| Commit message (Collapse) | Author | Age |
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Sound a sketch makes leaves for the device about one pull after draw, and
SDL opens 1024-sample pulls at 48 kHz (21 ms); 256 leaves in about 6.5 ms,
but the audio server then runs every client at that size, and drops out
under load. The only way to ask was SDL_AUDIO_DEVICE_SAMPLE_FRAMES in the
environment. Now an app declares `pub const audio_pull = 256;` beside its
name; abi.Info carries it (0, SDL's own, when it is not declared), and the
host sets it as SDL's hint just before it opens the device, at normal
priority, so the environment variable still wins. It is read once, at
start; the web host has no device to open and ignores it. The sound
overlay's header says what the device took: OUTPUT ch1 / ... / pull 256.
Test: audio, on SDL's dummy driver: the device opened with nothing asked
takes 1024 at once, asked 256 takes 256, and with 512 in SDL's copy of the
environment takes 512 though 256 was asked. test-audio: 48/48 steps, 9/9
tests. README: the app table's row, and a paragraph in sound with the
trade-off.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_016Y9MNBECMWcJD5pdS4SR6T
|
| |\ \
| | |
| | |
| | |
| | |
| | |
| | |
| | |
| | | |
objects
The Store's deletion with shared-page keyframes and the 64 MiB cap; f.audioBuffer and in.lead with spheres on them; the embedded-SVG image, Typst's root, part, and the manual example. No conflicts; zig build test: 179/179 steps, 140/140 tests; spheres 75/75, micscope 67/67.
Co-Authored-By: Claude Fable 5.2 <[email protected]>
Claude-Session: https://claude.ai/code/session_016Y9MNBECMWcJD5pdS4SR6T
|
| | |/
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
A compute shader's output was heard only after the sketch read it back
(f.download, a frame later) and pushed it behind a lead of its own
choosing; spheres kept 100 ms, and a change it made left for the device
96 ms later on average (measured, dummy driver, 60 Hz).
f.audioBuffer(key, buffer, .{ .count, .offset }) streams `count` f32 a
shader wrote: the host copies them back while it runs the frame's
commands, waiting on the copy's fence (Executor.read: 0.2 ms on average
here, 2 ms at the 99th percentile on a loaded machine), and queues them
on the stream, so nothing is in flight between frames and the sketch
keeps no books. Input.lead says how far ahead to make sound: the slowest
of the last 64 frames' worth of samples (timed between the ends of the
frames' commands, where the pushes are), plus the device's pull (what
SDL takes from a stream at once, read from the device: 1024 here), plus
half a pull for each underrun, a sample less each frame after; capped at
8192. A sketch pushes `in.lead -| in.stream(key).queued` a frame, and
its change leaves after about what is queued ahead of it, a pull: 19 to
27 ms on average with SDL's 1024-sample pulls, 6.3 to 6.5 ms with 256
(SDL_AUDIO_DEVICE_SAMPLE_FRAMES; 17 under heavy load). Stream.underruns
counts the pushes that found a stream empty though the frame before had
fed it (a gap).
The voices keep their 2048 samples: mixed to the lead, a one-shot longer
than it is mixed over several frames, and voices.script's expect-audio
and expect-no-audio, which look one frame at a time, failed.
On the web, audio.buffer reads the buffer back with mapAsync and queues
it on arrival, in order, counting it as queued meanwhile; in.lead stays
at 100 ms there.
The measurements (a probe patched into the host for the purpose, kept
out of the tree) are in the 02-pardes-code scratch, .scratch/sketch/gpusynth.
Tests: the lead's arithmetic (audio.zig, in test-audio's filter); spheres'
play script checks its stream is fed. zig build test: 164/164 steps,
49/49 tests run (the rest cached); zig build test-audio: 48/48 steps, with
the voices sketch's wasm, so web.zig compiles; voices.script passes four
runs in four. web/host.js: node --check only, no browser run.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_016Y9MNBECMWcJD5pdS4SR6T
|
| | |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
first
A save grew without bound. Now Sessions.max_save is 64 MiB and
layout.max_state 32 MiB, so history always has room (the compile error
says so: "State is N bytes, and may be 32 MiB, half of a save, so
history has room: ...").
A commit that leaves the file past the cap trims it (Sessions.trim): one
transaction drops history oldest first, a keyframe by removing its pages
(its frame, batch and time stay, so keyframe numbers, keyBefore and
marks by index still resolve) and a batch of frames by emptying its
slot, until the keyframes' written pages and the batches left count
three quarters of the cap; then the file is compacted into `<path>.new`
and renamed over it, and the run goes on in the new file (compact now
reopens). A dropped keyframe's pages may still be shared by a later one,
so a round can free less than it counted: the next round lowers the
count by the overshoot, four rounds at most. Never dropped: the present
(the newest keyframe and the batches from its own on), the keyframe of
the newest mark of each name, the keyframe of the newest clip of each
name and every batch holding one of its frames. A session left with
nothing loses its keyframes, batches and stretches; its summary stays.
Asked for, what was dropped is error.HistoryDropped, which the console
says in words ("That part of the history is gone: a save keeps 64 MiB,
the newest.").
Marks are pinned under the cap: a mark of a new name needs room for a
whole State, history trimmed to make it; without, it is refused,
error.SaveFull, the console saying how to make room: save over a mark's
name (the older one then goes with history) or start another save. A
mark over an existing name is always taken. (The spec's "room" rule,
refusing only after a trim failed, let one mark take the file past the
cap; room for a whole State keeps the cap hard, at the price of
refusing early for a big State.)
Builds that do not record (`history` false, from Info.record) use the
same mechanism with the history limit zero: every commit holding a
keyframe drops all it may (no compaction unless past the cap), and the
exit compacts, so the disk holds at most the last minute of input while
it runs and the present and the marks after. The exit compaction a
recording build made past twice its contents plus 64 MiB is gone: the
cap's compaction does that work.
Measured (Debug, unit test): a trim of a 64 MiB file to 48 MiB takes
165-187 ms (two rounds of drop and compaction, the first ending 0.5%
over); a no-history trim at a keyframe, 0-5 ms. ReleaseFast not
measured.
The console's `marks` reads each mark's keyframe fields, not its State:
an older mark of a name may have gone with history (listed "dropped"
when its session went whole). Docs: record means "keep history" in
App.Options, sketch.Build and abi.Info.
Tests: zig build test, 168/168 steps, 137/137 tests. New, in
Sessions.zig: 100 keyframes of 1 MiB each with marks at the 5th and
20th: the file never past the cap, the oldest keyframes and frames
HistoryDropped, both marks and the newest keyframe whole, the present's
frames readable; without history, 20 keyframes and a mark: only the
mark and the present stay, the file compacted at exit under three times
the State and the present reopening whole; 2 MiB marks until one is
refused (SaveFull, the file under the cap), then one over an old name
taken and the file under the cap, every mark whole. micscope 67/67,
10/10; spheres 75/75, 3/3; noc e2e 292/292; the drat-meout copy 87/87,
30/30.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_016Y9MNBECMWcJD5pdS4SR6T
|
| |/
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A sketch's data is State: flat bytes, what grows has its capacity in its
type. The Store, a heap in a file mapped at 0x2000_0000_0000 with 64 GiB
reserved, its power-of-two allocator, its root adopt and migration, and
its test, are gone from memory.zig; so are Context.store and store_root,
Info.store, Memory.store, the Store declaration's glue, Init.store,
Frame.store and Console.store, the web host's store in wasm memory, the
host's StoreFile wiring and `<state>.store` file, the `.store` suffix in
moving an old state into .sketchbook/, the build's tests' `rm` of it, and
layout.check's `pointers` flag (State is the one checked memory now).
Its one real user, micscope, moved off it in the change before. Old
`.state.store` files are not read, converted or deleted, and stay
ignored by .gitignore. A sketch that still declares a Store gets no
error for the declaration: nothing reads it, and a `*Store` parameter is
a compile error (no argument of that type).
layout.of(T, .state) refuses a State bigger than layout.max_state, 64 MiB,
one constant: "State is N bytes, and a save holds at most 64 MiB: lower
a capacity (an array, an sk.List), or keep what need not be saved in
Session." The pointer-hash gap the spec found (a Store pointer's
element type changing kept the old bytes) is gone with the only
persisted memory that held pointers.
README: no Store row, allocator row, migration paragraph, file bullet or
known limits; draw and command lose their store parameter; State's row
says 64 MiB at most; data that grows and is the user's work goes in a
List in State or, bigger than a save, in files the sketch appends to.
Tests: zig build test, 166/166 steps (the store's unit test gone; new: a
65-MiB State fails to compile with the message above, by expect_errors).
micscope, spheres, noc's e2e and drat-meout's test unchanged and green.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_016Y9MNBECMWcJD5pdS4SR6T
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Voice.rate is the voice's playback speed and so its pitch: 2 is an octave
up and half as long, 0 holds it, and changing it does not restart the
voice. The mixer keeps a fractional cursor and resamples by linear
interpolation between a sample and the next (the last toward silence, or
the first when looping): the least code that sounds right, a little dull
far from 1; at rate 1 the samples come out exactly. An infinite or NaN
rate ends the voice instead of crashing the conversion.
Voice.group (0-7) puts a voice under a group's gain, which draw declares
each frame with f.gain(group, value) (Command.gain); a group not given
one plays at 1, so it is declared state and replays like the voices. The
mixer glides each group toward its gain by 1/480 a sample (10 ms for a
full swing), so a gain set in one step does not click. Music ducking
under speech is then the sketch lowering one number; nothing ducks by
itself. The web host ignores gains as it ignores voices.
Test: half and double speed against the interpolated samples, rate 0
holding, an infinite rate ending; a looping voice in group 1 at gain 0.5
gliding there and staying while group 0's is untouched, and going back to
1 when the gain is no longer declared. test-audio also builds the voices
sketch for the web, which compiles the web host's newSound and gain. The
README lists rate, group and f.gain with what they do.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_016Y9MNBECMWcJD5pdS4SR6T
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
02-dont-rat-me-out's synth.zig, a pure and deterministic port of
DrPetter's sfxr, moves into the library as src/sketch/sfxr.zig, exposed as
sk.sfxr: render(params, rate, out) gives a coin, blip, laser or explosion
as samples for host.newSound. Its knobs count in samples at 44.1 kHz, so
the one change is the rate: every length (envelope, arpeggio, repeat,
phaser offset, period) grows by rate/44.1 kHz and every per-sample step
(slides, duty sweep, vibrato, filter coefficients and their ramps)
shrinks by it, to first order for the slides and filters; at 44.1 kHz
each factor is 1 and the samples are the original's to the bit. Two
spellings change without changing a sample: the phaser buffer is
@splat (Zig 0.17 has no ** on arrays) and the phaser delay is clamped
before it is converted, so a long sweep cannot overflow it.
Test: four parameter sets that between them turn every knob render the
lengths and Wyhash of the samples the original rendered at 44.1 kHz (run
once, outside the repo), and at 48 kHz a sound as long within 1 %.
test/voices.zig makes its blip with sfxr in draw, at in.sample_rate, and
is the README's example: a blip in six lines.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_016Y9MNBECMWcJD5pdS4SR6T
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Live play was recorded only after someone typed record N, into 16 fixed
slots. Now every run is a session in one xitdb database per sketch, beside
its state file (<name>.xitdb): an append-only log of the frames live
update was fed, one commit a second of session time, and keyframes of the
State at start, every minute, on jumps, before and after reloads and at
exit. A kill -9 loses at most the uncommitted second (test-kill).
Whether a build records travels in the sketch: App.Options.record, null
meaning Debug, reaching the host as Info.record. Scripted runs from fresh
state record in memory, exports not at all. Each stretch of a session
keeps its build, layout and Info.source: the repository revision (one jj
or git call per build, build_info), zig version, backend and mode.
Sessions.zig takes plain data (bytes, events, strings, a path, an Io), so
a second host can drive it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_016Y9MNBECMWcJD5pdS4SR6T
|
| |
|
|
|
|
|
|
| |
abi.hash now covers the zig version, codegen backend and optimize mode: Zig's
own calling convention (Host.VTable, every Allocator) differs between them,
so a Debug host loading a ReleaseFast or LLVM library faulted on its first
draw. Libraries export sketch_build, and the host's refusal says what
differs and what to run instead.
|
| |
|
|
|
|
|
|
|
|
| |
executor pardes shares
`f.postProcess(source, target, .{...})` records a fullscreen pass that samples a completed render into a distinct target image: colour at fragment binding 0 and, with `sample_depth`, depth at 1, which needs the source's latest pass to have depth and one sample (missing depth, multisampled depth and feedback loops are rejected). The scene format's version 4 carries it, the executor records it on the caller's command buffer and owns the sampler, the web host does the same with its texture and sampler pairs, and native profiling reports the work as `postprocess`. The README says how to use it.
docs/pardes-rendering.md describes the executor as pardes uses it for addon scenes: a borrowed SDL GPU device, `record` into the caller's command buffer, `run`, `output` and `postProcess`.
The working copy as it stood on 2026-11-16; message written by Claude from the diff, at the author's request.
|
| | |
|
| |
|
|
| |
Draw native GPU frames through the same bounded canvas used by headless capture. Backdrop commands reuse one snapshot and shader state, with an unchanged fallback on unsupported renderers. Verified through native debugger lenses, restart, and repeated font zoom checks.
|
| |
|
|
|
|
|
|
| |
canim keeps its table of places on the frame rather than in a process
global, so two players at once each have their own.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_016Y9MNCKgYCGiANnEwcEQxa
|
| |
|
|
| |
(f.textScaled), steady under a moving camera
|
| |
|
|
| |
host.textSizeIn, sk.Face)
|
| | |
|
| |
|
|
| |
spectrum, spectrogram), audio codecs and fonts, web host
|
| |
|
|
|
|
|
|
|
|
|
| |
- input.started: true on the first update of a run, never after a reload,
for a sketch to stop what should not resume from saved state.
- A console line that errors fails its script. Sketch commands answer
unknown, done or failed (abi.Outcome) instead of a bool that hid errors.
- addScript also takes the hot-reloadable library, run through the generic
host, so scripts can reload a real build.
- Sounds that do not load are .none, not an empty clip.
- sk_tag reads an enum field's tag name from the web build, for tests.
|
|
|
The root is the game. The platform library moves to lib/sketchbook (with
serve_web.py, which serves any sketch's web build, as a tool of it), and
vecanim to sketches/vecanim, still its own package on the same library.
|