sna: Emit points using the BLT primitive when appropriate

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

4 files changed, 151 insertions, 19 deletions

diff --git a/src/sna/sna_accel.c b/src/sna/sna_accel.c
index 6359de1d..aff7b343 100644
--- a/src/sna/sna_accel.c
+++ b/src/sna/sna_accel.c
@@ -6586,10 +6586,14 @@ sna_poly_point__fill(DrawablePtr drawable, GCPtr gc,
 	DDXPointRec last;
 
 	DBG(("%s: count=%d\n", __FUNCTION__, n));
+	if (n == 0)
+		return;
 
 	last.x = drawable->x + data->dx;
 	last.y = drawable->y + data->dy;
-	while (n) {
+	if (op->points && mode != CoordModePrevious) {
+		op->points(data->sna, op, last.x, last.y, pt, n);
+	} else do {
 		BoxRec *b = box;
 		unsigned nbox = n;
 		if (nbox > ARRAY_SIZE(box))
@@ -6608,7 +6612,7 @@ sna_poly_point__fill(DrawablePtr drawable, GCPtr gc,
 			b++;
 		} while (--nbox);
 		op->boxes(data->sna, op, box, b - box);
-	}
+	} while (n);
 }
 
 static void
@@ -8177,26 +8181,30 @@ sna_poly_point_blt(DrawablePtr drawable,
 
 		assert_pixmap_contains_points(pixmap, pt, n, last.x, last.y);
 		sna_damage_add_points(damage, pt, n, last.x, last.y);
-		do {
-			unsigned nbox = n;
-			if (nbox > ARRAY_SIZE(box))
-				nbox = ARRAY_SIZE(box);
-			n -= nbox;
+		if (fill.points && mode != CoordModePrevious) {
+			fill.points(sna, &fill, last.x, last.y, pt, n);
+		} else {
 			do {
-				*(DDXPointRec *)b = *pt++;
+				unsigned nbox = n;
+				if (nbox > ARRAY_SIZE(box))
+					nbox = ARRAY_SIZE(box);
+				n -= nbox;
+				do {
+					*(DDXPointRec *)b = *pt++;
 
-				b->x1 += last.x;
-				b->y1 += last.y;
-				if (mode == CoordModePrevious)
-					last = *(DDXPointRec *)b;
+					b->x1 += last.x;
+					b->y1 += last.y;
+					if (mode == CoordModePrevious)
+						last = *(DDXPointRec *)b;
 
-				b->x2 = b->x1 + 1;
-				b->y2 = b->y1 + 1;
-				b++;
-			} while (--nbox);
-			fill.boxes(sna, &fill, box, b - box);
-			b = box;
-		} while (n);
+					b->x2 = b->x1 + 1;
+					b->y2 = b->y1 + 1;
+					b++;
+				} while (--nbox);
+				fill.boxes(sna, &fill, box, b - box);
+				b = box;
+			} while (n);
+		}
 	} else {
 		RegionPtr clip = gc->pCompositeClip;
 
diff --git a/src/sna/sna_blt.c b/src/sna/sna_blt.c
index 4bbcdbcc..26a09443 100644
--- a/src/sna/sna_blt.c
+++ b/src/sna/sna_blt.c
@@ -2862,6 +2862,123 @@ fastcall static void sna_blt_fill_op_boxes(struct sna *sna,
 	_sna_blt_fill_boxes(sna, &op->base.u.blt, box, nbox);
 }
 
+static inline uint64_t pt_add(const DDXPointRec *pt, int16_t dx, int16_t dy)
+{
+	union {
+		DDXPointRec pt;
+		uint32_t i;
+	} u;
+
+	u.pt.x = pt->x + dx;
+	u.pt.y = pt->y + dy;
+
+	return XY_PIXEL_BLT | (uint64_t)u.i<<32;
+}
+
+fastcall static void sna_blt_fill_op_points(struct sna *sna,
+					    const struct sna_fill_op *op,
+					    int16_t dx, int16_t dy,
+					    const DDXPointRec *p, int n)
+{
+	const struct sna_blt_state *blt = &op->base.u.blt;
+	struct kgem *kgem = &sna->kgem;
+
+	DBG(("%s: %08x x %d\n", __FUNCTION__, blt->pixel, npoints));
+
+	if (sna->blt_state.fill_bo != op->base.u.blt.bo[0]->unique_id) {
+		sna_blt_fill_begin(sna, blt);
+
+		sna->blt_state.fill_bo = blt->bo[0]->unique_id;
+		sna->blt_state.fill_pixel = blt->pixel;
+		sna->blt_state.fill_alu = blt->alu;
+	}
+
+	if (!kgem_check_batch(kgem, 2))
+		sna_blt_fill_begin(sna, blt);
+
+	do {
+		uint32_t *b = kgem->batch + kgem->nbatch;
+		int n_this_time;
+
+		assert(sna->kgem.mode == KGEM_BLT);
+		n_this_time = n;
+		if (2*n_this_time > kgem->surface - kgem->nbatch - KGEM_BATCH_RESERVED)
+			n_this_time = (kgem->surface - kgem->nbatch - KGEM_BATCH_RESERVED) / 2;
+		assert(n_this_time);
+		n -= n_this_time;
+
+		kgem->nbatch += 2 * n_this_time;
+		assert(kgem->nbatch < kgem->surface);
+
+		if ((dx|dy) == 0) {
+			while (n_this_time >= 8) {
+				*((uint64_t *)b + 0) = pt_add(p+0, 0, 0);
+				*((uint64_t *)b + 1) = pt_add(p+1, 0, 0);
+				*((uint64_t *)b + 2) = pt_add(p+2, 0, 0);
+				*((uint64_t *)b + 3) = pt_add(p+3, 0, 0);
+				*((uint64_t *)b + 4) = pt_add(p+4, 0, 0);
+				*((uint64_t *)b + 5) = pt_add(p+5, 0, 0);
+				*((uint64_t *)b + 6) = pt_add(p+6, 0, 0);
+				*((uint64_t *)b + 7) = pt_add(p+7, 0, 0);
+				b += 16;
+				n_this_time -= 8;
+				p += 8;
+			}
+			if (n_this_time & 4) {
+				*((uint64_t *)b + 0) = pt_add(p+0, 0, 0);
+				*((uint64_t *)b + 1) = pt_add(p+1, 0, 0);
+				*((uint64_t *)b + 2) = pt_add(p+2, 0, 0);
+				*((uint64_t *)b + 3) = pt_add(p+3, 0, 0);
+				b += 8;
+				p += 4;
+			}
+			if (n_this_time & 2) {
+				*((uint64_t *)b + 0) = pt_add(p+0, 0, 0);
+				*((uint64_t *)b + 1) = pt_add(p+1, 0, 0);
+				b += 4;
+				p += 2;
+			}
+			if (n_this_time & 1)
+				*((uint64_t *)b + 0) = pt_add(p++, 0, 0);
+		} else {
+			while (n_this_time >= 8) {
+				*((uint64_t *)b + 0) = pt_add(p+0, dx, dy);
+				*((uint64_t *)b + 1) = pt_add(p+1, dx, dy);
+				*((uint64_t *)b + 2) = pt_add(p+2, dx, dy);
+				*((uint64_t *)b + 3) = pt_add(p+3, dx, dy);
+				*((uint64_t *)b + 4) = pt_add(p+4, dx, dy);
+				*((uint64_t *)b + 5) = pt_add(p+5, dx, dy);
+				*((uint64_t *)b + 6) = pt_add(p+6, dx, dy);
+				*((uint64_t *)b + 7) = pt_add(p+7, dx, dy);
+				b += 16;
+				n_this_time -= 8;
+				p += 8;
+			}
+			if (n_this_time & 4) {
+				*((uint64_t *)b + 0) = pt_add(p+0, dx, dy);
+				*((uint64_t *)b + 1) = pt_add(p+1, dx, dy);
+				*((uint64_t *)b + 2) = pt_add(p+2, dx, dy);
+				*((uint64_t *)b + 3) = pt_add(p+3, dx, dy);
+				b += 8;
+				p += 8;
+			}
+			if (n_this_time & 2) {
+				*((uint64_t *)b + 0) = pt_add(p+0, dx, dy);
+				*((uint64_t *)b + 1) = pt_add(p+1, dx, dy);
+				b += 4;
+				p += 2;
+			}
+			if (n_this_time & 1)
+				*((uint64_t *)b + 0) = pt_add(p++, dx, dy);
+		}
+
+		if (!n)
+			return;
+
+		sna_blt_fill_begin(sna, blt);
+	} while (1);
+}
+
 bool sna_blt_fill(struct sna *sna, uint8_t alu,
 		  struct kgem_bo *bo, int bpp,
 		  uint32_t pixel,
@@ -2886,6 +3003,7 @@ bool sna_blt_fill(struct sna *sna, uint8_t alu,
 	fill->blt   = sna_blt_fill_op_blt;
 	fill->box   = sna_blt_fill_op_box;
 	fill->boxes = sna_blt_fill_op_boxes;
+	fill->points = sna_blt_fill_op_points;
 	fill->done  =
 		(void (*)(struct sna *, const struct sna_fill_op *))nop_done;
 	return true;
diff --git a/src/sna/sna_reg.h b/src/sna/sna_reg.h
index bda6ef67..92a1ae57 100644
--- a/src/sna/sna_reg.h
+++ b/src/sna/sna_reg.h
@@ -47,6 +47,7 @@
 #define XY_SETUP_BLT			(2<<29|0x01<<22)
 #define XY_SETUP_MONO_PATTERN_SL_BLT	(2<<29|0x11<<22)
 #define XY_SETUP_CLIP			(2<<29|0x03<<22|1)
+#define XY_PIXEL_BLT			(2<<29|0x24<<22)
 #define XY_SCANLINE_BLT			(2<<29|0x25<<22|1)
 #define XY_TEXT_IMMEDIATE_BLT		(2<<29|0x31<<22|(1<<16))
 #define XY_SRC_COPY_BLT_CMD		(2<<29|0x53<<22)
diff --git a/src/sna/sna_render.h b/src/sna/sna_render.h
index 67848b98..325b7cca 100644
--- a/src/sna/sna_render.h
+++ b/src/sna/sna_render.h
@@ -199,6 +199,11 @@ struct sna_fill_op {
 			       const struct sna_fill_op *op,
 			       const BoxRec *box,
 			       int count);
+	fastcall void (*points)(struct sna *sna,
+			       const struct sna_fill_op *op,
+			       int16_t dx, int16_t dy,
+			       const DDXPointRec *points,
+			       int count);
 	void (*done)(struct sna *sna, const struct sna_fill_op *op);
 };