diff options
| -rw-r--r-- | STATE-CURRENT.md | 39 |
1 files changed, 21 insertions, 18 deletions
diff --git a/STATE-CURRENT.md b/STATE-CURRENT.md index d81de74..954d1d9 100644 --- a/STATE-CURRENT.md +++ b/STATE-CURRENT.md @@ -1,23 +1,26 @@ # Open-IOV Current State -> **Snapshot: 2026-10-28 ~06:00 (physical-engine guest CUDA at 10 -> consecutive green clients on a fresh boot — `victory_mail`, -> `victory_matmul`, `victory_saxpy`, `victory_strided`, `victory_math`, -> `victory_reduce`, `victory_atomics`, `multikernel`, `vector_add` — with -> launch p50 ~650 ms below the recorded 684 ms baseline; the held-chain -> stall class root-caused and fixed (replay passes re-marking -> `op_tr_pending`, re-arming a 3 s hold that stalled the copy channel); the -> boot-to-boot variance root-caused (a per-client argument-base publication -> race: green boots resolve each client's cudaMalloc base via the -> FB-aperture directory walk, red boots return `none`/`rc=-2` on every -> registered directory and the launch fails closed); the vidmem-window fill -> decoder and the fail-closed defer-cap landed; GL now runs on the guest GPU -> (VASpace construction and 3D-engine selection walls fixed; next wall = the -> FBO class); the Looking Glass frame path is proven end-to-end (guest CUDA -> render -> KVMFR ivshmem -> host PNG frame); two guests with 512M + 1GiB -> emulated VRAM coexist at boot level; the 0.9.9 release-preview demo video -> v1 is recorded. Frontier = the remaining boot variance + streams classes, -> two-guest concurrent CUDA, GZDOOM (GL/Vulkan), Windows guests).** +> **Snapshot: 2026-10-28 ~23:30 (the guest-corruption panic class is +> root-caused and fixed — an r254-era ledger parity store clobbered the +> per-page copy landings and wrote client result data into reused guest +> pages, killing the guest kernel in the nvidia channel teardown; r627 +> guards it, five consecutive boots clean, and the best battery yet is +> 11/15 with zero hangs (mail, matmul, saxpy, strided, math, reduce, +> atomics, multikernel, vector_add, smiley, inplace_scale green). Launch +> p50 ~650 ms (below the 684 ms baseline); the held-chain stall class +> fixed; a refused launch now surfaces a CUDA error instead of spinning +> forever. r632 serves unresolved operand pages from the VA shadow lane to +> close the red-draw chain (walk=none -> H2D shadow -> fill fail-close -> +> degenerate refusal -> synthetic fault); the remaining top blockers are +> the release-target resolver (`release0 could not target the guest page` +> with pins present) and the deterministic wrongness class (streams2 / +> streams_events / memset2d / roundtrip). GL runs on the guest GPU (next +> wall = the FBO class); the Looking Glass frame path is proven end-to-end +> (guest CUDA render -> KVMFR ivshmem -> host PNG frame); two guests with +> 512M + 1GiB emulated VRAM coexist at boot level; the 0.9.9 release-preview +> demo video v1 is recorded and synced to files.vgpu.io. Frontier = +> release-target + wrongness + repeat-run degradation, then two-guest +> concurrent CUDA, GZDOOM (GL/Vulkan), Windows guests).** > This is a dated public > status page, not live state; it is refreshed on every progress check. Live > working notes are maintained outside this repository; the public |