sna/blt: Use SETUP_MONO to reduce the number of fill relocations

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

7 files changed, 301 insertions, 38 deletions

diff --git a/src/sna/gen2_render.c b/src/sna/gen2_render.c
index 43a7cad1..43aaef08 100644
--- a/src/sna/gen2_render.c
+++ b/src/sna/gen2_render.c
@@ -2003,6 +2003,87 @@ gen2_render_fill(struct sna *sna, uint8_t alu,
 	return TRUE;
 }
 
+static Bool
+gen2_render_fill_one_try_blt(struct sna *sna, PixmapPtr dst, struct kgem_bo *bo,
+			     uint32_t color,
+			     int16_t x1, int16_t y1, int16_t x2, int16_t y2,
+			     uint8_t alu)
+{
+	BoxRec box;
+
+	box.x1 = x1;
+	box.y1 = y1;
+	box.x2 = x2;
+	box.y2 = y2;
+
+	return sna_blt_fill_boxes(sna, alu,
+				  bo, dst->drawable.bitsPerPixel,
+				  color, &box, 1);
+}
+
+static Bool
+gen2_render_fill_one(struct sna *sna, PixmapPtr dst, struct kgem_bo *bo,
+		     uint32_t color,
+		     int16_t x1, int16_t y1,
+		     int16_t x2, int16_t y2,
+		     uint8_t alu)
+{
+	struct sna_composite_op tmp;
+
+#if NO_FILL_BOXES
+	return gen2_render_fill_one_try_blt(sna, dst, bo, color,
+					    x1, y1, x2, y2, alu);
+#endif
+
+	/* Prefer to use the BLT if already engaged */
+	if (sna->kgem.mode != KGEM_RENDER &&
+	    gen2_render_fill_one_try_blt(sna, dst, bo, color,
+					 x1, y1, x2, y2, alu))
+		return TRUE;
+
+	/* Must use the BLT if we can't RENDER... */
+	if (!(alu == GXcopy || alu == GXclear) ||
+	    dst->drawable.width > 2048 || dst->drawable.height > 2048 ||
+	    bo->pitch > 8192)
+		return gen2_render_fill_one_try_blt(sna, dst, bo, color,
+						    x1, y1, x2, y2, alu);
+
+	if (alu == GXclear)
+		color = 0;
+
+	memset(&tmp, 0, sizeof(tmp));
+	tmp.op = color == 0 ? PictOpClear : PictOpSrc;
+	tmp.dst.pixmap = dst;
+	tmp.dst.width = dst->drawable.width;
+	tmp.dst.height = dst->drawable.height;
+	tmp.dst.format = sna_format_for_depth(dst->drawable.depth);
+	tmp.dst.bo = bo;
+	tmp.floats_per_vertex = 2;
+
+	tmp.src.u.gen2.pixel =
+		sna_rgba_for_color(color, dst->drawable.depth);
+
+	if (!kgem_check_bo(&sna->kgem, bo, NULL)) {
+		kgem_submit(&sna->kgem);
+		if (gen2_render_fill_one_try_blt(sna, dst, bo, color,
+						 x1, y1, x2, y2, alu))
+			return TRUE;
+	}
+
+	gen2_emit_fill_state(sna, &tmp);
+	gen2_get_rectangles(sna, &tmp, 1);
+	DBG(("	(%d, %d), (%d, %d): %x\n", x1, y1, x2, y2, pixel));
+	OUT_VERTEX(x2);
+	OUT_VERTEX(y2);
+	OUT_VERTEX(x1);
+	OUT_VERTEX(y2);
+	OUT_VERTEX(x1);
+	OUT_VERTEX(y1);
+	gen2_vertex_flush(sna);
+
+	return TRUE;
+}
+
 static void
 gen2_render_copy_setup_source(struct sna_composite_channel *channel,
 			      PixmapPtr pixmap,
@@ -2297,8 +2378,8 @@ Bool gen2_render_init(struct sna *sna)
 	render->composite = gen2_render_composite;
 	render->composite_spans = gen2_render_composite_spans;
 	render->fill_boxes = gen2_render_fill_boxes;
-
 	render->fill = gen2_render_fill;
+	render->fill_one = gen2_render_fill_one;
 	render->copy = gen2_render_copy;
 	render->copy_boxes = gen2_render_copy_boxes;
 
diff --git a/src/sna/gen3_render.c b/src/sna/gen3_render.c
index bac03533..65176ed9 100644
--- a/src/sna/gen3_render.c
+++ b/src/sna/gen3_render.c
@@ -3659,6 +3659,90 @@ gen3_render_fill(struct sna *sna, uint8_t alu,
 	return TRUE;
 }
 
+
+static Bool
+gen3_render_fill_one_try_blt(struct sna *sna, PixmapPtr dst, struct kgem_bo *bo,
+			     uint32_t color,
+			     int16_t x1, int16_t y1, int16_t x2, int16_t y2,
+			     uint8_t alu)
+{
+	BoxRec box;
+
+	box.x1 = x1;
+	box.y1 = y1;
+	box.x2 = x2;
+	box.y2 = y2;
+
+	return sna_blt_fill_boxes(sna, alu,
+				  bo, dst->drawable.bitsPerPixel,
+				  color, &box, 1);
+}
+
+static Bool
+gen3_render_fill_one(struct sna *sna, PixmapPtr dst, struct kgem_bo *bo,
+		     uint32_t color,
+		     int16_t x1, int16_t y1,
+		     int16_t x2, int16_t y2,
+		     uint8_t alu)
+{
+	struct sna_composite_op tmp;
+
+#if NO_FILL_BOXES
+	return gen3_render_fill_one_try_blt(sna, dst, bo, color,
+					    x1, y1, x2, y2, alu);
+#endif
+
+	/* Prefer to use the BLT if already engaged */
+	if (sna->kgem.mode != KGEM_RENDER &&
+	    gen3_render_fill_one_try_blt(sna, dst, bo, color,
+					 x1, y1, x2, y2, alu))
+		return TRUE;
+
+	/* Must use the BLT if we can't RENDER... */
+	if (!(alu == GXcopy || alu == GXclear) ||
+	    dst->drawable.width > 2048 || dst->drawable.height > 2048 ||
+	    bo->pitch > 8192)
+		return gen3_render_fill_one_try_blt(sna, dst, bo, color,
+						    x1, y1, x2, y2, alu);
+
+	if (alu == GXclear)
+		color = 0;
+
+	memset(&tmp, 0, sizeof(tmp));
+	tmp.op = color == 0 ? PictOpClear : PictOpSrc;
+	tmp.dst.pixmap = dst;
+	tmp.dst.width = dst->drawable.width;
+	tmp.dst.height = dst->drawable.height;
+	tmp.dst.format = sna_format_for_depth(dst->drawable.depth);
+	tmp.dst.bo = bo;
+	tmp.floats_per_vertex = 2;
+
+	tmp.src.u.gen3.type = SHADER_CONSTANT;
+	tmp.src.u.gen3.mode =
+		sna_rgba_for_color(color, dst->drawable.depth);
+
+	if (!kgem_check_bo(&sna->kgem, bo, NULL)) {
+		kgem_submit(&sna->kgem);
+		if (gen3_render_fill_one_try_blt(sna, dst, bo, color,
+						 x1, y1, x2, y2, alu))
+			return TRUE;
+	}
+
+	gen3_emit_composite_state(sna, &tmp);
+	gen3_align_vertex(sna, &tmp);
+	gen3_get_rectangles(sna, &tmp, 1);
+	DBG(("	(%d, %d), (%d, %d): %x\n", x1, y1, x2, y2, pixel));
+	OUT_VERTEX(x2);
+	OUT_VERTEX(y2);
+	OUT_VERTEX(x1);
+	OUT_VERTEX(y2);
+	OUT_VERTEX(x1);
+	OUT_VERTEX(y1);
+	gen3_vertex_flush(sna);
+
+	return TRUE;
+}
+
 static void gen3_render_flush(struct sna *sna)
 {
 	gen3_vertex_finish(sna, TRUE);
@@ -3683,6 +3767,7 @@ Bool gen3_render_init(struct sna *sna)
 
 	render->fill_boxes = gen3_render_fill_boxes;
 	render->fill = gen3_render_fill;
+	render->fill_one = gen3_render_fill_one;
 
 	render->reset = gen3_render_reset;
 	render->flush = gen3_render_flush;
diff --git a/src/sna/sna_accel.c b/src/sna/sna_accel.c
index 1b43a8b6..a6e639ba 100644
--- a/src/sna/sna_accel.c
+++ b/src/sna/sna_accel.c
@@ -1650,6 +1650,29 @@ sna_spans_extents(DrawablePtr drawable, GCPtr gc,
 	return box_empty(&box);
 }
 
+static struct sna_damage **
+reduce_damage(DrawablePtr drawable,
+	      struct sna_damage **damage,
+	      const BoxRec *box)
+{
+	PixmapPtr pixmap = get_drawable_pixmap(drawable);
+	int16_t dx, dy;
+	BoxRec r;
+
+	if (*damage == NULL)
+		return damage;
+
+	get_drawable_deltas(drawable, pixmap, &dx, &dy);
+
+	r = *box;
+	r.x1 += dx; r.x2 += dx;
+	r.y1 += dy; r.y2 += dy;
+	if (sna_damage_contains_box(*damage, &r) == PIXMAN_REGION_IN)
+		return NULL;
+	else
+		return damage;
+}
+
 static void
 sna_poly_fill_rect(DrawablePtr draw, GCPtr gc, int n, xRectangle *rect);
 
@@ -1702,13 +1725,14 @@ sna_fill_spans(DrawablePtr drawable, GCPtr gc, int n,
 		if (sna_drawable_use_gpu_bo(drawable, &extents) &&
 		    sna_fill_spans_blt(drawable,
 				       priv->gpu_bo,
-				       priv->gpu_only ? NULL : &priv->gpu_damage,
+				       priv->gpu_only ? NULL : reduce_damage(drawable, &priv->gpu_damage, &extents),
 				       gc, n, pt, width, sorted))
 			return;
 
 		if (sna_drawable_use_cpu_bo(drawable, &extents) &&
 		    sna_fill_spans_blt(drawable,
-				       priv->cpu_bo, &priv->cpu_damage,
+				       priv->cpu_bo,
+				       reduce_damage(drawable, &priv->cpu_damage, &extents),
 				       gc, n, pt, width, sorted))
 			return;
 	} else if (gc->fillStyle == FillTiled) {
@@ -2556,6 +2580,33 @@ sna_poly_fill_rect_blt(DrawablePtr drawable,
 	DBG(("%s x %d [(%d, %d)+(%d, %d)...]\n",
 	     __FUNCTION__, n, rect->x, rect->y, rect->width, rect->height));
 
+	if (n == 1 && REGION_NUM_RECTS(clip) == 1) {
+		BoxPtr box = REGION_RECTS(clip);
+		BoxRec r;
+		bool success = true;
+
+		r.x1 = rect->x + drawable->x;
+		r.y1 = rect->y + drawable->y;
+		r.x2 = bound(r.x1, rect->width);
+		r.y2 = bound(r.y1, rect->height);
+		if (box_intersect(&r, box)) {
+			get_drawable_deltas(drawable, pixmap, &dx, &dy);
+			r.x1 += dx; r.y1 += dy;
+			r.x2 += dx; r.y2 += dy;
+			if (sna->render.fill_one(sna, pixmap, bo, pixel,
+						 r.x1, r.y1, r.x2, r.y2,
+						 gc->alu)) {
+				if (damage) {
+					assert_pixmap_contains_box(pixmap, &r);
+					sna_damage_add_box(damage, &r);
+				}
+			} else
+				success = false;
+		}
+
+		return success;
+	}
+
 	if (!sna_fill_init_blt(&fill, sna, pixmap, bo, gc->alu, pixel)) {
 		DBG(("%s: unsupported blt\n", __FUNCTION__));
 		return FALSE;
@@ -2935,14 +2986,14 @@ sna_poly_fill_rect(DrawablePtr draw, GCPtr gc, int n, xRectangle *rect)
 		if (sna_drawable_use_gpu_bo(draw, &extents) &&
 		    sna_poly_fill_rect_blt(draw,
 					   priv->gpu_bo,
-					   priv->gpu_only ? NULL : &priv->gpu_damage,
+					   priv->gpu_only ? NULL : reduce_damage(draw, &priv->gpu_damage, &extents),
 					   gc, n, rect))
 			return;
 
 		if (sna_drawable_use_cpu_bo(draw, &extents) &&
 		    sna_poly_fill_rect_blt(draw,
 					   priv->cpu_bo,
-					   &priv->cpu_damage,
+					   reduce_damage(draw, &priv->cpu_damage, &extents),
 					   gc, n, rect))
 			return;
 	} else if (gc->fillStyle == FillTiled) {
@@ -2953,14 +3004,14 @@ sna_poly_fill_rect(DrawablePtr draw, GCPtr gc, int n, xRectangle *rect)
 		if (sna_drawable_use_gpu_bo(draw, &extents) &&
 		    sna_poly_fill_rect_tiled(draw,
 					     priv->gpu_bo,
-					     priv->gpu_only ? NULL : &priv->gpu_damage,
+					     priv->gpu_only ? NULL : reduce_damage(draw, &priv->gpu_damage, &extents),
 					     gc, n, rect))
 			return;
 
 		if (sna_drawable_use_cpu_bo(draw, &extents) &&
 		    sna_poly_fill_rect_tiled(draw,
 					     priv->cpu_bo,
-					     &priv->cpu_damage,
+					     reduce_damage(draw, &priv->cpu_damage, &extents),
 					     gc, n, rect))
 			return;
 	}
diff --git a/src/sna/sna_blt.c b/src/sna/sna_blt.c
index 5abfb026..86388a8f 100644
--- a/src/sna/sna_blt.c
+++ b/src/sna/sna_blt.c
@@ -117,10 +117,9 @@ static bool sna_blt_fill_init(struct sna *sna,
 	struct kgem *kgem = &sna->kgem;
 	int pitch;
 
-
 	blt->bo[0] = bo;
 
-	blt->cmd = XY_COLOR_BLT_CMD;
+	blt->cmd = XY_SETUP_MONO_PATTERN_SL_BLT;
 	if (bpp == 32)
 		blt->cmd |= BLT_WRITE_ALPHA | BLT_WRITE_RGB;
 
@@ -133,7 +132,7 @@ static bool sna_blt_fill_init(struct sna *sna,
 		return FALSE;
 
 	blt->overwrites = alu == GXcopy || alu == GXclear;
-	blt->br13 = (fill_ROP[alu] << 16) | pitch;
+	blt->br13 = 1<<31 | (fill_ROP[alu] << 16) | pitch;
 	switch (bpp) {
 	default: assert(0);
 	case 32: blt->br13 |= 1 << 25; /* RGB8888 */
@@ -144,8 +143,29 @@ static bool sna_blt_fill_init(struct sna *sna,
 	blt->pixel = pixel;
 
 	kgem_set_mode(kgem, KGEM_BLT);
-	if (!kgem_check_bo_fenced(kgem, bo, NULL))
+	if (!kgem_check_bo_fenced(kgem, bo, NULL) ||
+	    !kgem_check_batch(kgem, 3)) {
 		_kgem_submit(kgem);
+		_kgem_set_mode(kgem, KGEM_BLT);
+	}
+
+	{
+		uint32_t *b = kgem->batch + kgem->nbatch;
+		b[0] = blt->cmd;
+		b[1] = blt->br13;
+		b[2] = 0;
+		b[3] = 0;
+		b[4] = kgem_add_reloc(kgem, kgem->nbatch + 4, bo,
+				      I915_GEM_DOMAIN_RENDER << 16 |
+				      I915_GEM_DOMAIN_RENDER |
+				      KGEM_RELOC_FENCED,
+				      0);
+		b[5] = pixel;
+		b[6] = pixel;
+		b[7] = 0;
+		b[8] = 0;
+		kgem->nbatch += 9;
+	}
 
 	return TRUE;
 }
@@ -165,36 +185,34 @@ static void sna_blt_fill_one(struct sna *sna,
 	assert(y >= 0);
 	assert((y+height) * blt->bo[0]->pitch <= blt->bo[0]->size);
 
-	/* All too frequently one blt completely overwrites the previous */
-	if (kgem->nbatch >= 6 &&
-	    blt->overwrites &&
-	    kgem->batch[kgem->nbatch-6] == blt->cmd &&
-	    kgem->batch[kgem->nbatch-4] == ((uint32_t)y << 16 | (uint16_t)x) &&
-	    kgem->batch[kgem->nbatch-3] == ((uint32_t)(y+height) << 16 | (uint16_t)(x+width)) &&
-	    kgem->reloc[kgem->nreloc-1].target_handle == blt->bo[0]->handle) {
-		DBG(("%s: replacing last fill\n", __FUNCTION__));
-		kgem->batch[kgem->nbatch-5] = blt->br13;
-		kgem->batch[kgem->nbatch-1] = blt->pixel;
-		return;
-	}
-
-	if (!kgem_check_batch(kgem,  6) ||
-	    kgem->nreloc + 1 > KGEM_RELOC_SIZE(kgem))
+	if (!kgem_check_batch(kgem, 3)) {
 		_kgem_submit(kgem);
+		_kgem_set_mode(kgem, KGEM_BLT);
+
+		b = kgem->batch + kgem->nbatch;
+		b[0] = blt->cmd;
+		b[1] = blt->br13;
+		b[2] = 0;
+		b[3] = 0;
+		b[4] = kgem_add_reloc(kgem, kgem->nbatch + 4, blt->bo[0],
+				      I915_GEM_DOMAIN_RENDER << 16 |
+				      I915_GEM_DOMAIN_RENDER |
+				      KGEM_RELOC_FENCED,
+				      0);
+		b[5] = blt->pixel;
+		b[6] = blt->pixel;
+		b[7] = 0;
+		b[8] = 0;
+		kgem->nbatch += 9;
+	}
 
 	b = kgem->batch + kgem->nbatch;
-	b[0] = blt->cmd;
-	b[1] = blt->br13;
-	b[2] = (y << 16) | x;
-	b[3] = ((y + height) << 16) | (x + width);
-	b[4] = kgem_add_reloc(kgem, kgem->nbatch + 4,
-			      blt->bo[0],
-			      I915_GEM_DOMAIN_RENDER << 16 |
-			      I915_GEM_DOMAIN_RENDER |
-			      KGEM_RELOC_FENCED,
-			      0);
-	b[5] = blt->pixel;
-	kgem->nbatch += 6;
+	b[0] = XY_SCANLINE_BLT;
+	if (kgem->gen >= 40 && blt->bo[0]->tiling)
+		b[0] |= 1 << 11;
+	b[1] = (y << 16) | x;
+	b[2] = ((y + height) << 16) | (x + width);
+	kgem->nbatch += 3;
 }
 
 static Bool sna_blt_copy_init(struct sna *sna,
diff --git a/src/sna/sna_reg.h b/src/sna/sna_reg.h
index f6e53979..f16c64dc 100644
--- a/src/sna/sna_reg.h
+++ b/src/sna/sna_reg.h
@@ -78,7 +78,10 @@
 
 #define COLOR_BLT_CMD		((2<<29)|(0x40<<22)|(0x3))
 #define XY_COLOR_BLT_CMD		((2<<29)|(0x50<<22)|(0x4))
+#define XY_SETUP_BLT_CMD		((2<<29)|(1<<22)|6)
+#define XY_SETUP_MONO_PATTERN_SL_BLT	((2<<29)|(0x11<<22)|7)
 #define XY_SETUP_CLIP_BLT_CMD		((2<<29)|(3<<22)|1)
+#define XY_SCANLINE_BLT			((2<<29)|(0x25<<22)|1)
 #define XY_SRC_COPY_BLT_CMD		((2<<29)|(0x53<<22)|6)
 #define SRC_COPY_BLT_CMD		((2<<29)|(0x43<<22)|0x4)
 #define XY_PAT_BLT_IMMEDIATE		((2<<29)|(0x72<<22))
diff --git a/src/sna/sna_render.c b/src/sna/sna_render.c
index 292853a0..7ab0de1d 100644
--- a/src/sna/sna_render.c
+++ b/src/sna/sna_render.c
@@ -179,6 +179,26 @@ no_render_fill(struct sna *sna, uint8_t alu,
 			    tmp);
 }
 
+static Bool
+no_render_fill_one(struct sna *sna, PixmapPtr dst, struct kgem_bo *bo,
+		   uint32_t color,
+		   int16_t x1, int16_t y1, int16_t x2, int16_t y2,
+		   uint8_t alu)
+{
+	BoxRec box;
+
+	box.x1 = x1;
+	box.y1 = y1;
+	box.x2 = x2;
+	box.y2 = y2;
+
+	DBG(("%s (alu=%d, color=%08x) (%d,%d), (%d, %d)\n",
+	     __FUNCTION__, alu, color, x1, y1, x2, y2));
+	return sna_blt_fill_boxes(sna, alu,
+				  bo, dst->drawable.bitsPerPixel,
+				  color, &box, 1);
+}
+
 static void no_render_reset(struct sna *sna)
 {
 	(void)sna;
@@ -216,6 +236,7 @@ void no_render_init(struct sna *sna)
 
 	render->fill_boxes = no_render_fill_boxes;
 	render->fill = no_render_fill;
+	render->fill_one = no_render_fill_one;
 
 	render->reset = no_render_reset;
 	render->flush = no_render_flush;
diff --git a/src/sna/sna_render.h b/src/sna/sna_render.h
index 80a9d00d..e46f027a 100644
--- a/src/sna/sna_render.h
+++ b/src/sna/sna_render.h
@@ -201,6 +201,10 @@ struct sna_render {
 		     PixmapPtr dst, struct kgem_bo *dst_bo,
 		     uint32_t color,
 		     struct sna_fill_op *tmp);
+	Bool (*fill_one)(struct sna *sna, PixmapPtr dst, struct kgem_bo *dst_bo,
+			 uint32_t color,
+			 int16_t x1, int16_t y1, int16_t x2, int16_t y2,
+			 uint8_t alu);
 
 	Bool (*copy_boxes)(struct sna *sna, uint8_t alu,
 			   PixmapPtr src, struct kgem_bo *src_bo, int16_t src_dx, int16_t src_dy,