1/*
2 * Copyright © 2018 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 */
23
24#ifndef INTEL_PERF_H
25#define INTEL_PERF_H
26
27#include <stdio.h>
28#include <stdbool.h>
29#include <stdint.h>
30#include <string.h>
31
32#if defined(MAJOR_IN_SYSMACROS)
33#include <sys/sysmacros.h>
34#elif defined(MAJOR_IN_MKDEV)
35#include <sys/mkdev.h>
36#endif
37
38#include "util/hash_table.h"
39#include "compiler/glsl/list.h"
40#include "util/ralloc.h"
41
42#include "drm-uapi/i915_drm.h"
43
44#ifdef __cplusplus
45extern "C" {
46#endif
47
48struct intel_device_info;
49
50struct intel_perf_config;
51struct intel_perf_query_info;
52
53#define INTEL_PERF_INVALID_CTX_ID (0xffffffff)
54
55enum intel_perf_counter_type {
56   INTEL_PERF_COUNTER_TYPE_EVENT,
57   INTEL_PERF_COUNTER_TYPE_DURATION_NORM,
58   INTEL_PERF_COUNTER_TYPE_DURATION_RAW,
59   INTEL_PERF_COUNTER_TYPE_THROUGHPUT,
60   INTEL_PERF_COUNTER_TYPE_RAW,
61   INTEL_PERF_COUNTER_TYPE_TIMESTAMP,
62};
63
64enum intel_perf_counter_data_type {
65   INTEL_PERF_COUNTER_DATA_TYPE_BOOL32,
66   INTEL_PERF_COUNTER_DATA_TYPE_UINT32,
67   INTEL_PERF_COUNTER_DATA_TYPE_UINT64,
68   INTEL_PERF_COUNTER_DATA_TYPE_FLOAT,
69   INTEL_PERF_COUNTER_DATA_TYPE_DOUBLE,
70};
71
72enum intel_perf_counter_units {
73   /* size */
74   INTEL_PERF_COUNTER_UNITS_BYTES,
75
76   /* frequency */
77   INTEL_PERF_COUNTER_UNITS_HZ,
78
79   /* time */
80   INTEL_PERF_COUNTER_UNITS_NS,
81   INTEL_PERF_COUNTER_UNITS_US,
82
83   /**/
84   INTEL_PERF_COUNTER_UNITS_PIXELS,
85   INTEL_PERF_COUNTER_UNITS_TEXELS,
86   INTEL_PERF_COUNTER_UNITS_THREADS,
87   INTEL_PERF_COUNTER_UNITS_PERCENT,
88
89   /* events */
90   INTEL_PERF_COUNTER_UNITS_MESSAGES,
91   INTEL_PERF_COUNTER_UNITS_NUMBER,
92   INTEL_PERF_COUNTER_UNITS_CYCLES,
93   INTEL_PERF_COUNTER_UNITS_EVENTS,
94   INTEL_PERF_COUNTER_UNITS_UTILIZATION,
95
96   /**/
97   INTEL_PERF_COUNTER_UNITS_EU_SENDS_TO_L3_CACHE_LINES,
98   INTEL_PERF_COUNTER_UNITS_EU_ATOMIC_REQUESTS_TO_L3_CACHE_LINES,
99   INTEL_PERF_COUNTER_UNITS_EU_REQUESTS_TO_L3_CACHE_LINES,
100   INTEL_PERF_COUNTER_UNITS_EU_BYTES_PER_L3_CACHE_LINE,
101
102   INTEL_PERF_COUNTER_UNITS_MAX
103};
104
105struct intel_pipeline_stat {
106   uint32_t reg;
107   uint32_t numerator;
108   uint32_t denominator;
109};
110
111/*
112 * The largest OA formats we can use include:
113 * For Haswell:
114 *   1 timestamp, 45 A counters, 8 B counters and 8 C counters.
115 * For Gfx8+
116 *   1 timestamp, 1 clock, 36 A counters, 8 B counters and 8 C counters
117 *
118 * Plus 2 PERF_CNT registers and 1 RPSTAT register.
119 */
120#define MAX_OA_REPORT_COUNTERS (62 + 2 + 1)
121
122/*
123 * When currently allocate only one page for pipeline statistics queries. Here
124 * we derived the maximum number of counters for that amount.
125 */
126#define STATS_BO_SIZE               4096
127#define STATS_BO_END_OFFSET_BYTES   (STATS_BO_SIZE / 2)
128#define MAX_STAT_COUNTERS           (STATS_BO_END_OFFSET_BYTES / 8)
129
130#define I915_PERF_OA_SAMPLE_SIZE (8 +   /* drm_i915_perf_record_header */ \
131                                  256)  /* OA counter report */
132
133struct intel_perf_query_result {
134   /**
135    * Storage for the final accumulated OA counters.
136    */
137   uint64_t accumulator[MAX_OA_REPORT_COUNTERS];
138
139   /**
140    * Hw ID used by the context on which the query was running.
141    */
142   uint32_t hw_id;
143
144   /**
145    * Number of reports accumulated to produce the results.
146    */
147   uint32_t reports_accumulated;
148
149   /**
150    * Frequency in the slices of the GT at the begin and end of the
151    * query.
152    */
153   uint64_t slice_frequency[2];
154
155   /**
156    * Frequency in the unslice of the GT at the begin and end of the
157    * query.
158    */
159   uint64_t unslice_frequency[2];
160
161   /**
162    * Frequency of the whole GT at the begin and end of the query.
163    */
164   uint64_t gt_frequency[2];
165
166   /**
167    * Timestamp of the query.
168    */
169   uint64_t begin_timestamp;
170
171   /**
172    * Whether the query was interrupted by another workload (aka preemption).
173    */
174   bool query_disjoint;
175};
176
177struct intel_perf_query_counter {
178   const char *name;
179   const char *desc;
180   const char *symbol_name;
181   const char *category;
182   enum intel_perf_counter_type type;
183   enum intel_perf_counter_data_type data_type;
184   enum intel_perf_counter_units units;
185   uint64_t raw_max;
186   size_t offset;
187
188   union {
189      uint64_t (*oa_counter_read_uint64)(struct intel_perf_config *perf,
190                                         const struct intel_perf_query_info *query,
191                                         const struct intel_perf_query_result *results);
192      float (*oa_counter_read_float)(struct intel_perf_config *perf,
193                                     const struct intel_perf_query_info *query,
194                                     const struct intel_perf_query_result *results);
195      struct intel_pipeline_stat pipeline_stat;
196   };
197};
198
199struct intel_perf_query_register_prog {
200   uint32_t reg;
201   uint32_t val;
202};
203
204/* Register programming for a given query */
205struct intel_perf_registers {
206   const struct intel_perf_query_register_prog *flex_regs;
207   uint32_t n_flex_regs;
208
209   const struct intel_perf_query_register_prog *mux_regs;
210   uint32_t n_mux_regs;
211
212   const struct intel_perf_query_register_prog *b_counter_regs;
213   uint32_t n_b_counter_regs;
214};
215
216struct intel_perf_query_info {
217   struct intel_perf_config *perf;
218
219   enum intel_perf_query_type {
220      INTEL_PERF_QUERY_TYPE_OA,
221      INTEL_PERF_QUERY_TYPE_RAW,
222      INTEL_PERF_QUERY_TYPE_PIPELINE,
223   } kind;
224   const char *name;
225   const char *symbol_name;
226   const char *guid;
227   struct intel_perf_query_counter *counters;
228   int n_counters;
229   int max_counters;
230   size_t data_size;
231
232   /* OA specific */
233   uint64_t oa_metrics_set_id;
234   int oa_format;
235
236   /* For indexing into the accumulator[] ... */
237   int gpu_time_offset;
238   int gpu_clock_offset;
239   int a_offset;
240   int b_offset;
241   int c_offset;
242   int perfcnt_offset;
243   int rpstat_offset;
244
245   struct intel_perf_registers config;
246};
247
248/* When not using the MI_RPC command, this structure describes the list of
249 * register offsets as well as their storage location so that they can be
250 * stored through a series of MI_SRM commands and accumulated with
251 * intel_perf_query_result_accumulate_snapshots().
252 */
253struct intel_perf_query_field_layout {
254   /* Alignment for the layout */
255   uint32_t alignment;
256
257   /* Size of the whole layout */
258   uint32_t size;
259
260   uint32_t n_fields;
261
262   struct intel_perf_query_field {
263      /* MMIO location of this register */
264      uint16_t mmio_offset;
265
266      /* Location of this register in the storage */
267      uint16_t location;
268
269      /* Type of register, for accumulation (see intel_perf_query_info:*_offset
270       * fields)
271       */
272      enum intel_perf_query_field_type {
273         INTEL_PERF_QUERY_FIELD_TYPE_MI_RPC,
274         INTEL_PERF_QUERY_FIELD_TYPE_SRM_PERFCNT,
275         INTEL_PERF_QUERY_FIELD_TYPE_SRM_RPSTAT,
276         INTEL_PERF_QUERY_FIELD_TYPE_SRM_OA_B,
277         INTEL_PERF_QUERY_FIELD_TYPE_SRM_OA_C,
278      } type;
279
280      /* Index of register in the given type (for instance A31 or B2,
281       * etc...)
282       */
283      uint8_t index;
284
285      /* 4, 8 or 256 */
286      uint16_t size;
287
288      /* If not 0, mask to apply to the register value. */
289      uint64_t mask;
290   } *fields;
291};
292
293struct intel_perf_query_counter_info {
294   struct intel_perf_query_counter *counter;
295
296   uint64_t query_mask;
297
298   /**
299    * Each counter can be a part of many groups, each time at different index.
300    * This struct stores one of those locations.
301    */
302   struct {
303      int group_idx; /* query/group number */
304      int counter_idx; /* index inside of query/group */
305   } location;
306};
307
308struct intel_perf_config {
309   /* Whether i915 has DRM_I915_QUERY_PERF_CONFIG support. */
310   bool i915_query_supported;
311
312   /* Version of the i915-perf subsystem, refer to i915_drm.h. */
313   int i915_perf_version;
314
315   /* Powergating configuration for the running the query. */
316   struct drm_i915_gem_context_param_sseu sseu;
317
318   struct intel_perf_query_info *queries;
319   int n_queries;
320
321   struct intel_perf_query_counter_info *counter_infos;
322   int n_counters;
323
324   struct intel_perf_query_field_layout query_layout;
325
326   /* Variables referenced in the XML meta data for OA performance
327    * counters, e.g in the normalization equations.
328    *
329    * All uint64_t for consistent operand types in generated code
330    */
331   struct {
332      uint64_t timestamp_frequency; /** $GpuTimestampFrequency */
333      uint64_t n_eus;               /** $EuCoresTotalCount */
334      uint64_t n_eu_slices;         /** $EuSlicesTotalCount */
335      uint64_t n_eu_sub_slices;     /** $EuSubslicesTotalCount */
336      uint64_t eu_threads_count;    /** $EuThreadsCount */
337      uint64_t slice_mask;          /** $SliceMask */
338      uint64_t subslice_mask;       /** $SubsliceMask */
339      uint64_t gt_min_freq;         /** $GpuMinFrequency */
340      uint64_t gt_max_freq;         /** $GpuMaxFrequency */
341      uint64_t revision;            /** $SkuRevisionId */
342      bool     query_mode;          /** $QueryMode */
343   } sys_vars;
344
345   /* OA metric sets, indexed by GUID, as know by Mesa at build time, to
346    * cross-reference with the GUIDs of configs advertised by the kernel at
347    * runtime
348    */
349   struct hash_table *oa_metrics_table;
350
351   /* When MDAPI hasn't configured the metric we need to use by the time the
352    * query begins, this OA metric is used as a fallback.
353    */
354   uint64_t fallback_raw_oa_metric;
355
356   /* Whether we have support for this platform. If true && n_queries == 0,
357    * this means we will not be able to use i915-perf because of it is in
358    * paranoid mode.
359    */
360   bool platform_supported;
361
362   /* Location of the device's sysfs entry. */
363   char sysfs_dev_dir[256];
364
365   struct {
366      void *(*bo_alloc)(void *bufmgr, const char *name, uint64_t size);
367      void (*bo_unreference)(void *bo);
368      void *(*bo_map)(void *ctx, void *bo, unsigned flags);
369      void (*bo_unmap)(void *bo);
370      bool (*batch_references)(void *batch, void *bo);
371      void (*bo_wait_rendering)(void *bo);
372      int (*bo_busy)(void *bo);
373      void (*emit_stall_at_pixel_scoreboard)(void *ctx);
374      void (*emit_mi_report_perf_count)(void *ctx,
375                                        void *bo,
376                                        uint32_t offset_in_bytes,
377                                        uint32_t report_id);
378      void (*batchbuffer_flush)(void *ctx,
379                                const char *file, int line);
380      void (*store_register_mem)(void *ctx, void *bo, uint32_t reg, uint32_t reg_size, uint32_t offset);
381
382   } vtbl;
383};
384
385struct intel_perf_counter_pass {
386   struct intel_perf_query_info *query;
387   struct intel_perf_query_counter *counter;
388   uint32_t pass;
389};
390
391/** Initialize the intel_perf_config object for a given device.
392 *
393 *    include_pipeline_statistics : Whether to add a pipeline statistic query
394 *                                  intel_perf_query_info object
395 *
396 *    use_register_snapshots : Whether the queries should include counters
397 *                             that rely on register snapshots using command
398 *                             streamer instructions (not possible when using
399 *                             only the OA buffer data).
400 */
401void intel_perf_init_metrics(struct intel_perf_config *perf_cfg,
402                             const struct intel_device_info *devinfo,
403                             int drm_fd,
404                             bool include_pipeline_statistics,
405                             bool use_register_snapshots);
406
407/** Query i915 for a metric id using guid.
408 */
409bool intel_perf_load_metric_id(struct intel_perf_config *perf_cfg,
410                               const char *guid,
411                               uint64_t *metric_id);
412
413/** Load a configuation's content from i915 using a guid.
414 */
415struct intel_perf_registers *intel_perf_load_configuration(struct intel_perf_config *perf_cfg,
416                                                           int fd, const char *guid);
417
418/** Store a configuration into i915 using guid and return a new metric id.
419 *
420 * If guid is NULL, then a generated one will be provided by hashing the
421 * content of the configuration.
422 */
423uint64_t intel_perf_store_configuration(struct intel_perf_config *perf_cfg, int fd,
424                                        const struct intel_perf_registers *config,
425                                        const char *guid);
426
427/** Read the slice/unslice frequency from 2 OA reports and store then into
428 *  result.
429 */
430void intel_perf_query_result_read_frequencies(struct intel_perf_query_result *result,
431                                              const struct intel_device_info *devinfo,
432                                              const uint32_t *start,
433                                              const uint32_t *end);
434
435/** Store the GT frequency as reported by the RPSTAT register.
436 */
437void intel_perf_query_result_read_gt_frequency(struct intel_perf_query_result *result,
438                                               const struct intel_device_info *devinfo,
439                                               const uint32_t start,
440                                               const uint32_t end);
441
442/** Store PERFCNT registers values.
443 */
444void intel_perf_query_result_read_perfcnts(struct intel_perf_query_result *result,
445                                           const struct intel_perf_query_info *query,
446                                           const uint64_t *start,
447                                           const uint64_t *end);
448
449/** Accumulate the delta between 2 OA reports into result for a given query.
450 */
451void intel_perf_query_result_accumulate(struct intel_perf_query_result *result,
452                                        const struct intel_perf_query_info *query,
453                                        const struct intel_device_info *devinfo,
454                                        const uint32_t *start,
455                                        const uint32_t *end);
456
457/** Accumulate the delta between 2 snapshots of OA perf registers (layout
458 * should match description specified through intel_perf_query_register_layout).
459 */
460void intel_perf_query_result_accumulate_fields(struct intel_perf_query_result *result,
461                                               const struct intel_perf_query_info *query,
462                                               const struct intel_device_info *devinfo,
463                                               const void *start,
464                                               const void *end,
465                                               bool no_oa_accumulate);
466
467void intel_perf_query_result_clear(struct intel_perf_query_result *result);
468
469/** Debug helper printing out query data.
470 */
471void intel_perf_query_result_print_fields(const struct intel_perf_query_info *query,
472                                          const struct intel_device_info *devinfo,
473                                          const void *data);
474
475static inline size_t
476intel_perf_query_counter_get_size(const struct intel_perf_query_counter *counter)
477{
478   switch (counter->data_type) {
479   case INTEL_PERF_COUNTER_DATA_TYPE_BOOL32:
480      return sizeof(uint32_t);
481   case INTEL_PERF_COUNTER_DATA_TYPE_UINT32:
482      return sizeof(uint32_t);
483   case INTEL_PERF_COUNTER_DATA_TYPE_UINT64:
484      return sizeof(uint64_t);
485   case INTEL_PERF_COUNTER_DATA_TYPE_FLOAT:
486      return sizeof(float);
487   case INTEL_PERF_COUNTER_DATA_TYPE_DOUBLE:
488      return sizeof(double);
489   default:
490      unreachable("invalid counter data type");
491   }
492}
493
494static inline struct intel_perf_config *
495intel_perf_new(void *ctx)
496{
497   struct intel_perf_config *perf = rzalloc(ctx, struct intel_perf_config);
498   return perf;
499}
500
501/** Whether we have the ability to hold off preemption on a batch so we don't
502 * have to look at the OA buffer to subtract unrelated workloads off the
503 * values captured through MI_* commands.
504 */
505static inline bool
506intel_perf_has_hold_preemption(const struct intel_perf_config *perf)
507{
508   return perf->i915_perf_version >= 3;
509}
510
511/** Whether we have the ability to lock EU array power configuration for the
512 * duration of the performance recording. This is useful on Gfx11 where the HW
513 * architecture requires half the EU for particular workloads.
514 */
515static inline bool
516intel_perf_has_global_sseu(const struct intel_perf_config *perf)
517{
518   return perf->i915_perf_version >= 4;
519}
520
521uint32_t intel_perf_get_n_passes(struct intel_perf_config *perf,
522                                 const uint32_t *counter_indices,
523                                 uint32_t counter_indices_count,
524                                 struct intel_perf_query_info **pass_queries);
525void intel_perf_get_counters_passes(struct intel_perf_config *perf,
526                                    const uint32_t *counter_indices,
527                                    uint32_t counter_indices_count,
528                                    struct intel_perf_counter_pass *counter_pass);
529
530#ifdef __cplusplus
531} // extern "C"
532#endif
533
534#endif /* INTEL_PERF_H */
535