@mdbase-dev/sdk 0.0.0-stage → 0.1.0-beta.153

package/node_modules/@noble/hashes/src/blake2.ts

@@ -0,0 +1,648 @@
+/**
+ * blake2b (64-bit) & blake2s (8 to 32-bit) hash functions.
+ * b could have been faster, but there is no fast u64 in js, so s is 1.5x faster.
+ * @module
+ */
+import { BSIGMA, G1s, G2s } from './_blake.ts';
+import { SHA256_IV } from './_md.ts';
+import * as u64 from './_u64.ts';
+// prettier-ignore
+import {
+  abytes, aexists, anumber, aoutput,
+  checkOpts,
+  clean, copyBytes, createHasher,
+  swap32IfBE, swap8IfBE,
+  u32,
+  type CHash,
+  type Hash,
+  type TArg,
+  type TRet
+} from './utils.ts';
+
+// Unrealized speed-up: G1b/G2b call the tiny u64.* helpers (rotations, split adds) across
+// the module boundary, which V8 does not inline — u64.add's {h,l} return even allocates
+// per call. File-local copies of the helpers measured ~15-20% faster blake2b on Node 24.
+// Kept deduplicated in _u64 for auditability. Also doesn't matter for bundles
+// which don't do `import`.
+
+/**
+ * Blake hash options.
+ * `dkLen` is output length. `key` is used in MAC mode. `salt` is used in
+ * KDF mode.
+ */
+export type Blake2Opts = {
+  /** Desired digest length in bytes. RFC 7693 uses 1..64 for blake2b and 1..32 for blake2s. */
+  dkLen?: number;
+  /** Optional MAC key. */
+  key?: Uint8Array;
+  /** Optional salt mixed into initialization. */
+  salt?: Uint8Array;
+  /** Optional personalization bytes. */
+  personalization?: Uint8Array;
+};
+
+// Same IV words as `SHA512_IV`, but endian-swapped into LE u32 low/high halves
+// for the BLAKE2b u64 helpers below.
+const B2B_IV = /* @__PURE__ */ Uint32Array.from([
+  0xf3bcc908, 0x6a09e667, 0x84caa73b, 0xbb67ae85, 0xfe94f82b, 0x3c6ef372, 0x5f1d36f1, 0xa54ff53a,
+  0xade682d1, 0x510e527f, 0x2b3e6c1f, 0x9b05688c, 0xfb41bd6b, 0x1f83d9ab, 0x137e2179, 0x5be0cd19,
+]);
+// Shared synchronous BLAKE2b work vector as LE u32 low/high halves.
+const BBUF = /* @__PURE__ */ new Uint32Array(32);
+
+// BLAKE2b G mix split into two half-rounds over LE u32 low/high limbs.
+function G1b(a: number, b: number, c: number, d: number, msg: TArg<Uint32Array>, x: number) {
+  // NOTE: V is LE here
+  const Xl = msg[x], Xh = msg[x + 1]; // prettier-ignore
+  let Al = BBUF[2 * a], Ah = BBUF[2 * a + 1]; // prettier-ignore
+  let Bl = BBUF[2 * b], Bh = BBUF[2 * b + 1]; // prettier-ignore
+  let Cl = BBUF[2 * c], Ch = BBUF[2 * c + 1]; // prettier-ignore
+  let Dl = BBUF[2 * d], Dh = BBUF[2 * d + 1]; // prettier-ignore
+  // v[a] = (v[a] + v[b] + x) | 0;
+  const ll = u64.add3L(Al, Bl, Xl);
+  Ah = u64.add3H(ll, Ah, Bh, Xh);
+  Al = ll | 0;
+  // v[d] = rotr(v[d] ^ v[a], 32)
+  let xh = Dh ^ Ah, xl = Dl ^ Al; // prettier-ignore
+  Dh = u64.rotr32H(xh, xl);
+  Dl = u64.rotr32L(xh, xl);
+  // v[c] = (v[c] + v[d]) | 0;
+  ({ h: Ch, l: Cl } = u64.add(Ch, Cl, Dh, Dl));
+  // v[b] = rotr(v[b] ^ v[c], 24)
+  xh = Bh ^ Ch;
+  xl = Bl ^ Cl;
+  Bh = u64.rotrSH(xh, xl, 24);
+  Bl = u64.rotrSL(xh, xl, 24);
+  BBUF[2 * a] = Al;
+  BBUF[2 * a + 1] = Ah;
+  BBUF[2 * b] = Bl;
+  BBUF[2 * b + 1] = Bh;
+  BBUF[2 * c] = Cl;
+  BBUF[2 * c + 1] = Ch;
+  BBUF[2 * d] = Dl;
+  BBUF[2 * d + 1] = Dh;
+}
+
+// Second half-round of the same LE-limb BLAKE2b G mix; `x` is the message word offset.
+function G2b(a: number, b: number, c: number, d: number, msg: TArg<Uint32Array>, x: number) {
+  // NOTE: V is LE here
+  const Xl = msg[x], Xh = msg[x + 1]; // prettier-ignore
+  let Al = BBUF[2 * a], Ah = BBUF[2 * a + 1]; // prettier-ignore
+  let Bl = BBUF[2 * b], Bh = BBUF[2 * b + 1]; // prettier-ignore
+  let Cl = BBUF[2 * c], Ch = BBUF[2 * c + 1]; // prettier-ignore
+  let Dl = BBUF[2 * d], Dh = BBUF[2 * d + 1]; // prettier-ignore
+  // v[a] = (v[a] + v[b] + x) | 0;
+  const ll = u64.add3L(Al, Bl, Xl);
+  Ah = u64.add3H(ll, Ah, Bh, Xh);
+  Al = ll | 0;
+  // v[d] = rotr(v[d] ^ v[a], 16)
+  let xh = Dh ^ Ah, xl = Dl ^ Al; // prettier-ignore
+  Dh = u64.rotrSH(xh, xl, 16);
+  Dl = u64.rotrSL(xh, xl, 16);
+  // v[c] = (v[c] + v[d]) | 0;
+  ({ h: Ch, l: Cl } = u64.add(Ch, Cl, Dh, Dl));
+  // v[b] = rotr(v[b] ^ v[c], 63)
+  xh = Bh ^ Ch;
+  xl = Bl ^ Cl;
+  Bh = u64.rotrBH(xh, xl, 63);
+  Bl = u64.rotrBL(xh, xl, 63);
+  BBUF[2 * a] = Al;
+  BBUF[2 * a + 1] = Ah;
+  BBUF[2 * b] = Bl;
+  BBUF[2 * b + 1] = Bh;
+  BBUF[2 * c] = Cl;
+  BBUF[2 * c + 1] = Ch;
+  BBUF[2 * d] = Dl;
+  BBUF[2 * d + 1] = Dh;
+}
+
+function checkBlake2Opts(
+  outputLen: number,
+  opts: TArg<Blake2Opts | undefined> = {},
+  keyLen: number,
+  saltLen: number,
+  persLen: number
+) {
+  anumber(keyLen);
+  // RFC 7693 §2.1 requires digest length nn in 1..keyLen (keyLen doubles as
+  // the per-variant max for both key and digest lengths: 64 for b, 32 for s).
+  if (outputLen <= 0 || outputLen > keyLen)
+    throw new Error('"dkLen" must be 1..' + keyLen + ', got ' + outputLen);
+  const { key, salt, personalization } = opts;
+  // This API uses `undefined` for the RFC 7693 `kk = 0` case, so a provided key must be non-empty.
+  if (key !== undefined && (key.length < 1 || key.length > keyLen))
+    throw new Error('"key" expected to be undefined or of length=1..' + keyLen);
+  if (salt !== undefined) abytes(salt, saltLen, 'salt');
+  if (personalization !== undefined) abytes(personalization, persLen, 'personalization');
+}
+
+/** Internal base class for BLAKE2. */
+export abstract class _BLAKE2<T extends _BLAKE2<T>> implements Hash<T> {
+  protected abstract compress(msg: Uint32Array, offset: number, isLast: boolean): void;
+  protected abstract get(): number[];
+  protected abstract set(...args: number[]): void;
+  abstract destroy(): void;
+  protected buffer: Uint8Array;
+  protected buffer32: Uint32Array;
+  protected finished = false;
+  protected destroyed = false;
+  protected length: number = 0;
+  protected pos: number = 0;
+  readonly blockLen: number;
+  readonly outputLen: number;
+  readonly canXOF: boolean = false;
+
+  constructor(blockLen: number, outputLen: number) {
+    anumber(blockLen);
+    anumber(outputLen);
+    this.blockLen = blockLen;
+    this.outputLen = outputLen;
+    this.buffer = new Uint8Array(blockLen);
+    this.buffer32 = u32(this.buffer);
+  }
+  update(data: TArg<Uint8Array>): this {
+    aexists(this);
+    abytes(data);
+    // Main difference with other hashes: there is flag for last block,
+    // so we cannot process current block before we know that there
+    // is the next one. This significantly complicates logic and reduces ability
+    // to do zero-copy processing
+    const { blockLen, buffer, buffer32 } = this;
+    const len = data.length;
+    const offset = data.byteOffset;
+    const buf = data.buffer;
+    for (let pos = 0; pos < len;) {
+      // If buffer is full and we still have input (don't process last block, same as blake2s)
+      if (this.pos === blockLen) {
+        swap32IfBE(buffer32);
+        this.compress(buffer32, 0, false);
+        swap32IfBE(buffer32);
+        this.pos = 0;
+      }
+      const take = Math.min(blockLen - this.pos, len - pos);
+      const dataOffset = offset + pos;
+      // Zero-copy only for full, 4-byte-aligned, non-final blocks.
+      if (take === blockLen && !(dataOffset % 4) && pos + take < len) {
+        const data32 = new Uint32Array(buf, dataOffset, Math.floor((len - pos) / 4));
+        swap32IfBE(data32);
+        for (let pos32 = 0; pos + blockLen < len; pos32 += buffer32.length, pos += blockLen) {
+          this.length += blockLen;
+          this.compress(data32, pos32, false);
+        }
+        swap32IfBE(data32);
+        continue;
+      }
+      // When the whole input is buffered in one go (common for short messages), passing `data`
+      // directly avoids allocating a subarray view.
+      buffer.set(pos === 0 && take === len ? data : data.subarray(pos, pos + take), this.pos);
+      this.pos += take;
+      this.length += take;
+      pos += take;
+    }
+    return this;
+  }
+  digestInto(out: TArg<Uint8Array>): void {
+    aexists(this);
+    aoutput(out, this);
+    // Reject unaligned views explicitly instead of hiding them behind a full scratch copy.
+    if (out.byteOffset & 3)
+      throw new RangeError('"output" expected 4-byte aligned byteOffset, got ' + out.byteOffset);
+    const { pos, buffer32 } = this;
+    this.finished = true;
+    // Padding
+    this.buffer.fill(0, pos);
+    swap32IfBE(buffer32);
+    this.compress(buffer32, 0, true);
+    swap32IfBE(buffer32);
+    const state = this.get();
+    // digest() passes our own `buffer` as `out`; reuse its cached u32 view instead of allocating.
+    const out32 = out === this.buffer ? buffer32 : u32(out);
+    const full = Math.floor(this.outputLen / 4);
+    for (let i = 0; i < full; i++) out32[i] = swap8IfBE(state[i]);
+    const tail = this.outputLen % 4;
+    if (!tail) return;
+    const off = full * 4;
+    const word = state[full];
+    for (let i = 0; i < tail; i++) out[off + i] = word >>> (8 * i);
+  }
+  digest(): TRet<Uint8Array> {
+    const { buffer, outputLen } = this;
+    this.digestInto(buffer);
+    // Return a copy so callers do not alias the instance scratch buffer used during finalization.
+    const res = buffer.slice(0, outputLen);
+    this.destroy();
+    return res as TRet<Uint8Array>;
+  }
+  _cloneInto(to?: T): T {
+    const { buffer, length, finished, destroyed, outputLen, pos } = this;
+    // Recreate only `dkLen`; key/salt/personalization are already absorbed into the copied state.
+    to ||= new (this.constructor as any)({ dkLen: outputLen }) as T;
+    to.set(...this.get());
+    // Last-block-aware lazy compression keeps the pending block live even when full.
+    to.buffer.set(buffer);
+    to.destroyed = destroyed;
+    to.finished = finished;
+    to.length = length;
+    to.pos = pos;
+    // @ts-ignore
+    to.outputLen = outputLen;
+    return to;
+  }
+  clone(): T {
+    return this._cloneInto();
+  }
+}
+
+/** Internal blake2b hash class with state stored as LE u32 low/high halves. */
+export class _BLAKE2b extends _BLAKE2<_BLAKE2b> {
+  // Same IV words as SHA-512 / BLAKE2b, encoded as LE u32 low/high halves.
+  private v0l = B2B_IV[0] | 0;
+  private v0h = B2B_IV[1] | 0;
+  private v1l = B2B_IV[2] | 0;
+  private v1h = B2B_IV[3] | 0;
+  private v2l = B2B_IV[4] | 0;
+  private v2h = B2B_IV[5] | 0;
+  private v3l = B2B_IV[6] | 0;
+  private v3h = B2B_IV[7] | 0;
+  private v4l = B2B_IV[8] | 0;
+  private v4h = B2B_IV[9] | 0;
+  private v5l = B2B_IV[10] | 0;
+  private v5h = B2B_IV[11] | 0;
+  private v6l = B2B_IV[12] | 0;
+  private v6h = B2B_IV[13] | 0;
+  private v7l = B2B_IV[14] | 0;
+  private v7h = B2B_IV[15] | 0;
+
+  constructor(opts: Blake2Opts = {}) {
+    opts = checkOpts({}, opts);
+    const olen = opts.dkLen === undefined ? 64 : opts.dkLen;
+    super(128, olen);
+    checkBlake2Opts(olen, opts, 64, 16, 16);
+    let { key, personalization, salt } = opts;
+    let keyLength = 0;
+    if (key !== undefined) {
+      abytes(key, undefined, 'key');
+      keyLength = key.length;
+    }
+    // RFC 7693 §2.5: xor `p[0] = 0x0101kknn` into the low 32 bits of `h[0]`;
+    // the high 32 bits stay at `IV[0]`.
+    this.v0l ^= this.outputLen | (keyLength << 8) | (0x01 << 16) | (0x01 << 24);
+    if (salt !== undefined) {
+      abytes(salt, undefined, 'salt');
+      // Copy: u32() would throw on views with byteOffset not divisible by 4.
+      const slt = u32(copyBytes(salt));
+      this.v4l ^= swap8IfBE(slt[0]);
+      this.v4h ^= swap8IfBE(slt[1]);
+      this.v5l ^= swap8IfBE(slt[2]);
+      this.v5h ^= swap8IfBE(slt[3]);
+    }
+    if (personalization !== undefined) {
+      abytes(personalization, undefined, 'personalization');
+      // Copy: u32() would throw on views with byteOffset not divisible by 4.
+      const pers = u32(copyBytes(personalization));
+      this.v6l ^= swap8IfBE(pers[0]);
+      this.v6h ^= swap8IfBE(pers[1]);
+      this.v7l ^= swap8IfBE(pers[2]);
+      this.v7h ^= swap8IfBE(pers[3]);
+    }
+    if (key !== undefined) {
+      // Pad to blockLen and update
+      const tmp = new Uint8Array(this.blockLen);
+      tmp.set(key);
+      this.update(tmp);
+      // The padded copy holds key material; buffer/state keep what they need.
+      clean(tmp);
+    }
+  }
+  // prettier-ignore
+  protected get(): [
+    number, number, number, number, number, number, number, number,
+    number, number, number, number, number, number, number, number
+  ] {
+    let { v0l, v0h, v1l, v1h, v2l, v2h, v3l, v3h, v4l, v4h, v5l, v5h, v6l, v6h, v7l, v7h } = this;
+    return [v0l, v0h, v1l, v1h, v2l, v2h, v3l, v3h, v4l, v4h, v5l, v5h, v6l, v6h, v7l, v7h];
+  }
+  // prettier-ignore
+  protected set(
+    v0l: number, v0h: number, v1l: number, v1h: number,
+    v2l: number, v2h: number, v3l: number, v3h: number,
+    v4l: number, v4h: number, v5l: number, v5h: number,
+    v6l: number, v6h: number, v7l: number, v7h: number
+  ): void {
+    this.v0l = v0l | 0;
+    this.v0h = v0h | 0;
+    this.v1l = v1l | 0;
+    this.v1h = v1h | 0;
+    this.v2l = v2l | 0;
+    this.v2h = v2h | 0;
+    this.v3l = v3l | 0;
+    this.v3h = v3h | 0;
+    this.v4l = v4l | 0;
+    this.v4h = v4h | 0;
+    this.v5l = v5l | 0;
+    this.v5h = v5h | 0;
+    this.v6l = v6l | 0;
+    this.v6h = v6h | 0;
+    this.v7l = v7l | 0;
+    this.v7h = v7h | 0;
+  }
+  protected compress(msg: Uint32Array, offset: number, isLast: boolean): void {
+    // First half from state. Direct writes: get() would allocate an array +
+    // closure per block.
+    // prettier-ignore
+    const { v0l, v0h, v1l, v1h, v2l, v2h, v3l, v3h, v4l, v4h, v5l, v5h, v6l, v6h, v7l, v7h } = this;
+    // prettier-ignore
+    { BBUF[0] = v0l; BBUF[1] = v0h; BBUF[2] = v1l; BBUF[3] = v1h;
+      BBUF[4] = v2l; BBUF[5] = v2h; BBUF[6] = v3l; BBUF[7] = v3h;
+      BBUF[8] = v4l; BBUF[9] = v4h; BBUF[10] = v5l; BBUF[11] = v5h;
+      BBUF[12] = v6l; BBUF[13] = v6h; BBUF[14] = v7l; BBUF[15] = v7h; }
+    BBUF.set(B2B_IV, 16); // Second half from IV.
+    const l = u64.fromNumL(this.length);
+    const h = u64.fromNumH(this.length);
+    BBUF[24] = B2B_IV[8] ^ l; // Low word of the offset.
+    BBUF[25] = B2B_IV[9] ^ h; // High word.
+    // Invert all bits for last block
+    if (isLast) {
+      BBUF[28] = ~BBUF[28];
+      BBUF[29] = ~BBUF[29];
+    }
+    let j = 0;
+    const s = BSIGMA;
+    // SIGMA selects 64-bit message words; multiply by 2 because `msg` stores
+    // each word as [low32, high32].
+    for (let i = 0; i < 12; i++) {
+      G1b(0, 4, 8, 12, msg, offset + 2 * s[j++]);
+      G2b(0, 4, 8, 12, msg, offset + 2 * s[j++]);
+      G1b(1, 5, 9, 13, msg, offset + 2 * s[j++]);
+      G2b(1, 5, 9, 13, msg, offset + 2 * s[j++]);
+      G1b(2, 6, 10, 14, msg, offset + 2 * s[j++]);
+      G2b(2, 6, 10, 14, msg, offset + 2 * s[j++]);
+      G1b(3, 7, 11, 15, msg, offset + 2 * s[j++]);
+      G2b(3, 7, 11, 15, msg, offset + 2 * s[j++]);
+
+      G1b(0, 5, 10, 15, msg, offset + 2 * s[j++]);
+      G2b(0, 5, 10, 15, msg, offset + 2 * s[j++]);
+      G1b(1, 6, 11, 12, msg, offset + 2 * s[j++]);
+      G2b(1, 6, 11, 12, msg, offset + 2 * s[j++]);
+      G1b(2, 7, 8, 13, msg, offset + 2 * s[j++]);
+      G2b(2, 7, 8, 13, msg, offset + 2 * s[j++]);
+      G1b(3, 4, 9, 14, msg, offset + 2 * s[j++]);
+      G2b(3, 4, 9, 14, msg, offset + 2 * s[j++]);
+    }
+    this.v0l ^= BBUF[0] ^ BBUF[16];
+    this.v0h ^= BBUF[1] ^ BBUF[17];
+    this.v1l ^= BBUF[2] ^ BBUF[18];
+    this.v1h ^= BBUF[3] ^ BBUF[19];
+    this.v2l ^= BBUF[4] ^ BBUF[20];
+    this.v2h ^= BBUF[5] ^ BBUF[21];
+    this.v3l ^= BBUF[6] ^ BBUF[22];
+    this.v3h ^= BBUF[7] ^ BBUF[23];
+    this.v4l ^= BBUF[8] ^ BBUF[24];
+    this.v4h ^= BBUF[9] ^ BBUF[25];
+    this.v5l ^= BBUF[10] ^ BBUF[26];
+    this.v5h ^= BBUF[11] ^ BBUF[27];
+    this.v6l ^= BBUF[12] ^ BBUF[28];
+    this.v6h ^= BBUF[13] ^ BBUF[29];
+    this.v7l ^= BBUF[14] ^ BBUF[30];
+    this.v7h ^= BBUF[15] ^ BBUF[31];
+    clean(BBUF);
+  }
+  destroy(): void {
+    this.destroyed = true;
+    clean(this.buffer32);
+    this.set(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
+  }
+}
+
+/**
+ * Blake2b hash function. 64-bit. 1.5x slower than blake2s in JS.
+ * @param msg - message that would be hashed
+ * @param opts - Optional output, MAC, salt, and personalization settings.
+ *   `dkLen` must be 1..64 bytes; `salt` and `personalization`, if present,
+ *   must be 16 bytes each. See {@link Blake2Opts}.
+ * @returns Digest bytes.
+ * @example
+ * Hash a message with Blake2b.
+ * ```ts
+ * blake2b(new Uint8Array([97, 98, 99]));
+ * ```
+ * @example
+ * Hash a message with Blake2b while selecting output, MAC, salt, and personalization settings.
+ * ```ts
+ * blake2b(new Uint8Array([97, 98, 99]), {
+ *   dkLen: 32,
+ *   key: new Uint8Array(32),
+ *   salt: new Uint8Array(16),
+ *   personalization: new Uint8Array(16),
+ * });
+ * ```
+ */
+export const blake2b: TRet<CHash<_BLAKE2b, Blake2Opts>> = /* @__PURE__ */ createHasher(
+  (opts) => new _BLAKE2b(opts)
+);
+
+// =================
+// Blake2S
+// =================
+
+/** Internal type, 16 numbers. */
+// prettier-ignore
+export type _Num16 = {
+  v0: number; v1: number; v2: number; v3: number;
+  v4: number; v5: number; v6: number; v7: number;
+  v8: number; v9: number; v10: number; v11: number;
+  v12: number; v13: number; v14: number; v15: number;
+};
+
+/**
+ * BLAKE2-compress core method.
+ * Runs only the round function over a caller-supplied local vector; callers initialize `v0..v15`
+ * and apply the final `h[i] ^= v[i] ^ v[i + 8]` fold themselves.
+ * @param s - flattened sigma schedule bytes
+ * @param offset - starting word offset inside `msg`, not a byte offset
+ * @param msg - message words
+ * @param rounds - round count to execute
+ * @param v0 - state word 0
+ * @param v1 - state word 1
+ * @param v2 - state word 2
+ * @param v3 - state word 3
+ * @param v4 - state word 4
+ * @param v5 - state word 5
+ * @param v6 - state word 6
+ * @param v7 - state word 7
+ * @param v8 - state word 8
+ * @param v9 - state word 9
+ * @param v10 - state word 10
+ * @param v11 - state word 11
+ * @param v12 - state word 12
+ * @param v13 - state word 13
+ * @param v14 - state word 14
+ * @param v15 - state word 15
+ * @returns Updated compression state words.
+ * @example
+ * Run the BLAKE2 compression core on zeroed state and message words.
+ * ```ts
+ * import { _compress } from '@noble/hashes/blake2.js';
+ * const state = _compress(
+ *   new Uint8Array(16),
+ *   0,
+ *   new Uint32Array(16),
+ *   1,
+ *   0, 0, 0, 0, 0, 0, 0, 0,
+ *   0, 0, 0, 0, 0, 0, 0, 0
+ * );
+ * state.v0;
+ * ```
+ */
+// prettier-ignore
+export function _compress(
+  s: TArg<Uint8Array>, offset: number, msg: TArg<Uint32Array>, rounds: number,
+  v0: number, v1: number, v2: number, v3: number, v4: number, v5: number, v6: number, v7: number,
+  v8: number, v9: number, v10: number, v11: number, v12: number, v13: number, v14: number, v15: number,
+): _Num16 {
+  let j = 0;
+  for (let i = 0; i < rounds; i++) {
+    ({ a: v0, b: v4, c: v8, d: v12 } = G1s(v0, v4, v8, v12, msg[offset + s[j++]]));
+    ({ a: v0, b: v4, c: v8, d: v12 } = G2s(v0, v4, v8, v12, msg[offset + s[j++]]));
+    ({ a: v1, b: v5, c: v9, d: v13 } = G1s(v1, v5, v9, v13, msg[offset + s[j++]]));
+    ({ a: v1, b: v5, c: v9, d: v13 } = G2s(v1, v5, v9, v13, msg[offset + s[j++]]));
+    ({ a: v2, b: v6, c: v10, d: v14 } = G1s(v2, v6, v10, v14, msg[offset + s[j++]]));
+    ({ a: v2, b: v6, c: v10, d: v14 } = G2s(v2, v6, v10, v14, msg[offset + s[j++]]));
+    ({ a: v3, b: v7, c: v11, d: v15 } = G1s(v3, v7, v11, v15, msg[offset + s[j++]]));
+    ({ a: v3, b: v7, c: v11, d: v15 } = G2s(v3, v7, v11, v15, msg[offset + s[j++]]));
+
+    ({ a: v0, b: v5, c: v10, d: v15 } = G1s(v0, v5, v10, v15, msg[offset + s[j++]]));
+    ({ a: v0, b: v5, c: v10, d: v15 } = G2s(v0, v5, v10, v15, msg[offset + s[j++]]));
+    ({ a: v1, b: v6, c: v11, d: v12 } = G1s(v1, v6, v11, v12, msg[offset + s[j++]]));
+    ({ a: v1, b: v6, c: v11, d: v12 } = G2s(v1, v6, v11, v12, msg[offset + s[j++]]));
+    ({ a: v2, b: v7, c: v8, d: v13 } = G1s(v2, v7, v8, v13, msg[offset + s[j++]]));
+    ({ a: v2, b: v7, c: v8, d: v13 } = G2s(v2, v7, v8, v13, msg[offset + s[j++]]));
+    ({ a: v3, b: v4, c: v9, d: v14 } = G1s(v3, v4, v9, v14, msg[offset + s[j++]]));
+    ({ a: v3, b: v4, c: v9, d: v14 } = G2s(v3, v4, v9, v14, msg[offset + s[j++]]));
+  }
+  return { v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15 };
+}
+
+// Blake2s reuses the SHA-256 IV words as-is.
+const B2S_IV = /* @__PURE__ */ SHA256_IV.slice();
+
+/** Internal blake2s hash class. */
+export class _BLAKE2s extends _BLAKE2<_BLAKE2s> {
+  // Internal state, same as SHA-256
+  private v0 = B2S_IV[0] | 0;
+  private v1 = B2S_IV[1] | 0;
+  private v2 = B2S_IV[2] | 0;
+  private v3 = B2S_IV[3] | 0;
+  private v4 = B2S_IV[4] | 0;
+  private v5 = B2S_IV[5] | 0;
+  private v6 = B2S_IV[6] | 0;
+  private v7 = B2S_IV[7] | 0;
+
+  constructor(opts: Blake2Opts = {}) {
+    opts = checkOpts({}, opts);
+    const olen = opts.dkLen === undefined ? 32 : opts.dkLen;
+    super(64, olen);
+    checkBlake2Opts(olen, opts, 32, 8, 8);
+    let { key, personalization, salt } = opts;
+    let keyLength = 0;
+    if (key !== undefined) {
+      abytes(key, undefined, 'key');
+      keyLength = key.length;
+    }
+    // RFC 7693 §2.5: xor `p[0] = 0x0101kknn` directly into `h[0]`, since
+    // BLAKE2s stores each state word as one `u32`.
+    this.v0 ^= this.outputLen | (keyLength << 8) | (0x01 << 16) | (0x01 << 24);
+    if (salt !== undefined) {
+      abytes(salt, undefined, 'salt');
+      // Copy: u32() would throw on views with byteOffset not divisible by 4.
+      const slt = u32(copyBytes(salt as Uint8Array));
+      this.v4 ^= swap8IfBE(slt[0]);
+      this.v5 ^= swap8IfBE(slt[1]);
+    }
+    if (personalization !== undefined) {
+      abytes(personalization, undefined, 'personalization');
+      // Copy: u32() would throw on views with byteOffset not divisible by 4.
+      const pers = u32(copyBytes(personalization as Uint8Array));
+      this.v6 ^= swap8IfBE(pers[0]);
+      this.v7 ^= swap8IfBE(pers[1]);
+    }
+    if (key !== undefined) {
+      // Pad to blockLen and update
+      const tmp = new Uint8Array(this.blockLen);
+      tmp.set(key);
+      this.update(tmp);
+      // The padded copy holds key material; buffer/state keep what they need.
+      clean(tmp);
+    }
+  }
+  protected get(): [number, number, number, number, number, number, number, number] {
+    const { v0, v1, v2, v3, v4, v5, v6, v7 } = this;
+    return [v0, v1, v2, v3, v4, v5, v6, v7];
+  }
+  // prettier-ignore
+  protected set(
+    v0: number, v1: number, v2: number, v3: number, v4: number, v5: number, v6: number, v7: number
+  ): void {
+    this.v0 = v0 | 0;
+    this.v1 = v1 | 0;
+    this.v2 = v2 | 0;
+    this.v3 = v3 | 0;
+    this.v4 = v4 | 0;
+    this.v5 = v5 | 0;
+    this.v6 = v6 | 0;
+    this.v7 = v7 | 0;
+  }
+  protected compress(msg: Uint32Array, offset: number, isLast: boolean): void {
+    const l = u64.fromNumL(this.length);
+    const h = u64.fromNumH(this.length);
+    // Seed v8..v15 from the IV, xor the low/high 32-bit byte counter into
+    // v12/v13, and invert v14 on the final block.
+    // prettier-ignore
+    const { v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15 } =
+      _compress(
+        BSIGMA, offset, msg, 10,
+        this.v0, this.v1, this.v2, this.v3, this.v4, this.v5, this.v6, this.v7,
+        B2S_IV[0], B2S_IV[1], B2S_IV[2], B2S_IV[3], l ^ B2S_IV[4], h ^ B2S_IV[5], isLast ? ~B2S_IV[6] : B2S_IV[6], B2S_IV[7]
+      );
+    this.v0 ^= v0 ^ v8;
+    this.v1 ^= v1 ^ v9;
+    this.v2 ^= v2 ^ v10;
+    this.v3 ^= v3 ^ v11;
+    this.v4 ^= v4 ^ v12;
+    this.v5 ^= v5 ^ v13;
+    this.v6 ^= v6 ^ v14;
+    this.v7 ^= v7 ^ v15;
+  }
+  destroy(): void {
+    this.destroyed = true;
+    clean(this.buffer32);
+    this.set(0, 0, 0, 0, 0, 0, 0, 0);
+  }
+}
+
+/**
+ * Blake2s hash function. Focuses on 8-bit to 32-bit platforms. 1.5x faster than blake2b in JS.
+ * @param msg - message that would be hashed
+ * @param opts - Optional output, MAC, salt, and personalization settings.
+ *   `dkLen` must be 1..32 bytes; `salt` and `personalization`, if present,
+ *   must be 8 bytes each. See {@link Blake2Opts}.
+ * @returns Digest bytes.
+ * @example
+ * Hash a message with Blake2s.
+ * ```ts
+ * blake2s(new Uint8Array([97, 98, 99]));
+ * ```
+ * @example
+ * Hash a message with Blake2s while selecting output, MAC, salt, and personalization settings.
+ * ```ts
+ * blake2s(new Uint8Array([97, 98, 99]), {
+ *   dkLen: 16,
+ *   key: new Uint8Array(32),
+ *   salt: new Uint8Array(8),
+ *   personalization: new Uint8Array(8),
+ * });
+ * ```
+ */
+export const blake2s: TRet<CHash<_BLAKE2s, Blake2Opts>> = /* @__PURE__ */ createHasher(
+  (opts) => new _BLAKE2s(opts)
+);