Skip to content

Xendroid+: Gears of War 3 options, 21 to 30 fps at the fountain (#99) - #139

Merged
phforner0 merged 1 commit into
mainfrom
perf/game-quirks-batch2
Nov 18, 2026
Merged

phforner0 merged 1 commit into
mainfrom
perf/game-quirks-batch2

Conversation

@phforner0

Copy link
Copy Markdown
Owner

Gears of War 3 (0x4D5308AB) volta ao seu teto de 30 fps no primeiro checkpoint jogável (fonte do prólogo, ~2520 draws/frame), caindo de 44,9 para 28,2 ms/frame de GPU (−37,1%) num POCO F7 (Adreno 825, Turnip V38, e4003ba66).

Depois de 2026-10-05, o arredondamento para zero em software nos vertex shaders (#99) tinha somado ~13,6 ms/frame aos passes de cena e sombras projetadas da Unreal Engine 3, derrubando essa cena de 30 para ~21,3 fps. As quatro opções promovidas recuperam os 30 fps sem perda visual ou geométrica nas regiões medidas.

Configuração (0x4D5308AB, janelas de 40 s com timestamps por passe) fps (relógio) GPU ms/frame Passe 1280x2048 Passe 1120x2352
Base main (e4003ba66, início / fim da série) 21,21 / 21,98 45,20 / 44,55 29,17 / 29,91 ms 3,25 / 3,38 ms
+ alu_round_toward_zero = false (isolada) 27,82 31,25 18,76 ms 1,33 ms
+ spirv_vs_relaxed_math = 1 (isolada) 23,68 40,80 25,53 ms 3,53 ms
+ spirv_texture_implicit_lod = true (isolada) 22,02 43,46 28,97 ms 3,33 ms
+ vulkan_texture_load_to_image = true (intercalada) 21,79 44,54 29,17 ms 3,25 ms
Combinadas 29,33 28,24 15,33 ms 1,01 ms

O que muda

  1. game_quirks.cc (0x4D5308AB — Gears of War 3): adiciona alu_round_toward_zero = false, spirv_vs_relaxed_math = 1, spirv_texture_implicit_lod = true e vulkan_texture_load_to_image = true junto às duas quirks existentes do título (vulkan_depth_4x_as_1x e vulkan_samples_as_pixels_simple_ps).
  2. performance-tests/gears-of-war-3-results.md: relatório completo da rodada A/B no aparelho, decomposição por passe (VkPassTime / VkMiscTime / VkMemexport), tabela de integridade visual por região (dLuma, MAE, RMSE, energia de borda) e justificativa técnica das opções rejeitadas.

Opções testadas e mantidas fora das quirks

  • spirv_structured_forward_jumps = true: 44,90 ms (+0,02 ms, neutra).
  • vulkan_memexport_barriers_by_range = true: +0,54 ms, 0,0 barreiras elididas (read_after_export = 0,0).
  • vulkan_samples_as_pixels_coarse + vulkan_fragment_shading_rate: −0,05 ms, 6298 tiles inalterados (as sombras 4x já são cobertas por vulkan_depth_4x_as_1x + vulkan_samples_as_pixels_simple_ps).
  • vulkan_resolve_clear_in_guest_pass = true: 46,04 ms (+0,80 ms).
  • render_target_7e3_as_r11g11b10 = true (43,92 ms, −0,96 ms) e occlusion_query = 'fake' (43,15 ms, −1,73 ms): descartadas por serem com perda/inexatas (flare extra no topo do céu com fake) e desnecessárias, pois o conjunto seguro/verificado já alcança o limite de 30 fps (28,24 ms mesmo com timestamps por passe ligados).

Limites e não testado

  • Campanha longa, veículos e multiplayer sem o RTZ em software: se surgir geometria piscando ou esticando em outro capítulo, alu_round_toward_zero = false é a primeira quirk a suspeitar.
  • Medições feitas sobre e4003ba66 antes do merge do PR [GPU/Vulkan] Guard Vulkan 1.0-1.2 extensions, port Edge rendering fixes, and optimize software RTZ #135.
  • A fila dos demais títulos (Red Dead Redemption, Naruto, Forza Horizon e Crysis 3) foi pausada quando a bateria fora da tomada atingiu o piso de segurança (19%) ao término do controle final g3-b8-base-end.

Gears of War 3 returns to its 30 fps cap in its opening checkpoint on a POCO F7 (Adreno 825), cutting GPU frame time by 37% (44.9 to 28.2 ms).

  • Per-game vertex math, implicit texture LOD, and direct texture load settings recover the speed lost to software round-toward-zero emulation (Performance regression on version 0.2.1 #99) with the same image in the tested scene.
  • The first launch after updating compiles the game's shaders again, so some objects may briefly pop in.

⚠️ Still open

  • Tested in the first playable checkpoint; please report any flickering or stretched geometry in later chapters.

Gears of War 3 volta ao seu limite de 30 fps no primeiro checkpoint jogável num POCO F7 (Adreno 825), reduzindo o tempo de GPU em 37% (de 44,9 para 28,2 ms por frame).

  • Ajustes por jogo na matemática de vertex shaders, no LOD implícito de texturas e na carga direta de texturas recuperam a velocidade perdida com o arredondamento para zero em software (Performance regression on version 0.2.1 #99), mantendo a mesma imagem na cena medida.
  • A primeira abertura depois de atualizar compila de novo os shaders do jogo, então alguns objetos podem levar alguns instantes para aparecer.

⚠️ Ainda em aberto

  • Testado no primeiro checkpoint jogável; avise se aparecer geometria piscando ou esticando em capítulos seguintes.

Software round-toward-zero in vertex shaders pushed Gears of War 3's
first playable checkpoint (~2520 draws, 65 passes, 6298 tiles a frame)
from its 30 fps cap down to 21.3-22.0 fps (44.6-45.2 ms of GPU a frame
with pass timestamps on a POCO F7, Adreno 825, Turnip V38).

Enable alu_round_toward_zero=false (-13.6 ms), spirv_vs_relaxed_math=1
(-4.1 ms), spirv_texture_implicit_lod=true (-1.4 ms) and
vulkan_texture_load_to_image=true (-0.68 ms interleaved) for 0x4D5308AB:
together 28.24 ms of GPU a frame (-37.1%, 29.3-30 fps even with pass
timestamps enabled) with matching sky, horizon, cobblestone ground,
fountain and HUD. Record the full A/B breakdown and rejected candidates
in performance-tests/gears-of-war-3-results.md.
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 18, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-18T07:01:02.526184Z 0f0d9d2 PR opened
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0f0d9d25c9

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

// 1.01 ms), 29.3-30 fps even with pass timestamps on, the same sky,
// horizon, cobblestone ground and fountain. Memexport barriers by range,
// structured forward jumps and coarse 4x-as-1x were neutral here.
{0x4D5308AB, "alu_round_toward_zero", false,

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Preserve native RTZ on capable Vulkan devices

When the host advertises shaderRoundingModeRTZFloat32, this setting disables native RTZ as well as the software emulation: RoundsTowardZero() checks this cvar before the hardware capability, so affected vertex operations use round-to-nearest instead of Xenos rounding and can produce different geometry. The report only measured the software-emulation cost on a POCO F7; avoid disabling the exact hardware path on capable devices.

Useful? React with 👍 / 👎.

"no SM3 zero-multiply emulation in vertex shaders"},
{0x4D5308AB, "spirv_texture_implicit_lod", true,
"2D texture fetches with the host's LOD"},
{0x4D5308AB, "vulkan_texture_load_to_image", true,

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Leave texture-load aliases to the device gate

This per-game true bypasses the startup auto-detection, which enables storage aliases only on Turnip Adreno 750+; on older Adreno or other drivers, the texture-cache code will still create eligible textures with storage usage, which the cvar documentation warns can reduce compression and slow sampling. The measurements cover only a POCO F7, so avoid forcing this path on devices outside that tested range.

Useful? React with 👍 / 👎.

@phforner0
phforner0 merged commit 64615f8 into main Nov 18, 2026
5 checks passed
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