Choose FSR 4.1 or a built-in FSR 3.1 from one Insert menu row

N3oTraX commented

2026-11-08 23:41:56 +00:00

(Migrated from github.com)

One Insert menu row chooses the upscaler: Off, Auto, FSR 4.1, FSR 3.1, changed at once while playing. FSR 3.1 needs no external library, so cards and systems without FSR 4 still get a temporal upscaler.

Stacked on hookshot-amd#2 (the Linux port, on the current #828, which sits on #849 and #850) and including #822 (with #821) at 2008ac7. GitHub shows all of those commits until they land; the last two commits are new here, plus the merge of #822.

What changed

FSR 3.1 runs on the inputs #828 already builds: the same colour, depth, camera motion and jitter, with no second input path.
The passes are recompiled for those inputs instead of converting them: HDR_COLOR_INPUT=0, INVERTED_DEPTH=0, low-resolution motion, not jittered. Near, far and field of view come from the camera used for the motion. tests/test_fsr3_shaders.py pins the options.
Menu: the AMD FSR 4.1 row becomes FSR upscaler. Auto takes FSR 4.1 when the library is present with a model for the card, else FSR 3.1, else the game's upscale. The row shows the engine in use, and why a forced choice cannot run. If FSR 4.1 fails in play, FSR 3.1 takes over.
Compatibility: RX_FSR4=0/1 and fsr4= still read as Off and Auto. New: RX_UPSCALER=off|auto|fsr4|fsr3 and upscaler=.
Fixes to #822 found on the way:
- images of an old size were destroyed while work in flight could still use them, and the new ones stayed UNDEFINED;
- the no-wave pyramids were never selected;
- hos_oot_camera and hos_oot_view_projection are weak, so the renderer still loads with a host built without them.

One conflict with #822, in the network dispatch log in vk_draw.cpp: both behaviours are kept.

How it was tested


GPU, driver	RX 7900 XTX (RDNA3), RADV Mesa 26.2.6
System	Linux 7.3.1, Hyprland
Game	v131072, build `e0f3ccfa`, Ultrawide on
Result	FSR 4.1.1 INT8 and FSR 3.1 each drew `5120x2160 -> 5120x2160`, switched live several times from the menu. FSR 3.1 reported `camera near 0.100 far 25040.7`, and all 11 passes were created. No crash or device lost in the log
Checks	`tests/test_fsr3_shaders.py`: 11 modules valid with `spirv-val`. `tests/run_all.py fsr menu upscale resolution`: 7 passed, 1 skipped (`test_upscale_contract`, no Vulkan headers found; passes by hand). Full suite: 98 passed, 13 skipped, 1 failed (`test_mem_trim`, which fails the same way on the commit below this work)

Image quality was not compared and fps was not measured.

Run again after the rebase (6095c4d, full game rebuilt): same GPU and driver, FSR 4.1.1 INT8 and FSR 3.1 each drew 5120x2160 -> 5120x2160, switched live several times from the menu, no crash or device lost in the log.

The run used RX_VK_NO_WARM=1, RX_VK_NO_PREDICT=1 and RX_VK_PRECOMPILE=0 because of #803, and RX_VK_ANISO=16.

Not tested

RDNA2 and RDNA4, and the real FSR 4.1 failure path to FSR 3.1.
Windows: not compiled or run here, including the weak symbols.
The earlier RX_VK_FSR=1 route now runs on the same LDR shaders and was not revalidated.

🤖 Generated with Claude Code

One Insert menu row chooses the upscaler: **Off, Auto, FSR 4.1, FSR 3.1**, changed at once while playing. FSR 3.1 needs no external library, so cards and systems without FSR 4 still get a temporal upscaler. > **Stacked on [hookshot-amd#2](https://github.com/madebyjordan/hookshot-amd/pull/2)** (the Linux port, on the current #828, which sits on #849 and #850) and including #822 (with #821) at `2008ac7`. GitHub shows all of those commits until they land; the last two commits are new here, plus the merge of #822. ## What changed - **FSR 3.1 runs on the inputs #828 already builds:** the same colour, depth, camera motion and jitter, with no second input path. - **The passes are recompiled for those inputs** instead of converting them: `HDR_COLOR_INPUT=0`, `INVERTED_DEPTH=0`, low-resolution motion, not jittered. Near, far and field of view come from the camera used for the motion. `tests/test_fsr3_shaders.py` pins the options. - **Menu:** the `AMD FSR 4.1` row becomes `FSR upscaler`. Auto takes FSR 4.1 when the library is present with a model for the card, else FSR 3.1, else the game's upscale. The row shows the engine in use, and why a forced choice cannot run. If FSR 4.1 fails in play, FSR 3.1 takes over. - **Compatibility:** `RX_FSR4=0/1` and `fsr4=` still read as Off and Auto. New: `RX_UPSCALER=off|auto|fsr4|fsr3` and `upscaler=`. - **Fixes to #822 found on the way:** - images of an old size were destroyed while work in flight could still use them, and the new ones stayed `UNDEFINED`; - the no-wave pyramids were never selected; - `hos_oot_camera` and `hos_oot_view_projection` are weak, so the renderer still loads with a host built without them. One conflict with #822, in the network dispatch log in `vk_draw.cpp`: both behaviours are kept. ## How it was tested | | | |---|---| | GPU, driver | RX 7900 XTX (RDNA3), RADV Mesa 26.2.6 | | System | Linux 7.3.1, Hyprland | | Game | v131072, build `e0f3ccfa`, Ultrawide on | | Result | FSR 4.1.1 INT8 and FSR 3.1 each drew `5120x2160 -> 5120x2160`, switched live several times from the menu. FSR 3.1 reported `camera near 0.100 far 25040.7`, and all 11 passes were created. No crash or device lost in the log | | Checks | `tests/test_fsr3_shaders.py`: 11 modules valid with `spirv-val`. `tests/run_all.py fsr menu upscale resolution`: 7 passed, 1 skipped (`test_upscale_contract`, no Vulkan headers found; passes by hand). Full suite: 98 passed, 13 skipped, 1 failed (`test_mem_trim`, which fails the same way on the commit below this work) | **Image quality was not compared and fps was not measured.** **Run again after the rebase (`6095c4d`, full game rebuilt):** same GPU and driver, FSR 4.1.1 INT8 and FSR 3.1 each drew `5120x2160 -> 5120x2160`, switched live several times from the menu, no crash or device lost in the log. The run used `RX_VK_NO_WARM=1`, `RX_VK_NO_PREDICT=1` and `RX_VK_PRECOMPILE=0` because of #803, and `RX_VK_ANISO=16`. ## Not tested - RDNA2 and RDNA4, and the real FSR 4.1 failure path to FSR 3.1. - Windows: not compiled or run here, including the weak symbols. - The earlier `RX_VK_FSR=1` route now runs on the same LDR shaders and was not revalidated. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

NotRexed commented

2026-11-09 05:30:45 +00:00

(Migrated from github.com)

Thank you! This is stacked on #857 (closed), #828, #822 and #821, and conflicts with main. Closing to keep the queue clear — please rebase on current main and reopen; thanks! As separate, smaller PRs if possible.

N3oTraX commented

2026-11-09 08:03:03 +00:00

(Migrated from github.com)

Thank you! This is stacked on #857 (closed), #828, #822 and #821, and conflicts with main. Closing to keep the queue clear — please rebase on current main and reopen; thanks! As separate, smaller PRs if possible.

I will build a rebased PR during this day. Thanks for your feedback.

> Thank you! This is stacked on #857 (closed), #828, #822 and #821, and conflicts with main. Closing to keep the queue clear — please rebase on current main and reopen; thanks! As separate, smaller PRs if possible. I will build a rebased PR during this day. Thanks for your feedback.

N3oTraX commented

2026-11-09 11:17:52 +00:00

(Migrated from github.com)

@NotRexed Done ! On top of 2.0.7.

Rows
Columns

Choose FSR 4.1 or a built-in FSR 3.1 from one Insert menu row #871

What changed

How it was tested

Not tested