summaryrefslogtreecommitdiffstats
path: root/wiki/Project-Plan.md
diff options
context:
space:
mode:
Diffstat (limited to 'wiki/Project-Plan.md')
-rw-r--r--wiki/Project-Plan.md33
1 files changed, 21 insertions, 12 deletions
diff --git a/wiki/Project-Plan.md b/wiki/Project-Plan.md
index 102359f..889ca87 100644
--- a/wiki/Project-Plan.md
+++ b/wiki/Project-Plan.md
@@ -15,7 +15,7 @@
> [`../docs/06-portability/03-device-validation-and-porting-sequence.md`](../docs/06-portability/03-device-validation-and-porting-sequence.md);
> the docs tree wins on disagreement.
-Status: 2026-11-15, tree `master @ 88d5005a`. Tags: `[done]`, `[active]`,
+Status: 2026-11-16, tree `master @ a06d4b13`. Tags: `[done]`, `[active]`,
`[ahead]` — the plan's own state vocabulary, distinct from the goal tags.
## 1. How the plan works
@@ -40,11 +40,11 @@ itself:
## 2. The phases
-| Phase | What it delivers | Exit gate | State (2026-11-15) |
+| Phase | What it delivers | Exit gate | State (2026-11-16) |
|---|---|---|---|
| **P0 — Platform** | Emulated vGPU device, GSP boot and GSP-RM emulation, stock guest driver initialisation | Driver-API smoke green (`ALL OK`) = evaluation rung 2 | `[done]` |
| **P1 — First execution (compute)** | Guest CUDA kernels executing on the physical GPU, with engine-authored completions; then hardening | Verified clients + engine-authored completion = rung 3 core; hardening: staged-transfer class, every-boot acceptance, per-guest isolation | `[active]` — execution green; hardening in progress |
-| **P2 — Graphics execution** | OpenGL/Vulkan draw commands executed by the physical 3D pipeline and delivered as readable pixels | The graphics ladder: M-GL1 client context-init executes; M-GL2 `glxgears` non-black; M-GL3 GZDoom GL non-black through the mdev | `[active]` — between M-GL1 and M-GL2 with M-GL2's off-screen half proven: the display/KMS and DDX replay machinery work, X + stock GLX are green, the first non-black GPU-rendered frames and videos are captured, the X DDX/ScreenInit stability walls are closed, captured GL sections execute on the engine, the direct-GL submission is fully decoded (macro-driven, no classic draw header), the client's draw-state push reaches the engine, the MME pack completes, the released replay executes `DRAW_IAB` (`mac=0x22`, c788; TCS-off DS clean, synthetic viewport, chained release), the DS stage image is authored and live for both vertex slots (`st0=st1=020a0481`), and the saved-window release carrier deterministically replays the held window once the pack is resident; the remaining walls are the vertex-buffer identity, the render-target store/alias visibility, the pack-completion lottery, the stores, the window/screen present, and the per-boot binding/X first-generation start lottery |
+| **P2 — Graphics execution** | OpenGL/Vulkan draw commands executed by the physical 3D pipeline and delivered as readable pixels | The graphics ladder: M-GL1 client context-init executes; M-GL2 `glxgears` non-black; M-GL3 GZDoom GL non-black through the mdev | `[active]` — between M-GL1 and M-GL2 with M-GL2's off-screen half proven: the display/KMS and DDX replay machinery work, X + stock GLX are green, the first non-black GPU-rendered frames and videos are captured, the X DDX/ScreenInit stability walls are closed, captured GL sections execute on the engine, the direct-GL submission is fully decoded (macro-driven, no classic draw header), the released replay executes `DRAW_IAB` (`mac=0x22`, c788), and the 11-16 wave proved the host-side chain deterministic end to end (real frame pushes admitted with a persisted colour-target bind, a 40-chunk full-window replay and a 128-slot ring; a clean DS `DS-FAST2 ds_after=80000000` at the live-derived GS map; a golden 2799-word MME pack; CALL hold/carry fire/SA re-arm with the latch kept across the channel recycle). The decisive finding is guest-side: across 36 boots the client never submitted a non-null `CALL(34)` draw macro, and the same cycles logged guest EGL/dri2 failures and OOM kills; the remaining walls are a real guest draw submission, the vertex/store, the window/screen present, and the per-boot X first-generation start lottery. A GZDoom video was separately delivered through the guest's software GL path (Mesa llvmpipe) — not a vGPU rendering claim |
| **P3 — Sustained and multi-tenant** | Frame rates without assistance; multiple VMs sharing the GPU with measured fairness; soak-stable operation | Graphics M-GL4 (60 fps, no forcing) and M-GL5 (two VMs, two mdevs); CUDA battery + soak; per-guest execution isolation = rungs 4/5 | `[ahead]` |
| **P4 — Breadth** | Video (NVENC/NVDEC), Windows/DirectX guests, live migration, non-KVM hypervisor transports | Per-case bring-up gates plus the matrix's evidence expectations | `[ahead]` |
| **P5 — Scale and portability** | New GPU generations and driver families carried forward; published performance programme | A profile + measured delta per version, validated on hardware; A/B numbers published per the measurement contract | `[ahead]` |
@@ -77,15 +77,24 @@ Ranked, in dependency order:
`MME-DRAW-MOVE` reorder, TCS-off (`45dbcdcd`) clears the DS latch, the
synthetic viewport (`dab4a02c`) and the chained release (`3481e62d`)
complete the replay, and the body decodes with the public Mesa MME
- toolchain (`mme_tu104_sim`: 8 draws of 3 vertices). The named walls are
- the vertex-buffer identity (no captured client push binds the stream; the
- eng71 scan candidate is a heuristic, with eng74 `VS-BLOB` dumping it) and
- the render-target store/alias visibility (the big-surface alias pages do
- not yet enter `GUEST-WIT`; eng74 records them), the
- pack-completion lottery, the stores, the window/screen present path
- (glxgears' window is black and a window draw wedges X), the per-boot
- binding/X first-generation start lottery (even `modeNum` required; now the
- dominant throughput limiter), and then non-black window frames.
+ toolchain (`mme_tu104_sim`: 8 draws of 3 vertices). The 11-16 wave proved the **host-side chain deterministic end to end**:
+real frame pushes admitted (persisted colour-target bind, draw markers inside
+counted runs, a 14.8-thousand-word window replayed as 40 chunks, a 128-slot
+candidate ring), a clean DS on the beef draw lane (`DS-FAST2
+ds_after=80000000`, c963) at the live-derived GS map, a golden 2799-word MME
+pack on capture-less boots, and CALL hold/carry fire/SA re-arm with the
+prelude latch kept across the channel recycle. The **decisive finding is
+guest-side**: across 36 boots (c990-c1019) the client never submitted a
+non-null `CALL(34)` draw macro — every `0x3910/0x3914` hit is an interior
+method write with `data=0x0`, no CALL header exists in the admitted windows,
+and the same cycles logged guest EGL/dri2 failures and OOM kills. The named
+walls are therefore a real guest draw submission first, then the
+vertex/store, the window/screen present path (glxgears' window is black and
+a window draw wedges X), and the per-boot X first-generation start lottery
+(even `modeNum` required; the dominant throughput limiter). A GZDoom video
+was separately delivered through the guest's **software GL** path (Mesa
+llvmpipe) as a guaranteed-quality deliverable; it is not a vGPU rendering
+claim.
Full detail:
[`Graphics-Execution.md`](Graphics-Execution.md) §6.
2. **Compute hardening (P1).** The staged-transfer class (failure modes proven,