Performance improvements for benchmarks

[GabrielMajeri]

This pull request contains improvements to the performance of the example crates in this repository.

fannkuch_redux

• I noticed a small performance reduction because LLVM was emitting memory accesses to the self.odd field, insted of storing it in a register.
  Not sure why this happens, LLVM should know we borrow self mutably & have exclusive access to self and be able to cache its value locally.

• Overall, about 1.5% better performance.

stencil

See issue #95 for the full discussion.

• Now we're using mul_adde (FMA) where possible.
• Improved performance of Data::reinit by using unsafe memory accesses. Not part of the benchmark, but it was slow and showing up a lot in the profiler.
• The parallel code is slower than the SIMD code now:

simd: 175 ms
simd+par: 279 ms
ispc: 176 ms
ispc+tasks: 373 ms

I guess that's... good?

Failing tests

• I've updated the testing code to use approximative float comparisons, with an error no larger than 10^-4.

aobench

• Added mul_adde to most places, matching ISPC's assembly.
• Inlined some functions which weren't getting inlined before.
• Overall, when using the system math library (instead of comparing against ISPC's built-in one) Rust performance is better than with ISPC.

[gnzlbg]

I think this can be updated to use the new fmae function right?

[GabrielMajeri]

@gnzlbg I was going to wait for #123 to land so I can use mul_adde .

I'm not planning to get this merged however, not until we manage to beat ISPC in aobench at least.

[GabrielMajeri]

I'd decided to compare oranges to oranges and use libm for both ISPC and Rust, to avoid comparing using libsleef vs. ISPC's inlined math functions.

The results are in:

Rust: 4170 ms
ISPC: 4532 ms

So Rust is faster than ISPC. I don't know if it's due to a lack of inlining (#126), or some issue with sleef, but it's clear there's not much more we can do to optimize the Rust code itself. We're only limited by the speed of math functions.

[gnzlbg]

Same here for avx +- fma

[gnzlbg]

Note that ISPC inline math functions are not as precise as sleef, libm, SVML, etc.

Let me know when you finish the nitpicks and this is green so that we can merge this. I think it looks really good. I'll run the benchmarks on some systems and update the readme's afterwards (you can also add the results for your own systems in a follow up PR if you want).

[gnzlbg]

Also: THANKS for doing this!

[GabrielMajeri]

Alright, just fixed a little big issue with the RNG: those functions were not marked as #[inline(always)], so the target_feature attribute would end up not applying to them, so we would end up with AVX2 code calling into a function built with SSE2 instructions.

---

I think I've addressed most of the issues I could, some final bechmark results here (with libm, in order from slowest to fastest):

ispc: 4542 ms
vector: 4355 ms
tiled: 4117 ms

vector_par:  1233 ms
ispc tasks: 1227 ms
tiled_par: 1170 ms

---

Side note: would it be fair to modify build.rs to force ISPC to always use the system's math library? After all, this repository is not about benchmarking's ISPC fast math library

[gnzlbg]

Alright, just fixed a little big issue with the RNG:

Damn, this is a know problem, and we have no lints against this yet :(

Side note: would it be fair to modify build.rs to force ISPC to always use the system's math library?

I'd prefer the default to be ispc's default, but we can add cargo features (e.g. ispc_svml, ispc_libm, ispc_builtin_math...) that select a different math library in the build.rs.

I'd prefer if you do this in a different PR, because it applies to many benchmarks and I'd like to merge this ;)

[GabrielMajeri]

@gnzlbg Clippy failed, but I don't think it's related to this PR.

I'd prefer if you do this in a different PR, because it applies to many benchmarks and I'd like to merge this ;)

Alright, I'll see if I can't get cross-language inlining to work, maybe we can use sleef to beat ISPC's math library. Are you on IRC ?

[gnzlbg]

Clippy failed, but I don't think it's related to this PR.

I'll fix that :)

Alright, I'll see if I can't get cross-language inlining to work, maybe we can use sleef to beat ISPC's math library.

I have a branch where I started doing this the other day. I can upload it if you want to pick it up from there.

[gnzlbg]

@GabrielMajeri i'm gnzlbg on IRC