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 rc_constant_list *constants = &fs->shader->code.constants;
 173     struct r300_constant_buffer *buf = (struct r300_constant_buffer*)state;
 174     unsigned i, count = fs->shader->externals_count;
 175     CS_LOCALS(r300);
 176
 177     if (count == 0)
 178         return;
 179
 180     BEGIN_CS(size);
 181     OUT_CS_REG_SEQ(R300_PFS_PARAM_0_X, count * 4);
 182     for(i = 0; i < count; ++i) {
 183         const float *data;
 184         assert(constants->Constants[i].Type == RC_CONSTANT_EXTERNAL);
 185         data = buf->constants[i];
 186         OUT_CS(pack_float24(data[0]));
 187         OUT_CS(pack_float24(data[1]));
 188         OUT_CS(pack_float24(data[2]));
 189         OUT_CS(pack_float24(data[3]));
 190     }
 191     END_CS;
 192 }
 193
 194 void r300_emit_fs_rc_constant_state(struct r300_context* r300, unsigned size, void *state)
 195 {
 196     struct r300_fragment_shader *fs = r300_fs(r300);
 197     struct rc_constant_list *constants = &fs->shader->code.constants;
 198     unsigned i;
 199     unsigned count = fs->shader->rc_state_count;
 200     unsigned first = fs->shader->externals_count;
 201     unsigned end = constants->Count;
 202     CS_LOCALS(r300);
 203
 204     if (count == 0)
 205         return;
 206
 207     BEGIN_CS(size);
 208     for(i = first; i < end; ++i) {
 209         if (constants->Constants[i].Type == RC_CONSTANT_STATE) {
 210             const float *data =
 211                     get_rc_constant_state(r300, &constants->Constants[i]);
 212
 213             OUT_CS_REG_SEQ(R300_PFS_PARAM_0_X + i * 16, 4);
 214             OUT_CS(pack_float24(data[0]));
 215             OUT_CS(pack_float24(data[1]));
 216             OUT_CS(pack_float24(data[2]));
 217             OUT_CS(pack_float24(data[3]));
 218         }
 219     }
 220     END_CS;
 221 }
 222
 223 void r500_emit_fs(struct r300_context* r300, unsigned size, void *state)
 224 {
 225     struct r300_fragment_shader *fs = r300_fs(r300);
 226     CS_LOCALS(r300);
 227
 228     WRITE_CS_TABLE(fs->shader->cb_code, fs->shader->cb_code_size);
 229 }
 230
 231 void r500_emit_fs_constants(struct r300_context* r300, unsigned size, void *state)
 232 {
 233     struct r300_fragment_shader *fs = r300_fs(r300);
 234     struct rc_constant_list *constants = &fs->shader->code.constants;
 235     struct r300_constant_buffer *buf = (struct r300_constant_buffer*)state;
 236     unsigned i, count = fs->shader->externals_count;
 237     CS_LOCALS(r300);
 238
 239     if (count == 0)
 240         return;
 241
 242     BEGIN_CS(size);
 243     OUT_CS_REG(R500_GA_US_VECTOR_INDEX, R500_GA_US_VECTOR_INDEX_TYPE_CONST);
 244     OUT_CS_ONE_REG(R500_GA_US_VECTOR_DATA, count * 4);
 245     for(i = 0; i < count; ++i) {
 246         assert(constants->Constants[i].Type == RC_CONSTANT_EXTERNAL);
 247     }
 248     OUT_CS_TABLE(buf->constants, count * 4);
 249     END_CS;
 250 }
 251
 252 void r500_emit_fs_rc_constant_state(struct r300_context* r300, unsigned size, void *state)
 253 {
 254     struct r300_fragment_shader *fs = r300_fs(r300);
 255     struct rc_constant_list *constants = &fs->shader->code.constants;
 256     unsigned i;
 257     unsigned count = fs->shader->rc_state_count;
 258     unsigned first = fs->shader->externals_count;
 259     unsigned end = constants->Count;
 260     CS_LOCALS(r300);
 261
 262     if (count == 0)
 263         return;
 264
 265     BEGIN_CS(size);
 266     for(i = first; i < end; ++i) {
 267         if (constants->Constants[i].Type == RC_CONSTANT_STATE) {
 268             const float *data =
 269                     get_rc_constant_state(r300, &constants->Constants[i]);
 270
 271             OUT_CS_REG(R500_GA_US_VECTOR_INDEX,
 272                        R500_GA_US_VECTOR_INDEX_TYPE_CONST |
 273                        (i & R500_GA_US_VECTOR_INDEX_MASK));
 274             OUT_CS_ONE_REG(R500_GA_US_VECTOR_DATA, 4);
 275             OUT_CS_TABLE(data, 4);
 276         }
 277     }
 278     END_CS;
 279 }
 280
 281 void r300_emit_fb_state(struct r300_context* r300, unsigned size, void* state)
 282 {
 283     struct pipe_framebuffer_state* fb = (struct pipe_framebuffer_state*)state;
 284     struct r300_texture* tex;
 285     struct pipe_surface* surf;
 286     int i;
 287     CS_LOCALS(r300);
 288
 289     BEGIN_CS(size);
 290
 291     /* Flush and free renderbuffer caches. */
 292     OUT_CS_REG(R300_RB3D_DSTCACHE_CTLSTAT,
 293         R300_RB3D_DSTCACHE_CTLSTAT_DC_FREE_FREE_3D_TAGS |
 294         R300_RB3D_DSTCACHE_CTLSTAT_DC_FLUSH_FLUSH_DIRTY_3D);
 295     OUT_CS_REG(R300_ZB_ZCACHE_CTLSTAT,
 296         R300_ZB_ZCACHE_CTLSTAT_ZC_FLUSH_FLUSH_AND_FREE |
 297         R300_ZB_ZCACHE_CTLSTAT_ZC_FREE_FREE);
 298
 299     /* NUM_MULTIWRITES replicates COLOR[0] to all colorbuffers, which is not
 300      * what we usually want. */
 301     if (r300->screen->caps.is_r500) {
 302         OUT_CS_REG(R300_RB3D_CCTL,
 303             R300_RB3D_CCTL_INDEPENDENT_COLORFORMAT_ENABLE_ENABLE);
 304     } else {
 305         OUT_CS_REG(R300_RB3D_CCTL, 0);
 306     }
 307
 308     /* Set up colorbuffers. */
 309     for (i = 0; i < fb->nr_cbufs; i++) {
 310         surf = fb->cbufs[i];
 311         tex = r300_texture(surf->texture);
 312         assert(tex && tex->buffer && "cbuf is marked, but NULL!");
 313
 314         OUT_CS_REG_SEQ(R300_RB3D_COLOROFFSET0 + (4 * i), 1);
 315         OUT_CS_TEX_RELOC(tex, surf->offset, 0, tex->domain, 0);
 316
 317         OUT_CS_REG_SEQ(R300_RB3D_COLORPITCH0 + (4 * i), 1);
 318         OUT_CS_TEX_RELOC(tex, tex->fb_state.colorpitch[surf->level],
 319                      0, tex->domain, 0);
 320
 321         OUT_CS_REG(R300_US_OUT_FMT_0 + (4 * i), tex->fb_state.us_out_fmt);
 322     }
 323     for (; i < 4; i++) {
 324         OUT_CS_REG(R300_US_OUT_FMT_0 + (4 * i), R300_US_OUT_FMT_UNUSED);
 325     }
 326
 327     /* Set up a zbuffer. */
 328     if (fb->zsbuf) {
 329         surf = fb->zsbuf;
 330         tex = r300_texture(surf->texture);
 331         assert(tex && tex->buffer && "zsbuf is marked, but NULL!");
 332
 333         OUT_CS_REG_SEQ(R300_ZB_DEPTHOFFSET, 1);
 334         OUT_CS_TEX_RELOC(tex, surf->offset, 0, tex->domain, 0);
 335
 336         OUT_CS_REG(R300_ZB_FORMAT, tex->fb_state.zb_format);
 337
 338         OUT_CS_REG_SEQ(R300_ZB_DEPTHPITCH, 1);
 339         OUT_CS_TEX_RELOC(tex, tex->fb_state.depthpitch[surf->level],
 340                      0, tex->domain, 0);
 341     }
 342
 343     OUT_CS_REG_SEQ(R300_SC_SCISSORS_TL, 2);
 344     if (r300->screen->caps.is_r500) {
 345         OUT_CS(0);
 346         OUT_CS(((fb->width  - 1) << R300_SCISSORS_X_SHIFT) |
 347                ((fb->height - 1) << R300_SCISSORS_Y_SHIFT));
 348     } else {
 349         OUT_CS((1440 << R300_SCISSORS_X_SHIFT) |
 350                (1440 << R300_SCISSORS_Y_SHIFT));
 351         OUT_CS(((fb->width  + 1440-1) << R300_SCISSORS_X_SHIFT) |
 352                ((fb->height + 1440-1) << R300_SCISSORS_Y_SHIFT));
 353     }
 354     END_CS;
 355 }
 356
 357 void r300_emit_query_start(struct r300_context *r300, unsigned size, void*state)
 358 {
 359     struct r300_query *query = r300->query_current;
 360     CS_LOCALS(r300);
 361
 362     if (!query)
 363         return;
 364
 365     BEGIN_CS(size);
 366     if (r300->screen->caps.family == CHIP_FAMILY_RV530) {
 367         OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_ALL);
 368     } else {
 369         OUT_CS_REG(R300_SU_REG_DEST, R300_RASTER_PIPE_SELECT_ALL);
 370     }
 371     OUT_CS_REG(R300_ZB_ZPASS_DATA, 0);
 372     END_CS;
 373     query->begin_emitted = TRUE;
 374 }
 375
 376
 377 static void r300_emit_query_end_frag_pipes(struct r300_context *r300,
 378                                            struct r300_query *query)
 379 {
 380     struct r300_capabilities* caps = &r300->screen->caps;
 381     CS_LOCALS(r300);
 382
 383     assert(caps->num_frag_pipes);
 384
 385     BEGIN_CS(6 * caps->num_frag_pipes + 2);
 386     /* I'm not so sure I like this switch, but it's hard to be elegant
 387      * when there's so many special cases...
 388      *
 389      * So here's the basic idea. For each pipe, enable writes to it only,
 390      * then put out the relocation for ZPASS_ADDR, taking into account a
 391      * 4-byte offset for each pipe. RV380 and older are special; they have
 392      * only two pipes, and the second pipe's enable is on bit 3, not bit 1,
 393      * so there's a chipset cap for that. */
 394     switch (caps->num_frag_pipes) {
 395         case 4:
 396             /* pipe 3 only */
 397             OUT_CS_REG(R300_SU_REG_DEST, 1 << 3);
 398             OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 399             OUT_CS_BUF_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 3),
 400                     0, r300_buffer(r300->oqbo)->domain, 0);
 401         case 3:
 402             /* pipe 2 only */
 403             OUT_CS_REG(R300_SU_REG_DEST, 1 << 2);
 404             OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 405             OUT_CS_BUF_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 2),
 406                     0, r300_buffer(r300->oqbo)->domain, 0);
 407         case 2:
 408             /* pipe 1 only */
 409             /* As mentioned above, accomodate RV380 and older. */
 410             OUT_CS_REG(R300_SU_REG_DEST,
 411                     1 << (caps->high_second_pipe ? 3 : 1));
 412             OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 413             OUT_CS_BUF_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 1),
 414                     0, r300_buffer(r300->oqbo)->domain, 0);
 415         case 1:
 416             /* pipe 0 only */
 417             OUT_CS_REG(R300_SU_REG_DEST, 1 << 0);
 418             OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 419             OUT_CS_BUF_RELOC(r300->oqbo, query->offset + (sizeof(uint32_t) * 0),
 420                     0, r300_buffer(r300->oqbo)->domain, 0);
 421             break;
 422         default:
 423             fprintf(stderr, "r300: Implementation error: Chipset reports %d"
 424                     " pixel pipes!\n", caps->num_frag_pipes);
 425             abort();
 426     }
 427
 428     /* And, finally, reset it to normal... */
 429     OUT_CS_REG(R300_SU_REG_DEST, 0xF);
 430     END_CS;
 431 }
 432
 433 static void rv530_emit_query_end_single_z(struct r300_context *r300,
 434                                           struct r300_query *query)
 435 {
 436     CS_LOCALS(r300);
 437
 438     BEGIN_CS(8);
 439     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_0);
 440     OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 441     OUT_CS_BUF_RELOC(r300->oqbo, query->offset, 0, r300_buffer(r300->oqbo)->domain, 0);
 442     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_ALL);
 443     END_CS;
 444 }
 445
 446 static void rv530_emit_query_end_double_z(struct r300_context *r300,
 447                                           struct r300_query *query)
 448 {
 449     CS_LOCALS(r300);
 450
 451     BEGIN_CS(14);
 452     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_0);
 453     OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 454     OUT_CS_BUF_RELOC(r300->oqbo, query->offset, 0, r300_buffer(r300->oqbo)->domain, 0);
 455     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_1);
 456     OUT_CS_REG_SEQ(R300_ZB_ZPASS_ADDR, 1);
 457     OUT_CS_BUF_RELOC(r300->oqbo, query->offset + sizeof(uint32_t), 0, r300_buffer(r300->oqbo)->domain, 0);
 458     OUT_CS_REG(RV530_FG_ZBREG_DEST, RV530_FG_ZBREG_DEST_PIPE_SELECT_ALL);
 459     END_CS;
 460 }
 461
 462 void r300_emit_query_end(struct r300_context* r300)
 463 {
 464     struct r300_capabilities *caps = &r300->screen->caps;
 465     struct r300_query *query = r300->query_current;
 466
 467     if (!query)
 468         return;
 469
 470     if (query->begin_emitted == FALSE)
 471         return;
 472
 473     if (caps->family == CHIP_FAMILY_RV530) {
 474         if (caps->num_z_pipes == 2)
 475             rv530_emit_query_end_double_z(r300, query);
 476         else
 477             rv530_emit_query_end_single_z(r300, query);
 478     } else
 479         r300_emit_query_end_frag_pipes(r300, query);
 480
 481     query->begin_emitted = FALSE;
 482 }
 483
 484 void r300_emit_rs_state(struct r300_context* r300, unsigned size, void* state)
 485 {
 486     struct r300_rs_state* rs = (struct r300_rs_state*)state;
 487     float scale, offset;
 488     CS_LOCALS(r300);
 489
 490     BEGIN_CS(size);
 491     OUT_CS_REG(R300_VAP_CNTL_STATUS, rs->vap_control_status);
 492
 493     OUT_CS_REG(R300_GB_AA_CONFIG, rs->antialiasing_config);
 494
 495     OUT_CS_REG(R300_GA_POINT_SIZE, rs->point_size);
 496     OUT_CS_REG_SEQ(R300_GA_POINT_MINMAX, 2);
 497     OUT_CS(rs->point_minmax);
 498     OUT_CS(rs->line_control);
 499
 500     if (rs->polygon_offset_enable) {
 501         scale = rs->depth_scale * 12;
 502         offset = rs->depth_offset;
 503
 504         switch (r300->zbuffer_bpp) {
 505             case 16:
 506                 offset *= 4;
 507                 break;
 508             case 24:
 509                 offset *= 2;
 510                 break;
 511         }
 512
 513         OUT_CS_REG_SEQ(R300_SU_POLY_OFFSET_FRONT_SCALE, 4);
 514         OUT_CS_32F(scale);
 515         OUT_CS_32F(offset);
 516         OUT_CS_32F(scale);
 517         OUT_CS_32F(offset);
 518     }
 519
 520     OUT_CS_REG_SEQ(R300_SU_POLY_OFFSET_ENABLE, 2);
 521     OUT_CS(rs->polygon_offset_enable);
 522     OUT_CS(rs->cull_mode);
 523     OUT_CS_REG(R300_GA_LINE_STIPPLE_CONFIG, rs->line_stipple_config);
 524     OUT_CS_REG(R300_GA_LINE_STIPPLE_VALUE, rs->line_stipple_value);
 525     OUT_CS_REG(R300_GA_POLY_MODE, rs->polygon_mode);
 526     OUT_CS_REG(R300_SC_CLIP_RULE, rs->clip_rule);
 527     OUT_CS_REG(R300_GB_ENABLE, rs->stuffing_enable);
 528     OUT_CS_REG_SEQ(R300_GA_POINT_S0, 4);
 529     OUT_CS_32F(rs->point_texcoord_left);
 530     OUT_CS_32F(rs->point_texcoord_bottom);
 531     OUT_CS_32F(rs->point_texcoord_right);
 532     OUT_CS_32F(rs->point_texcoord_top);
 533     END_CS;
 534 }
 535
 536 void r300_emit_rs_block_state(struct r300_context* r300,
 537                               unsigned size, void* state)
 538 {
 539     struct r300_rs_block* rs = (struct r300_rs_block*)state;
 540     unsigned i;
 541     /* It's the same for both INST and IP tables */
 542     unsigned count = (rs->inst_count & R300_RS_INST_COUNT_MASK) + 1;
 543     CS_LOCALS(r300);
 544
 545     if (SCREEN_DBG_ON(r300->screen, DBG_DRAW)) {
 546         r500_dump_rs_block(rs);
 547     }
 548
 549     DBG(r300, DBG_DRAW, "r300: RS emit:\n");
 550
 551     BEGIN_CS(size);
 552     OUT_CS_REG_SEQ(R300_VAP_VTX_STATE_CNTL, 2);
 553     OUT_CS(rs->vap_vtx_state_cntl);
 554     OUT_CS(rs->vap_vsm_vtx_assm);
 555     OUT_CS_REG_SEQ(R300_VAP_OUTPUT_VTX_FMT_0, 2);
 556     OUT_CS(rs->vap_out_vtx_fmt[0]);
 557     OUT_CS(rs->vap_out_vtx_fmt[1]);
 558
 559     if (r300->screen->caps.is_r500) {
 560         OUT_CS_REG_SEQ(R500_RS_IP_0, count);
 561     } else {
 562         OUT_CS_REG_SEQ(R300_RS_IP_0, count);
 563     }
 564     OUT_CS_TABLE(rs->ip, count);
 565     for (i = 0; i < count; i++) {
 566         DBG(r300, DBG_DRAW, "    : ip %d: 0x%08x\n", i, rs->ip[i]);
 567     }
 568
 569     OUT_CS_REG_SEQ(R300_RS_COUNT, 2);
 570     OUT_CS(rs->count);
 571     OUT_CS(rs->inst_count);
 572
 573     if (r300->screen->caps.is_r500) {
 574         OUT_CS_REG_SEQ(R500_RS_INST_0, count);
 575     } else {
 576         OUT_CS_REG_SEQ(R300_RS_INST_0, count);
 577     }
 578     OUT_CS_TABLE(rs->inst, count);
 579     for (i = 0; i < count; i++) {
 580         DBG(r300, DBG_DRAW, "    : inst %d: 0x%08x\n", i, rs->inst[i]);
 581     }
 582
 583     DBG(r300, DBG_DRAW, "    : count: 0x%08x inst_count: 0x%08x\n",
 584         rs->count, rs->inst_count);
 585
 586     END_CS;
 587 }
 588
 589 void r300_emit_scissor_state(struct r300_context* r300,
 590                              unsigned size, void* state)
 591 {
 592     struct pipe_scissor_state* scissor = (struct pipe_scissor_state*)state;
 593     CS_LOCALS(r300);
 594
 595     BEGIN_CS(size);
 596     OUT_CS_REG_SEQ(R300_SC_CLIPRECT_TL_0, 2);
 597     if (r300->screen->caps.is_r500) {
 598         OUT_CS((scissor->minx << R300_CLIPRECT_X_SHIFT) |
 599                (scissor->miny << R300_CLIPRECT_Y_SHIFT));
 600         OUT_CS(((scissor->maxx - 1) << R300_CLIPRECT_X_SHIFT) |
 601                ((scissor->maxy - 1) << R300_CLIPRECT_Y_SHIFT));
 602     } else {
 603         OUT_CS(((scissor->minx + 1440) << R300_CLIPRECT_X_SHIFT) |
 604                ((scissor->miny + 1440) << R300_CLIPRECT_Y_SHIFT));
 605         OUT_CS(((scissor->maxx + 1440-1) << R300_CLIPRECT_X_SHIFT) |
 606                ((scissor->maxy + 1440-1) << R300_CLIPRECT_Y_SHIFT));
 607     }
 608     END_CS;
 609 }
 610
 611 void r300_emit_textures_state(struct r300_context *r300,
 612                               unsigned size, void *state)
 613 {
 614     struct r300_textures_state *allstate = (struct r300_textures_state*)state;
 615     struct r300_texture_sampler_state *texstate;
 616     struct r300_texture *tex;
 617     unsigned i;
 618     CS_LOCALS(r300);
 619
 620     BEGIN_CS(size);
 621     OUT_CS_REG(R300_TX_ENABLE, allstate->tx_enable);
 622
 623     for (i = 0; i < allstate->count; i++) {
 624         if ((1 << i) & allstate->tx_enable) {
 625             texstate = &allstate->regs[i];
 626             tex = r300_texture(allstate->sampler_views[i]->base.texture);
 627
 628             OUT_CS_REG(R300_TX_FILTER0_0 + (i * 4), texstate->filter0);
 629             OUT_CS_REG(R300_TX_FILTER1_0 + (i * 4), texstate->filter1);
 630             OUT_CS_REG(R300_TX_BORDER_COLOR_0 + (i * 4),
 631                        texstate->border_color);
 632
 633             OUT_CS_REG(R300_TX_FORMAT0_0 + (i * 4), texstate->format.format0);
 634             OUT_CS_REG(R300_TX_FORMAT1_0 + (i * 4), texstate->format.format1);
 635             OUT_CS_REG(R300_TX_FORMAT2_0 + (i * 4), texstate->format.format2);
 636
 637             OUT_CS_REG_SEQ(R300_TX_OFFSET_0 + (i * 4), 1);
 638             OUT_CS_TEX_RELOC(tex, texstate->format.tile_config, tex->domain,
 639                              0, 0);
 640         }
 641     }
 642     END_CS;
 643 }
 644
 645 void r300_emit_aos(struct r300_context* r300, int offset, boolean indexed)
 646 {
 647     struct pipe_vertex_buffer *vb1, *vb2, *vbuf = r300->vertex_buffer;
 648     struct pipe_vertex_element *velem = r300->velems->velem;
 649     struct r300_buffer *buf;
 650     int i;
 651     unsigned *hw_format_size = r300->velems->hw_format_size;
 652     unsigned size1, size2, aos_count = r300->velems->count;
 653     unsigned packet_size = (aos_count * 3 + 1) / 2;
 654     CS_LOCALS(r300);
 655
 656     BEGIN_CS(2 + packet_size + aos_count * 2);
 657     OUT_CS_PKT3(R300_PACKET3_3D_LOAD_VBPNTR, packet_size);
 658     OUT_CS(aos_count | (!indexed ? R300_VC_FORCE_PREFETCH : 0));
 659
 660     for (i = 0; i < aos_count - 1; i += 2) {
 661         vb1 = &vbuf[velem[i].vertex_buffer_index];
 662         vb2 = &vbuf[velem[i+1].vertex_buffer_index];
 663         size1 = hw_format_size[i];
 664         size2 = hw_format_size[i+1];
 665
 666         OUT_CS(R300_VBPNTR_SIZE0(size1) | R300_VBPNTR_STRIDE0(vb1->stride) |
 667                R300_VBPNTR_SIZE1(size2) | R300_VBPNTR_STRIDE1(vb2->stride));
 668         OUT_CS(vb1->buffer_offset + velem[i].src_offset   + offset * vb1->stride);
 669         OUT_CS(vb2->buffer_offset + velem[i+1].src_offset + offset * vb2->stride);
 670     }
 671
 672     if (aos_count & 1) {
 673         vb1 = &vbuf[velem[i].vertex_buffer_index];
 674         size1 = hw_format_size[i];
 675
 676         OUT_CS(R300_VBPNTR_SIZE0(size1) | R300_VBPNTR_STRIDE0(vb1->stride));
 677         OUT_CS(vb1->buffer_offset + velem[i].src_offset + offset * vb1->stride);
 678     }
 679
 680     for (i = 0; i < aos_count; i++) {
 681         buf = r300_buffer(vbuf[velem[i].vertex_buffer_index].buffer);
 682         OUT_CS_BUF_RELOC_NO_OFFSET(&buf->b.b, buf->domain, 0, 0);
 683     }
 684     END_CS;
 685 }
 686
 687 void r300_emit_aos_swtcl(struct r300_context *r300, boolean indexed)
 688 {
 689     CS_LOCALS(r300);
 690
 691     DBG(r300, DBG_DRAW, "r300: Preparing vertex buffer %p for render, "
 692             "vertex size %d\n", r300->vbo,
 693             r300->vertex_info.size);
 694     /* Set the pointer to our vertex buffer. The emitted values are this:
 695      * PACKET3 [3D_LOAD_VBPNTR]
 696      * COUNT   [1]
 697      * FORMAT  [size | stride << 8]
 698      * OFFSET  [offset into BO]
 699      * VBPNTR  [relocated BO]
 700      */
 701     BEGIN_CS(7);
 702     OUT_CS_PKT3(R300_PACKET3_3D_LOAD_VBPNTR, 3);
 703     OUT_CS(1 | (!indexed ? R300_VC_FORCE_PREFETCH : 0));
 704     OUT_CS(r300->vertex_info.size |
 705             (r300->vertex_info.size << 8));
 706     OUT_CS(r300->vbo_offset);
 707     OUT_CS_BUF_RELOC(r300->vbo, 0, r300_buffer(r300->vbo)->domain, 0, 0);
 708     END_CS;
 709 }
 710
 711 void r300_emit_vertex_stream_state(struct r300_context* r300,
 712                                    unsigned size, void* state)
 713 {
 714     struct r300_vertex_stream_state *streams =
 715         (struct r300_vertex_stream_state*)state;
 716     unsigned i;
 717     CS_LOCALS(r300);
 718
 719     DBG(r300, DBG_DRAW, "r300: PSC emit:\n");
 720
 721     BEGIN_CS(size);
 722     OUT_CS_REG_SEQ(R300_VAP_PROG_STREAM_CNTL_0, streams->count);
 723     OUT_CS_TABLE(streams->vap_prog_stream_cntl, streams->count);
 724     for (i = 0; i < streams->count; i++) {
 725         DBG(r300, DBG_DRAW, "    : prog_stream_cntl%d: 0x%08x\n", i,
 726                streams->vap_prog_stream_cntl[i]);
 727     }
 728     OUT_CS_REG_SEQ(R300_VAP_PROG_STREAM_CNTL_EXT_0, streams->count);
 729     OUT_CS_TABLE(streams->vap_prog_stream_cntl_ext, streams->count);
 730     for (i = 0; i < streams->count; i++) {
 731         DBG(r300, DBG_DRAW, "    : prog_stream_cntl_ext%d: 0x%08x\n", i,
 732                streams->vap_prog_stream_cntl_ext[i]);
 733     }
 734     END_CS;
 735 }
 736
 737 void r300_emit_pvs_flush(struct r300_context* r300, unsigned size, void* state)
 738 {
 739     CS_LOCALS(r300);
 740
 741     BEGIN_CS(size);
 742     OUT_CS_REG(R300_VAP_PVS_STATE_FLUSH_REG, 0x0);
 743     END_CS;
 744 }
 745
 746 void r300_emit_vs_state(struct r300_context* r300, unsigned size, void* state)
 747 {
 748     struct r300_vertex_shader* vs = (struct r300_vertex_shader*)state;
 749     struct r300_vertex_program_code* code = &vs->code;
 750     struct r300_screen* r300screen = r300->screen;
 751     unsigned instruction_count = code->length / 4;
 752     unsigned i;
 753
 754     unsigned vtx_mem_size = r300screen->caps.is_r500 ? 128 : 72;
 755     unsigned input_count = MAX2(util_bitcount(code->InputsRead), 1);
 756     unsigned output_count = MAX2(util_bitcount(code->OutputsWritten), 1);
 757     unsigned temp_count = MAX2(code->num_temporaries, 1);
 758
 759     unsigned pvs_num_slots = MIN3(vtx_mem_size / input_count,
 760                                   vtx_mem_size / output_count, 10);
 761     unsigned pvs_num_controllers = MIN2(vtx_mem_size / temp_count, 6);
 762
 763     unsigned imm_first = vs->externals_count;
 764     unsigned imm_end = vs->code.constants.Count;
 765     unsigned imm_count = vs->immediates_count;
 766
 767     CS_LOCALS(r300);
 768
 769     BEGIN_CS(size);
 770     /* R300_VAP_PVS_CODE_CNTL_0
 771      * R300_VAP_PVS_CONST_CNTL
 772      * R300_VAP_PVS_CODE_CNTL_1
 773      * See the r5xx docs for instructions on how to use these. */
 774     OUT_CS_REG_SEQ(R300_VAP_PVS_CODE_CNTL_0, 3);
 775     OUT_CS(R300_PVS_FIRST_INST(0) |
 776             R300_PVS_XYZW_VALID_INST(instruction_count - 1) |
 777             R300_PVS_LAST_INST(instruction_count - 1));
 778     OUT_CS(R300_PVS_MAX_CONST_ADDR(code->constants.Count - 1));
 779     OUT_CS(instruction_count - 1);
 780
 781     OUT_CS_REG(R300_VAP_PVS_VECTOR_INDX_REG, 0);
 782     OUT_CS_ONE_REG(R300_VAP_PVS_UPLOAD_DATA, code->length);
 783     OUT_CS_TABLE(code->body.d, code->length);
 784
 785     OUT_CS_REG(R300_VAP_CNTL, R300_PVS_NUM_SLOTS(pvs_num_slots) |
 786             R300_PVS_NUM_CNTLRS(pvs_num_controllers) |
 787             R300_PVS_NUM_FPUS(r300screen->caps.num_vert_fpus) |
 788             R300_PVS_VF_MAX_VTX_NUM(12) |
 789             (r300screen->caps.is_r500 ? R500_TCL_STATE_OPTIMIZATION : 0));
 790
 791     /* Emit immediates. */
 792     if (imm_count) {
 793         OUT_CS_REG(R300_VAP_PVS_VECTOR_INDX_REG,
 794                    (r300->screen->caps.is_r500 ?
 795                    R500_PVS_CONST_START : R300_PVS_CONST_START) +
 796                    imm_first);
 797         OUT_CS_ONE_REG(R300_VAP_PVS_UPLOAD_DATA, imm_count * 4);
 798         for (i = imm_first; i < imm_end; i++) {
 799             const float *data = vs->code.constants.Constants[i].u.Immediate;
 800             OUT_CS_TABLE(data, 4);
 801         }
 802     }
 803     END_CS;
 804 }
 805
 806 void r300_emit_vs_constants(struct r300_context* r300,
 807                             unsigned size, void *state)
 808 {
 809     unsigned count =
 810         ((struct r300_vertex_shader*)r300->vs_state.state)->externals_count;
 811     struct r300_constant_buffer *buf = (struct r300_constant_buffer*)state;
 812     CS_LOCALS(r300);
 813
 814     if (!count)
 815         return;
 816
 817     BEGIN_CS(size);
 818     OUT_CS_REG(R300_VAP_PVS_VECTOR_INDX_REG,
 819                (r300->screen->caps.is_r500 ?
 820                R500_PVS_CONST_START : R300_PVS_CONST_START));
 821     OUT_CS_ONE_REG(R300_VAP_PVS_UPLOAD_DATA, count * 4);
 822     OUT_CS_TABLE(buf->constants, count * 4);
 823     END_CS;
 824 }
 825
 826 void r300_emit_viewport_state(struct r300_context* r300,
 827                               unsigned size, void* state)
 828 {
 829     struct r300_viewport_state* viewport = (struct r300_viewport_state*)state;
 830     CS_LOCALS(r300);
 831
 832     BEGIN_CS(size);
 833     OUT_CS_REG_SEQ(R300_SE_VPORT_XSCALE, 6);
 834     OUT_CS_32F(viewport->xscale);
 835     OUT_CS_32F(viewport->xoffset);
 836     OUT_CS_32F(viewport->yscale);
 837     OUT_CS_32F(viewport->yoffset);
 838     OUT_CS_32F(viewport->zscale);
 839     OUT_CS_32F(viewport->zoffset);
 840     OUT_CS_REG(R300_VAP_VTE_CNTL, viewport->vte_control);
 841     END_CS;
 842 }
 843
 844 void r300_emit_ztop_state(struct r300_context* r300,
 845                           unsigned size, void* state)
 846 {
 847     struct r300_ztop_state* ztop = (struct r300_ztop_state*)state;
 848     CS_LOCALS(r300);
 849
 850     BEGIN_CS(size);
 851     OUT_CS_REG(R300_ZB_ZTOP, ztop->z_buffer_top);
 852     END_CS;
 853 }
 854
 855 void r300_emit_texture_cache_inval(struct r300_context* r300, unsigned size, void* state)
 856 {
 857     CS_LOCALS(r300);
 858
 859     BEGIN_CS(size);
 860     OUT_CS_REG(R300_TX_INVALTAGS, 0);
 861     END_CS;
 862 }
 863
 864 void r300_emit_buffer_validate(struct r300_context *r300,
 865                                boolean do_validate_vertex_buffers,
 866                                struct pipe_resource *index_buffer)
 867 {
 868     struct pipe_framebuffer_state* fb =
 869         (struct pipe_framebuffer_state*)r300->fb_state.state;
 870     struct r300_textures_state *texstate =
 871         (struct r300_textures_state*)r300->textures_state.state;
 872     struct r300_texture* tex;
 873     struct pipe_vertex_buffer *vbuf = r300->vertex_buffer;
 874     struct pipe_vertex_element *velem = r300->velems->velem;
 875     struct pipe_resource *pbuf;
 876     unsigned i;
 877     boolean invalid = FALSE;
 878
 879     /* upload buffers first */
 880     if (r300->screen->caps.has_tcl && r300->any_user_vbs) {
 881         r300_upload_user_buffers(r300);
 882         r300->any_user_vbs = false;
 883     }
 884
 885     /* Clean out BOs. */
 886     r300->rws->reset_bos(r300->rws);
 887
 888 validate:
 889     /* Color buffers... */
 890     for (i = 0; i < fb->nr_cbufs; i++) {
 891         tex = r300_texture(fb->cbufs[i]->texture);
 892         assert(tex && tex->buffer && "cbuf is marked, but NULL!");
 893         if (!r300_add_texture(r300->rws, tex, 0, tex->domain)) {
 894             r300->context.flush(&r300->context, 0, NULL);
 895             goto validate;
 896         }
 897     }
 898     /* ...depth buffer... */
 899     if (fb->zsbuf) {
 900         tex = r300_texture(fb->zsbuf->texture);
 901         assert(tex && tex->buffer && "zsbuf is marked, but NULL!");
 902         if (!r300_add_texture(r300->rws, tex,
 903                               0, tex->domain)) {
 904             r300->context.flush(&r300->context, 0, NULL);
 905             goto validate;
 906         }
 907     }
 908     /* ...textures... */
 909     for (i = 0; i < texstate->count; i++) {
 910         if (!(texstate->tx_enable & (1 << i))) {
 911             continue;
 912         }
 913
 914         tex = r300_texture(texstate->sampler_views[i]->base.texture);
 915         if (!r300_add_texture(r300->rws, tex, tex->domain, 0)) {
 916             r300->context.flush(&r300->context, 0, NULL);
 917             goto validate;
 918         }
 919     }
 920     /* ...occlusion query buffer... */
 921     if (r300->query_start.dirty ||
 922         (r300->query_current && r300->query_current->begin_emitted)) {
 923         if (!r300_add_buffer(r300->rws, r300->oqbo,
 924                              0, r300_buffer(r300->oqbo)->domain)) {
 925             r300->context.flush(&r300->context, 0, NULL);
 926             goto validate;
 927         }
 928     }
 929     /* ...vertex buffer for SWTCL path... */
 930     if (r300->vbo) {
 931         if (!r300_add_buffer(r300->rws, r300->vbo,
 932                              r300_buffer(r300->vbo)->domain, 0)) {
 933             r300->context.flush(&r300->context, 0, NULL);
 934             goto validate;
 935         }
 936     }
 937     /* ...vertex buffers for HWTCL path... */
 938     if (do_validate_vertex_buffers) {
 939         for (i = 0; i < r300->velems->count; i++) {
 940             pbuf = vbuf[velem[i].vertex_buffer_index].buffer;
 941
 942             if (!r300_add_buffer(r300->rws, pbuf,
 943                                  r300_buffer(pbuf)->domain, 0)) {
 944                 r300->context.flush(&r300->context, 0, NULL);
 945                 goto validate;
 946             }
 947         }
 948     }
 949     /* ...and index buffer for HWTCL path. */
 950     if (index_buffer) {
 951         if (!r300_add_buffer(r300->rws, index_buffer,
 952                              r300_buffer(index_buffer)->domain, 0)) {
 953             r300->context.flush(&r300->context, 0, NULL);
 954             goto validate;
 955         }
 956     }
 957     if (!r300->rws->validate(r300->rws)) {
 958         r300->context.flush(&r300->context, 0, NULL);
 959         if (invalid) {
 960             /* Well, hell. */
 961             fprintf(stderr, "r300: Stuck in validation loop, gonna quit now.\n");
 962             abort();
 963         }
 964         invalid = TRUE;
 965         goto validate;
 966     }
 967 }
 968
 969 unsigned r300_get_num_dirty_dwords(struct r300_context *r300)
 970 {
 971     struct r300_atom* atom;
 972     unsigned dwords = 0;
 973
 974     foreach(atom, &r300->atom_list) {
 975         if (atom->dirty) {
 976             dwords += atom->size;
 977         }
 978     }
 979
 980     /* let's reserve some more, just in case */
 981     dwords += 32;
 982
 983     return dwords;
 984 }
 985
 986 /* Emit all dirty state. */
 987 void r300_emit_dirty_state(struct r300_context* r300)
 988 {
 989     struct r300_atom* atom;
 990
 991     foreach(atom, &r300->atom_list) {
 992         if (atom->dirty) {
 993             atom->emit(r300, atom->size, atom->state);
 994             if (SCREEN_DBG_ON(r300->screen, DBG_STATS)) {
 995                 atom->counter++;
 996             }
 997             atom->dirty = FALSE;
 998         }
 999     }
1000
1001     r300->dirty_hw++;
1002 }