docs(tested): robust buffer access off and on on an RTX 4090, plain, with DLSS Frame Generation and with the renderer's own #460

Open
Jiro113 wants to merge 1 commit from Jiro113/Longshot:docs/tested-robust-buffers-4090 into main
Contributor

Why

Longshot asks robust buffer access for the whole device by default (the Debug tab's Robust buffers row, robust_buffers, RX_VK_ROBUST), and nobody had timed what that costs. Asking it for some pipelines only (VK_EXT_pipeline_robustness, in Vulkan 1.4's core) was claimed in the team chat as a way to win that cost back, so the cost was measured first: 24 runs on an RTX 4090 resolve a cost of at most about 0.8 % of GPU time (plain, and in Kokiri Forest with DLSS Frame Generation), and the two frame generation series cannot exclude up to about 3 % in Hyrule Field. The switch was not written.

What changed

  • docs: 24 files under docs/tested/, one per run (2026-11-14-robust-buffers-native-jiro113-run01 to run10, -dlssg- run01 to run08, -framegen- run01 to run06); the frozen table is untouched and no assembled view is committed. No code changes: the per-pipeline switch was not written.

What the runs found. Each run starts the game once with RX_VK_ROBUST=0 or =1, stands still in Kokiri Forest and then in Hyrule Field (child) for 15 s and then 90 s measured, at 5120x1440. GPU time is the mean of the log's GPU time lines over the window (the log prints one decimal; a window holds about 70 lines); each cell is the median of the runs' windows, with their mean and count in parentheses. Lower is better.

Plain: Render resolution 100 % (the scene drawn at 5120x1440), no DLSS or FSR, no frame generation, V-Sync Auto (mailbox), 10 runs.

GPU time, ms a frame main @ 0caf0618, RX_VK_ROBUST=0 main @ 0caf0618, RX_VK_ROBUST=1 Δ medians (means)
Kokiri Forest 6.172 (6.168, n=5) 6.190 (6.216, n=4) +0.29 % (+0.77 %)
Hyrule Field 7.355 (7.354, n=5) 7.380 (7.398, n=6) +0.34 % (+0.60 %)

DLSS Quality (3414x960 to 5120x1440) and DLSS Frame Generation 2x, V-Sync Off, 8 runs.

GPU time, ms a frame main @ 5e072fc8, RX_VK_ROBUST=0 main @ 5e072fc8, RX_VK_ROBUST=1 Δ medians (means)
Kokiri Forest 9.361 (9.366, n=4) 9.421 (9.428, n=4) +0.64 % (+0.67 %)
Hyrule Field 10.593 (10.595, n=4) 10.816 (10.729, n=4) +2.11 % (+1.27 %)

The renderer's own frame generation 2x, Render resolution 100 %, no DLSS or FSR, V-Sync Off, 6 runs.

GPU time, ms a frame main @ 5e072fc8, RX_VK_ROBUST=0 main @ 5e072fc8, RX_VK_ROBUST=1 Δ medians (means)
Kokiri Forest 7.608 (7.641, n=3) 7.647 (7.646, n=3) +0.51 % (+0.07 %)
Hyrule Field 8.911 (8.976, n=3) 9.124 (9.073, n=3) +2.39 % (+1.09 %)

In each table the baseline is the RX_VK_ROBUST=0 column. A switch on one build, so the comparison is the switch off against on, not main against a change.

How far these go:

  • Plain: 0.3 to 0.8 % more GPU time with robustness on, at the edge of what the runs tell apart. The off windows span 6.122 to 6.199 ms and 7.304 to 7.399 ms by themselves (1.3 %). Neighbour runs, on against off: Kokiri +0.34 %, -0.08 %, +2.15 %, -0.06 %; Hyrule +0.29 %, -0.24 %, -0.23 %, +2.51 %, +1.19 %. The two over 2 % are both run 7, the series' slowest (its Hyrule p99 is 17.3 ms against about 13.4). The fps moved by 0.2 % or less either way (Kokiri 108.9 off and 108.7 on, Hyrule 93.1 and 93.2, medians).
  • DLSS Frame Generation: about +0.7 % in Kokiri Forest, the same sign in all four neighbour pairs (+1.24 %, +0.59 %, +0.63 %, +0.25 %). Hyrule Field's +2.1 % is not resolved: its pairs are +0.34 %, +5.41 %, -0.38 %, -0.14 %, and the +5.41 % is run 4, a disturbed window (other programs used 5.0 cores) on the wrong side of a step. GPU time rose about 4.5 % in both arms between runs 4 and 5 (Kokiri 9.12 to 9.23 ms, then 9.55 to 9.65), and three of the on arm's four Hyrule windows are after it or disturbed. Within the halves Hyrule is +0.3 % (without the disturbed window), then -0.3 %.
  • The renderer's own frame generation: not resolved at about 3 %. Neighbour pairs: Kokiri -3.24 %, +3.00 %, +0.51 %; Hyrule -0.39 %, +3.29 %, +0.43 %. Runs 1 to 3 were slower than runs 4 to 6 in both arms (Hyrule 9.12 to 9.16 ms, then 8.86 to 8.95), and the order off, on, on, off, on, off puts two on runs in the slow half.
  • What the runs resolve is at most about 0.8 % (plain, and Kokiri Forest with DLSS Frame Generation). In Hyrule Field the two frame generation series vary 2 to 3 % from run to run and cannot exclude a cost up to about 3 %. A per-pipeline switch could win back no more than the whole device's robustness costs.
  • With robustness off the game started and stood still 12 times with no device lost. That is not a test of what robust buffer access guards against.

Flags and settings

None: no flag is added or changed. The runs use the existing RX_VK_ROBUST (the options file's robust_buffers: 0 Auto (on), 1 Off, 2 On), and the default stays on.

Measurements

A docs change: measured by its check (CONTRIBUTING.md, "Measurements": docs, test output on main and with this PR).

main: 5e072fc8 (2026-11-14)
this PR: docs/tested-robust-buffers-4090 @ f78bfe72
Hardware: RTX 4090 24 GB, NVIDIA 621.24, Core i9-13900K, 64 GB, Windows 11 Home 26300
Scene: none (game-free checks)
Method: both checks run twice on main and three times on this branch, alternating, on the same PC, with llvm-mingw's compilers on the PATH; the table gives the counts that came back each time but one (below)
How to reproduce: python -I -B tests/test_tested_configurations.py, python -I -B tests/run_all.py, python -I scripts/tested_configurations.py

main @ 5e072fc8 this PR @ f78bfe72
tests/test_tested_configurations.py 25 tests, 1 failure 25 tests, the same 1 failure
rows in scripts/tested_configurations.py's view dated 2026 163 187 (the 24 new files)
tests/run_all.py 510 passed, 8 failed, 43 skipped (both runs) 510 passed, 8 failed, 43 skipped (two runs of three; the other: 509 and 9)
this PR:
FAIL: test_cli_ascii_pipe_emits_the_arrow_containing_source_as_utf8 (__main__.Fragments.test_cli_ascii_pipe_emits_the_arrow_containing_source_as_utf8)
Ran 25 tests in 0.349s
FAILED (failures=1)

510 passed, 8 failed, 43 skipped in 117 s (windows, 32 at once)
509 passed, 9 failed, 43 skipped in 115 s (windows, 32 at once)
510 passed, 8 failed, 43 skipped in 118 s (windows, 32 at once)

main:
Ran 25 tests in 0.665s
FAILED (failures=1)

510 passed, 8 failed, 43 skipped in 121 s (windows, 32 at once)
510 passed, 8 failed, 43 skipped in 123 s (windows, 32 at once)

The same eight fail on main on this Windows PC: test_bake_portability, test_bug_report, test_framegen_resources, test_lift_cache, test_module_ids, test_oot_state_tap, test_pad_layout and test_tested_configurations. The last one's failure is its stand-in for a redirected pipe, which writes \r\n on Windows where the test expects \n; it builds its own files and reads none of this PR's. The test that reads the repository's fragments, test_repository_fragments_validate_and_source_suffix_is_preserved, passes with the 24 new files.

The ninth failure, in one run of three on this branch, is tests/test_control.py: its 27 cases passed and its test program then exited with 3221225477 (an access violation). It passed three times of three run by itself and in the next whole run; a docs file is not something it reads.

The runs' method. One PC, one afternoon (12:15 to 15:40 local), a 2560x720 window on a 5120x1440 screen at 120 Hz, SDR; Quality preset Auto (this PC), anisotropic 16x, Reflex Off (no NVIDIA Reflex: on line in any log); HOS_FPS_MAX=1000, HOS_PRESENT_TRACE on. The scene's size is read from the textures each run made: 5120x1440 in the plain and the renderer's own series, 3414x960 with DLSS.

  • Builds: development runtimes of two full builds of main (scripts/dev_packs.py rt, llvm-mingw, no pack over them): 0caf0618 made without the NVIDIA DLSS SDK for the plain series, 5e072fc8 made with it for the two others. The same runtime in both arms of a series.
  • A run: the same save copied fresh, scripts/longshot_ctl.py launch --run, Continue, longshot_ctl.py teleport KokiriForest_Child --arrive 60, 15 s to settle, 90 s measured with nothing sent to the game, then the same in HyruleField_Child.
  • The arms: RX_VK_ROBUST=0 and =1, interleaved (plain: on, off, on, off, off, on, on, off, off, on; DLSS Frame Generation: on, off, off, on, off, on, on, off; the renderer's own: off, on, on, off, on, off). The log says vk: robust buffer access on in every on run and in no off run.
  • The series' switches: DLSS Frame Generation RX_VK_DLSS=1 RX_VK_FRAMEGEN_MULT=2 RX_VK_IMMEDIATE=1 (DLSS Quality 3414x960 to 5120x1440; the log's of the last 600 presents ... made by DLSS-G lines say 599 to 600 in every window); the renderer's own RX_VK_DLSS=0 RX_VK_DLSS_FG=0 RX_VK_FRAMEGEN_MULT=2 RX_VK_IMMEDIATE=1 (no NGX line in the log; its frame interpolation lines count 31,066 to 33,379 frames shown on time over a run's two windows and 0 to 2 late). V-Sync Off is not the default: V-Sync Auto holds DLSS Frame Generation to the screen's rate.
  • A window is called disturbed when other programs used more than 4.0 cores during it, a limit set before the frame generation series ran: run 4's Hyrule Field (5.0) and run 8's Kokiri Forest (6.3) of the DLSS series, both kept in the tables above and marked in their files. No pipeline was made during a window.
  • One teleport did not take: the plain series' run 6 measured Hyrule Field twice, so its robustness-on arm has 4 Kokiri windows and 6 Hyrule ones.

From the middle of a Hyrule Field window, one run of each arm (the log's counters; with either maker the frame interpolation line counts the frames shown):

plain, RX_VK_ROBUST=0 (run 2)
vk: 24000 presents, 93.9 fps; a frame: 1491 draws (0 skipped), 159 dispatches, 149 passes, 130 barriers, 6 submits, 0.0 uploads (0.0 MB), 80 ub updates (56 overlays), GPU-bound wait 0.0 ms; 24744 pipelines (240331 ms), 0 programs (0 failed), 1684 textures (4281 MB)
vk: GPU time 7.3 ms a frame (the longest submission 2.0 ms); 0.0 textures made a frame
plain, RX_VK_ROBUST=1 (run 3)
vk: 24240 presents, 92.4 fps; a frame: 1497 draws (0 skipped), 159 dispatches, 149 passes, 130 barriers, 6 submits, 0.0 uploads (0.0 MB), 80 ub updates (56 overlays), GPU-bound wait 0.0 ms; 24744 pipelines (239851 ms), 0 programs (0 failed), 1696 textures (4264 MB)
vk: GPU time 7.4 ms a frame (the longest submission 2.0 ms); 0.0 textures made a frame

DLSS Frame Generation, RX_VK_ROBUST=0 (run 2)
vk: 19920 presents, 79.8 fps; a frame: 1502 draws (0 skipped), 159 dispatches, 150 passes, 130 barriers, 6 submits, 0.0 uploads (0.0 MB), 80 ub updates (56 overlays), GPU-bound wait 0.0 ms; 7272 pipelines (14896 ms), 0 programs (0 failed), 2077 textures (4918 MB)
vk: GPU time 10.5 ms a frame (the longest submission 3.8 ms); 0.0 textures made a frame
vk: frame interpolation: 810 frames shown on time, 0 late (over 2 ms); shown apart: <7 ms 807, 7-9 1, 9-12 2, 12-17 0, 17-25 0, over 25 0
vk: DLSS Frame Gen (NGX): of the last 600 presents 600 made by DLSS-G (0 told to reset, 600 with the camera still, 0 NGX said not to show); the renderer's own made the rest: 0 without new inputs (no DLSS/DLAA this frame), 0 after a gap, 0 asked more than 1 frames, 0 refused
DLSS Frame Generation, RX_VK_ROBUST=1 (run 1)
vk: 19560 presents, 82.8 fps; a frame: 1498 draws (0 skipped), 159 dispatches, 149 passes, 130 barriers, 6 submits, 0.0 uploads (0.0 MB), 80 ub updates (56 overlays), GPU-bound wait 0.0 ms; 7272 pipelines (14494 ms), 0 programs (0 failed), 2060 textures (4867 MB)
vk: GPU time 10.1 ms a frame (the longest submission 3.0 ms); 0.0 textures made a frame
vk: frame interpolation: 830 frames shown on time, 0 late (over 2 ms); shown apart: <7 ms 826, 7-9 3, 9-12 1, 12-17 0, 17-25 0, over 25 0
vk: DLSS Frame Gen (NGX): of the last 600 presents 600 made by DLSS-G (0 told to reset, 600 with the camera still, 0 NGX said not to show); the renderer's own made the rest: 0 without new inputs (no DLSS/DLAA this frame), 0 after a gap, 0 asked more than 1 frames, 0 refused

the renderer's own frame generation, RX_VK_ROBUST=0 (run 4)
vk: 20880 presents, 83.0 fps; a frame: 1497 draws (0 skipped), 159 dispatches, 149 passes, 130 barriers, 6 submits, 0.0 uploads (0.0 MB), 80 ub updates (56 overlays), GPU-bound wait 0.0 ms; 7271 pipelines (14812 ms), 0 programs (0 failed), 1671 textures (4234 MB)
vk: GPU time 8.9 ms a frame (the longest submission 2.8 ms); 0.0 textures made a frame
vk: frame interpolation: 818 frames shown on time, 0 late (over 2 ms); shown apart: <7 ms 806, 7-9 9, 9-12 3, 12-17 0, 17-25 0, over 25 0
the renderer's own frame generation, RX_VK_ROBUST=1 (run 5)
vk: 20880 presents, 85.5 fps; a frame: 1499 draws (0 skipped), 159 dispatches, 150 passes, 130 barriers, 6 submits, 0.0 uploads (0.0 MB), 80 ub updates (56 overlays), GPU-bound wait 0.0 ms; 7270 pipelines (14010 ms), 0 programs (0 failed), 1697 textures (4346 MB)
vk: GPU time 9.0 ms a frame (the longest submission 2.7 ms); 0.0 textures made a frame
vk: frame interpolation: 860 frames shown on time, 0 late (over 2 ms); shown apart: <7 ms 855, 7-9 3, 9-12 1, 12-17 1, 17-25 0, over 25 0

Testing

Tested: the game-free checks above, on Windows 11 with an RTX 4090 (NVIDIA 621.24). The 24 runs are in the game: Kokiri Forest and Hyrule Field (child), standing still, on the same PC; each file's numbers are written from its run's log and present trace by a script, not typed.

Not tested: the website's catalogue built from these files. What the runs leave open: any GPU that isn't NVIDIA's, a card with less than 24 GB, Linux, FSR and the other upscalers without frame generation (with RX_VK_DLSS=0 the Auto preset rendered at 100 %), a moving camera or a fight, and what robust buffer access guards against when it is off.

Needs other testers: None for this PR, which only records runs. The cost on AMD and Intel drivers is unmeasured; no request is open for it.

Screenshots

Nothing on screen.

Risk and rollback

Docs only: nothing the game or the build reads. Reverting the commit removes the 24 files.

Reviewer notes

  • Runs measured and left out, with no file: the first two of the plain series (the first starts of that runtime: pipelines were still being made during the windows), two warm-up runs, and two 45 s trials before the frame generation series.
  • The two frame generation series vary more from run to run (2 to 3 %) than the plain one (about 1 %). Other programs were running throughout: 2.5 to 3.9 cores in the frame generation series' undisturbed windows, 2.2 to 3.4 in the plain runs 7 to 10 (not recorded in runs 1 to 6).
  • Seen and not followed up: in the DLSS Frame Generation series the four robustness-on starts had 7 or 8 frames over 100 ms between the menu and the first window (1.7 to 2.4 s in all), the four off starts 5 or 6 (1.1 to 1.3 s). The plain series (3 to 6 on, 5 or 6 off) and the renderer's own frame generation (4 to 7 on, 5 or 6 off) show no such difference.
  • The claim this answers: "Claimed: per-pipeline buffer robustness" in the team chat on 2026-11-14. It is released with this PR; nothing was written for it.

Issues

None.

Checklist

For everyone

  • One topic.
  • The title is a Conventional Commit (type(scope): description).
  • Rebased on current main.
  • python -I -B tests/run_all.py run and passing (any failure that also fails on main is named above).
  • Tested in the game, or what wasn't is named under Testing.
  • Testing filled in: what was tested, what wasn't, and what needs other testers (with test requests).
  • Measured: main and this PR at named commits, on the same PC, scene and settings, at least 2 interleaved runs each, the median reported.
  • No game files, dumps, build output or generated files.
  • third_party/ untouched (a new folder with a PROVENANCE.md instead).
  • The license is OK: GPL-3.0-or-later, nothing incompatible and nothing decompiled.

If applicable

  • AI assistance recorded with Assisted-by: / Co-Authored-By: trailers.
  • A file in docs/tested/ for each new measured run (format in docs/tested-configurations.md; no appended row or assembled view).
  • Docs updated.

🤖 Generated with Claude Code

## Why Longshot asks robust buffer access for the whole device by default (the Debug tab's Robust buffers row, `robust_buffers`, `RX_VK_ROBUST`), and nobody had timed what that costs. Asking it for some pipelines only (`VK_EXT_pipeline_robustness`, in Vulkan 1.4's core) was claimed in the team chat as a way to win that cost back, so the cost was measured first: 24 runs on an RTX 4090 resolve a cost of at most about 0.8 % of GPU time (plain, and in Kokiri Forest with DLSS Frame Generation), and the two frame generation series cannot exclude up to about 3 % in Hyrule Field. The switch was not written. ## What changed - docs: 24 files under `docs/tested/`, one per run (`2026-11-14-robust-buffers-native-jiro113-run01` to `run10`, `-dlssg-` `run01` to `run08`, `-framegen-` `run01` to `run06`); the frozen table is untouched and no assembled view is committed. No code changes: the per-pipeline switch was not written. **What the runs found.** Each run starts the game once with `RX_VK_ROBUST=0` or `=1`, stands still in Kokiri Forest and then in Hyrule Field (child) for 15 s and then 90 s measured, at 5120x1440. GPU time is the mean of the log's `GPU time` lines over the window (the log prints one decimal; a window holds about 70 lines); each cell is the median of the runs' windows, with their mean and count in parentheses. Lower is better. Plain: Render resolution 100 % (the scene drawn at 5120x1440), no DLSS or FSR, no frame generation, V-Sync Auto (mailbox), 10 runs. | GPU time, ms a frame | main @ 0caf0618, `RX_VK_ROBUST=0` | main @ 0caf0618, `RX_VK_ROBUST=1` | Δ medians (means) | |---|---|---|---| | Kokiri Forest | 6.172 (6.168, n=5) | 6.190 (6.216, n=4) | +0.29 % (+0.77 %) | | Hyrule Field | 7.355 (7.354, n=5) | 7.380 (7.398, n=6) | +0.34 % (+0.60 %) | DLSS Quality (3414x960 to 5120x1440) and DLSS Frame Generation 2x, V-Sync Off, 8 runs. | GPU time, ms a frame | main @ 5e072fc8, `RX_VK_ROBUST=0` | main @ 5e072fc8, `RX_VK_ROBUST=1` | Δ medians (means) | |---|---|---|---| | Kokiri Forest | 9.361 (9.366, n=4) | 9.421 (9.428, n=4) | +0.64 % (+0.67 %) | | Hyrule Field | 10.593 (10.595, n=4) | 10.816 (10.729, n=4) | +2.11 % (+1.27 %) | The renderer's own frame generation 2x, Render resolution 100 %, no DLSS or FSR, V-Sync Off, 6 runs. | GPU time, ms a frame | main @ 5e072fc8, `RX_VK_ROBUST=0` | main @ 5e072fc8, `RX_VK_ROBUST=1` | Δ medians (means) | |---|---|---|---| | Kokiri Forest | 7.608 (7.641, n=3) | 7.647 (7.646, n=3) | +0.51 % (+0.07 %) | | Hyrule Field | 8.911 (8.976, n=3) | 9.124 (9.073, n=3) | +2.39 % (+1.09 %) | In each table the baseline is the `RX_VK_ROBUST=0` column. A switch on one build, so the comparison is the switch off against on, not main against a change. How far these go: - **Plain: 0.3 to 0.8 % more GPU time with robustness on, at the edge of what the runs tell apart.** The off windows span 6.122 to 6.199 ms and 7.304 to 7.399 ms by themselves (1.3 %). Neighbour runs, on against off: Kokiri +0.34 %, -0.08 %, +2.15 %, -0.06 %; Hyrule +0.29 %, -0.24 %, -0.23 %, +2.51 %, +1.19 %. The two over 2 % are both run 7, the series' slowest (its Hyrule p99 is 17.3 ms against about 13.4). The fps moved by 0.2 % or less either way (Kokiri 108.9 off and 108.7 on, Hyrule 93.1 and 93.2, medians). - **DLSS Frame Generation: about +0.7 % in Kokiri Forest, the same sign in all four neighbour pairs** (+1.24 %, +0.59 %, +0.63 %, +0.25 %). Hyrule Field's +2.1 % is not resolved: its pairs are +0.34 %, +5.41 %, -0.38 %, -0.14 %, and the +5.41 % is run 4, a disturbed window (other programs used 5.0 cores) on the wrong side of a step. GPU time rose about 4.5 % in both arms between runs 4 and 5 (Kokiri 9.12 to 9.23 ms, then 9.55 to 9.65), and three of the on arm's four Hyrule windows are after it or disturbed. Within the halves Hyrule is +0.3 % (without the disturbed window), then -0.3 %. - **The renderer's own frame generation: not resolved at about 3 %.** Neighbour pairs: Kokiri -3.24 %, +3.00 %, +0.51 %; Hyrule -0.39 %, +3.29 %, +0.43 %. Runs 1 to 3 were slower than runs 4 to 6 in both arms (Hyrule 9.12 to 9.16 ms, then 8.86 to 8.95), and the order off, on, on, off, on, off puts two on runs in the slow half. - **What the runs resolve is at most about 0.8 %** (plain, and Kokiri Forest with DLSS Frame Generation). In Hyrule Field the two frame generation series vary 2 to 3 % from run to run and cannot exclude a cost up to about 3 %. A per-pipeline switch could win back no more than the whole device's robustness costs. - With robustness off the game started and stood still 12 times with no device lost. That is not a test of what robust buffer access guards against. ## Flags and settings None: no flag is added or changed. The runs use the existing `RX_VK_ROBUST` (the options file's `robust_buffers`: 0 Auto (on), 1 Off, 2 On), and the default stays on. ## Measurements A docs change: measured by its check (CONTRIBUTING.md, "Measurements": `docs`, test output on main and with this PR). main: 5e072fc8 (2026-11-14) this PR: docs/tested-robust-buffers-4090 @ f78bfe72 Hardware: RTX 4090 24 GB, NVIDIA 621.24, Core i9-13900K, 64 GB, Windows 11 Home 26300 Scene: none (game-free checks) Method: both checks run twice on main and three times on this branch, alternating, on the same PC, with llvm-mingw's compilers on the PATH; the table gives the counts that came back each time but one (below) How to reproduce: `python -I -B tests/test_tested_configurations.py`, `python -I -B tests/run_all.py`, `python -I scripts/tested_configurations.py` | | main @ 5e072fc8 | this PR @ f78bfe72 | |---|---|---| | `tests/test_tested_configurations.py` | 25 tests, 1 failure | 25 tests, the same 1 failure | | rows in `scripts/tested_configurations.py`'s view dated 2026 | 163 | 187 (the 24 new files) | | `tests/run_all.py` | 510 passed, 8 failed, 43 skipped (both runs) | 510 passed, 8 failed, 43 skipped (two runs of three; the other: 509 and 9) | ``` this PR: FAIL: test_cli_ascii_pipe_emits_the_arrow_containing_source_as_utf8 (__main__.Fragments.test_cli_ascii_pipe_emits_the_arrow_containing_source_as_utf8) Ran 25 tests in 0.349s FAILED (failures=1) 510 passed, 8 failed, 43 skipped in 117 s (windows, 32 at once) 509 passed, 9 failed, 43 skipped in 115 s (windows, 32 at once) 510 passed, 8 failed, 43 skipped in 118 s (windows, 32 at once) main: Ran 25 tests in 0.665s FAILED (failures=1) 510 passed, 8 failed, 43 skipped in 121 s (windows, 32 at once) 510 passed, 8 failed, 43 skipped in 123 s (windows, 32 at once) ``` The same eight fail on main on this Windows PC: `test_bake_portability`, `test_bug_report`, `test_framegen_resources`, `test_lift_cache`, `test_module_ids`, `test_oot_state_tap`, `test_pad_layout` and `test_tested_configurations`. The last one's failure is its stand-in for a redirected pipe, which writes `\r\n` on Windows where the test expects `\n`; it builds its own files and reads none of this PR's. The test that reads the repository's fragments, `test_repository_fragments_validate_and_source_suffix_is_preserved`, passes with the 24 new files. The ninth failure, in one run of three on this branch, is `tests/test_control.py`: its 27 cases passed and its test program then exited with 3221225477 (an access violation). It passed three times of three run by itself and in the next whole run; a docs file is not something it reads. **The runs' method.** One PC, one afternoon (12:15 to 15:40 local), a 2560x720 window on a 5120x1440 screen at 120 Hz, SDR; Quality preset Auto (this PC), anisotropic 16x, Reflex Off (no `NVIDIA Reflex: on` line in any log); `HOS_FPS_MAX=1000`, `HOS_PRESENT_TRACE` on. The scene's size is read from the textures each run made: 5120x1440 in the plain and the renderer's own series, 3414x960 with DLSS. - Builds: development runtimes of two full builds of main (`scripts/dev_packs.py rt`, llvm-mingw, no pack over them): 0caf0618 made without the NVIDIA DLSS SDK for the plain series, 5e072fc8 made with it for the two others. The same runtime in both arms of a series. - A run: the same save copied fresh, `scripts/longshot_ctl.py launch --run`, Continue, `longshot_ctl.py teleport KokiriForest_Child --arrive 60`, 15 s to settle, 90 s measured with nothing sent to the game, then the same in `HyruleField_Child`. - The arms: `RX_VK_ROBUST=0` and `=1`, interleaved (plain: on, off, on, off, off, on, on, off, off, on; DLSS Frame Generation: on, off, off, on, off, on, on, off; the renderer's own: off, on, on, off, on, off). The log says `vk: robust buffer access on` in every on run and in no off run. - The series' switches: DLSS Frame Generation `RX_VK_DLSS=1 RX_VK_FRAMEGEN_MULT=2 RX_VK_IMMEDIATE=1` (DLSS Quality 3414x960 to 5120x1440; the log's `of the last 600 presents ... made by DLSS-G` lines say 599 to 600 in every window); the renderer's own `RX_VK_DLSS=0 RX_VK_DLSS_FG=0 RX_VK_FRAMEGEN_MULT=2 RX_VK_IMMEDIATE=1` (no NGX line in the log; its `frame interpolation` lines count 31,066 to 33,379 frames shown on time over a run's two windows and 0 to 2 late). V-Sync Off is not the default: V-Sync Auto holds DLSS Frame Generation to the screen's rate. - A window is called disturbed when other programs used more than 4.0 cores during it, a limit set before the frame generation series ran: run 4's Hyrule Field (5.0) and run 8's Kokiri Forest (6.3) of the DLSS series, both kept in the tables above and marked in their files. No pipeline was made during a window. - One teleport did not take: the plain series' run 6 measured Hyrule Field twice, so its robustness-on arm has 4 Kokiri windows and 6 Hyrule ones. From the middle of a Hyrule Field window, one run of each arm (the log's counters; with either maker the `frame interpolation` line counts the frames shown): ``` plain, RX_VK_ROBUST=0 (run 2) vk: 24000 presents, 93.9 fps; a frame: 1491 draws (0 skipped), 159 dispatches, 149 passes, 130 barriers, 6 submits, 0.0 uploads (0.0 MB), 80 ub updates (56 overlays), GPU-bound wait 0.0 ms; 24744 pipelines (240331 ms), 0 programs (0 failed), 1684 textures (4281 MB) vk: GPU time 7.3 ms a frame (the longest submission 2.0 ms); 0.0 textures made a frame plain, RX_VK_ROBUST=1 (run 3) vk: 24240 presents, 92.4 fps; a frame: 1497 draws (0 skipped), 159 dispatches, 149 passes, 130 barriers, 6 submits, 0.0 uploads (0.0 MB), 80 ub updates (56 overlays), GPU-bound wait 0.0 ms; 24744 pipelines (239851 ms), 0 programs (0 failed), 1696 textures (4264 MB) vk: GPU time 7.4 ms a frame (the longest submission 2.0 ms); 0.0 textures made a frame DLSS Frame Generation, RX_VK_ROBUST=0 (run 2) vk: 19920 presents, 79.8 fps; a frame: 1502 draws (0 skipped), 159 dispatches, 150 passes, 130 barriers, 6 submits, 0.0 uploads (0.0 MB), 80 ub updates (56 overlays), GPU-bound wait 0.0 ms; 7272 pipelines (14896 ms), 0 programs (0 failed), 2077 textures (4918 MB) vk: GPU time 10.5 ms a frame (the longest submission 3.8 ms); 0.0 textures made a frame vk: frame interpolation: 810 frames shown on time, 0 late (over 2 ms); shown apart: <7 ms 807, 7-9 1, 9-12 2, 12-17 0, 17-25 0, over 25 0 vk: DLSS Frame Gen (NGX): of the last 600 presents 600 made by DLSS-G (0 told to reset, 600 with the camera still, 0 NGX said not to show); the renderer's own made the rest: 0 without new inputs (no DLSS/DLAA this frame), 0 after a gap, 0 asked more than 1 frames, 0 refused DLSS Frame Generation, RX_VK_ROBUST=1 (run 1) vk: 19560 presents, 82.8 fps; a frame: 1498 draws (0 skipped), 159 dispatches, 149 passes, 130 barriers, 6 submits, 0.0 uploads (0.0 MB), 80 ub updates (56 overlays), GPU-bound wait 0.0 ms; 7272 pipelines (14494 ms), 0 programs (0 failed), 2060 textures (4867 MB) vk: GPU time 10.1 ms a frame (the longest submission 3.0 ms); 0.0 textures made a frame vk: frame interpolation: 830 frames shown on time, 0 late (over 2 ms); shown apart: <7 ms 826, 7-9 3, 9-12 1, 12-17 0, 17-25 0, over 25 0 vk: DLSS Frame Gen (NGX): of the last 600 presents 600 made by DLSS-G (0 told to reset, 600 with the camera still, 0 NGX said not to show); the renderer's own made the rest: 0 without new inputs (no DLSS/DLAA this frame), 0 after a gap, 0 asked more than 1 frames, 0 refused the renderer's own frame generation, RX_VK_ROBUST=0 (run 4) vk: 20880 presents, 83.0 fps; a frame: 1497 draws (0 skipped), 159 dispatches, 149 passes, 130 barriers, 6 submits, 0.0 uploads (0.0 MB), 80 ub updates (56 overlays), GPU-bound wait 0.0 ms; 7271 pipelines (14812 ms), 0 programs (0 failed), 1671 textures (4234 MB) vk: GPU time 8.9 ms a frame (the longest submission 2.8 ms); 0.0 textures made a frame vk: frame interpolation: 818 frames shown on time, 0 late (over 2 ms); shown apart: <7 ms 806, 7-9 9, 9-12 3, 12-17 0, 17-25 0, over 25 0 the renderer's own frame generation, RX_VK_ROBUST=1 (run 5) vk: 20880 presents, 85.5 fps; a frame: 1499 draws (0 skipped), 159 dispatches, 150 passes, 130 barriers, 6 submits, 0.0 uploads (0.0 MB), 80 ub updates (56 overlays), GPU-bound wait 0.0 ms; 7270 pipelines (14010 ms), 0 programs (0 failed), 1697 textures (4346 MB) vk: GPU time 9.0 ms a frame (the longest submission 2.7 ms); 0.0 textures made a frame vk: frame interpolation: 860 frames shown on time, 0 late (over 2 ms); shown apart: <7 ms 855, 7-9 3, 9-12 1, 12-17 1, 17-25 0, over 25 0 ``` ## Testing **Tested:** the game-free checks above, on Windows 11 with an RTX 4090 (NVIDIA 621.24). The 24 runs are in the game: Kokiri Forest and Hyrule Field (child), standing still, on the same PC; each file's numbers are written from its run's log and present trace by a script, not typed. **Not tested:** the website's catalogue built from these files. What the runs leave open: any GPU that isn't NVIDIA's, a card with less than 24 GB, Linux, FSR and the other upscalers without frame generation (with `RX_VK_DLSS=0` the Auto preset rendered at 100 %), a moving camera or a fight, and what robust buffer access guards against when it is off. **Needs other testers:** None for this PR, which only records runs. The cost on AMD and Intel drivers is unmeasured; no request is open for it. ## Screenshots Nothing on screen. ## Risk and rollback Docs only: nothing the game or the build reads. Reverting the commit removes the 24 files. ## Reviewer notes - Runs measured and left out, with no file: the first two of the plain series (the first starts of that runtime: pipelines were still being made during the windows), two warm-up runs, and two 45 s trials before the frame generation series. - The two frame generation series vary more from run to run (2 to 3 %) than the plain one (about 1 %). Other programs were running throughout: 2.5 to 3.9 cores in the frame generation series' undisturbed windows, 2.2 to 3.4 in the plain runs 7 to 10 (not recorded in runs 1 to 6). - Seen and not followed up: in the DLSS Frame Generation series the four robustness-on starts had 7 or 8 frames over 100 ms between the menu and the first window (1.7 to 2.4 s in all), the four off starts 5 or 6 (1.1 to 1.3 s). The plain series (3 to 6 on, 5 or 6 off) and the renderer's own frame generation (4 to 7 on, 5 or 6 off) show no such difference. - The claim this answers: "Claimed: per-pipeline buffer robustness" in the team chat on 2026-11-14. It is released with this PR; nothing was written for it. ## Issues None. ## Checklist ### For everyone - [x] One topic. - [x] The title is a Conventional Commit (`type(scope): description`). - [x] Rebased on current `main`. - [x] `python -I -B tests/run_all.py` run and passing (any failure that also fails on `main` is named above). - [x] Tested in the game, or what wasn't is named under Testing. - [x] Testing filled in: what was tested, what wasn't, and what needs other testers (with test requests). - [x] Measured: main and this PR at named commits, on the same PC, scene and settings, at least 2 interleaved runs each, the median reported. - [x] No game files, dumps, build output or generated files. - [x] `third_party/` untouched (a new folder with a `PROVENANCE.md` instead). - [x] The license is OK: GPL-3.0-or-later, nothing incompatible and nothing decompiled. ### If applicable - [x] AI assistance recorded with `Assisted-by:` / `Co-Authored-By:` trailers. - [x] A file in `docs/tested/` for each new measured run (format in `docs/tested-configurations.md`; no appended row or assembled view). - [x] Docs updated. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
docs(tested): robust buffer access off and on on an RTX 4090, plain, with DLSS Frame Generation and with the renderer's own
Some checks are pending
CI / Conventional Commit title (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Blocked by required conditions
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Blocked by required conditions
CI / Workflow lint (actionlint) (pull_request) Blocked by required conditions
Third-party parts / The same as upstream (pull_request) Blocked by required conditions
Third-party parts / As their notes say (pull_request) Blocked by required conditions
Website / Website on Cloudflare (pull_request) Blocked by required conditions
PR description / Description follows the template (pull_request_target) Successful in 6s
Website / Website on Cloudflare (pull_request_target) Has been skipped
f78bfe721e
24 runs of main on one PC on 2026-11-14 (12:15 to 15:40 local), standing
still in Kokiri Forest and Hyrule Field at 5120x1440, RX_VK_ROBUST=0 and
=1 interleaved: 10 plain (main 0caf0618), 8 with DLSS Quality and DLSS
Frame Generation 2x and 6 with the renderer's own frame generation 2x
(main 5e072fc8).

Asking robust buffer access for the whole device costs at most about
0.8 % of GPU time where the runs resolve it: the medians are 0.3 % apart
in the plain series (the means 0.6 to 0.8 %) and about 0.7 % in Kokiri
Forest with DLSS Frame Generation. The two frame generation series vary
2 to 3 % from run to run in Hyrule Field, with the arms on both sides of
each other, and cannot exclude a cost up to about 3 % there.

One file per run; the frozen table is untouched and no assembled view is
committed.

Co-Authored-By: Claude Fable 5.2 <[email protected]>
Jiro113 changed title from WIP: docs(tested): robust buffer access off and on on an RTX 4090, plain, with DLSS Frame Generation and with the renderer's own to docs(tested): robust buffer access off and on on an RTX 4090, plain, with DLSS Frame Generation and with the renderer's own 2026-11-14 15:11:57 +00:00
Author
Contributor

Added to the description: a few log lines (counters only) from the middle of a Hyrule Field window, one run of each arm in each series, so the numbers in the tables can be read against the lines they come from. Nothing else changed; the branch is still at f78bfe72.

Added to the description: a few log lines (counters only) from the middle of a Hyrule Field window, one run of each arm in each series, so the numbers in the tables can be read against the lines they come from. Nothing else changed; the branch is still at f78bfe72.
Contributor

Taking an independent review of the reported run records, pair arithmetic and bounded claims at f78bfe721e. I will validate the fragments and methodology; I am not repeating the RTX4090 sessions. Matrix broadcast is queued for the next permitted hourly slot.

Taking an independent review of the reported run records, pair arithmetic and bounded claims at f78bfe721e18d930580ccf31bd461ac94a80e89c. I will validate the fragments and methodology; I am not repeating the RTX4090 sessions. Matrix broadcast is queued for the next permitted hourly slot.
monsterlag approved these changes 2026-11-14 16:53:45 +00:00
monsterlag left a comment

Independent docs/data review at f78bfe721e: no blocking findings.

All24 new fragments validate with the existing repository check (1 passed, no skips); the frozen catalogue and code are untouched. Recomputed the six arm counts, means, medians and neighbour-pair deltas from the committed three-decimal window means, matching the reported precision. The two Hyrule windows in native run6 explain its arm count and reproduce the run-level -0.23% pair. Methods distinguish window means from run medians and native rendering from NGX/renderer frame generation; disturbances, drift, teleport failure and exclusions are disclosed.

Nonblocking wording note: the0.8% /3% ranges describe observed central differences and spread, rather than statistical upper confidence bounds. Keeping that distinction explicit avoids implying a guaranteed maximum cost.

I did not independently repeat the GPU sessions or inspect their private raw logs, build the website, run the full suite, or validate other hardware/drivers. This approval covers the committed records, arithmetic and bounded methodology; it is not new GPU evidence or a robustness-off safety result.

Independent docs/data review at f78bfe721e18d930580ccf31bd461ac94a80e89c: no blocking findings. All24 new fragments validate with the existing repository check (1 passed, no skips); the frozen catalogue and code are untouched. Recomputed the six arm counts, means, medians and neighbour-pair deltas from the committed three-decimal window means, matching the reported precision. The two Hyrule windows in native run6 explain its arm count and reproduce the run-level -0.23% pair. Methods distinguish window means from run medians and native rendering from NGX/renderer frame generation; disturbances, drift, teleport failure and exclusions are disclosed. Nonblocking wording note: the0.8% /3% ranges describe observed central differences and spread, rather than statistical upper confidence bounds. Keeping that distinction explicit avoids implying a guaranteed maximum cost. I did not independently repeat the GPU sessions or inspect their private raw logs, build the website, run the full suite, or validate other hardware/drivers. This approval covers the committed records, arithmetic and bounded methodology; it is not new GPU evidence or a robustness-off safety result.
Some checks are pending
CI / Conventional Commit title (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request) Blocked by required conditions
CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request) Blocked by required conditions
CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request) Blocked by required conditions
CI / Workflow lint (actionlint) (pull_request) Blocked by required conditions
Third-party parts / The same as upstream (pull_request) Blocked by required conditions
Third-party parts / As their notes say (pull_request) Blocked by required conditions
Website / Website on Cloudflare (pull_request) Blocked by required conditions
PR description / Description follows the template (pull_request_target) Successful in 6s
Website / Website on Cloudflare (pull_request_target) Has been skipped
This pull request doesn't have enough approvals yet. 0 of 1 approvals granted.
Some workflows are waiting to be reviewed.
This branch is out-of-date with the base branch
You are not authorized to merge this pull request.
View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.
git fetch -u 'docs/tested-robust-buffers-4090:Jiro113-docs/tested-robust-buffers-4090'
git switch 'Jiro113-docs/tested-robust-buffers-4090'
Sign in to join this conversation.
No reviewers
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Longshot!460
No description provided.