Released async shader can permanently inhibit frame elision

[hkf57]

Description

One Path of Building instance can enter a severe, focus-gated GPU/UI stutter state on Windows. The displayed frame remains pixel-identical, but a focused 3840x1600 window repeatedly redraws and saturates the ANGLE/D3D11 path.

This was reproduced live with PoB 2.66.2 and SimpleGraphic 2.5-3b1a346 on an RTX 3080 Ti.

Measured behavior

State	PoB GPU 3D	PoB CPU
affected instance, focused, maximized (3856x1568)	80.5% avg / 89.7% max	89.4% avg
affected instance, unfocused	no measurable PoB GPU engine	4.0% avg
affected instance, minimized	no measurable PoB GPU engine	negligible
affected instance, focused, windowed (1296x759)	1.26% avg	24.9% avg
affected instance, near-fullscreen windowed (3800x1500)	79.4% avg	99.4% avg
affected instance, focused at temporary 60 Hz	81.8% avg	83.6% avg
second PoB instance, same executable/runtime, focused/maximized	2.57% avg	33.1% avg

Two consecutive captures of the affected window were byte-for-byte identical. Changing PoB tabs did not reduce load while focused. The workload stops when the window loses focus and scales strongly with framebuffer size.

The NVIDIA per-application profile correctly targets the exact executable and disables G-SYNC for it. DWM remained light during the bad sample, and there were no display-driver resets or relevant application errors. Both PoB instances had the same RTSS hook loaded, so the overlay is not the primary discriminator.

Confirmed renderer state

SimpleGraphic's built-in Layers diagnostic showed:

• 122 out of 122 frames drawn
• Elide identical frames enabled
• Elision inhibited continuously true

This rules out a changing command-buffer hash as the immediate trigger. A first candidate that stabilized raw command padding also passed the Windows build but made no performance difference in a portable A/B test.

Root cause

r_shaderHnd_c::~r_shaderHnd_c() aborts an asynchronous texture load when its shader's reference count reaches zero. Cancellation can leave the texture in a non-DONE state (INIT or PENDING_UPLOAD), and PurgeShaders() retains non-DONE shader slots.

r_renderer_c::PumpShaders() currently scans every retained slot. It therefore treats an already released, unusable shader as an active incomplete texture and sets inhibitElision on every frame. EndFrame() then clears lastFrameHash, forcing the full draw queue to replay continuously while PoB is focused.

Proposed fix

Only active shader slots (refCount > 0) should inhibit frame elision while their textures load. Released shaders cannot be used for rendering and must not globally disable elision.

PR #109 contains the one-line correction. The Windows CI build passed, and portable A/B validation on the reproducing build reduced focused/maximized GPU 3D load from 80.7% average to 6.27% while clearing the renderer's inhibition flag (22 out of 1020 frames drawn).

Environment

• Windows build 26100
• NVIDIA RTX 3080 Ti, driver 610.47
• Active display: 3840x1600 at 144 Hz
• PoB 2.66.2
• SimpleGraphic 2.5-3b1a346
• Loaded path: SimpleGraphic -> GLFW 3.4 -> EGL/GLES (ANGLE) -> D3D11/DXGI

[hkf57]

Portable A/B result: the first candidate (66e4af1, zeroing command allocations) compiled successfully but did not improve the problem. With the affected copied build focused and maximized, the patched runtime measured 80.7% average GPU (83.8% max) and 188% process CPU. Focus was verified for the complete sample. This rules out uninitialized command padding as the live trigger; PR #109 remains draft while the elision-inhibited versus changing-command-stream cases are separated.

[hkf57]

Confirmed by portable A/B with corrected commit dad6573 (SimpleGraphic 2.5-dad6573).

• Windows 2022 CI: passed (run 30789884304)
• Test window was maximized and its PID was verified foreground for the complete counter sample
• GPU 3D fell from 80.7% average with the stock/first-candidate runtime to 6.27% average / 8.82% max (about 92% lower)
• Before: 122 out of 122 frames drawn, elision enabled, Elision inhibited true
• After: 22 out of 1020 frames drawn, elision enabled, Elision inhibited false

This directly confirms that a released non-DONE shader slot was holding the global inhibition flag, and that excluding zero-reference shader slots restores identical-frame elision. PR #109 is ready for review.