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

@@ -995,6 +995,9 @@ static int i915_getparam(struct drm_device dev, void data, · 994 unmodified lines
995	995	case I915_PARAM_HAS_EXEC_NO_RELOC:
996	996	value = 1;
997	997	break;
	998	+ case I915_PARAM_HAS_EXEC_HANDLE_LUT:
	999	+ value = 1;
	1000	+ break;
998	1001	default:
999	1002	DRM_DEBUG_DRIVER("Unknown parameter %d\n",
1000	1003	param->param);

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

@@ -36,24 +36,40 @@ · 35 unmodified lines
36	36	struct eb_objects {
37	37	struct list_head objects;
38	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	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;
46		− int count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
47		− BUILD_BUG_ON(!is_power_of_2(PAGE_SIZE / sizeof(struct hlist_head)));
48		− while (count > size)
49		− count >>= 1;
50		− eb = kzalloc(count*sizeof(struct hlist_head) +
51		− sizeof(struct eb_objects),
52		− GFP_KERNEL);
53		− if (eb == NULL)
54		− return eb;
55		−
56		− eb->and = count - 1;
	48	+ struct eb_objects *eb = NULL;
	49	+
	50	+ if (args->flags & I915_EXEC_HANDLE_LUT) {
	51	+ int size = args->buffer_count;
	52	+ size = sizeof(struct drm_i915_gem_object );
	53	+ size += sizeof(struct eb_objects);
	54	+ eb = kmalloc(size, GFP_TEMPORARY \| __GFP_NOWARN \| __GFP_NORETRY);
	55	+ }
	56	+
	57	+ if (eb == NULL) {
	58	+ int size = args->buffer_count;
	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	73	INIT_LIST_HEAD(&eb->objects);
58	74	return eb;
59	75	}
@@ -61,26 +77,20 @@ eb_create(int size) · 1 unmodified line
61	77	static void
62	78	eb_reset(struct eb_objects *eb)
63	79	{
64		− memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
65		−}
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]);
	80	+ if (eb->and >= 0)
	81	+ memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
72	82	}
73	83
74	84	static int
75	85	eb_lookup_objects(struct eb_objects *eb,
76	86	struct drm_i915_gem_exec_object2 *exec,
77		− int count,
	87	+ const struct drm_i915_gem_execbuffer2 *args,
78	88	struct drm_file *file)
79	89	{
80	90	int i;
81	91
82	92	spin_lock(&file->table_lock);
83		− for (i = 0; i < count; i++) {
	93	+ for (i = 0; i < args->buffer_count; i++) {
84	94	struct drm_i915_gem_object *obj;
85	95
86	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	111	drm_gem_object_reference(&obj->base);
102	112	list_add_tail(&obj->exec_list, &eb->objects);
103	113
104		− obj->exec_handle = exec[i].handle;
105	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	124	spin_unlock(&file->table_lock);
109	125
@@ -113,18 +129,24 @@ eb_lookup_objects(struct eb_objects *eb, · 3 unmodified lines
113	129	static struct drm_i915_gem_object *
114	130	eb_get_object(struct eb_objects *eb, unsigned long handle)
115	131	{
116		− struct hlist_head *head;
117		− struct hlist_node *node;
118		− struct drm_i915_gem_object *obj;
	132	+ if (eb->and < 0) {
	133	+ if (handle >= -eb->and)
	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];
121		− hlist_for_each(node, head) {
122		− obj = hlist_entry(node, struct drm_i915_gem_object, exec_node);
123		− if (obj->exec_handle == handle)
124		− return obj;
125		− }
	140	+ head = &eb->buckets[handle & eb->and];
	141	+ hlist_for_each(node, head) {
	142	+ struct drm_i915_gem_object *obj;
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	152	static void
@@ -615,7 +637,7 @@ i915_gem_execbuffer_relocate_slow(struct drm_device *dev, · 484 unmodified lines
615	637
616	638	/* reacquire the objects */
617	639	eb_reset(eb);
618		− ret = eb_lookup_objects(eb, exec, count, file);
	640	+ ret = eb_lookup_objects(eb, exec, args, file);
619	641	if (ret)
620	642	goto err;
621	643
@@ -919,7 +941,7 @@ i915_gem_do_execbuffer(struct drm_device dev, void data, · 297 unmodified lines
919	941	goto pre_mutex_err;
920	942	}
921	943
922		− eb = eb_create(args->buffer_count);
	944	+ eb = eb_create(args);
923	945	if (eb == NULL) {
924	946	mutex_unlock(&dev->struct_mutex);
925	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	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	953	if (ret)
932	954	goto err;
933	955

include/uapi/drm/i915_drm.h+7−1

@@ -309,6 +309,7 @@ typedef struct drm_i915_irq_wait { · 308 unmodified lines
309	309	#define I915_PARAM_HAS_SECURE_BATCHES 23
310	310	#define I915_PARAM_HAS_PINNED_BATCHES 24
311	311	#define I915_PARAM_HAS_EXEC_NO_RELOC 25
	312	+#define I915_PARAM_HAS_EXEC_HANDLE_LUT 26
312	313
313	314	typedef struct drm_i915_getparam {
314	315	int param;
@@ -699,7 +700,12 @@ struct drm_i915_gem_execbuffer2 { · 384 unmodified lines
699	700	*/
700	701	#define I915_EXEC_NO_RELOC (1<<11)
701	702
702		−#define __I915_EXEC_UNKNOWN_FLAGS -(I915_EXEC_NO_RELOC<<1)
	703	+/** Use the reloc.handle as an index into the exec object array rather
	704	+ * than as the per-file handle.
	705	+ */
	706	+#define I915_EXEC_HANDLE_LUT (1<<12)
	707	+
	708	+#define __I915_EXEC_UNKNOWN_FLAGS -(I915_EXEC_HANDLE_LUT<<1)
703	709
704	710	#define I915_EXEC_CONTEXT_ID_MASK (0xffffffff)
705	711	#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