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