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 "util/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       unreachable("Unsupported target in brwNewProgram()");
 131    }
 132 }
 133
 134 static void brwDeleteProgram( struct gl_context *ctx,
 135                               struct gl_program *prog )
 136 {
 137    _mesa_delete_program( ctx, prog );
 138 }
 139
 140
 141 static GLboolean
 142 brwIsProgramNative(struct gl_context *ctx,
 143                    GLenum target,
 144                    struct gl_program *prog)
 145 {
 146    return true;
 147 }
 148
 149 static GLboolean
 150 brwProgramStringNotify(struct gl_context *ctx,
 151                        GLenum target,
 152                        struct gl_program *prog)
 153 {
 154    struct brw_context *brw = brw_context(ctx);
 155
 156    switch (target) {
 157    case GL_FRAGMENT_PROGRAM_ARB: {
 158       struct gl_fragment_program *fprog = (struct gl_fragment_program *) prog;
 159       struct brw_fragment_program *newFP = brw_fragment_program(fprog);
 160       const struct brw_fragment_program *curFP =
 161          brw_fragment_program_const(brw->fragment_program);
 162
 163       if (newFP == curFP)
 164          brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
 165       newFP->id = get_new_program_id(brw->intelScreen);
 166       break;
 167    }
 168    case GL_VERTEX_PROGRAM_ARB: {
 169       struct gl_vertex_program *vprog = (struct gl_vertex_program *) prog;
 170       struct brw_vertex_program *newVP = brw_vertex_program(vprog);
 171       const struct brw_vertex_program *curVP =
 172          brw_vertex_program_const(brw->vertex_program);
 173
 174       if (newVP == curVP)
 175          brw->state.dirty.brw |= BRW_NEW_VERTEX_PROGRAM;
 176       if (newVP->program.IsPositionInvariant) {
 177          _mesa_insert_mvp_code(ctx, &newVP->program);
 178       }
 179       newVP->id = get_new_program_id(brw->intelScreen);
 180
 181       /* Also tell tnl about it:
 182        */
 183       _tnl_program_string(ctx, target, prog);
 184       break;
 185    }
 186    default:
 187       /*
 188        * driver->ProgramStringNotify is only called for ARB programs, fixed
 189        * function vertex programs, and ir_to_mesa (which isn't used by the
 190        * i965 back-end).  Therefore, even after geometry shaders are added,
 191        * this function should only ever be called with a target of
 192        * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
 193        */
 194       unreachable("Unexpected target in brwProgramStringNotify");
 195    }
 196
 197    brw_add_texrect_params(prog);
 198
 199    return true;
 200 }
 201
 202 void
 203 brw_add_texrect_params(struct gl_program *prog)
 204 {
 205    for (int texunit = 0; texunit < BRW_MAX_TEX_UNIT; texunit++) {
 206       if (!(prog->TexturesUsed[texunit] & (1 << TEXTURE_RECT_INDEX)))
 207          continue;
 208
 209       int tokens[STATE_LENGTH] = {
 210          STATE_INTERNAL,
 211          STATE_TEXRECT_SCALE,
 212          texunit,
 213          0,
 214          0
 215       };
 216
 217       _mesa_add_state_reference(prog->Parameters, (gl_state_index *)tokens);
 218    }
 219 }
 220
 221 /* Per-thread scratch space is a power-of-two multiple of 1KB. */
 222 int
 223 brw_get_scratch_size(int size)
 224 {
 225    int i;
 226
 227    for (i = 1024; i < size; i *= 2)
 228       ;
 229
 230    return i;
 231 }
 232
 233 void
 234 brw_get_scratch_bo(struct brw_context *brw,
 235                    drm_intel_bo **scratch_bo, int size)
 236 {
 237    drm_intel_bo *old_bo = *scratch_bo;
 238
 239    if (old_bo && old_bo->size < size) {
 240       drm_intel_bo_unreference(old_bo);
 241       old_bo = NULL;
 242    }
 243
 244    if (!old_bo) {
 245       *scratch_bo = drm_intel_bo_alloc(brw->bufmgr, "scratch bo", size, 4096);
 246    }
 247 }
 248
 249 void brwInitFragProgFuncs( struct dd_function_table *functions )
 250 {
 251    assert(functions->ProgramStringNotify == _tnl_program_string);
 252
 253    functions->BindProgram = brwBindProgram;
 254    functions->NewProgram = brwNewProgram;
 255    functions->DeleteProgram = brwDeleteProgram;
 256    functions->IsProgramNative = brwIsProgramNative;
 257    functions->ProgramStringNotify = brwProgramStringNotify;
 258
 259    functions->NewShader = brw_new_shader;
 260    functions->NewShaderProgram = brw_new_shader_program;
 261    functions->LinkShader = brw_link_shader;
 262 }
 263
 264 void
 265 brw_init_shader_time(struct brw_context *brw)
 266 {
 267    const int max_entries = 4096;
 268    brw->shader_time.bo = drm_intel_bo_alloc(brw->bufmgr, "shader time",
 269                                             max_entries * SHADER_TIME_STRIDE,
 270                                             4096);
 271    brw->shader_time.shader_programs = rzalloc_array(brw, struct gl_shader_program *,
 272                                                     max_entries);
 273    brw->shader_time.programs = rzalloc_array(brw, struct gl_program *,
 274                                              max_entries);
 275    brw->shader_time.types = rzalloc_array(brw, enum shader_time_shader_type,
 276                                           max_entries);
 277    brw->shader_time.cumulative = rzalloc_array(brw, uint64_t,
 278                                                max_entries);
 279    brw->shader_time.max_entries = max_entries;
 280 }
 281
 282 static int
 283 compare_time(const void *a, const void *b)
 284 {
 285    uint64_t * const *a_val = a;
 286    uint64_t * const *b_val = b;
 287
 288    /* We don't just subtract because we're turning the value to an int. */
 289    if (**a_val < **b_val)
 290       return -1;
 291    else if (**a_val == **b_val)
 292       return 0;
 293    else
 294       return 1;
 295 }
 296
 297 static void
 298 get_written_and_reset(struct brw_context *brw, int i,
 299                       uint64_t *written, uint64_t *reset)
 300 {
 301    enum shader_time_shader_type type = brw->shader_time.types[i];
 302    assert(type == ST_VS || type == ST_GS || type == ST_FS8 || type == ST_FS16);
 303
 304    /* Find where we recorded written and reset. */
 305    int wi, ri;
 306
 307    for (wi = i; brw->shader_time.types[wi] != type + 1; wi++)
 308       ;
 309
 310    for (ri = i; brw->shader_time.types[ri] != type + 2; ri++)
 311       ;
 312
 313    *written = brw->shader_time.cumulative[wi];
 314    *reset = brw->shader_time.cumulative[ri];
 315 }
 316
 317 static void
 318 print_shader_time_line(const char *stage, const char *name,
 319                        int shader_num, uint64_t time, uint64_t total)
 320 {
 321    fprintf(stderr, "%-6s%-18s", stage, name);
 322
 323    if (shader_num != -1)
 324       fprintf(stderr, "%4d: ", shader_num);
 325    else
 326       fprintf(stderr, "    : ");
 327
 328    fprintf(stderr, "%16lld (%7.2f Gcycles)      %4.1f%%\n",
 329            (long long)time,
 330            (double)time / 1000000000.0,
 331            (double)time / total * 100.0);
 332 }
 333
 334 static void
 335 brw_report_shader_time(struct brw_context *brw)
 336 {
 337    if (!brw->shader_time.bo || !brw->shader_time.num_entries)
 338       return;
 339
 340    uint64_t scaled[brw->shader_time.num_entries];
 341    uint64_t *sorted[brw->shader_time.num_entries];
 342    uint64_t total_by_type[ST_FS16 + 1];
 343    memset(total_by_type, 0, sizeof(total_by_type));
 344    double total = 0;
 345    for (int i = 0; i < brw->shader_time.num_entries; i++) {
 346       uint64_t written = 0, reset = 0;
 347       enum shader_time_shader_type type = brw->shader_time.types[i];
 348
 349       sorted[i] = &scaled[i];
 350
 351       switch (type) {
 352       case ST_VS_WRITTEN:
 353       case ST_VS_RESET:
 354       case ST_GS_WRITTEN:
 355       case ST_GS_RESET:
 356       case ST_FS8_WRITTEN:
 357       case ST_FS8_RESET:
 358       case ST_FS16_WRITTEN:
 359       case ST_FS16_RESET:
 360          /* We'll handle these when along with the time. */
 361          scaled[i] = 0;
 362          continue;
 363
 364       case ST_VS:
 365       case ST_GS:
 366       case ST_FS8:
 367       case ST_FS16:
 368          get_written_and_reset(brw, i, &written, &reset);
 369          break;
 370
 371       default:
 372          /* I sometimes want to print things that aren't the 3 shader times.
 373           * Just print the sum in that case.
 374           */
 375          written = 1;
 376          reset = 0;
 377          break;
 378       }
 379
 380       uint64_t time = brw->shader_time.cumulative[i];
 381       if (written) {
 382          scaled[i] = time / written * (written + reset);
 383       } else {
 384          scaled[i] = time;
 385       }
 386
 387       switch (type) {
 388       case ST_VS:
 389       case ST_GS:
 390       case ST_FS8:
 391       case ST_FS16:
 392          total_by_type[type] += scaled[i];
 393          break;
 394       default:
 395          break;
 396       }
 397
 398       total += scaled[i];
 399    }
 400
 401    if (total == 0) {
 402       fprintf(stderr, "No shader time collected yet\n");
 403       return;
 404    }
 405
 406    qsort(sorted, brw->shader_time.num_entries, sizeof(sorted[0]), compare_time);
 407
 408    fprintf(stderr, "\n");
 409    fprintf(stderr, "type          ID                  cycles spent                   %% of total\n");
 410    for (int s = 0; s < brw->shader_time.num_entries; s++) {
 411       const char *shader_name;
 412       const char *stage;
 413       /* Work back from the sorted pointers times to a time to print. */
 414       int i = sorted[s] - scaled;
 415       struct gl_shader_program *prog = brw->shader_time.shader_programs[i];
 416
 417       if (scaled[i] == 0)
 418          continue;
 419
 420       int shader_num = -1;
 421       if (prog) {
 422          shader_num = prog->Name;
 423
 424          /* The fixed function fragment shader generates GLSL IR with a Name
 425           * of 0, and nothing else does.
 426           */
 427          if (prog->Label) {
 428             shader_name = prog->Label;
 429          } else if (shader_num == 0 &&
 430              (brw->shader_time.types[i] == ST_FS8 ||
 431               brw->shader_time.types[i] == ST_FS16)) {
 432             shader_name = "ff";
 433             shader_num = -1;
 434          } else {
 435             shader_name = "glsl";
 436          }
 437       } else if (brw->shader_time.programs[i]) {
 438          shader_num = brw->shader_time.programs[i]->Id;
 439          if (shader_num == 0) {
 440             shader_name = "ff";
 441             shader_num = -1;
 442          } else {
 443             shader_name = "prog";
 444          }
 445       } else {
 446          shader_name = "other";
 447       }
 448
 449       switch (brw->shader_time.types[i]) {
 450       case ST_VS:
 451          stage = "vs";
 452          break;
 453       case ST_GS:
 454          stage = "gs";
 455          break;
 456       case ST_FS8:
 457          stage = "fs8";
 458          break;
 459       case ST_FS16:
 460          stage = "fs16";
 461          break;
 462       default:
 463          stage = "other";
 464          break;
 465       }
 466
 467       print_shader_time_line(stage, shader_name, shader_num,
 468                              scaled[i], total);
 469    }
 470
 471    fprintf(stderr, "\n");
 472    print_shader_time_line("total", "vs", -1, total_by_type[ST_VS], total);
 473    print_shader_time_line("total", "gs", -1, total_by_type[ST_GS], total);
 474    print_shader_time_line("total", "fs8", -1, total_by_type[ST_FS8], total);
 475    print_shader_time_line("total", "fs16", -1, total_by_type[ST_FS16], total);
 476 }
 477
 478 static void
 479 brw_collect_shader_time(struct brw_context *brw)
 480 {
 481    if (!brw->shader_time.bo)
 482       return;
 483
 484    /* This probably stalls on the last rendering.  We could fix that by
 485     * delaying reading the reports, but it doesn't look like it's a big
 486     * overhead compared to the cost of tracking the time in the first place.
 487     */
 488    drm_intel_bo_map(brw->shader_time.bo, true);
 489
 490    uint32_t *times = brw->shader_time.bo->virtual;
 491
 492    for (int i = 0; i < brw->shader_time.num_entries; i++) {
 493       brw->shader_time.cumulative[i] += times[i * SHADER_TIME_STRIDE / 4];
 494    }
 495
 496    /* Zero the BO out to clear it out for our next collection.
 497     */
 498    memset(times, 0, brw->shader_time.bo->size);
 499    drm_intel_bo_unmap(brw->shader_time.bo);
 500 }
 501
 502 void
 503 brw_collect_and_report_shader_time(struct brw_context *brw)
 504 {
 505    brw_collect_shader_time(brw);
 506
 507    if (brw->shader_time.report_time == 0 ||
 508        get_time() - brw->shader_time.report_time >= 1.0) {
 509       brw_report_shader_time(brw);
 510       brw->shader_time.report_time = get_time();
 511    }
 512 }
 513
 514 /**
 515  * Chooses an index in the shader_time buffer and sets up tracking information
 516  * for our printouts.
 517  *
 518  * Note that this holds on to references to the underlying programs, which may
 519  * change their lifetimes compared to normal operation.
 520  */
 521 int
 522 brw_get_shader_time_index(struct brw_context *brw,
 523                           struct gl_shader_program *shader_prog,
 524                           struct gl_program *prog,
 525                           enum shader_time_shader_type type)
 526 {
 527    struct gl_context *ctx = &brw->ctx;
 528
 529    int shader_time_index = brw->shader_time.num_entries++;
 530    assert(shader_time_index < brw->shader_time.max_entries);
 531    brw->shader_time.types[shader_time_index] = type;
 532
 533    _mesa_reference_shader_program(ctx,
 534                                   &brw->shader_time.shader_programs[shader_time_index],
 535                                   shader_prog);
 536
 537    _mesa_reference_program(ctx,
 538                            &brw->shader_time.programs[shader_time_index],
 539                            prog);
 540
 541    return shader_time_index;
 542 }
 543
 544 void
 545 brw_destroy_shader_time(struct brw_context *brw)
 546 {
 547    drm_intel_bo_unreference(brw->shader_time.bo);
 548    brw->shader_time.bo = NULL;
 549 }
 550
 551 void
 552 brw_mark_surface_used(struct brw_stage_prog_data *prog_data,
 553                       unsigned surf_index)
 554 {
 555    assert(surf_index < BRW_MAX_SURFACES);
 556
 557    prog_data->binding_table.size_bytes =
 558       MAX2(prog_data->binding_table.size_bytes, (surf_index + 1) * 4);
 559 }
 560
 561 bool
 562 brw_stage_prog_data_compare(const struct brw_stage_prog_data *a,
 563                             const struct brw_stage_prog_data *b)
 564 {
 565    /* Compare all the struct up to the pointers. */
 566    if (memcmp(a, b, offsetof(struct brw_stage_prog_data, param)))
 567       return false;
 568
 569    if (memcmp(a->param, b->param, a->nr_params * sizeof(void *)))
 570       return false;
 571
 572    if (memcmp(a->pull_param, b->pull_param, a->nr_pull_params * sizeof(void *)))
 573       return false;
 574
 575    return true;
 576 }
 577
 578 void
 579 brw_stage_prog_data_free(const void *p)
 580 {
 581    struct brw_stage_prog_data *prog_data = (struct brw_stage_prog_data *)p;
 582
 583    ralloc_free(prog_data->param);
 584    ralloc_free(prog_data->pull_param);
 585 }
 586
 587 void
 588 brw_dump_ir(struct brw_context *brw, const char *stage,
 589             struct gl_shader_program *shader_prog,
 590             struct gl_shader *shader, struct gl_program *prog)
 591 {
 592    if (shader_prog) {
 593       fprintf(stderr,
 594               "GLSL IR for native %s shader %d:\n", stage, shader_prog->Name);
 595       _mesa_print_ir(stderr, shader->ir, NULL);
 596       fprintf(stderr, "\n\n");
 597    } else {
 598       fprintf(stderr, "ARB_%s_program %d ir for native %s shader\n",
 599               stage, prog->Id, stage);
 600       _mesa_print_program(prog);
 601    }
 602 }