diff options
Diffstat (limited to 'docs/04-implementation')
| -rw-r--r-- | docs/04-implementation/04-real-render-walls.md | 33 |
1 files changed, 33 insertions, 0 deletions
diff --git a/docs/04-implementation/04-real-render-walls.md b/docs/04-implementation/04-real-render-walls.md index ec1b75f..cc16b04 100644 --- a/docs/04-implementation/04-real-render-walls.md +++ b/docs/04-implementation/04-real-render-walls.md @@ -1181,3 +1181,36 @@ vertices) and the window capture is black (`xwd -id`, non-black 0). The remaining work is to capture the constant-buffer *bulk* transfer itself (the staging buffer to the vidmem CB page), not the method-stream patches. + +## Session 155 cont 8 (2026-11-06): DS SPH bitmap semantics and the BEGIN/END placement wall + +Two results worth keeping: + +1. **The DS SPH bitmap is `stage*4+check` over the six GF100 pipeline stages** + (VSA, VSB, TCS, TES, GS, PS). The six per-stage SPH start registers + (`DS_HWW_ESR_SPH_INFO_STAGE0..5`, 0x405c00..0x405c14) and the live + `DS-FAST` stage VAs plus the fetched SPH header types prove the stage + index: stage1 0x…900 is a V4T1 `SHADER_TYPE=1` VTG image, stage2 0x…200 a + V4T1 `TYPE=2` tess-init image, stage5 0x…0 a V4T2 `TYPE=5` pixel image. + The four checks per stage are the four SPH-V4 validation groups: header + type/version, mpipe/stage binding, attribute map (IMAP/OMAP + register + count + stream formats), and stream-out/topology. A clean boot with only + VSB/PS bound reported `a0040a06` = SPH1|2 (VSA checks 1/2), SPH9|11 (TCS + checks 1/3), SPH18 (GS check2) - i.e. the unbound stages are the failures. + Enabling the TCS bind added SPH10 inside the TCS group (r1554c), which is + the live proof of the `stage*4+check` split. + +2. **Synthesizing the missing VSA/GS stage records clears those bits but + their method run lands inside BEGIN/END.** Binding VSA/GS to the client's + own VTG image (`mask=0x35`) moved the bitmap from `a0040a06` to + `900400a0` (SPH1|2|9|11 gone; SPH5|7|18 remain), then the next window + raised `Class Error Offset 0x2018 Data 019a0900 EC 0x0d` - + `ILLEGAL_METHOD_IN_BEGIN_END`. The value is accepted; the placement is + the wall (the synth block is prepended to a window whose previous END was + dropped). Next: make the stage-record emission BEGIN/END-safe (emit after + an explicit END when the FE begin state is still open) and keep the + client's own TYPE/REGISTER_COUNT-before-address order. + +`SET_SPH_VERSION`/`CHECK_SPH_VERSION` (0x16a4/0x16a8) is not a lever for +this bitmap: blob-derived, `0x10004` and `0x40004` all produce the same DS +value on the same boot. |