drm/i915: Use the reloc.handle as an index into the execbuffer array

Using copywinwin10 as an example that is dependent upon emitting a lot of relocations (2 per operation), we see improvements of: c2d/gm45: 618000.0/sec to 623000.0/sec. i3-330m: 748000.0/sec to 789000.0/sec. (measured relative to a baseline with neither optimisations applied).

Chris Wilson committed 13 years ago

Commit eef90cc1 parent ed5982e3 files · +71 −40Browse files

Unified · Split

3 files

drivers/gpu/drm/i915/i915_dma.c +3 −0
drivers/gpu/drm/i915/i915_gem_execbuffer.c +61 −39
include/uapi/drm/i915_drm.h +7 −1

drivers/gpu/drm/i915/i915_dma.c+3−0

 	case I915_PARAM_HAS_EXEC_NO_RELOC:
 		value = 1;
 		break;
+	case I915_PARAM_HAS_EXEC_HANDLE_LUT:
+		value = 1;
+		break;
 	default:
 		DRM_DEBUG_DRIVER("Unknown parameter %d\n",
 				 param->param);

drivers/gpu/drm/i915/i915_gem_execbuffer.c+61−39

@@ -36,24 +36,40 @@ · 35 unmodified lines
36	struct eb_objects {	36	struct eb_objects {
37	struct list_head objects;	37	struct list_head objects;
38	int and;	38	int and;
39	struct hlist_head buckets[0];	39	union {
		40	struct drm_i915_gem_object *lut[0];
		41	struct hlist_head buckets[0];
		42	};
40	};	43	};
41		44
42	static struct eb_objects *	45	static struct eb_objects *
43	eb_create(int size)	46	eb_create(struct drm_i915_gem_execbuffer2 *args)
44	{	47	{
45	struct eb_objects *eb;	48	struct eb_objects *eb = NULL;
46	int count = PAGE_SIZE / sizeof(struct hlist_head) / 2;	49
47	BUILD_BUG_ON(!is_power_of_2(PAGE_SIZE / sizeof(struct hlist_head)));	50	if (args->flags & I915_EXEC_HANDLE_LUT) {
48	while (count > size)	51	int size = args->buffer_count;
49	count >>= 1;	52	size = sizeof(struct drm_i915_gem_object );
50	eb = kzalloc(count*sizeof(struct hlist_head) +	53	size += sizeof(struct eb_objects);
51	sizeof(struct eb_objects),	54	eb = kmalloc(size, GFP_TEMPORARY \| __GFP_NOWARN \| __GFP_NORETRY);
52	GFP_KERNEL);	55	}
53	if (eb == NULL)	56
54	return eb;	57	if (eb == NULL) {
55		58	int size = args->buffer_count;
56	eb->and = count - 1;	59	int count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
		60	BUILD_BUG_ON(!is_power_of_2(PAGE_SIZE / sizeof(struct hlist_head)));
		61	while (count > 2*size)
		62	count >>= 1;
		63	eb = kzalloc(count*sizeof(struct hlist_head) +
		64	sizeof(struct eb_objects),
		65	GFP_TEMPORARY);
		66	if (eb == NULL)
		67	return eb;
		68
		69	eb->and = count - 1;
		70	} else
		71	eb->and = -args->buffer_count;
		72
57	INIT_LIST_HEAD(&eb->objects);	73	INIT_LIST_HEAD(&eb->objects);
58	return eb;	74	return eb;
59	}	75	}
@@ -61,26 +77,20 @@ eb_create(int size) · 1 unmodified line
61	static void	77	static void
62	eb_reset(struct eb_objects *eb)	78	eb_reset(struct eb_objects *eb)
63	{	79	{
64	memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));	80	if (eb->and >= 0)
65	}	81	memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
66
67	static void
68	eb_add_object(struct eb_objects eb, struct drm_i915_gem_object obj)
69	{
70	hlist_add_head(&obj->exec_node,
71	&eb->buckets[obj->exec_handle & eb->and]);
72	}	82	}
73		83
74	static int	84	static int
75	eb_lookup_objects(struct eb_objects *eb,	85	eb_lookup_objects(struct eb_objects *eb,
76	struct drm_i915_gem_exec_object2 *exec,	86	struct drm_i915_gem_exec_object2 *exec,
77	int count,	87	const struct drm_i915_gem_execbuffer2 *args,
78	struct drm_file *file)	88	struct drm_file *file)
79	{	89	{
80	int i;	90	int i;
81		91
82	spin_lock(&file->table_lock);	92	spin_lock(&file->table_lock);
83	for (i = 0; i < count; i++) {	93	for (i = 0; i < args->buffer_count; i++) {
84	struct drm_i915_gem_object *obj;	94	struct drm_i915_gem_object *obj;
85		95
86	obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle));	96	obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle));
@@ -101,9 +111,15 @@ eb_lookup_objects(struct eb_objects *eb, · 14 unmodified lines
101	drm_gem_object_reference(&obj->base);	111	drm_gem_object_reference(&obj->base);
102	list_add_tail(&obj->exec_list, &eb->objects);	112	list_add_tail(&obj->exec_list, &eb->objects);
103		113
104	obj->exec_handle = exec[i].handle;
105	obj->exec_entry = &exec[i];	114	obj->exec_entry = &exec[i];
106	eb_add_object(eb, obj);	115	if (eb->and < 0) {
		116	eb->lut[i] = obj;
		117	} else {
		118	uint32_t handle = args->flags & I915_EXEC_HANDLE_LUT ? i : exec[i].handle;
		119	obj->exec_handle = handle;
		120	hlist_add_head(&obj->exec_node,
		121	&eb->buckets[handle & eb->and]);
		122	}
107	}	123	}
108	spin_unlock(&file->table_lock);	124	spin_unlock(&file->table_lock);
109		125
@@ -113,18 +129,24 @@ eb_lookup_objects(struct eb_objects *eb, · 3 unmodified lines
113	static struct drm_i915_gem_object *	129	static struct drm_i915_gem_object *
114	eb_get_object(struct eb_objects *eb, unsigned long handle)	130	eb_get_object(struct eb_objects *eb, unsigned long handle)
115	{	131	{
116	struct hlist_head *head;	132	if (eb->and < 0) {
117	struct hlist_node *node;	133	if (handle >= -eb->and)
118	struct drm_i915_gem_object *obj;	134	return NULL;
		135	return eb->lut[handle];
		136	} else {
		137	struct hlist_head *head;
		138	struct hlist_node *node;
119		139
120	head = &eb->buckets[handle & eb->and];	140	head = &eb->buckets[handle & eb->and];
121	hlist_for_each(node, head) {	141	hlist_for_each(node, head) {
122	obj = hlist_entry(node, struct drm_i915_gem_object, exec_node);	142	struct drm_i915_gem_object *obj;
123	if (obj->exec_handle == handle)
124	return obj;
125	}
126		143
127	return NULL;	144	obj = hlist_entry(node, struct drm_i915_gem_object, exec_node);
		145	if (obj->exec_handle == handle)
		146	return obj;
		147	}
		148	return NULL;
		149	}
128	}	150	}
129		151
130	static void	152	static void
@@ -615,7 +637,7 @@ i915_gem_execbuffer_relocate_slow(struct drm_device *dev, · 484 unmodified lines
615		637
616	/* reacquire the objects */	638	/* reacquire the objects */
617	eb_reset(eb);	639	eb_reset(eb);
618	ret = eb_lookup_objects(eb, exec, count, file);	640	ret = eb_lookup_objects(eb, exec, args, file);
619	if (ret)	641	if (ret)
620	goto err;	642	goto err;
621		643
@@ -919,7 +941,7 @@ i915_gem_do_execbuffer(struct drm_device dev, void data, · 297 unmodified lines
919	goto pre_mutex_err;	941	goto pre_mutex_err;
920	}	942	}
921		943
922	eb = eb_create(args->buffer_count);	944	eb = eb_create(args);
923	if (eb == NULL) {	945	if (eb == NULL) {
924	mutex_unlock(&dev->struct_mutex);	946	mutex_unlock(&dev->struct_mutex);
925	ret = -ENOMEM;	947	ret = -ENOMEM;
@@ -927,7 +949,7 @@ i915_gem_do_execbuffer(struct drm_device dev, void data, · 1 unmodified line
927	}	949	}
928		950
929	/* Look up object handles */	951	/* Look up object handles */
930	ret = eb_lookup_objects(eb, exec, args->buffer_count, file);	952	ret = eb_lookup_objects(eb, exec, args, file);
931	if (ret)	953	if (ret)
932	goto err;	954	goto err;
933		955

include/uapi/drm/i915_drm.h+7−1

 #define I915_PARAM_HAS_SECURE_BATCHES	 23
 #define I915_PARAM_HAS_PINNED_BATCHES	 24
 #define I915_PARAM_HAS_EXEC_NO_RELOC	 25
+#define I915_PARAM_HAS_EXEC_HANDLE_LUT   26
 typedef struct drm_i915_getparam {
 	int param;
  */
 #define I915_EXEC_NO_RELOC		(1<<11)
-#define __I915_EXEC_UNKNOWN_FLAGS -(I915_EXEC_NO_RELOC<<1)
+/** Use the reloc.handle as an index into the exec object array rather
+ * than as the per-file handle.
+ */
+#define I915_EXEC_HANDLE_LUT		(1<<12)
+#define __I915_EXEC_UNKNOWN_FLAGS -(I915_EXEC_HANDLE_LUT<<1)
 #define I915_EXEC_CONTEXT_ID_MASK	(0xffffffff)
 #define i915_execbuffer2_set_context_id(eb2, context) \

drm/i915: Use the reloc.handle as an index into the execbuffer array

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code