[FFmpeg-devel] [PATCH 09/11] avutil/half2float: use native _Float16 if available

[Timo Rothenpieler]

On 10.08.2022 23:03, Andreas Rheinhardt wrote:
> Timo Rothenpieler:
>> _Float16 support was available on arm/aarch64 for a while, and with gcc
>> 12 was enabled on x86 as long as SSE2 is supported.
>>
>> If the target arch supports f16c, gcc emits fairly efficient assembly,
>> taking advantage of it. This is the case on x86-64-v3 or higher.
>> Without f16c, it emulates it in software using sse2 instructions.
> 
> How is the performance of this emulation compared to our current code?
> And how is the native _Float16 performance compared to the current code?

The performance of the sse2 emulation is actually surprisingly poor, in 
a quick test:

./ffmpeg -s 512x512 -f rawvideo -pix_fmt rgbaf16 -i /dev/zero -vf 
format=yuv444p -f null -

_Float16 full SSE2 emulation:
frame=50074 fps=848 q=-0.0 size=N/A time=00:33:22.96 bitrate=N/A speed=33.9x

_Float16 f16c accelerated (Zen2, --cpu=znver2):
frame=50636 fps=1965 q=-0.0 Lsize=N/A time=00:33:45.40 bitrate=N/A 
speed=78.6x

classic half2float full software implementation:
frame=49926 fps=1605 q=-0.0 Lsize=N/A time=00:33:17.00 bitrate=N/A 
speed=64.2x

Unfortunately I don't see a good way to runtime-detect the presence of 
f16c without going full self-written assembly, which would diminish the 
compilers ability to take advantage of f16c only ever operating on 4 or 
8 values at a time.
But the HAVE_FLOAT16 checks could be paired with a check for __F16C__, 
which seems to universally be the established define for "the code is 
being built f16c optimizations".

That at least avoids the case of the apparently quite slow sse2 emulation.