src/gallium/drivers/nvc0/nvc0_query.c

   1 /*
   2  * Copyright 2011 Nouveau Project
   3  *
   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:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  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 BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
  18  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
  19  * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  20  * SOFTWARE.
  21  *
  22  * Authors: Christoph Bumiller
  23  */
  24
  25 #include "nvc0_context.h"
  26 #include "nouveau/nv_object.xml.h"
  27
  28 struct nvc0_query {
  29    uint32_t *data;
  30    uint16_t type;
  31    uint16_t index;
  32    uint32_t sequence;
  33    struct nouveau_bo *bo;
  34    uint32_t base;
  35    uint32_t offset; /* base + i * rotate */
  36    boolean ready;
  37    boolean active;
  38    boolean is64bit;
  39    uint8_t rotate;
  40    int nesting; /* only used for occlusion queries */
  41    struct nouveau_mm_allocation *mm;
  42 };
  43
  44 #define NVC0_QUERY_ALLOC_SPACE 256
  45
  46 static INLINE struct nvc0_query *
  47 nvc0_query(struct pipe_query *pipe)
  48 {
  49    return (struct nvc0_query *)pipe;
  50 }
  51
  52 static boolean
  53 nvc0_query_allocate(struct nvc0_context *nvc0, struct nvc0_query *q, int size)
  54 {
  55    struct nvc0_screen *screen = nvc0->screen;
  56    int ret;
  57
  58    if (q->bo) {
  59       nouveau_bo_ref(NULL, &q->bo);
  60       if (q->mm) {
  61          if (q->ready)
  62             nouveau_mm_free(q->mm);
  63          else
  64             nouveau_fence_work(screen->base.fence.current,
  65                                nouveau_mm_free_work, q->mm);
  66       }
  67    }
  68    if (size) {
  69       q->mm = nouveau_mm_allocate(screen->base.mm_GART, size, &q->bo, &q->base);
  70       if (!q->bo)
  71          return FALSE;
  72       q->offset = q->base;
  73
  74       ret = nouveau_bo_map_range(q->bo, q->base, size, NOUVEAU_BO_RD |
  75                                  NOUVEAU_BO_NOSYNC);
  76       if (ret) {
  77          nvc0_query_allocate(nvc0, q, 0);
  78          return FALSE;
  79       }
  80       q->data = q->bo->map;
  81       nouveau_bo_unmap(q->bo);
  82    }
  83    return TRUE;
  84 }
  85
  86 static void
  87 nvc0_query_destroy(struct pipe_context *pipe, struct pipe_query *pq)
  88 {
  89    nvc0_query_allocate(nvc0_context(pipe), nvc0_query(pq), 0);
  90    FREE(nvc0_query(pq));
  91 }
  92
  93 static struct pipe_query *
  94 nvc0_query_create(struct pipe_context *pipe, unsigned type)
  95 {
  96    struct nvc0_context *nvc0 = nvc0_context(pipe);
  97    struct nvc0_query *q;
  98    unsigned space = NVC0_QUERY_ALLOC_SPACE;
  99
 100    q = CALLOC_STRUCT(nvc0_query);
 101    if (!q)
 102       return NULL;
 103
 104    switch (type) {
 105    case PIPE_QUERY_OCCLUSION_COUNTER:
 106    case PIPE_QUERY_OCCLUSION_PREDICATE:
 107       q->rotate = 32;
 108       space = NVC0_QUERY_ALLOC_SPACE;
 109       break;
 110    case PIPE_QUERY_PIPELINE_STATISTICS:
 111       q->is64bit = TRUE;
 112       space = 512;
 113       break;
 114    case PIPE_QUERY_SO_STATISTICS:
 115    case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 116       q->is64bit = TRUE;
 117       space = 64;
 118       break;
 119    case PIPE_QUERY_TIME_ELAPSED:
 120    case PIPE_QUERY_TIMESTAMP:
 121    case PIPE_QUERY_TIMESTAMP_DISJOINT:
 122    case PIPE_QUERY_GPU_FINISHED:
 123    case PIPE_QUERY_PRIMITIVES_GENERATED:
 124    case PIPE_QUERY_PRIMITIVES_EMITTED:
 125       space = 32;
 126       break;
 127    case NVC0_QUERY_TFB_BUFFER_OFFSET:
 128       space = 16;
 129       break;
 130    default:
 131       FREE(q);
 132       return NULL;
 133    }
 134    if (!nvc0_query_allocate(nvc0, q, space)) {
 135       FREE(q);
 136       return NULL;
 137    }
 138
 139    q->type = type;
 140
 141    if (q->rotate) {
 142       /* we advance before query_begin ! */
 143       q->offset -= q->rotate;
 144       q->data -= q->rotate / sizeof(*q->data);
 145    } else
 146    if (!q->is64bit)
 147       q->data[0] = 0; /* initialize sequence */
 148
 149    return (struct pipe_query *)q;
 150 }
 151
 152 static void
 153 nvc0_query_get(struct nouveau_channel *chan, struct nvc0_query *q,
 154                unsigned offset, uint32_t get)
 155 {
 156    offset += q->offset;
 157
 158    MARK_RING (chan, 5, 2);
 159    BEGIN_RING(chan, RING_3D(QUERY_ADDRESS_HIGH), 4);
 160    OUT_RELOCh(chan, q->bo, offset, NOUVEAU_BO_GART | NOUVEAU_BO_WR);
 161    OUT_RELOCl(chan, q->bo, offset, NOUVEAU_BO_GART | NOUVEAU_BO_WR);
 162    OUT_RING  (chan, q->sequence);
 163    OUT_RING  (chan, get);
 164 }
 165
 166 static void
 167 nvc0_query_rotate(struct nvc0_context *nvc0, struct nvc0_query *q)
 168 {
 169    q->offset += q->rotate;
 170    q->data += q->rotate / sizeof(*q->data);
 171    if (q->offset - q->base == NVC0_QUERY_ALLOC_SPACE)
 172       nvc0_query_allocate(nvc0, q, NVC0_QUERY_ALLOC_SPACE);
 173 }
 174
 175 static void
 176 nvc0_query_begin(struct pipe_context *pipe, struct pipe_query *pq)
 177 {
 178    struct nvc0_context *nvc0 = nvc0_context(pipe);
 179    struct nouveau_channel *chan = nvc0->screen->base.channel;
 180    struct nvc0_query *q = nvc0_query(pq);
 181
 182    /* For occlusion queries we have to change the storage, because a previous
 183     * query might set the initial render conition to FALSE even *after* we re-
 184     * initialized it to TRUE.
 185     */
 186    if (q->rotate) {
 187       nvc0_query_rotate(nvc0, q);
 188
 189       /* XXX: can we do this with the GPU, and sync with respect to a previous
 190        *  query ?
 191        */
 192       q->data[0] = q->sequence; /* initialize sequence */
 193       q->data[1] = 1; /* initial render condition = TRUE */
 194       q->data[4] = q->sequence + 1; /* for comparison COND_MODE */
 195       q->data[5] = 0;
 196    }
 197    q->sequence++;
 198
 199    switch (q->type) {
 200    case PIPE_QUERY_OCCLUSION_COUNTER:
 201    case PIPE_QUERY_OCCLUSION_PREDICATE:
 202       q->nesting = nvc0->screen->num_occlusion_queries_active++;
 203       if (q->nesting) {
 204          nvc0_query_get(chan, q, 0x10, 0x0100f002);
 205       } else {
 206          BEGIN_RING(chan, RING_3D(COUNTER_RESET), 1);
 207          OUT_RING  (chan, NVC0_3D_COUNTER_RESET_SAMPLECNT);
 208          IMMED_RING(chan, RING_3D(SAMPLECNT_ENABLE), 1);
 209       }
 210       break;
 211    case PIPE_QUERY_PRIMITIVES_GENERATED:
 212       nvc0_query_get(chan, q, 0x10, 0x06805002 | (q->index << 5));
 213       break;
 214    case PIPE_QUERY_PRIMITIVES_EMITTED:
 215       nvc0_query_get(chan, q, 0x10, 0x05805002 | (q->index << 5));
 216       break;
 217    case PIPE_QUERY_SO_STATISTICS:
 218       nvc0_query_get(chan, q, 0x20, 0x05805002 | (q->index << 5));
 219       nvc0_query_get(chan, q, 0x30, 0x06805002 | (q->index << 5));
 220       break;
 221    case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 222       nvc0_query_get(chan, q, 0x10, 0x03005002 | (q->index << 5));
 223       break;
 224    case PIPE_QUERY_TIMESTAMP_DISJOINT:
 225    case PIPE_QUERY_TIME_ELAPSED:
 226       nvc0_query_get(chan, q, 0x10, 0x00005002);
 227       break;
 228    case PIPE_QUERY_PIPELINE_STATISTICS:
 229       nvc0_query_get(chan, q, 0xc0 + 0x00, 0x00801002); /* VFETCH, VERTICES */
 230       nvc0_query_get(chan, q, 0xc0 + 0x10, 0x01801002); /* VFETCH, PRIMS */
 231       nvc0_query_get(chan, q, 0xc0 + 0x20, 0x02802002); /* VP, LAUNCHES */
 232       nvc0_query_get(chan, q, 0xc0 + 0x30, 0x03806002); /* GP, LAUNCHES */
 233       nvc0_query_get(chan, q, 0xc0 + 0x40, 0x04806002); /* GP, PRIMS_OUT */
 234       nvc0_query_get(chan, q, 0xc0 + 0x50, 0x07804002); /* RAST, PRIMS_IN */
 235       nvc0_query_get(chan, q, 0xc0 + 0x60, 0x08804002); /* RAST, PRIMS_OUT */
 236       nvc0_query_get(chan, q, 0xc0 + 0x70, 0x0980a002); /* ROP, PIXELS */
 237       nvc0_query_get(chan, q, 0xc0 + 0x80, 0x0d808002); /* TCP, LAUNCHES */
 238       nvc0_query_get(chan, q, 0xc0 + 0x90, 0x0e809002); /* TEP, LAUNCHES */
 239       break;
 240    default:
 241       break;
 242    }
 243    q->ready = FALSE;
 244    q->active = TRUE;
 245 }
 246
 247 static void
 248 nvc0_query_end(struct pipe_context *pipe, struct pipe_query *pq)
 249 {
 250    struct nvc0_context *nvc0 = nvc0_context(pipe);
 251    struct nouveau_channel *chan = nvc0->screen->base.channel;
 252    struct nvc0_query *q = nvc0_query(pq);
 253
 254    if (!q->active) {
 255       /* some queries don't require 'begin' to be called (e.g. GPU_FINISHED) */
 256       if (q->rotate)
 257          nvc0_query_rotate(nvc0, q);
 258       q->sequence++;
 259    }
 260    q->ready = FALSE;
 261    q->active = FALSE;
 262
 263    switch (q->type) {
 264    case PIPE_QUERY_OCCLUSION_COUNTER:
 265    case PIPE_QUERY_OCCLUSION_PREDICATE:
 266       nvc0_query_get(chan, q, 0, 0x0100f002);
 267       if (--nvc0->screen->num_occlusion_queries_active == 0)
 268          IMMED_RING(chan, RING_3D(SAMPLECNT_ENABLE), 0);
 269       break;
 270    case PIPE_QUERY_PRIMITIVES_GENERATED:
 271       nvc0_query_get(chan, q, 0, 0x06805002 | (q->index << 5));
 272       break;
 273    case PIPE_QUERY_PRIMITIVES_EMITTED:
 274       nvc0_query_get(chan, q, 0, 0x05805002 | (q->index << 5));
 275       break;
 276    case PIPE_QUERY_SO_STATISTICS:
 277       nvc0_query_get(chan, q, 0x00, 0x05805002 | (q->index << 5));
 278       nvc0_query_get(chan, q, 0x10, 0x06805002 | (q->index << 5));
 279       break;
 280    case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 281       /* TODO: How do we sum over all streams for render condition ? */
 282       /* PRIMS_DROPPED doesn't write sequence, use a ZERO query to sync on */
 283       nvc0_query_get(chan, q, 0x00, 0x03005002 | (q->index << 5));
 284       nvc0_query_get(chan, q, 0x20, 0x00005002);
 285       break;
 286    case PIPE_QUERY_TIMESTAMP:
 287    case PIPE_QUERY_TIMESTAMP_DISJOINT:
 288    case PIPE_QUERY_TIME_ELAPSED:
 289       nvc0_query_get(chan, q, 0, 0x00005002);
 290       break;
 291    case PIPE_QUERY_GPU_FINISHED:
 292       nvc0_query_get(chan, q, 0, 0x1000f010);
 293       break;
 294    case PIPE_QUERY_PIPELINE_STATISTICS:
 295       nvc0_query_get(chan, q, 0x00, 0x00801002); /* VFETCH, VERTICES */
 296       nvc0_query_get(chan, q, 0x10, 0x01801002); /* VFETCH, PRIMS */
 297       nvc0_query_get(chan, q, 0x20, 0x02802002); /* VP, LAUNCHES */
 298       nvc0_query_get(chan, q, 0x30, 0x03806002); /* GP, LAUNCHES */
 299       nvc0_query_get(chan, q, 0x40, 0x04806002); /* GP, PRIMS_OUT */
 300       nvc0_query_get(chan, q, 0x50, 0x07804002); /* RAST, PRIMS_IN */
 301       nvc0_query_get(chan, q, 0x60, 0x08804002); /* RAST, PRIMS_OUT */
 302       nvc0_query_get(chan, q, 0x70, 0x0980a002); /* ROP, PIXELS */
 303       nvc0_query_get(chan, q, 0x80, 0x0d808002); /* TCP, LAUNCHES */
 304       nvc0_query_get(chan, q, 0x90, 0x0e809002); /* TEP, LAUNCHES */
 305       break;
 306    case NVC0_QUERY_TFB_BUFFER_OFFSET:
 307       /* indexed by TFB buffer instead of by vertex stream */
 308       nvc0_query_get(chan, q, 0x00, 0x0d005002 | (q->index << 5));
 309       break;
 310    default:
 311       assert(0);
 312       break;
 313    }
 314 }
 315
 316 static INLINE boolean
 317 nvc0_query_ready(struct nvc0_query *q)
 318 {
 319    if (q->is64bit) {
 320       if (nouveau_bo_map(q->bo, NOUVEAU_BO_RD | NOUVEAU_BO_NOWAIT))
 321          return FALSE;
 322       nouveau_bo_unmap(q->bo);
 323       return TRUE;
 324    } else {
 325       return q->data[0] == q->sequence;
 326    }
 327 }
 328
 329 static INLINE boolean
 330 nvc0_query_wait(struct nvc0_query *q)
 331 {
 332    int ret = nouveau_bo_map(q->bo, NOUVEAU_BO_RD);
 333    if (ret)
 334       return FALSE;
 335    nouveau_bo_unmap(q->bo);
 336    return TRUE;
 337 }
 338
 339 static boolean
 340 nvc0_query_result(struct pipe_context *pipe, struct pipe_query *pq,
 341                   boolean wait, union pipe_query_result *result)
 342 {
 343    struct nvc0_query *q = nvc0_query(pq);
 344    uint64_t *res64 = (uint64_t*)result;
 345    uint32_t *res32 = (uint32_t*)result;
 346    boolean *res8 = (boolean*)result;
 347    uint64_t *data64 = (uint64_t *)q->data;
 348    unsigned i;
 349
 350    if (!q->ready) /* update ? */
 351       q->ready = nvc0_query_ready(q);
 352    if (!q->ready) {
 353       struct nouveau_channel *chan = nvc0_context(pipe)->screen->base.channel;
 354       if (!wait) {
 355          if (nouveau_bo_pending(q->bo) & NOUVEAU_BO_WR) /* for daft apps */
 356             FIRE_RING(chan);
 357          return FALSE;
 358       }
 359       if (!nvc0_query_wait(q))
 360          return FALSE;
 361    }
 362    q->ready = TRUE;
 363
 364    switch (q->type) {
 365    case PIPE_QUERY_GPU_FINISHED:
 366       res8[0] = TRUE;
 367       break;
 368    case PIPE_QUERY_OCCLUSION_COUNTER: /* u32 sequence, u32 count, u64 time */
 369       res64[0] = q->data[1] - q->data[5];
 370       break;
 371    case PIPE_QUERY_OCCLUSION_PREDICATE:
 372       res8[0] = q->data[1] != q->data[5];
 373       break;
 374    case PIPE_QUERY_PRIMITIVES_GENERATED: /* u64 count, u64 time */
 375    case PIPE_QUERY_PRIMITIVES_EMITTED: /* u64 count, u64 time */
 376       res64[0] = data64[0] - data64[2];
 377       break;
 378    case PIPE_QUERY_SO_STATISTICS:
 379       res64[0] = data64[0] - data64[4];
 380       res64[1] = data64[2] - data64[6];
 381       break;
 382    case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 383       res8[0] = data64[0] != data64[2];
 384       break;
 385    case PIPE_QUERY_TIMESTAMP:
 386       res64[0] = data64[1];
 387       break;
 388    case PIPE_QUERY_TIMESTAMP_DISJOINT: /* u32 sequence, u32 0, u64 time */
 389       res64[0] = 1000000000;
 390       res8[8] = (data64[1] == data64[3]) ? FALSE : TRUE;
 391       break;
 392    case PIPE_QUERY_TIME_ELAPSED:
 393       res64[0] = data64[1] - data64[3];
 394       break;
 395    case PIPE_QUERY_PIPELINE_STATISTICS:
 396       for (i = 0; i < 10; ++i)
 397          res64[i] = data64[i * 2] - data64[24 + i * 2];
 398       break;
 399    case NVC0_QUERY_TFB_BUFFER_OFFSET:
 400       res32[0] = q->data[1];
 401       break;
 402    default:
 403       return FALSE;
 404    }
 405
 406    return TRUE;
 407 }
 408
 409 void
 410 nvc0_query_fifo_wait(struct nouveau_channel *chan, struct pipe_query *pq)
 411 {
 412    struct nvc0_query *q = nvc0_query(pq);
 413    unsigned offset = q->offset;
 414
 415    if (q->type == PIPE_QUERY_SO_OVERFLOW_PREDICATE) offset += 0x20;
 416
 417    MARK_RING (chan, 5, 2);
 418    BEGIN_RING(chan, RING_3D_(NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH), 4);
 419    OUT_RELOCh(chan, q->bo, offset, NOUVEAU_BO_GART | NOUVEAU_BO_RD);
 420    OUT_RELOCl(chan, q->bo, offset, NOUVEAU_BO_GART | NOUVEAU_BO_RD);
 421    OUT_RING  (chan, q->sequence);
 422    OUT_RING  (chan, (1 << 12) |
 423               NV84_SUBCHAN_SEMAPHORE_TRIGGER_ACQUIRE_EQUAL);
 424 }
 425
 426 static void
 427 nvc0_render_condition(struct pipe_context *pipe,
 428                       struct pipe_query *pq, uint mode)
 429 {
 430    struct nvc0_context *nvc0 = nvc0_context(pipe);
 431    struct nouveau_channel *chan = nvc0->screen->base.channel;
 432    struct nvc0_query *q;
 433    uint32_t cond;
 434    boolean negated = FALSE;
 435    boolean wait =
 436       mode != PIPE_RENDER_COND_NO_WAIT &&
 437       mode != PIPE_RENDER_COND_BY_REGION_NO_WAIT;
 438
 439    if (!pq) {
 440       IMMED_RING(chan, RING_3D(COND_MODE), NVC0_3D_COND_MODE_ALWAYS);
 441       return;
 442    }
 443    q = nvc0_query(pq);
 444
 445    /* NOTE: comparison of 2 queries only works if both have completed */
 446    switch (q->type) {
 447    case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 448       cond = negated ? NVC0_3D_COND_MODE_EQUAL :
 449                        NVC0_3D_COND_MODE_NOT_EQUAL;
 450       wait = TRUE;
 451       break;
 452    case PIPE_QUERY_OCCLUSION_COUNTER:
 453    case PIPE_QUERY_OCCLUSION_PREDICATE:
 454       if (likely(!negated)) {
 455          if (unlikely(q->nesting))
 456             cond = wait ? NVC0_3D_COND_MODE_NOT_EQUAL :
 457                           NVC0_3D_COND_MODE_ALWAYS;
 458          else
 459             cond = NVC0_3D_COND_MODE_RES_NON_ZERO;
 460       } else {
 461          cond = wait ? NVC0_3D_COND_MODE_EQUAL : NVC0_3D_COND_MODE_ALWAYS;
 462       }
 463       break;
 464    default:
 465       assert(!"render condition query not a predicate");
 466       mode = NVC0_3D_COND_MODE_ALWAYS;
 467       break;
 468    }
 469
 470    if (wait)
 471       nvc0_query_fifo_wait(chan, pq);
 472
 473    MARK_RING (chan, 4, 2);
 474    BEGIN_RING(chan, RING_3D(COND_ADDRESS_HIGH), 3);
 475    OUT_RELOCh(chan, q->bo, q->offset, NOUVEAU_BO_GART | NOUVEAU_BO_RD);
 476    OUT_RELOCl(chan, q->bo, q->offset, NOUVEAU_BO_GART | NOUVEAU_BO_RD);
 477    OUT_RING  (chan, cond);
 478 }
 479
 480 void
 481 nvc0_query_pushbuf_submit(struct nouveau_channel *chan,
 482                           struct pipe_query *pq, unsigned result_offset)
 483 {
 484    struct nvc0_query *q = nvc0_query(pq);
 485
 486 #define NVC0_IB_ENTRY_1_NO_PREFETCH (1 << (31 - 8))
 487
 488    nouveau_pushbuf_submit(chan, q->bo, q->offset + result_offset, 4 |
 489                           NVC0_IB_ENTRY_1_NO_PREFETCH);
 490 }
 491
 492 void
 493 nvc0_so_target_save_offset(struct pipe_context *pipe,
 494                            struct pipe_stream_output_target *ptarg,
 495                            unsigned index, boolean *serialize)
 496 {
 497    struct nvc0_so_target *targ = nvc0_so_target(ptarg);
 498
 499    if (*serialize) {
 500       struct nouveau_channel *chan = nvc0_context(pipe)->screen->base.channel;
 501       *serialize = FALSE;
 502       IMMED_RING(chan, RING_3D(SERIALIZE), 0);
 503    }
 504
 505    nvc0_query(targ->pq)->index = index;
 506
 507    nvc0_query_end(pipe, targ->pq);
 508 }
 509
 510 void
 511 nvc0_init_query_functions(struct nvc0_context *nvc0)
 512 {
 513    struct pipe_context *pipe = &nvc0->base.pipe;
 514
 515    pipe->create_query = nvc0_query_create;
 516    pipe->destroy_query = nvc0_query_destroy;
 517    pipe->begin_query = nvc0_query_begin;
 518    pipe->end_query = nvc0_query_end;
 519    pipe->get_query_result = nvc0_query_result;
 520    pipe->render_condition = nvc0_render_condition;
 521 }