vtgl

Commit: 6f9a8436cf0cc98b4c2c944cc4c2124b69507270
Parent: 60fa5e65f199607d3edda83635fc75f7906201c7
Author: Randy Palamar
Date:   Wed, 23 Oct 2024 21:48:47 -0600

remove nasty selection check in fb_clear_region

vis likes to emit CSI that calls this and it was quite expensive
largely due to the stupid branching by checking the selection for
every cell. Checking beforehand with a proper inclusion check
saves about 50% of the cycles. It is still roughly 10x slower than
it should be but this is because of the tricky pointer business.
It may need to be rexplored later.

Diffstat:
M
terminal.c
 | 
37
+++++++++++++++++++++++++++++--------
M
util.h
 | 
2
+-
M
vtgl.c
 | 
11
+++++++++--

3 files changed, 39 insertions(+), 11 deletions(-)
diff --git a/terminal.c b/terminal.c
@@ -149,6 +149,21 @@ is_selected(Selection *s, i32 x, i32 y)
 	return result;
 }
 
+static b32
+selection_intersects_region(Selection *s, uv2 tl, uv2 br)
+{
+	/* TODO: maybe this can be further simplified (eg. with a k-map) */
+	Range r = s->range;
+	b32 valid   = is_valid_range(r);
+	b32 whole   = r.start.y <  tl.y && r.end.y   >  br.y;
+	b32 start_x = r.start.x >= tl.x && r.start.x <= br.x;
+	b32 start_y = r.start.y >= tl.y && r.start.y <= br.y;
+	b32 end_x   = r.end.x   >= tl.x && r.end.x   <= br.x;
+	b32 end_y   = r.end.y   >= tl.y && r.end.y   <= br.y;
+	b32 result  = valid && (whole || (start_y && start_x) || (end_y && end_x));
+	return result;
+}
+
 static void
 fb_clear_region(Term *t, u32 r1, u32 r2, u32 c1, u32 c2)
 {
@@ -169,16 +184,20 @@ fb_clear_region(Term *t, u32 r1, u32 r2, u32 c1, u32 c2)
 	CLAMP(r1, 0, t->size.h - 1);
 	CLAMP(r2, 0, t->size.h - 1);
 
-	TermView *tv = t->views + t->view_idx;
-	for (u32 r = r1; r <= r2; r++) {
-		for (u32 c = c1; c <= c2; c++) {
-			tv->fb.rows[r][c].style = t->cursor.style;
-			tv->fb.rows[r][c].cp    = ' ';
-			/* TODO: this shouldn't be in this loop; make a region intersection test */
-			if (is_selected(&t->selection, c, r))
-				selection_clear(&t->selection);
+	uv2 top_left     = {.x = c1, .y = r1};
+	uv2 bottom_right = {.x = c2, .y = r2};
+	if (selection_intersects_region(&t->selection, top_left, bottom_right))
+		selection_clear(&t->selection);
+
+	Row *rows = t->views[t->view_idx].fb.rows;
+	CellStyle cursor_style = t->cursor.style;
+	for (u32 r = top_left.y; r <= bottom_right.y; r++) {
+		for (u32 c = top_left.x; c <= bottom_right.x; c++) {
+			rows[r][c].style = cursor_style;
+			rows[r][c].cp    = ' ';
 		}
 	}
+
 	END_TIMED_BLOCK();
 }
 
@@ -1010,6 +1029,7 @@ validate_osc(Term *t, s8 *raw)
 static enum escape_moves_cursor_result
 check_if_csi_moves_cursor(Term *t, s8 *raw)
 {
+	BEGIN_TIMED_BLOCK();
 	enum escape_moves_cursor_result result = EMC_NORMAL_RETURN;
 	CSI csi = {0};
 	if (parse_csi(raw, &csi) == -1)
@@ -1039,6 +1059,7 @@ check_if_csi_moves_cursor(Term *t, s8 *raw)
 	if (mode != (t->mode & TM_ALTSCREEN))
 		result = EMC_SWAPPED_SCREEN;
 
+	END_TIMED_BLOCK();
 	return result;
 }
 
diff --git a/util.h b/util.h
@@ -75,7 +75,7 @@ typedef union {
 	f32 E[2];
 } v2;
 
-typedef union {
+typedef __attribute__((aligned(16))) union {
 	struct { f32 x, y, z, w; };
 	struct { f32 r, g, b, a; };
 	f32 E[4];
diff --git a/vtgl.c b/vtgl.c
@@ -215,6 +215,8 @@ get_render_push_buffer_idx(RenderCtx *rc, u32 count)
 static void
 push_rect_full(RenderCtx *rc, Rect r, v4 colour, v2 min_tex_coord, v2 max_tex_coord)
 {
+	BEGIN_TIMED_BLOCK();
+
 	u32 idx = get_render_push_buffer_idx(rc, 4);
 	v2 start = r.pos;
 	v2 end   = {.x = r.pos.x + r.size.w, .y = r.pos.y + r.size.h};
@@ -234,6 +236,8 @@ push_rect_full(RenderCtx *rc, Rect r, v4 colour, v2 min_tex_coord, v2 max_tex_co
 	rpb->colours[idx + 1] = colour;
 	rpb->colours[idx + 2] = colour;
 	rpb->colours[idx + 3] = colour;
+
+	END_TIMED_BLOCK();
 }
 
 static void
@@ -261,6 +265,7 @@ push_rect(RenderCtx *rc, Rect r, v4 colour)
 static v2
 push_s8(RenderCtx *rc, v2 pos, v4 colour, u32 font_id, s8 s)
 {
+	BEGIN_TIMED_BLOCK();
 	CachedGlyph *cg;
 	v2 start, end, text_size = {0};
 	v2 scale = {.x = 1.0f / rc->gl->glyph_bitmap_dim.x, .y = 1.0f / rc->gl->glyph_bitmap_dim.y};
@@ -287,6 +292,8 @@ push_s8(RenderCtx *rc, v2 pos, v4 colour, u32 font_id, s8 s)
 	text_size.x -= cg->advance;
 	text_size.w += cg->width;
 
+	END_TIMED_BLOCK();
+
 	return text_size;
 }
 
@@ -974,8 +981,8 @@ draw_debug_overlay(Term *t, RenderCtx *rc)
 		DebugRecord *dr = debug_records + i;
 
 		s8 txt    = buf;
-		txt.len   = snprintf((char *)txt.data, buf.len, "%29s: %5u %4s %11.02f cycs/hit",
-		                     dr->function_name, hits[i], hits[i] > 1? "hits" : "hit",
+		txt.len   = snprintf((char *)txt.data, buf.len, "%29s: %9lu cycs %5u %4s %11.02f cycs/hit",
+		                     dr->function_name, cycs[i], hits[i], hits[i] > 1? "hits" : "hit",
 		                     (f32)cycs[i]/(f32)hits[i]);
 		txt_pos.y = (u32)(txt_pos.y - line_height - line_pad);
 		v2 ts = push_s8(rc, txt_pos, fg, font_id, txt);