Add WASM SIMD C implementation

[mimi89999]

I tested it in NodeJS and it seems 2x slower 😞

Update: It's faster!

[mimi89999]

The bad performance was caused by a small bug. Now the WASM SIMD implementation is barely faster than the portable one.

[monoid]

What is your result with -O3?

[mimi89999]

What is your result with -O3?

Exactly the same

[mimi89999]

Here is my benchmark. It's very primitive:

#include "blake3.h"

#include <stdio.h>
#include <string.h>

int main() {
  char hello[] = "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";

  uint8_t output[BLAKE3_OUT_LEN];

  for (unsigned int i = 0; i < 100000000; i++) {
    blake3_hasher hasher;
    blake3_hasher_init(&hasher);

    blake3_hasher_update(&hasher, hello, strlen(hello));

    blake3_hasher_finalize(&hasher, output, BLAKE3_OUT_LEN);
  }

  for (size_t i = 0; i < BLAKE3_OUT_LEN; i++) {
    printf("%02x", output[i]);
  }
  printf("\n");

  return 0;
}

michel@debian:~/git/BLAKE3/c$ emcc -O2 test.c blake3.c blake3_dispatch.c blake3_portable.c -o test.js
michel@debian:~/git/BLAKE3/c$ time nodejs test.js 
b0e1430ecdd18b09c5834a6ddbbc741dcf01df6ed0a1552cb8a7c0eeb31de404

real	2m27,537s
user	2m27,510s
sys	0m0,037s
michel@debian:~/git/BLAKE3/c$ emcc -O2 -msimd128 test.c blake3.c blake3_dispatch.c blake3_portable.c blake3_wasm32_simd.c -o test.js
michel@debian:~/git/BLAKE3/c$ time nodejs test.js 
b0e1430ecdd18b09c5834a6ddbbc741dcf01df6ed0a1552cb8a7c0eeb31de404

real	1m56,135s
user	1m56,128s
sys	0m0,017s

[monoid]

Try moving strlen out of the for loop.

[mimi89999]

@monoid That should not affect the difference between the two implementations. Could you check how it compares with the Rust WASM and Rust WASM SIMD implementations?

[monoid]

Same problem with Rust implementation. One needs longer data to see a difference. Start with something around 8-16Kb.

[mimi89999]

I see that the performance is near native on x86 compared to BLAKE3 with SSE2 only.

[monoid]

Sounds great! Meanwhile, I've managed to start a quickcheck to compare your implementation with a native one.

[mimi89999]

Here are some benchmarks:

michel@debian:~/git/BLAKE3/c$ emcc -O2 -msimd128 -sSTACK_SIZE=1114112 test.c blake3.c blake3_dispatch.c blake3_portable.c blake3_wasm32_simd.c -o test.js
michel@debian:~/git/BLAKE3/c$ nodejs test.js 
b5358909f8bed53f55bf9324e290e9a5a585de8b0239d18040e9d3b0c7e8f9cf
Time: 1.950000
Data size: 1048576
Rounds: 1000
Time per round: 0.001950
michel@debian:~/git/BLAKE3/c$ emcc -O2 -sSTACK_SIZE=1114112 test.c blake3.c blake3_dispatch.c blake3_portable.c -o test.js
michel@debian:~/git/BLAKE3/c$ nodejs test.js 
b5358909f8bed53f55bf9324e290e9a5a585de8b0239d18040e9d3b0c7e8f9cf
Time: 2.253000
Data size: 1048576
Rounds: 1000
Time per round: 0.002253
michel@debian:~/git/BLAKE3/c$ vim test.c
michel@debian:~/git/BLAKE3/c$ emcc -O2 -msimd128 -sSTACK_SIZE=1114112 test.c blake3.c blake3_dispatch.c blake3_portable.c blake3_wasm32_simd.c -o test.js
michel@debian:~/git/BLAKE3/c$ nodejs test.js 
b5358909f8bed53f55bf9324e290e9a5a585de8b0239d18040e9d3b0c7e8f9cf
Time: 19.468000
Data size: 1048576
Rounds: 10000
Time per round: 0.001947
michel@debian:~/git/BLAKE3/c$ emcc -O2 -sSTACK_SIZE=1114112 test.c blake3.c blake3_dispatch.c blake3_portable.c -o test.js
michel@debian:~/git/BLAKE3/c$ nodejs test.js 
b5358909f8bed53f55bf9324e290e9a5a585de8b0239d18040e9d3b0c7e8f9cf
Time: 22.336000
Data size: 1048576
Rounds: 10000
Time per round: 0.002234

[mimi89999]

@oconnor663 After some benchmarking, I can't replicate the huge performance improvements of using SIMD instructions on my machine. Nevertheless, it's still a 15% improvement. I think that it's worth adding it. Maybe on other machines, the difference will be much bigger. 🤔