But admit this boost is only seen in ‘an obscure filter’.
Just goes to show that when you work on a project that’s so fundamentally useful that it doesn’t require constant pivoting, you can not only address all your tech debt, but actually get around to those optimizations you brainstormed.
These guys are true assembly wizards
AVX512, SIMD
It’s not just “handwritten assembly”, it’s all intrinsics, again. The reason a lot of tech that needs to use fast matrix algebra (or fast numeric math in general) tries to use the same small set of libraries, tightly optimized to use those optimized instruction sets.
Assembly pff…
The future is vibe coding! If it had been made with Claudepilot with the right types-question guy, it would be 10,000,000 times speedier.
What does ‘handwritten assembly code’ mean? I assume they’re not writing it in cursive. Apparently some code is compiler created so it sounds like this is a synonym for code written by a person? There must be a better term for describing this than ‘handwritten’.
It means someone wrote actual assembly code instead of writing C code and relying on the compiler to generare optimal assembly for them.
Thank you 🙏
Yes, I believe you understand correctly. The vast majority of all native code is written by other computer programs (compilers). Handwritten assembly code usually means code that was instead produced more directly by a human programmer, most probably with the help of an assembler rather than compiling from a higher-level language.
Here’s an example of inline assembly in C++. You can write assembly inside your higher level code for performance optimizations to just doing really specific things that you can only really do at an assembly level. I’ve never done it before but it definitely is cool when people do it.
Assembly Shudders
I wish they made a 100x leap to the AV1 encoder. Even on my 12900k it runs at like 0.001x real time, which is… well, unusably slow.
Are you using ‘libaom-av1’ or ‘libsvtav1’? Libaom-av1 is very known for being very slow.
I was under the impression that libsvtav1 was still underdeveloped, but turns out the ffmpeg documentation for it was just lacking. Looks to be pretty good quality and even supports two-pass (which the documentation doesn’t mention).
such a powerful tool
