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