1 /*
2 * Copyright © 2014 Rob Clark <[email protected]>
3 * SPDX-License-Identifier: MIT
4 *
5 * Authors:
6 * Rob Clark <[email protected]>
7 */
8
9 #ifndef FREEDRENO_QUERY_HW_H_
10 #define FREEDRENO_QUERY_HW_H_
11
12 #include "util/list.h"
13
14 #include "freedreno_context.h"
15 #include "freedreno_query.h"
16
17 /*
18 * HW Queries:
19 *
20 * See: https://gitlab.freedesktop.org/freedreno/freedreno/-/wikis/Queries#hardware-queries
21 *
22 * Hardware queries will be specific to gpu generation, but they need
23 * some common infrastructure for triggering start/stop samples at
24 * various points (for example, to exclude mem2gmem/gmem2mem or clear)
25 * as well as per tile tracking.
26 *
27 * NOTE: in at least some cases hw writes sample values to memory addr
28 * specified in some register. So we don't really have the option to
29 * just sample the same counter multiple times for multiple different
30 * queries with the same query_type. So we cache per sample provider
31 * the most recent sample since the last draw. This way multiple
32 * sample periods for multiple queries can reference the same sample.
33 *
34 * fd_hw_sample_provider:
35 * - one per query type, registered/implemented by gpu generation
36 * specific code
37 * - can construct fd_hw_samples on demand
38 * - most recent sample (since last draw) cached so multiple
39 * different queries can ref the same sample
40 *
41 * fd_hw_sample:
42 * - abstracts one snapshot of counter value(s) across N tiles
43 * - backing object not allocated until submit time when number
44 * of samples and number of tiles is known
45 *
46 * fd_hw_sample_period:
47 * - consists of start and stop sample
48 * - a query accumulates a list of sample periods
49 * - the query result is the sum of the sample periods
50 */
51
52 struct fd_hw_sample_provider {
53 unsigned query_type;
54
55 /* Set if the provider should still count while !ctx->active_queries */
56 bool always;
57
58 /* Optional hook for enabling a counter. Guaranteed to happen
59 * at least once before the first ->get_sample() in a batch.
60 */
61 void (*enable)(struct fd_context *ctx, struct fd_ringbuffer *ring) dt;
62
63 /* when a new sample is required, emit appropriate cmdstream
64 * and return a sample object:
65 */
66 struct fd_hw_sample *(*get_sample)(struct fd_batch *batch,
67 struct fd_ringbuffer *ring)dt;
68
69 /* accumulate the results from specified sample period: */
70 void (*accumulate_result)(struct fd_context *ctx, const void *start,
71 const void *end, union pipe_query_result *result);
72 };
73
74 struct fd_hw_sample {
75 struct pipe_reference reference; /* keep this first */
76
77 /* offset and size of the sample are know at the time the
78 * sample is constructed.
79 */
80 uint32_t size;
81 uint32_t offset;
82
83 /* backing object, offset/stride/etc are determined not when
84 * the sample is constructed, but when the batch is submitted.
85 * This way we can defer allocation until total # of requested
86 * samples, and total # of tiles, is known.
87 */
88 struct pipe_resource *prsc;
89 uint32_t num_tiles;
90 uint32_t tile_stride;
91 };
92
93 struct fd_hw_sample_period;
94
95 struct fd_hw_query {
96 struct fd_query base;
97
98 const struct fd_hw_sample_provider *provider;
99
100 /* list of fd_hw_sample_periods: */
101 struct list_head periods;
102
103 /* if active and not paused, the current sample period (not
104 * yet added to current_periods):
105 */
106 struct fd_hw_sample_period *period;
107
108 struct list_head list; /* list-node in batch->active_queries */
109 };
110
111 static inline struct fd_hw_query *
fd_hw_query(struct fd_query * q)112 fd_hw_query(struct fd_query *q)
113 {
114 return (struct fd_hw_query *)q;
115 }
116
117 struct fd_query *fd_hw_create_query(struct fd_context *ctx, unsigned query_type,
118 unsigned index);
119 /* helper for sample providers: */
120 struct fd_hw_sample *fd_hw_sample_init(struct fd_batch *batch, uint32_t size);
121 /* don't call directly, use fd_hw_sample_reference() */
122 void __fd_hw_sample_destroy(struct fd_context *ctx, struct fd_hw_sample *samp);
123 void fd_hw_query_prepare(struct fd_batch *batch, uint32_t num_tiles) assert_dt;
124 void fd_hw_query_prepare_tile(struct fd_batch *batch, uint32_t n,
125 struct fd_ringbuffer *ring) assert_dt;
126 void fd_hw_query_update_batch(struct fd_batch *batch, bool end_batch) assert_dt;
127 void fd_hw_query_enable(struct fd_batch *batch,
128 struct fd_ringbuffer *ring) assert_dt;
129 void
130 fd_hw_query_register_provider(struct pipe_context *pctx,
131 const struct fd_hw_sample_provider *provider);
132 void fd_hw_query_init(struct pipe_context *pctx);
133 void fd_hw_query_fini(struct pipe_context *pctx);
134
135 static inline void
fd_hw_sample_reference(struct fd_context * ctx,struct fd_hw_sample ** ptr,struct fd_hw_sample * samp)136 fd_hw_sample_reference(struct fd_context *ctx, struct fd_hw_sample **ptr,
137 struct fd_hw_sample *samp)
138 {
139 struct fd_hw_sample *old_samp = *ptr;
140
141 if (pipe_reference(&(*ptr)->reference, &samp->reference))
142 __fd_hw_sample_destroy(ctx, old_samp);
143 *ptr = samp;
144 }
145
146 #endif /* FREEDRENO_QUERY_HW_H_ */
147