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 "glheader.h"
  31 #include "context.h"
  32 #include "state.h"
  33 #include "api_validate.h"
  34 #include "enums.h"
  35
  36 #include "brw_draw.h"
  37 #include "brw_defines.h"
  38 #include "brw_context.h"
  39 #include "brw_aub.h"
  40 #include "brw_state.h"
  41 #include "brw_fallback.h"
  42
  43 #include "intel_ioctl.h"
  44 #include "intel_batchbuffer.h"
  45 #include "intel_buffer_objects.h"
  46
  47 #include "tnl/tnl.h"
  48 #include "vbo/vbo_context.h"
  49
  50
  51
  52
  53 static GLuint hw_prim[GL_POLYGON+1] = {
  54    _3DPRIM_POINTLIST,
  55    _3DPRIM_LINELIST,
  56    _3DPRIM_LINELOOP,
  57    _3DPRIM_LINESTRIP,
  58    _3DPRIM_TRILIST,
  59    _3DPRIM_TRISTRIP,
  60    _3DPRIM_TRIFAN,
  61    _3DPRIM_QUADLIST,
  62    _3DPRIM_QUADSTRIP,
  63    _3DPRIM_POLYGON
  64 };
  65
  66
  67 static const GLenum reduced_prim[GL_POLYGON+1] = {
  68    GL_POINTS,
  69    GL_LINES,
  70    GL_LINES,
  71    GL_LINES,
  72    GL_TRIANGLES,
  73    GL_TRIANGLES,
  74    GL_TRIANGLES,
  75    GL_TRIANGLES,
  76    GL_TRIANGLES,
  77    GL_TRIANGLES
  78 };
  79
  80
  81 /* When the primitive changes, set a state bit and re-validate.  Not
  82  * the nicest and would rather deal with this by having all the
  83  * programs be immune to the active primitive (ie. cope with all
  84  * possibilities).  That may not be realistic however.
  85  */
  86 static GLuint brw_set_prim(struct brw_context *brw, GLenum prim)
  87 {
  88    if (INTEL_DEBUG & DEBUG_PRIMS)
  89       _mesa_printf("PRIM: %s\n", _mesa_lookup_enum_by_nr(prim));
  90
  91    /* Slight optimization to avoid the GS program when not needed:
  92     */
  93    if (prim == GL_QUAD_STRIP &&
  94        brw->attribs.Light->ShadeModel != GL_FLAT &&
  95        brw->attribs.Polygon->FrontMode == GL_FILL &&
  96        brw->attribs.Polygon->BackMode == GL_FILL)
  97       prim = GL_TRIANGLE_STRIP;
  98
  99    if (prim != brw->primitive) {
 100       brw->primitive = prim;
 101       brw->state.dirty.brw |= BRW_NEW_PRIMITIVE;
 102
 103       if (reduced_prim[prim] != brw->intel.reduced_primitive) {
 104          brw->intel.reduced_primitive = reduced_prim[prim];
 105          brw->state.dirty.brw |= BRW_NEW_REDUCED_PRIMITIVE;
 106       }
 107
 108       brw_validate_state(brw);
 109    }
 110
 111    return hw_prim[prim];
 112 }
 113
 114
 115 static GLuint trim(GLenum prim, GLuint length)
 116 {
 117    if (prim == GL_QUAD_STRIP)
 118       return length > 3 ? (length - length % 2) : 0;
 119    else if (prim == GL_QUADS)
 120       return length - length % 4;
 121    else
 122       return length;
 123 }
 124
 125
 126 static void brw_emit_cliprect( struct brw_context *brw,
 127                                const drm_clip_rect_t *rect )
 128 {
 129    struct brw_drawrect bdr;
 130
 131    bdr.header.opcode = CMD_DRAW_RECT;
 132    bdr.header.length = sizeof(bdr)/4 - 2;
 133    bdr.xmin = rect->x1;
 134    bdr.xmax = rect->x2 - 1;
 135    bdr.ymin = rect->y1;
 136    bdr.ymax = rect->y2 - 1;
 137    bdr.xorg = brw->intel.drawX;
 138    bdr.yorg = brw->intel.drawY;
 139
 140    intel_batchbuffer_data( brw->intel.batch, &bdr, sizeof(bdr),
 141                            INTEL_BATCH_NO_CLIPRECTS);
 142 }
 143
 144
 145 static void brw_emit_prim( struct brw_context *brw,
 146                            const struct _mesa_prim *prim )
 147
 148 {
 149    struct brw_3d_primitive prim_packet;
 150
 151    if (INTEL_DEBUG & DEBUG_PRIMS)
 152       _mesa_printf("PRIM: %s %d %d\n", _mesa_lookup_enum_by_nr(prim->mode),
 153                    prim->start, prim->count);
 154
 155    prim_packet.header.opcode = CMD_3D_PRIM;
 156    prim_packet.header.length = sizeof(prim_packet)/4 - 2;
 157    prim_packet.header.pad = 0;
 158    prim_packet.header.topology = brw_set_prim(brw, prim->mode);
 159    prim_packet.header.indexed = prim->indexed;
 160
 161    prim_packet.verts_per_instance = trim(prim->mode, prim->count);
 162    prim_packet.start_vert_location = prim->start;
 163    prim_packet.instance_count = 1;
 164    prim_packet.start_instance_location = 0;
 165    prim_packet.base_vert_location = 0;
 166
 167    if (prim_packet.verts_per_instance) {
 168       intel_batchbuffer_data( brw->intel.batch, &prim_packet, sizeof(prim_packet),
 169                               INTEL_BATCH_NO_CLIPRECTS);
 170    }
 171 }
 172
 173 static void brw_merge_inputs( struct brw_context *brw,
 174                        const struct gl_client_array *arrays[])
 175 {
 176    struct brw_vertex_element *inputs = brw->vb.inputs;
 177    struct brw_vertex_info old = brw->vb.info;
 178    GLuint i;
 179
 180    memset(inputs, 0, sizeof(*inputs));
 181    memset(&brw->vb.info, 0, sizeof(brw->vb.info));
 182
 183    for (i = 0; i < VERT_ATTRIB_MAX; i++) {
 184       brw->vb.inputs[i].glarray = arrays[i];
 185
 186       if (arrays[i]->StrideB != 0)
 187          brw->vb.info.varying |= 1 << i;
 188
 189       brw->vb.info.sizes[i/16] |= (inputs[i].glarray->Size - 1) << ((i%16) * 2);
 190    }
 191
 192    /* Raise statechanges if input sizes and varying have changed:
 193     */
 194    if (memcmp(brw->vb.info.sizes, old.sizes, sizeof(old.sizes)) != 0)
 195       brw->state.dirty.brw |= BRW_NEW_INPUT_DIMENSIONS;
 196
 197    if (brw->vb.info.varying != old.varying)
 198       brw->state.dirty.brw |= BRW_NEW_INPUT_VARYING;
 199 }
 200
 201 /* XXX: could split the primitive list to fallback only on the
 202  * non-conformant primitives.
 203  */
 204 static GLboolean check_fallbacks( struct brw_context *brw,
 205                                   const struct _mesa_prim *prim,
 206                                   GLuint nr_prims )
 207 {
 208    GLuint i;
 209
 210    if (!brw->intel.strict_conformance)
 211       return GL_FALSE;
 212
 213    if (brw->attribs.Polygon->SmoothFlag) {
 214       for (i = 0; i < nr_prims; i++)
 215          if (reduced_prim[prim[i].mode] == GL_TRIANGLES)
 216             return GL_TRUE;
 217    }
 218
 219    /* BRW hardware will do AA lines, but they are non-conformant it
 220     * seems.  TBD whether we keep this fallback:
 221     */
 222    if (brw->attribs.Line->SmoothFlag) {
 223       for (i = 0; i < nr_prims; i++)
 224          if (reduced_prim[prim[i].mode] == GL_LINES)
 225             return GL_TRUE;
 226    }
 227
 228    /* Stipple -- these fallbacks could be resolved with a little
 229     * bit of work?
 230     */
 231    if (brw->attribs.Line->StippleFlag) {
 232       for (i = 0; i < nr_prims; i++) {
 233          /* GS doesn't get enough information to know when to reset
 234           * the stipple counter?!?
 235           */
 236          if (prim[i].mode == GL_LINE_LOOP)
 237             return GL_TRUE;
 238
 239          if (prim[i].mode == GL_POLYGON &&
 240              (brw->attribs.Polygon->FrontMode == GL_LINE ||
 241               brw->attribs.Polygon->BackMode == GL_LINE))
 242             return GL_TRUE;
 243       }
 244    }
 245
 246
 247    if (brw->attribs.Point->SmoothFlag) {
 248       for (i = 0; i < nr_prims; i++)
 249          if (prim[i].mode == GL_POINTS)
 250             return GL_TRUE;
 251    }
 252
 253    return GL_FALSE;
 254 }
 255
 256 /* May fail if out of video memory for texture or vbo upload, or on
 257  * fallback conditions.
 258  */
 259 static GLboolean brw_try_draw_prims( GLcontext *ctx,
 260                                      const struct gl_client_array *arrays[],
 261                                      const struct _mesa_prim *prim,
 262                                      GLuint nr_prims,
 263                                      const struct _mesa_index_buffer *ib,
 264                                      GLuint min_index,
 265                                      GLuint max_index )
 266 {
 267    struct intel_context *intel = intel_context(ctx);
 268    struct brw_context *brw = brw_context(ctx);
 269    GLboolean retval = GL_FALSE;
 270    GLuint i, j;
 271
 272    if (ctx->NewState)
 273       _mesa_update_state( ctx );
 274
 275    /* Bind all inputs, derive varying and size information:
 276     */
 277    brw_merge_inputs( brw, arrays );
 278
 279    /* Have to validate state quite late.  Will rebuild tnl_program,
 280     * which depends on varying information.
 281     *
 282     * Note this is where brw->vs->prog_data.inputs_read is calculated,
 283     * so can't access it earlier.
 284     */
 285
 286    LOCK_HARDWARE(intel);
 287
 288    if (brw->intel.numClipRects == 0) {
 289       assert(intel->batch->ptr == intel->batch->map + intel->batch->offset);
 290       UNLOCK_HARDWARE(intel);
 291       return GL_TRUE;
 292    }
 293
 294    {
 295       /* Set the first primitive early, ahead of validate_state:
 296        */
 297       brw_set_prim(brw, prim[0].mode);
 298
 299       /* XXX:  Need to separate validate and upload of state.
 300        */
 301       brw_validate_state( brw );
 302
 303       /* Various fallback checks:
 304        */
 305       if (brw->intel.Fallback)
 306          goto out;
 307
 308       if (check_fallbacks( brw, prim, nr_prims ))
 309          goto out;
 310
 311       /* Upload index, vertex data:
 312        */
 313       if (ib)
 314          brw_upload_indices( brw, ib );
 315
 316       if (!brw_upload_vertices( brw, min_index, max_index)) {
 317          goto out;
 318       }
 319
 320       /* For single cliprect, state is already emitted:
 321        */
 322       if (brw->intel.numClipRects == 1) {
 323          for (i = 0; i < nr_prims; i++) {
 324             brw_emit_prim(brw, &prim[i]);
 325          }
 326       }
 327       else {
 328          /* Otherwise, explicitly do the cliprects at this point:
 329           */
 330          for (j = 0; j < brw->intel.numClipRects; j++) {
 331             brw_emit_cliprect(brw, &brw->intel.pClipRects[j]);
 332
 333             /* Emit prims to batchbuffer:
 334              */
 335             for (i = 0; i < nr_prims; i++) {
 336                brw_emit_prim(brw, &prim[i]);
 337             }
 338          }
 339       }
 340
 341       intel->need_flush = GL_TRUE;
 342       retval = GL_TRUE;
 343    }
 344
 345  out:
 346
 347    /* Currently have to do this to synchronize with the map/unmap of
 348     * the vertex buffer in brw_exec_api.c.  Not sure if there is any
 349     * way around this, as not every flush is due to a buffer filling
 350     * up.
 351     */
 352    if (!intel_batchbuffer_flush( brw->intel.batch )) {
 353       DBG("%s intel_batchbuffer_flush failed\n", __FUNCTION__);
 354       retval = GL_FALSE;
 355    }
 356
 357    if (retval && intel->thrashing) {
 358       bmSetFence(intel);
 359    }
 360
 361    /* Free any old data so it doesn't clog up texture memory - we
 362     * won't be referencing it again.
 363     */
 364    while (brw->vb.upload.wrap != brw->vb.upload.buf) {
 365       ctx->Driver.BufferData(ctx,
 366                              GL_ARRAY_BUFFER_ARB,
 367                              BRW_UPLOAD_INIT_SIZE,
 368                              NULL,
 369                              GL_DYNAMIC_DRAW_ARB,
 370                              brw->vb.upload.vbo[brw->vb.upload.wrap]);
 371       brw->vb.upload.wrap++;
 372       brw->vb.upload.wrap %= BRW_NR_UPLOAD_BUFS;
 373    }
 374
 375    UNLOCK_HARDWARE(intel);
 376
 377    if (!retval)
 378       DBG("%s failed\n", __FUNCTION__);
 379
 380    return retval;
 381 }
 382
 383 static GLboolean brw_need_rebase( GLcontext *ctx,
 384                                   const struct gl_client_array *arrays[],
 385                                   const struct _mesa_index_buffer *ib,
 386                                   GLuint min_index )
 387 {
 388    if (min_index == 0)
 389       return GL_FALSE;
 390
 391    if (ib) {
 392       if (!vbo_all_varyings_in_vbos(arrays))
 393          return GL_TRUE;
 394       else
 395          return GL_FALSE;
 396    }
 397    else {
 398       /* Hmm.  This isn't quite what I wanted.  BRW can actually
 399        * handle the mixed case well enough that we shouldn't need to
 400        * rebase.  However, it's probably not very common, nor hugely
 401        * expensive to do it this way:
 402        */
 403       if (!vbo_all_varyings_in_vbos(arrays))
 404          return GL_TRUE;
 405       else
 406          return GL_FALSE;
 407    }
 408 }
 409
 410
 411 void brw_draw_prims( GLcontext *ctx,
 412                      const struct gl_client_array *arrays[],
 413                      const struct _mesa_prim *prim,
 414                      GLuint nr_prims,
 415                      const struct _mesa_index_buffer *ib,
 416                      GLuint min_index,
 417                      GLuint max_index )
 418 {
 419    struct intel_context *intel = intel_context(ctx);
 420    GLboolean retval;
 421
 422    /* Decide if we want to rebase.  If so we end up recursing once
 423     * only into this function.
 424     */
 425    if (brw_need_rebase( ctx, arrays, ib, min_index )) {
 426       vbo_rebase_prims( ctx, arrays,
 427                         prim, nr_prims,
 428                         ib, min_index, max_index,
 429                         brw_draw_prims );
 430
 431       return;
 432    }
 433
 434
 435    /* Make a first attempt at drawing:
 436     */
 437    retval = brw_try_draw_prims(ctx, arrays, prim, nr_prims, ib, min_index, max_index);
 438
 439
 440    /* This looks like out-of-memory but potentially we have
 441     * situation where there is enough memory but it has become
 442     * fragmented.  Clear out all heaps and start from scratch by
 443     * faking a contended lock event:  (done elsewhere)
 444     */
 445    if (!retval && !intel->Fallback && bmError(intel)) {
 446       DBG("retrying\n");
 447       /* Then try a second time only to upload textures and draw the
 448        * primitives:
 449        */
 450       retval = brw_try_draw_prims(ctx, arrays, prim, nr_prims, ib, min_index, max_index);
 451    }
 452
 453    /* Otherwise, we really are out of memory.  Pass the drawing
 454     * command to the software tnl module and which will in turn call
 455     * swrast to do the drawing.
 456     */
 457    if (!retval) {
 458       _tnl_draw_prims(ctx, arrays, prim, nr_prims, ib, min_index, max_index);
 459    }
 460
 461    if (intel->aub_file && (INTEL_DEBUG & DEBUG_SYNC)) {
 462       intelFinish( &intel->ctx );
 463       intel->aub_wrap = 1;
 464    }
 465 }
 466
 467
 468 static void brw_invalidate_vbo_cb( struct intel_context *intel, void *ptr )
 469 {
 470    /* nothing to do, we don't rely on the contents being preserved */
 471 }
 472
 473
 474 void brw_draw_init( struct brw_context *brw )
 475 {
 476    GLcontext *ctx = &brw->intel.ctx;
 477    struct vbo_context *vbo = vbo_context(ctx);
 478    GLuint i;
 479
 480    /* Register our drawing function:
 481     */
 482    vbo->draw_prims = brw_draw_prims;
 483
 484    brw->vb.upload.size = BRW_UPLOAD_INIT_SIZE;
 485
 486    for (i = 0; i < BRW_NR_UPLOAD_BUFS; i++) {
 487       brw->vb.upload.vbo[i] = ctx->Driver.NewBufferObject(ctx, 1, GL_ARRAY_BUFFER_ARB);
 488
 489       /* NOTE:  These are set to no-backing-store.
 490        */
 491       bmBufferSetInvalidateCB(&brw->intel,
 492                               intel_bufferobj_buffer(intel_buffer_object(brw->vb.upload.vbo[i])),
 493                               brw_invalidate_vbo_cb,
 494                               &brw->intel,
 495                               GL_TRUE);
 496    }
 497
 498    ctx->Driver.BufferData( ctx,
 499                            GL_ARRAY_BUFFER_ARB,
 500                            BRW_UPLOAD_INIT_SIZE,
 501                            NULL,
 502                            GL_DYNAMIC_DRAW_ARB,
 503                            brw->vb.upload.vbo[0] );
 504 }
 505
 506 void brw_draw_destroy( struct brw_context *brw )
 507 {
 508    GLcontext *ctx = &brw->intel.ctx;
 509    GLuint i;
 510
 511    for (i = 0; i < BRW_NR_UPLOAD_BUFS; i++)
 512       ctx->Driver.DeleteBuffer(ctx, brw->vb.upload.vbo[i]);
 513 }