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"
29 #include "nvc0/nvc0_query_sw.h"
31 #include "nv_object.xml.h"
32 #include "nvc0/nve4_compute.xml.h"
33 #include "nvc0/nvc0_compute.xml.h"
35 #define NVC0_QUERY_STATE_READY 0
36 #define NVC0_QUERY_STATE_ACTIVE 1
37 #define NVC0_QUERY_STATE_ENDED 2
38 #define NVC0_QUERY_STATE_FLUSHED 3
40 #define NVC0_QUERY_ALLOC_SPACE 256
42 static boolean
nvc0_hw_sm_query_begin(struct nvc0_context
*,
44 static void nvc0_hw_sm_query_end(struct nvc0_context
*, struct nvc0_query
*);
45 static boolean
nvc0_hw_sm_query_result(struct nvc0_context
*,
46 struct nvc0_query
*, void *, boolean
);
49 nvc0_query_allocate(struct nvc0_context
*nvc0
, struct nvc0_query
*q
, int size
)
51 struct nvc0_screen
*screen
= nvc0
->screen
;
55 nouveau_bo_ref(NULL
, &q
->bo
);
57 if (q
->state
== NVC0_QUERY_STATE_READY
)
58 nouveau_mm_free(q
->mm
);
60 nouveau_fence_work(screen
->base
.fence
.current
,
61 nouveau_mm_free_work
, q
->mm
);
65 q
->mm
= nouveau_mm_allocate(screen
->base
.mm_GART
, size
, &q
->bo
, &q
->base
);
70 ret
= nouveau_bo_map(q
->bo
, 0, screen
->base
.client
);
72 nvc0_query_allocate(nvc0
, q
, 0);
75 q
->data
= (uint32_t *)((uint8_t *)q
->bo
->map
+ q
->base
);
81 nvc0_query_destroy(struct pipe_context
*pipe
, struct pipe_query
*pq
)
83 nvc0_query_allocate(nvc0_context(pipe
), nvc0_query(pq
), 0);
84 nouveau_fence_ref(NULL
, &nvc0_query(pq
)->fence
);
88 static struct pipe_query
*
89 nvc0_query_create(struct pipe_context
*pipe
, unsigned type
, unsigned index
)
91 struct nvc0_context
*nvc0
= nvc0_context(pipe
);
93 unsigned space
= NVC0_QUERY_ALLOC_SPACE
;
95 q
= nvc0_sw_create_query(nvc0
, type
, index
);
97 return (struct pipe_query
*)q
;
99 q
= CALLOC_STRUCT(nvc0_query
);
104 case PIPE_QUERY_OCCLUSION_COUNTER
:
105 case PIPE_QUERY_OCCLUSION_PREDICATE
:
107 space
= NVC0_QUERY_ALLOC_SPACE
;
109 case PIPE_QUERY_PIPELINE_STATISTICS
:
113 case PIPE_QUERY_SO_STATISTICS
:
114 case PIPE_QUERY_SO_OVERFLOW_PREDICATE
:
118 case PIPE_QUERY_PRIMITIVES_GENERATED
:
119 case PIPE_QUERY_PRIMITIVES_EMITTED
:
124 case PIPE_QUERY_TIME_ELAPSED
:
125 case PIPE_QUERY_TIMESTAMP
:
126 case PIPE_QUERY_TIMESTAMP_DISJOINT
:
127 case PIPE_QUERY_GPU_FINISHED
:
130 case NVC0_QUERY_TFB_BUFFER_OFFSET
:
134 if (nvc0
->screen
->base
.device
->drm_version
>= 0x01000101) {
135 if (type
>= NVE4_HW_SM_QUERY(0) && type
<= NVE4_HW_SM_QUERY_LAST
) {
157 * [50] = WS0.sequence
158 * [54] = WS1.sequence
159 * [58] = WS2.sequence
160 * [5c] = WS3.sequence
162 space
= (4 * 4 + 4 + 4) * nvc0
->screen
->mp_count
* sizeof(uint32_t);
165 if (type
>= NVC0_HW_SM_QUERY(0) && type
<= NVC0_HW_SM_QUERY_LAST
) {
177 space
= (8 + 1) * nvc0
->screen
->mp_count
* sizeof(uint32_t);
181 debug_printf("invalid query type: %u\n", type
);
185 if (!nvc0_query_allocate(nvc0
, q
, space
)) {
193 /* we advance before query_begin ! */
194 q
->offset
-= q
->rotate
;
195 q
->data
-= q
->rotate
/ sizeof(*q
->data
);
198 q
->data
[0] = 0; /* initialize sequence */
200 return (struct pipe_query
*)q
;
204 nvc0_query_get(struct nouveau_pushbuf
*push
, struct nvc0_query
*q
,
205 unsigned offset
, uint32_t get
)
210 PUSH_REFN (push
, q
->bo
, NOUVEAU_BO_GART
| NOUVEAU_BO_WR
);
211 BEGIN_NVC0(push
, NVC0_3D(QUERY_ADDRESS_HIGH
), 4);
212 PUSH_DATAh(push
, q
->bo
->offset
+ offset
);
213 PUSH_DATA (push
, q
->bo
->offset
+ offset
);
214 PUSH_DATA (push
, q
->sequence
);
215 PUSH_DATA (push
, get
);
219 nvc0_query_rotate(struct nvc0_context
*nvc0
, struct nvc0_query
*q
)
221 q
->offset
+= q
->rotate
;
222 q
->data
+= q
->rotate
/ sizeof(*q
->data
);
223 if (q
->offset
- q
->base
== NVC0_QUERY_ALLOC_SPACE
)
224 nvc0_query_allocate(nvc0
, q
, NVC0_QUERY_ALLOC_SPACE
);
228 nvc0_query_begin(struct pipe_context
*pipe
, struct pipe_query
*pq
)
230 struct nvc0_context
*nvc0
= nvc0_context(pipe
);
231 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
232 struct nvc0_query
*q
= nvc0_query(pq
);
235 /* For occlusion queries we have to change the storage, because a previous
236 * query might set the initial render conition to false even *after* we re-
237 * initialized it to true.
240 nvc0_query_rotate(nvc0
, q
);
242 /* XXX: can we do this with the GPU, and sync with respect to a previous
245 q
->data
[0] = q
->sequence
; /* initialize sequence */
246 q
->data
[1] = 1; /* initial render condition = true */
247 q
->data
[4] = q
->sequence
+ 1; /* for comparison COND_MODE */
253 case PIPE_QUERY_OCCLUSION_COUNTER
:
254 case PIPE_QUERY_OCCLUSION_PREDICATE
:
255 q
->nesting
= nvc0
->screen
->num_occlusion_queries_active
++;
257 nvc0_query_get(push
, q
, 0x10, 0x0100f002);
260 BEGIN_NVC0(push
, NVC0_3D(COUNTER_RESET
), 1);
261 PUSH_DATA (push
, NVC0_3D_COUNTER_RESET_SAMPLECNT
);
262 IMMED_NVC0(push
, NVC0_3D(SAMPLECNT_ENABLE
), 1);
265 case PIPE_QUERY_PRIMITIVES_GENERATED
:
266 nvc0_query_get(push
, q
, 0x10, 0x09005002 | (q
->index
<< 5));
268 case PIPE_QUERY_PRIMITIVES_EMITTED
:
269 nvc0_query_get(push
, q
, 0x10, 0x05805002 | (q
->index
<< 5));
271 case PIPE_QUERY_SO_STATISTICS
:
272 nvc0_query_get(push
, q
, 0x20, 0x05805002 | (q
->index
<< 5));
273 nvc0_query_get(push
, q
, 0x30, 0x06805002 | (q
->index
<< 5));
275 case PIPE_QUERY_SO_OVERFLOW_PREDICATE
:
276 nvc0_query_get(push
, q
, 0x10, 0x03005002 | (q
->index
<< 5));
278 case PIPE_QUERY_TIME_ELAPSED
:
279 nvc0_query_get(push
, q
, 0x10, 0x00005002);
281 case PIPE_QUERY_PIPELINE_STATISTICS
:
282 nvc0_query_get(push
, q
, 0xc0 + 0x00, 0x00801002); /* VFETCH, VERTICES */
283 nvc0_query_get(push
, q
, 0xc0 + 0x10, 0x01801002); /* VFETCH, PRIMS */
284 nvc0_query_get(push
, q
, 0xc0 + 0x20, 0x02802002); /* VP, LAUNCHES */
285 nvc0_query_get(push
, q
, 0xc0 + 0x30, 0x03806002); /* GP, LAUNCHES */
286 nvc0_query_get(push
, q
, 0xc0 + 0x40, 0x04806002); /* GP, PRIMS_OUT */
287 nvc0_query_get(push
, q
, 0xc0 + 0x50, 0x07804002); /* RAST, PRIMS_IN */
288 nvc0_query_get(push
, q
, 0xc0 + 0x60, 0x08804002); /* RAST, PRIMS_OUT */
289 nvc0_query_get(push
, q
, 0xc0 + 0x70, 0x0980a002); /* ROP, PIXELS */
290 nvc0_query_get(push
, q
, 0xc0 + 0x80, 0x0d808002); /* TCP, LAUNCHES */
291 nvc0_query_get(push
, q
, 0xc0 + 0x90, 0x0e809002); /* TEP, LAUNCHES */
294 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
295 if (q
->type
>= NVC0_SW_QUERY_DRV_STAT(0) &&
296 q
->type
<= NVC0_SW_QUERY_DRV_STAT_LAST
) {
297 return q
->funcs
->begin_query(nvc0
, q
);
300 if ((q
->type
>= NVE4_HW_SM_QUERY(0) && q
->type
<= NVE4_HW_SM_QUERY_LAST
) ||
301 (q
->type
>= NVC0_HW_SM_QUERY(0) && q
->type
<= NVC0_HW_SM_QUERY_LAST
)) {
302 ret
= nvc0_hw_sm_query_begin(nvc0
, q
);
306 q
->state
= NVC0_QUERY_STATE_ACTIVE
;
311 nvc0_query_end(struct pipe_context
*pipe
, struct pipe_query
*pq
)
313 struct nvc0_context
*nvc0
= nvc0_context(pipe
);
314 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
315 struct nvc0_query
*q
= nvc0_query(pq
);
317 if (q
->state
!= NVC0_QUERY_STATE_ACTIVE
) {
318 /* some queries don't require 'begin' to be called (e.g. GPU_FINISHED) */
320 nvc0_query_rotate(nvc0
, q
);
323 q
->state
= NVC0_QUERY_STATE_ENDED
;
326 case PIPE_QUERY_OCCLUSION_COUNTER
:
327 case PIPE_QUERY_OCCLUSION_PREDICATE
:
328 nvc0_query_get(push
, q
, 0, 0x0100f002);
329 if (--nvc0
->screen
->num_occlusion_queries_active
== 0) {
331 IMMED_NVC0(push
, NVC0_3D(SAMPLECNT_ENABLE
), 0);
334 case PIPE_QUERY_PRIMITIVES_GENERATED
:
335 nvc0_query_get(push
, q
, 0, 0x09005002 | (q
->index
<< 5));
337 case PIPE_QUERY_PRIMITIVES_EMITTED
:
338 nvc0_query_get(push
, q
, 0, 0x05805002 | (q
->index
<< 5));
340 case PIPE_QUERY_SO_STATISTICS
:
341 nvc0_query_get(push
, q
, 0x00, 0x05805002 | (q
->index
<< 5));
342 nvc0_query_get(push
, q
, 0x10, 0x06805002 | (q
->index
<< 5));
344 case PIPE_QUERY_SO_OVERFLOW_PREDICATE
:
345 /* TODO: How do we sum over all streams for render condition ? */
346 /* PRIMS_DROPPED doesn't write sequence, use a ZERO query to sync on */
347 nvc0_query_get(push
, q
, 0x00, 0x03005002 | (q
->index
<< 5));
348 nvc0_query_get(push
, q
, 0x20, 0x00005002);
350 case PIPE_QUERY_TIMESTAMP
:
351 case PIPE_QUERY_TIME_ELAPSED
:
352 nvc0_query_get(push
, q
, 0, 0x00005002);
354 case PIPE_QUERY_GPU_FINISHED
:
355 nvc0_query_get(push
, q
, 0, 0x1000f010);
357 case PIPE_QUERY_PIPELINE_STATISTICS
:
358 nvc0_query_get(push
, q
, 0x00, 0x00801002); /* VFETCH, VERTICES */
359 nvc0_query_get(push
, q
, 0x10, 0x01801002); /* VFETCH, PRIMS */
360 nvc0_query_get(push
, q
, 0x20, 0x02802002); /* VP, LAUNCHES */
361 nvc0_query_get(push
, q
, 0x30, 0x03806002); /* GP, LAUNCHES */
362 nvc0_query_get(push
, q
, 0x40, 0x04806002); /* GP, PRIMS_OUT */
363 nvc0_query_get(push
, q
, 0x50, 0x07804002); /* RAST, PRIMS_IN */
364 nvc0_query_get(push
, q
, 0x60, 0x08804002); /* RAST, PRIMS_OUT */
365 nvc0_query_get(push
, q
, 0x70, 0x0980a002); /* ROP, PIXELS */
366 nvc0_query_get(push
, q
, 0x80, 0x0d808002); /* TCP, LAUNCHES */
367 nvc0_query_get(push
, q
, 0x90, 0x0e809002); /* TEP, LAUNCHES */
369 case NVC0_QUERY_TFB_BUFFER_OFFSET
:
370 /* indexed by TFB buffer instead of by vertex stream */
371 nvc0_query_get(push
, q
, 0x00, 0x0d005002 | (q
->index
<< 5));
373 case PIPE_QUERY_TIMESTAMP_DISJOINT
:
374 /* This query is not issued on GPU because disjoint is forced to false */
375 q
->state
= NVC0_QUERY_STATE_READY
;
378 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
379 if (q
->type
>= NVC0_SW_QUERY_DRV_STAT(0) &&
380 q
->type
<= NVC0_SW_QUERY_DRV_STAT_LAST
) {
381 q
->funcs
->end_query(nvc0
, q
);
385 if ((q
->type
>= NVE4_HW_SM_QUERY(0) && q
->type
<= NVE4_HW_SM_QUERY_LAST
) ||
386 (q
->type
>= NVC0_HW_SM_QUERY(0) && q
->type
<= NVC0_HW_SM_QUERY_LAST
)) {
387 nvc0_hw_sm_query_end(nvc0
, q
);
392 nouveau_fence_ref(nvc0
->screen
->base
.fence
.current
, &q
->fence
);
396 nvc0_query_update(struct nouveau_client
*cli
, struct nvc0_query
*q
)
399 if (nouveau_fence_signalled(q
->fence
))
400 q
->state
= NVC0_QUERY_STATE_READY
;
402 if (q
->data
[0] == q
->sequence
)
403 q
->state
= NVC0_QUERY_STATE_READY
;
408 nvc0_query_result(struct pipe_context
*pipe
, struct pipe_query
*pq
,
409 boolean wait
, union pipe_query_result
*result
)
411 struct nvc0_context
*nvc0
= nvc0_context(pipe
);
412 struct nvc0_query
*q
= nvc0_query(pq
);
413 uint64_t *res64
= (uint64_t*)result
;
414 uint32_t *res32
= (uint32_t*)result
;
415 uint8_t *res8
= (uint8_t*)result
;
416 uint64_t *data64
= (uint64_t *)q
->data
;
419 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
420 if (q
->type
>= NVC0_SW_QUERY_DRV_STAT(0) &&
421 q
->type
<= NVC0_SW_QUERY_DRV_STAT_LAST
) {
422 return q
->funcs
->get_query_result(nvc0
, q
, wait
, result
);
425 if ((q
->type
>= NVE4_HW_SM_QUERY(0) && q
->type
<= NVE4_HW_SM_QUERY_LAST
) ||
426 (q
->type
>= NVC0_HW_SM_QUERY(0) && q
->type
<= NVC0_HW_SM_QUERY_LAST
)) {
427 return nvc0_hw_sm_query_result(nvc0
, q
, result
, wait
);
430 if (q
->state
!= NVC0_QUERY_STATE_READY
)
431 nvc0_query_update(nvc0
->screen
->base
.client
, q
);
433 if (q
->state
!= NVC0_QUERY_STATE_READY
) {
435 if (q
->state
!= NVC0_QUERY_STATE_FLUSHED
) {
436 q
->state
= NVC0_QUERY_STATE_FLUSHED
;
437 /* flush for silly apps that spin on GL_QUERY_RESULT_AVAILABLE */
438 PUSH_KICK(nvc0
->base
.pushbuf
);
442 if (nouveau_bo_wait(q
->bo
, NOUVEAU_BO_RD
, nvc0
->screen
->base
.client
))
444 NOUVEAU_DRV_STAT(&nvc0
->screen
->base
, query_sync_count
, 1);
446 q
->state
= NVC0_QUERY_STATE_READY
;
449 case PIPE_QUERY_GPU_FINISHED
:
452 case PIPE_QUERY_OCCLUSION_COUNTER
: /* u32 sequence, u32 count, u64 time */
453 res64
[0] = q
->data
[1] - q
->data
[5];
455 case PIPE_QUERY_OCCLUSION_PREDICATE
:
456 res8
[0] = q
->data
[1] != q
->data
[5];
458 case PIPE_QUERY_PRIMITIVES_GENERATED
: /* u64 count, u64 time */
459 case PIPE_QUERY_PRIMITIVES_EMITTED
: /* u64 count, u64 time */
460 res64
[0] = data64
[0] - data64
[2];
462 case PIPE_QUERY_SO_STATISTICS
:
463 res64
[0] = data64
[0] - data64
[4];
464 res64
[1] = data64
[2] - data64
[6];
466 case PIPE_QUERY_SO_OVERFLOW_PREDICATE
:
467 res8
[0] = data64
[0] != data64
[2];
469 case PIPE_QUERY_TIMESTAMP
:
470 res64
[0] = data64
[1];
472 case PIPE_QUERY_TIMESTAMP_DISJOINT
:
473 res64
[0] = 1000000000;
476 case PIPE_QUERY_TIME_ELAPSED
:
477 res64
[0] = data64
[1] - data64
[3];
479 case PIPE_QUERY_PIPELINE_STATISTICS
:
480 for (i
= 0; i
< 10; ++i
)
481 res64
[i
] = data64
[i
* 2] - data64
[24 + i
* 2];
483 case NVC0_QUERY_TFB_BUFFER_OFFSET
:
484 res32
[0] = q
->data
[1];
487 assert(0); /* can't happen, we don't create queries with invalid type */
495 nvc0_query_fifo_wait(struct nouveau_pushbuf
*push
, struct nvc0_query
*q
)
497 unsigned offset
= q
->offset
;
499 if (q
->type
== PIPE_QUERY_SO_OVERFLOW_PREDICATE
) offset
+= 0x20;
502 PUSH_REFN (push
, q
->bo
, NOUVEAU_BO_GART
| NOUVEAU_BO_RD
);
503 BEGIN_NVC0(push
, SUBC_3D(NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH
), 4);
504 PUSH_DATAh(push
, q
->bo
->offset
+ offset
);
505 PUSH_DATA (push
, q
->bo
->offset
+ offset
);
506 PUSH_DATA (push
, q
->sequence
);
507 PUSH_DATA (push
, (1 << 12) |
508 NV84_SUBCHAN_SEMAPHORE_TRIGGER_ACQUIRE_EQUAL
);
512 nvc0_render_condition(struct pipe_context
*pipe
,
513 struct pipe_query
*pq
,
514 boolean condition
, uint mode
)
516 struct nvc0_context
*nvc0
= nvc0_context(pipe
);
517 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
518 struct nvc0_query
*q
;
521 mode
!= PIPE_RENDER_COND_NO_WAIT
&&
522 mode
!= PIPE_RENDER_COND_BY_REGION_NO_WAIT
;
525 cond
= NVC0_3D_COND_MODE_ALWAYS
;
529 /* NOTE: comparison of 2 queries only works if both have completed */
531 case PIPE_QUERY_SO_OVERFLOW_PREDICATE
:
532 cond
= condition
? NVC0_3D_COND_MODE_EQUAL
:
533 NVC0_3D_COND_MODE_NOT_EQUAL
;
536 case PIPE_QUERY_OCCLUSION_COUNTER
:
537 case PIPE_QUERY_OCCLUSION_PREDICATE
:
538 if (likely(!condition
)) {
539 if (unlikely(q
->nesting
))
540 cond
= wait
? NVC0_3D_COND_MODE_NOT_EQUAL
:
541 NVC0_3D_COND_MODE_ALWAYS
;
543 cond
= NVC0_3D_COND_MODE_RES_NON_ZERO
;
545 cond
= wait
? NVC0_3D_COND_MODE_EQUAL
: NVC0_3D_COND_MODE_ALWAYS
;
549 assert(!"render condition query not a predicate");
550 cond
= NVC0_3D_COND_MODE_ALWAYS
;
555 nvc0
->cond_query
= pq
;
556 nvc0
->cond_cond
= condition
;
557 nvc0
->cond_condmode
= cond
;
558 nvc0
->cond_mode
= mode
;
562 IMMED_NVC0(push
, NVC0_3D(COND_MODE
), cond
);
567 nvc0_query_fifo_wait(push
, q
);
570 PUSH_REFN (push
, q
->bo
, NOUVEAU_BO_GART
| NOUVEAU_BO_RD
);
571 BEGIN_NVC0(push
, NVC0_3D(COND_ADDRESS_HIGH
), 3);
572 PUSH_DATAh(push
, q
->bo
->offset
+ q
->offset
);
573 PUSH_DATA (push
, q
->bo
->offset
+ q
->offset
);
574 PUSH_DATA (push
, cond
);
575 BEGIN_NVC0(push
, NVC0_2D(COND_ADDRESS_HIGH
), 2);
576 PUSH_DATAh(push
, q
->bo
->offset
+ q
->offset
);
577 PUSH_DATA (push
, q
->bo
->offset
+ q
->offset
);
581 nvc0_query_pushbuf_submit(struct nouveau_pushbuf
*push
,
582 struct nvc0_query
*q
, unsigned result_offset
)
584 #define NVC0_IB_ENTRY_1_NO_PREFETCH (1 << (31 - 8))
586 PUSH_REFN(push
, q
->bo
, NOUVEAU_BO_RD
| NOUVEAU_BO_GART
);
587 nouveau_pushbuf_space(push
, 0, 0, 1);
588 nouveau_pushbuf_data(push
, q
->bo
, q
->offset
+ result_offset
, 4 |
589 NVC0_IB_ENTRY_1_NO_PREFETCH
);
592 /* === DRIVER STATISTICS === */
594 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
596 static const char *nvc0_sw_query_drv_stat_names
[] =
598 "drv-tex_obj_current_count",
599 "drv-tex_obj_current_bytes",
600 "drv-buf_obj_current_count",
601 "drv-buf_obj_current_bytes_vid",
602 "drv-buf_obj_current_bytes_sys",
603 "drv-tex_transfers_rd",
604 "drv-tex_transfers_wr",
605 "drv-tex_copy_count",
606 "drv-tex_blit_count",
607 "drv-tex_cache_flush_count",
608 "drv-buf_transfers_rd",
609 "drv-buf_transfers_wr",
610 "drv-buf_read_bytes_staging_vid",
611 "drv-buf_write_bytes_direct",
612 "drv-buf_write_bytes_staging_vid",
613 "drv-buf_write_bytes_staging_sys",
614 "drv-buf_copy_bytes",
615 "drv-buf_non_kernel_fence_sync_count",
616 "drv-any_non_kernel_fence_sync_count",
617 "drv-query_sync_count",
618 "drv-gpu_serialize_count",
619 "drv-draw_calls_array",
620 "drv-draw_calls_indexed",
621 "drv-draw_calls_fallback_count",
622 "drv-user_buffer_upload_bytes",
623 "drv-constbuf_upload_count",
624 "drv-constbuf_upload_bytes",
626 "drv-resource_validate_count"
629 #endif /* NOUVEAU_ENABLE_DRIVER_STATISTICS */
632 /* === PERFORMANCE MONITORING COUNTERS for NVE4+ === */
634 /* Code to read out MP counters: They are accessible via mmio, too, but let's
635 * just avoid mapping registers in userspace. We'd have to know which MPs are
636 * enabled/present, too, and that information is not presently exposed.
637 * We could add a kernel interface for it, but reading the counters like this
638 * has the advantage of being async (if get_result isn't called immediately).
640 static const uint64_t nve4_read_hw_sm_counters_code
[] =
642 /* sched 0x20 0x20 0x20 0x20 0x20 0x20 0x20
644 * mov b32 $r12 $physid
650 * sched 0x20 0x20 0x23 0x04 0x20 0x04 0x2b
654 * set $p0 0x1 eq u32 $r8 0x0
655 * mov b32 $r10 c0[0x0]
656 * ext u32 $r8 $r12 0x414
657 * mov b32 $r11 c0[0x4]
658 * sched 0x04 0x2e 0x04 0x20 0x20 0x28 0x04
659 * ext u32 $r9 $r12 0x208
661 * set $p1 0x1 eq u32 $r9 0x0
662 * mul $r8 u32 $r8 u32 96
663 * mul $r12 u32 $r9 u32 16
664 * mul $r13 u32 $r9 u32 4
665 * add b32 $r9 $r8 $r13
666 * sched 0x28 0x04 0x2c 0x04 0x2c 0x04 0x2c
667 * add b32 $r8 $r8 $r12
669 * add b32 $r10 $c $r10 $r8
671 * add b32 $r11 $r11 0x0 $c
672 * add b32 $r12 $c $r12 $r9
673 * st b128 wt g[$r10d] $r0q
674 * sched 0x4 0x2c 0x20 0x04 0x2e 0x00 0x00
675 * mov b32 $r0 c0[0x8]
676 * add b32 $r13 $r13 0x0 $c
677 * $p1 st b128 wt g[$r12d+0x40] $r4q
678 * st b32 wt g[$r12d+0x50] $r0
680 0x2202020202020207ULL
,
681 0x2c00000084021c04ULL
,
682 0x2c0000000c031c04ULL
,
683 0x2c00000010001c04ULL
,
684 0x2c00000014005c04ULL
,
685 0x2c00000018009c04ULL
,
686 0x2c0000001c00dc04ULL
,
687 0x2c00000020011c04ULL
,
688 0x22b0420042320207ULL
,
689 0x2c00000024015c04ULL
,
690 0x2c00000028019c04ULL
,
691 0x2c0000002c01dc04ULL
,
692 0x190e0000fc81dc03ULL
,
693 0x2800400000029de4ULL
,
694 0x7000c01050c21c03ULL
,
695 0x280040001002dde4ULL
,
696 0x204282020042e047ULL
,
697 0x7000c00820c25c03ULL
,
698 0x80000000000021e7ULL
,
699 0x190e0000fc93dc03ULL
,
700 0x1000000180821c02ULL
,
701 0x1000000040931c02ULL
,
702 0x1000000010935c02ULL
,
703 0x4800000034825c03ULL
,
704 0x22c042c042c04287ULL
,
705 0x4800000030821c03ULL
,
706 0x2800000028031de4ULL
,
707 0x4801000020a29c03ULL
,
708 0x280000002c035de4ULL
,
709 0x0800000000b2dc42ULL
,
710 0x4801000024c31c03ULL
,
711 0x9400000000a01fc5ULL
,
712 0x200002e04202c047ULL
,
713 0x2800400020001de4ULL
,
714 0x0800000000d35c42ULL
,
715 0x9400000100c107c5ULL
,
716 0x9400000140c01f85ULL
,
717 0x8000000000001de7ULL
720 /* NOTE: intentionally using the same names as NV */
721 static const char *nve4_pm_query_names
[] =
730 "global_ld_mem_divergence_replays",
731 "global_store_transaction",
732 "global_st_mem_divergence_replays",
739 "l1_global_load_hit",
740 "l1_global_load_miss",
742 "l1_local_load_miss",
743 "l1_local_store_hit",
744 "l1_local_store_miss",
745 "l1_shared_load_transactions",
746 "l1_shared_store_transactions",
748 "local_load_transactions",
750 "local_store_transactions",
760 "shared_load_replay",
762 "shared_store_replay",
765 "uncached_global_load_transaction",
767 /* metrics, i.e. functions of the MP counters */
768 "metric-ipc", /* inst_executed, clock */
769 "metric-ipac", /* inst_executed, active_cycles */
770 "metric-ipec", /* inst_executed, (bool)inst_executed */
771 "metric-achieved_occupancy", /* active_warps, active_cycles */
772 "metric-sm_efficiency", /* active_cycles, clock */
773 "metric-inst_replay_overhead" /* inst_issued, inst_executed */
776 /* For simplicity, we will allocate as many group slots as we allocate counter
777 * slots. This means that a single counter which wants to source from 2 groups
778 * will have to be declared as using 2 counter slots. This shouldn't really be
779 * a problem because such queries don't make much sense ... (unless someone is
782 struct nvc0_mp_counter_cfg
784 uint32_t func
: 16; /* mask or 4-bit logic op (depending on mode) */
785 uint32_t mode
: 4; /* LOGOP,B6,LOGOP_B6(_PULSE) */
786 uint32_t num_src
: 3; /* number of sources (1 - 6, only for NVC0:NVE4) */
787 uint32_t sig_dom
: 1; /* if 0, MP_PM_A (per warp-sched), if 1, MP_PM_B */
788 uint32_t sig_sel
: 8; /* signal group */
789 uint64_t src_sel
; /* signal selection for up to 6 sources (48 bit) */
792 #define NVC0_COUNTER_OPn_SUM 0
793 #define NVC0_COUNTER_OPn_OR 1
794 #define NVC0_COUNTER_OPn_AND 2
795 #define NVC0_COUNTER_OP2_REL_SUM_MM 3 /* (sum(ctr0) - sum(ctr1)) / sum(ctr0) */
796 #define NVC0_COUNTER_OP2_DIV_SUM_M0 4 /* sum(ctr0) / ctr1 of MP[0]) */
797 #define NVC0_COUNTER_OP2_AVG_DIV_MM 5 /* avg(ctr0 / ctr1) */
798 #define NVC0_COUNTER_OP2_AVG_DIV_M0 6 /* avg(ctr0) / ctr1 of MP[0]) */
800 struct nvc0_hw_sm_query_cfg
802 struct nvc0_mp_counter_cfg ctr
[4];
803 uint8_t num_counters
;
805 uint8_t norm
[2]; /* normalization num,denom */
808 #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 } }
809 #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 } }
810 #define _M2A(n, f0, m0, g0, s0, f1, m1, g1, s1, o, nu, dn) [NVE4_HW_SM_QUERY_METRIC_##n] = { { \
811 { f0, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m0, 0, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g0, s0 }, \
812 { f1, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m1, 0, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g1, s1 }, \
813 {}, {}, }, 2, NVC0_COUNTER_OP2_##o, { nu, dn } }
814 #define _M2B(n, f0, m0, g0, s0, f1, m1, g1, s1, o, nu, dn) [NVE4_HW_SM_QUERY_METRIC_##n] = { { \
815 { f0, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m0, 0, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g0, s0 }, \
816 { f1, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m1, 0, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g1, s1 }, \
817 {}, {}, }, 2, NVC0_COUNTER_OP2_##o, { nu, dn } }
818 #define _M2AB(n, f0, m0, g0, s0, f1, m1, g1, s1, o, nu, dn) [NVE4_HW_SM_QUERY_METRIC_##n] = { { \
819 { f0, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m0, 0, 0, NVE4_COMPUTE_MP_PM_A_SIGSEL_##g0, s0 }, \
820 { f1, NVE4_COMPUTE_MP_PM_FUNC_MODE_##m1, 0, 1, NVE4_COMPUTE_MP_PM_B_SIGSEL_##g1, s1 }, \
821 {}, {}, }, 2, NVC0_COUNTER_OP2_##o, { nu, dn } }
824 * active_warps: bit 0 alternates btw 0 and 1 for odd nr of warps
825 * inst_executed etc.: we only count a single warp scheduler
826 * metric-ipXc: we simply multiply by 4 to account for the 4 warp schedulers;
827 * this is inaccurate !
829 static const struct nvc0_hw_sm_query_cfg nve4_hw_sm_queries
[] =
831 _Q1B(ACTIVE_CYCLES
, 0x0001, B6
, WARP
, 0x00000000, 1, 1),
832 _Q1B(ACTIVE_WARPS
, 0x003f, B6
, WARP
, 0x31483104, 2, 1),
833 _Q1A(ATOM_COUNT
, 0x0001, B6
, BRANCH
, 0x00000000, 1, 1),
834 _Q1A(BRANCH
, 0x0001, B6
, BRANCH
, 0x0000000c, 1, 1),
835 _Q1A(DIVERGENT_BRANCH
, 0x0001, B6
, BRANCH
, 0x00000010, 1, 1),
836 _Q1A(GLD_REQUEST
, 0x0001, B6
, LDST
, 0x00000010, 1, 1),
837 _Q1B(GLD_MEM_DIV_REPLAY
, 0x0001, B6
, REPLAY
, 0x00000010, 1, 1),
838 _Q1B(GST_TRANSACTIONS
, 0x0001, B6
, MEM
, 0x00000004, 1, 1),
839 _Q1B(GST_MEM_DIV_REPLAY
, 0x0001, B6
, REPLAY
, 0x00000014, 1, 1),
840 _Q1A(GRED_COUNT
, 0x0001, B6
, BRANCH
, 0x00000008, 1, 1),
841 _Q1A(GST_REQUEST
, 0x0001, B6
, LDST
, 0x00000014, 1, 1),
842 _Q1A(INST_EXECUTED
, 0x0003, B6
, EXEC
, 0x00000398, 1, 1),
843 _Q1A(INST_ISSUED
, 0x0003, B6
, ISSUE
, 0x00000104, 1, 1),
844 _Q1A(INST_ISSUED1
, 0x0001, B6
, ISSUE
, 0x00000004, 1, 1),
845 _Q1A(INST_ISSUED2
, 0x0001, B6
, ISSUE
, 0x00000008, 1, 1),
846 _Q1B(L1_GLD_HIT
, 0x0001, B6
, L1
, 0x00000010, 1, 1),
847 _Q1B(L1_GLD_MISS
, 0x0001, B6
, L1
, 0x00000014, 1, 1),
848 _Q1B(L1_LOCAL_LD_HIT
, 0x0001, B6
, L1
, 0x00000000, 1, 1),
849 _Q1B(L1_LOCAL_LD_MISS
, 0x0001, B6
, L1
, 0x00000004, 1, 1),
850 _Q1B(L1_LOCAL_ST_HIT
, 0x0001, B6
, L1
, 0x00000008, 1, 1),
851 _Q1B(L1_LOCAL_ST_MISS
, 0x0001, B6
, L1
, 0x0000000c, 1, 1),
852 _Q1B(L1_SHARED_LD_TRANSACTIONS
, 0x0001, B6
, TRANSACTION
, 0x00000008, 1, 1),
853 _Q1B(L1_SHARED_ST_TRANSACTIONS
, 0x0001, B6
, TRANSACTION
, 0x0000000c, 1, 1),
854 _Q1A(LOCAL_LD
, 0x0001, B6
, LDST
, 0x00000008, 1, 1),
855 _Q1B(LOCAL_LD_TRANSACTIONS
, 0x0001, B6
, TRANSACTION
, 0x00000000, 1, 1),
856 _Q1A(LOCAL_ST
, 0x0001, B6
, LDST
, 0x0000000c, 1, 1),
857 _Q1B(LOCAL_ST_TRANSACTIONS
, 0x0001, B6
, TRANSACTION
, 0x00000004, 1, 1),
858 _Q1A(PROF_TRIGGER_0
, 0x0001, B6
, USER
, 0x00000000, 1, 1),
859 _Q1A(PROF_TRIGGER_1
, 0x0001, B6
, USER
, 0x00000004, 1, 1),
860 _Q1A(PROF_TRIGGER_2
, 0x0001, B6
, USER
, 0x00000008, 1, 1),
861 _Q1A(PROF_TRIGGER_3
, 0x0001, B6
, USER
, 0x0000000c, 1, 1),
862 _Q1A(PROF_TRIGGER_4
, 0x0001, B6
, USER
, 0x00000010, 1, 1),
863 _Q1A(PROF_TRIGGER_5
, 0x0001, B6
, USER
, 0x00000014, 1, 1),
864 _Q1A(PROF_TRIGGER_6
, 0x0001, B6
, USER
, 0x00000018, 1, 1),
865 _Q1A(PROF_TRIGGER_7
, 0x0001, B6
, USER
, 0x0000001c, 1, 1),
866 _Q1A(SHARED_LD
, 0x0001, B6
, LDST
, 0x00000000, 1, 1),
867 _Q1B(SHARED_LD_REPLAY
, 0x0001, B6
, REPLAY
, 0x00000008, 1, 1),
868 _Q1A(SHARED_ST
, 0x0001, B6
, LDST
, 0x00000004, 1, 1),
869 _Q1B(SHARED_ST_REPLAY
, 0x0001, B6
, REPLAY
, 0x0000000c, 1, 1),
870 _Q1B(SM_CTA_LAUNCHED
, 0x0001, B6
, WARP
, 0x0000001c, 1, 1),
871 _Q1A(THREADS_LAUNCHED
, 0x003f, B6
, LAUNCH
, 0x398a4188, 1, 1),
872 _Q1B(UNCACHED_GLD_TRANSACTIONS
, 0x0001, B6
, MEM
, 0x00000000, 1, 1),
873 _Q1A(WARPS_LAUNCHED
, 0x0001, B6
, LAUNCH
, 0x00000004, 1, 1),
874 _M2AB(IPC
, 0x3, B6
, EXEC
, 0x398, 0xffff, LOGOP
, WARP
, 0x0, DIV_SUM_M0
, 10, 1),
875 _M2AB(IPAC
, 0x3, B6
, EXEC
, 0x398, 0x1, B6
, WARP
, 0x0, AVG_DIV_MM
, 10, 1),
876 _M2A(IPEC
, 0x3, B6
, EXEC
, 0x398, 0xe, LOGOP
, EXEC
, 0x398, AVG_DIV_MM
, 10, 1),
877 _M2A(INST_REPLAY_OHEAD
, 0x3, B6
, ISSUE
, 0x104, 0x3, B6
, EXEC
, 0x398, REL_SUM_MM
, 100, 1),
878 _M2B(MP_OCCUPANCY
, 0x3f, B6
, WARP
, 0x31483104, 0x01, B6
, WARP
, 0x0, AVG_DIV_MM
, 200, 64),
879 _M2B(MP_EFFICIENCY
, 0x01, B6
, WARP
, 0x0, 0xffff, LOGOP
, WARP
, 0x0, AVG_DIV_M0
, 100, 1),
887 /* === PERFORMANCE MONITORING COUNTERS for NVC0:NVE4 === */
888 static const uint64_t nvc0_read_hw_sm_counters_code
[] =
891 * mov b32 $r9 $physid
900 * set $p0 0x1 eq u32 $r8 0x0
901 * mov b32 $r10 c0[0x0]
902 * mov b32 $r11 c0[0x4]
903 * ext u32 $r8 $r9 0x414
905 * mul $r8 u32 $r8 u32 36
906 * add b32 $r10 $c $r10 $r8
907 * add b32 $r11 $r11 0x0 $c
908 * mov b32 $r8 c0[0x8]
909 * st b128 wt g[$r10d+0x00] $r0q
910 * st b128 wt g[$r10d+0x10] $r4q
911 * st b32 wt g[$r10d+0x20] $r8
913 0x2c00000084021c04ULL
,
914 0x2c0000000c025c04ULL
,
915 0x2c00000010001c04ULL
,
916 0x2c00000014005c04ULL
,
917 0x2c00000018009c04ULL
,
918 0x2c0000001c00dc04ULL
,
919 0x2c00000020011c04ULL
,
920 0x2c00000024015c04ULL
,
921 0x2c00000028019c04ULL
,
922 0x2c0000002c01dc04ULL
,
923 0x190e0000fc81dc03ULL
,
924 0x2800400000029de4ULL
,
925 0x280040001002dde4ULL
,
926 0x7000c01050921c03ULL
,
927 0x80000000000021e7ULL
,
928 0x1000000090821c02ULL
,
929 0x4801000020a29c03ULL
,
930 0x0800000000b2dc42ULL
,
931 0x2800400020021de4ULL
,
932 0x9400000000a01fc5ULL
,
933 0x9400000040a11fc5ULL
,
934 0x9400000080a21f85ULL
,
935 0x8000000000001de7ULL
938 static const char *nvc0_pm_query_names
[] =
967 "thread_inst_executed_0",
968 "thread_inst_executed_1",
969 "thread_inst_executed_2",
970 "thread_inst_executed_3",
974 #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 } }
976 static const struct nvc0_hw_sm_query_cfg nvc0_hw_sm_queries
[] =
978 _Q(ACTIVE_CYCLES
, 0xaaaa, LOGOP
, 0x11, 1, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00),
979 _Q(ACTIVE_WARPS
, 0xaaaa, LOGOP
, 0x24, 6, 0x10, 0x21, 0x32, 0x43, 0x54, 0x65),
980 _Q(ATOM_COUNT
, 0xaaaa, LOGOP
, 0x63, 1, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00),
981 _Q(BRANCH
, 0xaaaa, LOGOP
, 0x1a, 2, 0x00, 0x11, 0x00, 0x00, 0x00, 0x00),
982 _Q(DIVERGENT_BRANCH
, 0xaaaa, LOGOP
, 0x19, 2, 0x20, 0x31, 0x00, 0x00, 0x00, 0x00),
983 _Q(GLD_REQUEST
, 0xaaaa, LOGOP
, 0x64, 1, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00),
984 _Q(GRED_COUNT
, 0xaaaa, LOGOP
, 0x63, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
985 _Q(GST_REQUEST
, 0xaaaa, LOGOP
, 0x64, 1, 0x60, 0x00, 0x00, 0x00, 0x00, 0x00),
986 _Q(INST_EXECUTED
, 0xaaaa, LOGOP
, 0x2d, 3, 0x00, 0x11, 0x22, 0x00, 0x00, 0x00),
987 _Q(INST_ISSUED1_0
, 0xaaaa, LOGOP
, 0x7e, 1, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00),
988 _Q(INST_ISSUED1_1
, 0xaaaa, LOGOP
, 0x7e, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
989 _Q(INST_ISSUED2_0
, 0xaaaa, LOGOP
, 0x7e, 1, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00),
990 _Q(INST_ISSUED2_1
, 0xaaaa, LOGOP
, 0x7e, 1, 0x50, 0x00, 0x00, 0x00, 0x00, 0x00),
991 _Q(LOCAL_LD
, 0xaaaa, LOGOP
, 0x64, 1, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00),
992 _Q(LOCAL_ST
, 0xaaaa, LOGOP
, 0x64, 1, 0x50, 0x00, 0x00, 0x00, 0x00, 0x00),
993 _Q(PROF_TRIGGER_0
, 0xaaaa, LOGOP
, 0x01, 1, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00),
994 _Q(PROF_TRIGGER_1
, 0xaaaa, LOGOP
, 0x01, 1, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00),
995 _Q(PROF_TRIGGER_2
, 0xaaaa, LOGOP
, 0x01, 1, 0x20, 0x00, 0x00, 0x00, 0x00, 0x00),
996 _Q(PROF_TRIGGER_3
, 0xaaaa, LOGOP
, 0x01, 1, 0x30, 0x00, 0x00, 0x00, 0x00, 0x00),
997 _Q(PROF_TRIGGER_4
, 0xaaaa, LOGOP
, 0x01, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
998 _Q(PROF_TRIGGER_5
, 0xaaaa, LOGOP
, 0x01, 1, 0x50, 0x00, 0x00, 0x00, 0x00, 0x00),
999 _Q(PROF_TRIGGER_6
, 0xaaaa, LOGOP
, 0x01, 1, 0x60, 0x00, 0x00, 0x00, 0x00, 0x00),
1000 _Q(PROF_TRIGGER_7
, 0xaaaa, LOGOP
, 0x01, 1, 0x70, 0x00, 0x00, 0x00, 0x00, 0x00),
1001 _Q(SHARED_LD
, 0xaaaa, LOGOP
, 0x64, 1, 0x10, 0x00, 0x00, 0x00, 0x00, 0x00),
1002 _Q(SHARED_ST
, 0xaaaa, LOGOP
, 0x64, 1, 0x40, 0x00, 0x00, 0x00, 0x00, 0x00),
1003 _Q(THREADS_LAUNCHED
, 0xaaaa, LOGOP
, 0x26, 6, 0x10, 0x21, 0x32, 0x43, 0x54, 0x65),
1004 _Q(TH_INST_EXECUTED_0
, 0xaaaa, LOGOP
, 0xa3, 6, 0x00, 0x11, 0x22, 0x33, 0x44, 0x55),
1005 _Q(TH_INST_EXECUTED_1
, 0xaaaa, LOGOP
, 0xa5, 6, 0x00, 0x11, 0x22, 0x33, 0x44, 0x55),
1006 _Q(TH_INST_EXECUTED_2
, 0xaaaa, LOGOP
, 0xa4, 6, 0x00, 0x11, 0x22, 0x33, 0x44, 0x55),
1007 _Q(TH_INST_EXECUTED_3
, 0xaaaa, LOGOP
, 0xa6, 6, 0x00, 0x11, 0x22, 0x33, 0x44, 0x55),
1008 _Q(WARPS_LAUNCHED
, 0xaaaa, LOGOP
, 0x26, 1, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00),
1013 static const struct nvc0_hw_sm_query_cfg
*
1014 nvc0_hw_sm_query_get_cfg(struct nvc0_context
*nvc0
, struct nvc0_query
*q
)
1016 struct nvc0_screen
*screen
= nvc0
->screen
;
1018 if (screen
->base
.class_3d
>= NVE4_3D_CLASS
)
1019 return &nve4_hw_sm_queries
[q
->type
- PIPE_QUERY_DRIVER_SPECIFIC
];
1020 return &nvc0_hw_sm_queries
[q
->type
- NVC0_HW_SM_QUERY(0)];
1024 nvc0_hw_sm_query_begin(struct nvc0_context
*nvc0
, struct nvc0_query
*q
)
1026 struct nvc0_screen
*screen
= nvc0
->screen
;
1027 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
1028 const bool is_nve4
= screen
->base
.class_3d
>= NVE4_3D_CLASS
;
1029 const struct nvc0_hw_sm_query_cfg
*cfg
;
1031 unsigned num_ab
[2] = { 0, 0 };
1033 cfg
= nvc0_hw_sm_query_get_cfg(nvc0
, q
);
1035 /* check if we have enough free counter slots */
1036 for (i
= 0; i
< cfg
->num_counters
; ++i
)
1037 num_ab
[cfg
->ctr
[i
].sig_dom
]++;
1039 if (screen
->pm
.num_hw_sm_active
[0] + num_ab
[0] > 4 ||
1040 screen
->pm
.num_hw_sm_active
[1] + num_ab
[1] > 4) {
1041 NOUVEAU_ERR("Not enough free MP counter slots !\n");
1045 assert(cfg
->num_counters
<= 4);
1046 PUSH_SPACE(push
, 4 * 8 * (is_nve4
? 1 : 6) + 6);
1048 if (!screen
->pm
.mp_counters_enabled
) {
1049 screen
->pm
.mp_counters_enabled
= true;
1050 BEGIN_NVC0(push
, SUBC_SW(0x06ac), 1);
1051 PUSH_DATA (push
, 0x1fcb);
1054 /* set sequence field to 0 (used to check if result is available) */
1055 for (i
= 0; i
< screen
->mp_count
; ++i
)
1056 q
->data
[i
* 10 + 10] = 0;
1058 for (i
= 0; i
< cfg
->num_counters
; ++i
) {
1059 const unsigned d
= cfg
->ctr
[i
].sig_dom
;
1061 if (!screen
->pm
.num_hw_sm_active
[d
]) {
1062 uint32_t m
= (1 << 22) | (1 << (7 + (8 * !d
)));
1063 if (screen
->pm
.num_hw_sm_active
[!d
])
1064 m
|= 1 << (7 + (8 * d
));
1065 BEGIN_NVC0(push
, SUBC_SW(0x0600), 1);
1066 PUSH_DATA (push
, m
);
1068 screen
->pm
.num_hw_sm_active
[d
]++;
1070 for (c
= d
* 4; c
< (d
* 4 + 4); ++c
) {
1071 if (!screen
->pm
.mp_counter
[c
]) {
1073 screen
->pm
.mp_counter
[c
] = (struct pipe_query
*)q
;
1077 assert(c
<= (d
* 4 + 3)); /* must succeed, already checked for space */
1079 /* configure and reset the counter(s) */
1082 BEGIN_NVC0(push
, NVE4_COMPUTE(MP_PM_A_SIGSEL(c
& 3)), 1);
1084 BEGIN_NVC0(push
, NVE4_COMPUTE(MP_PM_B_SIGSEL(c
& 3)), 1);
1085 PUSH_DATA (push
, cfg
->ctr
[i
].sig_sel
);
1086 BEGIN_NVC0(push
, NVE4_COMPUTE(MP_PM_SRCSEL(c
)), 1);
1087 PUSH_DATA (push
, cfg
->ctr
[i
].src_sel
+ 0x2108421 * (c
& 3));
1088 BEGIN_NVC0(push
, NVE4_COMPUTE(MP_PM_FUNC(c
)), 1);
1089 PUSH_DATA (push
, (cfg
->ctr
[i
].func
<< 4) | cfg
->ctr
[i
].mode
);
1090 BEGIN_NVC0(push
, NVE4_COMPUTE(MP_PM_SET(c
)), 1);
1091 PUSH_DATA (push
, 0);
1095 for (s
= 0; s
< cfg
->ctr
[i
].num_src
; s
++) {
1096 BEGIN_NVC0(push
, NVC0_COMPUTE(MP_PM_SIGSEL(s
)), 1);
1097 PUSH_DATA (push
, cfg
->ctr
[i
].sig_sel
);
1098 BEGIN_NVC0(push
, NVC0_COMPUTE(MP_PM_SRCSEL(s
)), 1);
1099 PUSH_DATA (push
, (cfg
->ctr
[i
].src_sel
>> (s
* 8)) & 0xff);
1100 BEGIN_NVC0(push
, NVC0_COMPUTE(MP_PM_OP(s
)), 1);
1101 PUSH_DATA (push
, (cfg
->ctr
[i
].func
<< 4) | cfg
->ctr
[i
].mode
);
1102 BEGIN_NVC0(push
, NVC0_COMPUTE(MP_PM_SET(s
)), 1);
1103 PUSH_DATA (push
, 0);
1111 nvc0_hw_sm_query_end(struct nvc0_context
*nvc0
, struct nvc0_query
*q
)
1113 struct nvc0_screen
*screen
= nvc0
->screen
;
1114 struct pipe_context
*pipe
= &nvc0
->base
.pipe
;
1115 struct nouveau_pushbuf
*push
= nvc0
->base
.pushbuf
;
1116 const bool is_nve4
= screen
->base
.class_3d
>= NVE4_3D_CLASS
;
1119 const uint block
[3] = { 32, is_nve4
? 4 : 1, 1 };
1120 const uint grid
[3] = { screen
->mp_count
, 1, 1 };
1122 const struct nvc0_hw_sm_query_cfg
*cfg
;
1124 cfg
= nvc0_hw_sm_query_get_cfg(nvc0
, q
);
1126 if (unlikely(!screen
->pm
.prog
)) {
1127 struct nvc0_program
*prog
= CALLOC_STRUCT(nvc0_program
);
1128 prog
->type
= PIPE_SHADER_COMPUTE
;
1129 prog
->translated
= true;
1130 prog
->num_gprs
= 14;
1131 prog
->parm_size
= 12;
1133 prog
->code
= (uint32_t *)nve4_read_hw_sm_counters_code
;
1134 prog
->code_size
= sizeof(nve4_read_hw_sm_counters_code
);
1136 prog
->code
= (uint32_t *)nvc0_read_hw_sm_counters_code
;
1137 prog
->code_size
= sizeof(nvc0_read_hw_sm_counters_code
);
1139 screen
->pm
.prog
= prog
;
1142 /* disable all counting */
1143 PUSH_SPACE(push
, 8);
1144 for (c
= 0; c
< 8; ++c
)
1145 if (screen
->pm
.mp_counter
[c
]) {
1147 IMMED_NVC0(push
, NVE4_COMPUTE(MP_PM_FUNC(c
)), 0);
1149 IMMED_NVC0(push
, NVC0_COMPUTE(MP_PM_OP(c
)), 0);
1152 /* release counters for this query */
1153 for (c
= 0; c
< 8; ++c
) {
1154 if (nvc0_query(screen
->pm
.mp_counter
[c
]) == q
) {
1155 screen
->pm
.num_hw_sm_active
[c
/ 4]--;
1156 screen
->pm
.mp_counter
[c
] = NULL
;
1160 BCTX_REFN_bo(nvc0
->bufctx_cp
, CP_QUERY
, NOUVEAU_BO_GART
| NOUVEAU_BO_WR
,
1163 PUSH_SPACE(push
, 1);
1164 IMMED_NVC0(push
, SUBC_COMPUTE(NV50_GRAPH_SERIALIZE
), 0);
1166 pipe
->bind_compute_state(pipe
, screen
->pm
.prog
);
1167 input
[0] = (q
->bo
->offset
+ q
->base
);
1168 input
[1] = (q
->bo
->offset
+ q
->base
) >> 32;
1169 input
[2] = q
->sequence
;
1170 pipe
->launch_grid(pipe
, block
, grid
, 0, input
);
1172 nouveau_bufctx_reset(nvc0
->bufctx_cp
, NVC0_BIND_CP_QUERY
);
1174 /* re-activate other counters */
1175 PUSH_SPACE(push
, 16);
1177 for (c
= 0; c
< 8; ++c
) {
1179 q
= nvc0_query(screen
->pm
.mp_counter
[c
]);
1182 cfg
= nvc0_hw_sm_query_get_cfg(nvc0
, q
);
1183 for (i
= 0; i
< cfg
->num_counters
; ++i
) {
1184 if (mask
& (1 << q
->ctr
[i
]))
1186 mask
|= 1 << q
->ctr
[i
];
1188 BEGIN_NVC0(push
, NVE4_COMPUTE(MP_PM_FUNC(q
->ctr
[i
])), 1);
1190 BEGIN_NVC0(push
, NVC0_COMPUTE(MP_PM_OP(q
->ctr
[i
])), 1);
1192 PUSH_DATA (push
, (cfg
->ctr
[i
].func
<< 4) | cfg
->ctr
[i
].mode
);
1198 nvc0_hw_sm_query_read_data(uint32_t count
[32][4],
1199 struct nvc0_context
*nvc0
, bool wait
,
1200 struct nvc0_query
*q
,
1201 const struct nvc0_hw_sm_query_cfg
*cfg
,
1206 for (p
= 0; p
< mp_count
; ++p
) {
1207 const unsigned b
= (0x24 / 4) * p
;
1209 for (c
= 0; c
< cfg
->num_counters
; ++c
) {
1210 if (q
->data
[b
+ 8] != q
->sequence
) {
1213 if (nouveau_bo_wait(q
->bo
, NOUVEAU_BO_RD
, nvc0
->base
.client
))
1216 count
[p
][c
] = q
->data
[b
+ q
->ctr
[c
]];
1223 nve4_hw_sm_query_read_data(uint32_t count
[32][4],
1224 struct nvc0_context
*nvc0
, bool wait
,
1225 struct nvc0_query
*q
,
1226 const struct nvc0_hw_sm_query_cfg
*cfg
,
1231 for (p
= 0; p
< mp_count
; ++p
) {
1232 const unsigned b
= (0x60 / 4) * p
;
1234 for (c
= 0; c
< cfg
->num_counters
; ++c
) {
1236 for (d
= 0; d
< ((q
->ctr
[c
] & ~3) ? 1 : 4); ++d
) {
1237 if (q
->data
[b
+ 20 + d
] != q
->sequence
) {
1240 if (nouveau_bo_wait(q
->bo
, NOUVEAU_BO_RD
, nvc0
->base
.client
))
1243 if (q
->ctr
[c
] & ~0x3)
1244 count
[p
][c
] = q
->data
[b
+ 16 + (q
->ctr
[c
] & 3)];
1246 count
[p
][c
] += q
->data
[b
+ d
* 4 + q
->ctr
[c
]];
1253 /* Metric calculations:
1254 * sum(x) ... sum of x over all MPs
1255 * avg(x) ... average of x over all MPs
1257 * IPC : sum(inst_executed) / clock
1258 * INST_REPLAY_OHEAD: (sum(inst_issued) - sum(inst_executed)) / sum(inst_issued)
1259 * MP_OCCUPANCY : avg((active_warps / 64) / active_cycles)
1260 * MP_EFFICIENCY : avg(active_cycles / clock)
1262 * NOTE: Interpretation of IPC requires knowledge of MP count.
1265 nvc0_hw_sm_query_result(struct nvc0_context
*nvc0
, struct nvc0_query
*q
,
1266 void *result
, boolean wait
)
1268 uint32_t count
[32][4];
1270 unsigned mp_count
= MIN2(nvc0
->screen
->mp_count_compute
, 32);
1272 const struct nvc0_hw_sm_query_cfg
*cfg
;
1275 cfg
= nvc0_hw_sm_query_get_cfg(nvc0
, q
);
1277 if (nvc0
->screen
->base
.class_3d
>= NVE4_3D_CLASS
)
1278 ret
= nve4_hw_sm_query_read_data(count
, nvc0
, wait
, q
, cfg
, mp_count
);
1280 ret
= nvc0_hw_sm_query_read_data(count
, nvc0
, wait
, q
, cfg
, mp_count
);
1284 if (cfg
->op
== NVC0_COUNTER_OPn_SUM
) {
1285 for (c
= 0; c
< cfg
->num_counters
; ++c
)
1286 for (p
= 0; p
< mp_count
; ++p
)
1287 value
+= count
[p
][c
];
1288 value
= (value
* cfg
->norm
[0]) / cfg
->norm
[1];
1290 if (cfg
->op
== NVC0_COUNTER_OPn_OR
) {
1292 for (c
= 0; c
< cfg
->num_counters
; ++c
)
1293 for (p
= 0; p
< mp_count
; ++p
)
1295 value
= ((uint64_t)v
* cfg
->norm
[0]) / cfg
->norm
[1];
1297 if (cfg
->op
== NVC0_COUNTER_OPn_AND
) {
1299 for (c
= 0; c
< cfg
->num_counters
; ++c
)
1300 for (p
= 0; p
< mp_count
; ++p
)
1302 value
= ((uint64_t)v
* cfg
->norm
[0]) / cfg
->norm
[1];
1304 if (cfg
->op
== NVC0_COUNTER_OP2_REL_SUM_MM
) {
1305 uint64_t v
[2] = { 0, 0 };
1306 for (p
= 0; p
< mp_count
; ++p
) {
1307 v
[0] += count
[p
][0];
1308 v
[1] += count
[p
][1];
1311 value
= ((v
[0] - v
[1]) * cfg
->norm
[0]) / (v
[0] * cfg
->norm
[1]);
1313 if (cfg
->op
== NVC0_COUNTER_OP2_DIV_SUM_M0
) {
1314 for (p
= 0; p
< mp_count
; ++p
)
1315 value
+= count
[p
][0];
1317 value
= (value
* cfg
->norm
[0]) / (count
[0][1] * cfg
->norm
[1]);
1321 if (cfg
->op
== NVC0_COUNTER_OP2_AVG_DIV_MM
) {
1322 unsigned mp_used
= 0;
1323 for (p
= 0; p
< mp_count
; ++p
, mp_used
+= !!count
[p
][0])
1325 value
+= (count
[p
][0] * cfg
->norm
[0]) / count
[p
][1];
1327 value
/= (uint64_t)mp_used
* cfg
->norm
[1];
1329 if (cfg
->op
== NVC0_COUNTER_OP2_AVG_DIV_M0
) {
1330 unsigned mp_used
= 0;
1331 for (p
= 0; p
< mp_count
; ++p
, mp_used
+= !!count
[p
][0])
1332 value
+= count
[p
][0];
1333 if (count
[0][1] && mp_used
) {
1334 value
*= cfg
->norm
[0];
1335 value
/= (uint64_t)count
[0][1] * mp_used
* cfg
->norm
[1];
1341 *(uint64_t *)result
= value
;
1346 nvc0_screen_get_driver_query_info(struct pipe_screen
*pscreen
,
1348 struct pipe_driver_query_info
*info
)
1350 struct nvc0_screen
*screen
= nvc0_screen(pscreen
);
1353 count
+= NVC0_SW_QUERY_DRV_STAT_COUNT
;
1355 if (screen
->base
.device
->drm_version
>= 0x01000101) {
1356 if (screen
->compute
) {
1357 if (screen
->base
.class_3d
== NVE4_3D_CLASS
) {
1358 count
+= NVE4_HW_SM_QUERY_COUNT
;
1360 if (screen
->base
.class_3d
< NVE4_3D_CLASS
) {
1361 /* NVC0_COMPUTE is not always enabled */
1362 count
+= NVC0_HW_SM_QUERY_COUNT
;
1370 /* Init default values. */
1371 info
->name
= "this_is_not_the_query_you_are_looking_for";
1372 info
->query_type
= 0xdeadd01d;
1373 info
->max_value
.u64
= 0;
1374 info
->type
= PIPE_DRIVER_QUERY_TYPE_UINT64
;
1375 info
->group_id
= -1;
1377 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
1378 if (id
< NVC0_SW_QUERY_DRV_STAT_COUNT
) {
1379 info
->name
= nvc0_sw_query_drv_stat_names
[id
];
1380 info
->query_type
= NVC0_SW_QUERY_DRV_STAT(id
);
1381 info
->max_value
.u64
= 0;
1382 if (strstr(info
->name
, "bytes"))
1383 info
->type
= PIPE_DRIVER_QUERY_TYPE_BYTES
;
1384 info
->group_id
= NVC0_SW_QUERY_DRV_STAT_GROUP
;
1389 if (screen
->compute
) {
1390 if (screen
->base
.class_3d
== NVE4_3D_CLASS
) {
1391 info
->name
= nve4_pm_query_names
[id
- NVC0_SW_QUERY_DRV_STAT_COUNT
];
1392 info
->query_type
= NVE4_HW_SM_QUERY(id
- NVC0_SW_QUERY_DRV_STAT_COUNT
);
1393 info
->max_value
.u64
=
1394 (id
< NVE4_HW_SM_QUERY_METRIC_MP_OCCUPANCY
) ? 0 : 100;
1395 info
->group_id
= NVC0_QUERY_MP_COUNTER_GROUP
;
1398 if (screen
->base
.class_3d
< NVE4_3D_CLASS
) {
1399 info
->name
= nvc0_pm_query_names
[id
- NVC0_SW_QUERY_DRV_STAT_COUNT
];
1400 info
->query_type
= NVC0_HW_SM_QUERY(id
- NVC0_SW_QUERY_DRV_STAT_COUNT
);
1401 info
->group_id
= NVC0_QUERY_MP_COUNTER_GROUP
;
1406 /* user asked for info about non-existing query */
1411 nvc0_screen_get_driver_query_group_info(struct pipe_screen
*pscreen
,
1413 struct pipe_driver_query_group_info
*info
)
1415 struct nvc0_screen
*screen
= nvc0_screen(pscreen
);
1418 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
1422 if (screen
->base
.device
->drm_version
>= 0x01000101) {
1423 if (screen
->compute
) {
1424 if (screen
->base
.class_3d
== NVE4_3D_CLASS
) {
1427 if (screen
->base
.class_3d
< NVE4_3D_CLASS
) {
1428 count
++; /* NVC0_COMPUTE is not always enabled */
1436 if (id
== NVC0_QUERY_MP_COUNTER_GROUP
) {
1437 if (screen
->compute
) {
1438 info
->name
= "MP counters";
1439 info
->type
= PIPE_DRIVER_QUERY_GROUP_TYPE_GPU
;
1441 if (screen
->base
.class_3d
== NVE4_3D_CLASS
) {
1442 info
->num_queries
= NVE4_HW_SM_QUERY_COUNT
;
1444 /* On NVE4+, each multiprocessor have 8 hardware counters separated
1445 * in two distinct domains, but we allow only one active query
1446 * simultaneously because some of them use more than one hardware
1447 * counter and this will result in an undefined behaviour. */
1448 info
->max_active_queries
= 1; /* TODO: handle multiple hw counters */
1451 if (screen
->base
.class_3d
< NVE4_3D_CLASS
) {
1452 info
->num_queries
= NVC0_HW_SM_QUERY_COUNT
;
1454 /* On NVC0:NVE4, each multiprocessor have 8 hardware counters
1455 * in a single domain. */
1456 info
->max_active_queries
= 8;
1461 #ifdef NOUVEAU_ENABLE_DRIVER_STATISTICS
1462 else if (id
== NVC0_SW_QUERY_DRV_STAT_GROUP
) {
1463 info
->name
= "Driver statistics";
1464 info
->type
= PIPE_DRIVER_QUERY_GROUP_TYPE_CPU
;
1465 info
->max_active_queries
= NVC0_SW_QUERY_DRV_STAT_COUNT
;
1466 info
->num_queries
= NVC0_SW_QUERY_DRV_STAT_COUNT
;
1471 /* user asked for info about non-existing query group */
1472 info
->name
= "this_is_not_the_query_group_you_are_looking_for";
1473 info
->max_active_queries
= 0;
1474 info
->num_queries
= 0;
1480 nvc0_init_query_functions(struct nvc0_context
*nvc0
)
1482 struct pipe_context
*pipe
= &nvc0
->base
.pipe
;
1484 pipe
->create_query
= nvc0_query_create
;
1485 pipe
->destroy_query
= nvc0_query_destroy
;
1486 pipe
->begin_query
= nvc0_query_begin
;
1487 pipe
->end_query
= nvc0_query_end
;
1488 pipe
->get_query_result
= nvc0_query_result
;
1489 pipe
->render_condition
= nvc0_render_condition
;