mirror of
https://github.com/RetroPortingToolKit/psxrecomp
synced 2026-10-28 10:59:16 -04:00
Alexandros Mandravillis
57 lines
3.0 KiB
Markdown
57 lines
3.0 KiB
Markdown
# Interlaced primitive drawing
|
|
|
|
The GPU preserves one row parity while drawing in 480-line interlaced mode
|
|
when GP0(E1h) prohibits drawing to the display area. The display origin and
|
|
current field select that parity. Progressive modes and the draw-to-display
|
|
override draw both parities. Fills, copies and VRAM transfers keep both fields.
|
|
|
|
`gpu_raster_skipped_row()` reads current GPU state for each primitive. The
|
|
renderer facade applies the field mask with native-row clip rectangles and
|
|
restores the original clip. This also preserves the corresponding rows in
|
|
scaled surfaces. Each clipped triangle retains its precision and perspective
|
|
metadata, which hardware backends otherwise consume on the first submission.
|
|
The facade counts each original perspective triangle once; repeated backend
|
|
metadata setup does not add diagnostic hits. No title address or title detection
|
|
selects this behavior.
|
|
|
|
Pro Pinball: Timeshock! USA SLUS-00639 repeatedly draws and reads a test pixel
|
|
at (1020,511) during startup. Frozen source 63446a28 always returned the drawn
|
|
pixel and stayed black. This correction lets the guest's existing field test
|
|
finish. The same disc, BIOS and settings then reach the table.
|
|
|
|
The field latch uses the runtime's existing vertical-blank model. This change
|
|
does not claim scanline-accurate timing. Row clipping can increase submissions
|
|
for large interlaced primitives; no general performance claim is made.
|
|
|
|
## Verification
|
|
|
|
Configure `runtime/` with `BUILD_TESTING=ON`, then run
|
|
`ctest --test-dir <build> -R gpu_interlace_render_test --output-on-failure`.
|
|
The fixture covers the mode predicate, origin/field parity, all nine primitive
|
|
families, clip restoration, transfers, mode changes, scaled output and
|
|
triangle metadata consumed by a backend. The original renderer fails the
|
|
same row-preservation assertions.
|
|
|
|
## Prior art
|
|
|
|
- [PSX-SPX GPU specification](https://psx-spx.consoledev.net/graphicsprocessingunitgpu/)
|
|
describes interlace affecting draw commands.
|
|
- [DuckStation difficult games](https://github.com/stenzek/duckstation/wiki/Difficult-to-Emulate-Games)
|
|
identifies required line skipping for Pro Pinball startup.
|
|
|
|
These sources supplied behavioral evidence. The implementation and regression
|
|
fixture were written independently. Exact title run receipts and route limits
|
|
are maintained by the Timeshock correction project.
|
|
|
|
Precise vertices use signed 16.16 positions. Row dispatch covers the union of
|
|
integer fallback bounds and precise vertex bounds, with one row for rounding.
|
|
Tests compare native and enhanced masked pixels with unmasked images when
|
|
precise positions cross the integer bounds, including negative coordinates.
|
|
A tiny precise triangle must submit three row clips rather than 256.
|
|
|
|
Flat overlay rectangles also submit one-row geometry while field masking is
|
|
active. This keeps the mask when a wide backend intentionally replaces its
|
|
scissor. Non-interlaced overlays retain the original geometry and policy.
|
|
A backend fixture that ignores the scissor tests both field parities, original
|
|
progressive behavior, and clip restoration. This is not Vulkan driver testing.
|