Measured demand (DeslanStudio, N3350)
Importing a 50 s stereo 48 kHz WAV costs 10.8 s in wav_read — the PCM16 byte decode is a per-sample interpreted loop (src/tools/wavio.eigs): two buf_gets, the s16 two's-complement fold, and / 32767 per sample, ~4.8 M times. The write path (wav_write) is the same shape in reverse. This is the byte-decode sibling of the #597 mix-down kernels, same contract: one C loop, VM-identical arithmetic.
Builtins
buf_from_pcm16le of [bytes, byte_off, count] → NEW float buffer of count samples decoded from little-endian signed 16-bit PCM at byte_off. Arithmetic exactly matches wavio's read loop: v = b0 + 256*b1; if v >= 32768: v -= 65536; sample = v / 32767 (num_guard per step, so the builtin is bit-identical to the interpreted loop).buf_to_pcm16le of [floats, off, count] → NEW byte buffer (2*count elements), exactly wavio's write loop: clamp to [-1, 1], v = round(x * 32767), if v < 0: v += 65536, low byte v - floor(v/256)*256 (the ds_fmod expansion), high byte floor(v/256).buf_deinterleave of [src, channel, nch, count?] → NEW buffer of every nch-th sample starting at index channel (frame-interleaved channel split, wavio's (i*nch + c) addressing). count defaults to the full available tail; explicit count is window-checked.
All three: pure compute over arguments → sandbox pure-compute allowlist; deterministic (nothing to tape); loud bounds per the #597 family (value / index_range raises, no silent truncation); available freestanding.
DoD
- Differential suite leg: each builtin exactly equals the interpreted wavio loop on seeded pseudo-random data (byte-identical doubles), including a decode→deinterleave composition vs wav_read's fused loop, and encode→decode round-trip parity with the interpreted round trip.
- Loud-bounds leg (bad windows, odd shapes, non-buffer, bad channel/nch).
- Sandbox allowlist entries + n=5 micro-bench (interp vs builtin) in the PR body.
- Full release + ASan suites; BUILTINS.md rows.
Measured demand (DeslanStudio, N3350)
Importing a 50 s stereo 48 kHz WAV costs 10.8 s in
wav_read— the PCM16 byte decode is a per-sample interpreted loop (src/tools/wavio.eigs): twobuf_gets, the s16 two's-complement fold, and/ 32767per sample, ~4.8 M times. The write path (wav_write) is the same shape in reverse. This is the byte-decode sibling of the #597 mix-down kernels, same contract: one C loop, VM-identical arithmetic.Builtins
buf_from_pcm16le of [bytes, byte_off, count]→ NEW float buffer ofcountsamples decoded from little-endian signed 16-bit PCM atbyte_off. Arithmetic exactly matches wavio's read loop:v = b0 + 256*b1; if v >= 32768: v -= 65536; sample = v / 32767(num_guard per step, so the builtin is bit-identical to the interpreted loop).buf_to_pcm16le of [floats, off, count]→ NEW byte buffer (2*countelements), exactly wavio's write loop: clamp to [-1, 1],v = round(x * 32767),if v < 0: v += 65536, low bytev - floor(v/256)*256(theds_fmodexpansion), high bytefloor(v/256).buf_deinterleave of [src, channel, nch, count?]→ NEW buffer of everynch-th sample starting at indexchannel(frame-interleaved channel split, wavio's(i*nch + c)addressing).countdefaults to the full available tail; explicitcountis window-checked.All three: pure compute over arguments → sandbox pure-compute allowlist; deterministic (nothing to tape); loud bounds per the #597 family (
value/index_rangeraises, no silent truncation); available freestanding.DoD