> My mistake in retrospect was using small arrays as part of a struct, which being immutable got replaced at each time step with a new struct requiring new arrays to be allocated and initialized. I would not have done that in c++, but julia puts my brain in matlab mode.
I see. Yes, it's an interesting design space where Julia makes both heap and stack allocations easy enough, so sometimes you just reach for the heap like in MATLAB mode. Hopefully Prem and Shuhei's work lands soon enough to stack allocate small non-escaping arrays so that user's done need to think about this.
> Alignment I'd assumed, but padding the struct instead of the tuple did nothing, so probably extra work to clear a piece of an simd load. Any insight on why avx availability didn't help would be appreciated. I did verify some avx instructions were in the asm it generated, so it knew, it just didn't use.
The major differences at this point seem to come down to GCC (g++) vs LLVM and proofs of aliasing. LLVM's auto-vectorizer isn't that great, and it seems to be able to prove 2 arrays are not aliasing less reliably. For the first part, some people have just improved the loop analysis code from the Julia side (https://github.com/JuliaSIMD/LoopVectorization.jl), forcing SIMD onto LLVM can help it make the right choices. But for the second part you do need to do `@simd ivdep for ...` (or use LoopVectorization.jl) to match some C++ examples. This is hopefully one of the things that the JET.jl and other new analysis passes can help with, along with the new effects system (see https://github.com/JuliaLang/julia/pull/43852, this is a pretty huge new compiler feature in v1.8, but right now it's manually specified and will take time before things like https://github.com/JuliaLang/julia/pull/44822 land and start to make it more pervasive). When that's all together, LLVM will have more ammo for proving things more effectively (pun intended).
I see. Yes, it's an interesting design space where Julia makes both heap and stack allocations easy enough, so sometimes you just reach for the heap like in MATLAB mode. Hopefully Prem and Shuhei's work lands soon enough to stack allocate small non-escaping arrays so that user's done need to think about this.
> Alignment I'd assumed, but padding the struct instead of the tuple did nothing, so probably extra work to clear a piece of an simd load. Any insight on why avx availability didn't help would be appreciated. I did verify some avx instructions were in the asm it generated, so it knew, it just didn't use.
The major differences at this point seem to come down to GCC (g++) vs LLVM and proofs of aliasing. LLVM's auto-vectorizer isn't that great, and it seems to be able to prove 2 arrays are not aliasing less reliably. For the first part, some people have just improved the loop analysis code from the Julia side (https://github.com/JuliaSIMD/LoopVectorization.jl), forcing SIMD onto LLVM can help it make the right choices. But for the second part you do need to do `@simd ivdep for ...` (or use LoopVectorization.jl) to match some C++ examples. This is hopefully one of the things that the JET.jl and other new analysis passes can help with, along with the new effects system (see https://github.com/JuliaLang/julia/pull/43852, this is a pretty huge new compiler feature in v1.8, but right now it's manually specified and will take time before things like https://github.com/JuliaLang/julia/pull/44822 land and start to make it more pervasive). When that's all together, LLVM will have more ammo for proving things more effectively (pun intended).