src/gallium/drivers/nouveau/nv50/nv50_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 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.
  21  *
  22  * Authors: Christoph Bumiller
  23  */
  24
  25 #define NV50_PUSH_EXPLICIT_SPACE_CHECKING
  26
  27 #include "nv50/nv50_context.h"
  28 #include "nv_object.xml.h"
  29
  30 #define NV50_QUERY_STATE_READY   0
  31 #define NV50_QUERY_STATE_ACTIVE  1
  32 #define NV50_QUERY_STATE_ENDED   2
  33 #define NV50_QUERY_STATE_FLUSHED 3
  34
  35 /* XXX: Nested queries, and simultaneous queries on multiple gallium contexts
  36  * (since we use only a single GPU channel per screen) will not work properly.
  37  *
  38  * The first is not that big of an issue because OpenGL does not allow nested
  39  * queries anyway.
  40  */
  41
  42 struct nv50_query {
  43    uint32_t *data;
  44    uint16_t type;
  45    uint16_t index;
  46    uint32_t sequence;
  47    struct nouveau_bo *bo;
  48    uint32_t base;
  49    uint32_t offset; /* base + i * 32 */
  50    uint8_t state;
  51    bool is64bit;
  52    int nesting; /* only used for occlusion queries */
  53    struct nouveau_mm_allocation *mm;
  54    struct nouveau_fence *fence;
  55 };
  56
  57 #define NV50_QUERY_ALLOC_SPACE 256
  58
  59 static INLINE struct nv50_query *
  60 nv50_query(struct pipe_query *pipe)
  61 {
  62    return (struct nv50_query *)pipe;
  63 }
  64
  65 static bool
  66 nv50_query_allocate(struct nv50_context *nv50, struct nv50_query *q, int size)
  67 {
  68    struct nv50_screen *screen = nv50->screen;
  69    int ret;
  70
  71    if (q->bo) {
  72       nouveau_bo_ref(NULL, &q->bo);
  73       if (q->mm) {
  74          if (q->state == NV50_QUERY_STATE_READY)
  75             nouveau_mm_free(q->mm);
  76          else
  77             nouveau_fence_work(screen->base.fence.current, nouveau_mm_free_work,
  78                                q->mm);
  79       }
  80    }
  81    if (size) {
  82       q->mm = nouveau_mm_allocate(screen->base.mm_GART, size, &q->bo, &q->base);
  83       if (!q->bo)
  84          return false;
  85       q->offset = q->base;
  86
  87       ret = nouveau_bo_map(q->bo, 0, screen->base.client);
  88       if (ret) {
  89          nv50_query_allocate(nv50, q, 0);
  90          return false;
  91       }
  92       q->data = (uint32_t *)((uint8_t *)q->bo->map + q->base);
  93    }
  94    return true;
  95 }
  96
  97 static void
  98 nv50_query_destroy(struct pipe_context *pipe, struct pipe_query *pq)
  99 {
 100    nv50_query_allocate(nv50_context(pipe), nv50_query(pq), 0);
 101    nouveau_fence_ref(NULL, &nv50_query(pq)->fence);
 102    FREE(nv50_query(pq));
 103 }
 104
 105 static struct pipe_query *
 106 nv50_query_create(struct pipe_context *pipe, unsigned type, unsigned index)
 107 {
 108    struct nv50_context *nv50 = nv50_context(pipe);
 109    struct nv50_query *q;
 110
 111    q = CALLOC_STRUCT(nv50_query);
 112    if (!q)
 113       return NULL;
 114
 115    if (!nv50_query_allocate(nv50, q, NV50_QUERY_ALLOC_SPACE)) {
 116       FREE(q);
 117       return NULL;
 118    }
 119
 120    q->is64bit = (type == PIPE_QUERY_PRIMITIVES_GENERATED ||
 121                  type == PIPE_QUERY_PRIMITIVES_EMITTED ||
 122                  type == PIPE_QUERY_SO_STATISTICS ||
 123                  type == PIPE_QUERY_PIPELINE_STATISTICS);
 124    q->type = type;
 125
 126    if (q->type == PIPE_QUERY_OCCLUSION_COUNTER) {
 127       q->offset -= 32;
 128       q->data -= 32 / sizeof(*q->data); /* we advance before query_begin ! */
 129    }
 130
 131    return (struct pipe_query *)q;
 132 }
 133
 134 static void
 135 nv50_query_get(struct nouveau_pushbuf *push, struct nv50_query *q,
 136                unsigned offset, uint32_t get)
 137 {
 138    offset += q->offset;
 139
 140    PUSH_SPACE(push, 5);
 141    PUSH_REFN (push, q->bo, NOUVEAU_BO_GART | NOUVEAU_BO_WR);
 142    BEGIN_NV04(push, NV50_3D(QUERY_ADDRESS_HIGH), 4);
 143    PUSH_DATAh(push, q->bo->offset + offset);
 144    PUSH_DATA (push, q->bo->offset + offset);
 145    PUSH_DATA (push, q->sequence);
 146    PUSH_DATA (push, get);
 147 }
 148
 149 static boolean
 150 nv50_query_begin(struct pipe_context *pipe, struct pipe_query *pq)
 151 {
 152    struct nv50_context *nv50 = nv50_context(pipe);
 153    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 154    struct nv50_query *q = nv50_query(pq);
 155
 156    /* For occlusion queries we have to change the storage, because a previous
 157     * query might set the initial render conition to false even *after* we re-
 158     * initialized it to true.
 159     */
 160    if (q->type == PIPE_QUERY_OCCLUSION_COUNTER) {
 161       q->offset += 32;
 162       q->data += 32 / sizeof(*q->data);
 163       if (q->offset - q->base == NV50_QUERY_ALLOC_SPACE)
 164          nv50_query_allocate(nv50, q, NV50_QUERY_ALLOC_SPACE);
 165
 166       /* XXX: can we do this with the GPU, and sync with respect to a previous
 167        *  query ?
 168        */
 169       q->data[0] = q->sequence; /* initialize sequence */
 170       q->data[1] = 1; /* initial render condition = true */
 171       q->data[4] = q->sequence + 1; /* for comparison COND_MODE */
 172       q->data[5] = 0;
 173    }
 174    if (!q->is64bit)
 175       q->data[0] = q->sequence++; /* the previously used one */
 176
 177    switch (q->type) {
 178    case PIPE_QUERY_OCCLUSION_COUNTER:
 179       q->nesting = nv50->screen->num_occlusion_queries_active++;
 180       if (q->nesting) {
 181          nv50_query_get(push, q, 0x10, 0x0100f002);
 182       } else {
 183          PUSH_SPACE(push, 4);
 184          BEGIN_NV04(push, NV50_3D(COUNTER_RESET), 1);
 185          PUSH_DATA (push, NV50_3D_COUNTER_RESET_SAMPLECNT);
 186          BEGIN_NV04(push, NV50_3D(SAMPLECNT_ENABLE), 1);
 187          PUSH_DATA (push, 1);
 188       }
 189       break;
 190    case PIPE_QUERY_PRIMITIVES_GENERATED:
 191       nv50_query_get(push, q, 0x10, 0x06805002);
 192       break;
 193    case PIPE_QUERY_PRIMITIVES_EMITTED:
 194       nv50_query_get(push, q, 0x10, 0x05805002);
 195       break;
 196    case PIPE_QUERY_SO_STATISTICS:
 197       nv50_query_get(push, q, 0x20, 0x05805002);
 198       nv50_query_get(push, q, 0x30, 0x06805002);
 199       break;
 200    case PIPE_QUERY_PIPELINE_STATISTICS:
 201       nv50_query_get(push, q, 0x80, 0x00801002); /* VFETCH, VERTICES */
 202       nv50_query_get(push, q, 0x90, 0x01801002); /* VFETCH, PRIMS */
 203       nv50_query_get(push, q, 0xa0, 0x02802002); /* VP, LAUNCHES */
 204       nv50_query_get(push, q, 0xb0, 0x03806002); /* GP, LAUNCHES */
 205       nv50_query_get(push, q, 0xc0, 0x04806002); /* GP, PRIMS_OUT */
 206       nv50_query_get(push, q, 0xd0, 0x07804002); /* RAST, PRIMS_IN */
 207       nv50_query_get(push, q, 0xe0, 0x08804002); /* RAST, PRIMS_OUT */
 208       nv50_query_get(push, q, 0xf0, 0x0980a002); /* ROP, PIXELS */
 209       break;
 210    case PIPE_QUERY_TIME_ELAPSED:
 211       nv50_query_get(push, q, 0x10, 0x00005002);
 212       break;
 213    default:
 214       break;
 215    }
 216    q->state = NV50_QUERY_STATE_ACTIVE;
 217    return true;
 218 }
 219
 220 static void
 221 nv50_query_end(struct pipe_context *pipe, struct pipe_query *pq)
 222 {
 223    struct nv50_context *nv50 = nv50_context(pipe);
 224    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 225    struct nv50_query *q = nv50_query(pq);
 226
 227    q->state = NV50_QUERY_STATE_ENDED;
 228
 229    switch (q->type) {
 230    case PIPE_QUERY_OCCLUSION_COUNTER:
 231       nv50_query_get(push, q, 0, 0x0100f002);
 232       if (--nv50->screen->num_occlusion_queries_active == 0) {
 233          PUSH_SPACE(push, 2);
 234          BEGIN_NV04(push, NV50_3D(SAMPLECNT_ENABLE), 1);
 235          PUSH_DATA (push, 0);
 236       }
 237       break;
 238    case PIPE_QUERY_PRIMITIVES_GENERATED:
 239       nv50_query_get(push, q, 0, 0x06805002);
 240       break;
 241    case PIPE_QUERY_PRIMITIVES_EMITTED:
 242       nv50_query_get(push, q, 0, 0x05805002);
 243       break;
 244    case PIPE_QUERY_SO_STATISTICS:
 245       nv50_query_get(push, q, 0x00, 0x05805002);
 246       nv50_query_get(push, q, 0x10, 0x06805002);
 247       break;
 248    case PIPE_QUERY_PIPELINE_STATISTICS:
 249       nv50_query_get(push, q, 0x00, 0x00801002); /* VFETCH, VERTICES */
 250       nv50_query_get(push, q, 0x10, 0x01801002); /* VFETCH, PRIMS */
 251       nv50_query_get(push, q, 0x20, 0x02802002); /* VP, LAUNCHES */
 252       nv50_query_get(push, q, 0x30, 0x03806002); /* GP, LAUNCHES */
 253       nv50_query_get(push, q, 0x40, 0x04806002); /* GP, PRIMS_OUT */
 254       nv50_query_get(push, q, 0x50, 0x07804002); /* RAST, PRIMS_IN */
 255       nv50_query_get(push, q, 0x60, 0x08804002); /* RAST, PRIMS_OUT */
 256       nv50_query_get(push, q, 0x70, 0x0980a002); /* ROP, PIXELS */
 257       break;
 258    case PIPE_QUERY_TIMESTAMP:
 259       q->sequence++;
 260       /* fall through */
 261    case PIPE_QUERY_TIME_ELAPSED:
 262       nv50_query_get(push, q, 0, 0x00005002);
 263       break;
 264    case PIPE_QUERY_GPU_FINISHED:
 265       q->sequence++;
 266       nv50_query_get(push, q, 0, 0x1000f010);
 267       break;
 268    case NVA0_QUERY_STREAM_OUTPUT_BUFFER_OFFSET:
 269       nv50_query_get(push, q, 0, 0x0d005002 | (q->index << 5));
 270       break;
 271    case PIPE_QUERY_TIMESTAMP_DISJOINT:
 272       /* This query is not issued on GPU because disjoint is forced to false */
 273       q->state = NV50_QUERY_STATE_READY;
 274       break;
 275    default:
 276       assert(0);
 277       break;
 278    }
 279
 280    if (q->is64bit)
 281       nouveau_fence_ref(nv50->screen->base.fence.current, &q->fence);
 282 }
 283
 284 static INLINE void
 285 nv50_query_update(struct nv50_query *q)
 286 {
 287    if (q->is64bit) {
 288       if (nouveau_fence_signalled(q->fence))
 289          q->state = NV50_QUERY_STATE_READY;
 290    } else {
 291       if (q->data[0] == q->sequence)
 292          q->state = NV50_QUERY_STATE_READY;
 293    }
 294 }
 295
 296 static boolean
 297 nv50_query_result(struct pipe_context *pipe, struct pipe_query *pq,
 298                   boolean wait, union pipe_query_result *result)
 299 {
 300    struct nv50_context *nv50 = nv50_context(pipe);
 301    struct nv50_query *q = nv50_query(pq);
 302    uint64_t *res64 = (uint64_t *)result;
 303    uint32_t *res32 = (uint32_t *)result;
 304    uint8_t *res8 = (uint8_t *)result;
 305    uint64_t *data64 = (uint64_t *)q->data;
 306    int i;
 307
 308    if (q->state != NV50_QUERY_STATE_READY)
 309       nv50_query_update(q);
 310
 311    if (q->state != NV50_QUERY_STATE_READY) {
 312       if (!wait) {
 313          /* for broken apps that spin on GL_QUERY_RESULT_AVAILABLE */
 314          if (q->state != NV50_QUERY_STATE_FLUSHED) {
 315             q->state = NV50_QUERY_STATE_FLUSHED;
 316             PUSH_KICK(nv50->base.pushbuf);
 317          }
 318          return false;
 319       }
 320       if (nouveau_bo_wait(q->bo, NOUVEAU_BO_RD, nv50->screen->base.client))
 321          return false;
 322    }
 323    q->state = NV50_QUERY_STATE_READY;
 324
 325    switch (q->type) {
 326    case PIPE_QUERY_GPU_FINISHED:
 327       res8[0] = true;
 328       break;
 329    case PIPE_QUERY_OCCLUSION_COUNTER: /* u32 sequence, u32 count, u64 time */
 330       res64[0] = q->data[1] - q->data[5];
 331       break;
 332    case PIPE_QUERY_PRIMITIVES_GENERATED: /* u64 count, u64 time */
 333    case PIPE_QUERY_PRIMITIVES_EMITTED: /* u64 count, u64 time */
 334       res64[0] = data64[0] - data64[2];
 335       break;
 336    case PIPE_QUERY_SO_STATISTICS:
 337       res64[0] = data64[0] - data64[4];
 338       res64[1] = data64[2] - data64[6];
 339       break;
 340    case PIPE_QUERY_PIPELINE_STATISTICS:
 341       for (i = 0; i < 8; ++i)
 342          res64[i] = data64[i * 2] - data64[16 + i * 2];
 343       break;
 344    case PIPE_QUERY_TIMESTAMP:
 345       res64[0] = data64[1];
 346       break;
 347    case PIPE_QUERY_TIMESTAMP_DISJOINT:
 348       res64[0] = 1000000000;
 349       res8[8] = false;
 350       break;
 351    case PIPE_QUERY_TIME_ELAPSED:
 352       res64[0] = data64[1] - data64[3];
 353       break;
 354    case NVA0_QUERY_STREAM_OUTPUT_BUFFER_OFFSET:
 355       res32[0] = q->data[1];
 356       break;
 357    default:
 358       return false;
 359    }
 360
 361    return true;
 362 }
 363
 364 void
 365 nv84_query_fifo_wait(struct nouveau_pushbuf *push, struct pipe_query *pq)
 366 {
 367    struct nv50_query *q = nv50_query(pq);
 368    unsigned offset = q->offset;
 369
 370    PUSH_SPACE(push, 5);
 371    PUSH_REFN (push, q->bo, NOUVEAU_BO_GART | NOUVEAU_BO_RD);
 372    BEGIN_NV04(push, SUBC_3D(NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH), 4);
 373    PUSH_DATAh(push, q->bo->offset + offset);
 374    PUSH_DATA (push, q->bo->offset + offset);
 375    PUSH_DATA (push, q->sequence);
 376    PUSH_DATA (push, NV84_SUBCHAN_SEMAPHORE_TRIGGER_ACQUIRE_EQUAL);
 377 }
 378
 379 static void
 380 nv50_render_condition(struct pipe_context *pipe,
 381                       struct pipe_query *pq,
 382                       boolean condition, uint mode)
 383 {
 384    struct nv50_context *nv50 = nv50_context(pipe);
 385    struct nouveau_pushbuf *push = nv50->base.pushbuf;
 386    struct nv50_query *q;
 387    uint32_t cond;
 388    bool wait =
 389       mode != PIPE_RENDER_COND_NO_WAIT &&
 390       mode != PIPE_RENDER_COND_BY_REGION_NO_WAIT;
 391
 392    if (!pq) {
 393       cond = NV50_3D_COND_MODE_ALWAYS;
 394    }
 395    else {
 396       q = nv50_query(pq);
 397       /* NOTE: comparison of 2 queries only works if both have completed */
 398       switch (q->type) {
 399       case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 400          cond = condition ? NV50_3D_COND_MODE_EQUAL :
 401                             NV50_3D_COND_MODE_NOT_EQUAL;
 402          wait = true;
 403          break;
 404       case PIPE_QUERY_OCCLUSION_COUNTER:
 405       case PIPE_QUERY_OCCLUSION_PREDICATE:
 406          if (likely(!condition)) {
 407             if (unlikely(q->nesting))
 408                cond = wait ? NV50_3D_COND_MODE_NOT_EQUAL :
 409                              NV50_3D_COND_MODE_ALWAYS;
 410             else
 411                cond = NV50_3D_COND_MODE_RES_NON_ZERO;
 412          } else {
 413             cond = wait ? NV50_3D_COND_MODE_EQUAL : NV50_3D_COND_MODE_ALWAYS;
 414          }
 415          break;
 416       default:
 417          assert(!"render condition query not a predicate");
 418          cond = NV50_3D_COND_MODE_ALWAYS;
 419          break;
 420       }
 421    }
 422
 423    nv50->cond_query = pq;
 424    nv50->cond_cond = condition;
 425    nv50->cond_condmode = cond;
 426    nv50->cond_mode = mode;
 427
 428    if (!pq) {
 429       PUSH_SPACE(push, 2);
 430       BEGIN_NV04(push, NV50_3D(COND_MODE), 1);
 431       PUSH_DATA (push, cond);
 432       return;
 433    }
 434
 435    PUSH_SPACE(push, 9);
 436
 437    if (wait) {
 438       BEGIN_NV04(push, SUBC_3D(NV50_GRAPH_SERIALIZE), 1);
 439       PUSH_DATA (push, 0);
 440    }
 441
 442    PUSH_REFN (push, q->bo, NOUVEAU_BO_GART | NOUVEAU_BO_RD);
 443    BEGIN_NV04(push, NV50_3D(COND_ADDRESS_HIGH), 3);
 444    PUSH_DATAh(push, q->bo->offset + q->offset);
 445    PUSH_DATA (push, q->bo->offset + q->offset);
 446    PUSH_DATA (push, cond);
 447
 448    BEGIN_NV04(push, NV50_2D(COND_ADDRESS_HIGH), 2);
 449    PUSH_DATAh(push, q->bo->offset + q->offset);
 450    PUSH_DATA (push, q->bo->offset + q->offset);
 451 }
 452
 453 void
 454 nv50_query_pushbuf_submit(struct nouveau_pushbuf *push,
 455                           struct pipe_query *pq, unsigned result_offset)
 456 {
 457    struct nv50_query *q = nv50_query(pq);
 458
 459    /* XXX: does this exist ? */
 460 #define NV50_IB_ENTRY_1_NO_PREFETCH (0 << (31 - 8))
 461
 462    PUSH_REFN(push, q->bo, NOUVEAU_BO_RD | NOUVEAU_BO_GART);
 463    nouveau_pushbuf_space(push, 0, 0, 1);
 464    nouveau_pushbuf_data(push, q->bo, q->offset + result_offset, 4 |
 465                         NV50_IB_ENTRY_1_NO_PREFETCH);
 466 }
 467
 468 void
 469 nva0_so_target_save_offset(struct pipe_context *pipe,
 470                            struct pipe_stream_output_target *ptarg,
 471                            unsigned index, bool serialize)
 472 {
 473    struct nv50_so_target *targ = nv50_so_target(ptarg);
 474
 475    if (serialize) {
 476       struct nouveau_pushbuf *push = nv50_context(pipe)->base.pushbuf;
 477       PUSH_SPACE(push, 2);
 478       BEGIN_NV04(push, SUBC_3D(NV50_GRAPH_SERIALIZE), 1);
 479       PUSH_DATA (push, 0);
 480    }
 481
 482    nv50_query(targ->pq)->index = index;
 483    nv50_query_end(pipe, targ->pq);
 484 }
 485
 486 void
 487 nv50_init_query_functions(struct nv50_context *nv50)
 488 {
 489    struct pipe_context *pipe = &nv50->base.pipe;
 490
 491    pipe->create_query = nv50_query_create;
 492    pipe->destroy_query = nv50_query_destroy;
 493    pipe->begin_query = nv50_query_begin;
 494    pipe->end_query = nv50_query_end;
 495    pipe->get_query_result = nv50_query_result;
 496    pipe->render_condition = nv50_render_condition;
 497 }