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 struct gl_program *brwNewProgram( struct gl_context *ctx,
  58                                       GLenum target,
  59                                       GLuint id )
  60 {
  61    struct brw_context *brw = brw_context(ctx);
  62
  63    switch (target) {
  64    case GL_VERTEX_PROGRAM_ARB: {
  65       struct brw_vertex_program *prog = CALLOC_STRUCT(brw_vertex_program);
  66       if (prog) {
  67          prog->id = get_new_program_id(brw->intelScreen);
  68
  69          return _mesa_init_vertex_program( ctx, &prog->program,
  70                                              target, id );
  71       }
  72       else
  73          return NULL;
  74    }
  75
  76    case GL_FRAGMENT_PROGRAM_ARB: {
  77       struct brw_fragment_program *prog = CALLOC_STRUCT(brw_fragment_program);
  78       if (prog) {
  79          prog->id = get_new_program_id(brw->intelScreen);
  80
  81          return _mesa_init_fragment_program( ctx, &prog->program,
  82                                              target, id );
  83       }
  84       else
  85          return NULL;
  86    }
  87
  88    case MESA_GEOMETRY_PROGRAM: {
  89       struct brw_geometry_program *prog = CALLOC_STRUCT(brw_geometry_program);
  90       if (prog) {
  91          prog->id = get_new_program_id(brw->intelScreen);
  92
  93          return _mesa_init_geometry_program(ctx, &prog->program, target, id);
  94       } else {
  95          return NULL;
  96       }
  97    }
  98
  99    case GL_COMPUTE_PROGRAM_NV: {
 100       struct brw_compute_program *prog = CALLOC_STRUCT(brw_compute_program);
 101       if (prog) {
 102          prog->id = get_new_program_id(brw->intelScreen);
 103
 104          return _mesa_init_compute_program(ctx, &prog->program, target, id);
 105       } else {
 106          return NULL;
 107       }
 108    }
 109
 110    default:
 111       unreachable("Unsupported target in brwNewProgram()");
 112    }
 113 }
 114
 115 static void brwDeleteProgram( struct gl_context *ctx,
 116                               struct gl_program *prog )
 117 {
 118    _mesa_delete_program( ctx, prog );
 119 }
 120
 121
 122 static GLboolean
 123 brwProgramStringNotify(struct gl_context *ctx,
 124                        GLenum target,
 125                        struct gl_program *prog)
 126 {
 127    struct brw_context *brw = brw_context(ctx);
 128
 129    switch (target) {
 130    case GL_FRAGMENT_PROGRAM_ARB: {
 131       struct gl_fragment_program *fprog = (struct gl_fragment_program *) prog;
 132       struct brw_fragment_program *newFP = brw_fragment_program(fprog);
 133       const struct brw_fragment_program *curFP =
 134          brw_fragment_program_const(brw->fragment_program);
 135
 136       if (newFP == curFP)
 137          brw->state.dirty.brw |= BRW_NEW_FRAGMENT_PROGRAM;
 138       newFP->id = get_new_program_id(brw->intelScreen);
 139       break;
 140    }
 141    case GL_VERTEX_PROGRAM_ARB: {
 142       struct gl_vertex_program *vprog = (struct gl_vertex_program *) prog;
 143       struct brw_vertex_program *newVP = brw_vertex_program(vprog);
 144       const struct brw_vertex_program *curVP =
 145          brw_vertex_program_const(brw->vertex_program);
 146
 147       if (newVP == curVP)
 148          brw->state.dirty.brw |= BRW_NEW_VERTEX_PROGRAM;
 149       if (newVP->program.IsPositionInvariant) {
 150          _mesa_insert_mvp_code(ctx, &newVP->program);
 151       }
 152       newVP->id = get_new_program_id(brw->intelScreen);
 153
 154       /* Also tell tnl about it:
 155        */
 156       _tnl_program_string(ctx, target, prog);
 157       break;
 158    }
 159    default:
 160       /*
 161        * driver->ProgramStringNotify is only called for ARB programs, fixed
 162        * function vertex programs, and ir_to_mesa (which isn't used by the
 163        * i965 back-end).  Therefore, even after geometry shaders are added,
 164        * this function should only ever be called with a target of
 165        * GL_VERTEX_PROGRAM_ARB or GL_FRAGMENT_PROGRAM_ARB.
 166        */
 167       unreachable("Unexpected target in brwProgramStringNotify");
 168    }
 169
 170    brw_add_texrect_params(prog);
 171
 172    return true;
 173 }
 174
 175 void
 176 brw_add_texrect_params(struct gl_program *prog)
 177 {
 178    for (int texunit = 0; texunit < BRW_MAX_TEX_UNIT; texunit++) {
 179       if (!(prog->TexturesUsed[texunit] & (1 << TEXTURE_RECT_INDEX)))
 180          continue;
 181
 182       int tokens[STATE_LENGTH] = {
 183          STATE_INTERNAL,
 184          STATE_TEXRECT_SCALE,
 185          texunit,
 186          0,
 187          0
 188       };
 189
 190       _mesa_add_state_reference(prog->Parameters, (gl_state_index *)tokens);
 191    }
 192 }
 193
 194 /* Per-thread scratch space is a power-of-two multiple of 1KB. */
 195 int
 196 brw_get_scratch_size(int size)
 197 {
 198    int i;
 199
 200    for (i = 1024; i < size; i *= 2)
 201       ;
 202
 203    return i;
 204 }
 205
 206 void
 207 brw_get_scratch_bo(struct brw_context *brw,
 208                    drm_intel_bo **scratch_bo, int size)
 209 {
 210    drm_intel_bo *old_bo = *scratch_bo;
 211
 212    if (old_bo && old_bo->size < size) {
 213       drm_intel_bo_unreference(old_bo);
 214       old_bo = NULL;
 215    }
 216
 217    if (!old_bo) {
 218       *scratch_bo = drm_intel_bo_alloc(brw->bufmgr, "scratch bo", size, 4096);
 219    }
 220 }
 221
 222 void brwInitFragProgFuncs( struct dd_function_table *functions )
 223 {
 224    assert(functions->ProgramStringNotify == _tnl_program_string);
 225
 226    functions->NewProgram = brwNewProgram;
 227    functions->DeleteProgram = brwDeleteProgram;
 228    functions->ProgramStringNotify = brwProgramStringNotify;
 229
 230    functions->NewShader = brw_new_shader;
 231    functions->LinkShader = brw_link_shader;
 232 }
 233
 234 void
 235 brw_init_shader_time(struct brw_context *brw)
 236 {
 237    const int max_entries = 4096;
 238    brw->shader_time.bo = drm_intel_bo_alloc(brw->bufmgr, "shader time",
 239                                             max_entries * SHADER_TIME_STRIDE,
 240                                             4096);
 241    brw->shader_time.shader_programs = rzalloc_array(brw, struct gl_shader_program *,
 242                                                     max_entries);
 243    brw->shader_time.programs = rzalloc_array(brw, struct gl_program *,
 244                                              max_entries);
 245    brw->shader_time.types = rzalloc_array(brw, enum shader_time_shader_type,
 246                                           max_entries);
 247    brw->shader_time.cumulative = rzalloc_array(brw, uint64_t,
 248                                                max_entries);
 249    brw->shader_time.max_entries = max_entries;
 250 }
 251
 252 static int
 253 compare_time(const void *a, const void *b)
 254 {
 255    uint64_t * const *a_val = a;
 256    uint64_t * const *b_val = b;
 257
 258    /* We don't just subtract because we're turning the value to an int. */
 259    if (**a_val < **b_val)
 260       return -1;
 261    else if (**a_val == **b_val)
 262       return 0;
 263    else
 264       return 1;
 265 }
 266
 267 static void
 268 get_written_and_reset(struct brw_context *brw, int i,
 269                       uint64_t *written, uint64_t *reset)
 270 {
 271    enum shader_time_shader_type type = brw->shader_time.types[i];
 272    assert(type == ST_VS || type == ST_GS || type == ST_FS8 || type == ST_FS16);
 273
 274    /* Find where we recorded written and reset. */
 275    int wi, ri;
 276
 277    for (wi = i; brw->shader_time.types[wi] != type + 1; wi++)
 278       ;
 279
 280    for (ri = i; brw->shader_time.types[ri] != type + 2; ri++)
 281       ;
 282
 283    *written = brw->shader_time.cumulative[wi];
 284    *reset = brw->shader_time.cumulative[ri];
 285 }
 286
 287 static void
 288 print_shader_time_line(const char *stage, const char *name,
 289                        int shader_num, uint64_t time, uint64_t total)
 290 {
 291    fprintf(stderr, "%-6s%-18s", stage, name);
 292
 293    if (shader_num != -1)
 294       fprintf(stderr, "%4d: ", shader_num);
 295    else
 296       fprintf(stderr, "    : ");
 297
 298    fprintf(stderr, "%16lld (%7.2f Gcycles)      %4.1f%%\n",
 299            (long long)time,
 300            (double)time / 1000000000.0,
 301            (double)time / total * 100.0);
 302 }
 303
 304 static void
 305 brw_report_shader_time(struct brw_context *brw)
 306 {
 307    if (!brw->shader_time.bo || !brw->shader_time.num_entries)
 308       return;
 309
 310    uint64_t scaled[brw->shader_time.num_entries];
 311    uint64_t *sorted[brw->shader_time.num_entries];
 312    uint64_t total_by_type[ST_FS16 + 1];
 313    memset(total_by_type, 0, sizeof(total_by_type));
 314    double total = 0;
 315    for (int i = 0; i < brw->shader_time.num_entries; i++) {
 316       uint64_t written = 0, reset = 0;
 317       enum shader_time_shader_type type = brw->shader_time.types[i];
 318
 319       sorted[i] = &scaled[i];
 320
 321       switch (type) {
 322       case ST_VS_WRITTEN:
 323       case ST_VS_RESET:
 324       case ST_GS_WRITTEN:
 325       case ST_GS_RESET:
 326       case ST_FS8_WRITTEN:
 327       case ST_FS8_RESET:
 328       case ST_FS16_WRITTEN:
 329       case ST_FS16_RESET:
 330          /* We'll handle these when along with the time. */
 331          scaled[i] = 0;
 332          continue;
 333
 334       case ST_VS:
 335       case ST_GS:
 336       case ST_FS8:
 337       case ST_FS16:
 338          get_written_and_reset(brw, i, &written, &reset);
 339          break;
 340
 341       default:
 342          /* I sometimes want to print things that aren't the 3 shader times.
 343           * Just print the sum in that case.
 344           */
 345          written = 1;
 346          reset = 0;
 347          break;
 348       }
 349
 350       uint64_t time = brw->shader_time.cumulative[i];
 351       if (written) {
 352          scaled[i] = time / written * (written + reset);
 353       } else {
 354          scaled[i] = time;
 355       }
 356
 357       switch (type) {
 358       case ST_VS:
 359       case ST_GS:
 360       case ST_FS8:
 361       case ST_FS16:
 362          total_by_type[type] += scaled[i];
 363          break;
 364       default:
 365          break;
 366       }
 367
 368       total += scaled[i];
 369    }
 370
 371    if (total == 0) {
 372       fprintf(stderr, "No shader time collected yet\n");
 373       return;
 374    }
 375
 376    qsort(sorted, brw->shader_time.num_entries, sizeof(sorted[0]), compare_time);
 377
 378    fprintf(stderr, "\n");
 379    fprintf(stderr, "type          ID                  cycles spent                   %% of total\n");
 380    for (int s = 0; s < brw->shader_time.num_entries; s++) {
 381       const char *shader_name;
 382       const char *stage;
 383       /* Work back from the sorted pointers times to a time to print. */
 384       int i = sorted[s] - scaled;
 385       struct gl_shader_program *prog = brw->shader_time.shader_programs[i];
 386
 387       if (scaled[i] == 0)
 388          continue;
 389
 390       int shader_num = -1;
 391       if (prog) {
 392          shader_num = prog->Name;
 393
 394          /* The fixed function fragment shader generates GLSL IR with a Name
 395           * of 0, and nothing else does.
 396           */
 397          if (prog->Label) {
 398             shader_name = prog->Label;
 399          } else if (shader_num == 0 &&
 400              (brw->shader_time.types[i] == ST_FS8 ||
 401               brw->shader_time.types[i] == ST_FS16)) {
 402             shader_name = "ff";
 403             shader_num = -1;
 404          } else {
 405             shader_name = "glsl";
 406          }
 407       } else if (brw->shader_time.programs[i]) {
 408          shader_num = brw->shader_time.programs[i]->Id;
 409          if (shader_num == 0) {
 410             shader_name = "ff";
 411             shader_num = -1;
 412          } else {
 413             shader_name = "prog";
 414          }
 415       } else {
 416          shader_name = "other";
 417       }
 418
 419       switch (brw->shader_time.types[i]) {
 420       case ST_VS:
 421          stage = "vs";
 422          break;
 423       case ST_GS:
 424          stage = "gs";
 425          break;
 426       case ST_FS8:
 427          stage = "fs8";
 428          break;
 429       case ST_FS16:
 430          stage = "fs16";
 431          break;
 432       default:
 433          stage = "other";
 434          break;
 435       }
 436
 437       print_shader_time_line(stage, shader_name, shader_num,
 438                              scaled[i], total);
 439    }
 440
 441    fprintf(stderr, "\n");
 442    print_shader_time_line("total", "vs", -1, total_by_type[ST_VS], total);
 443    print_shader_time_line("total", "gs", -1, total_by_type[ST_GS], total);
 444    print_shader_time_line("total", "fs8", -1, total_by_type[ST_FS8], total);
 445    print_shader_time_line("total", "fs16", -1, total_by_type[ST_FS16], total);
 446 }
 447
 448 static void
 449 brw_collect_shader_time(struct brw_context *brw)
 450 {
 451    if (!brw->shader_time.bo)
 452       return;
 453
 454    /* This probably stalls on the last rendering.  We could fix that by
 455     * delaying reading the reports, but it doesn't look like it's a big
 456     * overhead compared to the cost of tracking the time in the first place.
 457     */
 458    drm_intel_bo_map(brw->shader_time.bo, true);
 459
 460    uint32_t *times = brw->shader_time.bo->virtual;
 461
 462    for (int i = 0; i < brw->shader_time.num_entries; i++) {
 463       brw->shader_time.cumulative[i] += times[i * SHADER_TIME_STRIDE / 4];
 464    }
 465
 466    /* Zero the BO out to clear it out for our next collection.
 467     */
 468    memset(times, 0, brw->shader_time.bo->size);
 469    drm_intel_bo_unmap(brw->shader_time.bo);
 470 }
 471
 472 void
 473 brw_collect_and_report_shader_time(struct brw_context *brw)
 474 {
 475    brw_collect_shader_time(brw);
 476
 477    if (brw->shader_time.report_time == 0 ||
 478        get_time() - brw->shader_time.report_time >= 1.0) {
 479       brw_report_shader_time(brw);
 480       brw->shader_time.report_time = get_time();
 481    }
 482 }
 483
 484 /**
 485  * Chooses an index in the shader_time buffer and sets up tracking information
 486  * for our printouts.
 487  *
 488  * Note that this holds on to references to the underlying programs, which may
 489  * change their lifetimes compared to normal operation.
 490  */
 491 int
 492 brw_get_shader_time_index(struct brw_context *brw,
 493                           struct gl_shader_program *shader_prog,
 494                           struct gl_program *prog,
 495                           enum shader_time_shader_type type)
 496 {
 497    struct gl_context *ctx = &brw->ctx;
 498
 499    int shader_time_index = brw->shader_time.num_entries++;
 500    assert(shader_time_index < brw->shader_time.max_entries);
 501    brw->shader_time.types[shader_time_index] = type;
 502
 503    _mesa_reference_shader_program(ctx,
 504                                   &brw->shader_time.shader_programs[shader_time_index],
 505                                   shader_prog);
 506
 507    _mesa_reference_program(ctx,
 508                            &brw->shader_time.programs[shader_time_index],
 509                            prog);
 510
 511    return shader_time_index;
 512 }
 513
 514 void
 515 brw_destroy_shader_time(struct brw_context *brw)
 516 {
 517    drm_intel_bo_unreference(brw->shader_time.bo);
 518    brw->shader_time.bo = NULL;
 519 }
 520
 521 void
 522 brw_mark_surface_used(struct brw_stage_prog_data *prog_data,
 523                       unsigned surf_index)
 524 {
 525    assert(surf_index < BRW_MAX_SURFACES);
 526
 527    prog_data->binding_table.size_bytes =
 528       MAX2(prog_data->binding_table.size_bytes, (surf_index + 1) * 4);
 529 }
 530
 531 bool
 532 brw_stage_prog_data_compare(const struct brw_stage_prog_data *a,
 533                             const struct brw_stage_prog_data *b)
 534 {
 535    /* Compare all the struct up to the pointers. */
 536    if (memcmp(a, b, offsetof(struct brw_stage_prog_data, param)))
 537       return false;
 538
 539    if (memcmp(a->param, b->param, a->nr_params * sizeof(void *)))
 540       return false;
 541
 542    if (memcmp(a->pull_param, b->pull_param, a->nr_pull_params * sizeof(void *)))
 543       return false;
 544
 545    return true;
 546 }
 547
 548 void
 549 brw_stage_prog_data_free(const void *p)
 550 {
 551    struct brw_stage_prog_data *prog_data = (struct brw_stage_prog_data *)p;
 552
 553    ralloc_free(prog_data->param);
 554    ralloc_free(prog_data->pull_param);
 555 }
 556
 557 void
 558 brw_dump_ir(struct brw_context *brw, const char *stage,
 559             struct gl_shader_program *shader_prog,
 560             struct gl_shader *shader, struct gl_program *prog)
 561 {
 562    if (shader_prog) {
 563       fprintf(stderr,
 564               "GLSL IR for native %s shader %d:\n", stage, shader_prog->Name);
 565       _mesa_print_ir(stderr, shader->ir, NULL);
 566       fprintf(stderr, "\n\n");
 567    } else {
 568       fprintf(stderr, "ARB_%s_program %d ir for native %s shader\n",
 569               stage, prog->Id, stage);
 570       _mesa_print_program(prog);
 571    }
 572 }