`Buffer.asciiWrite` is 3x slower than `Buffer.set`

[ronag]

I would assume that Buffer.set and Buffer.asciiWrite should be roughly the same (they are both essentially a memcpy). However this is not the case.

Given that the speed is somewhat the same even when the string size grows I would assume that most goes to call overhead.

cpu: Apple M2 Pro
runtime: node v21.6.0 (arm64-darwin)

benchmark       time (avg)             (min … max)       p75       p99      p999
-------------------------------------------------- -----------------------------
• 8
-------------------------------------------------- -----------------------------
asciiWrite   34.29 ns/iter     (31.17 ns … 493 ns)  34.77 ns  40.71 ns  70.13 ns
loopWrite    10.27 ns/iter    (9.28 ns … 1'906 ns)  10.11 ns  14.97 ns  21.99 ns
bufWrite     13.31 ns/iter   (11.47 ns … 5'753 ns)  12.43 ns  15.65 ns  49.74 ns

summary for 8
  loopWrite
   1.3x faster than bufWrite
   3.34x faster than asciiWrite

• 16
-------------------------------------------------- -----------------------------
asciiWrite   33.56 ns/iter     (31.35 ns … 160 ns)  33.77 ns  39.75 ns  66.26 ns
loopWrite    20.99 ns/iter   (18.17 ns … 6'464 ns)  19.08 ns  26.37 ns   57.6 ns
bufWrite      13.7 ns/iter      (12 ns … 7'717 ns)  12.92 ns  16.48 ns     34 ns

summary for 16
  bufWrite
   1.53x faster than loopWrite
   2.45x faster than asciiWrite

• 32
-------------------------------------------------- -----------------------------
asciiWrite   37.25 ns/iter     (34.87 ns … 221 ns)  37.62 ns  43.68 ns  79.14 ns
loopWrite    92.46 ns/iter   (75.13 ns … 9'072 ns)  82.25 ns    228 ns  1'358 ns
bufWrite      13.6 ns/iter     (12.57 ns … 473 ns)  13.61 ns  17.01 ns  32.76 ns

summary for 32
  bufWrite
   2.74x faster than asciiWrite
   6.8x faster than loopWrite

• 64
-------------------------------------------------- -----------------------------
asciiWrite   37.03 ns/iter   (33.41 ns … 4'565 ns)  37.03 ns  43.38 ns  86.69 ns
loopWrite      166 ns/iter     (150 ns … 2'707 ns)    165 ns    183 ns  2'318 ns
bufWrite     12.52 ns/iter     (11.72 ns … 554 ns)  12.63 ns  15.83 ns  23.99 ns

summary for 64
  bufWrite
   2.96x faster than asciiWrite
   13.25x faster than loopWrite

import { bench, group, run } from 'mitata'

const BUF = Buffer.allocUnsafe(64).fill(88)
const BUF_BUF = new Array(64).fill(null)

function asciiWrite(str) {
  BUF.asciiWrite(str, 0)
}

function bufWrite(str, index) {
  BUF.set((BUF_BUF[index] ??= Buffer.from(str)))
}

function loopWrite(src) {
  for (let n = 0; n < src.length; n++) {
    BUF[n] = src.charCodeAt(n)
  }
}

const str8 = '01234567'
const str16 = '0123456789abcdef'
const str32 = '0123456789abcdef'.repeat(2)
const str64 = '0123456789abcdef'.repeat(4)

group('8', () => {
  bench('asciiWrite', () => asciiWrite(str8))
  bench('loopWrite', () => loopWrite(str8))
  bench('bufWrite', () => bufWrite(str8, 1))
})
group('16', () => {
  bench('asciiWrite', () => asciiWrite(str16))
  bench('loopWrite', () => loopWrite(str16))
  bench('bufWrite', () => bufWrite(str16, 2))
})
group('32', () => {
  bench('asciiWrite', () => asciiWrite(str32))
  bench('loopWrite', () => loopWrite(str32))
  bench('bufWrite', () => bufWrite(str32, 3))
})
group('64', () => {
  bench('asciiWrite', () => asciiWrite(str64))
  bench('loopWrite', () => loopWrite(str64))
  bench('bufWrite', () => bufWrite(str64, 4))
})

await run()

Refs: https://github.com/nodejs/performance/issues/168