src/gallium/drivers/r300/r300_emit.c

   1 /*
   2  * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
   3  * Copyright 2009 Marek Olšák <maraeo@gmail.com>
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * on the rights to use, copy, modify, merge, publish, distribute, sub
   9  * license, and/or sell copies of the Software, and to permit persons to whom
  10  * the Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  22  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
  23
  24 /* r300_emit: Functions for emitting state. */
  25
  26 #include "util/u_format.h"
  27 #include "util/u_math.h"
  28 #include "util/u_simple_list.h"
  29
  30 #include "r300_context.h"
  31 #include "r300_cs.h"
  32 #include "r300_emit.h"
  33 #include "r300_fs.h"
  34 #include "r300_screen.h"
  35 #include "r300_screen_buffer.h"
  36 #include "r300_vs.h"
  37
  38 void r300_emit_blend_state(struct r300_context* r300,
  39                            unsigned size, void* state)
  40 {
  41     struct r300_blend_state* blend = (struct r300_blend_state*)state;
  42     struct pipe_framebuffer_state* fb =
  43         (struct pipe_framebuffer_state*)r300->fb_state.state;
  44     CS_LOCALS(r300);
  45
  46     if (fb->nr_cbufs) {
  47         WRITE_CS_TABLE(blend->cb, size);
  48     } else {
  49         WRITE_CS_TABLE(blend->cb_no_readwrite, size);
  50     }
  51 }
  52
  53 void r300_emit_blend_color_state(struct r300_context* r300,
  54                                  unsigned size, void* state)
  55 {
  56     struct r300_blend_color_state* bc = (struct r300_blend_color_state*)state;
  57     CS_LOCALS(r300);
  58
  59     WRITE_CS_TABLE(bc->cb, size);
  60 }
  61
  62 void r300_emit_clip_state(struct r300_context* r300,
  63                           unsigned size, void* state)
  64 {
  65     struct r300_clip_state* clip = (struct r300_clip_state*)state;
  66     CS_LOCALS(r300);
  67
  68     WRITE_CS_TABLE(clip->cb, size);
  69 }
  70
  71 void r300_emit_dsa_state(struct r300_context* r300, unsigned size, void* state)
  72 {
  73     struct r300_dsa_state* dsa = (struct r300_dsa_state*)state;
  74     struct pipe_framebuffer_state* fb =
  75         (struct pipe_framebuffer_state*)r300->fb_state.state;
  76     CS_LOCALS(r300);
  77
  78     if (fb->zsbuf) {
  79         WRITE_CS_TABLE(&dsa->cb_begin, size);
  80     } else {
  81         WRITE_CS_TABLE(dsa->cb_no_readwrite, size);
  82     }
  83 }
  84
  85 static const float * get_rc_constant_state(
  86     struct r300_context * r300,
  87     struct rc_constant * constant)
  88 {
  89     struct r300_textures_state* texstate = r300->textures_state.state;
  90     static float vec[4] = { 0.0, 0.0, 0.0, 1.0 };
  91     struct pipe_resource *tex;
  92
  93     assert(constant->Type == RC_CONSTANT_STATE);
  94
  95     switch (constant->u.State[0]) {
  96         /* Factor for converting rectangle coords to
  97          * normalized coords. Should only show up on non-r500. */
  98         case RC_STATE_R300_TEXRECT_FACTOR:
  99             tex = texstate->sampler_views[constant->u.State[1]]->base.texture;
 100             vec[0] = 1.0 / tex->width0;
 101             vec[1] = 1.0 / tex->height0;
 102             break;
 103
 104         case RC_STATE_R300_VIEWPORT_SCALE:
 105             vec[0] = r300->viewport.scale[0];
 106             vec[1] = r300->viewport.scale[1];
 107             vec[2] = r300->viewport.scale[2];
 108             break;
 109
 110         case RC_STATE_R300_VIEWPORT_OFFSET:
 111             vec[0] = r300->viewport.translate[0];
 112             vec[1] = r300->viewport.translate[1];
 113             vec[2] = r300->viewport.translate[2];
 114             break;
 115
 116         default:
 117             fprintf(stderr, "r300: Implementation error: "
 118                 "Unknown RC_CONSTANT type %d\n", constant->u.State[0]);
 119     }
 120
 121     /* This should either be (0, 0, 0, 1), which should be a relatively safe
 122      * RGBA or STRQ value, or it could be one of the RC_CONSTANT_STATE
 123      * state factors. */
 124     return vec;
 125 }
 126
 127 /* Convert a normal single-precision float into the 7.16 format
 128  * used by the R300 fragment shader.
 129  */
 130 uint32_t pack_float24(float f)
 131 {
 132     union {
 133         float fl;
 134         uint32_t u;
 135     } u;
 136     float mantissa;
 137     int exponent;
 138     uint32_t float24 = 0;
 139
 140     if (f == 0.0)
 141         return 0;
 142
 143     u.fl = f;
 144
 145     mantissa = frexpf(f, &exponent);
 146
 147     /* Handle -ve */
 148     if (mantissa < 0) {
 149         float24 |= (1 << 23);
 150         mantissa = mantissa * -1.0;
 151     }
 152     /* Handle exponent, bias of 63 */
 153     exponent += 62;
 154     float24 |= (exponent << 16);
 155     /* Kill 7 LSB of mantissa */
 156     float24 |= (u.u & 0x7FFFFF) >> 7;
 157
 158     return float24;
 159 }
 160
 161 void r300_emit_fs(struct r300_context* r300, unsigned size, void *state)
 162 {
 163     struct r300_fragment_shader *fs = r300_fs(r300);
 164     CS_LOCALS(r300);
 165
 166     WRITE_CS_TABLE(fs->shader->cb_code, fs->shader->cb_code_size);
 167 }
 168
 169 void r300_emit_fs_constants(struct r300_context* r300, unsigned size, void *state)
 170 {
 171     struct r300_fragment_shader *fs = r300_fs(r300);
 172     struct r300_constant_buffer *buf = (struct r300_constant_buffer*)state;
 173     unsigned count = fs->shader->externals_count * 4;
 174     CS_LOCALS(r300);
 175
 176     if (count == 0)
 177         return;
 178
 179     BEGIN_CS(size);
 180     OUT_CS_REG_SEQ(R300_PFS_PARAM_0_X, count);
 181     OUT_CS_TABLE(buf->constants, count);
 182     END_CS;
 183 }
 184
 185 void r300_emit_fs_rc_constant_state(struct r300_context* r300, unsigned size, void *state)
 186 {
 187     struct r300_fragment_shader *fs = r300_fs(r300);
 188     struct rc_constant_list *constants = &fs->shader->code.constants;
 189     unsigned i;
 190     unsigned count = fs->shader->rc_state_count;
 191     unsigned first = fs->shader->externals_count;
 192     unsigned end = constants->Count;
 193     uint32_t cdata[4];
 194     unsigned j;
 195     CS_LOCALS(r300);
 196
 197     if (count == 0)
 198         return;
 199
 200     BEGIN_CS(size);
 201     for(i = first; i < end; ++i) {
 202         if (constants->Constants[i].Type == RC_CONSTANT_STATE) {
 203             const float *data =
 204                     get_rc_constant_state(r300, &constants->Constants[i]);
 205
 206             for (j = 0; j < 4; j++)
 207                 cdata[j] = pack_float24(data[j]);
 208
 209             OUT_CS_REG_SEQ(R300_PFS_PARAM_0_X + i * 16, 4);
 210             OUT_CS_TABLE(cdata, 4);
 211         }
 212     }
 213     END_CS;
 214 }
 215
 216 void r500_emit_fs(struct r300_context* r300, unsigned size, void *state)
 217 {
 218     struct r300_fragment_shader *fs = r300_fs(r300);
 219     CS_LOCALS(r300);
 220
 221     WRITE_CS_TABLE(fs->shader->cb_code, fs->shader->cb_code_size);
 222 }
 223
 224 void r500_emit_fs_constants(struct r300_context* r300, unsigned size, void *state)
 225 {
 226     struct r300_fragment_shader *fs = r300_fs(r300);
 227     struct r300_constant_buffer *buf = (struct r300_constant_buffer*)state;
 228     unsigned count = fs->shader->externals_count * 4;
 229     CS_LOCALS(r300);
 230
 231     if (count == 0)
 232         return;
 233
 234     BEGIN_CS(size);
 235     OUT_CS_REG(R500_GA_US_VECTOR_INDEX, R500_GA_US_VECTOR_INDEX_TYPE_CONST);
 236     OUT_CS_ONE_REG(R500_GA_US_VECTOR_DATA, count);
 237     OUT_CS_TABLE(buf->constants, count);
 238     END_CS;
 239 }
 240
 241 void r500_emit_fs_rc_constant_state(struct r300_context* r300, unsigned size, void *state)
 242 {
 243     struct r300_fragment_shader *fs = r300_fs(r300);
 244     struct rc_constant_list *constants = &fs->shader->code.constants;
 245     unsigned i;
 246     unsigned count = fs->shader->rc_state_count;
 247     unsigned first = fs->shader->externals_count;
 248     unsigned end = constants->Count;
 249     CS_LOCALS(r300);
 250
 251     if (count == 0)
 252         return;
 253
 254     BEGIN_CS(size);
 255     for(i = first; i < end; ++i) {
 256         if (constants->Constants[i].Type == RC_CONSTANT_STATE) {
 257             const float *data =
 258                     get_rc_constant_state(r300, &constants->Constants[i]);
 259
 260             OUT_CS_REG(R500_GA_US_VECTOR_INDEX,
 261                        R500_GA_US_VECTOR_INDEX_TYPE_CONST |
 262                        (i & R500_GA_US_VECTOR_INDEX_MASK));
 263             OUT_CS_ONE_REG(R500_GA_US_VECTOR_DATA, 4);
 264             OUT_CS_TABLE(data, 4);
 265         }
 266     }
 267     END_CS;
 268 }
 269
 270 void r300_emit_gpu_flush(struct r300_context *r300, unsigned size, void *state)
 271 {
 272     struct r300_gpu_flush *gpuflush = (struct r300_gpu_flush*)state;
 273     struct pipe_framebuffer_state* fb =
 274             (struct pipe_framebuffer_state*)r300->fb_state.state;
 275     CS_LOCALS(r300);
 276
 277     BEGIN_CS(size);
 278
 279     /* Set up scissors.
 280      * By writing to the SC registers, SC & US assert idle. */
 281     OUT_CS_REG_SEQ(R300_SC_SCISSORS_TL, 2);
 282     if (r300->screen->caps.is_r500) {
 283         OUT_CS(0);
 284         OUT_CS(((fb->width  - 1) << R300_SCISSORS_X_SHIFT) |
 285                ((fb->height - 1) << R300_SCISSORS_Y_SHIFT));
 286     } else {
 287         OUT_CS((1440 << R300_SCISSORS_X_SHIFT) |
 288                (1440 << R300_SCISSORS_Y_SHIFT));
 289         OUT_CS(((fb->width  + 1440-1) << R300_SCISSORS_X_SHIFT) |
 290                ((fb->height + 1440-1) << R300_SCISSORS_Y_SHIFT));
 291     }
 292
 293     /* Flush CB & ZB caches and wait until the 3D engine is idle and clean. */
 294     OUT_CS_TABLE(gpuflush->cb_flush_clean, 6);
 295     END_CS;
 296 }
 297
 298 void r300_emit_aa_state(struct r300_context *r300, unsigned size, void *state)
 299 {
 300     struct r300_aa_state *aa = (struct r300_aa_state*)state;
 301     CS_LOCALS(r300);
 302
 303     BEGIN_CS(size);
 304     OUT_CS_REG(R300_GB_AA_CONFIG, aa->aa_config);
 305
 306     if (aa->dest) {
 307         OUT_CS_REG_SEQ(R300_RB3D_AARESOLVE_OFFSET, 1);
 308         OUT_CS_RELOC(aa->dest->buffer, aa->dest->offset, 0, aa->dest->domain, 0);
 309
 310         OUT_CS_REG_SEQ(R300_RB3D_AARESOLVE_PITCH, 1);
 311         OUT_CS_RELOC(aa->dest->buffer, aa->dest->pitch, 0, aa->dest->domain, 0);
 312     }
 313
 314     OUT_CS_REG(R300_RB3D_AARESOLVE_CTL, aa->aaresolve_ctl);
 315     END_CS;
 316 }
 317
 318 void r300_emit_fb_state(struct r300_context* r300, unsigned size, void* state)
 319 {
 320     struct pipe_framebuffer_state* fb = (struct pipe_framebuffer_state*)state;
 321     struct r300_surface* surf;
 322     unsigned i;
 323     CS_LOCALS(r300);
 324
 325     BEGIN_CS(size);
 326
 327     /* NUM_MULTIWRITES replicates COLOR[0] to all colorbuffers, which is not
 328      * what we usually want. */
 329     if (r300->screen->caps.is_r500) {
 330         OUT_CS_REG(R300_RB3D_CCTL,
 331             R300_RB3D_CCTL_INDEPENDENT_COLORFORMAT_ENABLE_ENABLE);
 332     } else {
 333         OUT_CS_REG(R300_RB3D_CCTL, 0);
 334     }
 335
 336     /* Set up colorbuffers. */
 337     for (i = 0; i < fb->nr_cbufs; i++) {
 338         surf = r300_surface(fb->cbufs[i]);
 339
 340         OUT_CS_REG_SEQ(R300_RB3D_COLOROFFSET0 + (4 * i), 1);
 341         OUT_CS_RELOC(surf->buffer, surf->offset, 0, surf->domain, 0);
 342
 343         OUT_CS_REG_SEQ(R300_RB3D_COLORPITCH0 + (4 * i), 1);
 344         OUT_CS_RELOC(surf->buffer, surf->pitch, 0, surf->domain, 0);
 345     }
 346
 347     /* Set up a zbuffer. */
 348     if (fb->zsbuf) {
 349         surf = r300_surface(fb->zsbuf);
 350
 351         OUT_CS_REG(R300_ZB_FORMAT, surf->format);
 352
 353         OUT_CS_REG_SEQ(R300_ZB_DEPTHOFFSET, 1);
 354         OUT_CS_RELOC(surf->buffer, surf->offset, 0, surf->domain, 0);
 355
 356         OUT_CS_REG_SEQ(R300_ZB_DEPTHPITCH, 1);
 357         OUT_CS_RELOC(surf->buffer, surf->pitch, 0, surf->domain, 0);
 358
 359         /* HiZ RAM. */
 360         if (r300->screen->caps.has_hiz) {
 361             OUT_CS_REG(R300_ZB_HIZ_OFFSET, 0);
 362             OUT_CS_REG(R300_ZB_HIZ_PITCH, 0);
 363         }
 364
 365         /* Z Mask RAM. (compressed zbuffer) */
 366         OUT_CS_REG(R300_ZB_ZMASK_OFFSET, 0);
 367         OUT_CS_REG(R300_ZB_ZMASK_PITCH, 0);
 368     }
 369
 370     END_CS;
 371 }
 372
 373 void r300_emit_hyperz_state(struct r300_context *r300,
 374                             unsigned size, void *state)
 375 {
 376     CS_LOCALS(r300);
 377     WRITE_CS_TABLE(state, size);
 378 }
 379
 380 void r300_emit_fb_state_pipelined(struct r300_context *r300,
 381                                   unsigned size, void *state)
 382 {
 383     struct pipe_framebuffer_state* fb =
 384             (struct pipe_framebuffer_state*)r300->fb_state.state;
 385     unsigned i;
 386     CS_LOCALS(r300);
 387
 388     BEGIN_CS(size);
 389
 390     /* Colorbuffer format in the US block.
 391      * (must be written after unpipelined regs) */
 392     OUT_CS_REG_SEQ(R300_US_OUT_FMT_0, 4);
 393     for (i = 0; i < fb->nr_cbufs; i++) {
 394         OUT_CS(r300_surface(fb->cbufs[i])->format);
 395     }
 396     for (; i < 4; i++) {
 397         OUT_CS(R300_US_OUT_FMT_UNUSED);
 398     }
 399
 400     /* Multisampling. Depends on framebuffer sample count.
 401      * These are pipelined regs and as such cannot be moved
 402      * to the AA state. */
 403     if (r300->rws->get_value(r300->rws, R300_VID_DRM_2_3_0)) {
 404         unsigned mspos0 = 0x66666666;
 405         unsigned mspos1 = 0x6666666;
 406
 407         if (fb->nr_cbufs && fb->cbufs[0]->texture->nr_samples > 1) {
 408             /* Subsample placement. These may not be optimal. */
 409             switch (fb->cbufs[0]->texture->nr_samples) {
 410                 case 2:
 411                     mspos0 = 0x33996633;
 412                     mspos1 = 0x6666663;
 413                     break;
 414                 case 3:
 415                     mspos0 = 0x33936933;
 416                     mspos1 = 0x6666663;
 417                     break;
 418                 case 4:
 419                     mspos0 = 0x33939933;
 420                     mspos1 = 0x3966663;
 421                     break;
 422                 case 6:
 423                     mspos0 = 0x22a2aa22;
 424                     mspos1 = 0x2a65672;
 425                     break;
 426                 default:
 427                     debug_printf("r300: Bad number of multisamples!\n");
 428             }
 429         }
 430
 431         OUT_CS_REG_SEQ(R300_GB_MSPOS0, 2);
 432         OUT_CS(mspos0);
 433         OUT_CS(mspos1);
 434     }
 435     END_CS;
 436 }
 437
 438 void r300_emit_query_start(struct r300_context *r300, unsigned size, void*state)
 439 {
 440     struct r300_query *query = r300->query_current;
 441     CS_LOCALS(r300);
 442
 443     if (!query)
 444         return;
 445
 446     BEGIN_CS(size);
 447     if (r300->screen->caps.family == CHIP_FAMILY_RV530) {
 448         OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_ALL);
 449     } else {
 450         OUT_CS_REG(R300_SU_REG_DEST, R300_RASTER_PIPE_SELECT_ALL);
 451     }
 452     OUT_CS_REG(R300_ZB_ZPASS_DATA, 0);
 453     END_CS;
 454     query->begin_emitted = TRUE;
 455     query->flushed = FALSE;
 456 }
 457
 458 static void r300_emit_query_end_frag_pipes(struct r300_context *r300,
 459                                            struct r300_query *query)
 460 {
 461     struct r300_capabilities* caps = &r300->screen->caps;
 462     struct r300_winsys_buffer *buf = r300->query_current->buffer;
 463     CS_LOCALS(r300);
 464
 465     assert(caps->num_frag_pipes);
 466
 467     BEGIN_CS(6 * caps->num_frag_pipes + 2);
 468     /* I'm not so sure I like this switch, but it's hard to be elegant
 469      * when there's so many special cases...
 470      *
 471      * So here's the basic idea. For each pipe, enable writes to it only,
 472      * then put out the relocation for ZPASS_ADDR, taking into account a
 473      * 4-byte offset for each pipe. RV380 and older are special; they have
 474      * only two pipes, and the second pipe's enable is on bit 3, not bit 1,
 475      * so there's a chipset cap for that. */
 476     switch (caps->num_frag_pipes) {
 477         case 4:
 478             /* pipe 3 only */
 479             OUT_CS_REG(R300_SU_REG_DEST, 1 << 3);
 480             OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 481             OUT_CS_RELOC(buf, (query->num_results + 3) * 4,
 482                     0, query->domain, 0);
 483         case 3:
 484             /* pipe 2 only */
 485             OUT_CS_REG(R300_SU_REG_DEST, 1 << 2);
 486             OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 487             OUT_CS_RELOC(buf, (query->num_results + 2) * 4,
 488                     0, query->domain, 0);
 489         case 2:
 490             /* pipe 1 only */
 491             /* As mentioned above, accomodate RV380 and older. */
 492             OUT_CS_REG(R300_SU_REG_DEST,
 493                     1 << (caps->high_second_pipe ? 3 : 1));
 494             OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 495             OUT_CS_RELOC(buf, (query->num_results + 1) * 4,
 496                     0, query->domain, 0);
 497         case 1:
 498             /* pipe 0 only */
 499             OUT_CS_REG(R300_SU_REG_DEST, 1 << 0);
 500             OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 501             OUT_CS_RELOC(buf, (query->num_results + 0) * 4,
 502                     0, query->domain, 0);
 503             break;
 504         default:
 505             fprintf(stderr, "r300: Implementation error: Chipset reports %d"
 506                     " pixel pipes!\n", caps->num_frag_pipes);
 507             abort();
 508     }
 509
 510     /* And, finally, reset it to normal... */
 511     OUT_CS_REG(R300_SU_REG_DEST, 0xF);
 512     END_CS;
 513 }
 514
 515 static void rv530_emit_query_end_single_z(struct r300_context *r300,
 516                                           struct r300_query *query)
 517 {
 518     struct r300_winsys_buffer *buf = r300->query_current->buffer;
 519     CS_LOCALS(r300);
 520
 521     BEGIN_CS(8);
 522     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_0);
 523     OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 524     OUT_CS_RELOC(buf, query->num_results * 4, 0, query->domain, 0);
 525     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_ALL);
 526     END_CS;
 527 }
 528
 529 static void rv530_emit_query_end_double_z(struct r300_context *r300,
 530                                           struct r300_query *query)
 531 {
 532     struct r300_winsys_buffer *buf = r300->query_current->buffer;
 533     CS_LOCALS(r300);
 534
 535     BEGIN_CS(14);
 536     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_0);
 537     OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 538     OUT_CS_RELOC(buf, (query->num_results + 0) * 4, 0, query->domain, 0);
 539     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_1);
 540     OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 541     OUT_CS_RELOC(buf, (query->num_results + 1) * 4, 0, query->domain, 0);
 542     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_ALL);
 543     END_CS;
 544 }
 545
 546 void r300_emit_query_end(struct r300_context* r300)
 547 {
 548     struct r300_capabilities *caps = &r300->screen->caps;
 549     struct r300_query *query = r300->query_current;
 550
 551     if (!query)
 552         return;
 553
 554     if (query->begin_emitted == FALSE)
 555         return;
 556
 557     if (caps->family == CHIP_FAMILY_RV530) {
 558         if (caps->num_z_pipes == 2)
 559             rv530_emit_query_end_double_z(r300, query);
 560         else
 561             rv530_emit_query_end_single_z(r300, query);
 562     } else
 563         r300_emit_query_end_frag_pipes(r300, query);
 564
 565     query->begin_emitted = FALSE;
 566     query->num_results += query->num_pipes;
 567
 568     /* XXX grab all the results and reset the counter. */
 569     if (query->num_results >= query->buffer_size / 4 - 4) {
 570         query->num_results = (query->buffer_size / 4) / 2;
 571         fprintf(stderr, "r300: Rewinding OQBO...\n");
 572     }
 573 }
 574
 575 void r300_emit_invariant_state(struct r300_context *r300,
 576                                unsigned size, void *state)
 577 {
 578     CS_LOCALS(r300);
 579     WRITE_CS_TABLE(state, size);
 580 }
 581
 582 void r300_emit_rs_state(struct r300_context* r300, unsigned size, void* state)
 583 {
 584     struct r300_rs_state* rs = state;
 585     CS_LOCALS(r300);
 586
 587     BEGIN_CS(size);
 588     OUT_CS_TABLE(rs->cb_main, 25);
 589     if (rs->polygon_offset_enable) {
 590         if (r300->zbuffer_bpp == 16) {
 591             OUT_CS_TABLE(rs->cb_poly_offset_zb16, 5);
 592         } else {
 593             OUT_CS_TABLE(rs->cb_poly_offset_zb24, 5);
 594         }
 595     }
 596     END_CS;
 597 }
 598
 599 void r300_emit_rs_block_state(struct r300_context* r300,
 600                               unsigned size, void* state)
 601 {
 602     struct r300_rs_block* rs = (struct r300_rs_block*)state;
 603     unsigned i;
 604     /* It's the same for both INST and IP tables */
 605     unsigned count = (rs->inst_count & R300_RS_INST_COUNT_MASK) + 1;
 606     CS_LOCALS(r300);
 607
 608     if (SCREEN_DBG_ON(r300->screen, DBG_DRAW)) {
 609         r500_dump_rs_block(rs);
 610
 611         fprintf(stderr, "r300: RS emit:\n");
 612
 613         for (i = 0; i < count; i++)
 614             fprintf(stderr, "    : ip %d: 0x%08x\n", i, rs->ip[i]);
 615
 616         for (i = 0; i < count; i++)
 617             fprintf(stderr, "    : inst %d: 0x%08x\n", i, rs->inst[i]);
 618
 619         fprintf(stderr, "    : count: 0x%08x inst_count: 0x%08x\n",
 620             rs->count, rs->inst_count);
 621     }
 622
 623     BEGIN_CS(size);
 624     OUT_CS_REG_SEQ(R300_VAP_VTX_STATE_CNTL, 2);
 625     OUT_CS(rs->vap_vtx_state_cntl);
 626     OUT_CS(rs->vap_vsm_vtx_assm);
 627     OUT_CS_REG_SEQ(R300_VAP_OUTPUT_VTX_FMT_0, 2);
 628     OUT_CS(rs->vap_out_vtx_fmt[0]);
 629     OUT_CS(rs->vap_out_vtx_fmt[1]);
 630
 631     if (r300->screen->caps.is_r500) {
 632         OUT_CS_REG_SEQ(R500_RS_IP_0, count);
 633     } else {
 634         OUT_CS_REG_SEQ(R300_RS_IP_0, count);
 635     }
 636     OUT_CS_TABLE(rs->ip, count);
 637
 638     OUT_CS_REG_SEQ(R300_RS_COUNT, 2);
 639     OUT_CS(rs->count);
 640     OUT_CS(rs->inst_count);
 641
 642     if (r300->screen->caps.is_r500) {
 643         OUT_CS_REG_SEQ(R500_RS_INST_0, count);
 644     } else {
 645         OUT_CS_REG_SEQ(R300_RS_INST_0, count);
 646     }
 647     OUT_CS_TABLE(rs->inst, count);
 648     END_CS;
 649 }
 650
 651 void r300_emit_scissor_state(struct r300_context* r300,
 652                              unsigned size, void* state)
 653 {
 654     struct pipe_scissor_state* scissor = (struct pipe_scissor_state*)state;
 655     CS_LOCALS(r300);
 656
 657     BEGIN_CS(size);
 658     OUT_CS_REG_SEQ(R300_SC_CLIPRECT_TL_0, 2);
 659     if (r300->screen->caps.is_r500) {
 660         OUT_CS((scissor->minx << R300_CLIPRECT_X_SHIFT) |
 661                (scissor->miny << R300_CLIPRECT_Y_SHIFT));
 662         OUT_CS(((scissor->maxx - 1) << R300_CLIPRECT_X_SHIFT) |
 663                ((scissor->maxy - 1) << R300_CLIPRECT_Y_SHIFT));
 664     } else {
 665         OUT_CS(((scissor->minx + 1440) << R300_CLIPRECT_X_SHIFT) |
 666                ((scissor->miny + 1440) << R300_CLIPRECT_Y_SHIFT));
 667         OUT_CS(((scissor->maxx + 1440-1) << R300_CLIPRECT_X_SHIFT) |
 668                ((scissor->maxy + 1440-1) << R300_CLIPRECT_Y_SHIFT));
 669     }
 670     END_CS;
 671 }
 672
 673 void r300_emit_textures_state(struct r300_context *r300,
 674                               unsigned size, void *state)
 675 {
 676     struct r300_textures_state *allstate = (struct r300_textures_state*)state;
 677     struct r300_texture_sampler_state *texstate;
 678     struct r300_texture *tex;
 679     unsigned i;
 680     CS_LOCALS(r300);
 681
 682     BEGIN_CS(size);
 683     OUT_CS_REG(R300_TX_ENABLE, allstate->tx_enable);
 684
 685     for (i = 0; i < allstate->count; i++) {
 686         if ((1 << i) & allstate->tx_enable) {
 687             texstate = &allstate->regs[i];
 688             tex = r300_texture(allstate->sampler_views[i]->base.texture);
 689
 690             OUT_CS_REG(R300_TX_FILTER0_0 + (i * 4), texstate->filter0);
 691             OUT_CS_REG(R300_TX_FILTER1_0 + (i * 4), texstate->filter1);
 692             OUT_CS_REG(R300_TX_BORDER_COLOR_0 + (i * 4),
 693                        texstate->border_color);
 694
 695             OUT_CS_REG(R300_TX_FORMAT0_0 + (i * 4), texstate->format.format0);
 696             OUT_CS_REG(R300_TX_FORMAT1_0 + (i * 4), texstate->format.format1);
 697             OUT_CS_REG(R300_TX_FORMAT2_0 + (i * 4), texstate->format.format2);
 698
 699             OUT_CS_REG_SEQ(R300_TX_OFFSET_0 + (i * 4), 1);
 700             OUT_CS_TEX_RELOC(tex, texstate->format.tile_config, tex->domain,
 701                              0, 0);
 702         }
 703     }
 704     END_CS;
 705 }
 706
 707 void r300_emit_aos(struct r300_context* r300, int offset, boolean indexed)
 708 {
 709     struct pipe_vertex_buffer *vb1, *vb2, *vbuf = r300->vertex_buffer;
 710     struct pipe_vertex_element *velem = r300->velems->velem;
 711     struct r300_buffer *buf;
 712     int i;
 713     unsigned *hw_format_size = r300->velems->hw_format_size;
 714     unsigned size1, size2, aos_count = r300->velems->count;
 715     unsigned packet_size = (aos_count * 3 + 1) / 2;
 716     CS_LOCALS(r300);
 717
 718     BEGIN_CS(2 + packet_size + aos_count * 2);
 719     OUT_CS_PKT3(R300_PACKET3_3D_LOAD_VBPNTR, packet_size);
 720     OUT_CS(aos_count | (!indexed ? R300_VC_FORCE_PREFETCH : 0));
 721
 722     for (i = 0; i < aos_count - 1; i += 2) {
 723         vb1 = &vbuf[velem[i].vertex_buffer_index];
 724         vb2 = &vbuf[velem[i+1].vertex_buffer_index];
 725         size1 = hw_format_size[i];
 726         size2 = hw_format_size[i+1];
 727
 728         OUT_CS(R300_VBPNTR_SIZE0(size1) | R300_VBPNTR_STRIDE0(vb1->stride) |
 729                R300_VBPNTR_SIZE1(size2) | R300_VBPNTR_STRIDE1(vb2->stride));
 730         OUT_CS(vb1->buffer_offset + velem[i].src_offset   + offset * vb1->stride);
 731         OUT_CS(vb2->buffer_offset + velem[i+1].src_offset + offset * vb2->stride);
 732     }
 733
 734     if (aos_count & 1) {
 735         vb1 = &vbuf[velem[i].vertex_buffer_index];
 736         size1 = hw_format_size[i];
 737
 738         OUT_CS(R300_VBPNTR_SIZE0(size1) | R300_VBPNTR_STRIDE0(vb1->stride));
 739         OUT_CS(vb1->buffer_offset + velem[i].src_offset + offset * vb1->stride);
 740     }
 741
 742     for (i = 0; i < aos_count; i++) {
 743         buf = r300_buffer(vbuf[velem[i].vertex_buffer_index].buffer);
 744         OUT_CS_BUF_RELOC_NO_OFFSET(&buf->b.b, buf->domain, 0, 0);
 745     }
 746     END_CS;
 747 }
 748
 749 void r300_emit_aos_swtcl(struct r300_context *r300, boolean indexed)
 750 {
 751     CS_LOCALS(r300);
 752
 753     DBG(r300, DBG_DRAW, "r300: Preparing vertex buffer %p for render, "
 754             "vertex size %d\n", r300->vbo,
 755             r300->vertex_info.size);
 756     /* Set the pointer to our vertex buffer. The emitted values are this:
 757      * PACKET3 [3D_LOAD_VBPNTR]
 758      * COUNT   [1]
 759      * FORMAT  [size | stride << 8]
 760      * OFFSET  [offset into BO]
 761      * VBPNTR  [relocated BO]
 762      */
 763     BEGIN_CS(7);
 764     OUT_CS_PKT3(R300_PACKET3_3D_LOAD_VBPNTR, 3);
 765     OUT_CS(1 | (!indexed ? R300_VC_FORCE_PREFETCH : 0));
 766     OUT_CS(r300->vertex_info.size |
 767             (r300->vertex_info.size << 8));
 768     OUT_CS(r300->vbo_offset);
 769     OUT_CS_BUF_RELOC(r300->vbo, 0, r300_buffer(r300->vbo)->domain, 0, 0);
 770     END_CS;
 771 }
 772
 773 void r300_emit_vertex_stream_state(struct r300_context* r300,
 774                                    unsigned size, void* state)
 775 {
 776     struct r300_vertex_stream_state *streams =
 777         (struct r300_vertex_stream_state*)state;
 778     unsigned i;
 779     CS_LOCALS(r300);
 780
 781     if (DBG_ON(r300, DBG_DRAW)) {
 782         fprintf(stderr, "r300: PSC emit:\n");
 783
 784         for (i = 0; i < streams->count; i++) {
 785             fprintf(stderr, "    : prog_stream_cntl%d: 0x%08x\n", i,
 786                    streams->vap_prog_stream_cntl[i]);
 787         }
 788
 789         for (i = 0; i < streams->count; i++) {
 790             fprintf(stderr, "    : prog_stream_cntl_ext%d: 0x%08x\n", i,
 791                    streams->vap_prog_stream_cntl_ext[i]);
 792         }
 793     }
 794
 795     BEGIN_CS(size);
 796     OUT_CS_REG_SEQ(R300_VAP_PROG_STREAM_CNTL_0, streams->count);
 797     OUT_CS_TABLE(streams->vap_prog_stream_cntl, streams->count);
 798     OUT_CS_REG_SEQ(R300_VAP_PROG_STREAM_CNTL_EXT_0, streams->count);
 799     OUT_CS_TABLE(streams->vap_prog_stream_cntl_ext, streams->count);
 800     END_CS;
 801 }
 802
 803 void r300_emit_pvs_flush(struct r300_context* r300, unsigned size, void* state)
 804 {
 805     CS_LOCALS(r300);
 806
 807     BEGIN_CS(size);
 808     OUT_CS_REG(R300_VAP_PVS_STATE_FLUSH_REG, 0x0);
 809     END_CS;
 810 }
 811
 812 void r300_emit_vap_invariant_state(struct r300_context *r300,
 813                                    unsigned size, void *state)
 814 {
 815     CS_LOCALS(r300);
 816     WRITE_CS_TABLE(state, size);
 817 }
 818
 819 void r300_emit_vs_state(struct r300_context* r300, unsigned size, void* state)
 820 {
 821     struct r300_vertex_shader* vs = (struct r300_vertex_shader*)state;
 822     struct r300_vertex_program_code* code = &vs->code;
 823     struct r300_screen* r300screen = r300->screen;
 824     unsigned instruction_count = code->length / 4;
 825     unsigned i;
 826
 827     unsigned vtx_mem_size = r300screen->caps.is_r500 ? 128 : 72;
 828     unsigned input_count = MAX2(util_bitcount(code->InputsRead), 1);
 829     unsigned output_count = MAX2(util_bitcount(code->OutputsWritten), 1);
 830     unsigned temp_count = MAX2(code->num_temporaries, 1);
 831
 832     unsigned pvs_num_slots = MIN3(vtx_mem_size / input_count,
 833                                   vtx_mem_size / output_count, 10);
 834     unsigned pvs_num_controllers = MIN2(vtx_mem_size / temp_count, 6);
 835
 836     unsigned imm_first = vs->externals_count;
 837     unsigned imm_end = vs->code.constants.Count;
 838     unsigned imm_count = vs->immediates_count;
 839
 840     CS_LOCALS(r300);
 841
 842     BEGIN_CS(size);
 843
 844     /* R300_VAP_PVS_CODE_CNTL_0
 845      * R300_VAP_PVS_CONST_CNTL
 846      * R300_VAP_PVS_CODE_CNTL_1
 847      * See the r5xx docs for instructions on how to use these. */
 848     OUT_CS_REG_SEQ(R300_VAP_PVS_CODE_CNTL_0, 3);
 849     OUT_CS(R300_PVS_FIRST_INST(0) |
 850             R300_PVS_XYZW_VALID_INST(instruction_count - 1) |
 851             R300_PVS_LAST_INST(instruction_count - 1));
 852     OUT_CS(R300_PVS_MAX_CONST_ADDR(code->constants.Count - 1));
 853     OUT_CS(instruction_count - 1);
 854
 855     OUT_CS_REG(R300_VAP_PVS_VECTOR_INDX_REG, 0);
 856     OUT_CS_ONE_REG(R300_VAP_PVS_UPLOAD_DATA, code->length);
 857     OUT_CS_TABLE(code->body.d, code->length);
 858
 859     OUT_CS_REG(R300_VAP_CNTL, R300_PVS_NUM_SLOTS(pvs_num_slots) |
 860             R300_PVS_NUM_CNTLRS(pvs_num_controllers) |
 861             R300_PVS_NUM_FPUS(r300screen->caps.num_vert_fpus) |
 862             R300_PVS_VF_MAX_VTX_NUM(12) |
 863             (r300screen->caps.is_r500 ? R500_TCL_STATE_OPTIMIZATION : 0));
 864
 865     /* Emit immediates. */
 866     if (imm_count) {
 867         OUT_CS_REG(R300_VAP_PVS_VECTOR_INDX_REG,
 868                    (r300->screen->caps.is_r500 ?
 869                    R500_PVS_CONST_START : R300_PVS_CONST_START) +
 870                    imm_first);
 871         OUT_CS_ONE_REG(R300_VAP_PVS_UPLOAD_DATA, imm_count * 4);
 872         for (i = imm_first; i < imm_end; i++) {
 873             const float *data = vs->code.constants.Constants[i].u.Immediate;
 874             OUT_CS_TABLE(data, 4);
 875         }
 876     }
 877     END_CS;
 878 }
 879
 880 void r300_emit_vs_constants(struct r300_context* r300,
 881                             unsigned size, void *state)
 882 {
 883     unsigned count =
 884         ((struct r300_vertex_shader*)r300->vs_state.state)->externals_count;
 885     struct r300_constant_buffer *buf = (struct r300_constant_buffer*)state;
 886     CS_LOCALS(r300);
 887
 888     if (!count)
 889         return;
 890
 891     BEGIN_CS(size);
 892     OUT_CS_REG(R300_VAP_PVS_VECTOR_INDX_REG,
 893                (r300->screen->caps.is_r500 ?
 894                R500_PVS_CONST_START : R300_PVS_CONST_START));
 895     OUT_CS_ONE_REG(R300_VAP_PVS_UPLOAD_DATA, count * 4);
 896     OUT_CS_TABLE(buf->constants, count * 4);
 897     END_CS;
 898 }
 899
 900 void r300_emit_viewport_state(struct r300_context* r300,
 901                               unsigned size, void* state)
 902 {
 903     struct r300_viewport_state* viewport = (struct r300_viewport_state*)state;
 904     CS_LOCALS(r300);
 905
 906     BEGIN_CS(size);
 907     OUT_CS_REG_SEQ(R300_SE_VPORT_XSCALE, 6);
 908     OUT_CS_TABLE(&viewport->xscale, 6);
 909     OUT_CS_REG(R300_VAP_VTE_CNTL, viewport->vte_control);
 910     END_CS;
 911 }
 912
 913 void r300_emit_ztop_state(struct r300_context* r300,
 914                           unsigned size, void* state)
 915 {
 916     struct r300_ztop_state* ztop = (struct r300_ztop_state*)state;
 917     CS_LOCALS(r300);
 918
 919     BEGIN_CS(size);
 920     OUT_CS_REG(R300_ZB_ZTOP, ztop->z_buffer_top);
 921     END_CS;
 922 }
 923
 924 void r300_emit_texture_cache_inval(struct r300_context* r300, unsigned size, void* state)
 925 {
 926     CS_LOCALS(r300);
 927
 928     BEGIN_CS(size);
 929     OUT_CS_REG(R300_TX_INVALTAGS, 0);
 930     END_CS;
 931 }
 932
 933 void r300_emit_buffer_validate(struct r300_context *r300,
 934                                boolean do_validate_vertex_buffers,
 935                                struct pipe_resource *index_buffer)
 936 {
 937     struct pipe_framebuffer_state* fb =
 938         (struct pipe_framebuffer_state*)r300->fb_state.state;
 939     struct r300_textures_state *texstate =
 940         (struct r300_textures_state*)r300->textures_state.state;
 941     struct r300_texture* tex;
 942     struct pipe_vertex_buffer *vbuf = r300->vertex_buffer;
 943     struct pipe_vertex_element *velem = r300->velems->velem;
 944     struct pipe_resource *pbuf;
 945     unsigned i;
 946     boolean invalid = FALSE;
 947
 948     /* upload buffers first */
 949     if (r300->screen->caps.has_tcl && r300->any_user_vbs) {
 950         r300_upload_user_buffers(r300);
 951         r300->any_user_vbs = false;
 952     }
 953
 954     /* Clean out BOs. */
 955     r300->rws->reset_bos(r300->rws);
 956
 957 validate:
 958     /* Color buffers... */
 959     for (i = 0; i < fb->nr_cbufs; i++) {
 960         tex = r300_texture(fb->cbufs[i]->texture);
 961         assert(tex && tex->buffer && "cbuf is marked, but NULL!");
 962         if (!r300_add_texture(r300->rws, tex, 0, tex->domain)) {
 963             r300->context.flush(&r300->context, 0, NULL);
 964             goto validate;
 965         }
 966     }
 967     /* ...depth buffer... */
 968     if (fb->zsbuf) {
 969         tex = r300_texture(fb->zsbuf->texture);
 970         assert(tex && tex->buffer && "zsbuf is marked, but NULL!");
 971         if (!r300_add_texture(r300->rws, tex,
 972                               0, tex->domain)) {
 973             r300->context.flush(&r300->context, 0, NULL);
 974             goto validate;
 975         }
 976     }
 977     /* ...textures... */
 978     for (i = 0; i < texstate->count; i++) {
 979         if (!(texstate->tx_enable & (1 << i))) {
 980             continue;
 981         }
 982
 983         tex = r300_texture(texstate->sampler_views[i]->base.texture);
 984         if (!r300_add_texture(r300->rws, tex, tex->domain, 0)) {
 985             r300->context.flush(&r300->context, 0, NULL);
 986             goto validate;
 987         }
 988     }
 989     /* ...occlusion query buffer... */
 990     if (r300->query_current) {
 991         if (!r300->rws->add_buffer(r300->rws, r300->query_current->buffer,
 992                                    0, r300->query_current->domain)) {
 993             r300->context.flush(&r300->context, 0, NULL);
 994             goto validate;
 995         }
 996     }
 997     /* ...vertex buffer for SWTCL path... */
 998     if (r300->vbo) {
 999         if (!r300_add_buffer(r300->rws, r300->vbo,
1000                              r300_buffer(r300->vbo)->domain, 0)) {
1001             r300->context.flush(&r300->context, 0, NULL);
1002             goto validate;
1003         }
1004     }
1005     /* ...vertex buffers for HWTCL path... */
1006     if (do_validate_vertex_buffers) {
1007         for (i = 0; i < r300->velems->count; i++) {
1008             pbuf = vbuf[velem[i].vertex_buffer_index].buffer;
1009
1010             if (!r300_add_buffer(r300->rws, pbuf,
1011                                  r300_buffer(pbuf)->domain, 0)) {
1012                 r300->context.flush(&r300->context, 0, NULL);
1013                 goto validate;
1014             }
1015         }
1016     }
1017     /* ...and index buffer for HWTCL path. */
1018     if (index_buffer) {
1019         if (!r300_add_buffer(r300->rws, index_buffer,
1020                              r300_buffer(index_buffer)->domain, 0)) {
1021             r300->context.flush(&r300->context, 0, NULL);
1022             goto validate;
1023         }
1024     }
1025     if (!r300->rws->validate(r300->rws)) {
1026         r300->context.flush(&r300->context, 0, NULL);
1027         if (invalid) {
1028             /* Well, hell. */
1029             fprintf(stderr, "r300: Stuck in validation loop, gonna quit now.\n");
1030             abort();
1031         }
1032         invalid = TRUE;
1033         goto validate;
1034     }
1035 }
1036
1037 unsigned r300_get_num_dirty_dwords(struct r300_context *r300)
1038 {
1039     struct r300_atom* atom;
1040     unsigned dwords = 0;
1041
1042     foreach(atom, &r300->atom_list) {
1043         if (atom->dirty) {
1044             dwords += atom->size;
1045         }
1046     }
1047
1048     /* let's reserve some more, just in case */
1049     dwords += 32;
1050
1051     return dwords;
1052 }
1053
1054 /* Emit all dirty state. */
1055 void r300_emit_dirty_state(struct r300_context* r300)
1056 {
1057     struct r300_atom* atom;
1058
1059     foreach(atom, &r300->atom_list) {
1060         if (atom->dirty) {
1061             atom->emit(r300, atom->size, atom->state);
1062             if (SCREEN_DBG_ON(r300->screen, DBG_STATS)) {
1063                 atom->counter++;
1064             }
1065             atom->dirty = FALSE;
1066         }
1067     }
1068
1069     r300->dirty_hw++;
1070 }