hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-12-02 21:24:07 +00:00
Code Issues Actions 1 Projects Releases Wiki Activity
3240 commits 3 branches 6 tags 11 MiB
Commit graph

465 commits

Author SHA1 Message Date
Ari Lemmetti 84222cf3e7 Replace old block extrapolation with more capable one. 
Separate paddings for different directions can be now specified.
 2021-03-08 22:36:04 +02:00
Pauli Oikkonen 816789c9f4 Allow fast coeff weights to be read from a file 2020-10-29 15:22:51 +02:00
Pauli Oikkonen 6799019db0 Move fast coeff table to transform.h 
Guess this is a more logical place for it
 2020-10-29 15:20:27 +02:00
Pauli Oikkonen 4712ce5f59 Round the fast coeff result instead of flooring 2020-10-29 15:20:27 +02:00
Pauli Oikkonen 0fb09c9920 New filtered coeff weight by QP values 2020-10-29 15:20:27 +02:00
Pauli Oikkonen 24d487f553 New weights for 12 <= QP <= 42 
Trained using MSU ultrafast settings now
 2020-10-29 15:20:27 +02:00
Pauli Oikkonen 3e1c6d84b8 Fix issues in fast coeff estimation 
Allow weight table to start from nonzero QP, and round weights to Q8.8
instead of flooring them
 2020-10-29 15:20:27 +02:00
Pauli Oikkonen 5f91bda762 Use newer data for fast coeff cost estimation 
Same training dataset, but this time only buckets 0...3 were used to
approximate the function, no sign/cg width bucket.
 2020-10-29 15:20:27 +02:00
Pauli Oikkonen 2abd733199 Use unsigned min() to correctly clip -32768 
If a coeff happens to be -32768 (0x8000), its 16-bit abs() is also
0x8000. It should ultimately be clipped to 3, so interpret absolute
values as unsigned instead to make that happen.
 2020-10-29 15:20:27 +02:00
Pauli Oikkonen b93b90c0d7 Implement new fast coeff cost estimator in AVX2 2020-10-29 15:20:27 +02:00
Pauli Oikkonen 2f74a112b3 Try first lookup table based fast coeff estimation 2020-10-29 15:20:27 +02:00
Pauli Oikkonen 780da4568a Exclude 8-bit-only code from 10-bit builds and use uint8_t instead of kvz_pixel for code that assumes 8-bit pixels 2020-09-02 17:46:33 +03:00
Jan Beich 1fa69c705d Rename truncate() from 30ce461d98 to avoid conflict with POSIX version 
strategies/avx2/dct-avx2.c:55:23: error: static declaration of 'truncate' follows non-static declaration
static INLINE __m256i truncate(__m256i v, __m256i debias, int32_t shift)
                      ^
/usr/include/stdio.h:448:6: note: previous declaration is here
int      truncate(const char *, __off_t);
         ^
 2020-04-22 16:09:42 +00:00
Ari Lemmetti f31dddc019 Bypass inverse quantization and inverse transform when trying early skip 2020-04-10 16:02:09 +03:00
Pauli Oikkonen 8617530b13 Use _mm_store_epi64 instead of _mm_cvtsi128_si64 
Fix 32-bit builds that tend to lack the cvt intrinsic. Hope it will be
optimized to a movq r64, xmm on modern platforms though
 2020-04-07 23:51:54 +03:00
Pauli Oikkonen a82966c0f5 Fix lacking _mm256_cvtss_f32 intrinsic on VS 
Cast __m256 into __m128 first, the XMM variant of the intrinsic has been
around for a long enough time to be supported
 2020-04-07 22:38:10 +03:00
Ari Lemmetti 901c25c0c8 Merge branch 'vaq' 2020-04-03 19:51:17 +03:00
Ari Lemmetti 51451be5ef Handle cases where the number of pixels is not divisible by 32 2020-04-03 19:37:47 +03:00
siivonek e5267f7706 Fix define for use with Visual Studio. 2020-04-03 15:11:01 +02:00
Pauli Oikkonen addc1c3ede Fix warning about potentially unused hsum_8x32b 
There's a lot of alternative options available, such as making it
globally visible with a kvz_ prefix, force inlining it, or anything.
This could be good too, hope it won't be compiled at all to translation
units where it's not used.
 2020-04-02 16:44:22 +03:00
siivonek 566680af7b Move function hsum to file where it is used to avoid errors. 2020-04-02 14:03:06 +02:00
siivonek 58be514e2a Fix pipeline error. 2020-04-02 13:50:08 +02:00
Pauli Oikkonen 99889dab15 Fix switch(bool) in picture-avx2.c 
It passes on GCC but warns on Clang
 2020-03-31 15:42:19 +03:00
Jaakko Laitinen af3d559d8d Let pu-depth be defined per gop-layer 2020-03-17 17:57:18 +02:00
Pauli Oikkonen 60e7956dc5 Disable inaccurate integer variance calculation for now 2020-03-02 19:18:55 +02:00
Pauli Oikkonen fc1b91335b Implement variance calculation in integer math 
Maybe this is a bit faster than FP, it's not accurate though
 2020-03-02 18:17:18 +02:00
Pauli Oikkonen 35c825c75f Move hsum_8x32b to avx2_common_functions 2020-02-27 17:52:17 +02:00
Pauli Oikkonen b00ac7d1c4 AVX2 version of buffer variance calculation 2020-02-25 15:57:56 +02:00
Pauli Oikkonen 1bd9c6dd93 Make a strategy out of pixel_var 2020-02-24 19:37:36 +02:00
Ari Lemmetti 3c7dd0752f Remove the broken "no mov" branch. 
Causes hash mismatches for example in SlideShow sequence.
 2020-02-03 15:26:31 +02:00
RLamm 30d5df40c5 Custom headers for the distributed coding 2020-01-29 15:54:49 +02:00
Pauli Oikkonen c3d9e97e9f Fix VS build 2019-12-12 18:34:55 +02:00
Pauli Oikkonen 7f238ca299 Remove debug print functions 
Whoops
 2019-12-12 18:19:31 +02:00
Pauli Oikkonen eefb5e50b3 De-inline pred_filtered_dc functions, shouldn't make much difference though 2019-12-12 17:30:00 +02:00
Pauli Oikkonen 169314de4f 32x32 filtered DC prediction in AVX2 2019-12-11 18:17:06 +02:00
Pauli Oikkonen fb2481b7e4 16x16 filtered DC implemented in AVX2 2019-12-10 15:54:50 +02:00
Pauli Oikkonen da370ea36d Implement AVX2 8x8 filtered DC algorithm 2019-11-28 14:10:10 +02:00
Pauli Oikkonen 5d9b7019ca Implement a 4x4 filtered DC pred function 2019-11-26 17:05:54 +02:00
Pauli Oikkonen f1485ab087 Start doing an arbitrary size filtered DC pred - maybe easier to just create separate functions for fixed block sizes? 2019-11-25 15:20:29 +02:00
Pauli Oikkonen 979d66031c Create a strategy out of intra_pred_filtered_dc 2019-11-19 14:50:31 +02:00
Pauli Oikkonen fa4bb86406 Optimize intra_pred_planar_avx2 for 4x4 blocks 2019-11-19 13:39:02 +02:00
Pauli Oikkonen 4761d228f9 Start to vectorize the 4x4 loop 2019-11-15 17:32:40 +02:00
Pauli Oikkonen 8d45ab4951 Stupidify the 4x4 planar loop for vectorization 2019-11-14 17:14:04 +02:00
Pauli Oikkonen 6d7a4f555c Also remove 16x16 (A * B^T)^T matrix multiply 
Can be done using (B * A^T) instead, it's the exact same
 2019-10-28 16:19:42 +02:00
Pauli Oikkonen 2c2deb2366 Tidy AVX2 32x32 matrix multiply 2019-10-28 16:19:42 +02:00
Pauli Oikkonen 98ad78b333 Tidy the old AVX2 32x32 matrix multiply 
It was actually a very good algorithm, just looked messy!
 2019-10-28 16:19:42 +02:00
Pauli Oikkonen 4a921cbdb5 Retain data as much in YMM registers as possible 
This seems to make it a whole lot quicker
 2019-10-28 16:19:42 +02:00
Pauli Oikkonen ac4d710e23 Unroll 32x32 matrix multiply, use all regs 2019-10-28 16:19:42 +02:00
Pauli Oikkonen a58608d0b8 Remove totally unnecessary (A * B^T)^T 32x32 multiply 2019-10-28 16:19:42 +02:00
Pauli Oikkonen 043f53539f Implement a streamlined matrix-multiply 32x32 DCT 2019-10-28 16:19:42 +02:00