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 "gen_perf_regs.h"
37 #include "perf/gen_perf_mdapi.h"
38 #include "perf/gen_perf_metrics.h"
40 #include "dev/gen_debug.h"
41 #include "dev/gen_device_info.h"
42 #include "util/bitscan.h"
43 #include "util/mesa-sha1.h"
44 #include "util/u_math.h"
46 #define FILE_DEBUG_FLAG DEBUG_PERFMON
47 #define MI_RPC_BO_SIZE 4096
48 #define MI_FREQ_START_OFFSET_BYTES (3072)
49 #define MI_RPC_BO_END_OFFSET_BYTES (MI_RPC_BO_SIZE / 2)
50 #define MI_FREQ_END_OFFSET_BYTES (3076)
52 #define INTEL_MASK(high, low) (((1u<<((high)-(low)+1))-1)<<(low))
54 #define GEN7_RPSTAT1 0xA01C
55 #define GEN7_RPSTAT1_CURR_GT_FREQ_SHIFT 7
56 #define GEN7_RPSTAT1_CURR_GT_FREQ_MASK INTEL_MASK(13, 7)
57 #define GEN7_RPSTAT1_PREV_GT_FREQ_SHIFT 0
58 #define GEN7_RPSTAT1_PREV_GT_FREQ_MASK INTEL_MASK(6, 0)
60 #define GEN9_RPSTAT0 0xA01C
61 #define GEN9_RPSTAT0_CURR_GT_FREQ_SHIFT 23
62 #define GEN9_RPSTAT0_CURR_GT_FREQ_MASK INTEL_MASK(31, 23)
63 #define GEN9_RPSTAT0_PREV_GT_FREQ_SHIFT 0
64 #define GEN9_RPSTAT0_PREV_GT_FREQ_MASK INTEL_MASK(8, 0)
66 #define GEN6_SO_PRIM_STORAGE_NEEDED 0x2280
67 #define GEN7_SO_PRIM_STORAGE_NEEDED(n) (0x5240 + (n) * 8)
68 #define GEN6_SO_NUM_PRIMS_WRITTEN 0x2288
69 #define GEN7_SO_NUM_PRIMS_WRITTEN(n) (0x5200 + (n) * 8)
71 #define MAP_READ (1 << 0)
72 #define MAP_WRITE (1 << 1)
75 * Periodic OA samples are read() into these buffer structures via the
76 * i915 perf kernel interface and appended to the
77 * perf_ctx->sample_buffers linked list. When we process the
78 * results of an OA metrics query we need to consider all the periodic
79 * samples between the Begin and End MI_REPORT_PERF_COUNT command
82 * 'Periodic' is a simplification as there are other automatic reports
83 * written by the hardware also buffered here.
85 * Considering three queries, A, B and C:
88 * ________________A_________________
90 * | ________B_________ _____C___________
93 * And an illustration of sample buffers read over this time frame:
94 * [HEAD ][ ][ ][ ][ ][ ][ ][ ][TAIL ]
96 * These nodes may hold samples for query A:
97 * [ ][ ][ A ][ A ][ A ][ A ][ A ][ ][ ]
99 * These nodes may hold samples for query B:
100 * [ ][ ][ B ][ B ][ B ][ ][ ][ ][ ]
102 * These nodes may hold samples for query C:
103 * [ ][ ][ ][ ][ ][ C ][ C ][ C ][ ]
105 * The illustration assumes we have an even distribution of periodic
106 * samples so all nodes have the same size plotted against time:
108 * Note, to simplify code, the list is never empty.
110 * With overlapping queries we can see that periodic OA reports may
111 * relate to multiple queries and care needs to be take to keep
112 * track of sample buffers until there are no queries that might
113 * depend on their contents.
115 * We use a node ref counting system where a reference ensures that a
116 * node and all following nodes can't be freed/recycled until the
117 * reference drops to zero.
119 * E.g. with a ref of one here:
120 * [ 0 ][ 0 ][ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
122 * These nodes could be freed or recycled ("reaped"):
125 * These must be preserved until the leading ref drops to zero:
126 * [ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
128 * When a query starts we take a reference on the current tail of
129 * the list, knowing that no already-buffered samples can possibly
130 * relate to the newly-started query. A pointer to this node is
131 * also saved in the query object's ->oa.samples_head.
133 * E.g. starting query A while there are two nodes in .sample_buffers:
134 * ________________A________
138 * ^_______ Add a reference and store pointer to node in
141 * Moving forward to when the B query starts with no new buffer nodes:
142 * (for reference, i915 perf reads() are only done when queries finish)
143 * ________________A_______
148 * ^_______ Add a reference and store pointer to
149 * node in B->oa.samples_head
151 * Once a query is finished, after an OA query has become 'Ready',
152 * once the End OA report has landed and after we we have processed
153 * all the intermediate periodic samples then we drop the
154 * ->oa.samples_head reference we took at the start.
156 * So when the B query has finished we have:
157 * ________________A________
158 * | ______B___________
160 * [ 0 ][ 1 ][ 0 ][ 0 ][ 0 ]
161 * ^_______ Drop B->oa.samples_head reference
163 * We still can't free these due to the A->oa.samples_head ref:
164 * [ 1 ][ 0 ][ 0 ][ 0 ]
166 * When the A query finishes: (note there's a new ref for C's samples_head)
167 * ________________A_________________
171 * [ 0 ][ 0 ][ 0 ][ 0 ][ 1 ][ 0 ][ 0 ]
172 * ^_______ Drop A->oa.samples_head reference
174 * And we can now reap these nodes up to the C->oa.samples_head:
175 * [ X ][ X ][ X ][ X ]
176 * keeping -> [ 1 ][ 0 ][ 0 ]
178 * We reap old sample buffers each time we finish processing an OA
179 * query by iterating the sample_buffers list from the head until we
180 * find a referenced node and stop.
182 * Reaped buffers move to a perfquery.free_sample_buffers list and
183 * when we come to read() we first look to recycle a buffer from the
184 * free_sample_buffers list before allocating a new buffer.
186 struct oa_sample_buf
{
187 struct exec_node link
;
190 uint8_t buf
[I915_PERF_OA_SAMPLE_SIZE
* 10];
191 uint32_t last_timestamp
;
195 * gen representation of a performance query object.
197 * NB: We want to keep this structure relatively lean considering that
198 * applications may expect to allocate enough objects to be able to
199 * query around all draw calls in a frame.
201 struct gen_perf_query_object
203 const struct gen_perf_query_info
*queryinfo
;
205 /* See query->kind to know which state below is in use... */
210 * BO containing OA counter snapshots at query Begin/End time.
215 * Address of mapped of @bo
220 * The MI_REPORT_PERF_COUNT command lets us specify a unique
221 * ID that will be reflected in the resulting OA report
222 * that's written by the GPU. This is the ID we're expecting
223 * in the begin report and the the end report should be
224 * @begin_report_id + 1.
229 * Reference the head of the brw->perfquery.sample_buffers
230 * list at the time that the query started (so we only need
231 * to look at nodes after this point when looking for samples
232 * related to this query)
234 * (See struct brw_oa_sample_buf description for more details)
236 struct exec_node
*samples_head
;
239 * false while in the unaccumulated_elements list, and set to
240 * true when the final, end MI_RPC snapshot has been
243 bool results_accumulated
;
246 * Frequency of the GT at begin and end of the query.
248 uint64_t gt_frequency
[2];
251 * Accumulated OA results between begin and end of the query.
253 struct gen_perf_query_result result
;
258 * BO containing starting and ending snapshots for the
259 * statistics counters.
266 struct gen_perf_context
{
267 struct gen_perf_config
*perf
;
269 void * ctx
; /* driver context (eg, brw_context) */
271 const struct gen_device_info
*devinfo
;
276 /* The i915 perf stream we open to setup + enable the OA counters */
279 /* An i915 perf stream fd gives exclusive access to the OA unit that will
280 * report counter snapshots for a specific counter set/profile in a
281 * specific layout/format so we can only start OA queries that are
282 * compatible with the currently open fd...
284 int current_oa_metrics_set_id
;
285 int current_oa_format
;
287 /* List of buffers containing OA reports */
288 struct exec_list sample_buffers
;
290 /* Cached list of empty sample buffers */
291 struct exec_list free_sample_buffers
;
293 int n_active_oa_queries
;
294 int n_active_pipeline_stats_queries
;
296 /* The number of queries depending on running OA counters which
297 * extends beyond brw_end_perf_query() since we need to wait until
298 * the last MI_RPC command has parsed by the GPU.
300 * Accurate accounting is important here as emitting an
301 * MI_REPORT_PERF_COUNT command while the OA unit is disabled will
302 * effectively hang the gpu.
306 /* To help catch an spurious problem with the hardware or perf
307 * forwarding samples, we emit each MI_REPORT_PERF_COUNT command
308 * with a unique ID that we can explicitly check for...
310 int next_query_start_report_id
;
313 * An array of queries whose results haven't yet been assembled
314 * based on the data in buffer objects.
316 * These may be active, or have already ended. However, the
317 * results have not been requested.
319 struct gen_perf_query_object
**unaccumulated
;
320 int unaccumulated_elements
;
321 int unaccumulated_array_size
;
323 /* The total number of query objects so we can relinquish
324 * our exclusive access to perf if the application deletes
325 * all of its objects. (NB: We only disable perf while
326 * there are no active queries)
328 int n_query_instances
;
331 const struct gen_perf_query_info
*
332 gen_perf_query_info(const struct gen_perf_query_object
*query
)
334 return query
->queryinfo
;
337 struct gen_perf_context
*
338 gen_perf_new_context(void *parent
)
340 struct gen_perf_context
*ctx
= rzalloc(parent
, struct gen_perf_context
);
342 fprintf(stderr
, "%s: failed to alloc context\n", __func__
);
346 struct gen_perf_config
*
347 gen_perf_config(struct gen_perf_context
*ctx
)
352 struct gen_perf_query_object
*
353 gen_perf_new_query(struct gen_perf_context
*perf_ctx
, unsigned query_index
)
355 const struct gen_perf_query_info
*query
=
356 &perf_ctx
->perf
->queries
[query_index
];
357 struct gen_perf_query_object
*obj
=
358 calloc(1, sizeof(struct gen_perf_query_object
));
363 obj
->queryinfo
= query
;
365 perf_ctx
->n_query_instances
++;
370 gen_perf_active_queries(struct gen_perf_context
*perf_ctx
,
371 const struct gen_perf_query_info
*query
)
373 assert(perf_ctx
->n_active_oa_queries
== 0 || perf_ctx
->n_active_pipeline_stats_queries
== 0);
375 switch (query
->kind
) {
376 case GEN_PERF_QUERY_TYPE_OA
:
377 case GEN_PERF_QUERY_TYPE_RAW
:
378 return perf_ctx
->n_active_oa_queries
;
381 case GEN_PERF_QUERY_TYPE_PIPELINE
:
382 return perf_ctx
->n_active_pipeline_stats_queries
;
386 unreachable("Unknown query type");
391 static inline uint64_t to_user_pointer(void *ptr
)
393 return (uintptr_t) ptr
;
397 get_sysfs_dev_dir(struct gen_perf_config
*perf
, int fd
)
402 struct dirent
*drm_entry
;
405 perf
->sysfs_dev_dir
[0] = '\0';
407 if (fstat(fd
, &sb
)) {
408 DBG("Failed to stat DRM fd\n");
412 maj
= major(sb
.st_rdev
);
413 min
= minor(sb
.st_rdev
);
415 if (!S_ISCHR(sb
.st_mode
)) {
416 DBG("DRM fd is not a character device as expected\n");
420 len
= snprintf(perf
->sysfs_dev_dir
,
421 sizeof(perf
->sysfs_dev_dir
),
422 "/sys/dev/char/%d:%d/device/drm", maj
, min
);
423 if (len
< 0 || len
>= sizeof(perf
->sysfs_dev_dir
)) {
424 DBG("Failed to concatenate sysfs path to drm device\n");
428 drmdir
= opendir(perf
->sysfs_dev_dir
);
430 DBG("Failed to open %s: %m\n", perf
->sysfs_dev_dir
);
434 while ((drm_entry
= readdir(drmdir
))) {
435 if ((drm_entry
->d_type
== DT_DIR
||
436 drm_entry
->d_type
== DT_LNK
) &&
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
;
553 if ((metric_entry
->d_type
!= DT_DIR
&&
554 metric_entry
->d_type
!= DT_LNK
) ||
555 metric_entry
->d_name
[0] == '.')
558 DBG("metric set: %s\n", metric_entry
->d_name
);
559 entry
= _mesa_hash_table_search(perf
->oa_metrics_table
,
560 metric_entry
->d_name
);
563 if (!gen_perf_load_metric_id(perf
, metric_entry
->d_name
, &id
)) {
564 DBG("Failed to read metric set id from %s: %m", buf
);
568 register_oa_config(perf
, (const struct gen_perf_query_info
*)entry
->data
, id
);
570 DBG("metric set not known by mesa (skipping)\n");
573 closedir(metricsdir
);
577 kernel_has_dynamic_config_support(struct gen_perf_config
*perf
, int fd
)
579 uint64_t invalid_config_id
= UINT64_MAX
;
581 return gen_ioctl(fd
, DRM_IOCTL_I915_PERF_REMOVE_CONFIG
,
582 &invalid_config_id
) < 0 && errno
== ENOENT
;
586 i915_query_items(struct gen_perf_config
*perf
, int fd
,
587 struct drm_i915_query_item
*items
, uint32_t n_items
)
589 struct drm_i915_query q
= {
590 .num_items
= n_items
,
591 .items_ptr
= to_user_pointer(items
),
593 return gen_ioctl(fd
, DRM_IOCTL_I915_QUERY
, &q
);
597 i915_query_perf_config_supported(struct gen_perf_config
*perf
, int fd
)
599 struct drm_i915_query_item item
= {
600 .query_id
= DRM_I915_QUERY_PERF_CONFIG
,
601 .flags
= DRM_I915_QUERY_PERF_CONFIG_LIST
,
604 return i915_query_items(perf
, fd
, &item
, 1) == 0 && item
.length
> 0;
608 i915_query_perf_config_data(struct gen_perf_config
*perf
,
609 int fd
, const char *guid
,
610 struct drm_i915_perf_oa_config
*config
)
613 struct drm_i915_query_perf_config query
;
614 struct drm_i915_perf_oa_config config
;
616 struct drm_i915_query_item item
= {
617 .query_id
= DRM_I915_QUERY_PERF_CONFIG
,
618 .flags
= DRM_I915_QUERY_PERF_CONFIG_DATA_FOR_UUID
,
619 .data_ptr
= to_user_pointer(&item_data
),
620 .length
= sizeof(item_data
),
623 memset(&item_data
, 0, sizeof(item_data
));
624 memcpy(item_data
.query
.uuid
, guid
, sizeof(item_data
.query
.uuid
));
625 memcpy(&item_data
.config
, config
, sizeof(item_data
.config
));
627 if (!(i915_query_items(perf
, fd
, &item
, 1) == 0 && item
.length
> 0))
630 memcpy(config
, &item_data
.config
, sizeof(item_data
.config
));
636 gen_perf_load_metric_id(struct gen_perf_config
*perf_cfg
,
640 char config_path
[280];
642 snprintf(config_path
, sizeof(config_path
), "%s/metrics/%s/id",
643 perf_cfg
->sysfs_dev_dir
, guid
);
645 /* Don't recreate already loaded configs. */
646 return read_file_uint64(config_path
, metric_id
);
650 i915_add_config(struct gen_perf_config
*perf
, int fd
,
651 const struct gen_perf_registers
*config
,
654 struct drm_i915_perf_oa_config i915_config
= { 0, };
656 memcpy(i915_config
.uuid
, guid
, sizeof(i915_config
.uuid
));
658 i915_config
.n_mux_regs
= config
->n_mux_regs
;
659 i915_config
.mux_regs_ptr
= to_user_pointer(config
->mux_regs
);
661 i915_config
.n_boolean_regs
= config
->n_b_counter_regs
;
662 i915_config
.boolean_regs_ptr
= to_user_pointer(config
->b_counter_regs
);
664 i915_config
.n_flex_regs
= config
->n_flex_regs
;
665 i915_config
.flex_regs_ptr
= to_user_pointer(config
->flex_regs
);
667 int ret
= gen_ioctl(fd
, DRM_IOCTL_I915_PERF_ADD_CONFIG
, &i915_config
);
668 return ret
> 0 ? ret
: 0;
672 init_oa_configs(struct gen_perf_config
*perf
, int fd
)
674 hash_table_foreach(perf
->oa_metrics_table
, entry
) {
675 const struct gen_perf_query_info
*query
= entry
->data
;
678 if (gen_perf_load_metric_id(perf
, query
->guid
, &config_id
)) {
679 DBG("metric set: %s (already loaded)\n", query
->guid
);
680 register_oa_config(perf
, query
, config_id
);
684 int ret
= i915_add_config(perf
, fd
, &query
->config
, query
->guid
);
686 DBG("Failed to load \"%s\" (%s) metrics set in kernel: %s\n",
687 query
->name
, query
->guid
, strerror(errno
));
691 register_oa_config(perf
, query
, ret
);
692 DBG("metric set: %s (added)\n", query
->guid
);
697 compute_topology_builtins(struct gen_perf_config
*perf
,
698 const struct gen_device_info
*devinfo
)
700 perf
->sys_vars
.slice_mask
= devinfo
->slice_masks
;
701 perf
->sys_vars
.n_eu_slices
= devinfo
->num_slices
;
703 for (int i
= 0; i
< sizeof(devinfo
->subslice_masks
[i
]); i
++) {
704 perf
->sys_vars
.n_eu_sub_slices
+=
705 __builtin_popcount(devinfo
->subslice_masks
[i
]);
708 for (int i
= 0; i
< sizeof(devinfo
->eu_masks
); i
++)
709 perf
->sys_vars
.n_eus
+= __builtin_popcount(devinfo
->eu_masks
[i
]);
711 perf
->sys_vars
.eu_threads_count
= devinfo
->num_thread_per_eu
;
713 /* The subslice mask builtin contains bits for all slices. Prior to Gen11
714 * it had groups of 3bits for each slice, on Gen11 it's 8bits for each
717 * Ideally equations would be updated to have a slice/subslice query
720 perf
->sys_vars
.subslice_mask
= 0;
722 int bits_per_subslice
= devinfo
->gen
== 11 ? 8 : 3;
724 for (int s
= 0; s
< util_last_bit(devinfo
->slice_masks
); s
++) {
725 for (int ss
= 0; ss
< (devinfo
->subslice_slice_stride
* 8); ss
++) {
726 if (gen_device_info_subslice_available(devinfo
, s
, ss
))
727 perf
->sys_vars
.subslice_mask
|= 1ULL << (s
* bits_per_subslice
+ ss
);
733 init_oa_sys_vars(struct gen_perf_config
*perf
, const struct gen_device_info
*devinfo
)
735 uint64_t min_freq_mhz
= 0, max_freq_mhz
= 0;
737 if (!read_sysfs_drm_device_file_uint64(perf
, "gt_min_freq_mhz", &min_freq_mhz
))
740 if (!read_sysfs_drm_device_file_uint64(perf
, "gt_max_freq_mhz", &max_freq_mhz
))
743 memset(&perf
->sys_vars
, 0, sizeof(perf
->sys_vars
));
744 perf
->sys_vars
.gt_min_freq
= min_freq_mhz
* 1000000;
745 perf
->sys_vars
.gt_max_freq
= max_freq_mhz
* 1000000;
746 perf
->sys_vars
.timestamp_frequency
= devinfo
->timestamp_frequency
;
747 perf
->sys_vars
.revision
= devinfo
->revision
;
748 compute_topology_builtins(perf
, devinfo
);
753 typedef void (*perf_register_oa_queries_t
)(struct gen_perf_config
*);
755 static perf_register_oa_queries_t
756 get_register_queries_function(const struct gen_device_info
*devinfo
)
758 if (devinfo
->is_haswell
)
759 return gen_oa_register_queries_hsw
;
760 if (devinfo
->is_cherryview
)
761 return gen_oa_register_queries_chv
;
762 if (devinfo
->is_broadwell
)
763 return gen_oa_register_queries_bdw
;
764 if (devinfo
->is_broxton
)
765 return gen_oa_register_queries_bxt
;
766 if (devinfo
->is_skylake
) {
767 if (devinfo
->gt
== 2)
768 return gen_oa_register_queries_sklgt2
;
769 if (devinfo
->gt
== 3)
770 return gen_oa_register_queries_sklgt3
;
771 if (devinfo
->gt
== 4)
772 return gen_oa_register_queries_sklgt4
;
774 if (devinfo
->is_kabylake
) {
775 if (devinfo
->gt
== 2)
776 return gen_oa_register_queries_kblgt2
;
777 if (devinfo
->gt
== 3)
778 return gen_oa_register_queries_kblgt3
;
780 if (devinfo
->is_geminilake
)
781 return gen_oa_register_queries_glk
;
782 if (devinfo
->is_coffeelake
) {
783 if (devinfo
->gt
== 2)
784 return gen_oa_register_queries_cflgt2
;
785 if (devinfo
->gt
== 3)
786 return gen_oa_register_queries_cflgt3
;
788 if (devinfo
->is_cannonlake
)
789 return gen_oa_register_queries_cnl
;
790 if (devinfo
->gen
== 11) {
791 if (devinfo
->is_elkhartlake
)
792 return gen_oa_register_queries_lkf
;
793 return gen_oa_register_queries_icl
;
795 if (devinfo
->gen
== 12)
796 return gen_oa_register_queries_tgl
;
802 add_stat_reg(struct gen_perf_query_info
*query
, uint32_t reg
,
803 uint32_t numerator
, uint32_t denominator
,
804 const char *name
, const char *description
)
806 struct gen_perf_query_counter
*counter
;
808 assert(query
->n_counters
< query
->max_counters
);
810 counter
= &query
->counters
[query
->n_counters
];
811 counter
->name
= name
;
812 counter
->desc
= description
;
813 counter
->type
= GEN_PERF_COUNTER_TYPE_RAW
;
814 counter
->data_type
= GEN_PERF_COUNTER_DATA_TYPE_UINT64
;
815 counter
->offset
= sizeof(uint64_t) * query
->n_counters
;
816 counter
->pipeline_stat
.reg
= reg
;
817 counter
->pipeline_stat
.numerator
= numerator
;
818 counter
->pipeline_stat
.denominator
= denominator
;
824 add_basic_stat_reg(struct gen_perf_query_info
*query
,
825 uint32_t reg
, const char *name
)
827 add_stat_reg(query
, reg
, 1, 1, name
, name
);
831 load_pipeline_statistic_metrics(struct gen_perf_config
*perf_cfg
,
832 const struct gen_device_info
*devinfo
)
834 struct gen_perf_query_info
*query
=
835 append_query_info(perf_cfg
, MAX_STAT_COUNTERS
);
837 query
->kind
= GEN_PERF_QUERY_TYPE_PIPELINE
;
838 query
->name
= "Pipeline Statistics Registers";
840 add_basic_stat_reg(query
, IA_VERTICES_COUNT
,
841 "N vertices submitted");
842 add_basic_stat_reg(query
, IA_PRIMITIVES_COUNT
,
843 "N primitives submitted");
844 add_basic_stat_reg(query
, VS_INVOCATION_COUNT
,
845 "N vertex shader invocations");
847 if (devinfo
->gen
== 6) {
848 add_stat_reg(query
, GEN6_SO_PRIM_STORAGE_NEEDED
, 1, 1,
849 "SO_PRIM_STORAGE_NEEDED",
850 "N geometry shader stream-out primitives (total)");
851 add_stat_reg(query
, GEN6_SO_NUM_PRIMS_WRITTEN
, 1, 1,
852 "SO_NUM_PRIMS_WRITTEN",
853 "N geometry shader stream-out primitives (written)");
855 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(0), 1, 1,
856 "SO_PRIM_STORAGE_NEEDED (Stream 0)",
857 "N stream-out (stream 0) primitives (total)");
858 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(1), 1, 1,
859 "SO_PRIM_STORAGE_NEEDED (Stream 1)",
860 "N stream-out (stream 1) primitives (total)");
861 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(2), 1, 1,
862 "SO_PRIM_STORAGE_NEEDED (Stream 2)",
863 "N stream-out (stream 2) primitives (total)");
864 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(3), 1, 1,
865 "SO_PRIM_STORAGE_NEEDED (Stream 3)",
866 "N stream-out (stream 3) primitives (total)");
867 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(0), 1, 1,
868 "SO_NUM_PRIMS_WRITTEN (Stream 0)",
869 "N stream-out (stream 0) primitives (written)");
870 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(1), 1, 1,
871 "SO_NUM_PRIMS_WRITTEN (Stream 1)",
872 "N stream-out (stream 1) primitives (written)");
873 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(2), 1, 1,
874 "SO_NUM_PRIMS_WRITTEN (Stream 2)",
875 "N stream-out (stream 2) primitives (written)");
876 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(3), 1, 1,
877 "SO_NUM_PRIMS_WRITTEN (Stream 3)",
878 "N stream-out (stream 3) primitives (written)");
881 add_basic_stat_reg(query
, HS_INVOCATION_COUNT
,
882 "N TCS shader invocations");
883 add_basic_stat_reg(query
, DS_INVOCATION_COUNT
,
884 "N TES shader invocations");
886 add_basic_stat_reg(query
, GS_INVOCATION_COUNT
,
887 "N geometry shader invocations");
888 add_basic_stat_reg(query
, GS_PRIMITIVES_COUNT
,
889 "N geometry shader primitives emitted");
891 add_basic_stat_reg(query
, CL_INVOCATION_COUNT
,
892 "N primitives entering clipping");
893 add_basic_stat_reg(query
, CL_PRIMITIVES_COUNT
,
894 "N primitives leaving clipping");
896 if (devinfo
->is_haswell
|| devinfo
->gen
== 8) {
897 add_stat_reg(query
, PS_INVOCATION_COUNT
, 1, 4,
898 "N fragment shader invocations",
899 "N fragment shader invocations");
901 add_basic_stat_reg(query
, PS_INVOCATION_COUNT
,
902 "N fragment shader invocations");
905 add_basic_stat_reg(query
, PS_DEPTH_COUNT
,
906 "N z-pass fragments");
908 if (devinfo
->gen
>= 7) {
909 add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
910 "N compute shader invocations");
913 query
->data_size
= sizeof(uint64_t) * query
->n_counters
;
917 load_oa_metrics(struct gen_perf_config
*perf
, int fd
,
918 const struct gen_device_info
*devinfo
)
920 perf_register_oa_queries_t oa_register
= get_register_queries_function(devinfo
);
921 bool i915_perf_oa_available
= false;
924 perf
->i915_query_supported
= i915_query_perf_config_supported(perf
, fd
);
926 /* The existence of this sysctl parameter implies the kernel supports
927 * the i915 perf interface.
929 if (stat("/proc/sys/dev/i915/perf_stream_paranoid", &sb
) == 0) {
931 /* If _paranoid == 1 then on Gen8+ we won't be able to access OA
932 * metrics unless running as root.
934 if (devinfo
->is_haswell
)
935 i915_perf_oa_available
= true;
937 uint64_t paranoid
= 1;
939 read_file_uint64("/proc/sys/dev/i915/perf_stream_paranoid", ¶noid
);
941 if (paranoid
== 0 || geteuid() == 0)
942 i915_perf_oa_available
= true;
946 if (!i915_perf_oa_available
||
948 !get_sysfs_dev_dir(perf
, fd
) ||
949 !init_oa_sys_vars(perf
, devinfo
))
952 perf
->oa_metrics_table
=
953 _mesa_hash_table_create(perf
, _mesa_key_hash_string
,
954 _mesa_key_string_equal
);
956 /* Index all the metric sets mesa knows about before looking to see what
957 * the kernel is advertising.
961 if (likely((INTEL_DEBUG
& DEBUG_NO_OACONFIG
) == 0) &&
962 kernel_has_dynamic_config_support(perf
, fd
))
963 init_oa_configs(perf
, fd
);
965 enumerate_sysfs_metrics(perf
);
970 struct gen_perf_registers
*
971 gen_perf_load_configuration(struct gen_perf_config
*perf_cfg
, int fd
, const char *guid
)
973 if (!perf_cfg
->i915_query_supported
)
976 struct drm_i915_perf_oa_config i915_config
= { 0, };
977 if (!i915_query_perf_config_data(perf_cfg
, fd
, guid
, &i915_config
))
980 struct gen_perf_registers
*config
= rzalloc(NULL
, struct gen_perf_registers
);
981 config
->n_flex_regs
= i915_config
.n_flex_regs
;
982 config
->flex_regs
= rzalloc_array(config
, struct gen_perf_query_register_prog
, config
->n_flex_regs
);
983 config
->n_mux_regs
= i915_config
.n_mux_regs
;
984 config
->mux_regs
= rzalloc_array(config
, struct gen_perf_query_register_prog
, config
->n_mux_regs
);
985 config
->n_b_counter_regs
= i915_config
.n_boolean_regs
;
986 config
->b_counter_regs
= rzalloc_array(config
, struct gen_perf_query_register_prog
, config
->n_b_counter_regs
);
989 * struct gen_perf_query_register_prog maps exactly to the tuple of
990 * (register offset, register value) returned by the i915.
992 i915_config
.flex_regs_ptr
= to_user_pointer(config
->flex_regs
);
993 i915_config
.mux_regs_ptr
= to_user_pointer(config
->mux_regs
);
994 i915_config
.boolean_regs_ptr
= to_user_pointer(config
->b_counter_regs
);
995 if (!i915_query_perf_config_data(perf_cfg
, fd
, guid
, &i915_config
)) {
1004 gen_perf_store_configuration(struct gen_perf_config
*perf_cfg
, int fd
,
1005 const struct gen_perf_registers
*config
,
1009 return i915_add_config(perf_cfg
, fd
, config
, guid
);
1011 struct mesa_sha1 sha1_ctx
;
1012 _mesa_sha1_init(&sha1_ctx
);
1014 if (config
->flex_regs
) {
1015 _mesa_sha1_update(&sha1_ctx
, config
->flex_regs
,
1016 sizeof(config
->flex_regs
[0]) *
1017 config
->n_flex_regs
);
1019 if (config
->mux_regs
) {
1020 _mesa_sha1_update(&sha1_ctx
, config
->mux_regs
,
1021 sizeof(config
->mux_regs
[0]) *
1022 config
->n_mux_regs
);
1024 if (config
->b_counter_regs
) {
1025 _mesa_sha1_update(&sha1_ctx
, config
->b_counter_regs
,
1026 sizeof(config
->b_counter_regs
[0]) *
1027 config
->n_b_counter_regs
);
1031 _mesa_sha1_final(&sha1_ctx
, hash
);
1033 char formatted_hash
[41];
1034 _mesa_sha1_format(formatted_hash
, hash
);
1036 char generated_guid
[37];
1037 snprintf(generated_guid
, sizeof(generated_guid
),
1038 "%.8s-%.4s-%.4s-%.4s-%.12s",
1039 &formatted_hash
[0], &formatted_hash
[8],
1040 &formatted_hash
[8 + 4], &formatted_hash
[8 + 4 + 4],
1041 &formatted_hash
[8 + 4 + 4 + 4]);
1043 /* Check if already present. */
1045 if (gen_perf_load_metric_id(perf_cfg
, generated_guid
, &id
))
1048 return i915_add_config(perf_cfg
, fd
, config
, generated_guid
);
1051 /* Accumulate 32bits OA counters */
1053 accumulate_uint32(const uint32_t *report0
,
1054 const uint32_t *report1
,
1055 uint64_t *accumulator
)
1057 *accumulator
+= (uint32_t)(*report1
- *report0
);
1060 /* Accumulate 40bits OA counters */
1062 accumulate_uint40(int a_index
,
1063 const uint32_t *report0
,
1064 const uint32_t *report1
,
1065 uint64_t *accumulator
)
1067 const uint8_t *high_bytes0
= (uint8_t *)(report0
+ 40);
1068 const uint8_t *high_bytes1
= (uint8_t *)(report1
+ 40);
1069 uint64_t high0
= (uint64_t)(high_bytes0
[a_index
]) << 32;
1070 uint64_t high1
= (uint64_t)(high_bytes1
[a_index
]) << 32;
1071 uint64_t value0
= report0
[a_index
+ 4] | high0
;
1072 uint64_t value1
= report1
[a_index
+ 4] | high1
;
1075 if (value0
> value1
)
1076 delta
= (1ULL << 40) + value1
- value0
;
1078 delta
= value1
- value0
;
1080 *accumulator
+= delta
;
1084 gen8_read_report_clock_ratios(const uint32_t *report
,
1085 uint64_t *slice_freq_hz
,
1086 uint64_t *unslice_freq_hz
)
1088 /* The lower 16bits of the RPT_ID field of the OA reports contains a
1089 * snapshot of the bits coming from the RP_FREQ_NORMAL register and is
1090 * divided this way :
1092 * RPT_ID[31:25]: RP_FREQ_NORMAL[20:14] (low squashed_slice_clock_frequency)
1093 * RPT_ID[10:9]: RP_FREQ_NORMAL[22:21] (high squashed_slice_clock_frequency)
1094 * RPT_ID[8:0]: RP_FREQ_NORMAL[31:23] (squashed_unslice_clock_frequency)
1096 * RP_FREQ_NORMAL[31:23]: Software Unslice Ratio Request
1097 * Multiple of 33.33MHz 2xclk (16 MHz 1xclk)
1099 * RP_FREQ_NORMAL[22:14]: Software Slice Ratio Request
1100 * Multiple of 33.33MHz 2xclk (16 MHz 1xclk)
1103 uint32_t unslice_freq
= report
[0] & 0x1ff;
1104 uint32_t slice_freq_low
= (report
[0] >> 25) & 0x7f;
1105 uint32_t slice_freq_high
= (report
[0] >> 9) & 0x3;
1106 uint32_t slice_freq
= slice_freq_low
| (slice_freq_high
<< 7);
1108 *slice_freq_hz
= slice_freq
* 16666667ULL;
1109 *unslice_freq_hz
= unslice_freq
* 16666667ULL;
1113 gen_perf_query_result_read_frequencies(struct gen_perf_query_result
*result
,
1114 const struct gen_device_info
*devinfo
,
1115 const uint32_t *start
,
1116 const uint32_t *end
)
1118 /* Slice/Unslice frequency is only available in the OA reports when the
1119 * "Disable OA reports due to clock ratio change" field in
1120 * OA_DEBUG_REGISTER is set to 1. This is how the kernel programs this
1121 * global register (see drivers/gpu/drm/i915/i915_perf.c)
1123 * Documentation says this should be available on Gen9+ but experimentation
1124 * shows that Gen8 reports similar values, so we enable it there too.
1126 if (devinfo
->gen
< 8)
1129 gen8_read_report_clock_ratios(start
,
1130 &result
->slice_frequency
[0],
1131 &result
->unslice_frequency
[0]);
1132 gen8_read_report_clock_ratios(end
,
1133 &result
->slice_frequency
[1],
1134 &result
->unslice_frequency
[1]);
1138 gen_perf_query_result_accumulate(struct gen_perf_query_result
*result
,
1139 const struct gen_perf_query_info
*query
,
1140 const uint32_t *start
,
1141 const uint32_t *end
)
1145 result
->hw_id
= start
[2];
1146 result
->reports_accumulated
++;
1148 switch (query
->oa_format
) {
1149 case I915_OA_FORMAT_A32u40_A4u32_B8_C8
:
1150 accumulate_uint32(start
+ 1, end
+ 1, result
->accumulator
+ idx
++); /* timestamp */
1151 accumulate_uint32(start
+ 3, end
+ 3, result
->accumulator
+ idx
++); /* clock */
1153 /* 32x 40bit A counters... */
1154 for (i
= 0; i
< 32; i
++)
1155 accumulate_uint40(i
, start
, end
, result
->accumulator
+ idx
++);
1157 /* 4x 32bit A counters... */
1158 for (i
= 0; i
< 4; i
++)
1159 accumulate_uint32(start
+ 36 + i
, end
+ 36 + i
, result
->accumulator
+ idx
++);
1161 /* 8x 32bit B counters + 8x 32bit C counters... */
1162 for (i
= 0; i
< 16; i
++)
1163 accumulate_uint32(start
+ 48 + i
, end
+ 48 + i
, result
->accumulator
+ idx
++);
1166 case I915_OA_FORMAT_A45_B8_C8
:
1167 accumulate_uint32(start
+ 1, end
+ 1, result
->accumulator
); /* timestamp */
1169 for (i
= 0; i
< 61; i
++)
1170 accumulate_uint32(start
+ 3 + i
, end
+ 3 + i
, result
->accumulator
+ 1 + i
);
1174 unreachable("Can't accumulate OA counters in unknown format");
1180 gen_perf_query_result_clear(struct gen_perf_query_result
*result
)
1182 memset(result
, 0, sizeof(*result
));
1183 result
->hw_id
= 0xffffffff; /* invalid */
1187 register_mdapi_statistic_query(struct gen_perf_config
*perf_cfg
,
1188 const struct gen_device_info
*devinfo
)
1190 if (!(devinfo
->gen
>= 7 && devinfo
->gen
<= 11))
1193 struct gen_perf_query_info
*query
=
1194 append_query_info(perf_cfg
, MAX_STAT_COUNTERS
);
1196 query
->kind
= GEN_PERF_QUERY_TYPE_PIPELINE
;
1197 query
->name
= "Intel_Raw_Pipeline_Statistics_Query";
1199 /* The order has to match mdapi_pipeline_metrics. */
1200 add_basic_stat_reg(query
, IA_VERTICES_COUNT
,
1201 "N vertices submitted");
1202 add_basic_stat_reg(query
, IA_PRIMITIVES_COUNT
,
1203 "N primitives submitted");
1204 add_basic_stat_reg(query
, VS_INVOCATION_COUNT
,
1205 "N vertex shader invocations");
1206 add_basic_stat_reg(query
, GS_INVOCATION_COUNT
,
1207 "N geometry shader invocations");
1208 add_basic_stat_reg(query
, GS_PRIMITIVES_COUNT
,
1209 "N geometry shader primitives emitted");
1210 add_basic_stat_reg(query
, CL_INVOCATION_COUNT
,
1211 "N primitives entering clipping");
1212 add_basic_stat_reg(query
, CL_PRIMITIVES_COUNT
,
1213 "N primitives leaving clipping");
1214 if (devinfo
->is_haswell
|| devinfo
->gen
== 8) {
1215 add_stat_reg(query
, PS_INVOCATION_COUNT
, 1, 4,
1216 "N fragment shader invocations",
1217 "N fragment shader invocations");
1219 add_basic_stat_reg(query
, PS_INVOCATION_COUNT
,
1220 "N fragment shader invocations");
1222 add_basic_stat_reg(query
, HS_INVOCATION_COUNT
,
1223 "N TCS shader invocations");
1224 add_basic_stat_reg(query
, DS_INVOCATION_COUNT
,
1225 "N TES shader invocations");
1226 if (devinfo
->gen
>= 7) {
1227 add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
1228 "N compute shader invocations");
1231 if (devinfo
->gen
>= 10) {
1232 /* Reuse existing CS invocation register until we can expose this new
1235 add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
1239 query
->data_size
= sizeof(uint64_t) * query
->n_counters
;
1243 fill_mdapi_perf_query_counter(struct gen_perf_query_info
*query
,
1245 uint32_t data_offset
,
1247 enum gen_perf_counter_data_type data_type
)
1249 struct gen_perf_query_counter
*counter
= &query
->counters
[query
->n_counters
];
1251 assert(query
->n_counters
<= query
->max_counters
);
1253 counter
->name
= name
;
1254 counter
->desc
= "Raw counter value";
1255 counter
->type
= GEN_PERF_COUNTER_TYPE_RAW
;
1256 counter
->data_type
= data_type
;
1257 counter
->offset
= data_offset
;
1259 query
->n_counters
++;
1261 assert(counter
->offset
+ gen_perf_query_counter_get_size(counter
) <= query
->data_size
);
1264 #define MDAPI_QUERY_ADD_COUNTER(query, struct_name, field_name, type_name) \
1265 fill_mdapi_perf_query_counter(query, #field_name, \
1266 (uint8_t *) &struct_name.field_name - \
1267 (uint8_t *) &struct_name, \
1268 sizeof(struct_name.field_name), \
1269 GEN_PERF_COUNTER_DATA_TYPE_##type_name)
1270 #define MDAPI_QUERY_ADD_ARRAY_COUNTER(ctx, query, struct_name, field_name, idx, type_name) \
1271 fill_mdapi_perf_query_counter(query, \
1272 ralloc_asprintf(ctx, "%s%i", #field_name, idx), \
1273 (uint8_t *) &struct_name.field_name[idx] - \
1274 (uint8_t *) &struct_name, \
1275 sizeof(struct_name.field_name[0]), \
1276 GEN_PERF_COUNTER_DATA_TYPE_##type_name)
1279 register_mdapi_oa_query(const struct gen_device_info
*devinfo
,
1280 struct gen_perf_config
*perf
)
1282 struct gen_perf_query_info
*query
= NULL
;
1284 /* MDAPI requires different structures for pretty much every generation
1285 * (right now we have definitions for gen 7 to 11).
1287 if (!(devinfo
->gen
>= 7 && devinfo
->gen
<= 11))
1290 switch (devinfo
->gen
) {
1292 query
= append_query_info(perf
, 1 + 45 + 16 + 7);
1293 query
->oa_format
= I915_OA_FORMAT_A45_B8_C8
;
1295 struct gen7_mdapi_metrics metric_data
;
1296 query
->data_size
= sizeof(metric_data
);
1298 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, TotalTime
, UINT64
);
1299 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.ACounters
); i
++) {
1300 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1301 metric_data
, ACounters
, i
, UINT64
);
1303 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.NOACounters
); i
++) {
1304 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1305 metric_data
, NOACounters
, i
, UINT64
);
1307 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter1
, UINT64
);
1308 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter2
, UINT64
);
1309 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SplitOccured
, BOOL32
);
1310 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequencyChanged
, BOOL32
);
1311 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequency
, UINT64
);
1312 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportId
, UINT32
);
1313 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportsCount
, UINT32
);
1317 query
= append_query_info(perf
, 2 + 36 + 16 + 16);
1318 query
->oa_format
= I915_OA_FORMAT_A32u40_A4u32_B8_C8
;
1320 struct gen8_mdapi_metrics metric_data
;
1321 query
->data_size
= sizeof(metric_data
);
1323 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, TotalTime
, UINT64
);
1324 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, GPUTicks
, UINT64
);
1325 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.OaCntr
); i
++) {
1326 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1327 metric_data
, OaCntr
, i
, UINT64
);
1329 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.NoaCntr
); i
++) {
1330 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1331 metric_data
, NoaCntr
, i
, UINT64
);
1333 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, BeginTimestamp
, UINT64
);
1334 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved1
, UINT64
);
1335 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved2
, UINT64
);
1336 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved3
, UINT32
);
1337 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, OverrunOccured
, BOOL32
);
1338 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerUser
, UINT64
);
1339 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerDriver
, UINT64
);
1340 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SliceFrequency
, UINT64
);
1341 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, UnsliceFrequency
, UINT64
);
1342 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter1
, UINT64
);
1343 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter2
, UINT64
);
1344 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SplitOccured
, BOOL32
);
1345 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequencyChanged
, BOOL32
);
1346 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequency
, UINT64
);
1347 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportId
, UINT32
);
1348 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportsCount
, UINT32
);
1354 query
= append_query_info(perf
, 2 + 36 + 16 + 16 + 16 + 2);
1355 query
->oa_format
= I915_OA_FORMAT_A32u40_A4u32_B8_C8
;
1357 struct gen9_mdapi_metrics metric_data
;
1358 query
->data_size
= sizeof(metric_data
);
1360 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, TotalTime
, UINT64
);
1361 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, GPUTicks
, UINT64
);
1362 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.OaCntr
); i
++) {
1363 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1364 metric_data
, OaCntr
, i
, UINT64
);
1366 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.NoaCntr
); i
++) {
1367 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1368 metric_data
, NoaCntr
, i
, UINT64
);
1370 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, BeginTimestamp
, UINT64
);
1371 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved1
, UINT64
);
1372 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved2
, UINT64
);
1373 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved3
, UINT32
);
1374 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, OverrunOccured
, BOOL32
);
1375 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerUser
, UINT64
);
1376 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerDriver
, UINT64
);
1377 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SliceFrequency
, UINT64
);
1378 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, UnsliceFrequency
, UINT64
);
1379 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter1
, UINT64
);
1380 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter2
, UINT64
);
1381 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SplitOccured
, BOOL32
);
1382 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequencyChanged
, BOOL32
);
1383 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequency
, UINT64
);
1384 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportId
, UINT32
);
1385 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportsCount
, UINT32
);
1386 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.UserCntr
); i
++) {
1387 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1388 metric_data
, UserCntr
, i
, UINT64
);
1390 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, UserCntrCfgId
, UINT32
);
1391 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved4
, UINT32
);
1395 unreachable("Unsupported gen");
1399 query
->kind
= GEN_PERF_QUERY_TYPE_RAW
;
1400 query
->name
= "Intel_Raw_Hardware_Counters_Set_0_Query";
1401 query
->guid
= GEN_PERF_QUERY_GUID_MDAPI
;
1404 /* Accumulation buffer offsets copied from an actual query... */
1405 const struct gen_perf_query_info
*copy_query
=
1408 query
->gpu_time_offset
= copy_query
->gpu_time_offset
;
1409 query
->gpu_clock_offset
= copy_query
->gpu_clock_offset
;
1410 query
->a_offset
= copy_query
->a_offset
;
1411 query
->b_offset
= copy_query
->b_offset
;
1412 query
->c_offset
= copy_query
->c_offset
;
1417 get_metric_id(struct gen_perf_config
*perf
,
1418 const struct gen_perf_query_info
*query
)
1420 /* These queries are know not to ever change, their config ID has been
1421 * loaded upon the first query creation. No need to look them up again.
1423 if (query
->kind
== GEN_PERF_QUERY_TYPE_OA
)
1424 return query
->oa_metrics_set_id
;
1426 assert(query
->kind
== GEN_PERF_QUERY_TYPE_RAW
);
1428 /* Raw queries can be reprogrammed up by an external application/library.
1429 * When a raw query is used for the first time it's id is set to a value !=
1430 * 0. When it stops being used the id returns to 0. No need to reload the
1431 * ID when it's already loaded.
1433 if (query
->oa_metrics_set_id
!= 0) {
1434 DBG("Raw query '%s' guid=%s using cached ID: %"PRIu64
"\n",
1435 query
->name
, query
->guid
, query
->oa_metrics_set_id
);
1436 return query
->oa_metrics_set_id
;
1439 struct gen_perf_query_info
*raw_query
= (struct gen_perf_query_info
*)query
;
1440 if (!gen_perf_load_metric_id(perf
, query
->guid
,
1441 &raw_query
->oa_metrics_set_id
)) {
1442 DBG("Unable to read query guid=%s ID, falling back to test config\n", query
->guid
);
1443 raw_query
->oa_metrics_set_id
= 1ULL;
1445 DBG("Raw query '%s'guid=%s loaded ID: %"PRIu64
"\n",
1446 query
->name
, query
->guid
, query
->oa_metrics_set_id
);
1448 return query
->oa_metrics_set_id
;
1451 static struct oa_sample_buf
*
1452 get_free_sample_buf(struct gen_perf_context
*perf_ctx
)
1454 struct exec_node
*node
= exec_list_pop_head(&perf_ctx
->free_sample_buffers
);
1455 struct oa_sample_buf
*buf
;
1458 buf
= exec_node_data(struct oa_sample_buf
, node
, link
);
1460 buf
= ralloc_size(perf_ctx
->perf
, sizeof(*buf
));
1462 exec_node_init(&buf
->link
);
1471 reap_old_sample_buffers(struct gen_perf_context
*perf_ctx
)
1473 struct exec_node
*tail_node
=
1474 exec_list_get_tail(&perf_ctx
->sample_buffers
);
1475 struct oa_sample_buf
*tail_buf
=
1476 exec_node_data(struct oa_sample_buf
, tail_node
, link
);
1478 /* Remove all old, unreferenced sample buffers walking forward from
1479 * the head of the list, except always leave at least one node in
1480 * the list so we always have a node to reference when we Begin
1483 foreach_list_typed_safe(struct oa_sample_buf
, buf
, link
,
1484 &perf_ctx
->sample_buffers
)
1486 if (buf
->refcount
== 0 && buf
!= tail_buf
) {
1487 exec_node_remove(&buf
->link
);
1488 exec_list_push_head(&perf_ctx
->free_sample_buffers
, &buf
->link
);
1495 free_sample_bufs(struct gen_perf_context
*perf_ctx
)
1497 foreach_list_typed_safe(struct oa_sample_buf
, buf
, link
,
1498 &perf_ctx
->free_sample_buffers
)
1501 exec_list_make_empty(&perf_ctx
->free_sample_buffers
);
1504 /******************************************************************************/
1507 * Emit MI_STORE_REGISTER_MEM commands to capture all of the
1508 * pipeline statistics for the performance query object.
1511 snapshot_statistics_registers(void *context
,
1512 struct gen_perf_config
*perf
,
1513 struct gen_perf_query_object
*obj
,
1514 uint32_t offset_in_bytes
)
1516 const struct gen_perf_query_info
*query
= obj
->queryinfo
;
1517 const int n_counters
= query
->n_counters
;
1519 for (int i
= 0; i
< n_counters
; i
++) {
1520 const struct gen_perf_query_counter
*counter
= &query
->counters
[i
];
1522 assert(counter
->data_type
== GEN_PERF_COUNTER_DATA_TYPE_UINT64
);
1524 perf
->vtbl
.store_register_mem64(context
, obj
->pipeline_stats
.bo
,
1525 counter
->pipeline_stat
.reg
,
1526 offset_in_bytes
+ i
* sizeof(uint64_t));
1531 gen_perf_close(struct gen_perf_context
*perfquery
,
1532 const struct gen_perf_query_info
*query
)
1534 if (perfquery
->oa_stream_fd
!= -1) {
1535 close(perfquery
->oa_stream_fd
);
1536 perfquery
->oa_stream_fd
= -1;
1538 if (query
->kind
== GEN_PERF_QUERY_TYPE_RAW
) {
1539 struct gen_perf_query_info
*raw_query
=
1540 (struct gen_perf_query_info
*) query
;
1541 raw_query
->oa_metrics_set_id
= 0;
1546 gen_perf_open(struct gen_perf_context
*perf_ctx
,
1549 int period_exponent
,
1553 uint64_t properties
[] = {
1554 /* Single context sampling */
1555 DRM_I915_PERF_PROP_CTX_HANDLE
, ctx_id
,
1557 /* Include OA reports in samples */
1558 DRM_I915_PERF_PROP_SAMPLE_OA
, true,
1560 /* OA unit configuration */
1561 DRM_I915_PERF_PROP_OA_METRICS_SET
, metrics_set_id
,
1562 DRM_I915_PERF_PROP_OA_FORMAT
, report_format
,
1563 DRM_I915_PERF_PROP_OA_EXPONENT
, period_exponent
,
1565 struct drm_i915_perf_open_param param
= {
1566 .flags
= I915_PERF_FLAG_FD_CLOEXEC
|
1567 I915_PERF_FLAG_FD_NONBLOCK
|
1568 I915_PERF_FLAG_DISABLED
,
1569 .num_properties
= ARRAY_SIZE(properties
) / 2,
1570 .properties_ptr
= (uintptr_t) properties
,
1572 int fd
= gen_ioctl(drm_fd
, DRM_IOCTL_I915_PERF_OPEN
, ¶m
);
1574 DBG("Error opening gen perf OA stream: %m\n");
1578 perf_ctx
->oa_stream_fd
= fd
;
1580 perf_ctx
->current_oa_metrics_set_id
= metrics_set_id
;
1581 perf_ctx
->current_oa_format
= report_format
;
1587 inc_n_users(struct gen_perf_context
*perf_ctx
)
1589 if (perf_ctx
->n_oa_users
== 0 &&
1590 gen_ioctl(perf_ctx
->oa_stream_fd
, I915_PERF_IOCTL_ENABLE
, 0) < 0)
1594 ++perf_ctx
->n_oa_users
;
1600 dec_n_users(struct gen_perf_context
*perf_ctx
)
1602 /* Disabling the i915 perf stream will effectively disable the OA
1603 * counters. Note it's important to be sure there are no outstanding
1604 * MI_RPC commands at this point since they could stall the CS
1605 * indefinitely once OACONTROL is disabled.
1607 --perf_ctx
->n_oa_users
;
1608 if (perf_ctx
->n_oa_users
== 0 &&
1609 gen_ioctl(perf_ctx
->oa_stream_fd
, I915_PERF_IOCTL_DISABLE
, 0) < 0)
1611 DBG("WARNING: Error disabling gen perf stream: %m\n");
1616 gen_perf_init_metrics(struct gen_perf_config
*perf_cfg
,
1617 const struct gen_device_info
*devinfo
,
1620 load_pipeline_statistic_metrics(perf_cfg
, devinfo
);
1621 register_mdapi_statistic_query(perf_cfg
, devinfo
);
1622 if (load_oa_metrics(perf_cfg
, drm_fd
, devinfo
))
1623 register_mdapi_oa_query(devinfo
, perf_cfg
);
1627 gen_perf_init_context(struct gen_perf_context
*perf_ctx
,
1628 struct gen_perf_config
*perf_cfg
,
1629 void * ctx
, /* driver context (eg, brw_context) */
1630 void * bufmgr
, /* eg brw_bufmgr */
1631 const struct gen_device_info
*devinfo
,
1635 perf_ctx
->perf
= perf_cfg
;
1636 perf_ctx
->ctx
= ctx
;
1637 perf_ctx
->bufmgr
= bufmgr
;
1638 perf_ctx
->drm_fd
= drm_fd
;
1639 perf_ctx
->hw_ctx
= hw_ctx
;
1640 perf_ctx
->devinfo
= devinfo
;
1642 perf_ctx
->unaccumulated
=
1643 ralloc_array(ctx
, struct gen_perf_query_object
*, 2);
1644 perf_ctx
->unaccumulated_elements
= 0;
1645 perf_ctx
->unaccumulated_array_size
= 2;
1647 exec_list_make_empty(&perf_ctx
->sample_buffers
);
1648 exec_list_make_empty(&perf_ctx
->free_sample_buffers
);
1650 /* It's convenient to guarantee that this linked list of sample
1651 * buffers is never empty so we add an empty head so when we
1652 * Begin an OA query we can always take a reference on a buffer
1655 struct oa_sample_buf
*buf
= get_free_sample_buf(perf_ctx
);
1656 exec_list_push_head(&perf_ctx
->sample_buffers
, &buf
->link
);
1658 perf_ctx
->oa_stream_fd
= -1;
1659 perf_ctx
->next_query_start_report_id
= 1000;
1663 * Add a query to the global list of "unaccumulated queries."
1665 * Queries are tracked here until all the associated OA reports have
1666 * been accumulated via accumulate_oa_reports() after the end
1667 * MI_REPORT_PERF_COUNT has landed in query->oa.bo.
1670 add_to_unaccumulated_query_list(struct gen_perf_context
*perf_ctx
,
1671 struct gen_perf_query_object
*obj
)
1673 if (perf_ctx
->unaccumulated_elements
>=
1674 perf_ctx
->unaccumulated_array_size
)
1676 perf_ctx
->unaccumulated_array_size
*= 1.5;
1677 perf_ctx
->unaccumulated
=
1678 reralloc(perf_ctx
->ctx
, perf_ctx
->unaccumulated
,
1679 struct gen_perf_query_object
*,
1680 perf_ctx
->unaccumulated_array_size
);
1683 perf_ctx
->unaccumulated
[perf_ctx
->unaccumulated_elements
++] = obj
;
1687 gen_perf_begin_query(struct gen_perf_context
*perf_ctx
,
1688 struct gen_perf_query_object
*query
)
1690 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1691 const struct gen_perf_query_info
*queryinfo
= query
->queryinfo
;
1693 /* XXX: We have to consider that the command parser unit that parses batch
1694 * buffer commands and is used to capture begin/end counter snapshots isn't
1695 * implicitly synchronized with what's currently running across other GPU
1696 * units (such as the EUs running shaders) that the performance counters are
1699 * The intention of performance queries is to measure the work associated
1700 * with commands between the begin/end delimiters and so for that to be the
1701 * case we need to explicitly synchronize the parsing of commands to capture
1702 * Begin/End counter snapshots with what's running across other parts of the
1705 * When the command parser reaches a Begin marker it effectively needs to
1706 * drain everything currently running on the GPU until the hardware is idle
1707 * before capturing the first snapshot of counters - otherwise the results
1708 * would also be measuring the effects of earlier commands.
1710 * When the command parser reaches an End marker it needs to stall until
1711 * everything currently running on the GPU has finished before capturing the
1712 * end snapshot - otherwise the results won't be a complete representation
1715 * Theoretically there could be opportunities to minimize how much of the
1716 * GPU pipeline is drained, or that we stall for, when we know what specific
1717 * units the performance counters being queried relate to but we don't
1718 * currently attempt to be clever here.
1720 * Note: with our current simple approach here then for back-to-back queries
1721 * we will redundantly emit duplicate commands to synchronize the command
1722 * streamer with the rest of the GPU pipeline, but we assume that in HW the
1723 * second synchronization is effectively a NOOP.
1725 * N.B. The final results are based on deltas of counters between (inside)
1726 * Begin/End markers so even though the total wall clock time of the
1727 * workload is stretched by larger pipeline bubbles the bubbles themselves
1728 * are generally invisible to the query results. Whether that's a good or a
1729 * bad thing depends on the use case. For a lower real-time impact while
1730 * capturing metrics then periodic sampling may be a better choice than
1731 * INTEL_performance_query.
1734 * This is our Begin synchronization point to drain current work on the
1735 * GPU before we capture our first counter snapshot...
1737 perf_cfg
->vtbl
.emit_mi_flush(perf_ctx
->ctx
);
1739 switch (queryinfo
->kind
) {
1740 case GEN_PERF_QUERY_TYPE_OA
:
1741 case GEN_PERF_QUERY_TYPE_RAW
: {
1743 /* Opening an i915 perf stream implies exclusive access to the OA unit
1744 * which will generate counter reports for a specific counter set with a
1745 * specific layout/format so we can't begin any OA based queries that
1746 * require a different counter set or format unless we get an opportunity
1747 * to close the stream and open a new one...
1749 uint64_t metric_id
= get_metric_id(perf_ctx
->perf
, queryinfo
);
1751 if (perf_ctx
->oa_stream_fd
!= -1 &&
1752 perf_ctx
->current_oa_metrics_set_id
!= metric_id
) {
1754 if (perf_ctx
->n_oa_users
!= 0) {
1755 DBG("WARNING: Begin failed already using perf config=%i/%"PRIu64
"\n",
1756 perf_ctx
->current_oa_metrics_set_id
, metric_id
);
1759 gen_perf_close(perf_ctx
, queryinfo
);
1762 /* If the OA counters aren't already on, enable them. */
1763 if (perf_ctx
->oa_stream_fd
== -1) {
1764 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
1766 /* The period_exponent gives a sampling period as follows:
1767 * sample_period = timestamp_period * 2^(period_exponent + 1)
1769 * The timestamps increments every 80ns (HSW), ~52ns (GEN9LP) or
1772 * The counter overflow period is derived from the EuActive counter
1773 * which reads a counter that increments by the number of clock
1774 * cycles multiplied by the number of EUs. It can be calculated as:
1776 * 2^(number of bits in A counter) / (n_eus * max_gen_freq * 2)
1778 * (E.g. 40 EUs @ 1GHz = ~53ms)
1780 * We select a sampling period inferior to that overflow period to
1781 * ensure we cannot see more than 1 counter overflow, otherwise we
1782 * could loose information.
1785 int a_counter_in_bits
= 32;
1786 if (devinfo
->gen
>= 8)
1787 a_counter_in_bits
= 40;
1789 uint64_t overflow_period
= pow(2, a_counter_in_bits
) / (perf_cfg
->sys_vars
.n_eus
*
1790 /* drop 1GHz freq to have units in nanoseconds */
1793 DBG("A counter overflow period: %"PRIu64
"ns, %"PRIu64
"ms (n_eus=%"PRIu64
")\n",
1794 overflow_period
, overflow_period
/ 1000000ul, perf_cfg
->sys_vars
.n_eus
);
1796 int period_exponent
= 0;
1797 uint64_t prev_sample_period
, next_sample_period
;
1798 for (int e
= 0; e
< 30; e
++) {
1799 prev_sample_period
= 1000000000ull * pow(2, e
+ 1) / devinfo
->timestamp_frequency
;
1800 next_sample_period
= 1000000000ull * pow(2, e
+ 2) / devinfo
->timestamp_frequency
;
1802 /* Take the previous sampling period, lower than the overflow
1805 if (prev_sample_period
< overflow_period
&&
1806 next_sample_period
> overflow_period
)
1807 period_exponent
= e
+ 1;
1810 if (period_exponent
== 0) {
1811 DBG("WARNING: enable to find a sampling exponent\n");
1815 DBG("OA sampling exponent: %i ~= %"PRIu64
"ms\n", period_exponent
,
1816 prev_sample_period
/ 1000000ul);
1818 if (!gen_perf_open(perf_ctx
, metric_id
, queryinfo
->oa_format
,
1819 period_exponent
, perf_ctx
->drm_fd
,
1823 assert(perf_ctx
->current_oa_metrics_set_id
== metric_id
&&
1824 perf_ctx
->current_oa_format
== queryinfo
->oa_format
);
1827 if (!inc_n_users(perf_ctx
)) {
1828 DBG("WARNING: Error enabling i915 perf stream: %m\n");
1833 perf_cfg
->vtbl
.bo_unreference(query
->oa
.bo
);
1834 query
->oa
.bo
= NULL
;
1837 query
->oa
.bo
= perf_cfg
->vtbl
.bo_alloc(perf_ctx
->bufmgr
,
1838 "perf. query OA MI_RPC bo",
1841 /* Pre-filling the BO helps debug whether writes landed. */
1842 void *map
= perf_cfg
->vtbl
.bo_map(perf_ctx
->ctx
, query
->oa
.bo
, MAP_WRITE
);
1843 memset(map
, 0x80, MI_RPC_BO_SIZE
);
1844 perf_cfg
->vtbl
.bo_unmap(query
->oa
.bo
);
1847 query
->oa
.begin_report_id
= perf_ctx
->next_query_start_report_id
;
1848 perf_ctx
->next_query_start_report_id
+= 2;
1850 /* We flush the batchbuffer here to minimize the chances that MI_RPC
1851 * delimiting commands end up in different batchbuffers. If that's the
1852 * case, the measurement will include the time it takes for the kernel
1853 * scheduler to load a new request into the hardware. This is manifested in
1854 * tools like frameretrace by spikes in the "GPU Core Clocks" counter.
1856 perf_cfg
->vtbl
.batchbuffer_flush(perf_ctx
->ctx
, __FILE__
, __LINE__
);
1858 /* Take a starting OA counter snapshot. */
1859 perf_cfg
->vtbl
.emit_mi_report_perf_count(perf_ctx
->ctx
, query
->oa
.bo
, 0,
1860 query
->oa
.begin_report_id
);
1861 perf_cfg
->vtbl
.capture_frequency_stat_register(perf_ctx
->ctx
, query
->oa
.bo
,
1862 MI_FREQ_START_OFFSET_BYTES
);
1864 ++perf_ctx
->n_active_oa_queries
;
1866 /* No already-buffered samples can possibly be associated with this query
1867 * so create a marker within the list of sample buffers enabling us to
1868 * easily ignore earlier samples when processing this query after
1871 assert(!exec_list_is_empty(&perf_ctx
->sample_buffers
));
1872 query
->oa
.samples_head
= exec_list_get_tail(&perf_ctx
->sample_buffers
);
1874 struct oa_sample_buf
*buf
=
1875 exec_node_data(struct oa_sample_buf
, query
->oa
.samples_head
, link
);
1877 /* This reference will ensure that future/following sample
1878 * buffers (that may relate to this query) can't be freed until
1879 * this drops to zero.
1883 gen_perf_query_result_clear(&query
->oa
.result
);
1884 query
->oa
.results_accumulated
= false;
1886 add_to_unaccumulated_query_list(perf_ctx
, query
);
1890 case GEN_PERF_QUERY_TYPE_PIPELINE
:
1891 if (query
->pipeline_stats
.bo
) {
1892 perf_cfg
->vtbl
.bo_unreference(query
->pipeline_stats
.bo
);
1893 query
->pipeline_stats
.bo
= NULL
;
1896 query
->pipeline_stats
.bo
=
1897 perf_cfg
->vtbl
.bo_alloc(perf_ctx
->bufmgr
,
1898 "perf. query pipeline stats bo",
1901 /* Take starting snapshots. */
1902 snapshot_statistics_registers(perf_ctx
->ctx
, perf_cfg
, query
, 0);
1904 ++perf_ctx
->n_active_pipeline_stats_queries
;
1908 unreachable("Unknown query type");
1916 gen_perf_end_query(struct gen_perf_context
*perf_ctx
,
1917 struct gen_perf_query_object
*query
)
1919 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1921 /* Ensure that the work associated with the queried commands will have
1922 * finished before taking our query end counter readings.
1924 * For more details see comment in brw_begin_perf_query for
1925 * corresponding flush.
1927 perf_cfg
->vtbl
.emit_mi_flush(perf_ctx
->ctx
);
1929 switch (query
->queryinfo
->kind
) {
1930 case GEN_PERF_QUERY_TYPE_OA
:
1931 case GEN_PERF_QUERY_TYPE_RAW
:
1933 /* NB: It's possible that the query will have already been marked
1934 * as 'accumulated' if an error was seen while reading samples
1935 * from perf. In this case we mustn't try and emit a closing
1936 * MI_RPC command in case the OA unit has already been disabled
1938 if (!query
->oa
.results_accumulated
) {
1939 /* Take an ending OA counter snapshot. */
1940 perf_cfg
->vtbl
.capture_frequency_stat_register(perf_ctx
->ctx
, query
->oa
.bo
,
1941 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
->ctx
, perf_cfg
, 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
= tail_buf
->last_timestamp
;
1985 struct oa_sample_buf
*buf
= get_free_sample_buf(perf_ctx
);
1989 while ((len
= read(perf_ctx
->oa_stream_fd
, buf
->buf
,
1990 sizeof(buf
->buf
))) < 0 && errno
== EINTR
)
1994 exec_list_push_tail(&perf_ctx
->free_sample_buffers
, &buf
->link
);
1997 if (errno
== EAGAIN
)
1998 return ((last_timestamp
- start_timestamp
) >=
1999 (end_timestamp
- start_timestamp
)) ?
2000 OA_READ_STATUS_FINISHED
:
2001 OA_READ_STATUS_UNFINISHED
;
2003 DBG("Error reading i915 perf samples: %m\n");
2006 DBG("Spurious EOF reading i915 perf samples\n");
2008 return OA_READ_STATUS_ERROR
;
2012 exec_list_push_tail(&perf_ctx
->sample_buffers
, &buf
->link
);
2014 /* Go through the reports and update the last timestamp. */
2016 while (offset
< buf
->len
) {
2017 const struct drm_i915_perf_record_header
*header
=
2018 (const struct drm_i915_perf_record_header
*) &buf
->buf
[offset
];
2019 uint32_t *report
= (uint32_t *) (header
+ 1);
2021 if (header
->type
== DRM_I915_PERF_RECORD_SAMPLE
)
2022 last_timestamp
= report
[1];
2024 offset
+= header
->size
;
2027 buf
->last_timestamp
= last_timestamp
;
2030 unreachable("not reached");
2031 return OA_READ_STATUS_ERROR
;
2035 * Try to read all the reports until either the delimiting timestamp
2036 * or an error arises.
2039 read_oa_samples_for_query(struct gen_perf_context
*perf_ctx
,
2040 struct gen_perf_query_object
*query
,
2041 void *current_batch
)
2046 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2048 /* We need the MI_REPORT_PERF_COUNT to land before we can start
2050 assert(!perf_cfg
->vtbl
.batch_references(current_batch
, query
->oa
.bo
) &&
2051 !perf_cfg
->vtbl
.bo_busy(query
->oa
.bo
));
2053 /* Map the BO once here and let accumulate_oa_reports() unmap
2055 if (query
->oa
.map
== NULL
)
2056 query
->oa
.map
= perf_cfg
->vtbl
.bo_map(perf_ctx
->ctx
, query
->oa
.bo
, MAP_READ
);
2058 start
= last
= query
->oa
.map
;
2059 end
= query
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
2061 if (start
[0] != query
->oa
.begin_report_id
) {
2062 DBG("Spurious start report id=%"PRIu32
"\n", start
[0]);
2065 if (end
[0] != (query
->oa
.begin_report_id
+ 1)) {
2066 DBG("Spurious end report id=%"PRIu32
"\n", end
[0]);
2070 /* Read the reports until the end timestamp. */
2071 switch (read_oa_samples_until(perf_ctx
, start
[1], end
[1])) {
2072 case OA_READ_STATUS_ERROR
:
2073 /* Fallthrough and let accumulate_oa_reports() deal with the
2075 case OA_READ_STATUS_FINISHED
:
2077 case OA_READ_STATUS_UNFINISHED
:
2081 unreachable("invalid read status");
2086 gen_perf_wait_query(struct gen_perf_context
*perf_ctx
,
2087 struct gen_perf_query_object
*query
,
2088 void *current_batch
)
2090 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2091 struct brw_bo
*bo
= NULL
;
2093 switch (query
->queryinfo
->kind
) {
2094 case GEN_PERF_QUERY_TYPE_OA
:
2095 case GEN_PERF_QUERY_TYPE_RAW
:
2099 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2100 bo
= query
->pipeline_stats
.bo
;
2104 unreachable("Unknown query type");
2111 /* If the current batch references our results bo then we need to
2114 if (perf_cfg
->vtbl
.batch_references(current_batch
, bo
))
2115 perf_cfg
->vtbl
.batchbuffer_flush(perf_ctx
->ctx
, __FILE__
, __LINE__
);
2117 perf_cfg
->vtbl
.bo_wait_rendering(bo
);
2119 /* Due to a race condition between the OA unit signaling report
2120 * availability and the report actually being written into memory,
2121 * we need to wait for all the reports to come in before we can
2124 if (query
->queryinfo
->kind
== GEN_PERF_QUERY_TYPE_OA
||
2125 query
->queryinfo
->kind
== GEN_PERF_QUERY_TYPE_RAW
) {
2126 while (!read_oa_samples_for_query(perf_ctx
, query
, current_batch
))
2132 gen_perf_is_query_ready(struct gen_perf_context
*perf_ctx
,
2133 struct gen_perf_query_object
*query
,
2134 void *current_batch
)
2136 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2138 switch (query
->queryinfo
->kind
) {
2139 case GEN_PERF_QUERY_TYPE_OA
:
2140 case GEN_PERF_QUERY_TYPE_RAW
:
2141 return (query
->oa
.results_accumulated
||
2143 !perf_cfg
->vtbl
.batch_references(current_batch
, query
->oa
.bo
) &&
2144 !perf_cfg
->vtbl
.bo_busy(query
->oa
.bo
) &&
2145 read_oa_samples_for_query(perf_ctx
, query
, current_batch
)));
2146 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2147 return (query
->pipeline_stats
.bo
&&
2148 !perf_cfg
->vtbl
.batch_references(current_batch
, query
->pipeline_stats
.bo
) &&
2149 !perf_cfg
->vtbl
.bo_busy(query
->pipeline_stats
.bo
));
2152 unreachable("Unknown query type");
2160 * Remove a query from the global list of unaccumulated queries once
2161 * after successfully accumulating the OA reports associated with the
2162 * query in accumulate_oa_reports() or when discarding unwanted query
2166 drop_from_unaccumulated_query_list(struct gen_perf_context
*perf_ctx
,
2167 struct gen_perf_query_object
*query
)
2169 for (int i
= 0; i
< perf_ctx
->unaccumulated_elements
; i
++) {
2170 if (perf_ctx
->unaccumulated
[i
] == query
) {
2171 int last_elt
= --perf_ctx
->unaccumulated_elements
;
2174 perf_ctx
->unaccumulated
[i
] = NULL
;
2176 perf_ctx
->unaccumulated
[i
] =
2177 perf_ctx
->unaccumulated
[last_elt
];
2184 /* Drop our samples_head reference so that associated periodic
2185 * sample data buffers can potentially be reaped if they aren't
2186 * referenced by any other queries...
2189 struct oa_sample_buf
*buf
=
2190 exec_node_data(struct oa_sample_buf
, query
->oa
.samples_head
, link
);
2192 assert(buf
->refcount
> 0);
2195 query
->oa
.samples_head
= NULL
;
2197 reap_old_sample_buffers(perf_ctx
);
2200 /* In general if we see anything spurious while accumulating results,
2201 * we don't try and continue accumulating the current query, hoping
2202 * for the best, we scrap anything outstanding, and then hope for the
2203 * best with new queries.
2206 discard_all_queries(struct gen_perf_context
*perf_ctx
)
2208 while (perf_ctx
->unaccumulated_elements
) {
2209 struct gen_perf_query_object
*query
= perf_ctx
->unaccumulated
[0];
2211 query
->oa
.results_accumulated
= true;
2212 drop_from_unaccumulated_query_list(perf_ctx
, query
);
2214 dec_n_users(perf_ctx
);
2219 * Accumulate raw OA counter values based on deltas between pairs of
2222 * Accumulation starts from the first report captured via
2223 * MI_REPORT_PERF_COUNT (MI_RPC) by brw_begin_perf_query() until the
2224 * last MI_RPC report requested by brw_end_perf_query(). Between these
2225 * two reports there may also some number of periodically sampled OA
2226 * reports collected via the i915 perf interface - depending on the
2227 * duration of the query.
2229 * These periodic snapshots help to ensure we handle counter overflow
2230 * correctly by being frequent enough to ensure we don't miss multiple
2231 * overflows of a counter between snapshots. For Gen8+ the i915 perf
2232 * snapshots provide the extra context-switch reports that let us
2233 * subtract out the progress of counters associated with other
2234 * contexts running on the system.
2237 accumulate_oa_reports(struct gen_perf_context
*perf_ctx
,
2238 struct gen_perf_query_object
*query
)
2240 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
2244 struct exec_node
*first_samples_node
;
2246 int out_duration
= 0;
2248 assert(query
->oa
.map
!= NULL
);
2250 start
= last
= query
->oa
.map
;
2251 end
= query
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
2253 if (start
[0] != query
->oa
.begin_report_id
) {
2254 DBG("Spurious start report id=%"PRIu32
"\n", start
[0]);
2257 if (end
[0] != (query
->oa
.begin_report_id
+ 1)) {
2258 DBG("Spurious end report id=%"PRIu32
"\n", end
[0]);
2262 /* On Gen12+ OA reports are sourced from per context counters, so we don't
2263 * ever have to look at the global OA buffer. Yey \o/
2265 if (perf_ctx
->devinfo
->gen
>= 12) {
2270 /* See if we have any periodic reports to accumulate too... */
2272 /* N.B. The oa.samples_head was set when the query began and
2273 * pointed to the tail of the perf_ctx->sample_buffers list at
2274 * the time the query started. Since the buffer existed before the
2275 * first MI_REPORT_PERF_COUNT command was emitted we therefore know
2276 * that no data in this particular node's buffer can possibly be
2277 * associated with the query - so skip ahead one...
2279 first_samples_node
= query
->oa
.samples_head
->next
;
2281 foreach_list_typed_from(struct oa_sample_buf
, buf
, link
,
2282 &perf_ctx
.sample_buffers
,
2287 while (offset
< buf
->len
) {
2288 const struct drm_i915_perf_record_header
*header
=
2289 (const struct drm_i915_perf_record_header
*)(buf
->buf
+ offset
);
2291 assert(header
->size
!= 0);
2292 assert(header
->size
<= buf
->len
);
2294 offset
+= header
->size
;
2296 switch (header
->type
) {
2297 case DRM_I915_PERF_RECORD_SAMPLE
: {
2298 uint32_t *report
= (uint32_t *)(header
+ 1);
2301 /* Ignore reports that come before the start marker.
2302 * (Note: takes care to allow overflow of 32bit timestamps)
2304 if (gen_device_info_timebase_scale(devinfo
,
2305 report
[1] - start
[1]) > 5000000000) {
2309 /* Ignore reports that come after the end marker.
2310 * (Note: takes care to allow overflow of 32bit timestamps)
2312 if (gen_device_info_timebase_scale(devinfo
,
2313 report
[1] - end
[1]) <= 5000000000) {
2317 /* For Gen8+ since the counters continue while other
2318 * contexts are running we need to discount any unrelated
2319 * deltas. The hardware automatically generates a report
2320 * on context switch which gives us a new reference point
2321 * to continuing adding deltas from.
2323 * For Haswell we can rely on the HW to stop the progress
2324 * of OA counters while any other context is acctive.
2326 if (devinfo
->gen
>= 8) {
2327 if (in_ctx
&& report
[2] != query
->oa
.result
.hw_id
) {
2328 DBG("i915 perf: Switch AWAY (observed by ID change)\n");
2331 } else if (in_ctx
== false && report
[2] == query
->oa
.result
.hw_id
) {
2332 DBG("i915 perf: Switch TO\n");
2335 /* From experimentation in IGT, we found that the OA unit
2336 * might label some report as "idle" (using an invalid
2337 * context ID), right after a report for a given context.
2338 * Deltas generated by those reports actually belong to the
2339 * previous context, even though they're not labelled as
2342 * We didn't *really* Switch AWAY in the case that we e.g.
2343 * saw a single periodic report while idle...
2345 if (out_duration
>= 1)
2347 } else if (in_ctx
) {
2348 assert(report
[2] == query
->oa
.result
.hw_id
);
2349 DBG("i915 perf: Continuation IN\n");
2351 assert(report
[2] != query
->oa
.result
.hw_id
);
2352 DBG("i915 perf: Continuation OUT\n");
2359 gen_perf_query_result_accumulate(&query
->oa
.result
,
2369 case DRM_I915_PERF_RECORD_OA_BUFFER_LOST
:
2370 DBG("i915 perf: OA error: all reports lost\n");
2372 case DRM_I915_PERF_RECORD_OA_REPORT_LOST
:
2373 DBG("i915 perf: OA report lost\n");
2381 gen_perf_query_result_accumulate(&query
->oa
.result
, query
->queryinfo
,
2384 query
->oa
.results_accumulated
= true;
2385 drop_from_unaccumulated_query_list(perf_ctx
, query
);
2386 dec_n_users(perf_ctx
);
2392 discard_all_queries(perf_ctx
);
2396 gen_perf_delete_query(struct gen_perf_context
*perf_ctx
,
2397 struct gen_perf_query_object
*query
)
2399 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2401 /* We can assume that the frontend waits for a query to complete
2402 * before ever calling into here, so we don't have to worry about
2403 * deleting an in-flight query object.
2405 switch (query
->queryinfo
->kind
) {
2406 case GEN_PERF_QUERY_TYPE_OA
:
2407 case GEN_PERF_QUERY_TYPE_RAW
:
2409 if (!query
->oa
.results_accumulated
) {
2410 drop_from_unaccumulated_query_list(perf_ctx
, query
);
2411 dec_n_users(perf_ctx
);
2414 perf_cfg
->vtbl
.bo_unreference(query
->oa
.bo
);
2415 query
->oa
.bo
= NULL
;
2418 query
->oa
.results_accumulated
= false;
2421 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2422 if (query
->pipeline_stats
.bo
) {
2423 perf_cfg
->vtbl
.bo_unreference(query
->pipeline_stats
.bo
);
2424 query
->pipeline_stats
.bo
= NULL
;
2429 unreachable("Unknown query type");
2433 /* As an indication that the INTEL_performance_query extension is no
2434 * longer in use, it's a good time to free our cache of sample
2435 * buffers and close any current i915-perf stream.
2437 if (--perf_ctx
->n_query_instances
== 0) {
2438 free_sample_bufs(perf_ctx
);
2439 gen_perf_close(perf_ctx
, query
->queryinfo
);
2445 #define GET_FIELD(word, field) (((word) & field ## _MASK) >> field ## _SHIFT)
2448 read_gt_frequency(struct gen_perf_context
*perf_ctx
,
2449 struct gen_perf_query_object
*obj
)
2451 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
2452 uint32_t start
= *((uint32_t *)(obj
->oa
.map
+ MI_FREQ_START_OFFSET_BYTES
)),
2453 end
= *((uint32_t *)(obj
->oa
.map
+ MI_FREQ_END_OFFSET_BYTES
));
2455 switch (devinfo
->gen
) {
2458 obj
->oa
.gt_frequency
[0] = GET_FIELD(start
, GEN7_RPSTAT1_CURR_GT_FREQ
) * 50ULL;
2459 obj
->oa
.gt_frequency
[1] = GET_FIELD(end
, GEN7_RPSTAT1_CURR_GT_FREQ
) * 50ULL;
2464 obj
->oa
.gt_frequency
[0] = GET_FIELD(start
, GEN9_RPSTAT0_CURR_GT_FREQ
) * 50ULL / 3ULL;
2465 obj
->oa
.gt_frequency
[1] = GET_FIELD(end
, GEN9_RPSTAT0_CURR_GT_FREQ
) * 50ULL / 3ULL;
2468 unreachable("unexpected gen");
2471 /* Put the numbers into Hz. */
2472 obj
->oa
.gt_frequency
[0] *= 1000000ULL;
2473 obj
->oa
.gt_frequency
[1] *= 1000000ULL;
2477 get_oa_counter_data(struct gen_perf_context
*perf_ctx
,
2478 struct gen_perf_query_object
*query
,
2482 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2483 const struct gen_perf_query_info
*queryinfo
= query
->queryinfo
;
2484 int n_counters
= queryinfo
->n_counters
;
2487 for (int i
= 0; i
< n_counters
; i
++) {
2488 const struct gen_perf_query_counter
*counter
= &queryinfo
->counters
[i
];
2489 uint64_t *out_uint64
;
2491 size_t counter_size
= gen_perf_query_counter_get_size(counter
);
2494 switch (counter
->data_type
) {
2495 case GEN_PERF_COUNTER_DATA_TYPE_UINT64
:
2496 out_uint64
= (uint64_t *)(data
+ counter
->offset
);
2498 counter
->oa_counter_read_uint64(perf_cfg
, queryinfo
,
2499 query
->oa
.result
.accumulator
);
2501 case GEN_PERF_COUNTER_DATA_TYPE_FLOAT
:
2502 out_float
= (float *)(data
+ counter
->offset
);
2504 counter
->oa_counter_read_float(perf_cfg
, queryinfo
,
2505 query
->oa
.result
.accumulator
);
2508 /* So far we aren't using uint32, double or bool32... */
2509 unreachable("unexpected counter data type");
2511 written
= counter
->offset
+ counter_size
;
2519 get_pipeline_stats_data(struct gen_perf_context
*perf_ctx
,
2520 struct gen_perf_query_object
*query
,
2525 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2526 const struct gen_perf_query_info
*queryinfo
= query
->queryinfo
;
2527 int n_counters
= queryinfo
->n_counters
;
2530 uint64_t *start
= perf_cfg
->vtbl
.bo_map(perf_ctx
->ctx
, query
->pipeline_stats
.bo
, MAP_READ
);
2531 uint64_t *end
= start
+ (STATS_BO_END_OFFSET_BYTES
/ sizeof(uint64_t));
2533 for (int i
= 0; i
< n_counters
; i
++) {
2534 const struct gen_perf_query_counter
*counter
= &queryinfo
->counters
[i
];
2535 uint64_t value
= end
[i
] - start
[i
];
2537 if (counter
->pipeline_stat
.numerator
!=
2538 counter
->pipeline_stat
.denominator
) {
2539 value
*= counter
->pipeline_stat
.numerator
;
2540 value
/= counter
->pipeline_stat
.denominator
;
2543 *((uint64_t *)p
) = value
;
2547 perf_cfg
->vtbl
.bo_unmap(query
->pipeline_stats
.bo
);
2553 gen_perf_get_query_data(struct gen_perf_context
*perf_ctx
,
2554 struct gen_perf_query_object
*query
,
2557 unsigned *bytes_written
)
2559 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2562 switch (query
->queryinfo
->kind
) {
2563 case GEN_PERF_QUERY_TYPE_OA
:
2564 case GEN_PERF_QUERY_TYPE_RAW
:
2565 if (!query
->oa
.results_accumulated
) {
2566 read_gt_frequency(perf_ctx
, query
);
2567 uint32_t *begin_report
= query
->oa
.map
;
2568 uint32_t *end_report
= query
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
2569 gen_perf_query_result_read_frequencies(&query
->oa
.result
,
2573 accumulate_oa_reports(perf_ctx
, query
);
2574 assert(query
->oa
.results_accumulated
);
2576 perf_cfg
->vtbl
.bo_unmap(query
->oa
.bo
);
2577 query
->oa
.map
= NULL
;
2579 if (query
->queryinfo
->kind
== GEN_PERF_QUERY_TYPE_OA
) {
2580 written
= get_oa_counter_data(perf_ctx
, query
, data_size
, (uint8_t *)data
);
2582 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
2584 written
= gen_perf_query_result_write_mdapi((uint8_t *)data
, data_size
,
2585 devinfo
, &query
->oa
.result
,
2586 query
->oa
.gt_frequency
[0],
2587 query
->oa
.gt_frequency
[1]);
2591 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2592 written
= get_pipeline_stats_data(perf_ctx
, query
, data_size
, (uint8_t *)data
);
2596 unreachable("Unknown query type");
2601 *bytes_written
= written
;
2605 gen_perf_dump_query_count(struct gen_perf_context
*perf_ctx
)
2607 DBG("Queries: (Open queries = %d, OA users = %d)\n",
2608 perf_ctx
->n_active_oa_queries
, perf_ctx
->n_oa_users
);
2612 gen_perf_dump_query(struct gen_perf_context
*ctx
,
2613 struct gen_perf_query_object
*obj
,
2614 void *current_batch
)
2616 switch (obj
->queryinfo
->kind
) {
2617 case GEN_PERF_QUERY_TYPE_OA
:
2618 case GEN_PERF_QUERY_TYPE_RAW
:
2619 DBG("BO: %-4s OA data: %-10s %-15s\n",
2620 obj
->oa
.bo
? "yes," : "no,",
2621 gen_perf_is_query_ready(ctx
, obj
, current_batch
) ? "ready," : "not ready,",
2622 obj
->oa
.results_accumulated
? "accumulated" : "not accumulated");
2624 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2626 obj
->pipeline_stats
.bo
? "yes" : "no");
2629 unreachable("Unknown query type");