This page preserves the earlier 16-patch comparison checkpoint against upstream projectM 4.2 development. Its images and source identities remain frozen. The current 37-patch catalog includes the later audit repairs, measured benefits and isolated real-preset close-ups. Older patch history remains in the evidence archive.
Images use a GPU-accelerated Android TV emulator with matched preset, textures, audio, time and seed. Zooms enlarge the same pixels without changing brightness. A middle column, where present, removes one patch to isolate its effect. Generated controls are labeled; a rejected render is an error panel, not a black screenshot.
Earlier images use our 13-patch snapshot 654815d8; 0014 uses the recorded 14-patch
snapshot 41ec3fc1. The image-comparison checkpoint stays frozen at main 120547f3, including0015; main’s later0016 is documented from its separate performance evidence. Earlier captures retain their original snapshot identities;Reference lines and0003,0011 and0015 now have locked15-patch comparisons.
The upstream renderer matches master e98fca85; capture adjustments and source
identities are recorded in the evidence record.
This consolidated patch contains several changes. The examples below cover line scaling, equation loading and resource handling separately.
In Geiss — 3D — Shockwaves, upstream's thin loops fade into the background at 4K. Our classic-line control also stays thin. Enabling reference-scaled lines keeps the loops broader and more visible; the close-up shows the stroke difference.
Fixed one-pixel lines shrink relative to the picture as resolution increases. Our optional quad renderer scales their width from a chosen reference size. This addresses the scaling requested in upstream #682. It uses miter joins and flat ends; rounded joins/caps are not implemented. Line evidence.
Upstream rejects 161.milk before producing a frame. Our library accepts its legacy split equation records and renders the preset. The reader retries those records using MilkDrop-compatible joining rules; genuinely rejected code remains reported. The left panel is the load error, not rendered output.
After 16 intentionally rejected fragment compilations, upstream leaves 16 vertex shaders alive; ours leaves none. The vertex shader succeeded before the fragment failed, so the patch deletes that unattached object before rethrowing the error. The normal preset pictures show rendering still works; the counters show the fix. Shader evidence.
The lower panels test first-read texture contents. A shared allocator deliberately fills new storage green: upstream leaves it there, while ours clears fresh and reused attachments to transparent black. This is a labeled diagnostic, not a naturally green artist preset. With pooling enabled, ours reuses one allocation instead of making two. Pooling retains storage and is off by default. Texture evidence.
Batching, caches, pass reductions and Native-trails replay also remain in 0001. Their deeper optimization/lifecycle comparisons are deferred to a follow-up. Patch source.
Flexi — madness portal becomes a plain red disc in upstream. Ours runs the
artist's yellow/blue composite with its fine circular detail. The shader declares
a local variable named sample; the old parser treats it as a reserved word,
rejects the shader and uses a fallback. The contextual-identifier fix allows it.
Single-patch control.
In ORB — Quicksand Lab, upstream uses the brighter cyan fallback. Our library
runs the authored five-tap filter, producing dimmer regions and stronger edges.
The shader fills five float4 elements from 20 scalar values; the translator must
group those values into vectors and initialize the array correctly.
Original-preset evidence.
This generated local-array control shows the same issue inside a function: a red fallback becomes the intended gradient. It tests local layout separately from Quicksand's global array.
The generated control writes q18 but still needs incoming q19=0.8.
Upstream loses the green component; ours preserves it. The writable copy now
starts from the whole incoming uniform bank instead of uninitialized storage.
In Martin — QBikal — Surface Turbulence, the shader scales time before star
helpers read it. Our initialized copy is shared with those helpers, restoring the
authored star/ray arrangement. The preset itself is unchanged.
Uniform evidence.
0002 also preserves float32 constants when emitting GLSL and supports the retained macro/postfix and implicit-input rules. Patch source.
Stahlregen — funky Blur uses a lone . in its zoom equation. Upstream rejects
it; a tolerant loader without 0003 omits that code and produces a different pattern.
Ours reads the dot as zero, matching the equation language's intended behavior.
The patch also gives each evaluator thread its own random state, preventing background preparation from advancing the foreground stream. That part is checked with thread/stream controls, rather than inferred from this picture. Evidence · Patch source. Locked15 image/failure receipts and complete evaluator control.
In widest swing, look at the large yellow/blue textured shape toward the top left. Its feedback sampling changes when a sampler left on texture unit zero interferes with the shape. Ours explicitly owns repeat/linear sampling for each main-textured fill, including replayed geometry. Single-patch comparison · Patch source.
In Flexi — a julia fractal for hexcollie embossed (Jelly), our image has different relief and highlights around the same fractal. The zoom makes the changed shading visible around the dark central shape.
The authored blur bounds become a narrow/inverted interval after clamping. A typo moved both repaired bounds downward, collapsing the interval and allowing division by zero. Ours moves the upper bound upward and keeps blur storage and decoding consistent; the artist's shader stays unchanged. Single-patch comparison · Patch source.
Hexcollie — This is where we begin stripped develops a different green center
and feedback structure. It uses negative zoom with exponent one. Sending that to
GLSL pow is undefined even though the intended result is simply the signed base.
Ours preserves that signed base and the reflected transform. Patch0015 carries this case into its broader CPU-power path.
Single-patch comparison ·
Patch source.
The Hexcollie — now entering the wormhole2 preset sets one initial wave mode,
then changes it through wave_mode=q8%7. The old renderer kept drawing the initial
mode. Ours follows the evaluated mode, changing the wave geometry and the feedback
it leaves behind. The matching zoom shows the changed structure.
Upstream/current frames ·
Patch source.
This generated control turns inversion on through an equation. Upstream keeps the orange shape on black; ours changes to the inverted blue shape on white. The renderer now consumes evaluated gamma, echo and filter flags each frame, including filters that start disabled. Custom composite shaders keep their own presentation rules. Patch source.
In Suksma — chemosynthetic nosferatu — rand tritex, compare the colored contours in the zoom. Small texture-upload byte differences grow through feedback. Our loader preserves the released SOIL2 premultiplication/rounding bytes, including opaque-channel rounding. This is a texture compatibility choice, not a general brightness boost. Byte and frame evidence · Patch source.
This is an enhancement for hosts switching custom packs. Both generated packs
contain aurora_ownership_core.png, but in separate texture directories: SOL's
image is orange and LUNA's is blue. At frame 20 the host changes directories before
loading LUNA. Upstream replaces SOL's emblem; ours keeps SOL's image until the
preset retires, including during a fade/reset.
Upstream uses the current global texture lookup. Ours retains the lookup that belonged to each preset when it loaded. Ordinary Cream of the Crop switching uses one shared directory and does not trigger this case. It is not a MilkDrop bug. Pack sequence and evidence · Patch source.
The EoS/Phat/PeterP/Sentinel/Aware witness uses rot=10000000. This labeled
coordinate-color control uses the same angle: upstream collapses the field to one
color, while ours preserves the rotation. GPU sine/cosine lost the transform on
the tested driver; ours computes both on the CPU after float conversion.
The original preset image is dark and upstream also rejects an equation, so it is kept as supporting evidence. The clear control above is generated, not the original artwork. Locked-source verification · Patch source.
In Amandio C — the green machine … btbam covers sepultura, pink/brown wedges become a green/yellow curved structure. Tiny custom shapes seed a gradient-driven feedback shader, so a small coverage error grows into a large pattern difference.
D3D9 and GLES sample different pixel-center grids. Ours translates each shape draw by half a destination pixel, preserving the authored position, radius and colors. The isolated comparison separates this correction from other changes in the full library. Patch source.
In DemonLD — Toxic water diffusion, compare the green/cyan contour rims and their neighboring colors in the zoom. An extra half-texel offset made GLES read between texels and blend colors that the shader expected to sample at their centers. Its nonlinear color operations amplify that unwanted averaging.
Ours removes the redundant offset. The single-patch comparison shows the contour change without other patch effects; the impulse control shows one bright texel staying intact instead of spreading across four. Patch source.
In Geiss — Hurricane, upstream and current-minus-0014 add a green cast despite
fShader=0. Ours respects the disabled tint, giving cooler, nearly neutral trails.
The patch also restores mode 1's alpha multiplier and open spiral, removing the
extra line that connected its ends.
The constant-color controls check disabled, half and full tint separately. The spiral close-up isolates the waveform change from tint. Current-source evidence · Patch source.
This generated control uses negative zoom with finite integer nested powers.
Upstream collapses the coordinate field; ours preserves its gradient. GLSL pow
is undefined for negative bases even when the CPU expression has a valid answer.
The patch evaluates that expression on the CPU after float conversion and reuses
its result when drawing the prepared mesh. Positive zoom is unchanged.
The original Great Tulip Majesty still enters fractional negative-power domains and its recorded frames remain unchanged. It must not be presented as fixed. Locked-source comparison and checks · Patch source.
Upstream rebuilt the warp and final-composite grids every frame because it never saved the viewport size used by its cache check. This patch saves the size and rebuilds/uploads static mesh data only when the viewport or grid changes.
On the AM6, Royal — Mashup (102) at1080p/30fps used11.8% less process CPU in two 40-second runs per side, with silence. GPU time and4K were not measured. Existing host controls check reused versus fresh mesh data; no on-device image comparison was run. This later main optimization is included without expanding the frozen 15-patch capture checkpoint. Performance evidence · Patch source.
Full-resolution images, inputs and verification retain repeated-frame checks, source/binary identities and failed attempts. The original PR input map provides the preset names and activation profiles used to guide remaining captures. The detailed working assessment keeps the component inventory and technical notes out of this review page.
The15 patch sections now have matched comparison evidence. Deeper optimization/lifecycle checks are deferred by agreement; final review and CI remain required before publication. The upstream capture admits GLES 3.0; our image workers disable the emulator's broken program-binary export. No original Windows/MilkDrop GPU screenshot was produced.
The candidate series retains released patches0001–0016 and audit repairs0017–0028 unchanged, then adds five independently revertible authored-fidelity fixes:
| Patch | Finding | Behavior |
|---|---|---|
| 0029 | I19 | Restore original line-wave sample budgets, capped by the reference-equivalent width, with a two-point floor. |
| 0030 | I22 | Submit authored discrete custom dots without interpolated midpoints; preserve a single-dot NaN sample input. |
| 0031 | I20 | Restore circle angular spacing and explicit closure before shared smoothing. |
| 0032 | I24 | Capture evaluated outline thickness per shape instance for authored and Native replay; preserve saved style outside the defined int32 truncation domain. |
| 0033 | I16 | Publish the actual previous completed warp field while vectors are hidden, preserving real fragment discard/output writes and per-preset ownership. |
The user accepts necessary authored work instead of the historical zero-added-cost gate for these five. I24 uses the small per-instance repair rather than experimental grouped geometry; I16 uses the actual warp producer rather than a synthetic UV-only shortcut. Earlier CPU warp trig, negative power, cached mesh uploads, prepared replay, float values, textures and Native/reference policies remain intact. No new API, dependency, asset or version change is introduced.
The candidate repairs18 audit IDs. Ten retained policies and five other deferred proposals remain as recorded in the immutable historical audit. Its original patch numbers, zero-cost decisions and captures are not relabelled as new acceptance.
Current qualification binds the baseline28 and candidate33 source/artifact identities. Source checks pass329 host,63 normal and sanitizer renderer controls,137 JVM and60 capture/helper tests plus24 subtests. Both ARM core-release/debug-APK builds pass. The focused Native run completed76 jobs across19 presets/fixtures;608 selected final-output RGB captures verify and repeat exactly within each role. This is source-derived GLES qualification, not Windows/D3D pixel equivalence, whole-corpus certification or measured physical-TV headroom. Active-canvas lifecycle passes four repeated jobs/120 boundary captures. Isolated combined cost verifies84 jobs/672 selected captures, with a largest overall added time of+.343ms and variable cycle deltas. Final PR review/CI remain open.
Patch0035 adds projectm_opengl_set_program_cache_enabled for the process-wide program binary cache from patch0001. Disabled, programs compile from source in their own context without the retrievable hint, export or glProgramBinary load; disabling waits for binary calls already under way on other threads; enabled again, the cache starts empty. The former Advanced › Shader binary cache setting was retired after it did not resolve the reported PowerVR GE9215 crashes. The app uses its normal enabled default; the core compatibility API remains. Not proposed upstream: the cache itself is TV-only.
I31 is a rendering-cost repair with unchanged checked images. A native Apple M4 Pro 3840×2160 comparison of the 34-patch engine against its exact 0019 ablation confirms three gamma draws become two at loaded gamma 2.0009999275. In the Standard profile, GPU elapsed time falls 2.307→1.847 ms (−19.9%) and completed-frame time 2.704→2.241 ms (−17.1%). All eight paired ABBA/BAAB blocks favor the patch; the gamma 2.0 inactive control has no resolved GPU/completed-frame effect. Three selected RGB frames match across roles and repeats within each of the Classic and Standard profiles.
This supports retaining 0019 on measured benefit, superseding the earlier inconclusive timing status. It establishes neither universal app FPS gains nor physical-TV performance; gain size depends on workload and hardware bottlenecks. The isolated proof retains 80 timed runs/28,800 measured frames, source/harness identities, controls, intervals and replay instructions.
A source-bound ProjectM 2.3.55 Fire TV report identifies a driver glBlitFramebuffer crash in feedback resizing. Patch0038 replaces PowerVR resize, Native-trails recombination and scaled-presentation blits with cached textured draws. It also isolates colour-zero routing, completes mixed-attachment resizing before copying, restores touched state and forgets resources safely after context loss. Other GPUs keep blits with explicit read-source detachment before texture retirement. Primary image formats already match; the separate motion-vector attachment keeps its data format.
The focused evidence binds the released ARM instruction bytes to the resize caller and records host GL and Android/GLES controls. Synthetic 512×288 RGBA copies match exactly, 4K scaling differs by at most one channel step, and nearest Native-trails copying is exact. The reporter confirmed on 2026-10-18 that the released fix resolves the affected Fire TV Stick 4K Max (2nd generation, AFTKRT, GE9215) crashes. It is included from ProjectM 2.3.56 and Milkbeat 1.0.13 and enabled automatically on all supported Imagination/PowerVR GPUs. See the device coverage and upgrade guide. No speedup is claimed. Patch source and comparison images.