sna: Pass render operation to flush and avoid the implicit flush-on-batch-end

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

7 files changed, 140 insertions, 173 deletions

diff --git a/src/sna/gen2_render.c b/src/sna/gen2_render.c
index 87f9e8ce..ef32f011 100644
--- a/src/sna/gen2_render.c
+++ b/src/sna/gen2_render.c
@@ -510,6 +510,7 @@ gen2_get_batch(struct sna *sna)
 		     __FUNCTION__, 30+40,
 		     sna->kgem.surface-sna->kgem.nbatch));
 		kgem_submit(&sna->kgem);
+		_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	}
 
 	if (sna->kgem.nreloc + 3 > KGEM_RELOC_SIZE(&sna->kgem)) {
@@ -518,6 +519,7 @@ gen2_get_batch(struct sna *sna)
 		     sna->kgem.nreloc + 3,
 		     (int)KGEM_RELOC_SIZE(&sna->kgem)));
 		kgem_submit(&sna->kgem);
+		_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	}
 
 	if (sna->kgem.nexec + 3 > KGEM_EXEC_SIZE(&sna->kgem)) {
@@ -526,6 +528,7 @@ gen2_get_batch(struct sna *sna)
 		     sna->kgem.nexec + 1,
 		     (int)KGEM_EXEC_SIZE(&sna->kgem)));
 		kgem_submit(&sna->kgem);
+		_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	}
 
 	if (sna->render_state.gen2.need_invariant)
@@ -936,7 +939,8 @@ static void gen2_magic_ca_pass(struct sna *sna,
 		*dst++ = *src++;
 }
 
-static void gen2_vertex_flush(struct sna *sna)
+static void gen2_vertex_flush(struct sna *sna,
+			      const struct sna_composite_op *op)
 {
 	if (sna->render.vertex_index == 0)
 		return;
@@ -944,8 +948,7 @@ static void gen2_vertex_flush(struct sna *sna)
 	sna->kgem.batch[sna->render_state.gen2.vertex_offset] |=
 		sna->render.vertex_index - 1;
 
-	if (sna->render.op)
-		gen2_magic_ca_pass(sna, sna->render.op);
+	gen2_magic_ca_pass(sna, op);
 
 	sna->render_state.gen2.vertex_offset = 0;
 	sna->render.vertex_index = 0;
@@ -971,7 +974,8 @@ inline static int gen2_get_rectangles(struct sna *sna,
 	DBG(("%s: want=%d, need=%d,size=%d, rem=%d\n",
 	     __FUNCTION__, want, need, size, rem));
 	if (rem < need + size) {
-		kgem_submit (&sna->kgem);
+		gen2_vertex_flush(sna, op);
+		kgem_submit(&sna->kgem);
 		return 0;
 	}
 
@@ -1073,9 +1077,7 @@ gen2_render_composite_boxes(struct sna *sna,
 static void gen2_render_composite_done(struct sna *sna,
 				       const struct sna_composite_op *op)
 {
-	gen2_vertex_flush(sna);
-	sna->render.op = NULL;
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	gen2_vertex_flush(sna, op);
 
 	if (op->mask.bo)
 		kgem_bo_destroy(&sna->kgem, op->mask.bo);
@@ -1790,8 +1792,6 @@ gen2_render_composite(struct sna *sna,
 	}
 
 	gen2_emit_composite_state(sna, tmp);
-
-	sna->render.op = tmp;
 	return TRUE;
 
 cleanup_src:
@@ -2105,11 +2105,10 @@ fastcall static void
 gen2_render_composite_spans_done(struct sna *sna,
 				 const struct sna_composite_spans_op *op)
 {
-	gen2_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
-
 	DBG(("%s()\n", __FUNCTION__));
 
+	gen2_vertex_flush(sna, &op->base);
+
 	if (op->base.src.bo)
 		kgem_bo_destroy(&sna->kgem, op->base.src.bo);
 
@@ -2436,8 +2435,7 @@ gen2_render_fill_boxes(struct sna *sna,
 		} while (--n_this_time);
 	} while (n);
 
-	gen2_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	gen2_vertex_flush(sna, &tmp);
 	return TRUE;
 }
 
@@ -2540,8 +2538,7 @@ gen2_render_fill_op_boxes(struct sna *sna,
 static void
 gen2_render_fill_op_done(struct sna *sna, const struct sna_fill_op *op)
 {
-	gen2_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	gen2_vertex_flush(sna, &op->base);
 }
 
 static Bool
@@ -2677,7 +2674,7 @@ gen2_render_fill_one(struct sna *sna, PixmapPtr dst, struct kgem_bo *bo,
 	VERTEX(y2);
 	VERTEX(x1);
 	VERTEX(y1);
-	gen2_vertex_flush(sna);
+	gen2_vertex_flush(sna, &tmp);
 
 	return TRUE;
 }
@@ -2862,8 +2859,7 @@ gen2_render_copy_boxes(struct sna *sna, uint8_t alu,
 		} while (--n_this_time);
 	} while (n);
 
-	gen2_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	gen2_vertex_flush(sna, &tmp);
 	return TRUE;
 }
 
@@ -2898,8 +2894,7 @@ gen2_render_copy_blt(struct sna *sna,
 static void
 gen2_render_copy_done(struct sna *sna, const struct sna_copy_op *op)
 {
-	gen2_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	gen2_vertex_flush(sna, &op->base);
 }
 
 static Bool
@@ -2984,7 +2979,7 @@ gen2_render_reset(struct sna *sna)
 static void
 gen2_render_flush(struct sna *sna)
 {
-	gen2_vertex_flush(sna);
+	assert(sna->render.vertex_index == 0);
 }
 
 static void
diff --git a/src/sna/gen3_render.c b/src/sna/gen3_render.c
index fc896d91..883e6223 100644
--- a/src/sna/gen3_render.c
+++ b/src/sna/gen3_render.c
@@ -1218,6 +1218,7 @@ gen3_get_batch(struct sna *sna)
 		     __FUNCTION__, 200,
 		     sna->kgem.surface-sna->kgem.nbatch));
 		kgem_submit(&sna->kgem);
+		_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	}
 
 	if (sna->kgem.nreloc > KGEM_RELOC_SIZE(&sna->kgem) - MAX_OBJECTS) {
@@ -1226,6 +1227,7 @@ gen3_get_batch(struct sna *sna)
 		     sna->kgem.nreloc,
 		     (int)KGEM_RELOC_SIZE(&sna->kgem) - MAX_OBJECTS));
 		kgem_submit(&sna->kgem);
+		_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	}
 
 	if (sna->kgem.nexec > KGEM_EXEC_SIZE(&sna->kgem) - MAX_OBJECTS - 1) {
@@ -1234,6 +1236,7 @@ gen3_get_batch(struct sna *sna)
 		     sna->kgem.nexec,
 		     (int)KGEM_EXEC_SIZE(&sna->kgem) - MAX_OBJECTS - 1));
 		kgem_submit(&sna->kgem);
+		_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	}
 
 	if (sna->render_state.gen3.need_invariant)
@@ -1492,9 +1495,7 @@ static void gen3_magic_ca_pass(struct sna *sna,
 
 static void gen3_vertex_flush(struct sna *sna)
 {
-	if (sna->render_state.gen3.vertex_offset == 0 ||
-	    sna->render.vertex_index == sna->render.vertex_start)
-		return;
+	assert(sna->render_state.gen3.vertex_offset);
 
 	DBG(("%s[%x] = %d\n", __FUNCTION__,
 	     4*sna->render_state.gen3.vertex_offset,
@@ -1506,9 +1507,6 @@ static void gen3_vertex_flush(struct sna *sna)
 	sna->kgem.batch[sna->render_state.gen3.vertex_offset + 1] =
 		sna->render.vertex_start;
 
-	if (sna->render.op)
-		gen3_magic_ca_pass(sna, sna->render.op);
-
 	sna->render_state.gen3.vertex_offset = 0;
 }
 
@@ -1516,9 +1514,7 @@ static int gen3_vertex_finish(struct sna *sna)
 {
 	struct kgem_bo *bo;
 
-	gen3_vertex_flush(sna);
-	if (!sna->render.vertex_used)
-		return sna->render.vertex_size;
+	assert(sna->render.vertex_used);
 
 	bo = sna->render.vbo;
 	if (bo) {
@@ -1560,8 +1556,6 @@ static void gen3_vertex_close(struct sna *sna)
 	struct kgem_bo *bo;
 	int delta = 0;
 
-	gen3_vertex_flush(sna);
-
 	if (!sna->render.vertex_used) {
 		assert(sna->render.vbo == NULL);
 		assert(sna->render.vertices == sna->render.vertex_data);
@@ -1656,9 +1650,15 @@ static bool gen3_rectangle_begin(struct sna *sna,
 	return true;
 }
 
-static int gen3_get_rectangles__flush(struct sna *sna, bool ca)
+static int gen3_get_rectangles__flush(struct sna *sna,
+				      const struct sna_composite_op *op)
 {
-	if (!kgem_check_batch(&sna->kgem, ca ? 105: 5))
+	if (sna->render_state.gen3.vertex_offset) {
+		gen3_vertex_flush(sna);
+		gen3_magic_ca_pass(sna, op);
+	}
+
+	if (!kgem_check_batch(&sna->kgem, op->need_magic_ca_pass ? 105: 5))
 		return 0;
 	if (sna->kgem.nexec > KGEM_EXEC_SIZE(&sna->kgem) - 2)
 		return 0;
@@ -1681,7 +1681,7 @@ inline static int gen3_get_rectangles(struct sna *sna,
 	if (op->floats_per_rect > rem) {
 		DBG(("flushing vbo for %s: %d < %d\n",
 		     __FUNCTION__, rem, op->floats_per_rect));
-		rem = gen3_get_rectangles__flush(sna, op->need_magic_ca_pass);
+		rem = gen3_get_rectangles__flush(sna, op);
 		if (rem == 0)
 			return 0;
 	}
@@ -1791,14 +1791,13 @@ static void
 gen3_render_composite_done(struct sna *sna,
 			   const struct sna_composite_op *op)
 {
-	assert(sna->render.op == op);
-
-	gen3_vertex_flush(sna);
-	sna->render.op = NULL;
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
-
 	DBG(("%s()\n", __FUNCTION__));
 
+	if (sna->render_state.gen3.vertex_offset) {
+		gen3_vertex_flush(sna);
+		gen3_magic_ca_pass(sna, op);
+	}
+
 	if (op->mask.bo)
 		kgem_bo_destroy(&sna->kgem, op->mask.bo);
 	if (op->src.bo)
@@ -2767,8 +2766,6 @@ gen3_render_composite(struct sna *sna,
 
 	gen3_emit_composite_state(sna, tmp);
 	gen3_align_vertex(sna, tmp);
-
-	sna->render.op = tmp;
 	return TRUE;
 
 cleanup_mask:
@@ -3069,8 +3066,8 @@ fastcall static void
 gen3_render_composite_spans_done(struct sna *sna,
 				 const struct sna_composite_spans_op *op)
 {
-	gen3_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	if (sna->render_state.gen3.vertex_offset)
+		gen3_vertex_flush(sna);
 
 	DBG(("%s()\n", __FUNCTION__));
 
@@ -3828,7 +3825,6 @@ gen3_render_copy_boxes(struct sna *sna, uint8_t alu,
 	} while (n);
 
 	gen3_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -3863,8 +3859,8 @@ gen3_render_copy_blt(struct sna *sna,
 static void
 gen3_render_copy_done(struct sna *sna, const struct sna_copy_op *op)
 {
-	gen3_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	if (sna->render_state.gen3.vertex_offset)
+		gen3_vertex_flush(sna);
 }
 
 static Bool
@@ -4103,7 +4099,6 @@ gen3_render_fill_boxes(struct sna *sna,
 	} while (n);
 
 	gen3_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -4175,8 +4170,8 @@ gen3_render_fill_op_boxes(struct sna *sna,
 static void
 gen3_render_fill_op_done(struct sna *sna, const struct sna_fill_op *op)
 {
-	gen3_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	if (sna->render_state.gen3.vertex_offset)
+		gen3_vertex_flush(sna);
 }
 
 static Bool
diff --git a/src/sna/gen4_render.c b/src/sna/gen4_render.c
index 942f8fb0..12281e3b 100644
--- a/src/sna/gen4_render.c
+++ b/src/sna/gen4_render.c
@@ -63,12 +63,12 @@
 #define NO_FILL_BOXES 0
 
 #if FLUSH_EVERY_VERTEX
-#define FLUSH() do { \
-	gen4_vertex_flush(sna); \
+#define FLUSH(OP) do { \
+	gen4_vertex_flush(sna, OP); \
 	OUT_BATCH(MI_FLUSH | MI_INHIBIT_RENDER_CACHE_FLUSH); \
 } while (0)
 #else
-#define FLUSH()
+#define FLUSH(OP)
 #endif
 
 #define GEN4_GRF_BLOCKS(nreg)    ((nreg + 15) / 16 - 1)
@@ -342,7 +342,8 @@ static void gen4_magic_ca_pass(struct sna *sna,
 	state->last_primitive = sna->kgem.nbatch;
 }
 
-static void gen4_vertex_flush(struct sna *sna)
+static void gen4_vertex_flush(struct sna *sna,
+			      const struct sna_composite_op *op)
 {
 	if (sna->render_state.gen4.vertex_offset == 0)
 		return;
@@ -354,8 +355,7 @@ static void gen4_vertex_flush(struct sna *sna)
 		sna->render.vertex_index - sna->render.vertex_start;
 	sna->render_state.gen4.vertex_offset = 0;
 
-	if (sna->render.op)
-		gen4_magic_ca_pass(sna, sna->render.op);
+	gen4_magic_ca_pass(sna, op);
 }
 
 static int gen4_vertex_finish(struct sna *sna)
@@ -363,9 +363,7 @@ static int gen4_vertex_finish(struct sna *sna)
 	struct kgem_bo *bo;
 	unsigned int i;
 
-	gen4_vertex_flush(sna);
-	if (!sna->render.vertex_used)
-		return sna->render.vertex_size;
+	assert(sna->render.vertex_used);
 
 	/* Note: we only need dword alignment (currently) */
 
@@ -417,7 +415,6 @@ static void gen4_vertex_close(struct sna *sna)
 	struct kgem_bo *bo;
 	unsigned int i, delta = 0;
 
-	gen4_vertex_flush(sna);
 	if (!sna->render.vertex_used) {
 		assert(sna->render.vbo == NULL);
 		assert(sna->render.vertices == sna->render.vertex_data);
@@ -1114,8 +1111,11 @@ static bool gen4_rectangle_begin(struct sna *sna,
 	return true;
 }
 
-static int gen4_get_rectangles__flush(struct sna *sna)
+static int gen4_get_rectangles__flush(struct sna *sna,
+				      const struct sna_composite_op *op)
 {
+	gen4_vertex_flush(sna, op);
+
 	if (!kgem_check_batch(&sna->kgem, 25))
 		return 0;
 	if (sna->kgem.nexec > KGEM_EXEC_SIZE(&sna->kgem) - 1)
@@ -1135,7 +1135,7 @@ inline static int gen4_get_rectangles(struct sna *sna,
 	if (rem < 3*op->floats_per_vertex) {
 		DBG(("flushing vbo for %s: %d < %d\n",
 		     __FUNCTION__, rem, 3*op->floats_per_vertex));
-		rem = gen4_get_rectangles__flush(sna);
+		rem = gen4_get_rectangles__flush(sna, op);
 		if (rem == 0)
 			return 0;
 	}
@@ -1262,6 +1262,7 @@ gen4_get_batch(struct sna *sna)
 		     __FUNCTION__, sna->kgem.surface - sna->kgem.nbatch,
 		     150, 4*8));
 		kgem_submit(&sna->kgem);
+		_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	}
 
 	if (sna->render_state.gen4.needs_invariant)
@@ -1513,7 +1514,7 @@ gen4_render_composite_blt(struct sna *sna,
 	op->prim_emit(sna, op, r);
 
 	/* XXX are the shaders fubar? */
-	FLUSH();
+	FLUSH(op);
 }
 
 fastcall static void
@@ -1751,7 +1752,7 @@ gen4_render_video(struct sna *sna,
 		OUT_VERTEX_F((box->x1 - dxo) * src_scale_x);
 		OUT_VERTEX_F((box->y1 - dyo) * src_scale_y);
 
-		FLUSH();
+		FLUSH(&tmp);
 
 		if (!DAMAGE_IS_ALL(priv->gpu_damage)) {
 			sna_damage_add_box(&priv->gpu_damage, &r);
@@ -1760,8 +1761,7 @@ gen4_render_video(struct sna *sna,
 		box++;
 	}
 
-	gen4_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	gen4_vertex_flush(sna, &tmp);
 	return TRUE;
 }
 
@@ -1876,12 +1876,10 @@ static void
 gen4_render_composite_done(struct sna *sna,
 			   const struct sna_composite_op *op)
 {
-	gen4_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
-	sna->render.op = NULL;
-
 	DBG(("%s()\n", __FUNCTION__));
 
+	gen4_vertex_flush(sna, op);
+
 	if (op->mask.bo)
 		kgem_bo_destroy(&sna->kgem, op->mask.bo);
 	if (op->src.bo)
@@ -2274,8 +2272,6 @@ gen4_render_composite(struct sna *sna,
 
 	gen4_bind_surfaces(sna, tmp);
 	gen4_align_vertex(sna, tmp);
-
-	sna->render.op = tmp;
 	return TRUE;
 
 cleanup_src:
@@ -2367,7 +2363,7 @@ gen4_render_copy_one(struct sna *sna,
 	OUT_VERTEX_F(sx*op->src.scale[0]);
 	OUT_VERTEX_F(sy*op->src.scale[1]);
 
-	FLUSH();
+	FLUSH(op);
 }
 
 static Bool
@@ -2452,8 +2448,6 @@ gen4_render_copy_boxes(struct sna *sna, uint8_t alu,
 				     box->x1, box->y1);
 		box++;
 	} while (--n);
-
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -2470,8 +2464,7 @@ gen4_render_copy_blt(struct sna *sna,
 static void
 gen4_render_copy_done(struct sna *sna, const struct sna_copy_op *op)
 {
-	gen4_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	gen4_vertex_flush(sna, &op->base);
 }
 
 static Bool
@@ -2599,7 +2592,7 @@ gen4_render_fill_rectangle(struct sna *sna,
 	OUT_VERTEX_F(0);
 	OUT_VERTEX_F(0);
 
-	FLUSH();
+	FLUSH(op);
 }
 
 static Bool
@@ -2706,7 +2699,6 @@ gen4_render_fill_boxes(struct sna *sna,
 	} while (--n);
 
 	kgem_bo_destroy(&sna->kgem, tmp.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -2744,9 +2736,8 @@ gen4_render_fill_op_boxes(struct sna *sna,
 static void
 gen4_render_fill_op_done(struct sna *sna, const struct sna_fill_op *op)
 {
-	gen4_vertex_flush(sna);
+	gen4_vertex_flush(sna, &op->base);
 	kgem_bo_destroy(&sna->kgem, op->base.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 }
 
 static Bool
@@ -2899,9 +2890,8 @@ gen4_render_fill_one(struct sna *sna, PixmapPtr dst, struct kgem_bo *bo,
 
 	gen4_render_fill_rectangle(sna, &tmp, x1, y1, x2 - x1, y2 - y1);
 
-	gen4_vertex_flush(sna);
+	gen4_vertex_flush(sna, &tmp);
 	kgem_bo_destroy(&sna->kgem, tmp.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 
 	return TRUE;
 }
diff --git a/src/sna/gen5_render.c b/src/sna/gen5_render.c
index 573478d3..560830e1 100644
--- a/src/sna/gen5_render.c
+++ b/src/sna/gen5_render.c
@@ -312,6 +312,8 @@ static void gen5_magic_ca_pass(struct sna *sna,
 	if (!op->need_magic_ca_pass)
 		return;
 
+	assert(sna->render.vertex_index > sna->render.vertex_start);
+
 	DBG(("%s: CA fixup\n", __FUNCTION__));
 
 	gen5_emit_pipelined_pointers
@@ -335,8 +337,8 @@ static void gen5_magic_ca_pass(struct sna *sna,
 
 static void gen5_vertex_flush(struct sna *sna)
 {
-	if (sna->render_state.gen5.vertex_offset == 0)
-		return;
+	assert(sna->render_state.gen5.vertex_offset);
+	assert(sna->render.vertex_index > sna->render.vertex_start);
 
 	DBG(("%s[%x] = %d\n", __FUNCTION__,
 	     4*sna->render_state.gen5.vertex_offset,
@@ -344,9 +346,6 @@ static void gen5_vertex_flush(struct sna *sna)
 	sna->kgem.batch[sna->render_state.gen5.vertex_offset] =
 		sna->render.vertex_index - sna->render.vertex_start;
 	sna->render_state.gen5.vertex_offset = 0;
-
-	if (sna->render.op)
-		gen5_magic_ca_pass(sna, sna->render.op);
 }
 
 static int gen5_vertex_finish(struct sna *sna)
@@ -354,9 +353,7 @@ static int gen5_vertex_finish(struct sna *sna)
 	struct kgem_bo *bo;
 	unsigned int i;
 
-	gen5_vertex_flush(sna);
-	if (!sna->render.vertex_used)
-		return sna->render.vertex_size;
+	assert(sna->render.vertex_used);
 
 	/* Note: we only need dword alignment (currently) */
 
@@ -414,7 +411,6 @@ static void gen5_vertex_close(struct sna *sna)
 	struct kgem_bo *bo;
 	unsigned int i, delta = 0;
 
-	gen5_vertex_flush(sna);
 	if (!sna->render.vertex_used) {
 		assert(sna->render.vbo == NULL);
 		assert(sna->render.vertices == sna->render.vertex_data);
@@ -1145,13 +1141,19 @@ static bool gen5_rectangle_begin(struct sna *sna,
 	return true;
 }
 
-static int gen5_get_rectangles__flush(struct sna *sna)
+static int gen5_get_rectangles__flush(struct sna *sna,
+				      const struct sna_composite_op *op)
 {
-	if (!kgem_check_batch(&sna->kgem, 25))
+	if (sna->render_state.gen5.vertex_offset) {
+		gen5_vertex_flush(sna);
+		gen5_magic_ca_pass(sna, op);
+	}
+
+	if (!kgem_check_batch(&sna->kgem, op->need_magic_ca_pass ? 20 : 6))
 		return 0;
 	if (sna->kgem.nexec > KGEM_EXEC_SIZE(&sna->kgem) - 1)
 		return 0;
-	if (sna->kgem.nreloc > KGEM_RELOC_SIZE(&sna->kgem) - 1)
+	if (sna->kgem.nreloc > KGEM_RELOC_SIZE(&sna->kgem) - 2)
 		return 0;
 
 	return gen5_vertex_finish(sna);
@@ -1166,7 +1168,7 @@ inline static int gen5_get_rectangles(struct sna *sna,
 	if (rem < op->floats_per_rect) {
 		DBG(("flushing vbo for %s: %d < %d\n",
 		     __FUNCTION__, rem, op->floats_per_rect));
-		rem = gen5_get_rectangles__flush(sna);
+		rem = gen5_get_rectangles__flush(sna, op);
 		if (rem == 0)
 			return 0;
 	}
@@ -1292,6 +1294,7 @@ gen5_get_batch(struct sna *sna)
 		     __FUNCTION__, sna->kgem.surface - sna->kgem.nbatch,
 		     150, 4*8));
 		kgem_submit(&sna->kgem);
+		_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	}
 
 	if (sna->render_state.gen5.needs_invariant)
@@ -1810,7 +1813,6 @@ gen5_render_video(struct sna *sna,
 	}
 
 	gen5_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -1919,9 +1921,10 @@ static void
 gen5_render_composite_done(struct sna *sna,
 			   const struct sna_composite_op *op)
 {
-	gen5_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
-	sna->render.op = NULL;
+	if (sna->render_state.gen5.vertex_offset) {
+		gen5_vertex_flush(sna);
+		gen5_magic_ca_pass(sna,op);
+	}
 
 	DBG(("%s()\n", __FUNCTION__));
 
@@ -2333,8 +2336,6 @@ gen5_render_composite(struct sna *sna,
 
 	gen5_bind_surfaces(sna, tmp);
 	gen5_align_vertex(sna, tmp);
-
-	sna->render.op = tmp;
 	return TRUE;
 
 cleanup_src:
@@ -2548,8 +2549,8 @@ fastcall static void
 gen5_render_composite_spans_done(struct sna *sna,
 				 const struct sna_composite_spans_op *op)
 {
-	gen5_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	if (sna->render_state.gen5.vertex_offset)
+		gen5_vertex_flush(sna);
 
 	DBG(("%s()\n", __FUNCTION__));
 
@@ -2796,7 +2797,6 @@ gen5_render_copy_boxes(struct sna *sna, uint8_t alu,
 	} while (n);
 
 	gen5_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -2832,8 +2832,8 @@ static void
 gen5_render_copy_done(struct sna *sna,
 		      const struct sna_copy_op *op)
 {
-	gen5_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	if (sna->render_state.gen5.vertex_offset)
+		gen5_vertex_flush(sna);
 
 	DBG(("%s()\n", __FUNCTION__));
 }
@@ -3071,7 +3071,6 @@ gen5_render_fill_boxes(struct sna *sna,
 
 	gen5_vertex_flush(sna);
 	kgem_bo_destroy(&sna->kgem, tmp.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -3164,9 +3163,9 @@ static void
 gen5_render_fill_op_done(struct sna *sna,
 			 const struct sna_fill_op *op)
 {
-	gen5_vertex_flush(sna);
+	if (sna->render_state.gen5.vertex_offset)
+		gen5_vertex_flush(sna);
 	kgem_bo_destroy(&sna->kgem, op->base.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 
 	DBG(("%s()\n", __FUNCTION__));
 }
@@ -3340,7 +3339,6 @@ gen5_render_fill_one(struct sna *sna, PixmapPtr dst, struct kgem_bo *bo,
 
 	gen5_vertex_flush(sna);
 	kgem_bo_destroy(&sna->kgem, tmp.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 
 	return TRUE;
 }
diff --git a/src/sna/gen6_render.c b/src/sna/gen6_render.c
index 1ec6f065..1e9d9304 100644
--- a/src/sna/gen6_render.c
+++ b/src/sna/gen6_render.c
@@ -896,8 +896,7 @@ static void gen6_magic_ca_pass(struct sna *sna,
 
 static void gen6_vertex_flush(struct sna *sna)
 {
-	if (sna->render_state.gen6.vertex_offset == 0)
-		return;
+	assert(sna->render_state.gen6.vertex_offset);
 
 	DBG(("%s[%x] = %d\n", __FUNCTION__,
 	     4*sna->render_state.gen6.vertex_offset,
@@ -905,9 +904,6 @@ static void gen6_vertex_flush(struct sna *sna)
 	sna->kgem.batch[sna->render_state.gen6.vertex_offset] =
 		sna->render.vertex_index - sna->render.vertex_start;
 	sna->render_state.gen6.vertex_offset = 0;
-
-	if (sna->render.op)
-		gen6_magic_ca_pass(sna, sna->render.op);
 }
 
 static int gen6_vertex_finish(struct sna *sna)
@@ -915,9 +911,7 @@ static int gen6_vertex_finish(struct sna *sna)
 	struct kgem_bo *bo;
 	unsigned int i;
 
-	gen6_vertex_flush(sna);
-	if (!sna->render.vertex_used)
-		return sna->render.vertex_size;
+	assert(sna->render.vertex_used);
 
 	/* Note: we only need dword alignment (currently) */
 
@@ -976,7 +970,6 @@ static void gen6_vertex_close(struct sna *sna)
 	struct kgem_bo *bo;
 	unsigned int i, delta = 0;
 
-	gen6_vertex_flush(sna);
 	if (!sna->render.vertex_used) {
 		assert(sna->render.vbo == NULL);
 		assert(sna->render.vertices == sna->render.vertex_data);
@@ -1541,13 +1534,19 @@ static bool gen6_rectangle_begin(struct sna *sna,
 	return true;
 }
 
-static int gen6_get_rectangles__flush(struct sna *sna, bool ca)
+static int gen6_get_rectangles__flush(struct sna *sna,
+				      const struct sna_composite_op *op)
 {
-	if (!kgem_check_batch(&sna->kgem, ca ? 65 : 5))
+	if (sna->render_state.gen6.vertex_offset) {
+		gen6_vertex_flush(sna);
+		gen6_magic_ca_pass(sna, op);
+	}
+
+	if (!kgem_check_batch(&sna->kgem, op->need_magic_ca_pass ? 65 : 5))
 		return 0;
 	if (sna->kgem.nexec > KGEM_EXEC_SIZE(&sna->kgem) - 1)
 		return 0;
-	if (sna->kgem.nreloc > KGEM_RELOC_SIZE(&sna->kgem) - 1)
+	if (sna->kgem.nreloc > KGEM_RELOC_SIZE(&sna->kgem) - 2)
 		return 0;
 
 	return gen6_vertex_finish(sna);
@@ -1562,7 +1561,7 @@ inline static int gen6_get_rectangles(struct sna *sna,
 	if (rem < op->floats_per_rect) {
 		DBG(("flushing vbo for %s: %d < %d\n",
 		     __FUNCTION__, rem, op->floats_per_rect));
-		rem = gen6_get_rectangles__flush(sna, op->need_magic_ca_pass);
+		rem = gen6_get_rectangles__flush(sna, op);
 		if (rem == 0)
 			return 0;
 	}
@@ -1613,6 +1612,7 @@ gen6_get_batch(struct sna *sna)
 		     __FUNCTION__, sna->kgem.surface - sna->kgem.nbatch,
 		     150, 4*8));
 		kgem_submit(&sna->kgem);
+		_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	}
 
 	if (sna->render_state.gen6.needs_invariant)
@@ -1666,6 +1666,7 @@ gen6_align_vertex(struct sna *sna, const struct sna_composite_op *op)
 	assert (sna->render_state.gen6.vertex_offset == 0);
 	if (op->floats_per_vertex != sna->render_state.gen6.floats_per_vertex) {
 		if (sna->render.vertex_size - sna->render.vertex_used < 2*op->floats_per_rect)
+			/* XXX propagate failure */
 			gen6_vertex_finish(sna);
 
 		DBG(("aligning vertex: was %d, now %d floats per vertex, %d->%d\n",
@@ -2000,7 +2001,6 @@ gen6_render_video(struct sna *sna,
 	}
 
 	gen6_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -2119,9 +2119,10 @@ static void gen6_composite_channel_convert(struct sna_composite_channel *channel
 static void gen6_render_composite_done(struct sna *sna,
 				       const struct sna_composite_op *op)
 {
-	gen6_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
-	sna->render.op = NULL;
+	if (sna->render_state.gen6.vertex_offset) {
+		gen6_vertex_flush(sna);
+		gen6_magic_ca_pass(sna, op);
+	}
 
 	if (op->mask.bo)
 		kgem_bo_destroy(&sna->kgem, op->mask.bo);
@@ -2546,8 +2547,6 @@ gen6_render_composite(struct sna *sna,
 
 	gen6_emit_composite_state(sna, tmp);
 	gen6_align_vertex(sna, tmp);
-
-	sna->render.op = tmp;
 	return TRUE;
 
 cleanup_src:
@@ -2841,8 +2840,8 @@ fastcall static void
 gen6_render_composite_spans_done(struct sna *sna,
 				 const struct sna_composite_spans_op *op)
 {
-	gen6_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	if (sna->render_state.gen6.vertex_offset)
+		gen6_vertex_flush(sna);
 
 	DBG(("%s()\n", __FUNCTION__));
 
@@ -3158,7 +3157,6 @@ gen6_render_copy_boxes(struct sna *sna, uint8_t alu,
 	} while (n);
 
 	gen6_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -3190,8 +3188,8 @@ gen6_render_copy_blt(struct sna *sna,
 static void
 gen6_render_copy_done(struct sna *sna, const struct sna_copy_op *op)
 {
-	gen6_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	if (sna->render_state.gen6.vertex_offset)
+		gen6_vertex_flush(sna);
 }
 
 static Bool
@@ -3439,7 +3437,6 @@ gen6_render_fill_boxes(struct sna *sna,
 
 	gen6_vertex_flush(sna);
 	kgem_bo_destroy(&sna->kgem, tmp.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -3531,9 +3528,9 @@ gen6_render_op_fill_boxes(struct sna *sna,
 static void
 gen6_render_op_fill_done(struct sna *sna, const struct sna_fill_op *op)
 {
-	gen6_vertex_flush(sna);
+	if (sna->render_state.gen6.vertex_offset)
+		gen6_vertex_flush(sna);
 	kgem_bo_destroy(&sna->kgem, op->base.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 }
 
 static Bool
@@ -3716,7 +3713,6 @@ gen6_render_fill_one(struct sna *sna, PixmapPtr dst, struct kgem_bo *bo,
 
 	gen6_vertex_flush(sna);
 	kgem_bo_destroy(&sna->kgem, tmp.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 
 	return TRUE;
 }
@@ -3812,7 +3808,6 @@ gen6_render_clear(struct sna *sna, PixmapPtr dst, struct kgem_bo *bo)
 
 	gen6_vertex_flush(sna);
 	kgem_bo_destroy(&sna->kgem, tmp.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 
 	return TRUE;
 }
diff --git a/src/sna/gen7_render.c b/src/sna/gen7_render.c
index 16c389c6..b14bc708 100644
--- a/src/sna/gen7_render.c
+++ b/src/sna/gen7_render.c
@@ -999,8 +999,7 @@ static void gen7_magic_ca_pass(struct sna *sna,
 
 static void gen7_vertex_flush(struct sna *sna)
 {
-	if (sna->render_state.gen7.vertex_offset == 0)
-		return;
+	assert(sna->render_state.gen7.vertex_offset);
 
 	DBG(("%s[%x] = %d\n", __FUNCTION__,
 	     4*sna->render_state.gen7.vertex_offset,
@@ -1008,9 +1007,6 @@ static void gen7_vertex_flush(struct sna *sna)
 	sna->kgem.batch[sna->render_state.gen7.vertex_offset] =
 		sna->render.vertex_index - sna->render.vertex_start;
 	sna->render_state.gen7.vertex_offset = 0;
-
-	if (sna->render.op)
-		gen7_magic_ca_pass(sna, sna->render.op);
 }
 
 static int gen7_vertex_finish(struct sna *sna)
@@ -1018,9 +1014,7 @@ static int gen7_vertex_finish(struct sna *sna)
 	struct kgem_bo *bo;
 	unsigned int i;
 
-	gen7_vertex_flush(sna);
-	if (!sna->render.vertex_used)
-		return sna->render.vertex_size;
+	assert(sna->render.vertex_used);
 
 	/* Note: we only need dword alignment (currently) */
 
@@ -1079,7 +1073,6 @@ static void gen7_vertex_close(struct sna *sna)
 	struct kgem_bo *bo;
 	unsigned int i, delta = 0;
 
-	gen7_vertex_flush(sna);
 	if (!sna->render.vertex_used) {
 		assert(sna->render.vbo == NULL);
 		assert(sna->render.vertices == sna->render.vertex_data);
@@ -1636,13 +1629,19 @@ static bool gen7_rectangle_begin(struct sna *sna,
 	return true;
 }
 
-static int gen7_get_rectangles__flush(struct sna *sna, bool ca)
+static int gen7_get_rectangles__flush(struct sna *sna,
+				      const struct sna_composite_op *op)
 {
-	if (!kgem_check_batch(&sna->kgem, ca ? 65 : 5))
+	if (sna->render_state.gen7.vertex_offset) {
+		gen7_vertex_flush(sna);
+		gen7_magic_ca_pass(sna, op);
+	}
+
+	if (!kgem_check_batch(&sna->kgem, op->need_magic_ca_pass ? 65 : 6))
 		return 0;
 	if (sna->kgem.nexec > KGEM_EXEC_SIZE(&sna->kgem) - 1)
 		return 0;
-	if (sna->kgem.nreloc > KGEM_RELOC_SIZE(&sna->kgem) - 1)
+	if (sna->kgem.nreloc > KGEM_RELOC_SIZE(&sna->kgem) - 2)
 		return 0;
 
 	return gen7_vertex_finish(sna);
@@ -1657,7 +1656,7 @@ inline static int gen7_get_rectangles(struct sna *sna,
 	if (rem < op->floats_per_rect) {
 		DBG(("flushing vbo for %s: %d < %d\n",
 		     __FUNCTION__, rem, op->floats_per_rect));
-		rem = gen7_get_rectangles__flush(sna, op->need_magic_ca_pass);
+		rem = gen7_get_rectangles__flush(sna, op);
 		if (rem == 0)
 			return 0;
 	}
@@ -1708,6 +1707,7 @@ gen7_get_batch(struct sna *sna)
 		     __FUNCTION__, sna->kgem.surface - sna->kgem.nbatch,
 		     150, 4*8));
 		kgem_submit(&sna->kgem);
+		_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	}
 
 	if (sna->render_state.gen7.needs_invariant)
@@ -2093,7 +2093,6 @@ gen7_render_video(struct sna *sna,
 	}
 
 	gen7_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -2212,9 +2211,10 @@ static void gen7_composite_channel_convert(struct sna_composite_channel *channel
 static void gen7_render_composite_done(struct sna *sna,
 				       const struct sna_composite_op *op)
 {
-	gen7_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
-	sna->render.op = NULL;
+	if (sna->render_state.gen7.vertex_offset) {
+		gen7_vertex_flush(sna);
+		gen7_magic_ca_pass(sna, op);
+	}
 
 	if (op->mask.bo)
 		kgem_bo_destroy(&sna->kgem, op->mask.bo);
@@ -2632,8 +2632,6 @@ gen7_render_composite(struct sna *sna,
 
 	gen7_emit_composite_state(sna, tmp);
 	gen7_align_vertex(sna, tmp);
-
-	sna->render.op = tmp;
 	return TRUE;
 
 cleanup_src:
@@ -2847,8 +2845,8 @@ fastcall static void
 gen7_render_composite_spans_done(struct sna *sna,
 				 const struct sna_composite_spans_op *op)
 {
-	gen7_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	if (sna->render_state.gen7.vertex_offset)
+		gen7_vertex_flush(sna);
 
 	DBG(("%s()\n", __FUNCTION__));
 
@@ -3118,7 +3116,6 @@ gen7_render_copy_boxes(struct sna *sna, uint8_t alu,
 	} while (n);
 
 	gen7_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -3150,8 +3147,8 @@ gen7_render_copy_blt(struct sna *sna,
 static void
 gen7_render_copy_done(struct sna *sna, const struct sna_copy_op *op)
 {
-	gen7_vertex_flush(sna);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
+	if (sna->render_state.gen7.vertex_offset)
+		gen7_vertex_flush(sna);
 }
 
 static Bool
@@ -3403,9 +3400,9 @@ gen7_render_fill_boxes(struct sna *sna,
 		} while (--n_this_time);
 	} while (n);
 
-	gen7_vertex_flush(sna);
+	if (sna->render_state.gen7.vertex_offset)
+		gen7_vertex_flush(sna);
 	kgem_bo_destroy(&sna->kgem, tmp.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 	return TRUE;
 }
 
@@ -3497,9 +3494,9 @@ gen7_render_fill_op_boxes(struct sna *sna,
 static void
 gen7_render_fill_op_done(struct sna *sna, const struct sna_fill_op *op)
 {
-	gen7_vertex_flush(sna);
+	if (sna->render_state.gen7.vertex_offset)
+		gen7_vertex_flush(sna);
 	kgem_bo_destroy(&sna->kgem, op->base.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 }
 
 static Bool
@@ -3682,7 +3679,6 @@ gen7_render_fill_one(struct sna *sna, PixmapPtr dst, struct kgem_bo *bo,
 
 	gen7_vertex_flush(sna);
 	kgem_bo_destroy(&sna->kgem, tmp.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 
 	return TRUE;
 }
@@ -3778,7 +3774,6 @@ gen7_render_clear(struct sna *sna, PixmapPtr dst, struct kgem_bo *bo)
 
 	gen7_vertex_flush(sna);
 	kgem_bo_destroy(&sna->kgem, tmp.src.bo);
-	_kgem_set_mode(&sna->kgem, KGEM_RENDER);
 
 	return TRUE;
 }
diff --git a/src/sna/sna_render.h b/src/sna/sna_render.h
index abb19dce..04543dec 100644
--- a/src/sna/sna_render.h
+++ b/src/sna/sna_render.h
@@ -278,7 +278,6 @@ struct sna_render {
 	float *vertices;
 
 	float vertex_data[1024];
-	const struct sna_composite_op *op;
 };
 
 struct gen2_render_state {