[j64esatxzdrhfg]

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 feat(hdr): HDR contrast row, also grading the game's native HDR #1941, fix(hdr): Native HDR scaled in 10 bits and shown as the game made it at 100% contrast #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 fix(hdr): Sharpness with HDR, and DLSS for the game's native HDR #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

[Joey-Magnus]

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, a0dfd28 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 a0dfd28 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 a0dfd28, 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]

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.

[j64esatxzdrhfg]

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.