mirror of
https://github.com/HayatoG/iss_pc.git
synced 2026-11-07 10:46:10 -10:00
e7b467e992
* Test: BC 1D on NVIDIA (TLG_BC_TINY=auto, promoted BC 1D answers dimension queries as 1D) Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm * Test script: occlusion and BP09 variants Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm * Launcher: Advanced screen with the GPU/runtime tuning parameters (switches, numbers, lists) 31 environment-only parameters (synchronisation, threads, readbacks, pipelines, image workarounds, diagnostics) with bilingual descriptions, Automatic/default handling, search, restore defaults and a copyable combination. Choices live in tlg.ini (adv_<name>) and reach the game as ISS_<name> only when changed; the Config: log block shows them. Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm * Test tools: one-round runner, monitor and analyzer for the parameter study Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm * GPU: bounded wait for a compiling pipeline (TLG_ASYNC_WAIT_MS) and a faster driver cache save A draw whose pipeline is still compiling in the background used to be skipped at once (missing ground or effects while the shader cache is cold). With TLG_ASYNC_WAIT_MS=N (default 0 = off) it first waits up to N ms for the compile, and the waits of one GPU frame add up to at most TLG_ASYNC_WAIT_FRAME_MS (default 8). The log gets an 'Async wait (total)' line every 5 s. The driver cache is saved every 60 s (TLG_DRIVER_CACHE_SAVE_S, was 180). Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm * GPU: bounded wait for a compiling pipeline (TLG_ASYNC_WAIT_MS) and a faster driver cache save A draw whose pipeline is still compiling in the background used to be skipped at once (missing ground or effects while the shader cache is cold). With TLG_ASYNC_WAIT_MS=N (default 0 = off) it first waits up to N ms for the compile, and the waits of one GPU frame add up to at most TLG_ASYNC_WAIT_FRAME_MS (default 8). The log gets an 'Async wait (total)' line every 5 s. The driver cache is saved every 60 s (TLG_DRIVER_CACHE_SAVE_S, was 180). Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm * Launcher Advanced: async wait and driver cache save parameters Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm * Tools: shader_index.py, an index of the OrbShdr shaders in the player's own dump Streaming scan of art/cache (.xpps, KCAP + OrbShdr), SQLite index hash -> file/offset/size/stage, lookup, verify against existing shader caches, stats. Reads the dump only; nothing of the game is copied or committed. Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm * GPU: region trace (TLG_REGION_TRACE=1), step A of the per-region pre-warm Writes region_trace.tsv in the GPU user folder: every .xpps package the game opens (O) and the first request of each graphics/compute pipeline (P), on one monotonic clock, so an offline pass can tell which pipelines each region needs before anything is built on top of it. Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm * GPU: the bounded wait also covers a shader still being translated A new shader is the usual reason for a skipped draw (its pipeline is only queued once the shader is translated), so GetProgram waits too, within the same per-draw and per-frame budget. Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm * GPU: region trace records whether a pipeline was ready; the async wait tally is a warning The log keeps warnings and above, so the 'Async wait (total)' line never showed. The region trace's P lines get a fifth column: r = the pipeline existed at its first request, m = it was missing. Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm * Docs and polish: Advanced screen, pre-warm notes, region trace parameter, test variant 8 Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm * Restore the third_party submodule entries (a worktree's symlinks had been committed) Co-Authored-By: Claude Sonnet 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_01BGrrT8Atu6mshYBUh1H1Mm --------- Co-authored-by: HayatoG <[email protected]> Co-authored-by: Claude Sonnet 5.5 <[email protected]>
2.1 KiB
2.1 KiB
Pré-aquecimento de shaders: o que sabemos e o que existe
Nota de trabalho (15/10/2026). O jogo traduz os shaders do PS4 (GCN) para Vulkan em tempo de execução; enquanto um shader ou pipeline novo compila em segundo plano, o desenho que dependia dele é pulado. É o chão, os prédios e os efeitos que somem por alguns quadros, principalmente com o cache de shaders frio.
O que o projeto faz hoje
- Cache em disco de pipelines e shaders, com a chave incluindo um hash dos fontes do recompilador, e um aquecimento em segundo plano no boot (quase 6000 pipelines em ~6 s).
- BP07 (desligado): desenhos importantes compilam na hora em vez de serem pulados.
- Espera curta (
ASYNC_WAIT_MS, desligada): um desenho cujo shader ou pipeline ainda compila espera até N ms (e no máximoASYNC_WAIT_FRAME_MSpor quadro) antes de ser pulado. O log mostra a linhaAsync wait (total)a cada 5 s. - Cache do driver salvo a cada 60 s (
DRIVER_CACHE_SAVE_S).
O que descobrimos
- O port do God of War III (mesma base) tem o mesmo mecanismo e admite o mesmo problema no README; não há pré-compilação nem ferramenta de aquecimento nele.
- Os shaders do jogo ficam sem compressão nos pacotes
.xppsdeart/cache(contêinerKCAP, cabeçalhoOrbShdr).tools/shader_index.pyindexa todos (10 374 únicos no jogo 1.07), e o hash do índice é o mesmo que o port usa. Tudo só na máquina do jogador; nada do jogo é copiado. - Uma pré-tradução completa sem o jogo rodando não é possível: cada shader vira várias versões conforme o estado do desenho, e esse estado só aparece em jogo.
- O jogo abre pacotes
bNN.xppspor região, e logo depois pedem-se rajadas de pipelines. O rastreio (TLG_REGION_TRACE=1) grava isso emregion_trace.tsvpara estudar o que cada região precisa.
Próximos passos (propostos)
- Medir com o rastreio e a espera ligados em áreas novas.
- Priorizar no aquecimento os pipelines vistos em cada região, no cache local do jogador.
- Só então, um manifesto de estados por região e tradução sem o jogo (trabalho grande e arriscado).