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 #ifndef HAVE_DIRENT_D_TYPE
33 #include <limits.h> // PATH_MAX
36 #include <drm-uapi/i915_drm.h>
38 #include "common/gen_gem.h"
40 #include "gen_perf_regs.h"
41 #include "perf/gen_perf_mdapi.h"
42 #include "perf/gen_perf_metrics.h"
44 #include "dev/gen_debug.h"
45 #include "dev/gen_device_info.h"
46 #include "util/bitscan.h"
47 #include "util/mesa-sha1.h"
48 #include "util/u_math.h"
50 #define FILE_DEBUG_FLAG DEBUG_PERFMON
51 #define MI_RPC_BO_SIZE 4096
52 #define MI_FREQ_START_OFFSET_BYTES (3072)
53 #define MI_RPC_BO_END_OFFSET_BYTES (MI_RPC_BO_SIZE / 2)
54 #define MI_FREQ_END_OFFSET_BYTES (3076)
56 #define MAP_READ (1 << 0)
57 #define MAP_WRITE (1 << 1)
59 #define OA_REPORT_INVALID_CTX_ID (0xffffffff)
62 * Periodic OA samples are read() into these buffer structures via the
63 * i915 perf kernel interface and appended to the
64 * perf_ctx->sample_buffers linked list. When we process the
65 * results of an OA metrics query we need to consider all the periodic
66 * samples between the Begin and End MI_REPORT_PERF_COUNT command
69 * 'Periodic' is a simplification as there are other automatic reports
70 * written by the hardware also buffered here.
72 * Considering three queries, A, B and C:
75 * ________________A_________________
77 * | ________B_________ _____C___________
80 * And an illustration of sample buffers read over this time frame:
81 * [HEAD ][ ][ ][ ][ ][ ][ ][ ][TAIL ]
83 * These nodes may hold samples for query A:
84 * [ ][ ][ A ][ A ][ A ][ A ][ A ][ ][ ]
86 * These nodes may hold samples for query B:
87 * [ ][ ][ B ][ B ][ B ][ ][ ][ ][ ]
89 * These nodes may hold samples for query C:
90 * [ ][ ][ ][ ][ ][ C ][ C ][ C ][ ]
92 * The illustration assumes we have an even distribution of periodic
93 * samples so all nodes have the same size plotted against time:
95 * Note, to simplify code, the list is never empty.
97 * With overlapping queries we can see that periodic OA reports may
98 * relate to multiple queries and care needs to be take to keep
99 * track of sample buffers until there are no queries that might
100 * depend on their contents.
102 * We use a node ref counting system where a reference ensures that a
103 * node and all following nodes can't be freed/recycled until the
104 * reference drops to zero.
106 * E.g. with a ref of one here:
107 * [ 0 ][ 0 ][ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
109 * These nodes could be freed or recycled ("reaped"):
112 * These must be preserved until the leading ref drops to zero:
113 * [ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
115 * When a query starts we take a reference on the current tail of
116 * the list, knowing that no already-buffered samples can possibly
117 * relate to the newly-started query. A pointer to this node is
118 * also saved in the query object's ->oa.samples_head.
120 * E.g. starting query A while there are two nodes in .sample_buffers:
121 * ________________A________
125 * ^_______ Add a reference and store pointer to node in
128 * Moving forward to when the B query starts with no new buffer nodes:
129 * (for reference, i915 perf reads() are only done when queries finish)
130 * ________________A_______
135 * ^_______ Add a reference and store pointer to
136 * node in B->oa.samples_head
138 * Once a query is finished, after an OA query has become 'Ready',
139 * once the End OA report has landed and after we we have processed
140 * all the intermediate periodic samples then we drop the
141 * ->oa.samples_head reference we took at the start.
143 * So when the B query has finished we have:
144 * ________________A________
145 * | ______B___________
147 * [ 0 ][ 1 ][ 0 ][ 0 ][ 0 ]
148 * ^_______ Drop B->oa.samples_head reference
150 * We still can't free these due to the A->oa.samples_head ref:
151 * [ 1 ][ 0 ][ 0 ][ 0 ]
153 * When the A query finishes: (note there's a new ref for C's samples_head)
154 * ________________A_________________
158 * [ 0 ][ 0 ][ 0 ][ 0 ][ 1 ][ 0 ][ 0 ]
159 * ^_______ Drop A->oa.samples_head reference
161 * And we can now reap these nodes up to the C->oa.samples_head:
162 * [ X ][ X ][ X ][ X ]
163 * keeping -> [ 1 ][ 0 ][ 0 ]
165 * We reap old sample buffers each time we finish processing an OA
166 * query by iterating the sample_buffers list from the head until we
167 * find a referenced node and stop.
169 * Reaped buffers move to a perfquery.free_sample_buffers list and
170 * when we come to read() we first look to recycle a buffer from the
171 * free_sample_buffers list before allocating a new buffer.
173 struct oa_sample_buf
{
174 struct exec_node link
;
177 uint8_t buf
[I915_PERF_OA_SAMPLE_SIZE
* 10];
178 uint32_t last_timestamp
;
182 * gen representation of a performance query object.
184 * NB: We want to keep this structure relatively lean considering that
185 * applications may expect to allocate enough objects to be able to
186 * query around all draw calls in a frame.
188 struct gen_perf_query_object
190 const struct gen_perf_query_info
*queryinfo
;
192 /* See query->kind to know which state below is in use... */
197 * BO containing OA counter snapshots at query Begin/End time.
202 * Address of mapped of @bo
207 * The MI_REPORT_PERF_COUNT command lets us specify a unique
208 * ID that will be reflected in the resulting OA report
209 * that's written by the GPU. This is the ID we're expecting
210 * in the begin report and the the end report should be
211 * @begin_report_id + 1.
216 * Reference the head of the brw->perfquery.sample_buffers
217 * list at the time that the query started (so we only need
218 * to look at nodes after this point when looking for samples
219 * related to this query)
221 * (See struct brw_oa_sample_buf description for more details)
223 struct exec_node
*samples_head
;
226 * false while in the unaccumulated_elements list, and set to
227 * true when the final, end MI_RPC snapshot has been
230 bool results_accumulated
;
233 * Frequency of the GT at begin and end of the query.
235 uint64_t gt_frequency
[2];
238 * Accumulated OA results between begin and end of the query.
240 struct gen_perf_query_result result
;
245 * BO containing starting and ending snapshots for the
246 * statistics counters.
253 struct gen_perf_context
{
254 struct gen_perf_config
*perf
;
256 void * ctx
; /* driver context (eg, brw_context) */
258 const struct gen_device_info
*devinfo
;
263 /* The i915 perf stream we open to setup + enable the OA counters */
266 /* An i915 perf stream fd gives exclusive access to the OA unit that will
267 * report counter snapshots for a specific counter set/profile in a
268 * specific layout/format so we can only start OA queries that are
269 * compatible with the currently open fd...
271 int current_oa_metrics_set_id
;
272 int current_oa_format
;
274 /* List of buffers containing OA reports */
275 struct exec_list sample_buffers
;
277 /* Cached list of empty sample buffers */
278 struct exec_list free_sample_buffers
;
280 int n_active_oa_queries
;
281 int n_active_pipeline_stats_queries
;
283 /* The number of queries depending on running OA counters which
284 * extends beyond brw_end_perf_query() since we need to wait until
285 * the last MI_RPC command has parsed by the GPU.
287 * Accurate accounting is important here as emitting an
288 * MI_REPORT_PERF_COUNT command while the OA unit is disabled will
289 * effectively hang the gpu.
293 /* To help catch an spurious problem with the hardware or perf
294 * forwarding samples, we emit each MI_REPORT_PERF_COUNT command
295 * with a unique ID that we can explicitly check for...
297 int next_query_start_report_id
;
300 * An array of queries whose results haven't yet been assembled
301 * based on the data in buffer objects.
303 * These may be active, or have already ended. However, the
304 * results have not been requested.
306 struct gen_perf_query_object
**unaccumulated
;
307 int unaccumulated_elements
;
308 int unaccumulated_array_size
;
310 /* The total number of query objects so we can relinquish
311 * our exclusive access to perf if the application deletes
312 * all of its objects. (NB: We only disable perf while
313 * there are no active queries)
315 int n_query_instances
;
318 const struct gen_perf_query_info
*
319 gen_perf_query_info(const struct gen_perf_query_object
*query
)
321 return query
->queryinfo
;
324 struct gen_perf_context
*
325 gen_perf_new_context(void *parent
)
327 struct gen_perf_context
*ctx
= rzalloc(parent
, struct gen_perf_context
);
329 fprintf(stderr
, "%s: failed to alloc context\n", __func__
);
333 struct gen_perf_config
*
334 gen_perf_config(struct gen_perf_context
*ctx
)
339 struct gen_perf_query_object
*
340 gen_perf_new_query(struct gen_perf_context
*perf_ctx
, unsigned query_index
)
342 const struct gen_perf_query_info
*query
=
343 &perf_ctx
->perf
->queries
[query_index
];
344 struct gen_perf_query_object
*obj
=
345 calloc(1, sizeof(struct gen_perf_query_object
));
350 obj
->queryinfo
= query
;
352 perf_ctx
->n_query_instances
++;
357 gen_perf_active_queries(struct gen_perf_context
*perf_ctx
,
358 const struct gen_perf_query_info
*query
)
360 assert(perf_ctx
->n_active_oa_queries
== 0 || perf_ctx
->n_active_pipeline_stats_queries
== 0);
362 switch (query
->kind
) {
363 case GEN_PERF_QUERY_TYPE_OA
:
364 case GEN_PERF_QUERY_TYPE_RAW
:
365 return perf_ctx
->n_active_oa_queries
;
368 case GEN_PERF_QUERY_TYPE_PIPELINE
:
369 return perf_ctx
->n_active_pipeline_stats_queries
;
373 unreachable("Unknown query type");
378 static inline uint64_t to_user_pointer(void *ptr
)
380 return (uintptr_t) ptr
;
384 is_dir_or_link(const struct dirent
*entry
, const char *parent_dir
)
386 #ifdef HAVE_DIRENT_D_TYPE
387 return entry
->d_type
== DT_DIR
|| entry
->d_type
== DT_LNK
;
390 char path
[PATH_MAX
+ 1];
391 snprintf(path
, sizeof(path
), "%s/%s", parent_dir
, entry
->d_name
);
393 return S_ISDIR(st
.st_mode
) || S_ISLNK(st
.st_mode
);
398 get_sysfs_dev_dir(struct gen_perf_config
*perf
, int fd
)
403 struct dirent
*drm_entry
;
406 perf
->sysfs_dev_dir
[0] = '\0';
408 if (fstat(fd
, &sb
)) {
409 DBG("Failed to stat DRM fd\n");
413 maj
= major(sb
.st_rdev
);
414 min
= minor(sb
.st_rdev
);
416 if (!S_ISCHR(sb
.st_mode
)) {
417 DBG("DRM fd is not a character device as expected\n");
421 len
= snprintf(perf
->sysfs_dev_dir
,
422 sizeof(perf
->sysfs_dev_dir
),
423 "/sys/dev/char/%d:%d/device/drm", maj
, min
);
424 if (len
< 0 || len
>= sizeof(perf
->sysfs_dev_dir
)) {
425 DBG("Failed to concatenate sysfs path to drm device\n");
429 drmdir
= opendir(perf
->sysfs_dev_dir
);
431 DBG("Failed to open %s: %m\n", perf
->sysfs_dev_dir
);
435 while ((drm_entry
= readdir(drmdir
))) {
436 if (is_dir_or_link(drm_entry
, perf
->sysfs_dev_dir
) &&
437 strncmp(drm_entry
->d_name
, "card", 4) == 0)
439 len
= snprintf(perf
->sysfs_dev_dir
,
440 sizeof(perf
->sysfs_dev_dir
),
441 "/sys/dev/char/%d:%d/device/drm/%s",
442 maj
, min
, drm_entry
->d_name
);
444 if (len
< 0 || len
>= sizeof(perf
->sysfs_dev_dir
))
453 DBG("Failed to find cardX directory under /sys/dev/char/%d:%d/device/drm\n",
460 read_file_uint64(const char *file
, uint64_t *val
)
468 while ((n
= read(fd
, buf
, sizeof (buf
) - 1)) < 0 &&
475 *val
= strtoull(buf
, NULL
, 0);
481 read_sysfs_drm_device_file_uint64(struct gen_perf_config
*perf
,
488 len
= snprintf(buf
, sizeof(buf
), "%s/%s", perf
->sysfs_dev_dir
, file
);
489 if (len
< 0 || len
>= sizeof(buf
)) {
490 DBG("Failed to concatenate sys filename to read u64 from\n");
494 return read_file_uint64(buf
, value
);
497 static inline struct gen_perf_query_info
*
498 append_query_info(struct gen_perf_config
*perf
, int max_counters
)
500 struct gen_perf_query_info
*query
;
502 perf
->queries
= reralloc(perf
, perf
->queries
,
503 struct gen_perf_query_info
,
505 query
= &perf
->queries
[perf
->n_queries
- 1];
506 memset(query
, 0, sizeof(*query
));
508 if (max_counters
> 0) {
509 query
->max_counters
= max_counters
;
511 rzalloc_array(perf
, struct gen_perf_query_counter
, max_counters
);
518 register_oa_config(struct gen_perf_config
*perf
,
519 const struct gen_perf_query_info
*query
,
522 struct gen_perf_query_info
*registered_query
= append_query_info(perf
, 0);
524 *registered_query
= *query
;
525 registered_query
->oa_metrics_set_id
= config_id
;
526 DBG("metric set registered: id = %" PRIu64
", guid = %s\n",
527 registered_query
->oa_metrics_set_id
, query
->guid
);
531 enumerate_sysfs_metrics(struct gen_perf_config
*perf
)
533 DIR *metricsdir
= NULL
;
534 struct dirent
*metric_entry
;
538 len
= snprintf(buf
, sizeof(buf
), "%s/metrics", perf
->sysfs_dev_dir
);
539 if (len
< 0 || len
>= sizeof(buf
)) {
540 DBG("Failed to concatenate path to sysfs metrics/ directory\n");
544 metricsdir
= opendir(buf
);
546 DBG("Failed to open %s: %m\n", buf
);
550 while ((metric_entry
= readdir(metricsdir
))) {
551 struct hash_entry
*entry
;
552 if (!is_dir_or_link(metric_entry
, buf
) ||
553 metric_entry
->d_name
[0] == '.')
556 DBG("metric set: %s\n", metric_entry
->d_name
);
557 entry
= _mesa_hash_table_search(perf
->oa_metrics_table
,
558 metric_entry
->d_name
);
561 if (!gen_perf_load_metric_id(perf
, metric_entry
->d_name
, &id
)) {
562 DBG("Failed to read metric set id from %s: %m", buf
);
566 register_oa_config(perf
, (const struct gen_perf_query_info
*)entry
->data
, id
);
568 DBG("metric set not known by mesa (skipping)\n");
571 closedir(metricsdir
);
575 kernel_has_dynamic_config_support(struct gen_perf_config
*perf
, int fd
)
577 uint64_t invalid_config_id
= UINT64_MAX
;
579 return gen_ioctl(fd
, DRM_IOCTL_I915_PERF_REMOVE_CONFIG
,
580 &invalid_config_id
) < 0 && errno
== ENOENT
;
584 i915_query_items(struct gen_perf_config
*perf
, int fd
,
585 struct drm_i915_query_item
*items
, uint32_t n_items
)
587 struct drm_i915_query q
= {
588 .num_items
= n_items
,
589 .items_ptr
= to_user_pointer(items
),
591 return gen_ioctl(fd
, DRM_IOCTL_I915_QUERY
, &q
);
595 i915_query_perf_config_supported(struct gen_perf_config
*perf
, int fd
)
597 struct drm_i915_query_item item
= {
598 .query_id
= DRM_I915_QUERY_PERF_CONFIG
,
599 .flags
= DRM_I915_QUERY_PERF_CONFIG_LIST
,
602 return i915_query_items(perf
, fd
, &item
, 1) == 0 && item
.length
> 0;
606 i915_query_perf_config_data(struct gen_perf_config
*perf
,
607 int fd
, const char *guid
,
608 struct drm_i915_perf_oa_config
*config
)
611 struct drm_i915_query_perf_config query
;
612 struct drm_i915_perf_oa_config config
;
614 struct drm_i915_query_item item
= {
615 .query_id
= DRM_I915_QUERY_PERF_CONFIG
,
616 .flags
= DRM_I915_QUERY_PERF_CONFIG_DATA_FOR_UUID
,
617 .data_ptr
= to_user_pointer(&item_data
),
618 .length
= sizeof(item_data
),
621 memset(&item_data
, 0, sizeof(item_data
));
622 memcpy(item_data
.query
.uuid
, guid
, sizeof(item_data
.query
.uuid
));
623 memcpy(&item_data
.config
, config
, sizeof(item_data
.config
));
625 if (!(i915_query_items(perf
, fd
, &item
, 1) == 0 && item
.length
> 0))
628 memcpy(config
, &item_data
.config
, sizeof(item_data
.config
));
634 gen_perf_load_metric_id(struct gen_perf_config
*perf_cfg
,
638 char config_path
[280];
640 snprintf(config_path
, sizeof(config_path
), "%s/metrics/%s/id",
641 perf_cfg
->sysfs_dev_dir
, guid
);
643 /* Don't recreate already loaded configs. */
644 return read_file_uint64(config_path
, metric_id
);
648 i915_add_config(struct gen_perf_config
*perf
, int fd
,
649 const struct gen_perf_registers
*config
,
652 struct drm_i915_perf_oa_config i915_config
= { 0, };
654 memcpy(i915_config
.uuid
, guid
, sizeof(i915_config
.uuid
));
656 i915_config
.n_mux_regs
= config
->n_mux_regs
;
657 i915_config
.mux_regs_ptr
= to_user_pointer(config
->mux_regs
);
659 i915_config
.n_boolean_regs
= config
->n_b_counter_regs
;
660 i915_config
.boolean_regs_ptr
= to_user_pointer(config
->b_counter_regs
);
662 i915_config
.n_flex_regs
= config
->n_flex_regs
;
663 i915_config
.flex_regs_ptr
= to_user_pointer(config
->flex_regs
);
665 int ret
= gen_ioctl(fd
, DRM_IOCTL_I915_PERF_ADD_CONFIG
, &i915_config
);
666 return ret
> 0 ? ret
: 0;
670 init_oa_configs(struct gen_perf_config
*perf
, int fd
)
672 hash_table_foreach(perf
->oa_metrics_table
, entry
) {
673 const struct gen_perf_query_info
*query
= entry
->data
;
676 if (gen_perf_load_metric_id(perf
, query
->guid
, &config_id
)) {
677 DBG("metric set: %s (already loaded)\n", query
->guid
);
678 register_oa_config(perf
, query
, config_id
);
682 int ret
= i915_add_config(perf
, fd
, &query
->config
, query
->guid
);
684 DBG("Failed to load \"%s\" (%s) metrics set in kernel: %s\n",
685 query
->name
, query
->guid
, strerror(errno
));
689 register_oa_config(perf
, query
, ret
);
690 DBG("metric set: %s (added)\n", query
->guid
);
695 compute_topology_builtins(struct gen_perf_config
*perf
,
696 const struct gen_device_info
*devinfo
)
698 perf
->sys_vars
.slice_mask
= devinfo
->slice_masks
;
699 perf
->sys_vars
.n_eu_slices
= devinfo
->num_slices
;
701 for (int i
= 0; i
< sizeof(devinfo
->subslice_masks
[i
]); i
++) {
702 perf
->sys_vars
.n_eu_sub_slices
+=
703 __builtin_popcount(devinfo
->subslice_masks
[i
]);
706 for (int i
= 0; i
< sizeof(devinfo
->eu_masks
); i
++)
707 perf
->sys_vars
.n_eus
+= __builtin_popcount(devinfo
->eu_masks
[i
]);
709 perf
->sys_vars
.eu_threads_count
= devinfo
->num_thread_per_eu
;
711 /* The subslice mask builtin contains bits for all slices. Prior to Gen11
712 * it had groups of 3bits for each slice, on Gen11 it's 8bits for each
715 * Ideally equations would be updated to have a slice/subslice query
718 perf
->sys_vars
.subslice_mask
= 0;
720 int bits_per_subslice
= devinfo
->gen
== 11 ? 8 : 3;
722 for (int s
= 0; s
< util_last_bit(devinfo
->slice_masks
); s
++) {
723 for (int ss
= 0; ss
< (devinfo
->subslice_slice_stride
* 8); ss
++) {
724 if (gen_device_info_subslice_available(devinfo
, s
, ss
))
725 perf
->sys_vars
.subslice_mask
|= 1ULL << (s
* bits_per_subslice
+ ss
);
731 init_oa_sys_vars(struct gen_perf_config
*perf
, const struct gen_device_info
*devinfo
)
733 uint64_t min_freq_mhz
= 0, max_freq_mhz
= 0;
735 if (!read_sysfs_drm_device_file_uint64(perf
, "gt_min_freq_mhz", &min_freq_mhz
))
738 if (!read_sysfs_drm_device_file_uint64(perf
, "gt_max_freq_mhz", &max_freq_mhz
))
741 memset(&perf
->sys_vars
, 0, sizeof(perf
->sys_vars
));
742 perf
->sys_vars
.gt_min_freq
= min_freq_mhz
* 1000000;
743 perf
->sys_vars
.gt_max_freq
= max_freq_mhz
* 1000000;
744 perf
->sys_vars
.timestamp_frequency
= devinfo
->timestamp_frequency
;
745 perf
->sys_vars
.revision
= devinfo
->revision
;
746 compute_topology_builtins(perf
, devinfo
);
751 typedef void (*perf_register_oa_queries_t
)(struct gen_perf_config
*);
753 static perf_register_oa_queries_t
754 get_register_queries_function(const struct gen_device_info
*devinfo
)
756 if (devinfo
->is_haswell
)
757 return gen_oa_register_queries_hsw
;
758 if (devinfo
->is_cherryview
)
759 return gen_oa_register_queries_chv
;
760 if (devinfo
->is_broadwell
)
761 return gen_oa_register_queries_bdw
;
762 if (devinfo
->is_broxton
)
763 return gen_oa_register_queries_bxt
;
764 if (devinfo
->is_skylake
) {
765 if (devinfo
->gt
== 2)
766 return gen_oa_register_queries_sklgt2
;
767 if (devinfo
->gt
== 3)
768 return gen_oa_register_queries_sklgt3
;
769 if (devinfo
->gt
== 4)
770 return gen_oa_register_queries_sklgt4
;
772 if (devinfo
->is_kabylake
) {
773 if (devinfo
->gt
== 2)
774 return gen_oa_register_queries_kblgt2
;
775 if (devinfo
->gt
== 3)
776 return gen_oa_register_queries_kblgt3
;
778 if (devinfo
->is_geminilake
)
779 return gen_oa_register_queries_glk
;
780 if (devinfo
->is_coffeelake
) {
781 if (devinfo
->gt
== 2)
782 return gen_oa_register_queries_cflgt2
;
783 if (devinfo
->gt
== 3)
784 return gen_oa_register_queries_cflgt3
;
786 if (devinfo
->is_cannonlake
)
787 return gen_oa_register_queries_cnl
;
788 if (devinfo
->gen
== 11) {
789 if (devinfo
->is_elkhartlake
)
790 return gen_oa_register_queries_lkf
;
791 return gen_oa_register_queries_icl
;
793 if (devinfo
->gen
== 12)
794 return gen_oa_register_queries_tgl
;
800 add_stat_reg(struct gen_perf_query_info
*query
, uint32_t reg
,
801 uint32_t numerator
, uint32_t denominator
,
802 const char *name
, const char *description
)
804 struct gen_perf_query_counter
*counter
;
806 assert(query
->n_counters
< query
->max_counters
);
808 counter
= &query
->counters
[query
->n_counters
];
809 counter
->name
= name
;
810 counter
->desc
= description
;
811 counter
->type
= GEN_PERF_COUNTER_TYPE_RAW
;
812 counter
->data_type
= GEN_PERF_COUNTER_DATA_TYPE_UINT64
;
813 counter
->offset
= sizeof(uint64_t) * query
->n_counters
;
814 counter
->pipeline_stat
.reg
= reg
;
815 counter
->pipeline_stat
.numerator
= numerator
;
816 counter
->pipeline_stat
.denominator
= denominator
;
822 add_basic_stat_reg(struct gen_perf_query_info
*query
,
823 uint32_t reg
, const char *name
)
825 add_stat_reg(query
, reg
, 1, 1, name
, name
);
829 load_pipeline_statistic_metrics(struct gen_perf_config
*perf_cfg
,
830 const struct gen_device_info
*devinfo
)
832 struct gen_perf_query_info
*query
=
833 append_query_info(perf_cfg
, MAX_STAT_COUNTERS
);
835 query
->kind
= GEN_PERF_QUERY_TYPE_PIPELINE
;
836 query
->name
= "Pipeline Statistics Registers";
838 add_basic_stat_reg(query
, IA_VERTICES_COUNT
,
839 "N vertices submitted");
840 add_basic_stat_reg(query
, IA_PRIMITIVES_COUNT
,
841 "N primitives submitted");
842 add_basic_stat_reg(query
, VS_INVOCATION_COUNT
,
843 "N vertex shader invocations");
845 if (devinfo
->gen
== 6) {
846 add_stat_reg(query
, GEN6_SO_PRIM_STORAGE_NEEDED
, 1, 1,
847 "SO_PRIM_STORAGE_NEEDED",
848 "N geometry shader stream-out primitives (total)");
849 add_stat_reg(query
, GEN6_SO_NUM_PRIMS_WRITTEN
, 1, 1,
850 "SO_NUM_PRIMS_WRITTEN",
851 "N geometry shader stream-out primitives (written)");
853 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(0), 1, 1,
854 "SO_PRIM_STORAGE_NEEDED (Stream 0)",
855 "N stream-out (stream 0) primitives (total)");
856 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(1), 1, 1,
857 "SO_PRIM_STORAGE_NEEDED (Stream 1)",
858 "N stream-out (stream 1) primitives (total)");
859 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(2), 1, 1,
860 "SO_PRIM_STORAGE_NEEDED (Stream 2)",
861 "N stream-out (stream 2) primitives (total)");
862 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(3), 1, 1,
863 "SO_PRIM_STORAGE_NEEDED (Stream 3)",
864 "N stream-out (stream 3) primitives (total)");
865 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(0), 1, 1,
866 "SO_NUM_PRIMS_WRITTEN (Stream 0)",
867 "N stream-out (stream 0) primitives (written)");
868 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(1), 1, 1,
869 "SO_NUM_PRIMS_WRITTEN (Stream 1)",
870 "N stream-out (stream 1) primitives (written)");
871 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(2), 1, 1,
872 "SO_NUM_PRIMS_WRITTEN (Stream 2)",
873 "N stream-out (stream 2) primitives (written)");
874 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(3), 1, 1,
875 "SO_NUM_PRIMS_WRITTEN (Stream 3)",
876 "N stream-out (stream 3) primitives (written)");
879 add_basic_stat_reg(query
, HS_INVOCATION_COUNT
,
880 "N TCS shader invocations");
881 add_basic_stat_reg(query
, DS_INVOCATION_COUNT
,
882 "N TES shader invocations");
884 add_basic_stat_reg(query
, GS_INVOCATION_COUNT
,
885 "N geometry shader invocations");
886 add_basic_stat_reg(query
, GS_PRIMITIVES_COUNT
,
887 "N geometry shader primitives emitted");
889 add_basic_stat_reg(query
, CL_INVOCATION_COUNT
,
890 "N primitives entering clipping");
891 add_basic_stat_reg(query
, CL_PRIMITIVES_COUNT
,
892 "N primitives leaving clipping");
894 if (devinfo
->is_haswell
|| devinfo
->gen
== 8) {
895 add_stat_reg(query
, PS_INVOCATION_COUNT
, 1, 4,
896 "N fragment shader invocations",
897 "N fragment shader invocations");
899 add_basic_stat_reg(query
, PS_INVOCATION_COUNT
,
900 "N fragment shader invocations");
903 add_basic_stat_reg(query
, PS_DEPTH_COUNT
,
904 "N z-pass fragments");
906 if (devinfo
->gen
>= 7) {
907 add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
908 "N compute shader invocations");
911 query
->data_size
= sizeof(uint64_t) * query
->n_counters
;
915 load_oa_metrics(struct gen_perf_config
*perf
, int fd
,
916 const struct gen_device_info
*devinfo
)
918 perf_register_oa_queries_t oa_register
= get_register_queries_function(devinfo
);
919 bool i915_perf_oa_available
= false;
922 perf
->i915_query_supported
= i915_query_perf_config_supported(perf
, fd
);
924 /* The existence of this sysctl parameter implies the kernel supports
925 * the i915 perf interface.
927 if (stat("/proc/sys/dev/i915/perf_stream_paranoid", &sb
) == 0) {
929 /* If _paranoid == 1 then on Gen8+ we won't be able to access OA
930 * metrics unless running as root.
932 if (devinfo
->is_haswell
)
933 i915_perf_oa_available
= true;
935 uint64_t paranoid
= 1;
937 read_file_uint64("/proc/sys/dev/i915/perf_stream_paranoid", ¶noid
);
939 if (paranoid
== 0 || geteuid() == 0)
940 i915_perf_oa_available
= true;
944 if (!i915_perf_oa_available
||
946 !get_sysfs_dev_dir(perf
, fd
) ||
947 !init_oa_sys_vars(perf
, devinfo
))
950 perf
->oa_metrics_table
=
951 _mesa_hash_table_create(perf
, _mesa_hash_string
,
952 _mesa_key_string_equal
);
954 /* Index all the metric sets mesa knows about before looking to see what
955 * the kernel is advertising.
959 if (likely((INTEL_DEBUG
& DEBUG_NO_OACONFIG
) == 0) &&
960 kernel_has_dynamic_config_support(perf
, fd
))
961 init_oa_configs(perf
, fd
);
963 enumerate_sysfs_metrics(perf
);
968 struct gen_perf_registers
*
969 gen_perf_load_configuration(struct gen_perf_config
*perf_cfg
, int fd
, const char *guid
)
971 if (!perf_cfg
->i915_query_supported
)
974 struct drm_i915_perf_oa_config i915_config
= { 0, };
975 if (!i915_query_perf_config_data(perf_cfg
, fd
, guid
, &i915_config
))
978 struct gen_perf_registers
*config
= rzalloc(NULL
, struct gen_perf_registers
);
979 config
->n_flex_regs
= i915_config
.n_flex_regs
;
980 config
->flex_regs
= rzalloc_array(config
, struct gen_perf_query_register_prog
, config
->n_flex_regs
);
981 config
->n_mux_regs
= i915_config
.n_mux_regs
;
982 config
->mux_regs
= rzalloc_array(config
, struct gen_perf_query_register_prog
, config
->n_mux_regs
);
983 config
->n_b_counter_regs
= i915_config
.n_boolean_regs
;
984 config
->b_counter_regs
= rzalloc_array(config
, struct gen_perf_query_register_prog
, config
->n_b_counter_regs
);
987 * struct gen_perf_query_register_prog maps exactly to the tuple of
988 * (register offset, register value) returned by the i915.
990 i915_config
.flex_regs_ptr
= to_user_pointer(config
->flex_regs
);
991 i915_config
.mux_regs_ptr
= to_user_pointer(config
->mux_regs
);
992 i915_config
.boolean_regs_ptr
= to_user_pointer(config
->b_counter_regs
);
993 if (!i915_query_perf_config_data(perf_cfg
, fd
, guid
, &i915_config
)) {
1002 gen_perf_store_configuration(struct gen_perf_config
*perf_cfg
, int fd
,
1003 const struct gen_perf_registers
*config
,
1007 return i915_add_config(perf_cfg
, fd
, config
, guid
);
1009 struct mesa_sha1 sha1_ctx
;
1010 _mesa_sha1_init(&sha1_ctx
);
1012 if (config
->flex_regs
) {
1013 _mesa_sha1_update(&sha1_ctx
, config
->flex_regs
,
1014 sizeof(config
->flex_regs
[0]) *
1015 config
->n_flex_regs
);
1017 if (config
->mux_regs
) {
1018 _mesa_sha1_update(&sha1_ctx
, config
->mux_regs
,
1019 sizeof(config
->mux_regs
[0]) *
1020 config
->n_mux_regs
);
1022 if (config
->b_counter_regs
) {
1023 _mesa_sha1_update(&sha1_ctx
, config
->b_counter_regs
,
1024 sizeof(config
->b_counter_regs
[0]) *
1025 config
->n_b_counter_regs
);
1029 _mesa_sha1_final(&sha1_ctx
, hash
);
1031 char formatted_hash
[41];
1032 _mesa_sha1_format(formatted_hash
, hash
);
1034 char generated_guid
[37];
1035 snprintf(generated_guid
, sizeof(generated_guid
),
1036 "%.8s-%.4s-%.4s-%.4s-%.12s",
1037 &formatted_hash
[0], &formatted_hash
[8],
1038 &formatted_hash
[8 + 4], &formatted_hash
[8 + 4 + 4],
1039 &formatted_hash
[8 + 4 + 4 + 4]);
1041 /* Check if already present. */
1043 if (gen_perf_load_metric_id(perf_cfg
, generated_guid
, &id
))
1046 return i915_add_config(perf_cfg
, fd
, config
, generated_guid
);
1049 /* Accumulate 32bits OA counters */
1051 accumulate_uint32(const uint32_t *report0
,
1052 const uint32_t *report1
,
1053 uint64_t *accumulator
)
1055 *accumulator
+= (uint32_t)(*report1
- *report0
);
1058 /* Accumulate 40bits OA counters */
1060 accumulate_uint40(int a_index
,
1061 const uint32_t *report0
,
1062 const uint32_t *report1
,
1063 uint64_t *accumulator
)
1065 const uint8_t *high_bytes0
= (uint8_t *)(report0
+ 40);
1066 const uint8_t *high_bytes1
= (uint8_t *)(report1
+ 40);
1067 uint64_t high0
= (uint64_t)(high_bytes0
[a_index
]) << 32;
1068 uint64_t high1
= (uint64_t)(high_bytes1
[a_index
]) << 32;
1069 uint64_t value0
= report0
[a_index
+ 4] | high0
;
1070 uint64_t value1
= report1
[a_index
+ 4] | high1
;
1073 if (value0
> value1
)
1074 delta
= (1ULL << 40) + value1
- value0
;
1076 delta
= value1
- value0
;
1078 *accumulator
+= delta
;
1082 gen8_read_report_clock_ratios(const uint32_t *report
,
1083 uint64_t *slice_freq_hz
,
1084 uint64_t *unslice_freq_hz
)
1086 /* The lower 16bits of the RPT_ID field of the OA reports contains a
1087 * snapshot of the bits coming from the RP_FREQ_NORMAL register and is
1088 * divided this way :
1090 * RPT_ID[31:25]: RP_FREQ_NORMAL[20:14] (low squashed_slice_clock_frequency)
1091 * RPT_ID[10:9]: RP_FREQ_NORMAL[22:21] (high squashed_slice_clock_frequency)
1092 * RPT_ID[8:0]: RP_FREQ_NORMAL[31:23] (squashed_unslice_clock_frequency)
1094 * RP_FREQ_NORMAL[31:23]: Software Unslice Ratio Request
1095 * Multiple of 33.33MHz 2xclk (16 MHz 1xclk)
1097 * RP_FREQ_NORMAL[22:14]: Software Slice Ratio Request
1098 * Multiple of 33.33MHz 2xclk (16 MHz 1xclk)
1101 uint32_t unslice_freq
= report
[0] & 0x1ff;
1102 uint32_t slice_freq_low
= (report
[0] >> 25) & 0x7f;
1103 uint32_t slice_freq_high
= (report
[0] >> 9) & 0x3;
1104 uint32_t slice_freq
= slice_freq_low
| (slice_freq_high
<< 7);
1106 *slice_freq_hz
= slice_freq
* 16666667ULL;
1107 *unslice_freq_hz
= unslice_freq
* 16666667ULL;
1111 gen_perf_query_result_read_frequencies(struct gen_perf_query_result
*result
,
1112 const struct gen_device_info
*devinfo
,
1113 const uint32_t *start
,
1114 const uint32_t *end
)
1116 /* Slice/Unslice frequency is only available in the OA reports when the
1117 * "Disable OA reports due to clock ratio change" field in
1118 * OA_DEBUG_REGISTER is set to 1. This is how the kernel programs this
1119 * global register (see drivers/gpu/drm/i915/i915_perf.c)
1121 * Documentation says this should be available on Gen9+ but experimentation
1122 * shows that Gen8 reports similar values, so we enable it there too.
1124 if (devinfo
->gen
< 8)
1127 gen8_read_report_clock_ratios(start
,
1128 &result
->slice_frequency
[0],
1129 &result
->unslice_frequency
[0]);
1130 gen8_read_report_clock_ratios(end
,
1131 &result
->slice_frequency
[1],
1132 &result
->unslice_frequency
[1]);
1136 gen_perf_query_result_accumulate(struct gen_perf_query_result
*result
,
1137 const struct gen_perf_query_info
*query
,
1138 const uint32_t *start
,
1139 const uint32_t *end
)
1143 if (result
->hw_id
== OA_REPORT_INVALID_CTX_ID
&&
1144 start
[2] != OA_REPORT_INVALID_CTX_ID
)
1145 result
->hw_id
= start
[2];
1146 if (result
->reports_accumulated
== 0)
1147 result
->begin_timestamp
= start
[1];
1148 result
->reports_accumulated
++;
1150 switch (query
->oa_format
) {
1151 case I915_OA_FORMAT_A32u40_A4u32_B8_C8
:
1152 accumulate_uint32(start
+ 1, end
+ 1, result
->accumulator
+ idx
++); /* timestamp */
1153 accumulate_uint32(start
+ 3, end
+ 3, result
->accumulator
+ idx
++); /* clock */
1155 /* 32x 40bit A counters... */
1156 for (i
= 0; i
< 32; i
++)
1157 accumulate_uint40(i
, start
, end
, result
->accumulator
+ idx
++);
1159 /* 4x 32bit A counters... */
1160 for (i
= 0; i
< 4; i
++)
1161 accumulate_uint32(start
+ 36 + i
, end
+ 36 + i
, result
->accumulator
+ idx
++);
1163 /* 8x 32bit B counters + 8x 32bit C counters... */
1164 for (i
= 0; i
< 16; i
++)
1165 accumulate_uint32(start
+ 48 + i
, end
+ 48 + i
, result
->accumulator
+ idx
++);
1168 case I915_OA_FORMAT_A45_B8_C8
:
1169 accumulate_uint32(start
+ 1, end
+ 1, result
->accumulator
); /* timestamp */
1171 for (i
= 0; i
< 61; i
++)
1172 accumulate_uint32(start
+ 3 + i
, end
+ 3 + i
, result
->accumulator
+ 1 + i
);
1176 unreachable("Can't accumulate OA counters in unknown format");
1182 gen_perf_query_result_clear(struct gen_perf_query_result
*result
)
1184 memset(result
, 0, sizeof(*result
));
1185 result
->hw_id
= OA_REPORT_INVALID_CTX_ID
; /* invalid */
1189 register_mdapi_statistic_query(struct gen_perf_config
*perf_cfg
,
1190 const struct gen_device_info
*devinfo
)
1192 if (!(devinfo
->gen
>= 7 && devinfo
->gen
<= 11))
1195 struct gen_perf_query_info
*query
=
1196 append_query_info(perf_cfg
, MAX_STAT_COUNTERS
);
1198 query
->kind
= GEN_PERF_QUERY_TYPE_PIPELINE
;
1199 query
->name
= "Intel_Raw_Pipeline_Statistics_Query";
1201 /* The order has to match mdapi_pipeline_metrics. */
1202 add_basic_stat_reg(query
, IA_VERTICES_COUNT
,
1203 "N vertices submitted");
1204 add_basic_stat_reg(query
, IA_PRIMITIVES_COUNT
,
1205 "N primitives submitted");
1206 add_basic_stat_reg(query
, VS_INVOCATION_COUNT
,
1207 "N vertex shader invocations");
1208 add_basic_stat_reg(query
, GS_INVOCATION_COUNT
,
1209 "N geometry shader invocations");
1210 add_basic_stat_reg(query
, GS_PRIMITIVES_COUNT
,
1211 "N geometry shader primitives emitted");
1212 add_basic_stat_reg(query
, CL_INVOCATION_COUNT
,
1213 "N primitives entering clipping");
1214 add_basic_stat_reg(query
, CL_PRIMITIVES_COUNT
,
1215 "N primitives leaving clipping");
1216 if (devinfo
->is_haswell
|| devinfo
->gen
== 8) {
1217 add_stat_reg(query
, PS_INVOCATION_COUNT
, 1, 4,
1218 "N fragment shader invocations",
1219 "N fragment shader invocations");
1221 add_basic_stat_reg(query
, PS_INVOCATION_COUNT
,
1222 "N fragment shader invocations");
1224 add_basic_stat_reg(query
, HS_INVOCATION_COUNT
,
1225 "N TCS shader invocations");
1226 add_basic_stat_reg(query
, DS_INVOCATION_COUNT
,
1227 "N TES shader invocations");
1228 if (devinfo
->gen
>= 7) {
1229 add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
1230 "N compute shader invocations");
1233 if (devinfo
->gen
>= 10) {
1234 /* Reuse existing CS invocation register until we can expose this new
1237 add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
1241 query
->data_size
= sizeof(uint64_t) * query
->n_counters
;
1245 fill_mdapi_perf_query_counter(struct gen_perf_query_info
*query
,
1247 uint32_t data_offset
,
1249 enum gen_perf_counter_data_type data_type
)
1251 struct gen_perf_query_counter
*counter
= &query
->counters
[query
->n_counters
];
1253 assert(query
->n_counters
<= query
->max_counters
);
1255 counter
->name
= name
;
1256 counter
->desc
= "Raw counter value";
1257 counter
->type
= GEN_PERF_COUNTER_TYPE_RAW
;
1258 counter
->data_type
= data_type
;
1259 counter
->offset
= data_offset
;
1261 query
->n_counters
++;
1263 assert(counter
->offset
+ gen_perf_query_counter_get_size(counter
) <= query
->data_size
);
1266 #define MDAPI_QUERY_ADD_COUNTER(query, struct_name, field_name, type_name) \
1267 fill_mdapi_perf_query_counter(query, #field_name, \
1268 (uint8_t *) &struct_name.field_name - \
1269 (uint8_t *) &struct_name, \
1270 sizeof(struct_name.field_name), \
1271 GEN_PERF_COUNTER_DATA_TYPE_##type_name)
1272 #define MDAPI_QUERY_ADD_ARRAY_COUNTER(ctx, query, struct_name, field_name, idx, type_name) \
1273 fill_mdapi_perf_query_counter(query, \
1274 ralloc_asprintf(ctx, "%s%i", #field_name, idx), \
1275 (uint8_t *) &struct_name.field_name[idx] - \
1276 (uint8_t *) &struct_name, \
1277 sizeof(struct_name.field_name[0]), \
1278 GEN_PERF_COUNTER_DATA_TYPE_##type_name)
1281 register_mdapi_oa_query(const struct gen_device_info
*devinfo
,
1282 struct gen_perf_config
*perf
)
1284 struct gen_perf_query_info
*query
= NULL
;
1286 /* MDAPI requires different structures for pretty much every generation
1287 * (right now we have definitions for gen 7 to 11).
1289 if (!(devinfo
->gen
>= 7 && devinfo
->gen
<= 11))
1292 switch (devinfo
->gen
) {
1294 query
= append_query_info(perf
, 1 + 45 + 16 + 7);
1295 query
->oa_format
= I915_OA_FORMAT_A45_B8_C8
;
1297 struct gen7_mdapi_metrics metric_data
;
1298 query
->data_size
= sizeof(metric_data
);
1300 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, TotalTime
, UINT64
);
1301 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.ACounters
); i
++) {
1302 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1303 metric_data
, ACounters
, i
, UINT64
);
1305 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.NOACounters
); i
++) {
1306 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1307 metric_data
, NOACounters
, i
, UINT64
);
1309 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter1
, UINT64
);
1310 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter2
, UINT64
);
1311 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SplitOccured
, BOOL32
);
1312 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequencyChanged
, BOOL32
);
1313 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequency
, UINT64
);
1314 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportId
, UINT32
);
1315 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportsCount
, UINT32
);
1319 query
= append_query_info(perf
, 2 + 36 + 16 + 16);
1320 query
->oa_format
= I915_OA_FORMAT_A32u40_A4u32_B8_C8
;
1322 struct gen8_mdapi_metrics metric_data
;
1323 query
->data_size
= sizeof(metric_data
);
1325 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, TotalTime
, UINT64
);
1326 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, GPUTicks
, UINT64
);
1327 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.OaCntr
); i
++) {
1328 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1329 metric_data
, OaCntr
, i
, UINT64
);
1331 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.NoaCntr
); i
++) {
1332 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1333 metric_data
, NoaCntr
, i
, UINT64
);
1335 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, BeginTimestamp
, UINT64
);
1336 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved1
, UINT64
);
1337 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved2
, UINT64
);
1338 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved3
, UINT32
);
1339 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, OverrunOccured
, BOOL32
);
1340 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerUser
, UINT64
);
1341 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerDriver
, UINT64
);
1342 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SliceFrequency
, UINT64
);
1343 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, UnsliceFrequency
, UINT64
);
1344 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter1
, UINT64
);
1345 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter2
, UINT64
);
1346 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SplitOccured
, BOOL32
);
1347 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequencyChanged
, BOOL32
);
1348 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequency
, UINT64
);
1349 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportId
, UINT32
);
1350 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportsCount
, UINT32
);
1356 query
= append_query_info(perf
, 2 + 36 + 16 + 16 + 16 + 2);
1357 query
->oa_format
= I915_OA_FORMAT_A32u40_A4u32_B8_C8
;
1359 struct gen9_mdapi_metrics metric_data
;
1360 query
->data_size
= sizeof(metric_data
);
1362 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, TotalTime
, UINT64
);
1363 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, GPUTicks
, UINT64
);
1364 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.OaCntr
); i
++) {
1365 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1366 metric_data
, OaCntr
, i
, UINT64
);
1368 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.NoaCntr
); i
++) {
1369 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1370 metric_data
, NoaCntr
, i
, UINT64
);
1372 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, BeginTimestamp
, UINT64
);
1373 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved1
, UINT64
);
1374 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved2
, UINT64
);
1375 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved3
, UINT32
);
1376 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, OverrunOccured
, BOOL32
);
1377 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerUser
, UINT64
);
1378 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerDriver
, UINT64
);
1379 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SliceFrequency
, UINT64
);
1380 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, UnsliceFrequency
, UINT64
);
1381 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter1
, UINT64
);
1382 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter2
, UINT64
);
1383 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SplitOccured
, BOOL32
);
1384 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequencyChanged
, BOOL32
);
1385 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequency
, UINT64
);
1386 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportId
, UINT32
);
1387 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportsCount
, UINT32
);
1388 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.UserCntr
); i
++) {
1389 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1390 metric_data
, UserCntr
, i
, UINT64
);
1392 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, UserCntrCfgId
, UINT32
);
1393 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved4
, UINT32
);
1397 unreachable("Unsupported gen");
1401 query
->kind
= GEN_PERF_QUERY_TYPE_RAW
;
1402 query
->name
= "Intel_Raw_Hardware_Counters_Set_0_Query";
1403 query
->guid
= GEN_PERF_QUERY_GUID_MDAPI
;
1406 /* Accumulation buffer offsets copied from an actual query... */
1407 const struct gen_perf_query_info
*copy_query
=
1410 query
->gpu_time_offset
= copy_query
->gpu_time_offset
;
1411 query
->gpu_clock_offset
= copy_query
->gpu_clock_offset
;
1412 query
->a_offset
= copy_query
->a_offset
;
1413 query
->b_offset
= copy_query
->b_offset
;
1414 query
->c_offset
= copy_query
->c_offset
;
1419 get_metric_id(struct gen_perf_config
*perf
,
1420 const struct gen_perf_query_info
*query
)
1422 /* These queries are know not to ever change, their config ID has been
1423 * loaded upon the first query creation. No need to look them up again.
1425 if (query
->kind
== GEN_PERF_QUERY_TYPE_OA
)
1426 return query
->oa_metrics_set_id
;
1428 assert(query
->kind
== GEN_PERF_QUERY_TYPE_RAW
);
1430 /* Raw queries can be reprogrammed up by an external application/library.
1431 * When a raw query is used for the first time it's id is set to a value !=
1432 * 0. When it stops being used the id returns to 0. No need to reload the
1433 * ID when it's already loaded.
1435 if (query
->oa_metrics_set_id
!= 0) {
1436 DBG("Raw query '%s' guid=%s using cached ID: %"PRIu64
"\n",
1437 query
->name
, query
->guid
, query
->oa_metrics_set_id
);
1438 return query
->oa_metrics_set_id
;
1441 struct gen_perf_query_info
*raw_query
= (struct gen_perf_query_info
*)query
;
1442 if (!gen_perf_load_metric_id(perf
, query
->guid
,
1443 &raw_query
->oa_metrics_set_id
)) {
1444 DBG("Unable to read query guid=%s ID, falling back to test config\n", query
->guid
);
1445 raw_query
->oa_metrics_set_id
= 1ULL;
1447 DBG("Raw query '%s'guid=%s loaded ID: %"PRIu64
"\n",
1448 query
->name
, query
->guid
, query
->oa_metrics_set_id
);
1450 return query
->oa_metrics_set_id
;
1453 static struct oa_sample_buf
*
1454 get_free_sample_buf(struct gen_perf_context
*perf_ctx
)
1456 struct exec_node
*node
= exec_list_pop_head(&perf_ctx
->free_sample_buffers
);
1457 struct oa_sample_buf
*buf
;
1460 buf
= exec_node_data(struct oa_sample_buf
, node
, link
);
1462 buf
= ralloc_size(perf_ctx
->perf
, sizeof(*buf
));
1464 exec_node_init(&buf
->link
);
1473 reap_old_sample_buffers(struct gen_perf_context
*perf_ctx
)
1475 struct exec_node
*tail_node
=
1476 exec_list_get_tail(&perf_ctx
->sample_buffers
);
1477 struct oa_sample_buf
*tail_buf
=
1478 exec_node_data(struct oa_sample_buf
, tail_node
, link
);
1480 /* Remove all old, unreferenced sample buffers walking forward from
1481 * the head of the list, except always leave at least one node in
1482 * the list so we always have a node to reference when we Begin
1485 foreach_list_typed_safe(struct oa_sample_buf
, buf
, link
,
1486 &perf_ctx
->sample_buffers
)
1488 if (buf
->refcount
== 0 && buf
!= tail_buf
) {
1489 exec_node_remove(&buf
->link
);
1490 exec_list_push_head(&perf_ctx
->free_sample_buffers
, &buf
->link
);
1497 free_sample_bufs(struct gen_perf_context
*perf_ctx
)
1499 foreach_list_typed_safe(struct oa_sample_buf
, buf
, link
,
1500 &perf_ctx
->free_sample_buffers
)
1503 exec_list_make_empty(&perf_ctx
->free_sample_buffers
);
1506 /******************************************************************************/
1509 * Emit MI_STORE_REGISTER_MEM commands to capture all of the
1510 * pipeline statistics for the performance query object.
1513 snapshot_statistics_registers(struct gen_perf_context
*ctx
,
1514 struct gen_perf_query_object
*obj
,
1515 uint32_t offset_in_bytes
)
1517 struct gen_perf_config
*perf
= ctx
->perf
;
1518 const struct gen_perf_query_info
*query
= obj
->queryinfo
;
1519 const int n_counters
= query
->n_counters
;
1521 for (int i
= 0; i
< n_counters
; i
++) {
1522 const struct gen_perf_query_counter
*counter
= &query
->counters
[i
];
1524 assert(counter
->data_type
== GEN_PERF_COUNTER_DATA_TYPE_UINT64
);
1526 perf
->vtbl
.store_register_mem(ctx
->ctx
, obj
->pipeline_stats
.bo
,
1527 counter
->pipeline_stat
.reg
, 8,
1528 offset_in_bytes
+ i
* sizeof(uint64_t));
1533 snapshot_freq_register(struct gen_perf_context
*ctx
,
1534 struct gen_perf_query_object
*query
,
1537 struct gen_perf_config
*perf
= ctx
->perf
;
1538 const struct gen_device_info
*devinfo
= ctx
->devinfo
;
1540 if (devinfo
->gen
== 8 && !devinfo
->is_cherryview
)
1541 perf
->vtbl
.store_register_mem(ctx
->ctx
, query
->oa
.bo
, GEN7_RPSTAT1
, 4, bo_offset
);
1542 else if (devinfo
->gen
>= 9)
1543 perf
->vtbl
.store_register_mem(ctx
->ctx
, query
->oa
.bo
, GEN9_RPSTAT0
, 4, bo_offset
);
1547 gen_perf_close(struct gen_perf_context
*perfquery
,
1548 const struct gen_perf_query_info
*query
)
1550 if (perfquery
->oa_stream_fd
!= -1) {
1551 close(perfquery
->oa_stream_fd
);
1552 perfquery
->oa_stream_fd
= -1;
1554 if (query
->kind
== GEN_PERF_QUERY_TYPE_RAW
) {
1555 struct gen_perf_query_info
*raw_query
=
1556 (struct gen_perf_query_info
*) query
;
1557 raw_query
->oa_metrics_set_id
= 0;
1562 gen_perf_open(struct gen_perf_context
*perf_ctx
,
1565 int period_exponent
,
1569 uint64_t properties
[] = {
1570 /* Single context sampling */
1571 DRM_I915_PERF_PROP_CTX_HANDLE
, ctx_id
,
1573 /* Include OA reports in samples */
1574 DRM_I915_PERF_PROP_SAMPLE_OA
, true,
1576 /* OA unit configuration */
1577 DRM_I915_PERF_PROP_OA_METRICS_SET
, metrics_set_id
,
1578 DRM_I915_PERF_PROP_OA_FORMAT
, report_format
,
1579 DRM_I915_PERF_PROP_OA_EXPONENT
, period_exponent
,
1581 struct drm_i915_perf_open_param param
= {
1582 .flags
= I915_PERF_FLAG_FD_CLOEXEC
|
1583 I915_PERF_FLAG_FD_NONBLOCK
|
1584 I915_PERF_FLAG_DISABLED
,
1585 .num_properties
= ARRAY_SIZE(properties
) / 2,
1586 .properties_ptr
= (uintptr_t) properties
,
1588 int fd
= gen_ioctl(drm_fd
, DRM_IOCTL_I915_PERF_OPEN
, ¶m
);
1590 DBG("Error opening gen perf OA stream: %m\n");
1594 perf_ctx
->oa_stream_fd
= fd
;
1596 perf_ctx
->current_oa_metrics_set_id
= metrics_set_id
;
1597 perf_ctx
->current_oa_format
= report_format
;
1603 inc_n_users(struct gen_perf_context
*perf_ctx
)
1605 if (perf_ctx
->n_oa_users
== 0 &&
1606 gen_ioctl(perf_ctx
->oa_stream_fd
, I915_PERF_IOCTL_ENABLE
, 0) < 0)
1610 ++perf_ctx
->n_oa_users
;
1616 dec_n_users(struct gen_perf_context
*perf_ctx
)
1618 /* Disabling the i915 perf stream will effectively disable the OA
1619 * counters. Note it's important to be sure there are no outstanding
1620 * MI_RPC commands at this point since they could stall the CS
1621 * indefinitely once OACONTROL is disabled.
1623 --perf_ctx
->n_oa_users
;
1624 if (perf_ctx
->n_oa_users
== 0 &&
1625 gen_ioctl(perf_ctx
->oa_stream_fd
, I915_PERF_IOCTL_DISABLE
, 0) < 0)
1627 DBG("WARNING: Error disabling gen perf stream: %m\n");
1632 gen_perf_init_metrics(struct gen_perf_config
*perf_cfg
,
1633 const struct gen_device_info
*devinfo
,
1636 load_pipeline_statistic_metrics(perf_cfg
, devinfo
);
1637 register_mdapi_statistic_query(perf_cfg
, devinfo
);
1638 if (load_oa_metrics(perf_cfg
, drm_fd
, devinfo
))
1639 register_mdapi_oa_query(devinfo
, perf_cfg
);
1643 gen_perf_init_context(struct gen_perf_context
*perf_ctx
,
1644 struct gen_perf_config
*perf_cfg
,
1645 void * ctx
, /* driver context (eg, brw_context) */
1646 void * bufmgr
, /* eg brw_bufmgr */
1647 const struct gen_device_info
*devinfo
,
1651 perf_ctx
->perf
= perf_cfg
;
1652 perf_ctx
->ctx
= ctx
;
1653 perf_ctx
->bufmgr
= bufmgr
;
1654 perf_ctx
->drm_fd
= drm_fd
;
1655 perf_ctx
->hw_ctx
= hw_ctx
;
1656 perf_ctx
->devinfo
= devinfo
;
1658 perf_ctx
->unaccumulated
=
1659 ralloc_array(ctx
, struct gen_perf_query_object
*, 2);
1660 perf_ctx
->unaccumulated_elements
= 0;
1661 perf_ctx
->unaccumulated_array_size
= 2;
1663 exec_list_make_empty(&perf_ctx
->sample_buffers
);
1664 exec_list_make_empty(&perf_ctx
->free_sample_buffers
);
1666 /* It's convenient to guarantee that this linked list of sample
1667 * buffers is never empty so we add an empty head so when we
1668 * Begin an OA query we can always take a reference on a buffer
1671 struct oa_sample_buf
*buf
= get_free_sample_buf(perf_ctx
);
1672 exec_list_push_head(&perf_ctx
->sample_buffers
, &buf
->link
);
1674 perf_ctx
->oa_stream_fd
= -1;
1675 perf_ctx
->next_query_start_report_id
= 1000;
1679 * Add a query to the global list of "unaccumulated queries."
1681 * Queries are tracked here until all the associated OA reports have
1682 * been accumulated via accumulate_oa_reports() after the end
1683 * MI_REPORT_PERF_COUNT has landed in query->oa.bo.
1686 add_to_unaccumulated_query_list(struct gen_perf_context
*perf_ctx
,
1687 struct gen_perf_query_object
*obj
)
1689 if (perf_ctx
->unaccumulated_elements
>=
1690 perf_ctx
->unaccumulated_array_size
)
1692 perf_ctx
->unaccumulated_array_size
*= 1.5;
1693 perf_ctx
->unaccumulated
=
1694 reralloc(perf_ctx
->ctx
, perf_ctx
->unaccumulated
,
1695 struct gen_perf_query_object
*,
1696 perf_ctx
->unaccumulated_array_size
);
1699 perf_ctx
->unaccumulated
[perf_ctx
->unaccumulated_elements
++] = obj
;
1703 gen_perf_begin_query(struct gen_perf_context
*perf_ctx
,
1704 struct gen_perf_query_object
*query
)
1706 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1707 const struct gen_perf_query_info
*queryinfo
= query
->queryinfo
;
1709 /* XXX: We have to consider that the command parser unit that parses batch
1710 * buffer commands and is used to capture begin/end counter snapshots isn't
1711 * implicitly synchronized with what's currently running across other GPU
1712 * units (such as the EUs running shaders) that the performance counters are
1715 * The intention of performance queries is to measure the work associated
1716 * with commands between the begin/end delimiters and so for that to be the
1717 * case we need to explicitly synchronize the parsing of commands to capture
1718 * Begin/End counter snapshots with what's running across other parts of the
1721 * When the command parser reaches a Begin marker it effectively needs to
1722 * drain everything currently running on the GPU until the hardware is idle
1723 * before capturing the first snapshot of counters - otherwise the results
1724 * would also be measuring the effects of earlier commands.
1726 * When the command parser reaches an End marker it needs to stall until
1727 * everything currently running on the GPU has finished before capturing the
1728 * end snapshot - otherwise the results won't be a complete representation
1731 * To achieve this, we stall the pipeline at pixel scoreboard (prevent any
1732 * additional work to be processed by the pipeline until all pixels of the
1733 * previous draw has be completed).
1735 * N.B. The final results are based on deltas of counters between (inside)
1736 * Begin/End markers so even though the total wall clock time of the
1737 * workload is stretched by larger pipeline bubbles the bubbles themselves
1738 * are generally invisible to the query results. Whether that's a good or a
1739 * bad thing depends on the use case. For a lower real-time impact while
1740 * capturing metrics then periodic sampling may be a better choice than
1741 * INTEL_performance_query.
1744 * This is our Begin synchronization point to drain current work on the
1745 * GPU before we capture our first counter snapshot...
1747 perf_cfg
->vtbl
.emit_stall_at_pixel_scoreboard(perf_ctx
->ctx
);
1749 switch (queryinfo
->kind
) {
1750 case GEN_PERF_QUERY_TYPE_OA
:
1751 case GEN_PERF_QUERY_TYPE_RAW
: {
1753 /* Opening an i915 perf stream implies exclusive access to the OA unit
1754 * which will generate counter reports for a specific counter set with a
1755 * specific layout/format so we can't begin any OA based queries that
1756 * require a different counter set or format unless we get an opportunity
1757 * to close the stream and open a new one...
1759 uint64_t metric_id
= get_metric_id(perf_ctx
->perf
, queryinfo
);
1761 if (perf_ctx
->oa_stream_fd
!= -1 &&
1762 perf_ctx
->current_oa_metrics_set_id
!= metric_id
) {
1764 if (perf_ctx
->n_oa_users
!= 0) {
1765 DBG("WARNING: Begin failed already using perf config=%i/%"PRIu64
"\n",
1766 perf_ctx
->current_oa_metrics_set_id
, metric_id
);
1769 gen_perf_close(perf_ctx
, queryinfo
);
1772 /* If the OA counters aren't already on, enable them. */
1773 if (perf_ctx
->oa_stream_fd
== -1) {
1774 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
1776 /* The period_exponent gives a sampling period as follows:
1777 * sample_period = timestamp_period * 2^(period_exponent + 1)
1779 * The timestamps increments every 80ns (HSW), ~52ns (GEN9LP) or
1782 * The counter overflow period is derived from the EuActive counter
1783 * which reads a counter that increments by the number of clock
1784 * cycles multiplied by the number of EUs. It can be calculated as:
1786 * 2^(number of bits in A counter) / (n_eus * max_gen_freq * 2)
1788 * (E.g. 40 EUs @ 1GHz = ~53ms)
1790 * We select a sampling period inferior to that overflow period to
1791 * ensure we cannot see more than 1 counter overflow, otherwise we
1792 * could loose information.
1795 int a_counter_in_bits
= 32;
1796 if (devinfo
->gen
>= 8)
1797 a_counter_in_bits
= 40;
1799 uint64_t overflow_period
= pow(2, a_counter_in_bits
) / (perf_cfg
->sys_vars
.n_eus
*
1800 /* drop 1GHz freq to have units in nanoseconds */
1803 DBG("A counter overflow period: %"PRIu64
"ns, %"PRIu64
"ms (n_eus=%"PRIu64
")\n",
1804 overflow_period
, overflow_period
/ 1000000ul, perf_cfg
->sys_vars
.n_eus
);
1806 int period_exponent
= 0;
1807 uint64_t prev_sample_period
, next_sample_period
;
1808 for (int e
= 0; e
< 30; e
++) {
1809 prev_sample_period
= 1000000000ull * pow(2, e
+ 1) / devinfo
->timestamp_frequency
;
1810 next_sample_period
= 1000000000ull * pow(2, e
+ 2) / devinfo
->timestamp_frequency
;
1812 /* Take the previous sampling period, lower than the overflow
1815 if (prev_sample_period
< overflow_period
&&
1816 next_sample_period
> overflow_period
)
1817 period_exponent
= e
+ 1;
1820 if (period_exponent
== 0) {
1821 DBG("WARNING: enable to find a sampling exponent\n");
1825 DBG("OA sampling exponent: %i ~= %"PRIu64
"ms\n", period_exponent
,
1826 prev_sample_period
/ 1000000ul);
1828 if (!gen_perf_open(perf_ctx
, metric_id
, queryinfo
->oa_format
,
1829 period_exponent
, perf_ctx
->drm_fd
,
1833 assert(perf_ctx
->current_oa_metrics_set_id
== metric_id
&&
1834 perf_ctx
->current_oa_format
== queryinfo
->oa_format
);
1837 if (!inc_n_users(perf_ctx
)) {
1838 DBG("WARNING: Error enabling i915 perf stream: %m\n");
1843 perf_cfg
->vtbl
.bo_unreference(query
->oa
.bo
);
1844 query
->oa
.bo
= NULL
;
1847 query
->oa
.bo
= perf_cfg
->vtbl
.bo_alloc(perf_ctx
->bufmgr
,
1848 "perf. query OA MI_RPC bo",
1851 /* Pre-filling the BO helps debug whether writes landed. */
1852 void *map
= perf_cfg
->vtbl
.bo_map(perf_ctx
->ctx
, query
->oa
.bo
, MAP_WRITE
);
1853 memset(map
, 0x80, MI_RPC_BO_SIZE
);
1854 perf_cfg
->vtbl
.bo_unmap(query
->oa
.bo
);
1857 query
->oa
.begin_report_id
= perf_ctx
->next_query_start_report_id
;
1858 perf_ctx
->next_query_start_report_id
+= 2;
1860 /* Take a starting OA counter snapshot. */
1861 perf_cfg
->vtbl
.emit_mi_report_perf_count(perf_ctx
->ctx
, query
->oa
.bo
, 0,
1862 query
->oa
.begin_report_id
);
1863 snapshot_freq_register(perf_ctx
, query
, MI_FREQ_START_OFFSET_BYTES
);
1865 ++perf_ctx
->n_active_oa_queries
;
1867 /* No already-buffered samples can possibly be associated with this query
1868 * so create a marker within the list of sample buffers enabling us to
1869 * easily ignore earlier samples when processing this query after
1872 assert(!exec_list_is_empty(&perf_ctx
->sample_buffers
));
1873 query
->oa
.samples_head
= exec_list_get_tail(&perf_ctx
->sample_buffers
);
1875 struct oa_sample_buf
*buf
=
1876 exec_node_data(struct oa_sample_buf
, query
->oa
.samples_head
, link
);
1878 /* This reference will ensure that future/following sample
1879 * buffers (that may relate to this query) can't be freed until
1880 * this drops to zero.
1884 gen_perf_query_result_clear(&query
->oa
.result
);
1885 query
->oa
.results_accumulated
= false;
1887 add_to_unaccumulated_query_list(perf_ctx
, query
);
1891 case GEN_PERF_QUERY_TYPE_PIPELINE
:
1892 if (query
->pipeline_stats
.bo
) {
1893 perf_cfg
->vtbl
.bo_unreference(query
->pipeline_stats
.bo
);
1894 query
->pipeline_stats
.bo
= NULL
;
1897 query
->pipeline_stats
.bo
=
1898 perf_cfg
->vtbl
.bo_alloc(perf_ctx
->bufmgr
,
1899 "perf. query pipeline stats bo",
1902 /* Take starting snapshots. */
1903 snapshot_statistics_registers(perf_ctx
, query
, 0);
1905 ++perf_ctx
->n_active_pipeline_stats_queries
;
1909 unreachable("Unknown query type");
1917 gen_perf_end_query(struct gen_perf_context
*perf_ctx
,
1918 struct gen_perf_query_object
*query
)
1920 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1922 /* Ensure that the work associated with the queried commands will have
1923 * finished before taking our query end counter readings.
1925 * For more details see comment in brw_begin_perf_query for
1926 * corresponding flush.
1928 perf_cfg
->vtbl
.emit_stall_at_pixel_scoreboard(perf_ctx
->ctx
);
1930 switch (query
->queryinfo
->kind
) {
1931 case GEN_PERF_QUERY_TYPE_OA
:
1932 case GEN_PERF_QUERY_TYPE_RAW
:
1934 /* NB: It's possible that the query will have already been marked
1935 * as 'accumulated' if an error was seen while reading samples
1936 * from perf. In this case we mustn't try and emit a closing
1937 * MI_RPC command in case the OA unit has already been disabled
1939 if (!query
->oa
.results_accumulated
) {
1940 /* Take an ending OA counter snapshot. */
1941 snapshot_freq_register(perf_ctx
, query
, MI_FREQ_END_OFFSET_BYTES
);
1942 perf_cfg
->vtbl
.emit_mi_report_perf_count(perf_ctx
->ctx
, query
->oa
.bo
,
1943 MI_RPC_BO_END_OFFSET_BYTES
,
1944 query
->oa
.begin_report_id
+ 1);
1947 --perf_ctx
->n_active_oa_queries
;
1949 /* NB: even though the query has now ended, it can't be accumulated
1950 * until the end MI_REPORT_PERF_COUNT snapshot has been written
1955 case GEN_PERF_QUERY_TYPE_PIPELINE
:
1956 snapshot_statistics_registers(perf_ctx
, query
,
1957 STATS_BO_END_OFFSET_BYTES
);
1958 --perf_ctx
->n_active_pipeline_stats_queries
;
1962 unreachable("Unknown query type");
1968 OA_READ_STATUS_ERROR
,
1969 OA_READ_STATUS_UNFINISHED
,
1970 OA_READ_STATUS_FINISHED
,
1973 static enum OaReadStatus
1974 read_oa_samples_until(struct gen_perf_context
*perf_ctx
,
1975 uint32_t start_timestamp
,
1976 uint32_t end_timestamp
)
1978 struct exec_node
*tail_node
=
1979 exec_list_get_tail(&perf_ctx
->sample_buffers
);
1980 struct oa_sample_buf
*tail_buf
=
1981 exec_node_data(struct oa_sample_buf
, tail_node
, link
);
1982 uint32_t last_timestamp
=
1983 tail_buf
->len
== 0 ? start_timestamp
: tail_buf
->last_timestamp
;
1986 struct oa_sample_buf
*buf
= get_free_sample_buf(perf_ctx
);
1990 while ((len
= read(perf_ctx
->oa_stream_fd
, buf
->buf
,
1991 sizeof(buf
->buf
))) < 0 && errno
== EINTR
)
1995 exec_list_push_tail(&perf_ctx
->free_sample_buffers
, &buf
->link
);
1998 if (errno
== EAGAIN
) {
1999 return ((last_timestamp
- start_timestamp
) < INT32_MAX
&&
2000 (last_timestamp
- start_timestamp
) >=
2001 (end_timestamp
- start_timestamp
)) ?
2002 OA_READ_STATUS_FINISHED
:
2003 OA_READ_STATUS_UNFINISHED
;
2005 DBG("Error reading i915 perf samples: %m\n");
2008 DBG("Spurious EOF reading i915 perf samples\n");
2010 return OA_READ_STATUS_ERROR
;
2014 exec_list_push_tail(&perf_ctx
->sample_buffers
, &buf
->link
);
2016 /* Go through the reports and update the last timestamp. */
2018 while (offset
< buf
->len
) {
2019 const struct drm_i915_perf_record_header
*header
=
2020 (const struct drm_i915_perf_record_header
*) &buf
->buf
[offset
];
2021 uint32_t *report
= (uint32_t *) (header
+ 1);
2023 if (header
->type
== DRM_I915_PERF_RECORD_SAMPLE
)
2024 last_timestamp
= report
[1];
2026 offset
+= header
->size
;
2029 buf
->last_timestamp
= last_timestamp
;
2032 unreachable("not reached");
2033 return OA_READ_STATUS_ERROR
;
2037 * Try to read all the reports until either the delimiting timestamp
2038 * or an error arises.
2041 read_oa_samples_for_query(struct gen_perf_context
*perf_ctx
,
2042 struct gen_perf_query_object
*query
,
2043 void *current_batch
)
2048 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2050 /* We need the MI_REPORT_PERF_COUNT to land before we can start
2052 assert(!perf_cfg
->vtbl
.batch_references(current_batch
, query
->oa
.bo
) &&
2053 !perf_cfg
->vtbl
.bo_busy(query
->oa
.bo
));
2055 /* Map the BO once here and let accumulate_oa_reports() unmap
2057 if (query
->oa
.map
== NULL
)
2058 query
->oa
.map
= perf_cfg
->vtbl
.bo_map(perf_ctx
->ctx
, query
->oa
.bo
, MAP_READ
);
2060 start
= last
= query
->oa
.map
;
2061 end
= query
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
2063 if (start
[0] != query
->oa
.begin_report_id
) {
2064 DBG("Spurious start report id=%"PRIu32
"\n", start
[0]);
2067 if (end
[0] != (query
->oa
.begin_report_id
+ 1)) {
2068 DBG("Spurious end report id=%"PRIu32
"\n", end
[0]);
2072 /* Read the reports until the end timestamp. */
2073 switch (read_oa_samples_until(perf_ctx
, start
[1], end
[1])) {
2074 case OA_READ_STATUS_ERROR
:
2075 /* Fallthrough and let accumulate_oa_reports() deal with the
2077 case OA_READ_STATUS_FINISHED
:
2079 case OA_READ_STATUS_UNFINISHED
:
2083 unreachable("invalid read status");
2088 gen_perf_wait_query(struct gen_perf_context
*perf_ctx
,
2089 struct gen_perf_query_object
*query
,
2090 void *current_batch
)
2092 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2093 struct brw_bo
*bo
= NULL
;
2095 switch (query
->queryinfo
->kind
) {
2096 case GEN_PERF_QUERY_TYPE_OA
:
2097 case GEN_PERF_QUERY_TYPE_RAW
:
2101 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2102 bo
= query
->pipeline_stats
.bo
;
2106 unreachable("Unknown query type");
2113 /* If the current batch references our results bo then we need to
2116 if (perf_cfg
->vtbl
.batch_references(current_batch
, bo
))
2117 perf_cfg
->vtbl
.batchbuffer_flush(perf_ctx
->ctx
, __FILE__
, __LINE__
);
2119 perf_cfg
->vtbl
.bo_wait_rendering(bo
);
2121 /* Due to a race condition between the OA unit signaling report
2122 * availability and the report actually being written into memory,
2123 * we need to wait for all the reports to come in before we can
2126 if (query
->queryinfo
->kind
== GEN_PERF_QUERY_TYPE_OA
||
2127 query
->queryinfo
->kind
== GEN_PERF_QUERY_TYPE_RAW
) {
2128 while (!read_oa_samples_for_query(perf_ctx
, query
, current_batch
))
2134 gen_perf_is_query_ready(struct gen_perf_context
*perf_ctx
,
2135 struct gen_perf_query_object
*query
,
2136 void *current_batch
)
2138 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2140 switch (query
->queryinfo
->kind
) {
2141 case GEN_PERF_QUERY_TYPE_OA
:
2142 case GEN_PERF_QUERY_TYPE_RAW
:
2143 return (query
->oa
.results_accumulated
||
2145 !perf_cfg
->vtbl
.batch_references(current_batch
, query
->oa
.bo
) &&
2146 !perf_cfg
->vtbl
.bo_busy(query
->oa
.bo
) &&
2147 read_oa_samples_for_query(perf_ctx
, query
, current_batch
)));
2148 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2149 return (query
->pipeline_stats
.bo
&&
2150 !perf_cfg
->vtbl
.batch_references(current_batch
, query
->pipeline_stats
.bo
) &&
2151 !perf_cfg
->vtbl
.bo_busy(query
->pipeline_stats
.bo
));
2154 unreachable("Unknown query type");
2162 * Remove a query from the global list of unaccumulated queries once
2163 * after successfully accumulating the OA reports associated with the
2164 * query in accumulate_oa_reports() or when discarding unwanted query
2168 drop_from_unaccumulated_query_list(struct gen_perf_context
*perf_ctx
,
2169 struct gen_perf_query_object
*query
)
2171 for (int i
= 0; i
< perf_ctx
->unaccumulated_elements
; i
++) {
2172 if (perf_ctx
->unaccumulated
[i
] == query
) {
2173 int last_elt
= --perf_ctx
->unaccumulated_elements
;
2176 perf_ctx
->unaccumulated
[i
] = NULL
;
2178 perf_ctx
->unaccumulated
[i
] =
2179 perf_ctx
->unaccumulated
[last_elt
];
2186 /* Drop our samples_head reference so that associated periodic
2187 * sample data buffers can potentially be reaped if they aren't
2188 * referenced by any other queries...
2191 struct oa_sample_buf
*buf
=
2192 exec_node_data(struct oa_sample_buf
, query
->oa
.samples_head
, link
);
2194 assert(buf
->refcount
> 0);
2197 query
->oa
.samples_head
= NULL
;
2199 reap_old_sample_buffers(perf_ctx
);
2202 /* In general if we see anything spurious while accumulating results,
2203 * we don't try and continue accumulating the current query, hoping
2204 * for the best, we scrap anything outstanding, and then hope for the
2205 * best with new queries.
2208 discard_all_queries(struct gen_perf_context
*perf_ctx
)
2210 while (perf_ctx
->unaccumulated_elements
) {
2211 struct gen_perf_query_object
*query
= perf_ctx
->unaccumulated
[0];
2213 query
->oa
.results_accumulated
= true;
2214 drop_from_unaccumulated_query_list(perf_ctx
, query
);
2216 dec_n_users(perf_ctx
);
2220 /* Looks for the validity bit of context ID (dword 2) of an OA report. */
2222 oa_report_ctx_id_valid(const struct gen_device_info
*devinfo
,
2223 const uint32_t *report
)
2225 assert(devinfo
->gen
>= 8);
2226 if (devinfo
->gen
== 8)
2227 return (report
[0] & (1 << 25)) != 0;
2228 return (report
[0] & (1 << 16)) != 0;
2232 * Accumulate raw OA counter values based on deltas between pairs of
2235 * Accumulation starts from the first report captured via
2236 * MI_REPORT_PERF_COUNT (MI_RPC) by brw_begin_perf_query() until the
2237 * last MI_RPC report requested by brw_end_perf_query(). Between these
2238 * two reports there may also some number of periodically sampled OA
2239 * reports collected via the i915 perf interface - depending on the
2240 * duration of the query.
2242 * These periodic snapshots help to ensure we handle counter overflow
2243 * correctly by being frequent enough to ensure we don't miss multiple
2244 * overflows of a counter between snapshots. For Gen8+ the i915 perf
2245 * snapshots provide the extra context-switch reports that let us
2246 * subtract out the progress of counters associated with other
2247 * contexts running on the system.
2250 accumulate_oa_reports(struct gen_perf_context
*perf_ctx
,
2251 struct gen_perf_query_object
*query
)
2253 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
2257 struct exec_node
*first_samples_node
;
2258 bool last_report_ctx_match
= true;
2259 int out_duration
= 0;
2261 assert(query
->oa
.map
!= NULL
);
2263 start
= last
= query
->oa
.map
;
2264 end
= query
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
2266 if (start
[0] != query
->oa
.begin_report_id
) {
2267 DBG("Spurious start report id=%"PRIu32
"\n", start
[0]);
2270 if (end
[0] != (query
->oa
.begin_report_id
+ 1)) {
2271 DBG("Spurious end report id=%"PRIu32
"\n", end
[0]);
2275 /* On Gen12+ OA reports are sourced from per context counters, so we don't
2276 * ever have to look at the global OA buffer. Yey \o/
2278 if (perf_ctx
->devinfo
->gen
>= 12) {
2283 /* See if we have any periodic reports to accumulate too... */
2285 /* N.B. The oa.samples_head was set when the query began and
2286 * pointed to the tail of the perf_ctx->sample_buffers list at
2287 * the time the query started. Since the buffer existed before the
2288 * first MI_REPORT_PERF_COUNT command was emitted we therefore know
2289 * that no data in this particular node's buffer can possibly be
2290 * associated with the query - so skip ahead one...
2292 first_samples_node
= query
->oa
.samples_head
->next
;
2294 foreach_list_typed_from(struct oa_sample_buf
, buf
, link
,
2295 &perf_ctx
->sample_buffers
,
2300 while (offset
< buf
->len
) {
2301 const struct drm_i915_perf_record_header
*header
=
2302 (const struct drm_i915_perf_record_header
*)(buf
->buf
+ offset
);
2304 assert(header
->size
!= 0);
2305 assert(header
->size
<= buf
->len
);
2307 offset
+= header
->size
;
2309 switch (header
->type
) {
2310 case DRM_I915_PERF_RECORD_SAMPLE
: {
2311 uint32_t *report
= (uint32_t *)(header
+ 1);
2312 bool report_ctx_match
= true;
2315 /* Ignore reports that come before the start marker.
2316 * (Note: takes care to allow overflow of 32bit timestamps)
2318 if (gen_device_info_timebase_scale(devinfo
,
2319 report
[1] - start
[1]) > 5000000000) {
2323 /* Ignore reports that come after the end marker.
2324 * (Note: takes care to allow overflow of 32bit timestamps)
2326 if (gen_device_info_timebase_scale(devinfo
,
2327 report
[1] - end
[1]) <= 5000000000) {
2331 /* For Gen8+ since the counters continue while other
2332 * contexts are running we need to discount any unrelated
2333 * deltas. The hardware automatically generates a report
2334 * on context switch which gives us a new reference point
2335 * to continuing adding deltas from.
2337 * For Haswell we can rely on the HW to stop the progress
2338 * of OA counters while any other context is acctive.
2340 if (devinfo
->gen
>= 8) {
2341 /* Consider that the current report matches our context only if
2342 * the report says the report ID is valid.
2344 report_ctx_match
= oa_report_ctx_id_valid(devinfo
, report
) &&
2345 report
[2] == start
[2];
2346 if (report_ctx_match
)
2351 /* Only add the delta between <last, report> if the last report
2352 * was clearly identified as our context, or if we have at most
2353 * 1 report without a matching ID.
2355 * The OA unit will sometimes label reports with an invalid
2356 * context ID when i915 rewrites the execlist submit register
2357 * with the same context as the one currently running. This
2358 * happens when i915 wants to notify the HW of ringbuffer tail
2359 * register update. We have to consider this report as part of
2360 * our context as the 3d pipeline behind the OACS unit is still
2361 * processing the operations started at the previous execlist
2364 add
= last_report_ctx_match
&& out_duration
< 2;
2368 gen_perf_query_result_accumulate(&query
->oa
.result
,
2372 /* We're not adding the delta because we've identified it's not
2373 * for the context we filter for. We can consider that the
2376 query
->oa
.result
.query_disjoint
= true;
2380 last_report_ctx_match
= report_ctx_match
;
2385 case DRM_I915_PERF_RECORD_OA_BUFFER_LOST
:
2386 DBG("i915 perf: OA error: all reports lost\n");
2388 case DRM_I915_PERF_RECORD_OA_REPORT_LOST
:
2389 DBG("i915 perf: OA report lost\n");
2397 gen_perf_query_result_accumulate(&query
->oa
.result
, query
->queryinfo
,
2400 query
->oa
.results_accumulated
= true;
2401 drop_from_unaccumulated_query_list(perf_ctx
, query
);
2402 dec_n_users(perf_ctx
);
2408 discard_all_queries(perf_ctx
);
2412 gen_perf_delete_query(struct gen_perf_context
*perf_ctx
,
2413 struct gen_perf_query_object
*query
)
2415 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2417 /* We can assume that the frontend waits for a query to complete
2418 * before ever calling into here, so we don't have to worry about
2419 * deleting an in-flight query object.
2421 switch (query
->queryinfo
->kind
) {
2422 case GEN_PERF_QUERY_TYPE_OA
:
2423 case GEN_PERF_QUERY_TYPE_RAW
:
2425 if (!query
->oa
.results_accumulated
) {
2426 drop_from_unaccumulated_query_list(perf_ctx
, query
);
2427 dec_n_users(perf_ctx
);
2430 perf_cfg
->vtbl
.bo_unreference(query
->oa
.bo
);
2431 query
->oa
.bo
= NULL
;
2434 query
->oa
.results_accumulated
= false;
2437 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2438 if (query
->pipeline_stats
.bo
) {
2439 perf_cfg
->vtbl
.bo_unreference(query
->pipeline_stats
.bo
);
2440 query
->pipeline_stats
.bo
= NULL
;
2445 unreachable("Unknown query type");
2449 /* As an indication that the INTEL_performance_query extension is no
2450 * longer in use, it's a good time to free our cache of sample
2451 * buffers and close any current i915-perf stream.
2453 if (--perf_ctx
->n_query_instances
== 0) {
2454 free_sample_bufs(perf_ctx
);
2455 gen_perf_close(perf_ctx
, query
->queryinfo
);
2461 #define GET_FIELD(word, field) (((word) & field ## _MASK) >> field ## _SHIFT)
2464 read_gt_frequency(struct gen_perf_context
*perf_ctx
,
2465 struct gen_perf_query_object
*obj
)
2467 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
2468 uint32_t start
= *((uint32_t *)(obj
->oa
.map
+ MI_FREQ_START_OFFSET_BYTES
)),
2469 end
= *((uint32_t *)(obj
->oa
.map
+ MI_FREQ_END_OFFSET_BYTES
));
2471 switch (devinfo
->gen
) {
2474 obj
->oa
.gt_frequency
[0] = GET_FIELD(start
, GEN7_RPSTAT1_CURR_GT_FREQ
) * 50ULL;
2475 obj
->oa
.gt_frequency
[1] = GET_FIELD(end
, GEN7_RPSTAT1_CURR_GT_FREQ
) * 50ULL;
2480 obj
->oa
.gt_frequency
[0] = GET_FIELD(start
, GEN9_RPSTAT0_CURR_GT_FREQ
) * 50ULL / 3ULL;
2481 obj
->oa
.gt_frequency
[1] = GET_FIELD(end
, GEN9_RPSTAT0_CURR_GT_FREQ
) * 50ULL / 3ULL;
2484 unreachable("unexpected gen");
2487 /* Put the numbers into Hz. */
2488 obj
->oa
.gt_frequency
[0] *= 1000000ULL;
2489 obj
->oa
.gt_frequency
[1] *= 1000000ULL;
2493 get_oa_counter_data(struct gen_perf_context
*perf_ctx
,
2494 struct gen_perf_query_object
*query
,
2498 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2499 const struct gen_perf_query_info
*queryinfo
= query
->queryinfo
;
2500 int n_counters
= queryinfo
->n_counters
;
2503 for (int i
= 0; i
< n_counters
; i
++) {
2504 const struct gen_perf_query_counter
*counter
= &queryinfo
->counters
[i
];
2505 uint64_t *out_uint64
;
2507 size_t counter_size
= gen_perf_query_counter_get_size(counter
);
2510 switch (counter
->data_type
) {
2511 case GEN_PERF_COUNTER_DATA_TYPE_UINT64
:
2512 out_uint64
= (uint64_t *)(data
+ counter
->offset
);
2514 counter
->oa_counter_read_uint64(perf_cfg
, queryinfo
,
2515 query
->oa
.result
.accumulator
);
2517 case GEN_PERF_COUNTER_DATA_TYPE_FLOAT
:
2518 out_float
= (float *)(data
+ counter
->offset
);
2520 counter
->oa_counter_read_float(perf_cfg
, queryinfo
,
2521 query
->oa
.result
.accumulator
);
2524 /* So far we aren't using uint32, double or bool32... */
2525 unreachable("unexpected counter data type");
2527 written
= counter
->offset
+ counter_size
;
2535 get_pipeline_stats_data(struct gen_perf_context
*perf_ctx
,
2536 struct gen_perf_query_object
*query
,
2541 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2542 const struct gen_perf_query_info
*queryinfo
= query
->queryinfo
;
2543 int n_counters
= queryinfo
->n_counters
;
2546 uint64_t *start
= perf_cfg
->vtbl
.bo_map(perf_ctx
->ctx
, query
->pipeline_stats
.bo
, MAP_READ
);
2547 uint64_t *end
= start
+ (STATS_BO_END_OFFSET_BYTES
/ sizeof(uint64_t));
2549 for (int i
= 0; i
< n_counters
; i
++) {
2550 const struct gen_perf_query_counter
*counter
= &queryinfo
->counters
[i
];
2551 uint64_t value
= end
[i
] - start
[i
];
2553 if (counter
->pipeline_stat
.numerator
!=
2554 counter
->pipeline_stat
.denominator
) {
2555 value
*= counter
->pipeline_stat
.numerator
;
2556 value
/= counter
->pipeline_stat
.denominator
;
2559 *((uint64_t *)p
) = value
;
2563 perf_cfg
->vtbl
.bo_unmap(query
->pipeline_stats
.bo
);
2569 gen_perf_get_query_data(struct gen_perf_context
*perf_ctx
,
2570 struct gen_perf_query_object
*query
,
2573 unsigned *bytes_written
)
2575 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2578 switch (query
->queryinfo
->kind
) {
2579 case GEN_PERF_QUERY_TYPE_OA
:
2580 case GEN_PERF_QUERY_TYPE_RAW
:
2581 if (!query
->oa
.results_accumulated
) {
2582 read_gt_frequency(perf_ctx
, query
);
2583 uint32_t *begin_report
= query
->oa
.map
;
2584 uint32_t *end_report
= query
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
2585 gen_perf_query_result_read_frequencies(&query
->oa
.result
,
2589 accumulate_oa_reports(perf_ctx
, query
);
2590 assert(query
->oa
.results_accumulated
);
2592 perf_cfg
->vtbl
.bo_unmap(query
->oa
.bo
);
2593 query
->oa
.map
= NULL
;
2595 if (query
->queryinfo
->kind
== GEN_PERF_QUERY_TYPE_OA
) {
2596 written
= get_oa_counter_data(perf_ctx
, query
, data_size
, (uint8_t *)data
);
2598 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
2600 written
= gen_perf_query_result_write_mdapi((uint8_t *)data
, data_size
,
2601 devinfo
, &query
->oa
.result
,
2602 query
->oa
.gt_frequency
[0],
2603 query
->oa
.gt_frequency
[1]);
2607 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2608 written
= get_pipeline_stats_data(perf_ctx
, query
, data_size
, (uint8_t *)data
);
2612 unreachable("Unknown query type");
2617 *bytes_written
= written
;
2621 gen_perf_dump_query_count(struct gen_perf_context
*perf_ctx
)
2623 DBG("Queries: (Open queries = %d, OA users = %d)\n",
2624 perf_ctx
->n_active_oa_queries
, perf_ctx
->n_oa_users
);
2628 gen_perf_dump_query(struct gen_perf_context
*ctx
,
2629 struct gen_perf_query_object
*obj
,
2630 void *current_batch
)
2632 switch (obj
->queryinfo
->kind
) {
2633 case GEN_PERF_QUERY_TYPE_OA
:
2634 case GEN_PERF_QUERY_TYPE_RAW
:
2635 DBG("BO: %-4s OA data: %-10s %-15s\n",
2636 obj
->oa
.bo
? "yes," : "no,",
2637 gen_perf_is_query_ready(ctx
, obj
, current_batch
) ? "ready," : "not ready,",
2638 obj
->oa
.results_accumulated
? "accumulated" : "not accumulated");
2640 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2642 obj
->pipeline_stats
.bo
? "yes" : "no");
2645 unreachable("Unknown query type");