feat(fps): the render size, upscaler and AA in the FPS counter's footer #1657
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!1657
Loading…
Reference in a new issue
No description provided.
Delete branch "ballotstomper/Longshot:feat/fps-chart-resolution"
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?
Summary
The FPS counter's footer named the CPU, GPU and driver, but not how the frame was made. It now also shows the resolution the
3D was drawn at (and the picture's), and what upscaled or anti-aliased it, as the last frame drew them:
The lines report what drew, not what the menu asks for. Each path that draws the gameplay upscale publishes a record after it
has recorded the picture: the built-in network, its TAAU stand-in (FXAA) and its typed and blit fallbacks, DLSS, DLAA (before
the bloom too), FSR 3.1 on and off the chain, FSR 4, XeSS, and Off's stretch. So the game's dynamic resolution (Auto) shows its
live size, an engine that failed shows as
Built-in (DLSS not running), and a menu or loading screen showsResolution: not in gameplay. The wording, for every setting:Upscaling: Built-in/AA: Built-inUpscaling: Built-in fallback, FXAA/…, no AA/Upscaling: linear blit, no AAUpscaling: DLSS Quality,Upscaling: FSR 3.1 83%,Upscaling: XeSS Balanced,… (Auto)when set to AutoAA: DLAA,AA: FSR 3.1,AA: FSR 4,AA: XeSSUpscaling: off (stretched)/Upscaling: off, no AAUpscaling: Built-in (DLSS not running)Line 1 is always
Resolution WxH to WxH (N%), orResolution WxH (100%)at full size. The stop name comes from the share theframe drew at. The words stay English on purpose (screenshots and bug reports), as a comment in
upscale_route.hsays.Also: a Render lines row on the FPS counter details page turns the two lines off (the CPU and GPU lines row keeps its key and
meaning); "Shaders built (4 s)" and "Draws left out (4 s)" lose the "(4 s)" that ran into their values, and now go through
translation.
Read first:
gpu/vk/upscale_route.h(the record, the publish and read calls,footer_lines), then where each path publishes(
route_publishandroute_network_drawninvk_draw.cpp,route_drawninvk_temporal.cpp, the end offsr3_upscale),and
fps_render_linesinvk_menu.cpp. The upscalers' own state (camera, output size, gates) is main's, untouched.Testing
Tested: Linux, RX 6750 XT, RADV Mesa 26.2.7, Ryzen 5 5600X, at
8fcaae2c, the FPS counter at Dev, 2560x1440: Built-in atAuto (
Resolution 1920x1080 to 2560x1440 (75%),Upscaling: Built-in fallback, FXAA: this card runs the 75% network as theTAAU stand-in), FSR 3.1 at 83% and at Quality, Off at 67% and at 100%, Off at Auto (its size follows the dynamic resolution),
and a menu (
Resolution: not in gameplay,Upscaling: Built-in (chosen)). Screenshots, main and this head: #1657 (comment)Checks:
tests/test_fps_stats.pyhas 20 new game-free cases for the footer's text (every route, Auto, stops, a queued DLSSframe while switching to DLAA, the 40-character width).
test_menu_ini_keys,test_upscale_fallback,test_taau_env_lookup,test_psychov35_off,test_dlss_pre_bloompass.run_all.py --changed origin/main: 216 pass, 34 skipped, 19 fail, and each ofthose 19 fails on main @
6e1521a7too: test_bug_report_defaults, test_crt_options, test_damage_debug_menu, test_dev_ab,test_dev_packs, test_device_lost_exit, test_frame_overlap, test_linux_pressure, test_menu_module_slots, test_menu_modules,
test_menu_sections, test_mesh_conv, test_pipeline_play_cache, test_pipelines_outlive_cache_format, test_prepare_cache_save,
test_program_cache_one_variant, test_settings_table, test_sprint, test_upscaler_rows.
Still to test:
Upscaling: DLSS Quality/… (Auto)with the size following the dynamic resolution, andAA: DLAAat(100%)(@Beanstalk nvidia, the FPS counter'sfooter with DLSS and DLAA)
Upscaling: off (stretched)(@Beanstalk steamdeck,the FPS counter's footer)
Upscaling: FSR 4 Quality/Upscaling: XeSS Balanced(@Beanstalk windows, the FPS counter's footer)
Measurements
Display only, nothing to measure: with the footer's render lines hidden (the default: the counter is off), each drawing path does
one relaxed atomic load and returns; while they show, one locked copy per gameplay upscale and a read 4 times a second.
Risks
What could break: a path that draws the gameplay upscale without publishing would leave the footer at
Resolution: not in gameplaywhile playing (the counter's display only). Turning it off: the Render lines row. After updating: a newrender_lines=line inlongshot_options.ini(on by default); nothing else changes.Checklist
Ready for review
main; the title is a Conventional Commit.mise run testpasses (any failure that also fails onmainis named). (The checks reading the changed files; the 19failures are main's own, named in Testing.)
main, or Measurements says why there is nothing to measure.third_party/untouched; nothing GPL-incompatible.When it applies
tests/that reads changed source is updated, not weakened.Still to test.)
mainand here, for a change that shows on screen.Assisted-by:/Co-Authored-By:trailers. (Stated here instead: Assisted-by Claude Opus5.5 (Anthropic), through Claude Code.)
Ready to merge
main, measured again or a comment says why the numbers still hold.Tested in game: RX 6750 XT, RADV Mesa 26.2.7, Linux: FPS counter at Dev, FSR 3.1 at 83% and at native size show the render size and upscaler lines; Shaders built / Draws left out without (4 s)
Screenshots added: before (main) and after (this PR), the FPS counter at Dev.
mooki referenced this pull request2026-11-19 02:34:02 +00:00
Thanks, the footer is a useful place for this, and the "(4 s)" fix is clean in all 14 languages. One thing blocks it: with the default Upscaler, the render line is wrong in play.
Blocking
With Game's own, the default, the render line says
Render -all through play (gpu/vk/vk_temporal.cpp:1661and:1679)note_drawnis only called insidetemporal_upscale. Nothing calls that with Game's own:temporal_stand_in(gpu/vk/vk_draw.cpp) returns before it unlesstemporal_wanted() || temporal_off() || temporal_fg_inputs().temporal_rejected_networkhas the same gate at its call site.temporal_note_scene_drawhas it too, so there is no camera either.So with Upscaler Game's own and any Render resolution, which is most players,
temporal_render_sizealways returns false. The footer then readsRender -/Upscaler the game's network, AA the game's network. That is the screenshot case the PR is for.Off has the same hole at the fixed shares (67%, 58%, 50%, 33%). Their networks aren't
knownkeys, andtemporal_upscalereturns intostretchatif (mode == ur::GAME)before the!knownbranch that holds the secondnote_drawn. So only Off at 75% and 100% (the 4/3 and full-size keys) get a size. FSR 3.1 withRX_VK_FSR3_CHAIN=0is the same:fsr3_stand_indraws it andtemporal_upscalenever runs.The game runs you did were with FSR 3.1 on the chain, the one path that does reach
temporal_upscale.Fix: record the size where the gameplay upscale is found, whatever the Upscaler, before those gates. For example, a size-only note in
temporal_stand_inforgameplay_network(p.key), and for the shape-matched fixed shares, the same scene-size checkstretchmakes. Then check it in the game with Game's own at Auto and at 67%, and with Off at 50%.Suggestions
cpu_and_gpu_lines=line (gpu/vk/vk_menu.cpp:1723):save()doesn't treat this key as its own, so the old line goes intoothersand stays in the options file for good. Someone who had turned the lines off also gets them back.menu_initalready reads a renamed key this way (old_dynres). A few lines would readcpu_and_gpu_linesintofps_dev_show[11]and[12]when the new keys are missing, and add it to theminekeys insave()so it is dropped.gpu/vk/vk_menu.cpp:4957): the "not running" text repeats itself.Upscaler the game's network (FSR 4 not running), AA the game's network (FSR 4 not running)is about 95 characters in a 280-point column (about 40 characters at 12 point). So everything after the first half is cut, AA included. A shorter form would fit, e.g.Upscaler the game's network (FSR 4 not running)with the AA part left off when it is the same. With Version and hardware at every level, the levels below Breakdown use the 208-point column. ThereRender 2134x1200 (83%) to 2560x1440is probably cut too: worth a screenshot at the FPS level. The Steam Deck is no worse than other screens, since the font and the column scale together.gpu/vk/vk_menu.cpp:4961) readsRender 2560x1440 (native)withUpscaler Off (stretched). "none (native)" would match the other native modes.menu_i18n.h.gpu/vk/vk_menu.cpp:5120):fps_render_linesruns every frame while the footer shows: twosnprintf, a few heap strings and two compares.shortenonly runs when the text changes. That costs very little and only at Dev (or with version-always), so it's fine. Keying it on the sizes,upscaler(),opts.resolutionand the engine-drew bit would skip it, but this isn't needed.The mode names look right for each engine: DLSS, FSR 3.1 and FSR 4 use the DLSS names with a share for the other stops, XeSS uses its own presets, DLAA and the Native AA modes show as native, and Auto shows as Auto. The
(N%)is the width ratio, which is fine.Measurements: display only, and none are needed. The per-frame work is a few string operations while the footer shows, and only at Dev or with Version and hardware at every level.
Testing: Linux RX 6750 XT with FSR 3.1 at 83% and at 100%. Still needed: Game's own and Off in the game (they show the bug above), and DLSS and DLAA on NVIDIA, as the PR already asks.
@ -1721,3 +1721,3 @@{TAB_FPS_DEV, "Queue submits line", &opts.fps_dev_show[9], 0, 1, 1, tx_onoff, kFpsDevLineHelp, nullptr, false, true},{TAB_FPS_DEV, "Write-backs line", &opts.fps_dev_show[10], 0, 1, 1, tx_onoff, kFpsDevLineHelp, nullptr, false, true},{TAB_FPS_DEV, "CPU and GPU lines", &opts.fps_dev_show[11], 0, 1, 1, tx_onoff, kFpsDevLineHelp, nullptr, false, true},{TAB_FPS_DEV, "CPU line", &opts.fps_dev_show[11], 0, 1, 1, tx_onoff, kFpsDevLineHelp, nullptr, false, true},The old
cpu_and_gpu_lines=line now goes intosave()'sothersand stays in the file for good. Reading it into[11]and[12]inmenu_init(asold_dynresdoes) and adding it tominewould carry the player's choice over and drop the line.@ -4929,0 +4958,4 @@} else if (up == ur::GAME) {scale = "the game's network", aa = "the game's network";} else if (up == ur::OFF) {scale = "Off (stretched)", aa = "none";Nit: at 100% this reads
Render WxH (native)next toOff (stretched).none (native)would match the other native modes.@ -4929,0 +4954,4 @@std::string scale, aa;const bool on_chain = up != ur::GAME && up != ur::OFF && (up != ur::FSR3 || fsr3_on_chain());if (in_play && on_chain && !temporal_engine_drew()) {scale = aa = std::string("the game's network (") + name[up] + " not running)";Nit: this makes a line of about 95 characters, so it is cut around halfway in the 280-point column.
Upscaler the game's network (X not running), with the AA part left off when it is the same, would fit.@ -1631,6 +1658,7 @@ bool temporal_fg_inputs() { return temporal_mode() == ur::GAME && fsr3fi::framesbool temporal_upscale(VkCommandBuffer cmd, uint64_t key, bool known, Texture* in, VkImageView in_view, Texture* out,VkImageView out_view, uint32_t out_level) {const int mode = temporal_mode();if (known) note_drawn(in, out, out_level); // (one of the gameplay networks: the FPS counter's render line)Nothing calls
temporal_upscalewith Game's own:temporal_stand_inandtemporal_rejected_networkstop first unlesstemporal_wanted() || temporal_off() || temporal_fg_inputs(). So with the default Upscaler this never runs and the footer saysRender -all through play. The size needs noting before those gates.@ -1648,6 +1676,7 @@ bool temporal_upscale(VkCommandBuffer cmd, uint64_t key, bool known, Texture* in// (one in the gameplay networks' formats whose input is not this frame's scene: counted for the status line)if (!c.valid || texture_frame_no() - c.frame > 1 || in->width != c.width || in->height != c.height)return s.status_passed++, false;note_drawn(in, out, out_level); // (the scene into the picture: the gameplay upscale)Off returns into
stretchatif (mode == ur::GAME)before this branch, so Off at the fixed shares (67/58/50/33%, notknownkeys) never notes a size either.@fix
I changed this pull request where the review's points could be done here (the rest needs runs, measurements or decisions): the new head is
23b982e061. I rebased it ontomain(6e1521a70a).Conflicts resolved
docs/user/graphics-and-controls.md,gpu/vk/menu/fps_counter.cppdocs/user/graphics-and-controls.md,gpu/vk/menu/fps_counter.cppChanged
Changed in part (the rest is for the author)
Nothing was built or run: check the change. It will be reviewed again.
Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.
acce79610923b982e061Recommendation: Request changes. The footer would make screenshots easier to interpret, but it can report the wrong upscaler or AA, and its display thread reads the replay thread’s frame counter without synchronization. The diagnostic also changes which camera the renderer can select, even with the FPS counter off.
Adds render dimensions, upscaler and AA to the FPS footer, splits the footer controls, and shortens two overlapping labels. The Linux FSR 3.1 run shows the benefit. The implementation adds work to the draw and dispatch paths and couples the overlay to state used by upscaling and frame generation.
@fixwill do (17)gpu/vk/vk_temporal.cpp:1019: With frame generation, display_main reaches draw_fps through present_image while do_present can advance the plain texture frame_no on the replay thread. (more on the line)gpu/vk/vk_temporal.cpp:957: The footer’s broader format rule now selects the live engine camera. (more on the line)gpu/vk/vk_temporal.cpp:1010: Size observation calls note_output before confirming the candidate is the gameplay upscale. (more on the line)gpu/vk/vk_menu.cpp:3155: The footer infers the producer from the selected option and a 30-frame engine check. (more on the line)gpu/vk/vk_draw.cpp:11870: The size is stamped before any path has drawn. (more on the line)gpu/vk/vk_menu.cpp:3152: Native wording comes from the current menu values. (more on the line)gpu/vk/menu/fps_counter.cpp:145: Removing cpu_and_gpu_lines breaks that key for control-bridge clients. (more on the line)gpu/vk/vk_menu.cpp:3252: The two default render rows add 32 points without checking the available height. (more on the line)gpu/vk/vk_menu.cpp:3327: The fit cache truncates from the end, where the output dimensions, AA, or failure reason appear. (more on the line)gpu/vk/vk_menu.cpp:3563: The two shortened legend labels go straight through Tab::draw, which bypasses translation, so the 13 updated entries never appear there. (more on the line)gpu/vk/menu/fps_counter.cpp:145: The three new settings labels have no translation entries. (more on the line)gpu/vk/vk_menu.cpp:3563: The overlap fix has no game-free regression test as required by CONTRIBUTING.md. (more on the line)gpu/vk/vk_temporal.cpp:959: Gate the footer’s observer on a thread-safe enabled state and reuse the compute bindings already resolved for the dispatch.docs/user/images/guide/manifest.json:915: The settings screenshot tracks menu_i18n.h, vk_menu.cpp, vk_menu.h and vk_temporal.h, which all changed. Record the required comparison and update reviewed-source provenance. Keep the original capture metadata if the depicted Graphics view still matches.For the author (2) (a run, a measurement or a decision)
gpu/vk/vk_temporal.cpp:959: The counter defaults to Off, but the new path can still take the camera mutex and read submitted uniforms on qualifying Game’s own draws, and dispatch repeats descriptor lookups after building them. (more on the line)Checks: Still running:
PR description / Description follows the template (pull_request_target),CI / Website build and checks (pull_request),CI / Conventional Commit title (pull_request),CI / Workflow lint (actionlint) (pull_request),Guide screenshots / Guide images and local UI source review (pull_request),Third-party parts / As their notes say (pull_request).More: does it help, other systems, switches, what was read
Does it help: Unproven on
23b982e0. The author reports that b650470e showed 2134x1200 to 2560x1440 at 83% and 2560x1440 at native with FSR 3.1 on Linux/RADV. That supports those two display cases. The reported unchanged frame rate has no timings, 1% lows, or repeated runs against6e1521a7. There is no game result for the new capture paths on this head.Other systems: Camera selection and output-size tracking feed temporal upscaling, Off, DLAA before bloom, actor motion and frame interpolation. With frame generation, the display thread draws the footer while replay advances. AMD RADV and Steam Deck can use the typed or linear fallbacks, NVIDIA can use the translated networks and DLSS/DLAA, and XeSS is a Windows path. The supplied run covers AMD RADV with FSR 3.1 on the shared chain. The menu keys also serve the control bridge, and the edited menu files are tracked inputs to the website’s settings screenshot.
Switches: No RX_, HOS_ or LONGSHOT_ switches changed. CPU and GPU lines, previously On, is replaced by CPU line, GPU and driver lines, and Render lines. All three default to On for fresh profiles, which is reasonable when the footer is visible and legible. The head copies a saved legacy value into all three absent keys, so a player who hid hardware also starts with Render lines Off. Keep the new render default independent and preserve the legacy hardware control for existing callers. FPS counter itself still defaults to Off.
Reviewed:
23b982e061against6e1521a70a: Read AGENTS.md, CONTRIBUTING.md, the Forgejo template, developer standards, the applicable renderer and menu architecture rules, and the guide screenshot maintenance rules. No docs/review-checklist.md or sibling Tooling checkout was present. Read all eight file diffs and all changed code, the FPS menu module in full, and the surrounding camera, dispatch, fallback, presentation, settings and control-bridge paths. Skimmed all 13 translation changes, the screenshot manifest and fixture, and related test sources. Read the five commits, supplied comments, stale review and CI snapshot. Unchanged portions of the large renderer and menu files outside those paths were not read in full. Screenshot attachments were unavailable offline.; nothing run (a code review).Checked and fine:
Testing: Read-only review. git diff --check was clean. No build, test, game run or project script was executed. The supplied CI shows six pending jobs and no reported failures. The saved-key mismatch and camera format assertion were found by reading their source.
Overlaps and merge order: None found in the supplied review history or local refs. Open Forgejo pull requests could not be checked offline.
Automated code review by BeanBot (gpt-6.2-sol), asked for by BeanBot, after its fix. Nothing was run; a maintainer decides.
@ -144,3 +144,2 @@.preset_exempt = true}).row({.key = "cpu_and_gpu_lines",.label = "CPU and GPU lines",.row({.key = "cpu_line",L7: Removing cpu_and_gpu_lines breaks that key for control-bridge clients. A script setting it now gets no such row, even though INI loading recognizes it. The saved-key guard also rejects the removed key and the three unregistered new keys. Preserve the old combined operation through a compatible key or alias, register the new keys in FROZEN, and cover legacy INI loading with explicit new values taking precedence. Apply the legacy choice to CPU and GPU only, since it never controlled render details.
L11 (minor): The three new settings labels have no translation entries. A localized details page therefore mixes English controls with translated help. Translate the settings labels in all 13 tables. The footer itself can keep its documented English wording for bug reports.
@ -11848,6 +11867,7 @@ void dispatch(VkCommandBuffer cmd, NvnState& s, uint32_t x, uint32_t y, uint32_t// (a gameplay upscale's picture, copied for frame generation's HUD mark once it is written, whoever writes it)uint32_t picture_level = 0;Texture* const picture = upscale_picture(s, p, &picture_level);note_render_size(s, p);L5: The size is stamped before any path has drawn. In upscale_stub the stamp even precedes the blit capability checks. A rejected gameplay network whose fallback cannot record can therefore keep refreshing the footer as if it wrote a picture. On the translated path the observer also ignores failed image-view lookups. Use the resolved bindings and publish the sample only after a recognized gameplay route records its output.
@ -3196,3 +3249,4 @@const bool render_lines = footer && opts.fps_dev_show[fs::kDevLines + 2];if (foot) h += 6 * s + foot * lh;const float edge = left ? c.safe : c.w - c.safe;const float top = bottom ? c.h - c.safe - h : under_banner && !tune_mode && page != 5 && !beside ? banner_h(c) + 10 * s : c.safe;L8 (minor): The two default render rows add 32 points without checking the available height. At 960 pixels tall with Medium counter size, a top-anchored Dev counter with the render-thread lane, frame-time row and six Dev rows grows from roughly 920 to 968 pixels before margins. The footer fits before the change and is clipped after it. Compact the footer or fit the layout to the available height, and check both top and bottom anchors.
@ -3128,0 +3149,4 @@static const char* const name[ur::kUpscalers] = {"Game's own", "DLSS", "DLAA", "FSR 3.1", "FSR 4", "FSR 4 Native AA", "Off","XeSS", "XeSS Native AA"};const int up = std::clamp(upscaler(), 0, ur::kUpscalers - 1);const bool fsr3_native = up == ur::FSR3 && ur::percent(opts.resolution) == 100;L6 (minor): Native wording comes from the current menu values. For example, FSR 3.1 at Auto can draw at full size while the footer says it is upscaling. When switching to DLAA, the chain can still draw queued smaller frames with DLSS while the footer says native DLAA. Use the applied mode and dimensions from the same frame. Label Auto or a quality stop as a setting if it is shown alongside the applied mode.
@ -3128,0 +3152,4 @@const bool fsr3_native = up == ur::FSR3 && ur::percent(opts.resolution) == 100;std::string scale, aa;const bool on_chain = up != ur::GAME && up != ur::OFF && (up != ur::FSR3 || fsr3_on_chain());if (in_play && on_chain && !temporal_engine_drew()) {L4: The footer infers the producer from the selected option and a 30-frame engine check. On AMD’s software-matrix route, Game’s own at 75% can use the typed stand-in with FXAA, yet the footer says the game’s network. With RX_VK_FSR3_CHAIN=0 it says FSR 3.1 even after its input preparation fails. Off can also fall back when its stretch cannot be recorded. Publish the successful route and its AA from the draw path, including the typed and linear fallbacks, and keep the requested setting separate if it is useful to show.
@ -3262,0 +3324,4 @@fps_render_lines(now);if (now[0] != fit_lines[0] || now[1] != fit_lines[1])fit_lines[0] = now[0], fit_lines[1] = now[1], fit[4] = shorten(now[0], gw), fit[5] = shorten(now[1], gw);}L9 (minor): The fit cache truncates from the end, where the output dimensions, AA, or failure reason appear. The 208-point FPS column with the footer enabled can lose the end of a scaled 4K size or XeSS Ultra Performance and its AA. Use compact wording or a footer-specific width so the status fields survive, and verify the FPS-level layout.
@ -3493,3 +3562,2 @@const ImU32 red = ui::rgba(240, 100, 85);cell(0, 0, false, "Shaders built (4 s)", fps_built_text, strcmp(fps_built_text, "0") ? red : dim);cell(1, 0, false, "Draws left out (4 s)", fps_left_text, strcmp(fps_left_text, "0") ? red : dim);cell(0, 0, false, "Shaders built", fps_built_text, strcmp(fps_built_text, "0") ? red : dim);L10 (minor): The two shortened legend labels go straight through Tab::draw, which bypasses translation, so the 13 updated entries never appear there. A Japanese or French user still sees the English labels. Translate these two labels at the call site and fit the localized text beside the value.
L12: The overlap fix has no game-free regression test as required by CONTRIBUTING.md. The translation checks would still pass if the old labels and their entries came back together. Add a layout check that measures the label and value bounds in the actual half-width cell and fails on the original overlap. Cover the new capture and fallback selection with authored inputs while fixing those paths.
@ -950,2 +956,4 @@// (also with the game's network: the FPS footer identifies fixed-share upscales by the scene's size)if (!render_size_scene_format(color0->format)) return;State& s = st();std::lock_guard lk(s.mu);L13 (minor): The counter defaults to Off, but the new path can still take the camera mutex and read submitted uniforms on qualifying Game’s own draws, and dispatch repeats descriptor lookups after building them. Compare base and fixed head in the same scene with the counter Off and Dev, including average FPS, 1% lows and frame times on a Steam Deck or another CPU-limited setup. The older RX 6750 XT screenshots do not measure this path.
@ -948,2 +956,2 @@// (and with no engine, for frame generation's inputs: temporal_fg_inputs)if (!temporal_scene_format(color0->format) || (!temporal_wanted() && !temporal_off() && !temporal_fg_inputs())) return;// (also with the game's network: the FPS footer identifies fixed-share upscales by the scene's size)if (!render_size_scene_format(color0->format)) return;L2: The footer’s broader format rule now selects the live engine camera. With PsychoV35 and RX_VK_PSYCHOV35_DLSS=0, an RGBA16F draw can take the largest-camera slot and reject a later, smaller packed scene that the engine could have used. draw_engine then rejects the selected camera’s format. Keep temporal_scene_format and the existing activation rules for the engine camera. Use a separate size-only observation for the footer. Preserve the camera admission coverage in tests/test_psychov35_dlss.py.
@ -986,0 +1016,4 @@bool temporal_render_size(uint32_t* rw, uint32_t* rh, uint32_t* ow, uint32_t* oh) {const State& s = st();const uint64_t at = s.drawn_frame.load(std::memory_order_relaxed), v = s.drawn_sizes.load(std::memory_order_relaxed);if (!at || !v || texture_frame_no() - at > 30) return false; // (no gameplay upscale for half a second: a menu, a movie)L1: With frame generation, display_main reaches draw_fps through present_image while do_present can advance the plain texture frame_no on the replay thread. The new age reads race that increment, and the two relaxed loads can pair sizes with a stamp from another update. fps_render_lines also reads live settings while replay can auto-tune the render share. Publish one coherent snapshot of size, age and applied route on the replay side and carry it with the frame or publish it through a synchronized handoff.
@ -986,0 +1007,4 @@if (ow < in->width || oh < in->height || (!known && ow <= in->width)) return;State& s = st();std::lock_guard lk(s.mu);note_output(out, out_level);L3: Size observation calls note_output before confirming the candidate is the gameplay upscale. For example, a rejected same-format dispatch can enter through upscale_stub, fail the scene-size match, and still replace out_w/out_h. The next frame uses those dimensions to admit scene draws and skinned motion. Keep the footer’s candidate dimensions separate and leave the renderer’s size hints to its existing consumers.
I just saw the comments earlier, so I'll now be doing the requested changes and then I'll make a push, further BeanBot intervention isn't required.
@HumanBean, thanks for the review. BeanBot's follow-up (
23b982e0) noted the size before the upscaler gates by widening them, whichits own review then showed changes the engines' camera and output size.
8fcaae2creplaces that approach:The footer reports what drew, published by the path that drew it.
gpu/vk/upscale_route.hholds one record: the route(built-in network, its stand-ins, an engine, Off's stretch), the engine, and the 3D's and the picture's sizes. Each path publishes
after it has recorded the picture:
route_network_drawnafter the network's own dispatch,upscale_stub's typed and blitfallbacks,
taau_stand_in,draw_engine(DLAA before the bloom too),stretch, and the end offsr3_upscale(FSR 3.1 off thechain).
vk_temporal.cppand the camera, output size and gates are main's again, plus those calls.On your points:
Built-in at Auto shows its live size, Off at 67% and at Auto show theirs, FSR 3.1 at 83% and Quality theirs (screenshots below).
cpu_and_gpu_lines=: kept as main has it, one row for the CPU, GPU and driver, with its key; onlyrender_linesis new(in FROZEN). Nothing to migrate, and the control bridge keeps its key.
Upscaling: Built-in (FSR 4 not running)); a test checks 40.Upscaling: off, no AA, withResolution 2560x1440 (100%); no line says "native" any more, so the methodline and the resolution line don't mix.
upscale_route.hsays so.second.
On BeanBot's review of
23b982e0:steady clock (no frame counter shared between threads); sizes and the engine come from the same frame, so Auto, a failed
engine, FSR 3.1 drawing at full size and queued DLSS frames while switching to DLAA say what drew. Nothing is noted unless the
footer shows its render lines.
i18n::trnow.tests/test_fps_stats.cpp, one per route and case.check_guide_screenshots.py --review-sourcesalready fails on main (37 files, most of them main's ownchanges); the settings screenshot shows the Graphics tab, which this doesn't change, and restamping it would sign off main's
changes too.
Screenshots (Linux, RX 6750 XT, the counter at Dev). Before, main: no render lines, and "Shaders built (4 s)":
After, at
8fcaae2c:@review
Recommendation: Request changes. The footer works on the demonstrated Linux routes, but it still reads mutable replay settings during frame generation, can report a dispatch with missing image views, and labels queued DLSS frames as DLAA. The test for the queued-frame case supplies the correct mode by hand, so it misses the producer bug.
The FPS footer adds render and output sizes and the upscaling or AA method. The new record keeps diagnostics out of the upscalers’ camera and output state. It adds observation on draw and dispatch paths, and presentation can still pair the latest replay record with an older picture.
@fixwill do (14)gpu/vk/vk_menu.cpp:3137: The display thread can format the footer while the replay thread writes opts.resolution during Auto preset reconciliation. menu_auto_tune takes menu_mu for the write, but menu_record does not hold that lock here. Generated presentation can also show a queued picture after replay publishes a newer route. Capture the settings needed by the footer on the producing side and carry the record with the picture through the present queue.gpu/vk/vk_draw.cpp:9241: For an unknown key, route_publish only checks dimensions and the input texture’s format. The blit caller supplies any successful upscale_stub, so a later non-gameplay blit from another packed texture the same size as the scene can replace the real route, even with a different destination format. Require a matching gameplay source and output, using the actual scene identity and the frame’s resolved bindings.gpu/vk/vk_draw.cpp:9683: texture_view_of can return a null view with a non-null Texture pointer. If the gameplay output view fails and the dispatch binds a dummy, this path can still report that the network wrote the real picture. Reuse successful descriptor bindings and their mip levels, and publish only when the real input and writable output were bound.gpu/vk/vk_temporal.cpp:1747: On a switch from DLSS to DLAA, waiting_for_native keeps draw_mode at DLSS for queued small scenes, but publication passes mode, so a scaled DLSS frame can read Upscaling: DLAA Quality. Publish draw_mode and distinguish AA followed by an output stretch from engine upscaling. Exercise the producer’s mode selection in the queued-frame test.gpu/vk/vk_menu.cpp:3216: When Version and hardware at every level is enabled and the open menu forces a top anchor below its banner, the clamp still budgets from c.safe. In a short window the new footer can extend below the safe bottom edge. Compute the available height from the actual anchor before dropping footer rows.gpu/vk/vk_menu.cpp:3294: At FPS level with the footer enabled, the 208-point fit can cut the output size or the failed-engine suffix. A Built-in fallback with FXAA and FSR 4 not running is longer than the tested plain Built-in case. Use compact forms or a footer width that preserves both fields, and check pixel widths with the actual font.gpu/vk/menu/fps_counter.cpp:153: Only Render lines remains new. A localized details page shows that row and its help in English. Add both strings through the existing menu translation mechanism while keeping the diagnostic footer text English.tests/test_fps_stats.cpp:170: The new cases format preselected samples. They would still pass with the live DLSS-to-DLAA producer bug, and restoring the old legend labels and translations would not fail them. Add authored checks for route admission and applied-mode selection, and a game-free legend test using the label and value bounds in the real half-width cell.gpu/vk/vk_draw.cpp:8305: Cache the enabled state per replay frame and reuse resolved dispatch bindings.docs/user/images/guide/manifest.json:915: The settings screenshot tracks menu_i18n.h, vk_menu.cpp, vk_menu.h and vk_temporal.h, which all changed. Record the required comparison and update reviewed-source provenance. Keep the original capture metadata if the depicted Graphics view still matches.gpu/vk/vk_menu.cpp:3137: The requested-engine text uses upscaler(), which returns Game’s own when temporal_mode() rejects an unavailable SDK. (more on the line)PR description / Description follows the template (pull_request_target)fails on this head and passes on the base branch: Failing after 5s.For the author (2) (a run, a measurement or a decision)
gpu/vk/vk_draw.cpp:8305: With the footer visible, wanted() reads the clock on eligible draw calls and note_scene reads it again on qualifying draws. The translated path repeats descriptor lookups. Compare base and HEAD with the counter Off and Dev in a CPU-limited scene, including average FPS, 1% lows and frame times.Since the last review (on
23b982e061)Resolved: L2 (The footer’s broader format rule now selects the live engine camera. With PsychoV35 and RX…); L3 (Size observation calls note_output before confirming the candidate is the gameplay upscale…); L7 (Removing cpu_and_gpu_lines breaks that key for control-bridge clients. A script setting it…); L10 (The two shortened legend labels go straight through Tab::draw, which bypasses translation,…); L16 (The description is stale for this head. It says the old key is ignored, describes the orig…)
Still open: L1 (The display thread can format the footer while the replay thread writes opts.resolution du…); L4 (For an unknown key, route_publish only checks dimensions and the input texture’s format. T…); L5 (texture_view_of can return a null view with a non-null Texture pointer. If the gameplay ou…); L6 (On a switch from DLSS to DLAA, waiting_for_native keeps draw_mode at DLSS for queued small…); L8 (When Version and hardware at every level is enabled and the open menu forces a top anchor…); L9 (At FPS level with the footer enabled, the 208-point fit can cut the output size or the fai…); L11 (Only Render lines remains new. A localized details page shows that row and its help in Eng…); L12 (The new cases format preselected samples. They would still pass with the live DLSS-to-DLAA…); L13 (With the footer visible, wanted() reads the clock on eligible draw calls and note_scene re…); L14 (The AMD Linux report covers the new head on one GPU. Record the remaining fixed-share and…); L15 (The settings screenshot tracks menu_i18n.h, vk_menu.cpp, vk_menu.h and vk_temporal.h, whic…); L17 (The PR declares AI assistance, but none of its five commits has the required Assisted-by o…)
New: L18 (The requested-engine text uses upscaler(), which returns Game’s own when temporal_mode() r…); L19 (The check
PR description / Description follows the template (pull_request_target)fails…)Checks: Still running:
CI / Conventional Commit title (pull_request),CI / Python source checks (ubuntu-24.04, part 1/2) (pull_request),CI / Python source checks (ubuntu-24.04, part 2/2) (pull_request),CI / SPIR-V fixtures (global narrow load/store, direct surface atomics, F2I immediate) (pull_request),CI / Workflow lint (actionlint) (pull_request),CI / Website build and checks (pull_request),Guide screenshots / Guide images and local UI source review (pull_request),Third-party parts / As their notes say (pull_request),Third-party parts / The same as upstream (pull_request),Website / Website on Cloudflare (pull_request).More: does it help, other systems, switches, what was read
Does it help: Yes for the display feature on the Linux RX 6750 XT with RADV Mesa 26.2.7. The author reports that
8fcaae2cshows 1920x1080 to 2560x1440 (75%) with the Built-in FXAA fallback at Auto, and reports FSR 3.1, Off and a menu transition. CPU and frame-pacing cost are unproven. No matched base and HEAD timing runs were supplied.Other systems: The observer runs on both OSes and across GPU vendors. The temporal camera, motion and output gates match main. NVIDIA DLSS/DLAA, FSR 4, XeSS, Steam Deck, Windows and generated-frame presentation still lack current-head evidence. The diff adds no save, audio, input or public renderer API changes.
Switches: Render lines adds the preset-exempt render_lines INI key, default On. The FPS counter still defaults Off. On is appropriate for an optional diagnostic footer once the report is reliable. cpu_and_gpu_lines keeps its existing key and behavior. No environment switches change.
Reviewed:
8fcaae2c4cagainst6e1521a70a: Read AGENTS.md, CONTRIBUTING.md, the Forgejo template, the applicable developer standards, the comments and reviews, and all six commit messages. Reviewed the base-to-HEAD changes and HEAD implementations around the ledger items, including descriptor resolution, engine selection, the presentation caller, auto-tuning, the menu key and the test fixtures. Read upscale_route.cpp and upscale_route.h in full. Reviewed changed regions and callers in the larger files, skimmed translation entries and screenshot provenance, and used the supplied run reports for evidence. Screenshot attachments were unavailable in the checkout.; nothing run (a code review).Checked and fine:
Testing: Read-only review. The author reports 20 new formatter cases and 216 passed, 34 skipped and 19 named baseline failures in the changed-source run. The description check fails on this head and the supplied CI snapshot shows other jobs still running. Tier 1 and Windows verification remain pending.
Overlaps and merge order: None identified in the supplied history.
Automated code review by BeanBot (gpt-6.2-sol), asked for by HumanBean. Nothing was run; a maintainer decides.
@ -3128,0 +3134,4 @@static void fps_render_lines(std::string out[2]) {upscale_route::Sample r;const bool drawn = upscale_route::latest(&r, fps_stats::kSecond / 2);upscale_route::footer_lines(drawn ? &r : nullptr, std::clamp(upscaler(), 0, ur::kUpscalers - 1), opts.resolution, out);L18 (minor): The requested-engine text uses upscaler(), which returns Game’s own when temporal_mode() rejects an unavailable SDK. A saved DLSS selection with NGX unavailable can therefore show plain Built-in. Capture the requested engine before capability checks and keep it separate from the engine that drew.
@fix
@fix
I didn't fix this pull request: it failed (Codex didn't finish with an answer (exit 1)). Nothing was pushed.
Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.
@fix
I didn't fix this pull request: it isn't open.
Automated fix by BeanBot (gpt-6.1-terra), asked for by HumanBean. Nothing was run; a maintainer decides.