2 * Copyright © 2013 Intel Corporation
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:
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
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
21 * DEALINGS IN THE SOFTWARE.
25 * \file brw_performance_query.c
27 * Implementation of the GL_INTEL_performance_query extension.
29 * Currently there are two possible counter sources exposed here:
31 * On Gen6+ hardware we have numerous 64bit Pipeline Statistics Registers
32 * that we can snapshot at the beginning and end of a query.
34 * On Gen7.5+ we have Observability Architecture counters which are
35 * covered in separate document from the rest of the PRMs. It is available at:
36 * https://01.org/linuxgraphics/documentation/driver-documentation-prms
37 * => 2013 Intel Core Processor Family => Observability Performance Counters
38 * (This one volume covers Sandybridge, Ivybridge, Baytrail, and Haswell,
39 * though notably we currently only support OA counters for Haswell+)
45 /* put before sys/types.h to silence glibc warnings */
47 #include <sys/mkdev.h>
49 #ifdef MAJOR_IN_SYSMACROS
50 #include <sys/sysmacros.h>
52 #include <sys/types.h>
56 #include <sys/ioctl.h>
61 #include "main/hash.h"
62 #include "main/macros.h"
63 #include "main/mtypes.h"
64 #include "main/performance_query.h"
66 #include "util/bitset.h"
67 #include "util/ralloc.h"
68 #include "util/hash_table.h"
69 #include "util/list.h"
71 #include "brw_context.h"
72 #include "brw_defines.h"
73 #include "brw_performance_query.h"
74 #include "brw_oa_metrics.h"
75 #include "intel_batchbuffer.h"
77 #define FILE_DEBUG_FLAG DEBUG_PERFMON
79 #define OAREPORT_REASON_MASK 0x3f
80 #define OAREPORT_REASON_SHIFT 19
81 #define OAREPORT_REASON_TIMER (1<<0)
82 #define OAREPORT_REASON_TRIGGER1 (1<<1)
83 #define OAREPORT_REASON_TRIGGER2 (1<<2)
84 #define OAREPORT_REASON_CTX_SWITCH (1<<3)
85 #define OAREPORT_REASON_GO_TRANSITION (1<<4)
87 #define I915_PERF_OA_SAMPLE_SIZE (8 + /* drm_i915_perf_record_header */ \
88 256) /* OA counter report */
91 * Periodic OA samples are read() into these buffer structures via the
92 * i915 perf kernel interface and appended to the
93 * brw->perfquery.sample_buffers linked list. When we process the
94 * results of an OA metrics query we need to consider all the periodic
95 * samples between the Begin and End MI_REPORT_PERF_COUNT command
98 * 'Periodic' is a simplification as there are other automatic reports
99 * written by the hardware also buffered here.
101 * Considering three queries, A, B and C:
104 * ________________A_________________
106 * | ________B_________ _____C___________
109 * And an illustration of sample buffers read over this time frame:
110 * [HEAD ][ ][ ][ ][ ][ ][ ][ ][TAIL ]
112 * These nodes may hold samples for query A:
113 * [ ][ ][ A ][ A ][ A ][ A ][ A ][ ][ ]
115 * These nodes may hold samples for query B:
116 * [ ][ ][ B ][ B ][ B ][ ][ ][ ][ ]
118 * These nodes may hold samples for query C:
119 * [ ][ ][ ][ ][ ][ C ][ C ][ C ][ ]
121 * The illustration assumes we have an even distribution of periodic
122 * samples so all nodes have the same size plotted against time:
124 * Note, to simplify code, the list is never empty.
126 * With overlapping queries we can see that periodic OA reports may
127 * relate to multiple queries and care needs to be take to keep
128 * track of sample buffers until there are no queries that might
129 * depend on their contents.
131 * We use a node ref counting system where a reference ensures that a
132 * node and all following nodes can't be freed/recycled until the
133 * reference drops to zero.
135 * E.g. with a ref of one here:
136 * [ 0 ][ 0 ][ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
138 * These nodes could be freed or recycled ("reaped"):
141 * These must be preserved until the leading ref drops to zero:
142 * [ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
144 * When a query starts we take a reference on the current tail of
145 * the list, knowing that no already-buffered samples can possibly
146 * relate to the newly-started query. A pointer to this node is
147 * also saved in the query object's ->oa.samples_head.
149 * E.g. starting query A while there are two nodes in .sample_buffers:
150 * ________________A________
154 * ^_______ Add a reference and store pointer to node in
157 * Moving forward to when the B query starts with no new buffer nodes:
158 * (for reference, i915 perf reads() are only done when queries finish)
159 * ________________A_______
164 * ^_______ Add a reference and store pointer to
165 * node in B->oa.samples_head
167 * Once a query is finished, after an OA query has become 'Ready',
168 * once the End OA report has landed and after we we have processed
169 * all the intermediate periodic samples then we drop the
170 * ->oa.samples_head reference we took at the start.
172 * So when the B query has finished we have:
173 * ________________A________
174 * | ______B___________
176 * [ 0 ][ 1 ][ 0 ][ 0 ][ 0 ]
177 * ^_______ Drop B->oa.samples_head reference
179 * We still can't free these due to the A->oa.samples_head ref:
180 * [ 1 ][ 0 ][ 0 ][ 0 ]
182 * When the A query finishes: (note there's a new ref for C's samples_head)
183 * ________________A_________________
187 * [ 0 ][ 0 ][ 0 ][ 0 ][ 1 ][ 0 ][ 0 ]
188 * ^_______ Drop A->oa.samples_head reference
190 * And we can now reap these nodes up to the C->oa.samples_head:
191 * [ X ][ X ][ X ][ X ]
192 * keeping -> [ 1 ][ 0 ][ 0 ]
194 * We reap old sample buffers each time we finish processing an OA
195 * query by iterating the sample_buffers list from the head until we
196 * find a referenced node and stop.
198 * Reaped buffers move to a perfquery.free_sample_buffers list and
199 * when we come to read() we first look to recycle a buffer from the
200 * free_sample_buffers list before allocating a new buffer.
202 struct brw_oa_sample_buf
{
203 struct exec_node link
;
206 uint8_t buf
[I915_PERF_OA_SAMPLE_SIZE
* 10];
207 uint32_t last_timestamp
;
210 /** Downcasting convenience macro. */
211 static inline struct brw_perf_query_object
*
212 brw_perf_query(struct gl_perf_query_object
*o
)
214 return (struct brw_perf_query_object
*) o
;
217 #define MI_RPC_BO_SIZE 4096
218 #define MI_RPC_BO_END_OFFSET_BYTES (MI_RPC_BO_SIZE / 2)
219 #define MI_FREQ_START_OFFSET_BYTES (3072)
220 #define MI_FREQ_END_OFFSET_BYTES (3076)
222 /******************************************************************************/
225 read_file_uint64(const char *file
, uint64_t *val
)
233 while ((n
= read(fd
, buf
, sizeof (buf
) - 1)) < 0 &&
240 *val
= strtoull(buf
, NULL
, 0);
246 read_sysfs_drm_device_file_uint64(struct brw_context
*brw
,
253 len
= snprintf(buf
, sizeof(buf
), "%s/%s",
254 brw
->perfquery
.sysfs_dev_dir
, file
);
255 if (len
< 0 || len
>= sizeof(buf
)) {
256 DBG("Failed to concatenate sys filename to read u64 from\n");
260 return read_file_uint64(buf
, value
);
263 /******************************************************************************/
266 brw_is_perf_query_ready(struct gl_context
*ctx
,
267 struct gl_perf_query_object
*o
);
270 brw_perf_query_get_metric_id(struct brw_context
*brw
,
271 const struct brw_perf_query_info
*query
)
273 /* These queries are know not to ever change, their config ID has been
274 * loaded upon the first query creation. No need to look them up again.
276 if (query
->kind
== OA_COUNTERS
)
277 return query
->oa_metrics_set_id
;
279 assert(query
->kind
== OA_COUNTERS_RAW
);
281 /* Raw queries can be reprogrammed up by an external application/library.
282 * When a raw query is used for the first time it's id is set to a value !=
283 * 0. When it stops being used the id returns to 0. No need to reload the
284 * ID when it's already loaded.
286 if (query
->oa_metrics_set_id
!= 0) {
287 DBG("Raw query '%s' guid=%s using cached ID: %"PRIu64
"\n",
288 query
->name
, query
->guid
, query
->oa_metrics_set_id
);
289 return query
->oa_metrics_set_id
;
292 char metric_id_file
[280];
293 snprintf(metric_id_file
, sizeof(metric_id_file
),
294 "%s/metrics/%s/id", brw
->perfquery
.sysfs_dev_dir
, query
->guid
);
296 struct brw_perf_query_info
*raw_query
= (struct brw_perf_query_info
*)query
;
297 if (!read_file_uint64(metric_id_file
, &raw_query
->oa_metrics_set_id
)) {
298 DBG("Unable to read query guid=%s ID, falling back to test config\n", query
->guid
);
299 raw_query
->oa_metrics_set_id
= 1ULL;
301 DBG("Raw query '%s'guid=%s loaded ID: %"PRIu64
"\n",
302 query
->name
, query
->guid
, query
->oa_metrics_set_id
);
304 return query
->oa_metrics_set_id
;
308 dump_perf_query_callback(GLuint id
, void *query_void
, void *brw_void
)
310 struct gl_context
*ctx
= brw_void
;
311 struct gl_perf_query_object
*o
= query_void
;
312 struct brw_perf_query_object
*obj
= query_void
;
314 switch (obj
->query
->kind
) {
316 case OA_COUNTERS_RAW
:
317 DBG("%4d: %-6s %-8s BO: %-4s OA data: %-10s %-15s\n",
319 o
->Used
? "Dirty," : "New,",
320 o
->Active
? "Active," : (o
->Ready
? "Ready," : "Pending,"),
321 obj
->oa
.bo
? "yes," : "no,",
322 brw_is_perf_query_ready(ctx
, o
) ? "ready," : "not ready,",
323 obj
->oa
.results_accumulated
? "accumulated" : "not accumulated");
326 DBG("%4d: %-6s %-8s BO: %-4s\n",
328 o
->Used
? "Dirty," : "New,",
329 o
->Active
? "Active," : (o
->Ready
? "Ready," : "Pending,"),
330 obj
->pipeline_stats
.bo
? "yes" : "no");
333 unreachable("Unknown query type");
339 dump_perf_queries(struct brw_context
*brw
)
341 struct gl_context
*ctx
= &brw
->ctx
;
342 DBG("Queries: (Open queries = %d, OA users = %d)\n",
343 brw
->perfquery
.n_active_oa_queries
, brw
->perfquery
.n_oa_users
);
344 _mesa_HashWalk(ctx
->PerfQuery
.Objects
, dump_perf_query_callback
, brw
);
347 /******************************************************************************/
349 static struct brw_oa_sample_buf
*
350 get_free_sample_buf(struct brw_context
*brw
)
352 struct exec_node
*node
= exec_list_pop_head(&brw
->perfquery
.free_sample_buffers
);
353 struct brw_oa_sample_buf
*buf
;
356 buf
= exec_node_data(struct brw_oa_sample_buf
, node
, link
);
358 buf
= ralloc_size(brw
, sizeof(*buf
));
360 exec_node_init(&buf
->link
);
369 reap_old_sample_buffers(struct brw_context
*brw
)
371 struct exec_node
*tail_node
=
372 exec_list_get_tail(&brw
->perfquery
.sample_buffers
);
373 struct brw_oa_sample_buf
*tail_buf
=
374 exec_node_data(struct brw_oa_sample_buf
, tail_node
, link
);
376 /* Remove all old, unreferenced sample buffers walking forward from
377 * the head of the list, except always leave at least one node in
378 * the list so we always have a node to reference when we Begin
381 foreach_list_typed_safe(struct brw_oa_sample_buf
, buf
, link
,
382 &brw
->perfquery
.sample_buffers
)
384 if (buf
->refcount
== 0 && buf
!= tail_buf
) {
385 exec_node_remove(&buf
->link
);
386 exec_list_push_head(&brw
->perfquery
.free_sample_buffers
, &buf
->link
);
393 free_sample_bufs(struct brw_context
*brw
)
395 foreach_list_typed_safe(struct brw_oa_sample_buf
, buf
, link
,
396 &brw
->perfquery
.free_sample_buffers
)
399 exec_list_make_empty(&brw
->perfquery
.free_sample_buffers
);
402 /******************************************************************************/
405 * Driver hook for glGetPerfQueryInfoINTEL().
408 brw_get_perf_query_info(struct gl_context
*ctx
,
409 unsigned query_index
,
415 struct brw_context
*brw
= brw_context(ctx
);
416 const struct brw_perf_query_info
*query
=
417 &brw
->perfquery
.queries
[query_index
];
420 *data_size
= query
->data_size
;
421 *n_counters
= query
->n_counters
;
423 switch (query
->kind
) {
425 case OA_COUNTERS_RAW
:
426 *n_active
= brw
->perfquery
.n_active_oa_queries
;
430 *n_active
= brw
->perfquery
.n_active_pipeline_stats_queries
;
434 unreachable("Unknown query type");
440 * Driver hook for glGetPerfCounterInfoINTEL().
443 brw_get_perf_counter_info(struct gl_context
*ctx
,
444 unsigned query_index
,
445 unsigned counter_index
,
451 GLuint
*data_type_enum
,
454 struct brw_context
*brw
= brw_context(ctx
);
455 const struct brw_perf_query_info
*query
=
456 &brw
->perfquery
.queries
[query_index
];
457 const struct brw_perf_query_counter
*counter
=
458 &query
->counters
[counter_index
];
460 *name
= counter
->name
;
461 *desc
= counter
->desc
;
462 *offset
= counter
->offset
;
463 *data_size
= counter
->size
;
464 *type_enum
= counter
->type
;
465 *data_type_enum
= counter
->data_type
;
466 *raw_max
= counter
->raw_max
;
469 /******************************************************************************/
472 * Emit MI_STORE_REGISTER_MEM commands to capture all of the
473 * pipeline statistics for the performance query object.
476 snapshot_statistics_registers(struct brw_context
*brw
,
477 struct brw_perf_query_object
*obj
,
478 uint32_t offset_in_bytes
)
480 const struct brw_perf_query_info
*query
= obj
->query
;
481 const int n_counters
= query
->n_counters
;
483 for (int i
= 0; i
< n_counters
; i
++) {
484 const struct brw_perf_query_counter
*counter
= &query
->counters
[i
];
486 assert(counter
->data_type
== GL_PERFQUERY_COUNTER_DATA_UINT64_INTEL
);
488 brw_store_register_mem64(brw
, obj
->pipeline_stats
.bo
,
489 counter
->pipeline_stat
.reg
,
490 offset_in_bytes
+ i
* sizeof(uint64_t));
495 * Add a query to the global list of "unaccumulated queries."
497 * Queries are tracked here until all the associated OA reports have
498 * been accumulated via accumulate_oa_reports() after the end
499 * MI_REPORT_PERF_COUNT has landed in query->oa.bo.
502 add_to_unaccumulated_query_list(struct brw_context
*brw
,
503 struct brw_perf_query_object
*obj
)
505 if (brw
->perfquery
.unaccumulated_elements
>=
506 brw
->perfquery
.unaccumulated_array_size
)
508 brw
->perfquery
.unaccumulated_array_size
*= 1.5;
509 brw
->perfquery
.unaccumulated
=
510 reralloc(brw
, brw
->perfquery
.unaccumulated
,
511 struct brw_perf_query_object
*,
512 brw
->perfquery
.unaccumulated_array_size
);
515 brw
->perfquery
.unaccumulated
[brw
->perfquery
.unaccumulated_elements
++] = obj
;
519 * Remove a query from the global list of unaccumulated queries once
520 * after successfully accumulating the OA reports associated with the
521 * query in accumulate_oa_reports() or when discarding unwanted query
525 drop_from_unaccumulated_query_list(struct brw_context
*brw
,
526 struct brw_perf_query_object
*obj
)
528 for (int i
= 0; i
< brw
->perfquery
.unaccumulated_elements
; i
++) {
529 if (brw
->perfquery
.unaccumulated
[i
] == obj
) {
530 int last_elt
= --brw
->perfquery
.unaccumulated_elements
;
533 brw
->perfquery
.unaccumulated
[i
] = NULL
;
535 brw
->perfquery
.unaccumulated
[i
] =
536 brw
->perfquery
.unaccumulated
[last_elt
];
543 /* Drop our samples_head reference so that associated periodic
544 * sample data buffers can potentially be reaped if they aren't
545 * referenced by any other queries...
548 struct brw_oa_sample_buf
*buf
=
549 exec_node_data(struct brw_oa_sample_buf
, obj
->oa
.samples_head
, link
);
551 assert(buf
->refcount
> 0);
554 obj
->oa
.samples_head
= NULL
;
556 reap_old_sample_buffers(brw
);
560 * Given pointers to starting and ending OA snapshots, add the deltas for each
561 * counter to the results.
564 add_deltas(struct brw_context
*brw
,
565 struct brw_perf_query_object
*obj
,
566 const uint32_t *start
,
569 const struct brw_perf_query_info
*query
= obj
->query
;
570 uint64_t *accumulator
= obj
->oa
.accumulator
;
574 obj
->oa
.reports_accumulated
++;
576 switch (query
->oa_format
) {
577 case I915_OA_FORMAT_A32u40_A4u32_B8_C8
:
578 brw_perf_query_accumulate_uint32(start
+ 1, end
+ 1, accumulator
+ idx
++); /* timestamp */
579 brw_perf_query_accumulate_uint32(start
+ 3, end
+ 3, accumulator
+ idx
++); /* clock */
581 /* 32x 40bit A counters... */
582 for (i
= 0; i
< 32; i
++)
583 brw_perf_query_accumulate_uint40(i
, start
, end
, accumulator
+ idx
++);
585 /* 4x 32bit A counters... */
586 for (i
= 0; i
< 4; i
++)
587 brw_perf_query_accumulate_uint32(start
+ 36 + i
, end
+ 36 + i
,
588 accumulator
+ idx
++);
590 /* 8x 32bit B counters + 8x 32bit C counters... */
591 for (i
= 0; i
< 16; i
++)
592 brw_perf_query_accumulate_uint32(start
+ 48 + i
, end
+ 48 + i
,
593 accumulator
+ idx
++);
596 case I915_OA_FORMAT_A45_B8_C8
:
597 brw_perf_query_accumulate_uint32(start
+ 1, end
+ 1, accumulator
); /* timestamp */
599 for (i
= 0; i
< 61; i
++)
600 brw_perf_query_accumulate_uint32(start
+ 3 + i
, end
+ 3 + i
, accumulator
+ 1 + i
);
604 unreachable("Can't accumulate OA counters in unknown format");
609 inc_n_oa_users(struct brw_context
*brw
)
611 if (brw
->perfquery
.n_oa_users
== 0 &&
612 drmIoctl(brw
->perfquery
.oa_stream_fd
,
613 I915_PERF_IOCTL_ENABLE
, 0) < 0)
617 ++brw
->perfquery
.n_oa_users
;
623 dec_n_oa_users(struct brw_context
*brw
)
625 /* Disabling the i915 perf stream will effectively disable the OA
626 * counters. Note it's important to be sure there are no outstanding
627 * MI_RPC commands at this point since they could stall the CS
628 * indefinitely once OACONTROL is disabled.
630 --brw
->perfquery
.n_oa_users
;
631 if (brw
->perfquery
.n_oa_users
== 0 &&
632 drmIoctl(brw
->perfquery
.oa_stream_fd
, I915_PERF_IOCTL_DISABLE
, 0) < 0)
634 DBG("WARNING: Error disabling i915 perf stream: %m\n");
638 /* In general if we see anything spurious while accumulating results,
639 * we don't try and continue accumulating the current query, hoping
640 * for the best, we scrap anything outstanding, and then hope for the
641 * best with new queries.
644 discard_all_queries(struct brw_context
*brw
)
646 while (brw
->perfquery
.unaccumulated_elements
) {
647 struct brw_perf_query_object
*obj
= brw
->perfquery
.unaccumulated
[0];
649 obj
->oa
.results_accumulated
= true;
650 drop_from_unaccumulated_query_list(brw
, brw
->perfquery
.unaccumulated
[0]);
657 OA_READ_STATUS_ERROR
,
658 OA_READ_STATUS_UNFINISHED
,
659 OA_READ_STATUS_FINISHED
,
662 static enum OaReadStatus
663 read_oa_samples_until(struct brw_context
*brw
,
664 uint32_t start_timestamp
,
665 uint32_t end_timestamp
)
667 struct exec_node
*tail_node
=
668 exec_list_get_tail(&brw
->perfquery
.sample_buffers
);
669 struct brw_oa_sample_buf
*tail_buf
=
670 exec_node_data(struct brw_oa_sample_buf
, tail_node
, link
);
671 uint32_t last_timestamp
= tail_buf
->last_timestamp
;
674 struct brw_oa_sample_buf
*buf
= get_free_sample_buf(brw
);
678 while ((len
= read(brw
->perfquery
.oa_stream_fd
, buf
->buf
,
679 sizeof(buf
->buf
))) < 0 && errno
== EINTR
)
683 exec_list_push_tail(&brw
->perfquery
.free_sample_buffers
, &buf
->link
);
687 return ((last_timestamp
- start_timestamp
) >=
688 (end_timestamp
- start_timestamp
)) ?
689 OA_READ_STATUS_FINISHED
:
690 OA_READ_STATUS_UNFINISHED
;
692 DBG("Error reading i915 perf samples: %m\n");
695 DBG("Spurious EOF reading i915 perf samples\n");
697 return OA_READ_STATUS_ERROR
;
701 exec_list_push_tail(&brw
->perfquery
.sample_buffers
, &buf
->link
);
703 /* Go through the reports and update the last timestamp. */
705 while (offset
< buf
->len
) {
706 const struct drm_i915_perf_record_header
*header
=
707 (const struct drm_i915_perf_record_header
*) &buf
->buf
[offset
];
708 uint32_t *report
= (uint32_t *) (header
+ 1);
710 if (header
->type
== DRM_I915_PERF_RECORD_SAMPLE
)
711 last_timestamp
= report
[1];
713 offset
+= header
->size
;
716 buf
->last_timestamp
= last_timestamp
;
719 unreachable("not reached");
720 return OA_READ_STATUS_ERROR
;
724 * Try to read all the reports until either the delimiting timestamp
725 * or an error arises.
728 read_oa_samples_for_query(struct brw_context
*brw
,
729 struct brw_perf_query_object
*obj
)
735 /* We need the MI_REPORT_PERF_COUNT to land before we can start
737 assert(!brw_batch_references(&brw
->batch
, obj
->oa
.bo
) &&
738 !brw_bo_busy(obj
->oa
.bo
));
740 /* Map the BO once here and let accumulate_oa_reports() unmap
742 if (obj
->oa
.map
== NULL
)
743 obj
->oa
.map
= brw_bo_map(brw
, obj
->oa
.bo
, MAP_READ
);
745 start
= last
= obj
->oa
.map
;
746 end
= obj
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
748 if (start
[0] != obj
->oa
.begin_report_id
) {
749 DBG("Spurious start report id=%"PRIu32
"\n", start
[0]);
752 if (end
[0] != (obj
->oa
.begin_report_id
+ 1)) {
753 DBG("Spurious end report id=%"PRIu32
"\n", end
[0]);
757 /* Read the reports until the end timestamp. */
758 switch (read_oa_samples_until(brw
, start
[1], end
[1])) {
759 case OA_READ_STATUS_ERROR
:
760 /* Fallthrough and let accumulate_oa_reports() deal with the
762 case OA_READ_STATUS_FINISHED
:
764 case OA_READ_STATUS_UNFINISHED
:
768 unreachable("invalid read status");
773 * Accumulate raw OA counter values based on deltas between pairs of
776 * Accumulation starts from the first report captured via
777 * MI_REPORT_PERF_COUNT (MI_RPC) by brw_begin_perf_query() until the
778 * last MI_RPC report requested by brw_end_perf_query(). Between these
779 * two reports there may also some number of periodically sampled OA
780 * reports collected via the i915 perf interface - depending on the
781 * duration of the query.
783 * These periodic snapshots help to ensure we handle counter overflow
784 * correctly by being frequent enough to ensure we don't miss multiple
785 * overflows of a counter between snapshots. For Gen8+ the i915 perf
786 * snapshots provide the extra context-switch reports that let us
787 * subtract out the progress of counters associated with other
788 * contexts running on the system.
791 accumulate_oa_reports(struct brw_context
*brw
,
792 struct brw_perf_query_object
*obj
)
794 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
795 struct gl_perf_query_object
*o
= &obj
->base
;
799 struct exec_node
*first_samples_node
;
801 int out_duration
= 0;
804 assert(obj
->oa
.map
!= NULL
);
806 start
= last
= obj
->oa
.map
;
807 end
= obj
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
809 if (start
[0] != obj
->oa
.begin_report_id
) {
810 DBG("Spurious start report id=%"PRIu32
"\n", start
[0]);
813 if (end
[0] != (obj
->oa
.begin_report_id
+ 1)) {
814 DBG("Spurious end report id=%"PRIu32
"\n", end
[0]);
818 obj
->oa
.hw_id
= start
[2];
820 /* See if we have any periodic reports to accumulate too... */
822 /* N.B. The oa.samples_head was set when the query began and
823 * pointed to the tail of the brw->perfquery.sample_buffers list at
824 * the time the query started. Since the buffer existed before the
825 * first MI_REPORT_PERF_COUNT command was emitted we therefore know
826 * that no data in this particular node's buffer can possibly be
827 * associated with the query - so skip ahead one...
829 first_samples_node
= obj
->oa
.samples_head
->next
;
831 foreach_list_typed_from(struct brw_oa_sample_buf
, buf
, link
,
832 &brw
->perfquery
.sample_buffers
,
837 while (offset
< buf
->len
) {
838 const struct drm_i915_perf_record_header
*header
=
839 (const struct drm_i915_perf_record_header
*)(buf
->buf
+ offset
);
841 assert(header
->size
!= 0);
842 assert(header
->size
<= buf
->len
);
844 offset
+= header
->size
;
846 switch (header
->type
) {
847 case DRM_I915_PERF_RECORD_SAMPLE
: {
848 uint32_t *report
= (uint32_t *)(header
+ 1);
851 /* Ignore reports that come before the start marker.
852 * (Note: takes care to allow overflow of 32bit timestamps)
854 if (brw_timebase_scale(brw
, report
[1] - start
[1]) > 5000000000)
857 /* Ignore reports that come after the end marker.
858 * (Note: takes care to allow overflow of 32bit timestamps)
860 if (brw_timebase_scale(brw
, report
[1] - end
[1]) <= 5000000000)
863 /* For Gen8+ since the counters continue while other
864 * contexts are running we need to discount any unrelated
865 * deltas. The hardware automatically generates a report
866 * on context switch which gives us a new reference point
867 * to continuing adding deltas from.
869 * For Haswell we can rely on the HW to stop the progress
870 * of OA counters while any other context is acctive.
872 if (devinfo
->gen
>= 8) {
873 if (in_ctx
&& report
[2] != obj
->oa
.hw_id
) {
874 DBG("i915 perf: Switch AWAY (observed by ID change)\n");
877 } else if (in_ctx
== false && report
[2] == obj
->oa
.hw_id
) {
878 DBG("i915 perf: Switch TO\n");
881 /* From experimentation in IGT, we found that the OA unit
882 * might label some report as "idle" (using an invalid
883 * context ID), right after a report for a given context.
884 * Deltas generated by those reports actually belong to the
885 * previous context, even though they're not labelled as
888 * We didn't *really* Switch AWAY in the case that we e.g.
889 * saw a single periodic report while idle...
891 if (out_duration
>= 1)
894 assert(report
[2] == obj
->oa
.hw_id
);
895 DBG("i915 perf: Continuation IN\n");
897 assert(report
[2] != obj
->oa
.hw_id
);
898 DBG("i915 perf: Continuation OUT\n");
905 add_deltas(brw
, obj
, last
, report
);
912 case DRM_I915_PERF_RECORD_OA_BUFFER_LOST
:
913 DBG("i915 perf: OA error: all reports lost\n");
915 case DRM_I915_PERF_RECORD_OA_REPORT_LOST
:
916 DBG("i915 perf: OA report lost\n");
924 add_deltas(brw
, obj
, last
, end
);
926 DBG("Marking %d accumulated - results gathered\n", o
->Id
);
928 obj
->oa
.results_accumulated
= true;
929 drop_from_unaccumulated_query_list(brw
, obj
);
936 discard_all_queries(brw
);
939 /******************************************************************************/
942 open_i915_perf_oa_stream(struct brw_context
*brw
,
949 uint64_t properties
[] = {
950 /* Single context sampling */
951 DRM_I915_PERF_PROP_CTX_HANDLE
, ctx_id
,
953 /* Include OA reports in samples */
954 DRM_I915_PERF_PROP_SAMPLE_OA
, true,
956 /* OA unit configuration */
957 DRM_I915_PERF_PROP_OA_METRICS_SET
, metrics_set_id
,
958 DRM_I915_PERF_PROP_OA_FORMAT
, report_format
,
959 DRM_I915_PERF_PROP_OA_EXPONENT
, period_exponent
,
961 struct drm_i915_perf_open_param param
= {
962 .flags
= I915_PERF_FLAG_FD_CLOEXEC
|
963 I915_PERF_FLAG_FD_NONBLOCK
|
964 I915_PERF_FLAG_DISABLED
,
965 .num_properties
= ARRAY_SIZE(properties
) / 2,
966 .properties_ptr
= (uintptr_t) properties
,
968 int fd
= drmIoctl(drm_fd
, DRM_IOCTL_I915_PERF_OPEN
, ¶m
);
970 DBG("Error opening i915 perf OA stream: %m\n");
974 brw
->perfquery
.oa_stream_fd
= fd
;
976 brw
->perfquery
.current_oa_metrics_set_id
= metrics_set_id
;
977 brw
->perfquery
.current_oa_format
= report_format
;
983 close_perf(struct brw_context
*brw
,
984 const struct brw_perf_query_info
*query
)
986 if (brw
->perfquery
.oa_stream_fd
!= -1) {
987 close(brw
->perfquery
.oa_stream_fd
);
988 brw
->perfquery
.oa_stream_fd
= -1;
990 if (query
->kind
== OA_COUNTERS_RAW
) {
991 struct brw_perf_query_info
*raw_query
=
992 (struct brw_perf_query_info
*) query
;
993 raw_query
->oa_metrics_set_id
= 0;
998 capture_frequency_stat_register(struct brw_context
*brw
,
1002 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1004 if (devinfo
->gen
>= 7 && devinfo
->gen
<= 8 &&
1005 !devinfo
->is_baytrail
&& !devinfo
->is_cherryview
) {
1006 brw_store_register_mem32(brw
, bo
, GEN7_RPSTAT1
, bo_offset
);
1007 } else if (devinfo
->gen
>= 9) {
1008 brw_store_register_mem32(brw
, bo
, GEN9_RPSTAT0
, bo_offset
);
1013 * Driver hook for glBeginPerfQueryINTEL().
1016 brw_begin_perf_query(struct gl_context
*ctx
,
1017 struct gl_perf_query_object
*o
)
1019 struct brw_context
*brw
= brw_context(ctx
);
1020 struct brw_perf_query_object
*obj
= brw_perf_query(o
);
1021 const struct brw_perf_query_info
*query
= obj
->query
;
1023 /* We can assume the frontend hides mistaken attempts to Begin a
1024 * query object multiple times before its End. Similarly if an
1025 * application reuses a query object before results have arrived
1026 * the frontend will wait for prior results so we don't need
1027 * to support abandoning in-flight results.
1030 assert(!o
->Used
|| o
->Ready
); /* no in-flight query to worry about */
1032 DBG("Begin(%d)\n", o
->Id
);
1034 /* XXX: We have to consider that the command parser unit that parses batch
1035 * buffer commands and is used to capture begin/end counter snapshots isn't
1036 * implicitly synchronized with what's currently running across other GPU
1037 * units (such as the EUs running shaders) that the performance counters are
1040 * The intention of performance queries is to measure the work associated
1041 * with commands between the begin/end delimiters and so for that to be the
1042 * case we need to explicitly synchronize the parsing of commands to capture
1043 * Begin/End counter snapshots with what's running across other parts of the
1046 * When the command parser reaches a Begin marker it effectively needs to
1047 * drain everything currently running on the GPU until the hardware is idle
1048 * before capturing the first snapshot of counters - otherwise the results
1049 * would also be measuring the effects of earlier commands.
1051 * When the command parser reaches an End marker it needs to stall until
1052 * everything currently running on the GPU has finished before capturing the
1053 * end snapshot - otherwise the results won't be a complete representation
1056 * Theoretically there could be opportunities to minimize how much of the
1057 * GPU pipeline is drained, or that we stall for, when we know what specific
1058 * units the performance counters being queried relate to but we don't
1059 * currently attempt to be clever here.
1061 * Note: with our current simple approach here then for back-to-back queries
1062 * we will redundantly emit duplicate commands to synchronize the command
1063 * streamer with the rest of the GPU pipeline, but we assume that in HW the
1064 * second synchronization is effectively a NOOP.
1066 * N.B. The final results are based on deltas of counters between (inside)
1067 * Begin/End markers so even though the total wall clock time of the
1068 * workload is stretched by larger pipeline bubbles the bubbles themselves
1069 * are generally invisible to the query results. Whether that's a good or a
1070 * bad thing depends on the use case. For a lower real-time impact while
1071 * capturing metrics then periodic sampling may be a better choice than
1072 * INTEL_performance_query.
1075 * This is our Begin synchronization point to drain current work on the
1076 * GPU before we capture our first counter snapshot...
1078 brw_emit_mi_flush(brw
);
1080 switch (query
->kind
) {
1082 case OA_COUNTERS_RAW
: {
1084 /* Opening an i915 perf stream implies exclusive access to the OA unit
1085 * which will generate counter reports for a specific counter set with a
1086 * specific layout/format so we can't begin any OA based queries that
1087 * require a different counter set or format unless we get an opportunity
1088 * to close the stream and open a new one...
1090 uint64_t metric_id
= brw_perf_query_get_metric_id(brw
, query
);
1092 if (brw
->perfquery
.oa_stream_fd
!= -1 &&
1093 brw
->perfquery
.current_oa_metrics_set_id
!= metric_id
) {
1095 if (brw
->perfquery
.n_oa_users
!= 0) {
1096 DBG("WARNING: Begin(%d) failed already using perf config=%i/%"PRIu64
"\n",
1097 o
->Id
, brw
->perfquery
.current_oa_metrics_set_id
, metric_id
);
1100 close_perf(brw
, query
);
1103 /* If the OA counters aren't already on, enable them. */
1104 if (brw
->perfquery
.oa_stream_fd
== -1) {
1105 __DRIscreen
*screen
= brw
->screen
->driScrnPriv
;
1106 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1108 /* The period_exponent gives a sampling period as follows:
1109 * sample_period = timestamp_period * 2^(period_exponent + 1)
1111 * The timestamps increments every 80ns (HSW), ~52ns (GEN9LP) or
1114 * The counter overflow period is derived from the EuActive counter
1115 * which reads a counter that increments by the number of clock
1116 * cycles multiplied by the number of EUs. It can be calculated as:
1118 * 2^(number of bits in A counter) / (n_eus * max_gen_freq * 2)
1120 * (E.g. 40 EUs @ 1GHz = ~53ms)
1122 * We select a sampling period inferior to that overflow period to
1123 * ensure we cannot see more than 1 counter overflow, otherwise we
1124 * could loose information.
1127 int a_counter_in_bits
= 32;
1128 if (devinfo
->gen
>= 8)
1129 a_counter_in_bits
= 40;
1131 uint64_t overflow_period
= pow(2, a_counter_in_bits
) /
1132 (brw
->perfquery
.sys_vars
.n_eus
*
1133 /* drop 1GHz freq to have units in nanoseconds */
1136 DBG("A counter overflow period: %"PRIu64
"ns, %"PRIu64
"ms (n_eus=%"PRIu64
")\n",
1137 overflow_period
, overflow_period
/ 1000000ul, brw
->perfquery
.sys_vars
.n_eus
);
1139 int period_exponent
= 0;
1140 uint64_t prev_sample_period
, next_sample_period
;
1141 for (int e
= 0; e
< 30; e
++) {
1142 prev_sample_period
= 1000000000ull * pow(2, e
+ 1) / devinfo
->timestamp_frequency
;
1143 next_sample_period
= 1000000000ull * pow(2, e
+ 2) / devinfo
->timestamp_frequency
;
1145 /* Take the previous sampling period, lower than the overflow
1148 if (prev_sample_period
< overflow_period
&&
1149 next_sample_period
> overflow_period
)
1150 period_exponent
= e
+ 1;
1153 if (period_exponent
== 0) {
1154 DBG("WARNING: enable to find a sampling exponent\n");
1158 DBG("OA sampling exponent: %i ~= %"PRIu64
"ms\n", period_exponent
,
1159 prev_sample_period
/ 1000000ul);
1161 if (!open_i915_perf_oa_stream(brw
,
1165 screen
->fd
, /* drm fd */
1169 assert(brw
->perfquery
.current_oa_metrics_set_id
== metric_id
&&
1170 brw
->perfquery
.current_oa_format
== query
->oa_format
);
1173 if (!inc_n_oa_users(brw
)) {
1174 DBG("WARNING: Error enabling i915 perf stream: %m\n");
1179 brw_bo_unreference(obj
->oa
.bo
);
1184 brw_bo_alloc(brw
->bufmgr
, "perf. query OA MI_RPC bo", MI_RPC_BO_SIZE
,
1187 /* Pre-filling the BO helps debug whether writes landed. */
1188 void *map
= brw_bo_map(brw
, obj
->oa
.bo
, MAP_WRITE
);
1189 memset(map
, 0x80, MI_RPC_BO_SIZE
);
1190 brw_bo_unmap(obj
->oa
.bo
);
1193 obj
->oa
.begin_report_id
= brw
->perfquery
.next_query_start_report_id
;
1194 brw
->perfquery
.next_query_start_report_id
+= 2;
1196 /* We flush the batchbuffer here to minimize the chances that MI_RPC
1197 * delimiting commands end up in different batchbuffers. If that's the
1198 * case, the measurement will include the time it takes for the kernel
1199 * scheduler to load a new request into the hardware. This is manifested in
1200 * tools like frameretrace by spikes in the "GPU Core Clocks" counter.
1202 intel_batchbuffer_flush(brw
);
1204 /* Take a starting OA counter snapshot. */
1205 brw
->vtbl
.emit_mi_report_perf_count(brw
, obj
->oa
.bo
, 0,
1206 obj
->oa
.begin_report_id
);
1207 capture_frequency_stat_register(brw
, obj
->oa
.bo
, MI_FREQ_START_OFFSET_BYTES
);
1209 ++brw
->perfquery
.n_active_oa_queries
;
1211 /* No already-buffered samples can possibly be associated with this query
1212 * so create a marker within the list of sample buffers enabling us to
1213 * easily ignore earlier samples when processing this query after
1216 assert(!exec_list_is_empty(&brw
->perfquery
.sample_buffers
));
1217 obj
->oa
.samples_head
= exec_list_get_tail(&brw
->perfquery
.sample_buffers
);
1219 struct brw_oa_sample_buf
*buf
=
1220 exec_node_data(struct brw_oa_sample_buf
, obj
->oa
.samples_head
, link
);
1222 /* This reference will ensure that future/following sample
1223 * buffers (that may relate to this query) can't be freed until
1224 * this drops to zero.
1228 obj
->oa
.hw_id
= 0xffffffff;
1229 memset(obj
->oa
.accumulator
, 0, sizeof(obj
->oa
.accumulator
));
1230 obj
->oa
.results_accumulated
= false;
1232 add_to_unaccumulated_query_list(brw
, obj
);
1236 case PIPELINE_STATS
:
1237 if (obj
->pipeline_stats
.bo
) {
1238 brw_bo_unreference(obj
->pipeline_stats
.bo
);
1239 obj
->pipeline_stats
.bo
= NULL
;
1242 obj
->pipeline_stats
.bo
=
1243 brw_bo_alloc(brw
->bufmgr
, "perf. query pipeline stats bo",
1244 STATS_BO_SIZE
, BRW_MEMZONE_OTHER
);
1246 /* Take starting snapshots. */
1247 snapshot_statistics_registers(brw
, obj
, 0);
1249 ++brw
->perfquery
.n_active_pipeline_stats_queries
;
1253 unreachable("Unknown query type");
1257 if (INTEL_DEBUG
& DEBUG_PERFMON
)
1258 dump_perf_queries(brw
);
1264 * Driver hook for glEndPerfQueryINTEL().
1267 brw_end_perf_query(struct gl_context
*ctx
,
1268 struct gl_perf_query_object
*o
)
1270 struct brw_context
*brw
= brw_context(ctx
);
1271 struct brw_perf_query_object
*obj
= brw_perf_query(o
);
1273 DBG("End(%d)\n", o
->Id
);
1275 /* Ensure that the work associated with the queried commands will have
1276 * finished before taking our query end counter readings.
1278 * For more details see comment in brw_begin_perf_query for
1279 * corresponding flush.
1281 brw_emit_mi_flush(brw
);
1283 switch (obj
->query
->kind
) {
1285 case OA_COUNTERS_RAW
:
1287 /* NB: It's possible that the query will have already been marked
1288 * as 'accumulated' if an error was seen while reading samples
1289 * from perf. In this case we mustn't try and emit a closing
1290 * MI_RPC command in case the OA unit has already been disabled
1292 if (!obj
->oa
.results_accumulated
) {
1293 /* Take an ending OA counter snapshot. */
1294 capture_frequency_stat_register(brw
, obj
->oa
.bo
, MI_FREQ_END_OFFSET_BYTES
);
1295 brw
->vtbl
.emit_mi_report_perf_count(brw
, obj
->oa
.bo
,
1296 MI_RPC_BO_END_OFFSET_BYTES
,
1297 obj
->oa
.begin_report_id
+ 1);
1300 --brw
->perfquery
.n_active_oa_queries
;
1302 /* NB: even though the query has now ended, it can't be accumulated
1303 * until the end MI_REPORT_PERF_COUNT snapshot has been written
1308 case PIPELINE_STATS
:
1309 snapshot_statistics_registers(brw
, obj
,
1310 STATS_BO_END_OFFSET_BYTES
);
1311 --brw
->perfquery
.n_active_pipeline_stats_queries
;
1315 unreachable("Unknown query type");
1321 brw_wait_perf_query(struct gl_context
*ctx
, struct gl_perf_query_object
*o
)
1323 struct brw_context
*brw
= brw_context(ctx
);
1324 struct brw_perf_query_object
*obj
= brw_perf_query(o
);
1325 struct brw_bo
*bo
= NULL
;
1329 switch (obj
->query
->kind
) {
1331 case OA_COUNTERS_RAW
:
1335 case PIPELINE_STATS
:
1336 bo
= obj
->pipeline_stats
.bo
;
1340 unreachable("Unknown query type");
1347 /* If the current batch references our results bo then we need to
1350 if (brw_batch_references(&brw
->batch
, bo
))
1351 intel_batchbuffer_flush(brw
);
1353 brw_bo_wait_rendering(bo
);
1355 /* Due to a race condition between the OA unit signaling report
1356 * availability and the report actually being written into memory,
1357 * we need to wait for all the reports to come in before we can
1360 if (obj
->query
->kind
== OA_COUNTERS
||
1361 obj
->query
->kind
== OA_COUNTERS_RAW
) {
1362 while (!read_oa_samples_for_query(brw
, obj
))
1368 brw_is_perf_query_ready(struct gl_context
*ctx
,
1369 struct gl_perf_query_object
*o
)
1371 struct brw_context
*brw
= brw_context(ctx
);
1372 struct brw_perf_query_object
*obj
= brw_perf_query(o
);
1377 switch (obj
->query
->kind
) {
1379 case OA_COUNTERS_RAW
:
1380 return (obj
->oa
.results_accumulated
||
1382 !brw_batch_references(&brw
->batch
, obj
->oa
.bo
) &&
1383 !brw_bo_busy(obj
->oa
.bo
) &&
1384 read_oa_samples_for_query(brw
, obj
)));
1385 case PIPELINE_STATS
:
1386 return (obj
->pipeline_stats
.bo
&&
1387 !brw_batch_references(&brw
->batch
, obj
->pipeline_stats
.bo
) &&
1388 !brw_bo_busy(obj
->pipeline_stats
.bo
));
1391 unreachable("Unknown query type");
1399 gen8_read_report_clock_ratios(const uint32_t *report
,
1400 uint64_t *slice_freq_hz
,
1401 uint64_t *unslice_freq_hz
)
1403 /* The lower 16bits of the RPT_ID field of the OA reports contains a
1404 * snapshot of the bits coming from the RP_FREQ_NORMAL register and is
1405 * divided this way :
1407 * RPT_ID[31:25]: RP_FREQ_NORMAL[20:14] (low squashed_slice_clock_frequency)
1408 * RPT_ID[10:9]: RP_FREQ_NORMAL[22:21] (high squashed_slice_clock_frequency)
1409 * RPT_ID[8:0]: RP_FREQ_NORMAL[31:23] (squashed_unslice_clock_frequency)
1411 * RP_FREQ_NORMAL[31:23]: Software Unslice Ratio Request
1412 * Multiple of 33.33MHz 2xclk (16 MHz 1xclk)
1414 * RP_FREQ_NORMAL[22:14]: Software Slice Ratio Request
1415 * Multiple of 33.33MHz 2xclk (16 MHz 1xclk)
1418 uint32_t unslice_freq
= report
[0] & 0x1ff;
1419 uint32_t slice_freq_low
= (report
[0] >> 25) & 0x7f;
1420 uint32_t slice_freq_high
= (report
[0] >> 9) & 0x3;
1421 uint32_t slice_freq
= slice_freq_low
| (slice_freq_high
<< 7);
1423 *slice_freq_hz
= slice_freq
* 16666667ULL;
1424 *unslice_freq_hz
= unslice_freq
* 16666667ULL;
1428 read_slice_unslice_frequencies(struct brw_context
*brw
,
1429 struct brw_perf_query_object
*obj
)
1431 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1432 uint32_t *begin_report
, *end_report
;
1434 /* Slice/Unslice frequency is only available in the OA reports when the
1435 * "Disable OA reports due to clock ratio change" field in
1436 * OA_DEBUG_REGISTER is set to 1. This is how the kernel programs this
1437 * global register (see drivers/gpu/drm/i915/i915_perf.c)
1439 * Documentation says this should be available on Gen9+ but experimentation
1440 * shows that Gen8 reports similar values, so we enable it there too.
1442 if (devinfo
->gen
< 8)
1445 begin_report
= obj
->oa
.map
;
1446 end_report
= obj
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
1448 gen8_read_report_clock_ratios(begin_report
,
1449 &obj
->oa
.slice_frequency
[0],
1450 &obj
->oa
.unslice_frequency
[0]);
1451 gen8_read_report_clock_ratios(end_report
,
1452 &obj
->oa
.slice_frequency
[1],
1453 &obj
->oa
.unslice_frequency
[1]);
1457 read_gt_frequency(struct brw_context
*brw
,
1458 struct brw_perf_query_object
*obj
)
1460 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1461 uint32_t start
= *((uint32_t *)(obj
->oa
.map
+ MI_FREQ_START_OFFSET_BYTES
)),
1462 end
= *((uint32_t *)(obj
->oa
.map
+ MI_FREQ_END_OFFSET_BYTES
));
1464 switch (devinfo
->gen
) {
1467 obj
->oa
.gt_frequency
[0] = GET_FIELD(start
, GEN7_RPSTAT1_CURR_GT_FREQ
) * 50ULL;
1468 obj
->oa
.gt_frequency
[1] = GET_FIELD(end
, GEN7_RPSTAT1_CURR_GT_FREQ
) * 50ULL;
1473 obj
->oa
.gt_frequency
[0] = GET_FIELD(start
, GEN9_RPSTAT0_CURR_GT_FREQ
) * 50ULL / 3ULL;
1474 obj
->oa
.gt_frequency
[1] = GET_FIELD(end
, GEN9_RPSTAT0_CURR_GT_FREQ
) * 50ULL / 3ULL;
1477 unreachable("unexpected gen");
1480 /* Put the numbers into Hz. */
1481 obj
->oa
.gt_frequency
[0] *= 1000000ULL;
1482 obj
->oa
.gt_frequency
[1] *= 1000000ULL;
1486 get_oa_counter_data(struct brw_context
*brw
,
1487 struct brw_perf_query_object
*obj
,
1491 const struct brw_perf_query_info
*query
= obj
->query
;
1492 int n_counters
= query
->n_counters
;
1495 for (int i
= 0; i
< n_counters
; i
++) {
1496 const struct brw_perf_query_counter
*counter
= &query
->counters
[i
];
1497 uint64_t *out_uint64
;
1500 if (counter
->size
) {
1501 switch (counter
->data_type
) {
1502 case GL_PERFQUERY_COUNTER_DATA_UINT64_INTEL
:
1503 out_uint64
= (uint64_t *)(data
+ counter
->offset
);
1504 *out_uint64
= counter
->oa_counter_read_uint64(brw
, query
,
1505 obj
->oa
.accumulator
);
1507 case GL_PERFQUERY_COUNTER_DATA_FLOAT_INTEL
:
1508 out_float
= (float *)(data
+ counter
->offset
);
1509 *out_float
= counter
->oa_counter_read_float(brw
, query
,
1510 obj
->oa
.accumulator
);
1513 /* So far we aren't using uint32, double or bool32... */
1514 unreachable("unexpected counter data type");
1516 written
= counter
->offset
+ counter
->size
;
1524 get_pipeline_stats_data(struct brw_context
*brw
,
1525 struct brw_perf_query_object
*obj
,
1530 const struct brw_perf_query_info
*query
= obj
->query
;
1531 int n_counters
= obj
->query
->n_counters
;
1534 uint64_t *start
= brw_bo_map(brw
, obj
->pipeline_stats
.bo
, MAP_READ
);
1535 uint64_t *end
= start
+ (STATS_BO_END_OFFSET_BYTES
/ sizeof(uint64_t));
1537 for (int i
= 0; i
< n_counters
; i
++) {
1538 const struct brw_perf_query_counter
*counter
= &query
->counters
[i
];
1539 uint64_t value
= end
[i
] - start
[i
];
1541 if (counter
->pipeline_stat
.numerator
!=
1542 counter
->pipeline_stat
.denominator
) {
1543 value
*= counter
->pipeline_stat
.numerator
;
1544 value
/= counter
->pipeline_stat
.denominator
;
1547 *((uint64_t *)p
) = value
;
1551 brw_bo_unmap(obj
->pipeline_stats
.bo
);
1557 * Driver hook for glGetPerfQueryDataINTEL().
1560 brw_get_perf_query_data(struct gl_context
*ctx
,
1561 struct gl_perf_query_object
*o
,
1564 GLuint
*bytes_written
)
1566 struct brw_context
*brw
= brw_context(ctx
);
1567 struct brw_perf_query_object
*obj
= brw_perf_query(o
);
1570 assert(brw_is_perf_query_ready(ctx
, o
));
1572 DBG("GetData(%d)\n", o
->Id
);
1574 if (INTEL_DEBUG
& DEBUG_PERFMON
)
1575 dump_perf_queries(brw
);
1577 /* We expect that the frontend only calls this hook when it knows
1578 * that results are available.
1582 switch (obj
->query
->kind
) {
1584 case OA_COUNTERS_RAW
:
1585 if (!obj
->oa
.results_accumulated
) {
1586 read_gt_frequency(brw
, obj
);
1587 read_slice_unslice_frequencies(brw
, obj
);
1588 accumulate_oa_reports(brw
, obj
);
1589 assert(obj
->oa
.results_accumulated
);
1591 brw_bo_unmap(obj
->oa
.bo
);
1594 if (obj
->query
->kind
== OA_COUNTERS
)
1595 written
= get_oa_counter_data(brw
, obj
, data_size
, (uint8_t *)data
);
1597 written
= brw_perf_query_get_mdapi_oa_data(brw
, obj
, data_size
, (uint8_t *)data
);
1600 case PIPELINE_STATS
:
1601 written
= get_pipeline_stats_data(brw
, obj
, data_size
, (uint8_t *)data
);
1605 unreachable("Unknown query type");
1610 *bytes_written
= written
;
1613 static struct gl_perf_query_object
*
1614 brw_new_perf_query_object(struct gl_context
*ctx
, unsigned query_index
)
1616 struct brw_context
*brw
= brw_context(ctx
);
1617 const struct brw_perf_query_info
*query
=
1618 &brw
->perfquery
.queries
[query_index
];
1619 struct brw_perf_query_object
*obj
=
1620 calloc(1, sizeof(struct brw_perf_query_object
));
1627 brw
->perfquery
.n_query_instances
++;
1633 * Driver hook for glDeletePerfQueryINTEL().
1636 brw_delete_perf_query(struct gl_context
*ctx
,
1637 struct gl_perf_query_object
*o
)
1639 struct brw_context
*brw
= brw_context(ctx
);
1640 struct brw_perf_query_object
*obj
= brw_perf_query(o
);
1642 /* We can assume that the frontend waits for a query to complete
1643 * before ever calling into here, so we don't have to worry about
1644 * deleting an in-flight query object.
1647 assert(!o
->Used
|| o
->Ready
);
1649 DBG("Delete(%d)\n", o
->Id
);
1651 switch (obj
->query
->kind
) {
1653 case OA_COUNTERS_RAW
:
1655 if (!obj
->oa
.results_accumulated
) {
1656 drop_from_unaccumulated_query_list(brw
, obj
);
1657 dec_n_oa_users(brw
);
1660 brw_bo_unreference(obj
->oa
.bo
);
1664 obj
->oa
.results_accumulated
= false;
1667 case PIPELINE_STATS
:
1668 if (obj
->pipeline_stats
.bo
) {
1669 brw_bo_unreference(obj
->pipeline_stats
.bo
);
1670 obj
->pipeline_stats
.bo
= NULL
;
1675 unreachable("Unknown query type");
1679 /* As an indication that the INTEL_performance_query extension is no
1680 * longer in use, it's a good time to free our cache of sample
1681 * buffers and close any current i915-perf stream.
1683 if (--brw
->perfquery
.n_query_instances
== 0) {
1684 free_sample_bufs(brw
);
1685 close_perf(brw
, obj
->query
);
1691 /******************************************************************************/
1694 init_pipeline_statistic_query_registers(struct brw_context
*brw
)
1696 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1697 struct brw_perf_query_info
*query
= brw_perf_query_append_query_info(brw
);
1699 query
->kind
= PIPELINE_STATS
;
1700 query
->name
= "Pipeline Statistics Registers";
1701 query
->n_counters
= 0;
1703 rzalloc_array(brw
, struct brw_perf_query_counter
, MAX_STAT_COUNTERS
);
1705 brw_perf_query_info_add_basic_stat_reg(query
, IA_VERTICES_COUNT
,
1706 "N vertices submitted");
1707 brw_perf_query_info_add_basic_stat_reg(query
, IA_PRIMITIVES_COUNT
,
1708 "N primitives submitted");
1709 brw_perf_query_info_add_basic_stat_reg(query
, VS_INVOCATION_COUNT
,
1710 "N vertex shader invocations");
1712 if (devinfo
->gen
== 6) {
1713 brw_perf_query_info_add_stat_reg(query
, GEN6_SO_PRIM_STORAGE_NEEDED
, 1, 1,
1714 "SO_PRIM_STORAGE_NEEDED",
1715 "N geometry shader stream-out primitives (total)");
1716 brw_perf_query_info_add_stat_reg(query
, GEN6_SO_NUM_PRIMS_WRITTEN
, 1, 1,
1717 "SO_NUM_PRIMS_WRITTEN",
1718 "N geometry shader stream-out primitives (written)");
1720 brw_perf_query_info_add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(0), 1, 1,
1721 "SO_PRIM_STORAGE_NEEDED (Stream 0)",
1722 "N stream-out (stream 0) primitives (total)");
1723 brw_perf_query_info_add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(1), 1, 1,
1724 "SO_PRIM_STORAGE_NEEDED (Stream 1)",
1725 "N stream-out (stream 1) primitives (total)");
1726 brw_perf_query_info_add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(2), 1, 1,
1727 "SO_PRIM_STORAGE_NEEDED (Stream 2)",
1728 "N stream-out (stream 2) primitives (total)");
1729 brw_perf_query_info_add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(3), 1, 1,
1730 "SO_PRIM_STORAGE_NEEDED (Stream 3)",
1731 "N stream-out (stream 3) primitives (total)");
1732 brw_perf_query_info_add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(0), 1, 1,
1733 "SO_NUM_PRIMS_WRITTEN (Stream 0)",
1734 "N stream-out (stream 0) primitives (written)");
1735 brw_perf_query_info_add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(1), 1, 1,
1736 "SO_NUM_PRIMS_WRITTEN (Stream 1)",
1737 "N stream-out (stream 1) primitives (written)");
1738 brw_perf_query_info_add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(2), 1, 1,
1739 "SO_NUM_PRIMS_WRITTEN (Stream 2)",
1740 "N stream-out (stream 2) primitives (written)");
1741 brw_perf_query_info_add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(3), 1, 1,
1742 "SO_NUM_PRIMS_WRITTEN (Stream 3)",
1743 "N stream-out (stream 3) primitives (written)");
1746 brw_perf_query_info_add_basic_stat_reg(query
, HS_INVOCATION_COUNT
,
1747 "N TCS shader invocations");
1748 brw_perf_query_info_add_basic_stat_reg(query
, DS_INVOCATION_COUNT
,
1749 "N TES shader invocations");
1751 brw_perf_query_info_add_basic_stat_reg(query
, GS_INVOCATION_COUNT
,
1752 "N geometry shader invocations");
1753 brw_perf_query_info_add_basic_stat_reg(query
, GS_PRIMITIVES_COUNT
,
1754 "N geometry shader primitives emitted");
1756 brw_perf_query_info_add_basic_stat_reg(query
, CL_INVOCATION_COUNT
,
1757 "N primitives entering clipping");
1758 brw_perf_query_info_add_basic_stat_reg(query
, CL_PRIMITIVES_COUNT
,
1759 "N primitives leaving clipping");
1761 if (devinfo
->is_haswell
|| devinfo
->gen
== 8)
1762 brw_perf_query_info_add_stat_reg(query
, PS_INVOCATION_COUNT
, 1, 4,
1763 "N fragment shader invocations",
1764 "N fragment shader invocations");
1766 brw_perf_query_info_add_basic_stat_reg(query
, PS_INVOCATION_COUNT
,
1767 "N fragment shader invocations");
1769 brw_perf_query_info_add_basic_stat_reg(query
, PS_DEPTH_COUNT
, "N z-pass fragments");
1771 if (devinfo
->gen
>= 7)
1772 brw_perf_query_info_add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
1773 "N compute shader invocations");
1775 query
->data_size
= sizeof(uint64_t) * query
->n_counters
;
1779 register_oa_config(struct brw_context
*brw
,
1780 const struct brw_perf_query_info
*query
,
1783 struct brw_perf_query_info
*registred_query
=
1784 brw_perf_query_append_query_info(brw
);
1786 *registred_query
= *query
;
1787 registred_query
->oa_metrics_set_id
= config_id
;
1788 DBG("metric set registred: id = %" PRIu64
", guid = %s\n",
1789 registred_query
->oa_metrics_set_id
, query
->guid
);
1793 enumerate_sysfs_metrics(struct brw_context
*brw
)
1796 DIR *metricsdir
= NULL
;
1797 struct dirent
*metric_entry
;
1800 len
= snprintf(buf
, sizeof(buf
), "%s/metrics", brw
->perfquery
.sysfs_dev_dir
);
1801 if (len
< 0 || len
>= sizeof(buf
)) {
1802 DBG("Failed to concatenate path to sysfs metrics/ directory\n");
1806 metricsdir
= opendir(buf
);
1808 DBG("Failed to open %s: %m\n", buf
);
1812 while ((metric_entry
= readdir(metricsdir
))) {
1813 struct hash_entry
*entry
;
1815 if ((metric_entry
->d_type
!= DT_DIR
&&
1816 metric_entry
->d_type
!= DT_LNK
) ||
1817 metric_entry
->d_name
[0] == '.')
1820 DBG("metric set: %s\n", metric_entry
->d_name
);
1821 entry
= _mesa_hash_table_search(brw
->perfquery
.oa_metrics_table
,
1822 metric_entry
->d_name
);
1826 len
= snprintf(buf
, sizeof(buf
), "%s/metrics/%s/id",
1827 brw
->perfquery
.sysfs_dev_dir
, metric_entry
->d_name
);
1828 if (len
< 0 || len
>= sizeof(buf
)) {
1829 DBG("Failed to concatenate path to sysfs metric id file\n");
1833 if (!read_file_uint64(buf
, &id
)) {
1834 DBG("Failed to read metric set id from %s: %m", buf
);
1838 register_oa_config(brw
, (const struct brw_perf_query_info
*)entry
->data
, id
);
1840 DBG("metric set not known by mesa (skipping)\n");
1843 closedir(metricsdir
);
1847 kernel_has_dynamic_config_support(struct brw_context
*brw
)
1849 __DRIscreen
*screen
= brw
->screen
->driScrnPriv
;
1850 struct hash_entry
*entry
;
1852 hash_table_foreach(brw
->perfquery
.oa_metrics_table
, entry
) {
1853 struct brw_perf_query_info
*query
= entry
->data
;
1854 char config_path
[280];
1857 snprintf(config_path
, sizeof(config_path
), "%s/metrics/%s/id",
1858 brw
->perfquery
.sysfs_dev_dir
, query
->guid
);
1860 /* Look for the test config, which we know we can't replace. */
1861 if (read_file_uint64(config_path
, &config_id
) && config_id
== 1) {
1862 return drmIoctl(screen
->fd
, DRM_IOCTL_I915_PERF_REMOVE_CONFIG
,
1863 &config_id
) < 0 && errno
== ENOENT
;
1871 init_oa_configs(struct brw_context
*brw
)
1873 __DRIscreen
*screen
= brw
->screen
->driScrnPriv
;
1874 struct hash_entry
*entry
;
1876 hash_table_foreach(brw
->perfquery
.oa_metrics_table
, entry
) {
1877 const struct brw_perf_query_info
*query
= entry
->data
;
1878 struct drm_i915_perf_oa_config config
;
1879 char config_path
[280];
1883 snprintf(config_path
, sizeof(config_path
), "%s/metrics/%s/id",
1884 brw
->perfquery
.sysfs_dev_dir
, query
->guid
);
1886 /* Don't recreate already loaded configs. */
1887 if (read_file_uint64(config_path
, &config_id
)) {
1888 DBG("metric set: %s (already loaded)\n", query
->guid
);
1889 register_oa_config(brw
, query
, config_id
);
1893 memset(&config
, 0, sizeof(config
));
1895 memcpy(config
.uuid
, query
->guid
, sizeof(config
.uuid
));
1897 config
.n_mux_regs
= query
->n_mux_regs
;
1898 config
.mux_regs_ptr
= (uintptr_t) query
->mux_regs
;
1900 config
.n_boolean_regs
= query
->n_b_counter_regs
;
1901 config
.boolean_regs_ptr
= (uintptr_t) query
->b_counter_regs
;
1903 config
.n_flex_regs
= query
->n_flex_regs
;
1904 config
.flex_regs_ptr
= (uintptr_t) query
->flex_regs
;
1906 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_I915_PERF_ADD_CONFIG
, &config
);
1908 DBG("Failed to load \"%s\" (%s) metrics set in kernel: %s\n",
1909 query
->name
, query
->guid
, strerror(errno
));
1913 register_oa_config(brw
, query
, ret
);
1914 DBG("metric set: %s (added)\n", query
->guid
);
1919 query_topology(struct brw_context
*brw
)
1921 __DRIscreen
*screen
= brw
->screen
->driScrnPriv
;
1922 struct drm_i915_query_item item
= {
1923 .query_id
= DRM_I915_QUERY_TOPOLOGY_INFO
,
1925 struct drm_i915_query query
= {
1927 .items_ptr
= (uintptr_t) &item
,
1930 if (drmIoctl(screen
->fd
, DRM_IOCTL_I915_QUERY
, &query
))
1933 struct drm_i915_query_topology_info
*topo_info
=
1934 (struct drm_i915_query_topology_info
*) calloc(1, item
.length
);
1935 item
.data_ptr
= (uintptr_t) topo_info
;
1937 if (drmIoctl(screen
->fd
, DRM_IOCTL_I915_QUERY
, &query
) ||
1941 gen_device_info_update_from_topology(&brw
->screen
->devinfo
,
1950 getparam_topology(struct brw_context
*brw
)
1952 __DRIscreen
*screen
= brw
->screen
->driScrnPriv
;
1953 drm_i915_getparam_t gp
;
1957 gp
.param
= I915_PARAM_SLICE_MASK
;
1958 gp
.value
= &slice_mask
;
1959 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_I915_GETPARAM
, &gp
);
1963 int subslice_mask
= 0;
1964 gp
.param
= I915_PARAM_SUBSLICE_MASK
;
1965 gp
.value
= &subslice_mask
;
1966 ret
= drmIoctl(screen
->fd
, DRM_IOCTL_I915_GETPARAM
, &gp
);
1970 gen_device_info_update_from_masks(&brw
->screen
->devinfo
,
1973 brw
->screen
->eu_total
);
1979 compute_topology_builtins(struct brw_context
*brw
)
1981 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
1983 brw
->perfquery
.sys_vars
.slice_mask
= devinfo
->slice_masks
;
1984 brw
->perfquery
.sys_vars
.n_eu_slices
= devinfo
->num_slices
;
1986 for (int i
= 0; i
< sizeof(devinfo
->subslice_masks
[i
]); i
++) {
1987 brw
->perfquery
.sys_vars
.n_eu_sub_slices
+=
1988 _mesa_bitcount(devinfo
->subslice_masks
[i
]);
1991 for (int i
= 0; i
< sizeof(devinfo
->eu_masks
); i
++)
1992 brw
->perfquery
.sys_vars
.n_eus
+= _mesa_bitcount(devinfo
->eu_masks
[i
]);
1994 brw
->perfquery
.sys_vars
.eu_threads_count
=
1995 brw
->perfquery
.sys_vars
.n_eus
* devinfo
->num_thread_per_eu
;
1997 /* At the moment the subslice mask builtin has groups of 3bits for each
2000 * Ideally equations would be updated to have a slice/subslice query
2001 * function/operator.
2003 brw
->perfquery
.sys_vars
.subslice_mask
= 0;
2004 for (int s
= 0; s
< util_last_bit(devinfo
->slice_masks
); s
++) {
2005 for (int ss
= 0; ss
< (devinfo
->subslice_slice_stride
* 8); ss
++) {
2006 if (gen_device_info_subslice_available(devinfo
, s
, ss
))
2007 brw
->perfquery
.sys_vars
.subslice_mask
|= 1UL << (s
* 3 + ss
);
2013 init_oa_sys_vars(struct brw_context
*brw
)
2015 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
2016 uint64_t min_freq_mhz
= 0, max_freq_mhz
= 0;
2017 __DRIscreen
*screen
= brw
->screen
->driScrnPriv
;
2019 if (!read_sysfs_drm_device_file_uint64(brw
, "gt_min_freq_mhz", &min_freq_mhz
))
2022 if (!read_sysfs_drm_device_file_uint64(brw
, "gt_max_freq_mhz", &max_freq_mhz
))
2025 if (!query_topology(brw
)) {
2026 /* We need the i915 query uAPI on CNL+ (kernel 4.17+). */
2027 if (devinfo
->gen
>= 10)
2030 if (!getparam_topology(brw
)) {
2031 /* We need the SLICE_MASK/SUBSLICE_MASK on gen8+ (kernel 4.13+). */
2032 if (devinfo
->gen
>= 8)
2035 /* On Haswell, the values are already computed for us in
2041 memset(&brw
->perfquery
.sys_vars
, 0, sizeof(brw
->perfquery
.sys_vars
));
2042 brw
->perfquery
.sys_vars
.gt_min_freq
= min_freq_mhz
* 1000000;
2043 brw
->perfquery
.sys_vars
.gt_max_freq
= max_freq_mhz
* 1000000;
2044 brw
->perfquery
.sys_vars
.timestamp_frequency
= devinfo
->timestamp_frequency
;
2045 brw
->perfquery
.sys_vars
.revision
= intel_device_get_revision(screen
->fd
);
2046 compute_topology_builtins(brw
);
2052 get_sysfs_dev_dir(struct brw_context
*brw
)
2054 __DRIscreen
*screen
= brw
->screen
->driScrnPriv
;
2058 struct dirent
*drm_entry
;
2061 brw
->perfquery
.sysfs_dev_dir
[0] = '\0';
2063 if (fstat(screen
->fd
, &sb
)) {
2064 DBG("Failed to stat DRM fd\n");
2068 maj
= major(sb
.st_rdev
);
2069 min
= minor(sb
.st_rdev
);
2071 if (!S_ISCHR(sb
.st_mode
)) {
2072 DBG("DRM fd is not a character device as expected\n");
2076 len
= snprintf(brw
->perfquery
.sysfs_dev_dir
,
2077 sizeof(brw
->perfquery
.sysfs_dev_dir
),
2078 "/sys/dev/char/%d:%d/device/drm", maj
, min
);
2079 if (len
< 0 || len
>= sizeof(brw
->perfquery
.sysfs_dev_dir
)) {
2080 DBG("Failed to concatenate sysfs path to drm device\n");
2084 drmdir
= opendir(brw
->perfquery
.sysfs_dev_dir
);
2086 DBG("Failed to open %s: %m\n", brw
->perfquery
.sysfs_dev_dir
);
2090 while ((drm_entry
= readdir(drmdir
))) {
2091 if ((drm_entry
->d_type
== DT_DIR
||
2092 drm_entry
->d_type
== DT_LNK
) &&
2093 strncmp(drm_entry
->d_name
, "card", 4) == 0)
2095 len
= snprintf(brw
->perfquery
.sysfs_dev_dir
,
2096 sizeof(brw
->perfquery
.sysfs_dev_dir
),
2097 "/sys/dev/char/%d:%d/device/drm/%s",
2098 maj
, min
, drm_entry
->d_name
);
2100 if (len
< 0 || len
>= sizeof(brw
->perfquery
.sysfs_dev_dir
))
2109 DBG("Failed to find cardX directory under /sys/dev/char/%d:%d/device/drm\n",
2115 typedef void (*perf_register_oa_queries_t
)(struct brw_context
*);
2117 static perf_register_oa_queries_t
2118 get_register_queries_function(const struct gen_device_info
*devinfo
)
2120 if (devinfo
->is_haswell
)
2121 return brw_oa_register_queries_hsw
;
2122 if (devinfo
->is_cherryview
)
2123 return brw_oa_register_queries_chv
;
2124 if (devinfo
->is_broadwell
)
2125 return brw_oa_register_queries_bdw
;
2126 if (devinfo
->is_broxton
)
2127 return brw_oa_register_queries_bxt
;
2128 if (devinfo
->is_skylake
) {
2129 if (devinfo
->gt
== 2)
2130 return brw_oa_register_queries_sklgt2
;
2131 if (devinfo
->gt
== 3)
2132 return brw_oa_register_queries_sklgt3
;
2133 if (devinfo
->gt
== 4)
2134 return brw_oa_register_queries_sklgt4
;
2136 if (devinfo
->is_kabylake
) {
2137 if (devinfo
->gt
== 2)
2138 return brw_oa_register_queries_kblgt2
;
2139 if (devinfo
->gt
== 3)
2140 return brw_oa_register_queries_kblgt3
;
2142 if (devinfo
->is_geminilake
)
2143 return brw_oa_register_queries_glk
;
2144 if (devinfo
->is_coffeelake
) {
2145 if (devinfo
->gt
== 2)
2146 return brw_oa_register_queries_cflgt2
;
2147 if (devinfo
->gt
== 3)
2148 return brw_oa_register_queries_cflgt3
;
2150 if (devinfo
->is_cannonlake
)
2151 return brw_oa_register_queries_cnl
;
2157 brw_init_perf_query_info(struct gl_context
*ctx
)
2159 struct brw_context
*brw
= brw_context(ctx
);
2160 const struct gen_device_info
*devinfo
= &brw
->screen
->devinfo
;
2161 bool i915_perf_oa_available
= false;
2163 perf_register_oa_queries_t oa_register
;
2165 if (brw
->perfquery
.n_queries
)
2166 return brw
->perfquery
.n_queries
;
2168 init_pipeline_statistic_query_registers(brw
);
2169 brw_perf_query_register_mdapi_statistic_query(brw
);
2171 oa_register
= get_register_queries_function(devinfo
);
2173 /* The existence of this sysctl parameter implies the kernel supports
2174 * the i915 perf interface.
2176 if (stat("/proc/sys/dev/i915/perf_stream_paranoid", &sb
) == 0) {
2178 /* If _paranoid == 1 then on Gen8+ we won't be able to access OA
2179 * metrics unless running as root.
2181 if (devinfo
->is_haswell
)
2182 i915_perf_oa_available
= true;
2184 uint64_t paranoid
= 1;
2186 read_file_uint64("/proc/sys/dev/i915/perf_stream_paranoid", ¶noid
);
2188 if (paranoid
== 0 || geteuid() == 0)
2189 i915_perf_oa_available
= true;
2193 if (i915_perf_oa_available
&&
2195 get_sysfs_dev_dir(brw
) &&
2196 init_oa_sys_vars(brw
))
2198 brw
->perfquery
.oa_metrics_table
=
2199 _mesa_hash_table_create(NULL
, _mesa_key_hash_string
,
2200 _mesa_key_string_equal
);
2202 /* Index all the metric sets mesa knows about before looking to see what
2203 * the kernel is advertising.
2207 if (likely((INTEL_DEBUG
& DEBUG_NO_OACONFIG
) == 0) &&
2208 kernel_has_dynamic_config_support(brw
))
2209 init_oa_configs(brw
);
2211 enumerate_sysfs_metrics(brw
);
2213 brw_perf_query_register_mdapi_oa_query(brw
);
2216 brw
->perfquery
.unaccumulated
=
2217 ralloc_array(brw
, struct brw_perf_query_object
*, 2);
2218 brw
->perfquery
.unaccumulated_elements
= 0;
2219 brw
->perfquery
.unaccumulated_array_size
= 2;
2221 exec_list_make_empty(&brw
->perfquery
.sample_buffers
);
2222 exec_list_make_empty(&brw
->perfquery
.free_sample_buffers
);
2224 /* It's convenient to guarantee that this linked list of sample
2225 * buffers is never empty so we add an empty head so when we
2226 * Begin an OA query we can always take a reference on a buffer
2229 struct brw_oa_sample_buf
*buf
= get_free_sample_buf(brw
);
2230 exec_list_push_head(&brw
->perfquery
.sample_buffers
, &buf
->link
);
2232 brw
->perfquery
.oa_stream_fd
= -1;
2234 brw
->perfquery
.next_query_start_report_id
= 1000;
2236 return brw
->perfquery
.n_queries
;
2240 brw_init_performance_queries(struct brw_context
*brw
)
2242 struct gl_context
*ctx
= &brw
->ctx
;
2244 ctx
->Driver
.InitPerfQueryInfo
= brw_init_perf_query_info
;
2245 ctx
->Driver
.GetPerfQueryInfo
= brw_get_perf_query_info
;
2246 ctx
->Driver
.GetPerfCounterInfo
= brw_get_perf_counter_info
;
2247 ctx
->Driver
.NewPerfQueryObject
= brw_new_perf_query_object
;
2248 ctx
->Driver
.DeletePerfQuery
= brw_delete_perf_query
;
2249 ctx
->Driver
.BeginPerfQuery
= brw_begin_perf_query
;
2250 ctx
->Driver
.EndPerfQuery
= brw_end_perf_query
;
2251 ctx
->Driver
.WaitPerfQuery
= brw_wait_perf_query
;
2252 ctx
->Driver
.IsPerfQueryReady
= brw_is_perf_query_ready
;
2253 ctx
->Driver
.GetPerfQueryData
= brw_get_perf_query_data
;