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       /* XXX: metaops passes null arrays */
 187       if (arrays[i]) {
 188          if (arrays[i]->StrideB != 0)
 189             brw->vb.info.varying |= 1 << i;
 190
 191          brw->vb.info.sizes[i/16] |= (inputs[i].glarray->Size - 1) << ((i%16) * 2);
 192       }
 193    }
 194
 195    /* Raise statechanges if input sizes and varying have changed:
 196     */
 197    if (memcmp(brw->vb.info.sizes, old.sizes, sizeof(old.sizes)) != 0)
 198       brw->state.dirty.brw |= BRW_NEW_INPUT_DIMENSIONS;
 199
 200    if (brw->vb.info.varying != old.varying)
 201       brw->state.dirty.brw |= BRW_NEW_INPUT_VARYING;
 202 }
 203
 204 /* XXX: could split the primitive list to fallback only on the
 205  * non-conformant primitives.
 206  */
 207 static GLboolean check_fallbacks( struct brw_context *brw,
 208                                   const struct _mesa_prim *prim,
 209                                   GLuint nr_prims )
 210 {
 211    GLuint i;
 212
 213    if (!brw->intel.strict_conformance)
 214       return GL_FALSE;
 215
 216    if (brw->attribs.Polygon->SmoothFlag) {
 217       for (i = 0; i < nr_prims; i++)
 218          if (reduced_prim[prim[i].mode] == GL_TRIANGLES)
 219             return GL_TRUE;
 220    }
 221
 222    /* BRW hardware will do AA lines, but they are non-conformant it
 223     * seems.  TBD whether we keep this fallback:
 224     */
 225    if (brw->attribs.Line->SmoothFlag) {
 226       for (i = 0; i < nr_prims; i++)
 227          if (reduced_prim[prim[i].mode] == GL_LINES)
 228             return GL_TRUE;
 229    }
 230
 231    /* Stipple -- these fallbacks could be resolved with a little
 232     * bit of work?
 233     */
 234    if (brw->attribs.Line->StippleFlag) {
 235       for (i = 0; i < nr_prims; i++) {
 236          /* GS doesn't get enough information to know when to reset
 237           * the stipple counter?!?
 238           */
 239          if (prim[i].mode == GL_LINE_LOOP)
 240             return GL_TRUE;
 241
 242          if (prim[i].mode == GL_POLYGON &&
 243              (brw->attribs.Polygon->FrontMode == GL_LINE ||
 244               brw->attribs.Polygon->BackMode == GL_LINE))
 245             return GL_TRUE;
 246       }
 247    }
 248
 249
 250    if (brw->attribs.Point->SmoothFlag) {
 251       for (i = 0; i < nr_prims; i++)
 252          if (prim[i].mode == GL_POINTS)
 253             return GL_TRUE;
 254    }
 255
 256    return GL_FALSE;
 257 }
 258
 259 /* May fail if out of video memory for texture or vbo upload, or on
 260  * fallback conditions.
 261  */
 262 static GLboolean brw_try_draw_prims( GLcontext *ctx,
 263                                      const struct gl_client_array *arrays[],
 264                                      const struct _mesa_prim *prim,
 265                                      GLuint nr_prims,
 266                                      const struct _mesa_index_buffer *ib,
 267                                      GLuint min_index,
 268                                      GLuint max_index )
 269 {
 270    struct intel_context *intel = intel_context(ctx);
 271    struct brw_context *brw = brw_context(ctx);
 272    GLboolean retval = GL_FALSE;
 273    GLuint i, j;
 274
 275    if (ctx->NewState)
 276       _mesa_update_state( ctx );
 277
 278    /* Bind all inputs, derive varying and size information:
 279     */
 280    brw_merge_inputs( brw, arrays );
 281
 282    /* Have to validate state quite late.  Will rebuild tnl_program,
 283     * which depends on varying information.
 284     *
 285     * Note this is where brw->vs->prog_data.inputs_read is calculated,
 286     * so can't access it earlier.
 287     */
 288
 289    LOCK_HARDWARE(intel);
 290
 291    if (brw->intel.numClipRects == 0) {
 292       assert(intel->batch->ptr == intel->batch->map + intel->batch->offset);
 293       UNLOCK_HARDWARE(intel);
 294       return GL_TRUE;
 295    }
 296
 297    {
 298       /* Set the first primitive early, ahead of validate_state:
 299        */
 300       brw_set_prim(brw, prim[0].mode);
 301
 302       /* XXX:  Need to separate validate and upload of state.
 303        */
 304       brw_validate_state( brw );
 305
 306       /* Various fallback checks:
 307        */
 308       if (brw->intel.Fallback)
 309          goto out;
 310
 311       if (check_fallbacks( brw, prim, nr_prims ))
 312          goto out;
 313
 314       /* Upload index, vertex data:
 315        */
 316       if (ib)
 317          brw_upload_indices( brw, ib );
 318
 319       if (!brw_upload_vertices( brw, min_index, max_index)) {
 320          goto out;
 321       }
 322
 323       /* For single cliprect, state is already emitted:
 324        */
 325       if (brw->intel.numClipRects == 1) {
 326          for (i = 0; i < nr_prims; i++) {
 327             brw_emit_prim(brw, &prim[i]);
 328          }
 329       }
 330       else {
 331          /* Otherwise, explicitly do the cliprects at this point:
 332           */
 333          for (j = 0; j < brw->intel.numClipRects; j++) {
 334             brw_emit_cliprect(brw, &brw->intel.pClipRects[j]);
 335
 336             /* Emit prims to batchbuffer:
 337              */
 338             for (i = 0; i < nr_prims; i++) {
 339                brw_emit_prim(brw, &prim[i]);
 340             }
 341          }
 342       }
 343
 344       intel->need_flush = GL_TRUE;
 345       retval = GL_TRUE;
 346    }
 347
 348  out:
 349
 350    /* Currently have to do this to synchronize with the map/unmap of
 351     * the vertex buffer in brw_exec_api.c.  Not sure if there is any
 352     * way around this, as not every flush is due to a buffer filling
 353     * up.
 354     */
 355    if (!intel_batchbuffer_flush( brw->intel.batch )) {
 356       DBG("%s intel_batchbuffer_flush failed\n", __FUNCTION__);
 357       retval = GL_FALSE;
 358    }
 359
 360    if (retval && intel->thrashing) {
 361       bmSetFence(intel);
 362    }
 363
 364    /* Free any old data so it doesn't clog up texture memory - we
 365     * won't be referencing it again.
 366     */
 367    while (brw->vb.upload.wrap != brw->vb.upload.buf) {
 368       ctx->Driver.BufferData(ctx,
 369                              GL_ARRAY_BUFFER_ARB,
 370                              BRW_UPLOAD_INIT_SIZE,
 371                              NULL,
 372                              GL_DYNAMIC_DRAW_ARB,
 373                              brw->vb.upload.vbo[brw->vb.upload.wrap]);
 374       brw->vb.upload.wrap++;
 375       brw->vb.upload.wrap %= BRW_NR_UPLOAD_BUFS;
 376    }
 377
 378    UNLOCK_HARDWARE(intel);
 379
 380    if (!retval)
 381       DBG("%s failed\n", __FUNCTION__);
 382
 383    return retval;
 384 }
 385
 386 static GLboolean brw_need_rebase( GLcontext *ctx,
 387                                   const struct gl_client_array *arrays[],
 388                                   const struct _mesa_index_buffer *ib,
 389                                   GLuint min_index )
 390 {
 391    if (min_index == 0)
 392       return GL_FALSE;
 393
 394    if (ib) {
 395       if (!vbo_all_varyings_in_vbos(arrays))
 396          return GL_TRUE;
 397       else
 398          return GL_FALSE;
 399    }
 400    else {
 401       /* Hmm.  This isn't quite what I wanted.  BRW can actually
 402        * handle the mixed case well enough that we shouldn't need to
 403        * rebase.  However, it's probably not very common, nor hugely
 404        * expensive to do it this way:
 405        */
 406       if (!vbo_all_varyings_in_vbos(arrays))
 407          return GL_TRUE;
 408       else
 409          return GL_FALSE;
 410    }
 411 }
 412
 413
 414 void brw_draw_prims( GLcontext *ctx,
 415                      const struct gl_client_array *arrays[],
 416                      const struct _mesa_prim *prim,
 417                      GLuint nr_prims,
 418                      const struct _mesa_index_buffer *ib,
 419                      GLuint min_index,
 420                      GLuint max_index )
 421 {
 422    struct intel_context *intel = intel_context(ctx);
 423    GLboolean retval;
 424
 425    /* Decide if we want to rebase.  If so we end up recursing once
 426     * only into this function.
 427     */
 428    if (brw_need_rebase( ctx, arrays, ib, min_index )) {
 429       vbo_rebase_prims( ctx, arrays,
 430                         prim, nr_prims,
 431                         ib, min_index, max_index,
 432                         brw_draw_prims );
 433
 434       return;
 435    }
 436
 437
 438    /* Make a first attempt at drawing:
 439     */
 440    retval = brw_try_draw_prims(ctx, arrays, prim, nr_prims, ib, min_index, max_index);
 441
 442
 443    /* This looks like out-of-memory but potentially we have
 444     * situation where there is enough memory but it has become
 445     * fragmented.  Clear out all heaps and start from scratch by
 446     * faking a contended lock event:  (done elsewhere)
 447     */
 448    if (!retval && !intel->Fallback && bmError(intel)) {
 449       DBG("retrying\n");
 450       /* Then try a second time only to upload textures and draw the
 451        * primitives:
 452        */
 453       retval = brw_try_draw_prims(ctx, arrays, prim, nr_prims, ib, min_index, max_index);
 454    }
 455
 456    /* Otherwise, we really are out of memory.  Pass the drawing
 457     * command to the software tnl module and which will in turn call
 458     * swrast to do the drawing.
 459     */
 460    if (!retval) {
 461       _tnl_draw_prims(ctx, arrays, prim, nr_prims, ib, min_index, max_index);
 462    }
 463
 464    if (intel->aub_file && (INTEL_DEBUG & DEBUG_SYNC)) {
 465       intelFinish( &intel->ctx );
 466       intel->aub_wrap = 1;
 467    }
 468 }
 469
 470
 471 static void brw_invalidate_vbo_cb( struct intel_context *intel, void *ptr )
 472 {
 473    /* nothing to do, we don't rely on the contents being preserved */
 474 }
 475
 476
 477 void brw_draw_init( struct brw_context *brw )
 478 {
 479    GLcontext *ctx = &brw->intel.ctx;
 480    struct vbo_context *vbo = vbo_context(ctx);
 481    GLuint i;
 482
 483    /* Register our drawing function:
 484     */
 485    vbo->draw_prims = brw_draw_prims;
 486
 487    brw->vb.upload.size = BRW_UPLOAD_INIT_SIZE;
 488
 489    for (i = 0; i < BRW_NR_UPLOAD_BUFS; i++) {
 490       brw->vb.upload.vbo[i] = ctx->Driver.NewBufferObject(ctx, 1, GL_ARRAY_BUFFER_ARB);
 491
 492       /* NOTE:  These are set to no-backing-store.
 493        */
 494       bmBufferSetInvalidateCB(&brw->intel,
 495                               intel_bufferobj_buffer(intel_buffer_object(brw->vb.upload.vbo[i])),
 496                               brw_invalidate_vbo_cb,
 497                               &brw->intel,
 498                               GL_TRUE);
 499    }
 500
 501    ctx->Driver.BufferData( ctx,
 502                            GL_ARRAY_BUFFER_ARB,
 503                            BRW_UPLOAD_INIT_SIZE,
 504                            NULL,
 505                            GL_DYNAMIC_DRAW_ARB,
 506                            brw->vb.upload.vbo[0] );
 507 }
 508
 509 void brw_draw_destroy( struct brw_context *brw )
 510 {
 511    GLcontext *ctx = &brw->intel.ctx;
 512    GLuint i;
 513
 514    for (i = 0; i < BRW_NR_UPLOAD_BUFS; i++)
 515       ctx->Driver.DeleteBuffer(ctx, brw->vb.upload.vbo[i]);
 516 }