Tolerance adjustment to mdf_inner_prod in the test (kernel is unchanged) to previous merge (!49). The float check measured the C-vs-RVV rounding difference as a fraction of the dot product's own result, which cancellation can drive to nearl…
Jean-Marc-pred-instability-results-20260805.zip @jmvalin I tested your patch on top of commit 3da9f7a6, and it fixes both issue #382 and our independent DTX-off reproduction. The second case uses one continuous encoder and a two-packet gap.…
SpacemiT K1 (VLEN=256): ┌──────┬─────────┬─────────┬─────────┐ │ pair │ main │ branch │ speedup │ ├──────┼─────────┼─────────┼─────────┤ │ 1 │ 4.536 s │ 4.050 s │ 1.120× │ ├──────┼─────────┼─────────┼─────────┤ │ 2 │ 4.505 s │ 4.067 s │ 1.1…
The SSE and SSE2 probes compiled their test programs with -msse/-msse2, but those flags were never added to the real compile lines. On x86-64 this goes unnoticed since SSE2 is part of the base ABI, but on i686 the probes succeed while the a…
To validate before and after, use this script: $ cat ci-i386.sh #!/bin/sh # Emulate the gitlab i686 runner: debian:bookworm on linux/386. set -ex apt-get update -qq apt-get install -y -qq ninja-build meson gcc git nasm > /dev/null cd /src r…
OK, can you give this a try? It seems to fix the issue, but I've done no real testing on it yet. Please report how it behaves.pred_instability_fix1.patch
Problem The input from xiph/opus#382 contains a low-level six-sample pattern during silence: [0, 0, 0, 0, 0, 8] at 8 kHz. One possible source is T1 CAS robbed-bit signaling, which uses the least significant bit of every sixth T1 frame. The
The page seeking code includes a special case for when we cannot find a page with a granuleposition smaller than the target, in which case it starts reading from the beginning of the stream. However, it did so by assuming that the page that…
mdf: vectorize the adaptive-filter weight update mdf: convert X-history shift loop into a single SPEEX_MOVE (all archs) End to end speed up for bench_echo: ┌─────────────────────────────────────┬─────────────┬─────────────┬─────────┐ │ boar…
The FFT and power spectrum in speex_echo_get_residual already ride the existing RVV kernels; hoist its two remaining elementwise loops into OVERRIDE_MDF_RESIDUAL_* hooks and add kernels for both builds. All four are bit-exact, including flo…
The columns are too narrow for addresses and UA strings to fit. This makes the table hard to read and long. I have miles of white space on both sides of the table, so the table is artificially restricting width. v2.4 was much more readable.…
RomanG replied to my request for help on the icecast-dev@xiph.org mailing list: "Re: bug in v2.50 - aliases incorrect behavior" 2026-07-27 11:52 pm (his reply below) I followed Romans instructions - that got the links working correctly. But…