diff options
Diffstat (limited to 'docs/04-implementation/04-real-render-walls.md')
| -rw-r--r-- | docs/04-implementation/04-real-render-walls.md | 34 |
1 files changed, 34 insertions, 0 deletions
diff --git a/docs/04-implementation/04-real-render-walls.md b/docs/04-implementation/04-real-render-walls.md index 74fca80..3cd370a 100644 --- a/docs/04-implementation/04-real-render-walls.md +++ b/docs/04-implementation/04-real-render-walls.md @@ -944,3 +944,37 @@ Three durable findings from the r1414-r1427 experiment series (host module unchanged (GATE-1 still black); the lane still poisons (`badf1201` -> WEDGE-RECOVER) after ~60 s, and the stores land on relocated-operand pages rather than the render target. + +## Session 154b (2026-11-03): the SPH wall - draw executes, DS rejects at SPH validation + +Follow-up state after the first store retirements, three further levers: + +1. **The macro "run bit" is a completion latch, not an arming gate.** The + engine's macro-status sampler reads back `0x4044b4`; a completed + slot-34 body latches `mac=0x22 mpc=637` with `run=0` (637 = the body's + last instruction). The earlier belief that the DRAW never ran was a + sampler artefact - the slot-34 body is a short straight-line program the + transition-gated logger never caught mid-flight. +2. **`sema_gl_sph_bind` (pipeline-stage bind emission) was 0 on every boot + to r1433 and is the bind lever.** Enabling it emits the per-window + pipeline-stage binds (`SPH-BIND emitted ... win-binds=2 j1-type=0x11 + initblk=17`), zeroes the previous SPH11 signature, and produces the first + observed slot-34 execution (`mac=0x22 run=1`) with live draw-surface + retirements (`retirement wait: 3/2/1 of 901`, OUT-PAGE-WIT on + 0x120700000-0x120705000 = draw-surface rows 614-618, operand/uniform + values). +3. **The residual wall is the DS stage validation.** DS-ESR decode: + bits 23:0 = SPH<n>_ERR, bits 29:24 = pipeline stage. `a2040800` = + SPH11|SPH18; with binds on, `a2040a00` = SPH9 (the vertex SPH fetched + from the window's own j=1 bind alias is rejected; the DS `region=` + readback is 0/0). Variants: `sema_sph_version_val=0x00040003` flips the + SPH9 attribution but not the wall; `sema_sph_region_legacy=1` emits the + legacy 0x1608/0x160c pair (base 0x121201000) but is refused at the method + level (`Cl 0000c597 Off 00001608`); `sema_sph_region_base=0x01500000` + propagates but the launch is refused by an MME class error + (`Cl 0000902d Off 0x2000`) before the DS stage is reached. + +Current best lineage: `golden=1 golden_force=1 hold_calls=1` + +IRAM-gate break/build 0 + `sa_rearm=0` + `sph_bind=1` (binds + draw +execution + draw-surface retirements; still no rasterized pixels - the +region record for the synthesized stage slots remains the named gap). |