add an implementation using Wasm SIMD

[oconnor663]

Rust just shipped version 1.54, which stabilizes Wasm SIMD intrinsics. We should port the SSE4.1 implementation to use these. First I should add Wasm to CI and make sure benchmarks work there.

[oconnor663]

Looks like the current easiest way to run benchmarks is:

cargo +nightly build --benches --release --target wasm32-wasi
wasmer ./target/wasm32-wasi/release/deps/bench-*.wasm -- --bench

(Though that wildcard doesn't do what I want when I build multiple versions of the bench binary. I need something simpler...)

At the moment, this requires getting rid of the line where we use page_size:

diff --git a/Cargo.toml b/Cargo.toml
index a43f000..eadf73c 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -89,7 +89,6 @@ crypto-mac = { version = "0.11.0", optional = true }
 
 [dev-dependencies]
 hex = "0.4.2"
-page_size = "0.4.1"
 rand = "0.8.0"
 rand_chacha = "0.3.0"
 reference_impl = { path = "./reference_impl" }
diff --git a/benches/bench.rs b/benches/bench.rs
index 5efb9e6..c53278e 100644
--- a/benches/bench.rs
+++ b/benches/bench.rs
@@ -25,7 +25,7 @@ pub struct RandomInput {
 impl RandomInput {
     pub fn new(b: &mut Bencher, len: usize) -> Self {
         b.bytes += len as u64;
-        let page_size: usize = page_size::get();
+        let page_size: usize = 65536;
         let mut buf = vec![0u8; len + page_size];
         let mut rng = rand::thread_rng();
         rng.fill_bytes(&mut buf);

But once Elzair/page_size_rs#3 lands upstream this workaround shouldn't be necessary anymore.

[monoid]

Hi, @oconnor663! Is there anything you would like me to improve in this PR?

[oconnor663]

Apologies, I've been distracted by other things. I'll make a calendar note to come back to this PR this weekend.

[monoid]

For history: I quickchecked the wasm32-simd implementation against a native one for 500 CPU-hours on M1 performance cores, that's about 1.5e12 samples. No mismatches were found.

[mimi89999]

Would it be possible to also add WASM SIMD support in the C implementation? Emscripten does support it https://emscripten.org/docs/porting/simd.html and the headers are actually in LLVM https://github.com/llvm/llvm-project/blob/main/clang/lib/Headers/wasm_simd128.h

[mimi89999]

It seems that Node modules like https://github.com/connor4312/blake3 and https://github.com/Daninet/hash-wasm are actually using the C implementation.

[oconnor663]

Would it be possible to also add WASM SIMD support in the C implementation?

I think that would be a good idea (though I'm probably not going to work on it myself). The performance boost was much larger than I was expecting.

[mimi89999]

Actually, I have it almost ready

[mimi89999]

I added it in #459, but it seems slower than the portable implementation.

[monoid]

@mimi89999 how do you compile it? I'd like to see the WASM generated.

Upd: I see you have fixed it.

[mimi89999]

@monoid I'm compiling it with emcc -O2 -msimd128 test.c blake3.c blake3_dispatch.c blake3_portable.c blake3_wasm32_simd.c -o test.js, but the performance still isn't that impressive. Let's keep the discussion under the pull request.