Skip to content

Performance: optimize sparse mel filterbank matmul #12

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
ysdede wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Performance: optimize sparse mel filterbank matmul #12

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions .jules/bolt.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,3 @@
## 2024-05-23 - Sparse Mel Filterbank
Learning: The Mel filterbank matrix is ~98.5% sparse (only ~500 non-zero elements out of ~32k), making dense matrix multiplication extremely inefficient.
Action: Always check for sparsity in fixed transform matrices (like Mel or DCT) and implement sparse iteration if sparsity > 90%.
31 changes: 29 additions & 2 deletions src/mel.js
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -261,6 +261,32 @@ export class MelSpectrogram {
this.hannWindow = createPaddedHannWindow(this.winLength, this.nFft);
this.twiddles = precomputeTwiddles(this.nFft);

// Precompute sparse matrix indices (optimize for ~98.5% sparsity)
this._fbStart = new Int32Array(this.nMels);
this._fbEnd = new Int32Array(this.nMels);
for (let m = 0; m < this.nMels; m++) {
let start = 0;
let end = this.nFreqBins;
Copy link

@kiloconnect kiloconnect bot Feb 21, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

SUGGESTION: The default value for end is this.nFreqBins, which means if a mel row were entirely zero (no non-zero element found in the second loop), the full range [0, nFreqBins) would be iterated — a safe conservative fallback. However, the comment above only documents the first loop's default. Consider adding a brief comment on the end default for symmetry and clarity:

Suggested change
let end = this.nFreqBins;
let end = this.nFreqBins; // default: full range (safe fallback if row is all-zero)

const offset = m * this.nFreqBins;

// Find first non-zero
for (let k = 0; k < this.nFreqBins; k++) {
if (this.melFilterbank[offset + k] !== 0) {
start = k;
break;
}
}
// Find last non-zero
for (let k = this.nFreqBins - 1; k >= 0; k--) {
if (this.melFilterbank[offset + k] !== 0) {
end = k + 1;
break;
}
}
this._fbStart[m] = start;
this._fbEnd[m] = end;
}

// Pre-allocate reusable buffers
this._fftRe = new Float64Array(this.nFft);
this._fftIm = new Float64Array(this.nFft);
Expand Down Expand Up @@ -315,7 +341,7 @@ export class MelSpectrogram {
// 4. STFT + Power + Mel + Log
const rawMel = new Float32Array(this.nMels * nFrames);
const { _fftRe: fftRe, _fftIm: fftIm, _powerBuf: powerBuf } = this;
const { hannWindow: window, melFilterbank: fb, nMels, twiddles: tw, nFft, nFreqBins, hopLength, logZeroGuard } = this;
const { hannWindow: window, melFilterbank: fb, nMels, twiddles: tw, nFft, nFreqBins, hopLength, logZeroGuard, _fbStart, _fbEnd } = this;

for (let t = 0; t < nFrames; t++) {
const offset = t * hopLength;
Expand All @@ -325,7 +351,8 @@ export class MelSpectrogram {
for (let m = 0; m < nMels; m++) {
let melVal = 0;
const fbOff = m * nFreqBins;
for (let k = 0; k < nFreqBins; k++) melVal += powerBuf[k] * fb[fbOff + k];
// Optimization: only iterate over non-zero filterbank elements
Copy link

@kiloconnect kiloconnect bot Feb 21, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

SUGGESTION: The optimization comment is accurate, but consider also noting that correctness relies on the filterbank having contiguous non-zero support per row (triangular filters). If a non-standard filterbank with non-contiguous non-zero elements were ever used, this sparse-range approach would silently skip interior zeros that happen to be zero but are surrounded by non-zeros — though for the Slaney triangular filterbank this is guaranteed safe. A one-line note would help future maintainers:

Suggested change
// Optimization: only iterate over non-zero filterbank elements
// Optimization: only iterate over non-zero filterbank elements (valid for contiguous triangular filters)

for (let k = _fbStart[m]; k < _fbEnd[m]; k++) melVal += powerBuf[k] * fb[fbOff + k];
rawMel[m * nFrames + t] = Math.log(melVal + logZeroGuard);
}
}
Expand Down