Skip to content

Release v0.7.0 - #1

Merged
FueledByRedBull merged 18 commits into
mainfrom
release/v0.7.0
Oct 23, 2026
Merged

FueledByRedBull merged 18 commits into
mainfrom
release/v0.7.0

Conversation

@FueledByRedBull

Copy link
Copy Markdown
Owner

Brings v0.7.0 to main: the simulation and performance work, plus a cleanup pass.

What changes for players:

  • Pools settle flat and still (ghost-fluid surface on a smoothed level set).
  • Sparks and embers are particles that follow the smoke, bounce off rock and die in water.
  • Fire no longer makes frames alternate heavy and light; the gas solve spans two ticks.
  • Packed sand collapses into dug holes like loose sand, damp sand keeps per-grain bonds when it wakes, slow creep hardens, and the sand brush has a pool budget.
  • F10 shows real pass timings; startup errors explain themselves.

Repository changes: AGENTS.md is no longer tracked, CI and release builds run the config-parsing test, docs no longer carry stale numbers, and the README has a download section.

How it was checked, on an RX 7800 XT with nothing else on the GPU:

  • ctest 103/103.
  • POWDER_REPLAY=all passes both repeats of every scene; the cleanup commits leave every replay signature byte-identical.
  • benchmark-realtime passes: combined about 62-63 FPS, p95 about 21 ms.
  • The exe runs alone from a fresh folder, and the package check accepts a release-shaped zip.

Not checked: live play in a window (including the F10 bars on screen), other GPUs and drivers, and Linux or MinGW builds locally.

- Water: ghost-fluid free surface on a smoothed density level set, so
  settled pools are still and level; the GPU arms the CG restart, which
  removes a blocking readback from most solves
- Gas: each pressure solve spans a tick pair, so frames carry even work
- Sand: packed piles collapse into dug holes like moving ones, damp
  cells wake with their grains' own bonds, slow creep hardens, and the
  sand brush has a deterministic pool budget
- Sparks and embers are GPU particles stepped per tick and drawn in lit
  frames; benchmarks ignore the real cursor
- Stop tracking AGENTS.md
Lead the quick start with the release ZIP and its checksum file, label
the build steps as building from source, and describe ROADMAP.md by its
new title.
PLAN now points at VALIDATION for test counts and timings. The algorithm
reference no longer carries a version in its title.
ROADMAP gets a current handoff and a new title; the v0.6.0 handoff is
marked historical. VALIDATION opens with the command reference, records
that the integration is committed as 12c1343, and moves the old water
review pass into the historical region. Process and tooling wording in the
notes is reworded without removing any evidence.
Record the last GLFW error description and append it to the window and
OpenGL 4.3 failure messages. After the textures and buffers are created,
report GL_OUT_OF_MEMORY with the water and sand pool sizes, after
dropping stale errors so an old one is not blamed on the allocations.
A failed glClientWaitSync or a null glMapBufferRange result in the
pressure loop was ignored and the mapped pointer was dereferenced.
Print the unknown name and the valid scenes to stderr. The exit code is
still 1.
Multiply the pool byte sizes in 64-bit instead of int so a larger pool
cannot overflow, and drop the literal 12.
The per-program name map hashes transparently, so CachedUniform no longer
builds a std::string for every call.
BrushConfig::clear_water and clear_gas were never set, so the clearGas
uniform and its block in spawn_full.comp could not run; erase still uses
clearWater. Initialize either completes or throws, so it returns void.
Also drop the smokeSize uniform no shader declares and a redundant
(void)region.
The hand-written list had drifted and omitted water-sparse-face.
The pass-timing bars only ever filled during replays; the interactive loop
issued the GL_TIME_ELAPSED queries every tick and never read them. With the
F10 view up, alternate two query sets per frame and read the previous one
once the GPU reports every query in it available, keeping the last values
until then so the frame never waits. Other views skip the readback.

Drop the unused no-argument ReadBackTimingQueries overload.
The test needs no OpenGL context, so it can run on the hosted runners
after each build and fail the job on a regression.
The water P2G weight comment now uses the shortcut: prefix.
The Windows CI runner crashes config-parsing before any buffered output reaches ctest; unbuffered stdout shows which check ran last.
On the GitHub Windows runner the parent crashed inside _spawnve while
starting the --empty child with a one-entry environment. Start the child
with CreateProcessA instead and give it this process's environment plus
the empty fixture, so it starts the way any other process does.
@FueledByRedBull
FueledByRedBull merged commit 0bcb329 into main Oct 23, 2026
2 checks passed
@FueledByRedBull
FueledByRedBull deleted the release/v0.7.0 branch October 23, 2026 17:05
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