1 /*
2 * Copyright (C) 2018 Jonathan Marek <jonathan@marek.ca>
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 * Jonathan Marek <jonathan@marek.ca>
25 * Rob Clark <robclark@freedesktop.org>
26 */
27
28 /* NOTE: perfcntrs are 48-bits but we only have 32-bit accumulate (?)
29 * so we work with 32-bits only. we accumulate start/stop separately,
30 * which differs from a5xx but works with only accumulate (no add/neg)
31 */
32
33 #include "freedreno_query_acc.h"
34 #include "freedreno_resource.h"
35
36 #include "fd2_context.h"
37 #include "fd2_query.h"
38
39 struct PACKED fd2_query_sample {
40 uint32_t start;
41 uint32_t stop;
42 };
43
44 /* offset of a single field of an array of fd2_query_sample: */
45 #define query_sample_idx(aq, idx, field) \
46 fd_resource((aq)->prsc)->bo, \
47 (idx * sizeof(struct fd2_query_sample)) + \
48 offsetof(struct fd2_query_sample, field), \
49 0, 0
50
51 /* offset of a single field of fd2_query_sample: */
52 #define query_sample(aq, field) query_sample_idx(aq, 0, field)
53
54 /*
55 * Performance Counter (batch) queries:
56 *
57 * Only one of these is active at a time, per design of the gallium
58 * batch_query API design. On perfcntr query tracks N query_types,
59 * each of which has a 'fd_batch_query_entry' that maps it back to
60 * the associated group and counter.
61 */
62
63 struct fd_batch_query_entry {
64 uint8_t gid; /* group-id */
65 uint8_t cid; /* countable-id within the group */
66 };
67
68 struct fd_batch_query_data {
69 struct fd_screen *screen;
70 unsigned num_query_entries;
71 struct fd_batch_query_entry query_entries[];
72 };
73
74 static void
perfcntr_resume(struct fd_acc_query * aq,struct fd_batch * batch)75 perfcntr_resume(struct fd_acc_query *aq, struct fd_batch *batch) assert_dt
76 {
77 struct fd_batch_query_data *data = aq->query_data;
78 struct fd_screen *screen = data->screen;
79 struct fd_ringbuffer *ring = batch->draw;
80
81 unsigned counters_per_group[screen->num_perfcntr_groups];
82 memset(counters_per_group, 0, sizeof(counters_per_group));
83
84 fd_wfi(batch, ring);
85
86 /* configure performance counters for the requested queries: */
87 for (unsigned i = 0; i < data->num_query_entries; i++) {
88 struct fd_batch_query_entry *entry = &data->query_entries[i];
89 const struct fd_perfcntr_group *g = &screen->perfcntr_groups[entry->gid];
90 unsigned counter_idx = counters_per_group[entry->gid]++;
91
92 assert(counter_idx < g->num_counters);
93
94 OUT_PKT0(ring, g->counters[counter_idx].select_reg, 1);
95 OUT_RING(ring, g->countables[entry->cid].selector);
96 }
97
98 memset(counters_per_group, 0, sizeof(counters_per_group));
99
100 /* and snapshot the start values */
101 for (unsigned i = 0; i < data->num_query_entries; i++) {
102 struct fd_batch_query_entry *entry = &data->query_entries[i];
103 const struct fd_perfcntr_group *g = &screen->perfcntr_groups[entry->gid];
104 unsigned counter_idx = counters_per_group[entry->gid]++;
105 const struct fd_perfcntr_counter *counter = &g->counters[counter_idx];
106
107 OUT_PKT3(ring, CP_REG_TO_MEM, 2);
108 OUT_RING(ring, counter->counter_reg_lo | CP_REG_TO_MEM_0_ACCUMULATE);
109 OUT_RELOC(ring, query_sample_idx(aq, i, start));
110 }
111 }
112
113 static void
perfcntr_pause(struct fd_acc_query * aq,struct fd_batch * batch)114 perfcntr_pause(struct fd_acc_query *aq, struct fd_batch *batch) assert_dt
115 {
116 struct fd_batch_query_data *data = aq->query_data;
117 struct fd_screen *screen = data->screen;
118 struct fd_ringbuffer *ring = batch->draw;
119
120 unsigned counters_per_group[screen->num_perfcntr_groups];
121 memset(counters_per_group, 0, sizeof(counters_per_group));
122
123 fd_wfi(batch, ring);
124
125 /* TODO do we need to bother to turn anything off? */
126
127 /* snapshot the end values: */
128 for (unsigned i = 0; i < data->num_query_entries; i++) {
129 struct fd_batch_query_entry *entry = &data->query_entries[i];
130 const struct fd_perfcntr_group *g = &screen->perfcntr_groups[entry->gid];
131 unsigned counter_idx = counters_per_group[entry->gid]++;
132 const struct fd_perfcntr_counter *counter = &g->counters[counter_idx];
133
134 OUT_PKT3(ring, CP_REG_TO_MEM, 2);
135 OUT_RING(ring, counter->counter_reg_lo | CP_REG_TO_MEM_0_ACCUMULATE);
136 OUT_RELOC(ring, query_sample_idx(aq, i, stop));
137 }
138 }
139
140 static void
perfcntr_accumulate_result(struct fd_acc_query * aq,void * buf,union pipe_query_result * result)141 perfcntr_accumulate_result(struct fd_acc_query *aq, void *buf,
142 union pipe_query_result *result)
143 {
144 struct fd_batch_query_data *data = aq->query_data;
145 struct fd2_query_sample *sp = buf;
146
147 for (unsigned i = 0; i < data->num_query_entries; i++)
148 result->batch[i].u64 = sp[i].stop - sp[i].start;
149 }
150
151 static const struct fd_acc_sample_provider perfcntr = {
152 .query_type = FD_QUERY_FIRST_PERFCNTR,
153 .always = true,
154 .resume = perfcntr_resume,
155 .pause = perfcntr_pause,
156 .result = perfcntr_accumulate_result,
157 };
158
159 static struct pipe_query *
fd2_create_batch_query(struct pipe_context * pctx,unsigned num_queries,unsigned * query_types)160 fd2_create_batch_query(struct pipe_context *pctx, unsigned num_queries,
161 unsigned *query_types)
162 {
163 struct fd_context *ctx = fd_context(pctx);
164 struct fd_screen *screen = ctx->screen;
165 struct fd_query *q;
166 struct fd_acc_query *aq;
167 struct fd_batch_query_data *data;
168
169 data = CALLOC_VARIANT_LENGTH_STRUCT(
170 fd_batch_query_data, num_queries * sizeof(data->query_entries[0]));
171
172 data->screen = screen;
173 data->num_query_entries = num_queries;
174
175 /* validate the requested query_types and ensure we don't try
176 * to request more query_types of a given group than we have
177 * counters:
178 */
179 unsigned counters_per_group[screen->num_perfcntr_groups];
180 memset(counters_per_group, 0, sizeof(counters_per_group));
181
182 for (unsigned i = 0; i < num_queries; i++) {
183 unsigned idx = query_types[i] - FD_QUERY_FIRST_PERFCNTR;
184
185 /* verify valid query_type, ie. is it actually a perfcntr? */
186 if ((query_types[i] < FD_QUERY_FIRST_PERFCNTR) ||
187 (idx >= screen->num_perfcntr_queries)) {
188 mesa_loge("invalid batch query query_type: %u", query_types[i]);
189 goto error;
190 }
191
192 struct fd_batch_query_entry *entry = &data->query_entries[i];
193 struct pipe_driver_query_info *pq = &screen->perfcntr_queries[idx];
194
195 entry->gid = pq->group_id;
196
197 /* the perfcntr_queries[] table flattens all the countables
198 * for each group in series, ie:
199 *
200 * (G0,C0), .., (G0,Cn), (G1,C0), .., (G1,Cm), ...
201 *
202 * So to find the countable index just step back through the
203 * table to find the first entry with the same group-id.
204 */
205 while (pq > screen->perfcntr_queries) {
206 pq--;
207 if (pq->group_id == entry->gid)
208 entry->cid++;
209 }
210
211 if (counters_per_group[entry->gid] >=
212 screen->perfcntr_groups[entry->gid].num_counters) {
213 mesa_loge("too many counters for group %u", entry->gid);
214 goto error;
215 }
216
217 counters_per_group[entry->gid]++;
218 }
219
220 q = fd_acc_create_query2(ctx, 0, 0, &perfcntr);
221 aq = fd_acc_query(q);
222
223 /* sample buffer size is based on # of queries: */
224 aq->size = num_queries * sizeof(struct fd2_query_sample);
225 aq->query_data = data;
226
227 return (struct pipe_query *)q;
228
229 error:
230 free(data);
231 return NULL;
232 }
233
234 void
fd2_query_context_init(struct pipe_context * pctx)235 fd2_query_context_init(struct pipe_context *pctx) disable_thread_safety_analysis
236 {
237 struct fd_context *ctx = fd_context(pctx);
238
239 ctx->create_query = fd_acc_create_query;
240 ctx->query_update_batch = fd_acc_query_update_batch;
241
242 pctx->create_batch_query = fd2_create_batch_query;
243 }
244