docs(tested): native float add on an RX 9070 XT (RADV) in Kakariko and on Death Mountain Trail #1088
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
3 participants
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
MagicBeanDev/Longshot!1088
Loading…
Reference in a new issue
No description provided.
Delete branch "GrimVisage/Longshot:docs/tested-native-fadd-rdna4"
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: docs · Risk: low · Tested on: Linux (Fedora 45), Ryzen 5 5600X, RX 9070 XT (RADV 26.2.7) · Needs: none
Why
The game's shaders add floats into storage buffers. Where the GPU's own float add is trusted it is used (
VK_EXT_shader_atomic_float); elsewhere, AMD included, the add runs as a compare-exchange loop. https://github.com/NotRexed/Hookshot/pull/2306 asked for a run on AMD, in Kakariko Village and on Death Mountain Trail, before AMD's default changes. This PR records that run from an RX 9070 XT (RDNA 4) on RADV,RX_VK_NATIVE_FADD=1against the loop, so the AMD default has numbers to go on. It changes no default. No issue or tracking issue on this tracker covers it (searched:RX_VK_NATIVE_FADD, "float add",[Tracking], "AMD Linux", "RADV"); #816 is the same switch for Intel's Panther Lake.What changed
Four run files under
docs/tested/, nothing else. They compare the compare-exchange loop (the default here) againstRX_VK_NATIVE_FADD=1in one build of main1764fed2, on a Ryzen 5 5600X with an RX 9070 XT on Fedora 45.docs/tested/2026-11-15-native-float-add-kakariko-dmt-grimvisage-run01.mdtorun04.md, one run each, in the format ofdocs/tested/README.md: run01 and run04 the loop, run02 and run03 native, each with its four parts' numbers.The runs these files record
One build of main
1764fed2(longshot abc-main1764 (source f6bc6f5f)), built as a player's (the builder AppImage's--cli), copied into two game folders, one a side: the program cache's stamp holds the float add's choice, so one folder switched back and forth would prepare its programs again in the measured runs. Both folders were prepared and warmed before the series. The log saidvk: float atomic add on storage buffers: compare-exchange loopin every loop run andvk: float atomic add on storage buffers: native (VK_EXT_shader_atomic_float), RX_VK_NATIVE_FADD=1in every native run.Each run: fullscreen 2560x1440 (
HOS_FULLSCREEN=1), SDR, the game's own upscaler at Render resolution 100% (Quality preset Auto (this PC); no FSR 4 library in the test folders), frame generation off, the FPS counter off, mesh shaders Automatic (Compute), Threaded rendering Auto (on), uncapped (HOS_FPS_MAX=240),RX_VK_AUTO_TUNE=0, the frame-time log on (RX_VK_FRAME_LOG), Disable camera centering On, one kept options file (HOS_OPTIONS_FILE), a fresh copy of one child save. In play, the debug teleport to Kakariko Village, then to Death Mountain Trail; in each, 15 s to settle, the clock set to 12:00 with the Debug tab's Set time (Debug tools on for the test run only,HOS_OOT_CHEATS=1; the log'scheats: requested 12:00checked), then 30 s standing with no input and 30 s turning the camera (right stick at half). A warm-up run of each side, then four measured runs interleaved loop, native, native, loop. Every run's log shows one swapchain size, 2560x1440, from start to end.Measurements
Verdict: nothing to measure in the game for this PR: docs only. The files pass their format check, and the game-free suite gives the same result as on main (CONTRIBUTING.md, "Measurements":
docs).Hardware: Ryzen 5 5600X, 31 GB, Fedora 45 (kernel 7.3.3)
Scene: none (game-free checks)
Method:
python -I -B tests/test_tested_configurations.py, andpython -I -B tests/run_all.pyon main and on this PRHow to reproduce: the two commands above from the repository root
f7922126c6f9f281tests/test_tested_configurations.pytests/run_all.pyWhat the runs show
Baseline: main @
1764fed2with the compare-exchange loop (the default on AMD), against the same build withRX_VK_NATIVE_FADD=1: the switch is on main already, so its effect is the switch, not a main-against-PR build.Within the spread. In Kakariko Village the runs are steady (a side's two runs within about 1 fps) and the two sides match: fps, 1% low, frame times and GPU time overlap. On Death Mountain Trail the runs vary far more than the two sides differ, on both sides alike (standing: loop 46.0 and 60.4 fps, native 57.7 and 49.8; turning the camera: loop 53.6 and 58.2, native 57.3 and 50.2), so that scene can't show a difference this small; its medians are given but are not a result. No device lost in any run.
Medians of the two runs a side; fps, 1% low, median and p99 frame time from
RX_VK_FRAME_LOGover each part's presents; the replay's submit time and GPU time the medians of the log's status lines in each part.1764fed21764fed2A status line of each side, in Kakariko standing (run01 and run02):
Flags and settings
None: this PR adds or changes no switch, menu row, ini key or build flag (the runs use
RX_VK_NATIVE_FADD, unchanged).Testing
Tested: the four files with
tests/test_tested_configurations.py(25 tests OK) and the whole game-free suite withtests/run_all.py(the same 13 failures as on mainf7922126), on Fedora 45. The runs themselves are what the files record.Not tested: the native float add on other AMD cards (RDNA 2 and 3, the APUs), on AMD's Windows driver, and in other scenes: this PC only.
Must be tested: None (docs only; the runs make no claim about other PCs).
Risks
What could break: nothing in the game: four Markdown files under
docs/tested/.Turning it off: revert the commit.
Compatibility: Nothing changes.
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.
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.docs/tested/for each new measured run (format indocs/tested-configurations.md).🤖 Generated with Claude Code
WIP: docs(tested): native float add on an RX 9070 XT (RADV) in Kakariko and on Death Mountain Trailto docs(tested): native float add on an RX 9070 XT (RADV) in Kakariko and on Death Mountain Trailc6f9f281b040d55c99bbCommit message reworded: a link that didn't belong in it is gone, and the same link is gone from the description. Same files, same base (
c6f9f281is now40d55c99).Basic code-only review of
40d55c99bb; informational feedback, with final approval left to another reviewer.No obvious defect found in this documentation-only change. The four independent run records follow the tested-catalogue format and retain the hardware, driver, build identity, effective switch, scene, settings and run order. The submitted per-run numbers support the description's cautious conclusion: Kakariko is within the observed spread, while Death Mountain Trail is too variable to establish a benefit. This does not justify changing AMD's default or extrapolating to other cards/drivers.
I also checked the renderer's feature selection and program-cache stamp: the description correctly distinguishes the compare-exchange default from explicitly enabled native float add, and explains using separately warmed folders. No runtime code or default changes here. The adjacent mesh-shader measurements are a separate experiment, not a dependency.
The author reports 25 format checks passing and the same 13 suite failures on both sides. I reviewed the submitted evidence and source; I did not independently reproduce the measurements. No tests or builds run.
Brought 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.