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/program.h"
  38 #include "program/programopt.h"
  39 #include "tnl/tnl.h"
  40 #include "glsl/ralloc.h"
  41
  42 #include "brw_context.h"
  43 #include "brw_wm.h"
  44
  45 static unsigned
  46 get_new_program_id(struct intel_screen *screen)
  47 {
  48    static pthread_mutex_t m = PTHREAD_MUTEX_INITIALIZER;
  49    pthread_mutex_lock(&m);
  50    unsigned id = screen->program_id++;
  51    pthread_mutex_unlock(&m);
  52    return id;
  53 }
  54
  55 static void brwBindProgram( struct gl_context *ctx,
  56                             GLenum target,
  57                             struct gl_program *prog )
  58 {
  59    struct brw_context *brw = brw_context(ctx);
  60
  61    switch (target) {
  62    case GL_VERTEX_PROGRAM_ARB:
  63       brw->state.dirty.brw |= BRW_NEW_VERTEX_PROGRAM;
  64       break;
  65    case MESA_GEOMETRY_PROGRAM:
  66       brw->state.dirty.brw |= BRW_NEW_GEOMETRY_PROGRAM;
  67       break;
  68    case GL_FRAGMENT_PROGRAM_ARB:
  69       brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
  70       break;
  71    }
  72 }
  73
  74 static struct gl_program *brwNewProgram( struct gl_context *ctx,
  75                                       GLenum target,
  76                                       GLuint id )
  77 {
  78    struct brw_context *brw = brw_context(ctx);
  79
  80    switch (target) {
  81    case GL_VERTEX_PROGRAM_ARB: {
  82       struct brw_vertex_program *prog = CALLOC_STRUCT(brw_vertex_program);
  83       if (prog) {
  84          prog->id = get_new_program_id(brw->intelScreen);
  85
  86          return _mesa_init_vertex_program( ctx, &prog->program,
  87                                              target, id );
  88       }
  89       else
  90          return NULL;
  91    }
  92
  93    case GL_FRAGMENT_PROGRAM_ARB: {
  94       struct brw_fragment_program *prog = CALLOC_STRUCT(brw_fragment_program);
  95       if (prog) {
  96          prog->id = get_new_program_id(brw->intelScreen);
  97
  98          return _mesa_init_fragment_program( ctx, &prog->program,
  99                                              target, id );
 100       }
 101       else
 102          return NULL;
 103    }
 104
 105    case MESA_GEOMETRY_PROGRAM: {
 106       struct brw_geometry_program *prog = CALLOC_STRUCT(brw_geometry_program);
 107       if (prog) {
 108          prog->id = get_new_program_id(brw->intelScreen);
 109
 110          return _mesa_init_geometry_program(ctx, &prog->program, target, id);
 111       } else {
 112          return NULL;
 113       }
 114    }
 115
 116    case GL_COMPUTE_PROGRAM_NV: {
 117       struct brw_compute_program *prog = CALLOC_STRUCT(brw_compute_program);
 118       if (prog) {
 119          prog->id = get_new_program_id(brw->intelScreen);
 120
 121          return _mesa_init_compute_program(ctx, &prog->program, target, id);
 122       } else {
 123          return NULL;
 124       }
 125    }
 126
 127    default:
 128       assert(!"Unsupported target in brwNewProgram()");
 129       return NULL;
 130    }
 131 }
 132
 133 static void brwDeleteProgram( struct gl_context *ctx,
 134                               struct gl_program *prog )
 135 {
 136    _mesa_delete_program( ctx, prog );
 137 }
 138
 139
 140 static GLboolean
 141 brwIsProgramNative(struct gl_context *ctx,
 142                    GLenum target,
 143                    struct gl_program *prog)
 144 {
 145    return true;
 146 }
 147
 148 static GLboolean
 149 brwProgramStringNotify(struct gl_context *ctx,
 150                        GLenum target,
 151                        struct gl_program *prog)
 152 {
 153    struct brw_context *brw = brw_context(ctx);
 154
 155    switch (target) {
 156    case GL_FRAGMENT_PROGRAM_ARB: {
 157       struct gl_fragment_program *fprog = (struct gl_fragment_program *) prog;
 158       struct brw_fragment_program *newFP = brw_fragment_program(fprog);
 159       const struct brw_fragment_program *curFP =
 160          brw_fragment_program_const(brw->fragment_program);
 161
 162       if (newFP == curFP)
 163          brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
 164       newFP->id = get_new_program_id(brw->intelScreen);
 165       break;
 166    }
 167    case GL_VERTEX_PROGRAM_ARB: {
 168       struct gl_vertex_program *vprog = (struct gl_vertex_program *) prog;
 169       struct brw_vertex_program *newVP = brw_vertex_program(vprog);
 170       const struct brw_vertex_program *curVP =
 171          brw_vertex_program_const(brw->vertex_program);
 172
 173       if (newVP == curVP)
 174          brw->state.dirty.brw |= BRW_NEW_VERTEX_PROGRAM;
 175       if (newVP->program.IsPositionInvariant) {
 176          _mesa_insert_mvp_code(ctx, &newVP->program);
 177       }
 178       newVP->id = get_new_program_id(brw->intelScreen);
 179
 180       /* Also tell tnl about it:
 181        */
 182       _tnl_program_string(ctx, target, prog);
 183       break;
 184    }
 185    default:
 186       /*
 187        * driver->ProgramStringNotify is only called for ARB programs, fixed
 188        * function vertex programs, and ir_to_mesa (which isn't used by the
 189        * i965 back-end).  Therefore, even after geometry shaders are added,
 190        * this function should only ever be called with a target of
 191        * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
 192        */
 193       assert(!"Unexpected target in brwProgramStringNotify");
 194       break;
 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    printf("%-6s%-6s", stage, name);
 322
 323    if (shader_num != -1)
 324       printf("%4d: ", shader_num);
 325    else
 326       printf("    : ");
 327
 328    printf("%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       printf("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    printf("\n");
 409    printf("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
 416       if (scaled[i] == 0)
 417          continue;
 418
 419       int shader_num = -1;
 420       if (brw->shader_time.shader_programs[i]) {
 421          shader_num = brw->shader_time.shader_programs[i]->Name;
 422
 423          /* The fixed function fragment shader generates GLSL IR with a Name
 424           * of 0, and nothing else does.
 425           */
 426          if (shader_num == 0 &&
 427              (brw->shader_time.types[i] == ST_FS8 ||
 428               brw->shader_time.types[i] == ST_FS16)) {
 429             shader_name = "ff";
 430             shader_num = -1;
 431          } else {
 432             shader_name = "glsl";
 433          }
 434       } else if (brw->shader_time.programs[i]) {
 435          shader_num = brw->shader_time.programs[i]->Id;
 436          if (shader_num == 0) {
 437             shader_name = "ff";
 438             shader_num = -1;
 439          } else {
 440             shader_name = "prog";
 441          }
 442       } else {
 443          shader_name = "other";
 444       }
 445
 446       switch (brw->shader_time.types[i]) {
 447       case ST_VS:
 448          stage = "vs";
 449          break;
 450       case ST_GS:
 451          stage = "gs";
 452          break;
 453       case ST_FS8:
 454          stage = "fs8";
 455          break;
 456       case ST_FS16:
 457          stage = "fs16";
 458          break;
 459       default:
 460          stage = "other";
 461          break;
 462       }
 463
 464       print_shader_time_line(stage, shader_name, shader_num,
 465                              scaled[i], total);
 466    }
 467
 468    printf("\n");
 469    print_shader_time_line("total", "vs", -1, total_by_type[ST_VS], total);
 470    print_shader_time_line("total", "gs", -1, total_by_type[ST_GS], total);
 471    print_shader_time_line("total", "fs8", -1, total_by_type[ST_FS8], total);
 472    print_shader_time_line("total", "fs16", -1, total_by_type[ST_FS16], total);
 473 }
 474
 475 static void
 476 brw_collect_shader_time(struct brw_context *brw)
 477 {
 478    if (!brw->shader_time.bo)
 479       return;
 480
 481    /* This probably stalls on the last rendering.  We could fix that by
 482     * delaying reading the reports, but it doesn't look like it's a big
 483     * overhead compared to the cost of tracking the time in the first place.
 484     */
 485    drm_intel_bo_map(brw->shader_time.bo, true);
 486
 487    uint32_t *times = brw->shader_time.bo->virtual;
 488
 489    for (int i = 0; i < brw->shader_time.num_entries; i++) {
 490       brw->shader_time.cumulative[i] += times[i * SHADER_TIME_STRIDE / 4];
 491    }
 492
 493    /* Zero the BO out to clear it out for our next collection.
 494     */
 495    memset(times, 0, brw->shader_time.bo->size);
 496    drm_intel_bo_unmap(brw->shader_time.bo);
 497 }
 498
 499 void
 500 brw_collect_and_report_shader_time(struct brw_context *brw)
 501 {
 502    brw_collect_shader_time(brw);
 503
 504    if (brw->shader_time.report_time == 0 ||
 505        get_time() - brw->shader_time.report_time >= 1.0) {
 506       brw_report_shader_time(brw);
 507       brw->shader_time.report_time = get_time();
 508    }
 509 }
 510
 511 /**
 512  * Chooses an index in the shader_time buffer and sets up tracking information
 513  * for our printouts.
 514  *
 515  * Note that this holds on to references to the underlying programs, which may
 516  * change their lifetimes compared to normal operation.
 517  */
 518 int
 519 brw_get_shader_time_index(struct brw_context *brw,
 520                           struct gl_shader_program *shader_prog,
 521                           struct gl_program *prog,
 522                           enum shader_time_shader_type type)
 523 {
 524    struct gl_context *ctx = &brw->ctx;
 525
 526    int shader_time_index = brw->shader_time.num_entries++;
 527    assert(shader_time_index < brw->shader_time.max_entries);
 528    brw->shader_time.types[shader_time_index] = type;
 529
 530    _mesa_reference_shader_program(ctx,
 531                                   &brw->shader_time.shader_programs[shader_time_index],
 532                                   shader_prog);
 533
 534    _mesa_reference_program(ctx,
 535                            &brw->shader_time.programs[shader_time_index],
 536                            prog);
 537
 538    return shader_time_index;
 539 }
 540
 541 void
 542 brw_destroy_shader_time(struct brw_context *brw)
 543 {
 544    drm_intel_bo_unreference(brw->shader_time.bo);
 545    brw->shader_time.bo = NULL;
 546 }
 547
 548 void
 549 brw_mark_surface_used(struct brw_stage_prog_data *prog_data,
 550                       unsigned surf_index)
 551 {
 552    assert(surf_index < BRW_MAX_SURFACES);
 553
 554    prog_data->binding_table.size_bytes =
 555       MAX2(prog_data->binding_table.size_bytes, (surf_index + 1) * 4);
 556 }
 557
 558 bool
 559 brw_stage_prog_data_compare(const struct brw_stage_prog_data *a,
 560                             const struct brw_stage_prog_data *b)
 561 {
 562    /* Compare all the struct up to the pointers. */
 563    if (memcmp(a, b, offsetof(struct brw_stage_prog_data, param)))
 564       return false;
 565
 566    if (memcmp(a->param, b->param, a->nr_params * sizeof(void *)))
 567       return false;
 568
 569    if (memcmp(a->pull_param, b->pull_param, a->nr_pull_params * sizeof(void *)))
 570       return false;
 571
 572    return true;
 573 }
 574
 575 void
 576 brw_stage_prog_data_free(const void *p)
 577 {
 578    struct brw_stage_prog_data *prog_data = (struct brw_stage_prog_data *)p;
 579
 580    ralloc_free(prog_data->param);
 581    ralloc_free(prog_data->pull_param);
 582 }