Provide arm64 neon optimized implementations for hscale16To19 with
filter sizes 4, 8 and X4.
The tests and benchmarks run on AWS Graviton 2 instances.
The results from a checkasm tool are shown below.
hscale_16_to_19__fs_4_dstW_512_c: 6216.0
hscale_16_to_19__fs_4_dstW_512_neon: 2257.0
hscale_16_to_19__fs_8_dstW_512_c: 10417.7
hscale_16_to_19__fs_8_dstW_512_neon: 3112.5
hscale_16_to_19__fs_12_dstW_512_c: 14890.5
hscale_16_to_19__fs_12_dstW_512_neon: 3899.0
hscale_16_to_19__fs_16_dstW_512_c: 19006.5
hscale_16_to_19__fs_16_dstW_512_neon: 5341.2
hscale_16_to_19__fs_32_dstW_512_c: 36629.5
hscale_16_to_19__fs_32_dstW_512_neon: 9502.7
hscale_16_to_19__fs_40_dstW_512_c: 45477.5
hscale_16_to_19__fs_40_dstW_512_neon: 11552.0
(Note, the checkasm tests for these functions haven't been
merged since they fail on x86.)
Signed-off-by: Hubert Mazur <hum@semihalf.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Add arm64 neon implementations for hscale 16 to 15 with filter
sizes 4, 8 and X4.
The tests and benchmarks run on AWS Graviton 2 instances.
The results from a checkasm tool are shown below.
hscale_16_to_15__fs_4_dstW_512_c: 6703.5
hscale_16_to_15__fs_4_dstW_512_neon: 2298.0
hscale_16_to_15__fs_8_dstW_512_c: 10983.0
hscale_16_to_15__fs_8_dstW_512_neon: 3216.5
hscale_16_to_15__fs_12_dstW_512_c: 15526.0
hscale_16_to_15__fs_12_dstW_512_neon: 3993.0
hscale_16_to_15__fs_16_dstW_512_c: 20183.5
hscale_16_to_15__fs_16_dstW_512_neon: 5369.7
hscale_16_to_15__fs_32_dstW_512_c: 39315.2
hscale_16_to_15__fs_32_dstW_512_neon: 9511.2
hscale_16_to_15__fs_40_dstW_512_c: 48995.7
hscale_16_to_15__fs_40_dstW_512_neon: 11570.0
(Note, the checkasm tests for these functions haven't been
merged since they fail on x86.)
Signed-off-by: Hubert Mazur <hum@semihalf.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Add arm64 neon implementations for hscale 8 to 19 with filter
sizes 4, 4X and 8. Both implementations are based on very similar ones
dedicated to hscale 8 to 15. The major changes refer to saving
the data - instead of writing the result as int16_t it is done
with int32_t.
These functions are heavily inspired on patches provided by J. Swinney
and M. Storsjö for hscale8to15 which were slightly adapted for
hscale8to19.
The tests and benchmarks run on AWS Graviton 2 instances. The results
from a checkasm tool shown below.
hscale_8_to_19__fs_4_dstW_512_c: 5663.2
hscale_8_to_19__fs_4_dstW_512_neon: 1259.7
hscale_8_to_19__fs_8_dstW_512_c: 9306.0
hscale_8_to_19__fs_8_dstW_512_neon: 2020.2
hscale_8_to_19__fs_12_dstW_512_c: 12932.7
hscale_8_to_19__fs_12_dstW_512_neon: 2462.5
hscale_8_to_19__fs_16_dstW_512_c: 16844.2
hscale_8_to_19__fs_16_dstW_512_neon: 4671.2
hscale_8_to_19__fs_32_dstW_512_c: 32803.7
hscale_8_to_19__fs_32_dstW_512_neon: 5474.2
hscale_8_to_19__fs_40_dstW_512_c: 40948.0
hscale_8_to_19__fs_40_dstW_512_neon: 6669.7
Signed-off-by: Hubert Mazur <hum@semihalf.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Treat the 32 bit stride registers as signed.
Alternatively, we could make the stride arguments ptrdiff_t instead
of int, and changing all of the assembly to operate on these
registers with their full 64 bit width, but that's a much larger
and more intrusive change (and risks missing some operation, which
would clamp the intermediates to 32 bit still).
Fixes: https://trac.ffmpeg.org/ticket/9985
Signed-off-by: Martin Storsjö <martin@martin.st>
This commit adds new code paths for vscale when filterSize is 2, 4, or
8. By using specialized code with unrolling to match the filterSize we
can improve performance.
On AWS c7g (Graviton 3, Neoverse V1) instances:
before after
yuv2yuvX_2_0_512_accurate_neon: 558.8 268.9
yuv2yuvX_4_0_512_accurate_neon: 637.5 434.9
yuv2yuvX_8_0_512_accurate_neon: 1144.8 806.2
yuv2yuvX_16_0_512_accurate_neon: 2080.5 1853.7
Signed-off-by: Jonathan Swinney <jswinney@amazon.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Use scalar times vector multiply accumlate instructions instead of
vector times vector to remove the need for replicating load instructions
which are slightly slower.
On AWS c7g (Graviton 3, Neoverse V1) instances:
yuv2yuvX_8_0_512_accurate_neon: 1144.8 987.4
yuv2yuvX_16_0_512_accurate_neon: 2080.5 1869.4
Signed-off-by: Jonathan Swinney <jswinney@amazon.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
This specialization handles the case where filtersize is 4 mod 8, e.g.
12, 20, etc. Aarch64 was previously using the c function for this case.
This implementation speeds up that case significantly.
hscale_8_to_15__fs_12_dstW_512_c: 6234.1
hscale_8_to_15__fs_12_dstW_512_neon: 1505.6
Signed-off-by: Jonathan Swinney <jswinney@amazon.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
This patch adds code to support specializations of the hscale function
and adds a specialization for filterSize == 4.
ff_hscale8to15_4_neon is a complete rewrite. Since the main bottleneck
here is loading the data from src, this data is loaded a whole block
ahead and stored back to the stack to be loaded again with ld4. This
arranges the data for most efficient use of the vector instructions and
removes the need for completion adds at the end. The number of
iterations of the C per iteration of the assembly is increased from 4 to
8, but because of the prefetching, there must be a special section
without prefetching when dstW < 16.
This improves speed on Graviton 2 (Neoverse N1) dramatically in the case
where previously fs=8 would have been required.
before: hscale_8_to_15__fs_8_dstW_512_neon: 1962.8
after : hscale_8_to_15__fs_4_dstW_512_neon: 1220.9
Signed-off-by: Jonathan Swinney <jswinney@amazon.com>
Signed-off-by: Martin Storsjö <martin@martin.st>
Some files currently rely on libavutil/cpu.h to include it for them;
yet said file won't use include it any more after the currently
deprecated functions are removed, so include attributes.h directly.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
We return 0 for this particular architecture but should instead be
returning the number of lines.
Fixes users who check the return value matches what they expect.
The NEON hscale function only supports X8 filter sizes and should only
be selected when these are being used. At the moment filterAlign is
set to 8 but in the future when extra NEON assembly for specific sizes is
added they will need to have checks here too.
The immediate usecase for this change is making the hscale checkasm
test easier and without NEON specific edge-cases (x86 already has these
guards).
Signed-off-by: Josh de Kock <josh@itanimul.li>
The x18 is a reserved platform register on Darwin and Windows.
x8/w8 seems to be unused in this function though (and same about
x10 and x14), so there's really no reason to use x18 here - just change
the uses of x18/w18 into x8/w8 instead without any further rewrites.
Signed-off-by: Martin Storsjö <martin@martin.st>
This patch rewrites the innermost loop of ff_yuv2planeX_8_neon to avoid zips and
horizontal adds by using fused multiply adds. The patch also uses ld1r to load
one element and replicate it across all lanes of the vector. The patch also
improves the clipping code by removing the shift right instructions and
performing the shift with the shift-right narrow instructions.
I see 8% difference on an m6g instance with neoverse-n1 CPUs:
$ ffmpeg -nostats -f lavfi -i testsrc2=4k:d=2 -vf bench=start,scale=1024x1024,bench=stop -f null -
before: t:0.014015 avg:0.014096 max:0.015018 min:0.013971
after: t:0.012985 avg:0.013013 max:0.013996 min:0.012818
Tested with `make check` on aarch64-linux.
Signed-off-by: Sebastian Pop <spop@amazon.com>
Reviewed-by: Clément Bœsch <u@pkh.me>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This patch implements ff_hscale_8_to_15_neon with NEON fused multiply accumulate
and bumps the vectorization factor from 2 to 4.
The speedup is of 25% on Graviton1 A1 instances based on A-72 cpus:
$ ffmpeg -nostats -f lavfi -i testsrc2=4k:d=2 -vf bench=start,scale=1024x1024,bench=stop -f null -
before: t:0.040303 avg:0.040287 max:0.040371 min:0.039214
after: t:0.032168 avg:0.032215 max:0.033081 min:0.032146
The speedup is of 39% on Graviton2 m6g instances based on Neoverse-N1 cpus:
$ ffmpeg -nostats -f lavfi -i testsrc2=4k:d=2 -vf bench=start,scale=1024x1024,bench=stop -f null -
before: t:0.019446 avg:0.019423 max:0.019493 min:0.019181
after: t:0.014015 avg:0.014096 max:0.015018 min:0.013971
Tested with `make check` on aarch64-linux.
Signed-off-by: Sebastian Pop <spop@amazon.com>
Reviewed-by: Jean-Baptiste Kempf <jb@videolan.org>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
y_offset and y_coeff being successive 32-bit integers, they are packed
into 8 bytes instead of 2x8 bytes.
See https://developer.apple.com/library/ios/documentation/Xcode/Conceptual/iPhoneOSABIReference/Articles/ARM64FunctionCallingConventions.html
> iOS diverges from Procedure Call Standard for the ARM 64-bit
> Architecture in several ways
[...]
> In the generic procedure call standard, all function arguments passed
> on the stack consume slots in multiples of 8 bytes. In iOS, this
> requirement is dropped, and values consume only the space required.
[...]
> Padding is still inserted on the stack to satisfy arguments’ alignment
> requirements.
