1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 2017 Etnaviv Project
4 * Copyright (C) 2017 Zodiac Inflight Innovations
5 */
6
7 #include "common.xml.h"
8 #include "etnaviv_gpu.h"
9 #include "etnaviv_perfmon.h"
10 #include "state_hi.xml.h"
11
12 struct etnaviv_pm_domain;
13
14 struct etnaviv_pm_signal {
15 char name[64];
16 u32 data;
17
18 u32 (*sample)(struct etnaviv_gpu *gpu,
19 const struct etnaviv_pm_domain *domain,
20 const struct etnaviv_pm_signal *signal);
21 };
22
23 struct etnaviv_pm_domain {
24 char name[64];
25
26 /* profile register */
27 u32 profile_read;
28 u32 profile_config;
29
30 u8 nr_signals;
31 const struct etnaviv_pm_signal *signal;
32 };
33
34 struct etnaviv_pm_domain_meta {
35 unsigned int feature;
36 const struct etnaviv_pm_domain *domains;
37 u32 nr_domains;
38 };
39
perf_reg_read(struct etnaviv_gpu * gpu,const struct etnaviv_pm_domain * domain,const struct etnaviv_pm_signal * signal)40 static u32 perf_reg_read(struct etnaviv_gpu *gpu,
41 const struct etnaviv_pm_domain *domain,
42 const struct etnaviv_pm_signal *signal)
43 {
44 gpu_write(gpu, domain->profile_config, signal->data);
45
46 return gpu_read(gpu, domain->profile_read);
47 }
48
pipe_reg_read(struct etnaviv_gpu * gpu,const struct etnaviv_pm_domain * domain,const struct etnaviv_pm_signal * signal)49 static u32 pipe_reg_read(struct etnaviv_gpu *gpu,
50 const struct etnaviv_pm_domain *domain,
51 const struct etnaviv_pm_signal *signal)
52 {
53 u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
54 u32 value = 0;
55 unsigned i;
56
57 for (i = 0; i < gpu->identity.pixel_pipes; i++) {
58 clock &= ~(VIVS_HI_CLOCK_CONTROL_DEBUG_PIXEL_PIPE__MASK);
59 clock |= VIVS_HI_CLOCK_CONTROL_DEBUG_PIXEL_PIPE(i);
60 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
61 gpu_write(gpu, domain->profile_config, signal->data);
62 value += gpu_read(gpu, domain->profile_read);
63 }
64
65 /* switch back to pixel pipe 0 to prevent GPU hang */
66 clock &= ~(VIVS_HI_CLOCK_CONTROL_DEBUG_PIXEL_PIPE__MASK);
67 clock |= VIVS_HI_CLOCK_CONTROL_DEBUG_PIXEL_PIPE(0);
68 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
69
70 return value;
71 }
72
hi_total_cycle_read(struct etnaviv_gpu * gpu,const struct etnaviv_pm_domain * domain,const struct etnaviv_pm_signal * signal)73 static u32 hi_total_cycle_read(struct etnaviv_gpu *gpu,
74 const struct etnaviv_pm_domain *domain,
75 const struct etnaviv_pm_signal *signal)
76 {
77 u32 reg = VIVS_HI_PROFILE_TOTAL_CYCLES;
78
79 if (gpu->identity.model == chipModel_GC880 ||
80 gpu->identity.model == chipModel_GC2000 ||
81 gpu->identity.model == chipModel_GC2100)
82 reg = VIVS_MC_PROFILE_CYCLE_COUNTER;
83
84 return gpu_read(gpu, reg);
85 }
86
hi_total_idle_cycle_read(struct etnaviv_gpu * gpu,const struct etnaviv_pm_domain * domain,const struct etnaviv_pm_signal * signal)87 static u32 hi_total_idle_cycle_read(struct etnaviv_gpu *gpu,
88 const struct etnaviv_pm_domain *domain,
89 const struct etnaviv_pm_signal *signal)
90 {
91 u32 reg = VIVS_HI_PROFILE_IDLE_CYCLES;
92
93 if (gpu->identity.model == chipModel_GC880 ||
94 gpu->identity.model == chipModel_GC2000 ||
95 gpu->identity.model == chipModel_GC2100)
96 reg = VIVS_HI_PROFILE_TOTAL_CYCLES;
97
98 return gpu_read(gpu, reg);
99 }
100
101 static const struct etnaviv_pm_domain doms_3d[] = {
102 {
103 .name = "HI",
104 .profile_read = VIVS_MC_PROFILE_HI_READ,
105 .profile_config = VIVS_MC_PROFILE_CONFIG2,
106 .nr_signals = 5,
107 .signal = (const struct etnaviv_pm_signal[]) {
108 {
109 "TOTAL_CYCLES",
110 0,
111 &hi_total_cycle_read
112 },
113 {
114 "IDLE_CYCLES",
115 0,
116 &hi_total_idle_cycle_read
117 },
118 {
119 "AXI_CYCLES_READ_REQUEST_STALLED",
120 VIVS_MC_PROFILE_CONFIG2_HI_AXI_CYCLES_READ_REQUEST_STALLED,
121 &perf_reg_read
122 },
123 {
124 "AXI_CYCLES_WRITE_REQUEST_STALLED",
125 VIVS_MC_PROFILE_CONFIG2_HI_AXI_CYCLES_WRITE_REQUEST_STALLED,
126 &perf_reg_read
127 },
128 {
129 "AXI_CYCLES_WRITE_DATA_STALLED",
130 VIVS_MC_PROFILE_CONFIG2_HI_AXI_CYCLES_WRITE_DATA_STALLED,
131 &perf_reg_read
132 }
133 }
134 },
135 {
136 .name = "PE",
137 .profile_read = VIVS_MC_PROFILE_PE_READ,
138 .profile_config = VIVS_MC_PROFILE_CONFIG0,
139 .nr_signals = 4,
140 .signal = (const struct etnaviv_pm_signal[]) {
141 {
142 "PIXEL_COUNT_KILLED_BY_COLOR_PIPE",
143 VIVS_MC_PROFILE_CONFIG0_PE_PIXEL_COUNT_KILLED_BY_COLOR_PIPE,
144 &pipe_reg_read
145 },
146 {
147 "PIXEL_COUNT_KILLED_BY_DEPTH_PIPE",
148 VIVS_MC_PROFILE_CONFIG0_PE_PIXEL_COUNT_KILLED_BY_DEPTH_PIPE,
149 &pipe_reg_read
150 },
151 {
152 "PIXEL_COUNT_DRAWN_BY_COLOR_PIPE",
153 VIVS_MC_PROFILE_CONFIG0_PE_PIXEL_COUNT_DRAWN_BY_COLOR_PIPE,
154 &pipe_reg_read
155 },
156 {
157 "PIXEL_COUNT_DRAWN_BY_DEPTH_PIPE",
158 VIVS_MC_PROFILE_CONFIG0_PE_PIXEL_COUNT_DRAWN_BY_DEPTH_PIPE,
159 &pipe_reg_read
160 }
161 }
162 },
163 {
164 .name = "SH",
165 .profile_read = VIVS_MC_PROFILE_SH_READ,
166 .profile_config = VIVS_MC_PROFILE_CONFIG0,
167 .nr_signals = 9,
168 .signal = (const struct etnaviv_pm_signal[]) {
169 {
170 "SHADER_CYCLES",
171 VIVS_MC_PROFILE_CONFIG0_SH_SHADER_CYCLES,
172 &perf_reg_read
173 },
174 {
175 "PS_INST_COUNTER",
176 VIVS_MC_PROFILE_CONFIG0_SH_PS_INST_COUNTER,
177 &perf_reg_read
178 },
179 {
180 "RENDERED_PIXEL_COUNTER",
181 VIVS_MC_PROFILE_CONFIG0_SH_RENDERED_PIXEL_COUNTER,
182 &perf_reg_read
183 },
184 {
185 "VS_INST_COUNTER",
186 VIVS_MC_PROFILE_CONFIG0_SH_VS_INST_COUNTER,
187 &pipe_reg_read
188 },
189 {
190 "RENDERED_VERTICE_COUNTER",
191 VIVS_MC_PROFILE_CONFIG0_SH_RENDERED_VERTICE_COUNTER,
192 &pipe_reg_read
193 },
194 {
195 "VTX_BRANCH_INST_COUNTER",
196 VIVS_MC_PROFILE_CONFIG0_SH_VTX_BRANCH_INST_COUNTER,
197 &pipe_reg_read
198 },
199 {
200 "VTX_TEXLD_INST_COUNTER",
201 VIVS_MC_PROFILE_CONFIG0_SH_VTX_TEXLD_INST_COUNTER,
202 &pipe_reg_read
203 },
204 {
205 "PXL_BRANCH_INST_COUNTER",
206 VIVS_MC_PROFILE_CONFIG0_SH_PXL_BRANCH_INST_COUNTER,
207 &pipe_reg_read
208 },
209 {
210 "PXL_TEXLD_INST_COUNTER",
211 VIVS_MC_PROFILE_CONFIG0_SH_PXL_TEXLD_INST_COUNTER,
212 &pipe_reg_read
213 }
214 }
215 },
216 {
217 .name = "PA",
218 .profile_read = VIVS_MC_PROFILE_PA_READ,
219 .profile_config = VIVS_MC_PROFILE_CONFIG1,
220 .nr_signals = 6,
221 .signal = (const struct etnaviv_pm_signal[]) {
222 {
223 "INPUT_VTX_COUNTER",
224 VIVS_MC_PROFILE_CONFIG1_PA_INPUT_VTX_COUNTER,
225 &perf_reg_read
226 },
227 {
228 "INPUT_PRIM_COUNTER",
229 VIVS_MC_PROFILE_CONFIG1_PA_INPUT_PRIM_COUNTER,
230 &perf_reg_read
231 },
232 {
233 "OUTPUT_PRIM_COUNTER",
234 VIVS_MC_PROFILE_CONFIG1_PA_OUTPUT_PRIM_COUNTER,
235 &perf_reg_read
236 },
237 {
238 "DEPTH_CLIPPED_COUNTER",
239 VIVS_MC_PROFILE_CONFIG1_PA_DEPTH_CLIPPED_COUNTER,
240 &pipe_reg_read
241 },
242 {
243 "TRIVIAL_REJECTED_COUNTER",
244 VIVS_MC_PROFILE_CONFIG1_PA_TRIVIAL_REJECTED_COUNTER,
245 &pipe_reg_read
246 },
247 {
248 "CULLED_COUNTER",
249 VIVS_MC_PROFILE_CONFIG1_PA_CULLED_COUNTER,
250 &pipe_reg_read
251 }
252 }
253 },
254 {
255 .name = "SE",
256 .profile_read = VIVS_MC_PROFILE_SE_READ,
257 .profile_config = VIVS_MC_PROFILE_CONFIG1,
258 .nr_signals = 2,
259 .signal = (const struct etnaviv_pm_signal[]) {
260 {
261 "CULLED_TRIANGLE_COUNT",
262 VIVS_MC_PROFILE_CONFIG1_SE_CULLED_TRIANGLE_COUNT,
263 &perf_reg_read
264 },
265 {
266 "CULLED_LINES_COUNT",
267 VIVS_MC_PROFILE_CONFIG1_SE_CULLED_LINES_COUNT,
268 &perf_reg_read
269 }
270 }
271 },
272 {
273 .name = "RA",
274 .profile_read = VIVS_MC_PROFILE_RA_READ,
275 .profile_config = VIVS_MC_PROFILE_CONFIG1,
276 .nr_signals = 7,
277 .signal = (const struct etnaviv_pm_signal[]) {
278 {
279 "VALID_PIXEL_COUNT",
280 VIVS_MC_PROFILE_CONFIG1_RA_VALID_PIXEL_COUNT,
281 &perf_reg_read
282 },
283 {
284 "TOTAL_QUAD_COUNT",
285 VIVS_MC_PROFILE_CONFIG1_RA_TOTAL_QUAD_COUNT,
286 &perf_reg_read
287 },
288 {
289 "VALID_QUAD_COUNT_AFTER_EARLY_Z",
290 VIVS_MC_PROFILE_CONFIG1_RA_VALID_QUAD_COUNT_AFTER_EARLY_Z,
291 &perf_reg_read
292 },
293 {
294 "TOTAL_PRIMITIVE_COUNT",
295 VIVS_MC_PROFILE_CONFIG1_RA_TOTAL_PRIMITIVE_COUNT,
296 &perf_reg_read
297 },
298 {
299 "PIPE_CACHE_MISS_COUNTER",
300 VIVS_MC_PROFILE_CONFIG1_RA_PIPE_CACHE_MISS_COUNTER,
301 &perf_reg_read
302 },
303 {
304 "PREFETCH_CACHE_MISS_COUNTER",
305 VIVS_MC_PROFILE_CONFIG1_RA_PREFETCH_CACHE_MISS_COUNTER,
306 &perf_reg_read
307 },
308 {
309 "CULLED_QUAD_COUNT",
310 VIVS_MC_PROFILE_CONFIG1_RA_CULLED_QUAD_COUNT,
311 &perf_reg_read
312 }
313 }
314 },
315 {
316 .name = "TX",
317 .profile_read = VIVS_MC_PROFILE_TX_READ,
318 .profile_config = VIVS_MC_PROFILE_CONFIG1,
319 .nr_signals = 9,
320 .signal = (const struct etnaviv_pm_signal[]) {
321 {
322 "TOTAL_BILINEAR_REQUESTS",
323 VIVS_MC_PROFILE_CONFIG1_TX_TOTAL_BILINEAR_REQUESTS,
324 &perf_reg_read
325 },
326 {
327 "TOTAL_TRILINEAR_REQUESTS",
328 VIVS_MC_PROFILE_CONFIG1_TX_TOTAL_TRILINEAR_REQUESTS,
329 &perf_reg_read
330 },
331 {
332 "TOTAL_DISCARDED_TEXTURE_REQUESTS",
333 VIVS_MC_PROFILE_CONFIG1_TX_TOTAL_DISCARDED_TEXTURE_REQUESTS,
334 &perf_reg_read
335 },
336 {
337 "TOTAL_TEXTURE_REQUESTS",
338 VIVS_MC_PROFILE_CONFIG1_TX_TOTAL_TEXTURE_REQUESTS,
339 &perf_reg_read
340 },
341 {
342 "MEM_READ_COUNT",
343 VIVS_MC_PROFILE_CONFIG1_TX_MEM_READ_COUNT,
344 &perf_reg_read
345 },
346 {
347 "MEM_READ_IN_8B_COUNT",
348 VIVS_MC_PROFILE_CONFIG1_TX_MEM_READ_IN_8B_COUNT,
349 &perf_reg_read
350 },
351 {
352 "CACHE_MISS_COUNT",
353 VIVS_MC_PROFILE_CONFIG1_TX_CACHE_MISS_COUNT,
354 &perf_reg_read
355 },
356 {
357 "CACHE_HIT_TEXEL_COUNT",
358 VIVS_MC_PROFILE_CONFIG1_TX_CACHE_HIT_TEXEL_COUNT,
359 &perf_reg_read
360 },
361 {
362 "CACHE_MISS_TEXEL_COUNT",
363 VIVS_MC_PROFILE_CONFIG1_TX_CACHE_MISS_TEXEL_COUNT,
364 &perf_reg_read
365 }
366 }
367 },
368 {
369 .name = "MC",
370 .profile_read = VIVS_MC_PROFILE_MC_READ,
371 .profile_config = VIVS_MC_PROFILE_CONFIG2,
372 .nr_signals = 3,
373 .signal = (const struct etnaviv_pm_signal[]) {
374 {
375 "TOTAL_READ_REQ_8B_FROM_PIPELINE",
376 VIVS_MC_PROFILE_CONFIG2_MC_TOTAL_READ_REQ_8B_FROM_PIPELINE,
377 &perf_reg_read
378 },
379 {
380 "TOTAL_READ_REQ_8B_FROM_IP",
381 VIVS_MC_PROFILE_CONFIG2_MC_TOTAL_READ_REQ_8B_FROM_IP,
382 &perf_reg_read
383 },
384 {
385 "TOTAL_WRITE_REQ_8B_FROM_PIPELINE",
386 VIVS_MC_PROFILE_CONFIG2_MC_TOTAL_WRITE_REQ_8B_FROM_PIPELINE,
387 &perf_reg_read
388 }
389 }
390 }
391 };
392
393 static const struct etnaviv_pm_domain doms_2d[] = {
394 {
395 .name = "PE",
396 .profile_read = VIVS_MC_PROFILE_PE_READ,
397 .profile_config = VIVS_MC_PROFILE_CONFIG0,
398 .nr_signals = 1,
399 .signal = (const struct etnaviv_pm_signal[]) {
400 {
401 "PIXELS_RENDERED_2D",
402 VIVS_MC_PROFILE_CONFIG0_PE_PIXELS_RENDERED_2D,
403 &pipe_reg_read
404 }
405 }
406 }
407 };
408
409 static const struct etnaviv_pm_domain doms_vg[] = {
410 };
411
412 static const struct etnaviv_pm_domain_meta doms_meta[] = {
413 {
414 .feature = chipFeatures_PIPE_3D,
415 .nr_domains = ARRAY_SIZE(doms_3d),
416 .domains = &doms_3d[0]
417 },
418 {
419 .feature = chipFeatures_PIPE_2D,
420 .nr_domains = ARRAY_SIZE(doms_2d),
421 .domains = &doms_2d[0]
422 },
423 {
424 .feature = chipFeatures_PIPE_VG,
425 .nr_domains = ARRAY_SIZE(doms_vg),
426 .domains = &doms_vg[0]
427 }
428 };
429
num_pm_domains(const struct etnaviv_gpu * gpu)430 static unsigned int num_pm_domains(const struct etnaviv_gpu *gpu)
431 {
432 unsigned int num = 0, i;
433
434 for (i = 0; i < ARRAY_SIZE(doms_meta); i++) {
435 const struct etnaviv_pm_domain_meta *meta = &doms_meta[i];
436
437 if (gpu->identity.features & meta->feature)
438 num += meta->nr_domains;
439 }
440
441 return num;
442 }
443
pm_domain(const struct etnaviv_gpu * gpu,unsigned int index)444 static const struct etnaviv_pm_domain *pm_domain(const struct etnaviv_gpu *gpu,
445 unsigned int index)
446 {
447 const struct etnaviv_pm_domain *domain = NULL;
448 unsigned int offset = 0, i;
449
450 for (i = 0; i < ARRAY_SIZE(doms_meta); i++) {
451 const struct etnaviv_pm_domain_meta *meta = &doms_meta[i];
452
453 if (!(gpu->identity.features & meta->feature))
454 continue;
455
456 if (index - offset >= meta->nr_domains) {
457 offset += meta->nr_domains;
458 continue;
459 }
460
461 domain = meta->domains + (index - offset);
462 }
463
464 return domain;
465 }
466
etnaviv_pm_query_dom(struct etnaviv_gpu * gpu,struct drm_etnaviv_pm_domain * domain)467 int etnaviv_pm_query_dom(struct etnaviv_gpu *gpu,
468 struct drm_etnaviv_pm_domain *domain)
469 {
470 const unsigned int nr_domains = num_pm_domains(gpu);
471 const struct etnaviv_pm_domain *dom;
472
473 if (domain->iter >= nr_domains)
474 return -EINVAL;
475
476 dom = pm_domain(gpu, domain->iter);
477 if (!dom)
478 return -EINVAL;
479
480 domain->id = domain->iter;
481 domain->nr_signals = dom->nr_signals;
482 strncpy(domain->name, dom->name, sizeof(domain->name));
483
484 domain->iter++;
485 if (domain->iter == nr_domains)
486 domain->iter = 0xff;
487
488 return 0;
489 }
490
etnaviv_pm_query_sig(struct etnaviv_gpu * gpu,struct drm_etnaviv_pm_signal * signal)491 int etnaviv_pm_query_sig(struct etnaviv_gpu *gpu,
492 struct drm_etnaviv_pm_signal *signal)
493 {
494 const unsigned int nr_domains = num_pm_domains(gpu);
495 const struct etnaviv_pm_domain *dom;
496 const struct etnaviv_pm_signal *sig;
497
498 if (signal->domain >= nr_domains)
499 return -EINVAL;
500
501 dom = pm_domain(gpu, signal->domain);
502 if (!dom)
503 return -EINVAL;
504
505 if (signal->iter >= dom->nr_signals)
506 return -EINVAL;
507
508 sig = &dom->signal[signal->iter];
509
510 signal->id = signal->iter;
511 strncpy(signal->name, sig->name, sizeof(signal->name));
512
513 signal->iter++;
514 if (signal->iter == dom->nr_signals)
515 signal->iter = 0xffff;
516
517 return 0;
518 }
519
etnaviv_pm_req_validate(const struct drm_etnaviv_gem_submit_pmr * r,u32 exec_state)520 int etnaviv_pm_req_validate(const struct drm_etnaviv_gem_submit_pmr *r,
521 u32 exec_state)
522 {
523 const struct etnaviv_pm_domain_meta *meta = &doms_meta[exec_state];
524 const struct etnaviv_pm_domain *dom;
525
526 if (r->domain >= meta->nr_domains)
527 return -EINVAL;
528
529 dom = meta->domains + r->domain;
530
531 if (r->signal >= dom->nr_signals)
532 return -EINVAL;
533
534 return 0;
535 }
536
etnaviv_perfmon_process(struct etnaviv_gpu * gpu,const struct etnaviv_perfmon_request * pmr,u32 exec_state)537 void etnaviv_perfmon_process(struct etnaviv_gpu *gpu,
538 const struct etnaviv_perfmon_request *pmr, u32 exec_state)
539 {
540 const struct etnaviv_pm_domain_meta *meta = &doms_meta[exec_state];
541 const struct etnaviv_pm_domain *dom;
542 const struct etnaviv_pm_signal *sig;
543 u32 *bo = pmr->bo_vma;
544 u32 val;
545
546 dom = meta->domains + pmr->domain;
547 sig = &dom->signal[pmr->signal];
548 val = sig->sample(gpu, dom, sig);
549
550 *(bo + pmr->offset) = val;
551 }
552
