Skip to content

launcher: expose frame pacing, draw pass ordering and tile optimisations - #316

Closed
Xipzer wants to merge 3 commits into
seedds:mainfrom
Xipzer:launcher-settings
Closed

Xipzer wants to merge 3 commits into
seedds:mainfrom
Xipzer:launcher-settings

Conversation

@Xipzer

@Xipzer Xipzer commented Oct 26, 2026

Copy link
Copy Markdown
Contributor

Four switches the host already reads were environment only. They change how a game runs on a given device, so they belong in the settings screen beside the frame rate.

DM_FPS_HOLD, frame pacing

Decides whether a late frame may raise the hold. On hardware sitting near the deadline, halving to 30 is the wrong call: it also lowers the GPU clock to the longer hold, so the hold does not come back. That is the trap the fps_hold comment already describes for Age of Calamity.

Breath of the Wild 1.9.0, iPad Pro 13" M4, same save and areas:

fps_hold frames at 60 fps step ups / downs late frames
on 90.6 % 952 / 919 164,930
off 100.0 % 0 / 0 6,221

Player description with it on: swinging between 60 and 30 with no relation to what was happening on screen, and a pause into the inventory restoring 60 for 20 to 30 seconds. With it off: a locked 60, worst reading 58, seen once.

Not holding frames also made far fewer of them late.

DM_NVN_BARRIER, draw pass ordering

game removed 8885 of 40014 draw-pass entry barriers in the same game. Kept at full by default.

DM_NVN_TBDR, tile optimisations

Stays off by default. Listed so its effect can be measured on a game that reads textures through bindless handles, which is where the inference cannot see the readers.

DM_RENDER_SCALE cap raised to 2.0

It was capped at 1.0 because the game's per-frame texture pool size is unknown. It still is, and the usable value is per game: Breath of the Wild 1.9.0 runs at 1.35 and dies at 1.50, SIGSEGV at main+0xed7128 addr 0x20 when the allocation fails and the game takes the null back.

The loader now warns above 1.0 instead of refusing, since the working range is useful and only measurement finds it. Happy to drop this part if you would rather keep the refusal.

Note this supersamples into the game's own window, it does not widen it: the game still composites into the size it gave nvnWindowInitialize.

Xipz added 3 commits October 26, 2026 17:14
…, and raise the render scale cap

Four switches the host already reads were environment only. They change how a
game runs on a given device, so they belong in the settings screen beside the
frame rate.

DM_FPS_HOLD decides whether a late frame may raise the hold. On hardware sitting
near the deadline, halving to 30 is the wrong call and also lowers the GPU clock
to the longer hold, so the hold does not come back (the trap the fps_hold comment
already describes for Age of Calamity). Measured on Breath of the Wild 1.9.0,
iPad Pro 13" M4: on, 90.6 % of frames at 60 fps with 952 step ups; off, 100.0 %
with 0. The player reported a locked 60 with a worst reading of 58.

DM_NVN_BARRIER=game removed 8885 of 40014 draw-pass entry barriers in the same
game. DM_NVN_TBDR stays off by default; it is listed so its effect can be
measured on a game that reads textures through bindless handles.

DM_RENDER_SCALE now goes to 2.0. It was capped at 1.0 because the game's
per-frame texture pool size is unknown. It still is, and the value is per game:
Breath of the Wild 1.9.0 runs at 1.35 and dies at 1.50 (SIGSEGV at main+0xed7128,
addr 0x20, when the allocation fails and the game takes the null back). The
working range is useful and only measurement finds it, so the loader warns
instead of refusing.
…ndless game

Measured on Breath of the Wild 1.9.0, iPad Pro 13" M4: DM_NVN_TBDR=1 halves the
frame rate. Two costs stack and neither is offset. DM_NVN_UB_DEFER is off in TBDR
mode because TBDR reorders passes, so the uniform buffer deferral is given up;
and 38% of this game's shaders read textures through bindless handles, which the
slot-based inference cannot see, so every attachment must be assumed live and the
store elision never happens.
Bisected on Breath of the Wild 1.9.0, iPad Pro 13" M4: 1.40 loads the field
with and without MetalFX, 1.45 and above does not.
@v0lumehi

Copy link
Copy Markdown
Contributor

Nice measurements, especially the fps_hold numbers on the M4 iPad. Some notes after comparing with current main:

  1. DM_FPS_HOLD is already on the settings screen: it came in with c13ddcf as "Hold frames when late" in the Performance section. This branch predates the single-screen launcher (the rows here lack the section column), so a rebase would show it twice. Your data might be a stronger case for something else, though: making fps_hold off the default in Breath of the Wild's profile, if it holds on iPhones under the thermal cap too. That could be its own small PR.
  2. DM_NVN_BARRIER and DM_NVN_TBDR read to me like developer switches rather than player settings: a wrong choice shows up as broken rendering, and TBDR halves BotW's frame rate by your own measurement. Maybe they belong in the docs rather than on the player-facing screen? Happy to defer to @seedds on that.
  3. Render scale above 1.0: since BotW crashes at 1.45, a player slider up to 2.0 can crash a game, and the warning only lands in the log. A per-profile ceiling (e.g. a render_scale_max key, 1.35 for BotW, 1.0 elsewhere) would keep the useful range without the crash.

@Xipzer

Xipzer commented Oct 26, 2026

Copy link
Copy Markdown
Contributor Author

All three points taken, and the same stale-main problem as #315: this branch predates c13ddcf, so most of it is already there.

  1. DM_FPS_HOLD. Agreed, dropping it from this branch. And you are right that the stronger claim is the default rather than the row: on an M4 iPad, Breath of the Wild went from 27.8% of frames at 60 fps to 100.0% with fps_hold off, with zero step ups or downs over the window. I will open that as its own one-line PR against the profile, and hold it until someone has the same numbers on an iPhone under the thermal cap, since that is the case where holding might still be the right default.

  2. DM_NVN_BARRIER and DM_NVN_TBDR. Agreed, these are developer switches. TBDR halving the frame rate is not a setting a player should be able to reach: the cause is that it disables DM_NVN_UB_DEFER and the slot-based inference cannot see through bindless handles, which is 38% of the shaders in that game. Dropping both from the player screen; I will put the measurement in the docs instead.

  3. Render scale ceiling. Agreed, and render_scale_max as a profile key is better than what I did. A slider that can crash the game is not a setting. Reworking this branch to be exactly two things: the per-profile ceiling, and one new column in settings.tsv.

The column is needs, <VAR>=<value>: a row is drawn only while the setting it names has that value. It makes a sub-setting out of a row that is meaningless on its own, so Rumble strength disappears when Game rumble is off, and an upscaler's output size disappears when the upscaler is off. With the screen now at 15 rows and growing, hiding the rows that cannot do anything seemed worth the one column. Happy to drop it if you would rather the screen stayed flat.

@Xipzer

Xipzer commented Oct 27, 2026

Copy link
Copy Markdown
Contributor Author

Closing: this branch predates the single-screen launcher and c13ddcf, so most of its diff is already in main. Reopened as two small PRs on current main, following your review point by point: #354 for the per-profile render scale ceiling (with a measured 1.35 for Breath of the Wild, from #303's diagnostic) and #355 for the needs column. DM_FPS_HOLD, DM_NVN_BARRIER and DM_NVN_TBDR are dropped from the player screen as you suggested; the fps_hold default will be its own PR once there are iPhone numbers under the thermal cap.

@Xipzer Xipzer closed this Oct 27, 2026
ltrdpper pushed a commit to ltrdpper/BananaNative that referenced this pull request Nov 16, 2026
…nslator with the nested-phi handle fix (Kirby - Nintendo Switch 2 Edition)' (seedds#316) from Wick3dPhant0m/BananaNative:translator-kirby-s2e into android

Reviewed-on: https://git.eden-emu.dev/moceure/BananaNative/pulls/316
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants