SeG | pytorch implementation of the paper : Self-Attention | Natural Language Processing library

The default -mtune=generic has -mprefer-vector-width=256, and -mavx2 doesn't change that.

znver1 implies -mprefer-vector-width=128, because that's all the native width of the HW. An instruction using 32-byte YMM vectors decodes to at least 2 uops, more if it's a lane-crossing shuffle. For simple vertical SIMD like this, 32-byte vectors would be ok; the pipeline handles 2-uop instructions efficiently. (And I think is 6 uops wide but only 5 instructions wide, so max front-end throughput isn't available using only 1-uop instructions). But when vectorization would require shuffling, e.g. with arrays of different element widths, GCC code-gen can get messier with 256-bit or wider.

And vmovdqa ymm0, ymm1 mov-elimination only works on the low 128-bit half on Zen1. Also, normally using 256-bit vectors would imply one should use vzeroupper afterwards, to avoid performance problems on other CPUs (but not Zen1).

I don't know how Zen1 handles misaligned 32-byte loads/stores where each 16-byte half is aligned but in separate cache lines. If that performs well, GCC might want to consider increasing the znver1 -mprefer-vector-width to 256. But wider vectors means more cleanup code if the size isn't known to be a multiple of the vector width.

Ideally GCC would be able to detect easy cases like this and use 256-bit vectors there. (Pure vertical, no mixing of element widths, constant size that's am multiple of 32 bytes.) At least on CPUs where that's fine: znver1, but not bdver2 for example where 256-bit stores are always slow due to a CPU design bug.

You can see the result of this choice in the way it vectorizes your first loop, the memset-like loop, with a vmovdqu [rdx], xmm0. https://godbolt.org/z/E5Tq7Gfzc

So given that GCC has decided to only use 128-bit vectors, which can only hold two uint64_t elements, it (rightly or wrongly) decides it wouldn't be worth using vpsllq / vpaddd to implement qword *5 as (v<<2) + v, vs. doing it with integer in one LEA instruction.

Almost certainly wrongly in this case, since it still requires a separate load and store for every element or pair of elements. (And loop overhead since GCC's default is not to unroll except with PGO, -fprofile-use. SIMD is like loop unrolling, especially on a CPU that handles 256-bit vectors as 2 separate uops.)

I'm not sure exactly what GCC means by "not vectorized: unsupported data-type". x86 doesn't have a SIMD uint64_t multiply instruction until AVX-512, so perhaps GCC assigns it a cost based on the general case of having to emulate it with multiple 32x32 => 64-bit pmuludq instructions and a bunch of shuffles. And it's only after it gets over that hump that it realizes that it's actually quite cheap for a constant like 5 with only 2 set bits?

That would explain GCC's decision-making process here, but I'm not sure it's exactly the right explanation. Still, these kinds of factors are what happen in a complex piece of machinery like a compiler. A skilled human can easily make smarter choices, but compilers just do sequences of optimization passes that don't always consider the big picture and all the details at the same time.

(The benchmarks in the question confirm that an actual Zen1 CPU gets a nearly 2x speedup, as expected from doing 2x uint64 in 6 uops vs. 1x in 5 uops with scalar. Or 4x uint64_t in 10 uops with 256-bit vectors, including two 128-bit stores which will be the throughput bottleneck along with the front-end.)

Even with -march=znver1 -O3 -mprefer-vector-width=256, we don't get the *= 5 loop vectorized with GCC9, 10, or 11, or current trunk. As you say, we do with -march=znver2. https://godbolt.org/z/dMTh7Wxcq

We do get vectorization with those options for uint32_t (even leaving the vector width at 128-bit). Scalar would cost 4 operations per vector uop (not instruction), regardless of 128 or 256-bit vectorization on Zen1, so this doesn't tell us whether *= is what makes the cost-model decide not to vectorize, or just the 2 vs. 4 elements per 128-bit internal uop.

With uint64_t, changing to arr[i] += arr[i]<<2; still doesn't vectorize, but arr[i] <<= 1; does. (https://godbolt.org/z/6PMn93Y5G). Even arr[i] <<= 2; and arr[i] += 123 in the same loop vectorize, to the same instructions that GCC thinks aren't worth it for vectorizing *= 5, just different operands, constant instead of the original vector again. (Scalar could still use one LEA). So clearly the cost-model isn't looking as far as final x86 asm machine instructions, but I don't know why arr[i] += arr[i] would be considered more expensive than arr[i] <<= 1; which is exactly the same thing.

GCC8 does vectorize your loop, even with 128-bit vector width: https://godbolt.org/z/5o6qjc7f6