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