The code originally pre-multiply by 2 the steps, causing the running sum
of the h factors to drift away due to the lack of precision. It quickly
causes an inaccuracy > 0.01.
I tried diverse approaches such as multiply by 2.0 (instead of adding
the value itself) without success.
I'm unable to bench the impact of this change, feel free to compare.
This commit fixes the incoming aacpsdsp tests.
Following is an alternative simplified function (matching the incoming
AArch64 code) that may be used:
function ff_ps_stereo_interpolate_neon, export=1
vld1.32 {q0}, [r2]
vld1.32 {q1}, [r3]
ldr r12, [sp]
vmov.f32 q8, q0
vmov.f32 q9, q1
vzip.32 q8, q0
vzip.32 q9, q1
1:
vld1.32 {d4}, [r0,:64]
vld1.32 {d6}, [r1,:64]
vadd.f32 q8, q8, q9
vadd.f32 q0, q0, q1
vmov.f32 d5, d4
vmov.f32 d7, d6
vmul.f32 q2, q2, q8
vmla.f32 q2, q3, q0
vst1.32 {d4}, [r0,:64]!
vst1.32 {d5}, [r1,:64]!
subs r12, r12, #1
bgt 1b
bx lr
endfunc
Instead, hardcode the use of the _arm implementation of add_pixels,
and use the C version for put_pixels (as no arm-optimized version
exists). Since there's separate implementations of idct{,_put,_add}
for neon, this has no practical impact on performance.
This work is sponsored by, and copyright, Google.
This reduces the code size of libavcodec/arm/vp9itxfm_16bpp_neon.o from
17500 to 14516 bytes.
This gives a small slowdown of a couple tens of cycles, up to around
150 cycles for the full case of the largest transform, but makes
it more feasible to add more optimized versions of these transforms.
Before: Cortex A7 A8 A9 A53
vp9_inv_dct_dct_16x16_sub4_add_10_neon: 4237.4 3561.5 3971.8 2525.3
vp9_inv_dct_dct_16x16_sub16_add_10_neon: 6371.9 5452.0 5779.3 3910.5
vp9_inv_dct_dct_32x32_sub4_add_10_neon: 22068.8 17867.5 19555.2 13871.6
vp9_inv_dct_dct_32x32_sub32_add_10_neon: 37268.9 38684.2 32314.2 23969.0
After:
vp9_inv_dct_dct_16x16_sub4_add_10_neon: 4375.1 3571.9 4283.8 2567.2
vp9_inv_dct_dct_16x16_sub16_add_10_neon: 6415.6 5578.9 5844.6 3948.3
vp9_inv_dct_dct_32x32_sub4_add_10_neon: 22653.7 18079.7 19603.7 13905.3
vp9_inv_dct_dct_32x32_sub32_add_10_neon: 37593.2 38862.2 32235.8 24070.9
Signed-off-by: Martin Storsjö <martin@martin.st>
Keep the idct32 coefficients in narrow form in q6-q7, and idct16
coefficients in lengthened 32 bit form in q0-q3. Avoid clobbering
q0-q3 in the pass1 function, and squeeze the idct16 coefficients
into q0-q1 in the pass2 function to avoid reloading them.
The idct16 coefficients are clobbered and reloaded within idct32_odd
though, since that turns out to be faster than narrowing them and
swapping them into q6-q7.
Before: Cortex A7 A8 A9 A53
vp9_inv_dct_dct_32x32_sub4_add_10_neon: 22653.8 18268.4 19598.0 14079.0
vp9_inv_dct_dct_32x32_sub32_add_10_neon: 37699.0 38665.2 32542.3 24472.2
After:
vp9_inv_dct_dct_32x32_sub4_add_10_neon: 22270.8 18159.3 19531.0 13865.0
vp9_inv_dct_dct_32x32_sub32_add_10_neon: 37523.3 37731.6 32181.7 24071.2
Signed-off-by: Martin Storsjö <martin@martin.st>
Align the second/third operands as they usually are.
Due to the wildly varying sizes of the written out operands
in aarch64 assembly, the column alignment is usually not as clear
as in arm assembly.
This is cherrypicked from libav commit
7995ebfad1.
Signed-off-by: Martin Storsjö <martin@martin.st>
In the half/quarter cases where we don't use the min_eob array, defer
loading the pointer until we know it will be needed.
This is cherrypicked from libav commit
3a0d5e206d.
Signed-off-by: Martin Storsjö <martin@martin.st>
This reduces the number of lines and reduces the duplication.
Also simplify the eob check for the half case.
If we are in the half case, we know we at least will need to do the
first three slices, we only need to check eob for the fourth one,
so we can hardcode the value to check against instead of loading
from the min_eob array.
Since at most one slice can be skipped in the first pass, we can
unroll the loop for filling zeros completely, as it was done for
the quarter case before.
This allows skipping loading the min_eob pointer when using the
quarter/half cases.
This is cherrypicked from libav commit
98ee855ae0.
Signed-off-by: Martin Storsjö <martin@martin.st>
This matches the order they are in the 16 bpp version.
There they are in this order, to make sure we access them in the
same order they are declared, easing loading only half of the
coefficients at a time.
This makes the 8 bpp version match the 16 bpp version better.
This is cherrypicked from libav commit
08074c092d.
Signed-off-by: Martin Storsjö <martin@martin.st>
All elements are used pairwise, except for the first one.
Previously, the 16th element was unused. Move the unused element
to the second slot, to make the later element pairs not split
across registers.
This simplifies loading only parts of the coefficients,
reducing the difference to the 16 bpp version.
This is cherrypicked from libav commit
de06bdfe6c.
Signed-off-by: Martin Storsjö <martin@martin.st>
The idct32x32 function actually pushed q4-q7 onto the stack even
though it didn't clobber them; there are plenty of registers that
can be used to allow keeping all the idct coefficients in registers
without having to reload different subsets of them at different
stages in the transform.
Since the idct16 core transform avoids clobbering q4-q7 (but clobbers
q2-q3 instead, to avoid needing to back up and restore q4-q7 at all
in the idct16 function), and the lanewise vmul needs a register in
the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5
while doing idct16.
While keeping these coefficients in registers, we still can skip pushing
q7.
Before: Cortex A7 A8 A9 A53
vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7
After:
vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8
This is cherrypicked from libav commit
402546a172.
Signed-off-by: Martin Storsjö <martin@martin.st>
For this case, with 8 inputs but only changing 4 of them, we can fit
all 16 input pixels into a q register, and still have enough temporary
registers for doing the loop filter.
The wd=8 filters would require too many temporary registers for
processing all 16 pixels at once though.
Before: Cortex A7 A8 A9 A53
vp9_loop_filter_mix2_v_44_16_neon: 289.7 256.2 237.5 181.2
After:
vp9_loop_filter_mix2_v_44_16_neon: 221.2 150.5 177.7 138.0
This is cherrypicked from libav commit
575e31e931.
Signed-off-by: Martin Storsjö <martin@martin.st>
This adds lots of extra .ifs, but speeds it up by a couple cycles,
by avoiding stalls.
This is cherrypicked from libav commit
e18c39005a.
Signed-off-by: Martin Storsjö <martin@martin.st>
Previously we first calculated hev, and then negated it.
Since we were able to schedule the negation in the middle
of another calculation, we don't see any gain in all cases.
Before: Cortex A7 A8 A9 A53 A53/AArch64
vp9_loop_filter_v_4_8_neon: 147.0 129.0 115.8 89.0 88.7
vp9_loop_filter_v_8_8_neon: 242.0 198.5 174.7 140.0 136.7
vp9_loop_filter_v_16_8_neon: 500.0 419.5 382.7 293.0 275.7
vp9_loop_filter_v_16_16_neon: 971.2 825.5 731.5 579.0 453.0
After:
vp9_loop_filter_v_4_8_neon: 143.0 127.7 114.8 88.0 87.7
vp9_loop_filter_v_8_8_neon: 241.0 197.2 173.7 140.0 136.7
vp9_loop_filter_v_16_8_neon: 497.0 419.5 379.7 293.0 275.7
vp9_loop_filter_v_16_16_neon: 965.2 818.7 731.4 579.0 452.0
This is cherrypicked from libav commit
e1f9de86f4.
Signed-off-by: Martin Storsjö <martin@martin.st>
This allows reusing the macro for a separate implementation of the
pass2 function.
This is cherrypicked from libav commit
47b3c2c18d.
Signed-off-by: Martin Storsjö <martin@martin.st>
This work is sponsored by, and copyright, Google.
This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from
15324 to 12388 bytes.
This gives a small slowdown of a couple tens of cycles, up to around
150 cycles for the full case of the largest transform, but makes
it more feasible to add more optimized versions of these transforms.
Before: Cortex A7 A8 A9 A53
vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1
vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3
vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2
vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2
After:
vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0
vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5
vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1
vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9
This is cherrypicked from libav commit
0331c3f5e8.
Signed-off-by: Martin Storsjö <martin@martin.st>
This work is sponsored by, and copyright, Google.
This is pretty much similar to the 8 bpp version, but in some senses
simpler. All input pixels are 16 bits, and all intermediates also fit
in 16 bits, so there's no lengthening/narrowing in the filter at all.
For the full 16 pixel wide filter, we can only process 4 pixels at a time
(using an implementation very much similar to the one for 8 bpp),
but we can do 8 pixels at a time for the 4 and 8 pixel wide filters with
a different implementation of the core filter.
Examples of relative speedup compared to the C version, from checkasm:
Cortex A7 A8 A9 A53
vp9_loop_filter_h_4_8_10bpp_neon: 1.83 2.16 1.40 2.09
vp9_loop_filter_h_8_8_10bpp_neon: 1.39 1.67 1.24 1.70
vp9_loop_filter_h_16_8_10bpp_neon: 1.56 1.47 1.10 1.81
vp9_loop_filter_h_16_16_10bpp_neon: 1.94 1.69 1.33 2.24
vp9_loop_filter_mix2_h_44_16_10bpp_neon: 2.01 2.27 1.67 2.39
vp9_loop_filter_mix2_h_48_16_10bpp_neon: 1.84 2.06 1.45 2.19
vp9_loop_filter_mix2_h_84_16_10bpp_neon: 1.89 2.20 1.47 2.29
vp9_loop_filter_mix2_h_88_16_10bpp_neon: 1.69 2.12 1.47 2.08
vp9_loop_filter_mix2_v_44_16_10bpp_neon: 3.16 3.98 2.50 4.05
vp9_loop_filter_mix2_v_48_16_10bpp_neon: 2.84 3.64 2.25 3.77
vp9_loop_filter_mix2_v_84_16_10bpp_neon: 2.65 3.45 2.16 3.54
vp9_loop_filter_mix2_v_88_16_10bpp_neon: 2.55 3.30 2.16 3.55
vp9_loop_filter_v_4_8_10bpp_neon: 2.85 3.97 2.24 3.68
vp9_loop_filter_v_8_8_10bpp_neon: 2.27 3.19 1.96 3.08
vp9_loop_filter_v_16_8_10bpp_neon: 3.42 2.74 2.26 4.40
vp9_loop_filter_v_16_16_10bpp_neon: 2.86 2.44 1.93 3.88
The speedup vs C code measured in checkasm is around 1.1-4x.
These numbers are quite inconclusive though, since the checkasm test
runs multiple filterings on top of each other, so later rounds might
end up with different codepaths (different decisions on which filter
to apply, based on input pixel differences).
Based on START_TIMER/STOP_TIMER wrapping around a few individual
functions, the speedup vs C code is around 2-4x.
Signed-off-by: Martin Storsjö <martin@martin.st>
This work is sponsored by, and copyright, Google.
This is structured similarly to the 8 bit version. In the 8 bit
version, the coefficients are 16 bits, and intermediates are 32 bits.
Here, the coefficients are 32 bit. For the 4x4 transforms for 10 bit
content, the intermediates also fit in 32 bits, but for all other
transforms (4x4 for 12 bit content, and 8x8 and larger for both 10
and 12 bit) the intermediates are 64 bit.
For the existing 8 bit case, the 8x8 transform fit all coefficients in
registers; for 10/12 bit, when the coefficients are 32 bit, the 8x8
transform also has to be done in slices of 4 pixels (just as 16x16 and
32x32 for 8 bit).
The slice width also shrinks from 4 elements to 2 elements in parallel
for the 16x16 and 32x32 cases.
The 16 bit coefficients from idct_coeffs and similar tables also need
to be lenghtened to 32 bit in order to be used in multiplication with
vectors with 32 bit elements. This leads to the fixed coefficient
vectors needing more space, leading to more cases where they have to
be reloaded within the transform (in iadst16).
This technically would need testing in checkasm for subpartitions
in increments of 2, but that slows down normal checkasm runs
excessively.
Examples of relative speedup compared to the C version, from checkasm:
Cortex A7 A8 A9 A53
vp9_inv_adst_adst_4x4_sub4_add_10_neon: 4.83 11.36 5.22 6.77
vp9_inv_adst_adst_8x8_sub8_add_10_neon: 4.12 7.60 4.06 4.84
vp9_inv_adst_adst_16x16_sub16_add_10_neon: 3.93 8.16 4.52 5.35
vp9_inv_dct_dct_4x4_sub1_add_10_neon: 1.36 2.57 1.41 1.61
vp9_inv_dct_dct_4x4_sub4_add_10_neon: 4.24 8.66 5.06 5.81
vp9_inv_dct_dct_8x8_sub1_add_10_neon: 2.63 4.18 1.68 2.87
vp9_inv_dct_dct_8x8_sub4_add_10_neon: 4.52 9.47 4.24 5.39
vp9_inv_dct_dct_8x8_sub8_add_10_neon: 3.45 7.34 3.45 4.30
vp9_inv_dct_dct_16x16_sub1_add_10_neon: 3.56 6.21 2.47 4.32
vp9_inv_dct_dct_16x16_sub2_add_10_neon: 5.68 12.73 5.28 7.07
vp9_inv_dct_dct_16x16_sub8_add_10_neon: 4.42 9.28 4.24 5.45
vp9_inv_dct_dct_16x16_sub16_add_10_neon: 3.41 7.29 3.35 4.19
vp9_inv_dct_dct_32x32_sub1_add_10_neon: 4.52 8.35 3.83 6.40
vp9_inv_dct_dct_32x32_sub2_add_10_neon: 5.86 13.19 6.14 7.04
vp9_inv_dct_dct_32x32_sub16_add_10_neon: 4.29 8.11 4.59 5.06
vp9_inv_dct_dct_32x32_sub32_add_10_neon: 3.31 5.70 3.56 3.84
vp9_inv_wht_wht_4x4_sub4_add_10_neon: 1.89 2.80 1.82 1.97
The speedup compared to the C functions is around 1.3 to 7x for the
full transforms, even higher for the smaller subpartitions.
Signed-off-by: Martin Storsjö <martin@martin.st>
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>
This work is sponsored by, and copyright, Google.
This is more in line with how it will be extended for more bitdepths.
Signed-off-by: Martin Storsjö <martin@martin.st>
This work is sponsored by, and copyright, Google.
Previously all subpartitions except the eob=1 (DC) case ran with
the same runtime:
Cortex A7 A8 A9 A53
vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0
vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3
By skipping individual 4x16 or 4x32 pixel slices in the first pass,
we reduce the runtime of these functions like this:
vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8
vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7
vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5
vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7
vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1
vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1
vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9
vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5
vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8
vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5
vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2
vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6
vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6
vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0
vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7
vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1
I.e. in general a very minor overhead for the full subpartition case due
to the additional loads and cmps, but a significant speedup for the cases
when we only need to process a small part of the actual input data.
In common VP9 content in a few inspected clips, 70-90% of the non-dc-only
16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left
8x8 or 16x16 subpartitions respectively.
This is cherrypicked from libav commit
9c8bc74c2b.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This avoids reloading them if they haven't been clobbered, if the
first pass also was idct.
This is similar to what was done in the aarch64 version.
This is cherrypicked from libav commit
3c87039a40.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Since the same parameter is used for both input and output,
the name inout is more fitting.
This matches the naming used below in the dmbutterfly macro.
This is cherrypicked from libav commit
79566ec8c7.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This is one instruction less for thumb, and only have got
1/2 arm/thumb specific instructions.
This is cherrypicked from libav commit
e5b0fc170f.
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This work is sponsored by, and copyright, Google.
Previously all subpartitions except the eob=1 (DC) case ran with
the same runtime:
Cortex A7 A8 A9 A53
vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0
vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3
By skipping individual 4x16 or 4x32 pixel slices in the first pass,
we reduce the runtime of these functions like this:
vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8
vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7
vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5
vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7
vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1
vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1
vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9
vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5
vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8
vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5
vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2
vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6
vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6
vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0
vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7
vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1
I.e. in general a very minor overhead for the full subpartition case due
to the additional loads and cmps, but a significant speedup for the cases
when we only need to process a small part of the actual input data.
In common VP9 content in a few inspected clips, 70-90% of the non-dc-only
16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left
8x8 or 16x16 subpartitions respectively.
Signed-off-by: Martin Storsjö <martin@martin.st>
This avoids reloading them if they haven't been clobbered, if the
first pass also was idct.
This is similar to what was done in the aarch64 version.
Signed-off-by: Martin Storsjö <martin@martin.st>
Since the same parameter is used for both input and output,
the name inout is more fitting.
This matches the naming used below in the dmbutterfly macro.
Signed-off-by: Martin Storsjö <martin@martin.st>
This work is sponsored by, and copyright, Google.
The implementation tries to have smart handling of cases
where no pixels need the full filtering for the 8/16 width
filters, skipping both calculation and writeback of the
unmodified pixels in those cases. The actual effect of this
is hard to test with checkasm though, since it tests the
full filtering, and the benefit depends on how many filtered
blocks use the shortcut.
Examples of relative speedup compared to the C version, from checkasm:
Cortex A7 A8 A9 A53
vp9_loop_filter_h_4_8_neon: 2.72 2.68 1.78 3.15
vp9_loop_filter_h_8_8_neon: 2.36 2.38 1.70 2.91
vp9_loop_filter_h_16_8_neon: 1.80 1.89 1.45 2.01
vp9_loop_filter_h_16_16_neon: 2.81 2.78 2.18 3.16
vp9_loop_filter_mix2_h_44_16_neon: 2.65 2.67 1.93 3.05
vp9_loop_filter_mix2_h_48_16_neon: 2.46 2.38 1.81 2.85
vp9_loop_filter_mix2_h_84_16_neon: 2.50 2.41 1.73 2.85
vp9_loop_filter_mix2_h_88_16_neon: 2.77 2.66 1.96 3.23
vp9_loop_filter_mix2_v_44_16_neon: 4.28 4.46 3.22 5.70
vp9_loop_filter_mix2_v_48_16_neon: 3.92 4.00 3.03 5.19
vp9_loop_filter_mix2_v_84_16_neon: 3.97 4.31 2.98 5.33
vp9_loop_filter_mix2_v_88_16_neon: 3.91 4.19 3.06 5.18
vp9_loop_filter_v_4_8_neon: 4.53 4.47 3.31 6.05
vp9_loop_filter_v_8_8_neon: 3.58 3.99 2.92 5.17
vp9_loop_filter_v_16_8_neon: 3.40 3.50 2.81 4.68
vp9_loop_filter_v_16_16_neon: 4.66 4.41 3.74 6.02
The speedup vs C code is around 2-6x. The numbers are quite
inconclusive though, since the checkasm test runs multiple filterings
on top of each other, so later rounds might end up with different
codepaths (different decisions on which filter to apply, based
on input pixel differences). Disabling the early-exit in the asm
doesn't give a fair comparison either though, since the C code
only does the necessary calcuations for each row.
Based on START_TIMER/STOP_TIMER wrapping around a few individual
functions, the speedup vs C code is around 4-9x.
This is pretty similar in runtime to the corresponding routines
in libvpx. (This is comparing vpx_lpf_vertical_16_neon,
vpx_lpf_horizontal_edge_8_neon and vpx_lpf_horizontal_edge_16_neon
to vp9_loop_filter_h_16_8_neon, vp9_loop_filter_v_16_8_neon
and vp9_loop_filter_v_16_16_neon - note that the naming of horizonal
and vertical is flipped between the libraries.)
In order to have stable, comparable numbers, the early exits in both
asm versions were disabled, forcing the full filtering codepath.
Cortex A7 A8 A9 A53
vp9_loop_filter_h_16_8_neon: 597.2 472.0 482.4 415.0
libvpx vpx_lpf_vertical_16_neon: 626.0 464.5 470.7 445.0
vp9_loop_filter_v_16_8_neon: 500.2 422.5 429.7 295.0
libvpx vpx_lpf_horizontal_edge_8_neon: 586.5 414.5 415.6 383.2
vp9_loop_filter_v_16_16_neon: 905.0 784.7 791.5 546.0
libvpx vpx_lpf_horizontal_edge_16_neon: 1060.2 751.7 743.5 685.2
Our version is consistently faster on on A7 and A53, marginally slower on
A8, and sometimes faster, sometimes slower on A9 (marginally slower in all
three tests in this particular test run).
This is an adapted cherry-pick from libav commit
dd299a2d6d.
Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
This work is sponsored by, and copyright, Google.
For the transforms up to 8x8, we can fit all the data (including
temporaries) in registers and just do a straightforward transform
of all the data. For 16x16, we do a transform of 4x16 pixels in
4 slices, using a temporary buffer. For 32x32, we transform 4x32
pixels at a time, in two steps of 4x16 pixels each.
Examples of relative speedup compared to the C version, from checkasm:
Cortex A7 A8 A9 A53
vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01
vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98
vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16
vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46
vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86
vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79
vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42
vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77
Thus, the speedup vs C code is around 3-6x.
This is mostly marginally faster than the corresponding routines
in libvpx on most cores, tested with their 32x32 idct (compared to
vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's
favour since their version doesn't clear the input buffer like ours
do (although the effect of that on the total runtime probably is
negligible.)
Cortex A7 A8 A9 A53
vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9
libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5
Only on the Cortex A8, the libvpx function is faster. On the other cores,
ours is slightly faster even though ours has got source block clearing
integrated.
This is an adapted cherry-pick from libav commits
a67ae67083 and
52d196fb30.
Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
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
ffbd1d2b00,
392caa65df,
557c1675cf and
11623217e3.
Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
Clément Bœsch
Ronald S. Bultje
Clément Bœsch
Martin Storsjö
Janne Grunau