arm: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible
authorMartin Storsjö <martin@martin.st>
Fri, 24 Feb 2017 15:39:00 +0000 (17:39 +0200)
committerMartin Storsjö <martin@martin.st>
Sun, 19 Mar 2017 20:54:33 +0000 (22:54 +0200)
commiteabc5abf949bb8cadafe33df73adacf87ee4c5e0
tree58c9971ac51a2c0f0736024740922db1a379271a
parentd564c9018f8a45c1f4c38f02844186545582531d
arm: vp9itxfm16: Do a simpler half/quarter idct16/idct32 when possible

This work is sponsored by, and copyright, Google.

This avoids loading and calculating coefficients that we know will
be zero, and avoids filling the temp buffer with zeros in places
where we know the second pass won't read.

This gives a pretty substantial speedup for the smaller subpartitions.

The code size increases from 14516 bytes to 22484 bytes.

The idct16/32_end macros are moved above the individual functions; the
instructions themselves are unchanged, but since new functions are added
at the same place where the code is moved from, the diff looks rather
messy.

Before:                                 Cortex A7       A8       A9      A53
vp9_inv_dct_dct_16x16_sub1_add_10_neon:     454.0    270.7    418.5    295.4
vp9_inv_dct_dct_16x16_sub2_add_10_neon:    3840.2   3244.8   3700.1   2337.9
vp9_inv_dct_dct_16x16_sub4_add_10_neon:    4212.5   3575.4   3996.9   2571.6
vp9_inv_dct_dct_16x16_sub8_add_10_neon:    5174.4   4270.5   4615.5   3031.9
vp9_inv_dct_dct_16x16_sub12_add_10_neon:   5676.0   4908.5   5226.5   3491.3
vp9_inv_dct_dct_16x16_sub16_add_10_neon:   6403.9   5589.0   5839.8   3948.5
vp9_inv_dct_dct_32x32_sub1_add_10_neon:    1710.7    944.7   1582.1   1045.4
vp9_inv_dct_dct_32x32_sub2_add_10_neon:   21040.7  16706.1  18687.7  13193.1
vp9_inv_dct_dct_32x32_sub4_add_10_neon:   22197.7  18282.7  19577.5  13918.6
vp9_inv_dct_dct_32x32_sub8_add_10_neon:   24511.5  20911.5  21472.5  15367.5
vp9_inv_dct_dct_32x32_sub12_add_10_neon:  26939.5  24264.3  23239.1  16830.3
vp9_inv_dct_dct_32x32_sub16_add_10_neon:  29419.5  26845.1  25020.6  18259.9
vp9_inv_dct_dct_32x32_sub20_add_10_neon:  31146.4  29633.5  26803.3  19721.7
vp9_inv_dct_dct_32x32_sub24_add_10_neon:  33376.3  32507.8  28642.4  21174.2
vp9_inv_dct_dct_32x32_sub28_add_10_neon:  35629.4  35439.6  30416.5  22625.7
vp9_inv_dct_dct_32x32_sub32_add_10_neon:  37269.9  37914.9  32271.9  24078.9

After:
vp9_inv_dct_dct_16x16_sub1_add_10_neon:     454.0    276.0    418.5    295.1
vp9_inv_dct_dct_16x16_sub2_add_10_neon:    2336.2   1886.0   2251.0   1458.6
vp9_inv_dct_dct_16x16_sub4_add_10_neon:    2531.0   2054.7   2402.8   1591.1
vp9_inv_dct_dct_16x16_sub8_add_10_neon:    3848.6   3491.1   3845.7   2554.8
vp9_inv_dct_dct_16x16_sub12_add_10_neon:   5703.8   4831.6   5230.8   3493.4
vp9_inv_dct_dct_16x16_sub16_add_10_neon:   6399.5   5567.0   5832.4   3951.5
vp9_inv_dct_dct_32x32_sub1_add_10_neon:    1722.1    938.5   1577.3   1044.5
vp9_inv_dct_dct_32x32_sub2_add_10_neon:   15003.5  11576.8  13105.8   9602.2
vp9_inv_dct_dct_32x32_sub4_add_10_neon:   15768.5  12677.2  13726.0  10138.1
vp9_inv_dct_dct_32x32_sub8_add_10_neon:   17278.8  14825.4  14907.5  11185.7
vp9_inv_dct_dct_32x32_sub12_add_10_neon:  22335.7  21544.5  20379.5  15019.8
vp9_inv_dct_dct_32x32_sub16_add_10_neon:  24165.6  23881.7  21938.6  16308.2
vp9_inv_dct_dct_32x32_sub20_add_10_neon:  31082.2  30860.9  26835.3  19711.3
vp9_inv_dct_dct_32x32_sub24_add_10_neon:  33102.6  31922.8  28638.3  21161.0
vp9_inv_dct_dct_32x32_sub28_add_10_neon:  35104.9  34867.5  30411.7  22621.2
vp9_inv_dct_dct_32x32_sub32_add_10_neon:  37438.1  39103.4  32217.8  24067.6

Signed-off-by: Martin Storsjö <martin@martin.st>
libavcodec/arm/vp9itxfm_16bpp_neon.S