#mizchi/simd

• @simdcore — drop-in faster equivalents of common moonbitlang/core idioms (sum, sort, Bytes search, UTF-8 encode/decode, JSON structural indexing, …). Same shape as the core idiom it replaces; results are identical on every target, only the throughput differs. Start here if you just want core to be faster.
• @simd_buffer — the SimdBuffer* buffer family for numeric / byte / image pipelines that own their data across many SIMD ops.

just test          # all 4 targets
just bench-wasm    # wasm benchmark
just bench-native  # native benchmark

#Packages & backend acceleration

#Install

"deps": {
  "mizchi/simd": "0.4.0"
}

import {
  "mizchi/simd/src/simdcore",
  "mizchi/simd/src/simdimage",
  "mizchi/simd/src/simd_buffer",
  "mizchi/simd/src/simdjson",
  "mizchi/simd/src/simdcodec",
  "mizchi/simd/src/simdhash",
}

#Quick start (recommended)

let a : FixedArray[Int] = [1, 2, 3, 4, 5, 6, 7, 8]
let buf = @simd_buffer.SimdBuffer::from_array(a)
let total = buf.sum()              // SIMD on wasm / wasm-gc / native, scalar on js
let out = @simd_buffer.SimdBuffer::make(buf.length())
@simd_buffer.SimdBuffer::add(buf, buf, out)
let back : FixedArray[Int] = out.to_array()

#What you get per target

js is scalar. MoonBit on the js backend has no SIMD escape hatch. SimdBuffer compiles and runs on js for API portability (same code across all four targets), but throughput-critical hot paths on js should keep data in native JS typed arrays and call back into wasm where SimdBuffer is actually accelerated.

#Buffer types and operations

• SimdBuffer (i32): sum, dot, add, sub, mul, neg, abs, min_elem, max_elem, eq, lt, gt, where_, saxpy, min, max, argmin, argmax, prod, count_nonzero, any, all, cumsum, cumprod, div, gather, scatter, sort, sort4, sort16, bitonic_merge8/16/32/64
• SimdBufferF32: add, sub, mul, div, sqrt, min_elem, max_elem, sum, dot
• SimdBufferF64: add, sub, mul, div, sqrt, min_elem, max_elem, sum, dot, mean, variance, matmul, gemv, transpose
• SimdBufferBytes: popcount, memcpy, memset, equal, find_byte, count_byte, is_ascii, to_lower_ascii, to_upper_ascii, validate_utf8 (structural), validate_utf8_strict (RFC 3629 — rejects overlong / surrogate / > U+10FFFF), adler32, base64_encode / base64_decode (+ _into in-place variants)
• 0.3.0 byte arithmetic (SimdBufferBytes): byte_add, byte_sub, byte_avg, sat_add, sat_sub, clamp, byte_sub_offset (PNG Sub-filter pattern)
• image / pixel ops now live in their own @simdimage package and operate on Bytes / FixedArray[Byte] directly (no SimdBuffer hop) — see the @simdimage section below
• 0.3.0 i32: SimdBuffer::array_equal(a, b, len) -> Bool — SIMD all-equal reduction
• SimdBufferRing: single-arena bump allocator. On wasm/wasm-gc it amortises memory.grow across sub-allocations (~120 µs → 8 ns per alloc). On native/js it's a thin shell — alloc_* just calls make because GC alloc is already cheap.
• Copy bridges: from_array / to_array / copy_from_array / copy_to_array for FixedArray ↔ SimdBuffer interop.

#@simdcore — faster moonbitlang/core equivalents

a.iter().fold(init=0, fn(x, y) { x + y })  // core
@simdcore.sum(a)                           // faster equivalent

a.iter().maximum()   ->  @simdcore.maximum(a)    // Int?
a.search(x)          ->  @simdcore.search(a, x)  // Int?
a.sort()             ->  @simdcore.sort(a)       // FixedArray[Int]

• FixedArray[Int]: sum, product, dot, maximum/minimum, search/contains, count_nonzero, fill, sort; element-wise add, sub, mul, neg, abs, saxpy.
• FixedArray[Double]: sum_f64, dot_f64, mean_f64, variance_f64; add_f64, sub_f64, mul_f64, div_f64, sqrt_f64, min_elem_f64, max_elem_f64.
• Bytes (zero-copy on wasm + native): bytes_equal, bytes_search/bytes_contains, bytes_count, bytes_is_ascii, bytes_index_of/bytes_contains_sub (substring), bytes_rindex (last byte). Native binds these to libc memchr / memmem / memrchr.
• FixedArray[Byte] (in-place): to_lower_ascii, to_upper_ascii.
• String ↔ UTF-8 (MoonBit's biggest FFI bottleneck): encode_utf8 / encode_utf8_into, is_ascii_string, decode_utf8_unsafe / decode_utf8_unsafe_intrinsic. ASCII fast path vectorises the UTF-16↔UTF-8 narrow/widen; non-ASCII delegates to @encoding/utf8 so the result always matches core.
• Array[Int] / Array[Double] bridge: to_fixedarray / of_fixedarray plus copy-in one-shots array_sum, array_sort, array_dot, array_mean_f64, …
• JSON structural indexing over Bytes: json_classify_structural (bitmap), json_structural_indices / _into (full pipeline).

#Headlines (wasm, V8, Apple Silicon)

#@simdimage — SIMD image / pixel ops

let rgb = Bytes::from_array(pixels)          // n*3 RGB bytes
let rgba : FixedArray[Byte] = FixedArray::make(n * 4, b'\x00')
@simdimage.rgb_to_rgba(rgb, 0xFF, rgba)      // expand to RGBA, opaque

#Bench (wasm, V8, 4096 px, vs a plain per-byte loop)

#@simdjson — JSON byte classification

let input = @simd_buffer.SimdBufferBytes::from_array(json_bytes)
let words = (input.length() + 31) / 32
let structural = @simd_buffer.SimdBuffer::make(words)
let quote_mask = @simd_buffer.SimdBuffer::make(words)
let indices = @simd_buffer.SimdBuffer::make(input.length())
let count = @simdjson.find_structural_indices_with_scratch(
  input, structural, quote_mask, indices,
)
// indices.get(0..count) now hold byte offsets of `{ } [ ] , :` outside any string

#Memory model gotcha (wasm-only)

let ring = @simd_buffer.SimdBufferRing::make(65536)
for input in inputs {
  ring.reset()
  let out = ring.alloc_bytes((input.length() + 2) / 3 * 4)
  @simd_buffer.SimdBufferBytes::base64_encode_into(input, out)
  // ... use out, then forget it ...
}

#Alternative: FixedArray-based API

let arr : FixedArray[Int] = [1, 2, 3, 4, 5, 6, 7, 8]
let total = @simd.sum_i32(arr)                   // wasm SIMD, scalar on others

#Speedup highlights

#wasm (V8, Apple Silicon, FixedArray-based API)

#wasm-gc (V8, Apple Silicon, SimdBuffer family)

#File structure

src/
  # FixedArray-API root package — imported as @simd
  simd_wasm_{i32,f64,f32,bytes,sort}.mbt   # wasm inline-WAT v128
  simd_native.mbt + simd_native_ffi.mbt    # native extern "C"
  simd_native.c                            # NEON / SSE2-baseline intrinsics
  simd_scalar.mbt                          # js + wasm-gc fallback
  internal/scalar*.mbt                     # shared scalar reference impls

  simdcore/                                # @simdcore — faster core equivalents
    simdcore.mbt                           # facade (i32 / f64 / Bytes / Array)
    simdcore_str_{wasm,fallback}.mbt       # String ↔ UTF-8
    simdcore_json_{wasm,native,fallback}.mbt + simdcore.c   # JSON indexing

  simdimage/                               # @simdimage — image / pixel ops
    simdimage.mbt                          # public API (Bytes / FixedArray[Byte])
    simdimage_wasm.mbt                     # wasm inline-WAT v128
    simdimage_fallback.mbt                 # wasm-gc + native + js scalar

  simdcodec/                               # @simdcodec — byte codecs (base64)
    base64_common.mbt                      # length helpers + encode/decode wrappers
    base64_wasm.mbt                        # wasm SIMD encode_into/decode_into
    base64_fallback.mbt                    # scalar tables + reference (all but native)
    base64_scalar.mbt                      # wasm-gc + js encode_into/decode_into
    base64_native.mbt + base64.c           # native C FFI (gcc scalar, LUT decode)

  simdhash/                                # @simdhash — SHA-256 / SHA-1 / MD5
    simdhash.mbt + sha1/md5/sha512.mbt   # scalar digests + public API
    simdhash_wasm.mbt                      # wasm 4-way inline-WAT (sha256/sha1/md5 _x4)
    simdhash_native.mbt + simdhash.c       # native SSE2/NEON 4-way (sha256/sha1)

  simdjson/                                # @simdjson — JSON byte classification
    simdjson_wasm.mbt                      # wasm + wasm-gc inline-WAT
    simdjson_scalar.mbt                    # native + js scalar

  simd_buffer/                             # @simd_buffer — portable API
    simd_buffer.mbt / _f32 / _f64 / _bytes / _sort / _ring / _copy.mbt
      # wasm + wasm-gc: linear-memory storage + inline-WAT v128
    simd_buffer_scalar.mbt
      # native + js: FixedArray storage + @internal / @simdcodec delegation
    simd_buffer_imports.mbt
      # cross-target import keep-alive

#Examples

• examples/cosine_similarity/ — vector / RAG search building block. Four implementations side by side (scalar, naive SIMD chained, precomputed-norm SIMD, fused inline-WAT) that double as a worked example of when chaining SIMD primitives helps and when it doesn't. precomputed and fused variants both land on the same memory-bandwidth × f64x2 ceiling (~4× scalar) — the fused variant demonstrates SimdBufferF64::raw_addr() as an escape hatch when the public ops can't compose efficiently.

#License