Hacker News Reader

Comment by snovv_crash | Hacker News Reader

I wonder whether it would be possible to add aftermarket AVX-512 instruction handling? Not for performance necessarily, but for compatibility.

View on HN

We have robust support for alternate code paths based on CPU flags after many years of this kind of thing. Is that really necessary or useful?

Custom microcode handling is a lot more brittle and chip- specific yet nearly equivalent to overloading in software your call to some avx512 op.

You're right, it isn't needed

The "last time" this was done was in the x87 days, where if the math coprocessor wasn't installed you could trap the corresponding interrupt and handle it to emulate the instructions.

This is done for MIPS FP emulation too - https://www.linux-mips.org/wiki/Floating_point#The_Linux_ker...

> The "last time" this was done was in the x87 days, where if the math coprocessor wasn't installed you could trap the corresponding interrupt and handle it to emulate the instructions.

Wasn't Hackintosh (or getting new OS X running on too old hardware) also using this technology to "support" CPUs without the newest SSEx instruction sets?

The tricky thing there would be those aren't just new instructions, they extend existing registers and add new ones. Where do you fit the extra bytes?

AMD already implements 256bit operations in terms of 2 128bit operations. Seems like going to 4 wouldn't be a stretch, at least for some of the simpler operations. Seems like a subset of operations would be possible.

I think Zen2 implements 256 bit instructions natively [0]. For AVX512, the new instructions [1] rather than the floating point arithmetic will be a problem IMHO. Emulating them with the microcode will be expensive and will provide no performance gains.

[0]: https://en.wikichip.org/wiki/amd/microarchitectures/zen_2

[1]: https://en.wikipedia.org/wiki/AVX-512

> AMD already implements 256bit operations in terms of 2 128bit operations.

Current Zen has actual 256-bit registers, it just doesn't have the execution units to process the whole register at once. It's not really the same thing.

Unlikely, IIRC you only had someting in the order of 32 three-instruction slots of memory.

I guess it would be easier to do that with virtualization. Advertise the capability to the VM, catch illegal instructions, emulate the missing pieces.

Reply on news.ycombinator.com