1. 16 Feb, 2018 17 commits
  2. 15 Feb, 2018 12 commits
    • Jingning Han's avatar
      Reduce blk_skip array size · ca3f0395
      Jingning Han authored
      Reduce the blk_skip array size to the maximum txb numbers in a
      coding block. This makes the speed 1 about 5% faster for smaller
      resolution clips.
      
      Change-Id: I1812dc4d403e08f0566ead8a14c6046e505bb849
      ca3f0395
    • Jingning Han's avatar
      Reduce txb_coeff_cost map size from 64 to 32 · f32f678d
      Jingning Han authored
      The maximum coding block size is 128 and the minimum tx size is
      4. Using 32 per dimension to keep the txb coeff cost should be
      safe.
      
      Change-Id: Ie44fc581037f0d8270caec64543454701159eec5
      f32f678d
    • Jingning Han's avatar
      Fix compiler errors in cfl.h · 5a338ee4
      Jingning Han authored
      Change-Id: Iadf001ae293ca50b805e8b8c569900f60a23943c
      5a338ee4
    • David Barker's avatar
      [NORMATIVE-DECODING] Add missing special case to has_top_right() · 3f0f1dfc
      David Barker authored
      Commit ea190906 fixed a major bug in has_top_right(), but missed
      out one special case:
      
      Consider a 128x128 block using 64x64 intra predictions. These
      intra predictions are applied in a 'Z' order, and so the bottom-left
      64x64 unit has pixels available from the top-right one. But we were
      mistakenly setting has_top_right() = 0 for this case.
      
      More generally, whichever transform unit has its top-right corner
      at the center of a 128x128 block, should have has_top_right() = 1.
      
      Fix this by introducing an explicit check for the special case.
      
      BUG=aomedia:1309
      
      Change-Id: I690a292be6c1755c76bd428be94ab953dd71fbd2
      3f0f1dfc
    • Yaowu Xu's avatar
      Remove CONFIG_TX64X64 · d3d4159f
      Yaowu Xu authored
      The experiment is fully adopted.
      
      Change-Id: I6cc80a2acf0c93c13b0e36e6f4a2378fe5ce33c3
      d3d4159f
    • Debargha Mukherjee's avatar
      Stop using VP9 convolve scheme in AV1 encoder. · 1fc3df55
      Debargha Mukherjee authored
      Discontinue all VP9 style convolve rounding operations in the non-normative
      parts of the encoder.
      
      The function av1_convolve_2d_sr_c is forced instead of SIMD versions
      of the same function, because of incompatibility when round_1 > 0.
      
      In the -DCONFIG_LOWPRECISION_BLEND=2 -DCONFIG_HIGHPRECISION_INTBUF=1
      setting, results on 15 frames of lowres (cpu-used=1) is -0.019% better.
      
      Change-Id: I72154bd896357c352c944fb2cd3b25bafafba46a
      1fc3df55
    • Dominic Symes's avatar
      film-grain: fix buffer overflow · aa5904ba
      Dominic Symes authored
      When bit_depth is 8 the copy_rect function was setting the size to
      2 bytes per sample. This causes a buffer overflow as each line copied
      in the loop is twice the number of bytes it should be and the last
      line writes off the end of the buffer.
      
      BUG=aomedia:1389
      
      Change-Id: Ib9fa11d1dd13806dedbce2cd47dd8d562007428d
      aa5904ba
    • David Barker's avatar
      [NORMATIVE-SYNTAX, intrabc] Disallow intrabc+superres · 218556e2
      David Barker authored
      Superres should not be used together with intrabc, for the same
      reasons why the other loop filters are disabled when intrabc is
      used. However, we actually read superres information first,
      so the simplest change is to only read allow_intrabc if superres is off.
      
      BUG=aomedia:1384
      
      Change-Id: I55c4c8ace72ab3fac9417da5cf803883be7efb84
      218556e2
    • Yaowu Xu's avatar
      align a stack array to prevent segfault · 1e2084a2
      Yaowu Xu authored
      BUG=aomedia:1383
      
      Change-Id: I8c3de068fd158b4706b578f7609c8ef939364525
      1e2084a2
    • Luc Trudeau's avatar
      [CFL] SSE2/AVX2 Versions of Sum and Subtract Average · 365f73bb
      Luc Trudeau authored
      Includes unit tests for conformance and speed.
      
      SSSE2/CFLSubAvgTest
      4x4: C time = 234 us, SIMD time = 152 us (~1.5x)
      8x8: C time = 664 us, SIMD time = 208 us (~3.2x)
      16x16: C time = 1687 us, SIMD time = 581 us (~2.9x)
      32x32: C time = 6118 us, SIMD time = 2119 us (~2.9x)
      
      AVX2/CFLSubAvgTest
      4x4: C time = 250 us, SIMD time = 221 us (~1.1x)
      8x8: C time = 683 us, SIMD time = 284 us (~2.4x)
      16x16: C time = 1727 us, SIMD time = 1091 us (~1.6x)
      32x32: C time = 6092 us, SIMD time = 2107 us (~2.9x)
      
      Change-Id: I44ffedc683829d2c16089854ac43d4ddb4415bcd
      365f73bb
    • Andrey Norkin's avatar
      [NORMATIVE] Film grain bug-fixes · 2e8ae05c
      Andrey Norkin authored
      BUG=aomedia:1366
      BUG=aomedia:1368
      
      Change-Id: I63f84dca86ca426b9c6927b056657741022d5f68
      2e8ae05c
    • Angie Chiang's avatar
      Simplify the for loop of reading lower levels · 1b1c4a3c
      Angie Chiang authored
      Move tcoeff reconstruction from the for loop of reading lower
      levels to the loop of reading signs
      
      This is to facilitate the potential simd optimization of
      reading lower levels
      
      Change-Id: I61e41ea95d30a948243fb13e3d708dd9860a058a
      1b1c4a3c
  3. 14 Feb, 2018 11 commits
    • David Barker's avatar
      [NORMATIVE-DECODING] Fix above/left chroma block selection · d3afdb90
      David Barker authored
      As pointed out by rsbultje, my previous patch to is_smooth()
      (a883e6ea) was not quite correct. This is because, when we're
      making a chroma prediction, the uv_mode for the above/left chroma
      predictions is not necessarily in above_mbmi/left_mbmi. Instead,
      it may be in any of several places, depending on subsampling and
      the values of mi_row/mi_col.
      
      The cleanest solution is to explicitly maintain pointers to the
      above and left chroma blocks. Then we can simply look at those
      pointers when we want to know the above or left uv_mode.
      
      Also include a bit of refactoring of get_filt_type: It seems
      to be recalculating what's already in xd->{above,left}_mi,
      so just use those directly.
      
      BUG=aomedia:1362
      
      Change-Id: I0230474a50d43b78cb587a2b553da9ca78cec0c6
      d3afdb90
    • Cheng Chen's avatar
      Fix jnt_comp in warp function · 5c4848d6
      Cheng Chen authored
      Fixed a bug in jnt_comp simd function.
      Re-write the function based on existing av1_warp_affine_ssse3.
      The function is almost the same as av1_warp_affine_ssse3, only
      difference is the weighted computation of final value.
      
      Change-Id: I65f2fadc9142f6c264a7d0e59250602636c9808b
      5c4848d6
    • Peng Bin's avatar
      Refactor inv_cos_bit for speedup · 28744b5c
      Peng Bin authored
      Replace the last parameter cos_bit for all 1D inv_txfm funcions with
      a macro define, as it is actually always equal to 12. By changing it
      to const value, compiler can do further optimization.
      
      Change-Id: If8a9fd99c7ac7eb6f485dafbce22b4803efda62e
      28744b5c
    • Linfeng Zhang's avatar
      Implement fdct4x8_new_sse2 and fadst4x8_new_sse2 · 043f4964
      Linfeng Zhang authored
      Change-Id: I9ab260c5ca31fe7e06bfc0f806893463c5255c45
      043f4964
    • Linfeng Zhang's avatar
      Implement fdct4x4_new_sse2 · 1fffc1f4
      Linfeng Zhang authored
      Change-Id: I5b905d54f7e551b15dac5f9cfd4e6d96e19d52c3
      1fffc1f4
    • Linfeng Zhang's avatar
      Enable av1_lowbd_fwd_txfm2d_4x8_sse2 · 686c5979
      Linfeng Zhang authored
      Forgot to put it in the function table.
      
      Change-Id: I86192158f74df84fca6c4a3dbda8b01659f6c9bd
      686c5979
    • Linfeng Zhang's avatar
      Implement av1_lowbd_fwd_txfm2d_8x4_sse2 · 7bd00743
      Linfeng Zhang authored
      So far the implemented av1_lowbd_fwd_#x#_sse2 provides 10% encoder
      speed up on speed 1.
      
      Change-Id: I3dab438c4498059262b065300743ba1519db64b4
      7bd00743
    • Michelle Findlay-Olynyk's avatar
      hash_based_trellis speed feature update · dea531da
      Michelle Findlay-Olynyk authored
      Add speed feature that uses hash tables to
      reuse deltas from previously found optimized
      coefficients in av1_optimize_txb. This skips some
      expensive optimize_txb calls.
      
      Work stopped (no current benefit):
      Current setup has worse speed (~1%), and is
      turned off by default in speed_features.c.
      Overall PSNR was ~0.1% worse. See
      go/hashBasedTrellis for feature details.
      
      Requires hash_me.
      
      Change-Id: Id12ada82984ea6682e2b6f27534fa57f3a3abb4a
      dea531da
    • Ola Hugosson's avatar
      [NORMATIVE] Enable CONFIG_EXT_INTRA_MOD2 by default · 63f1216a
      Ola Hugosson authored
      CONFIG_EXT_INTRA_MOD2 was provisionally adopted 16th of January
      
      Change-Id: I11f96cdff29a772813f0cc7cd4cae684e9a07448
      63f1216a
    • Peng Bin's avatar
      Refactor pair_set_epi16 for speedup · 8b8aaffc
      Peng Bin authored
      Use _mm_set1_epi32 instead of _mm_set_epi16, less instructions produced
      by compiler. This patch also removes the duplicate define of the same
      function.
      
      Speed test results:
      1. Unittest for each test cases in SSE2/AV1LbdInvTxfm2d shows 60%~80%
      speedup (except those case with TX_TYPE include iidentity)
      2. A brief speed test shows that with this CL, for speed1 encoder speeds up
      ~3% and decoder speeds up ~1.8%.
      (Baseline is 18976fa5)
      
      Change-Id: I2b0e12973fda05a21d6b6eb0f0efe11df6edfb84
      8b8aaffc
    • Yaowu Xu's avatar
      Remove unused variables · cbfffa8e
      Yaowu Xu authored
      Change-Id: I5290f94da6c1a0319357f84b2ec70b4331a0e4af
      cbfffa8e