32dd48f28ffe0c1362cf11bdadf0546d525ef382
3 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
32dd48f28f
|
Add a pluggable broad phase, and the incremental uniform grid behind it
akgl_Partitioner is a record of function pointers and an initializer, the same shape as the render and physics backends. `move` is its own slot rather than remove-then-insert, and that is the whole design: a uniform grid's move is a comparison and a return when the proxy has not left the cells it was in, which is the steady state for a walking actor and the permanent state for a static one. Spelling it as remove-and-insert would turn the incremental grid into the rebuild-every-frame tree that PERFORMANCE.md already measured and rejected. The grid is a dense 128x128 array of cell heads over the world, cells keyed on tile size, with two intrusive chains per entry -- one through the cell so a query can walk it, one through the proxy so removal unlinks a whole span without searching. Everything is an int16_t index rather than a pointer, so the structure is relocatable and clearing it is a memset. A proxy covering more cells than AKGL_COLLISION_GRID_MAX_SPAN spills onto one chain every query walks, which is what makes the cell pool's ceiling provable rather than hopeful. tests/partition.c compares every query against a linear scan over the same proxies and asserts containment in one direction, not equality. That asymmetry is the contract: over-reporting costs a narrowphase call, under-reporting is a wall an actor walks through. The suite is a table over partitioners so the same assertions run against every implementation, which is what the second one landing later will be held to. Three deliberate breaks, and two of them were not caught the first time: - Removing the min-cell rule so a shared-cell pair reports repeatedly: caught. - Dropping the unlink in `move`: **not caught**, initially. Stale entries do not produce wrong query answers, because the bounds re-check filters them out -- they leak the cell pool until the grid stops working, on a timescale no unit test reaches by accident. Counting live pool entries across 200 moves is the only thing that sees it, and now does: 505 entries against an expected 5. - Swapping floorf for a truncating cast: **not caught, and correctly so.** Everything below zero is clamped to cell 0 either way, so today the two are indistinguishable. The comment claiming truncation is a defect was overstating it and now says what is actually true: the clamp is the only thing hiding it, and a negative world origin or a relaxed clamp would make it real with no test standing in front of it. Co-Authored-By: Claude Code <noreply@anthropic.com> Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-Authored-By: Andrew Kesterson <andrew@aklabs.net> |
|||
|
cfe2d6a09c
|
Answer whether two shapes overlap, and how to undo it
akgl_collision_test takes two proxies and fills in a contact: a unit normal, a depth, and a point. Three paths, cheapest first. The proxies' bounds reject most pairs in four comparisons and no arithmetic, which is free because the bounds were computed when the proxy was synced. A box against a box is answered in closed form. That is not only cheaper than the general solver, it is *exact*, and the difference shows up where it matters most: an actor resting on a floor wants a normal of precisely (0, -1, 0), and an iterative solver converges to something like (0.0001, -0.99999, 0), which accumulates into a slow sideways creep. A tile game is almost entirely boxes, so this is the path that runs. Everything else goes to ccdMPRPenetration. MPR rather than GJK+EPA because MPR allocates nothing at all, converges in fewer iterations, and its one weakness -- a coarser contact *point* -- is on a field the blocking resolver never reads. EPA stays compiled and available behind the arena for a caller who one day wants an accurate manifold. The header says the point is approximate so nobody builds a damage falloff on it. The normal points out of the second shape and toward the first, so a caller moving `a` along it by `depth` separates the pair. libccd hands back the opposite convention; converting once here saves every resolver from remembering the sign, and getting it backwards would compile, would pass any "do these collide" test, and would drag actors into walls. There is a test that asserts the direction, and inverting the conversion turns it red. AKGL_COLLISION_TEST_PLANAR flattens the normal into the xy plane. The extrusion the setters apply already makes z the most expensive axis, so this is a net for a caller who set a depth by hand -- and the failure it catches is silent: the narrowphase reports a contact, the resolver pushes the actor into the screen, and the actor does not move on screen while remaining inside the floor. It also rescues the degenerate case, two shapes at exactly the same centre, where there is no planar direction at all and a zero-length normal would be a wall that stops nothing. Level authors stack things constantly. Three things the tests found rather than assumed: - The guard has two copies, one per path, and the box one is where it earns its place. Removing both is caught; removing only the iterative one is not, because the iterative solver is seeded from the line between the two centres and so converges to a planar answer on its own for a planar offset. That is measured and written down in the test rather than papered over with a fixture contrived to force it. - A concentric pair has no preferred sign on the degenerate axis, so the test asserts the magnitude of the normal rather than its direction. The first version asserted -1 and failed against an equally correct +1. - The box fast path and the iterative solver are two implementations of one answer, so they are driven over the same arrangements and required to agree on the decision and the axis. A shortcut nothing cross-checks is a shortcut waiting to diverge. Co-Authored-By: Claude Code <noreply@anthropic.com> Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-Authored-By: Andrew Kesterson <andrew@aklabs.net> |
|||
|
7e0e5a7841
|
Give libccd a static arena so libakgl still allocates nothing
libakgl does not call malloc at runtime. That is stated in seven places in the manual and is AGENTS.md's second standing rule, and every object comes from a fixed array in akgl/heap.h. libccd's EPA path does not know that: it builds its expanding polytope out of realloc and free, and ccdGJKPenetration documents a -2 return for when that fails. The alternative was to compile only the three files MPR needs and leave EPA out of the build. That works and it puts a copy of somebody else's source list in this repository, where it rots silently on the next submodule bump. This instead compiles all of libccd and points its allocator at a bump allocator over static BSS, which keeps the promise literally true -- and keeps EPA available rather than amputated, for the day a precise contact manifold is worth having. The arena is reset at the top of each query rather than freed block by block, so the lifetime is one narrowphase call, `free` is a no-op, and allocation is a pointer bump. Exhaustion returns NULL, which libccd already handles by unwinding to -2, which becomes AKGL_ERR_COLLISION naming the high-water mark -- a loud failure with a number in it rather than a silently missed collision, which would be a floor an actor falls through reported as success. One GJK/EPA box pair costs 7,264 bytes, measured and printed by the suite. The 64 KB ceiling is that with about nine times headroom, and the high-water mark is reported so the next reader can re-derive it rather than trust this line. The redirect lives in src/ccd_arena_shim.h, injected with -include, and not in CMake's COMPILE_DEFINITIONS. It was in COMPILE_DEFINITIONS first, and that is a mistake worth recording: CMake cannot carry a function-like macro through a -D, so it dropped __CCD_ALLOC_MEMORY without a diagnostic. libccd went on calling the C library's realloc while the shim's `free` quietly discarded the results -- strictly worse than doing nothing, and invisible, because it leaks rather than crashes. What caught it was insisting the test prove the wiring rather than the outcome. The first version asserted the arena balanced back to zero after a query, which turned out to be the wrong assertion for a different reason -- free is a no-op by design, so it cannot balance -- but a test that had merely checked "two boxes collide" would have passed throughout, against an allocator nothing was using. The suite now asserts what is actually provable: that a query allocates from the arena at all, and that the process survives, since glibc aborts when the real free(3) is handed a pointer it never issued. Also here: AKGL_ERR_COLLISION, with the name registered -- tests/error.c asserts the band and the names agree, and it caught the missing one immediately. -fvisibility=hidden and CCD_STATIC_DEFINE keep every ccd* symbol out of libakgl.so's dynamic table, which `nm -D` confirms is empty; AGENTS.md records a shipped defect where an exported `renderer` was preempted by a same-named symbol elsewhere, and a game linking a system libccd would hit exactly that. Co-Authored-By: Claude Code <noreply@anthropic.com> Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Co-Authored-By: Andrew Kesterson <andrew@aklabs.net> |
