)
PASS inter_sad: 9.081M x reg_sad(64x64):x86_asm_avx (1014 ticks, 1.014 sec)
2016-09-01 21:36:39 +03:00
Ari Lemmetti 28c4174d0e Fix incorrect shuffle parameters
_MM_SHUFFLE uses reverse order
2016-08-23 19:40:46 +03:00
Ari Lemmetti ce77bfa15b Replace KVZ_PERMUTE with _MM_SHUFFLE
The same exact macro already exists
2016-08-22 19:08:46 +03:00
Eemeli Kallio 99d8b9abeb Changed skip_rdoq name to kvz_skip_unnecessary_rdoq. Changed the order it uses when it goes through CGs and tuned its sum calculation. 2016-08-18 14:02:56 +03:00
Eemeli Kallio 1fb4755f31 Added rdoq-skip to quant-generic.c 2016-08-18 12:17:54 +03:00
Eemeli Kallio d20ac03ca2 Added --rdoq-skip option 2016-08-18 12:17:53 +03:00
Arttu Ylä-Outinen 2a946bd88e Rename encoder_state_t.global to frame
"Frame" is more accurate than "global" since when OWF is used, encoder
states for each frame have their own struct.
2016-08-10 13:22:36 +09:00
Arttu Ylä-Outinen 5fbb0a8c27 Fix includes 2016-08-10 13:05:40 +09:00
Ari Lemmetti 6bcba004ff Comment out to fix unused code error on clang. 2016-07-14 14:12:16 +03:00
Ari Lemmetti c0979ebdcb Implement AVX2 luma sampling 2016-07-14 12:53:02 +03:00
Ari Lemmetti 6244560426 Add avx2 strategy for kvz_filter_frac_blocks_luma. 2016-07-14 12:53:02 +03:00
Ari Lemmetti 9c4e9e049b Load only what is needed. Eliminate latency from hadds. 2016-07-14 12:53:01 +03:00
Ari Lemmetti fccfbd2f28 Add strategy for kvz_filter_frac_blocks_luma 2016-07-14 12:51:02 +03:00
Ari Lemmetti 2b0c8db349 Add quad satd for avx2 2016-07-14 12:50:24 +03:00
Ari Lemmetti 0ff69fd6f8 Add any size multi satd 2016-07-14 12:48:37 +03:00
Arttu Ylä-Outinen bf26661782 Add support for 4x4 blocks to SATD_ANY_SIZE.
Makes functions satd_any_size_generic and satd_any_size_8bit_avx2 work
on blocks whose width and/or height are not multiples of 8.
2016-06-16 18:53:17 +09:00
Ari Lemmetti 3107a93eaf Fix avx2 chroma sampling for amp 2016-05-17 14:09:57 +03:00
Ari Lemmetti efbdc5dade Utilize registers more efficiently for 8x8 and larger blocks 2016-04-21 13:26:38 +03:00
Ari Lemmetti 192cee95b2 Vectorize vertical filtering 2016-04-21 13:26:38 +03:00
Ari Lemmetti 0be35f72b8 Filter 4 pixels simultaneously in x direction 2016-04-21 13:26:38 +03:00
Ari Lemmetti 10484bda9f Make strategies out of fractional pixel sample functions 2016-04-21 13:26:38 +03:00
Ari Lemmetti 8247faf8e0 Remove 64-bit only instruction to fix 32-bit compilation. 2016-04-19 18:05:11 +03:00
Ari Lemmetti eb55d6b6b9 Fix writing over boundary. 2016-04-19 16:03:43 +03:00
Ari Lemmetti bcabc6fadd Remove pixel blit from strategies. Use memcpy instead. 2016-04-06 18:44:04 +03:00
Ari Koivula 61fc3e87ba Run include-what-you-use fix_includes.py fix_includes.py
The includes should make more sense now and not just happen to compile
due to headers included from other headers.

Used a modified version of IWYU. Modifications were to attribute int8_t
and so on to stdint.h instead of sys/types.h and immintrin.h instead of
more specific headers.

include-what-you-use 0.7 (git:b70df35)
based on clang version 3.9.0 (trunk 264728)
2016-04-01 17:46:55 +03:00
Ari Koivula 8908d85d66 Change all relative includes to absolute 2016-04-01 17:46:44 +03:00
Ari Koivula 4876879b82 Add IWYU pragmas 2016-03-31 12:33:34 +03:00
Ari Koivula 5b66578f71 Add kvz_ prefix to md5 functions
The non kvz_ symbols were being exported in the static lib, which got caught
by Travis tests.
2016-03-18 13:13:35 +02:00
Ari Koivula 4125218cfa Add --hash=md5
Add md5 through extras/libmd5 taken from HM with BSD license. It's
implemented as a generic strategy using the same interface as checksum,
so we can write a SIMD version if it seems necessary.
2016-03-18 05:23:57 +02:00
Ari Lemmetti e502292ba8 Remove old function 2016-03-16 20:18:55 +02:00
Ari Lemmetti c6cc96f5ec Optimize sao band ddistortion 2016-03-16 20:16:00 +02:00
Ari Lemmetti ab577f476f Optimize sao reconstruct color 2016-03-16 20:15:32 +02:00
Ari Lemmetti 48bfddf4ec Optimize calc sao edge dir 2016-03-16 20:14:50 +02:00
Ari Lemmetti ba69992941 Optimize sao edge ddistortion 2016-03-16 20:14:19 +02:00
Ari Lemmetti 941b6b3e27 Optimize calc eo cat 2016-03-16 20:13:30 +02:00
Ari Lemmetti 04fbb48a09 Add strategy for avx2. Copy generic functions there. 2016-03-16 20:13:15 +02:00
Ari Lemmetti 4e30a215d8 Create generic strategy for sao. 2016-03-16 20:11:15 +02:00
Ari Lemmetti 99e37ec235 Update old pixel type to the current one 2016-01-30 19:33:09 +02:00
Ari Koivula fa1af14637 Fix includes to include global.h first everywhere 2016-01-22 15:07:49 +02:00
Ari Lemmetti 44656aeb19 Remove useless calculation 2016-01-19 16:35:16 +02:00