VRAM: RT mask gradients are fp32 where fp16 suffices (32 MB); two DLSS-RR inputs written but unused (16 MB) #1433

Closed
opened 2026-11-04 06:03:36 +00:00 by sky_line · 1 comment
Owner

The RT mask grad and hist-grad images are rgba32f although the comment (llcleanrenderervk.cpp ~34028) and the shader header say fp16; their flags are exact ints below 2048. Change the C++ format and both shaders (31.6 MB at 4K). The DLSS-RR inputs mInReactive and mInTc are commented "(written, unused)" (llcleandlssrrvk.cpp ~202-203): stop allocating/writing them (15.8 MB). Medium-high / high confidence.

From the 4K VRAM audit (2026-11-04) prompted by #1422 (RTX 4070 Ti SUPER at 3840x2089 with RT + GI: the degrade ladder sat at rung 21/45, capping textures, mesh detail 1/4 and draw distance 35% while lighting kept full cost). Sizes are at 4K unless noted; COMPUTED = from formats in code, MEASURED = from logs.

The RT mask grad and hist-grad images are rgba32f although the comment (llcleanrenderervk.cpp ~34028) and the shader header say fp16; their flags are exact ints below 2048. Change the C++ format and both shaders (31.6 MB at 4K). The DLSS-RR inputs mInReactive and mInTc are commented "(written, unused)" (llcleandlssrrvk.cpp ~202-203): stop allocating/writing them (15.8 MB). Medium-high / high confidence. From the 4K VRAM audit (2026-11-04) prompted by #1422 (RTX 4070 Ti SUPER at 3840x2089 with RT + GI: the degrade ladder sat at rung 21/45, capping textures, mesh detail 1/4 and draw distance 35% while lighting kept full cost). Sizes are at 4K unless noted; COMPUTED = from formats in code, MEASURED = from logs.
sky_line added this to the Slipstream project 2026-11-04 06:03:36 +00:00
Collaborator

Both parts are already settled on main (2026-11-04), so nothing is left to do here:

  • DLSS-RR inputs: done in d5342b74b7. The reactive / T&C images are no longer allocated or written for DLSS (a 1x1 stub fills the bindings); FSR3 still writes its own.
  • fp16 grad / hist-grad: ruled out in c5abdc102d. The w channel packs two 12-bit timers (up to 2^24-1), and fp16 only holds integers exactly to 2048, so the images stay rgba32f. The two old comments that still say rgba16f (llcleanrenderervk.h ~590, llcleanrenderervk.cpp ~34581) are what the ticket was based on.

Closing.

Both parts are already settled on main (2026-11-04), so nothing is left to do here: - DLSS-RR inputs: done in d5342b74b75. The reactive / T&C images are no longer allocated or written for DLSS (a 1x1 stub fills the bindings); FSR3 still writes its own. - fp16 grad / hist-grad: ruled out in c5abdc102d3. The w channel packs two 12-bit timers (up to 2^24-1), and fp16 only holds integers exactly to 2048, so the images stay rgba32f. The two old comments that still say rgba16f (llcleanrenderervk.h ~590, llcleanrenderervk.cpp ~34581) are what the ticket was based on. Closing.
TheSlowly removed their assignment 2026-11-06 04:33:52 +00:00
TheSlowly 2026-11-06 04:33:52 +00:00
Sign in to join this conversation.
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
sky_line/Slipstream#1433
No description provided.