diff options
| author | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-10-20 10:09:53 -0400 |
|---|---|---|
| committer | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-10-20 10:09:53 -0400 |
| commit | ed91540a8d9bd3989c1d922db35d7733d6e9fdd9 (patch) | |
| tree | 55caa075067ed80f8d5bd969f1041c68aa74e7b0 /kernel/gsp_phys_gpu.c | |
| parent | cafc460bab4cab32968a6656242f34f2dc392486 (diff) | |
| download | open-iov-ed91540a8d9bd3989c1d922db35d7733d6e9fdd9.tar.gz open-iov-ed91540a8d9bd3989c1d922db35d7733d6e9fdd9.zip | |
gr: single scheduling edge (pcas default 0) - fixes double engine execution; retire-wait tracked-total fix; sema acquire wait gate per-entry + C56F op semantics (opt-in); fill-shape decode diagnostic
Diffstat (limited to 'kernel/gsp_phys_gpu.c')
| -rw-r--r-- | kernel/gsp_phys_gpu.c | 21 |
1 files changed, 14 insertions, 7 deletions
diff --git a/kernel/gsp_phys_gpu.c b/kernel/gsp_phys_gpu.c index e845493..e896664 100644 --- a/kernel/gsp_phys_gpu.c +++ b/kernel/gsp_phys_gpu.c @@ -638,13 +638,20 @@ int gsp_phys_rmcall_gr_qmdvirt_value(void) EXPORT_SYMBOL_GPL(gsp_phys_rmcall_gr_qmdvirt_value); /* - * Explicit-PCAS gate (default 1). The rebuilt launch carries the vendor's - * turingPushPcas(CU_PCAS_TYPE_COPY) pair (SEND_PCAS_A = QMD>>8 followed by - * SEND_SIGNALING_PCAS_B = INVALIDATE|SCHEDULE). The stock guest push does - * not emit these (it relies on the IQ2M window's implied copy PCAS); setting - * this to 0 emits the pure vendor window-only form for the A/B test. - */ -static int gsp_phys_rmcall_gr_pcas_param = 1; + * Explicit-PCAS gate (default 0 since r343 - the double-execution fix). + * The rebuilt launch used to carry the vendor's turingPushPcas + * (CU_PCAS_TYPE_COPY) pair (SEND_PCAS_A = QMD>>8 followed by + * SEND_SIGNALING_PCAS_B = INVALIDATE|SCHEDULE) ADDED to the inline-QMD run + * that the stock guest push already carries. The QMD's + * REQUIRE_SCHEDULING_PCAS bit is clear, so the inline run schedules by + * itself - the explicit pair was a SECOND scheduling edge and the same QMD + * was dispatched twice (launch_count: one atomicAdd +100 landed as +200; + * victory_reduce exactly 2x; scale/saxpy/memset2d doubled). The stock guest + * push has NO PCAS pair (captured push dumps); default 0 restores that + * single-edge form. 1 = re-add the pair (A/B debug only - with it on, the + * grid executes twice). + */ +static int gsp_phys_rmcall_gr_pcas_param; module_param_named(gsp_phys_rmcall_gr_pcas, gsp_phys_rmcall_gr_pcas_param, int, 0644); MODULE_PARM_DESC(gsp_phys_rmcall_gr_pcas, |