2 * Copyright 2011 Nouveau Project
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
22 * Authors: Christoph Bumiller
25 #define NVC0_PUSH_EXPLICIT_SPACE_CHECKING
27 #include "nvc0/nvc0_context.h"
28 #include "nvc0/nvc0_query.h"
30 #include "nv_object.xml.h"
31 #include "nvc0/nve4_compute.xml.h"
32 #include "nvc0/nvc0_compute.xml.h"
34 #define NVC0_QUERY_STATE_READY 0
35 #define NVC0_QUERY_STATE_ACTIVE 1
36 #define NVC0_QUERY_STATE_ENDED 2
37 #define NVC0_QUERY_STATE_FLUSHED 3
39 #define NVC0_QUERY_ALLOC_SPACE 256
41 static boolean
nvc0_hw_sm_query_begin(struct nvc0_context
*,
43 static void nvc0_hw_sm_query_end(struct nvc0_context
*, struct nvc0_query
*);
44 static boolean
nvc0_hw_sm_query_result(struct nvc0_context
*,
45 struct nvc0_query
*, void *, boolean
);
48 nvc0_query_allocate(struct nvc0_context
*nvc0
, struct nvc0_query
*q
, int size
)
50 struct nvc0_screen
*screen
= nvc0
->screen
;
54 nouveau_bo_ref(NULL
, &q
->bo
);
56 if (q
->state
== NVC0_QUERY_STATE_READY
)
57 nouveau_mm_free(q
->u
.mm
);
59 nouveau_fence_work(screen
->base
.fence
.current
,
60 nouveau_mm_free_work
, q
->u
.mm
);
64 q
->u
.mm
= nouveau_mm_allocate(screen
->base
.mm_GART
, size
, &q
->bo
, &q
->base
);
69 ret
= nouveau_bo_map(q
->bo
, 0, screen
->base
.client
);
71 nvc0_query_allocate(nvc0
, q
, 0);
74 q
->data
= (uint32_t *)((uint8_t *)q
->bo
->map
+ q
->base
);
80 nvc0_query_destroy(struct pipe_context
*pipe
, struct pipe_query
*pq
)
82 nvc0_query_allocate(nvc0_context(pipe
), nvc0_query(pq
), 0);
83 nouveau_fence_ref(NULL
, &nvc0_query(pq
)->fence
);
87 static struct pipe_query
*
88 nvc0_query_create(struct pipe_context
*pipe
, unsigned type
, unsigned index
)
90 struct nvc0_context
*nvc0
= nvc0_context(pipe
);
92 unsigned space
= NVC0_QUERY_ALLOC_SPACE
;
94 q
= CALLOC_STRUCT(nvc0_query
);
99 case PIPE_QUERY_OCCLUSION_COUNTER
:
100 case PIPE_QUERY_OCCLUSION_PREDICATE
:
102 space
= NVC0_QUERY_ALLOC_SPACE
;
104 case PIPE_QUERY_PIPELINE_STATISTICS
:
108 case PIPE_QUERY_SO_STATISTICS
:
109 case PIPE_QUERY_SO_OVERFLOW_PREDICATE
:
113 case PIPE_QUERY_PRIMITIVES_GENERATED
:
114 case PIPE_QUERY_PRIMITIVES_EMITTED
:
119 case PIPE_QUERY_TIME_ELAPSED
:
120 case PIPE_QUERY_TIMESTAMP
:
121 case PIPE_QUERY_TIMESTAMP_DISJOINT
:
122 case PIPE_QUERY_GPU_FINISHED
:
125 case NVC0_QUERY_TFB_BUFFER_OFFSET
:
129 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
130 if (type
>= NVC0_QUERY_DRV_STAT(0) && type
<= NVC0_QUERY_DRV_STAT_LAST
) {
133 q
->index
= type
- NVC0_QUERY_DRV_STAT(0);
137 if (nvc0
->screen
->base
.device
->drm_version
>= 0x01000101) {
138 if (type
>= NVE4_HW_SM_QUERY(0) && type
<= NVE4_HW_SM_QUERY_LAST
) {
160 * [50] = WS0.sequence
161 * [54] = WS1.sequence
162 * [58] = WS2.sequence
163 * [5c] = WS3.sequence
165 space
= (4 * 4 + 4 + 4) * nvc0
->screen
->mp_count
* sizeof(uint32_t);
168 if (type
>= NVC0_HW_SM_QUERY(0) && type
<= NVC0_HW_SM_QUERY_LAST
) {
180 space
= (8 + 1) * nvc0
->screen
->mp_count
* sizeof(uint32_t);
184 debug_printf("invalid query type: %u\n", type
);
188 if (!nvc0_query_allocate(nvc0
, q
, space
)) {
196 /* we advance before query_begin ! */
197 q
->offset
-= q
->rotate
;
198 q
->data
-= q
->rotate
/ sizeof(*q
->data
);
201 q
->data
[0] = 0; /* initialize sequence */
203 return (struct pipe_query
*)q
;
207 nvc0_query_get(struct nouveau_pushbuf
*push
, struct nvc0_query
*q
,
208 unsigned offset
, uint32_t get
)
213 PUSH_REFN (push
, q
->bo
, NOUVEAU_BO_GART
| NOUVEAU_BO_WR
);
214 BEGIN_NVC0(push
, NVC0_3D(QUERY_ADDRESS_HIGH
), 4);
215 PUSH_DATAh(push
, q
->bo
->offset
+ offset
);
216 PUSH_DATA (push
, q
->bo
->offset
+ offset
);
217 PUSH_DATA (push
, q
->sequence
);
218 PUSH_DATA (push
, get
);
222 nvc0_query_rotate(struct nvc0_context
*nvc0
, struct nvc0_query
*q
)
224 q
->offset
+= q
->rotate
;
225 q
->data
+= q
->rotate
/ sizeof(*q
->data
);
226 if (q
->offset
- q
->base
== NVC0_QUERY_ALLOC_SPACE
)
227 nvc0_query_allocate(nvc0
, q
, NVC0_QUERY_ALLOC_SPACE
);
231 nvc0_query_begin(struct pipe_context
*pipe
, struct pipe_query
*pq
)
233 struct nvc0_context
*nvc0
= nvc0_context(pipe
);
234 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
235 struct nvc0_query
*q
= nvc0_query(pq
);
238 /* For occlusion queries we have to change the storage, because a previous
239 * query might set the initial render conition to false even *after* we re-
240 * initialized it to true.
243 nvc0_query_rotate(nvc0
, q
);
245 /* XXX: can we do this with the GPU, and sync with respect to a previous
248 q
->data
[0] = q
->sequence
; /* initialize sequence */
249 q
->data
[1] = 1; /* initial render condition = true */
250 q
->data
[4] = q
->sequence
+ 1; /* for comparison COND_MODE */
256 case PIPE_QUERY_OCCLUSION_COUNTER
:
257 case PIPE_QUERY_OCCLUSION_PREDICATE
:
258 q
->nesting
= nvc0
->screen
->num_occlusion_queries_active
++;
260 nvc0_query_get(push
, q
, 0x10, 0x0100f002);
263 BEGIN_NVC0(push
, NVC0_3D(COUNTER_RESET
), 1);
264 PUSH_DATA (push
, NVC0_3D_COUNTER_RESET_SAMPLECNT
);
265 IMMED_NVC0(push
, NVC0_3D(SAMPLECNT_ENABLE
), 1);
268 case PIPE_QUERY_PRIMITIVES_GENERATED
:
269 nvc0_query_get(push
, q
, 0x10, 0x09005002 | (q
->index
<< 5));
271 case PIPE_QUERY_PRIMITIVES_EMITTED
:
272 nvc0_query_get(push
, q
, 0x10, 0x05805002 | (q
->index
<< 5));
274 case PIPE_QUERY_SO_STATISTICS
:
275 nvc0_query_get(push
, q
, 0x20, 0x05805002 | (q
->index
<< 5));
276 nvc0_query_get(push
, q
, 0x30, 0x06805002 | (q
->index
<< 5));
278 case PIPE_QUERY_SO_OVERFLOW_PREDICATE
:
279 nvc0_query_get(push
, q
, 0x10, 0x03005002 | (q
->index
<< 5));
281 case PIPE_QUERY_TIME_ELAPSED
:
282 nvc0_query_get(push
, q
, 0x10, 0x00005002);
284 case PIPE_QUERY_PIPELINE_STATISTICS
:
285 nvc0_query_get(push
, q
, 0xc0 + 0x00, 0x00801002); /* VFETCH, VERTICES */
286 nvc0_query_get(push
, q
, 0xc0 + 0x10, 0x01801002); /* VFETCH, PRIMS */
287 nvc0_query_get(push
, q
, 0xc0 + 0x20, 0x02802002); /* VP, LAUNCHES */
288 nvc0_query_get(push
, q
, 0xc0 + 0x30, 0x03806002); /* GP, LAUNCHES */
289 nvc0_query_get(push
, q
, 0xc0 + 0x40, 0x04806002); /* GP, PRIMS_OUT */
290 nvc0_query_get(push
, q
, 0xc0 + 0x50, 0x07804002); /* RAST, PRIMS_IN */
291 nvc0_query_get(push
, q
, 0xc0 + 0x60, 0x08804002); /* RAST, PRIMS_OUT */
292 nvc0_query_get(push
, q
, 0xc0 + 0x70, 0x0980a002); /* ROP, PIXELS */
293 nvc0_query_get(push
, q
, 0xc0 + 0x80, 0x0d808002); /* TCP, LAUNCHES */
294 nvc0_query_get(push
, q
, 0xc0 + 0x90, 0x0e809002); /* TEP, LAUNCHES */
297 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
298 if (q
->type
>= NVC0_QUERY_DRV_STAT(0) &&
299 q
->type
<= NVC0_QUERY_DRV_STAT_LAST
) {
301 q
->u
.value
= nvc0
->screen
->base
.stats
.v
[q
->index
];
306 if ((q
->type
>= NVE4_HW_SM_QUERY(0) && q
->type
<= NVE4_HW_SM_QUERY_LAST
) ||
307 (q
->type
>= NVC0_HW_SM_QUERY(0) && q
->type
<= NVC0_HW_SM_QUERY_LAST
)) {
308 ret
= nvc0_hw_sm_query_begin(nvc0
, q
);
312 q
->state
= NVC0_QUERY_STATE_ACTIVE
;
317 nvc0_query_end(struct pipe_context
*pipe
, struct pipe_query
*pq
)
319 struct nvc0_context
*nvc0
= nvc0_context(pipe
);
320 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
321 struct nvc0_query
*q
= nvc0_query(pq
);
323 if (q
->state
!= NVC0_QUERY_STATE_ACTIVE
) {
324 /* some queries don't require 'begin' to be called (e.g. GPU_FINISHED) */
326 nvc0_query_rotate(nvc0
, q
);
329 q
->state
= NVC0_QUERY_STATE_ENDED
;
332 case PIPE_QUERY_OCCLUSION_COUNTER
:
333 case PIPE_QUERY_OCCLUSION_PREDICATE
:
334 nvc0_query_get(push
, q
, 0, 0x0100f002);
335 if (--nvc0
->screen
->num_occlusion_queries_active
== 0) {
337 IMMED_NVC0(push
, NVC0_3D(SAMPLECNT_ENABLE
), 0);
340 case PIPE_QUERY_PRIMITIVES_GENERATED
:
341 nvc0_query_get(push
, q
, 0, 0x09005002 | (q
->index
<< 5));
343 case PIPE_QUERY_PRIMITIVES_EMITTED
:
344 nvc0_query_get(push
, q
, 0, 0x05805002 | (q
->index
<< 5));
346 case PIPE_QUERY_SO_STATISTICS
:
347 nvc0_query_get(push
, q
, 0x00, 0x05805002 | (q
->index
<< 5));
348 nvc0_query_get(push
, q
, 0x10, 0x06805002 | (q
->index
<< 5));
350 case PIPE_QUERY_SO_OVERFLOW_PREDICATE
:
351 /* TODO: How do we sum over all streams for render condition ? */
352 /* PRIMS_DROPPED doesn't write sequence, use a ZERO query to sync on */
353 nvc0_query_get(push
, q
, 0x00, 0x03005002 | (q
->index
<< 5));
354 nvc0_query_get(push
, q
, 0x20, 0x00005002);
356 case PIPE_QUERY_TIMESTAMP
:
357 case PIPE_QUERY_TIME_ELAPSED
:
358 nvc0_query_get(push
, q
, 0, 0x00005002);
360 case PIPE_QUERY_GPU_FINISHED
:
361 nvc0_query_get(push
, q
, 0, 0x1000f010);
363 case PIPE_QUERY_PIPELINE_STATISTICS
:
364 nvc0_query_get(push
, q
, 0x00, 0x00801002); /* VFETCH, VERTICES */
365 nvc0_query_get(push
, q
, 0x10, 0x01801002); /* VFETCH, PRIMS */
366 nvc0_query_get(push
, q
, 0x20, 0x02802002); /* VP, LAUNCHES */
367 nvc0_query_get(push
, q
, 0x30, 0x03806002); /* GP, LAUNCHES */
368 nvc0_query_get(push
, q
, 0x40, 0x04806002); /* GP, PRIMS_OUT */
369 nvc0_query_get(push
, q
, 0x50, 0x07804002); /* RAST, PRIMS_IN */
370 nvc0_query_get(push
, q
, 0x60, 0x08804002); /* RAST, PRIMS_OUT */
371 nvc0_query_get(push
, q
, 0x70, 0x0980a002); /* ROP, PIXELS */
372 nvc0_query_get(push
, q
, 0x80, 0x0d808002); /* TCP, LAUNCHES */
373 nvc0_query_get(push
, q
, 0x90, 0x0e809002); /* TEP, LAUNCHES */
375 case NVC0_QUERY_TFB_BUFFER_OFFSET
:
376 /* indexed by TFB buffer instead of by vertex stream */
377 nvc0_query_get(push
, q
, 0x00, 0x0d005002 | (q
->index
<< 5));
379 case PIPE_QUERY_TIMESTAMP_DISJOINT
:
380 /* This query is not issued on GPU because disjoint is forced to false */
381 q
->state
= NVC0_QUERY_STATE_READY
;
384 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
385 if (q
->type
>= NVC0_QUERY_DRV_STAT(0) &&
386 q
->type
<= NVC0_QUERY_DRV_STAT_LAST
) {
387 q
->u
.value
= nvc0
->screen
->base
.stats
.v
[q
->index
] - q
->u
.value
;
391 if ((q
->type
>= NVE4_HW_SM_QUERY(0) && q
->type
<= NVE4_HW_SM_QUERY_LAST
) ||
392 (q
->type
>= NVC0_HW_SM_QUERY(0) && q
->type
<= NVC0_HW_SM_QUERY_LAST
)) {
393 nvc0_hw_sm_query_end(nvc0
, q
);
398 nouveau_fence_ref(nvc0
->screen
->base
.fence
.current
, &q
->fence
);
402 nvc0_query_update(struct nouveau_client
*cli
, struct nvc0_query
*q
)
405 if (nouveau_fence_signalled(q
->fence
))
406 q
->state
= NVC0_QUERY_STATE_READY
;
408 if (q
->data
[0] == q
->sequence
)
409 q
->state
= NVC0_QUERY_STATE_READY
;
414 nvc0_query_result(struct pipe_context
*pipe
, struct pipe_query
*pq
,
415 boolean wait
, union pipe_query_result
*result
)
417 struct nvc0_context
*nvc0
= nvc0_context(pipe
);
418 struct nvc0_query
*q
= nvc0_query(pq
);
419 uint64_t *res64
= (uint64_t*)result
;
420 uint32_t *res32
= (uint32_t*)result
;
421 uint8_t *res8
= (uint8_t*)result
;
422 uint64_t *data64
= (uint64_t *)q
->data
;
425 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
426 if (q
->type
>= NVC0_QUERY_DRV_STAT(0) &&
427 q
->type
<= NVC0_QUERY_DRV_STAT_LAST
) {
428 res64
[0] = q
->u
.value
;
432 if ((q
->type
>= NVE4_HW_SM_QUERY(0) && q
->type
<= NVE4_HW_SM_QUERY_LAST
) ||
433 (q
->type
>= NVC0_HW_SM_QUERY(0) && q
->type
<= NVC0_HW_SM_QUERY_LAST
)) {
434 return nvc0_hw_sm_query_result(nvc0
, q
, result
, wait
);
437 if (q
->state
!= NVC0_QUERY_STATE_READY
)
438 nvc0_query_update(nvc0
->screen
->base
.client
, q
);
440 if (q
->state
!= NVC0_QUERY_STATE_READY
) {
442 if (q
->state
!= NVC0_QUERY_STATE_FLUSHED
) {
443 q
->state
= NVC0_QUERY_STATE_FLUSHED
;
444 /* flush for silly apps that spin on GL_QUERY_RESULT_AVAILABLE */
445 PUSH_KICK(nvc0
->base
.pushbuf
);
449 if (nouveau_bo_wait(q
->bo
, NOUVEAU_BO_RD
, nvc0
->screen
->base
.client
))
451 NOUVEAU_DRV_STAT(&nvc0
->screen
->base
, query_sync_count
, 1);
453 q
->state
= NVC0_QUERY_STATE_READY
;
456 case PIPE_QUERY_GPU_FINISHED
:
459 case PIPE_QUERY_OCCLUSION_COUNTER
: /* u32 sequence, u32 count, u64 time */
460 res64
[0] = q
->data
[1] - q
->data
[5];
462 case PIPE_QUERY_OCCLUSION_PREDICATE
:
463 res8
[0] = q
->data
[1] != q
->data
[5];
465 case PIPE_QUERY_PRIMITIVES_GENERATED
: /* u64 count, u64 time */
466 case PIPE_QUERY_PRIMITIVES_EMITTED
: /* u64 count, u64 time */
467 res64
[0] = data64
[0] - data64
[2];
469 case PIPE_QUERY_SO_STATISTICS
:
470 res64
[0] = data64
[0] - data64
[4];
471 res64
[1] = data64
[2] - data64
[6];
473 case PIPE_QUERY_SO_OVERFLOW_PREDICATE
:
474 res8
[0] = data64
[0] != data64
[2];
476 case PIPE_QUERY_TIMESTAMP
:
477 res64
[0] = data64
[1];
479 case PIPE_QUERY_TIMESTAMP_DISJOINT
:
480 res64
[0] = 1000000000;
483 case PIPE_QUERY_TIME_ELAPSED
:
484 res64
[0] = data64
[1] - data64
[3];
486 case PIPE_QUERY_PIPELINE_STATISTICS
:
487 for (i
= 0; i
< 10; ++i
)
488 res64
[i
] = data64
[i
* 2] - data64
[24 + i
* 2];
490 case NVC0_QUERY_TFB_BUFFER_OFFSET
:
491 res32
[0] = q
->data
[1];
494 assert(0); /* can't happen, we don't create queries with invalid type */
502 nvc0_query_fifo_wait(struct nouveau_pushbuf
*push
, struct nvc0_query
*q
)
504 unsigned offset
= q
->offset
;
506 if (q
->type
== PIPE_QUERY_SO_OVERFLOW_PREDICATE
) offset
+= 0x20;
509 PUSH_REFN (push
, q
->bo
, NOUVEAU_BO_GART
| NOUVEAU_BO_RD
);
510 BEGIN_NVC0(push
, SUBC_3D(NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH
), 4);
511 PUSH_DATAh(push
, q
->bo
->offset
+ offset
);
512 PUSH_DATA (push
, q
->bo
->offset
+ offset
);
513 PUSH_DATA (push
, q
->sequence
);
514 PUSH_DATA (push
, (1 << 12) |
515 NV84_SUBCHAN_SEMAPHORE_TRIGGER_ACQUIRE_EQUAL
);
519 nvc0_render_condition(struct pipe_context
*pipe
,
520 struct pipe_query
*pq
,
521 boolean condition
, uint mode
)
523 struct nvc0_context
*nvc0
= nvc0_context(pipe
);
524 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
525 struct nvc0_query
*q
;
528 mode
!= PIPE_RENDER_COND_NO_WAIT
&&
529 mode
!= PIPE_RENDER_COND_BY_REGION_NO_WAIT
;
532 cond
= NVC0_3D_COND_MODE_ALWAYS
;
536 /* NOTE: comparison of 2 queries only works if both have completed */
538 case PIPE_QUERY_SO_OVERFLOW_PREDICATE
:
539 cond
= condition
? NVC0_3D_COND_MODE_EQUAL
:
540 NVC0_3D_COND_MODE_NOT_EQUAL
;
543 case PIPE_QUERY_OCCLUSION_COUNTER
:
544 case PIPE_QUERY_OCCLUSION_PREDICATE
:
545 if (likely(!condition
)) {
546 if (unlikely(q
->nesting
))
547 cond
= wait
? NVC0_3D_COND_MODE_NOT_EQUAL
:
548 NVC0_3D_COND_MODE_ALWAYS
;
550 cond
= NVC0_3D_COND_MODE_RES_NON_ZERO
;
552 cond
= wait
? NVC0_3D_COND_MODE_EQUAL
: NVC0_3D_COND_MODE_ALWAYS
;
556 assert(!"render condition query not a predicate");
557 cond
= NVC0_3D_COND_MODE_ALWAYS
;
562 nvc0
->cond_query
= pq
;
563 nvc0
->cond_cond
= condition
;
564 nvc0
->cond_condmode
= cond
;
565 nvc0
->cond_mode
= mode
;
569 IMMED_NVC0(push
, NVC0_3D(COND_MODE
), cond
);
574 nvc0_query_fifo_wait(push
, q
);
577 PUSH_REFN (push
, q
->bo
, NOUVEAU_BO_GART
| NOUVEAU_BO_RD
);
578 BEGIN_NVC0(push
, NVC0_3D(COND_ADDRESS_HIGH
), 3);
579 PUSH_DATAh(push
, q
->bo
->offset
+ q
->offset
);
580 PUSH_DATA (push
, q
->bo
->offset
+ q
->offset
);
581 PUSH_DATA (push
, cond
);
582 BEGIN_NVC0(push
, NVC0_2D(COND_ADDRESS_HIGH
), 2);
583 PUSH_DATAh(push
, q
->bo
->offset
+ q
->offset
);
584 PUSH_DATA (push
, q
->bo
->offset
+ q
->offset
);
588 nvc0_query_pushbuf_submit(struct nouveau_pushbuf
*push
,
589 struct nvc0_query
*q
, unsigned result_offset
)
591 #define NVC0_IB_ENTRY_1_NO_PREFETCH (1 << (31 - 8))
593 PUSH_REFN(push
, q
->bo
, NOUVEAU_BO_RD
| NOUVEAU_BO_GART
);
594 nouveau_pushbuf_space(push
, 0, 0, 1);
595 nouveau_pushbuf_data(push
, q
->bo
, q
->offset
+ result_offset
, 4 |
596 NVC0_IB_ENTRY_1_NO_PREFETCH
);
600 nvc0_so_target_save_offset(struct pipe_context
*pipe
,
601 struct pipe_stream_output_target
*ptarg
,
602 unsigned index
, bool *serialize
)
604 struct nvc0_so_target
*targ
= nvc0_so_target(ptarg
);
608 PUSH_SPACE(nvc0_context(pipe
)->base
.pushbuf
, 1);
609 IMMED_NVC0(nvc0_context(pipe
)->base
.pushbuf
, NVC0_3D(SERIALIZE
), 0);
611 NOUVEAU_DRV_STAT(nouveau_screen(pipe
->screen
), gpu_serialize_count
, 1);
614 nvc0_query(targ
->pq
)->index
= index
;
616 nvc0_query_end(pipe
, targ
->pq
);
620 /* === DRIVER STATISTICS === */
622 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
624 static const char *nvc0_drv_stat_names
[] =
626 "drv-tex_obj_current_count",
627 "drv-tex_obj_current_bytes",
628 "drv-buf_obj_current_count",
629 "drv-buf_obj_current_bytes_vid",
630 "drv-buf_obj_current_bytes_sys",
631 "drv-tex_transfers_rd",
632 "drv-tex_transfers_wr",
633 "drv-tex_copy_count",
634 "drv-tex_blit_count",
635 "drv-tex_cache_flush_count",
636 "drv-buf_transfers_rd",
637 "drv-buf_transfers_wr",
638 "drv-buf_read_bytes_staging_vid",
639 "drv-buf_write_bytes_direct",
640 "drv-buf_write_bytes_staging_vid",
641 "drv-buf_write_bytes_staging_sys",
642 "drv-buf_copy_bytes",
643 "drv-buf_non_kernel_fence_sync_count",
644 "drv-any_non_kernel_fence_sync_count",
645 "drv-query_sync_count",
646 "drv-gpu_serialize_count",
647 "drv-draw_calls_array",
648 "drv-draw_calls_indexed",
649 "drv-draw_calls_fallback_count",
650 "drv-user_buffer_upload_bytes",
651 "drv-constbuf_upload_count",
652 "drv-constbuf_upload_bytes",
654 "drv-resource_validate_count"
657 #endif /* NOUVEAU_ENABLE_DRIVER_STATISTICS */
660 /* === PERFORMANCE MONITORING COUNTERS for NVE4+ === */
662 /* Code to read out MP counters: They are accessible via mmio, too, but let's
663 * just avoid mapping registers in userspace. We'd have to know which MPs are
664 * enabled/present, too, and that information is not presently exposed.
665 * We could add a kernel interface for it, but reading the counters like this
666 * has the advantage of being async (if get_result isn't called immediately).
668 static const uint64_t nve4_read_hw_sm_counters_code
[] =
670 /* sched 0x20 0x20 0x20 0x20 0x20 0x20 0x20
672 * mov b32 $r12 $physid
678 * sched 0x20 0x20 0x23 0x04 0x20 0x04 0x2b
682 * set $p0 0x1 eq u32 $r8 0x0
683 * mov b32 $r10 c0[0x0]
684 * ext u32 $r8 $r12 0x414
685 * mov b32 $r11 c0[0x4]
686 * sched 0x04 0x2e 0x04 0x20 0x20 0x28 0x04
687 * ext u32 $r9 $r12 0x208
689 * set $p1 0x1 eq u32 $r9 0x0
690 * mul $r8 u32 $r8 u32 96
691 * mul $r12 u32 $r9 u32 16
692 * mul $r13 u32 $r9 u32 4
693 * add b32 $r9 $r8 $r13
694 * sched 0x28 0x04 0x2c 0x04 0x2c 0x04 0x2c
695 * add b32 $r8 $r8 $r12
697 * add b32 $r10 $c $r10 $r8
699 * add b32 $r11 $r11 0x0 $c
700 * add b32 $r12 $c $r12 $r9
701 * st b128 wt g[$r10d] $r0q
702 * sched 0x4 0x2c 0x20 0x04 0x2e 0x00 0x00
703 * mov b32 $r0 c0[0x8]
704 * add b32 $r13 $r13 0x0 $c
705 * $p1 st b128 wt g[$r12d+0x40] $r4q
706 * st b32 wt g[$r12d+0x50] $r0
708 0x2202020202020207ULL
,
709 0x2c00000084021c04ULL
,
710 0x2c0000000c031c04ULL
,
711 0x2c00000010001c04ULL
,
712 0x2c00000014005c04ULL
,
713 0x2c00000018009c04ULL
,
714 0x2c0000001c00dc04ULL
,
715 0x2c00000020011c04ULL
,
716 0x22b0420042320207ULL
,
717 0x2c00000024015c04ULL
,
718 0x2c00000028019c04ULL
,
719 0x2c0000002c01dc04ULL
,
720 0x190e0000fc81dc03ULL
,
721 0x2800400000029de4ULL
,
722 0x7000c01050c21c03ULL
,
723 0x280040001002dde4ULL
,
724 0x204282020042e047ULL
,
725 0x7000c00820c25c03ULL
,
726 0x80000000000021e7ULL
,
727 0x190e0000fc93dc03ULL
,
728 0x1000000180821c02ULL
,
729 0x1000000040931c02ULL
,
730 0x1000000010935c02ULL
,
731 0x4800000034825c03ULL
,
732 0x22c042c042c04287ULL
,
733 0x4800000030821c03ULL
,
734 0x2800000028031de4ULL
,
735 0x4801000020a29c03ULL
,
736 0x280000002c035de4ULL
,
737 0x0800000000b2dc42ULL
,
738 0x4801000024c31c03ULL
,
739 0x9400000000a01fc5ULL
,
740 0x200002e04202c047ULL
,
741 0x2800400020001de4ULL
,
742 0x0800000000d35c42ULL
,
743 0x9400000100c107c5ULL
,
744 0x9400000140c01f85ULL
,
745 0x8000000000001de7ULL
748 /* NOTE: intentionally using the same names as NV */
749 static const char *nve4_pm_query_names
[] =
758 "global_ld_mem_divergence_replays",
759 "global_store_transaction",
760 "global_st_mem_divergence_replays",
767 "l1_global_load_hit",
768 "l1_global_load_miss",
770 "l1_local_load_miss",
771 "l1_local_store_hit",
772 "l1_local_store_miss",
773 "l1_shared_load_transactions",
774 "l1_shared_store_transactions",
776 "local_load_transactions",
778 "local_store_transactions",
788 "shared_load_replay",
790 "shared_store_replay",
793 "uncached_global_load_transaction",
795 /* metrics, i.e. functions of the MP counters */
796 "metric-ipc", /* inst_executed, clock */
797 "metric-ipac", /* inst_executed, active_cycles */
798 "metric-ipec", /* inst_executed, (bool)inst_executed */
799 "metric-achieved_occupancy", /* active_warps, active_cycles */
800 "metric-sm_efficiency", /* active_cycles, clock */
801 "metric-inst_replay_overhead" /* inst_issued, inst_executed */
804 /* For simplicity, we will allocate as many group slots as we allocate counter
805 * slots. This means that a single counter which wants to source from 2 groups
806 * will have to be declared as using 2 counter slots. This shouldn't really be
807 * a problem because such queries don't make much sense ... (unless someone is
810 struct nvc0_mp_counter_cfg
812 uint32_t func
: 16; /* mask or 4-bit logic op (depending on mode) */
813 uint32_t mode
: 4; /* LOGOP,B6,LOGOP_B6(_PULSE) */
814 uint32_t num_src
: 3; /* number of sources (1 - 6, only for NVC0:NVE4) */
815 uint32_t sig_dom
: 1; /* if 0, MP_PM_A (per warp-sched), if 1, MP_PM_B */
816 uint32_t sig_sel
: 8; /* signal group */
817 uint64_t src_sel
; /* signal selection for up to 6 sources (48 bit) */
820 #define NVC0_COUNTER_OPn_SUM 0
821 #define NVC0_COUNTER_OPn_OR 1
822 #define NVC0_COUNTER_OPn_AND 2
823 #define NVC0_COUNTER_OP2_REL_SUM_MM 3 /* (sum(ctr0) - sum(ctr1)) / sum(ctr0) */
824 #define NVC0_COUNTER_OP2_DIV_SUM_M0 4 /* sum(ctr0) / ctr1 of MP[0]) */
825 #define NVC0_COUNTER_OP2_AVG_DIV_MM 5 /* avg(ctr0 / ctr1) */
826 #define NVC0_COUNTER_OP2_AVG_DIV_M0 6 /* avg(ctr0) / ctr1 of MP[0]) */
828 struct nvc0_hw_sm_query_cfg
830 struct nvc0_mp_counter_cfg ctr
[4];
831 uint8_t num_counters
;
833 uint8_t norm
[2]; /* normalization num,denom */
836 #define _Q1A(n, f, m, g, s, nu, dn) [NVE4_HW_SM_QUERY_##n] = { { { f, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m, 0, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g, s }, {}, {}, {} }, 1, NVC0_COUNTER_OPn_SUM, { nu, dn } }
837 #define _Q1B(n, f, m, g, s, nu, dn) [NVE4_HW_SM_QUERY_##n] = { { { f, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m, 0, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g, s }, {}, {}, {} }, 1, NVC0_COUNTER_OPn_SUM, { nu, dn } }
838 #define _M2A(n, f0, m0, g0, s0, f1, m1, g1, s1, o, nu, dn) [NVE4_HW_SM_QUERY_METRIC_##n] = { { \
839 { f0, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m0, 0, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g0, s0 }, \
840 { f1, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m1, 0, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g1, s1 }, \
841 {}, {}, }, 2, NVC0_COUNTER_OP2_##o, { nu, dn } }
842 #define _M2B(n, f0, m0, g0, s0, f1, m1, g1, s1, o, nu, dn) [NVE4_HW_SM_QUERY_METRIC_##n] = { { \
843 { f0, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m0, 0, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g0, s0 }, \
844 { f1, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m1, 0, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g1, s1 }, \
845 {}, {}, }, 2, NVC0_COUNTER_OP2_##o, { nu, dn } }
846 #define _M2AB(n, f0, m0, g0, s0, f1, m1, g1, s1, o, nu, dn) [NVE4_HW_SM_QUERY_METRIC_##n] = { { \
847 { f0, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m0, 0, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g0, s0 }, \
848 { f1, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m1, 0, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g1, s1 }, \
849 {}, {}, }, 2, NVC0_COUNTER_OP2_##o, { nu, dn } }
852 * active_warps: bit 0 alternates btw 0 and 1 for odd nr of warps
853 * inst_executed etc.: we only count a single warp scheduler
854 * metric-ipXc: we simply multiply by 4 to account for the 4 warp schedulers;
855 * this is inaccurate !
857 static const struct nvc0_hw_sm_query_cfg nve4_hw_sm_queries
[] =
859 _Q1B(ACTIVE_CYCLES
, 0x0001, B6
, WARP
, 0x00000000, 1, 1),
860 _Q1B(ACTIVE_WARPS
, 0x003f, B6
, WARP
, 0x31483104, 2, 1),
861 _Q1A(ATOM_COUNT
, 0x0001, B6
, BRANCH
, 0x00000000, 1, 1),
862 _Q1A(BRANCH
, 0x0001, B6
, BRANCH
, 0x0000000c, 1, 1),
863 _Q1A(DIVERGENT_BRANCH
, 0x0001, B6
, BRANCH
, 0x00000010, 1, 1),
864 _Q1A(GLD_REQUEST
, 0x0001, B6
, LDST
, 0x00000010, 1, 1),
865 _Q1B(GLD_MEM_DIV_REPLAY
, 0x0001, B6
, REPLAY
, 0x00000010, 1, 1),
866 _Q1B(GST_TRANSACTIONS
, 0x0001, B6
, MEM
, 0x00000004, 1, 1),
867 _Q1B(GST_MEM_DIV_REPLAY
, 0x0001, B6
, REPLAY
, 0x00000014, 1, 1),
868 _Q1A(GRED_COUNT
, 0x0001, B6
, BRANCH
, 0x00000008, 1, 1),
869 _Q1A(GST_REQUEST
, 0x0001, B6
, LDST
, 0x00000014, 1, 1),
870 _Q1A(INST_EXECUTED
, 0x0003, B6
, EXEC
, 0x00000398, 1, 1),
871 _Q1A(INST_ISSUED
, 0x0003, B6
, ISSUE
, 0x00000104, 1, 1),
872 _Q1A(INST_ISSUED1
, 0x0001, B6
, ISSUE
, 0x00000004, 1, 1),
873 _Q1A(INST_ISSUED2
, 0x0001, B6
, ISSUE
, 0x00000008, 1, 1),
874 _Q1B(L1_GLD_HIT
, 0x0001, B6
, L1
, 0x00000010, 1, 1),
875 _Q1B(L1_GLD_MISS
, 0x0001, B6
, L1
, 0x00000014, 1, 1),
876 _Q1B(L1_LOCAL_LD_HIT
, 0x0001, B6
, L1
, 0x00000000, 1, 1),
877 _Q1B(L1_LOCAL_LD_MISS
, 0x0001, B6
, L1
, 0x00000004, 1, 1),
878 _Q1B(L1_LOCAL_ST_HIT
, 0x0001, B6
, L1
, 0x00000008, 1, 1),
879 _Q1B(L1_LOCAL_ST_MISS
, 0x0001, B6
, L1
, 0x0000000c, 1, 1),
880 _Q1B(L1_SHARED_LD_TRANSACTIONS
, 0x0001, B6
, TRANSACTION
, 0x00000008, 1, 1),
881 _Q1B(L1_SHARED_ST_TRANSACTIONS
, 0x0001, B6
, TRANSACTION
, 0x0000000c, 1, 1),
882 _Q1A(LOCAL_LD
, 0x0001, B6
, LDST
, 0x00000008, 1, 1),
883 _Q1B(LOCAL_LD_TRANSACTIONS
, 0x0001, B6
, TRANSACTION
, 0x00000000, 1, 1),
884 _Q1A(LOCAL_ST
, 0x0001, B6
, LDST
, 0x0000000c, 1, 1),
885 _Q1B(LOCAL_ST_TRANSACTIONS
, 0x0001, B6
, TRANSACTION
, 0x00000004, 1, 1),
886 _Q1A(PROF_TRIGGER_0
, 0x0001, B6
, USER
, 0x00000000, 1, 1),
887 _Q1A(PROF_TRIGGER_1
, 0x0001, B6
, USER
, 0x00000004, 1, 1),
888 _Q1A(PROF_TRIGGER_2
, 0x0001, B6
, USER
, 0x00000008, 1, 1),
889 _Q1A(PROF_TRIGGER_3
, 0x0001, B6
, USER
, 0x0000000c, 1, 1),
890 _Q1A(PROF_TRIGGER_4
, 0x0001, B6
, USER
, 0x00000010, 1, 1),
891 _Q1A(PROF_TRIGGER_5
, 0x0001, B6
, USER
, 0x00000014, 1, 1),
892 _Q1A(PROF_TRIGGER_6
, 0x0001, B6
, USER
, 0x00000018, 1, 1),
893 _Q1A(PROF_TRIGGER_7
, 0x0001, B6
, USER
, 0x0000001c, 1, 1),
894 _Q1A(SHARED_LD
, 0x0001, B6
, LDST
, 0x00000000, 1, 1),
895 _Q1B(SHARED_LD_REPLAY
, 0x0001, B6
, REPLAY
, 0x00000008, 1, 1),
896 _Q1A(SHARED_ST
, 0x0001, B6
, LDST
, 0x00000004, 1, 1),
897 _Q1B(SHARED_ST_REPLAY
, 0x0001, B6
, REPLAY
, 0x0000000c, 1, 1),
898 _Q1B(SM_CTA_LAUNCHED
, 0x0001, B6
, WARP
, 0x0000001c, 1, 1),
899 _Q1A(THREADS_LAUNCHED
, 0x003f, B6
, LAUNCH
, 0x398a4188, 1, 1),
900 _Q1B(UNCACHED_GLD_TRANSACTIONS
, 0x0001, B6
, MEM
, 0x00000000, 1, 1),
901 _Q1A(WARPS_LAUNCHED
, 0x0001, B6
, LAUNCH
, 0x00000004, 1, 1),
902 _M2AB(IPC
, 0x3, B6
, EXEC
, 0x398, 0xffff, LOGOP
, WARP
, 0x0, DIV_SUM_M0
, 10, 1),
903 _M2AB(IPAC
, 0x3, B6
, EXEC
, 0x398, 0x1, B6
, WARP
, 0x0, AVG_DIV_MM
, 10, 1),
904 _M2A(IPEC
, 0x3, B6
, EXEC
, 0x398, 0xe, LOGOP
, EXEC
, 0x398, AVG_DIV_MM
, 10, 1),
905 _M2A(INST_REPLAY_OHEAD
, 0x3, B6
, ISSUE
, 0x104, 0x3, B6
, EXEC
, 0x398, REL_SUM_MM
, 100, 1),
906 _M2B(MP_OCCUPANCY
, 0x3f, B6
, WARP
, 0x31483104, 0x01, B6
, WARP
, 0x0, AVG_DIV_MM
, 200, 64),
907 _M2B(MP_EFFICIENCY
, 0x01, B6
, WARP
, 0x0, 0xffff, LOGOP
, WARP
, 0x0, AVG_DIV_M0
, 100, 1),
915 /* === PERFORMANCE MONITORING COUNTERS for NVC0:NVE4 === */
916 static const uint64_t nvc0_read_hw_sm_counters_code
[] =
919 * mov b32 $r9 $physid
928 * set $p0 0x1 eq u32 $r8 0x0
929 * mov b32 $r10 c0[0x0]
930 * mov b32 $r11 c0[0x4]
931 * ext u32 $r8 $r9 0x414
933 * mul $r8 u32 $r8 u32 36
934 * add b32 $r10 $c $r10 $r8
935 * add b32 $r11 $r11 0x0 $c
936 * mov b32 $r8 c0[0x8]
937 * st b128 wt g[$r10d+0x00] $r0q
938 * st b128 wt g[$r10d+0x10] $r4q
939 * st b32 wt g[$r10d+0x20] $r8
941 0x2c00000084021c04ULL
,
942 0x2c0000000c025c04ULL
,
943 0x2c00000010001c04ULL
,
944 0x2c00000014005c04ULL
,
945 0x2c00000018009c04ULL
,
946 0x2c0000001c00dc04ULL
,
947 0x2c00000020011c04ULL
,
948 0x2c00000024015c04ULL
,
949 0x2c00000028019c04ULL
,
950 0x2c0000002c01dc04ULL
,
951 0x190e0000fc81dc03ULL
,
952 0x2800400000029de4ULL
,
953 0x280040001002dde4ULL
,
954 0x7000c01050921c03ULL
,
955 0x80000000000021e7ULL
,
956 0x1000000090821c02ULL
,
957 0x4801000020a29c03ULL
,
958 0x0800000000b2dc42ULL
,
959 0x2800400020021de4ULL
,
960 0x9400000000a01fc5ULL
,
961 0x9400000040a11fc5ULL
,
962 0x9400000080a21f85ULL
,
963 0x8000000000001de7ULL
966 static const char *nvc0_pm_query_names
[] =
995 "thread_inst_executed_0",
996 "thread_inst_executed_1",
997 "thread_inst_executed_2",
998 "thread_inst_executed_3",
1002 #define _Q(n, f, m, g, c, s0, s1, s2, s3, s4, s5) [NVC0_HW_SM_QUERY_##n] = { { { f, NVC0_COMPUTE_MP_PM_OP_MODE_##m, c, 0, g, s0|(s1 << 8)|(s2 << 16)|(s3 << 24)|(s4##ULL << 32)|(s5##ULL << 40) }, {}, {}, {} }, 1, NVC0_COUNTER_OPn_SUM, { 1, 1 } }
1004 static const struct nvc0_hw_sm_query_cfg nvc0_hw_sm_queries
[] =
1006 _Q(ACTIVE_CYCLES
, 0xaaaa, LOGOP
, 0x11, 1, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00),
1007 _Q(ACTIVE_WARPS
, 0xaaaa, LOGOP
, 0x24, 6, 0x10, 0x21, 0x32, 0x43, 0x54, 0x65),
1008 _Q(ATOM_COUNT
, 0xaaaa, LOGOP
, 0x63, 1, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00),
1009 _Q(BRANCH
, 0xaaaa, LOGOP
, 0x1a, 2, 0x00, 0x11, 0x00, 0x00, 0x00, 0x00),
1010 _Q(DIVERGENT_BRANCH
, 0xaaaa, LOGOP
, 0x19, 2, 0x20, 0x31, 0x00, 0x00, 0x00, 0x00),
1011 _Q(GLD_REQUEST
, 0xaaaa, LOGOP
, 0x64, 1, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00),
1012 _Q(GRED_COUNT
, 0xaaaa, LOGOP
, 0x63, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
1013 _Q(GST_REQUEST
, 0xaaaa, LOGOP
, 0x64, 1, 0x60, 0x00, 0x00, 0x00, 0x00, 0x00),
1014 _Q(INST_EXECUTED
, 0xaaaa, LOGOP
, 0x2d, 3, 0x00, 0x11, 0x22, 0x00, 0x00, 0x00),
1015 _Q(INST_ISSUED1_0
, 0xaaaa, LOGOP
, 0x7e, 1, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00),
1016 _Q(INST_ISSUED1_1
, 0xaaaa, LOGOP
, 0x7e, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
1017 _Q(INST_ISSUED2_0
, 0xaaaa, LOGOP
, 0x7e, 1, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00),
1018 _Q(INST_ISSUED2_1
, 0xaaaa, LOGOP
, 0x7e, 1, 0x50, 0x00, 0x00, 0x00, 0x00, 0x00),
1019 _Q(LOCAL_LD
, 0xaaaa, LOGOP
, 0x64, 1, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00),
1020 _Q(LOCAL_ST
, 0xaaaa, LOGOP
, 0x64, 1, 0x50, 0x00, 0x00, 0x00, 0x00, 0x00),
1021 _Q(PROF_TRIGGER_0
, 0xaaaa, LOGOP
, 0x01, 1, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00),
1022 _Q(PROF_TRIGGER_1
, 0xaaaa, LOGOP
, 0x01, 1, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00),
1023 _Q(PROF_TRIGGER_2
, 0xaaaa, LOGOP
, 0x01, 1, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00),
1024 _Q(PROF_TRIGGER_3
, 0xaaaa, LOGOP
, 0x01, 1, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00),
1025 _Q(PROF_TRIGGER_4
, 0xaaaa, LOGOP
, 0x01, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
1026 _Q(PROF_TRIGGER_5
, 0xaaaa, LOGOP
, 0x01, 1, 0x50, 0x00, 0x00, 0x00, 0x00, 0x00),
1027 _Q(PROF_TRIGGER_6
, 0xaaaa, LOGOP
, 0x01, 1, 0x60, 0x00, 0x00, 0x00, 0x00, 0x00),
1028 _Q(PROF_TRIGGER_7
, 0xaaaa, LOGOP
, 0x01, 1, 0x70, 0x00, 0x00, 0x00, 0x00, 0x00),
1029 _Q(SHARED_LD
, 0xaaaa, LOGOP
, 0x64, 1, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00),
1030 _Q(SHARED_ST
, 0xaaaa, LOGOP
, 0x64, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
1031 _Q(THREADS_LAUNCHED
, 0xaaaa, LOGOP
, 0x26, 6, 0x10, 0x21, 0x32, 0x43, 0x54, 0x65),
1032 _Q(TH_INST_EXECUTED_0
, 0xaaaa, LOGOP
, 0xa3, 6, 0x00, 0x11, 0x22, 0x33, 0x44, 0x55),
1033 _Q(TH_INST_EXECUTED_1
, 0xaaaa, LOGOP
, 0xa5, 6, 0x00, 0x11, 0x22, 0x33, 0x44, 0x55),
1034 _Q(TH_INST_EXECUTED_2
, 0xaaaa, LOGOP
, 0xa4, 6, 0x00, 0x11, 0x22, 0x33, 0x44, 0x55),
1035 _Q(TH_INST_EXECUTED_3
, 0xaaaa, LOGOP
, 0xa6, 6, 0x00, 0x11, 0x22, 0x33, 0x44, 0x55),
1036 _Q(WARPS_LAUNCHED
, 0xaaaa, LOGOP
, 0x26, 1, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00),
1041 static const struct nvc0_hw_sm_query_cfg
*
1042 nvc0_hw_sm_query_get_cfg(struct nvc0_context
*nvc0
, struct nvc0_query
*q
)
1044 struct nvc0_screen
*screen
= nvc0
->screen
;
1046 if (screen
->base
.class_3d
>= NVE4_3D_CLASS
)
1047 return &nve4_hw_sm_queries
[q
->type
- PIPE_QUERY_DRIVER_SPECIFIC
];
1048 return &nvc0_hw_sm_queries
[q
->type
- NVC0_HW_SM_QUERY(0)];
1052 nvc0_hw_sm_query_begin(struct nvc0_context
*nvc0
, struct nvc0_query
*q
)
1054 struct nvc0_screen
*screen
= nvc0
->screen
;
1055 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
1056 const bool is_nve4
= screen
->base
.class_3d
>= NVE4_3D_CLASS
;
1057 const struct nvc0_hw_sm_query_cfg
*cfg
;
1059 unsigned num_ab
[2] = { 0, 0 };
1061 cfg
= nvc0_hw_sm_query_get_cfg(nvc0
, q
);
1063 /* check if we have enough free counter slots */
1064 for (i
= 0; i
< cfg
->num_counters
; ++i
)
1065 num_ab
[cfg
->ctr
[i
].sig_dom
]++;
1067 if (screen
->pm
.num_hw_sm_active
[0] + num_ab
[0] > 4 ||
1068 screen
->pm
.num_hw_sm_active
[1] + num_ab
[1] > 4) {
1069 NOUVEAU_ERR("Not enough free MP counter slots !\n");
1073 assert(cfg
->num_counters
<= 4);
1074 PUSH_SPACE(push
, 4 * 8 * (is_nve4
? 1 : 6) + 6);
1076 if (!screen
->pm
.mp_counters_enabled
) {
1077 screen
->pm
.mp_counters_enabled
= true;
1078 BEGIN_NVC0(push
, SUBC_SW(0x06ac), 1);
1079 PUSH_DATA (push
, 0x1fcb);
1082 /* set sequence field to 0 (used to check if result is available) */
1083 for (i
= 0; i
< screen
->mp_count
; ++i
)
1084 q
->data
[i
* 10 + 10] = 0;
1086 for (i
= 0; i
< cfg
->num_counters
; ++i
) {
1087 const unsigned d
= cfg
->ctr
[i
].sig_dom
;
1089 if (!screen
->pm
.num_hw_sm_active
[d
]) {
1090 uint32_t m
= (1 << 22) | (1 << (7 + (8 * !d
)));
1091 if (screen
->pm
.num_hw_sm_active
[!d
])
1092 m
|= 1 << (7 + (8 * d
));
1093 BEGIN_NVC0(push
, SUBC_SW(0x0600), 1);
1094 PUSH_DATA (push
, m
);
1096 screen
->pm
.num_hw_sm_active
[d
]++;
1098 for (c
= d
* 4; c
< (d
* 4 + 4); ++c
) {
1099 if (!screen
->pm
.mp_counter
[c
]) {
1101 screen
->pm
.mp_counter
[c
] = (struct pipe_query
*)q
;
1105 assert(c
<= (d
* 4 + 3)); /* must succeed, already checked for space */
1107 /* configure and reset the counter(s) */
1110 BEGIN_NVC0(push
, NVE4_COMPUTE(MP_PM_A_SIGSEL(c
& 3)), 1);
1112 BEGIN_NVC0(push
, NVE4_COMPUTE(MP_PM_B_SIGSEL(c
& 3)), 1);
1113 PUSH_DATA (push
, cfg
->ctr
[i
].sig_sel
);
1114 BEGIN_NVC0(push
, NVE4_COMPUTE(MP_PM_SRCSEL(c
)), 1);
1115 PUSH_DATA (push
, cfg
->ctr
[i
].src_sel
+ 0x2108421 * (c
& 3));
1116 BEGIN_NVC0(push
, NVE4_COMPUTE(MP_PM_FUNC(c
)), 1);
1117 PUSH_DATA (push
, (cfg
->ctr
[i
].func
<< 4) | cfg
->ctr
[i
].mode
);
1118 BEGIN_NVC0(push
, NVE4_COMPUTE(MP_PM_SET(c
)), 1);
1119 PUSH_DATA (push
, 0);
1123 for (s
= 0; s
< cfg
->ctr
[i
].num_src
; s
++) {
1124 BEGIN_NVC0(push
, NVC0_COMPUTE(MP_PM_SIGSEL(s
)), 1);
1125 PUSH_DATA (push
, cfg
->ctr
[i
].sig_sel
);
1126 BEGIN_NVC0(push
, NVC0_COMPUTE(MP_PM_SRCSEL(s
)), 1);
1127 PUSH_DATA (push
, (cfg
->ctr
[i
].src_sel
>> (s
* 8)) & 0xff);
1128 BEGIN_NVC0(push
, NVC0_COMPUTE(MP_PM_OP(s
)), 1);
1129 PUSH_DATA (push
, (cfg
->ctr
[i
].func
<< 4) | cfg
->ctr
[i
].mode
);
1130 BEGIN_NVC0(push
, NVC0_COMPUTE(MP_PM_SET(s
)), 1);
1131 PUSH_DATA (push
, 0);
1139 nvc0_hw_sm_query_end(struct nvc0_context
*nvc0
, struct nvc0_query
*q
)
1141 struct nvc0_screen
*screen
= nvc0
->screen
;
1142 struct pipe_context
*pipe
= &nvc0
->base
.pipe
;
1143 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
1144 const bool is_nve4
= screen
->base
.class_3d
>= NVE4_3D_CLASS
;
1147 const uint block
[3] = { 32, is_nve4
? 4 : 1, 1 };
1148 const uint grid
[3] = { screen
->mp_count
, 1, 1 };
1150 const struct nvc0_hw_sm_query_cfg
*cfg
;
1152 cfg
= nvc0_hw_sm_query_get_cfg(nvc0
, q
);
1154 if (unlikely(!screen
->pm
.prog
)) {
1155 struct nvc0_program
*prog
= CALLOC_STRUCT(nvc0_program
);
1156 prog
->type
= PIPE_SHADER_COMPUTE
;
1157 prog
->translated
= true;
1158 prog
->num_gprs
= 14;
1159 prog
->parm_size
= 12;
1161 prog
->code
= (uint32_t *)nve4_read_hw_sm_counters_code
;
1162 prog
->code_size
= sizeof(nve4_read_hw_sm_counters_code
);
1164 prog
->code
= (uint32_t *)nvc0_read_hw_sm_counters_code
;
1165 prog
->code_size
= sizeof(nvc0_read_hw_sm_counters_code
);
1167 screen
->pm
.prog
= prog
;
1170 /* disable all counting */
1171 PUSH_SPACE(push
, 8);
1172 for (c
= 0; c
< 8; ++c
)
1173 if (screen
->pm
.mp_counter
[c
]) {
1175 IMMED_NVC0(push
, NVE4_COMPUTE(MP_PM_FUNC(c
)), 0);
1177 IMMED_NVC0(push
, NVC0_COMPUTE(MP_PM_OP(c
)), 0);
1180 /* release counters for this query */
1181 for (c
= 0; c
< 8; ++c
) {
1182 if (nvc0_query(screen
->pm
.mp_counter
[c
]) == q
) {
1183 screen
->pm
.num_hw_sm_active
[c
/ 4]--;
1184 screen
->pm
.mp_counter
[c
] = NULL
;
1188 BCTX_REFN_bo(nvc0
->bufctx_cp
, CP_QUERY
, NOUVEAU_BO_GART
| NOUVEAU_BO_WR
,
1191 PUSH_SPACE(push
, 1);
1192 IMMED_NVC0(push
, SUBC_COMPUTE(NV50_GRAPH_SERIALIZE
), 0);
1194 pipe
->bind_compute_state(pipe
, screen
->pm
.prog
);
1195 input
[0] = (q
->bo
->offset
+ q
->base
);
1196 input
[1] = (q
->bo
->offset
+ q
->base
) >> 32;
1197 input
[2] = q
->sequence
;
1198 pipe
->launch_grid(pipe
, block
, grid
, 0, input
);
1200 nouveau_bufctx_reset(nvc0
->bufctx_cp
, NVC0_BIND_CP_QUERY
);
1202 /* re-activate other counters */
1203 PUSH_SPACE(push
, 16);
1205 for (c
= 0; c
< 8; ++c
) {
1207 q
= nvc0_query(screen
->pm
.mp_counter
[c
]);
1210 cfg
= nvc0_hw_sm_query_get_cfg(nvc0
, q
);
1211 for (i
= 0; i
< cfg
->num_counters
; ++i
) {
1212 if (mask
& (1 << q
->ctr
[i
]))
1214 mask
|= 1 << q
->ctr
[i
];
1216 BEGIN_NVC0(push
, NVE4_COMPUTE(MP_PM_FUNC(q
->ctr
[i
])), 1);
1218 BEGIN_NVC0(push
, NVC0_COMPUTE(MP_PM_OP(q
->ctr
[i
])), 1);
1220 PUSH_DATA (push
, (cfg
->ctr
[i
].func
<< 4) | cfg
->ctr
[i
].mode
);
1226 nvc0_hw_sm_query_read_data(uint32_t count
[32][4],
1227 struct nvc0_context
*nvc0
, bool wait
,
1228 struct nvc0_query
*q
,
1229 const struct nvc0_hw_sm_query_cfg
*cfg
,
1234 for (p
= 0; p
< mp_count
; ++p
) {
1235 const unsigned b
= (0x24 / 4) * p
;
1237 for (c
= 0; c
< cfg
->num_counters
; ++c
) {
1238 if (q
->data
[b
+ 8] != q
->sequence
) {
1241 if (nouveau_bo_wait(q
->bo
, NOUVEAU_BO_RD
, nvc0
->base
.client
))
1244 count
[p
][c
] = q
->data
[b
+ q
->ctr
[c
]];
1251 nve4_hw_sm_query_read_data(uint32_t count
[32][4],
1252 struct nvc0_context
*nvc0
, bool wait
,
1253 struct nvc0_query
*q
,
1254 const struct nvc0_hw_sm_query_cfg
*cfg
,
1259 for (p
= 0; p
< mp_count
; ++p
) {
1260 const unsigned b
= (0x60 / 4) * p
;
1262 for (c
= 0; c
< cfg
->num_counters
; ++c
) {
1264 for (d
= 0; d
< ((q
->ctr
[c
] & ~3) ? 1 : 4); ++d
) {
1265 if (q
->data
[b
+ 20 + d
] != q
->sequence
) {
1268 if (nouveau_bo_wait(q
->bo
, NOUVEAU_BO_RD
, nvc0
->base
.client
))
1271 if (q
->ctr
[c
] & ~0x3)
1272 count
[p
][c
] = q
->data
[b
+ 16 + (q
->ctr
[c
] & 3)];
1274 count
[p
][c
] += q
->data
[b
+ d
* 4 + q
->ctr
[c
]];
1281 /* Metric calculations:
1282 * sum(x) ... sum of x over all MPs
1283 * avg(x) ... average of x over all MPs
1285 * IPC : sum(inst_executed) / clock
1286 * INST_REPLAY_OHEAD: (sum(inst_issued) - sum(inst_executed)) / sum(inst_issued)
1287 * MP_OCCUPANCY : avg((active_warps / 64) / active_cycles)
1288 * MP_EFFICIENCY : avg(active_cycles / clock)
1290 * NOTE: Interpretation of IPC requires knowledge of MP count.
1293 nvc0_hw_sm_query_result(struct nvc0_context
*nvc0
, struct nvc0_query
*q
,
1294 void *result
, boolean wait
)
1296 uint32_t count
[32][4];
1298 unsigned mp_count
= MIN2(nvc0
->screen
->mp_count_compute
, 32);
1300 const struct nvc0_hw_sm_query_cfg
*cfg
;
1303 cfg
= nvc0_hw_sm_query_get_cfg(nvc0
, q
);
1305 if (nvc0
->screen
->base
.class_3d
>= NVE4_3D_CLASS
)
1306 ret
= nve4_hw_sm_query_read_data(count
, nvc0
, wait
, q
, cfg
, mp_count
);
1308 ret
= nvc0_hw_sm_query_read_data(count
, nvc0
, wait
, q
, cfg
, mp_count
);
1312 if (cfg
->op
== NVC0_COUNTER_OPn_SUM
) {
1313 for (c
= 0; c
< cfg
->num_counters
; ++c
)
1314 for (p
= 0; p
< mp_count
; ++p
)
1315 value
+= count
[p
][c
];
1316 value
= (value
* cfg
->norm
[0]) / cfg
->norm
[1];
1318 if (cfg
->op
== NVC0_COUNTER_OPn_OR
) {
1320 for (c
= 0; c
< cfg
->num_counters
; ++c
)
1321 for (p
= 0; p
< mp_count
; ++p
)
1323 value
= ((uint64_t)v
* cfg
->norm
[0]) / cfg
->norm
[1];
1325 if (cfg
->op
== NVC0_COUNTER_OPn_AND
) {
1327 for (c
= 0; c
< cfg
->num_counters
; ++c
)
1328 for (p
= 0; p
< mp_count
; ++p
)
1330 value
= ((uint64_t)v
* cfg
->norm
[0]) / cfg
->norm
[1];
1332 if (cfg
->op
== NVC0_COUNTER_OP2_REL_SUM_MM
) {
1333 uint64_t v
[2] = { 0, 0 };
1334 for (p
= 0; p
< mp_count
; ++p
) {
1335 v
[0] += count
[p
][0];
1336 v
[1] += count
[p
][1];
1339 value
= ((v
[0] - v
[1]) * cfg
->norm
[0]) / (v
[0] * cfg
->norm
[1]);
1341 if (cfg
->op
== NVC0_COUNTER_OP2_DIV_SUM_M0
) {
1342 for (p
= 0; p
< mp_count
; ++p
)
1343 value
+= count
[p
][0];
1345 value
= (value
* cfg
->norm
[0]) / (count
[0][1] * cfg
->norm
[1]);
1349 if (cfg
->op
== NVC0_COUNTER_OP2_AVG_DIV_MM
) {
1350 unsigned mp_used
= 0;
1351 for (p
= 0; p
< mp_count
; ++p
, mp_used
+= !!count
[p
][0])
1353 value
+= (count
[p
][0] * cfg
->norm
[0]) / count
[p
][1];
1355 value
/= (uint64_t)mp_used
* cfg
->norm
[1];
1357 if (cfg
->op
== NVC0_COUNTER_OP2_AVG_DIV_M0
) {
1358 unsigned mp_used
= 0;
1359 for (p
= 0; p
< mp_count
; ++p
, mp_used
+= !!count
[p
][0])
1360 value
+= count
[p
][0];
1361 if (count
[0][1] && mp_used
) {
1362 value
*= cfg
->norm
[0];
1363 value
/= (uint64_t)count
[0][1] * mp_used
* cfg
->norm
[1];
1369 *(uint64_t *)result
= value
;
1374 nvc0_screen_get_driver_query_info(struct pipe_screen
*pscreen
,
1376 struct pipe_driver_query_info
*info
)
1378 struct nvc0_screen
*screen
= nvc0_screen(pscreen
);
1381 count
+= NVC0_QUERY_DRV_STAT_COUNT
;
1383 if (screen
->base
.device
->drm_version
>= 0x01000101) {
1384 if (screen
->compute
) {
1385 if (screen
->base
.class_3d
== NVE4_3D_CLASS
) {
1386 count
+= NVE4_HW_SM_QUERY_COUNT
;
1388 if (screen
->base
.class_3d
< NVE4_3D_CLASS
) {
1389 /* NVC0_COMPUTE is not always enabled */
1390 count
+= NVC0_HW_SM_QUERY_COUNT
;
1398 /* Init default values. */
1399 info
->name
= "this_is_not_the_query_you_are_looking_for";
1400 info
->query_type
= 0xdeadd01d;
1401 info
->max_value
.u64
= 0;
1402 info
->type
= PIPE_DRIVER_QUERY_TYPE_UINT64
;
1403 info
->group_id
= -1;
1405 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
1406 if (id
< NVC0_QUERY_DRV_STAT_COUNT
) {
1407 info
->name
= nvc0_drv_stat_names
[id
];
1408 info
->query_type
= NVC0_QUERY_DRV_STAT(id
);
1409 info
->max_value
.u64
= 0;
1410 if (strstr(info
->name
, "bytes"))
1411 info
->type
= PIPE_DRIVER_QUERY_TYPE_BYTES
;
1412 info
->group_id
= NVC0_QUERY_DRV_STAT_GROUP
;
1417 if (screen
->compute
) {
1418 if (screen
->base
.class_3d
== NVE4_3D_CLASS
) {
1419 info
->name
= nve4_pm_query_names
[id
- NVC0_QUERY_DRV_STAT_COUNT
];
1420 info
->query_type
= NVE4_HW_SM_QUERY(id
- NVC0_QUERY_DRV_STAT_COUNT
);
1421 info
->max_value
.u64
=
1422 (id
< NVE4_HW_SM_QUERY_METRIC_MP_OCCUPANCY
) ? 0 : 100;
1423 info
->group_id
= NVC0_QUERY_MP_COUNTER_GROUP
;
1426 if (screen
->base
.class_3d
< NVE4_3D_CLASS
) {
1427 info
->name
= nvc0_pm_query_names
[id
- NVC0_QUERY_DRV_STAT_COUNT
];
1428 info
->query_type
= NVC0_HW_SM_QUERY(id
- NVC0_QUERY_DRV_STAT_COUNT
);
1429 info
->group_id
= NVC0_QUERY_MP_COUNTER_GROUP
;
1434 /* user asked for info about non-existing query */
1439 nvc0_screen_get_driver_query_group_info(struct pipe_screen
*pscreen
,
1441 struct pipe_driver_query_group_info
*info
)
1443 struct nvc0_screen
*screen
= nvc0_screen(pscreen
);
1446 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
1450 if (screen
->base
.device
->drm_version
>= 0x01000101) {
1451 if (screen
->compute
) {
1452 if (screen
->base
.class_3d
== NVE4_3D_CLASS
) {
1455 if (screen
->base
.class_3d
< NVE4_3D_CLASS
) {
1456 count
++; /* NVC0_COMPUTE is not always enabled */
1464 if (id
== NVC0_QUERY_MP_COUNTER_GROUP
) {
1465 if (screen
->compute
) {
1466 info
->name
= "MP counters";
1467 info
->type
= PIPE_DRIVER_QUERY_GROUP_TYPE_GPU
;
1469 if (screen
->base
.class_3d
== NVE4_3D_CLASS
) {
1470 info
->num_queries
= NVE4_HW_SM_QUERY_COUNT
;
1472 /* On NVE4+, each multiprocessor have 8 hardware counters separated
1473 * in two distinct domains, but we allow only one active query
1474 * simultaneously because some of them use more than one hardware
1475 * counter and this will result in an undefined behaviour. */
1476 info
->max_active_queries
= 1; /* TODO: handle multiple hw counters */
1479 if (screen
->base
.class_3d
< NVE4_3D_CLASS
) {
1480 info
->num_queries
= NVC0_HW_SM_QUERY_COUNT
;
1482 /* On NVC0:NVE4, each multiprocessor have 8 hardware counters
1483 * in a single domain. */
1484 info
->max_active_queries
= 8;
1489 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
1490 else if (id
== NVC0_QUERY_DRV_STAT_GROUP
) {
1491 info
->name
= "Driver statistics";
1492 info
->type
= PIPE_DRIVER_QUERY_GROUP_TYPE_CPU
;
1493 info
->max_active_queries
= NVC0_QUERY_DRV_STAT_COUNT
;
1494 info
->num_queries
= NVC0_QUERY_DRV_STAT_COUNT
;
1499 /* user asked for info about non-existing query group */
1500 info
->name
= "this_is_not_the_query_group_you_are_looking_for";
1501 info
->max_active_queries
= 0;
1502 info
->num_queries
= 0;
1508 nvc0_init_query_functions(struct nvc0_context
*nvc0
)
1510 struct pipe_context
*pipe
= &nvc0
->base
.pipe
;
1512 pipe
->create_query
= nvc0_query_create
;
1513 pipe
->destroy_query
= nvc0_query_destroy
;
1514 pipe
->begin_query
= nvc0_query_begin
;
1515 pipe
->end_query
= nvc0_query_end
;
1516 pipe
->get_query_result
= nvc0_query_result
;
1517 pipe
->render_condition
= nvc0_render_condition
;