src/mesa/drivers/dri/i965/brw_program.c

   1 /*
   2  Copyright (C) Intel Corp.  2006.  All Rights Reserved.
   3  Intel funded Tungsten Graphics to
   4  develop this 3D driver.
   5
   6  Permission is hereby granted, free of charge, to any person obtaining
   7  a copy of this software and associated documentation files (the
   8  "Software"), to deal in the Software without restriction, including
   9  without limitation the rights to use, copy, modify, merge, publish,
  10  distribute, sublicense, and/or sell copies of the Software, and to
  11  permit persons to whom the Software is furnished to do so, subject to
  12  the following conditions:
  13
  14  The above copyright notice and this permission notice (including the
  15  next paragraph) shall be included in all copies or substantial
  16  portions of the Software.
  17
  18  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  19  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  21  IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
  22  LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  23  OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  24  WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25
  26  **********************************************************************/
  27  /*
  28   * Authors:
  29   *   Keith Whitwell <keithw@vmware.com>
  30   */
  31
  32 #include <pthread.h>
  33 #include "main/imports.h"
  34 #include "main/enums.h"
  35 #include "main/shaderobj.h"
  36 #include "program/prog_parameter.h"
  37 #include "program/prog_print.h"
  38 #include "program/program.h"
  39 #include "program/programopt.h"
  40 #include "tnl/tnl.h"
  41 #include "glsl/ralloc.h"
  42 #include "glsl/ir.h"
  43
  44 #include "brw_context.h"
  45 #include "brw_wm.h"
  46
  47 static unsigned
  48 get_new_program_id(struct intel_screen *screen)
  49 {
  50    static pthread_mutex_t m = PTHREAD_MUTEX_INITIALIZER;
  51    pthread_mutex_lock(&m);
  52    unsigned id = screen->program_id++;
  53    pthread_mutex_unlock(&m);
  54    return id;
  55 }
  56
  57 static void brwBindProgram( struct gl_context *ctx,
  58                             GLenum target,
  59                             struct gl_program *prog )
  60 {
  61    struct brw_context *brw = brw_context(ctx);
  62
  63    switch (target) {
  64    case GL_VERTEX_PROGRAM_ARB:
  65       brw->state.dirty.brw |= BRW_NEW_VERTEX_PROGRAM;
  66       break;
  67    case MESA_GEOMETRY_PROGRAM:
  68       brw->state.dirty.brw |= BRW_NEW_GEOMETRY_PROGRAM;
  69       break;
  70    case GL_FRAGMENT_PROGRAM_ARB:
  71       brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
  72       break;
  73    }
  74 }
  75
  76 static struct gl_program *brwNewProgram( struct gl_context *ctx,
  77                                       GLenum target,
  78                                       GLuint id )
  79 {
  80    struct brw_context *brw = brw_context(ctx);
  81
  82    switch (target) {
  83    case GL_VERTEX_PROGRAM_ARB: {
  84       struct brw_vertex_program *prog = CALLOC_STRUCT(brw_vertex_program);
  85       if (prog) {
  86          prog->id = get_new_program_id(brw->intelScreen);
  87
  88          return _mesa_init_vertex_program( ctx, &prog->program,
  89                                              target, id );
  90       }
  91       else
  92          return NULL;
  93    }
  94
  95    case GL_FRAGMENT_PROGRAM_ARB: {
  96       struct brw_fragment_program *prog = CALLOC_STRUCT(brw_fragment_program);
  97       if (prog) {
  98          prog->id = get_new_program_id(brw->intelScreen);
  99
 100          return _mesa_init_fragment_program( ctx, &prog->program,
 101                                              target, id );
 102       }
 103       else
 104          return NULL;
 105    }
 106
 107    case MESA_GEOMETRY_PROGRAM: {
 108       struct brw_geometry_program *prog = CALLOC_STRUCT(brw_geometry_program);
 109       if (prog) {
 110          prog->id = get_new_program_id(brw->intelScreen);
 111
 112          return _mesa_init_geometry_program(ctx, &prog->program, target, id);
 113       } else {
 114          return NULL;
 115       }
 116    }
 117
 118    case GL_COMPUTE_PROGRAM_NV: {
 119       struct brw_compute_program *prog = CALLOC_STRUCT(brw_compute_program);
 120       if (prog) {
 121          prog->id = get_new_program_id(brw->intelScreen);
 122
 123          return _mesa_init_compute_program(ctx, &prog->program, target, id);
 124       } else {
 125          return NULL;
 126       }
 127    }
 128
 129    default:
 130       assert(!"Unsupported target in brwNewProgram()");
 131       return NULL;
 132    }
 133 }
 134
 135 static void brwDeleteProgram( struct gl_context *ctx,
 136                               struct gl_program *prog )
 137 {
 138    _mesa_delete_program( ctx, prog );
 139 }
 140
 141
 142 static GLboolean
 143 brwIsProgramNative(struct gl_context *ctx,
 144                    GLenum target,
 145                    struct gl_program *prog)
 146 {
 147    return true;
 148 }
 149
 150 static GLboolean
 151 brwProgramStringNotify(struct gl_context *ctx,
 152                        GLenum target,
 153                        struct gl_program *prog)
 154 {
 155    struct brw_context *brw = brw_context(ctx);
 156
 157    switch (target) {
 158    case GL_FRAGMENT_PROGRAM_ARB: {
 159       struct gl_fragment_program *fprog = (struct gl_fragment_program *) prog;
 160       struct brw_fragment_program *newFP = brw_fragment_program(fprog);
 161       const struct brw_fragment_program *curFP =
 162          brw_fragment_program_const(brw->fragment_program);
 163
 164       if (newFP == curFP)
 165          brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
 166       newFP->id = get_new_program_id(brw->intelScreen);
 167       break;
 168    }
 169    case GL_VERTEX_PROGRAM_ARB: {
 170       struct gl_vertex_program *vprog = (struct gl_vertex_program *) prog;
 171       struct brw_vertex_program *newVP = brw_vertex_program(vprog);
 172       const struct brw_vertex_program *curVP =
 173          brw_vertex_program_const(brw->vertex_program);
 174
 175       if (newVP == curVP)
 176          brw->state.dirty.brw |= BRW_NEW_VERTEX_PROGRAM;
 177       if (newVP->program.IsPositionInvariant) {
 178          _mesa_insert_mvp_code(ctx, &newVP->program);
 179       }
 180       newVP->id = get_new_program_id(brw->intelScreen);
 181
 182       /* Also tell tnl about it:
 183        */
 184       _tnl_program_string(ctx, target, prog);
 185       break;
 186    }
 187    default:
 188       /*
 189        * driver->ProgramStringNotify is only called for ARB programs, fixed
 190        * function vertex programs, and ir_to_mesa (which isn't used by the
 191        * i965 back-end).  Therefore, even after geometry shaders are added,
 192        * this function should only ever be called with a target of
 193        * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
 194        */
 195       assert(!"Unexpected target in brwProgramStringNotify");
 196       break;
 197    }
 198
 199    brw_add_texrect_params(prog);
 200
 201    return true;
 202 }
 203
 204 void
 205 brw_add_texrect_params(struct gl_program *prog)
 206 {
 207    for (int texunit = 0; texunit < BRW_MAX_TEX_UNIT; texunit++) {
 208       if (!(prog->TexturesUsed[texunit] & (1 << TEXTURE_RECT_INDEX)))
 209          continue;
 210
 211       int tokens[STATE_LENGTH] = {
 212          STATE_INTERNAL,
 213          STATE_TEXRECT_SCALE,
 214          texunit,
 215          0,
 216          0
 217       };
 218
 219       _mesa_add_state_reference(prog->Parameters, (gl_state_index *)tokens);
 220    }
 221 }
 222
 223 /* Per-thread scratch space is a power-of-two multiple of 1KB. */
 224 int
 225 brw_get_scratch_size(int size)
 226 {
 227    int i;
 228
 229    for (i = 1024; i < size; i *= 2)
 230       ;
 231
 232    return i;
 233 }
 234
 235 void
 236 brw_get_scratch_bo(struct brw_context *brw,
 237                    drm_intel_bo **scratch_bo, int size)
 238 {
 239    drm_intel_bo *old_bo = *scratch_bo;
 240
 241    if (old_bo && old_bo->size < size) {
 242       drm_intel_bo_unreference(old_bo);
 243       old_bo = NULL;
 244    }
 245
 246    if (!old_bo) {
 247       *scratch_bo = drm_intel_bo_alloc(brw->bufmgr, "scratch bo", size, 4096);
 248    }
 249 }
 250
 251 void brwInitFragProgFuncs( struct dd_function_table *functions )
 252 {
 253    assert(functions->ProgramStringNotify == _tnl_program_string);
 254
 255    functions->BindProgram = brwBindProgram;
 256    functions->NewProgram = brwNewProgram;
 257    functions->DeleteProgram = brwDeleteProgram;
 258    functions->IsProgramNative = brwIsProgramNative;
 259    functions->ProgramStringNotify = brwProgramStringNotify;
 260
 261    functions->NewShader = brw_new_shader;
 262    functions->NewShaderProgram = brw_new_shader_program;
 263    functions->LinkShader = brw_link_shader;
 264 }
 265
 266 void
 267 brw_init_shader_time(struct brw_context *brw)
 268 {
 269    const int max_entries = 4096;
 270    brw->shader_time.bo = drm_intel_bo_alloc(brw->bufmgr, "shader time",
 271                                             max_entries * SHADER_TIME_STRIDE,
 272                                             4096);
 273    brw->shader_time.shader_programs = rzalloc_array(brw, struct gl_shader_program *,
 274                                                     max_entries);
 275    brw->shader_time.programs = rzalloc_array(brw, struct gl_program *,
 276                                              max_entries);
 277    brw->shader_time.types = rzalloc_array(brw, enum shader_time_shader_type,
 278                                           max_entries);
 279    brw->shader_time.cumulative = rzalloc_array(brw, uint64_t,
 280                                                max_entries);
 281    brw->shader_time.max_entries = max_entries;
 282 }
 283
 284 static int
 285 compare_time(const void *a, const void *b)
 286 {
 287    uint64_t * const *a_val = a;
 288    uint64_t * const *b_val = b;
 289
 290    /* We don't just subtract because we're turning the value to an int. */
 291    if (**a_val < **b_val)
 292       return -1;
 293    else if (**a_val == **b_val)
 294       return 0;
 295    else
 296       return 1;
 297 }
 298
 299 static void
 300 get_written_and_reset(struct brw_context *brw, int i,
 301                       uint64_t *written, uint64_t *reset)
 302 {
 303    enum shader_time_shader_type type = brw->shader_time.types[i];
 304    assert(type == ST_VS || type == ST_GS || type == ST_FS8 || type == ST_FS16);
 305
 306    /* Find where we recorded written and reset. */
 307    int wi, ri;
 308
 309    for (wi = i; brw->shader_time.types[wi] != type + 1; wi++)
 310       ;
 311
 312    for (ri = i; brw->shader_time.types[ri] != type + 2; ri++)
 313       ;
 314
 315    *written = brw->shader_time.cumulative[wi];
 316    *reset = brw->shader_time.cumulative[ri];
 317 }
 318
 319 static void
 320 print_shader_time_line(const char *stage, const char *name,
 321                        int shader_num, uint64_t time, uint64_t total)
 322 {
 323    fprintf(stderr, "%-6s%-18s", stage, name);
 324
 325    if (shader_num != -1)
 326       fprintf(stderr, "%4d: ", shader_num);
 327    else
 328       fprintf(stderr, "    : ");
 329
 330    fprintf(stderr, "%16lld (%7.2f Gcycles)      %4.1f%%\n",
 331            (long long)time,
 332            (double)time / 1000000000.0,
 333            (double)time / total * 100.0);
 334 }
 335
 336 static void
 337 brw_report_shader_time(struct brw_context *brw)
 338 {
 339    if (!brw->shader_time.bo || !brw->shader_time.num_entries)
 340       return;
 341
 342    uint64_t scaled[brw->shader_time.num_entries];
 343    uint64_t *sorted[brw->shader_time.num_entries];
 344    uint64_t total_by_type[ST_FS16 + 1];
 345    memset(total_by_type, 0, sizeof(total_by_type));
 346    double total = 0;
 347    for (int i = 0; i < brw->shader_time.num_entries; i++) {
 348       uint64_t written = 0, reset = 0;
 349       enum shader_time_shader_type type = brw->shader_time.types[i];
 350
 351       sorted[i] = &scaled[i];
 352
 353       switch (type) {
 354       case ST_VS_WRITTEN:
 355       case ST_VS_RESET:
 356       case ST_GS_WRITTEN:
 357       case ST_GS_RESET:
 358       case ST_FS8_WRITTEN:
 359       case ST_FS8_RESET:
 360       case ST_FS16_WRITTEN:
 361       case ST_FS16_RESET:
 362          /* We'll handle these when along with the time. */
 363          scaled[i] = 0;
 364          continue;
 365
 366       case ST_VS:
 367       case ST_GS:
 368       case ST_FS8:
 369       case ST_FS16:
 370          get_written_and_reset(brw, i, &written, &reset);
 371          break;
 372
 373       default:
 374          /* I sometimes want to print things that aren't the 3 shader times.
 375           * Just print the sum in that case.
 376           */
 377          written = 1;
 378          reset = 0;
 379          break;
 380       }
 381
 382       uint64_t time = brw->shader_time.cumulative[i];
 383       if (written) {
 384          scaled[i] = time / written * (written + reset);
 385       } else {
 386          scaled[i] = time;
 387       }
 388
 389       switch (type) {
 390       case ST_VS:
 391       case ST_GS:
 392       case ST_FS8:
 393       case ST_FS16:
 394          total_by_type[type] += scaled[i];
 395          break;
 396       default:
 397          break;
 398       }
 399
 400       total += scaled[i];
 401    }
 402
 403    if (total == 0) {
 404       fprintf(stderr, "No shader time collected yet\n");
 405       return;
 406    }
 407
 408    qsort(sorted, brw->shader_time.num_entries, sizeof(sorted[0]), compare_time);
 409
 410    fprintf(stderr, "\n");
 411    fprintf(stderr, "type          ID                  cycles spent                   %% of total\n");
 412    for (int s = 0; s < brw->shader_time.num_entries; s++) {
 413       const char *shader_name;
 414       const char *stage;
 415       /* Work back from the sorted pointers times to a time to print. */
 416       int i = sorted[s] - scaled;
 417       struct gl_shader_program *prog = brw->shader_time.shader_programs[i];
 418
 419       if (scaled[i] == 0)
 420          continue;
 421
 422       int shader_num = -1;
 423       if (prog) {
 424          shader_num = prog->Name;
 425
 426          /* The fixed function fragment shader generates GLSL IR with a Name
 427           * of 0, and nothing else does.
 428           */
 429          if (prog->Label) {
 430             shader_name = prog->Label;
 431          } else if (shader_num == 0 &&
 432              (brw->shader_time.types[i] == ST_FS8 ||
 433               brw->shader_time.types[i] == ST_FS16)) {
 434             shader_name = "ff";
 435             shader_num = -1;
 436          } else {
 437             shader_name = "glsl";
 438          }
 439       } else if (brw->shader_time.programs[i]) {
 440          shader_num = brw->shader_time.programs[i]->Id;
 441          if (shader_num == 0) {
 442             shader_name = "ff";
 443             shader_num = -1;
 444          } else {
 445             shader_name = "prog";
 446          }
 447       } else {
 448          shader_name = "other";
 449       }
 450
 451       switch (brw->shader_time.types[i]) {
 452       case ST_VS:
 453          stage = "vs";
 454          break;
 455       case ST_GS:
 456          stage = "gs";
 457          break;
 458       case ST_FS8:
 459          stage = "fs8";
 460          break;
 461       case ST_FS16:
 462          stage = "fs16";
 463          break;
 464       default:
 465          stage = "other";
 466          break;
 467       }
 468
 469       print_shader_time_line(stage, shader_name, shader_num,
 470                              scaled[i], total);
 471    }
 472
 473    fprintf(stderr, "\n");
 474    print_shader_time_line("total", "vs", -1, total_by_type[ST_VS], total);
 475    print_shader_time_line("total", "gs", -1, total_by_type[ST_GS], total);
 476    print_shader_time_line("total", "fs8", -1, total_by_type[ST_FS8], total);
 477    print_shader_time_line("total", "fs16", -1, total_by_type[ST_FS16], total);
 478 }
 479
 480 static void
 481 brw_collect_shader_time(struct brw_context *brw)
 482 {
 483    if (!brw->shader_time.bo)
 484       return;
 485
 486    /* This probably stalls on the last rendering.  We could fix that by
 487     * delaying reading the reports, but it doesn't look like it's a big
 488     * overhead compared to the cost of tracking the time in the first place.
 489     */
 490    drm_intel_bo_map(brw->shader_time.bo, true);
 491
 492    uint32_t *times = brw->shader_time.bo->virtual;
 493
 494    for (int i = 0; i < brw->shader_time.num_entries; i++) {
 495       brw->shader_time.cumulative[i] += times[i * SHADER_TIME_STRIDE / 4];
 496    }
 497
 498    /* Zero the BO out to clear it out for our next collection.
 499     */
 500    memset(times, 0, brw->shader_time.bo->size);
 501    drm_intel_bo_unmap(brw->shader_time.bo);
 502 }
 503
 504 void
 505 brw_collect_and_report_shader_time(struct brw_context *brw)
 506 {
 507    brw_collect_shader_time(brw);
 508
 509    if (brw->shader_time.report_time == 0 ||
 510        get_time() - brw->shader_time.report_time >= 1.0) {
 511       brw_report_shader_time(brw);
 512       brw->shader_time.report_time = get_time();
 513    }
 514 }
 515
 516 /**
 517  * Chooses an index in the shader_time buffer and sets up tracking information
 518  * for our printouts.
 519  *
 520  * Note that this holds on to references to the underlying programs, which may
 521  * change their lifetimes compared to normal operation.
 522  */
 523 int
 524 brw_get_shader_time_index(struct brw_context *brw,
 525                           struct gl_shader_program *shader_prog,
 526                           struct gl_program *prog,
 527                           enum shader_time_shader_type type)
 528 {
 529    struct gl_context *ctx = &brw->ctx;
 530
 531    int shader_time_index = brw->shader_time.num_entries++;
 532    assert(shader_time_index < brw->shader_time.max_entries);
 533    brw->shader_time.types[shader_time_index] = type;
 534
 535    _mesa_reference_shader_program(ctx,
 536                                   &brw->shader_time.shader_programs[shader_time_index],
 537                                   shader_prog);
 538
 539    _mesa_reference_program(ctx,
 540                            &brw->shader_time.programs[shader_time_index],
 541                            prog);
 542
 543    return shader_time_index;
 544 }
 545
 546 void
 547 brw_destroy_shader_time(struct brw_context *brw)
 548 {
 549    drm_intel_bo_unreference(brw->shader_time.bo);
 550    brw->shader_time.bo = NULL;
 551 }
 552
 553 void
 554 brw_mark_surface_used(struct brw_stage_prog_data *prog_data,
 555                       unsigned surf_index)
 556 {
 557    assert(surf_index < BRW_MAX_SURFACES);
 558
 559    prog_data->binding_table.size_bytes =
 560       MAX2(prog_data->binding_table.size_bytes, (surf_index + 1) * 4);
 561 }
 562
 563 bool
 564 brw_stage_prog_data_compare(const struct brw_stage_prog_data *a,
 565                             const struct brw_stage_prog_data *b)
 566 {
 567    /* Compare all the struct up to the pointers. */
 568    if (memcmp(a, b, offsetof(struct brw_stage_prog_data, param)))
 569       return false;
 570
 571    if (memcmp(a->param, b->param, a->nr_params * sizeof(void *)))
 572       return false;
 573
 574    if (memcmp(a->pull_param, b->pull_param, a->nr_pull_params * sizeof(void *)))
 575       return false;
 576
 577    return true;
 578 }
 579
 580 void
 581 brw_stage_prog_data_free(const void *p)
 582 {
 583    struct brw_stage_prog_data *prog_data = (struct brw_stage_prog_data *)p;
 584
 585    ralloc_free(prog_data->param);
 586    ralloc_free(prog_data->pull_param);
 587 }
 588
 589 void
 590 brw_dump_ir(struct brw_context *brw, const char *stage,
 591             struct gl_shader_program *shader_prog,
 592             struct gl_shader *shader, struct gl_program *prog)
 593 {
 594    if (shader_prog) {
 595       fprintf(stderr,
 596               "GLSL IR for native %s shader %d:\n", stage, shader_prog->Name);
 597       _mesa_print_ir(stderr, shader->ir, NULL);
 598       fprintf(stderr, "\n\n");
 599    } else {
 600       fprintf(stderr, "ARB_%s_program %d ir for native %s shader\n",
 601               stage, prog->Id, stage);
 602       _mesa_print_program(prog);
 603    }
 604 }