src/mesa/drivers/dri/i965/brw_draw.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include <stdlib.h>
  29
  30 #include "main/glheader.h"
  31 #include "main/context.h"
  32 #include "main/state.h"
  33 #include "main/api_validate.h"
  34 #include "main/enums.h"
  35
  36 #include "brw_draw.h"
  37 #include "brw_defines.h"
  38 #include "brw_context.h"
  39 #include "brw_state.h"
  40 #include "brw_fallback.h"
  41
  42 #include "intel_batchbuffer.h"
  43 #include "intel_buffer_objects.h"
  44
  45 #include "tnl/tnl.h"
  46 #include "vbo/vbo_context.h"
  47 #include "swrast/swrast.h"
  48 #include "swrast_setup/swrast_setup.h"
  49
  50 #define FILE_DEBUG_FLAG DEBUG_BATCH
  51
  52 static GLuint prim_to_hw_prim[GL_POLYGON+1] = {
  53    _3DPRIM_POINTLIST,
  54    _3DPRIM_LINELIST,
  55    _3DPRIM_LINELOOP,
  56    _3DPRIM_LINESTRIP,
  57    _3DPRIM_TRILIST,
  58    _3DPRIM_TRISTRIP,
  59    _3DPRIM_TRIFAN,
  60    _3DPRIM_QUADLIST,
  61    _3DPRIM_QUADSTRIP,
  62    _3DPRIM_POLYGON
  63 };
  64
  65
  66 static const GLenum reduced_prim[GL_POLYGON+1] = {
  67    GL_POINTS,
  68    GL_LINES,
  69    GL_LINES,
  70    GL_LINES,
  71    GL_TRIANGLES,
  72    GL_TRIANGLES,
  73    GL_TRIANGLES,
  74    GL_TRIANGLES,
  75    GL_TRIANGLES,
  76    GL_TRIANGLES
  77 };
  78
  79
  80 /* When the primitive changes, set a state bit and re-validate.  Not
  81  * the nicest and would rather deal with this by having all the
  82  * programs be immune to the active primitive (ie. cope with all
  83  * possibilities).  That may not be realistic however.
  84  */
  85 static GLuint brw_set_prim(struct brw_context *brw, GLenum prim)
  86 {
  87    GLcontext *ctx = &brw->intel.ctx;
  88
  89    if (INTEL_DEBUG & DEBUG_PRIMS)
  90       _mesa_printf("PRIM: %s\n", _mesa_lookup_enum_by_nr(prim));
  91
  92    /* Slight optimization to avoid the GS program when not needed:
  93     */
  94    if (prim == GL_QUAD_STRIP &&
  95        ctx->Light.ShadeModel != GL_FLAT &&
  96        ctx->Polygon.FrontMode == GL_FILL &&
  97        ctx->Polygon.BackMode == GL_FILL)
  98       prim = GL_TRIANGLE_STRIP;
  99
 100    if (prim != brw->primitive) {
 101       brw->primitive = prim;
 102       brw->state.dirty.brw |= BRW_NEW_PRIMITIVE;
 103
 104       if (reduced_prim[prim] != brw->intel.reduced_primitive) {
 105          brw->intel.reduced_primitive = reduced_prim[prim];
 106          brw->state.dirty.brw |= BRW_NEW_REDUCED_PRIMITIVE;
 107       }
 108    }
 109
 110    return prim_to_hw_prim[prim];
 111 }
 112
 113
 114 static GLuint trim(GLenum prim, GLuint length)
 115 {
 116    if (prim == GL_QUAD_STRIP)
 117       return length > 3 ? (length - length % 2) : 0;
 118    else if (prim == GL_QUADS)
 119       return length - length % 4;
 120    else
 121       return length;
 122 }
 123
 124
 125 static void brw_emit_prim(struct brw_context *brw,
 126                           const struct _mesa_prim *prim,
 127                           uint32_t hw_prim)
 128 {
 129    struct brw_3d_primitive prim_packet;
 130    struct intel_context *intel = &brw->intel;
 131
 132    if (INTEL_DEBUG & DEBUG_PRIMS)
 133       _mesa_printf("PRIM: %s %d %d\n", _mesa_lookup_enum_by_nr(prim->mode),
 134                    prim->start, prim->count);
 135
 136    prim_packet.header.opcode = CMD_3D_PRIM;
 137    prim_packet.header.length = sizeof(prim_packet)/4 - 2;
 138    prim_packet.header.pad = 0;
 139    prim_packet.header.topology = hw_prim;
 140    prim_packet.header.indexed = prim->indexed;
 141
 142    prim_packet.verts_per_instance = trim(prim->mode, prim->count);
 143    prim_packet.start_vert_location = prim->start;
 144    prim_packet.instance_count = 1;
 145    prim_packet.start_instance_location = 0;
 146    prim_packet.base_vert_location = 0;
 147
 148    /* Can't wrap here, since we rely on the validated state. */
 149    brw->no_batch_wrap = GL_TRUE;
 150
 151    /* If we're set to always flush, do it before and after the primitive emit.
 152     * We want to catch both missed flushes that hurt instruction/state cache
 153     * and missed flushes of the render cache as it heads to other parts of
 154     * the besides the draw code.
 155     */
 156    if (intel->always_flush_cache) {
 157       BEGIN_BATCH(1, IGNORE_CLIPRECTS);
 158       OUT_BATCH(intel->vtbl.flush_cmd());
 159       ADVANCE_BATCH();
 160    }
 161    if (prim_packet.verts_per_instance) {
 162       intel_batchbuffer_data( brw->intel.batch, &prim_packet,
 163                               sizeof(prim_packet), LOOP_CLIPRECTS);
 164    }
 165    if (intel->always_flush_cache) {
 166       BEGIN_BATCH(1, IGNORE_CLIPRECTS);
 167       OUT_BATCH(intel->vtbl.flush_cmd());
 168       ADVANCE_BATCH();
 169    }
 170
 171    brw->no_batch_wrap = GL_FALSE;
 172 }
 173
 174 static void brw_merge_inputs( struct brw_context *brw,
 175                        const struct gl_client_array *arrays[])
 176 {
 177    struct brw_vertex_info old = brw->vb.info;
 178    GLuint i;
 179
 180    for (i = 0; i < VERT_ATTRIB_MAX; i++)
 181       dri_bo_unreference(brw->vb.inputs[i].bo);
 182
 183    memset(&brw->vb.inputs, 0, sizeof(brw->vb.inputs));
 184    memset(&brw->vb.info, 0, sizeof(brw->vb.info));
 185
 186    for (i = 0; i < VERT_ATTRIB_MAX; i++) {
 187       brw->vb.inputs[i].glarray = arrays[i];
 188       brw->vb.inputs[i].attrib = (gl_vert_attrib) i;
 189
 190       if (arrays[i]->StrideB != 0)
 191          brw->vb.info.varying |= 1 << i;
 192
 193          brw->vb.info.sizes[i/16] |= (brw->vb.inputs[i].glarray->Size - 1) <<
 194             ((i%16) * 2);
 195    }
 196
 197    /* Raise statechanges if input sizes and varying have changed:
 198     */
 199    if (memcmp(brw->vb.info.sizes, old.sizes, sizeof(old.sizes)) != 0)
 200       brw->state.dirty.brw |= BRW_NEW_INPUT_DIMENSIONS;
 201
 202    if (brw->vb.info.varying != old.varying)
 203       brw->state.dirty.brw |= BRW_NEW_INPUT_VARYING;
 204 }
 205
 206 /* XXX: could split the primitive list to fallback only on the
 207  * non-conformant primitives.
 208  */
 209 static GLboolean check_fallbacks( struct brw_context *brw,
 210                                   const struct _mesa_prim *prim,
 211                                   GLuint nr_prims )
 212 {
 213    GLcontext *ctx = &brw->intel.ctx;
 214    GLuint i;
 215
 216    /* If we don't require strict OpenGL conformance, never
 217     * use fallbacks.  If we're forcing fallbacks, always
 218     * use fallfacks.
 219     */
 220    if (brw->intel.conformance_mode == 0)
 221       return GL_FALSE;
 222
 223    if (brw->intel.conformance_mode == 2)
 224       return GL_TRUE;
 225
 226    if (ctx->Polygon.SmoothFlag) {
 227       for (i = 0; i < nr_prims; i++)
 228          if (reduced_prim[prim[i].mode] == GL_TRIANGLES)
 229             return GL_TRUE;
 230    }
 231
 232    /* BRW hardware will do AA lines, but they are non-conformant it
 233     * seems.  TBD whether we keep this fallback:
 234     */
 235    if (ctx->Line.SmoothFlag) {
 236       for (i = 0; i < nr_prims; i++)
 237          if (reduced_prim[prim[i].mode] == GL_LINES)
 238             return GL_TRUE;
 239    }
 240
 241    /* Stipple -- these fallbacks could be resolved with a little
 242     * bit of work?
 243     */
 244    if (ctx->Line.StippleFlag) {
 245       for (i = 0; i < nr_prims; i++) {
 246          /* GS doesn't get enough information to know when to reset
 247           * the stipple counter?!?
 248           */
 249          if (prim[i].mode == GL_LINE_LOOP || prim[i].mode == GL_LINE_STRIP)
 250             return GL_TRUE;
 251
 252          if (prim[i].mode == GL_POLYGON &&
 253              (ctx->Polygon.FrontMode == GL_LINE ||
 254               ctx->Polygon.BackMode == GL_LINE))
 255             return GL_TRUE;
 256       }
 257    }
 258
 259    if (ctx->Point.SmoothFlag) {
 260       for (i = 0; i < nr_prims; i++)
 261          if (prim[i].mode == GL_POINTS)
 262             return GL_TRUE;
 263    }
 264
 265    /* BRW hardware doesn't handle GL_CLAMP texturing correctly;
 266     * brw_wm_sampler_state:translate_wrap_mode() treats GL_CLAMP
 267     * as GL_CLAMP_TO_EDGE instead.  If we're using GL_CLAMP, and
 268     * we want strict conformance, force the fallback.
 269     * Right now, we only do this for 2D textures.
 270     */
 271    {
 272       int u;
 273       for (u = 0; u < ctx->Const.MaxTextureCoordUnits; u++) {
 274          struct gl_texture_unit *texUnit = &ctx->Texture.Unit[u];
 275          if (texUnit->Enabled) {
 276             if (texUnit->Enabled & TEXTURE_1D_BIT) {
 277                if (texUnit->CurrentTex[TEXTURE_1D_INDEX]->WrapS == GL_CLAMP) {
 278                    return GL_TRUE;
 279                }
 280             }
 281             if (texUnit->Enabled & TEXTURE_2D_BIT) {
 282                if (texUnit->CurrentTex[TEXTURE_2D_INDEX]->WrapS == GL_CLAMP ||
 283                    texUnit->CurrentTex[TEXTURE_2D_INDEX]->WrapT == GL_CLAMP) {
 284                    return GL_TRUE;
 285                }
 286             }
 287             if (texUnit->Enabled & TEXTURE_3D_BIT) {
 288                if (texUnit->CurrentTex[TEXTURE_3D_INDEX]->WrapS == GL_CLAMP ||
 289                    texUnit->CurrentTex[TEXTURE_3D_INDEX]->WrapT == GL_CLAMP ||
 290                    texUnit->CurrentTex[TEXTURE_3D_INDEX]->WrapR == GL_CLAMP) {
 291                    return GL_TRUE;
 292                }
 293             }
 294          }
 295       }
 296    }
 297
 298    /* Nothing stopping us from the fast path now */
 299    return GL_FALSE;
 300 }
 301
 302 /* May fail if out of video memory for texture or vbo upload, or on
 303  * fallback conditions.
 304  */
 305 static GLboolean brw_try_draw_prims( GLcontext *ctx,
 306                                      const struct gl_client_array *arrays[],
 307                                      const struct _mesa_prim *prim,
 308                                      GLuint nr_prims,
 309                                      const struct _mesa_index_buffer *ib,
 310                                      GLuint min_index,
 311                                      GLuint max_index )
 312 {
 313    struct intel_context *intel = intel_context(ctx);
 314    struct brw_context *brw = brw_context(ctx);
 315    GLboolean retval = GL_FALSE;
 316    GLboolean warn = GL_FALSE;
 317    GLboolean first_time = GL_TRUE;
 318    GLuint i;
 319
 320    if (ctx->NewState)
 321       _mesa_update_state( ctx );
 322
 323    /* We have to validate the textures *before* checking for fallbacks;
 324     * otherwise, the software fallback won't be able to rely on the
 325     * texture state, the firstLevel and lastLevel fields won't be
 326     * set in the intel texture object (they'll both be 0), and the
 327     * software fallback will segfault if it attempts to access any
 328     * texture level other than level 0.
 329     */
 330    brw_validate_textures( brw );
 331
 332    if (check_fallbacks(brw, prim, nr_prims))
 333       return GL_FALSE;
 334
 335    /* Bind all inputs, derive varying and size information:
 336     */
 337    brw_merge_inputs( brw, arrays );
 338
 339    brw->ib.ib = ib;
 340    brw->state.dirty.brw |= BRW_NEW_INDICES;
 341
 342    brw->vb.min_index = min_index;
 343    brw->vb.max_index = max_index;
 344    brw->state.dirty.brw |= BRW_NEW_VERTICES;
 345
 346    /* Have to validate state quite late.  Will rebuild tnl_program,
 347     * which depends on varying information.
 348     *
 349     * Note this is where brw->vs->prog_data.inputs_read is calculated,
 350     * so can't access it earlier.
 351     */
 352
 353    LOCK_HARDWARE(intel);
 354
 355    if (!intel->constant_cliprect && intel->driDrawable->numClipRects == 0) {
 356       UNLOCK_HARDWARE(intel);
 357       return GL_TRUE;
 358    }
 359
 360    for (i = 0; i < nr_prims; i++) {
 361       uint32_t hw_prim;
 362
 363       /* Flush the batch if it's approaching full, so that we don't wrap while
 364        * we've got validated state that needs to be in the same batch as the
 365        * primitives.  This fraction is just a guess (minimal full state plus
 366        * a primitive is around 512 bytes), and would be better if we had
 367        * an upper bound of how much we might emit in a single
 368        * brw_try_draw_prims().
 369        */
 370       intel_batchbuffer_require_space(intel->batch, intel->batch->size / 4,
 371                                       LOOP_CLIPRECTS);
 372
 373       hw_prim = brw_set_prim(brw, prim[i].mode);
 374
 375       if (first_time || (brw->state.dirty.brw & BRW_NEW_PRIMITIVE)) {
 376          first_time = GL_FALSE;
 377
 378          brw_validate_state(brw);
 379
 380          /* Various fallback checks:  */
 381          if (brw->intel.Fallback)
 382             goto out;
 383
 384          /* Check that we can fit our state in with our existing batchbuffer, or
 385           * flush otherwise.
 386           */
 387          if (dri_bufmgr_check_aperture_space(brw->state.validated_bos,
 388                                              brw->state.validated_bo_count)) {
 389             static GLboolean warned;
 390             intel_batchbuffer_flush(intel->batch);
 391
 392             /* Validate the state after we flushed the batch (which would have
 393              * changed the set of dirty state).  If we still fail to
 394              * check_aperture, warn of what's happening, but attempt to continue
 395              * on since it may succeed anyway, and the user would probably rather
 396              * see a failure and a warning than a fallback.
 397              */
 398             brw_validate_state(brw);
 399             if (!warned &&
 400                 dri_bufmgr_check_aperture_space(brw->state.validated_bos,
 401                                                 brw->state.validated_bo_count)) {
 402                warn = GL_TRUE;
 403                warned = GL_TRUE;
 404             }
 405          }
 406
 407          brw_upload_state(brw);
 408       }
 409
 410       brw_emit_prim(brw, &prim[i], hw_prim);
 411
 412       retval = GL_TRUE;
 413    }
 414
 415    if (intel->always_flush_batch)
 416       intel_batchbuffer_flush(intel->batch);
 417  out:
 418    UNLOCK_HARDWARE(intel);
 419
 420    if (warn)
 421       fprintf(stderr, "i965: Single primitive emit potentially exceeded "
 422               "available aperture space\n");
 423
 424    if (!retval)
 425       DBG("%s failed\n", __FUNCTION__);
 426
 427    return retval;
 428 }
 429
 430 static GLboolean brw_need_rebase( GLcontext *ctx,
 431                                   const struct gl_client_array *arrays[],
 432                                   const struct _mesa_index_buffer *ib,
 433                                   GLuint min_index )
 434 {
 435    if (min_index == 0)
 436       return GL_FALSE;
 437
 438    if (ib) {
 439       if (!vbo_all_varyings_in_vbos(arrays))
 440          return GL_TRUE;
 441       else
 442          return GL_FALSE;
 443    }
 444    else {
 445       /* Hmm.  This isn't quite what I wanted.  BRW can actually
 446        * handle the mixed case well enough that we shouldn't need to
 447        * rebase.  However, it's probably not very common, nor hugely
 448        * expensive to do it this way:
 449        */
 450       if (!vbo_all_varyings_in_vbos(arrays))
 451          return GL_TRUE;
 452       else
 453          return GL_FALSE;
 454    }
 455 }
 456
 457
 458 void brw_draw_prims( GLcontext *ctx,
 459                      const struct gl_client_array *arrays[],
 460                      const struct _mesa_prim *prim,
 461                      GLuint nr_prims,
 462                      const struct _mesa_index_buffer *ib,
 463                      GLuint min_index,
 464                      GLuint max_index )
 465 {
 466    GLboolean retval;
 467
 468    /* Decide if we want to rebase.  If so we end up recursing once
 469     * only into this function.
 470     */
 471    if (brw_need_rebase( ctx, arrays, ib, min_index )) {
 472       vbo_rebase_prims( ctx, arrays,
 473                         prim, nr_prims,
 474                         ib, min_index, max_index,
 475                         brw_draw_prims );
 476
 477       return;
 478    }
 479
 480    /* Make a first attempt at drawing:
 481     */
 482    retval = brw_try_draw_prims(ctx, arrays, prim, nr_prims, ib, min_index, max_index);
 483
 484    /* Otherwise, we really are out of memory.  Pass the drawing
 485     * command to the software tnl module and which will in turn call
 486     * swrast to do the drawing.
 487     */
 488    if (!retval) {
 489        _swsetup_Wakeup(ctx);
 490       _tnl_draw_prims(ctx, arrays, prim, nr_prims, ib, min_index, max_index);
 491    }
 492
 493 }
 494
 495 void brw_draw_init( struct brw_context *brw )
 496 {
 497    GLcontext *ctx = &brw->intel.ctx;
 498    struct vbo_context *vbo = vbo_context(ctx);
 499
 500    /* Register our drawing function:
 501     */
 502    vbo->draw_prims = brw_draw_prims;
 503 }
 504
 505 void brw_draw_destroy( struct brw_context *brw )
 506 {
 507    int i;
 508
 509    if (brw->vb.upload.bo != NULL) {
 510       dri_bo_unreference(brw->vb.upload.bo);
 511       brw->vb.upload.bo = NULL;
 512    }
 513
 514    for (i = 0; i < VERT_ATTRIB_MAX; i++) {
 515       dri_bo_unreference(brw->vb.inputs[i].bo);
 516       brw->vb.inputs[i].bo = NULL;
 517    }
 518
 519    dri_bo_unreference(brw->ib.bo);
 520    brw->ib.bo = NULL;
 521 }