2 * Copyright © 2018 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
26 #include <sys/types.h>
32 #include <drm-uapi/i915_drm.h>
34 #include "common/gen_gem.h"
36 #include "perf/gen_perf_mdapi.h"
37 #include "perf/gen_perf_metrics.h"
39 #include "dev/gen_debug.h"
40 #include "dev/gen_device_info.h"
41 #include "util/bitscan.h"
42 #include "util/u_math.h"
44 #define FILE_DEBUG_FLAG DEBUG_PERFMON
45 #define MI_RPC_BO_SIZE 4096
46 #define MI_FREQ_START_OFFSET_BYTES (3072)
47 #define MI_RPC_BO_END_OFFSET_BYTES (MI_RPC_BO_SIZE / 2)
48 #define MI_FREQ_END_OFFSET_BYTES (3076)
50 #define INTEL_MASK(high, low) (((1u<<((high)-(low)+1))-1)<<(low))
52 #define GEN7_RPSTAT1 0xA01C
53 #define GEN7_RPSTAT1_CURR_GT_FREQ_SHIFT 7
54 #define GEN7_RPSTAT1_CURR_GT_FREQ_MASK INTEL_MASK(13, 7)
55 #define GEN7_RPSTAT1_PREV_GT_FREQ_SHIFT 0
56 #define GEN7_RPSTAT1_PREV_GT_FREQ_MASK INTEL_MASK(6, 0)
58 #define GEN9_RPSTAT0 0xA01C
59 #define GEN9_RPSTAT0_CURR_GT_FREQ_SHIFT 23
60 #define GEN9_RPSTAT0_CURR_GT_FREQ_MASK INTEL_MASK(31, 23)
61 #define GEN9_RPSTAT0_PREV_GT_FREQ_SHIFT 0
62 #define GEN9_RPSTAT0_PREV_GT_FREQ_MASK INTEL_MASK(8, 0)
64 #define GEN6_SO_PRIM_STORAGE_NEEDED 0x2280
65 #define GEN7_SO_PRIM_STORAGE_NEEDED(n) (0x5240 + (n) * 8)
66 #define GEN6_SO_NUM_PRIMS_WRITTEN 0x2288
67 #define GEN7_SO_NUM_PRIMS_WRITTEN(n) (0x5200 + (n) * 8)
69 #define MAP_READ (1 << 0)
70 #define MAP_WRITE (1 << 1)
73 * Periodic OA samples are read() into these buffer structures via the
74 * i915 perf kernel interface and appended to the
75 * perf_ctx->sample_buffers linked list. When we process the
76 * results of an OA metrics query we need to consider all the periodic
77 * samples between the Begin and End MI_REPORT_PERF_COUNT command
80 * 'Periodic' is a simplification as there are other automatic reports
81 * written by the hardware also buffered here.
83 * Considering three queries, A, B and C:
86 * ________________A_________________
88 * | ________B_________ _____C___________
91 * And an illustration of sample buffers read over this time frame:
92 * [HEAD ][ ][ ][ ][ ][ ][ ][ ][TAIL ]
94 * These nodes may hold samples for query A:
95 * [ ][ ][ A ][ A ][ A ][ A ][ A ][ ][ ]
97 * These nodes may hold samples for query B:
98 * [ ][ ][ B ][ B ][ B ][ ][ ][ ][ ]
100 * These nodes may hold samples for query C:
101 * [ ][ ][ ][ ][ ][ C ][ C ][ C ][ ]
103 * The illustration assumes we have an even distribution of periodic
104 * samples so all nodes have the same size plotted against time:
106 * Note, to simplify code, the list is never empty.
108 * With overlapping queries we can see that periodic OA reports may
109 * relate to multiple queries and care needs to be take to keep
110 * track of sample buffers until there are no queries that might
111 * depend on their contents.
113 * We use a node ref counting system where a reference ensures that a
114 * node and all following nodes can't be freed/recycled until the
115 * reference drops to zero.
117 * E.g. with a ref of one here:
118 * [ 0 ][ 0 ][ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
120 * These nodes could be freed or recycled ("reaped"):
123 * These must be preserved until the leading ref drops to zero:
124 * [ 1 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ][ 0 ]
126 * When a query starts we take a reference on the current tail of
127 * the list, knowing that no already-buffered samples can possibly
128 * relate to the newly-started query. A pointer to this node is
129 * also saved in the query object's ->oa.samples_head.
131 * E.g. starting query A while there are two nodes in .sample_buffers:
132 * ________________A________
136 * ^_______ Add a reference and store pointer to node in
139 * Moving forward to when the B query starts with no new buffer nodes:
140 * (for reference, i915 perf reads() are only done when queries finish)
141 * ________________A_______
146 * ^_______ Add a reference and store pointer to
147 * node in B->oa.samples_head
149 * Once a query is finished, after an OA query has become 'Ready',
150 * once the End OA report has landed and after we we have processed
151 * all the intermediate periodic samples then we drop the
152 * ->oa.samples_head reference we took at the start.
154 * So when the B query has finished we have:
155 * ________________A________
156 * | ______B___________
158 * [ 0 ][ 1 ][ 0 ][ 0 ][ 0 ]
159 * ^_______ Drop B->oa.samples_head reference
161 * We still can't free these due to the A->oa.samples_head ref:
162 * [ 1 ][ 0 ][ 0 ][ 0 ]
164 * When the A query finishes: (note there's a new ref for C's samples_head)
165 * ________________A_________________
169 * [ 0 ][ 0 ][ 0 ][ 0 ][ 1 ][ 0 ][ 0 ]
170 * ^_______ Drop A->oa.samples_head reference
172 * And we can now reap these nodes up to the C->oa.samples_head:
173 * [ X ][ X ][ X ][ X ]
174 * keeping -> [ 1 ][ 0 ][ 0 ]
176 * We reap old sample buffers each time we finish processing an OA
177 * query by iterating the sample_buffers list from the head until we
178 * find a referenced node and stop.
180 * Reaped buffers move to a perfquery.free_sample_buffers list and
181 * when we come to read() we first look to recycle a buffer from the
182 * free_sample_buffers list before allocating a new buffer.
184 struct oa_sample_buf
{
185 struct exec_node link
;
188 uint8_t buf
[I915_PERF_OA_SAMPLE_SIZE
* 10];
189 uint32_t last_timestamp
;
192 struct gen_perf_query_object
*
193 gen_perf_new_query(struct gen_perf_context
*perf_ctx
, unsigned query_index
)
195 const struct gen_perf_query_info
*query
=
196 &perf_ctx
->perf
->queries
[query_index
];
197 struct gen_perf_query_object
*obj
=
198 calloc(1, sizeof(struct gen_perf_query_object
));
203 obj
->queryinfo
= query
;
205 perf_ctx
->n_query_instances
++;
210 get_sysfs_dev_dir(struct gen_perf_config
*perf
, int fd
)
215 struct dirent
*drm_entry
;
218 perf
->sysfs_dev_dir
[0] = '\0';
220 if (fstat(fd
, &sb
)) {
221 DBG("Failed to stat DRM fd\n");
225 maj
= major(sb
.st_rdev
);
226 min
= minor(sb
.st_rdev
);
228 if (!S_ISCHR(sb
.st_mode
)) {
229 DBG("DRM fd is not a character device as expected\n");
233 len
= snprintf(perf
->sysfs_dev_dir
,
234 sizeof(perf
->sysfs_dev_dir
),
235 "/sys/dev/char/%d:%d/device/drm", maj
, min
);
236 if (len
< 0 || len
>= sizeof(perf
->sysfs_dev_dir
)) {
237 DBG("Failed to concatenate sysfs path to drm device\n");
241 drmdir
= opendir(perf
->sysfs_dev_dir
);
243 DBG("Failed to open %s: %m\n", perf
->sysfs_dev_dir
);
247 while ((drm_entry
= readdir(drmdir
))) {
248 if ((drm_entry
->d_type
== DT_DIR
||
249 drm_entry
->d_type
== DT_LNK
) &&
250 strncmp(drm_entry
->d_name
, "card", 4) == 0)
252 len
= snprintf(perf
->sysfs_dev_dir
,
253 sizeof(perf
->sysfs_dev_dir
),
254 "/sys/dev/char/%d:%d/device/drm/%s",
255 maj
, min
, drm_entry
->d_name
);
257 if (len
< 0 || len
>= sizeof(perf
->sysfs_dev_dir
))
266 DBG("Failed to find cardX directory under /sys/dev/char/%d:%d/device/drm\n",
273 read_file_uint64(const char *file
, uint64_t *val
)
281 while ((n
= read(fd
, buf
, sizeof (buf
) - 1)) < 0 &&
288 *val
= strtoull(buf
, NULL
, 0);
294 read_sysfs_drm_device_file_uint64(struct gen_perf_config
*perf
,
301 len
= snprintf(buf
, sizeof(buf
), "%s/%s", perf
->sysfs_dev_dir
, file
);
302 if (len
< 0 || len
>= sizeof(buf
)) {
303 DBG("Failed to concatenate sys filename to read u64 from\n");
307 return read_file_uint64(buf
, value
);
310 static inline struct gen_perf_query_info
*
311 append_query_info(struct gen_perf_config
*perf
, int max_counters
)
313 struct gen_perf_query_info
*query
;
315 perf
->queries
= reralloc(perf
, perf
->queries
,
316 struct gen_perf_query_info
,
318 query
= &perf
->queries
[perf
->n_queries
- 1];
319 memset(query
, 0, sizeof(*query
));
321 if (max_counters
> 0) {
322 query
->max_counters
= max_counters
;
324 rzalloc_array(perf
, struct gen_perf_query_counter
, max_counters
);
331 register_oa_config(struct gen_perf_config
*perf
,
332 const struct gen_perf_query_info
*query
,
335 struct gen_perf_query_info
*registred_query
= append_query_info(perf
, 0);
337 *registred_query
= *query
;
338 registred_query
->oa_metrics_set_id
= config_id
;
339 DBG("metric set registred: id = %" PRIu64
", guid = %s\n",
340 registred_query
->oa_metrics_set_id
, query
->guid
);
344 enumerate_sysfs_metrics(struct gen_perf_config
*perf
)
346 DIR *metricsdir
= NULL
;
347 struct dirent
*metric_entry
;
351 len
= snprintf(buf
, sizeof(buf
), "%s/metrics", perf
->sysfs_dev_dir
);
352 if (len
< 0 || len
>= sizeof(buf
)) {
353 DBG("Failed to concatenate path to sysfs metrics/ directory\n");
357 metricsdir
= opendir(buf
);
359 DBG("Failed to open %s: %m\n", buf
);
363 while ((metric_entry
= readdir(metricsdir
))) {
364 struct hash_entry
*entry
;
366 if ((metric_entry
->d_type
!= DT_DIR
&&
367 metric_entry
->d_type
!= DT_LNK
) ||
368 metric_entry
->d_name
[0] == '.')
371 DBG("metric set: %s\n", metric_entry
->d_name
);
372 entry
= _mesa_hash_table_search(perf
->oa_metrics_table
,
373 metric_entry
->d_name
);
377 len
= snprintf(buf
, sizeof(buf
), "%s/metrics/%s/id",
378 perf
->sysfs_dev_dir
, metric_entry
->d_name
);
379 if (len
< 0 || len
>= sizeof(buf
)) {
380 DBG("Failed to concatenate path to sysfs metric id file\n");
384 if (!read_file_uint64(buf
, &id
)) {
385 DBG("Failed to read metric set id from %s: %m", buf
);
389 register_oa_config(perf
, (const struct gen_perf_query_info
*)entry
->data
, id
);
391 DBG("metric set not known by mesa (skipping)\n");
394 closedir(metricsdir
);
398 kernel_has_dynamic_config_support(struct gen_perf_config
*perf
, int fd
)
400 uint64_t invalid_config_id
= UINT64_MAX
;
402 return gen_ioctl(fd
, DRM_IOCTL_I915_PERF_REMOVE_CONFIG
,
403 &invalid_config_id
) < 0 && errno
== ENOENT
;
407 load_metric_id(struct gen_perf_config
*perf
, const char *guid
,
410 char config_path
[280];
412 snprintf(config_path
, sizeof(config_path
), "%s/metrics/%s/id",
413 perf
->sysfs_dev_dir
, guid
);
415 /* Don't recreate already loaded configs. */
416 return read_file_uint64(config_path
, metric_id
);
420 init_oa_configs(struct gen_perf_config
*perf
, int fd
)
422 hash_table_foreach(perf
->oa_metrics_table
, entry
) {
423 const struct gen_perf_query_info
*query
= entry
->data
;
424 struct drm_i915_perf_oa_config config
;
428 if (load_metric_id(perf
, query
->guid
, &config_id
)) {
429 DBG("metric set: %s (already loaded)\n", query
->guid
);
430 register_oa_config(perf
, query
, config_id
);
434 memset(&config
, 0, sizeof(config
));
436 memcpy(config
.uuid
, query
->guid
, sizeof(config
.uuid
));
438 config
.n_mux_regs
= query
->n_mux_regs
;
439 config
.mux_regs_ptr
= (uintptr_t) query
->mux_regs
;
441 config
.n_boolean_regs
= query
->n_b_counter_regs
;
442 config
.boolean_regs_ptr
= (uintptr_t) query
->b_counter_regs
;
444 config
.n_flex_regs
= query
->n_flex_regs
;
445 config
.flex_regs_ptr
= (uintptr_t) query
->flex_regs
;
447 ret
= gen_ioctl(fd
, DRM_IOCTL_I915_PERF_ADD_CONFIG
, &config
);
449 DBG("Failed to load \"%s\" (%s) metrics set in kernel: %s\n",
450 query
->name
, query
->guid
, strerror(errno
));
454 register_oa_config(perf
, query
, ret
);
455 DBG("metric set: %s (added)\n", query
->guid
);
460 compute_topology_builtins(struct gen_perf_config
*perf
,
461 const struct gen_device_info
*devinfo
)
463 perf
->sys_vars
.slice_mask
= devinfo
->slice_masks
;
464 perf
->sys_vars
.n_eu_slices
= devinfo
->num_slices
;
466 for (int i
= 0; i
< sizeof(devinfo
->subslice_masks
[i
]); i
++) {
467 perf
->sys_vars
.n_eu_sub_slices
+=
468 __builtin_popcount(devinfo
->subslice_masks
[i
]);
471 for (int i
= 0; i
< sizeof(devinfo
->eu_masks
); i
++)
472 perf
->sys_vars
.n_eus
+= __builtin_popcount(devinfo
->eu_masks
[i
]);
474 perf
->sys_vars
.eu_threads_count
= devinfo
->num_thread_per_eu
;
476 /* The subslice mask builtin contains bits for all slices. Prior to Gen11
477 * it had groups of 3bits for each slice, on Gen11 it's 8bits for each
480 * Ideally equations would be updated to have a slice/subslice query
483 perf
->sys_vars
.subslice_mask
= 0;
485 int bits_per_subslice
= devinfo
->gen
== 11 ? 8 : 3;
487 for (int s
= 0; s
< util_last_bit(devinfo
->slice_masks
); s
++) {
488 for (int ss
= 0; ss
< (devinfo
->subslice_slice_stride
* 8); ss
++) {
489 if (gen_device_info_subslice_available(devinfo
, s
, ss
))
490 perf
->sys_vars
.subslice_mask
|= 1ULL << (s
* bits_per_subslice
+ ss
);
496 init_oa_sys_vars(struct gen_perf_config
*perf
, const struct gen_device_info
*devinfo
)
498 uint64_t min_freq_mhz
= 0, max_freq_mhz
= 0;
500 if (!read_sysfs_drm_device_file_uint64(perf
, "gt_min_freq_mhz", &min_freq_mhz
))
503 if (!read_sysfs_drm_device_file_uint64(perf
, "gt_max_freq_mhz", &max_freq_mhz
))
506 memset(&perf
->sys_vars
, 0, sizeof(perf
->sys_vars
));
507 perf
->sys_vars
.gt_min_freq
= min_freq_mhz
* 1000000;
508 perf
->sys_vars
.gt_max_freq
= max_freq_mhz
* 1000000;
509 perf
->sys_vars
.timestamp_frequency
= devinfo
->timestamp_frequency
;
510 perf
->sys_vars
.revision
= devinfo
->revision
;
511 compute_topology_builtins(perf
, devinfo
);
516 typedef void (*perf_register_oa_queries_t
)(struct gen_perf_config
*);
518 static perf_register_oa_queries_t
519 get_register_queries_function(const struct gen_device_info
*devinfo
)
521 if (devinfo
->is_haswell
)
522 return gen_oa_register_queries_hsw
;
523 if (devinfo
->is_cherryview
)
524 return gen_oa_register_queries_chv
;
525 if (devinfo
->is_broadwell
)
526 return gen_oa_register_queries_bdw
;
527 if (devinfo
->is_broxton
)
528 return gen_oa_register_queries_bxt
;
529 if (devinfo
->is_skylake
) {
530 if (devinfo
->gt
== 2)
531 return gen_oa_register_queries_sklgt2
;
532 if (devinfo
->gt
== 3)
533 return gen_oa_register_queries_sklgt3
;
534 if (devinfo
->gt
== 4)
535 return gen_oa_register_queries_sklgt4
;
537 if (devinfo
->is_kabylake
) {
538 if (devinfo
->gt
== 2)
539 return gen_oa_register_queries_kblgt2
;
540 if (devinfo
->gt
== 3)
541 return gen_oa_register_queries_kblgt3
;
543 if (devinfo
->is_geminilake
)
544 return gen_oa_register_queries_glk
;
545 if (devinfo
->is_coffeelake
) {
546 if (devinfo
->gt
== 2)
547 return gen_oa_register_queries_cflgt2
;
548 if (devinfo
->gt
== 3)
549 return gen_oa_register_queries_cflgt3
;
551 if (devinfo
->is_cannonlake
)
552 return gen_oa_register_queries_cnl
;
553 if (devinfo
->gen
== 11)
554 return gen_oa_register_queries_icl
;
560 add_stat_reg(struct gen_perf_query_info
*query
, uint32_t reg
,
561 uint32_t numerator
, uint32_t denominator
,
562 const char *name
, const char *description
)
564 struct gen_perf_query_counter
*counter
;
566 assert(query
->n_counters
< query
->max_counters
);
568 counter
= &query
->counters
[query
->n_counters
];
569 counter
->name
= name
;
570 counter
->desc
= description
;
571 counter
->type
= GEN_PERF_COUNTER_TYPE_RAW
;
572 counter
->data_type
= GEN_PERF_COUNTER_DATA_TYPE_UINT64
;
573 counter
->offset
= sizeof(uint64_t) * query
->n_counters
;
574 counter
->pipeline_stat
.reg
= reg
;
575 counter
->pipeline_stat
.numerator
= numerator
;
576 counter
->pipeline_stat
.denominator
= denominator
;
582 add_basic_stat_reg(struct gen_perf_query_info
*query
,
583 uint32_t reg
, const char *name
)
585 add_stat_reg(query
, reg
, 1, 1, name
, name
);
589 load_pipeline_statistic_metrics(struct gen_perf_config
*perf_cfg
,
590 const struct gen_device_info
*devinfo
)
592 struct gen_perf_query_info
*query
=
593 append_query_info(perf_cfg
, MAX_STAT_COUNTERS
);
595 query
->kind
= GEN_PERF_QUERY_TYPE_PIPELINE
;
596 query
->name
= "Pipeline Statistics Registers";
598 add_basic_stat_reg(query
, IA_VERTICES_COUNT
,
599 "N vertices submitted");
600 add_basic_stat_reg(query
, IA_PRIMITIVES_COUNT
,
601 "N primitives submitted");
602 add_basic_stat_reg(query
, VS_INVOCATION_COUNT
,
603 "N vertex shader invocations");
605 if (devinfo
->gen
== 6) {
606 add_stat_reg(query
, GEN6_SO_PRIM_STORAGE_NEEDED
, 1, 1,
607 "SO_PRIM_STORAGE_NEEDED",
608 "N geometry shader stream-out primitives (total)");
609 add_stat_reg(query
, GEN6_SO_NUM_PRIMS_WRITTEN
, 1, 1,
610 "SO_NUM_PRIMS_WRITTEN",
611 "N geometry shader stream-out primitives (written)");
613 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(0), 1, 1,
614 "SO_PRIM_STORAGE_NEEDED (Stream 0)",
615 "N stream-out (stream 0) primitives (total)");
616 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(1), 1, 1,
617 "SO_PRIM_STORAGE_NEEDED (Stream 1)",
618 "N stream-out (stream 1) primitives (total)");
619 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(2), 1, 1,
620 "SO_PRIM_STORAGE_NEEDED (Stream 2)",
621 "N stream-out (stream 2) primitives (total)");
622 add_stat_reg(query
, GEN7_SO_PRIM_STORAGE_NEEDED(3), 1, 1,
623 "SO_PRIM_STORAGE_NEEDED (Stream 3)",
624 "N stream-out (stream 3) primitives (total)");
625 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(0), 1, 1,
626 "SO_NUM_PRIMS_WRITTEN (Stream 0)",
627 "N stream-out (stream 0) primitives (written)");
628 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(1), 1, 1,
629 "SO_NUM_PRIMS_WRITTEN (Stream 1)",
630 "N stream-out (stream 1) primitives (written)");
631 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(2), 1, 1,
632 "SO_NUM_PRIMS_WRITTEN (Stream 2)",
633 "N stream-out (stream 2) primitives (written)");
634 add_stat_reg(query
, GEN7_SO_NUM_PRIMS_WRITTEN(3), 1, 1,
635 "SO_NUM_PRIMS_WRITTEN (Stream 3)",
636 "N stream-out (stream 3) primitives (written)");
639 add_basic_stat_reg(query
, HS_INVOCATION_COUNT
,
640 "N TCS shader invocations");
641 add_basic_stat_reg(query
, DS_INVOCATION_COUNT
,
642 "N TES shader invocations");
644 add_basic_stat_reg(query
, GS_INVOCATION_COUNT
,
645 "N geometry shader invocations");
646 add_basic_stat_reg(query
, GS_PRIMITIVES_COUNT
,
647 "N geometry shader primitives emitted");
649 add_basic_stat_reg(query
, CL_INVOCATION_COUNT
,
650 "N primitives entering clipping");
651 add_basic_stat_reg(query
, CL_PRIMITIVES_COUNT
,
652 "N primitives leaving clipping");
654 if (devinfo
->is_haswell
|| devinfo
->gen
== 8) {
655 add_stat_reg(query
, PS_INVOCATION_COUNT
, 1, 4,
656 "N fragment shader invocations",
657 "N fragment shader invocations");
659 add_basic_stat_reg(query
, PS_INVOCATION_COUNT
,
660 "N fragment shader invocations");
663 add_basic_stat_reg(query
, PS_DEPTH_COUNT
,
664 "N z-pass fragments");
666 if (devinfo
->gen
>= 7) {
667 add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
668 "N compute shader invocations");
671 query
->data_size
= sizeof(uint64_t) * query
->n_counters
;
675 load_oa_metrics(struct gen_perf_config
*perf
, int fd
,
676 const struct gen_device_info
*devinfo
)
678 perf_register_oa_queries_t oa_register
= get_register_queries_function(devinfo
);
679 bool i915_perf_oa_available
= false;
682 /* The existence of this sysctl parameter implies the kernel supports
683 * the i915 perf interface.
685 if (stat("/proc/sys/dev/i915/perf_stream_paranoid", &sb
) == 0) {
687 /* If _paranoid == 1 then on Gen8+ we won't be able to access OA
688 * metrics unless running as root.
690 if (devinfo
->is_haswell
)
691 i915_perf_oa_available
= true;
693 uint64_t paranoid
= 1;
695 read_file_uint64("/proc/sys/dev/i915/perf_stream_paranoid", ¶noid
);
697 if (paranoid
== 0 || geteuid() == 0)
698 i915_perf_oa_available
= true;
702 if (!i915_perf_oa_available
||
704 !get_sysfs_dev_dir(perf
, fd
) ||
705 !init_oa_sys_vars(perf
, devinfo
))
708 perf
->oa_metrics_table
=
709 _mesa_hash_table_create(perf
, _mesa_key_hash_string
,
710 _mesa_key_string_equal
);
712 /* Index all the metric sets mesa knows about before looking to see what
713 * the kernel is advertising.
717 if (likely((INTEL_DEBUG
& DEBUG_NO_OACONFIG
) == 0) &&
718 kernel_has_dynamic_config_support(perf
, fd
))
719 init_oa_configs(perf
, fd
);
721 enumerate_sysfs_metrics(perf
);
726 /* Accumulate 32bits OA counters */
728 accumulate_uint32(const uint32_t *report0
,
729 const uint32_t *report1
,
730 uint64_t *accumulator
)
732 *accumulator
+= (uint32_t)(*report1
- *report0
);
735 /* Accumulate 40bits OA counters */
737 accumulate_uint40(int a_index
,
738 const uint32_t *report0
,
739 const uint32_t *report1
,
740 uint64_t *accumulator
)
742 const uint8_t *high_bytes0
= (uint8_t *)(report0
+ 40);
743 const uint8_t *high_bytes1
= (uint8_t *)(report1
+ 40);
744 uint64_t high0
= (uint64_t)(high_bytes0
[a_index
]) << 32;
745 uint64_t high1
= (uint64_t)(high_bytes1
[a_index
]) << 32;
746 uint64_t value0
= report0
[a_index
+ 4] | high0
;
747 uint64_t value1
= report1
[a_index
+ 4] | high1
;
751 delta
= (1ULL << 40) + value1
- value0
;
753 delta
= value1
- value0
;
755 *accumulator
+= delta
;
759 gen8_read_report_clock_ratios(const uint32_t *report
,
760 uint64_t *slice_freq_hz
,
761 uint64_t *unslice_freq_hz
)
763 /* The lower 16bits of the RPT_ID field of the OA reports contains a
764 * snapshot of the bits coming from the RP_FREQ_NORMAL register and is
767 * RPT_ID[31:25]: RP_FREQ_NORMAL[20:14] (low squashed_slice_clock_frequency)
768 * RPT_ID[10:9]: RP_FREQ_NORMAL[22:21] (high squashed_slice_clock_frequency)
769 * RPT_ID[8:0]: RP_FREQ_NORMAL[31:23] (squashed_unslice_clock_frequency)
771 * RP_FREQ_NORMAL[31:23]: Software Unslice Ratio Request
772 * Multiple of 33.33MHz 2xclk (16 MHz 1xclk)
774 * RP_FREQ_NORMAL[22:14]: Software Slice Ratio Request
775 * Multiple of 33.33MHz 2xclk (16 MHz 1xclk)
778 uint32_t unslice_freq
= report
[0] & 0x1ff;
779 uint32_t slice_freq_low
= (report
[0] >> 25) & 0x7f;
780 uint32_t slice_freq_high
= (report
[0] >> 9) & 0x3;
781 uint32_t slice_freq
= slice_freq_low
| (slice_freq_high
<< 7);
783 *slice_freq_hz
= slice_freq
* 16666667ULL;
784 *unslice_freq_hz
= unslice_freq
* 16666667ULL;
788 query_result_read_frequencies(struct gen_perf_query_result
*result
,
789 const struct gen_device_info
*devinfo
,
790 const uint32_t *start
,
793 /* Slice/Unslice frequency is only available in the OA reports when the
794 * "Disable OA reports due to clock ratio change" field in
795 * OA_DEBUG_REGISTER is set to 1. This is how the kernel programs this
796 * global register (see drivers/gpu/drm/i915/i915_perf.c)
798 * Documentation says this should be available on Gen9+ but experimentation
799 * shows that Gen8 reports similar values, so we enable it there too.
801 if (devinfo
->gen
< 8)
804 gen8_read_report_clock_ratios(start
,
805 &result
->slice_frequency
[0],
806 &result
->unslice_frequency
[0]);
807 gen8_read_report_clock_ratios(end
,
808 &result
->slice_frequency
[1],
809 &result
->unslice_frequency
[1]);
813 query_result_accumulate(struct gen_perf_query_result
*result
,
814 const struct gen_perf_query_info
*query
,
815 const uint32_t *start
,
820 result
->hw_id
= start
[2];
821 result
->reports_accumulated
++;
823 switch (query
->oa_format
) {
824 case I915_OA_FORMAT_A32u40_A4u32_B8_C8
:
825 accumulate_uint32(start
+ 1, end
+ 1, result
->accumulator
+ idx
++); /* timestamp */
826 accumulate_uint32(start
+ 3, end
+ 3, result
->accumulator
+ idx
++); /* clock */
828 /* 32x 40bit A counters... */
829 for (i
= 0; i
< 32; i
++)
830 accumulate_uint40(i
, start
, end
, result
->accumulator
+ idx
++);
832 /* 4x 32bit A counters... */
833 for (i
= 0; i
< 4; i
++)
834 accumulate_uint32(start
+ 36 + i
, end
+ 36 + i
, result
->accumulator
+ idx
++);
836 /* 8x 32bit B counters + 8x 32bit C counters... */
837 for (i
= 0; i
< 16; i
++)
838 accumulate_uint32(start
+ 48 + i
, end
+ 48 + i
, result
->accumulator
+ idx
++);
841 case I915_OA_FORMAT_A45_B8_C8
:
842 accumulate_uint32(start
+ 1, end
+ 1, result
->accumulator
); /* timestamp */
844 for (i
= 0; i
< 61; i
++)
845 accumulate_uint32(start
+ 3 + i
, end
+ 3 + i
, result
->accumulator
+ 1 + i
);
849 unreachable("Can't accumulate OA counters in unknown format");
855 query_result_clear(struct gen_perf_query_result
*result
)
857 memset(result
, 0, sizeof(*result
));
858 result
->hw_id
= 0xffffffff; /* invalid */
862 register_mdapi_statistic_query(struct gen_perf_config
*perf_cfg
,
863 const struct gen_device_info
*devinfo
)
865 if (!(devinfo
->gen
>= 7 && devinfo
->gen
<= 11))
868 struct gen_perf_query_info
*query
=
869 append_query_info(perf_cfg
, MAX_STAT_COUNTERS
);
871 query
->kind
= GEN_PERF_QUERY_TYPE_PIPELINE
;
872 query
->name
= "Intel_Raw_Pipeline_Statistics_Query";
874 /* The order has to match mdapi_pipeline_metrics. */
875 add_basic_stat_reg(query
, IA_VERTICES_COUNT
,
876 "N vertices submitted");
877 add_basic_stat_reg(query
, IA_PRIMITIVES_COUNT
,
878 "N primitives submitted");
879 add_basic_stat_reg(query
, VS_INVOCATION_COUNT
,
880 "N vertex shader invocations");
881 add_basic_stat_reg(query
, GS_INVOCATION_COUNT
,
882 "N geometry shader invocations");
883 add_basic_stat_reg(query
, GS_PRIMITIVES_COUNT
,
884 "N geometry shader primitives emitted");
885 add_basic_stat_reg(query
, CL_INVOCATION_COUNT
,
886 "N primitives entering clipping");
887 add_basic_stat_reg(query
, CL_PRIMITIVES_COUNT
,
888 "N primitives leaving clipping");
889 if (devinfo
->is_haswell
|| devinfo
->gen
== 8) {
890 add_stat_reg(query
, PS_INVOCATION_COUNT
, 1, 4,
891 "N fragment shader invocations",
892 "N fragment shader invocations");
894 add_basic_stat_reg(query
, PS_INVOCATION_COUNT
,
895 "N fragment shader invocations");
897 add_basic_stat_reg(query
, HS_INVOCATION_COUNT
,
898 "N TCS shader invocations");
899 add_basic_stat_reg(query
, DS_INVOCATION_COUNT
,
900 "N TES shader invocations");
901 if (devinfo
->gen
>= 7) {
902 add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
903 "N compute shader invocations");
906 if (devinfo
->gen
>= 10) {
907 /* Reuse existing CS invocation register until we can expose this new
910 add_basic_stat_reg(query
, CS_INVOCATION_COUNT
,
914 query
->data_size
= sizeof(uint64_t) * query
->n_counters
;
918 fill_mdapi_perf_query_counter(struct gen_perf_query_info
*query
,
920 uint32_t data_offset
,
922 enum gen_perf_counter_data_type data_type
)
924 struct gen_perf_query_counter
*counter
= &query
->counters
[query
->n_counters
];
926 assert(query
->n_counters
<= query
->max_counters
);
928 counter
->name
= name
;
929 counter
->desc
= "Raw counter value";
930 counter
->type
= GEN_PERF_COUNTER_TYPE_RAW
;
931 counter
->data_type
= data_type
;
932 counter
->offset
= data_offset
;
936 assert(counter
->offset
+ gen_perf_query_counter_get_size(counter
) <= query
->data_size
);
939 #define MDAPI_QUERY_ADD_COUNTER(query, struct_name, field_name, type_name) \
940 fill_mdapi_perf_query_counter(query, #field_name, \
941 (uint8_t *) &struct_name.field_name - \
942 (uint8_t *) &struct_name, \
943 sizeof(struct_name.field_name), \
944 GEN_PERF_COUNTER_DATA_TYPE_##type_name)
945 #define MDAPI_QUERY_ADD_ARRAY_COUNTER(ctx, query, struct_name, field_name, idx, type_name) \
946 fill_mdapi_perf_query_counter(query, \
947 ralloc_asprintf(ctx, "%s%i", #field_name, idx), \
948 (uint8_t *) &struct_name.field_name[idx] - \
949 (uint8_t *) &struct_name, \
950 sizeof(struct_name.field_name[0]), \
951 GEN_PERF_COUNTER_DATA_TYPE_##type_name)
954 register_mdapi_oa_query(const struct gen_device_info
*devinfo
,
955 struct gen_perf_config
*perf
)
957 struct gen_perf_query_info
*query
= NULL
;
959 /* MDAPI requires different structures for pretty much every generation
960 * (right now we have definitions for gen 7 to 11).
962 if (!(devinfo
->gen
>= 7 && devinfo
->gen
<= 11))
965 switch (devinfo
->gen
) {
967 query
= append_query_info(perf
, 1 + 45 + 16 + 7);
968 query
->oa_format
= I915_OA_FORMAT_A45_B8_C8
;
970 struct gen7_mdapi_metrics metric_data
;
971 query
->data_size
= sizeof(metric_data
);
973 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, TotalTime
, UINT64
);
974 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.ACounters
); i
++) {
975 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
976 metric_data
, ACounters
, i
, UINT64
);
978 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.NOACounters
); i
++) {
979 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
980 metric_data
, NOACounters
, i
, UINT64
);
982 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter1
, UINT64
);
983 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter2
, UINT64
);
984 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SplitOccured
, BOOL32
);
985 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequencyChanged
, BOOL32
);
986 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequency
, UINT64
);
987 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportId
, UINT32
);
988 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportsCount
, UINT32
);
992 query
= append_query_info(perf
, 2 + 36 + 16 + 16);
993 query
->oa_format
= I915_OA_FORMAT_A32u40_A4u32_B8_C8
;
995 struct gen8_mdapi_metrics metric_data
;
996 query
->data_size
= sizeof(metric_data
);
998 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, TotalTime
, UINT64
);
999 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, GPUTicks
, UINT64
);
1000 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.OaCntr
); i
++) {
1001 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1002 metric_data
, OaCntr
, i
, UINT64
);
1004 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.NoaCntr
); i
++) {
1005 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1006 metric_data
, NoaCntr
, i
, UINT64
);
1008 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, BeginTimestamp
, UINT64
);
1009 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved1
, UINT64
);
1010 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved2
, UINT64
);
1011 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved3
, UINT32
);
1012 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, OverrunOccured
, BOOL32
);
1013 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerUser
, UINT64
);
1014 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerDriver
, UINT64
);
1015 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SliceFrequency
, UINT64
);
1016 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, UnsliceFrequency
, UINT64
);
1017 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter1
, UINT64
);
1018 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter2
, UINT64
);
1019 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SplitOccured
, BOOL32
);
1020 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequencyChanged
, BOOL32
);
1021 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequency
, UINT64
);
1022 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportId
, UINT32
);
1023 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportsCount
, UINT32
);
1029 query
= append_query_info(perf
, 2 + 36 + 16 + 16 + 16 + 2);
1030 query
->oa_format
= I915_OA_FORMAT_A32u40_A4u32_B8_C8
;
1032 struct gen9_mdapi_metrics metric_data
;
1033 query
->data_size
= sizeof(metric_data
);
1035 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, TotalTime
, UINT64
);
1036 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, GPUTicks
, UINT64
);
1037 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.OaCntr
); i
++) {
1038 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1039 metric_data
, OaCntr
, i
, UINT64
);
1041 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.NoaCntr
); i
++) {
1042 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1043 metric_data
, NoaCntr
, i
, UINT64
);
1045 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, BeginTimestamp
, UINT64
);
1046 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved1
, UINT64
);
1047 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved2
, UINT64
);
1048 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved3
, UINT32
);
1049 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, OverrunOccured
, BOOL32
);
1050 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerUser
, UINT64
);
1051 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, MarkerDriver
, UINT64
);
1052 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SliceFrequency
, UINT64
);
1053 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, UnsliceFrequency
, UINT64
);
1054 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter1
, UINT64
);
1055 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, PerfCounter2
, UINT64
);
1056 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, SplitOccured
, BOOL32
);
1057 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequencyChanged
, BOOL32
);
1058 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, CoreFrequency
, UINT64
);
1059 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportId
, UINT32
);
1060 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, ReportsCount
, UINT32
);
1061 for (int i
= 0; i
< ARRAY_SIZE(metric_data
.UserCntr
); i
++) {
1062 MDAPI_QUERY_ADD_ARRAY_COUNTER(perf
->queries
, query
,
1063 metric_data
, UserCntr
, i
, UINT64
);
1065 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, UserCntrCfgId
, UINT32
);
1066 MDAPI_QUERY_ADD_COUNTER(query
, metric_data
, Reserved4
, UINT32
);
1070 unreachable("Unsupported gen");
1074 query
->kind
= GEN_PERF_QUERY_TYPE_RAW
;
1075 query
->name
= "Intel_Raw_Hardware_Counters_Set_0_Query";
1076 query
->guid
= GEN_PERF_QUERY_GUID_MDAPI
;
1079 /* Accumulation buffer offsets copied from an actual query... */
1080 const struct gen_perf_query_info
*copy_query
=
1083 query
->gpu_time_offset
= copy_query
->gpu_time_offset
;
1084 query
->gpu_clock_offset
= copy_query
->gpu_clock_offset
;
1085 query
->a_offset
= copy_query
->a_offset
;
1086 query
->b_offset
= copy_query
->b_offset
;
1087 query
->c_offset
= copy_query
->c_offset
;
1092 get_metric_id(struct gen_perf_config
*perf
,
1093 const struct gen_perf_query_info
*query
)
1095 /* These queries are know not to ever change, their config ID has been
1096 * loaded upon the first query creation. No need to look them up again.
1098 if (query
->kind
== GEN_PERF_QUERY_TYPE_OA
)
1099 return query
->oa_metrics_set_id
;
1101 assert(query
->kind
== GEN_PERF_QUERY_TYPE_RAW
);
1103 /* Raw queries can be reprogrammed up by an external application/library.
1104 * When a raw query is used for the first time it's id is set to a value !=
1105 * 0. When it stops being used the id returns to 0. No need to reload the
1106 * ID when it's already loaded.
1108 if (query
->oa_metrics_set_id
!= 0) {
1109 DBG("Raw query '%s' guid=%s using cached ID: %"PRIu64
"\n",
1110 query
->name
, query
->guid
, query
->oa_metrics_set_id
);
1111 return query
->oa_metrics_set_id
;
1114 struct gen_perf_query_info
*raw_query
= (struct gen_perf_query_info
*)query
;
1115 if (!load_metric_id(perf
, query
->guid
,
1116 &raw_query
->oa_metrics_set_id
)) {
1117 DBG("Unable to read query guid=%s ID, falling back to test config\n", query
->guid
);
1118 raw_query
->oa_metrics_set_id
= 1ULL;
1120 DBG("Raw query '%s'guid=%s loaded ID: %"PRIu64
"\n",
1121 query
->name
, query
->guid
, query
->oa_metrics_set_id
);
1123 return query
->oa_metrics_set_id
;
1126 static struct oa_sample_buf
*
1127 get_free_sample_buf(struct gen_perf_context
*perf_ctx
)
1129 struct exec_node
*node
= exec_list_pop_head(&perf_ctx
->free_sample_buffers
);
1130 struct oa_sample_buf
*buf
;
1133 buf
= exec_node_data(struct oa_sample_buf
, node
, link
);
1135 buf
= ralloc_size(perf_ctx
->perf
, sizeof(*buf
));
1137 exec_node_init(&buf
->link
);
1146 reap_old_sample_buffers(struct gen_perf_context
*perf_ctx
)
1148 struct exec_node
*tail_node
=
1149 exec_list_get_tail(&perf_ctx
->sample_buffers
);
1150 struct oa_sample_buf
*tail_buf
=
1151 exec_node_data(struct oa_sample_buf
, tail_node
, link
);
1153 /* Remove all old, unreferenced sample buffers walking forward from
1154 * the head of the list, except always leave at least one node in
1155 * the list so we always have a node to reference when we Begin
1158 foreach_list_typed_safe(struct oa_sample_buf
, buf
, link
,
1159 &perf_ctx
->sample_buffers
)
1161 if (buf
->refcount
== 0 && buf
!= tail_buf
) {
1162 exec_node_remove(&buf
->link
);
1163 exec_list_push_head(&perf_ctx
->free_sample_buffers
, &buf
->link
);
1170 free_sample_bufs(struct gen_perf_context
*perf_ctx
)
1172 foreach_list_typed_safe(struct oa_sample_buf
, buf
, link
,
1173 &perf_ctx
->free_sample_buffers
)
1176 exec_list_make_empty(&perf_ctx
->free_sample_buffers
);
1179 /******************************************************************************/
1182 * Emit MI_STORE_REGISTER_MEM commands to capture all of the
1183 * pipeline statistics for the performance query object.
1186 snapshot_statistics_registers(void *context
,
1187 struct gen_perf_config
*perf
,
1188 struct gen_perf_query_object
*obj
,
1189 uint32_t offset_in_bytes
)
1191 const struct gen_perf_query_info
*query
= obj
->queryinfo
;
1192 const int n_counters
= query
->n_counters
;
1194 for (int i
= 0; i
< n_counters
; i
++) {
1195 const struct gen_perf_query_counter
*counter
= &query
->counters
[i
];
1197 assert(counter
->data_type
== GEN_PERF_COUNTER_DATA_TYPE_UINT64
);
1199 perf
->vtbl
.store_register_mem64(context
, obj
->pipeline_stats
.bo
,
1200 counter
->pipeline_stat
.reg
,
1201 offset_in_bytes
+ i
* sizeof(uint64_t));
1206 gen_perf_close(struct gen_perf_context
*perfquery
,
1207 const struct gen_perf_query_info
*query
)
1209 if (perfquery
->oa_stream_fd
!= -1) {
1210 close(perfquery
->oa_stream_fd
);
1211 perfquery
->oa_stream_fd
= -1;
1213 if (query
->kind
== GEN_PERF_QUERY_TYPE_RAW
) {
1214 struct gen_perf_query_info
*raw_query
=
1215 (struct gen_perf_query_info
*) query
;
1216 raw_query
->oa_metrics_set_id
= 0;
1221 gen_perf_open(struct gen_perf_context
*perf_ctx
,
1224 int period_exponent
,
1228 uint64_t properties
[] = {
1229 /* Single context sampling */
1230 DRM_I915_PERF_PROP_CTX_HANDLE
, ctx_id
,
1232 /* Include OA reports in samples */
1233 DRM_I915_PERF_PROP_SAMPLE_OA
, true,
1235 /* OA unit configuration */
1236 DRM_I915_PERF_PROP_OA_METRICS_SET
, metrics_set_id
,
1237 DRM_I915_PERF_PROP_OA_FORMAT
, report_format
,
1238 DRM_I915_PERF_PROP_OA_EXPONENT
, period_exponent
,
1240 struct drm_i915_perf_open_param param
= {
1241 .flags
= I915_PERF_FLAG_FD_CLOEXEC
|
1242 I915_PERF_FLAG_FD_NONBLOCK
|
1243 I915_PERF_FLAG_DISABLED
,
1244 .num_properties
= ARRAY_SIZE(properties
) / 2,
1245 .properties_ptr
= (uintptr_t) properties
,
1247 int fd
= gen_ioctl(drm_fd
, DRM_IOCTL_I915_PERF_OPEN
, ¶m
);
1249 DBG("Error opening gen perf OA stream: %m\n");
1253 perf_ctx
->oa_stream_fd
= fd
;
1255 perf_ctx
->current_oa_metrics_set_id
= metrics_set_id
;
1256 perf_ctx
->current_oa_format
= report_format
;
1262 inc_n_users(struct gen_perf_context
*perf_ctx
)
1264 if (perf_ctx
->n_oa_users
== 0 &&
1265 gen_ioctl(perf_ctx
->oa_stream_fd
, I915_PERF_IOCTL_ENABLE
, 0) < 0)
1269 ++perf_ctx
->n_oa_users
;
1275 dec_n_users(struct gen_perf_context
*perf_ctx
)
1277 /* Disabling the i915 perf stream will effectively disable the OA
1278 * counters. Note it's important to be sure there are no outstanding
1279 * MI_RPC commands at this point since they could stall the CS
1280 * indefinitely once OACONTROL is disabled.
1282 --perf_ctx
->n_oa_users
;
1283 if (perf_ctx
->n_oa_users
== 0 &&
1284 gen_ioctl(perf_ctx
->oa_stream_fd
, I915_PERF_IOCTL_DISABLE
, 0) < 0)
1286 DBG("WARNING: Error disabling gen perf stream: %m\n");
1291 gen_perf_init_metrics(struct gen_perf_config
*perf_cfg
,
1292 const struct gen_device_info
*devinfo
,
1295 load_pipeline_statistic_metrics(perf_cfg
, devinfo
);
1296 register_mdapi_statistic_query(perf_cfg
, devinfo
);
1297 if (load_oa_metrics(perf_cfg
, drm_fd
, devinfo
))
1298 register_mdapi_oa_query(devinfo
, perf_cfg
);
1302 gen_perf_init_context(struct gen_perf_context
*perf_ctx
,
1303 struct gen_perf_config
*perf_cfg
,
1304 void * ctx
, /* driver context (eg, brw_context) */
1305 void * bufmgr
, /* eg brw_bufmgr */
1306 const struct gen_device_info
*devinfo
,
1310 perf_ctx
->perf
= perf_cfg
;
1311 perf_ctx
->ctx
= ctx
;
1312 perf_ctx
->bufmgr
= bufmgr
;
1313 perf_ctx
->drm_fd
= drm_fd
;
1314 perf_ctx
->hw_ctx
= hw_ctx
;
1315 perf_ctx
->devinfo
= devinfo
;
1317 perf_ctx
->unaccumulated
=
1318 ralloc_array(ctx
, struct gen_perf_query_object
*, 2);
1319 perf_ctx
->unaccumulated_elements
= 0;
1320 perf_ctx
->unaccumulated_array_size
= 2;
1322 exec_list_make_empty(&perf_ctx
->sample_buffers
);
1323 exec_list_make_empty(&perf_ctx
->free_sample_buffers
);
1325 /* It's convenient to guarantee that this linked list of sample
1326 * buffers is never empty so we add an empty head so when we
1327 * Begin an OA query we can always take a reference on a buffer
1330 struct oa_sample_buf
*buf
= get_free_sample_buf(perf_ctx
);
1331 exec_list_push_head(&perf_ctx
->sample_buffers
, &buf
->link
);
1333 perf_ctx
->oa_stream_fd
= -1;
1334 perf_ctx
->next_query_start_report_id
= 1000;
1338 * Add a query to the global list of "unaccumulated queries."
1340 * Queries are tracked here until all the associated OA reports have
1341 * been accumulated via accumulate_oa_reports() after the end
1342 * MI_REPORT_PERF_COUNT has landed in query->oa.bo.
1345 add_to_unaccumulated_query_list(struct gen_perf_context
*perf_ctx
,
1346 struct gen_perf_query_object
*obj
)
1348 if (perf_ctx
->unaccumulated_elements
>=
1349 perf_ctx
->unaccumulated_array_size
)
1351 perf_ctx
->unaccumulated_array_size
*= 1.5;
1352 perf_ctx
->unaccumulated
=
1353 reralloc(perf_ctx
->ctx
, perf_ctx
->unaccumulated
,
1354 struct gen_perf_query_object
*,
1355 perf_ctx
->unaccumulated_array_size
);
1358 perf_ctx
->unaccumulated
[perf_ctx
->unaccumulated_elements
++] = obj
;
1362 gen_perf_begin_query(struct gen_perf_context
*perf_ctx
,
1363 struct gen_perf_query_object
*query
)
1365 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1366 const struct gen_perf_query_info
*queryinfo
= query
->queryinfo
;
1368 /* XXX: We have to consider that the command parser unit that parses batch
1369 * buffer commands and is used to capture begin/end counter snapshots isn't
1370 * implicitly synchronized with what's currently running across other GPU
1371 * units (such as the EUs running shaders) that the performance counters are
1374 * The intention of performance queries is to measure the work associated
1375 * with commands between the begin/end delimiters and so for that to be the
1376 * case we need to explicitly synchronize the parsing of commands to capture
1377 * Begin/End counter snapshots with what's running across other parts of the
1380 * When the command parser reaches a Begin marker it effectively needs to
1381 * drain everything currently running on the GPU until the hardware is idle
1382 * before capturing the first snapshot of counters - otherwise the results
1383 * would also be measuring the effects of earlier commands.
1385 * When the command parser reaches an End marker it needs to stall until
1386 * everything currently running on the GPU has finished before capturing the
1387 * end snapshot - otherwise the results won't be a complete representation
1390 * Theoretically there could be opportunities to minimize how much of the
1391 * GPU pipeline is drained, or that we stall for, when we know what specific
1392 * units the performance counters being queried relate to but we don't
1393 * currently attempt to be clever here.
1395 * Note: with our current simple approach here then for back-to-back queries
1396 * we will redundantly emit duplicate commands to synchronize the command
1397 * streamer with the rest of the GPU pipeline, but we assume that in HW the
1398 * second synchronization is effectively a NOOP.
1400 * N.B. The final results are based on deltas of counters between (inside)
1401 * Begin/End markers so even though the total wall clock time of the
1402 * workload is stretched by larger pipeline bubbles the bubbles themselves
1403 * are generally invisible to the query results. Whether that's a good or a
1404 * bad thing depends on the use case. For a lower real-time impact while
1405 * capturing metrics then periodic sampling may be a better choice than
1406 * INTEL_performance_query.
1409 * This is our Begin synchronization point to drain current work on the
1410 * GPU before we capture our first counter snapshot...
1412 perf_cfg
->vtbl
.emit_mi_flush(perf_ctx
->ctx
);
1414 switch (queryinfo
->kind
) {
1415 case GEN_PERF_QUERY_TYPE_OA
:
1416 case GEN_PERF_QUERY_TYPE_RAW
: {
1418 /* Opening an i915 perf stream implies exclusive access to the OA unit
1419 * which will generate counter reports for a specific counter set with a
1420 * specific layout/format so we can't begin any OA based queries that
1421 * require a different counter set or format unless we get an opportunity
1422 * to close the stream and open a new one...
1424 uint64_t metric_id
= get_metric_id(perf_ctx
->perf
, queryinfo
);
1426 if (perf_ctx
->oa_stream_fd
!= -1 &&
1427 perf_ctx
->current_oa_metrics_set_id
!= metric_id
) {
1429 if (perf_ctx
->n_oa_users
!= 0) {
1430 DBG("WARNING: Begin failed already using perf config=%i/%"PRIu64
"\n",
1431 perf_ctx
->current_oa_metrics_set_id
, metric_id
);
1434 gen_perf_close(perf_ctx
, queryinfo
);
1437 /* If the OA counters aren't already on, enable them. */
1438 if (perf_ctx
->oa_stream_fd
== -1) {
1439 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
1441 /* The period_exponent gives a sampling period as follows:
1442 * sample_period = timestamp_period * 2^(period_exponent + 1)
1444 * The timestamps increments every 80ns (HSW), ~52ns (GEN9LP) or
1447 * The counter overflow period is derived from the EuActive counter
1448 * which reads a counter that increments by the number of clock
1449 * cycles multiplied by the number of EUs. It can be calculated as:
1451 * 2^(number of bits in A counter) / (n_eus * max_gen_freq * 2)
1453 * (E.g. 40 EUs @ 1GHz = ~53ms)
1455 * We select a sampling period inferior to that overflow period to
1456 * ensure we cannot see more than 1 counter overflow, otherwise we
1457 * could loose information.
1460 int a_counter_in_bits
= 32;
1461 if (devinfo
->gen
>= 8)
1462 a_counter_in_bits
= 40;
1464 uint64_t overflow_period
= pow(2, a_counter_in_bits
) / (perf_cfg
->sys_vars
.n_eus
*
1465 /* drop 1GHz freq to have units in nanoseconds */
1468 DBG("A counter overflow period: %"PRIu64
"ns, %"PRIu64
"ms (n_eus=%"PRIu64
")\n",
1469 overflow_period
, overflow_period
/ 1000000ul, perf_cfg
->sys_vars
.n_eus
);
1471 int period_exponent
= 0;
1472 uint64_t prev_sample_period
, next_sample_period
;
1473 for (int e
= 0; e
< 30; e
++) {
1474 prev_sample_period
= 1000000000ull * pow(2, e
+ 1) / devinfo
->timestamp_frequency
;
1475 next_sample_period
= 1000000000ull * pow(2, e
+ 2) / devinfo
->timestamp_frequency
;
1477 /* Take the previous sampling period, lower than the overflow
1480 if (prev_sample_period
< overflow_period
&&
1481 next_sample_period
> overflow_period
)
1482 period_exponent
= e
+ 1;
1485 if (period_exponent
== 0) {
1486 DBG("WARNING: enable to find a sampling exponent\n");
1490 DBG("OA sampling exponent: %i ~= %"PRIu64
"ms\n", period_exponent
,
1491 prev_sample_period
/ 1000000ul);
1493 if (!gen_perf_open(perf_ctx
, metric_id
, queryinfo
->oa_format
,
1494 period_exponent
, perf_ctx
->drm_fd
,
1498 assert(perf_ctx
->current_oa_metrics_set_id
== metric_id
&&
1499 perf_ctx
->current_oa_format
== queryinfo
->oa_format
);
1502 if (!inc_n_users(perf_ctx
)) {
1503 DBG("WARNING: Error enabling i915 perf stream: %m\n");
1508 perf_cfg
->vtbl
.bo_unreference(query
->oa
.bo
);
1509 query
->oa
.bo
= NULL
;
1512 query
->oa
.bo
= perf_cfg
->vtbl
.bo_alloc(perf_ctx
->bufmgr
,
1513 "perf. query OA MI_RPC bo",
1516 /* Pre-filling the BO helps debug whether writes landed. */
1517 void *map
= perf_cfg
->vtbl
.bo_map(perf_ctx
->ctx
, query
->oa
.bo
, MAP_WRITE
);
1518 memset(map
, 0x80, MI_RPC_BO_SIZE
);
1519 perf_cfg
->vtbl
.bo_unmap(query
->oa
.bo
);
1522 query
->oa
.begin_report_id
= perf_ctx
->next_query_start_report_id
;
1523 perf_ctx
->next_query_start_report_id
+= 2;
1525 /* We flush the batchbuffer here to minimize the chances that MI_RPC
1526 * delimiting commands end up in different batchbuffers. If that's the
1527 * case, the measurement will include the time it takes for the kernel
1528 * scheduler to load a new request into the hardware. This is manifested in
1529 * tools like frameretrace by spikes in the "GPU Core Clocks" counter.
1531 perf_cfg
->vtbl
.batchbuffer_flush(perf_ctx
->ctx
, __FILE__
, __LINE__
);
1533 /* Take a starting OA counter snapshot. */
1534 perf_cfg
->vtbl
.emit_mi_report_perf_count(perf_ctx
->ctx
, query
->oa
.bo
, 0,
1535 query
->oa
.begin_report_id
);
1536 perf_cfg
->vtbl
.capture_frequency_stat_register(perf_ctx
->ctx
, query
->oa
.bo
,
1537 MI_FREQ_START_OFFSET_BYTES
);
1539 ++perf_ctx
->n_active_oa_queries
;
1541 /* No already-buffered samples can possibly be associated with this query
1542 * so create a marker within the list of sample buffers enabling us to
1543 * easily ignore earlier samples when processing this query after
1546 assert(!exec_list_is_empty(&perf_ctx
->sample_buffers
));
1547 query
->oa
.samples_head
= exec_list_get_tail(&perf_ctx
->sample_buffers
);
1549 struct oa_sample_buf
*buf
=
1550 exec_node_data(struct oa_sample_buf
, query
->oa
.samples_head
, link
);
1552 /* This reference will ensure that future/following sample
1553 * buffers (that may relate to this query) can't be freed until
1554 * this drops to zero.
1558 query_result_clear(&query
->oa
.result
);
1559 query
->oa
.results_accumulated
= false;
1561 add_to_unaccumulated_query_list(perf_ctx
, query
);
1565 case GEN_PERF_QUERY_TYPE_PIPELINE
:
1566 if (query
->pipeline_stats
.bo
) {
1567 perf_cfg
->vtbl
.bo_unreference(query
->pipeline_stats
.bo
);
1568 query
->pipeline_stats
.bo
= NULL
;
1571 query
->pipeline_stats
.bo
=
1572 perf_cfg
->vtbl
.bo_alloc(perf_ctx
->bufmgr
,
1573 "perf. query pipeline stats bo",
1576 /* Take starting snapshots. */
1577 snapshot_statistics_registers(perf_ctx
->ctx
, perf_cfg
, query
, 0);
1579 ++perf_ctx
->n_active_pipeline_stats_queries
;
1583 unreachable("Unknown query type");
1591 gen_perf_end_query(struct gen_perf_context
*perf_ctx
,
1592 struct gen_perf_query_object
*query
)
1594 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1596 /* Ensure that the work associated with the queried commands will have
1597 * finished before taking our query end counter readings.
1599 * For more details see comment in brw_begin_perf_query for
1600 * corresponding flush.
1602 perf_cfg
->vtbl
.emit_mi_flush(perf_ctx
->ctx
);
1604 switch (query
->queryinfo
->kind
) {
1605 case GEN_PERF_QUERY_TYPE_OA
:
1606 case GEN_PERF_QUERY_TYPE_RAW
:
1608 /* NB: It's possible that the query will have already been marked
1609 * as 'accumulated' if an error was seen while reading samples
1610 * from perf. In this case we mustn't try and emit a closing
1611 * MI_RPC command in case the OA unit has already been disabled
1613 if (!query
->oa
.results_accumulated
) {
1614 /* Take an ending OA counter snapshot. */
1615 perf_cfg
->vtbl
.capture_frequency_stat_register(perf_ctx
->ctx
, query
->oa
.bo
,
1616 MI_FREQ_END_OFFSET_BYTES
);
1617 perf_cfg
->vtbl
.emit_mi_report_perf_count(perf_ctx
->ctx
, query
->oa
.bo
,
1618 MI_RPC_BO_END_OFFSET_BYTES
,
1619 query
->oa
.begin_report_id
+ 1);
1622 --perf_ctx
->n_active_oa_queries
;
1624 /* NB: even though the query has now ended, it can't be accumulated
1625 * until the end MI_REPORT_PERF_COUNT snapshot has been written
1630 case GEN_PERF_QUERY_TYPE_PIPELINE
:
1631 snapshot_statistics_registers(perf_ctx
->ctx
, perf_cfg
, query
,
1632 STATS_BO_END_OFFSET_BYTES
);
1633 --perf_ctx
->n_active_pipeline_stats_queries
;
1637 unreachable("Unknown query type");
1643 OA_READ_STATUS_ERROR
,
1644 OA_READ_STATUS_UNFINISHED
,
1645 OA_READ_STATUS_FINISHED
,
1648 static enum OaReadStatus
1649 read_oa_samples_until(struct gen_perf_context
*perf_ctx
,
1650 uint32_t start_timestamp
,
1651 uint32_t end_timestamp
)
1653 struct exec_node
*tail_node
=
1654 exec_list_get_tail(&perf_ctx
->sample_buffers
);
1655 struct oa_sample_buf
*tail_buf
=
1656 exec_node_data(struct oa_sample_buf
, tail_node
, link
);
1657 uint32_t last_timestamp
= tail_buf
->last_timestamp
;
1660 struct oa_sample_buf
*buf
= get_free_sample_buf(perf_ctx
);
1664 while ((len
= read(perf_ctx
->oa_stream_fd
, buf
->buf
,
1665 sizeof(buf
->buf
))) < 0 && errno
== EINTR
)
1669 exec_list_push_tail(&perf_ctx
->free_sample_buffers
, &buf
->link
);
1672 if (errno
== EAGAIN
)
1673 return ((last_timestamp
- start_timestamp
) >=
1674 (end_timestamp
- start_timestamp
)) ?
1675 OA_READ_STATUS_FINISHED
:
1676 OA_READ_STATUS_UNFINISHED
;
1678 DBG("Error reading i915 perf samples: %m\n");
1681 DBG("Spurious EOF reading i915 perf samples\n");
1683 return OA_READ_STATUS_ERROR
;
1687 exec_list_push_tail(&perf_ctx
->sample_buffers
, &buf
->link
);
1689 /* Go through the reports and update the last timestamp. */
1691 while (offset
< buf
->len
) {
1692 const struct drm_i915_perf_record_header
*header
=
1693 (const struct drm_i915_perf_record_header
*) &buf
->buf
[offset
];
1694 uint32_t *report
= (uint32_t *) (header
+ 1);
1696 if (header
->type
== DRM_I915_PERF_RECORD_SAMPLE
)
1697 last_timestamp
= report
[1];
1699 offset
+= header
->size
;
1702 buf
->last_timestamp
= last_timestamp
;
1705 unreachable("not reached");
1706 return OA_READ_STATUS_ERROR
;
1710 * Try to read all the reports until either the delimiting timestamp
1711 * or an error arises.
1714 read_oa_samples_for_query(struct gen_perf_context
*perf_ctx
,
1715 struct gen_perf_query_object
*query
,
1716 void *current_batch
)
1721 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1723 /* We need the MI_REPORT_PERF_COUNT to land before we can start
1725 assert(!perf_cfg
->vtbl
.batch_references(current_batch
, query
->oa
.bo
) &&
1726 !perf_cfg
->vtbl
.bo_busy(query
->oa
.bo
));
1728 /* Map the BO once here and let accumulate_oa_reports() unmap
1730 if (query
->oa
.map
== NULL
)
1731 query
->oa
.map
= perf_cfg
->vtbl
.bo_map(perf_ctx
->ctx
, query
->oa
.bo
, MAP_READ
);
1733 start
= last
= query
->oa
.map
;
1734 end
= query
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
1736 if (start
[0] != query
->oa
.begin_report_id
) {
1737 DBG("Spurious start report id=%"PRIu32
"\n", start
[0]);
1740 if (end
[0] != (query
->oa
.begin_report_id
+ 1)) {
1741 DBG("Spurious end report id=%"PRIu32
"\n", end
[0]);
1745 /* Read the reports until the end timestamp. */
1746 switch (read_oa_samples_until(perf_ctx
, start
[1], end
[1])) {
1747 case OA_READ_STATUS_ERROR
:
1748 /* Fallthrough and let accumulate_oa_reports() deal with the
1750 case OA_READ_STATUS_FINISHED
:
1752 case OA_READ_STATUS_UNFINISHED
:
1756 unreachable("invalid read status");
1761 gen_perf_wait_query(struct gen_perf_context
*perf_ctx
,
1762 struct gen_perf_query_object
*query
,
1763 void *current_batch
)
1765 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1766 struct brw_bo
*bo
= NULL
;
1768 switch (query
->queryinfo
->kind
) {
1769 case GEN_PERF_QUERY_TYPE_OA
:
1770 case GEN_PERF_QUERY_TYPE_RAW
:
1774 case GEN_PERF_QUERY_TYPE_PIPELINE
:
1775 bo
= query
->pipeline_stats
.bo
;
1779 unreachable("Unknown query type");
1786 /* If the current batch references our results bo then we need to
1789 if (perf_cfg
->vtbl
.batch_references(current_batch
, bo
))
1790 perf_cfg
->vtbl
.batchbuffer_flush(perf_ctx
->ctx
, __FILE__
, __LINE__
);
1792 perf_cfg
->vtbl
.bo_wait_rendering(bo
);
1794 /* Due to a race condition between the OA unit signaling report
1795 * availability and the report actually being written into memory,
1796 * we need to wait for all the reports to come in before we can
1799 if (query
->queryinfo
->kind
== GEN_PERF_QUERY_TYPE_OA
||
1800 query
->queryinfo
->kind
== GEN_PERF_QUERY_TYPE_RAW
) {
1801 while (!read_oa_samples_for_query(perf_ctx
, query
, current_batch
))
1807 gen_perf_is_query_ready(struct gen_perf_context
*perf_ctx
,
1808 struct gen_perf_query_object
*query
,
1809 void *current_batch
)
1811 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
1813 switch (query
->queryinfo
->kind
) {
1814 case GEN_PERF_QUERY_TYPE_OA
:
1815 case GEN_PERF_QUERY_TYPE_RAW
:
1816 return (query
->oa
.results_accumulated
||
1818 !perf_cfg
->vtbl
.batch_references(current_batch
, query
->oa
.bo
) &&
1819 !perf_cfg
->vtbl
.bo_busy(query
->oa
.bo
) &&
1820 read_oa_samples_for_query(perf_ctx
, query
, current_batch
)));
1821 case GEN_PERF_QUERY_TYPE_PIPELINE
:
1822 return (query
->pipeline_stats
.bo
&&
1823 !perf_cfg
->vtbl
.batch_references(current_batch
, query
->pipeline_stats
.bo
) &&
1824 !perf_cfg
->vtbl
.bo_busy(query
->pipeline_stats
.bo
));
1827 unreachable("Unknown query type");
1835 * Remove a query from the global list of unaccumulated queries once
1836 * after successfully accumulating the OA reports associated with the
1837 * query in accumulate_oa_reports() or when discarding unwanted query
1841 drop_from_unaccumulated_query_list(struct gen_perf_context
*perf_ctx
,
1842 struct gen_perf_query_object
*query
)
1844 for (int i
= 0; i
< perf_ctx
->unaccumulated_elements
; i
++) {
1845 if (perf_ctx
->unaccumulated
[i
] == query
) {
1846 int last_elt
= --perf_ctx
->unaccumulated_elements
;
1849 perf_ctx
->unaccumulated
[i
] = NULL
;
1851 perf_ctx
->unaccumulated
[i
] =
1852 perf_ctx
->unaccumulated
[last_elt
];
1859 /* Drop our samples_head reference so that associated periodic
1860 * sample data buffers can potentially be reaped if they aren't
1861 * referenced by any other queries...
1864 struct oa_sample_buf
*buf
=
1865 exec_node_data(struct oa_sample_buf
, query
->oa
.samples_head
, link
);
1867 assert(buf
->refcount
> 0);
1870 query
->oa
.samples_head
= NULL
;
1872 reap_old_sample_buffers(perf_ctx
);
1875 /* In general if we see anything spurious while accumulating results,
1876 * we don't try and continue accumulating the current query, hoping
1877 * for the best, we scrap anything outstanding, and then hope for the
1878 * best with new queries.
1881 discard_all_queries(struct gen_perf_context
*perf_ctx
)
1883 while (perf_ctx
->unaccumulated_elements
) {
1884 struct gen_perf_query_object
*query
= perf_ctx
->unaccumulated
[0];
1886 query
->oa
.results_accumulated
= true;
1887 drop_from_unaccumulated_query_list(perf_ctx
, query
);
1889 dec_n_users(perf_ctx
);
1894 * Accumulate raw OA counter values based on deltas between pairs of
1897 * Accumulation starts from the first report captured via
1898 * MI_REPORT_PERF_COUNT (MI_RPC) by brw_begin_perf_query() until the
1899 * last MI_RPC report requested by brw_end_perf_query(). Between these
1900 * two reports there may also some number of periodically sampled OA
1901 * reports collected via the i915 perf interface - depending on the
1902 * duration of the query.
1904 * These periodic snapshots help to ensure we handle counter overflow
1905 * correctly by being frequent enough to ensure we don't miss multiple
1906 * overflows of a counter between snapshots. For Gen8+ the i915 perf
1907 * snapshots provide the extra context-switch reports that let us
1908 * subtract out the progress of counters associated with other
1909 * contexts running on the system.
1912 accumulate_oa_reports(struct gen_perf_context
*perf_ctx
,
1913 struct gen_perf_query_object
*query
)
1915 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
1919 struct exec_node
*first_samples_node
;
1921 int out_duration
= 0;
1923 assert(query
->oa
.map
!= NULL
);
1925 start
= last
= query
->oa
.map
;
1926 end
= query
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
1928 if (start
[0] != query
->oa
.begin_report_id
) {
1929 DBG("Spurious start report id=%"PRIu32
"\n", start
[0]);
1932 if (end
[0] != (query
->oa
.begin_report_id
+ 1)) {
1933 DBG("Spurious end report id=%"PRIu32
"\n", end
[0]);
1937 /* See if we have any periodic reports to accumulate too... */
1939 /* N.B. The oa.samples_head was set when the query began and
1940 * pointed to the tail of the perf_ctx->sample_buffers list at
1941 * the time the query started. Since the buffer existed before the
1942 * first MI_REPORT_PERF_COUNT command was emitted we therefore know
1943 * that no data in this particular node's buffer can possibly be
1944 * associated with the query - so skip ahead one...
1946 first_samples_node
= query
->oa
.samples_head
->next
;
1948 foreach_list_typed_from(struct oa_sample_buf
, buf
, link
,
1949 &perf_ctx
.sample_buffers
,
1954 while (offset
< buf
->len
) {
1955 const struct drm_i915_perf_record_header
*header
=
1956 (const struct drm_i915_perf_record_header
*)(buf
->buf
+ offset
);
1958 assert(header
->size
!= 0);
1959 assert(header
->size
<= buf
->len
);
1961 offset
+= header
->size
;
1963 switch (header
->type
) {
1964 case DRM_I915_PERF_RECORD_SAMPLE
: {
1965 uint32_t *report
= (uint32_t *)(header
+ 1);
1968 /* Ignore reports that come before the start marker.
1969 * (Note: takes care to allow overflow of 32bit timestamps)
1971 if (gen_device_info_timebase_scale(devinfo
,
1972 report
[1] - start
[1]) > 5000000000) {
1976 /* Ignore reports that come after the end marker.
1977 * (Note: takes care to allow overflow of 32bit timestamps)
1979 if (gen_device_info_timebase_scale(devinfo
,
1980 report
[1] - end
[1]) <= 5000000000) {
1984 /* For Gen8+ since the counters continue while other
1985 * contexts are running we need to discount any unrelated
1986 * deltas. The hardware automatically generates a report
1987 * on context switch which gives us a new reference point
1988 * to continuing adding deltas from.
1990 * For Haswell we can rely on the HW to stop the progress
1991 * of OA counters while any other context is acctive.
1993 if (devinfo
->gen
>= 8) {
1994 if (in_ctx
&& report
[2] != query
->oa
.result
.hw_id
) {
1995 DBG("i915 perf: Switch AWAY (observed by ID change)\n");
1998 } else if (in_ctx
== false && report
[2] == query
->oa
.result
.hw_id
) {
1999 DBG("i915 perf: Switch TO\n");
2002 /* From experimentation in IGT, we found that the OA unit
2003 * might label some report as "idle" (using an invalid
2004 * context ID), right after a report for a given context.
2005 * Deltas generated by those reports actually belong to the
2006 * previous context, even though they're not labelled as
2009 * We didn't *really* Switch AWAY in the case that we e.g.
2010 * saw a single periodic report while idle...
2012 if (out_duration
>= 1)
2014 } else if (in_ctx
) {
2015 assert(report
[2] == query
->oa
.result
.hw_id
);
2016 DBG("i915 perf: Continuation IN\n");
2018 assert(report
[2] != query
->oa
.result
.hw_id
);
2019 DBG("i915 perf: Continuation OUT\n");
2026 query_result_accumulate(&query
->oa
.result
, query
->queryinfo
,
2035 case DRM_I915_PERF_RECORD_OA_BUFFER_LOST
:
2036 DBG("i915 perf: OA error: all reports lost\n");
2038 case DRM_I915_PERF_RECORD_OA_REPORT_LOST
:
2039 DBG("i915 perf: OA report lost\n");
2047 query_result_accumulate(&query
->oa
.result
, query
->queryinfo
,
2050 query
->oa
.results_accumulated
= true;
2051 drop_from_unaccumulated_query_list(perf_ctx
, query
);
2052 dec_n_users(perf_ctx
);
2058 discard_all_queries(perf_ctx
);
2062 gen_perf_delete_query(struct gen_perf_context
*perf_ctx
,
2063 struct gen_perf_query_object
*query
)
2065 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2067 /* We can assume that the frontend waits for a query to complete
2068 * before ever calling into here, so we don't have to worry about
2069 * deleting an in-flight query object.
2071 switch (query
->queryinfo
->kind
) {
2072 case GEN_PERF_QUERY_TYPE_OA
:
2073 case GEN_PERF_QUERY_TYPE_RAW
:
2075 if (!query
->oa
.results_accumulated
) {
2076 drop_from_unaccumulated_query_list(perf_ctx
, query
);
2077 dec_n_users(perf_ctx
);
2080 perf_cfg
->vtbl
.bo_unreference(query
->oa
.bo
);
2081 query
->oa
.bo
= NULL
;
2084 query
->oa
.results_accumulated
= false;
2087 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2088 if (query
->pipeline_stats
.bo
) {
2089 perf_cfg
->vtbl
.bo_unreference(query
->pipeline_stats
.bo
);
2090 query
->pipeline_stats
.bo
= NULL
;
2095 unreachable("Unknown query type");
2099 /* As an indication that the INTEL_performance_query extension is no
2100 * longer in use, it's a good time to free our cache of sample
2101 * buffers and close any current i915-perf stream.
2103 if (--perf_ctx
->n_query_instances
== 0) {
2104 free_sample_bufs(perf_ctx
);
2105 gen_perf_close(perf_ctx
, query
->queryinfo
);
2111 #define GET_FIELD(word, field) (((word) & field ## _MASK) >> field ## _SHIFT)
2114 read_gt_frequency(struct gen_perf_context
*perf_ctx
,
2115 struct gen_perf_query_object
*obj
)
2117 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
2118 uint32_t start
= *((uint32_t *)(obj
->oa
.map
+ MI_FREQ_START_OFFSET_BYTES
)),
2119 end
= *((uint32_t *)(obj
->oa
.map
+ MI_FREQ_END_OFFSET_BYTES
));
2121 switch (devinfo
->gen
) {
2124 obj
->oa
.gt_frequency
[0] = GET_FIELD(start
, GEN7_RPSTAT1_CURR_GT_FREQ
) * 50ULL;
2125 obj
->oa
.gt_frequency
[1] = GET_FIELD(end
, GEN7_RPSTAT1_CURR_GT_FREQ
) * 50ULL;
2130 obj
->oa
.gt_frequency
[0] = GET_FIELD(start
, GEN9_RPSTAT0_CURR_GT_FREQ
) * 50ULL / 3ULL;
2131 obj
->oa
.gt_frequency
[1] = GET_FIELD(end
, GEN9_RPSTAT0_CURR_GT_FREQ
) * 50ULL / 3ULL;
2134 unreachable("unexpected gen");
2137 /* Put the numbers into Hz. */
2138 obj
->oa
.gt_frequency
[0] *= 1000000ULL;
2139 obj
->oa
.gt_frequency
[1] *= 1000000ULL;
2143 get_oa_counter_data(struct gen_perf_context
*perf_ctx
,
2144 struct gen_perf_query_object
*query
,
2148 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2149 const struct gen_perf_query_info
*queryinfo
= query
->queryinfo
;
2150 int n_counters
= queryinfo
->n_counters
;
2153 for (int i
= 0; i
< n_counters
; i
++) {
2154 const struct gen_perf_query_counter
*counter
= &queryinfo
->counters
[i
];
2155 uint64_t *out_uint64
;
2157 size_t counter_size
= gen_perf_query_counter_get_size(counter
);
2160 switch (counter
->data_type
) {
2161 case GEN_PERF_COUNTER_DATA_TYPE_UINT64
:
2162 out_uint64
= (uint64_t *)(data
+ counter
->offset
);
2164 counter
->oa_counter_read_uint64(perf_cfg
, queryinfo
,
2165 query
->oa
.result
.accumulator
);
2167 case GEN_PERF_COUNTER_DATA_TYPE_FLOAT
:
2168 out_float
= (float *)(data
+ counter
->offset
);
2170 counter
->oa_counter_read_float(perf_cfg
, queryinfo
,
2171 query
->oa
.result
.accumulator
);
2174 /* So far we aren't using uint32, double or bool32... */
2175 unreachable("unexpected counter data type");
2177 written
= counter
->offset
+ counter_size
;
2185 get_pipeline_stats_data(struct gen_perf_context
*perf_ctx
,
2186 struct gen_perf_query_object
*query
,
2191 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2192 const struct gen_perf_query_info
*queryinfo
= query
->queryinfo
;
2193 int n_counters
= queryinfo
->n_counters
;
2196 uint64_t *start
= perf_cfg
->vtbl
.bo_map(perf_ctx
->ctx
, query
->pipeline_stats
.bo
, MAP_READ
);
2197 uint64_t *end
= start
+ (STATS_BO_END_OFFSET_BYTES
/ sizeof(uint64_t));
2199 for (int i
= 0; i
< n_counters
; i
++) {
2200 const struct gen_perf_query_counter
*counter
= &queryinfo
->counters
[i
];
2201 uint64_t value
= end
[i
] - start
[i
];
2203 if (counter
->pipeline_stat
.numerator
!=
2204 counter
->pipeline_stat
.denominator
) {
2205 value
*= counter
->pipeline_stat
.numerator
;
2206 value
/= counter
->pipeline_stat
.denominator
;
2209 *((uint64_t *)p
) = value
;
2213 perf_cfg
->vtbl
.bo_unmap(query
->pipeline_stats
.bo
);
2219 gen_perf_get_query_data(struct gen_perf_context
*perf_ctx
,
2220 struct gen_perf_query_object
*query
,
2223 unsigned *bytes_written
)
2225 struct gen_perf_config
*perf_cfg
= perf_ctx
->perf
;
2228 switch (query
->queryinfo
->kind
) {
2229 case GEN_PERF_QUERY_TYPE_OA
:
2230 case GEN_PERF_QUERY_TYPE_RAW
:
2231 if (!query
->oa
.results_accumulated
) {
2232 read_gt_frequency(perf_ctx
, query
);
2233 uint32_t *begin_report
= query
->oa
.map
;
2234 uint32_t *end_report
= query
->oa
.map
+ MI_RPC_BO_END_OFFSET_BYTES
;
2235 query_result_read_frequencies(&query
->oa
.result
,
2239 accumulate_oa_reports(perf_ctx
, query
);
2240 assert(query
->oa
.results_accumulated
);
2242 perf_cfg
->vtbl
.bo_unmap(query
->oa
.bo
);
2243 query
->oa
.map
= NULL
;
2245 if (query
->queryinfo
->kind
== GEN_PERF_QUERY_TYPE_OA
) {
2246 written
= get_oa_counter_data(perf_ctx
, query
, data_size
, (uint8_t *)data
);
2248 const struct gen_device_info
*devinfo
= perf_ctx
->devinfo
;
2250 written
= gen_perf_query_result_write_mdapi((uint8_t *)data
, data_size
,
2251 devinfo
, &query
->oa
.result
,
2252 query
->oa
.gt_frequency
[0],
2253 query
->oa
.gt_frequency
[1]);
2257 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2258 written
= get_pipeline_stats_data(perf_ctx
, query
, data_size
, (uint8_t *)data
);
2262 unreachable("Unknown query type");
2267 *bytes_written
= written
;
2271 gen_perf_dump_query_count(struct gen_perf_context
*perf_ctx
)
2273 DBG("Queries: (Open queries = %d, OA users = %d)\n",
2274 perf_ctx
->n_active_oa_queries
, perf_ctx
->n_oa_users
);
2278 gen_perf_dump_query(struct gen_perf_context
*ctx
,
2279 struct gen_perf_query_object
*obj
,
2280 void *current_batch
)
2282 switch (obj
->queryinfo
->kind
) {
2283 case GEN_PERF_QUERY_TYPE_OA
:
2284 case GEN_PERF_QUERY_TYPE_RAW
:
2285 DBG("BO: %-4s OA data: %-10s %-15s\n",
2286 obj
->oa
.bo
? "yes," : "no,",
2287 gen_perf_is_query_ready(ctx
, obj
, current_batch
) ? "ready," : "not ready,",
2288 obj
->oa
.results_accumulated
? "accumulated" : "not accumulated");
2290 case GEN_PERF_QUERY_TYPE_PIPELINE
:
2292 obj
->pipeline_stats
.bo
? "yes" : "no");
2295 unreachable("Unknown query type");