wasmerio/c-wasm-simd128-example

Example C++ repo emitting Wasm SIMD 128 instructions

c-wasm-simd128-example's Issues

Analysis of benchmark

Nice work!

I looked into this benchmark with @tlively now. It's interesting, but we're not sure what's going on here. The main issue is that the SIMD code emitted by the LLVM wasm backend is quite poor, e.g., here is horizontal_add:

 (func $horizontal_add\28float\20vector\5b4\5d\29 (; 28 ;) (type $7) (param $0 v128) (result f32)
  (f32.add
   (f32x4.extract_lane 0
    (local.tee $0
     (f32x4.add
      (local.get $0)
      (v8x16.shuffle 8 9 10 11 12 13 14 15 0 0 0 0 0 0 0 0
       (local.get $0)
       (local.get $0)
      )
     )
    )
   )
   (f32x4.extract_lane 1
    (local.get $0)
   )
  )
 )

That looks quite painful as really just 3 scalar adds are needed here :( But this seems like a limitation of the current wasm SIMD spec. Indeed, testing on V8, the SIMD version is significantly slower (80%) than the scalar version.

So it is puzzling that you see such a big speedup in the wasmer LLVM backend when running on the SIMD wasm. We would guess that LLVM gets this unoptimal SIMD wasm code and autovectorizes it into something fast. That is slightly puzzling though as in simple code like this LLVM might also be able to do the same for the non-SIMD wasm as well. In other words, it seems like the speedup you see is due to the SIMD wasm having slightly better "hinting" that the wasmer LLVM backend takes advantage of?

cc @sunfishcode who may have ideas on what's going on, and may be interested in this example of a wasm VM doing very powerful backend optimizations, that at least some other VMs do not expect.

Scattering factor

Just curious, noticed that scatter velocity component between two particles is inverse proportional to distance to the power or 4, since distance function seems to return a square of the distance (as it lacks a square root) and update squares its output. Wondering if that is intentional, I know that not every Wasm runtime implements SIMD square root.