Wait 8 ms once for a pipeline met for the first time, and carry the compute variants in the pipeline list.

A draw that met a new pipeline waited 3 ms for its fast link, in every frame, which is just under what the link takes with the stage libraries in the driver's cache: in a first-visit session on a new install 66 of 87 such draws gave up at the end of the wait and their frame was held back. The wait is now 8 ms for each pipeline, counted from its first miss and shared by all its draws, and 16 ms for a frame, so a pipeline the driver has to compile costs one frame the wait and the frames after it nothing.

The pipeline list tool also takes the compute variants a game folder dispatched, which no pipeline record names and which were translated in the middle of play, and reads a list made before as one of its inputs.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
This commit is contained in:
chrg6
2026-11-12 08:25:42 -05:00
parent d31c3a1036
commit 9f7c8ae810
8 changed files with 89 additions and 22 deletions

View File

@@ -161,10 +161,14 @@ of them.
What this step leaves for play, and the shared pipeline list. The step translates every shader's default variant
and the shader list's variants and builds each one's stage library or compute pipeline. A whole graphics pipeline
is built ahead at the start of a game folder that has drawn it before (`pipeline-keys.bin`) or whose release
listed it (below); a pipeline met for the first time is linked from its stage libraries in 2 to 6 ms, except a mesh
listed it (below); a pipeline met for the first time is linked from its stage libraries in 2 to 8 ms, which its
first draw waits for (`pipe_link_draw_ms`, 8 ms from the pipeline's first miss; `runtime/vk_exec_draw.c`
`link_budget`), except a mesh
pipeline, which has no libraries and compiles whole (0.2 to 1.1 s, its draws left out of the picture meanwhile),
and many of them on a first visit add up to hitches. `tools/pipeline_list.py merge` folds the `pipeline-keys.bin`
and `shader-answers.bin` of game folders that were played into a pipeline list; `prepare_shaders.py` reads it from
and `shader-answers.bin` of game folders that were played, and lists made before, into a pipeline list, with the
compute variants those folders dispatched (no pipeline record names a compute shader, and a variant the shader list
lacks is translated on the frame that first dispatches it); `prepare_shaders.py` reads it from
`tools\data\pipeline_list.bin` (`data\pipeline_list.bin` in a release folder, or `--pipeline-list`), adds the
shader variants it names to the seed and writes its records next to the game as `seed-pipeline-keys.bin`, and
`--prepare-shaders` builds every one of them into the pipeline cache and the driver's (`runtime/vk_pipeline.c`

View File

@@ -18,12 +18,22 @@ for the launcher; a GPU with Vulkan 1.3 and `VK_EXT_external_memory_host` (curre
## Faster
- **A new install no longer builds its pipelines while you play.** A release now carries a list of the graphics
pipelines the game is known to draw, 8,715 of them in this one. The install builds all of them while it prepares the
pipelines the game is known to draw, 8,849 of them in this one. The install builds all of them while it prepares the
shaders, and every start builds them before play, as it already did for the pipelines your own install had drawn.
Before, a new install met each pipeline in play first: in Hyrule Field one ten-second stretch of a first visit had
1,769 draws wait for their pipelines, 461 ms in all, the longest frame 88 ms. In a first walk through Kakariko
Village on fresh caches no draw waited for a pipeline with the list, where 634 pipelines were missed without it, and
the first minute had 10 frames over 25 ms instead of 137. `pipe_seed = 0` in nx2navi.ini leaves the list out.
- **A pipeline the list does not have costs its frame less.** Such a pipeline is linked from the shader libraries the
install prepared, and its first draw waits for that. The wait was 3 ms for each draw, just under what the link
takes: in a first-visit session on a new install 66 of 87 such draws gave up at the end of it, and their frame was
then not shown, so the picture before it stayed for one frame more. Of the 67 pipelines that session met outside
the list, 61 were ready within 8 ms. The wait is now 8 ms, counted once for each pipeline from the first draw that
needed it, so a pipeline the driver has to compile costs one frame that much and the frames after it nothing
(`pipe_link_draw_ms`, and `pipe_link_ms` for all the pipelines of one frame, now 16).
- **Compute shader variants are prepared too.** The list now carries the 629 compute shader variants met in play. No
pipeline record names a compute shader, so a variant the shader list lacked was translated on the frame that first
used it, 30 and 38 ms in the two cases of that session.
- **The draw replay runs on several CPU cores by default** (`replay_workers = 3`; contributed by epicmeme546): a
frame's large groups of draws are recorded on three threads beside the one that submits them. On an RX 7900 XT in
Hyrule Field it gave 51.7 instead of 43.6 fps standing. `replay_workers = 0` records on one thread as before; try