2 * Copyright © 2018 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 DEALINGS
26 #include <sys/types.h>
32 #include <drm-uapi/i915_drm.h>
34 #include "common/gen_gem.h"
36 #include "perf/gen_perf_mdapi.h"
37 #include "perf/gen_perf_metrics.h"
39 #include "dev/gen_debug.h"
40 #include "dev/gen_device_info.h"
41 #include "util/bitscan.h"
42 #include "util/u_math.h"
44 #define FILE_DEBUG_FLAG DEBUG_PERFMON
45 #define MI_RPC_BO_SIZE 4096
46 #define MI_FREQ_START_OFFSET_BYTES (3072)
47 #define MI_RPC_BO_END_OFFSET_BYTES (MI_RPC_BO_SIZE / 2)
48 #define MI_FREQ_END_OFFSET_BYTES (3076)
50 #define INTEL_MASK(high, low) (((1u<<((high)-(low)+1))-1)<<(low))
52 #define GEN7_RPSTAT1 0xA01C
53 #define GEN7_RPSTAT1_CURR_GT_FREQ_SHIFT 7
54 #define GEN7_RPSTAT1_CURR_GT_FREQ_MASK INTEL_MASK(13, 7)
55 #define GEN7_RPSTAT1_PREV_GT_FREQ_SHIFT 0
56 #define GEN7_RPSTAT1_PREV_GT_FREQ_MASK INTEL_MASK(6, 0)
58 #define GEN9_RPSTAT0 0xA01C
59 #define GEN9_RPSTAT0_CURR_GT_FREQ_SHIFT 23
60 #define GEN9_RPSTAT0_CURR_GT_FREQ_MASK INTEL_MASK(31, 23)
61 #define GEN9_RPSTAT0_PREV_GT_FREQ_SHIFT 0
62 #define GEN9_RPSTAT0_PREV_GT_FREQ_MASK INTEL_MASK(8, 0)
64 #define GEN6_SO_PRIM_STORAGE_NEEDED 0x2280
65 #define GEN7_SO_PRIM_STORAGE_NEEDED(n) (0x5240 + (n) * 8)
66 #define GEN6_SO_NUM_PRIMS_WRITTEN 0x2288
67 #define GEN7_SO_NUM_PRIMS_WRITTEN(n) (0x5200 + (n) * 8)
69 #define MAP_READ (1 << 0)
70 #define MAP_WRITE (1 << 1)
73 * Periodic OA samples are read() into these buffer structures via the
74 * i915 perf kernel interface and appended to the
75 * perf_ctx->sample_buffers linked list. When we process the
76 * results of an OA metrics query we need to consider all the periodic
77 * samples between the Begin and End MI_REPORT_PERF_COUNT command
80 * 'Periodic' is a simplification as there are other automatic reports
81 * written by the hardware also buffered here.
83 * Considering three queries, A, B and C:
86 * ________________A_________________
88 * | ________B_________ _____C___________
91 * And an illustration of sample buffers read over this time frame:
92 * [HEAD ][ ][ ][ ][ ][ ][ ][ ][TAIL ]
94 * These nodes may hold samples for query A:
95 * [ ][ ][ A ][ A ][ A ][ A ][ A ][ ][ ]
97 * These nodes may hold samples for query B:
98 * [ ][ ][ B ][ B ][ B ][ ][ ][ ][ ]
100 * These nodes may hold samples for query C:
101 * [ ][ ][ ][ ][ ][ C ][ C ][ C ][ ]
103 * The illustration assumes we have an even distribution of periodic
104 * samples so all nodes have the same size plotted against time:
106 * Note, to simplify code, the list is never empty.
108 * With overlapping queries we can see that periodic OA reports may
109 * relate to multiple queries and care needs to be take to keep
110 * track of sample buffers until there are no queries that might
111 * depend on their contents.
113 * We use a node ref counting system where a reference ensures that a
114 * node and all following nodes can't be freed/recycled until the
115 * reference drops to zero.
117 * E.g. with a ref of one here:
118 * [ 0 ][ 0 ][ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
120 * These nodes could be freed or recycled ("reaped"):
123 * These must be preserved until the leading ref drops to zero:
124 * [ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
126 * When a query starts we take a reference on the current tail of
127 * the list, knowing that no already-buffered samples can possibly
128 * relate to the newly-started query. A pointer to this node is
129 * also saved in the query object's ->oa.samples_head.
131 * E.g. starting query A while there are two nodes in .sample_buffers:
132 * ________________A________
136 * ^_______ Add a reference and store pointer to node in
139 * Moving forward to when the B query starts with no new buffer nodes:
140 * (for reference, i915 perf reads() are only done when queries finish)
141 * ________________A_______
146 * ^_______ Add a reference and store pointer to
147 * node in B->oa.samples_head
149 * Once a query is finished, after an OA query has become 'Ready',
150 * once the End OA report has landed and after we we have processed
151 * all the intermediate periodic samples then we drop the
152 * ->oa.samples_head reference we took at the start.
154 * So when the B query has finished we have:
155 * ________________A________
156 * | ______B___________
158 * [ 0 ][ 1 ][ 0 ][ 0 ][ 0 ]
159 * ^_______ Drop B->oa.samples_head reference
161 * We still can't free these due to the A->oa.samples_head ref:
162 * [ 1 ][ 0 ][ 0 ][ 0 ]
164 * When the A query finishes: (note there's a new ref for C's samples_head)
165 * ________________A_________________
169 * [ 0 ][ 0 ][ 0 ][ 0 ][ 1 ][ 0 ][ 0 ]
170 * ^_______ Drop A->oa.samples_head reference
172 * And we can now reap these nodes up to the C->oa.samples_head:
173 * [ X ][ X ][ X ][ X ]
174 * keeping -> [ 1 ][ 0 ][ 0 ]
176 * We reap old sample buffers each time we finish processing an OA
177 * query by iterating the sample_buffers list from the head until we
178 * find a referenced node and stop.
180 * Reaped buffers move to a perfquery.free_sample_buffers list and
181 * when we come to read() we first look to recycle a buffer from the
182 * free_sample_buffers list before allocating a new buffer.
184 struct oa_sample_buf
{
185 struct exec_node link
;
188 uint8_t buf
[I915_PERF_OA_SAMPLE_SIZE
* 10];
189 uint32_t last_timestamp
;
192 struct gen_perf_context
{
193 struct gen_perf_config
*perf
;
195 void * ctx
; /* driver context (eg, brw_context) */
197 const struct gen_device_info
*devinfo
;
202 /* The i915 perf stream we open to setup + enable the OA counters */
205 /* An i915 perf stream fd gives exclusive access to the OA unit that will
206 * report counter snapshots for a specific counter set/profile in a
207 * specific layout/format so we can only start OA queries that are
208 * compatible with the currently open fd...
210 int current_oa_metrics_set_id
;
211 int current_oa_format
;
213 /* List of buffers containing OA reports */
214 struct exec_list sample_buffers
;
216 /* Cached list of empty sample buffers */
217 struct exec_list free_sample_buffers
;
219 int n_active_oa_queries
;
220 int n_active_pipeline_stats_queries
;
222 /* The number of queries depending on running OA counters which
223 * extends beyond brw_end_perf_query() since we need to wait until
224 * the last MI_RPC command has parsed by the GPU.
226 * Accurate accounting is important here as emitting an
227 * MI_REPORT_PERF_COUNT command while the OA unit is disabled will
228 * effectively hang the gpu.
232 /* To help catch an spurious problem with the hardware or perf
233 * forwarding samples, we emit each MI_REPORT_PERF_COUNT command
234 * with a unique ID that we can explicitly check for...
236 int next_query_start_report_id
;
239 * An array of queries whose results haven't yet been assembled
240 * based on the data in buffer objects.
242 * These may be active, or have already ended. However, the
243 * results have not been requested.
245 struct gen_perf_query_object
**unaccumulated
;
246 int unaccumulated_elements
;
247 int unaccumulated_array_size
;
249 /* The total number of query objects so we can relinquish
250 * our exclusive access to perf if the application deletes
251 * all of its objects. (NB: We only disable perf while
252 * there are no active queries)
254 int n_query_instances
;
257 struct gen_perf_context
*
258 gen_perf_new_context(void *parent
)
260 struct gen_perf_context
*ctx
= rzalloc(parent
, struct gen_perf_context
);
262 fprintf(stderr
, "%s: failed to alloc context\n", __func__
);
266 struct gen_perf_config
*
267 gen_perf_config(struct gen_perf_context
*ctx
)
272 struct gen_perf_query_object
*
273 gen_perf_new_query(struct gen_perf_context
*perf_ctx
, unsigned query_index
)
275 const struct gen_perf_query_info
*query
=
276 &perf_ctx
->perf
->queries
[query_index
];
277 struct gen_perf_query_object
*obj
=
278 calloc(1, sizeof(struct gen_perf_query_object
));
283 obj
->queryinfo
= query
;
285 perf_ctx
->n_query_instances
++;
290 gen_perf_active_queries(struct gen_perf_context
*perf_ctx
,
291 const struct gen_perf_query_info
*query
)
293 assert(perf_ctx
->n_active_oa_queries
== 0 || perf_ctx
->n_active_pipeline_stats_queries
== 0);
295 switch (query
->kind
) {
296 case GEN_PERF_QUERY_TYPE_OA
:
297 case GEN_PERF_QUERY_TYPE_RAW
:
298 return perf_ctx
->n_active_oa_queries
;
301 case GEN_PERF_QUERY_TYPE_PIPELINE
:
302 return perf_ctx
->n_active_pipeline_stats_queries
;
306 unreachable("Unknown query type");
312 get_sysfs_dev_dir(struct gen_perf_config
*perf
, int fd
)
317 struct dirent
*drm_entry
;
320 perf
->sysfs_dev_dir
[0] = '\0';
322 if (fstat(fd
, &sb
)) {
323 DBG("Failed to stat DRM fd\n");
327 maj
= major(sb
.st_rdev
);
328 min
= minor(sb
.st_rdev
);
330 if (!S_ISCHR(sb
.st_mode
)) {
331 DBG("DRM fd is not a character device as expected\n");
335 len
= snprintf(perf
->sysfs_dev_dir
,
336 sizeof(perf
->sysfs_dev_dir
),
337 "/sys/dev/char/%d:%d/device/drm", maj
, min
);
338 if (len
< 0 || len
>= sizeof(perf
->sysfs_dev_dir
)) {
339 DBG("Failed to concatenate sysfs path to drm device\n");
343 drmdir
= opendir(perf
->sysfs_dev_dir
);
345 DBG("Failed to open %s: %m\n", perf
->sysfs_dev_dir
);
349 while ((drm_entry
= readdir(drmdir
))) {
350 if ((drm_entry
->d_type
== DT_DIR
||
351 drm_entry
->d_type
== DT_LNK
) &&
352 strncmp(drm_entry
->d_name
, "card", 4) == 0)
354 len
= snprintf(perf
->sysfs_dev_dir
,
355 sizeof(perf
->sysfs_dev_dir
),
356 "/sys/dev/char/%d:%d/device/drm/%s",
357 maj
, min
, drm_entry
->d_name
);
359 if (len
< 0 || len
>= sizeof(perf
->sysfs_dev_dir
))
368 DBG("Failed to find cardX directory under /sys/dev/char/%d:%d/device/drm\n",
375 read_file_uint64(const char *file
, uint64_t *val
)
383 while ((n
= read(fd
, buf
, sizeof (buf
) - 1)) < 0 &&
390 *val
= strtoull(buf
, NULL
, 0);
396 read_sysfs_drm_device_file_uint64(struct gen_perf_config
*perf
,
403 len
= snprintf(buf
, sizeof(buf
), "%s/%s", perf
->sysfs_dev_dir
, file
);
404 if (len
< 0 || len
>= sizeof(buf
)) {
405 DBG("Failed to concatenate sys filename to read u64 from\n");
409 return read_file_uint64(buf
, value
);
412 static inline struct gen_perf_query_info
*
413 append_query_info(struct gen_perf_config
*perf
, int max_counters
)
415 struct gen_perf_query_info
*query
;
417 perf
->queries
= reralloc(perf
, perf
->queries
,
418 struct gen_perf_query_info
,
420 query
= &perf
->queries
[perf
->n_queries
- 1];
421 memset(query
, 0, sizeof(*query
));
423 if (max_counters
> 0) {
424 query
->max_counters
= max_counters
;
426 rzalloc_array(perf
, struct gen_perf_query_counter
, max_counters
);
433 register_oa_config(struct gen_perf_config
*perf
,
434 const struct gen_perf_query_info
*query
,
437 struct gen_perf_query_info
*registred_query
= append_query_info(perf
, 0);
439 *registred_query
= *query
;
440 registred_query
->oa_metrics_set_id
= config_id
;
441 DBG("metric set registred: id = %" PRIu64
", guid = %s\n",
442 registred_query
->oa_metrics_set_id
, query
->guid
);
446 enumerate_sysfs_metrics(struct gen_perf_config
*perf
)
448 DIR *metricsdir
= NULL
;
449 struct dirent
*metric_entry
;
453 len
= snprintf(buf
, sizeof(buf
), "%s/metrics", perf
->sysfs_dev_dir
);
454 if (len
< 0 || len
>= sizeof(buf
)) {
455 DBG("Failed to concatenate path to sysfs metrics/ directory\n");
459 metricsdir
= opendir(buf
);
461 DBG("Failed to open %s: %m\n", buf
);
465 while ((metric_entry
= readdir(metricsdir
))) {
466 struct hash_entry
*entry
;
468 if ((metric_entry
->d_type
!= DT_DIR
&&
469 metric_entry
->d_type
!= DT_LNK
) ||
470 metric_entry
->d_name
[0] == '.')
473 DBG("metric set: %s\n", metric_entry
->d_name
);
474 entry
= _mesa_hash_table_search(perf
->oa_metrics_table
,
475 metric_entry
->d_name
);
479 len
= snprintf(buf
, sizeof(buf
), "%s/metrics/%s/id",
480 perf
->sysfs_dev_dir
, metric_entry
->d_name
);
481 if (len
< 0 || len
>= sizeof(buf
)) {
482 DBG("Failed to concatenate path to sysfs metric id file\n");
486 if (!read_file_uint64(buf
, &id
)) {
487 DBG("Failed to read metric set id from %s: %m", buf
);
491 register_oa_config(perf
, (const struct gen_perf_query_info
*)entry
->data
, id
);
493 DBG("metric set not known by mesa (skipping)\n");
496 closedir(metricsdir
);
500 kernel_has_dynamic_config_support(struct gen_perf_config
*perf
, int fd
)
502 uint64_t invalid_config_id
= UINT64_MAX
;
504 return gen_ioctl(fd
, DRM_IOCTL_I915_PERF_REMOVE_CONFIG
,
505 &invalid_config_id
) < 0 && errno
== ENOENT
;
509 load_metric_id(struct gen_perf_config
*perf
, const char *guid
,
512 char config_path
[280];
514 snprintf(config_path
, sizeof(config_path
), "%s/metrics/%s/id",
515 perf
->sysfs_dev_dir
, guid
);
517 /* Don't recreate already loaded configs. */
518 return read_file_uint64(config_path
, metric_id
);
522 init_oa_configs(struct gen_perf_config
*perf
, int fd
)
524 hash_table_foreach(perf
->oa_metrics_table
, entry
) {
525 const struct gen_perf_query_info
*query
= entry
->data
;
526 struct drm_i915_perf_oa_config config
;
530 if (load_metric_id(perf
, query
->guid
, &config_id
)) {
531 DBG("metric set: %s (already loaded)\n", query
->guid
);
532 register_oa_config(perf
, query
, config_id
);
536 memset(&config
, 0, sizeof(config
));
538 memcpy(config
.uuid
, query
->guid
, sizeof(config
.uuid
));
540 config
.n_mux_regs
= query
->n_mux_regs
;
541 config
.mux_regs_ptr
= (uintptr_t) query
->mux_regs
;
543 config
.n_boolean_regs
= query
->n_b_counter_regs
;
544 config
.boolean_regs_ptr
= (uintptr_t) query
->b_counter_regs
;
546 config
.n_flex_regs
= query
->n_flex_regs
;
547 config
.flex_regs_ptr
= (uintptr_t) query
->flex_regs
;
549 ret
= gen_ioctl(fd
, DRM_IOCTL_I915_PERF_ADD_CONFIG
, &config
);
551 DBG("Failed to load \"%s\" (%s) metrics set in kernel: %s\n",
552 query
->name
, query
->guid
, strerror(errno
));
556 register_oa_config(perf
, query
, ret
);
557 DBG("metric set: %s (added)\n", query
->guid
);
562 compute_topology_builtins(struct gen_perf_config
*perf
,
563 const struct gen_device_info
*devinfo
)
565 perf
->sys_vars
.slice_mask
= devinfo
->slice_masks
;
566 perf
->sys_vars
.n_eu_slices
= devinfo
->num_slices
;
568 for (int i
= 0; i
< sizeof(devinfo
->subslice_masks
[i
]); i
++) {
569 perf
->sys_vars
.n_eu_sub_slices
+=
570 __builtin_popcount(devinfo
->subslice_masks
[i
]);
573 for (int i
= 0; i
< sizeof(devinfo
->eu_masks
); i
++)
574 perf
->sys_vars
.n_eus
+= __builtin_popcount(devinfo
->eu_masks
[i
]);
576 perf
->sys_vars
.eu_threads_count
= devinfo
->num_thread_per_eu
;
578 /* The subslice mask builtin contains bits for all slices. Prior to Gen11
579 * it had groups of 3bits for each slice, on Gen11 it's 8bits for each
582 * Ideally equations would be updated to have a slice/subslice query
585 perf
->sys_vars
.subslice_mask
= 0;
587 int bits_per_subslice
= devinfo
->gen
== 11 ? 8 : 3;
589 for (int s
= 0; s
< util_last_bit(devinfo
->slice_masks
); s
++) {
590 for (int ss
= 0; ss
< (devinfo
->subslice_slice_stride
* 8); ss
++) {
591 if (gen_device_info_subslice_available(devinfo
, s
, ss
))
592 perf
->sys_vars
.subslice_mask
|= 1ULL << (s
* bits_per_subslice
+ ss
);
598 init_oa_sys_vars(struct gen_perf_config
*perf
, const struct gen_device_info
*devinfo
)
600 uint64_t min_freq_mhz
= 0, max_freq_mhz
= 0;
602 if (!read_sysfs_drm_device_file_uint64(perf
, "gt_min_freq_mhz", &min_freq_mhz
))
605 if (!read_sysfs_drm_device_file_uint64(perf
, "gt_max_freq_mhz", &max_freq_mhz
))
608 memset(&perf
->sys_vars
, 0, sizeof(perf
->sys_vars
));
609 perf
->sys_vars
.gt_min_freq
= min_freq_mhz
* 1000000;
610 perf
->sys_vars
.gt_max_freq
= max_freq_mhz
* 1000000;
611 perf
->sys_vars
.timestamp_frequency
= devinfo
->timestamp_frequency
;
612 perf
->sys_vars
.revision
= devinfo
->revision
;
613 compute_topology_builtins(perf
, devinfo
);
618 typedef void (*perf_register_oa_queries_t
)(struct gen_perf_config
*);
620 static perf_register_oa_queries_t
621 get_register_queries_function(const struct gen_device_info
*devinfo
)
623 if (devinfo
->is_haswell
)
624 return gen_oa_register_queries_hsw
;
625 if (devinfo
->is_cherryview
)
626 return gen_oa_register_queries_chv
;
627 if (devinfo
->is_broadwell
)
628 return gen_oa_register_queries_bdw
;
629 if (devinfo
->is_broxton
)
630 return gen_oa_register_queries_bxt
;
631 if (devinfo
->is_skylake
) {
632 if (devinfo
->gt
== 2)
633 return gen_oa_register_queries_sklgt2
;
634 if (devinfo
->gt
== 3)
635 return gen_oa_register_queries_sklgt3
;
636 if (devinfo
->gt
== 4)
637 return gen_oa_register_queries_sklgt4
;
639 if (devinfo
->is_kabylake
) {
640 if (devinfo
->gt
== 2)
641 return gen_oa_register_queries_kblgt2
;
642 if (devinfo
->gt
== 3)
643 return gen_oa_register_queries_kblgt3
;
645 if (devinfo
->is_geminilake
)
646 return gen_oa_register_queries_glk
;
647 if (devinfo
->is_coffeelake
) {
648 if (devinfo
->gt
== 2)
649 return gen_oa_register_queries_cflgt2
;
650 if (devinfo
->gt
== 3)
651 return gen_oa_register_queries_cflgt3
;
653 if (devinfo
->is_cannonlake
)
654 return gen_oa_register_queries_cnl
;
655 if (devinfo
->gen
== 11)
656 return gen_oa_register_queries_icl
;
662 add_stat_reg(struct gen_perf_query_info
*query
, uint32_t reg
,
663 uint32_t numerator
, uint32_t denominator
,
664 const char *name
, const char *description
)
666 struct gen_perf_query_counter
*counter
;
668 assert(query
->n_counters
< query
->max_counters
);
670 counter
= &query
->counters
[query
->n_counters
];
671 counter
->name
= name
;
672 counter
->desc
= description
;
673 counter
->type
= GEN_PERF_COUNTER_TYPE_RAW
;
674 counter
->data_type
= GEN_PERF_COUNTER_DATA_TYPE_UINT64
;
675 counter
->offset
= sizeof(uint64_t) * query
->n_counters
;
676 counter
->pipeline_stat
.reg
= reg
;
677 counter
->pipeline_stat
.numerator
= numerator
;
678 counter
->pipeline_stat
.denominator
= denominator
;
684 add_basic_stat_reg(struct gen_perf_query_info
*query
,
685 uint32_t reg
, const char *name
)
687 add_stat_reg(query
, reg
, 1, 1, name
, name
);
691 load_pipeline_statistic_metrics(struct gen_perf_config
*perf_cfg
,
692 const struct gen_device_info
*devinfo
)
694 struct gen_perf_query_info
*query
=
695 append_query_info(perf_cfg
, MAX_STAT_COUNTERS
);
697 query
->kind
= GEN_PERF_QUERY_TYPE_PIPELINE
;
698 query
->name
= "Pipeline Statistics Registers";
700 add_basic_stat_reg(query
, IA_VERTICES_COUNT
,
701 "N vertices submitted");
702 add_basic_stat_reg(query
, IA_PRIMITIVES_COUNT
,
703 "N primitives submitted");
704 add_basic_stat_reg(query
, VS_INVOCATION_COUNT
,
705 "N vertex shader invocations");
707 if (devinfo
->gen
== 6) {
708 add_stat_reg(query
, GEN6_SO_PRIM_STORAGE_NEEDED
, 1, 1,
709 "SO_PRIM_STORAGE_NEEDED",
710 "N geometry shader stream-out primitives (total)");
711 add_stat_reg(query
, GEN6_SO_NUM_PRIMS_WRITTEN
, 1, 1,
712 "SO_NUM_PRIMS_WRITTEN",
713 "N geometry shader stream-out primitives (written)");
715 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(0), 1, 1,
716 "SO_PRIM_STORAGE_NEEDED (Stream 0)",
717 "N stream-out (stream 0) primitives (total)");
718 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(1), 1, 1,
719 "SO_PRIM_STORAGE_NEEDED (Stream 1)",
720 "N stream-out (stream 1) primitives (total)");
721 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(2), 1, 1,
722 "SO_PRIM_STORAGE_NEEDED (Stream 2)",
723 "N stream-out (stream 2) primitives (total)");
724 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(3), 1, 1,
725 "SO_PRIM_STORAGE_NEEDED (Stream 3)",
726 "N stream-out (stream 3) primitives (total)");
727 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(0), 1, 1,
728 "SO_NUM_PRIMS_WRITTEN (Stream 0)",
729 "N stream-out (stream 0) primitives (written)");
730 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(1), 1, 1,
731 "SO_NUM_PRIMS_WRITTEN (Stream 1)",
732 "N stream-out (stream 1) primitives (written)");
733 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(2), 1, 1,
734 "SO_NUM_PRIMS_WRITTEN (Stream 2)",
735 "N stream-out (stream 2) primitives (written)");
736 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(3), 1, 1,
737 "SO_NUM_PRIMS_WRITTEN (Stream 3)",
738 "N stream-out (stream 3) primitives (written)");
741 add_basic_stat_reg(query
, HS_INVOCATION_COUNT
,
742 "N TCS shader invocations");
743 add_basic_stat_reg(query
, DS_INVOCATION_COUNT
,
744 "N TES shader invocations");
746 add_basic_stat_reg(query
, GS_INVOCATION_COUNT
,
747 "N geometry shader invocations");
748 add_basic_stat_reg(query
, GS_PRIMITIVES_COUNT
,
749 "N geometry shader primitives emitted");
751 add_basic_stat_reg(query
, CL_INVOCATION_COUNT
,
752 "N primitives entering clipping");
753 add_basic_stat_reg(query
, CL_PRIMITIVES_COUNT
,
754 "N primitives leaving clipping");
756 if (devinfo
->is_haswell
|| devinfo
->gen
== 8) {
757 add_stat_reg(query
, PS_INVOCATION_COUNT
, 1, 4,
758 "N fragment shader invocations",
759 "N fragment shader invocations");
761 add_basic_stat_reg(query
, PS_INVOCATION_COUNT
,
762 "N fragment shader invocations");
765 add_basic_stat_reg(query
, PS_DEPTH_COUNT
,
766 "N z-pass fragments");
768 if (devinfo
->gen
>= 7) {
769 add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
770 "N compute shader invocations");
773 query
->data_size
= sizeof(uint64_t) * query
->n_counters
;
777 load_oa_metrics(struct gen_perf_config
*perf
, int fd
,
778 const struct gen_device_info
*devinfo
)
780 perf_register_oa_queries_t oa_register
= get_register_queries_function(devinfo
);
781 bool i915_perf_oa_available
= false;
784 /* The existence of this sysctl parameter implies the kernel supports
785 * the i915 perf interface.
787 if (stat("/proc/sys/dev/i915/perf_stream_paranoid", &sb
) == 0) {
789 /* If _paranoid == 1 then on Gen8+ we won't be able to access OA
790 * metrics unless running as root.
792 if (devinfo
->is_haswell
)
793 i915_perf_oa_available
= true;
795 uint64_t paranoid
= 1;
797 read_file_uint64("/proc/sys/dev/i915/perf_stream_paranoid", ¶noid
);
799 if (paranoid
== 0 || geteuid() == 0)
800 i915_perf_oa_available
= true;
804 if (!i915_perf_oa_available
||
806 !get_sysfs_dev_dir(perf
, fd
) ||
807 !init_oa_sys_vars(perf
, devinfo
))
810 perf
->oa_metrics_table
=
811 _mesa_hash_table_create(perf
, _mesa_key_hash_string
,
812 _mesa_key_string_equal
);
814 /* Index all the metric sets mesa knows about before looking to see what
815 * the kernel is advertising.
819 if (likely((INTEL_DEBUG
& DEBUG_NO_OACONFIG
) == 0) &&
820 kernel_has_dynamic_config_support(perf
, fd
))
821 init_oa_configs(perf
, fd
);
823 enumerate_sysfs_metrics(perf
);
828 /* Accumulate 32bits OA counters */
830 accumulate_uint32(const uint32_t *report0
,
831 const uint32_t *report1
,
832 uint64_t *accumulator
)
834 *accumulator
+= (uint32_t)(*report1
- *report0
);
837 /* Accumulate 40bits OA counters */
839 accumulate_uint40(int a_index
,
840 const uint32_t *report0
,
841 const uint32_t *report1
,
842 uint64_t *accumulator
)
844 const uint8_t *high_bytes0
= (uint8_t *)(report0
+ 40);
845 const uint8_t *high_bytes1
= (uint8_t *)(report1
+ 40);
846 uint64_t high0
= (uint64_t)(high_bytes0
[a_index
]) << 32;
847 uint64_t high1
= (uint64_t)(high_bytes1
[a_index
]) << 32;
848 uint64_t value0
= report0
[a_index
+ 4] | high0
;
849 uint64_t value1
= report1
[a_index
+ 4] | high1
;
853 delta
= (1ULL << 40) + value1
- value0
;
855 delta
= value1
- value0
;
857 *accumulator
+= delta
;
861 gen8_read_report_clock_ratios(const uint32_t *report
,
862 uint64_t *slice_freq_hz
,
863 uint64_t *unslice_freq_hz
)
865 /* The lower 16bits of the RPT_ID field of the OA reports contains a
866 * snapshot of the bits coming from the RP_FREQ_NORMAL register and is
869 * RPT_ID[31:25]: RP_FREQ_NORMAL[20:14] (low squashed_slice_clock_frequency)
870 * RPT_ID[10:9]: RP_FREQ_NORMAL[22:21] (high squashed_slice_clock_frequency)
871 * RPT_ID[8:0]: RP_FREQ_NORMAL[31:23] (squashed_unslice_clock_frequency)
873 * RP_FREQ_NORMAL[31:23]: Software Unslice Ratio Request
874 * Multiple of 33.33MHz 2xclk (16 MHz 1xclk)
876 * RP_FREQ_NORMAL[22:14]: Software Slice Ratio Request
877 * Multiple of 33.33MHz 2xclk (16 MHz 1xclk)
880 uint32_t unslice_freq
= report
[0] & 0x1ff;
881 uint32_t slice_freq_low
= (report
[0] >> 25) & 0x7f;
882 uint32_t slice_freq_high
= (report
[0] >> 9) & 0x3;
883 uint32_t slice_freq
= slice_freq_low
| (slice_freq_high
<< 7);
885 *slice_freq_hz
= slice_freq
* 16666667ULL;
886 *unslice_freq_hz
= unslice_freq
* 16666667ULL;
890 query_result_read_frequencies(struct gen_perf_query_result
*result
,
891 const struct gen_device_info
*devinfo
,
892 const uint32_t *start
,
895 /* Slice/Unslice frequency is only available in the OA reports when the
896 * "Disable OA reports due to clock ratio change" field in
897 * OA_DEBUG_REGISTER is set to 1. This is how the kernel programs this
898 * global register (see drivers/gpu/drm/i915/i915_perf.c)
900 * Documentation says this should be available on Gen9+ but experimentation
901 * shows that Gen8 reports similar values, so we enable it there too.
903 if (devinfo
->gen
< 8)
906 gen8_read_report_clock_ratios(start
,
907 &result
->slice_frequency
[0],
908 &result
->unslice_frequency
[0]);
909 gen8_read_report_clock_ratios(end
,
910 &result
->slice_frequency
[1],
911 &result
->unslice_frequency
[1]);
915 query_result_accumulate(struct gen_perf_query_result
*result
,
916 const struct gen_perf_query_info
*query
,
917 const uint32_t *start
,
922 result
->hw_id
= start
[2];
923 result
->reports_accumulated
++;
925 switch (query
->oa_format
) {
926 case I915_OA_FORMAT_A32u40_A4u32_B8_C8
:
927 accumulate_uint32(start
+ 1, end
+ 1, result
->accumulator
+ idx
++); /* timestamp */
928 accumulate_uint32(start
+ 3, end
+ 3, result
->accumulator
+ idx
++); /* clock */
930 /* 32x 40bit A counters... */
931 for (i
= 0; i
< 32; i
++)
932 accumulate_uint40(i
, start
, end
, result
->accumulator
+ idx
++);
934 /* 4x 32bit A counters... */
935 for (i
= 0; i
< 4; i
++)
936 accumulate_uint32(start
+ 36 + i
, end
+ 36 + i
, result
->accumulator
+ idx
++);
938 /* 8x 32bit B counters + 8x 32bit C counters... */
939 for (i
= 0; i
< 16; i
++)
940 accumulate_uint32(start
+ 48 + i
, end
+ 48 + i
, result
->accumulator
+ idx
++);
943 case I915_OA_FORMAT_A45_B8_C8
:
944 accumulate_uint32(start
+ 1, end
+ 1, result
->accumulator
); /* timestamp */
946 for (i
= 0; i
< 61; i
++)
947 accumulate_uint32(start
+ 3 + i
, end
+ 3 + i
, result
->accumulator
+ 1 + i
);
951 unreachable("Can't accumulate OA counters in unknown format");
957 query_result_clear(struct gen_perf_query_result
*result
)
959 memset(result
, 0, sizeof(*result
));
960 result
->hw_id
= 0xffffffff; /* invalid */
964 register_mdapi_statistic_query(struct gen_perf_config
*perf_cfg
,
965 const struct gen_device_info
*devinfo
)
967 if (!(devinfo
->gen
>= 7 && devinfo
->gen
<= 11))
970 struct gen_perf_query_info
*query
=
971 append_query_info(perf_cfg
, MAX_STAT_COUNTERS
);
973 query
->kind
= GEN_PERF_QUERY_TYPE_PIPELINE
;
974 query
->name
= "Intel_Raw_Pipeline_Statistics_Query";
976 /* The order has to match mdapi_pipeline_metrics. */
977 add_basic_stat_reg(query
, IA_VERTICES_COUNT
,
978 "N vertices submitted");
979 add_basic_stat_reg(query
, IA_PRIMITIVES_COUNT
,
980 "N primitives submitted");
981 add_basic_stat_reg(query
, VS_INVOCATION_COUNT
,
982 "N vertex shader invocations");
983 add_basic_stat_reg(query
, GS_INVOCATION_COUNT
,
984 "N geometry shader invocations");
985 add_basic_stat_reg(query
, GS_PRIMITIVES_COUNT
,
986 "N geometry shader primitives emitted");
987 add_basic_stat_reg(query
, CL_INVOCATION_COUNT
,
988 "N primitives entering clipping");
989 add_basic_stat_reg(query
, CL_PRIMITIVES_COUNT
,
990 "N primitives leaving clipping");
991 if (devinfo
->is_haswell
|| devinfo
->gen
== 8) {
992 add_stat_reg(query
, PS_INVOCATION_COUNT
, 1, 4,
993 "N fragment shader invocations",
994 "N fragment shader invocations");
996 add_basic_stat_reg(query
, PS_INVOCATION_COUNT
,
997 "N fragment shader invocations");
999 add_basic_stat_reg(query
, HS_INVOCATION_COUNT
,
1000 "N TCS shader invocations");
1001 add_basic_stat_reg(query
, DS_INVOCATION_COUNT
,
1002 "N TES shader invocations");
1003 if (devinfo
->gen
>= 7) {
1004 add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
1005 "N compute shader invocations");
1008 if (devinfo
->gen
>= 10) {
1009 /* Reuse existing CS invocation register until we can expose this new
1012 add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
1016 query
->data_size
= sizeof(uint64_t) * query
->n_counters
;
1020 fill_mdapi_perf_query_counter(struct gen_perf_query_info
*query
,
1022 uint32_t data_offset
,
1024 enum gen_perf_counter_data_type data_type
)
1026 struct gen_perf_query_counter
*counter
= &query
->counters
[query
->n_counters
];
1028 assert(query
->n_counters
<= query
->max_counters
);
1030 counter
->name
= name
;
1031 counter
->desc
= "Raw counter value";
1032 counter
->type
= GEN_PERF_COUNTER_TYPE_RAW
;
1033 counter
->data_type
= data_type
;
1034 counter
->offset
= data_offset
;
1036 query
->n_counters
++;
1038 assert(counter
->offset
+ gen_perf_query_counter_get_size(counter
) <= query
->data_size
);
1041 #define MDAPI_QUERY_ADD_COUNTER(query, struct_name, field_name, type_name) \
1042 fill_mdapi_perf_query_counter(query, #field_name, \
1043 (uint8_t *) &struct_name.field_name - \
1044 (uint8_t *) &struct_name, \
1045 sizeof(struct_name.field_name), \
1046 GEN_PERF_COUNTER_DATA_TYPE_##type_name)
1047 #define MDAPI_QUERY_ADD_ARRAY_COUNTER(ctx, query, struct_name, field_name, idx, type_name) \
1048 fill_mdapi_perf_query_counter(query, \
1049 ralloc_asprintf(ctx, "%s%i", #field_name, idx), \
1050 (uint8_t *) &struct_name.field_name[idx] - \
1051 (uint8_t *) &struct_name, \
1052 sizeof(struct_name.field_name[0]), \
1053 GEN_PERF_COUNTER_DATA_TYPE_##type_name)
1056 register_mdapi_oa_query(const struct gen_device_info
*devinfo
,
1057 struct gen_perf_config
*perf
)
1059 struct gen_perf_query_info
*query
= NULL
;
1061 /* MDAPI requires different structures for pretty much every generation
1062 * (right now we have definitions for gen 7 to 11).
1064 if (!(devinfo
->gen
>= 7 && devinfo
->gen
<= 11))
1067 switch (devinfo
->gen
) {
1069 query
= append_query_info(perf
, 1 + 45 + 16 + 7);
1070 query
->oa_format
= I915_OA_FORMAT_A45_B8_C8
;
1072 struct gen7_mdapi_metrics metric_data
;
1073 query
->data_size
= sizeof(metric_data
);
1075 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, TotalTime
, UINT64
);
1076 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.ACounters
); i
++) {
1077 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1078 metric_data
, ACounters
, i
, UINT64
);
1080 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.NOACounters
); i
++) {
1081 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1082 metric_data
, NOACounters
, i
, UINT64
);
1084 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter1
, UINT64
);
1085 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter2
, UINT64
);
1086 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SplitOccured
, BOOL32
);
1087 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequencyChanged
, BOOL32
);
1088 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequency
, UINT64
);
1089 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportId
, UINT32
);
1090 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportsCount
, UINT32
);
1094 query
= append_query_info(perf
, 2 + 36 + 16 + 16);
1095 query
->oa_format
= I915_OA_FORMAT_A32u40_A4u32_B8_C8
;
1097 struct gen8_mdapi_metrics metric_data
;
1098 query
->data_size
= sizeof(metric_data
);
1100 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, TotalTime
, UINT64
);
1101 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, GPUTicks
, UINT64
);
1102 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.OaCntr
); i
++) {
1103 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1104 metric_data
, OaCntr
, i
, UINT64
);
1106 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.NoaCntr
); i
++) {
1107 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1108 metric_data
, NoaCntr
, i
, UINT64
);
1110 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, BeginTimestamp
, UINT64
);
1111 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved1
, UINT64
);
1112 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved2
, UINT64
);
1113 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved3
, UINT32
);
1114 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, OverrunOccured
, BOOL32
);
1115 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerUser
, UINT64
);
1116 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerDriver
, UINT64
);
1117 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SliceFrequency
, UINT64
);
1118 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, UnsliceFrequency
, UINT64
);
1119 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter1
, UINT64
);
1120 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter2
, UINT64
);
1121 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SplitOccured
, BOOL32
);
1122 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequencyChanged
, BOOL32
);
1123 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequency
, UINT64
);
1124 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportId
, UINT32
);
1125 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportsCount
, UINT32
);
1131 query
= append_query_info(perf
, 2 + 36 + 16 + 16 + 16 + 2);
1132 query
->oa_format
= I915_OA_FORMAT_A32u40_A4u32_B8_C8
;
1134 struct gen9_mdapi_metrics metric_data
;
1135 query
->data_size
= sizeof(metric_data
);
1137 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, TotalTime
, UINT64
);
1138 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, GPUTicks
, UINT64
);
1139 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.OaCntr
); i
++) {
1140 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1141 metric_data
, OaCntr
, i
, UINT64
);
1143 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.NoaCntr
); i
++) {
1144 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1145 metric_data
, NoaCntr
, i
, UINT64
);
1147 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, BeginTimestamp
, UINT64
);
1148 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved1
, UINT64
);
1149 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved2
, UINT64
);
1150 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved3
, UINT32
);
1151 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, OverrunOccured
, BOOL32
);
1152 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerUser
, UINT64
);
1153 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerDriver
, UINT64
);
1154 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SliceFrequency
, UINT64
);
1155 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, UnsliceFrequency
, UINT64
);
1156 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter1
, UINT64
);
1157 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter2
, UINT64
);
1158 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SplitOccured
, BOOL32
);
1159 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequencyChanged
, BOOL32
);
1160 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequency
, UINT64
);
1161 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportId
, UINT32
);
1162 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportsCount
, UINT32
);
1163 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.UserCntr
); i
++) {
1164 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1165 metric_data
, UserCntr
, i
, UINT64
);
1167 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, UserCntrCfgId
, UINT32
);
1168 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved4
, UINT32
);
1172 unreachable("Unsupported gen");
1176 query
->kind
= GEN_PERF_QUERY_TYPE_RAW
;
1177 query
->name
= "Intel_Raw_Hardware_Counters_Set_0_Query";
1178 query
->guid
= GEN_PERF_QUERY_GUID_MDAPI
;
1181 /* Accumulation buffer offsets copied from an actual query... */
1182 const struct gen_perf_query_info
*copy_query
=
1185 query
->gpu_time_offset
= copy_query
->gpu_time_offset
;
1186 query
->gpu_clock_offset
= copy_query
->gpu_clock_offset
;
1187 query
->a_offset
= copy_query
->a_offset
;
1188 query
->b_offset
= copy_query
->b_offset
;
1189 query
->c_offset
= copy_query
->c_offset
;
1194 get_metric_id(struct gen_perf_config
*perf
,
1195 const struct gen_perf_query_info
*query
)
1197 /* These queries are know not to ever change, their config ID has been
1198 * loaded upon the first query creation. No need to look them up again.
1200 if (query
->kind
== GEN_PERF_QUERY_TYPE_OA
)
1201 return query
->oa_metrics_set_id
;
1203 assert(query
->kind
== GEN_PERF_QUERY_TYPE_RAW
);
1205 /* Raw queries can be reprogrammed up by an external application/library.
1206 * When a raw query is used for the first time it's id is set to a value !=
1207 * 0. When it stops being used the id returns to 0. No need to reload the
1208 * ID when it's already loaded.
1210 if (query
->oa_metrics_set_id
!= 0) {
1211 DBG("Raw query '%s' guid=%s using cached ID: %"PRIu64
"\n",
1212 query
->name
, query
->guid
, query
->oa_metrics_set_id
);
1213 return query
->oa_metrics_set_id
;
1216 struct gen_perf_query_info
*raw_query
= (struct gen_perf_query_info
*)query
;
1217 if (!load_metric_id(perf
, query
->guid
,
1218 &raw_query
->oa_metrics_set_id
)) {
1219 DBG("Unable to read query guid=%s ID, falling back to test config\n", query
->guid
);
1220 raw_query
->oa_metrics_set_id
= 1ULL;
1222 DBG("Raw query '%s'guid=%s loaded ID: %"PRIu64
"\n",
1223 query
->name
, query
->guid
, query
->oa_metrics_set_id
);
1225 return query
->oa_metrics_set_id
;
1228 static struct oa_sample_buf
*
1229 get_free_sample_buf(struct gen_perf_context
*perf_ctx
)
1231 struct exec_node
*node
= exec_list_pop_head(&perf_ctx
->free_sample_buffers
);
1232 struct oa_sample_buf
*buf
;
1235 buf
= exec_node_data(struct oa_sample_buf
, node
, link
);
1237 buf
= ralloc_size(perf_ctx
->perf
, sizeof(*buf
));
1239 exec_node_init(&buf
->link
);
1248 reap_old_sample_buffers(struct gen_perf_context
*perf_ctx
)
1250 struct exec_node
*tail_node
=
1251 exec_list_get_tail(&perf_ctx
->sample_buffers
);
1252 struct oa_sample_buf
*tail_buf
=
1253 exec_node_data(struct oa_sample_buf
, tail_node
, link
);
1255 /* Remove all old, unreferenced sample buffers walking forward from
1256 * the head of the list, except always leave at least one node in
1257 * the list so we always have a node to reference when we Begin
1260 foreach_list_typed_safe(struct oa_sample_buf
, buf
, link
,
1261 &perf_ctx
->sample_buffers
)
1263 if (buf
->refcount
== 0 && buf
!= tail_buf
) {
1264 exec_node_remove(&buf
->link
);
1265 exec_list_push_head(&perf_ctx
->free_sample_buffers
, &buf
->link
);
1272 free_sample_bufs(struct gen_perf_context
*perf_ctx
)
1274 foreach_list_typed_safe(struct oa_sample_buf
, buf
, link
,
1275 &perf_ctx
->free_sample_buffers
)
1278 exec_list_make_empty(&perf_ctx
->free_sample_buffers
);
1281 /******************************************************************************/
1284 * Emit MI_STORE_REGISTER_MEM commands to capture all of the
1285 * pipeline statistics for the performance query object.
1288 snapshot_statistics_registers(void *context
,
1289 struct gen_perf_config
*perf
,
1290 struct gen_perf_query_object
*obj
,
1291 uint32_t offset_in_bytes
)
1293 const struct gen_perf_query_info
*query
= obj
->queryinfo
;
1294 const int n_counters
= query
->n_counters
;
1296 for (int i
= 0; i
< n_counters
; i
++) {
1297 const struct gen_perf_query_counter
*counter
= &query
->counters
[i
];
1299 assert(counter
->data_type
== GEN_PERF_COUNTER_DATA_TYPE_UINT64
);
1301 perf
->vtbl
.store_register_mem64(context
, obj
->pipeline_stats
.bo
,
1302 counter
->pipeline_stat
.reg
,
1303 offset_in_bytes
+ i
* sizeof(uint64_t));
1308 gen_perf_close(struct gen_perf_context
*perfquery
,
1309 const struct gen_perf_query_info
*query
)
1311 if (perfquery
->oa_stream_fd
!= -1) {
1312 close(perfquery
->oa_stream_fd
);
1313 perfquery
->oa_stream_fd
= -1;
1315 if (query
->kind
== GEN_PERF_QUERY_TYPE_RAW
) {
1316 struct gen_perf_query_info
*raw_query
=
1317 (struct gen_perf_query_info
*) query
;
1318 raw_query
->oa_metrics_set_id
= 0;
1323 gen_perf_open(struct gen_perf_context
*perf_ctx
,
1326 int period_exponent
,
1330 uint64_t properties
[] = {
1331 /* Single context sampling */
1332 DRM_I915_PERF_PROP_CTX_HANDLE
, ctx_id
,
1334 /* Include OA reports in samples */
1335 DRM_I915_PERF_PROP_SAMPLE_OA
, true,
1337 /* OA unit configuration */
1338 DRM_I915_PERF_PROP_OA_METRICS_SET
, metrics_set_id
,
1339 DRM_I915_PERF_PROP_OA_FORMAT
, report_format
,
1340 DRM_I915_PERF_PROP_OA_EXPONENT
, period_exponent
,
1342 struct drm_i915_perf_open_param param
= {
1343 .flags
= I915_PERF_FLAG_FD_CLOEXEC
|
1344 I915_PERF_FLAG_FD_NONBLOCK
|
1345 I915_PERF_FLAG_DISABLED
,
1346 .num_properties
= ARRAY_SIZE(properties
) / 2,
1347 .properties_ptr
= (uintptr_t) properties
,
1349 int fd
= gen_ioctl(drm_fd
, DRM_IOCTL_I915_PERF_OPEN
, ¶m
);
1351 DBG("Error opening gen perf OA stream: %m\n");
1355 perf_ctx
->oa_stream_fd
= fd
;
1357 perf_ctx
->current_oa_metrics_set_id
= metrics_set_id
;
1358 perf_ctx
->current_oa_format
= report_format
;
1364 inc_n_users(struct gen_perf_context
*perf_ctx
)
1366 if (perf_ctx
->n_oa_users
== 0 &&
1367 gen_ioctl(perf_ctx
->oa_stream_fd
, I915_PERF_IOCTL_ENABLE
, 0) < 0)
1371 ++perf_ctx
->n_oa_users
;
1377 dec_n_users(struct gen_perf_context
*perf_ctx
)
1379 /* Disabling the i915 perf stream will effectively disable the OA
1380 * counters. Note it's important to be sure there are no outstanding
1381 * MI_RPC commands at this point since they could stall the CS
1382 * indefinitely once OACONTROL is disabled.
1384 --perf_ctx
->n_oa_users
;
1385 if (perf_ctx
->n_oa_users
== 0 &&
1386 gen_ioctl(perf_ctx
->oa_stream_fd
, I915_PERF_IOCTL_DISABLE
, 0) < 0)
1388 DBG("WARNING: Error disabling gen perf stream: %m\n");
1393 gen_perf_init_metrics(struct gen_perf_config
*perf_cfg
,
1394 const struct gen_device_info
*devinfo
,
1397 load_pipeline_statistic_metrics(perf_cfg
, devinfo
);
1398 register_mdapi_statistic_query(perf_cfg
, devinfo
);
1399 if (load_oa_metrics(perf_cfg
, drm_fd
, devinfo
))
1400 register_mdapi_oa_query(devinfo
, perf_cfg
);
1404 gen_perf_init_context(struct gen_perf_context
*perf_ctx
,
1405 struct gen_perf_config
*perf_cfg
,
1406 void * ctx
, /* driver context (eg, brw_context) */
1407 void * bufmgr
, /* eg brw_bufmgr */
1408 const struct gen_device_info
*devinfo
,
1412 perf_ctx
->perf
= perf_cfg
;
1413 perf_ctx
->ctx
= ctx
;
1414 perf_ctx
->bufmgr
= bufmgr
;
1415 perf_ctx
->drm_fd
= drm_fd
;
1416 perf_ctx
->hw_ctx
= hw_ctx
;
1417 perf_ctx
->devinfo
= devinfo
;
1419 perf_ctx
->unaccumulated
=
1420 ralloc_array(ctx
, struct gen_perf_query_object
*, 2);
1421 perf_ctx
->unaccumulated_elements
= 0;
1422 perf_ctx
->unaccumulated_array_size
= 2;
1424 exec_list_make_empty(&perf_ctx
->sample_buffers
);
1425 exec_list_make_empty(&perf_ctx
->free_sample_buffers
);
1427 /* It's convenient to guarantee that this linked list of sample
1428 * buffers is never empty so we add an empty head so when we
1429 * Begin an OA query we can always take a reference on a buffer
1432 struct oa_sample_buf
*buf
= get_free_sample_buf(perf_ctx
);
1433 exec_list_push_head(&perf_ctx
->sample_buffers
, &buf
->link
);
1435 perf_ctx
->oa_stream_fd
= -1;
1436 perf_ctx
->next_query_start_report_id
= 1000;
1440 * Add a query to the global list of "unaccumulated queries."
1442 * Queries are tracked here until all the associated OA reports have
1443 * been accumulated via accumulate_oa_reports() after the end
1444 * MI_REPORT_PERF_COUNT has landed in query->oa.bo.
1447 add_to_unaccumulated_query_list(struct gen_perf_context
*perf_ctx
,
1448 struct gen_perf_query_object
*obj
)
1450 if (perf_ctx
->unaccumulated_elements
>=
1451 perf_ctx
->unaccumulated_array_size
)
1453 perf_ctx
->unaccumulated_array_size
*= 1.5;
1454 perf_ctx
->unaccumulated
=
1455 reralloc(perf_ctx
->ctx
, perf_ctx
->unaccumulated
,
1456 struct gen_perf_query_object
*,
1457 perf_ctx
->unaccumulated_array_size
);
1460 perf_ctx
->unaccumulated
[perf_ctx
->unaccumulated_elements
++] = obj
;
1464 gen_perf_begin_query(struct gen_perf_context
*perf_ctx
,
1465 struct gen_perf_query_object
*query
)
1467 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1468 const struct gen_perf_query_info
*queryinfo
= query
->queryinfo
;
1470 /* XXX: We have to consider that the command parser unit that parses batch
1471 * buffer commands and is used to capture begin/end counter snapshots isn't
1472 * implicitly synchronized with what's currently running across other GPU
1473 * units (such as the EUs running shaders) that the performance counters are
1476 * The intention of performance queries is to measure the work associated
1477 * with commands between the begin/end delimiters and so for that to be the
1478 * case we need to explicitly synchronize the parsing of commands to capture
1479 * Begin/End counter snapshots with what's running across other parts of the
1482 * When the command parser reaches a Begin marker it effectively needs to
1483 * drain everything currently running on the GPU until the hardware is idle
1484 * before capturing the first snapshot of counters - otherwise the results
1485 * would also be measuring the effects of earlier commands.
1487 * When the command parser reaches an End marker it needs to stall until
1488 * everything currently running on the GPU has finished before capturing the
1489 * end snapshot - otherwise the results won't be a complete representation
1492 * Theoretically there could be opportunities to minimize how much of the
1493 * GPU pipeline is drained, or that we stall for, when we know what specific
1494 * units the performance counters being queried relate to but we don't
1495 * currently attempt to be clever here.
1497 * Note: with our current simple approach here then for back-to-back queries
1498 * we will redundantly emit duplicate commands to synchronize the command
1499 * streamer with the rest of the GPU pipeline, but we assume that in HW the
1500 * second synchronization is effectively a NOOP.
1502 * N.B. The final results are based on deltas of counters between (inside)
1503 * Begin/End markers so even though the total wall clock time of the
1504 * workload is stretched by larger pipeline bubbles the bubbles themselves
1505 * are generally invisible to the query results. Whether that's a good or a
1506 * bad thing depends on the use case. For a lower real-time impact while
1507 * capturing metrics then periodic sampling may be a better choice than
1508 * INTEL_performance_query.
1511 * This is our Begin synchronization point to drain current work on the
1512 * GPU before we capture our first counter snapshot...
1514 perf_cfg
->vtbl
.emit_mi_flush(perf_ctx
->ctx
);
1516 switch (queryinfo
->kind
) {
1517 case GEN_PERF_QUERY_TYPE_OA
:
1518 case GEN_PERF_QUERY_TYPE_RAW
: {
1520 /* Opening an i915 perf stream implies exclusive access to the OA unit
1521 * which will generate counter reports for a specific counter set with a
1522 * specific layout/format so we can't begin any OA based queries that
1523 * require a different counter set or format unless we get an opportunity
1524 * to close the stream and open a new one...
1526 uint64_t metric_id
= get_metric_id(perf_ctx
->perf
, queryinfo
);
1528 if (perf_ctx
->oa_stream_fd
!= -1 &&
1529 perf_ctx
->current_oa_metrics_set_id
!= metric_id
) {
1531 if (perf_ctx
->n_oa_users
!= 0) {
1532 DBG("WARNING: Begin failed already using perf config=%i/%"PRIu64
"\n",
1533 perf_ctx
->current_oa_metrics_set_id
, metric_id
);
1536 gen_perf_close(perf_ctx
, queryinfo
);
1539 /* If the OA counters aren't already on, enable them. */
1540 if (perf_ctx
->oa_stream_fd
== -1) {
1541 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
1543 /* The period_exponent gives a sampling period as follows:
1544 * sample_period = timestamp_period * 2^(period_exponent + 1)
1546 * The timestamps increments every 80ns (HSW), ~52ns (GEN9LP) or
1549 * The counter overflow period is derived from the EuActive counter
1550 * which reads a counter that increments by the number of clock
1551 * cycles multiplied by the number of EUs. It can be calculated as:
1553 * 2^(number of bits in A counter) / (n_eus * max_gen_freq * 2)
1555 * (E.g. 40 EUs @ 1GHz = ~53ms)
1557 * We select a sampling period inferior to that overflow period to
1558 * ensure we cannot see more than 1 counter overflow, otherwise we
1559 * could loose information.
1562 int a_counter_in_bits
= 32;
1563 if (devinfo
->gen
>= 8)
1564 a_counter_in_bits
= 40;
1566 uint64_t overflow_period
= pow(2, a_counter_in_bits
) / (perf_cfg
->sys_vars
.n_eus
*
1567 /* drop 1GHz freq to have units in nanoseconds */
1570 DBG("A counter overflow period: %"PRIu64
"ns, %"PRIu64
"ms (n_eus=%"PRIu64
")\n",
1571 overflow_period
, overflow_period
/ 1000000ul, perf_cfg
->sys_vars
.n_eus
);
1573 int period_exponent
= 0;
1574 uint64_t prev_sample_period
, next_sample_period
;
1575 for (int e
= 0; e
< 30; e
++) {
1576 prev_sample_period
= 1000000000ull * pow(2, e
+ 1) / devinfo
->timestamp_frequency
;
1577 next_sample_period
= 1000000000ull * pow(2, e
+ 2) / devinfo
->timestamp_frequency
;
1579 /* Take the previous sampling period, lower than the overflow
1582 if (prev_sample_period
< overflow_period
&&
1583 next_sample_period
> overflow_period
)
1584 period_exponent
= e
+ 1;
1587 if (period_exponent
== 0) {
1588 DBG("WARNING: enable to find a sampling exponent\n");
1592 DBG("OA sampling exponent: %i ~= %"PRIu64
"ms\n", period_exponent
,
1593 prev_sample_period
/ 1000000ul);
1595 if (!gen_perf_open(perf_ctx
, metric_id
, queryinfo
->oa_format
,
1596 period_exponent
, perf_ctx
->drm_fd
,
1600 assert(perf_ctx
->current_oa_metrics_set_id
== metric_id
&&
1601 perf_ctx
->current_oa_format
== queryinfo
->oa_format
);
1604 if (!inc_n_users(perf_ctx
)) {
1605 DBG("WARNING: Error enabling i915 perf stream: %m\n");
1610 perf_cfg
->vtbl
.bo_unreference(query
->oa
.bo
);
1611 query
->oa
.bo
= NULL
;
1614 query
->oa
.bo
= perf_cfg
->vtbl
.bo_alloc(perf_ctx
->bufmgr
,
1615 "perf. query OA MI_RPC bo",
1618 /* Pre-filling the BO helps debug whether writes landed. */
1619 void *map
= perf_cfg
->vtbl
.bo_map(perf_ctx
->ctx
, query
->oa
.bo
, MAP_WRITE
);
1620 memset(map
, 0x80, MI_RPC_BO_SIZE
);
1621 perf_cfg
->vtbl
.bo_unmap(query
->oa
.bo
);
1624 query
->oa
.begin_report_id
= perf_ctx
->next_query_start_report_id
;
1625 perf_ctx
->next_query_start_report_id
+= 2;
1627 /* We flush the batchbuffer here to minimize the chances that MI_RPC
1628 * delimiting commands end up in different batchbuffers. If that's the
1629 * case, the measurement will include the time it takes for the kernel
1630 * scheduler to load a new request into the hardware. This is manifested in
1631 * tools like frameretrace by spikes in the "GPU Core Clocks" counter.
1633 perf_cfg
->vtbl
.batchbuffer_flush(perf_ctx
->ctx
, __FILE__
, __LINE__
);
1635 /* Take a starting OA counter snapshot. */
1636 perf_cfg
->vtbl
.emit_mi_report_perf_count(perf_ctx
->ctx
, query
->oa
.bo
, 0,
1637 query
->oa
.begin_report_id
);
1638 perf_cfg
->vtbl
.capture_frequency_stat_register(perf_ctx
->ctx
, query
->oa
.bo
,
1639 MI_FREQ_START_OFFSET_BYTES
);
1641 ++perf_ctx
->n_active_oa_queries
;
1643 /* No already-buffered samples can possibly be associated with this query
1644 * so create a marker within the list of sample buffers enabling us to
1645 * easily ignore earlier samples when processing this query after
1648 assert(!exec_list_is_empty(&perf_ctx
->sample_buffers
));
1649 query
->oa
.samples_head
= exec_list_get_tail(&perf_ctx
->sample_buffers
);
1651 struct oa_sample_buf
*buf
=
1652 exec_node_data(struct oa_sample_buf
, query
->oa
.samples_head
, link
);
1654 /* This reference will ensure that future/following sample
1655 * buffers (that may relate to this query) can't be freed until
1656 * this drops to zero.
1660 query_result_clear(&query
->oa
.result
);
1661 query
->oa
.results_accumulated
= false;
1663 add_to_unaccumulated_query_list(perf_ctx
, query
);
1667 case GEN_PERF_QUERY_TYPE_PIPELINE
:
1668 if (query
->pipeline_stats
.bo
) {
1669 perf_cfg
->vtbl
.bo_unreference(query
->pipeline_stats
.bo
);
1670 query
->pipeline_stats
.bo
= NULL
;
1673 query
->pipeline_stats
.bo
=
1674 perf_cfg
->vtbl
.bo_alloc(perf_ctx
->bufmgr
,
1675 "perf. query pipeline stats bo",
1678 /* Take starting snapshots. */
1679 snapshot_statistics_registers(perf_ctx
->ctx
, perf_cfg
, query
, 0);
1681 ++perf_ctx
->n_active_pipeline_stats_queries
;
1685 unreachable("Unknown query type");
1693 gen_perf_end_query(struct gen_perf_context
*perf_ctx
,
1694 struct gen_perf_query_object
*query
)
1696 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1698 /* Ensure that the work associated with the queried commands will have
1699 * finished before taking our query end counter readings.
1701 * For more details see comment in brw_begin_perf_query for
1702 * corresponding flush.
1704 perf_cfg
->vtbl
.emit_mi_flush(perf_ctx
->ctx
);
1706 switch (query
->queryinfo
->kind
) {
1707 case GEN_PERF_QUERY_TYPE_OA
:
1708 case GEN_PERF_QUERY_TYPE_RAW
:
1710 /* NB: It's possible that the query will have already been marked
1711 * as 'accumulated' if an error was seen while reading samples
1712 * from perf. In this case we mustn't try and emit a closing
1713 * MI_RPC command in case the OA unit has already been disabled
1715 if (!query
->oa
.results_accumulated
) {
1716 /* Take an ending OA counter snapshot. */
1717 perf_cfg
->vtbl
.capture_frequency_stat_register(perf_ctx
->ctx
, query
->oa
.bo
,
1718 MI_FREQ_END_OFFSET_BYTES
);
1719 perf_cfg
->vtbl
.emit_mi_report_perf_count(perf_ctx
->ctx
, query
->oa
.bo
,
1720 MI_RPC_BO_END_OFFSET_BYTES
,
1721 query
->oa
.begin_report_id
+ 1);
1724 --perf_ctx
->n_active_oa_queries
;
1726 /* NB: even though the query has now ended, it can't be accumulated
1727 * until the end MI_REPORT_PERF_COUNT snapshot has been written
1732 case GEN_PERF_QUERY_TYPE_PIPELINE
:
1733 snapshot_statistics_registers(perf_ctx
->ctx
, perf_cfg
, query
,
1734 STATS_BO_END_OFFSET_BYTES
);
1735 --perf_ctx
->n_active_pipeline_stats_queries
;
1739 unreachable("Unknown query type");
1745 OA_READ_STATUS_ERROR
,
1746 OA_READ_STATUS_UNFINISHED
,
1747 OA_READ_STATUS_FINISHED
,
1750 static enum OaReadStatus
1751 read_oa_samples_until(struct gen_perf_context
*perf_ctx
,
1752 uint32_t start_timestamp
,
1753 uint32_t end_timestamp
)
1755 struct exec_node
*tail_node
=
1756 exec_list_get_tail(&perf_ctx
->sample_buffers
);
1757 struct oa_sample_buf
*tail_buf
=
1758 exec_node_data(struct oa_sample_buf
, tail_node
, link
);
1759 uint32_t last_timestamp
= tail_buf
->last_timestamp
;
1762 struct oa_sample_buf
*buf
= get_free_sample_buf(perf_ctx
);
1766 while ((len
= read(perf_ctx
->oa_stream_fd
, buf
->buf
,
1767 sizeof(buf
->buf
))) < 0 && errno
== EINTR
)
1771 exec_list_push_tail(&perf_ctx
->free_sample_buffers
, &buf
->link
);
1774 if (errno
== EAGAIN
)
1775 return ((last_timestamp
- start_timestamp
) >=
1776 (end_timestamp
- start_timestamp
)) ?
1777 OA_READ_STATUS_FINISHED
:
1778 OA_READ_STATUS_UNFINISHED
;
1780 DBG("Error reading i915 perf samples: %m\n");
1783 DBG("Spurious EOF reading i915 perf samples\n");
1785 return OA_READ_STATUS_ERROR
;
1789 exec_list_push_tail(&perf_ctx
->sample_buffers
, &buf
->link
);
1791 /* Go through the reports and update the last timestamp. */
1793 while (offset
< buf
->len
) {
1794 const struct drm_i915_perf_record_header
*header
=
1795 (const struct drm_i915_perf_record_header
*) &buf
->buf
[offset
];
1796 uint32_t *report
= (uint32_t *) (header
+ 1);
1798 if (header
->type
== DRM_I915_PERF_RECORD_SAMPLE
)
1799 last_timestamp
= report
[1];
1801 offset
+= header
->size
;
1804 buf
->last_timestamp
= last_timestamp
;
1807 unreachable("not reached");
1808 return OA_READ_STATUS_ERROR
;
1812 * Try to read all the reports until either the delimiting timestamp
1813 * or an error arises.
1816 read_oa_samples_for_query(struct gen_perf_context
*perf_ctx
,
1817 struct gen_perf_query_object
*query
,
1818 void *current_batch
)
1823 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1825 /* We need the MI_REPORT_PERF_COUNT to land before we can start
1827 assert(!perf_cfg
->vtbl
.batch_references(current_batch
, query
->oa
.bo
) &&
1828 !perf_cfg
->vtbl
.bo_busy(query
->oa
.bo
));
1830 /* Map the BO once here and let accumulate_oa_reports() unmap
1832 if (query
->oa
.map
== NULL
)
1833 query
->oa
.map
= perf_cfg
->vtbl
.bo_map(perf_ctx
->ctx
, query
->oa
.bo
, MAP_READ
);
1835 start
= last
= query
->oa
.map
;
1836 end
= query
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
1838 if (start
[0] != query
->oa
.begin_report_id
) {
1839 DBG("Spurious start report id=%"PRIu32
"\n", start
[0]);
1842 if (end
[0] != (query
->oa
.begin_report_id
+ 1)) {
1843 DBG("Spurious end report id=%"PRIu32
"\n", end
[0]);
1847 /* Read the reports until the end timestamp. */
1848 switch (read_oa_samples_until(perf_ctx
, start
[1], end
[1])) {
1849 case OA_READ_STATUS_ERROR
:
1850 /* Fallthrough and let accumulate_oa_reports() deal with the
1852 case OA_READ_STATUS_FINISHED
:
1854 case OA_READ_STATUS_UNFINISHED
:
1858 unreachable("invalid read status");
1863 gen_perf_wait_query(struct gen_perf_context
*perf_ctx
,
1864 struct gen_perf_query_object
*query
,
1865 void *current_batch
)
1867 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1868 struct brw_bo
*bo
= NULL
;
1870 switch (query
->queryinfo
->kind
) {
1871 case GEN_PERF_QUERY_TYPE_OA
:
1872 case GEN_PERF_QUERY_TYPE_RAW
:
1876 case GEN_PERF_QUERY_TYPE_PIPELINE
:
1877 bo
= query
->pipeline_stats
.bo
;
1881 unreachable("Unknown query type");
1888 /* If the current batch references our results bo then we need to
1891 if (perf_cfg
->vtbl
.batch_references(current_batch
, bo
))
1892 perf_cfg
->vtbl
.batchbuffer_flush(perf_ctx
->ctx
, __FILE__
, __LINE__
);
1894 perf_cfg
->vtbl
.bo_wait_rendering(bo
);
1896 /* Due to a race condition between the OA unit signaling report
1897 * availability and the report actually being written into memory,
1898 * we need to wait for all the reports to come in before we can
1901 if (query
->queryinfo
->kind
== GEN_PERF_QUERY_TYPE_OA
||
1902 query
->queryinfo
->kind
== GEN_PERF_QUERY_TYPE_RAW
) {
1903 while (!read_oa_samples_for_query(perf_ctx
, query
, current_batch
))
1909 gen_perf_is_query_ready(struct gen_perf_context
*perf_ctx
,
1910 struct gen_perf_query_object
*query
,
1911 void *current_batch
)
1913 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1915 switch (query
->queryinfo
->kind
) {
1916 case GEN_PERF_QUERY_TYPE_OA
:
1917 case GEN_PERF_QUERY_TYPE_RAW
:
1918 return (query
->oa
.results_accumulated
||
1920 !perf_cfg
->vtbl
.batch_references(current_batch
, query
->oa
.bo
) &&
1921 !perf_cfg
->vtbl
.bo_busy(query
->oa
.bo
) &&
1922 read_oa_samples_for_query(perf_ctx
, query
, current_batch
)));
1923 case GEN_PERF_QUERY_TYPE_PIPELINE
:
1924 return (query
->pipeline_stats
.bo
&&
1925 !perf_cfg
->vtbl
.batch_references(current_batch
, query
->pipeline_stats
.bo
) &&
1926 !perf_cfg
->vtbl
.bo_busy(query
->pipeline_stats
.bo
));
1929 unreachable("Unknown query type");
1937 * Remove a query from the global list of unaccumulated queries once
1938 * after successfully accumulating the OA reports associated with the
1939 * query in accumulate_oa_reports() or when discarding unwanted query
1943 drop_from_unaccumulated_query_list(struct gen_perf_context
*perf_ctx
,
1944 struct gen_perf_query_object
*query
)
1946 for (int i
= 0; i
< perf_ctx
->unaccumulated_elements
; i
++) {
1947 if (perf_ctx
->unaccumulated
[i
] == query
) {
1948 int last_elt
= --perf_ctx
->unaccumulated_elements
;
1951 perf_ctx
->unaccumulated
[i
] = NULL
;
1953 perf_ctx
->unaccumulated
[i
] =
1954 perf_ctx
->unaccumulated
[last_elt
];
1961 /* Drop our samples_head reference so that associated periodic
1962 * sample data buffers can potentially be reaped if they aren't
1963 * referenced by any other queries...
1966 struct oa_sample_buf
*buf
=
1967 exec_node_data(struct oa_sample_buf
, query
->oa
.samples_head
, link
);
1969 assert(buf
->refcount
> 0);
1972 query
->oa
.samples_head
= NULL
;
1974 reap_old_sample_buffers(perf_ctx
);
1977 /* In general if we see anything spurious while accumulating results,
1978 * we don't try and continue accumulating the current query, hoping
1979 * for the best, we scrap anything outstanding, and then hope for the
1980 * best with new queries.
1983 discard_all_queries(struct gen_perf_context
*perf_ctx
)
1985 while (perf_ctx
->unaccumulated_elements
) {
1986 struct gen_perf_query_object
*query
= perf_ctx
->unaccumulated
[0];
1988 query
->oa
.results_accumulated
= true;
1989 drop_from_unaccumulated_query_list(perf_ctx
, query
);
1991 dec_n_users(perf_ctx
);
1996 * Accumulate raw OA counter values based on deltas between pairs of
1999 * Accumulation starts from the first report captured via
2000 * MI_REPORT_PERF_COUNT (MI_RPC) by brw_begin_perf_query() until the
2001 * last MI_RPC report requested by brw_end_perf_query(). Between these
2002 * two reports there may also some number of periodically sampled OA
2003 * reports collected via the i915 perf interface - depending on the
2004 * duration of the query.
2006 * These periodic snapshots help to ensure we handle counter overflow
2007 * correctly by being frequent enough to ensure we don't miss multiple
2008 * overflows of a counter between snapshots. For Gen8+ the i915 perf
2009 * snapshots provide the extra context-switch reports that let us
2010 * subtract out the progress of counters associated with other
2011 * contexts running on the system.
2014 accumulate_oa_reports(struct gen_perf_context
*perf_ctx
,
2015 struct gen_perf_query_object
*query
)
2017 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
2021 struct exec_node
*first_samples_node
;
2023 int out_duration
= 0;
2025 assert(query
->oa
.map
!= NULL
);
2027 start
= last
= query
->oa
.map
;
2028 end
= query
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
2030 if (start
[0] != query
->oa
.begin_report_id
) {
2031 DBG("Spurious start report id=%"PRIu32
"\n", start
[0]);
2034 if (end
[0] != (query
->oa
.begin_report_id
+ 1)) {
2035 DBG("Spurious end report id=%"PRIu32
"\n", end
[0]);
2039 /* See if we have any periodic reports to accumulate too... */
2041 /* N.B. The oa.samples_head was set when the query began and
2042 * pointed to the tail of the perf_ctx->sample_buffers list at
2043 * the time the query started. Since the buffer existed before the
2044 * first MI_REPORT_PERF_COUNT command was emitted we therefore know
2045 * that no data in this particular node's buffer can possibly be
2046 * associated with the query - so skip ahead one...
2048 first_samples_node
= query
->oa
.samples_head
->next
;
2050 foreach_list_typed_from(struct oa_sample_buf
, buf
, link
,
2051 &perf_ctx
.sample_buffers
,
2056 while (offset
< buf
->len
) {
2057 const struct drm_i915_perf_record_header
*header
=
2058 (const struct drm_i915_perf_record_header
*)(buf
->buf
+ offset
);
2060 assert(header
->size
!= 0);
2061 assert(header
->size
<= buf
->len
);
2063 offset
+= header
->size
;
2065 switch (header
->type
) {
2066 case DRM_I915_PERF_RECORD_SAMPLE
: {
2067 uint32_t *report
= (uint32_t *)(header
+ 1);
2070 /* Ignore reports that come before the start marker.
2071 * (Note: takes care to allow overflow of 32bit timestamps)
2073 if (gen_device_info_timebase_scale(devinfo
,
2074 report
[1] - start
[1]) > 5000000000) {
2078 /* Ignore reports that come after the end marker.
2079 * (Note: takes care to allow overflow of 32bit timestamps)
2081 if (gen_device_info_timebase_scale(devinfo
,
2082 report
[1] - end
[1]) <= 5000000000) {
2086 /* For Gen8+ since the counters continue while other
2087 * contexts are running we need to discount any unrelated
2088 * deltas. The hardware automatically generates a report
2089 * on context switch which gives us a new reference point
2090 * to continuing adding deltas from.
2092 * For Haswell we can rely on the HW to stop the progress
2093 * of OA counters while any other context is acctive.
2095 if (devinfo
->gen
>= 8) {
2096 if (in_ctx
&& report
[2] != query
->oa
.result
.hw_id
) {
2097 DBG("i915 perf: Switch AWAY (observed by ID change)\n");
2100 } else if (in_ctx
== false && report
[2] == query
->oa
.result
.hw_id
) {
2101 DBG("i915 perf: Switch TO\n");
2104 /* From experimentation in IGT, we found that the OA unit
2105 * might label some report as "idle" (using an invalid
2106 * context ID), right after a report for a given context.
2107 * Deltas generated by those reports actually belong to the
2108 * previous context, even though they're not labelled as
2111 * We didn't *really* Switch AWAY in the case that we e.g.
2112 * saw a single periodic report while idle...
2114 if (out_duration
>= 1)
2116 } else if (in_ctx
) {
2117 assert(report
[2] == query
->oa
.result
.hw_id
);
2118 DBG("i915 perf: Continuation IN\n");
2120 assert(report
[2] != query
->oa
.result
.hw_id
);
2121 DBG("i915 perf: Continuation OUT\n");
2128 query_result_accumulate(&query
->oa
.result
, query
->queryinfo
,
2137 case DRM_I915_PERF_RECORD_OA_BUFFER_LOST
:
2138 DBG("i915 perf: OA error: all reports lost\n");
2140 case DRM_I915_PERF_RECORD_OA_REPORT_LOST
:
2141 DBG("i915 perf: OA report lost\n");
2149 query_result_accumulate(&query
->oa
.result
, query
->queryinfo
,
2152 query
->oa
.results_accumulated
= true;
2153 drop_from_unaccumulated_query_list(perf_ctx
, query
);
2154 dec_n_users(perf_ctx
);
2160 discard_all_queries(perf_ctx
);
2164 gen_perf_delete_query(struct gen_perf_context
*perf_ctx
,
2165 struct gen_perf_query_object
*query
)
2167 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2169 /* We can assume that the frontend waits for a query to complete
2170 * before ever calling into here, so we don't have to worry about
2171 * deleting an in-flight query object.
2173 switch (query
->queryinfo
->kind
) {
2174 case GEN_PERF_QUERY_TYPE_OA
:
2175 case GEN_PERF_QUERY_TYPE_RAW
:
2177 if (!query
->oa
.results_accumulated
) {
2178 drop_from_unaccumulated_query_list(perf_ctx
, query
);
2179 dec_n_users(perf_ctx
);
2182 perf_cfg
->vtbl
.bo_unreference(query
->oa
.bo
);
2183 query
->oa
.bo
= NULL
;
2186 query
->oa
.results_accumulated
= false;
2189 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2190 if (query
->pipeline_stats
.bo
) {
2191 perf_cfg
->vtbl
.bo_unreference(query
->pipeline_stats
.bo
);
2192 query
->pipeline_stats
.bo
= NULL
;
2197 unreachable("Unknown query type");
2201 /* As an indication that the INTEL_performance_query extension is no
2202 * longer in use, it's a good time to free our cache of sample
2203 * buffers and close any current i915-perf stream.
2205 if (--perf_ctx
->n_query_instances
== 0) {
2206 free_sample_bufs(perf_ctx
);
2207 gen_perf_close(perf_ctx
, query
->queryinfo
);
2213 #define GET_FIELD(word, field) (((word) & field ## _MASK) >> field ## _SHIFT)
2216 read_gt_frequency(struct gen_perf_context
*perf_ctx
,
2217 struct gen_perf_query_object
*obj
)
2219 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
2220 uint32_t start
= *((uint32_t *)(obj
->oa
.map
+ MI_FREQ_START_OFFSET_BYTES
)),
2221 end
= *((uint32_t *)(obj
->oa
.map
+ MI_FREQ_END_OFFSET_BYTES
));
2223 switch (devinfo
->gen
) {
2226 obj
->oa
.gt_frequency
[0] = GET_FIELD(start
, GEN7_RPSTAT1_CURR_GT_FREQ
) * 50ULL;
2227 obj
->oa
.gt_frequency
[1] = GET_FIELD(end
, GEN7_RPSTAT1_CURR_GT_FREQ
) * 50ULL;
2232 obj
->oa
.gt_frequency
[0] = GET_FIELD(start
, GEN9_RPSTAT0_CURR_GT_FREQ
) * 50ULL / 3ULL;
2233 obj
->oa
.gt_frequency
[1] = GET_FIELD(end
, GEN9_RPSTAT0_CURR_GT_FREQ
) * 50ULL / 3ULL;
2236 unreachable("unexpected gen");
2239 /* Put the numbers into Hz. */
2240 obj
->oa
.gt_frequency
[0] *= 1000000ULL;
2241 obj
->oa
.gt_frequency
[1] *= 1000000ULL;
2245 get_oa_counter_data(struct gen_perf_context
*perf_ctx
,
2246 struct gen_perf_query_object
*query
,
2250 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2251 const struct gen_perf_query_info
*queryinfo
= query
->queryinfo
;
2252 int n_counters
= queryinfo
->n_counters
;
2255 for (int i
= 0; i
< n_counters
; i
++) {
2256 const struct gen_perf_query_counter
*counter
= &queryinfo
->counters
[i
];
2257 uint64_t *out_uint64
;
2259 size_t counter_size
= gen_perf_query_counter_get_size(counter
);
2262 switch (counter
->data_type
) {
2263 case GEN_PERF_COUNTER_DATA_TYPE_UINT64
:
2264 out_uint64
= (uint64_t *)(data
+ counter
->offset
);
2266 counter
->oa_counter_read_uint64(perf_cfg
, queryinfo
,
2267 query
->oa
.result
.accumulator
);
2269 case GEN_PERF_COUNTER_DATA_TYPE_FLOAT
:
2270 out_float
= (float *)(data
+ counter
->offset
);
2272 counter
->oa_counter_read_float(perf_cfg
, queryinfo
,
2273 query
->oa
.result
.accumulator
);
2276 /* So far we aren't using uint32, double or bool32... */
2277 unreachable("unexpected counter data type");
2279 written
= counter
->offset
+ counter_size
;
2287 get_pipeline_stats_data(struct gen_perf_context
*perf_ctx
,
2288 struct gen_perf_query_object
*query
,
2293 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2294 const struct gen_perf_query_info
*queryinfo
= query
->queryinfo
;
2295 int n_counters
= queryinfo
->n_counters
;
2298 uint64_t *start
= perf_cfg
->vtbl
.bo_map(perf_ctx
->ctx
, query
->pipeline_stats
.bo
, MAP_READ
);
2299 uint64_t *end
= start
+ (STATS_BO_END_OFFSET_BYTES
/ sizeof(uint64_t));
2301 for (int i
= 0; i
< n_counters
; i
++) {
2302 const struct gen_perf_query_counter
*counter
= &queryinfo
->counters
[i
];
2303 uint64_t value
= end
[i
] - start
[i
];
2305 if (counter
->pipeline_stat
.numerator
!=
2306 counter
->pipeline_stat
.denominator
) {
2307 value
*= counter
->pipeline_stat
.numerator
;
2308 value
/= counter
->pipeline_stat
.denominator
;
2311 *((uint64_t *)p
) = value
;
2315 perf_cfg
->vtbl
.bo_unmap(query
->pipeline_stats
.bo
);
2321 gen_perf_get_query_data(struct gen_perf_context
*perf_ctx
,
2322 struct gen_perf_query_object
*query
,
2325 unsigned *bytes_written
)
2327 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2330 switch (query
->queryinfo
->kind
) {
2331 case GEN_PERF_QUERY_TYPE_OA
:
2332 case GEN_PERF_QUERY_TYPE_RAW
:
2333 if (!query
->oa
.results_accumulated
) {
2334 read_gt_frequency(perf_ctx
, query
);
2335 uint32_t *begin_report
= query
->oa
.map
;
2336 uint32_t *end_report
= query
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
2337 query_result_read_frequencies(&query
->oa
.result
,
2341 accumulate_oa_reports(perf_ctx
, query
);
2342 assert(query
->oa
.results_accumulated
);
2344 perf_cfg
->vtbl
.bo_unmap(query
->oa
.bo
);
2345 query
->oa
.map
= NULL
;
2347 if (query
->queryinfo
->kind
== GEN_PERF_QUERY_TYPE_OA
) {
2348 written
= get_oa_counter_data(perf_ctx
, query
, data_size
, (uint8_t *)data
);
2350 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
2352 written
= gen_perf_query_result_write_mdapi((uint8_t *)data
, data_size
,
2353 devinfo
, &query
->oa
.result
,
2354 query
->oa
.gt_frequency
[0],
2355 query
->oa
.gt_frequency
[1]);
2359 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2360 written
= get_pipeline_stats_data(perf_ctx
, query
, data_size
, (uint8_t *)data
);
2364 unreachable("Unknown query type");
2369 *bytes_written
= written
;
2373 gen_perf_dump_query_count(struct gen_perf_context
*perf_ctx
)
2375 DBG("Queries: (Open queries = %d, OA users = %d)\n",
2376 perf_ctx
->n_active_oa_queries
, perf_ctx
->n_oa_users
);
2380 gen_perf_dump_query(struct gen_perf_context
*ctx
,
2381 struct gen_perf_query_object
*obj
,
2382 void *current_batch
)
2384 switch (obj
->queryinfo
->kind
) {
2385 case GEN_PERF_QUERY_TYPE_OA
:
2386 case GEN_PERF_QUERY_TYPE_RAW
:
2387 DBG("BO: %-4s OA data: %-10s %-15s\n",
2388 obj
->oa
.bo
? "yes," : "no,",
2389 gen_perf_is_query_ready(ctx
, obj
, current_batch
) ? "ready," : "not ready,",
2390 obj
->oa
.results_accumulated
? "accumulated" : "not accumulated");
2392 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2394 obj
->pipeline_stats
.bo
? "yes" : "no");
2397 unreachable("Unknown query type");