Skip to content

A moving dial's second capture runs on a second core - #4

Merged
rations merged 1 commit into
masterfrom
parallel-processing
Oct 3, 2026
Merged

rations merged 1 commit into
masterfrom
parallel-processing

Conversation

@rations

@rations rations commented Oct 3, 2026

Copy link
Copy Markdown
Owner

While the sounding channel's dial moves, CrossfadeEngine runs two models per chunk: branch A, and branch B one capture up, crossfaded by weight. On one thread that doubles the amp's cost for as long as the dial moves. That was the xrun heard when turning a dial, in the plug-in under a DAW and in both standalones alike. No host splits one plug-in's work across threads.

The two calls are independent: two different nam::DSP instances, read-only shared input, separate outputs. core/branchrunner.{h,cpp} gives the engine a second thread to hand branch B to:

  • RT posts B, runs A, then takes B back with one compare-exchange if the runner has not claimed it yet, and runs it itself. A late, starved or absent runner therefore costs the serial path and never more;
  • RT waits only for a job the runner has already started, which is bounded by one model call on at most kChunk samples. The wait spins on an acquire load with a pause instruction: no syscall, no lock;
  • the runner is woken by a semaphore post, which does not block. It sleeps between jobs and costs nothing while the dial is still;
  • it is never less real-time than the audio thread, or it claims nothing. On Linux it copies the audio thread's scheduling; under JACK it came up SCHED_FIFO 65, like the client thread. On Windows it joins MMCSS "Pro Audio", loading avrt.dll at run time so the plug-in gains no import. macOS runs serially for now;
  • it allocates nothing after start(), arms flush-to-zero as the audio thread does, and is joined in ChannelRack::stop().

The engine's single input-calibration scratch buffer is split in two. It was shared "because the branches are consumed in turn"; once they overlap, B's scaling would overwrite what A was reading.

RATIONS_SERIAL_BRANCHES in the environment keeps the runner off, so the same binary can be measured both ways and there is an escape hatch.

Measured with rations_jackcheck --sweep-gain at 128 frames and 48 kHz, the same binary with and without RATIONS_SERIAL_BRANCHES, on 10-capture banks and on a 24-capture bank, stomping every 1200 ms and every 250 ms:

                 serial (14 runs)    branch runner (15 runs)

worst block 54-100% of period 33-55%
mean DSP load 31-47% 26-31%
xruns 2 0

Verified:

  • rations_switchcheck gains a proof that renders a whole-run dial sweep serial and parallel and requires the two to be bit-identical. They are, with input calibration on and off. It also requires the runner to have run branches (about 2,500 per run), or the comparison would be serial against serial. Both products pass;
  • rations_racecheck under a ThreadSanitizer build (NAMP_ENABLE_TSAN, its deliberate-race harness flagged) is clean with the runner running about 11,700 branches, and with --republish. It now fails if the runner ran none;
  • rations_rtcheck in both products gains a section that counts allocations with the sounding dial sweeping. The counters are process-wide, so the runner's thread is inside the count. Zero allocations, and it fails if the runner ran nothing;
  • namp-rack with a full rack (two VST3s before the amp, a VST3 and an LV2 after it): 0 xruns over 90 s with the dial still, at the same mean load as before;
  • phase-gate.sh passes on build and build-rack; the SDK validator gives 47/47 on both bundles; the Windows cross builds compile, and the plug-in's DLL import list is unchanged. The macOS branch was syntax- checked on Linux with linux undefined; the CI build is its gate.

Not verified: the Windows path on a real Windows machine under ASIO, and a dial sweep in namp-rack with hosted plug-ins loaded (namp-rack has no scripted dial sweep; the amp-only measurement above stands in for it).

While the sounding channel's dial moves, CrossfadeEngine runs two models
per chunk: branch A, and branch B one capture up, crossfaded by weight.
On one thread that doubles the amp's cost for as long as the dial moves.
That was the xrun heard when turning a dial, in the plug-in under a DAW and
in both standalones alike. No host splits one plug-in's work across threads.

The two calls are independent: two different nam::DSP instances, read-only
shared input, separate outputs. core/branchrunner.{h,cpp} gives the engine
a second thread to hand branch B to:

  * RT posts B, runs A, then takes B back with one compare-exchange if the
    runner has not claimed it yet, and runs it itself. A late, starved or
    absent runner therefore costs the serial path and never more;
  * RT waits only for a job the runner has already started, which is
    bounded by one model call on at most kChunk samples. The wait spins on
    an acquire load with a pause instruction: no syscall, no lock;
  * the runner is woken by a semaphore post, which does not block. It
    sleeps between jobs and costs nothing while the dial is still;
  * it is never less real-time than the audio thread, or it claims nothing.
    On Linux it copies the audio thread's scheduling; under JACK it came up
    SCHED_FIFO 65, like the client thread. On Windows it joins MMCSS "Pro
    Audio", loading avrt.dll at run time so the plug-in gains no import.
    macOS runs serially for now;
  * it allocates nothing after start(), arms flush-to-zero as the audio
    thread does, and is joined in ChannelRack::stop().

The engine's single input-calibration scratch buffer is split in two. It
was shared "because the branches are consumed in turn"; once they overlap,
B's scaling would overwrite what A was reading.

RATIONS_SERIAL_BRANCHES in the environment keeps the runner off, so the
same binary can be measured both ways and there is an escape hatch.

Measured with rations_jackcheck --sweep-gain at 128 frames and 48 kHz,
the same binary with and without RATIONS_SERIAL_BRANCHES, on 10-capture
banks and on a 24-capture bank, stomping every 1200 ms and every 250 ms:

                     serial (14 runs)    branch runner (15 runs)
  worst block        54-100% of period   33-55%
  mean DSP load      31-47%              26-31%
  xruns              2                   0

Verified:
  * rations_switchcheck gains a proof that renders a whole-run dial sweep
    serial and parallel and requires the two to be bit-identical. They
    are, with input calibration on and off. It also requires the runner
    to have run branches (about 2,500 per run), or the comparison would be
    serial against serial. Both products pass;
  * rations_racecheck under a ThreadSanitizer build (NAMP_ENABLE_TSAN, its
    deliberate-race harness flagged) is clean with the runner running about
    11,700 branches, and with --republish. It now fails if the runner ran
    none;
  * rations_rtcheck in both products gains a section that counts
    allocations with the sounding dial sweeping. The counters are
    process-wide, so the runner's thread is inside the count. Zero
    allocations, and it fails if the runner ran nothing;
  * namp-rack with a full rack (two VST3s before the amp, a VST3 and an
    LV2 after it): 0 xruns over 90 s with the dial still, at the same
    mean load as before;
  * phase-gate.sh passes on build and build-rack; the SDK validator gives
    47/47 on both bundles; the Windows cross builds compile, and the
    plug-in's DLL import list is unchanged. The macOS branch was syntax-
    checked on Linux with __linux__ undefined; the CI build is its gate.

Not verified: the Windows path on a real Windows machine under ASIO, and a
dial sweep in namp-rack with hosted plug-ins loaded (namp-rack has no
scripted dial sweep; the amp-only measurement above stands in for it).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@rations
rations merged commit cceb7bd into master Oct 3, 2026
3 of 5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant