projects / mesa.git / blob
? search:


1dd9249d57cfc05d482785f19fb970433d601e87
[mesa.git] / src / gallium / drivers / radeonsi / si_query.c
1 /*
2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3 * Copyright 2014 Marek Olšák <marek.olsak@amd.com>
4 * Copyright 2018 Advanced Micro Devices, Inc.
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * on the rights to use, copy, modify, merge, publish, distribute, sub
11 * license, and/or sell copies of the Software, and to permit persons to whom
12 * the Software is furnished to do so, subject to the following conditions:
13 *
14 * The above copyright notice and this permission notice (including the next
15 * paragraph) shall be included in all copies or substantial portions of the
16 * Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
22 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 */
26
27 #include "si_pipe.h"
28 #include "si_query.h"
29 #include "util/u_memory.h"
30 #include "util/u_upload_mgr.h"
31 #include "util/os_time.h"
32 #include "util/u_suballoc.h"
33 #include "amd/common/sid.h"
34
35 #define SI_MAX_STREAMS 4
36
37 static const struct si_query_ops query_hw_ops;
38
39 struct si_hw_query_params {
40 unsigned start_offset;
41 unsigned end_offset;
42 unsigned fence_offset;
43 unsigned pair_stride;
44 unsigned pair_count;
45 };
46
47 /* Queries without buffer handling or suspend/resume. */
48 struct si_query_sw {
49 struct si_query b;
50
51 uint64_t begin_result;
52 uint64_t end_result;
53
54 uint64_t begin_time;
55 uint64_t end_time;
56
57 /* Fence for GPU_FINISHED. */
58 struct pipe_fence_handle *fence;
59 };
60
61 static void si_query_sw_destroy(struct si_screen *sscreen,
62 struct si_query *squery)
63 {
64 struct si_query_sw *query = (struct si_query_sw *)squery;
65
66 sscreen->b.fence_reference(&sscreen->b, &query->fence, NULL);
67 FREE(query);
68 }
69
70 static enum radeon_value_id winsys_id_from_type(unsigned type)
71 {
72 switch (type) {
73 case SI_QUERY_REQUESTED_VRAM: return RADEON_REQUESTED_VRAM_MEMORY;
74 case SI_QUERY_REQUESTED_GTT: return RADEON_REQUESTED_GTT_MEMORY;
75 case SI_QUERY_MAPPED_VRAM: return RADEON_MAPPED_VRAM;
76 case SI_QUERY_MAPPED_GTT: return RADEON_MAPPED_GTT;
77 case SI_QUERY_BUFFER_WAIT_TIME: return RADEON_BUFFER_WAIT_TIME_NS;
78 case SI_QUERY_NUM_MAPPED_BUFFERS: return RADEON_NUM_MAPPED_BUFFERS;
79 case SI_QUERY_NUM_GFX_IBS: return RADEON_NUM_GFX_IBS;
80 case SI_QUERY_NUM_SDMA_IBS: return RADEON_NUM_SDMA_IBS;
81 case SI_QUERY_GFX_BO_LIST_SIZE: return RADEON_GFX_BO_LIST_COUNTER;
82 case SI_QUERY_GFX_IB_SIZE: return RADEON_GFX_IB_SIZE_COUNTER;
83 case SI_QUERY_NUM_BYTES_MOVED: return RADEON_NUM_BYTES_MOVED;
84 case SI_QUERY_NUM_EVICTIONS: return RADEON_NUM_EVICTIONS;
85 case SI_QUERY_NUM_VRAM_CPU_PAGE_FAULTS: return RADEON_NUM_VRAM_CPU_PAGE_FAULTS;
86 case SI_QUERY_VRAM_USAGE: return RADEON_VRAM_USAGE;
87 case SI_QUERY_VRAM_VIS_USAGE: return RADEON_VRAM_VIS_USAGE;
88 case SI_QUERY_GTT_USAGE: return RADEON_GTT_USAGE;
89 case SI_QUERY_GPU_TEMPERATURE: return RADEON_GPU_TEMPERATURE;
90 case SI_QUERY_CURRENT_GPU_SCLK: return RADEON_CURRENT_SCLK;
91 case SI_QUERY_CURRENT_GPU_MCLK: return RADEON_CURRENT_MCLK;
92 case SI_QUERY_CS_THREAD_BUSY: return RADEON_CS_THREAD_TIME;
93 default: unreachable("query type does not correspond to winsys id");
94 }
95 }
96
97 static int64_t si_finish_dma_get_cpu_time(struct si_context *sctx)
98 {
99 struct pipe_fence_handle *fence = NULL;
100
101 si_flush_dma_cs(sctx, 0, &fence);
102 if (fence) {
103 sctx->ws->fence_wait(sctx->ws, fence, PIPE_TIMEOUT_INFINITE);
104 sctx->ws->fence_reference(&fence, NULL);
105 }
106
107 return os_time_get_nano();
108 }
109
110 static bool si_query_sw_begin(struct si_context *sctx,
111 struct si_query *squery)
112 {
113 struct si_query_sw *query = (struct si_query_sw *)squery;
114 enum radeon_value_id ws_id;
115
116 switch(query->b.type) {
117 case PIPE_QUERY_TIMESTAMP_DISJOINT:
118 case PIPE_QUERY_GPU_FINISHED:
119 break;
120 case SI_QUERY_TIME_ELAPSED_SDMA_SI:
121 query->begin_result = si_finish_dma_get_cpu_time(sctx);
122 break;
123 case SI_QUERY_DRAW_CALLS:
124 query->begin_result = sctx->num_draw_calls;
125 break;
126 case SI_QUERY_DECOMPRESS_CALLS:
127 query->begin_result = sctx->num_decompress_calls;
128 break;
129 case SI_QUERY_MRT_DRAW_CALLS:
130 query->begin_result = sctx->num_mrt_draw_calls;
131 break;
132 case SI_QUERY_PRIM_RESTART_CALLS:
133 query->begin_result = sctx->num_prim_restart_calls;
134 break;
135 case SI_QUERY_SPILL_DRAW_CALLS:
136 query->begin_result = sctx->num_spill_draw_calls;
137 break;
138 case SI_QUERY_COMPUTE_CALLS:
139 query->begin_result = sctx->num_compute_calls;
140 break;
141 case SI_QUERY_SPILL_COMPUTE_CALLS:
142 query->begin_result = sctx->num_spill_compute_calls;
143 break;
144 case SI_QUERY_DMA_CALLS:
145 query->begin_result = sctx->num_dma_calls;
146 break;
147 case SI_QUERY_CP_DMA_CALLS:
148 query->begin_result = sctx->num_cp_dma_calls;
149 break;
150 case SI_QUERY_NUM_VS_FLUSHES:
151 query->begin_result = sctx->num_vs_flushes;
152 break;
153 case SI_QUERY_NUM_PS_FLUSHES:
154 query->begin_result = sctx->num_ps_flushes;
155 break;
156 case SI_QUERY_NUM_CS_FLUSHES:
157 query->begin_result = sctx->num_cs_flushes;
158 break;
159 case SI_QUERY_NUM_CB_CACHE_FLUSHES:
160 query->begin_result = sctx->num_cb_cache_flushes;
161 break;
162 case SI_QUERY_NUM_DB_CACHE_FLUSHES:
163 query->begin_result = sctx->num_db_cache_flushes;
164 break;
165 case SI_QUERY_NUM_L2_INVALIDATES:
166 query->begin_result = sctx->num_L2_invalidates;
167 break;
168 case SI_QUERY_NUM_L2_WRITEBACKS:
169 query->begin_result = sctx->num_L2_writebacks;
170 break;
171 case SI_QUERY_NUM_RESIDENT_HANDLES:
172 query->begin_result = sctx->num_resident_handles;
173 break;
174 case SI_QUERY_TC_OFFLOADED_SLOTS:
175 query->begin_result = sctx->tc ? sctx->tc->num_offloaded_slots : 0;
176 break;
177 case SI_QUERY_TC_DIRECT_SLOTS:
178 query->begin_result = sctx->tc ? sctx->tc->num_direct_slots : 0;
179 break;
180 case SI_QUERY_TC_NUM_SYNCS:
181 query->begin_result = sctx->tc ? sctx->tc->num_syncs : 0;
182 break;
183 case SI_QUERY_REQUESTED_VRAM:
184 case SI_QUERY_REQUESTED_GTT:
185 case SI_QUERY_MAPPED_VRAM:
186 case SI_QUERY_MAPPED_GTT:
187 case SI_QUERY_VRAM_USAGE:
188 case SI_QUERY_VRAM_VIS_USAGE:
189 case SI_QUERY_GTT_USAGE:
190 case SI_QUERY_GPU_TEMPERATURE:
191 case SI_QUERY_CURRENT_GPU_SCLK:
192 case SI_QUERY_CURRENT_GPU_MCLK:
193 case SI_QUERY_BACK_BUFFER_PS_DRAW_RATIO:
194 case SI_QUERY_NUM_MAPPED_BUFFERS:
195 query->begin_result = 0;
196 break;
197 case SI_QUERY_BUFFER_WAIT_TIME:
198 case SI_QUERY_GFX_IB_SIZE:
199 case SI_QUERY_NUM_GFX_IBS:
200 case SI_QUERY_NUM_SDMA_IBS:
201 case SI_QUERY_NUM_BYTES_MOVED:
202 case SI_QUERY_NUM_EVICTIONS:
203 case SI_QUERY_NUM_VRAM_CPU_PAGE_FAULTS: {
204 enum radeon_value_id ws_id = winsys_id_from_type(query->b.type);
205 query->begin_result = sctx->ws->query_value(sctx->ws, ws_id);
206 break;
207 }
208 case SI_QUERY_GFX_BO_LIST_SIZE:
209 ws_id = winsys_id_from_type(query->b.type);
210 query->begin_result = sctx->ws->query_value(sctx->ws, ws_id);
211 query->begin_time = sctx->ws->query_value(sctx->ws,
212 RADEON_NUM_GFX_IBS);
213 break;
214 case SI_QUERY_CS_THREAD_BUSY:
215 ws_id = winsys_id_from_type(query->b.type);
216 query->begin_result = sctx->ws->query_value(sctx->ws, ws_id);
217 query->begin_time = os_time_get_nano();
218 break;
219 case SI_QUERY_GALLIUM_THREAD_BUSY:
220 query->begin_result =
221 sctx->tc ? util_queue_get_thread_time_nano(&sctx->tc->queue, 0) : 0;
222 query->begin_time = os_time_get_nano();
223 break;
224 case SI_QUERY_GPU_LOAD:
225 case SI_QUERY_GPU_SHADERS_BUSY:
226 case SI_QUERY_GPU_TA_BUSY:
227 case SI_QUERY_GPU_GDS_BUSY:
228 case SI_QUERY_GPU_VGT_BUSY:
229 case SI_QUERY_GPU_IA_BUSY:
230 case SI_QUERY_GPU_SX_BUSY:
231 case SI_QUERY_GPU_WD_BUSY:
232 case SI_QUERY_GPU_BCI_BUSY:
233 case SI_QUERY_GPU_SC_BUSY:
234 case SI_QUERY_GPU_PA_BUSY:
235 case SI_QUERY_GPU_DB_BUSY:
236 case SI_QUERY_GPU_CP_BUSY:
237 case SI_QUERY_GPU_CB_BUSY:
238 case SI_QUERY_GPU_SDMA_BUSY:
239 case SI_QUERY_GPU_PFP_BUSY:
240 case SI_QUERY_GPU_MEQ_BUSY:
241 case SI_QUERY_GPU_ME_BUSY:
242 case SI_QUERY_GPU_SURF_SYNC_BUSY:
243 case SI_QUERY_GPU_CP_DMA_BUSY:
244 case SI_QUERY_GPU_SCRATCH_RAM_BUSY:
245 query->begin_result = si_begin_counter(sctx->screen,
246 query->b.type);
247 break;
248 case SI_QUERY_NUM_COMPILATIONS:
249 query->begin_result = p_atomic_read(&sctx->screen->num_compilations);
250 break;
251 case SI_QUERY_NUM_SHADERS_CREATED:
252 query->begin_result = p_atomic_read(&sctx->screen->num_shaders_created);
253 break;
254 case SI_QUERY_NUM_SHADER_CACHE_HITS:
255 query->begin_result =
256 p_atomic_read(&sctx->screen->num_shader_cache_hits);
257 break;
258 case SI_QUERY_GPIN_ASIC_ID:
259 case SI_QUERY_GPIN_NUM_SIMD:
260 case SI_QUERY_GPIN_NUM_RB:
261 case SI_QUERY_GPIN_NUM_SPI:
262 case SI_QUERY_GPIN_NUM_SE:
263 break;
264 default:
265 unreachable("si_query_sw_begin: bad query type");
266 }
267
268 return true;
269 }
270
271 static bool si_query_sw_end(struct si_context *sctx,
272 struct si_query *squery)
273 {
274 struct si_query_sw *query = (struct si_query_sw *)squery;
275 enum radeon_value_id ws_id;
276
277 switch(query->b.type) {
278 case PIPE_QUERY_TIMESTAMP_DISJOINT:
279 break;
280 case PIPE_QUERY_GPU_FINISHED:
281 sctx->b.flush(&sctx->b, &query->fence, PIPE_FLUSH_DEFERRED);
282 break;
283 case SI_QUERY_TIME_ELAPSED_SDMA_SI:
284 query->end_result = si_finish_dma_get_cpu_time(sctx);
285 break;
286 case SI_QUERY_DRAW_CALLS:
287 query->end_result = sctx->num_draw_calls;
288 break;
289 case SI_QUERY_DECOMPRESS_CALLS:
290 query->end_result = sctx->num_decompress_calls;
291 break;
292 case SI_QUERY_MRT_DRAW_CALLS:
293 query->end_result = sctx->num_mrt_draw_calls;
294 break;
295 case SI_QUERY_PRIM_RESTART_CALLS:
296 query->end_result = sctx->num_prim_restart_calls;
297 break;
298 case SI_QUERY_SPILL_DRAW_CALLS:
299 query->end_result = sctx->num_spill_draw_calls;
300 break;
301 case SI_QUERY_COMPUTE_CALLS:
302 query->end_result = sctx->num_compute_calls;
303 break;
304 case SI_QUERY_SPILL_COMPUTE_CALLS:
305 query->end_result = sctx->num_spill_compute_calls;
306 break;
307 case SI_QUERY_DMA_CALLS:
308 query->end_result = sctx->num_dma_calls;
309 break;
310 case SI_QUERY_CP_DMA_CALLS:
311 query->end_result = sctx->num_cp_dma_calls;
312 break;
313 case SI_QUERY_NUM_VS_FLUSHES:
314 query->end_result = sctx->num_vs_flushes;
315 break;
316 case SI_QUERY_NUM_PS_FLUSHES:
317 query->end_result = sctx->num_ps_flushes;
318 break;
319 case SI_QUERY_NUM_CS_FLUSHES:
320 query->end_result = sctx->num_cs_flushes;
321 break;
322 case SI_QUERY_NUM_CB_CACHE_FLUSHES:
323 query->end_result = sctx->num_cb_cache_flushes;
324 break;
325 case SI_QUERY_NUM_DB_CACHE_FLUSHES:
326 query->end_result = sctx->num_db_cache_flushes;
327 break;
328 case SI_QUERY_NUM_L2_INVALIDATES:
329 query->end_result = sctx->num_L2_invalidates;
330 break;
331 case SI_QUERY_NUM_L2_WRITEBACKS:
332 query->end_result = sctx->num_L2_writebacks;
333 break;
334 case SI_QUERY_NUM_RESIDENT_HANDLES:
335 query->end_result = sctx->num_resident_handles;
336 break;
337 case SI_QUERY_TC_OFFLOADED_SLOTS:
338 query->end_result = sctx->tc ? sctx->tc->num_offloaded_slots : 0;
339 break;
340 case SI_QUERY_TC_DIRECT_SLOTS:
341 query->end_result = sctx->tc ? sctx->tc->num_direct_slots : 0;
342 break;
343 case SI_QUERY_TC_NUM_SYNCS:
344 query->end_result = sctx->tc ? sctx->tc->num_syncs : 0;
345 break;
346 case SI_QUERY_REQUESTED_VRAM:
347 case SI_QUERY_REQUESTED_GTT:
348 case SI_QUERY_MAPPED_VRAM:
349 case SI_QUERY_MAPPED_GTT:
350 case SI_QUERY_VRAM_USAGE:
351 case SI_QUERY_VRAM_VIS_USAGE:
352 case SI_QUERY_GTT_USAGE:
353 case SI_QUERY_GPU_TEMPERATURE:
354 case SI_QUERY_CURRENT_GPU_SCLK:
355 case SI_QUERY_CURRENT_GPU_MCLK:
356 case SI_QUERY_BUFFER_WAIT_TIME:
357 case SI_QUERY_GFX_IB_SIZE:
358 case SI_QUERY_NUM_MAPPED_BUFFERS:
359 case SI_QUERY_NUM_GFX_IBS:
360 case SI_QUERY_NUM_SDMA_IBS:
361 case SI_QUERY_NUM_BYTES_MOVED:
362 case SI_QUERY_NUM_EVICTIONS:
363 case SI_QUERY_NUM_VRAM_CPU_PAGE_FAULTS: {
364 enum radeon_value_id ws_id = winsys_id_from_type(query->b.type);
365 query->end_result = sctx->ws->query_value(sctx->ws, ws_id);
366 break;
367 }
368 case SI_QUERY_GFX_BO_LIST_SIZE:
369 ws_id = winsys_id_from_type(query->b.type);
370 query->end_result = sctx->ws->query_value(sctx->ws, ws_id);
371 query->end_time = sctx->ws->query_value(sctx->ws,
372 RADEON_NUM_GFX_IBS);
373 break;
374 case SI_QUERY_CS_THREAD_BUSY:
375 ws_id = winsys_id_from_type(query->b.type);
376 query->end_result = sctx->ws->query_value(sctx->ws, ws_id);
377 query->end_time = os_time_get_nano();
378 break;
379 case SI_QUERY_GALLIUM_THREAD_BUSY:
380 query->end_result =
381 sctx->tc ? util_queue_get_thread_time_nano(&sctx->tc->queue, 0) : 0;
382 query->end_time = os_time_get_nano();
383 break;
384 case SI_QUERY_GPU_LOAD:
385 case SI_QUERY_GPU_SHADERS_BUSY:
386 case SI_QUERY_GPU_TA_BUSY:
387 case SI_QUERY_GPU_GDS_BUSY:
388 case SI_QUERY_GPU_VGT_BUSY:
389 case SI_QUERY_GPU_IA_BUSY:
390 case SI_QUERY_GPU_SX_BUSY:
391 case SI_QUERY_GPU_WD_BUSY:
392 case SI_QUERY_GPU_BCI_BUSY:
393 case SI_QUERY_GPU_SC_BUSY:
394 case SI_QUERY_GPU_PA_BUSY:
395 case SI_QUERY_GPU_DB_BUSY:
396 case SI_QUERY_GPU_CP_BUSY:
397 case SI_QUERY_GPU_CB_BUSY:
398 case SI_QUERY_GPU_SDMA_BUSY:
399 case SI_QUERY_GPU_PFP_BUSY:
400 case SI_QUERY_GPU_MEQ_BUSY:
401 case SI_QUERY_GPU_ME_BUSY:
402 case SI_QUERY_GPU_SURF_SYNC_BUSY:
403 case SI_QUERY_GPU_CP_DMA_BUSY:
404 case SI_QUERY_GPU_SCRATCH_RAM_BUSY:
405 query->end_result = si_end_counter(sctx->screen,
406 query->b.type,
407 query->begin_result);
408 query->begin_result = 0;
409 break;
410 case SI_QUERY_NUM_COMPILATIONS:
411 query->end_result = p_atomic_read(&sctx->screen->num_compilations);
412 break;
413 case SI_QUERY_NUM_SHADERS_CREATED:
414 query->end_result = p_atomic_read(&sctx->screen->num_shaders_created);
415 break;
416 case SI_QUERY_BACK_BUFFER_PS_DRAW_RATIO:
417 query->end_result = sctx->last_tex_ps_draw_ratio;
418 break;
419 case SI_QUERY_NUM_SHADER_CACHE_HITS:
420 query->end_result =
421 p_atomic_read(&sctx->screen->num_shader_cache_hits);
422 break;
423 case SI_QUERY_GPIN_ASIC_ID:
424 case SI_QUERY_GPIN_NUM_SIMD:
425 case SI_QUERY_GPIN_NUM_RB:
426 case SI_QUERY_GPIN_NUM_SPI:
427 case SI_QUERY_GPIN_NUM_SE:
428 break;
429 default:
430 unreachable("si_query_sw_end: bad query type");
431 }
432
433 return true;
434 }
435
436 static bool si_query_sw_get_result(struct si_context *sctx,
437 struct si_query *squery,
438 bool wait,
439 union pipe_query_result *result)
440 {
441 struct si_query_sw *query = (struct si_query_sw *)squery;
442
443 switch (query->b.type) {
444 case PIPE_QUERY_TIMESTAMP_DISJOINT:
445 /* Convert from cycles per millisecond to cycles per second (Hz). */
446 result->timestamp_disjoint.frequency =
447 (uint64_t)sctx->screen->info.clock_crystal_freq * 1000;
448 result->timestamp_disjoint.disjoint = false;
449 return true;
450 case PIPE_QUERY_GPU_FINISHED: {
451 struct pipe_screen *screen = sctx->b.screen;
452 struct pipe_context *ctx = squery->b.flushed ? NULL : &sctx->b;
453
454 result->b = screen->fence_finish(screen, ctx, query->fence,
455 wait ? PIPE_TIMEOUT_INFINITE : 0);
456 return result->b;
457 }
458
459 case SI_QUERY_GFX_BO_LIST_SIZE:
460 result->u64 = (query->end_result - query->begin_result) /
461 (query->end_time - query->begin_time);
462 return true;
463 case SI_QUERY_CS_THREAD_BUSY:
464 case SI_QUERY_GALLIUM_THREAD_BUSY:
465 result->u64 = (query->end_result - query->begin_result) * 100 /
466 (query->end_time - query->begin_time);
467 return true;
468 case SI_QUERY_GPIN_ASIC_ID:
469 result->u32 = 0;
470 return true;
471 case SI_QUERY_GPIN_NUM_SIMD:
472 result->u32 = sctx->screen->info.num_good_compute_units;
473 return true;
474 case SI_QUERY_GPIN_NUM_RB:
475 result->u32 = sctx->screen->info.num_render_backends;
476 return true;
477 case SI_QUERY_GPIN_NUM_SPI:
478 result->u32 = 1; /* all supported chips have one SPI per SE */
479 return true;
480 case SI_QUERY_GPIN_NUM_SE:
481 result->u32 = sctx->screen->info.max_se;
482 return true;
483 }
484
485 result->u64 = query->end_result - query->begin_result;
486
487 switch (query->b.type) {
488 case SI_QUERY_BUFFER_WAIT_TIME:
489 case SI_QUERY_GPU_TEMPERATURE:
490 result->u64 /= 1000;
491 break;
492 case SI_QUERY_CURRENT_GPU_SCLK:
493 case SI_QUERY_CURRENT_GPU_MCLK:
494 result->u64 *= 1000000;
495 break;
496 }
497
498 return true;
499 }
500
501
502 static const struct si_query_ops sw_query_ops = {
503 .destroy = si_query_sw_destroy,
504 .begin = si_query_sw_begin,
505 .end = si_query_sw_end,
506 .get_result = si_query_sw_get_result,
507 .get_result_resource = NULL
508 };
509
510 static struct pipe_query *si_query_sw_create(unsigned query_type)
511 {
512 struct si_query_sw *query;
513
514 query = CALLOC_STRUCT(si_query_sw);
515 if (!query)
516 return NULL;
517
518 query->b.type = query_type;
519 query->b.ops = &sw_query_ops;
520
521 return (struct pipe_query *)query;
522 }
523
524 void si_query_buffer_destroy(struct si_screen *sscreen, struct si_query_buffer *buffer)
525 {
526 struct si_query_buffer *prev = buffer->previous;
527
528 /* Release all query buffers. */
529 while (prev) {
530 struct si_query_buffer *qbuf = prev;
531 prev = prev->previous;
532 si_resource_reference(&qbuf->buf, NULL);
533 FREE(qbuf);
534 }
535
536 si_resource_reference(&buffer->buf, NULL);
537 }
538
539 void si_query_buffer_reset(struct si_context *sctx, struct si_query_buffer *buffer)
540 {
541 /* Discard all query buffers except for the oldest. */
542 while (buffer->previous) {
543 struct si_query_buffer *qbuf = buffer->previous;
544 buffer->previous = qbuf->previous;
545
546 si_resource_reference(&buffer->buf, NULL);
547 buffer->buf = qbuf->buf; /* move ownership */
548 FREE(qbuf);
549 }
550 buffer->results_end = 0;
551
552 if (!buffer->buf)
553 return;
554
555 /* Discard even the oldest buffer if it can't be mapped without a stall. */
556 if (si_rings_is_buffer_referenced(sctx, buffer->buf->buf, RADEON_USAGE_READWRITE) ||
557 !sctx->ws->buffer_wait(buffer->buf->buf, 0, RADEON_USAGE_READWRITE)) {
558 si_resource_reference(&buffer->buf, NULL);
559 } else {
560 buffer->unprepared = true;
561 }
562 }
563
564 bool si_query_buffer_alloc(struct si_context *sctx, struct si_query_buffer *buffer,
565 bool (*prepare_buffer)(struct si_context *, struct si_query_buffer*),
566 unsigned size)
567 {
568 bool unprepared = buffer->unprepared;
569 buffer->unprepared = false;
570
571 if (!buffer->buf || buffer->results_end + size > buffer->buf->b.b.width0) {
572 if (buffer->buf) {
573 struct si_query_buffer *qbuf = MALLOC_STRUCT(si_query_buffer);
574 memcpy(qbuf, buffer, sizeof(*qbuf));
575 buffer->previous = qbuf;
576 }
577 buffer->results_end = 0;
578
579 /* Queries are normally read by the CPU after
580 * being written by the gpu, hence staging is probably a good
581 * usage pattern.
582 */
583 struct si_screen *screen = sctx->screen;
584 unsigned buf_size = MAX2(size, screen->info.min_alloc_size);
585 buffer->buf = si_resource(
586 pipe_buffer_create(&screen->b, 0, PIPE_USAGE_STAGING, buf_size));
587 if (unlikely(!buffer->buf))
588 return false;
589 unprepared = true;
590 }
591
592 if (unprepared && prepare_buffer) {
593 if (unlikely(!prepare_buffer(sctx, buffer))) {
594 si_resource_reference(&buffer->buf, NULL);
595 return false;
596 }
597 }
598
599 return true;
600 }
601
602
603 void si_query_hw_destroy(struct si_screen *sscreen,
604 struct si_query *squery)
605 {
606 struct si_query_hw *query = (struct si_query_hw *)squery;
607
608 si_query_buffer_destroy(sscreen, &query->buffer);
609 si_resource_reference(&query->workaround_buf, NULL);
610 FREE(squery);
611 }
612
613 static bool si_query_hw_prepare_buffer(struct si_context *sctx,
614 struct si_query_buffer *qbuf)
615 {
616 static const struct si_query_hw si_query_hw_s;
617 struct si_query_hw *query = container_of(qbuf, &si_query_hw_s, buffer);
618 struct si_screen *screen = sctx->screen;
619
620 /* The caller ensures that the buffer is currently unused by the GPU. */
621 uint32_t *results = screen->ws->buffer_map(qbuf->buf->buf, NULL,
622 PIPE_TRANSFER_WRITE |
623 PIPE_TRANSFER_UNSYNCHRONIZED);
624 if (!results)
625 return false;
626
627 memset(results, 0, qbuf->buf->b.b.width0);
628
629 if (query->b.type == PIPE_QUERY_OCCLUSION_COUNTER ||
630 query->b.type == PIPE_QUERY_OCCLUSION_PREDICATE ||
631 query->b.type == PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE) {
632 unsigned max_rbs = screen->info.num_render_backends;
633 unsigned enabled_rb_mask = screen->info.enabled_rb_mask;
634 unsigned num_results;
635 unsigned i, j;
636
637 /* Set top bits for unused backends. */
638 num_results = qbuf->buf->b.b.width0 / query->result_size;
639 for (j = 0; j < num_results; j++) {
640 for (i = 0; i < max_rbs; i++) {
641 if (!(enabled_rb_mask & (1<<i))) {
642 results[(i * 4)+1] = 0x80000000;
643 results[(i * 4)+3] = 0x80000000;
644 }
645 }
646 results += 4 * max_rbs;
647 }
648 }
649
650 return true;
651 }
652
653 static void si_query_hw_get_result_resource(struct si_context *sctx,
654 struct si_query *squery,
655 bool wait,
656 enum pipe_query_value_type result_type,
657 int index,
658 struct pipe_resource *resource,
659 unsigned offset);
660
661 static void si_query_hw_do_emit_start(struct si_context *sctx,
662 struct si_query_hw *query,
663 struct si_resource *buffer,
664 uint64_t va);
665 static void si_query_hw_do_emit_stop(struct si_context *sctx,
666 struct si_query_hw *query,
667 struct si_resource *buffer,
668 uint64_t va);
669 static void si_query_hw_add_result(struct si_screen *sscreen,
670 struct si_query_hw *, void *buffer,
671 union pipe_query_result *result);
672 static void si_query_hw_clear_result(struct si_query_hw *,
673 union pipe_query_result *);
674
675 static struct si_query_hw_ops query_hw_default_hw_ops = {
676 .prepare_buffer = si_query_hw_prepare_buffer,
677 .emit_start = si_query_hw_do_emit_start,
678 .emit_stop = si_query_hw_do_emit_stop,
679 .clear_result = si_query_hw_clear_result,
680 .add_result = si_query_hw_add_result,
681 };
682
683 static struct pipe_query *si_query_hw_create(struct si_screen *sscreen,
684 unsigned query_type,
685 unsigned index)
686 {
687 struct si_query_hw *query = CALLOC_STRUCT(si_query_hw);
688 if (!query)
689 return NULL;
690
691 query->b.type = query_type;
692 query->b.ops = &query_hw_ops;
693 query->ops = &query_hw_default_hw_ops;
694
695 switch (query_type) {
696 case PIPE_QUERY_OCCLUSION_COUNTER:
697 case PIPE_QUERY_OCCLUSION_PREDICATE:
698 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
699 query->result_size = 16 * sscreen->info.num_render_backends;
700 query->result_size += 16; /* for the fence + alignment */
701 query->b.num_cs_dw_suspend = 6 + si_cp_write_fence_dwords(sscreen);
702 break;
703 case SI_QUERY_TIME_ELAPSED_SDMA:
704 /* GET_GLOBAL_TIMESTAMP only works if the offset is a multiple of 32. */
705 query->result_size = 64;
706 break;
707 case PIPE_QUERY_TIME_ELAPSED:
708 query->result_size = 24;
709 query->b.num_cs_dw_suspend = 8 + si_cp_write_fence_dwords(sscreen);
710 break;
711 case PIPE_QUERY_TIMESTAMP:
712 query->result_size = 16;
713 query->b.num_cs_dw_suspend = 8 + si_cp_write_fence_dwords(sscreen);
714 query->flags = SI_QUERY_HW_FLAG_NO_START;
715 break;
716 case PIPE_QUERY_PRIMITIVES_EMITTED:
717 case PIPE_QUERY_PRIMITIVES_GENERATED:
718 case PIPE_QUERY_SO_STATISTICS:
719 case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
720 /* NumPrimitivesWritten, PrimitiveStorageNeeded. */
721 query->result_size = 32;
722 query->b.num_cs_dw_suspend = 6;
723 query->stream = index;
724 break;
725 case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
726 /* NumPrimitivesWritten, PrimitiveStorageNeeded. */
727 query->result_size = 32 * SI_MAX_STREAMS;
728 query->b.num_cs_dw_suspend = 6 * SI_MAX_STREAMS;
729 break;
730 case PIPE_QUERY_PIPELINE_STATISTICS:
731 /* 11 values on GCN. */
732 query->result_size = 11 * 16;
733 query->result_size += 8; /* for the fence + alignment */
734 query->b.num_cs_dw_suspend = 6 + si_cp_write_fence_dwords(sscreen);
735 break;
736 default:
737 assert(0);
738 FREE(query);
739 return NULL;
740 }
741
742 return (struct pipe_query *)query;
743 }
744
745 static void si_update_occlusion_query_state(struct si_context *sctx,
746 unsigned type, int diff)
747 {
748 if (type == PIPE_QUERY_OCCLUSION_COUNTER ||
749 type == PIPE_QUERY_OCCLUSION_PREDICATE ||
750 type == PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE) {
751 bool old_enable = sctx->num_occlusion_queries != 0;
752 bool old_perfect_enable =
753 sctx->num_perfect_occlusion_queries != 0;
754 bool enable, perfect_enable;
755
756 sctx->num_occlusion_queries += diff;
757 assert(sctx->num_occlusion_queries >= 0);
758
759 if (type != PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE) {
760 sctx->num_perfect_occlusion_queries += diff;
761 assert(sctx->num_perfect_occlusion_queries >= 0);
762 }
763
764 enable = sctx->num_occlusion_queries != 0;
765 perfect_enable = sctx->num_perfect_occlusion_queries != 0;
766
767 if (enable != old_enable || perfect_enable != old_perfect_enable) {
768 si_set_occlusion_query_state(sctx, old_perfect_enable);
769 }
770 }
771 }
772
773 static unsigned event_type_for_stream(unsigned stream)
774 {
775 switch (stream) {
776 default:
777 case 0: return V_028A90_SAMPLE_STREAMOUTSTATS;
778 case 1: return V_028A90_SAMPLE_STREAMOUTSTATS1;
779 case 2: return V_028A90_SAMPLE_STREAMOUTSTATS2;
780 case 3: return V_028A90_SAMPLE_STREAMOUTSTATS3;
781 }
782 }
783
784 static void emit_sample_streamout(struct radeon_cmdbuf *cs, uint64_t va,
785 unsigned stream)
786 {
787 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
788 radeon_emit(cs, EVENT_TYPE(event_type_for_stream(stream)) | EVENT_INDEX(3));
789 radeon_emit(cs, va);
790 radeon_emit(cs, va >> 32);
791 }
792
793 static void si_query_hw_do_emit_start(struct si_context *sctx,
794 struct si_query_hw *query,
795 struct si_resource *buffer,
796 uint64_t va)
797 {
798 struct radeon_cmdbuf *cs = sctx->gfx_cs;
799
800 switch (query->b.type) {
801 case SI_QUERY_TIME_ELAPSED_SDMA:
802 si_dma_emit_timestamp(sctx, buffer, va - buffer->gpu_address);
803 return;
804 case PIPE_QUERY_OCCLUSION_COUNTER:
805 case PIPE_QUERY_OCCLUSION_PREDICATE:
806 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
807 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
808 radeon_emit(cs, EVENT_TYPE(V_028A90_ZPASS_DONE) | EVENT_INDEX(1));
809 radeon_emit(cs, va);
810 radeon_emit(cs, va >> 32);
811 break;
812 case PIPE_QUERY_PRIMITIVES_EMITTED:
813 case PIPE_QUERY_PRIMITIVES_GENERATED:
814 case PIPE_QUERY_SO_STATISTICS:
815 case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
816 emit_sample_streamout(cs, va, query->stream);
817 break;
818 case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
819 for (unsigned stream = 0; stream < SI_MAX_STREAMS; ++stream)
820 emit_sample_streamout(cs, va + 32 * stream, stream);
821 break;
822 case PIPE_QUERY_TIME_ELAPSED:
823 si_cp_release_mem(sctx, cs, V_028A90_BOTTOM_OF_PIPE_TS, 0,
824 EOP_DST_SEL_MEM, EOP_INT_SEL_NONE,
825 EOP_DATA_SEL_TIMESTAMP, NULL, va,
826 0, query->b.type);
827 break;
828 case PIPE_QUERY_PIPELINE_STATISTICS:
829 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
830 radeon_emit(cs, EVENT_TYPE(V_028A90_SAMPLE_PIPELINESTAT) | EVENT_INDEX(2));
831 radeon_emit(cs, va);
832 radeon_emit(cs, va >> 32);
833 break;
834 default:
835 assert(0);
836 }
837 radeon_add_to_buffer_list(sctx, sctx->gfx_cs, query->buffer.buf, RADEON_USAGE_WRITE,
838 RADEON_PRIO_QUERY);
839 }
840
841 static void si_query_hw_emit_start(struct si_context *sctx,
842 struct si_query_hw *query)
843 {
844 uint64_t va;
845
846 if (!si_query_buffer_alloc(sctx, &query->buffer, query->ops->prepare_buffer,
847 query->result_size))
848 return;
849
850 si_update_occlusion_query_state(sctx, query->b.type, 1);
851 si_update_prims_generated_query_state(sctx, query->b.type, 1);
852
853 if (query->b.type == PIPE_QUERY_PIPELINE_STATISTICS)
854 sctx->num_pipeline_stat_queries++;
855
856 if (query->b.type != SI_QUERY_TIME_ELAPSED_SDMA)
857 si_need_gfx_cs_space(sctx);
858
859 va = query->buffer.buf->gpu_address + query->buffer.results_end;
860 query->ops->emit_start(sctx, query, query->buffer.buf, va);
861 }
862
863 static void si_query_hw_do_emit_stop(struct si_context *sctx,
864 struct si_query_hw *query,
865 struct si_resource *buffer,
866 uint64_t va)
867 {
868 struct radeon_cmdbuf *cs = sctx->gfx_cs;
869 uint64_t fence_va = 0;
870
871 switch (query->b.type) {
872 case SI_QUERY_TIME_ELAPSED_SDMA:
873 si_dma_emit_timestamp(sctx, buffer, va + 32 - buffer->gpu_address);
874 return;
875 case PIPE_QUERY_OCCLUSION_COUNTER:
876 case PIPE_QUERY_OCCLUSION_PREDICATE:
877 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
878 va += 8;
879 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
880 radeon_emit(cs, EVENT_TYPE(V_028A90_ZPASS_DONE) | EVENT_INDEX(1));
881 radeon_emit(cs, va);
882 radeon_emit(cs, va >> 32);
883
884 fence_va = va + sctx->screen->info.num_render_backends * 16 - 8;
885 break;
886 case PIPE_QUERY_PRIMITIVES_EMITTED:
887 case PIPE_QUERY_PRIMITIVES_GENERATED:
888 case PIPE_QUERY_SO_STATISTICS:
889 case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
890 va += 16;
891 emit_sample_streamout(cs, va, query->stream);
892 break;
893 case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
894 va += 16;
895 for (unsigned stream = 0; stream < SI_MAX_STREAMS; ++stream)
896 emit_sample_streamout(cs, va + 32 * stream, stream);
897 break;
898 case PIPE_QUERY_TIME_ELAPSED:
899 va += 8;
900 /* fall through */
901 case PIPE_QUERY_TIMESTAMP:
902 si_cp_release_mem(sctx, cs, V_028A90_BOTTOM_OF_PIPE_TS, 0,
903 EOP_DST_SEL_MEM, EOP_INT_SEL_NONE,
904 EOP_DATA_SEL_TIMESTAMP, NULL, va,
905 0, query->b.type);
906 fence_va = va + 8;
907 break;
908 case PIPE_QUERY_PIPELINE_STATISTICS: {
909 unsigned sample_size = (query->result_size - 8) / 2;
910
911 va += sample_size;
912 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
913 radeon_emit(cs, EVENT_TYPE(V_028A90_SAMPLE_PIPELINESTAT) | EVENT_INDEX(2));
914 radeon_emit(cs, va);
915 radeon_emit(cs, va >> 32);
916
917 fence_va = va + sample_size;
918 break;
919 }
920 default:
921 assert(0);
922 }
923 radeon_add_to_buffer_list(sctx, sctx->gfx_cs, query->buffer.buf, RADEON_USAGE_WRITE,
924 RADEON_PRIO_QUERY);
925
926 if (fence_va) {
927 si_cp_release_mem(sctx, cs, V_028A90_BOTTOM_OF_PIPE_TS, 0,
928 EOP_DST_SEL_MEM, EOP_INT_SEL_NONE,
929 EOP_DATA_SEL_VALUE_32BIT,
930 query->buffer.buf, fence_va, 0x80000000,
931 query->b.type);
932 }
933 }
934
935 static void si_query_hw_emit_stop(struct si_context *sctx,
936 struct si_query_hw *query)
937 {
938 uint64_t va;
939
940 /* The queries which need begin already called this in begin_query. */
941 if (query->flags & SI_QUERY_HW_FLAG_NO_START) {
942 si_need_gfx_cs_space(sctx);
943 if (!si_query_buffer_alloc(sctx, &query->buffer, query->ops->prepare_buffer,
944 query->result_size))
945 return;
946 }
947
948 if (!query->buffer.buf)
949 return; // previous buffer allocation failure
950
951 /* emit end query */
952 va = query->buffer.buf->gpu_address + query->buffer.results_end;
953
954 query->ops->emit_stop(sctx, query, query->buffer.buf, va);
955
956 query->buffer.results_end += query->result_size;
957
958 si_update_occlusion_query_state(sctx, query->b.type, -1);
959 si_update_prims_generated_query_state(sctx, query->b.type, -1);
960
961 if (query->b.type == PIPE_QUERY_PIPELINE_STATISTICS)
962 sctx->num_pipeline_stat_queries--;
963 }
964
965 static void emit_set_predicate(struct si_context *ctx,
966 struct si_resource *buf, uint64_t va,
967 uint32_t op)
968 {
969 struct radeon_cmdbuf *cs = ctx->gfx_cs;
970
971 if (ctx->chip_class >= GFX9) {
972 radeon_emit(cs, PKT3(PKT3_SET_PREDICATION, 2, 0));
973 radeon_emit(cs, op);
974 radeon_emit(cs, va);
975 radeon_emit(cs, va >> 32);
976 } else {
977 radeon_emit(cs, PKT3(PKT3_SET_PREDICATION, 1, 0));
978 radeon_emit(cs, va);
979 radeon_emit(cs, op | ((va >> 32) & 0xFF));
980 }
981 radeon_add_to_buffer_list(ctx, ctx->gfx_cs, buf, RADEON_USAGE_READ,
982 RADEON_PRIO_QUERY);
983 }
984
985 static void si_emit_query_predication(struct si_context *ctx)
986 {
987 struct si_query_hw *query = (struct si_query_hw *)ctx->render_cond;
988 struct si_query_buffer *qbuf;
989 uint32_t op;
990 bool flag_wait, invert;
991
992 if (!query)
993 return;
994
995 invert = ctx->render_cond_invert;
996 flag_wait = ctx->render_cond_mode == PIPE_RENDER_COND_WAIT ||
997 ctx->render_cond_mode == PIPE_RENDER_COND_BY_REGION_WAIT;
998
999 if (query->workaround_buf) {
1000 op = PRED_OP(PREDICATION_OP_BOOL64);
1001 } else {
1002 switch (query->b.type) {
1003 case PIPE_QUERY_OCCLUSION_COUNTER:
1004 case PIPE_QUERY_OCCLUSION_PREDICATE:
1005 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
1006 op = PRED_OP(PREDICATION_OP_ZPASS);
1007 break;
1008 case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
1009 case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
1010 op = PRED_OP(PREDICATION_OP_PRIMCOUNT);
1011 invert = !invert;
1012 break;
1013 default:
1014 assert(0);
1015 return;
1016 }
1017 }
1018
1019 /* if true then invert, see GL_ARB_conditional_render_inverted */
1020 if (invert)
1021 op |= PREDICATION_DRAW_NOT_VISIBLE; /* Draw if not visible or overflow */
1022 else
1023 op |= PREDICATION_DRAW_VISIBLE; /* Draw if visible or no overflow */
1024
1025 /* Use the value written by compute shader as a workaround. Note that
1026 * the wait flag does not apply in this predication mode.
1027 *
1028 * The shader outputs the result value to L2. Workarounds only affect GFX8
1029 * and later, where the CP reads data from L2, so we don't need an
1030 * additional flush.
1031 */
1032 if (query->workaround_buf) {
1033 uint64_t va = query->workaround_buf->gpu_address + query->workaround_offset;
1034 emit_set_predicate(ctx, query->workaround_buf, va, op);
1035 return;
1036 }
1037
1038 op |= flag_wait ? PREDICATION_HINT_WAIT : PREDICATION_HINT_NOWAIT_DRAW;
1039
1040 /* emit predicate packets for all data blocks */
1041 for (qbuf = &query->buffer; qbuf; qbuf = qbuf->previous) {
1042 unsigned results_base = 0;
1043 uint64_t va_base = qbuf->buf->gpu_address;
1044
1045 while (results_base < qbuf->results_end) {
1046 uint64_t va = va_base + results_base;
1047
1048 if (query->b.type == PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE) {
1049 for (unsigned stream = 0; stream < SI_MAX_STREAMS; ++stream) {
1050 emit_set_predicate(ctx, qbuf->buf, va + 32 * stream, op);
1051
1052 /* set CONTINUE bit for all packets except the first */
1053 op |= PREDICATION_CONTINUE;
1054 }
1055 } else {
1056 emit_set_predicate(ctx, qbuf->buf, va, op);
1057 op |= PREDICATION_CONTINUE;
1058 }
1059
1060 results_base += query->result_size;
1061 }
1062 }
1063 }
1064
1065 static struct pipe_query *si_create_query(struct pipe_context *ctx, unsigned query_type, unsigned index)
1066 {
1067 struct si_screen *sscreen =
1068 (struct si_screen *)ctx->screen;
1069
1070 if (query_type == PIPE_QUERY_TIMESTAMP_DISJOINT ||
1071 query_type == PIPE_QUERY_GPU_FINISHED ||
1072 (query_type >= PIPE_QUERY_DRIVER_SPECIFIC &&
1073 query_type != SI_QUERY_TIME_ELAPSED_SDMA))
1074 return si_query_sw_create(query_type);
1075
1076 return si_query_hw_create(sscreen, query_type, index);
1077 }
1078
1079 static void si_destroy_query(struct pipe_context *ctx, struct pipe_query *query)
1080 {
1081 struct si_context *sctx = (struct si_context *)ctx;
1082 struct si_query *squery = (struct si_query *)query;
1083
1084 squery->ops->destroy(sctx->screen, squery);
1085 }
1086
1087 static boolean si_begin_query(struct pipe_context *ctx,
1088 struct pipe_query *query)
1089 {
1090 struct si_context *sctx = (struct si_context *)ctx;
1091 struct si_query *squery = (struct si_query *)query;
1092
1093 return squery->ops->begin(sctx, squery);
1094 }
1095
1096 bool si_query_hw_begin(struct si_context *sctx,
1097 struct si_query *squery)
1098 {
1099 struct si_query_hw *query = (struct si_query_hw *)squery;
1100
1101 if (query->flags & SI_QUERY_HW_FLAG_NO_START) {
1102 assert(0);
1103 return false;
1104 }
1105
1106 if (!(query->flags & SI_QUERY_HW_FLAG_BEGIN_RESUMES))
1107 si_query_buffer_reset(sctx, &query->buffer);
1108
1109 si_resource_reference(&query->workaround_buf, NULL);
1110
1111 si_query_hw_emit_start(sctx, query);
1112 if (!query->buffer.buf)
1113 return false;
1114
1115 LIST_ADDTAIL(&query->b.active_list, &sctx->active_queries);
1116 sctx->num_cs_dw_queries_suspend += query->b.num_cs_dw_suspend;
1117 return true;
1118 }
1119
1120 static bool si_end_query(struct pipe_context *ctx, struct pipe_query *query)
1121 {
1122 struct si_context *sctx = (struct si_context *)ctx;
1123 struct si_query *squery = (struct si_query *)query;
1124
1125 return squery->ops->end(sctx, squery);
1126 }
1127
1128 bool si_query_hw_end(struct si_context *sctx,
1129 struct si_query *squery)
1130 {
1131 struct si_query_hw *query = (struct si_query_hw *)squery;
1132
1133 if (query->flags & SI_QUERY_HW_FLAG_NO_START)
1134 si_query_buffer_reset(sctx, &query->buffer);
1135
1136 si_query_hw_emit_stop(sctx, query);
1137
1138 if (!(query->flags & SI_QUERY_HW_FLAG_NO_START)) {
1139 LIST_DELINIT(&query->b.active_list);
1140 sctx->num_cs_dw_queries_suspend -= query->b.num_cs_dw_suspend;
1141 }
1142
1143 if (!query->buffer.buf)
1144 return false;
1145
1146 return true;
1147 }
1148
1149 static void si_get_hw_query_params(struct si_context *sctx,
1150 struct si_query_hw *squery, int index,
1151 struct si_hw_query_params *params)
1152 {
1153 unsigned max_rbs = sctx->screen->info.num_render_backends;
1154
1155 params->pair_stride = 0;
1156 params->pair_count = 1;
1157
1158 switch (squery->b.type) {
1159 case PIPE_QUERY_OCCLUSION_COUNTER:
1160 case PIPE_QUERY_OCCLUSION_PREDICATE:
1161 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
1162 params->start_offset = 0;
1163 params->end_offset = 8;
1164 params->fence_offset = max_rbs * 16;
1165 params->pair_stride = 16;
1166 params->pair_count = max_rbs;
1167 break;
1168 case PIPE_QUERY_TIME_ELAPSED:
1169 params->start_offset = 0;
1170 params->end_offset = 8;
1171 params->fence_offset = 16;
1172 break;
1173 case PIPE_QUERY_TIMESTAMP:
1174 params->start_offset = 0;
1175 params->end_offset = 0;
1176 params->fence_offset = 8;
1177 break;
1178 case PIPE_QUERY_PRIMITIVES_EMITTED:
1179 params->start_offset = 8;
1180 params->end_offset = 24;
1181 params->fence_offset = params->end_offset + 4;
1182 break;
1183 case PIPE_QUERY_PRIMITIVES_GENERATED:
1184 params->start_offset = 0;
1185 params->end_offset = 16;
1186 params->fence_offset = params->end_offset + 4;
1187 break;
1188 case PIPE_QUERY_SO_STATISTICS:
1189 params->start_offset = 8 - index * 8;
1190 params->end_offset = 24 - index * 8;
1191 params->fence_offset = params->end_offset + 4;
1192 break;
1193 case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
1194 params->pair_count = SI_MAX_STREAMS;
1195 params->pair_stride = 32;
1196 case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
1197 params->start_offset = 0;
1198 params->end_offset = 16;
1199
1200 /* We can re-use the high dword of the last 64-bit value as a
1201 * fence: it is initialized as 0, and the high bit is set by
1202 * the write of the streamout stats event.
1203 */
1204 params->fence_offset = squery->result_size - 4;
1205 break;
1206 case PIPE_QUERY_PIPELINE_STATISTICS:
1207 {
1208 static const unsigned offsets[] = {56, 48, 24, 32, 40, 16, 8, 0, 64, 72, 80};
1209 params->start_offset = offsets[index];
1210 params->end_offset = 88 + offsets[index];
1211 params->fence_offset = 2 * 88;
1212 break;
1213 }
1214 default:
1215 unreachable("si_get_hw_query_params unsupported");
1216 }
1217 }
1218
1219 static unsigned si_query_read_result(void *map, unsigned start_index, unsigned end_index,
1220 bool test_status_bit)
1221 {
1222 uint32_t *current_result = (uint32_t*)map;
1223 uint64_t start, end;
1224
1225 start = (uint64_t)current_result[start_index] |
1226 (uint64_t)current_result[start_index+1] << 32;
1227 end = (uint64_t)current_result[end_index] |
1228 (uint64_t)current_result[end_index+1] << 32;
1229
1230 if (!test_status_bit ||
1231 ((start & 0x8000000000000000UL) && (end & 0x8000000000000000UL))) {
1232 return end - start;
1233 }
1234 return 0;
1235 }
1236
1237 static void si_query_hw_add_result(struct si_screen *sscreen,
1238 struct si_query_hw *query,
1239 void *buffer,
1240 union pipe_query_result *result)
1241 {
1242 unsigned max_rbs = sscreen->info.num_render_backends;
1243
1244 switch (query->b.type) {
1245 case PIPE_QUERY_OCCLUSION_COUNTER: {
1246 for (unsigned i = 0; i < max_rbs; ++i) {
1247 unsigned results_base = i * 16;
1248 result->u64 +=
1249 si_query_read_result(buffer + results_base, 0, 2, true);
1250 }
1251 break;
1252 }
1253 case PIPE_QUERY_OCCLUSION_PREDICATE:
1254 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE: {
1255 for (unsigned i = 0; i < max_rbs; ++i) {
1256 unsigned results_base = i * 16;
1257 result->b = result->b ||
1258 si_query_read_result(buffer + results_base, 0, 2, true) != 0;
1259 }
1260 break;
1261 }
1262 case PIPE_QUERY_TIME_ELAPSED:
1263 result->u64 += si_query_read_result(buffer, 0, 2, false);
1264 break;
1265 case SI_QUERY_TIME_ELAPSED_SDMA:
1266 result->u64 += si_query_read_result(buffer, 0, 32/4, false);
1267 break;
1268 case PIPE_QUERY_TIMESTAMP:
1269 result->u64 = *(uint64_t*)buffer;
1270 break;
1271 case PIPE_QUERY_PRIMITIVES_EMITTED:
1272 /* SAMPLE_STREAMOUTSTATS stores this structure:
1273 * {
1274 * u64 NumPrimitivesWritten;
1275 * u64 PrimitiveStorageNeeded;
1276 * }
1277 * We only need NumPrimitivesWritten here. */
1278 result->u64 += si_query_read_result(buffer, 2, 6, true);
1279 break;
1280 case PIPE_QUERY_PRIMITIVES_GENERATED:
1281 /* Here we read PrimitiveStorageNeeded. */
1282 result->u64 += si_query_read_result(buffer, 0, 4, true);
1283 break;
1284 case PIPE_QUERY_SO_STATISTICS:
1285 result->so_statistics.num_primitives_written +=
1286 si_query_read_result(buffer, 2, 6, true);
1287 result->so_statistics.primitives_storage_needed +=
1288 si_query_read_result(buffer, 0, 4, true);
1289 break;
1290 case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
1291 result->b = result->b ||
1292 si_query_read_result(buffer, 2, 6, true) !=
1293 si_query_read_result(buffer, 0, 4, true);
1294 break;
1295 case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
1296 for (unsigned stream = 0; stream < SI_MAX_STREAMS; ++stream) {
1297 result->b = result->b ||
1298 si_query_read_result(buffer, 2, 6, true) !=
1299 si_query_read_result(buffer, 0, 4, true);
1300 buffer = (char *)buffer + 32;
1301 }
1302 break;
1303 case PIPE_QUERY_PIPELINE_STATISTICS:
1304 result->pipeline_statistics.ps_invocations +=
1305 si_query_read_result(buffer, 0, 22, false);
1306 result->pipeline_statistics.c_primitives +=
1307 si_query_read_result(buffer, 2, 24, false);
1308 result->pipeline_statistics.c_invocations +=
1309 si_query_read_result(buffer, 4, 26, false);
1310 result->pipeline_statistics.vs_invocations +=
1311 si_query_read_result(buffer, 6, 28, false);
1312 result->pipeline_statistics.gs_invocations +=
1313 si_query_read_result(buffer, 8, 30, false);
1314 result->pipeline_statistics.gs_primitives +=
1315 si_query_read_result(buffer, 10, 32, false);
1316 result->pipeline_statistics.ia_primitives +=
1317 si_query_read_result(buffer, 12, 34, false);
1318 result->pipeline_statistics.ia_vertices +=
1319 si_query_read_result(buffer, 14, 36, false);
1320 result->pipeline_statistics.hs_invocations +=
1321 si_query_read_result(buffer, 16, 38, false);
1322 result->pipeline_statistics.ds_invocations +=
1323 si_query_read_result(buffer, 18, 40, false);
1324 result->pipeline_statistics.cs_invocations +=
1325 si_query_read_result(buffer, 20, 42, false);
1326 #if 0 /* for testing */
1327 printf("Pipeline stats: IA verts=%llu, IA prims=%llu, VS=%llu, HS=%llu, "
1328 "DS=%llu, GS=%llu, GS prims=%llu, Clipper=%llu, "
1329 "Clipper prims=%llu, PS=%llu, CS=%llu\n",
1330 result->pipeline_statistics.ia_vertices,
1331 result->pipeline_statistics.ia_primitives,
1332 result->pipeline_statistics.vs_invocations,
1333 result->pipeline_statistics.hs_invocations,
1334 result->pipeline_statistics.ds_invocations,
1335 result->pipeline_statistics.gs_invocations,
1336 result->pipeline_statistics.gs_primitives,
1337 result->pipeline_statistics.c_invocations,
1338 result->pipeline_statistics.c_primitives,
1339 result->pipeline_statistics.ps_invocations,
1340 result->pipeline_statistics.cs_invocations);
1341 #endif
1342 break;
1343 default:
1344 assert(0);
1345 }
1346 }
1347
1348 void si_query_hw_suspend(struct si_context *sctx, struct si_query *query)
1349 {
1350 si_query_hw_emit_stop(sctx, (struct si_query_hw *)query);
1351 }
1352
1353 void si_query_hw_resume(struct si_context *sctx, struct si_query *query)
1354 {
1355 si_query_hw_emit_start(sctx, (struct si_query_hw *)query);
1356 }
1357
1358 static const struct si_query_ops query_hw_ops = {
1359 .destroy = si_query_hw_destroy,
1360 .begin = si_query_hw_begin,
1361 .end = si_query_hw_end,
1362 .get_result = si_query_hw_get_result,
1363 .get_result_resource = si_query_hw_get_result_resource,
1364
1365 .suspend = si_query_hw_suspend,
1366 .resume = si_query_hw_resume,
1367 };
1368
1369 static boolean si_get_query_result(struct pipe_context *ctx,
1370 struct pipe_query *query, boolean wait,
1371 union pipe_query_result *result)
1372 {
1373 struct si_context *sctx = (struct si_context *)ctx;
1374 struct si_query *squery = (struct si_query *)query;
1375
1376 return squery->ops->get_result(sctx, squery, wait, result);
1377 }
1378
1379 static void si_get_query_result_resource(struct pipe_context *ctx,
1380 struct pipe_query *query,
1381 boolean wait,
1382 enum pipe_query_value_type result_type,
1383 int index,
1384 struct pipe_resource *resource,
1385 unsigned offset)
1386 {
1387 struct si_context *sctx = (struct si_context *)ctx;
1388 struct si_query *squery = (struct si_query *)query;
1389
1390 squery->ops->get_result_resource(sctx, squery, wait, result_type, index,
1391 resource, offset);
1392 }
1393
1394 static void si_query_hw_clear_result(struct si_query_hw *query,
1395 union pipe_query_result *result)
1396 {
1397 util_query_clear_result(result, query->b.type);
1398 }
1399
1400 bool si_query_hw_get_result(struct si_context *sctx,
1401 struct si_query *squery,
1402 bool wait, union pipe_query_result *result)
1403 {
1404 struct si_screen *sscreen = sctx->screen;
1405 struct si_query_hw *query = (struct si_query_hw *)squery;
1406 struct si_query_buffer *qbuf;
1407
1408 query->ops->clear_result(query, result);
1409
1410 for (qbuf = &query->buffer; qbuf; qbuf = qbuf->previous) {
1411 unsigned usage = PIPE_TRANSFER_READ |
1412 (wait ? 0 : PIPE_TRANSFER_DONTBLOCK);
1413 unsigned results_base = 0;
1414 void *map;
1415
1416 if (squery->b.flushed)
1417 map = sctx->ws->buffer_map(qbuf->buf->buf, NULL, usage);
1418 else
1419 map = si_buffer_map_sync_with_rings(sctx, qbuf->buf, usage);
1420
1421 if (!map)
1422 return false;
1423
1424 while (results_base != qbuf->results_end) {
1425 query->ops->add_result(sscreen, query, map + results_base,
1426 result);
1427 results_base += query->result_size;
1428 }
1429 }
1430
1431 /* Convert the time to expected units. */
1432 if (squery->type == PIPE_QUERY_TIME_ELAPSED ||
1433 squery->type == SI_QUERY_TIME_ELAPSED_SDMA ||
1434 squery->type == PIPE_QUERY_TIMESTAMP) {
1435 result->u64 = (1000000 * result->u64) / sscreen->info.clock_crystal_freq;
1436 }
1437 return true;
1438 }
1439
1440 static void si_restore_qbo_state(struct si_context *sctx,
1441 struct si_qbo_state *st)
1442 {
1443 sctx->b.bind_compute_state(&sctx->b, st->saved_compute);
1444
1445 sctx->b.set_constant_buffer(&sctx->b, PIPE_SHADER_COMPUTE, 0, &st->saved_const0);
1446 pipe_resource_reference(&st->saved_const0.buffer, NULL);
1447
1448 sctx->b.set_shader_buffers(&sctx->b, PIPE_SHADER_COMPUTE, 0, 3, st->saved_ssbo,
1449 st->saved_ssbo_writable_mask);
1450 for (unsigned i = 0; i < 3; ++i)
1451 pipe_resource_reference(&st->saved_ssbo[i].buffer, NULL);
1452 }
1453
1454 static void si_query_hw_get_result_resource(struct si_context *sctx,
1455 struct si_query *squery,
1456 bool wait,
1457 enum pipe_query_value_type result_type,
1458 int index,
1459 struct pipe_resource *resource,
1460 unsigned offset)
1461 {
1462 struct si_query_hw *query = (struct si_query_hw *)squery;
1463 struct si_query_buffer *qbuf;
1464 struct si_query_buffer *qbuf_prev;
1465 struct pipe_resource *tmp_buffer = NULL;
1466 unsigned tmp_buffer_offset = 0;
1467 struct si_qbo_state saved_state = {};
1468 struct pipe_grid_info grid = {};
1469 struct pipe_constant_buffer constant_buffer = {};
1470 struct pipe_shader_buffer ssbo[3];
1471 struct si_hw_query_params params;
1472 struct {
1473 uint32_t end_offset;
1474 uint32_t result_stride;
1475 uint32_t result_count;
1476 uint32_t config;
1477 uint32_t fence_offset;
1478 uint32_t pair_stride;
1479 uint32_t pair_count;
1480 } consts;
1481
1482 if (!sctx->query_result_shader) {
1483 sctx->query_result_shader = si_create_query_result_cs(sctx);
1484 if (!sctx->query_result_shader)
1485 return;
1486 }
1487
1488 if (query->buffer.previous) {
1489 u_suballocator_alloc(sctx->allocator_zeroed_memory, 16, 16,
1490 &tmp_buffer_offset, &tmp_buffer);
1491 if (!tmp_buffer)
1492 return;
1493 }
1494
1495 si_save_qbo_state(sctx, &saved_state);
1496
1497 si_get_hw_query_params(sctx, query, index >= 0 ? index : 0, &params);
1498 consts.end_offset = params.end_offset - params.start_offset;
1499 consts.fence_offset = params.fence_offset - params.start_offset;
1500 consts.result_stride = query->result_size;
1501 consts.pair_stride = params.pair_stride;
1502 consts.pair_count = params.pair_count;
1503
1504 constant_buffer.buffer_size = sizeof(consts);
1505 constant_buffer.user_buffer = &consts;
1506
1507 ssbo[1].buffer = tmp_buffer;
1508 ssbo[1].buffer_offset = tmp_buffer_offset;
1509 ssbo[1].buffer_size = 16;
1510
1511 ssbo[2] = ssbo[1];
1512
1513 sctx->b.bind_compute_state(&sctx->b, sctx->query_result_shader);
1514
1515 grid.block[0] = 1;
1516 grid.block[1] = 1;
1517 grid.block[2] = 1;
1518 grid.grid[0] = 1;
1519 grid.grid[1] = 1;
1520 grid.grid[2] = 1;
1521
1522 consts.config = 0;
1523 if (index < 0)
1524 consts.config |= 4;
1525 if (query->b.type == PIPE_QUERY_OCCLUSION_PREDICATE ||
1526 query->b.type == PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE)
1527 consts.config |= 8;
1528 else if (query->b.type == PIPE_QUERY_SO_OVERFLOW_PREDICATE ||
1529 query->b.type == PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE)
1530 consts.config |= 8 | 256;
1531 else if (query->b.type == PIPE_QUERY_TIMESTAMP ||
1532 query->b.type == PIPE_QUERY_TIME_ELAPSED)
1533 consts.config |= 32;
1534
1535 switch (result_type) {
1536 case PIPE_QUERY_TYPE_U64:
1537 case PIPE_QUERY_TYPE_I64:
1538 consts.config |= 64;
1539 break;
1540 case PIPE_QUERY_TYPE_I32:
1541 consts.config |= 128;
1542 break;
1543 case PIPE_QUERY_TYPE_U32:
1544 break;
1545 }
1546
1547 sctx->flags |= sctx->screen->barrier_flags.cp_to_L2;
1548
1549 for (qbuf = &query->buffer; qbuf; qbuf = qbuf_prev) {
1550 if (query->b.type != PIPE_QUERY_TIMESTAMP) {
1551 qbuf_prev = qbuf->previous;
1552 consts.result_count = qbuf->results_end / query->result_size;
1553 consts.config &= ~3;
1554 if (qbuf != &query->buffer)
1555 consts.config |= 1;
1556 if (qbuf->previous)
1557 consts.config |= 2;
1558 } else {
1559 /* Only read the last timestamp. */
1560 qbuf_prev = NULL;
1561 consts.result_count = 0;
1562 consts.config |= 16;
1563 params.start_offset += qbuf->results_end - query->result_size;
1564 }
1565
1566 sctx->b.set_constant_buffer(&sctx->b, PIPE_SHADER_COMPUTE, 0, &constant_buffer);
1567
1568 ssbo[0].buffer = &qbuf->buf->b.b;
1569 ssbo[0].buffer_offset = params.start_offset;
1570 ssbo[0].buffer_size = qbuf->results_end - params.start_offset;
1571
1572 if (!qbuf->previous) {
1573 ssbo[2].buffer = resource;
1574 ssbo[2].buffer_offset = offset;
1575 ssbo[2].buffer_size = 8;
1576
1577 si_resource(resource)->TC_L2_dirty = true;
1578 }
1579
1580 sctx->b.set_shader_buffers(&sctx->b, PIPE_SHADER_COMPUTE, 0, 3, ssbo,
1581 1 << 2);
1582
1583 if (wait && qbuf == &query->buffer) {
1584 uint64_t va;
1585
1586 /* Wait for result availability. Wait only for readiness
1587 * of the last entry, since the fence writes should be
1588 * serialized in the CP.
1589 */
1590 va = qbuf->buf->gpu_address + qbuf->results_end - query->result_size;
1591 va += params.fence_offset;
1592
1593 si_cp_wait_mem(sctx, sctx->gfx_cs, va, 0x80000000,
1594 0x80000000, WAIT_REG_MEM_EQUAL);
1595 }
1596
1597 sctx->b.launch_grid(&sctx->b, &grid);
1598 sctx->flags |= SI_CONTEXT_CS_PARTIAL_FLUSH;
1599 }
1600
1601 si_restore_qbo_state(sctx, &saved_state);
1602 pipe_resource_reference(&tmp_buffer, NULL);
1603 }
1604
1605 static void si_render_condition(struct pipe_context *ctx,
1606 struct pipe_query *query,
1607 boolean condition,
1608 enum pipe_render_cond_flag mode)
1609 {
1610 struct si_context *sctx = (struct si_context *)ctx;
1611 struct si_query_hw *squery = (struct si_query_hw *)query;
1612 struct si_atom *atom = &sctx->atoms.s.render_cond;
1613
1614 if (query) {
1615 bool needs_workaround = false;
1616
1617 /* There was a firmware regression in GFX8 which causes successive
1618 * SET_PREDICATION packets to give the wrong answer for
1619 * non-inverted stream overflow predication.
1620 */
1621 if (((sctx->chip_class == GFX8 && sctx->screen->info.pfp_fw_feature < 49) ||
1622 (sctx->chip_class == GFX9 && sctx->screen->info.pfp_fw_feature < 38)) &&
1623 !condition &&
1624 (squery->b.type == PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE ||
1625 (squery->b.type == PIPE_QUERY_SO_OVERFLOW_PREDICATE &&
1626 (squery->buffer.previous ||
1627 squery->buffer.results_end > squery->result_size)))) {
1628 needs_workaround = true;
1629 }
1630
1631 if (needs_workaround && !squery->workaround_buf) {
1632 bool old_force_off = sctx->render_cond_force_off;
1633 sctx->render_cond_force_off = true;
1634
1635 u_suballocator_alloc(
1636 sctx->allocator_zeroed_memory, 8, 8,
1637 &squery->workaround_offset,
1638 (struct pipe_resource **)&squery->workaround_buf);
1639
1640 /* Reset to NULL to avoid a redundant SET_PREDICATION
1641 * from launching the compute grid.
1642 */
1643 sctx->render_cond = NULL;
1644
1645 ctx->get_query_result_resource(
1646 ctx, query, true, PIPE_QUERY_TYPE_U64, 0,
1647 &squery->workaround_buf->b.b, squery->workaround_offset);
1648
1649 /* Settings this in the render cond atom is too late,
1650 * so set it here. */
1651 sctx->flags |= sctx->screen->barrier_flags.L2_to_cp |
1652 SI_CONTEXT_FLUSH_FOR_RENDER_COND;
1653
1654 sctx->render_cond_force_off = old_force_off;
1655 }
1656 }
1657
1658 sctx->render_cond = query;
1659 sctx->render_cond_invert = condition;
1660 sctx->render_cond_mode = mode;
1661
1662 si_set_atom_dirty(sctx, atom, query != NULL);
1663 }
1664
1665 void si_suspend_queries(struct si_context *sctx)
1666 {
1667 struct si_query *query;
1668
1669 LIST_FOR_EACH_ENTRY(query, &sctx->active_queries, active_list)
1670 query->ops->suspend(sctx, query);
1671 }
1672
1673 void si_resume_queries(struct si_context *sctx)
1674 {
1675 struct si_query *query;
1676
1677 /* Check CS space here. Resuming must not be interrupted by flushes. */
1678 si_need_gfx_cs_space(sctx);
1679
1680 LIST_FOR_EACH_ENTRY(query, &sctx->active_queries, active_list)
1681 query->ops->resume(sctx, query);
1682 }
1683
1684 #define XFULL(name_, query_type_, type_, result_type_, group_id_) \
1685 { \
1686 .name = name_, \
1687 .query_type = SI_QUERY_##query_type_, \
1688 .type = PIPE_DRIVER_QUERY_TYPE_##type_, \
1689 .result_type = PIPE_DRIVER_QUERY_RESULT_TYPE_##result_type_, \
1690 .group_id = group_id_ \
1691 }
1692
1693 #define X(name_, query_type_, type_, result_type_) \
1694 XFULL(name_, query_type_, type_, result_type_, ~(unsigned)0)
1695
1696 #define XG(group_, name_, query_type_, type_, result_type_) \
1697 XFULL(name_, query_type_, type_, result_type_, SI_QUERY_GROUP_##group_)
1698
1699 static struct pipe_driver_query_info si_driver_query_list[] = {
1700 X("num-compilations", NUM_COMPILATIONS, UINT64, CUMULATIVE),
1701 X("num-shaders-created", NUM_SHADERS_CREATED, UINT64, CUMULATIVE),
1702 X("num-shader-cache-hits", NUM_SHADER_CACHE_HITS, UINT64, CUMULATIVE),
1703 X("draw-calls", DRAW_CALLS, UINT64, AVERAGE),
1704 X("decompress-calls", DECOMPRESS_CALLS, UINT64, AVERAGE),
1705 X("MRT-draw-calls", MRT_DRAW_CALLS, UINT64, AVERAGE),
1706 X("prim-restart-calls", PRIM_RESTART_CALLS, UINT64, AVERAGE),
1707 X("spill-draw-calls", SPILL_DRAW_CALLS, UINT64, AVERAGE),
1708 X("compute-calls", COMPUTE_CALLS, UINT64, AVERAGE),
1709 X("spill-compute-calls", SPILL_COMPUTE_CALLS, UINT64, AVERAGE),
1710 X("dma-calls", DMA_CALLS, UINT64, AVERAGE),
1711 X("cp-dma-calls", CP_DMA_CALLS, UINT64, AVERAGE),
1712 X("num-vs-flushes", NUM_VS_FLUSHES, UINT64, AVERAGE),
1713 X("num-ps-flushes", NUM_PS_FLUSHES, UINT64, AVERAGE),
1714 X("num-cs-flushes", NUM_CS_FLUSHES, UINT64, AVERAGE),
1715 X("num-CB-cache-flushes", NUM_CB_CACHE_FLUSHES, UINT64, AVERAGE),
1716 X("num-DB-cache-flushes", NUM_DB_CACHE_FLUSHES, UINT64, AVERAGE),
1717 X("num-L2-invalidates", NUM_L2_INVALIDATES, UINT64, AVERAGE),
1718 X("num-L2-writebacks", NUM_L2_WRITEBACKS, UINT64, AVERAGE),
1719 X("num-resident-handles", NUM_RESIDENT_HANDLES, UINT64, AVERAGE),
1720 X("tc-offloaded-slots", TC_OFFLOADED_SLOTS, UINT64, AVERAGE),
1721 X("tc-direct-slots", TC_DIRECT_SLOTS, UINT64, AVERAGE),
1722 X("tc-num-syncs", TC_NUM_SYNCS, UINT64, AVERAGE),
1723 X("CS-thread-busy", CS_THREAD_BUSY, UINT64, AVERAGE),
1724 X("gallium-thread-busy", GALLIUM_THREAD_BUSY, UINT64, AVERAGE),
1725 X("requested-VRAM", REQUESTED_VRAM, BYTES, AVERAGE),
1726 X("requested-GTT", REQUESTED_GTT, BYTES, AVERAGE),
1727 X("mapped-VRAM", MAPPED_VRAM, BYTES, AVERAGE),
1728 X("mapped-GTT", MAPPED_GTT, BYTES, AVERAGE),
1729 X("buffer-wait-time", BUFFER_WAIT_TIME, MICROSECONDS, CUMULATIVE),
1730 X("num-mapped-buffers", NUM_MAPPED_BUFFERS, UINT64, AVERAGE),
1731 X("num-GFX-IBs", NUM_GFX_IBS, UINT64, AVERAGE),
1732 X("num-SDMA-IBs", NUM_SDMA_IBS, UINT64, AVERAGE),
1733 X("GFX-BO-list-size", GFX_BO_LIST_SIZE, UINT64, AVERAGE),
1734 X("GFX-IB-size", GFX_IB_SIZE, UINT64, AVERAGE),
1735 X("num-bytes-moved", NUM_BYTES_MOVED, BYTES, CUMULATIVE),
1736 X("num-evictions", NUM_EVICTIONS, UINT64, CUMULATIVE),
1737 X("VRAM-CPU-page-faults", NUM_VRAM_CPU_PAGE_FAULTS, UINT64, CUMULATIVE),
1738 X("VRAM-usage", VRAM_USAGE, BYTES, AVERAGE),
1739 X("VRAM-vis-usage", VRAM_VIS_USAGE, BYTES, AVERAGE),
1740 X("GTT-usage", GTT_USAGE, BYTES, AVERAGE),
1741 X("back-buffer-ps-draw-ratio", BACK_BUFFER_PS_DRAW_RATIO, UINT64, AVERAGE),
1742
1743 /* GPIN queries are for the benefit of old versions of GPUPerfStudio,
1744 * which use it as a fallback path to detect the GPU type.
1745 *
1746 * Note: The names of these queries are significant for GPUPerfStudio
1747 * (and possibly their order as well). */
1748 XG(GPIN, "GPIN_000", GPIN_ASIC_ID, UINT, AVERAGE),
1749 XG(GPIN, "GPIN_001", GPIN_NUM_SIMD, UINT, AVERAGE),
1750 XG(GPIN, "GPIN_002", GPIN_NUM_RB, UINT, AVERAGE),
1751 XG(GPIN, "GPIN_003", GPIN_NUM_SPI, UINT, AVERAGE),
1752 XG(GPIN, "GPIN_004", GPIN_NUM_SE, UINT, AVERAGE),
1753
1754 X("temperature", GPU_TEMPERATURE, UINT64, AVERAGE),
1755 X("shader-clock", CURRENT_GPU_SCLK, HZ, AVERAGE),
1756 X("memory-clock", CURRENT_GPU_MCLK, HZ, AVERAGE),
1757
1758 /* The following queries must be at the end of the list because their
1759 * availability is adjusted dynamically based on the DRM version. */
1760 X("GPU-load", GPU_LOAD, UINT64, AVERAGE),
1761 X("GPU-shaders-busy", GPU_SHADERS_BUSY, UINT64, AVERAGE),
1762 X("GPU-ta-busy", GPU_TA_BUSY, UINT64, AVERAGE),
1763 X("GPU-gds-busy", GPU_GDS_BUSY, UINT64, AVERAGE),
1764 X("GPU-vgt-busy", GPU_VGT_BUSY, UINT64, AVERAGE),
1765 X("GPU-ia-busy", GPU_IA_BUSY, UINT64, AVERAGE),
1766 X("GPU-sx-busy", GPU_SX_BUSY, UINT64, AVERAGE),
1767 X("GPU-wd-busy", GPU_WD_BUSY, UINT64, AVERAGE),
1768 X("GPU-bci-busy", GPU_BCI_BUSY, UINT64, AVERAGE),
1769 X("GPU-sc-busy", GPU_SC_BUSY, UINT64, AVERAGE),
1770 X("GPU-pa-busy", GPU_PA_BUSY, UINT64, AVERAGE),
1771 X("GPU-db-busy", GPU_DB_BUSY, UINT64, AVERAGE),
1772 X("GPU-cp-busy", GPU_CP_BUSY, UINT64, AVERAGE),
1773 X("GPU-cb-busy", GPU_CB_BUSY, UINT64, AVERAGE),
1774
1775 /* SRBM_STATUS2 */
1776 X("GPU-sdma-busy", GPU_SDMA_BUSY, UINT64, AVERAGE),
1777
1778 /* CP_STAT */
1779 X("GPU-pfp-busy", GPU_PFP_BUSY, UINT64, AVERAGE),
1780 X("GPU-meq-busy", GPU_MEQ_BUSY, UINT64, AVERAGE),
1781 X("GPU-me-busy", GPU_ME_BUSY, UINT64, AVERAGE),
1782 X("GPU-surf-sync-busy", GPU_SURF_SYNC_BUSY, UINT64, AVERAGE),
1783 X("GPU-cp-dma-busy", GPU_CP_DMA_BUSY, UINT64, AVERAGE),
1784 X("GPU-scratch-ram-busy", GPU_SCRATCH_RAM_BUSY, UINT64, AVERAGE),
1785 };
1786
1787 #undef X
1788 #undef XG
1789 #undef XFULL
1790
1791 static unsigned si_get_num_queries(struct si_screen *sscreen)
1792 {
1793 /* amdgpu */
1794 if (sscreen->info.drm_major == 3) {
1795 if (sscreen->info.chip_class >= GFX8)
1796 return ARRAY_SIZE(si_driver_query_list);
1797 else
1798 return ARRAY_SIZE(si_driver_query_list) - 7;
1799 }
1800
1801 /* radeon */
1802 if (sscreen->info.has_read_registers_query) {
1803 if (sscreen->info.chip_class == GFX7)
1804 return ARRAY_SIZE(si_driver_query_list) - 6;
1805 else
1806 return ARRAY_SIZE(si_driver_query_list) - 7;
1807 }
1808
1809 return ARRAY_SIZE(si_driver_query_list) - 21;
1810 }
1811
1812 static int si_get_driver_query_info(struct pipe_screen *screen,
1813 unsigned index,
1814 struct pipe_driver_query_info *info)
1815 {
1816 struct si_screen *sscreen = (struct si_screen*)screen;
1817 unsigned num_queries = si_get_num_queries(sscreen);
1818
1819 if (!info) {
1820 unsigned num_perfcounters =
1821 si_get_perfcounter_info(sscreen, 0, NULL);
1822
1823 return num_queries + num_perfcounters;
1824 }
1825
1826 if (index >= num_queries)
1827 return si_get_perfcounter_info(sscreen, index - num_queries, info);
1828
1829 *info = si_driver_query_list[index];
1830
1831 switch (info->query_type) {
1832 case SI_QUERY_REQUESTED_VRAM:
1833 case SI_QUERY_VRAM_USAGE:
1834 case SI_QUERY_MAPPED_VRAM:
1835 info->max_value.u64 = sscreen->info.vram_size;
1836 break;
1837 case SI_QUERY_REQUESTED_GTT:
1838 case SI_QUERY_GTT_USAGE:
1839 case SI_QUERY_MAPPED_GTT:
1840 info->max_value.u64 = sscreen->info.gart_size;
1841 break;
1842 case SI_QUERY_GPU_TEMPERATURE:
1843 info->max_value.u64 = 125;
1844 break;
1845 case SI_QUERY_VRAM_VIS_USAGE:
1846 info->max_value.u64 = sscreen->info.vram_vis_size;
1847 break;
1848 }
1849
1850 if (info->group_id != ~(unsigned)0 && sscreen->perfcounters)
1851 info->group_id += sscreen->perfcounters->num_groups;
1852
1853 return 1;
1854 }
1855
1856 /* Note: Unfortunately, GPUPerfStudio hardcodes the order of hardware
1857 * performance counter groups, so be careful when changing this and related
1858 * functions.
1859 */
1860 static int si_get_driver_query_group_info(struct pipe_screen *screen,
1861 unsigned index,
1862 struct pipe_driver_query_group_info *info)
1863 {
1864 struct si_screen *sscreen = (struct si_screen *)screen;
1865 unsigned num_pc_groups = 0;
1866
1867 if (sscreen->perfcounters)
1868 num_pc_groups = sscreen->perfcounters->num_groups;
1869
1870 if (!info)
1871 return num_pc_groups + SI_NUM_SW_QUERY_GROUPS;
1872
1873 if (index < num_pc_groups)
1874 return si_get_perfcounter_group_info(sscreen, index, info);
1875
1876 index -= num_pc_groups;
1877 if (index >= SI_NUM_SW_QUERY_GROUPS)
1878 return 0;
1879
1880 info->name = "GPIN";
1881 info->max_active_queries = 5;
1882 info->num_queries = 5;
1883 return 1;
1884 }
1885
1886 void si_init_query_functions(struct si_context *sctx)
1887 {
1888 sctx->b.create_query = si_create_query;
1889 sctx->b.create_batch_query = si_create_batch_query;
1890 sctx->b.destroy_query = si_destroy_query;
1891 sctx->b.begin_query = si_begin_query;
1892 sctx->b.end_query = si_end_query;
1893 sctx->b.get_query_result = si_get_query_result;
1894 sctx->b.get_query_result_resource = si_get_query_result_resource;
1895 sctx->atoms.s.render_cond.emit = si_emit_query_predication;
1896
1897 if (((struct si_screen*)sctx->b.screen)->info.num_render_backends > 0)
1898 sctx->b.render_condition = si_render_condition;
1899
1900 LIST_INITHEAD(&sctx->active_queries);
1901 }
1902
1903 void si_init_screen_query_functions(struct si_screen *sscreen)
1904 {
1905 sscreen->b.get_driver_query_info = si_get_driver_query_info;
1906 sscreen->b.get_driver_query_group_info = si_get_driver_query_group_info;
1907 }