src/mesa/drivers/dri/r300/r300_render.c

   1 /**************************************************************************
   2
   3 Copyright (C) 2004 Nicolai Haehnle.
   4
   5 All Rights Reserved.
   6
   7 Permission is hereby granted, free of charge, to any person obtaining a
   8 copy of this software and associated documentation files (the "Software"),
   9 to deal in the Software without restriction, including without limitation
  10 on the rights to use, copy, modify, merge, publish, distribute, sub
  11 license, and/or sell copies of the Software, and to permit persons to whom
  12 the Software is furnished to do so, subject to the following conditions:
  13
  14 The above copyright notice and this permission notice (including the next
  15 paragraph) shall be included in all copies or substantial portions of the
  16 Software.
  17
  18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  19 IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  20 FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  21 ATI, VA LINUX SYSTEMS AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  22 DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  23 OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  24 USE OR OTHER DEALINGS IN THE SOFTWARE.
  25
  26 **************************************************************************/
  27
  28 /**
  29  * \file
  30  *
  31  * \brief R300 Render (Vertex Buffer Implementation)
  32  *
  33  * The immediate implementation has been removed from CVS in favor of the vertex
  34  * buffer implementation.
  35  *
  36  * The render functions are called by the pipeline manager to render a batch of
  37  * primitives. They return TRUE to pass on to the next stage (i.e. software
  38  * rasterization) or FALSE to indicate that the pipeline has finished after
  39  * rendering something.
  40  *
  41  * When falling back to software TCL still attempt to use hardware
  42  * rasterization.
  43  *
  44  * I am not sure that the cache related registers are setup correctly, but
  45  * obviously this does work... Further investigation is needed.
  46  *
  47  * \author Nicolai Haehnle <prefect_@gmx.net>
  48  *
  49  * \todo Add immediate implementation back? Perhaps this is useful if there are
  50  * no bugs...
  51  */
  52
  53 #include "r300_render.h"
  54
  55 #include "main/glheader.h"
  56 #include "main/imports.h"
  57 #include "main/enums.h"
  58 #include "main/macros.h"
  59 #include "main/context.h"
  60 #include "main/dd.h"
  61 #include "main/simple_list.h"
  62 #include "main/api_arrayelt.h"
  63 #include "swrast/swrast.h"
  64 #include "swrast_setup/swrast_setup.h"
  65 #include "vbo/vbo.h"
  66 #include "vbo/vbo_split.h"
  67 #include "r300_context.h"
  68 #include "r300_state.h"
  69 #include "r300_reg.h"
  70 #include "r300_emit.h"
  71 #include "r300_swtcl.h"
  72
  73 /**
  74  * \brief Convert a OpenGL primitive type into a R300 primitive type.
  75  */
  76 int r300PrimitiveType(r300ContextPtr rmesa, int prim)
  77 {
  78         switch (prim & PRIM_MODE_MASK) {
  79         case GL_POINTS:
  80                 return R300_VAP_VF_CNTL__PRIM_POINTS;
  81                 break;
  82         case GL_LINES:
  83                 return R300_VAP_VF_CNTL__PRIM_LINES;
  84                 break;
  85         case GL_LINE_STRIP:
  86                 return R300_VAP_VF_CNTL__PRIM_LINE_STRIP;
  87                 break;
  88         case GL_LINE_LOOP:
  89                 return R300_VAP_VF_CNTL__PRIM_LINE_LOOP;
  90                 break;
  91         case GL_TRIANGLES:
  92                 return R300_VAP_VF_CNTL__PRIM_TRIANGLES;
  93                 break;
  94         case GL_TRIANGLE_STRIP:
  95                 return R300_VAP_VF_CNTL__PRIM_TRIANGLE_STRIP;
  96                 break;
  97         case GL_TRIANGLE_FAN:
  98                 return R300_VAP_VF_CNTL__PRIM_TRIANGLE_FAN;
  99                 break;
 100         case GL_QUADS:
 101                 return R300_VAP_VF_CNTL__PRIM_QUADS;
 102                 break;
 103         case GL_QUAD_STRIP:
 104                 return R300_VAP_VF_CNTL__PRIM_QUAD_STRIP;
 105                 break;
 106         case GL_POLYGON:
 107                 return R300_VAP_VF_CNTL__PRIM_POLYGON;
 108                 break;
 109         default:
 110                 assert(0);
 111                 return -1;
 112                 break;
 113         }
 114 }
 115
 116 int r300NumVerts(r300ContextPtr rmesa, int num_verts, int prim)
 117 {
 118         int verts_off = 0;
 119
 120         switch (prim & PRIM_MODE_MASK) {
 121         case GL_POINTS:
 122                 verts_off = 0;
 123                 break;
 124         case GL_LINES:
 125                 verts_off = num_verts % 2;
 126                 break;
 127         case GL_LINE_STRIP:
 128                 if (num_verts < 2)
 129                         verts_off = num_verts;
 130                 break;
 131         case GL_LINE_LOOP:
 132                 if (num_verts < 2)
 133                         verts_off = num_verts;
 134                 break;
 135         case GL_TRIANGLES:
 136                 verts_off = num_verts % 3;
 137                 break;
 138         case GL_TRIANGLE_STRIP:
 139                 if (num_verts < 3)
 140                         verts_off = num_verts;
 141                 break;
 142         case GL_TRIANGLE_FAN:
 143                 if (num_verts < 3)
 144                         verts_off = num_verts;
 145                 break;
 146         case GL_QUADS:
 147                 verts_off = num_verts % 4;
 148                 break;
 149         case GL_QUAD_STRIP:
 150                 if (num_verts < 4)
 151                         verts_off = num_verts;
 152                 else
 153                         verts_off = num_verts % 2;
 154                 break;
 155         case GL_POLYGON:
 156                 if (num_verts < 3)
 157                         verts_off = num_verts;
 158                 break;
 159         default:
 160                 assert(0);
 161                 return -1;
 162                 break;
 163         }
 164
 165         return num_verts - verts_off;
 166 }
 167
 168 static void r300FireEB(r300ContextPtr rmesa, int vertex_count, int type, int offset)
 169 {
 170         BATCH_LOCALS(&rmesa->radeon);
 171         int size;
 172
 173         /* offset is in indices */
 174         BEGIN_BATCH(10);
 175         OUT_BATCH_PACKET3(R300_PACKET3_3D_DRAW_INDX_2, 0);
 176         if (rmesa->ind_buf.is_32bit) {
 177                 /* convert to bytes */
 178                 offset *= 4;
 179                 size = vertex_count;
 180                 OUT_BATCH(R300_VAP_VF_CNTL__PRIM_WALK_INDICES |
 181                   (vertex_count << 16) | type |
 182                   R300_VAP_VF_CNTL__INDEX_SIZE_32bit);
 183         } else {
 184                 /* convert to bytes */
 185                 offset *= 2;
 186                 size = (vertex_count + 1) >> 1;
 187                 OUT_BATCH(R300_VAP_VF_CNTL__PRIM_WALK_INDICES |
 188                    (vertex_count << 16) | type);
 189         }
 190
 191         if (!rmesa->radeon.radeonScreen->kernel_mm) {
 192                 OUT_BATCH_PACKET3(R300_PACKET3_INDX_BUFFER, 2);
 193                 OUT_BATCH(R300_INDX_BUFFER_ONE_REG_WR | (0 << R300_INDX_BUFFER_SKIP_SHIFT) |
 194                                  (R300_VAP_PORT_IDX0 >> 2));
 195                 OUT_BATCH_RELOC(0, rmesa->ind_buf.bo, rmesa->ind_buf.bo_offset + offset, RADEON_GEM_DOMAIN_GTT, 0, 0);
 196                 OUT_BATCH(size);
 197         } else {
 198                 OUT_BATCH_PACKET3(R300_PACKET3_INDX_BUFFER, 2);
 199                 OUT_BATCH(R300_INDX_BUFFER_ONE_REG_WR | (0 << R300_INDX_BUFFER_SKIP_SHIFT) |
 200                                  (R300_VAP_PORT_IDX0 >> 2));
 201                 OUT_BATCH(rmesa->ind_buf.bo_offset + offset);
 202                 OUT_BATCH(size);
 203                 radeon_cs_write_reloc(rmesa->radeon.cmdbuf.cs,
 204                                       rmesa->ind_buf.bo, RADEON_GEM_DOMAIN_GTT, 0, 0);
 205         }
 206         END_BATCH();
 207 }
 208
 209 static void r300EmitAOS(r300ContextPtr rmesa, GLuint nr, GLuint offset)
 210 {
 211         BATCH_LOCALS(&rmesa->radeon);
 212         uint32_t voffset;
 213         int sz = 1 + (nr >> 1) * 3 + (nr & 1) * 2;
 214         int i;
 215
 216         if (RADEON_DEBUG & RADEON_VERTS)
 217                 fprintf(stderr, "%s: nr=%d, ofs=0x%08x\n", __FUNCTION__, nr,
 218                         offset);
 219
 220         if (!rmesa->radeon.radeonScreen->kernel_mm) {
 221                 BEGIN_BATCH(sz+2+(nr * 2));
 222                 OUT_BATCH_PACKET3(R300_PACKET3_3D_LOAD_VBPNTR, sz - 1);
 223                 OUT_BATCH(nr);
 224
 225                 for (i = 0; i + 1 < nr; i += 2) {
 226                         OUT_BATCH((rmesa->radeon.tcl.aos[i].components << 0) |
 227                                   (rmesa->radeon.tcl.aos[i].stride << 8) |
 228                                   (rmesa->radeon.tcl.aos[i + 1].components << 16) |
 229                                   (rmesa->radeon.tcl.aos[i + 1].stride << 24));
 230
 231                         voffset =  rmesa->radeon.tcl.aos[i + 0].offset +
 232                                 offset * 4 * rmesa->radeon.tcl.aos[i + 0].stride;
 233                         OUT_BATCH_RELOC(voffset,
 234                                         rmesa->radeon.tcl.aos[i].bo,
 235                                         voffset,
 236                                         RADEON_GEM_DOMAIN_GTT,
 237                                         0, 0);
 238                         voffset =  rmesa->radeon.tcl.aos[i + 1].offset +
 239                           offset * 4 * rmesa->radeon.tcl.aos[i + 1].stride;
 240                         OUT_BATCH_RELOC(voffset,
 241                                         rmesa->radeon.tcl.aos[i+1].bo,
 242                                         voffset,
 243                                         RADEON_GEM_DOMAIN_GTT,
 244                                         0, 0);
 245                 }
 246
 247                 if (nr & 1) {
 248                         OUT_BATCH((rmesa->radeon.tcl.aos[nr - 1].components << 0) |
 249                                   (rmesa->radeon.tcl.aos[nr - 1].stride << 8));
 250                         voffset =  rmesa->radeon.tcl.aos[nr - 1].offset +
 251                                 offset * 4 * rmesa->radeon.tcl.aos[nr - 1].stride;
 252                         OUT_BATCH_RELOC(voffset,
 253                                         rmesa->radeon.tcl.aos[nr - 1].bo,
 254                                         voffset,
 255                                         RADEON_GEM_DOMAIN_GTT,
 256                                         0, 0);
 257                 }
 258                 END_BATCH();
 259         } else {
 260
 261                 BEGIN_BATCH(sz+2+(nr * 2));
 262                 OUT_BATCH_PACKET3(R300_PACKET3_3D_LOAD_VBPNTR, sz - 1);
 263                 OUT_BATCH(nr);
 264
 265                 for (i = 0; i + 1 < nr; i += 2) {
 266                         OUT_BATCH((rmesa->radeon.tcl.aos[i].components << 0) |
 267                                   (rmesa->radeon.tcl.aos[i].stride << 8) |
 268                                   (rmesa->radeon.tcl.aos[i + 1].components << 16) |
 269                                   (rmesa->radeon.tcl.aos[i + 1].stride << 24));
 270
 271                         voffset =  rmesa->radeon.tcl.aos[i + 0].offset +
 272                                 offset * 4 * rmesa->radeon.tcl.aos[i + 0].stride;
 273                         OUT_BATCH(voffset);
 274                         voffset =  rmesa->radeon.tcl.aos[i + 1].offset +
 275                                 offset * 4 * rmesa->radeon.tcl.aos[i + 1].stride;
 276                         OUT_BATCH(voffset);
 277                 }
 278
 279                 if (nr & 1) {
 280                         OUT_BATCH((rmesa->radeon.tcl.aos[nr - 1].components << 0) |
 281                           (rmesa->radeon.tcl.aos[nr - 1].stride << 8));
 282                         voffset =  rmesa->radeon.tcl.aos[nr - 1].offset +
 283                                 offset * 4 * rmesa->radeon.tcl.aos[nr - 1].stride;
 284                         OUT_BATCH(voffset);
 285                 }
 286                 for (i = 0; i + 1 < nr; i += 2) {
 287                         voffset =  rmesa->radeon.tcl.aos[i + 0].offset +
 288                                 offset * 4 * rmesa->radeon.tcl.aos[i + 0].stride;
 289                         radeon_cs_write_reloc(rmesa->radeon.cmdbuf.cs,
 290                                               rmesa->radeon.tcl.aos[i+0].bo,
 291                                               RADEON_GEM_DOMAIN_GTT,
 292                                               0, 0);
 293                         voffset =  rmesa->radeon.tcl.aos[i + 1].offset +
 294                                 offset * 4 * rmesa->radeon.tcl.aos[i + 1].stride;
 295                         radeon_cs_write_reloc(rmesa->radeon.cmdbuf.cs,
 296                                               rmesa->radeon.tcl.aos[i+1].bo,
 297                                               RADEON_GEM_DOMAIN_GTT,
 298                                               0, 0);
 299                 }
 300                 if (nr & 1) {
 301                         voffset =  rmesa->radeon.tcl.aos[nr - 1].offset +
 302                                 offset * 4 * rmesa->radeon.tcl.aos[nr - 1].stride;
 303                         radeon_cs_write_reloc(rmesa->radeon.cmdbuf.cs,
 304                                               rmesa->radeon.tcl.aos[nr-1].bo,
 305                                               RADEON_GEM_DOMAIN_GTT,
 306                                               0, 0);
 307                 }
 308                 END_BATCH();
 309         }
 310
 311 }
 312
 313 static void r300FireAOS(r300ContextPtr rmesa, int vertex_count, int type)
 314 {
 315         BATCH_LOCALS(&rmesa->radeon);
 316
 317         r300_emit_scissor(rmesa->radeon.glCtx);
 318         BEGIN_BATCH(3);
 319         OUT_BATCH_PACKET3(R300_PACKET3_3D_DRAW_VBUF_2, 0);
 320         OUT_BATCH(R300_VAP_VF_CNTL__PRIM_WALK_VERTEX_LIST | (vertex_count << 16) | type);
 321         END_BATCH();
 322 }
 323
 324 void r300RunRenderPrimitive(struct gl_context * ctx, int start, int end, int prim)
 325 {
 326         r300ContextPtr rmesa = R300_CONTEXT(ctx);
 327         BATCH_LOCALS(&rmesa->radeon);
 328         int type, num_verts;
 329
 330         type = r300PrimitiveType(rmesa, prim);
 331         num_verts = r300NumVerts(rmesa, end - start, prim);
 332
 333         if (type < 0 || num_verts <= 0)
 334                 return;
 335
 336         if (rmesa->ind_buf.bo) {
 337                 GLuint first, incr, offset = 0;
 338
 339                 if (!split_prim_inplace(prim & PRIM_MODE_MASK, &first, &incr) &&
 340                         num_verts > 65500) {
 341                         WARN_ONCE("Fixme: can't handle spliting prim %d\n", prim);
 342                         return;
 343                 }
 344
 345
 346                 r300EmitAOS(rmesa, rmesa->radeon.tcl.aos_count, 0);
 347                 if (rmesa->radeon.radeonScreen->kernel_mm) {
 348                         BEGIN_BATCH_NO_AUTOSTATE(2);
 349                         OUT_BATCH_REGSEQ(R300_VAP_VF_MAX_VTX_INDX, 1);
 350                         OUT_BATCH(rmesa->radeon.tcl.aos[0].count);
 351                         END_BATCH();
 352                 }
 353
 354                 r300_emit_scissor(rmesa->radeon.glCtx);
 355                 while (num_verts > 0) {
 356                         int nr;
 357                         int align;
 358
 359                         nr = MIN2(num_verts, 65535);
 360                         nr -= (nr - first) % incr;
 361
 362                         /* get alignment for IB correct */
 363                         if (nr != num_verts) {
 364                                 do {
 365                                     align = nr * (rmesa->ind_buf.is_32bit ? 4 : 2);
 366                                     if (align % 4)
 367                                         nr -= incr;
 368                                 } while(align % 4);
 369                                 if (nr <= 0) {
 370                                         WARN_ONCE("did the impossible happen? we never aligned nr to dword\n");
 371                                         return;
 372                                 }
 373
 374                         }
 375                         r300FireEB(rmesa, nr, type, offset);
 376
 377                         num_verts -= nr;
 378                         offset += nr;
 379                 }
 380
 381         } else {
 382                 GLuint first, incr, offset = 0;
 383
 384                 if (!split_prim_inplace(prim & PRIM_MODE_MASK, &first, &incr) &&
 385                         num_verts > 65535) {
 386                         WARN_ONCE("Fixme: can't handle spliting prim %d\n", prim);
 387                         return;
 388                 }
 389
 390                 if (rmesa->radeon.radeonScreen->kernel_mm) {
 391                         BEGIN_BATCH_NO_AUTOSTATE(2);
 392                         OUT_BATCH_REGSEQ(R300_VAP_VF_MAX_VTX_INDX, 1);
 393                         OUT_BATCH(rmesa->radeon.tcl.aos[0].count);
 394                         END_BATCH();
 395                 }
 396
 397                 r300_emit_scissor(rmesa->radeon.glCtx);
 398                 while (num_verts > 0) {
 399                         int nr;
 400                         nr = MIN2(num_verts, 65535);
 401                         nr -= (nr - first) % incr;
 402                         r300EmitAOS(rmesa, rmesa->radeon.tcl.aos_count, start + offset);
 403                         r300FireAOS(rmesa, nr, type);
 404                         num_verts -= nr;
 405                         offset += nr;
 406                 }
 407         }
 408         COMMIT_BATCH();
 409 }
 410
 411 static const char *getFallbackString(r300ContextPtr rmesa, uint32_t bit)
 412 {
 413         static char common_fallback_str[32];
 414         switch (bit) {
 415                 case R300_FALLBACK_VERTEX_PROGRAM :
 416                         return "vertex program";
 417                 case R300_FALLBACK_LINE_SMOOTH:
 418                         return "smooth lines";
 419                 case R300_FALLBACK_POINT_SMOOTH:
 420                         return "smooth points";
 421                 case R300_FALLBACK_POLYGON_SMOOTH:
 422                         return "smooth polygons";
 423                 case R300_FALLBACK_LINE_STIPPLE:
 424                         return "line stipple";
 425                 case R300_FALLBACK_POLYGON_STIPPLE:
 426                         return "polygon stipple";
 427                 case R300_FALLBACK_STENCIL_TWOSIDE:
 428                         return "two-sided stencil";
 429                 case R300_FALLBACK_RENDER_MODE:
 430                         return "render mode != GL_RENDER";
 431                 case R300_FALLBACK_FRAGMENT_PROGRAM:
 432                         return "fragment program";
 433                 case R300_FALLBACK_RADEON_COMMON:
 434                         snprintf(common_fallback_str, 32, "radeon common 0x%08x", rmesa->radeon.Fallback);
 435                         return common_fallback_str;
 436                 case R300_FALLBACK_AOS_LIMIT:
 437                         return "aos limit";
 438                 case R300_FALLBACK_INVALID_BUFFERS:
 439                         return "invalid buffers";
 440                 default:
 441                         return "unknown";
 442         }
 443 }
 444
 445 void r300SwitchFallback(struct gl_context *ctx, uint32_t bit, GLboolean mode)
 446 {
 447         TNLcontext *tnl = TNL_CONTEXT(ctx);
 448         r300ContextPtr rmesa = R300_CONTEXT(ctx);
 449         uint32_t old_fallback = rmesa->fallback;
 450         static uint32_t fallback_warn = 0;
 451
 452         if (mode) {
 453                 if ((fallback_warn & bit) == 0) {
 454                         if (RADEON_DEBUG & RADEON_FALLBACKS)
 455                                 fprintf(stderr, "WARNING! Falling back to software for %s\n", getFallbackString(rmesa, bit));
 456                         fallback_warn |= bit;
 457                 }
 458                 rmesa->fallback |= bit;
 459
 460                 /* update only if we change from no tcl fallbacks to some tcl fallbacks */
 461                 if (rmesa->options.hw_tcl_enabled) {
 462                         if (((old_fallback & R300_TCL_FALLBACK_MASK) == 0) &&
 463                                 ((bit & R300_TCL_FALLBACK_MASK) > 0)) {
 464                                 R300_STATECHANGE(rmesa, vap_cntl_status);
 465                                 rmesa->hw.vap_cntl_status.cmd[1] |= R300_VAP_TCL_BYPASS;
 466                         }
 467                 }
 468
 469                 /* update only if we change from no raster fallbacks to some raster fallbacks */
 470                 if (((old_fallback & R300_RASTER_FALLBACK_MASK) == 0) &&
 471                         ((bit & R300_RASTER_FALLBACK_MASK) > 0)) {
 472
 473                         radeon_firevertices(&rmesa->radeon);
 474                         rmesa->radeon.swtcl.RenderIndex = ~0;
 475                         _swsetup_Wakeup( ctx );
 476                 }
 477         } else {
 478                 rmesa->fallback &= ~bit;
 479
 480                 /* update only if we have disabled all tcl fallbacks */
 481                 if (rmesa->options.hw_tcl_enabled) {
 482                         if ((old_fallback & R300_TCL_FALLBACK_MASK) == bit) {
 483                                 R300_STATECHANGE(rmesa, vap_cntl_status);
 484                                 rmesa->hw.vap_cntl_status.cmd[1] &= ~R300_VAP_TCL_BYPASS;
 485                         }
 486                 }
 487
 488                 /* update only if we have disabled all raster fallbacks */
 489                 if ((old_fallback & R300_RASTER_FALLBACK_MASK) == bit) {
 490                         _swrast_flush( ctx );
 491
 492                         tnl->Driver.Render.Start = r300RenderStart;
 493                         tnl->Driver.Render.Finish = r300RenderFinish;
 494                         tnl->Driver.Render.PrimitiveNotify = r300RenderPrimitive;
 495                         tnl->Driver.Render.ResetLineStipple = r300ResetLineStipple;
 496                         tnl->Driver.Render.BuildVertices = _tnl_build_vertices;
 497                         tnl->Driver.Render.CopyPV = _tnl_copy_pv;
 498                         tnl->Driver.Render.Interp = _tnl_interp;
 499
 500                         _tnl_invalidate_vertex_state( ctx, ~0 );
 501                         _tnl_invalidate_vertices( ctx, ~0 );
 502                 }
 503         }
 504
 505 }