This is an automated email from the git hooks/post-receive script.
git pushed a commit to branch span-gl-clean
in repository efl.
View the commit online.
commit c5bceadff0559422c9f204bc7a7ee76b2123f26b
Author: [email protected] <[email protected]>
AuthorDate: Tue May 5 11:14:05 2026 -0600
docs(evas_ector_gl): explain why span pipe-flush quad cap is latent
The 18-vertex / 3-quad cap in span_shader_pipe_flush() is intentional
but cannot be profitably raised due to the merge predicate in
evas_gl_common_context_span_push(). Every VG shape occupies its own
atlas tile, making span_fill_off_tx/ty unique per shape. The merge
predicate requires all ~40 per-shape uniforms to match, so distinct
shapes never coalesce in practice—array.num is always 6 (one quad) at
flush time.
The ndc[18*2] buffer and nverts clamp are latent safety guards for a
merge scenario that cannot occur with current per-shape atlas allocation.
Raising the cap would waste stack and reduce zero Mesa uniform-stream
rebuilds (since no merges happen). Two actual paths to coalescing are
documented as future work: (a) move atlas offsets into per-vertex
attributes so merge can match on shader identity alone, or (b) use
instanced rendering with pre-sorted shape batches.
This closes the V3D / Pi 4 perf investigation with a documented "no
action needed at this layer" conclusion so future contributors don't
re-investigate.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
---
.../engines/gl_generic/evas_ector_gl_span_shader.c | 33 ++++++++++++++++++++--
1 file changed, 30 insertions(+), 3 deletions(-)
diff --git a/src/modules/evas/engines/gl_generic/evas_ector_gl_span_shader.c b/src/modules/evas/engines/gl_generic/evas_ector_gl_span_shader.c
index 211375cd2f..3574d31bca 100644
--- a/src/modules/evas/engines/gl_generic/evas_ector_gl_span_shader.c
+++ b/src/modules/evas/engines/gl_generic/evas_ector_gl_span_shader.c
@@ -1261,9 +1261,36 @@ span_shader_pipe_flush(Evas_Engine_GL_Context *gc, int pipe_idx, int gw, int gh)
uint32_t mul_col = gc->pipe[pipe_idx].shader.span_mul_col;
int max_spans = gc->pipe[pipe_idx].shader.span_max_spans;
int vi, nverts = n;
- /* 18 vertices = max 3 merged quads (6 verts each). In practice
- * span_push creates one quad per push, but the pipe merge logic
- * can accumulate up to 3 quads in a single pipe entry. */
+ /* 18 vertices = 3 quads (6 verts each) — the design intent was to allow
+ * up to 3 same-uniform pushes to coalesce into one draw call.
+ *
+ * WHY THIS CAP IS INTENTIONAL AND CANNOT BE PROFITABLY RAISED:
+ *
+ * The merge predicate in evas_gl_common_context_span_push()
+ * (gl_common/evas_gl_context.c:2109–2157) requires ALL ~40 per-shape
+ * uniform fields to be bit-identical before two pushes may share a pipe
+ * entry. Every distinct VG shape occupies its own tile in the span-buffer
+ * atlas, so span_fill_off_tx / span_fill_off_ty differ for every shape.
+ * Additionally, gradient parameters (grad_a … grad_rdy) and actual_max_spans
+ * are shape-specific. Consequently no two distinct SHD_SPAN pushes ever
+ * pass the merge predicate: each push creates a fresh pipe entry with
+ * exactly 6 vertices. array.num is therefore always 6 when flush is called.
+ *
+ * The ndc[18*2] buffer and the "nverts > 18" clamp below are therefore
+ * latent safety guards for a merge scenario that cannot occur with the
+ * current per-shape atlas allocation scheme. Raising the cap to, say,
+ * 16 quads would cost 96*2*4 = 768 bytes of extra stack and reduce the
+ * maximum V3D uniform-stream rebuilds by zero (no merges happen).
+ *
+ * To actually reduce per-draw overhead on V3D (Cortex-A72 + Mesa V3D):
+ * 1. Relax the merge predicate to allow pushes that differ only in
+ * geometry (x/y/w/h) but share all uniform state — requires moving
+ * the span-buffer atlas offsets into per-vertex attributes rather
+ * than uniforms, which is a non-trivial shader rework.
+ * 2. Alternatively, pre-sort VG shapes so those sharing the same
+ * fill/stroke parameters are batched, then share a single atlas tile
+ * via instanced rendering.
+ * Until one of those changes lands, keep this at 18 (3 quads). */
GLfloat ndc[18 * 2];
GLfloat *dst;
--
To stop receiving notification emails like this one, please contact
the administrator of this repository.
