sna/gen3+: Keep the vertex buffer resident between batches

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

4 files changed, 173 insertions, 78 deletions

diff --git a/src/sna/gen3_render.c b/src/sna/gen3_render.c
index b50d0671..8f597cfa 100644
--- a/src/sna/gen3_render.c
+++ b/src/sna/gen3_render.c
@@ -1612,20 +1612,33 @@ static int gen3_vertex_finish(struct sna *sna)
 
 static void gen3_vertex_close(struct sna *sna)
 {
-	struct kgem_bo *bo;
-	int delta = 0;
+	struct kgem_bo *bo, *free_bo = NULL;
+	unsigned int delta = 0;
+
+	assert(sna->render_state.gen3.vertex_offset == 0);
 
-	if (!sna->render.vertex_used) {
+	DBG(("%s: used=%d, vbo active? %d\n",
+	     __FUNCTION__, sna->render.vertex_used, sna->render.vbo != NULL));
+
+	if (sna->render.vertex_used == 0) {
 		assert(sna->render.vbo == NULL);
 		assert(sna->render.vertices == sna->render.vertex_data);
 		assert(sna->render.vertex_size == ARRAY_SIZE(sna->render.vertex_data));
 		return;
 	}
 
-	DBG(("%s: used=%d\n", __FUNCTION__, sna->render.vertex_used));
-
 	bo = sna->render.vbo;
-	if (bo == NULL) {
+	if (bo) {
+		if (IS_CPU_MAP(bo->map) ||
+		    sna->render.vertex_size - sna->render.vertex_used < 64) {
+			DBG(("%s: discarding vbo (was CPU mapped)\n",
+			     __FUNCTION__));
+			sna->render.vbo = NULL;
+			sna->render.vertices = sna->render.vertex_data;
+			sna->render.vertex_size = ARRAY_SIZE(sna->render.vertex_data);
+			free_bo = bo;
+		}
+	} else {
 		if (sna->kgem.nbatch + sna->render.vertex_used <= sna->kgem.surface) {
 			DBG(("%s: copy to batch: %d @ %d\n", __FUNCTION__,
 			     sna->render.vertex_used, sna->kgem.nbatch));
@@ -1636,36 +1649,37 @@ static void gen3_vertex_close(struct sna *sna)
 			bo = NULL;
 			sna->kgem.nbatch += sna->render.vertex_used;
 		} else {
-			bo = kgem_create_linear(&sna->kgem,
-						4*sna->render.vertex_used);
-			if (bo && !kgem_bo_write(&sna->kgem, bo,
-						 sna->render.vertex_data,
-						 4*sna->render.vertex_used)) {
-				kgem_bo_destroy(&sna->kgem, bo);
-				goto reset;
-			}
 			DBG(("%s: new vbo: %d\n", __FUNCTION__,
 			     sna->render.vertex_used));
+			bo = kgem_create_linear(&sna->kgem,
+						4*sna->render.vertex_used);
+			if (bo)
+				kgem_bo_write(&sna->kgem, bo,
+					      sna->render.vertex_data,
+					      4*sna->render.vertex_used);
+			free_bo = bo;
 		}
 	}
 
 	DBG(("%s: reloc = %d\n", __FUNCTION__,
 	     sna->render.vertex_reloc[0]));
 
-	sna->kgem.batch[sna->render.vertex_reloc[0]] =
-		kgem_add_reloc(&sna->kgem, sna->render.vertex_reloc[0],
-			       bo, I915_GEM_DOMAIN_VERTEX << 16, delta);
-	if (bo)
-		kgem_bo_destroy(&sna->kgem, bo);
+	if (sna->render.vertex_reloc[0]) {
+		sna->kgem.batch[sna->render.vertex_reloc[0]] =
+			kgem_add_reloc(&sna->kgem, sna->render.vertex_reloc[0],
+				       bo, I915_GEM_DOMAIN_VERTEX << 16, delta);
+		sna->render.vertex_reloc[0] = 0;
+	}
 
-reset:
-	sna->render.vertex_reloc[0] = 0;
-	sna->render.vertex_used = 0;
-	sna->render.vertex_index = 0;
+	if (sna->render.vbo == NULL) {
+		sna->render.vertex_used = 0;
+		sna->render.vertex_index = 0;
+		assert(sna->render.vertices == sna->render.vertex_data);
+		assert(sna->render.vertex_size == ARRAY_SIZE(sna->render.vertex_data));
+	}
 
-	sna->render.vbo = NULL;
-	sna->render.vertices = sna->render.vertex_data;
-	sna->render.vertex_size = ARRAY_SIZE(sna->render.vertex_data);
+	if (free_bo)
+		kgem_bo_destroy(&sna->kgem, free_bo);
 }
 
 static bool gen3_rectangle_begin(struct sna *sna,
@@ -1885,10 +1899,23 @@ gen3_render_reset(struct sna *sna)
 	state->last_floats_per_vertex = 0;
 	state->last_vertex_offset = 0;
 	state->vertex_offset = 0;
+}
 
-	assert(sna->render.vertex_used == 0);
-	assert(sna->render.vertex_index == 0);
-	assert(sna->render.vertex_reloc[0] == 0);
+static void
+gen3_render_retire(struct kgem *kgem)
+{
+	struct sna *sna;
+
+	sna = container_of(kgem, struct sna, kgem);
+	if (!kgem->need_retire && kgem->nbatch == 0 && sna->render.vbo) {
+		DBG(("%s: discarding vbo\n", __FUNCTION__));
+		kgem_bo_destroy(kgem, sna->render.vbo);
+		sna->render.vbo = NULL;
+		sna->render.vertices = sna->render.vertex_data;
+		sna->render.vertex_size = ARRAY_SIZE(sna->render.vertex_data);
+		sna->render.vertex_used = 0;
+		sna->render.vertex_index = 0;
+	}
 }
 
 static Bool gen3_composite_channel_set_format(struct sna_composite_channel *channel,
@@ -4466,5 +4493,7 @@ Bool gen3_render_init(struct sna *sna)
 
 	render->max_3d_size = MAX_3D_SIZE;
 	render->max_3d_pitch = MAX_3D_PITCH;
+
+	sna->kgem.retire = gen3_render_retire;
 	return TRUE;
 }
diff --git a/src/sna/gen5_render.c b/src/sna/gen5_render.c
index a80ce0a4..bcba0d87 100644
--- a/src/sna/gen5_render.c
+++ b/src/sna/gen5_render.c
@@ -410,11 +410,14 @@ static int gen5_vertex_finish(struct sna *sna)
 
 static void gen5_vertex_close(struct sna *sna)
 {
-	struct kgem_bo *bo;
+	struct kgem_bo *bo, *free_bo = NULL;
 	unsigned int i, delta = 0;
 
 	assert(sna->render_state.gen5.vertex_offset == 0);
 
+	DBG(("%s: used=%d, vbo active? %d\n",
+	     __FUNCTION__, sna->render.vertex_used, sna->render.vbo != NULL));
+
 	if (!sna->render.vertex_used) {
 		assert(sna->render.vbo == NULL);
 		assert(sna->render.vertices == sna->render.vertex_data);
@@ -422,10 +425,18 @@ static void gen5_vertex_close(struct sna *sna)
 		return;
 	}
 
-	DBG(("%s: used=%d\n", __FUNCTION__, sna->render.vertex_used));
-
 	bo = sna->render.vbo;
-	if (bo == NULL) {
+	if (bo) {
+		if (IS_CPU_MAP(bo->map) ||
+		    sna->render.vertex_size - sna->render.vertex_used < 64) {
+			DBG(("%s: discarding vbo (was CPU mapped)\n",
+			     __FUNCTION__));
+			sna->render.vbo = NULL;
+			sna->render.vertices = sna->render.vertex_data;
+			sna->render.vertex_size = ARRAY_SIZE(sna->render.vertex_data);
+			free_bo = bo;
+		}
+	} else {
 		if (sna->kgem.nbatch + sna->render.vertex_used <= sna->kgem.surface) {
 			DBG(("%s: copy to batch: %d @ %d\n", __FUNCTION__,
 			     sna->render.vertex_used, sna->kgem.nbatch));
@@ -441,10 +452,11 @@ static void gen5_vertex_close(struct sna *sna)
 						 sna->render.vertex_data,
 						 4*sna->render.vertex_used)) {
 				kgem_bo_destroy(&sna->kgem, bo);
-				goto reset;
+				bo = NULL;
 			}
 			DBG(("%s: new vbo: %d\n", __FUNCTION__,
 			     sna->render.vertex_used));
+			free_bo = bo;
 		}
 	}
 
@@ -469,17 +481,13 @@ static void gen5_vertex_close(struct sna *sna)
 		}
 	}
 
-	if (bo)
-		kgem_bo_destroy(&sna->kgem, bo);
-
-reset:
-	sna->render.vertex_used = 0;
-	sna->render.vertex_index = 0;
-	sna->render_state.gen5.vb_id = 0;
+	if (sna->render.vbo == NULL) {
+		sna->render.vertex_used = 0;
+		sna->render.vertex_index = 0;
+	}
 
-	sna->render.vbo = NULL;
-	sna->render.vertices = sna->render.vertex_data;
-	sna->render.vertex_size = ARRAY_SIZE(sna->render.vertex_data);
+	if (free_bo)
+		kgem_bo_destroy(&sna->kgem, free_bo);
 }
 
 static uint32_t gen5_get_blend(int op,
@@ -3470,6 +3478,23 @@ gen5_render_context_switch(struct kgem *kgem,
 	}
 }
 
+static void
+gen5_render_retire(struct kgem *kgem)
+{
+	struct sna *sna;
+
+	sna = container_of(kgem, struct sna, kgem);
+	if (!kgem->need_retire && kgem->nbatch == 0 && sna->render.vbo) {
+		DBG(("%s: discarding vbo\n", __FUNCTION__));
+		kgem_bo_destroy(kgem, sna->render.vbo);
+		sna->render.vbo = NULL;
+		sna->render.vertices = sna->render.vertex_data;
+		sna->render.vertex_size = ARRAY_SIZE(sna->render.vertex_data);
+		sna->render.vertex_used = 0;
+		sna->render.vertex_index = 0;
+	}
+}
+
 static void gen5_render_reset(struct sna *sna)
 {
 	sna->render_state.gen5.needs_invariant = TRUE;
@@ -3730,6 +3755,7 @@ Bool gen5_render_init(struct sna *sna)
 		return FALSE;
 
 	sna->kgem.context_switch = gen5_render_context_switch;
+	sna->kgem.retire = gen5_render_retire;
 
 	sna->render.composite = gen5_render_composite;
 #if !NO_COMPOSITE_SPANS
diff --git a/src/sna/gen6_render.c b/src/sna/gen6_render.c
index b69b3a21..439fb521 100644
--- a/src/sna/gen6_render.c
+++ b/src/sna/gen6_render.c
@@ -989,9 +989,14 @@ static int gen6_vertex_finish(struct sna *sna)
 
 static void gen6_vertex_close(struct sna *sna)
 {
-	struct kgem_bo *bo;
+	struct kgem_bo *bo, *free_bo = NULL;
 	unsigned int i, delta = 0;
 
+	assert(sna->render_state.gen6.vertex_offset == 0);
+
+	DBG(("%s: used=%d, vbo active? %d\n",
+	     __FUNCTION__, sna->render.vertex_used, sna->render.vbo != NULL));
+
 	if (!sna->render.vertex_used) {
 		assert(sna->render.vbo == NULL);
 		assert(sna->render.vertices == sna->render.vertex_data);
@@ -999,13 +1004,16 @@ static void gen6_vertex_close(struct sna *sna)
 		return;
 	}
 
-	DBG(("%s: used=%d / %d\n", __FUNCTION__,
-	     sna->render.vertex_used, sna->render.vertex_size));
-
 	bo = sna->render.vbo;
-	if (bo == NULL) {
-		assert(sna->render.vertices == sna->render.vertex_data);
-		assert(sna->render.vertex_used < ARRAY_SIZE(sna->render.vertex_data));
+	if (bo) {
+		if (sna->render.vertex_size - sna->render.vertex_used < 64) {
+			DBG(("%s: discarding vbo (full)\n", __FUNCTION__));
+			sna->render.vbo = NULL;
+			sna->render.vertices = sna->render.vertex_data;
+			sna->render.vertex_size = ARRAY_SIZE(sna->render.vertex_data);
+			free_bo = bo;
+		}
+	} else {
 		if (sna->kgem.nbatch + sna->render.vertex_used <= sna->kgem.surface) {
 			DBG(("%s: copy to batch: %d @ %d\n", __FUNCTION__,
 			     sna->render.vertex_used, sna->kgem.nbatch));
@@ -1021,10 +1029,11 @@ static void gen6_vertex_close(struct sna *sna)
 						 sna->render.vertex_data,
 						 4*sna->render.vertex_used)) {
 				kgem_bo_destroy(&sna->kgem, bo);
-				goto reset;
+				bo = NULL;
 			}
 			DBG(("%s: new vbo: %d\n", __FUNCTION__,
 			     sna->render.vertex_used));
+			free_bo = bo;
 		}
 	}
 
@@ -1049,17 +1058,15 @@ static void gen6_vertex_close(struct sna *sna)
 		}
 	}
 
-	if (bo)
-		kgem_bo_destroy(&sna->kgem, bo);
-
-reset:
-	sna->render.vertex_used = 0;
-	sna->render.vertex_index = 0;
-	sna->render_state.gen6.vb_id = 0;
+	if (sna->render.vbo == NULL) {
+		sna->render.vertex_used = 0;
+		sna->render.vertex_index = 0;
+		assert(sna->render.vertices == sna->render.vertex_data);
+		assert(sna->render.vertex_size == ARRAY_SIZE(sna->render.vertex_data));
+	}
 
-	sna->render.vbo = NULL;
-	sna->render.vertices = sna->render.vertex_data;
-	sna->render.vertex_size = ARRAY_SIZE(sna->render.vertex_data);
+	if (free_bo)
+		kgem_bo_destroy(&sna->kgem, free_bo);
 }
 
 typedef struct gen6_surface_state_padded {
@@ -4095,8 +4102,21 @@ gen6_render_context_switch(struct kgem *kgem,
 static void
 gen6_render_retire(struct kgem *kgem)
 {
+	struct sna *sna;
+
 	if (kgem->ring && (kgem->has_semaphores || !kgem->need_retire))
 		kgem->ring = kgem->mode;
+
+	sna = container_of(kgem, struct sna, kgem);
+	if (!kgem->need_retire && kgem->nbatch == 0 && sna->render.vbo) {
+		DBG(("%s: discarding vbo\n", __FUNCTION__));
+		kgem_bo_destroy(kgem, sna->render.vbo);
+		sna->render.vbo = NULL;
+		sna->render.vertices = sna->render.vertex_data;
+		sna->render.vertex_size = ARRAY_SIZE(sna->render.vertex_data);
+		sna->render.vertex_used = 0;
+		sna->render.vertex_index = 0;
+	}
 }
 
 static void gen6_render_reset(struct sna *sna)
diff --git a/src/sna/gen7_render.c b/src/sna/gen7_render.c
index 0d913f6f..e3d9757d 100644
--- a/src/sna/gen7_render.c
+++ b/src/sna/gen7_render.c
@@ -1086,9 +1086,14 @@ static int gen7_vertex_finish(struct sna *sna)
 
 static void gen7_vertex_close(struct sna *sna)
 {
-	struct kgem_bo *bo;
+	struct kgem_bo *bo, *free_bo = NULL;
 	unsigned int i, delta = 0;
 
+	assert(sna->render_state.gen7.vertex_offset == 0);
+
+	DBG(("%s: used=%d, vbo active? %d\n",
+	     __FUNCTION__, sna->render.vertex_used, sna->render.vbo != NULL));
+
 	if (!sna->render.vertex_used) {
 		assert(sna->render.vbo == NULL);
 		assert(sna->render.vertices == sna->render.vertex_data);
@@ -1096,11 +1101,16 @@ static void gen7_vertex_close(struct sna *sna)
 		return;
 	}
 
-	DBG(("%s: used=%d / %d\n", __FUNCTION__,
-	     sna->render.vertex_used, sna->render.vertex_size));
-
 	bo = sna->render.vbo;
-	if (bo == NULL) {
+	if (bo) {
+		if (sna->render.vertex_size - sna->render.vertex_used < 64) {
+			DBG(("%s: discarding vbo (full)\n", __FUNCTION__));
+			sna->render.vbo = NULL;
+			sna->render.vertices = sna->render.vertex_data;
+			sna->render.vertex_size = ARRAY_SIZE(sna->render.vertex_data);
+			free_bo = bo;
+		}
+	} else {
 		if (sna->kgem.nbatch + sna->render.vertex_used <= sna->kgem.surface) {
 			DBG(("%s: copy to batch: %d @ %d\n", __FUNCTION__,
 			     sna->render.vertex_used, sna->kgem.nbatch));
@@ -1116,10 +1126,11 @@ static void gen7_vertex_close(struct sna *sna)
 						 sna->render.vertex_data,
 						 4*sna->render.vertex_used)) {
 				kgem_bo_destroy(&sna->kgem, bo);
-				goto reset;
+				bo = NULL;
 			}
 			DBG(("%s: new vbo: %d\n", __FUNCTION__,
 			     sna->render.vertex_used));
+			free_bo = bo;
 		}
 	}
 
@@ -1144,17 +1155,13 @@ static void gen7_vertex_close(struct sna *sna)
 		}
 	}
 
-	if (bo)
-		kgem_bo_destroy(&sna->kgem, bo);
-
-reset:
-	sna->render.vertex_used = 0;
-	sna->render.vertex_index = 0;
-	sna->render_state.gen7.vb_id = 0;
+	if (sna->render.vbo == NULL) {
+		sna->render.vertex_used = 0;
+		sna->render.vertex_index = 0;
+	}
 
-	sna->render.vbo = NULL;
-	sna->render.vertices = sna->render.vertex_data;
-	sna->render.vertex_size = ARRAY_SIZE(sna->render.vertex_data);
+	if (free_bo)
+		kgem_bo_destroy(&sna->kgem, free_bo);
 }
 
 static void null_create(struct sna_static_stream *stream)
@@ -4080,8 +4087,21 @@ gen7_render_context_switch(struct kgem *kgem,
 static void
 gen7_render_retire(struct kgem *kgem)
 {
+	struct sna *sna;
+
 	if (kgem->ring && (kgem->has_semaphores || !kgem->need_retire))
 		kgem->ring = kgem->mode;
+
+	sna = container_of(kgem, struct sna, kgem);
+	if (!kgem->need_retire && kgem->nbatch == 0 && sna->render.vbo) {
+		DBG(("%s: discarding vbo\n", __FUNCTION__));
+		kgem_bo_destroy(kgem, sna->render.vbo);
+		sna->render.vbo = NULL;
+		sna->render.vertices = sna->render.vertex_data;
+		sna->render.vertex_size = ARRAY_SIZE(sna->render.vertex_data);
+		sna->render.vertex_used = 0;
+		sna->render.vertex_index = 0;
+	}
 }
 
 static void gen7_render_reset(struct sna *sna)