mirror of
https://git.triforceproject.com/triforce/source
synced 2026-11-19 16:20:01 -04:00
Triforce v0.2.0
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
36
docs/presentation-geometry-trace.md
Normal file
36
docs/presentation-geometry-trace.md
Normal file
@@ -0,0 +1,36 @@
|
||||
# Effective presentation geometry diagnostics
|
||||
|
||||
Set `RX_VK_PRESENT_LOG=1` before launching to log presentation geometry changes.
|
||||
Unset it or set it to `0` to disable it. This writes ordinary diagnostic lines;
|
||||
it does not enable frame-log image readback or change rendering/settings.
|
||||
|
||||
Each sample reports texture size, supplied crop request, clipped content size,
|
||||
swapchain output size, the actual destination rectangle, observed host aspect,
|
||||
sampled-view availability, and the route selected by presentation:
|
||||
|
||||
- `direct`: the source blit was selected, including fallback when filtering returned no image.
|
||||
- `scaled`: the scaling function returned an image and its result was selected.
|
||||
- `clear`: presentation had no source and selected its black clear.
|
||||
- `hdr`: the HDR conversion function returned an image selected for presentation.
|
||||
A failed converter falls back to the scaled/direct route. Native PQ input bypasses conversion.
|
||||
|
||||
Filter, sharpness and frame-generation values are labeled `*-request`: configuration
|
||||
does not establish that a feature ran successfully. The route is a CPU command-recording
|
||||
decision, not a successful GPU present result. The host-aspect field reuses the exact
|
||||
aspect passed to the destination fit, without a second callback read. No-source frames
|
||||
report it unavailable, since fitting deliberately does not query the callback then.
|
||||
This diagnostic does not alter aspect policy.
|
||||
An unavailable/invalid aspect is explicit; invalid values are normalized for stable deduplication.
|
||||
Requests accompany geometry/route changes but do not themselves trigger a line, so dragging
|
||||
a sharpness slider cannot consume the geometry-change budget while the effective route stays the same.
|
||||
|
||||
Repeated unchanged samples are suppressed independently for game, interpolated (`made`)
|
||||
and startup/clear frames, so alternating ordinary/interpolated sources does not spam a
|
||||
line per frame. Up to 256 geometry changes are logged per process, followed by one
|
||||
suppression notice. Restart to reset the limit. Continuously changing dynamic resolution
|
||||
can reach that limit; this is a bounded diagnostic, not a full timeline capture.
|
||||
|
||||
`tests/test_present_geometry_trace.py` executes the production presentation-route and
|
||||
trace blocks with CPU command spies. It checks effective crop/rectangle/output/options,
|
||||
resize, direct fallback/scaled/clear routes, independent frame kinds, disabled/zero
|
||||
configuration, invalid aspects and the output bound. No GPU or game data is required.
|
||||
Reference in New Issue
Block a user