build(release): the pipeline tour's list from a tour of main f637b42d, 10993 pipelines #957
No reviewers
Labels
No labels
area/accessibility
area/audio
area/input
area/install
area/menu
area/renderer
area/saves
area/tooling
area/upscaling
beanstalk: approved
beanstalk: failed
beanstalk: needs approval
beanstalk: running
beanstalk: tested
ci: approved
good first issue
gpu/amd
gpu/igpu
gpu/intel
gpu/nvidia
help wanted
kind
bug
kind
chore
kind
docs
kind
feature
kind
performance
kind
question
needs measurements
platform/linux
platform/macos
platform/steam-deck
platform/windows
preview
priority
critical
priority
high
priority
low
regression
risk
high
risk
low
risk
medium
status
blocked
status
confirmed
status
duplicate
status
invalid
status
needs-info
status
wontfix
ci: approved
No milestone
No project
No assignees
2 participants
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
MagicBeanDev/Longshot!957
Loading…
Reference in a new issue
No description provided.
Delete branch "jabbaxipil/Longshot:tour-list"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Type: build · Risk: low · Tested on: Linux (Ubuntu 26.04.1), RTX 4070 · Needs: none
Why
assets/pipeline_states/tour.bin.xzis still the list from the draft tour (Hookshot's PR 2048) on main of 2026-11-10. The pipeline tour that main has now (#111) has never published one. The pipeline README says the tour's list is replaced whole by each tour, made only bytour --publish. This is that: the whole tour on current main, published by the tour itself. The list then matches the build it ships with.What changed
The tour's list and its note, replaced by
python -I scripts/pipeline_tour.py tour --game <game folder> --out <tour dir> --name play --publish. No other file changes.assets/pipeline_states/tour.bin.xz: 10,993 pipelines (9,637 graphics, 790 mesh, 566 compute) of 5,332 programs, from 45 games over the four parts (opening, child overworld, dungeons, adult overworld), from a new save, in 54 minutes of runs. Against the list it replaces (11,227, of which 0 compute), it adds 4,509 pipelines and lacks 4,743.assets/pipeline_states/tour.md: as--publishwrote it, except the GPU line. The tour wrote "NVIDIA GeForce RTX 4070 (Vulkan, native)", because it takes the line from the bridge's renderer, which names no driver. I ran the samenx2/pipeline_states.py tourcommandpublish()builds (the 45 runs' lists in its order, the same build, settings and route) with "NVIDIA GeForce RTX 4070, driver 615.71.09 (Linux)", which the runs' logs show (vk: driver: NVIDIA 615.71.09). The list it wrote is byte-identical to the published one.nx2/pipeline_states.py merge) goes from 63,920 to 63,717 pipelines: 2,222 that no source had before, and 2,425 that only the old tour's list held. In this build of main the game finds more of the list, not fewer: 32,905 noted pipelines of 20,495 programs at the start, against 32,390 of 20,411, and 2,211 compute programs against 2,199 (Measurements). Most of what drops out is pipelines this build no longer asks for, as the README warns of lists from older builds.The tour was run on an ext4 drive, so with #898's change to its game copy applied locally. That change touches only how the game folder is copied, not the runs or the lists.
Measurements
Verdict: the build prepares 515 more pipelines (+1.6%) and keeps 12 more compute programs ready. The cost is memory: about 100 to 150 MB more in play, in runs that don't overlap. Start-up times are within the spread between runs.
Before: main @
a8a30c22After: this PR @
37b3bc85Both columns run the same game build: main at
a8a30c22(this PR's base), built from source withnx2/make_game.pyand tools.lock's Linux tools. The only difference islongshot_pipeline_states_shipped.bin, merged on main and on this branch. This PR changes only that file's sources, so swapping it is the comparison.Hardware: Linux (Ubuntu 26.04.1, kernel 7.0), RTX 4070 12 GB, driver 615.71.09, Ryzen 7 5800X3D, 64 GB, NVMe
Scene: a save in Zora's Domain (child), headless, the build's own default options (the Auto preset on this PC: 2560x1440 output, the game's upscaler, no DLSS (a build without it), frame generation off, HDR off, no frame rate cap);
RX_VK_LOAD=createMethod: a warm-up round of each, then two rounds interleaved (main, PR, main, PR). Each round: no notes, no Vulkan pipeline cache and no
longshot_prepared.bin(cache/longshot_pipelines.bin*,cache/longshot_vk_pipelines*), then the build's step (longshot --prepare-shaders --compile-shaders most), then a first start and a later start. Each start loads the save into play through the control bridge (A once every ~1.5 s from the title until the state is play and not loading) and is measured 30 s into play. Times are by wall clock, so launch to play includes up to ~1.5 s of the presses' spacing. The precompile and load lines are from the game's log. Memory is the process's private bytes (smaps_rollup Private_Clean + Private_Dirty) and its resident set (VmRSS), 30 s into play. The driver's shader cache was shared and warm in both arms (the warm-up rounds filled it).How to reproduce:
python -I nx2/pipeline_states.py merge --out <file>on main and on this branch, each copied over a game folder'slongshot_pipeline_states_shipped.bin, withcache/longshot_pipelines.bin*,cache/longshot_vk_pipelines*andcache/longshot_prepared.binremoved first.a8a30c2237b3bc85The same measurement on a build of
f637b42d(before the rebase ontoa8a30c22, whose translator stamp differs) gave the same counts (+515 noted pipelines, +12 compute programs) and memory in the same direction.Flags and settings
None.
Testing
Tested: Linux (Ubuntu 26.04.1), RTX 4070, driver 615.71.09.
python -I nx2/pipeline_states.py check: passes (merged: 63,717 pipelines, 2,211 compute programs, underCOMPUTE_MOST).python -I -B tests/run_all.pyat37b3bc85: 545 passed, 20 failed, 80 skipped. The same 20 fail on maina8a30c22here:test_appimage_player_build,test_bake_portability(nozstandardhere),test_bug_report,test_crash_dialog(clang 21 rejects-x none) andtest_crt_optionstest_env_on,test_hardware_report,test_lift_cleanup_error,test_linux_preflightandtest_linux_updatetest_make_game_cache,test_missing_nup,test_mods_size_table,test_options_budget_onceandtest_pipeline_tour(the stale snapshot that #898 fixes)test_readback_late,test_run_all_changed,test_vk_occlusion_queries,test_vk_pgoandtest_walk_keya8a30c22, and the same onf637b42d: every start reached play in Zora's Domain at 59 to 60 fps and quit with exit 0.Not tested:
HKSTATE2), which the README says prepares the pipelines of other GPUs, upscalers and output sizes too. Here, only an RTX 4070 at the measured settings.Must be tested: None.
Risks
What could break: the 2,425 pipelines that only the old tour's list held are no longer prepared before the title. In this build most of them aren't asked for (the noted count rises by 515), but a few may be, and would then compile the first time the game meets them, as before that list. The game holds about 100 to 150 MB more in play here (Measurements).
Turning it off: revert the commit.
Compatibility: a build from this PR prepares the new merged list at its first start, as any list change does. Saves, settings and the caches' formats don't change.
Checklist
Everyone
type(scope): description) of the status line's type.can't be run by hand is in Testing's Not testable by hand, with why.
in Risks. Not ticked yet: Memory cost is measured only on one RTX 4070 under Linux, and Testing says other GPUs, settings and Windows were not measured.
main.python -I -B tests/run_all.pypasses (any failure that also fails onmainis named).third_party/untouched (a new folder with aPROVENANCE.mdinstead).When it applies
Assisted-by:/Co-Authored-By:trailers.🤖 Generated with Claude Code
f637b42d, 10993 pipelinesMeasured again on a build of this PR's base, main
a8a30c22, after the rebase (its translator stamp differs from f637b42d's). The pipeline counts are the same (+515 noted pipelines, +12 compute programs). The first start's precompile is now 5.2 s on both sides, so the +0.9 s seen onf637b42ddid not repeat. Memory in play is about 100 to 150 MB higher with this list, in runs that don't overlap; the description's Measurements and Risks now say so.WIP: build(release): the pipeline tour's list from a tour of main f637b42d, 10993 pipelinesto build(release): the pipeline tour's list from a tour of main f637b42d, 10993 pipelines37b3bc858d249c371fe2Brought up to the ready-for-review standard (#1454): the checklist's Everyone list has two new boxes, added here from what this PR's description and comments show (please check they match what you know).
View command line instructions
Checkout
From your project repository, check out a new branch and test the changes.