1. 06 Oct, 2017 3 commits
  2. 05 Oct, 2017 1 commit
  3. 04 Oct, 2017 1 commit
    • Yi Luo's avatar
      Lowbd TM_PRED intra pred avx2 optimization · 237cf1b2
      Yi Luo authored
      For block width >= 16, avx2 can further speedup the
      TM_PREM intra prediction.
      
      Function speedup on i7-6700:
      Predictor  avx2 v. ssse3
      16x8       ~1.6x
      16x16      ~1.8x
      16x32      ~1.9x
      32x16      ~1.9x
      32x32      ~1.9x
      
      Change-Id: I62c20bd7628f52251b0c051b99a9b738ee44f7e6
      237cf1b2
  4. 29 Sep, 2017 3 commits
    • Yi Luo's avatar
      Lowbd TM_PRED intrapred ssse3 optimization · a0f66fc0
      Yi Luo authored
      Function speedup (i7-6700)
      Predictor  ssse3 v. C
      4x4        ~2.1x
      4x8        ~2.4x
      8x4        ~4.1x
      8x8        ~5.4x
      8x16       ~6.1x
      16x8       ~5.9x
      16x16      ~6.4x
      16x32      ~6.7x
      32x16      ~7.4x
      32x32      ~8.0x
      
      Change-Id: I52b8ebf8193e76f4ea1137cbad5ad7fa109d86d8
      a0f66fc0
    • Rupert Swarbrick's avatar
      Add 32x128/128x32 block sizes · 2fa6e1ce
      Rupert Swarbrick authored
      Change-Id: Ieb28f40d85e4db4af33648c32c406dd2931ceb89
      2fa6e1ce
    • Yi Luo's avatar
      Lowbd intrapred DC/TOP/LEFT/128/V/H avx2 · 23c61903
      Yi Luo authored
      For prediction block width equal to 32, avx2 can further speedup
      the prediction function (i7-6700):
      
      32x32     avx2 v. sse2
      DC        ~1.4x
      top       ~1.5x
      left      ~1.4x
      128       ~1.5x
      v         ~1.6x
      h         ~1.2x
      
      32x16     avx2 v. sse2
      DC        ~2.2x
      top       ~1.7x
      left      ~1.6x
      128       ~1.8x
      v         ~1.9x
      
      Note: 32x16 H_PRED on avx2 does not run faster enough than sse2 yet.
      
      Change-Id: I145ed504d1b3ea9df283b94927be66a2c6f81225
      23c61903
  5. 28 Sep, 2017 1 commit
    • Yi Luo's avatar
      Lowbd rectangle V/H intra pred sse2 optimization · 0c0fd1e5
      Yi Luo authored
      Function speedup sse2 v. C
      Predictor  V_PRED  H_PRED
      4x8        ~1.7x   ~1.8x
      8x4        ~1.8x   ~2.2x
      8x16       ~1.5x   ~1.4x
      16x8       ~1.9x   ~1.3x
      16x32      ~1.6x   ~1.4x
      32x16      ~2.0x   ~1.9x
      
      This patch disables speed tests to save Jenkins build
      time. Developer can manually enable them by using,
      --gtest_also_run_disabled_test flag in test command line.
      
      Change-Id: I81eaee5e8afc55275c7507c99774f78cc9e49f9a
      0c0fd1e5
  6. 27 Sep, 2017 1 commit
    • Yi Luo's avatar
      Lowbd rect intrapred DC/LEFT/TOP/128 sse2 optimization · 39bdf36a
      Yi Luo authored
      Add lowbd unit test functionality to intrapred_test.cc
      Function speedup against C (i7-6700):
      Predictor   DC     LEFT   TOP    128
      4x8        ~1.4x  ~1.4x  ~1.7x  ~1.9x
      8x4        ~1.2x  ~1.6x  ~1.6x  ~2.6x
      8x16       ~1.4x  ~1.3x  ~1.4x  ~2.1x
      16x8       ~2.0x  ~1.8x  ~2.3x  ~2.1x
      16x32      ~2.0x  ~1.9x  ~1.8x  ~2.2x
      32x16      ~2.0x  ~2.0x  ~1.9x  ~2.2x
      
      Change-Id: I33db512020ca3c6853a9205a8079f3d00134f584
      39bdf36a
  7. 22 Sep, 2017 1 commit
    • Yi Luo's avatar
      Highbd rectangle intrapred V/DC sse2 optimization · bdddf33a
      Yi Luo authored
      Function speedup (i7-6700),  sse2 verse C:
      Predictor      V_PRED    DC_PRED
      4x8            ~1.5x     ~4.9x
      8x4            ~2.5x     ~4.8x
      8x16           ~1.9x     ~9.1x
      16x8           ~1.9x     ~4.4x
      16x32          ~2.1x     ~5.8x
      32x16          ~2.0x     ~3.6x
      
      Change-Id: I6deffd0637e57ee5d0bd533502f5705148c4cdd4
      bdddf33a
  8. 19 Sep, 2017 1 commit
    • Yi Luo's avatar
      Highbd intrapred DC_LEFT/TOP/128 sse2 optimization · bbf6186e
      Yi Luo authored
      Also extend intra pred speed test to rectangular block.
      Speedup (i7-6700)
      predictor      sse2 v. C
      left 4x4       ~5.6x
      top  4x4       ~7.2x
      128  4x4       ~6.9x
      left 4x8       ~7.7x
      top  4x8       ~10.1x
      128  4x8       ~10.0x
      
      left 8x4       ~8.1x
      top  8x4       ~9.1x
      128  8x4       ~10.1x
      left 8x8       ~10.3x
      top  8x8       ~13.6x
      128  8x8       ~14.8x
      left 8x16      ~12.6x
      top  8x16      ~14.0x
      128  8x16      ~15.5x
      
      left 16x8      ~6.3x
      top  16x8      ~7.0x
      128  16x8      ~6.5x
      left 16x16     ~6.5x
      top  16x16     ~7.1x
      128  16x16     ~8.2x
      left 16x32     ~5.1x
      top  16x32     ~6.4x
      128  16x32     ~5.6x
      
      left 32x16     ~4.2x
      top  32x16     ~4.3x
      128  32x16     ~4.5x
      left 32x32     ~3.8x
      top  32x32     ~3.7x
      128  32x32     ~3.9x
      
      Change-Id: Ie7fcc85b9ded3030ee904623c40e9edeec1695ae
      bbf6186e
  9. 18 Sep, 2017 1 commit
    • Yi Luo's avatar
      Highbd intra pred H_PRED sse2 optimization · 23b9b317
      Yi Luo authored
      sse2 v. C speedup:
      4x4   ~8.0x
      8x8   ~8.2x
      16x16 ~6.5x
      32x32 ~3.8x
      Blocksize:
      4x4, 4x8, 8x4, 8x8, 8x16, 16x8, 16x16, 16x32, 32x16, 32x32
      Square blocksize code is from libvpx:
      "30d9a1916 vpxdsp: [x86] add highbd_h_predictor functions",
      Credit goes to Scott LaVarnway. Speed tests do not support
      rectangular blocksize yet.
      
      Change-Id: I9a1f24aecab8de94f8ea59ec8748fe3537d721ae
      23b9b317
  10. 07 Sep, 2017 1 commit
    • Yi Luo's avatar
      Lowbd parallel_deblocking sse2 optimization · ea8a0d52
      Yi Luo authored
      Baseline + parallel_deblocking:
      
      - Passed unit tests *SSE2/Loop8Test6*, *AVX2/Loop8Test6*.
      - 1080p, 25 frames, profile=0, encoding/decoding, output match.
      - Decoder frame rate increases from 54.15 to 65.84.
      
      Change-Id: I55938c94961066594f4b9080192c7268c19d9bf9
      ea8a0d52
  11. 05 Sep, 2017 1 commit
    • Rupert Swarbrick's avatar
      Define missing subtract_xxx functions in highbd_subtract_sse2.c · 4b5c2bb4
      Rupert Swarbrick authored
      Also, get rid of the boilerplate code using some macros. STACK_V(h,f) means
      "call f twice, stacking vertically at an offset of h". STACK_H(w,f)
      means "call f twice, stacking horizontally at an offset of w".
      
      Note that functions like subtract_128x64 are now only defined when the
      equivalent block sizes (e.g. BLOCK_128x64) are defined. As such, we
      have to fix up subtract_test.cc so it doesn't try to call
      aom_highbd_subtract_block_sse2 with unsupported sizes.
      
      BUG=aomedia:684
      
      Change-Id: I5b0fefe70e4083786d11d25cdd5dcf02823bae7b
      4b5c2bb4
  12. 30 Aug, 2017 1 commit
    • Yi Luo's avatar
      Highbd parallel_deblocking sse2 optimization · 6f5569f3
      Yi Luo authored
      - Decoder speed improves ~13.7% (baseline + parallel_deblocking).
      - Highbd loopfilter AVX2 version works when this experiment is
        disabled.
      
      Change-Id: I5d56b137a1d52236a4735656c370d57ef71ae043
      6f5569f3
  13. 15 Aug, 2017 1 commit
    • Urvang Joshi's avatar
      Remove ALT_INTRA flag. · 93b543ab
      Urvang Joshi authored
      This experiment has been adopted as it has been cleared by Tapas.
      
      Change-Id: I0682face60f62dd43091efa0a92d09d846396850
      93b543ab
  14. 11 Aug, 2017 1 commit
    • Yi Luo's avatar
      Simplify pixel clamping in highbitdepth loop filter · 099b1221
      Yi Luo authored
      The constants used in pixel clamping is based on bitdepth.
      Their calculation is moved outside pixel clamping and does
      only once. This achieves about <2% speed improvement on
      decoder.
      
      Change-Id: I48dcaebe04a3478962c3b6568d247a23b47a89d4
      099b1221
  15. 10 Aug, 2017 1 commit
    • Yi Luo's avatar
      Highbd loop filter AVX2 · 6ae0054c
      Yi Luo authored
      - Speed test (ms) on i7-6700, Linux x86_64
        FUNCTION             SSE2    AVX2
        horizontal_edge_16   55      28
        vertical_16_dual     84      47
        horizontal_4_dual    27      13
        horizontal_8_dual    36      15
        vertical_4_dual      38      25
        vertical_8_dual      44      27
      - Decoder frame rate improves around 1.2% - 2.8%.
      
      Change-Id: I9c4123869bac9b6d32e626173c2a8e7eb0cf49e7
      6ae0054c
  16. 04 Aug, 2017 1 commit
  17. 14 Jul, 2017 1 commit
  18. 13 Jul, 2017 1 commit
    • Peter de Rivaz's avatar
      Prevent negative variance in 10/12-bit. · e0794b51
      Peter de Rivaz authored
      Due to rounding the variance can become negative.
      Similar behaviour has been fixed in commit b7e7c044.
      I have made the corresponding changes to the masked_variance code.
      
      BUG=aomedia:646
      
      Change-Id: I21ed5a3f58080def08ac7ab9a77cbcad90318c99
      e0794b51
  19. 12 Jul, 2017 1 commit
    • Rupert Swarbrick's avatar
      ext-partition-types: Add 4:1 partitions · 93c39e91
      Rupert Swarbrick authored
      This patch adds support for 4:1 rectangular blocks to various common
      data arrays, and adds new partition types to the EXT_PARTITION_TYPES
      experiment which will use them.
      
      This patch has the following restrictions, which can be lifted in
      future patches:
      
        * ext-partition-types is incompatible with fp_mb_stats and supertx
          for the moment
      
        * Currently only 32x32 superblocks can use the new partition types
      
      There's a slightly odd restriction about when we allow
      PARTITION_HORZ_4 or PARTITION_VERT_4. Since these both live in the
      EXT_PARTITION_TYPES CDF, read_partition() can only return them if both
      has_rows and has_cols is true. This means that at least half of the
      width and height of the block must be visible. It might be nice to
      relax that restriction but that would imply a change to how we encode
      partition types, which seems already to be in a state of flux, so
      maybe it's better to wait until that has settled down.
      
      Change-Id: Id7fc3fd0f762f35f63b3d3e3bf4e07c245c7b4fa
      93c39e91
  20. 11 Jul, 2017 1 commit
  21. 29 Jun, 2017 1 commit
  22. 28 Jun, 2017 3 commits
  23. 27 Jun, 2017 1 commit
    • Yi Luo's avatar
      Fix inv txfm low/high bitdepth selection logic · 51281095
      Yi Luo authored
      We are going to have several commits to setup new low/high
      bitdepth data path selection logic. This patch is for inverse
      transform. Let me summarize the ideas as following.
      
      - For low/high bitdepth selection, encoder depends on
        input configuration, e.g., video sequence bitdepth,
        profile. Decoder depends on input bitstream. This has
        nothing to do with compiler/build  configuration.
      
      - Typical encoder usage for sampling format 4:2:0.
        1) 8-bit video sequence:
         a) --profile=0
         Fastest encoding/decoding pipeline on speedup.
      
         b) --profile=2 --bit-depth=10
         Image pixels are left shifted by 2 bits. It
         employs 16-bit reference frame buffer and has high
         calculation precision. It usually enjoys higher
         compression performance.
      
        2) 10/12-bit video sequence (HDR):
         --profile=2 --bit-depth=10/12
      
      - Transform coefficient type:
        Lowbitdepth:  int16_t
        Highbitdepth: int32_t
      
      - The type, tran_low_t is still used in codebase,
        Which is int32_t, defining the data path capacity.
        Naturally, it is high bitdepth.
      
      Eventually we shall remove the configuration flags,
      CONFIG_HIGHBITDEPTH/CONFIG_LOWBITDEPTH, and seperate
      low and high bitdepth data path. Two data paths co-exist
      in the same build environment.
      
      Change-Id: I35c06d4d4f19ebf80d909168fdddbae57c3cc884
      51281095
  24. 22 Jun, 2017 1 commit
    • Yi Luo's avatar
      Add avx2 highbd_quantize_b · 193422e7
      Yi Luo authored
      - First pass encoding time reduces ~10.9% on i7-6700
        at 100 frames, 1080p.
      - avx2 works for coeff number >= 8 cases; coeff number < 8
        case will be implemented by sse2.
      - Unit test is added type B/FP/DC.
      
      Change-Id: Ibe5b7807c64e6dfc2d59c470ed50a6e8ca94ef7c
      193422e7
  25. 19 Jun, 2017 1 commit
    • Timothy B. Terriberry's avatar
      encoder: Remove 64x upsampled reference buffers · 5d24b6f0
      Timothy B. Terriberry authored
      They do not handle border extension correctly (interpolation and
      border extension do not commute unless you upsample into the
      border), nor do they handle crop dimensions that are not a multiple
      of 8 (the upsampled version is not sufficiently large), in addition
      to using massive amounts of memory and being a criminal waste of
      cache (1 byte used for every 8 bytes fetched).
      
      This commit reimplements use_upsampled_references by computing the
      subpixel samples on the fly. This implementation not only corrects
      the border handling, but is also faster, while maintaining the
      same quality.
      
      HL AWCY results are basically noise:
          PSNR | PSNR HVS |   SSIM | MS SSIM | CIEDE 2000
        0.0188 |   0.0187 | 0.0045 |  0.0063 |     0.0228
      
      Change-Id: I7527db9f83b87a7bb8b35342f7e6457cd0bef9cd
      5d24b6f0
  26. 09 Jun, 2017 1 commit
    • David Barker's avatar
      Fix Windows x86 build with --enable-ext-inter · dab3e99b
      David Barker authored and James Zern's avatar James Zern committed
      The Windows calling convention pushes any __m128i type arguments
      after the 3rd (4th on x86-64) onto the stack. But on x86,
      stack-allocated arguments are not guaranteed to be aligned to
      a multiple of their natural alignment, leading to compile errors.
      
      We fix this by making the functions which take >3 __m128i arguments
      instead take pointers. Since the functions are marked INLINE, the
      extra memory operations should optimize out.
      
      BUG=aomedia:587
      
      Change-Id: I0cb2831fd12aded6f2821c037365386e6183ba5c
      dab3e99b
  27. 08 Jun, 2017 2 commits
    • Frederic Barbier's avatar
      Cleanup dead fwd transform functions · d405f8a6
      Frederic Barbier authored
      Cleanup related wrappers and unit-tests.
      
      Change-Id: I2d37a8c80de63dbeaef584e3d5fa842c0b2ee6db
      d405f8a6
    • Sarah Parker's avatar
      Remove deprecated high-bitdepth functions · 31c66502
      Sarah Parker authored
      This unifies the codepath for high-bitdepth transforms and deletes
      all calls to the old deprecated versions. This required reworking
      the way 1d configurations are combined in order to support rectangular
      transforms.
      
      There is one remaining codepath that calls the deprecated 4x4 hbd
      transform from encoder/encodemb.c. I need to take a closer look
      at what is happening there and will leave that for a followup
      since this change has already gotten so large.
      
      lowres 10 bit: -0.035%
      lowres 12 bit: 0.021%
      
      BUG=aomedia:524
      
      Change-Id: I34cdeaed2461ed7942364147cef10d7d21e3779c
      31c66502
  28. 30 May, 2017 1 commit
  29. 27 May, 2017 1 commit
    • Debargha Mukherjee's avatar
      High precision Wiener filter rework · 11cf46f4
      Debargha Mukherjee authored
      Implements the high precision Wiener filter with an offset
      to reduce the error due to saturation without increasing
      the number of bits needed for intermediate precision.
      
      Also turns the high precision filter on.
      
      Change-Id: I34037a5746a6a89c5fce67753c1b027749085edf
      11cf46f4
  30. 26 May, 2017 1 commit
    • David Barker's avatar
      ext-inter: Vectorize new masked SAD/SSE functions · 0aa39ff0
      David Barker authored
      We would expect that these new functions would be slower than
      the old masked SAD/SSE functions, as they do additional work
      (blending two inputs and comparing to a third, rather than
      just comparing two inputs).
      
      This is true for the SAD functions, which are about 50% slower
      (depending on block size and bit depth). However, the sub-pixel
      SSE functions are comparable to the old speed for the accelerated
      special cases (xoffset or yoffset = 0 or 4), and are
      between 40-90% faster for the generic case.
      
      Change-Id: I1a296ed8fc9e3edc313a6add516ff76b17cd3e9f
      0aa39ff0
  31. 23 May, 2017 2 commits
    • David Barker's avatar
      Vectorize high-precision convolve filter · 5d34e6a7
      David Barker authored
      Add SSE2 lowbd and SSSE3 highbd versions of the filters
      introduced in https://aomedia-review.googlesource.com/c/11962/ .
      
      These filters are equivalent in speed to the SSE2 implementations
      of the regular convolve filter. The average time to filter a
      64x64 block is:
      
      lowbd C: 52us
      lowbd SSE2: 5.6us
      highbd C: 53us
      highbd SSSE3: 5.8us
      
      Also add a correctness test based on the warp filter tests.
      
      Change-Id: Ia0d81100e8a414bbfc2b5f664d751cf24765299e
      5d34e6a7
    • David Barker's avatar
      ext-inter: Delete dead code · 0f3c94e1
      David Barker authored
      Patches https://aomedia-review.googlesource.com/c/11987/
      and https://aomedia-review.googlesource.com/c/11988/
      replaced the old masked motion search pipeline with
      a new one which uses different SAD/SSE functions.
      This resulted in a lot of dead code.
      
      This patch removes the now-dead code. Note that this
      includes vectorized SAD/SSE functions, which will need
      to be rewritten at some point for the new pipeline. It
      also includes the masked_compound_variance_* functions
      since these turned out not to be used by the new pipeline.
      
      To help with the later addition of vectorized functions, the
      masked_sad/variance_test.cc files are kept but are modified
      to work with the new functions. The tests are then disabled
      until we actually have the vectorized functions.
      
      Change-Id: I61b686abd14bba5280bed94e1be62eb74ea23d89
      0f3c94e1
  32. 16 May, 2017 1 commit