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
  29 #include "main/glheader.h"
  30 #include "main/context.h"
  31 #include "main/state.h"
  32 #include "main/enums.h"
  33 #include "tnl/tnl.h"
  34 #include "vbo/vbo_context.h"
  35 #include "swrast/swrast.h"
  36 #include "swrast_setup/swrast_setup.h"
  37
  38 #include "brw_draw.h"
  39 #include "brw_defines.h"
  40 #include "brw_context.h"
  41 #include "brw_state.h"
  42
  43 #include "intel_batchbuffer.h"
  44
  45 #define FILE_DEBUG_FLAG DEBUG_PRIMS
  46
  47 static GLuint prim_to_hw_prim[GL_POLYGON+1] = {
  48    _3DPRIM_POINTLIST,
  49    _3DPRIM_LINELIST,
  50    _3DPRIM_LINELOOP,
  51    _3DPRIM_LINESTRIP,
  52    _3DPRIM_TRILIST,
  53    _3DPRIM_TRISTRIP,
  54    _3DPRIM_TRIFAN,
  55    _3DPRIM_QUADLIST,
  56    _3DPRIM_QUADSTRIP,
  57    _3DPRIM_POLYGON
  58 };
  59
  60
  61 static const GLenum reduced_prim[GL_POLYGON+1] = {
  62    GL_POINTS,
  63    GL_LINES,
  64    GL_LINES,
  65    GL_LINES,
  66    GL_TRIANGLES,
  67    GL_TRIANGLES,
  68    GL_TRIANGLES,
  69    GL_TRIANGLES,
  70    GL_TRIANGLES,
  71    GL_TRIANGLES
  72 };
  73
  74
  75 /* When the primitive changes, set a state bit and re-validate.  Not
  76  * the nicest and would rather deal with this by having all the
  77  * programs be immune to the active primitive (ie. cope with all
  78  * possibilities).  That may not be realistic however.
  79  */
  80 static GLuint brw_set_prim(struct brw_context *brw,
  81                            const struct _mesa_prim *prim)
  82 {
  83    struct gl_context *ctx = &brw->intel.ctx;
  84    GLenum mode = prim->mode;
  85
  86    DBG("PRIM: %s\n", _mesa_lookup_enum_by_nr(prim->mode));
  87
  88    /* Slight optimization to avoid the GS program when not needed:
  89     */
  90    if (mode == GL_QUAD_STRIP &&
  91        ctx->Light.ShadeModel != GL_FLAT &&
  92        ctx->Polygon.FrontMode == GL_FILL &&
  93        ctx->Polygon.BackMode == GL_FILL)
  94       mode = GL_TRIANGLE_STRIP;
  95
  96    if (prim->mode == GL_QUADS && prim->count == 4 &&
  97        ctx->Light.ShadeModel != GL_FLAT &&
  98        ctx->Polygon.FrontMode == GL_FILL &&
  99        ctx->Polygon.BackMode == GL_FILL) {
 100       mode = GL_TRIANGLE_FAN;
 101    }
 102
 103    if (mode != brw->primitive) {
 104       brw->primitive = mode;
 105       brw->state.dirty.brw |= BRW_NEW_PRIMITIVE;
 106
 107       if (reduced_prim[mode] != brw->intel.reduced_primitive) {
 108          brw->intel.reduced_primitive = reduced_prim[mode];
 109          brw->state.dirty.brw |= BRW_NEW_REDUCED_PRIMITIVE;
 110       }
 111    }
 112
 113    return prim_to_hw_prim[mode];
 114 }
 115
 116
 117 static GLuint trim(GLenum prim, GLuint length)
 118 {
 119    if (prim == GL_QUAD_STRIP)
 120       return length > 3 ? (length - length % 2) : 0;
 121    else if (prim == GL_QUADS)
 122       return length - length % 4;
 123    else
 124       return length;
 125 }
 126
 127
 128 static void brw_emit_prim(struct brw_context *brw,
 129                           const struct _mesa_prim *prim,
 130                           uint32_t hw_prim)
 131 {
 132    struct intel_context *intel = &brw->intel;
 133    int verts_per_instance;
 134    int vertex_access_type;
 135    int start_vertex_location;
 136    int base_vertex_location;
 137
 138    DBG("PRIM: %s %d %d\n", _mesa_lookup_enum_by_nr(prim->mode),
 139        prim->start, prim->count);
 140
 141    start_vertex_location = prim->start;
 142    base_vertex_location = prim->basevertex;
 143    if (prim->indexed) {
 144       vertex_access_type = GEN4_3DPRIM_VERTEXBUFFER_ACCESS_RANDOM;
 145       start_vertex_location += brw->ib.start_vertex_offset;
 146       base_vertex_location += brw->vb.start_vertex_bias;
 147    } else {
 148       vertex_access_type = GEN4_3DPRIM_VERTEXBUFFER_ACCESS_SEQUENTIAL;
 149       start_vertex_location += brw->vb.start_vertex_bias;
 150    }
 151
 152    verts_per_instance = trim(prim->mode, prim->count);
 153
 154    /* If nothing to emit, just return. */
 155    if (verts_per_instance == 0)
 156       return;
 157
 158    /* If we're set to always flush, do it before and after the primitive emit.
 159     * We want to catch both missed flushes that hurt instruction/state cache
 160     * and missed flushes of the render cache as it heads to other parts of
 161     * the besides the draw code.
 162     */
 163    if (intel->always_flush_cache) {
 164       intel_batchbuffer_emit_mi_flush(intel);
 165    }
 166
 167    BEGIN_BATCH(6);
 168    OUT_BATCH(CMD_3D_PRIM << 16 | (6 - 2) |
 169              hw_prim << GEN4_3DPRIM_TOPOLOGY_TYPE_SHIFT |
 170              vertex_access_type);
 171    OUT_BATCH(verts_per_instance);
 172    OUT_BATCH(start_vertex_location);
 173    OUT_BATCH(1); // instance count
 174    OUT_BATCH(0); // start instance location
 175    OUT_BATCH(base_vertex_location);
 176    ADVANCE_BATCH();
 177
 178    if (intel->always_flush_cache) {
 179       intel_batchbuffer_emit_mi_flush(intel);
 180    }
 181 }
 182
 183 static void brw_merge_inputs( struct brw_context *brw,
 184                        const struct gl_client_array *arrays[])
 185 {
 186    struct brw_vertex_info old = brw->vb.info;
 187    GLuint i;
 188
 189    for (i = 0; i < brw->vb.nr_buffers; i++) {
 190       drm_intel_bo_unreference(brw->vb.buffers[i].bo);
 191       brw->vb.buffers[i].bo = NULL;
 192    }
 193    brw->vb.nr_buffers = 0;
 194
 195    memset(&brw->vb.info, 0, sizeof(brw->vb.info));
 196
 197    for (i = 0; i < VERT_ATTRIB_MAX; i++) {
 198       brw->vb.inputs[i].buffer = -1;
 199       brw->vb.inputs[i].glarray = arrays[i];
 200       brw->vb.inputs[i].attrib = (gl_vert_attrib) i;
 201
 202       if (arrays[i]->StrideB != 0)
 203          brw->vb.info.sizes[i/16] |= (brw->vb.inputs[i].glarray->Size - 1) <<
 204             ((i%16) * 2);
 205    }
 206
 207    /* Raise statechanges if input sizes have changed. */
 208    if (memcmp(brw->vb.info.sizes, old.sizes, sizeof(old.sizes)) != 0)
 209       brw->state.dirty.brw |= BRW_NEW_INPUT_DIMENSIONS;
 210 }
 211
 212 /* XXX: could split the primitive list to fallback only on the
 213  * non-conformant primitives.
 214  */
 215 static GLboolean check_fallbacks( struct brw_context *brw,
 216                                   const struct _mesa_prim *prim,
 217                                   GLuint nr_prims )
 218 {
 219    struct gl_context *ctx = &brw->intel.ctx;
 220    GLuint i;
 221
 222    /* If we don't require strict OpenGL conformance, never
 223     * use fallbacks.  If we're forcing fallbacks, always
 224     * use fallfacks.
 225     */
 226    if (brw->intel.conformance_mode == 0)
 227       return GL_FALSE;
 228
 229    if (brw->intel.conformance_mode == 2)
 230       return GL_TRUE;
 231
 232    if (ctx->Polygon.SmoothFlag) {
 233       for (i = 0; i < nr_prims; i++)
 234          if (reduced_prim[prim[i].mode] == GL_TRIANGLES)
 235             return GL_TRUE;
 236    }
 237
 238    /* BRW hardware will do AA lines, but they are non-conformant it
 239     * seems.  TBD whether we keep this fallback:
 240     */
 241    if (ctx->Line.SmoothFlag) {
 242       for (i = 0; i < nr_prims; i++)
 243          if (reduced_prim[prim[i].mode] == GL_LINES)
 244             return GL_TRUE;
 245    }
 246
 247    /* Stipple -- these fallbacks could be resolved with a little
 248     * bit of work?
 249     */
 250    if (ctx->Line.StippleFlag) {
 251       for (i = 0; i < nr_prims; i++) {
 252          /* GS doesn't get enough information to know when to reset
 253           * the stipple counter?!?
 254           */
 255          if (prim[i].mode == GL_LINE_LOOP || prim[i].mode == GL_LINE_STRIP)
 256             return GL_TRUE;
 257
 258          if (prim[i].mode == GL_POLYGON &&
 259              (ctx->Polygon.FrontMode == GL_LINE ||
 260               ctx->Polygon.BackMode == GL_LINE))
 261             return GL_TRUE;
 262       }
 263    }
 264
 265    if (ctx->Point.SmoothFlag) {
 266       for (i = 0; i < nr_prims; i++)
 267          if (prim[i].mode == GL_POINTS)
 268             return GL_TRUE;
 269    }
 270
 271    /* BRW hardware doesn't handle GL_CLAMP texturing correctly;
 272     * brw_wm_sampler_state:translate_wrap_mode() treats GL_CLAMP
 273     * as GL_CLAMP_TO_EDGE instead.  If we're using GL_CLAMP, and
 274     * we want strict conformance, force the fallback.
 275     * Right now, we only do this for 2D textures.
 276     */
 277    {
 278       int u;
 279       for (u = 0; u < ctx->Const.MaxTextureCoordUnits; u++) {
 280          struct gl_texture_unit *texUnit = &ctx->Texture.Unit[u];
 281          if (texUnit->Enabled) {
 282             if (texUnit->Enabled & TEXTURE_1D_BIT) {
 283                if (texUnit->CurrentTex[TEXTURE_1D_INDEX]->Sampler.WrapS == GL_CLAMP) {
 284                    return GL_TRUE;
 285                }
 286             }
 287             if (texUnit->Enabled & TEXTURE_2D_BIT) {
 288                if (texUnit->CurrentTex[TEXTURE_2D_INDEX]->Sampler.WrapS == GL_CLAMP ||
 289                    texUnit->CurrentTex[TEXTURE_2D_INDEX]->Sampler.WrapT == GL_CLAMP) {
 290                    return GL_TRUE;
 291                }
 292             }
 293             if (texUnit->Enabled & TEXTURE_3D_BIT) {
 294                if (texUnit->CurrentTex[TEXTURE_3D_INDEX]->Sampler.WrapS == GL_CLAMP ||
 295                    texUnit->CurrentTex[TEXTURE_3D_INDEX]->Sampler.WrapT == GL_CLAMP ||
 296                    texUnit->CurrentTex[TEXTURE_3D_INDEX]->Sampler.WrapR == GL_CLAMP) {
 297                    return GL_TRUE;
 298                }
 299             }
 300          }
 301       }
 302    }
 303
 304    /* Nothing stopping us from the fast path now */
 305    return GL_FALSE;
 306 }
 307
 308 /* May fail if out of video memory for texture or vbo upload, or on
 309  * fallback conditions.
 310  */
 311 static GLboolean brw_try_draw_prims( struct gl_context *ctx,
 312                                      const struct gl_client_array *arrays[],
 313                                      const struct _mesa_prim *prim,
 314                                      GLuint nr_prims,
 315                                      const struct _mesa_index_buffer *ib,
 316                                      GLuint min_index,
 317                                      GLuint max_index )
 318 {
 319    struct intel_context *intel = intel_context(ctx);
 320    struct brw_context *brw = brw_context(ctx);
 321    GLboolean retval = GL_FALSE;
 322    GLboolean warn = GL_FALSE;
 323    GLuint i;
 324
 325    if (ctx->NewState)
 326       _mesa_update_state( ctx );
 327
 328    /* We have to validate the textures *before* checking for fallbacks;
 329     * otherwise, the software fallback won't be able to rely on the
 330     * texture state, the firstLevel and lastLevel fields won't be
 331     * set in the intel texture object (they'll both be 0), and the
 332     * software fallback will segfault if it attempts to access any
 333     * texture level other than level 0.
 334     */
 335    brw_validate_textures( brw );
 336
 337    if (check_fallbacks(brw, prim, nr_prims))
 338       return GL_FALSE;
 339
 340    /* Bind all inputs, derive varying and size information:
 341     */
 342    brw_merge_inputs( brw, arrays );
 343
 344    brw->ib.ib = ib;
 345    brw->state.dirty.brw |= BRW_NEW_INDICES;
 346
 347    brw->vb.min_index = min_index;
 348    brw->vb.max_index = max_index;
 349    brw->state.dirty.brw |= BRW_NEW_VERTICES;
 350
 351    /* Have to validate state quite late.  Will rebuild tnl_program,
 352     * which depends on varying information.
 353     *
 354     * Note this is where brw->vs->prog_data.inputs_read is calculated,
 355     * so can't access it earlier.
 356     */
 357
 358    intel_prepare_render(intel);
 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, 1024, false);
 371
 372       hw_prim = brw_set_prim(brw, &prim[i]);
 373       if (brw->state.dirty.brw) {
 374          brw_validate_state(brw);
 375
 376          /* Various fallback checks:  */
 377          if (brw->intel.Fallback)
 378             goto out;
 379
 380          /* Check that we can fit our state in with our existing batchbuffer, or
 381           * flush otherwise.
 382           */
 383          if (dri_bufmgr_check_aperture_space(brw->state.validated_bos,
 384                                              brw->state.validated_bo_count)) {
 385             static GLboolean warned;
 386             intel_batchbuffer_flush(intel);
 387
 388             /* Validate the state after we flushed the batch (which would have
 389              * changed the set of dirty state).  If we still fail to
 390              * check_aperture, warn of what's happening, but attempt to continue
 391              * on since it may succeed anyway, and the user would probably rather
 392              * see a failure and a warning than a fallback.
 393              */
 394             brw_validate_state(brw);
 395             if (!warned &&
 396                 dri_bufmgr_check_aperture_space(brw->state.validated_bos,
 397                                                 brw->state.validated_bo_count)) {
 398                warn = GL_TRUE;
 399                warned = GL_TRUE;
 400             }
 401          }
 402
 403          intel->no_batch_wrap = GL_TRUE;
 404          brw_upload_state(brw);
 405       }
 406
 407       brw_emit_prim(brw, &prim[i], hw_prim);
 408
 409       intel->no_batch_wrap = GL_FALSE;
 410
 411       retval = GL_TRUE;
 412    }
 413
 414    if (intel->always_flush_batch)
 415       intel_batchbuffer_flush(intel);
 416  out:
 417
 418    brw_state_cache_check_size(brw);
 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 void brw_draw_prims( struct gl_context *ctx,
 431                      const struct gl_client_array *arrays[],
 432                      const struct _mesa_prim *prim,
 433                      GLuint nr_prims,
 434                      const struct _mesa_index_buffer *ib,
 435                      GLboolean index_bounds_valid,
 436                      GLuint min_index,
 437                      GLuint max_index )
 438 {
 439    GLboolean retval;
 440
 441    if (!vbo_all_varyings_in_vbos(arrays)) {
 442       if (!index_bounds_valid)
 443          vbo_get_minmax_index(ctx, prim, ib, &min_index, &max_index);
 444
 445       /* Decide if we want to rebase.  If so we end up recursing once
 446        * only into this function.
 447        */
 448       if (min_index != 0 && !vbo_any_varyings_in_vbos(arrays)) {
 449          vbo_rebase_prims(ctx, arrays,
 450                           prim, nr_prims,
 451                           ib, min_index, max_index,
 452                           brw_draw_prims );
 453          return;
 454       }
 455    }
 456
 457    /* Make a first attempt at drawing:
 458     */
 459    retval = brw_try_draw_prims(ctx, arrays, prim, nr_prims, ib, min_index, max_index);
 460
 461    /* Otherwise, we really are out of memory.  Pass the drawing
 462     * command to the software tnl module and which will in turn call
 463     * swrast to do the drawing.
 464     */
 465    if (!retval) {
 466        _swsetup_Wakeup(ctx);
 467       _tnl_draw_prims(ctx, arrays, prim, nr_prims, ib, min_index, max_index);
 468    }
 469
 470 }
 471
 472 void brw_draw_init( struct brw_context *brw )
 473 {
 474    struct gl_context *ctx = &brw->intel.ctx;
 475    struct vbo_context *vbo = vbo_context(ctx);
 476    int i;
 477
 478    /* Register our drawing function:
 479     */
 480    vbo->draw_prims = brw_draw_prims;
 481
 482    for (i = 0; i < VERT_ATTRIB_MAX; i++)
 483       brw->vb.inputs[i].buffer = -1;
 484    brw->vb.nr_buffers = 0;
 485    brw->vb.nr_enabled = 0;
 486 }
 487
 488 void brw_draw_destroy( struct brw_context *brw )
 489 {
 490    int i;
 491
 492    for (i = 0; i < brw->vb.nr_buffers; i++) {
 493       drm_intel_bo_unreference(brw->vb.buffers[i].bo);
 494       brw->vb.buffers[i].bo = NULL;
 495    }
 496    brw->vb.nr_buffers = 0;
 497
 498    for (i = 0; i < brw->vb.nr_enabled; i++) {
 499       brw->vb.enabled[i]->buffer = -1;
 500    }
 501    brw->vb.nr_enabled = 0;
 502
 503    drm_intel_bo_unreference(brw->ib.bo);
 504    brw->ib.bo = NULL;
 505 }