Recherche avancée

Sur d’autres sites (7423)

• arm : vp9 : Add NEON optimizations of VP9 MC functions

  14 novembre 2016, par Martin Storsjö

  arm : vp9 : Add NEON optimizations of VP9 MC functions
  This work is sponsored by, and copyright, Google.
  The filter coefficients are signed values, where the product of the
  
  multiplication with one individual filter coefficient doesn’t
  
  overflow a 16 bit signed value (the largest filter coefficient is
  
  127). But when the products are accumulated, the resulting sum can
  
  overflow the 16 bit signed range. Instead of accumulating in 32 bit,
  
  we accumulate the largest product (either index 3 or 4) last with a
  
  saturated addition.
  (The VP8 MC asm does something similar, but slightly simpler, by
  
  accumulating each half of the filter separately. In the VP9 MC
  
  filters, each half of the filter can also overflow though, so the
  
  largest component has to be handled individually.)
  Examples of relative speedup compared to the C version, from checkasm :
  
                         Cortex      A7     A8     A9    A53
  
  vp9_avg4_neon :                   1.71   1.15   1.42   1.49
  
  vp9_avg8_neon :                   2.51   3.63   3.14   2.58
  
  vp9_avg16_neon :                  2.95   6.76   3.01   2.84
  
  vp9_avg32_neon :                  3.29   6.64   2.85   3.00
  
  vp9_avg64_neon :                  3.47   6.67   3.14   2.80
  
  vp9_avg_8tap_smooth_4h_neon :     3.22   4.73   2.76   4.67
  
  vp9_avg_8tap_smooth_4hv_neon :    3.67   4.76   3.28   4.71
  
  vp9_avg_8tap_smooth_4v_neon :     5.52   7.60   4.60   6.31
  
  vp9_avg_8tap_smooth_8h_neon :     6.22   9.04   5.12   9.32
  
  vp9_avg_8tap_smooth_8hv_neon :    6.38   8.21   5.72   8.17
  
  vp9_avg_8tap_smooth_8v_neon :     9.22  12.66   8.15  11.10
  
  vp9_avg_8tap_smooth_64h_neon :    7.02  10.23   5.54  11.58
  
  vp9_avg_8tap_smooth_64hv_neon :   6.76   9.46   5.93   9.40
  
  vp9_avg_8tap_smooth_64v_neon :   10.76  14.13   9.46  13.37
  
  vp9_put4_neon :                   1.11   1.47   1.00   1.21
  
  vp9_put8_neon :                   1.23   2.17   1.94   1.48
  
  vp9_put16_neon :                  1.63   4.02   1.73   1.97
  
  vp9_put32_neon :                  1.56   4.92   2.00   1.96
  
  vp9_put64_neon :                  2.10   5.28   2.03   2.35
  
  vp9_put_8tap_smooth_4h_neon :     3.11   4.35   2.63   4.35
  
  vp9_put_8tap_smooth_4hv_neon :    3.67   4.69   3.25   4.71
  
  vp9_put_8tap_smooth_4v_neon :     5.45   7.27   4.49   6.52
  
  vp9_put_8tap_smooth_8h_neon :     5.97   8.18   4.81   8.56
  
  vp9_put_8tap_smooth_8hv_neon :    6.39   7.90   5.64   8.15
  
  vp9_put_8tap_smooth_8v_neon :     9.03  11.84   8.07  11.51
  
  vp9_put_8tap_smooth_64h_neon :    6.78   9.48   4.88  10.89
  
  vp9_put_8tap_smooth_64hv_neon :   6.99   8.87   5.94   9.56
  
  vp9_put_8tap_smooth_64v_neon :   10.69  13.30   9.43  14.34
  For the larger 8tap filters, the speedup vs C code is around 5-14x.
  This is significantly faster than libvpx’s implementation of the same
  
  functions, at least when comparing the put_8tap_smooth_64 functions
  
  (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from
  
  libvpx).
  Absolute runtimes from checkasm :
  
                            Cortex      A7        A8        A9       A53
  
  vp9_put_8tap_smooth_64h_neon :    20150.3   14489.4   19733.6   10863.7
  
  libvpx vpx_convolve8_horiz_neon : 52623.3   19736.4   21907.7   25027.7
  vp9_put_8tap_smooth_64v_neon :    14455.0   12303.9   13746.4    9628.9
  
  libvpx vpx_convolve8_vert_neon :  42090.0   17706.2   17659.9   16941.2
  Thus, on the A9, the horizontal filter is only marginally faster than
  
  libvpx, while our version is significantly faster on the other cores,
  
  and the vertical filter is significantly faster on all cores. The
  
  difference is especially large on the A7.
  The libvpx implementation does the accumulation in 32 bit, which
  
  probably explains most of the differences.
  This is an adapted cherry-pick from libav commits
  
  ffbd1d2b0002576ef0d976a41ff959c635373fdc,
  
  392caa65df3efa8b2d48a80f08a6af4892c61c08,
  
  557c1675cf0e803b2fee43b4c8b58433842c84d0 and
  
  11623217e3c9b859daee544e31acdd0821b61039.
  Signed-off-by : Ronald S. Bultje <rsbultje@gmail.com>
  

  • [DH] libavcodec/arm/Makefile
  • [DH] libavcodec/arm/vp9dsp_init_arm.c
  • [DH] libavcodec/arm/vp9mc_neon.S
  • [DH] libavcodec/vp9.c
  • [DH] libavcodec/vp9dsp.c
  • [DH] libavcodec/vp9dsp.h

• avcodec/ffv1_parser : Rename close to ffv1_close

  19 février, par Zhao Zhili

  avcodec/ffv1_parser : Rename close to ffv1_close
  This fixed wasm checkasm failure
  $ wasm-tools validate tests/checkasm/checkasm
  
  error : wasisdk ://v25.0/build/sysroot/wasi-libc-wasm32-wasip1-threads/libc-top-half/musl/src/stdio/__stdio_close.c:24:9 function `__stdio_close` failed to validate
  Caused by :
  
      0 : func 4581 failed to validate
  
      1 : type mismatch : expected i32 but nothing on stack (at offset 0x43b770)
  Since close is declared as static function, it's more like a bug
  
  in wasi sdk, but we can workaround it easily.
  Signed-off-by : Zhao Zhili <zhilizhao@tencent.com>
  
  Reviewed-by : James Almer <jamrial@gmail.com>
  

  • [DH] libavcodec/ffv1_parser.c

• How exactly does this Jaccard distance logic (get_jaccarddist) in FFmpeg work ?

  23 mars 2024, par stilloo

  I see this C code in FFmpeg source signature_lookup file for Jaccard distance in method get_jaccarddist.

  


  If you see in the code below that is calculating for 5 BagOfWords.

  


  Now for two BagOfWords with say nothing in common intersection would be 0, that means jaccarddist would be 0, so ideally its a bad pair but this if condition jaccarddist >= sc->thworddist, would be false isn't it but we want it to be true as its bad pair ?

  


  What are the default values of sc->thworddist ?

  


  Also surprised why jaccarddist is int and not float/double ?

  


  /**
 * calculates the jaccard distance and evaluates a pair of coarse signatures as good
 * @return 0 if pair is bad, 1 otherwise
 */
static int get_jaccarddist(SignatureContext *sc, CoarseSignature *first, CoarseSignature *second)
{
    int jaccarddist, i, composdist = 0, cwthcount = 0;
    for (i = 0; i < 5; i++) {
        if ((jaccarddist = intersection_word(first->data[i], second->data[i])) > 0) {
            jaccarddist /= union_word(first->data[i], second->data[i]);
        }
        if (jaccarddist >= sc->thworddist) {
            if (++cwthcount > 2) {
                /* more than half (5/2) of distances are too wide */
                return 0;
            }
        }
        composdist += jaccarddist;
        if (composdist > sc->thcomposdist) {
            return 0;
        }
    }
    return 1;
}