Artemking4 found the root cause of the bone 0 problem in IMesh:DrawSkinned: in shaderapidx8.cpp, bone 0 is also loaded as the model matrix, so rendering breaks unless it is reset afterwards. Wrapping the call in cam.PushModelMatrix and the matching pop fixes it, and he asks for bone 0 to be exposed.

ArtyomChyornyj2035 reported 0 server FPS and "Connection problem" warnings on entity-heavy maps (x86-64 build 26168646).

In #next-update, Redox asked to enable gmod_mcore_test by default on main and dev too before x86-64 goes away. On 24 Oct Darth replied it's already 1 in both x86-64 and dev, but not main as far as he knows.

In #developer-talk, Redox shared his gmod_perfutils script that detours physics object functions to find expensive calls, after full metal sergeant 228 said one of his functions might be an issue.

artemking4 commented on the 2016 AVX2 request that SSE2 can be enabled safely since LuaJIT already requires it.

In #lua-help, A1steaksa asked what causes Render queue too large, flushing! when rendering an IMesh. Rubat explained the multicore rendering queue is too large: over 32 MB of vertices or indices, or over 15 MB of data in the queue. He named mat_report_queue_status for stats. full metal sergeant 228 said he hit it with too many particle systems.

Commit 280e7a on main fixes regressions in entity search (including Barnacles), allows three bone weights for procedural meshes, adds OnChangedFromMin/Max to math_counter and substantially changes Autorefresh error handling.

wrefgtzweve asked for more VPROF nodes in PhysFrame, often a top-5 cost. Done on 19 October (commit 280e7a).

wrefgtzweve suggested that a per-player max networking range could be done by checking distance in CServerGameEnts::CheckTransmit before IsInPVS.

In #developer-talk, Nak showed his LuaOS running DOOM at decent resolution with sound. A JIT-friendly self-patching block cache took it from 0.9M to 8M instructions per second; later a GTA beta loaded.

In #developer-talk, Nak released a NikNaks version with BSP physics parsing (not on the Workshop yet). Later he said BSP parsing is up to ~2x faster, BitBuffer 2–3x and mesh generation 3–5x.

In #developer-talk, Harry said he multithreaded SV_ComputeClientPacks, increasing performance by over 50% and running a solid 64 tick with 128 players. He is looking into going past 128 players; he expects the entity limit to be the main problem, since it can't be raised without client changes.

marchc1 asked for more functions like Entity:CopyBoneMatrix for hot paths. Grocel linked the :Left() request; artemking4 argued for an optional dst argument on all getters instead.

Vitroze opened a PR removing an unnecessary Think hook from the DOF post-processing effect, found with GProfiler. It was merged and shipped in commit 693484.