MagicBeanDev/Hookshot

Archived

Watch

Fork 0

fix(framegen): FSR 3.1 interpolation built with the SDK's name for low-resolution motion vectors #1943

Merged

j64esatxzdrhfg merged 1 commit from fix/fsr3fi-render-size-blur into main

2026-11-11 19:14:27 +00:00

j64esatxzdrhfg commented

2026-11-10 21:23:33 +00:00

(Migrated from github.com)

With FSR 3.1 frame interpolation (RX_VK_FSR_FG=1) and the 3D drawn smaller than the picture, every made frame is blurred and smeared in the top-left render-size corner. This happens with DLSS or FSR 4, or with the game's own upscale. The cause is a misspelt shader option: the interpolation pass that reads the game's motion was built for motion at the display size.

What changed:

scripts/fsr3_shaders.py passes -DFFX_FRAMEINTERPOLATION_OPTION_LOW_RES_MOTION_VECTORS=1, the name the SDK reads (ffx_frameinterpolation_reconstruct_dilated_velocity_and_previous_depth.h:106, and LOW_RES_ in AMD's CMakeCompileFrameinterpolationShaders.txt). Before, it passed the upscaler's spelling, ..._LOW_RESOLUTION_..., which the SDK ignores. So fi_reconstruct_and_dilate read the render-size motion at ComputeHrPosFromLrPos(), a render pixel times display/render.
gpu/vk/shaders/fsr3_fi_reconstruct_and_dilate_comp_spv.h is regenerated with glslang 16.6.0 (CI's pin). It is the only pass that reads the option, and its bindings and capabilities are unchanged. The other 27 FSR headers recompile byte-identical, apart from glslang's version comment, and are left alone.
gpu/vk/vk_fsr3fi.h: the Inputs comment named the same wrong macro.
New tests/test_fsr3_shader_options.py (needs no glslang): every -D the script passes for the upscaler, interpolation and optical flow must be tested by an #if in the SDK files those passes compile from, and every option the interpolation passes test must be set. It fails on main with exactly this name. tests/test_fsr3_shaders.py could not catch it, because the headers matched their misspelt sources.
The renderer is unchanged: FSR keeps the render-size depth and motion. The first version's display-size blits, RX_VK_FSR_FI_RENDER_SIZE and its readme note are gone.

Testing:

scripts/fsr3_shaders.py --check with glslang 16.6.0: all 28 headers match.
tests/run_all.py fsr3 fsr3fi framegen_resources upscale_independent on Windows: 4 passed. test_fsr3_shaders is skipped there because spirv-val is missing; run directly without it, all 28 match.
vk_fsr3fi.cpp and vk_framegen.cpp compile-checked. Based on main d7a7306.
In game: RTX 5080, DLSS Balanced to 3840x2160, frame generation 2x with RX_VK_FSR_FG=1, in a build of main (d7a7306) with this fix. FSR 3.1 made 508 of the last 600 made frames, from render-size depth and motion (2240x1260), and the captures below are sharp in the top-left corner, where main's made frames are smeared.

Not tested: AMD/Intel, FSR 4 as the upscaler, Linux.

Screenshots (a made frame each, DLSS Balanced to 3840x2160)

The before shot is main's behaviour: a made frame, smeared in the top-left render-size corner (HUD hearts doubled). The after shots are this fix, with the render-size inputs.

Before (main)	After (this PR)

Top-left corner at the same scale, main and two captures with this PR:

HDR10 captures, tone-mapped to SDR for this page (paper white to white, highlights rolled off up to 800 nits).

🤖 Generated with Claude Code

With FSR 3.1 frame interpolation (`RX_VK_FSR_FG=1`) and the 3D drawn smaller than the picture, every made frame is blurred and smeared in the top-left render-size corner. This happens with DLSS or FSR 4, or with the game's own upscale. The cause is a misspelt shader option: the interpolation pass that reads the game's motion was built for motion at the display size. **What changed:** - `scripts/fsr3_shaders.py` passes `-DFFX_FRAMEINTERPOLATION_OPTION_LOW_RES_MOTION_VECTORS=1`, the name the SDK reads (`ffx_frameinterpolation_reconstruct_dilated_velocity_and_previous_depth.h:106`, and `LOW_RES_` in AMD's `CMakeCompileFrameinterpolationShaders.txt`). Before, it passed the upscaler's spelling, `..._LOW_RESOLUTION_...`, which the SDK ignores. So `fi_reconstruct_and_dilate` read the render-size motion at `ComputeHrPosFromLrPos()`, a render pixel times display/render. - `gpu/vk/shaders/fsr3_fi_reconstruct_and_dilate_comp_spv.h` is regenerated with glslang 16.6.0 (CI's pin). It is the only pass that reads the option, and its bindings and capabilities are unchanged. The other 27 FSR headers recompile byte-identical, apart from glslang's version comment, and are left alone. - `gpu/vk/vk_fsr3fi.h`: the `Inputs` comment named the same wrong macro. - New `tests/test_fsr3_shader_options.py` (needs no glslang): every `-D` the script passes for the upscaler, interpolation and optical flow must be tested by an `#if` in the SDK files those passes compile from, and every option the interpolation passes test must be set. It fails on main with exactly this name. `tests/test_fsr3_shaders.py` could not catch it, because the headers matched their misspelt sources. - The renderer is unchanged: FSR keeps the render-size depth and motion. The first version's display-size blits, `RX_VK_FSR_FI_RENDER_SIZE` and its readme note are gone. **Testing:** - `scripts/fsr3_shaders.py --check` with glslang 16.6.0: all 28 headers match. - `tests/run_all.py fsr3 fsr3fi framegen_resources upscale_independent` on Windows: 4 passed. `test_fsr3_shaders` is skipped there because spirv-val is missing; run directly without it, all 28 match. - `vk_fsr3fi.cpp` and `vk_framegen.cpp` compile-checked. Based on main d7a7306. - In game: RTX 5080, DLSS Balanced to 3840x2160, frame generation 2x with `RX_VK_FSR_FG=1`, in a build of main (d7a7306) with this fix. FSR 3.1 made 508 of the last 600 made frames, from render-size depth and motion (2240x1260), and the captures below are sharp in the top-left corner, where main's made frames are smeared. **Not tested:** AMD/Intel, FSR 4 as the upscaler, Linux. **Screenshots** (a made frame each, DLSS Balanced to 3840x2160) The before shot is main's behaviour: a made frame, smeared in the top-left render-size corner (HUD hearts doubled). The after shots are this fix, with the render-size inputs. | Before (main) | After (this PR) | |---|---| | ![Made frame on main: smeared in the top-left render-size corner](https://raw.githubusercontent.com/j64esatxzdrhfg/Hookshot/pr-screenshots/fsr3fi-blur/before.jpg) | ![With the define fix: sharp](https://raw.githubusercontent.com/j64esatxzdrhfg/Hookshot/pr-screenshots/fsr3fi-blur/after-define-fix.jpg) | Top-left corner at the same scale, main and two captures with this PR: ![Top-left corner: main, then this PR twice](https://raw.githubusercontent.com/j64esatxzdrhfg/Hookshot/pr-screenshots/fsr3fi-blur/top-left-main-vs-fix.jpg) _HDR10 captures, tone-mapped to SDR for this page (paper white to white, highlights rolled off up to 800 nits)._ 🤖 Generated with [Claude Code](https://claude.com/claude-code)

hckrman101 (Migrated from github.com) reviewed

2026-11-11 02:20:20 +00:00

hckrman101 (Migrated from github.com) left a comment

Thanks, the before/after captures make the symptom clear. I think I found the root cause, though, and it's a one-line build fix rather than a workaround in the renderer.

The real bug is a misspelled shader define. scripts/fsr3_shaders.py compiles the interpolation passes with -DFFX_FRAMEINTERPOLATION_OPTION_LOW_RESOLUTION_MOTION_VECTORS=1, but the vendored SDK checks a different name: ffx_frameinterpolation_reconstruct_dilated_velocity_and_previous_depth.h:106 has #if FFX_FRAMEINTERPOLATION_OPTION_LOW_RES_MOTION_VECTORS (it's also LOW_RES_ in AMD's CMakeCompileFrameinterpolationShaders.txt). The upscaler's option really is ..._LOW_RESOLUTION_..., which is probably where the mix-up came from. Because of that, fi_reconstruct_and_dilate is built with the display-resolution motion branch. That branch reads the input motion at ComputeHrPosFromLrPos(iPxLrPos), i.e. render pixel × display/render, out of a render-size motion image. That's exactly the "one of its passes reads a render-size input in the picture's pixels" the PR describes. Blitting depth and motion up to the display size and setting render_width = uw makes that mapping the identity, which explains why the screenshots come out sharp. It works around the miscompiled pass, though, and doesn't fix it.

Fixing the define (and regenerating fi_reconstruct_and_dilate's _spv.h, the only pass that reads it) should give the same result without the costs this PR adds:

every FSR "render-size" pass (reconstruct/dilate, setup, previous depth, game vector field and its pyramid, disocclusion) and their internal images now run at the full picture size: at DLSS Performance into 4K that's 4x the pixels for those passes;
two more display-size images (~66 MB at 4K) plus two full-frame blits and two full barriers each made frame;
a RX_VK_FSR_FI_RENDER_SIZE env var and a readme note for what is really a build bug.

Smaller things, if you do keep this approach:

If make_img fails, fi_scaled.w stays 0, so every following frame does device_wait_idle_locked() and tries the allocation again. The other scaled images in this file work the same way, so it's minor, but a one-shot "failed" flag would avoid stalling every frame.
The vk_fsr3fi.h comment on Inputs names the same wrong macro and should change either way.

No overlap with the rest of the series that I can see: #1944 touches vk_fsr3fi.cpp/vk_framegen.cpp but not this path, and #1940–#1942 don't touch fsr3fi_frame. No CI has run on this branch yet.

Could you try the define fix (keeping the render-size inputs) with the same DLSS Balanced to 3840x2160 setup? If it's as sharp as your "after" shot, the PR can shrink to the script change, the regenerated header and maybe a test that checks the define against the name in the SDK header.

Thanks, the before/after captures make the symptom clear. I think I found the root cause, though, and it's a one-line build fix rather than a workaround in the renderer. **The real bug is a misspelled shader define.** `scripts/fsr3_shaders.py` compiles the interpolation passes with `-DFFX_FRAMEINTERPOLATION_OPTION_LOW_RESOLUTION_MOTION_VECTORS=1`, but the vendored SDK checks a different name: `ffx_frameinterpolation_reconstruct_dilated_velocity_and_previous_depth.h:106` has `#if FFX_FRAMEINTERPOLATION_OPTION_LOW_RES_MOTION_VECTORS` (it's also `LOW_RES_` in AMD's `CMakeCompileFrameinterpolationShaders.txt`). The upscaler's option really is `..._LOW_RESOLUTION_...`, which is probably where the mix-up came from. Because of that, `fi_reconstruct_and_dilate` is built with the display-resolution motion branch. That branch reads the input motion at `ComputeHrPosFromLrPos(iPxLrPos)`, i.e. render pixel × display/render, out of a render-size motion image. That's exactly the "one of its passes reads a render-size input in the picture's pixels" the PR describes. Blitting depth and motion up to the display size and setting `render_width = uw` makes that mapping the identity, which explains why the screenshots come out sharp. It works around the miscompiled pass, though, and doesn't fix it. Fixing the define (and regenerating `fi_reconstruct_and_dilate`'s `_spv.h`, the only pass that reads it) should give the same result without the costs this PR adds: - every FSR "render-size" pass (reconstruct/dilate, setup, previous depth, game vector field and its pyramid, disocclusion) and their internal images now run at the full picture size: at DLSS Performance into 4K that's 4x the pixels for those passes; - two more display-size images (~66 MB at 4K) plus two full-frame blits and two full barriers each made frame; - a `RX_VK_FSR_FI_RENDER_SIZE` env var and a readme note for what is really a build bug. Smaller things, if you do keep this approach: - If `make_img` fails, `fi_scaled.w` stays 0, so every following frame does `device_wait_idle_locked()` and tries the allocation again. The other scaled images in this file work the same way, so it's minor, but a one-shot "failed" flag would avoid stalling every frame. - The `vk_fsr3fi.h` comment on `Inputs` names the same wrong macro and should change either way. No overlap with the rest of the series that I can see: #1944 touches `vk_fsr3fi.cpp`/`vk_framegen.cpp` but not this path, and #1940–#1942 don't touch `fsr3fi_frame`. No CI has run on this branch yet. Could you try the define fix (keeping the render-size inputs) with the same DLSS Balanced to 3840x2160 setup? If it's as sharp as your "after" shot, the PR can shrink to the script change, the regenerated header and maybe a test that checks the define against the name in the SDK header.

gpu/vk/vk_framegen.cpp

This works because, with render_width == width, the miscompiled reconstruct-and-dilate pass's ComputeHrPosFromLrPos becomes the identity. The pass itself is built with the wrong option name (LOW_RESOLUTION_ vs the SDK's LOW_RES_, see the review body). I'd fix that in scripts/fsr3_shaders.py and drop this block, which runs all of FSR's render-size passes at display size.

This works because, with `render_width == width`, the miscompiled reconstruct-and-dilate pass's `ComputeHrPosFromLrPos` becomes the identity. The pass itself is built with the wrong option name (`LOW_RESOLUTION_` vs the SDK's `LOW_RES_`, see the review body). I'd fix that in `scripts/fsr3_shaders.py` and drop this block, which runs all of FSR's render-size passes at display size.

j64esatxzdrhfg commented

2026-11-11 10:39:29 +00:00

(Migrated from github.com)

Thanks, that was the bug. I replaced the workaround with the define fix and kept the render-size inputs:

scripts/fsr3_shaders.py now passes FFX_FRAMEINTERPOLATION_OPTION_LOW_RES_MOTION_VECTORS.
fsr3_fi_reconstruct_and_dilate_comp_spv.h is regenerated with glslang 16.6.0. The other 27 headers recompile byte-identical, apart from the version comment.
The Inputs comment in vk_fsr3fi.h is fixed.
tests/test_fsr3_shader_options.py checks every -D the script passes against the #ifs in the SDK files the passes compile from. It fails on main with exactly this name.

The blits, RX_VK_FSR_FI_RENDER_SIZE, the readme note and the fi_scaled retry are gone with it.

Checked in the game with the same setup (DLSS Balanced to 3840x2160, 2x, RX_VK_FSR_FG=1): FSR 3.1 made 508 of the last 600 made frames from the render-size inputs, and the top-left corner is as sharp as the earlier "after" shot. The screenshots in the description are now of this build, with a crop of the corner next to main's.

Thanks, that was the bug. I replaced the workaround with the define fix and kept the render-size inputs: - `scripts/fsr3_shaders.py` now passes `FFX_FRAMEINTERPOLATION_OPTION_LOW_RES_MOTION_VECTORS`. - `fsr3_fi_reconstruct_and_dilate_comp_spv.h` is regenerated with glslang 16.6.0. The other 27 headers recompile byte-identical, apart from the version comment. - The `Inputs` comment in `vk_fsr3fi.h` is fixed. - `tests/test_fsr3_shader_options.py` checks every `-D` the script passes against the `#if`s in the SDK files the passes compile from. It fails on main with exactly this name. The blits, `RX_VK_FSR_FI_RENDER_SIZE`, the readme note and the `fi_scaled` retry are gone with it. Checked in the game with the same setup (DLSS Balanced to 3840x2160, 2x, `RX_VK_FSR_FG=1`): FSR 3.1 made 508 of the last 600 made frames from the render-size inputs, and the top-left corner is as sharp as the earlier "after" shot. The screenshots in the description are now of this build, with a crop of the corner next to main's.

j64esatxzdrhfg (Migrated from github.com) reviewed

2026-11-11 10:39:30 +00:00

gpu/vk/vk_framegen.cpp

Dropped. The script now uses the SDK's LOW_RES_ name, so FSR keeps the render-size inputs.

Dropped. The script now uses the SDK's `LOW_RES_` name, so FSR keeps the render-size inputs.

hckrman101 (Migrated from github.com) reviewed

2026-11-11 12:03:16 +00:00

hckrman101 (Migrated from github.com) left a comment

Thanks, this is exactly the fix I hoped for. All of my earlier points are addressed:

The define is now the SDK's FFX_FRAMEINTERPOLATION_OPTION_LOW_RES_MOTION_VECTORS, and FSR keeps the render-size inputs. The display-size blits, RX_VK_FSR_FI_RENDER_SIZE, the readme note and the fi_scaled retry are all gone (the inline thread is answered).
The Inputs comment in vk_fsr3fi.h is fixed.
tests/test_fsr3_shader_options.py is a good guard. I checked it against the vendored SDK: every -D in the three option lists is read by an #if in the compiled GLSL/headers, and the CMake list's extra FI names (*_SAMPLERS_USE_DATA_HALF, JITTER_MOTION_VECTORS, ...) don't show up because only .h/.glsl files are scanned. It needs no RUN_ALL_* declarations, so run_all.py picks it up as is.
Only fi_reconstruct_and_dilate's header changes, which matches the SDK: that's the only pass with a LOW_RES_ #if. The new in-game check and the corner crops look convincing.

Nit (optional): compiled_files() scans every GPU header for each family, so a name from one family passed to another (say an FFX_FRAMEINTERPOLATION_OPTION_* in the upscaler's list) would still pass the first test. Narrowing each family to its own subfolder plus the shared headers would catch that too. Fine to leave as is.

CI: only the title checks have run. The build and test workflows (Linux AppImage, game-free source regressions, Windows builder) are waiting for a maintainer to approve them. The branch merges cleanly.

A question, not something this PR needs: would you be up for adding FSR 4 frame generation, here or in a follow-up? AMD's SDK v2.4.0, the release --fetch-fsr-runtime pins (#1866), also ships amd_fidelityfx_framegeneration_vk.dll with "FSR Frame Generation (ML)" 4.0.2 on Vulkan 1.2+. According to AMD's docs it's Windows 11 and RX 9000 (RDNA 4) only; other GPUs get 3.1.x. The catch is the reason main made FSR 3.1 interpolation opt-in (#1635: the HUD moves with the scene), and #1831 says FSR 4's frame generation misbehaved the same way in play. AMD's answer to that is a HUD-less colour input or a separate UI surface. So it's only worth doing if it can be fed a HUD-less picture and avoids that problem. If you know a route that does, it'd be very welcome.

Thanks, this is exactly the fix I hoped for. All of my earlier points are addressed: - The define is now the SDK's `FFX_FRAMEINTERPOLATION_OPTION_LOW_RES_MOTION_VECTORS`, and FSR keeps the render-size inputs. The display-size blits, `RX_VK_FSR_FI_RENDER_SIZE`, the readme note and the `fi_scaled` retry are all gone (the inline thread is answered). - The `Inputs` comment in `vk_fsr3fi.h` is fixed. - `tests/test_fsr3_shader_options.py` is a good guard. I checked it against the vendored SDK: every `-D` in the three option lists is read by an `#if` in the compiled GLSL/headers, and the CMake list's extra FI names (`*_SAMPLERS_USE_DATA_HALF`, `JITTER_MOTION_VECTORS`, ...) don't show up because only `.h`/`.glsl` files are scanned. It needs no `RUN_ALL_*` declarations, so `run_all.py` picks it up as is. - Only `fi_reconstruct_and_dilate`'s header changes, which matches the SDK: that's the only pass with a `LOW_RES_` `#if`. The new in-game check and the corner crops look convincing. Nit (optional): `compiled_files()` scans every GPU header for each family, so a name from one family passed to another (say an `FFX_FRAMEINTERPOLATION_OPTION_*` in the upscaler's list) would still pass the first test. Narrowing each family to its own subfolder plus the shared headers would catch that too. Fine to leave as is. CI: only the title checks have run. The build and test workflows (Linux AppImage, game-free source regressions, Windows builder) are waiting for a maintainer to approve them. The branch merges cleanly. A question, not something this PR needs: would you be up for adding FSR 4 frame generation, here or in a follow-up? AMD's SDK v2.4.0, the release `--fetch-fsr-runtime` pins (#1866), also ships `amd_fidelityfx_framegeneration_vk.dll` with "FSR Frame Generation (ML)" 4.0.2 on Vulkan 1.2+. According to AMD's docs it's Windows 11 and RX 9000 (RDNA 4) only; other GPUs get 3.1.x. The catch is the reason main made FSR 3.1 interpolation opt-in (#1635: the HUD moves with the scene), and #1831 says FSR 4's frame generation misbehaved the same way in play. AMD's answer to that is a HUD-less colour input or a separate UI surface. So it's only worth doing if it can be fed a HUD-less picture and avoids that problem. If you know a route that does, it'd be very welcome.

jarynclouatre commented

2026-11-11 13:26:42 +00:00

(Migrated from github.com)

Same fix tested on a 4070 Ti SUPER (Windows, 2x, 2048x1152 from a 67% render): on main the right third of every made frame is the game's frame unmoved; with the define fixed the whole frame moves, as at 100% render. The soft HUD in made frames is unchanged, at 100% too.

j64esatxzdrhfg commented

2026-11-11 15:07:26 +00:00

(Migrated from github.com)

Thanks! Took the nit: compiled_files() now follows each family's includes from its pass GLSL (beside the including file, then the GPU headers' folder and its family folders, as the script's -I paths do), so each family is checked against its own files plus the shared ones (gpu/, spd/, and fsr1/ for the upscaler). An upscaler option put into FI_OPTIONS now fails the first test; main's LOW_RESOLUTION_ spelling still fails all three. Rebased onto current main.

Thanks! Took the nit: `compiled_files()` now follows each family's includes from its pass GLSL (beside the including file, then the GPU headers' folder and its family folders, as the script's `-I` paths do), so each family is checked against its own files plus the shared ones (`gpu/`, `spd/`, and `fsr1/` for the upscaler). An upscaler option put into `FI_OPTIONS` now fails the first test; main's `LOW_RESOLUTION_` spelling still fails all three. Rebased onto current main.

j64esatxzdrhfg commented

2026-11-11 15:07:28 +00:00

(Migrated from github.com)

@jarynclouatre Thanks for confirming it on a 4070 Ti SUPER at a 67% render, that's a good second data point. The soft HUD in made frames is a separate thing: OoT draws its HUD into the same picture FSR 3.1 interpolates, so it moves with the scene's motion (the readme's reason this path is opt-in), at any render size.

mooki000 commented

2026-11-11 16:45:29 +00:00

(Migrated from github.com)

Helps on Linux as well, which the description lists as not tested: on main, made frames in motion have an edge at two thirds of the width and height, and with this change they move as a whole.

Build: Arch Linux, RTX 4090, NVIDIA 615.78.08, Hyprland (Wayland), a 1920x1080 window at 60 Hz. Main bdfe09c against bdfe09c with this change on top. I built 93a25f2f; the gpu/ and scripts/ change of the current head 0a043192 has the same patch id. Between bdfe09c and your base 49f75dc the frame generation code differs only by an HDR image usage flag.

Scene: output 3840x2160, render resolution 67 % (2560x1440), FSR 3 as the upscaler, frame generation 2x, RX_VK_FSR_FG=1. Kokiri Forest from a save, Link running while the camera turns (input script). Made frames captured with RX_VK_SHOT_MADE=1 at 1920x1080, one every 120 game frames. Both builds log fsr3fi: 3840x2160 from 2560x1440 inputs, flow 480x270, pyramid 1920x1080 with 11 levels and 595 of the last 600 made frames by FSR 3.1, 0 without its inputs.

Run	Build	What	Start	Load before -> after	Pairs from present 1680 (in motion)	Edge on x=1280, made / game frame (median)	Edge on y=720, made / game frame (median)
wr-fi-base	`bdfe09c`	main, Link and camera still	07:46	106.15 -> 127.07	8 (0)	- (none in motion)	- (none in motion)
wr-fim-base	`bdfe09c`	main, camera turning	09:32	11.01 -> 7.72	26 (17)	1.42 / 1.04	1.47 / 1.00
wr-fim-1943	3625f7fc	#1943, camera turning	08:01	73.28 -> 107.78	10 (10)	1.03 / 1.00	0.98 / 1.02

"Edge" is how much the picture changes across the line x=1280 (top 720 rows) or y=720 (left 1280 columns), divided by the change across the lines next to it: 1.00 means the line looks like its neighbours. A pair is in motion when the made frame differs from the game frame after it by more than 2 of 255 luma levels on average. On main the made frames have an edge on both lines that the game's frames do not have; with this change they do not.

Main: a made frame while the camera turns, with an edge at the red marks (two thirds of the width).
Main: the game's frame after it, with no edge there.
Main: a made frame while Link runs. The left two thirds are smeared, the right third is not.
This PR: a made frame while Link runs. The whole picture moves together, with no edge.

New test at 0a043192: tests/test_fsr3_shader_options.py passes (3 tests). With main's scripts/fsr3_shaders.py, 3 of 3 fail ('-DFFX_FRAMEINTERPOLATION_OPTION_LOW_RES_MOTION_VECTORS=1' not found in [...]). At 93a25f2f, scripts/fsr3_shaders.py --check with glslang 16.6.0 printed fsr3 shaders: 28 match, and the header is the same at 0a043192.

Not tested: DLSS as the upscaler (this build has no DLSS), frame generation Auto, Windows, AMD, and the whole tests/run_all.py. The two runs in motion were not taken at the same load (11 to 8 on main, 73 to 108 with this PR), so I compare only the edges and give no frame rates.

Helps on Linux as well, which the description lists as not tested: on main, made frames in motion have an edge at two thirds of the width and height, and with this change they move as a whole. **Build:** Arch Linux, RTX 4090, NVIDIA 615.78.08, Hyprland (Wayland), a 1920x1080 window at 60 Hz. Main bdfe09c against bdfe09c with this change on top. I built 93a25f2f; the `gpu/` and `scripts/` change of the current head 0a043192 has the same patch id. Between bdfe09c and your base 49f75dc the frame generation code differs only by an HDR image usage flag. **Scene:** output 3840x2160, render resolution 67 % (2560x1440), FSR 3 as the upscaler, frame generation 2x, `RX_VK_FSR_FG=1`. Kokiri Forest from a save, Link running while the camera turns (input script). Made frames captured with `RX_VK_SHOT_MADE=1` at 1920x1080, one every 120 game frames. Both builds log `fsr3fi: 3840x2160 from 2560x1440 inputs, flow 480x270, pyramid 1920x1080 with 11 levels` and `595 of the last 600 made frames by FSR 3.1, 0 without its inputs`. | Run | Build | What | Start | Load before -> after | Pairs from present 1680 (in motion) | Edge on x=1280, made / game frame (median) | Edge on y=720, made / game frame (median) | |---|---|---|---|---|---|---|---| | wr-fi-base | bdfe09c | main, Link and camera still | 07:46 | 106.15 -> 127.07 | 8 (0) | - (none in motion) | - (none in motion) | | wr-fim-base | bdfe09c | main, camera turning | 09:32 | 11.01 -> 7.72 | 26 (17) | 1.42 / 1.04 | 1.47 / 1.00 | | wr-fim-1943 | 3625f7fc | #1943, camera turning | 08:01 | 73.28 -> 107.78 | 10 (10) | 1.03 / 1.00 | 0.98 / 1.02 | "Edge" is how much the picture changes across the line x=1280 (top 720 rows) or y=720 (left 1280 columns), divided by the change across the lines next to it: 1.00 means the line looks like its neighbours. A pair is in motion when the made frame differs from the game frame after it by more than 2 of 255 luma levels on average. On main the made frames have an edge on both lines that the game's frames do not have; with this change they do not. ![Made frames on main and with this PR, the two-thirds lines marked](https://gist.githubusercontent.com/mooki000/b86bd3f35969803fb032f71f035844f8/raw/ee9a2d54d35f405b52a2f6304418290995f71fa7/hookshot-fsr3fi-made-frames.jpg) 1. Main: a made frame while the camera turns, with an edge at the red marks (two thirds of the width). 2. Main: the game's frame after it, with no edge there. 3. Main: a made frame while Link runs. The left two thirds are smeared, the right third is not. 4. This PR: a made frame while Link runs. The whole picture moves together, with no edge. **New test at 0a043192:** `tests/test_fsr3_shader_options.py` passes (3 tests). With main's `scripts/fsr3_shaders.py`, 3 of 3 fail (`'-DFFX_FRAMEINTERPOLATION_OPTION_LOW_RES_MOTION_VECTORS=1' not found in [...]`). At 93a25f2f, `scripts/fsr3_shaders.py --check` with glslang 16.6.0 printed `fsr3 shaders: 28 match`, and the header is the same at 0a043192. **Not tested:** DLSS as the upscaler (this build has no DLSS), frame generation Auto, Windows, AMD, and the whole `tests/run_all.py`. The two runs in motion were not taken at the same load (11 to 8 on main, 73 to 108 with this PR), so I compare only the edges and give no frame rates.

hckrman101 (Migrated from github.com) reviewed

2026-11-11 16:50:00 +00:00

hckrman101 (Migrated from github.com) left a comment

Thanks for taking the nit. compiled_files() now follows each family's includes from its own pass GLSL, so an option passed to the wrong family fails the first test. I read it against the vendored SDK: the lookup order matches the script's -I paths closely enough (it also searches the other family folders, but no header name repeats across them, so it can't pick a wrong file), and a header it can't find just goes unscanned, which can only make the test stricter, never let a wrong name through. Since 93a25f2, the regenerated SPIR-V header, vk_fsr3fi.h and scripts/fsr3_shaders.py are byte-identical; only the test changed, plus the rebase.

Two more in-game confirmations since then, both showing the same thing: @jarynclouatre on a 4070 Ti SUPER on Windows at a 67% render, and @mooki000 on Linux (RTX 4090, Wayland) with a measured edge at two thirds of the frame on main that goes away with this PR. So the "not tested on Linux" in the description is covered now. The soft HUD in made frames is the known #1635 behaviour, as you said, not something this PR changes.

CI: only the title check has run. The build and test workflows (game-free source regressions, Linux AppImage, Windows builder) are waiting for a maintainer to approve them. The branch merges cleanly.

On my earlier question about FSR 4 frame generation: I didn't see an answer. No pressure, and it isn't needed for this PR. If you're interested in a follow-up, it would need a HUD-less colour input or a separate UI layer to get around the #1635 problem.

Thanks for taking the nit. `compiled_files()` now follows each family's includes from its own pass GLSL, so an option passed to the wrong family fails the first test. I read it against the vendored SDK: the lookup order matches the script's `-I` paths closely enough (it also searches the other family folders, but no header name repeats across them, so it can't pick a wrong file), and a header it can't find just goes unscanned, which can only make the test stricter, never let a wrong name through. Since 93a25f2, the regenerated SPIR-V header, `vk_fsr3fi.h` and `scripts/fsr3_shaders.py` are byte-identical; only the test changed, plus the rebase. Two more in-game confirmations since then, both showing the same thing: @jarynclouatre on a 4070 Ti SUPER on Windows at a 67% render, and @mooki000 on Linux (RTX 4090, Wayland) with a measured edge at two thirds of the frame on main that goes away with this PR. So the "not tested on Linux" in the description is covered now. The soft HUD in made frames is the known #1635 behaviour, as you said, not something this PR changes. CI: only the title check has run. The build and test workflows (game-free source regressions, Linux AppImage, Windows builder) are waiting for a maintainer to approve them. The branch merges cleanly. On my earlier question about FSR 4 frame generation: I didn't see an answer. No pressure, and it isn't needed for this PR. If you're interested in a follow-up, it would need a HUD-less colour input or a separate UI layer to get around the #1635 problem.

hckrman101 (Migrated from github.com) approved these changes

2026-11-11 19:14:18 +00:00

hckrman101 (Migrated from github.com) left a comment

Approving. Nothing has changed since my last look at 0a04319. The low-res motion-vector option now uses the SDK's own name, the test follows each family's includes, and jarynclouatre (Windows, 4070 Ti SUPER) and mooki000 (Linux, RTX 4090) both confirmed the seam is gone in made frames. The soft HUD is the known #1635 behaviour, not this PR. The FSR 4 frame generation question isn't needed here, so that can be a separate follow-up.

Only the title check has run on this head so far, so the build and test workflows need approving and a green run before merging.

Approving. Nothing has changed since my last look at 0a04319. The low-res motion-vector option now uses the SDK's own name, the test follows each family's includes, and jarynclouatre (Windows, 4070 Ti SUPER) and mooki000 (Linux, RTX 4090) both confirmed the seam is gone in made frames. The soft HUD is the known #1635 behaviour, not this PR. The FSR 4 frame generation question isn't needed here, so that can be a separate follow-up. Only the title check has run on this head so far, so the build and test workflows need approving and a green run before merging.