Entry 48

Translucent models showed their hidden layers: FB_ONLY drew colour before Z

2026-09-23rendertest

Filescrates/piney-gs/src/lib.rscrates/piney-gs/src/convert.rs

The user reported that transparency on the desktop looked wrong. 47 ruled out the title's lighting. The cause was in the GPU renderer: the way it drew an alpha test whose failing pixels keep their colour but not their Z (AFAIL = FB_ONLY).

Finding it

Earlier comparisons ran the GPU and the CPU GS model (piney_desktop::soft) in separate runs. This time one desktop frame went through both renderers:

• Frame 545. The frame is taken after the cursor has moved. 1,520 pixels differed, all in the icon column: the icons were near white on the GPU and translucent on the CPU.
• Narrowing it down. Drawing one model at a time left six icon models, all translucent (mmat alpha 0.75, 0.5, 0.2).
• One pixel of one icon. On the CPU, the icon's nearer layer (Z 5591) drew 0xff at alpha 96 (191 over black), and a farther layer (Z 5587) failed the Z test. The GPU gave 239: that farther layer blended on top.

The cause

The desktop's models draw with the alpha test GEQUAL (the mmat's reference, usually 0) and FB_ONLY, with Z writes on.

• The GS. It writes each pixel's Z as it draws. A model's nearer layer therefore hides the layers drawn after it, back faces included.
• The GPU. It split every FB_ONLY draw into two passes: first colour for the whole draw with Z writes off, then a Z-only pass. In the colour pass nothing had written Z yet, so every layer of the model passed the test and blended in.

The fix

• Pass order. An FB_ONLY draw now draws its passing pixels first, with colour and Z, in drawing order as the GS does. A second pass adds the failing pixels' colour without Z (a new Params bit, 19, discards the passing pixels there).
• No split when nothing can fail. A test that cannot fail (ALWAYS, or GEQUAL 0) is drawn as one pass.

The result differs from the GS only where a failing pixel meets a passing one drawn after it in the same draw.

Checked

• One desktop frame through both renderers. Pixels that differ between the GPU and the CPU GS model, before and after the fix:

  frame	before	after
  100	not taken	2
  400	not taken	18
  545	1,520	87

  The remaining pixels are single pixels on triangle edges and one row at the selection bar.

• The GS tests. fb_only_hides_later_farther_layers draws a nearer sprite, then a farther one at alpha 0x60, in one draw. The farther one is hidden: 0xbf over black, as one layer gives. piney-gs has 13 tests, all passing, and clippy and fmt are clean.