A visual record of the in-house Direct3D 11 renderer plugin (gl11) being built up from
nothing to match the legacy Direct3D 9 reference (gl05), one milestone at a time.
Remember the days of just staring at a half-magenta / half-black screen? 😄 The magenta was the fallback pixel shader firing on geometry that was drawing correctly — we just couldn't shade it yet. Everything below is the climb out of that hole.
The client boots on the modern toolchain and renders real geometry: character select and a Tatooine exterior. No correct shading yet, but the scene graph, meshes, and camera work.
| Character select | Tatooine |
|---|---|
 |
 |
First honest side-by-side of the two renderers on the same spots — the gap analysis that drove the rest of the work (lighting, gamma, texel seams).
| D3D9 reference (interior) | D3D11 (interior) |
|---|---|
 |
 |
The big one: real asset pixel shaders (PSRC recompile lane) driving the character, 9/9 binds, parity with the D3D9 reference. This is where D3D11 stopped being a toy.
| D3D9 reference | D3D11 (asset-PS lane) |
|---|---|
 |
 |
Whole-world depth buffering, per-vertex Gouraud lighting (tfcl/tfcsl reauthoring), and a
generic vertex-shader fallback so legacy //asm programs draw at all (skippedNullVS
38927 → 0). Interiors and sky finally render.
A D3D11-only skinning regression: a few skinned-mesh vertices on robes/dresses/hands flung to garbage positions. The screenshots below are the diagnostic — each one a deliberate experiment that moved us toward the root cause.
The bug vs the reference:
| D3D9 — clean robe (target) | D3D11 — the spikes |
|---|---|
 |
 |
The debugging arc:
| Experiment | What it proved |
|---|---|
 |
Zeroing the dynamic-VB ring on discard sent the bad verts to object origin (the feet/floor) — proving they read undefined ring memory, not random corruption. |
 |
Forcing WRITE_DISCARD every lock largely killed the spikes but dropped whole meshes (stormtrooper hands) — proving the append model is required: meshes coexist at different ring offsets, drawn deferred. |
 |
The D3D9-faithful ms_used accounting fix took the per-frame ring wraps from 370 → 0, but unmasked a deeper deferred-draw/rename coupling (full-screen flicker). → CONSULT-20. |
Root cause: the D3D11 dynamic-VB ring charges ms_used by the upper-bound locked
length in lock(), while D3D9 charges the actual used count in unlock(). One UI lock
(CuiLayerRenderer, which locks the entire remaining ring) maxes ms_used → the shared ring
renames dozens of times per frame → deferred-draw skinned meshes read a stale rename → spikes.
Capture-invisible because RenderDoc serializes the CPU/GPU and the rename race vanishes.
The arc: magenta fallback geometry → correct shading → world depth + lighting → character parity → chasing the last skinning hazard. Months of incremental wins, captured frame by frame.