2 * Copyright 2011 Christoph Bumiller
3 * Copyright 2015 Samuel Pitoiset
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:
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
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.
24 #define NV50_PUSH_EXPLICIT_SPACE_CHECKING
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"
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.
35 * The first is not that big of an issue because OpenGL does not allow nested
39 #define NV50_HW_QUERY_ALLOC_SPACE 256
42 nv50_hw_query_allocate(struct nv50_context
*nv50
, struct nv50_query
*q
,
45 struct nv50_screen
*screen
= nv50
->screen
;
46 struct nv50_hw_query
*hq
= nv50_hw_query(q
);
50 nouveau_bo_ref(NULL
, &hq
->bo
);
52 if (hq
->state
== NV50_HW_QUERY_STATE_READY
)
53 nouveau_mm_free(hq
->mm
);
55 nouveau_fence_work(screen
->base
.fence
.current
,
56 nouveau_mm_free_work
, hq
->mm
);
60 hq
->mm
= nouveau_mm_allocate(screen
->base
.mm_GART
, size
,
61 &hq
->bo
, &hq
->base_offset
);
64 hq
->offset
= hq
->base_offset
;
66 ret
= nouveau_bo_map(hq
->bo
, 0, screen
->base
.client
);
68 nv50_hw_query_allocate(nv50
, q
, 0);
71 hq
->data
= (uint32_t *)((uint8_t *)hq
->bo
->map
+ hq
->base_offset
);
77 nv50_hw_query_get(struct nouveau_pushbuf
*push
, struct nv50_query
*q
,
78 unsigned offset
, uint32_t get
)
80 struct nv50_hw_query
*hq
= nv50_hw_query(q
);
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
);
94 nv50_hw_query_update(struct nv50_query
*q
)
96 struct nv50_hw_query
*hq
= nv50_hw_query(q
);
99 if (nouveau_fence_signalled(hq
->fence
))
100 hq
->state
= NV50_HW_QUERY_STATE_READY
;
102 if (hq
->data
[0] == hq
->sequence
)
103 hq
->state
= NV50_HW_QUERY_STATE_READY
;
108 nv50_hw_destroy_query(struct nv50_context
*nv50
, struct nv50_query
*q
)
110 struct nv50_hw_query
*hq
= nv50_hw_query(q
);
112 if (hq
->funcs
&& hq
->funcs
->destroy_query
) {
113 hq
->funcs
->destroy_query(nv50
, hq
);
117 nv50_hw_query_allocate(nv50
, q
, 0);
118 nouveau_fence_ref(NULL
, &hq
->fence
);
123 nv50_hw_begin_query(struct nv50_context
*nv50
, struct nv50_query
*q
)
125 struct nouveau_pushbuf
*push
= nv50
->base
.pushbuf
;
126 struct nv50_hw_query
*hq
= nv50_hw_query(q
);
128 if (hq
->funcs
&& hq
->funcs
->begin_query
)
129 return hq
->funcs
->begin_query(nv50
, hq
);
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.
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
);
141 /* XXX: can we do this with the GPU, and sync with respect to a previous
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 */
150 hq
->data
[0] = hq
->sequence
++; /* the previously used one */
153 case PIPE_QUERY_OCCLUSION_COUNTER
:
154 case PIPE_QUERY_OCCLUSION_PREDICATE
:
155 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE
:
156 if (nv50
->screen
->num_occlusion_queries_active
++) {
157 nv50_hw_query_get(push
, q
, 0x10, 0x0100f002);
160 BEGIN_NV04(push
, NV50_3D(COUNTER_RESET
), 1);
161 PUSH_DATA (push
, NV50_3D_COUNTER_RESET_SAMPLECNT
);
162 BEGIN_NV04(push
, NV50_3D(SAMPLECNT_ENABLE
), 1);
166 case PIPE_QUERY_PRIMITIVES_GENERATED
:
167 nv50_hw_query_get(push
, q
, 0x10, 0x06805002);
169 case PIPE_QUERY_PRIMITIVES_EMITTED
:
170 nv50_hw_query_get(push
, q
, 0x10, 0x05805002);
172 case PIPE_QUERY_SO_STATISTICS
:
173 nv50_hw_query_get(push
, q
, 0x20, 0x05805002);
174 nv50_hw_query_get(push
, q
, 0x30, 0x06805002);
176 case PIPE_QUERY_PIPELINE_STATISTICS
:
177 nv50_hw_query_get(push
, q
, 0x80, 0x00801002); /* VFETCH, VERTICES */
178 nv50_hw_query_get(push
, q
, 0x90, 0x01801002); /* VFETCH, PRIMS */
179 nv50_hw_query_get(push
, q
, 0xa0, 0x02802002); /* VP, LAUNCHES */
180 nv50_hw_query_get(push
, q
, 0xb0, 0x03806002); /* GP, LAUNCHES */
181 nv50_hw_query_get(push
, q
, 0xc0, 0x04806002); /* GP, PRIMS_OUT */
182 nv50_hw_query_get(push
, q
, 0xd0, 0x07804002); /* RAST, PRIMS_IN */
183 nv50_hw_query_get(push
, q
, 0xe0, 0x08804002); /* RAST, PRIMS_OUT */
184 nv50_hw_query_get(push
, q
, 0xf0, 0x0980a002); /* ROP, PIXELS */
186 case PIPE_QUERY_TIME_ELAPSED
:
187 nv50_hw_query_get(push
, q
, 0x10, 0x00005002);
193 hq
->state
= NV50_HW_QUERY_STATE_ACTIVE
;
198 nv50_hw_end_query(struct nv50_context
*nv50
, struct nv50_query
*q
)
200 struct nouveau_pushbuf
*push
= nv50
->base
.pushbuf
;
201 struct nv50_hw_query
*hq
= nv50_hw_query(q
);
203 if (hq
->funcs
&& hq
->funcs
->end_query
) {
204 hq
->funcs
->end_query(nv50
, hq
);
208 hq
->state
= NV50_HW_QUERY_STATE_ENDED
;
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) {
217 BEGIN_NV04(push
, NV50_3D(SAMPLECNT_ENABLE
), 1);
221 case PIPE_QUERY_PRIMITIVES_GENERATED
:
222 nv50_hw_query_get(push
, q
, 0, 0x06805002);
224 case PIPE_QUERY_PRIMITIVES_EMITTED
:
225 nv50_hw_query_get(push
, q
, 0, 0x05805002);
227 case PIPE_QUERY_SO_STATISTICS
:
228 nv50_hw_query_get(push
, q
, 0x00, 0x05805002);
229 nv50_hw_query_get(push
, q
, 0x10, 0x06805002);
231 case PIPE_QUERY_PIPELINE_STATISTICS
:
232 nv50_hw_query_get(push
, q
, 0x00, 0x00801002); /* VFETCH, VERTICES */
233 nv50_hw_query_get(push
, q
, 0x10, 0x01801002); /* VFETCH, PRIMS */
234 nv50_hw_query_get(push
, q
, 0x20, 0x02802002); /* VP, LAUNCHES */
235 nv50_hw_query_get(push
, q
, 0x30, 0x03806002); /* GP, LAUNCHES */
236 nv50_hw_query_get(push
, q
, 0x40, 0x04806002); /* GP, PRIMS_OUT */
237 nv50_hw_query_get(push
, q
, 0x50, 0x07804002); /* RAST, PRIMS_IN */
238 nv50_hw_query_get(push
, q
, 0x60, 0x08804002); /* RAST, PRIMS_OUT */
239 nv50_hw_query_get(push
, q
, 0x70, 0x0980a002); /* ROP, PIXELS */
241 case PIPE_QUERY_TIMESTAMP
:
244 case PIPE_QUERY_TIME_ELAPSED
:
245 nv50_hw_query_get(push
, q
, 0, 0x00005002);
247 case PIPE_QUERY_GPU_FINISHED
:
249 nv50_hw_query_get(push
, q
, 0, 0x1000f010);
251 case NVA0_HW_QUERY_STREAM_OUTPUT_BUFFER_OFFSET
:
253 nv50_hw_query_get(push
, q
, 0, 0x0d005002 | (q
->index
<< 5));
255 case PIPE_QUERY_TIMESTAMP_DISJOINT
:
256 /* This query is not issued on GPU because disjoint is forced to false */
257 hq
->state
= NV50_HW_QUERY_STATE_READY
;
264 nouveau_fence_ref(nv50
->screen
->base
.fence
.current
, &hq
->fence
);
268 nv50_hw_get_query_result(struct nv50_context
*nv50
, struct nv50_query
*q
,
269 boolean wait
, union pipe_query_result
*result
)
271 struct nv50_hw_query
*hq
= nv50_hw_query(q
);
272 uint64_t *res64
= (uint64_t *)result
;
273 uint32_t *res32
= (uint32_t *)result
;
274 uint8_t *res8
= (uint8_t *)result
;
275 uint64_t *data64
= (uint64_t *)hq
->data
;
278 if (hq
->funcs
&& hq
->funcs
->get_query_result
)
279 return hq
->funcs
->get_query_result(nv50
, hq
, wait
, result
);
281 if (hq
->state
!= NV50_HW_QUERY_STATE_READY
)
282 nv50_hw_query_update(q
);
284 if (hq
->state
!= NV50_HW_QUERY_STATE_READY
) {
286 /* for broken apps that spin on GL_QUERY_RESULT_AVAILABLE */
287 if (hq
->state
!= NV50_HW_QUERY_STATE_FLUSHED
) {
288 hq
->state
= NV50_HW_QUERY_STATE_FLUSHED
;
289 PUSH_KICK(nv50
->base
.pushbuf
);
293 if (nouveau_bo_wait(hq
->bo
, NOUVEAU_BO_RD
, nv50
->screen
->base
.client
))
296 hq
->state
= NV50_HW_QUERY_STATE_READY
;
299 case PIPE_QUERY_GPU_FINISHED
:
302 case PIPE_QUERY_OCCLUSION_COUNTER
: /* u32 sequence, u32 count, u64 time */
303 res64
[0] = hq
->data
[1] - hq
->data
[5];
305 case PIPE_QUERY_OCCLUSION_PREDICATE
:
306 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE
:
307 res8
[0] = hq
->data
[1] != hq
->data
[5];
309 case PIPE_QUERY_PRIMITIVES_GENERATED
: /* u64 count, u64 time */
310 case PIPE_QUERY_PRIMITIVES_EMITTED
: /* u64 count, u64 time */
311 res64
[0] = data64
[0] - data64
[2];
313 case PIPE_QUERY_SO_STATISTICS
:
314 res64
[0] = data64
[0] - data64
[4];
315 res64
[1] = data64
[2] - data64
[6];
317 case PIPE_QUERY_PIPELINE_STATISTICS
:
318 for (i
= 0; i
< 8; ++i
)
319 res64
[i
] = data64
[i
* 2] - data64
[16 + i
* 2];
321 case PIPE_QUERY_TIMESTAMP
:
322 res64
[0] = data64
[1];
324 case PIPE_QUERY_TIMESTAMP_DISJOINT
:
325 res64
[0] = 1000000000;
328 case PIPE_QUERY_TIME_ELAPSED
:
329 res64
[0] = data64
[1] - data64
[3];
331 case NVA0_HW_QUERY_STREAM_OUTPUT_BUFFER_OFFSET
:
332 res32
[0] = hq
->data
[1];
342 static const struct nv50_query_funcs hw_query_funcs
= {
343 .destroy_query
= nv50_hw_destroy_query
,
344 .begin_query
= nv50_hw_begin_query
,
345 .end_query
= nv50_hw_end_query
,
346 .get_query_result
= nv50_hw_get_query_result
,
350 nv50_hw_create_query(struct nv50_context
*nv50
, unsigned type
, unsigned index
)
352 struct nv50_hw_query
*hq
;
353 struct nv50_query
*q
;
355 hq
= nv50_hw_sm_create_query(nv50
, type
);
357 hq
->base
.funcs
= &hw_query_funcs
;
358 return (struct nv50_query
*)hq
;
361 hq
= nv50_hw_metric_create_query(nv50
, type
);
363 hq
->base
.funcs
= &hw_query_funcs
;
364 return (struct nv50_query
*)hq
;
367 hq
= CALLOC_STRUCT(nv50_hw_query
);
372 q
->funcs
= &hw_query_funcs
;
376 case PIPE_QUERY_OCCLUSION_COUNTER
:
377 case PIPE_QUERY_OCCLUSION_PREDICATE
:
378 case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE
:
381 case PIPE_QUERY_PRIMITIVES_GENERATED
:
382 case PIPE_QUERY_PRIMITIVES_EMITTED
:
383 case PIPE_QUERY_SO_STATISTICS
:
384 case PIPE_QUERY_PIPELINE_STATISTICS
:
387 case PIPE_QUERY_TIME_ELAPSED
:
388 case PIPE_QUERY_TIMESTAMP
:
389 case PIPE_QUERY_TIMESTAMP_DISJOINT
:
390 case PIPE_QUERY_GPU_FINISHED
:
391 case NVA0_HW_QUERY_STREAM_OUTPUT_BUFFER_OFFSET
:
394 debug_printf("invalid query type: %u\n", type
);
399 if (!nv50_hw_query_allocate(nv50
, q
, NV50_HW_QUERY_ALLOC_SPACE
)) {
405 /* we advance before query_begin ! */
406 hq
->offset
-= hq
->rotate
;
407 hq
->data
-= hq
->rotate
/ sizeof(*hq
->data
);
414 nv50_hw_get_driver_query_info(struct nv50_screen
*screen
, unsigned id
,
415 struct pipe_driver_query_info
*info
)
417 int num_hw_sm_queries
= 0, num_hw_metric_queries
= 0;
419 num_hw_sm_queries
= nv50_hw_sm_get_driver_query_info(screen
, 0, NULL
);
420 num_hw_metric_queries
=
421 nv50_hw_metric_get_driver_query_info(screen
, 0, NULL
);
424 return num_hw_sm_queries
+ num_hw_metric_queries
;
426 if (id
< num_hw_sm_queries
)
427 return nv50_hw_sm_get_driver_query_info(screen
, id
, info
);
429 return nv50_hw_metric_get_driver_query_info(screen
,
430 id
- num_hw_sm_queries
, info
);
434 nv50_hw_query_pushbuf_submit(struct nouveau_pushbuf
*push
, uint16_t method
,
435 struct nv50_query
*q
, unsigned result_offset
)
437 struct nv50_hw_query
*hq
= nv50_hw_query(q
);
439 nv50_hw_query_update(q
);
440 if (hq
->state
!= NV50_HW_QUERY_STATE_READY
)
441 nouveau_bo_wait(hq
->bo
, NOUVEAU_BO_RD
, push
->client
);
442 hq
->state
= NV50_HW_QUERY_STATE_READY
;
444 BEGIN_NV04(push
, SUBC_3D(method
), 1);
445 PUSH_DATA (push
, hq
->data
[result_offset
/ 4]);
449 nv84_hw_query_fifo_wait(struct nouveau_pushbuf
*push
, struct nv50_query
*q
)
451 struct nv50_hw_query
*hq
= nv50_hw_query(q
);
452 unsigned offset
= hq
->offset
;
455 PUSH_REFN (push
, hq
->bo
, NOUVEAU_BO_GART
| NOUVEAU_BO_RD
);
456 BEGIN_NV04(push
, SUBC_3D(NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH
), 4);
457 PUSH_DATAh(push
, hq
->bo
->offset
+ offset
);
458 PUSH_DATA (push
, hq
->bo
->offset
+ offset
);
459 PUSH_DATA (push
, hq
->sequence
);
460 PUSH_DATA (push
, NV84_SUBCHAN_SEMAPHORE_TRIGGER_ACQUIRE_EQUAL
);