Dav1d 0.9.2 Released With More SSSE3, SSE4, AVX2, NEON Optimizations

[pegasus]

That's a weird assumption.

Why? I'd also like to know what is the oldest cpu capable of running these modern things ... would be a great input to all people recycling "old" computers and giving them a second life. For example, what is the oldest pc running modern software capable of playing lets say HD youtube stream? Or whatever else you want to define as minimum level of acceptance on today's internet ...

[brad0]

Why?

Just because I have the latest and greatest doesn't mean I want apps running inefficiently. This is common sense.

[grok]

Why? I'd also like to know what is the oldest cpu capable of running these modern things ... would be a great input to all people recycling "old" computers and giving them a second life. For example, what is the oldest pc running modern software capable of playing lets say HD youtube stream? Or whatever else you want to define as minimum level of acceptance on today's internet ...

Typically, you get a CPU that can do everything you need it to, except it can't play 1080p60 and such. Don't then, 720p30 is enough. Note that SSE2 is mandatory on firefox, SSE3 on chrome. So, a lot of hardware is out. You can also ask youtube to send you h264 (using software means).
Well, let's cook a new PC "spec" : if it's got HDMI output it's good enough (although, Core2Duo with VGA-only Intel GPU is good enough too)

Need AV1 360p, 480p files or streams to test on crappy computers, but I think these streams will still be h264 in practice (or vp9 on youtube). Not too hyped on AV1 - in the context of a desktop, where the user needs a h264 decoder be installed else most videos don't work.

[flashmozzg]

Why? I'd also like to know what is the oldest cpu capable of running these modern things ... would be a great input to all people recycling "old" computers and giving them a second life. For example, what is the oldest pc running modern software capable of playing lets say HD youtube stream? Or whatever else you want to define as minimum level of acceptance on today's internet ...

IRC, old Core 2 Quads were able to handle 1080p60fps with the latest updates.

[schmidtbag]

That's a weird assumption.

Why? High-end systems don't struggle to do un-optimized video decoding. They might not do it efficiently, but most people don't buy such systems to consume media efficiently. So, knowing how low-end CPUs handle it is important, because that's what you're more likely to find in a HTPC.

[schmidtbag]

Low end CPUs won't have extensions like AVX2.

I don't know of any modern x86 CPU lacking AVX2. Even Tiger Lake Pentiums get it. But let's say you're right anyway: the whole point of what this article is reporting on is so CPUs lacking AVX2 will get further optimizations.

[brad0]

Why? High-end systems don't struggle to do un-optimized video decoding. They might not do it efficiently, but most people don't buy such systems to consume media efficiently. So, knowing how low-end CPUs handle it is important, because that's what you're more likely to find in a HTPC.

That doesn't mean you go out and write poorly written inefficient software. That is common sense. But that is seriously lacking nowadays.

[schmidtbag]

That doesn't mean you go out and write poorly written inefficient software. That is common sense. But that is seriously lacking nowadays.

I know that... in what way did anything I say suggest otherwise? The whole point of what I said is that highly-optimized code allows for low-end systems to do things they otherwise can't or would at least struggle with. So, it's worth seeing what a low-end CPU can do.

[andreano]

Just because I have the latest and greatest doesn't mean I want apps running inefficiently.

They aren't replacing AVX2 with SSSE3, if that's what you thought. They have multiple versions of each function that they want to optimize, each with a "plain C" version that they must have for compatibility, and are setting those function pointers at initialization time or something. In this scheme, adding yet another variant of such a function has no runtime overhead to speak of.

The difference is that computers that don't have AVX2 can run the SSSE3 instead of the C code, which helps massively. That's why those computers are what's really interesting in this case.

I'm writing this on an AMD Llano that will be 10 years this winter, and it is one of those. Remember that AVX2 came later on the AMD side. Its performance is good for anything I need, except AV1 decoding (drops frames above 480p, whereas VP9 is fine up to 720p).