From 2ab53299097ed4ef255672407ac4b463a9a91bc4 Mon Sep 17 00:00:00 2001 From: Valiantsin Dzerakh Date: Sun, 15 Nov 2026 08:54:22 +0300 Subject: [PATCH] The NVN probe's Cpu gets the FP result check threshold too Its static Cpu ran guest code with a zeroed threshold, which lets +-smallest normal pass the check unexamined. Co-Authored-By: Claude Opus 5.5 --- runtime/nvn_probe.c | 1 + 1 file changed, 1 insertion(+) diff --git a/runtime/nvn_probe.c b/runtime/nvn_probe.c index 8431d61a..a3996aed 100644 --- a/runtime/nvn_probe.c +++ b/runtime/nvn_probe.c @@ -919,6 +919,7 @@ int probe_nvn(void) { if (child && *child) return run_as_child(child); // The probe checks the null back end whatever TOTK_GPU says; the choice is made once, at the first object call. ((RtConfig *)rt_config())->gpu = "null"; + rt_fpcr_threshold(&cpu); // the FP result check threshold of a fresh Cpu (cpu.h): a zeroed one is wrong, not slow nvn_procs_register(); hostfn_seal(); kern_run_on_guest_main(probe_main);