From 8e91c95d13a857b1b386fa391716b33d3fe5ef1b Mon Sep 17 00:00:00 2001 From: chrg6 Date: Tue, 10 Nov 2026 03:27:14 -0500 Subject: [PATCH] Record what the scripted runs showed about the DLSS release rule, and that the model may matter and not the size. Co-Authored-By: Claude Fable 5.2 (cherry picked from commit bf734de7dcbe20a1a582b82e0b00492af26b3580) --- runtime/vk_dlaa.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/runtime/vk_dlaa.c b/runtime/vk_dlaa.c index 3d02dae6..aa70e8b0 100644 --- a/runtime/vk_dlaa.c +++ b/runtime/vk_dlaa.c @@ -522,8 +522,11 @@ static int image2d(uint32_t w, uint32_t h, VkFormat f, VkImageUsageFlags usage, // Giving the kept feature its own sizes again at each evaluation changed nothing. A feature made while a smaller one // is alive leaves the smaller one as it was (the pause screen's after the gameplay one, at every first pause). So when // a feature is made, the alive ones with a larger input are released first, and after a release since the last -// creation all of them (release_kept); their images stay and each is made again when its size is next evaluated. With -// the second half alone, none of 5 runs in which the gameplay feature had been freed was striped. +// creation all of them (release_kept); their images stay and each is made again when its size is next evaluated. +// Scripted on one build, each way with the rule and with 2: the long pause and the mode change were striped without it +// and clean with it. The input size may not be what matters: in every striped case the feature made was of another +// model (performance's or ultra performance's) than the pause screen's, and the one run in which balanced's feature, +// of the pause screen's model, was made again beside it without the rule was clean. The rule covers both readings. // 1 (default) idle features are freed // 0 idle features are never freed (about 0.5 GB more video memory after the first pause) // 2 idle features are freed and nothing is released when a feature is made (as before; shows the fault)