[danlark1]

Great work!

Popular narrative that NEON does not have a move mask alternative. Some time ago I published an article to simulate popular bit packing use cases with NEON with 1-2 instructions. This does not include unpacking cases but can be great for real world applications like compare+find, compare+iterate, compare+test.

https://community.arm.com/arm-community-blogs/b/servers-and-...

[Sesse__]

I never understood why they couldn't just include a movmskb instruction to begin with. It's massively useful for integer tasks, not expensive to implement as far as I know, and the vshrn+mov trick often requires an extra instruction either in front or after (in addition to just being, well, pretty obscure).

NEON in general is a bit sad, really; it's built around the idea of being implementable with a 64-bit ALU, and it shows. And SVE support is pretty much non-existent on the client.

[ashtonsix]

Not having (1 << k) - 1 as a single instruction sucks when it HAS to be in a hot loop, but you can usually hoist this to the loop prolougue: my stuff uses dummy inline assembly hints to force compilers to do this `asm volatile("" : "+w"(m));`.

I personally think calibrating ARM's ISA on smaller VL was a good choice: you get much better IPC. You also have an almost-complete absence of support for 8-bit elements with x86 ISAs, so elements per instruction is tied. And NEON kind-of-ish makes up for its small VL with multi-register TBL/TBX and LDP/STP.

Also: AVX512 is just as non-existent on clients as SVE2; although not really relevant for the server-side targets I'm optimising for (mostly OLAP).

[TinkersW]

8 bit is not absent in x86 SIMD, it is a slightly less covered than 32 & 16 bit, but you can fully implement all the common 8 bit ops and most are 1 instruction(with AVX2). There are even various horizontal ops on 8 bit values(avg, dot etc).

Also AVX512 is way more common than SVE2, all Zen4 & Zen5 support it.

[dzaima]

More specifically, basically the only absent 8-bit ops that have 32-bit equivalents in AVX2 are shifts and multiplies. Shifts are quite annoying (though, with a uniform shift they can be emulated on AVX-512 via GFNI abuse in 1 instr), multiplies are rather rare (though note that there is vpmaddubsw for an 8-bit→16-bit multiply-add). There's even a case of the opposite - saturating add/sub exist for 8-bit and 16-bit ints, but not wider.

[namibj]

GFNI is distinct from AVX-512; it was merely introduced in cores that also had AVX-512.

[Sesse__]

Does _any_ SIMD instruction set have (1 << k) - 1 as a single instruction?

[camel-cdr]

Not sure in which context this is used, but you can do -1 << k in most ISAs but that still requires a bit-not. But if you want to use the value in a bitwise instruction, then there are often variants that can invert the input operand.

E.g. in RVV instead of vand.vv(a, vadd.vi(vsll.vv(1,k),-1)) you could do vandn.vv(vsll.vv(-1,k))

AVX-512 can do this with any binary or ternary bitwise logic function via vpternlog

[Sesse__]

I don't know either; I was talking about the lack of PMOVMSKB in NEON and then the comment I replied to started talking about (1 << k) - 1 not being a single instruction sucks. Which I don't think it is in any non-NEON set either.

[ashtonsix]

Nice article! I personally find the ARM ISA far more cohesive than x86's: far less historical quirks. I also really appreciate the ubiquity of support for 8-bit elements in ARM and the absence of SMT (make performance much more predictable).