intel/perf: report query split for mdapi
[mesa.git] / src / intel / perf / gen_perf.h
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 GEN_PERF_H
25 #define GEN_PERF_H
26
27 #include <stdio.h>
28 #include <stdint.h>
29 #include <string.h>
30
31 #if defined(MAJOR_IN_SYSMACROS)
32 #include <sys/sysmacros.h>
33 #elif defined(MAJOR_IN_MKDEV)
34 #include <sys/mkdev.h>
35 #endif
36
37 #include "util/hash_table.h"
38 #include "compiler/glsl/list.h"
39 #include "util/ralloc.h"
40
41 struct gen_device_info;
42
43 struct gen_perf_config;
44 struct gen_perf_query_info;
45
46 enum gen_perf_counter_type {
47 GEN_PERF_COUNTER_TYPE_EVENT,
48 GEN_PERF_COUNTER_TYPE_DURATION_NORM,
49 GEN_PERF_COUNTER_TYPE_DURATION_RAW,
50 GEN_PERF_COUNTER_TYPE_THROUGHPUT,
51 GEN_PERF_COUNTER_TYPE_RAW,
52 GEN_PERF_COUNTER_TYPE_TIMESTAMP,
53 };
54
55 enum gen_perf_counter_data_type {
56 GEN_PERF_COUNTER_DATA_TYPE_BOOL32,
57 GEN_PERF_COUNTER_DATA_TYPE_UINT32,
58 GEN_PERF_COUNTER_DATA_TYPE_UINT64,
59 GEN_PERF_COUNTER_DATA_TYPE_FLOAT,
60 GEN_PERF_COUNTER_DATA_TYPE_DOUBLE,
61 };
62
63 struct gen_pipeline_stat {
64 uint32_t reg;
65 uint32_t numerator;
66 uint32_t denominator;
67 };
68
69 /*
70 * The largest OA formats we can use include:
71 * For Haswell:
72 * 1 timestamp, 45 A counters, 8 B counters and 8 C counters.
73 * For Gen8+
74 * 1 timestamp, 1 clock, 36 A counters, 8 B counters and 8 C counters
75 */
76 #define MAX_OA_REPORT_COUNTERS 62
77
78 /*
79 * When currently allocate only one page for pipeline statistics queries. Here
80 * we derived the maximum number of counters for that amount.
81 */
82 #define STATS_BO_SIZE 4096
83 #define STATS_BO_END_OFFSET_BYTES (STATS_BO_SIZE / 2)
84 #define MAX_STAT_COUNTERS (STATS_BO_END_OFFSET_BYTES / 8)
85
86 #define I915_PERF_OA_SAMPLE_SIZE (8 + /* drm_i915_perf_record_header */ \
87 256) /* OA counter report */
88
89 struct gen_perf_query_result {
90 /**
91 * Storage for the final accumulated OA counters.
92 */
93 uint64_t accumulator[MAX_OA_REPORT_COUNTERS];
94
95 /**
96 * Hw ID used by the context on which the query was running.
97 */
98 uint32_t hw_id;
99
100 /**
101 * Number of reports accumulated to produce the results.
102 */
103 uint32_t reports_accumulated;
104
105 /**
106 * Frequency in the slices of the GT at the begin and end of the
107 * query.
108 */
109 uint64_t slice_frequency[2];
110
111 /**
112 * Frequency in the unslice of the GT at the begin and end of the
113 * query.
114 */
115 uint64_t unslice_frequency[2];
116
117 /**
118 * Timestamp of the query.
119 */
120 uint64_t begin_timestamp;
121
122 /**
123 * Whether the query was interrupted by another workload (aka preemption).
124 */
125 bool query_disjoint;
126 };
127
128 struct gen_perf_query_counter {
129 const char *name;
130 const char *desc;
131 enum gen_perf_counter_type type;
132 enum gen_perf_counter_data_type data_type;
133 uint64_t raw_max;
134 size_t offset;
135
136 union {
137 uint64_t (*oa_counter_read_uint64)(struct gen_perf_config *perf,
138 const struct gen_perf_query_info *query,
139 const uint64_t *accumulator);
140 float (*oa_counter_read_float)(struct gen_perf_config *perf,
141 const struct gen_perf_query_info *query,
142 const uint64_t *accumulator);
143 struct gen_pipeline_stat pipeline_stat;
144 };
145 };
146
147 struct gen_perf_query_register_prog {
148 uint32_t reg;
149 uint32_t val;
150 };
151
152 /* Register programming for a given query */
153 struct gen_perf_registers {
154 struct gen_perf_query_register_prog *flex_regs;
155 uint32_t n_flex_regs;
156
157 struct gen_perf_query_register_prog *mux_regs;
158 uint32_t n_mux_regs;
159
160 struct gen_perf_query_register_prog *b_counter_regs;
161 uint32_t n_b_counter_regs;
162 };
163
164 struct gen_perf_query_info {
165 enum gen_perf_query_type {
166 GEN_PERF_QUERY_TYPE_OA,
167 GEN_PERF_QUERY_TYPE_RAW,
168 GEN_PERF_QUERY_TYPE_PIPELINE,
169 } kind;
170 const char *name;
171 const char *guid;
172 struct gen_perf_query_counter *counters;
173 int n_counters;
174 int max_counters;
175 size_t data_size;
176
177 /* OA specific */
178 uint64_t oa_metrics_set_id;
179 int oa_format;
180
181 /* For indexing into the accumulator[] ... */
182 int gpu_time_offset;
183 int gpu_clock_offset;
184 int a_offset;
185 int b_offset;
186 int c_offset;
187
188 struct gen_perf_registers config;
189 };
190
191 struct gen_perf_config {
192 bool i915_query_supported;
193
194 struct gen_perf_query_info *queries;
195 int n_queries;
196
197 /* Variables referenced in the XML meta data for OA performance
198 * counters, e.g in the normalization equations.
199 *
200 * All uint64_t for consistent operand types in generated code
201 */
202 struct {
203 uint64_t timestamp_frequency; /** $GpuTimestampFrequency */
204 uint64_t n_eus; /** $EuCoresTotalCount */
205 uint64_t n_eu_slices; /** $EuSlicesTotalCount */
206 uint64_t n_eu_sub_slices; /** $EuSubslicesTotalCount */
207 uint64_t eu_threads_count; /** $EuThreadsCount */
208 uint64_t slice_mask; /** $SliceMask */
209 uint64_t subslice_mask; /** $SubsliceMask */
210 uint64_t gt_min_freq; /** $GpuMinFrequency */
211 uint64_t gt_max_freq; /** $GpuMaxFrequency */
212 uint64_t revision; /** $SkuRevisionId */
213 } sys_vars;
214
215 /* OA metric sets, indexed by GUID, as know by Mesa at build time, to
216 * cross-reference with the GUIDs of configs advertised by the kernel at
217 * runtime
218 */
219 struct hash_table *oa_metrics_table;
220
221 /* Location of the device's sysfs entry. */
222 char sysfs_dev_dir[256];
223
224 struct {
225 void *(*bo_alloc)(void *bufmgr, const char *name, uint64_t size);
226 void (*bo_unreference)(void *bo);
227 void *(*bo_map)(void *ctx, void *bo, unsigned flags);
228 void (*bo_unmap)(void *bo);
229 bool (*batch_references)(void *batch, void *bo);
230 void (*bo_wait_rendering)(void *bo);
231 int (*bo_busy)(void *bo);
232 void (*emit_stall_at_pixel_scoreboard)(void *ctx);
233 void (*emit_mi_report_perf_count)(void *ctx,
234 void *bo,
235 uint32_t offset_in_bytes,
236 uint32_t report_id);
237 void (*batchbuffer_flush)(void *ctx,
238 const char *file, int line);
239 void (*store_register_mem)(void *ctx, void *bo, uint32_t reg, uint32_t reg_size, uint32_t offset);
240
241 } vtbl;
242 };
243
244 struct gen_perf_query_object;
245 const struct gen_perf_query_info* gen_perf_query_info(const struct gen_perf_query_object *);
246
247 void gen_perf_init_metrics(struct gen_perf_config *perf_cfg,
248 const struct gen_device_info *devinfo,
249 int drm_fd);
250
251 /** Query i915 for a metric id using guid.
252 */
253 bool gen_perf_load_metric_id(struct gen_perf_config *perf_cfg,
254 const char *guid,
255 uint64_t *metric_id);
256
257 /** Load a configuation's content from i915 using a guid.
258 */
259 struct gen_perf_registers *gen_perf_load_configuration(struct gen_perf_config *perf_cfg,
260 int fd, const char *guid);
261
262 /** Store a configuration into i915 using guid and return a new metric id.
263 *
264 * If guid is NULL, then a generated one will be provided by hashing the
265 * content of the configuration.
266 */
267 uint64_t gen_perf_store_configuration(struct gen_perf_config *perf_cfg, int fd,
268 const struct gen_perf_registers *config,
269 const char *guid);
270
271 /** Read the slice/unslice frequency from 2 OA reports and store then into
272 * result.
273 */
274 void gen_perf_query_result_read_frequencies(struct gen_perf_query_result *result,
275 const struct gen_device_info *devinfo,
276 const uint32_t *start,
277 const uint32_t *end);
278 /** Accumulate the delta between 2 OA reports into result for a given query.
279 */
280 void gen_perf_query_result_accumulate(struct gen_perf_query_result *result,
281 const struct gen_perf_query_info *query,
282 const uint32_t *start,
283 const uint32_t *end);
284 void gen_perf_query_result_clear(struct gen_perf_query_result *result);
285
286 struct gen_perf_context;
287 struct gen_perf_context *gen_perf_new_context(void *parent);
288
289 void gen_perf_init_context(struct gen_perf_context *perf_ctx,
290 struct gen_perf_config *perf_cfg,
291 void * ctx, /* driver context (eg, brw_context) */
292 void * bufmgr, /* eg brw_bufmgr */
293 const struct gen_device_info *devinfo,
294 uint32_t hw_ctx,
295 int drm_fd);
296
297 struct gen_perf_config *gen_perf_config(struct gen_perf_context *ctx);
298
299 int gen_perf_active_queries(struct gen_perf_context *perf_ctx,
300 const struct gen_perf_query_info *query);
301
302 static inline size_t
303 gen_perf_query_counter_get_size(const struct gen_perf_query_counter *counter)
304 {
305 switch (counter->data_type) {
306 case GEN_PERF_COUNTER_DATA_TYPE_BOOL32:
307 return sizeof(uint32_t);
308 case GEN_PERF_COUNTER_DATA_TYPE_UINT32:
309 return sizeof(uint32_t);
310 case GEN_PERF_COUNTER_DATA_TYPE_UINT64:
311 return sizeof(uint64_t);
312 case GEN_PERF_COUNTER_DATA_TYPE_FLOAT:
313 return sizeof(float);
314 case GEN_PERF_COUNTER_DATA_TYPE_DOUBLE:
315 return sizeof(double);
316 default:
317 unreachable("invalid counter data type");
318 }
319 }
320
321 static inline struct gen_perf_config *
322 gen_perf_new(void *ctx)
323 {
324 struct gen_perf_config *perf = rzalloc(ctx, struct gen_perf_config);
325 return perf;
326 }
327
328 struct gen_perf_query_object *
329 gen_perf_new_query(struct gen_perf_context *, unsigned query_index);
330
331
332 bool gen_perf_begin_query(struct gen_perf_context *perf_ctx,
333 struct gen_perf_query_object *query);
334 void gen_perf_end_query(struct gen_perf_context *perf_ctx,
335 struct gen_perf_query_object *query);
336 void gen_perf_wait_query(struct gen_perf_context *perf_ctx,
337 struct gen_perf_query_object *query,
338 void *current_batch);
339 bool gen_perf_is_query_ready(struct gen_perf_context *perf_ctx,
340 struct gen_perf_query_object *query,
341 void *current_batch);
342 void gen_perf_delete_query(struct gen_perf_context *perf_ctx,
343 struct gen_perf_query_object *query);
344 void gen_perf_get_query_data(struct gen_perf_context *perf_ctx,
345 struct gen_perf_query_object *query,
346 int data_size,
347 unsigned *data,
348 unsigned *bytes_written);
349
350 void gen_perf_dump_query_count(struct gen_perf_context *perf_ctx);
351 void gen_perf_dump_query(struct gen_perf_context *perf_ctx,
352 struct gen_perf_query_object *obj,
353 void *current_batch);
354
355 #endif /* GEN_PERF_H */