arm: Add NEON optimizations for 10 and 12 bit vp9 MC
authorMartin Storsjö <martin@martin.st>
Thu, 8 Dec 2016 21:35:31 +0000 (23:35 +0200)
committerMartin Storsjö <martin@martin.st>
Tue, 24 Jan 2017 20:35:50 +0000 (22:35 +0200)
commita4d4bad75c0bf73ed9cb3ed2973a45425ec9ebc5
treebd26294d0c4174da497922d4ad658d4c1ecc12e5
parentcda9a3e80b74160524ce484c0a984afebbbae71c
arm: Add NEON optimizations for 10 and 12 bit vp9 MC

This work is sponsored by, and copyright, Google.

The plain pixel put/copy functions are used from the 8 bit version,
for the double size (e.g. put16 uses ff_vp9_copy32_neon), and a new
copy128 is added.

Compared with the 8 bit version, the filters can no longer use the
trick to accumulate in 16 bit with only saturation at the end, but now
the accumulators need to be 32 bit. This avoids the need to keep track
of which filter index is the largest though, reducing the size of the
executable code for these filters.

For the horizontal filters, we only do 4 or 8 pixels wide in parallel
(while doing two rows at a time), since we don't have enough register
space to filter 16 pixels wide.

For the vertical filters, we still do 4 and 8 pixels in parallel just
as in the 8 bit case, but we need to store the output after every 2
rows instead of after every 4 rows.

Examples of relative speedup compared to the C version, from checkasm:
                               Cortex    A7     A8     A9    A53
vp9_avg4_10bpp_neon:                   2.25   2.44   3.05   2.16
vp9_avg8_10bpp_neon:                   3.66   8.48   3.86   3.50
vp9_avg16_10bpp_neon:                  3.39   8.26   3.37   2.72
vp9_avg32_10bpp_neon:                  4.03  10.20   4.07   3.42
vp9_avg64_10bpp_neon:                  4.15  10.01   4.13   3.70
vp9_avg_8tap_smooth_4h_10bpp_neon:     3.38   6.22   3.41   4.75
vp9_avg_8tap_smooth_4hv_10bpp_neon:    3.89   6.39   4.30   5.32
vp9_avg_8tap_smooth_4v_10bpp_neon:     5.32   9.73   6.34   7.31
vp9_avg_8tap_smooth_8h_10bpp_neon:     4.45   9.40   4.68   6.87
vp9_avg_8tap_smooth_8hv_10bpp_neon:    4.64   8.91   5.44   6.47
vp9_avg_8tap_smooth_8v_10bpp_neon:     6.44  13.42   8.68   8.79
vp9_avg_8tap_smooth_64h_10bpp_neon:    4.66   9.02   4.84   7.71
vp9_avg_8tap_smooth_64hv_10bpp_neon:   4.61   9.14   4.92   7.10
vp9_avg_8tap_smooth_64v_10bpp_neon:    6.90  14.13   9.57  10.41
vp9_put4_10bpp_neon:                   1.33   1.46   2.09   1.33
vp9_put8_10bpp_neon:                   1.57   3.42   1.83   1.84
vp9_put16_10bpp_neon:                  1.55   4.78   2.17   1.89
vp9_put32_10bpp_neon:                  2.06   5.35   2.14   2.30
vp9_put64_10bpp_neon:                  3.00   2.41   1.95   1.66
vp9_put_8tap_smooth_4h_10bpp_neon:     3.19   5.81   3.31   4.63
vp9_put_8tap_smooth_4hv_10bpp_neon:    3.86   6.22   4.32   5.21
vp9_put_8tap_smooth_4v_10bpp_neon:     5.40   9.77   6.08   7.21
vp9_put_8tap_smooth_8h_10bpp_neon:     4.22   8.41   4.46   6.63
vp9_put_8tap_smooth_8hv_10bpp_neon:    4.56   8.51   5.39   6.25
vp9_put_8tap_smooth_8v_10bpp_neon:     6.60  12.43   8.17   8.89
vp9_put_8tap_smooth_64h_10bpp_neon:    4.41   8.59   4.54   7.49
vp9_put_8tap_smooth_64hv_10bpp_neon:   4.43   8.58   5.34   6.63
vp9_put_8tap_smooth_64v_10bpp_neon:    7.26  13.92   9.27  10.92

For the larger 8tap filters, the speedup vs C code is around 4-14x.

Signed-off-by: Martin Storsjö <martin@martin.st>
libavcodec/arm/Makefile
libavcodec/arm/vp9dsp_init.h [new file with mode: 0644]
libavcodec/arm/vp9dsp_init_10bpp_arm.c [new file with mode: 0644]
libavcodec/arm/vp9dsp_init_12bpp_arm.c [new file with mode: 0644]
libavcodec/arm/vp9dsp_init_16bpp_arm_template.c [new file with mode: 0644]
libavcodec/arm/vp9dsp_init_arm.c
libavcodec/arm/vp9mc_16bpp_neon.S [new file with mode: 0644]