1 /*
2 * Copyright 2011 Christoph Bumiller
3 * Copyright 2015 Samuel Pitoiset
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the 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
19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21 * OTHER DEALINGS IN THE SOFTWARE.
22 */
23
24 #define NV50_PUSH_EXPLICIT_SPACE_CHECKING
25
26 #include "nv50/nv50_context.h"
27 #include "nv50/nv50_query_hw.h"
28 #include "nv50/nv50_query_hw_metric.h"
29 #include "nv50/nv50_query_hw_sm.h"
30 #include "nv_object.xml.h"
31
32 /* XXX: Nested queries, and simultaneous queries on multiple gallium contexts
33 * (since we use only a single GPU channel per screen) will not work properly.
34 *
35 * The first is not that big of an issue because OpenGL does not allow nested
36 * queries anyway.
37 */
38
39 #define NV50_HW_QUERY_ALLOC_SPACE 256
40
41 bool
nv50_hw_query_allocate(struct nv50_context * nv50,struct nv50_query * q,int size)42 nv50_hw_query_allocate(struct nv50_context *nv50, struct nv50_query *q,
43 int size)
44 {
45 struct nv50_screen *screen = nv50->screen;
46 struct nv50_hw_query *hq = nv50_hw_query(q);
47 int ret;
48
49 if (hq->bo) {
50 nouveau_bo_ref(NULL, &hq->bo);
51 if (hq->mm) {
52 if (hq->state == NV50_HW_QUERY_STATE_READY)
53 nouveau_mm_free(hq->mm);
54 else
55 nouveau_fence_work(screen->base.fence.current,
56 nouveau_mm_free_work, hq->mm);
57 }
58 }
59 if (size) {
60 hq->mm = nouveau_mm_allocate(screen->base.mm_GART, size,
61 &hq->bo, &hq->base_offset);
62 if (!hq->bo)
63 return false;
64 hq->offset = hq->base_offset;
65
66 ret = nouveau_bo_map(hq->bo, 0, screen->base.client);
67 if (ret) {
68 nv50_hw_query_allocate(nv50, q, 0);
69 return false;
70 }
71 hq->data = (uint32_t *)((uint8_t *)hq->bo->map + hq->base_offset);
72 }
73 return true;
74 }
75
76 static void
nv50_hw_query_get(struct nouveau_pushbuf * push,struct nv50_query * q,unsigned offset,uint32_t get)77 nv50_hw_query_get(struct nouveau_pushbuf *push, struct nv50_query *q,
78 unsigned offset, uint32_t get)
79 {
80 struct nv50_hw_query *hq = nv50_hw_query(q);
81
82 offset += hq->offset;
83
84 PUSH_SPACE(push, 5);
85 PUSH_REFN (push, hq->bo, NOUVEAU_BO_GART | NOUVEAU_BO_WR);
86 BEGIN_NV04(push, NV50_3D(QUERY_ADDRESS_HIGH), 4);
87 PUSH_DATAh(push, hq->bo->offset + offset);
88 PUSH_DATA (push, hq->bo->offset + offset);
89 PUSH_DATA (push, hq->sequence);
90 PUSH_DATA (push, get);
91 }
92
93 static inline void
nv50_hw_query_update(struct nv50_query * q)94 nv50_hw_query_update(struct nv50_query *q)
95 {
96 struct nv50_hw_query *hq = nv50_hw_query(q);
97
98 if (hq->is64bit) {
99 if (nouveau_fence_signalled(hq->fence))
100 hq->state = NV50_HW_QUERY_STATE_READY;
101 } else {
102 if (hq->data[0] == hq->sequence)
103 hq->state = NV50_HW_QUERY_STATE_READY;
104 }
105 }
106
107 static void
nv50_hw_destroy_query(struct nv50_context * nv50,struct nv50_query * q)108 nv50_hw_destroy_query(struct nv50_context *nv50, struct nv50_query *q)
109 {
110 struct nv50_hw_query *hq = nv50_hw_query(q);
111
112 if (hq->funcs && hq->funcs->destroy_query) {
113 hq->funcs->destroy_query(nv50, hq);
114 return;
115 }
116
117 nv50_hw_query_allocate(nv50, q, 0);
118 nouveau_fence_ref(NULL, &hq->fence);
119 FREE(hq);
120 }
121
122 static bool
nv50_hw_begin_query(struct nv50_context * nv50,struct nv50_query * q)123 nv50_hw_begin_query(struct nv50_context *nv50, struct nv50_query *q)
124 {
125 struct nouveau_pushbuf *push = nv50->base.pushbuf;
126 struct nv50_hw_query *hq = nv50_hw_query(q);
127
128 if (hq->funcs && hq->funcs->begin_query)
129 return hq->funcs->begin_query(nv50, hq);
130
131 /* For occlusion queries we have to change the storage, because a previous
132 * query might set the initial render condition to false even *after* we re-
133 * initialized it to true.
134 */
135 if (hq->rotate) {
136 hq->offset += hq->rotate;
137 hq->data += hq->rotate / sizeof(*hq->data);
138 if (hq->offset - hq->base_offset == NV50_HW_QUERY_ALLOC_SPACE)
139 nv50_hw_query_allocate(nv50, q, NV50_HW_QUERY_ALLOC_SPACE);
140
141 /* XXX: can we do this with the GPU, and sync with respect to a previous
142 * query ?
143 */
144 hq->data[0] = hq->sequence; /* initialize sequence */
145 hq->data[1] = 1; /* initial render condition = true */
146 hq->data[4] = hq->sequence + 1; /* for comparison COND_MODE */
147 hq->data[5] = 0;
148 }
149 hq->sequence++;
150
151 switch (q->type) {
152 case PIPE_QUERY_OCCLUSION_COUNTER:
153 case PIPE_QUERY_OCCLUSION_PREDICATE:
154 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
155 if (nv50->screen->num_occlusion_queries_active++) {
156 nv50_hw_query_get(push, q, 0x10, 0x0100f002);
157 } else {
158 PUSH_SPACE(push, 4);
159 BEGIN_NV04(push, NV50_3D(COUNTER_RESET), 1);
160 PUSH_DATA (push, NV50_3D_COUNTER_RESET_SAMPLECNT);
161 BEGIN_NV04(push, NV50_3D(SAMPLECNT_ENABLE), 1);
162 PUSH_DATA (push, 1);
163 }
164 break;
165 case PIPE_QUERY_PRIMITIVES_GENERATED:
166 nv50_hw_query_get(push, q, 0x20, 0x06805002);
167 break;
168 case PIPE_QUERY_PRIMITIVES_EMITTED:
169 nv50_hw_query_get(push, q, 0x20, 0x05805002);
170 break;
171 case PIPE_QUERY_SO_STATISTICS:
172 nv50_hw_query_get(push, q, 0x30, 0x05805002);
173 nv50_hw_query_get(push, q, 0x40, 0x06805002);
174 break;
175 case PIPE_QUERY_PIPELINE_STATISTICS:
176 nv50_hw_query_get(push, q, 0x90, 0x00801002); /* VFETCH, VERTICES */
177 nv50_hw_query_get(push, q, 0xa0, 0x01801002); /* VFETCH, PRIMS */
178 nv50_hw_query_get(push, q, 0xb0, 0x02802002); /* VP, LAUNCHES */
179 nv50_hw_query_get(push, q, 0xc0, 0x03806002); /* GP, LAUNCHES */
180 nv50_hw_query_get(push, q, 0xd0, 0x04806002); /* GP, PRIMS_OUT */
181 nv50_hw_query_get(push, q, 0xe0, 0x07804002); /* RAST, PRIMS_IN */
182 nv50_hw_query_get(push, q, 0xf0, 0x08804002); /* RAST, PRIMS_OUT */
183 nv50_hw_query_get(push, q, 0x100, 0x0980a002); /* ROP, PIXELS */
184 ((uint64_t *)hq->data)[2 * 0x11] = nv50->compute_invocations;
185 break;
186 case PIPE_QUERY_TIME_ELAPSED:
187 nv50_hw_query_get(push, q, 0x10, 0x00005002);
188 break;
189 default:
190 assert(0);
191 return false;
192 }
193 hq->state = NV50_HW_QUERY_STATE_ACTIVE;
194 return true;
195 }
196
197 static void
nv50_hw_end_query(struct nv50_context * nv50,struct nv50_query * q)198 nv50_hw_end_query(struct nv50_context *nv50, struct nv50_query *q)
199 {
200 struct nouveau_pushbuf *push = nv50->base.pushbuf;
201 struct nv50_hw_query *hq = nv50_hw_query(q);
202
203 if (hq->funcs && hq->funcs->end_query) {
204 hq->funcs->end_query(nv50, hq);
205 return;
206 }
207
208 hq->state = NV50_HW_QUERY_STATE_ENDED;
209
210 switch (q->type) {
211 case PIPE_QUERY_OCCLUSION_COUNTER:
212 case PIPE_QUERY_OCCLUSION_PREDICATE:
213 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
214 nv50_hw_query_get(push, q, 0, 0x0100f002);
215 if (--nv50->screen->num_occlusion_queries_active == 0) {
216 PUSH_SPACE(push, 2);
217 BEGIN_NV04(push, NV50_3D(SAMPLECNT_ENABLE), 1);
218 PUSH_DATA (push, 0);
219 }
220 break;
221 case PIPE_QUERY_PRIMITIVES_GENERATED:
222 nv50_hw_query_get(push, q, 0x10, 0x06805002);
223 nv50_hw_query_get(push, q, 0x00, 0x00005010);
224 break;
225 case PIPE_QUERY_PRIMITIVES_EMITTED:
226 nv50_hw_query_get(push, q, 0x10, 0x05805002);
227 nv50_hw_query_get(push, q, 0x00, 0x00005010);
228 break;
229 case PIPE_QUERY_SO_STATISTICS:
230 nv50_hw_query_get(push, q, 0x10, 0x05805002);
231 nv50_hw_query_get(push, q, 0x20, 0x06805002);
232 nv50_hw_query_get(push, q, 0x00, 0x00005010);
233 break;
234 case PIPE_QUERY_PIPELINE_STATISTICS:
235 nv50_hw_query_get(push, q, 0x00, 0x00801002); /* VFETCH, VERTICES */
236 nv50_hw_query_get(push, q, 0x10, 0x01801002); /* VFETCH, PRIMS */
237 nv50_hw_query_get(push, q, 0x20, 0x02802002); /* VP, LAUNCHES */
238 nv50_hw_query_get(push, q, 0x30, 0x03806002); /* GP, LAUNCHES */
239 nv50_hw_query_get(push, q, 0x40, 0x04806002); /* GP, PRIMS_OUT */
240 nv50_hw_query_get(push, q, 0x50, 0x07804002); /* RAST, PRIMS_IN */
241 nv50_hw_query_get(push, q, 0x60, 0x08804002); /* RAST, PRIMS_OUT */
242 nv50_hw_query_get(push, q, 0x70, 0x0980a002); /* ROP, PIXELS */
243 ((uint64_t *)hq->data)[2 * 0x8] = nv50->compute_invocations;
244 break;
245 case PIPE_QUERY_TIMESTAMP:
246 hq->sequence++;
247 FALLTHROUGH;
248 case PIPE_QUERY_TIME_ELAPSED:
249 nv50_hw_query_get(push, q, 0, 0x00005002);
250 break;
251 case PIPE_QUERY_GPU_FINISHED:
252 hq->sequence++;
253 nv50_hw_query_get(push, q, 0, 0x1000f010);
254 break;
255 case NVA0_HW_QUERY_STREAM_OUTPUT_BUFFER_OFFSET:
256 hq->sequence++;
257 nv50_hw_query_get(push, q, 0, 0x0d005002 | (q->index << 5));
258 break;
259 case PIPE_QUERY_TIMESTAMP_DISJOINT:
260 /* This query is not issued on GPU because disjoint is forced to false */
261 hq->state = NV50_HW_QUERY_STATE_READY;
262 break;
263 default:
264 assert(0);
265 break;
266 }
267 if (hq->is64bit)
268 nouveau_fence_ref(nv50->screen->base.fence.current, &hq->fence);
269 }
270
271 static bool
nv50_hw_get_query_result(struct nv50_context * nv50,struct nv50_query * q,bool wait,union pipe_query_result * result)272 nv50_hw_get_query_result(struct nv50_context *nv50, struct nv50_query *q,
273 bool wait, union pipe_query_result *result)
274 {
275 struct nv50_hw_query *hq = nv50_hw_query(q);
276 uint64_t *res64 = (uint64_t *)result;
277 uint32_t *res32 = (uint32_t *)result;
278 uint8_t *res8 = (uint8_t *)result;
279 uint64_t *data64 = (uint64_t *)hq->data;
280 int i;
281
282 if (hq->funcs && hq->funcs->get_query_result)
283 return hq->funcs->get_query_result(nv50, hq, wait, result);
284
285 if (hq->state != NV50_HW_QUERY_STATE_READY)
286 nv50_hw_query_update(q);
287
288 if (hq->state != NV50_HW_QUERY_STATE_READY) {
289 if (!wait) {
290 /* for broken apps that spin on GL_QUERY_RESULT_AVAILABLE */
291 if (hq->state != NV50_HW_QUERY_STATE_FLUSHED) {
292 hq->state = NV50_HW_QUERY_STATE_FLUSHED;
293 PUSH_KICK(nv50->base.pushbuf);
294 }
295 return false;
296 }
297 if (nouveau_bo_wait(hq->bo, NOUVEAU_BO_RD, nv50->screen->base.client))
298 return false;
299 }
300 hq->state = NV50_HW_QUERY_STATE_READY;
301
302 switch (q->type) {
303 case PIPE_QUERY_GPU_FINISHED:
304 res8[0] = true;
305 break;
306 case PIPE_QUERY_OCCLUSION_COUNTER: /* u32 sequence, u32 count, u64 time */
307 res64[0] = hq->data[1] - hq->data[5];
308 break;
309 case PIPE_QUERY_OCCLUSION_PREDICATE:
310 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
311 res8[0] = hq->data[1] != hq->data[5];
312 break;
313 case PIPE_QUERY_PRIMITIVES_GENERATED: /* u64 count, u64 time */
314 case PIPE_QUERY_PRIMITIVES_EMITTED: /* u64 count, u64 time */
315 res64[0] = data64[2] - data64[4];
316 break;
317 case PIPE_QUERY_SO_STATISTICS:
318 res64[0] = data64[2] - data64[6];
319 res64[1] = data64[4] - data64[8];
320 break;
321 case PIPE_QUERY_PIPELINE_STATISTICS:
322 for (i = 0; i < 8; ++i)
323 res64[i] = data64[i * 2] - data64[18 + i * 2];
324 result->pipeline_statistics.cs_invocations = data64[i * 2] - data64[18 + i * 2];
325 break;
326 case PIPE_QUERY_TIMESTAMP:
327 res64[0] = data64[1];
328 break;
329 case PIPE_QUERY_TIMESTAMP_DISJOINT:
330 res64[0] = 1000000000;
331 res8[8] = false;
332 break;
333 case PIPE_QUERY_TIME_ELAPSED:
334 res64[0] = data64[1] - data64[3];
335 break;
336 case NVA0_HW_QUERY_STREAM_OUTPUT_BUFFER_OFFSET:
337 res32[0] = hq->data[1];
338 break;
339 default:
340 assert(0);
341 return false;
342 }
343
344 return true;
345 }
346
347 static const struct nv50_query_funcs hw_query_funcs = {
348 .destroy_query = nv50_hw_destroy_query,
349 .begin_query = nv50_hw_begin_query,
350 .end_query = nv50_hw_end_query,
351 .get_query_result = nv50_hw_get_query_result,
352 };
353
354 struct nv50_query *
nv50_hw_create_query(struct nv50_context * nv50,unsigned type,unsigned index)355 nv50_hw_create_query(struct nv50_context *nv50, unsigned type, unsigned index)
356 {
357 struct nv50_hw_query *hq;
358 struct nv50_query *q;
359 unsigned space = NV50_HW_QUERY_ALLOC_SPACE;
360
361 hq = nv50_hw_sm_create_query(nv50, type);
362 if (hq) {
363 hq->base.funcs = &hw_query_funcs;
364 return (struct nv50_query *)hq;
365 }
366
367 hq = nv50_hw_metric_create_query(nv50, type);
368 if (hq) {
369 hq->base.funcs = &hw_query_funcs;
370 return (struct nv50_query *)hq;
371 }
372
373 hq = CALLOC_STRUCT(nv50_hw_query);
374 if (!hq)
375 return NULL;
376
377 q = &hq->base;
378 q->funcs = &hw_query_funcs;
379 q->type = type;
380
381 switch (q->type) {
382 case PIPE_QUERY_OCCLUSION_COUNTER:
383 case PIPE_QUERY_OCCLUSION_PREDICATE:
384 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
385 hq->rotate = 32;
386 break;
387 case PIPE_QUERY_PRIMITIVES_GENERATED:
388 case PIPE_QUERY_PRIMITIVES_EMITTED:
389 space = 32 + 16; /* separate fence value written here */
390 break;
391 case PIPE_QUERY_SO_STATISTICS:
392 space = 64 + 16; /* separate fence value written here */
393 break;
394 case PIPE_QUERY_PIPELINE_STATISTICS:
395 hq->is64bit = true;
396 space = 9 * 2 * 16; /* 9 values, start/end, 16-bytes each */
397 break;
398 case PIPE_QUERY_TIME_ELAPSED:
399 case PIPE_QUERY_TIMESTAMP:
400 case PIPE_QUERY_TIMESTAMP_DISJOINT:
401 case PIPE_QUERY_GPU_FINISHED:
402 space = 32;
403 break;
404 case NVA0_HW_QUERY_STREAM_OUTPUT_BUFFER_OFFSET:
405 space = 16;
406 break;
407 default:
408 debug_printf("invalid query type: %u\n", type);
409 FREE(q);
410 return NULL;
411 }
412
413 if (!nv50_hw_query_allocate(nv50, q, space)) {
414 FREE(hq);
415 return NULL;
416 }
417
418 if (hq->rotate) {
419 /* we advance before query_begin ! */
420 hq->offset -= hq->rotate;
421 hq->data -= hq->rotate / sizeof(*hq->data);
422 } else
423 if (!hq->is64bit)
424 hq->data[0] = 0; /* initialize sequence */
425
426 return q;
427 }
428
429 int
nv50_hw_get_driver_query_info(struct nv50_screen * screen,unsigned id,struct pipe_driver_query_info * info)430 nv50_hw_get_driver_query_info(struct nv50_screen *screen, unsigned id,
431 struct pipe_driver_query_info *info)
432 {
433 int num_hw_sm_queries = 0, num_hw_metric_queries = 0;
434
435 num_hw_sm_queries = nv50_hw_sm_get_driver_query_info(screen, 0, NULL);
436 num_hw_metric_queries =
437 nv50_hw_metric_get_driver_query_info(screen, 0, NULL);
438
439 if (!info)
440 return num_hw_sm_queries + num_hw_metric_queries;
441
442 if (id < num_hw_sm_queries)
443 return nv50_hw_sm_get_driver_query_info(screen, id, info);
444
445 return nv50_hw_metric_get_driver_query_info(screen,
446 id - num_hw_sm_queries, info);
447 }
448
449 void
nv50_hw_query_pushbuf_submit(struct nouveau_pushbuf * push,uint16_t method,struct nv50_query * q,unsigned result_offset)450 nv50_hw_query_pushbuf_submit(struct nouveau_pushbuf *push, uint16_t method,
451 struct nv50_query *q, unsigned result_offset)
452 {
453 struct nv50_hw_query *hq = nv50_hw_query(q);
454
455 nv50_hw_query_update(q);
456 if (hq->state != NV50_HW_QUERY_STATE_READY)
457 nouveau_bo_wait(hq->bo, NOUVEAU_BO_RD, push->client);
458 hq->state = NV50_HW_QUERY_STATE_READY;
459
460 BEGIN_NV04(push, SUBC_3D(method), 1);
461 PUSH_DATA (push, hq->data[result_offset / 4]);
462 }
463
464 void
nv84_hw_query_fifo_wait(struct nouveau_pushbuf * push,struct nv50_query * q)465 nv84_hw_query_fifo_wait(struct nouveau_pushbuf *push, struct nv50_query *q)
466 {
467 struct nv50_hw_query *hq = nv50_hw_query(q);
468 unsigned offset = hq->offset;
469
470 assert(!hq->is64bit);
471
472 PUSH_SPACE(push, 5);
473 PUSH_REFN (push, hq->bo, NOUVEAU_BO_GART | NOUVEAU_BO_RD);
474 BEGIN_NV04(push, SUBC_3D(NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH), 4);
475 PUSH_DATAh(push, hq->bo->offset + offset);
476 PUSH_DATA (push, hq->bo->offset + offset);
477 PUSH_DATA (push, hq->sequence);
478 PUSH_DATA (push, NV84_SUBCHAN_SEMAPHORE_TRIGGER_ACQUIRE_EQUAL);
479 }
480