Skip to content

DFN3 WASM optimization: tract 0.21 -> 0.22.1 + WASM SIMD kernel kit #1

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
czoli1976 wants to merge 10 commits into
base: main
Choose a base branch
from
Open

DFN3 WASM optimization: tract 0.21 -> 0.22.1 + WASM SIMD kernel kit #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
8ddf19e
DFN3 WASM optimization: bump tract 0.21 -> 0.22.1, route tract-linalg…
czoli1976 Apr 28, 2026
5308f75
workspace Cargo.toml: repoint tract-linalg patch at GH fork (git source)
czoli1976 Apr 28, 2026
8153d23
workspace Cargo.toml: drop 8x4 from kit comment (kernel was removed u…
czoli1976 Apr 28, 2026
c90060c
workspace Cargo.toml: pin tract-linalg patch to rev b82d1f0 (instead …
czoli1976 Apr 28, 2026
d089c96
workspace Cargo.toml: bump tract-linalg patch rev to d925624
czoli1976 Apr 28, 2026
d222c7b
fix(deps): vendor tract-linalg-0.22.1 with kernel kit (path-based [pa…
czoli1976 Apr 28, 2026
a7fa8de
perf(wasm): SIMD-vectorize compute_band_corr inner loop
czoli1976 May 3, 2026
4895b00
perf(wasm): SIMD-vectorize 3 more inference DSP loops
czoli1976 May 3, 2026
1084fe3
perf(wasm): SIMD-vectorize 2 more DSP loops (band_unit_norm + _t)
czoli1976 May 3, 2026
143b040
perf(wasm): SIMD-vectorize 3 more frame_synthesis loops
czoli1976 May 4, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
The table of contents is too big for display.
Diff view
Diff view
  •  
  •  
  •  
1,720 changes: 917 additions & 803 deletions Cargo.lock
View file
Open in desktop

Large diffs are not rendered by default.

12 changes: 12 additions & 0 deletions Cargo.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,3 +25,15 @@ debug = false
lto = "fat"
strip = true
panic = "abort"

# Patched tract-linalg with the WASM SIMD kernel kit:
# wasm_f32_4x4 (existing) + wasm_f32_4x1 / 8x1 / 16x1 (new GEMV variants)
# + wasm_f32_8x8 (new MM variant) + per-M dispatcher in Ops::mmv_f32.
# Cumulative impact on DFN3: RTF 0.1290 -> 0.0516 (-60%, 2.5x faster), bit-identical audio.
# Source: vendor/tract-linalg-0.22.1 (self-contained crate, version-matched 0.22.1).
# Lineage: cherry-picked from czoli1976/tract@kernel-kit-on-v0.22.1 — kernel kit
# from sonos/tract#2164 (merged) applied onto sonos/tract v0.22.1 tag.
[patch.crates-io]
tract-linalg = { path = "vendor/tract-linalg-0.22.1" }


8 changes: 4 additions & 4 deletions libDF/Cargo.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -120,10 +120,10 @@ claxon = { version = "^0.4", optional = true }
env_logger = { version = "0.11", optional = true }
clap = { version = "4.0", optional = true, features = ["derive"] }
rust-ini = { version = "^0.21", optional = true }
tract-core = { version = "^0.21.4", optional = true }
tract-onnx = { version = "^0.21.4", optional = true }
tract-pulse = { version = "^0.21.4", optional = true }
tract-hir = { version = "^0.21.4", optional = true }
tract-core = { version = "=0.22.1", optional = true }
tract-onnx = { version = "=0.22.1", optional = true }
tract-pulse = { version = "=0.22.1", optional = true }
tract-hir = { version = "=0.22.1", optional = true }
flate2 = { version = "1.0.24", optional = true }
tar = { version = "0.4.38", optional = true }
wasm-bindgen = { version = "0.2.87", optional = true }
Expand Down
2 changes: 1 addition & 1 deletion libDF/src/bin/enhance_wav.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,7 @@ use std::{path::PathBuf, process::exit, time::Instant};
use anyhow::Result;
use clap::{Parser, ValueHint};
use df::{tract::*, transforms::resample, wav_utils::*};
use ndarray::{prelude::*, Axis};
use tract_core::ndarray::{self, prelude::*, Axis};

#[cfg(all(
not(windows),
Expand Down
Loading