dav1d 0.5.1 Boosts AV1 Video Decode For Older CPUs by 40~50%

[archsway]

to stop working it has to start working. and it can't until you buy one of the three 5 year old SoC's with upstream hardware video decoding support (but you still need to use a special video player as the driver doesn't use any of the standard APIs.)

If you're lucky, as well as MJPEG you also get h264 support.

Fixed that for you.

EDIT: And hopefully you picked a board without USB controllers that cause kernel panics if you don't insert the cable just so.

[pal666]

hardware decoding does not stop working after 3 years

to stop working it has to start working. and it can't until you buy new hardware

[microcode]

Using NEON media instructions (aka ARMv7 and later) only. That's significant even if it is not a true ASIC hardware decoder, and it's not really true "software" decode, any more than AES-NI (crypto acceleration instructions in x86 CPUs) is "software crypto".

AES-NI is still software crypto, just like if you added a single-instruction ChaCha quarter round it would still be software. Sure, it's specialized, but AES-NI instructions can be, and in fact are, used for things other than AES encryption. And far more than these examples, packed SIMD (SSE2, NEON) is extremely standard, and very general-purpose. In fact, SSE2 is the only common way that floating point operations are used on x86.

[starshipeleven]

Quite naive if you think there's that few.

Quite naive if you think there's so much.

[schmidtbag]

Yeah, there are dozens, dozens of them, I tell you.

Quite naive if you think there's that few.

[starshipeleven]

Note how we routinely decode VP9 in software on phones.

Using NEON media instructions (aka ARMv7 and later) only. That's significant even if it is not a true ASIC hardware decoder, and it's not really true "software" decode, any more than AES-NI (crypto acceleration instructions in x86 CPUs) is "software crypto".

[starshipeleven]

Well, anyone who keeps a smartphone more than 3 years might disagree with you.

hardware decoding does not stop working after 3 years

[starshipeleven]

Tell that to nearly every ARM Linux user out there with a modern kernel who isn't using a Broadcom or Nvidia chip...

Yeah, there are dozens, dozens of them, I tell you.

[Orphis]

No sane person would do software encoding or decoding of video on typical ARM hardware. These chips have hardware support for video for a good reason. Optimizing for ARM is purely academic.

There are legitimate uses of software AV1. It's not only used for 4k content, but also low bitrate streams. If you try doing anything real-time but you are network-bound, why not use more CPU to alleviate that issue? It's all a tradeoff and perfectly valid sometimes.

[andreano]

Oh my, this release is meant for me! Not only have I a 7 year old AMD desktop CPU that doesn't even support SSSSE3, I also have a RPi4 running Raspbian, you know, with armv7 userland. Not sure which one is fastest.

I've tried Dav1d on Aarch64 (Pico-Pi iMX8M), and it beat my (almost as old) x86-64 laptop:
https://openbenchmarking.org/result/...SK-AMDE1120009