feat(framegen): AMD FSR 4 and AMD FSR 3.1 as Frame generation by choices #2665

Open

notquiteog wants to merge 1 commit from notquiteog/feat/framegen-by-amd into main

notquiteog commented

2026-11-12 04:48:49 +00:00

(Migrated from github.com)

The AMD FSR 4 half has not been run: AMD's runtime is Windows DLLs and this PC is Linux with an NVIDIA card. It is up for review as it is, with that said plainly; it needs a run on Windows with an AMD card before it merges, and I would be glad of one (testing wanted).

What it does

Row value	What makes the frames	Where
0 Auto	as today (DLSS Frame Generation where it runs, else the renderer's own)
1 Built-in	as today
2 AMD FSR 4	AMD's frame generation through AMD's runtime (`vk_ffx.cpp`), from the scene's depth, motion and camera; every made frame of a set in one dispatch	Windows, built with `--fsr`, AMD's DLLs beside the game
3 AMD FSR 3.1	the interpolation already in the tree (`vk_fsr3fi.cpp`), chosen in the menu at 2x	wherever it can run

framegen_methods.h is the choice, pure and tested: a value that does not run on this PC is stepped over by the arrows and shown as Auto when it comes from a saved file, and a set a method cannot make whole goes to the renderer's own, never to nothing. DLSS Frame Generation reads the row as it is shown (framegen_by_auto), so a saved AMD value on an NVIDIA PC still gets DLSS-G.
vk_ffx.cpp hands AMD's runtime our own images on our own command buffer (its "manual" dispatch, no swapchain of AMD's): Hookshot keeps presenting and pacing. A present it did not make is told to it as a reset. A context that cannot be made turns the choice off for the run and the row goes back to Auto.
camera_vectors.h: the camera's position and axes in the world from the projection and view-projection, which AMD's runtime takes.
The three frame generation headers are added to third_party/fidelityfx beside the four already there (MIT, same pinned SDK; the README there says which). AMD's frame generation DLL is not fetched by scripts/ffx_sdk.py yet: the readme says to put it beside the loader.
No change to vk_api.h.

How it was tested

tests/run_all.py (Linux), with main at 65a4d874 merged in: 500 passed, 2 failed, 22 skipped. Two checks fail on this branch exactly as they fail on main at 65a4d874 without it: test_readme_menu_rows (the Skip opening row has no readme line; #2857 is open for it) and test_program_cache_one_variant (its fixture lacks the zstd cache's names since #1630). New: test_framegen_methods, test_camera_vectors.
The Vulkan back end builds on Linux (clang, with NGX and --fsr). vk_ffx.cpp's Windows side and the other touched files pass clang++ --target=x86_64-w64-mingw32 -fsyntax-only with -DHOOKSHOT_FSR.
In the game with this branch's back end in a game built from main at 65a4d874 (run again after that merge): RTX 5060 Ti 16 GB, driver 615.71.09, Linux (COSMIC, Wayland), Kokiri Forest, 2048x1152 output, DLSS, V-Sync Off, 50 to 60 fps of game frames, no error or lost device in the log.
- Auto at 2x: 600 of 600 presents made by DLSS-G.
- AMD FSR 3.1 at 2x: fsr3fi: 600 of the last 600 made frames by FSR 3.1, 0 without its inputs; at 3x the renderer's own makes the set and the scene's inputs are no longer kept.
- AMD FSR 4 asked on this PC: the row steps over it to Built-in.
- Back to Auto: DLSS-G's feature is made again, 599 of 600.

Not tested

AMD FSR 4's frame generation at run time: loading AMD's DLLs, the provider query, the context, the dispatches, the image layouts, how the made frames look and how they pace. Compiled and read only.
Windows at all, AMD and Intel cards.
HDR with either AMD method.

Known and left

RX_VK_FG_METHOD=fsr3 does not start FSR 3.1 by itself (fsr3fi::asked() reads the row, and RX_VK_FSR_FG goes over both); the row does.
ffx::shutdown() is not called at exit.
The two new hints are English in all 13 menu languages.
As #1635 found, FSR 3.1's interpolation has no HUD-less colour here, so the HUD moves with the scene at 2x: that is why it stays a choice and not Auto's.

🤖 Generated with Claude Code

The Frame generation by row offers Auto and Built-in only: a player with an AMD card has no way to pick AMD's own frame generation from the menu, and FSR 3.1's interpolation is reachable only through `RX_VK_FSR_FG=1`. This adds both as choices of the row. It is the "FidelityFX FSR 4 fallback as a small, separate PR" asked for when #1831 was closed. **The AMD FSR 4 half has not been run**: AMD's runtime is Windows DLLs and this PC is Linux with an NVIDIA card. It is up for review as it is, with that said plainly; it needs a run on Windows with an AMD card before it merges, and I would be glad of one (`testing wanted`). ## What it does | Row value | What makes the frames | Where | |---|---|---| | 0 Auto | as today (DLSS Frame Generation where it runs, else the renderer's own) | | | 1 Built-in | as today | | | 2 AMD FSR 4 | AMD's frame generation through AMD's runtime (`vk_ffx.cpp`), from the scene's depth, motion and camera; every made frame of a set in one dispatch | Windows, built with `--fsr`, AMD's DLLs beside the game | | 3 AMD FSR 3.1 | the interpolation already in the tree (`vk_fsr3fi.cpp`), chosen in the menu at 2x | wherever it can run | - `framegen_methods.h` is the choice, pure and tested: a value that does not run on this PC is stepped over by the arrows and shown as Auto when it comes from a saved file, and a set a method cannot make whole goes to the renderer's own, never to nothing. DLSS Frame Generation reads the row as it is shown (`framegen_by_auto`), so a saved AMD value on an NVIDIA PC still gets DLSS-G. - `vk_ffx.cpp` hands AMD's runtime our own images on our own command buffer (its "manual" dispatch, no swapchain of AMD's): Hookshot keeps presenting and pacing. A present it did not make is told to it as a reset. A context that cannot be made turns the choice off for the run and the row goes back to Auto. - `camera_vectors.h`: the camera's position and axes in the world from the projection and view-projection, which AMD's runtime takes. - The three frame generation headers are added to `third_party/fidelityfx` beside the four already there (MIT, same pinned SDK; the README there says which). AMD's frame generation DLL is not fetched by `scripts/ffx_sdk.py` yet: the readme says to put it beside the loader. - No change to `vk_api.h`. ## How it was tested - `tests/run_all.py` (Linux), with main at 65a4d874 merged in: 500 passed, 2 failed, 22 skipped. Two checks fail on this branch exactly as they fail on main at 65a4d874 without it: `test_readme_menu_rows` (the Skip opening row has no readme line; #2857 is open for it) and `test_program_cache_one_variant` (its fixture lacks the zstd cache's names since #1630). New: `test_framegen_methods`, `test_camera_vectors`. - The Vulkan back end builds on Linux (clang, with NGX and `--fsr`). `vk_ffx.cpp`'s Windows side and the other touched files pass `clang++ --target=x86_64-w64-mingw32 -fsyntax-only` with `-DHOOKSHOT_FSR`. - In the game with this branch's back end in a game built from main at 65a4d874 (run again after that merge): RTX 5060 Ti 16 GB, driver 615.71.09, Linux (COSMIC, Wayland), Kokiri Forest, 2048x1152 output, DLSS, V-Sync Off, 50 to 60 fps of game frames, no error or lost device in the log. - Auto at 2x: 600 of 600 presents made by DLSS-G. - AMD FSR 3.1 at 2x: `fsr3fi: 600 of the last 600 made frames by FSR 3.1, 0 without its inputs`; at 3x the renderer's own makes the set and the scene's inputs are no longer kept. - AMD FSR 4 asked on this PC: the row steps over it to Built-in. - Back to Auto: DLSS-G's feature is made again, 599 of 600. ## Not tested - **AMD FSR 4's frame generation at run time**: loading AMD's DLLs, the provider query, the context, the dispatches, the image layouts, how the made frames look and how they pace. Compiled and read only. - Windows at all, AMD and Intel cards. - HDR with either AMD method. ## Known and left - `RX_VK_FG_METHOD=fsr3` does not start FSR 3.1 by itself (`fsr3fi::asked()` reads the row, and `RX_VK_FSR_FG` goes over both); the row does. - `ffx::shutdown()` is not called at exit. - The two new hints are English in all 13 menu languages. - As #1635 found, FSR 3.1's interpolation has no HUD-less colour here, so the HUD moves with the scene at 2x: that is why it stays a choice and not Auto's. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

github-sync changed title from ~~WIP: feat(framegen): AMD FSR 4 and AMD FSR 3.1 as Frame generation by choices~~ to feat(framegen): AMD FSR 4 and AMD FSR 3.1 as Frame generation by choices

2026-11-12 06:24:03 +00:00

github-sync referenced this pull request

2026-11-12 13:38:15 +00:00

feat(menu): Frame generation by and the FPS counter name who makes the frames #2788

github-sync referenced this pull request

2026-11-12 14:43:51 +00:00

fix(vk): DLSS comes back after a feature that was not made, and frame generation takes the multiplier its sets were made for #2658

github-sync referenced this pull request

2026-11-12 14:48:23 +00:00

feat(menu): Frame generation by and the FPS counter name who makes the frames #2788

github-sync referenced this pull request

2026-11-12 16:54:38 +00:00

fix(vk): keep ultrawide loading fades full width #2851

github-sync referenced this pull request

2026-11-12 16:55:32 +00:00

fix(vk): DLSS comes back after a feature that was not made, and frame generation takes the multiplier its sets were made for #2658

github-sync referenced this pull request from a commit

2026-11-12 19:00:50 +00:00

feat(menu): Frame generation by and the FPS counter name who makes the frames (#2788)

github-sync commented

2026-11-12 19:01:50 +00:00

Owner

@nullifyshot commented on GitHub · 2026-11-12 18:19 UTC · original

Hi, I know this isn't part of your PR, but I recommend changing "Auto" to DLSS, and move it over so 0 becomes Built-In and 1 becomes DLSS. We had discussions about it w/ the person who implemented DLSS, and that's what we like to do. We would like to basically hit two birds with one stone in this case. Thank you!

> [@nullifyshot](https://github.com/nullifyshot) commented on GitHub · 2026-11-12 18:19 UTC · [original](https://github.com/NotRexed/Hookshot/pull/2665#issuecomment-6832138939) Hi, I know this isn't part of your PR, but I recommend changing "Auto" to DLSS, and move it over so 0 becomes Built-In and 1 becomes DLSS. We had discussions about it w/ the person who implemented DLSS, and that's what we like to do. We would like to basically hit two birds with one stone in this case. Thank you!

Commenting is not possible because the repository is archived.

No reviewers