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 5a67a378ce44f02c2ab6be906892722154a5bd3e
Author: [email protected] <[email protected]>
AuthorDate: Tue May 5 11:09:38 2026 -0600
perf(evas_ector_gl): use glClear+scissor for FBO sub-region clear
Replaces the SHD_RECT pipe-clear-then-immediate-flush with direct
glClear(GL_COLOR_BUFFER_BIT) under glScissor. This avoids tile buffer
inefficiencies on tile-based GPUs (Broadcom V3D 4.2, ARM Mali) where the
previous immediate flush forced a tile store to LPDDR4 plus a tile reload
before each span batch.
On desktop GPUs the change is neutral; on Raspberry Pi 4, the tile-reload
overhead measured around 1 GB/s of additional bandwidth per round-trip. At
6.4 GB/s shared LPDDR4, this is a measurable fraction of the frame budget.
Since glClear is not a pipe entry, no pipe-reorder race exists between
SHD_RECT and SHD_SPAN; we flush once before the clear to drain any prior
pipe content from earlier shapes, then disable GL_SCISSOR_TEST.
The Evas GL state cache fields (clip, cx/cy/cw/ch, cur_tex, render_op)
are explicitly invalidated so the next pipe flush re-issues the correct
glEnable/glDisable(GL_SCISSOR_TEST) for clipped pipes. Without this, a
subsequent clipped pipe would desynchronize the cache: the pipe-flush loop
decides whether to enable/disable scissor by comparing shader.clip against
gc->state.current.clip, and a stale cache value would skip the enable branch,
silently rendering with scissor disabled.
Verification: Expedite VG benchmarks produce byte-identical PNG output vs
baseline; ector 13/0/0; evas 120/0/0.
Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
---
src/modules/evas/engines/gl_generic/evas_engine.c | 44 +++++++++++++++--------
1 file changed, 29 insertions(+), 15 deletions(-)
diff --git a/src/modules/evas/engines/gl_generic/evas_engine.c b/src/modules/evas/engines/gl_generic/evas_engine.c
index d53ef00746..18d9e71a8e 100644
--- a/src/modules/evas/engines/gl_generic/evas_engine.c
+++ b/src/modules/evas/engines/gl_generic/evas_engine.c
@@ -3111,21 +3111,35 @@ eng_ector_end(void *engine,
int ox = glim->tex ? glim->tex->x : 0;
int oy = glim->tex ? glim->tex->y : 0;
- /* Clear FBO sub-region to transparent. Flush immediately
- * so the clear is executed before any span draws — otherwise
- * the SHD_RECT pipe and SHD_SPAN pipes may be reordered
- * during batch flush, causing the clear to overwrite the
- * span content on some frames. */
- {
- RGBA_Draw_Context *dc_save = gc->dc;
- gc->dc = evas_common_draw_context_new();
- evas_common_draw_context_set_color(gc->dc, 0, 0, 0, 0);
- evas_common_draw_context_set_render_op(gc->dc, EVAS_RENDER_COPY);
- evas_gl_common_rect_draw(gc, 0, 0, w, h);
- evas_common_draw_context_free(gc->dc);
- gc->dc = dc_save;
- evas_gl_common_context_flush(gc);
- }
+ /* Clear FBO sub-region to transparent via glClear + scissor.
+ *
+ * On tile-based GPUs (Broadcom V3D / ARM Mali) glClear is
+ * a tile-buffer flag — no main-memory traffic — whereas the
+ * previous SHD_RECT + immediate flush forced a tile store
+ * and reload before the span draws. The clear happens
+ * outside the pipe system, so no pipe-reorder races; we
+ * still flush any pending pipe content first to be safe.
+ *
+ * glScissor uses the same (ox, oy, w, h) as glViewport:
+ * _evas_gl_common_viewport_set already sets the viewport to
+ * (tex->x, tex->y, w, h) for atlas sub-rects, so scissor
+ * coords are in framebuffer space, matching GL bottom-left.
+ *
+ * The Evas GL state cache is invalidated on textures and
+ * render op below so subsequent draws restore them as needed. */
+ evas_gl_common_context_flush(gc);
+ glEnable(GL_SCISSOR_TEST);
+ glScissor(ox, oy, w, h);
+ glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
+ glClear(GL_COLOR_BUFFER_BIT);
+ glDisable(GL_SCISSOR_TEST);
+ gc->state.current.clip = 0;
+ gc->state.current.cx = 0;
+ gc->state.current.cy = 0;
+ gc->state.current.cw = 0;
+ gc->state.current.ch = 0;
+ gc->state.current.cur_tex = -1;
+ gc->state.current.render_op = -1;
/* Per-shape draw loop.
*
--
To stop receiving notification emails like this one, please contact
the administrator of this repository.
