MagicBeanDev/Hookshot

Archived

Watch

Fork 0

feat(dlss): DLSS for the game's native HDR, given linear light #2243

Closed

j64esatxzdrhfg wants to merge 3 commits from j64esatxzdrhfg/feat/dlss-native-hdr into main

j64esatxzdrhfg commented

2026-11-11 10:40:02 +00:00

(Migrated from github.com)

Builds on #1942 (which builds on #1941): this PR is the last commit. It is the DLSS-for-Native part that was split out of #1940, reworked as reviewed there: keyed on what the game presents, and given linear light.

What changed

vk_dlss.cpp, vk_temporal.cpp/.h: while present_native_pq says the game presents its own PQ picture, DLSS's feature is made with IsHDR | AutoExposure and made again when that changes. The gate is the presented picture, not the option and an HDR10 swapchain; with Native picked and an SDR picture shown, the scene is LDR. The engine reports it through TemporalEngine::hdr; FSR 4 has none and keeps the LDR pair. present_native_pq moves out of vk_replay.cpp's anonymous namespace for that.
The chain then runs a second pair of its passes:
- dlss_prep.comp's kSquare (constant 1) squares the square-root encoded scene into linear light, which NVIDIA documents for IsHDR and from which AutoExposure works out its exposure.
- Its kColorMax (constant 0) is 16384 in place of 1, well inside the colour image's RGBA16F.
- dlss_out.comp's kSquare (constant 0) is off for it: the network's picture is linear already and is written as it is, at the same output scale as the LDR pair.
shaders/dlss_prep_comp_spv.h, dlss_out_comp_spv.h: glslang 16.6.0, --target-env vulkan1.3 (the command tests/test_dlss_actor.py checks them with).
RX_VK_DLSS_HDR=0 keeps the LDR feature and the clamped colour with Native too, for comparisons (docs/development.md).
Tests: test_hdr_encode.py checks the two constants in each shader, the HDR pipelines and their specialization data, that temporal_upscale binds the HDR pair for an engine whose hdr() says so, and that prepare keys the feature on present_native_pq (and the switch), not on the option.

Testing

In game: RTX 5080, 3840x2160 HDR10, 800 nits peak, 330 nits paper white, Native at 100% contrast, DLSS Balanced, in a build of main (d7a7306) with #1941, #1942 and this PR. Same spot, the Deku Tree's light shaft:
- With RX_VK_DLSS_HDR=0, as on main, the brightest light was 328 nits, which is the paper white. The 99.9th percentile was 328.
- With this change it reached 749 nits (99.9th percentile 654), and the light's structure is back.
- The log says HDR colour, linear (the game's native HDR) when the feature is made.
- This answers the review's question on #1940 about the largest scene value: in this scene the game's grade does go well past the paper white once DLSS stops clamping.
vk_dlss.cpp compiles with and without NGX; the touched files compile with the build's clang. tests/run_all.py with hdr_encode, dlss_actor, dlss_checks, upscale_independent, framegen_resources, present_, temporal_history, status_line and menu pass, except recorder_sync and present_semaphores, which fail the same way on main (this PC's toolchain).

Not tested: other NVIDIA generations, DLAA with Native, displays with another peak.

Screenshots (Native at 100%, DLSS Balanced; HDR10 captures tone-mapped to SDR for this page, paper white to white, highlights rolled off up to 800 nits)

Before (`RX_VK_DLSS_HDR=0`, as on main): 328 nits at most	After: 749 nits at most

The hole of light at full size, before and after:

🤖 Generated with Claude Code

With HDR on Native and DLSS as the upscaler, every highlight is cut at the paper white: DLSS's feature is made for LDR colour and the chain's inputs pass clamps the scene to 1 before DLSS sees it. The hole of light in the Deku Tree turns into a flat disc. > Builds on #1942 (which builds on #1941): this PR is the last commit. It is the DLSS-for-Native part that was split out of #1940, reworked as reviewed there: keyed on what the game presents, and given linear light. **What changed** - `vk_dlss.cpp`, `vk_temporal.cpp/.h`: while `present_native_pq` says the game presents its own PQ picture, DLSS's feature is made with `IsHDR | AutoExposure` and made again when that changes. The gate is the presented picture, not the option and an HDR10 swapchain; with Native picked and an SDR picture shown, the scene is LDR. The engine reports it through `TemporalEngine::hdr`; FSR 4 has none and keeps the LDR pair. `present_native_pq` moves out of `vk_replay.cpp`'s anonymous namespace for that. - The chain then runs a second pair of its passes: - `dlss_prep.comp`'s `kSquare` (constant 1) squares the square-root encoded scene into linear light, which NVIDIA documents for `IsHDR` and from which `AutoExposure` works out its exposure. - Its `kColorMax` (constant 0) is 16384 in place of 1, well inside the colour image's RGBA16F. - `dlss_out.comp`'s `kSquare` (constant 0) is off for it: the network's picture is linear already and is written as it is, at the same output scale as the LDR pair. - `shaders/dlss_prep_comp_spv.h`, `dlss_out_comp_spv.h`: glslang 16.6.0, `--target-env vulkan1.3` (the command `tests/test_dlss_actor.py` checks them with). - `RX_VK_DLSS_HDR=0` keeps the LDR feature and the clamped colour with Native too, for comparisons (docs/development.md). - Tests: `test_hdr_encode.py` checks the two constants in each shader, the HDR pipelines and their specialization data, that `temporal_upscale` binds the HDR pair for an engine whose `hdr()` says so, and that `prepare` keys the feature on `present_native_pq` (and the switch), not on the option. **Testing** - In game: RTX 5080, 3840x2160 HDR10, 800 nits peak, 330 nits paper white, Native at 100% contrast, DLSS Balanced, in a build of main (d7a7306) with #1941, #1942 and this PR. Same spot, the Deku Tree's light shaft: - With `RX_VK_DLSS_HDR=0`, as on main, the brightest light was **328 nits**, which is the paper white. The 99.9th percentile was 328. - With this change it reached **749 nits** (99.9th percentile 654), and the light's structure is back. - The log says `HDR colour, linear (the game's native HDR)` when the feature is made. - This answers the review's question on #1940 about the largest scene value: in this scene the game's grade does go well past the paper white once DLSS stops clamping. - `vk_dlss.cpp` compiles with and without NGX; the touched files compile with the build's clang. `tests/run_all.py` with `hdr_encode`, `dlss_actor`, `dlss_checks`, `upscale_independent`, `framegen_resources`, `present_`, `temporal_history`, `status_line` and `menu` pass, except `recorder_sync` and `present_semaphores`, which fail the same way on main (this PC's toolchain). **Not tested:** other NVIDIA generations, DLAA with Native, displays with another peak. **Screenshots** (Native at 100%, DLSS Balanced; HDR10 captures tone-mapped to SDR for this page, paper white to white, highlights rolled off up to 800 nits) | Before (`RX_VK_DLSS_HDR=0`, as on main): 328 nits at most | After: 749 nits at most | |---|---| | ![DLSS with Native HDR before: the light clipped at the paper white](https://raw.githubusercontent.com/j64esatxzdrhfg/Hookshot/pr-screenshots/dlss-native/before.jpg) | ![DLSS with Native HDR after: the light graded up to 749 nits](https://raw.githubusercontent.com/j64esatxzdrhfg/Hookshot/pr-screenshots/dlss-native/after.jpg) | The hole of light at full size, before and after: ![The light shaft: a flat disc before, its structure after](https://raw.githubusercontent.com/j64esatxzdrhfg/Hookshot/pr-screenshots/dlss-native/light-before-after.jpg) 🤖 Generated with [Claude Code](https://claude.com/claude-code)

Joey-Magnus (Migrated from github.com) reviewed

2026-11-11 20:07:25 +00:00

Joey-Magnus (Migrated from github.com) left a comment

I think this one is already merged, by another route. #2342 (FSR 4 for the game's native HDR) was stacked on this PR, and its squash merge, a0dfd28b on main (11 Nov), carried the whole stack: #1941, #1942 and this commit. I checked at this PR's head against main:

the shaders and their SPIR-V headers (dlss_prep, dlss_out, hdr_encode, fg_scale, the new fg_scale_pq_comp_spv.h) and vk_temporal.h are byte-identical on main;
the changes this PR makes to vk_dlss.cpp, vk_temporal.cpp, vk_replay.cpp and vk_framegen.cpp are exactly the ones a0dfd28b made;
tests/test_hdr_encode.py differs only by #2342's own check_fsr4 additions;
RX_VK_DLSS_HDR=0 and the present_native_pq gate in prepare are on main (gpu/vk/vk_dlss.cpp:308-310).

That is also why GitHub shows conflicts in vk_replay.cpp, docs/development.md and tests/test_hdr_encode.py: main has the same lines plus what came after. Unless there is something in this branch I missed that isn't in a0dfd28b, this PR can be closed as merged through #2342, and #1941 and #1942 probably too (worth a quick check of each before closing). If a follow-up is still needed, a fresh branch from main is easier than rebasing this stack.

One note for the changelog: the release notes take their line from the squash title, so the DLSS part shipped under feat(fsr4): .... If DLSS for native HDR should appear in the release notes, it needs a line added there by hand.

I think this one is already merged, by another route. #2342 (FSR 4 for the game's native HDR) was stacked on this PR, and its squash merge, a0dfd28b on main (11 Nov), carried the whole stack: #1941, #1942 and this commit. I checked at this PR's head against main: - the shaders and their SPIR-V headers (`dlss_prep`, `dlss_out`, `hdr_encode`, `fg_scale`, the new `fg_scale_pq_comp_spv.h`) and `vk_temporal.h` are byte-identical on main; - the changes this PR makes to `vk_dlss.cpp`, `vk_temporal.cpp`, `vk_replay.cpp` and `vk_framegen.cpp` are exactly the ones a0dfd28b made; - `tests/test_hdr_encode.py` differs only by #2342's own `check_fsr4` additions; - `RX_VK_DLSS_HDR=0` and the `present_native_pq` gate in `prepare` are on main (`gpu/vk/vk_dlss.cpp:308-310`). That is also why GitHub shows conflicts in `vk_replay.cpp`, `docs/development.md` and `tests/test_hdr_encode.py`: main has the same lines plus what came after. Unless there is something in this branch I missed that isn't in a0dfd28b, this PR can be closed as merged through #2342, and #1941 and #1942 probably too (worth a quick check of each before closing). If a follow-up is still needed, a fresh branch from main is easier than rebasing this stack. One note for the changelog: the release notes take their line from the squash title, so the DLSS part shipped under `feat(fsr4): ...`. If DLSS for native HDR should appear in the release notes, it needs a line added there by hand.

Joey-Magnus (Migrated from github.com) reviewed

2026-11-11 21:24:29 +00:00

Joey-Magnus (Migrated from github.com) left a comment

This looks already merged: #2342's squash (a0dfd28b) shipped it. All nine shader and SPIR-V header files and vk_temporal.h are identical to main, and a test merge leaves no difference in the files that merged cleanly. I think this can be closed. Since it went out under a feat(fsr4) title, DLSS with native HDR probably needs a hand-written line in the next release notes.

This looks already merged: #2342's squash (`a0dfd28b`) shipped it. All nine shader and SPIR-V header files and `vk_temporal.h` are identical to main, and a test merge leaves no difference in the files that merged cleanly. I think this can be closed. Since it went out under a `feat(fsr4)` title, DLSS with native HDR probably needs a hand-written line in the next release notes.

hckrman101 commented

2026-11-11 22:55:46 +00:00

(Migrated from github.com)

This now conflicts with main in docs/development.md, gpu/vk/shaders/dlss_prep.comp, gpu/vk/shaders/dlss_prep_comp_spv.h, gpu/vk/vk_menu.cpp, gpu/vk/vk_menu.h, gpu/vk/vk_replay.cpp, gpu/vk/vk_temporal.cpp, gpu/vk/vk_temporal.h, tests/test_hdr_encode.py. Could you merge main into the branch (or rebase) when you get a chance? I'll take another look once it's up to date.

This now conflicts with main in `docs/development.md`, `gpu/vk/shaders/dlss_prep.comp`, `gpu/vk/shaders/dlss_prep_comp_spv.h`, `gpu/vk/vk_menu.cpp`, `gpu/vk/vk_menu.h`, `gpu/vk/vk_replay.cpp`, `gpu/vk/vk_temporal.cpp`, `gpu/vk/vk_temporal.h`, `tests/test_hdr_encode.py`. Could you merge main into the branch (or rebase) when you get a chance? I'll take another look once it's up to date.

github-sync referenced this pull request

2026-11-12 08:30:51 +00:00

fix(vk): support PsychoV35 HDR with DLSS and DLAA #2707

github-sync referenced this pull request

2026-11-12 09:32:06 +00:00

feat(hdr): HDR contrast row, also grading the game's native HDR #1941

github-sync closed this pull request

2026-11-12 09:32:10 +00:00

github-sync commented

2026-11-12 09:32:12 +00:00

Owner

@j64esatxzdrhfg commented on GitHub · 2026-11-12 08:33 UTC · original

Closing as landed: #2342's squash merge (a0dfd28) carried this commit, RX_VK_DLSS_HDR included. Thanks @Joey-Magnus for checking. Agreed on the release notes: since it went out under a feat(fsr4) title, "DLSS with Native HDR gets the game's HDR colour as linear light (highlights no longer clipped at the paper white)" deserves its own line.

> [@j64esatxzdrhfg](https://github.com/j64esatxzdrhfg) commented on GitHub · 2026-11-12 08:33 UTC · [original](https://github.com/NotRexed/Hookshot/pull/2243#issuecomment-6818775868) Closing as landed: #2342's squash merge (a0dfd28) carried this commit, `RX_VK_DLSS_HDR` included. Thanks @Joey-Magnus for checking. Agreed on the release notes: since it went out under a `feat(fsr4)` title, "DLSS with Native HDR gets the game's HDR colour as linear light (highlights no longer clipped at the paper white)" deserves its own line.

Commenting is not possible because the repository is archived.

No reviewers