3 ms·

Lemire and collaborators often write in C++ intrinsics, or thin platform-specific wrappers on top of them. I count ~8 different implementations [1], which demo

by janwas 2y ago

Lemire and collaborators often write in C++ intrinsics, or thin platform-specific wrappers on top of them. I count ~8 different implementations [1], which demonstrates considerable commitment :) Personally, I prefer to write once with portable intrinsics. https://github.com/simdutf/simdutf/tree/1d5b5cd2b60850954df59caa8691ce10ebf04cfe/src https://github.com/simdutf/simdutf/tree/1d5b5cd2b60850954df5...

mananaysiempre 2y ago
I don’t think that really works beyond 128 bits. AVX/AVX2 with its weird tiered 2×128-bit structure occasionally forces you down some really awkward paths. AVX-512 with its masks is powerful and quite elegant, but the idioms it enables don’t really translate to other ISAs. And even in 128-bit land, the difference between good code for SSE and NEON can be quite significant, all because of a single instruction (PMOVMSKB) the latter lacks[1]. Portable instructions and “scalable” length-agnostic vectors are usually fine for straightforward mathy (type-1[2]) SIMD code. But real SIMD trickery, the one that starts as soon as you take your variable byte shuffle out of the stable, is rarely so kind. [1] https://branchfree.org/2019/04/01/fitting-my-head-through-the-arm-holes-or-two-sequences-to-substitute-for-the-missing-pmovmskb-instruction-on-arm-neon/ https://branchfree.org/2019/04/01/fitting-my-head-through-th... [2] https://branchfree.org/2024/06/09/a-draft-taxonomy-of-simd-usage/ https://branchfree.org/2024/06/09/a-draft-taxonomy-of-simd-u...
- janwas 2y ago
  Valid points, but I am doing exactly that full-time for the past years :) Including compression and quicksort, which are nontrivial uses of SIMD. The 2x128 is handled by adopting AVX2 semantics for our InterleaveLower/Upper op, and when we want the extra fixup, there is also InterleaveWholeLower. I agree AVX-512 is awesome, it's my favorite target. Some of its useful ops (compress[store]) are fine to emulate. What idioms did you have in mind that don't translate? Spot on about PMOVMSKB, that's a pain point. We provide several ops to replace it with more relaxed semantics that are easier to emulate on other platforms, for example AllTrue, CountTrue, AllFalse. (I think DanilaK's NEON VSHRN trick is faster than the one from your first link?)

by janwas 2y ago

mananaysiempre 2y ago

I don’t think that really works beyond 128 bits. AVX/AVX2 with its weird tiered 2×128-bit structure occasionally forces you down some really awkward paths. AVX-512 with its masks is powerful and quite elegant, but the idioms it enables don’t really translate to other ISAs. And even in 128-bit land, the difference between good code for SSE and NEON can be quite significant, all because of a single instruction (PMOVMSKB) the latter lacks[1]. Portable instructions and “scalable” length-agnostic vectors are usually fine for straightforward mathy (type-1[2]) SIMD code. But real SIMD trickery, the one that starts as soon as you take your variable byte shuffle out of the stable, is rarely so kind. [1] https://branchfree.org/2019/04/01/fitting-my-head-through-the-arm-holes-or-two-sequences-to-substitute-for-the-missing-pmovmskb-instruction-on-arm-neon/ https://branchfree.org/2019/04/01/fitting-my-head-through-th... [2] https://branchfree.org/2024/06/09/a-draft-taxonomy-of-simd-usage/ https://branchfree.org/2024/06/09/a-draft-taxonomy-of-simd-u...

janwas 2y ago
Valid points, but I am doing exactly that full-time for the past years :) Including compression and quicksort, which are nontrivial uses of SIMD. The 2x128 is handled by adopting AVX2 semantics for our InterleaveLower/Upper op, and when we want the extra fixup, there is also InterleaveWholeLower. I agree AVX-512 is awesome, it's my favorite target. Some of its useful ops (compress[store]) are fine to emulate. What idioms did you have in mind that don't translate? Spot on about PMOVMSKB, that's a pain point. We provide several ops to replace it with more relaxed semantics that are easier to emulate on other platforms, for example AllTrue, CountTrue, AllFalse. (I think DanilaK's NEON VSHRN trick is faster than the one from your first link?)