hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-11-26 19:04:06 +00:00
Code Issues Actions Projects Releases Wiki Activity
4463 commits 3 branches 6 tags 9.2 MiB
Commit graph

4463 commits

This branch
This branch
All branches
Author SHA1 Message Date
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 760bd0397d Pad the image buffer by 64 bytes from both ends 
This will be necessary for an efficient and straightforward
implementation of hor_sad for blocks over 16 pixels wide, because they
cannot use the shuffle trick because inter-lane shuffling is so hard to
do
 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 ca94ae9529 Handle extrapolated blocks with unmodified width using optimized_sad pointer 2019-02-04 20:41:40 +02:00
Pauli Oikkonen 91b30c7064 Tidy up kvz_image_calc_sad 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 007fb7ae19 Un-break tests 
Pass the NULL optimized function pointer from the test function, it
should still forward execution to width-specific SAD implementations
 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 bed93fb7f5 Merge branch 'sad-avx2' 2019-01-10 17:48:09 +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
Ari Lemmetti cd818db724 Add missing quantization and residual in cost calculation (inter rd=2). 2018-12-21 15:55:29 +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 3b635309a1 Add new files to Visual Studio project 2018-12-18 20:48:41 +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