src/mesa/drivers/dri/intel/intel_regions.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2006 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 /* Provide additional functionality on top of bufmgr buffers:
  29  *   - 2d semantics and blit operations
  30  *   - refcounting of buffers for multiple images in a buffer.
  31  *   - refcounting of buffer mappings.
  32  *   - some logic for moving the buffers to the best memory pools for
  33  *     given operations.
  34  *
  35  * Most of this is to make it easier to implement the fixed-layout
  36  * mipmap tree required by intel hardware in the face of GL's
  37  * programming interface where each image can be specifed in random
  38  * order and it isn't clear what layout the tree should have until the
  39  * last moment.
  40  */
  41
  42 #include "intel_context.h"
  43 #include "intel_regions.h"
  44 #include "intel_blit.h"
  45 #include "intel_buffer_objects.h"
  46 #include "dri_bufmgr.h"
  47 #include "intel_bufmgr_ttm.h"
  48 #include "intel_batchbuffer.h"
  49
  50 #define FILE_DEBUG_FLAG DEBUG_REGION
  51
  52 void
  53 intel_region_idle(struct intel_context *intel, struct intel_region *region)
  54 {
  55    DBG("%s\n", __FUNCTION__);
  56    /* XXX: Using this function is likely bogus -- it ought to only have been
  57     * used before a map, anyway, but leave this cheap implementation of it
  58     * for now.
  59     */
  60    if (region && region->buffer) {
  61       /* Mapping it for read will ensure that any acceleration to the region
  62        * would have landed already.
  63        */
  64       dri_bo_map(region->buffer, GL_TRUE);
  65       dri_bo_unmap(region->buffer);
  66    }
  67 }
  68
  69 /* XXX: Thread safety?
  70  */
  71 GLubyte *
  72 intel_region_map(struct intel_context *intel, struct intel_region *region)
  73 {
  74    DBG("%s\n", __FUNCTION__);
  75    if (!region->map_refcount++) {
  76       if (region->pbo)
  77          intel_region_cow(intel, region);
  78
  79       dri_bo_map(region->buffer, GL_TRUE);
  80       region->map = region->buffer->virtual;
  81    }
  82
  83    return region->map;
  84 }
  85
  86 void
  87 intel_region_unmap(struct intel_context *intel, struct intel_region *region)
  88 {
  89    DBG("%s\n", __FUNCTION__);
  90    if (!--region->map_refcount) {
  91       dri_bo_unmap(region->buffer);
  92       region->map = NULL;
  93    }
  94 }
  95
  96 struct intel_region *
  97 intel_region_alloc(struct intel_context *intel,
  98                    GLuint cpp, GLuint pitch, GLuint height)
  99 {
 100    struct intel_region *region = calloc(sizeof(*region), 1);
 101
 102    DBG("%s\n", __FUNCTION__);
 103
 104    region->cpp = cpp;
 105    region->pitch = pitch;
 106    region->height = height;     /* needed? */
 107    region->refcount = 1;
 108
 109    region->buffer = dri_bo_alloc(intel->bufmgr, "region",
 110                                  pitch * cpp * height, 64, DRM_BO_FLAG_MEM_TT);
 111    return region;
 112 }
 113
 114 void
 115 intel_region_reference(struct intel_region **dst, struct intel_region *src)
 116 {
 117    assert(*dst == NULL);
 118    if (src) {
 119       src->refcount++;
 120       *dst = src;
 121    }
 122 }
 123
 124 void
 125 intel_region_release(struct intel_region **region)
 126 {
 127    if (!*region)
 128       return;
 129
 130    DBG("%s %d\n", __FUNCTION__, (*region)->refcount - 1);
 131
 132    ASSERT((*region)->refcount > 0);
 133    (*region)->refcount--;
 134
 135    if ((*region)->refcount == 0) {
 136       assert((*region)->map_refcount == 0);
 137
 138       if ((*region)->pbo)
 139          (*region)->pbo->region = NULL;
 140       (*region)->pbo = NULL;
 141       dri_bo_unreference((*region)->buffer);
 142       free(*region);
 143    }
 144    *region = NULL;
 145 }
 146
 147 /*
 148  * XXX Move this into core Mesa?
 149  */
 150 static void
 151 _mesa_copy_rect(GLubyte * dst,
 152                 GLuint cpp,
 153                 GLuint dst_pitch,
 154                 GLuint dst_x,
 155                 GLuint dst_y,
 156                 GLuint width,
 157                 GLuint height,
 158                 const GLubyte * src,
 159                 GLuint src_pitch, GLuint src_x, GLuint src_y)
 160 {
 161    GLuint i;
 162
 163    dst_pitch *= cpp;
 164    src_pitch *= cpp;
 165    dst += dst_x * cpp;
 166    src += src_x * cpp;
 167    dst += dst_y * dst_pitch;
 168    src += src_y * dst_pitch;
 169    width *= cpp;
 170
 171    if (width == dst_pitch && width == src_pitch)
 172       memcpy(dst, src, height * width);
 173    else {
 174       for (i = 0; i < height; i++) {
 175          memcpy(dst, src, width);
 176          dst += dst_pitch;
 177          src += src_pitch;
 178       }
 179    }
 180 }
 181
 182
 183 /* Upload data to a rectangular sub-region.  Lots of choices how to do this:
 184  *
 185  * - memcpy by span to current destination
 186  * - upload data as new buffer and blit
 187  *
 188  * Currently always memcpy.
 189  */
 190 void
 191 intel_region_data(struct intel_context *intel,
 192                   struct intel_region *dst,
 193                   GLuint dst_offset,
 194                   GLuint dstx, GLuint dsty,
 195                   const void *src, GLuint src_pitch,
 196                   GLuint srcx, GLuint srcy, GLuint width, GLuint height)
 197 {
 198    DBG("%s\n", __FUNCTION__);
 199
 200    if (intel == NULL)
 201       return;
 202
 203    if (dst->pbo) {
 204       if (dstx == 0 &&
 205           dsty == 0 && width == dst->pitch && height == dst->height)
 206          intel_region_release_pbo(intel, dst);
 207       else
 208          intel_region_cow(intel, dst);
 209    }
 210
 211
 212    LOCK_HARDWARE(intel);
 213
 214    _mesa_copy_rect(intel_region_map(intel, dst) + dst_offset,
 215                    dst->cpp,
 216                    dst->pitch,
 217                    dstx, dsty, width, height, src, src_pitch, srcx, srcy);
 218
 219    intel_region_unmap(intel, dst);
 220
 221    UNLOCK_HARDWARE(intel);
 222
 223 }
 224
 225 /* Copy rectangular sub-regions. Need better logic about when to
 226  * push buffers into AGP - will currently do so whenever possible.
 227  */
 228 void
 229 intel_region_copy(struct intel_context *intel,
 230                   struct intel_region *dst,
 231                   GLuint dst_offset,
 232                   GLuint dstx, GLuint dsty,
 233                   struct intel_region *src,
 234                   GLuint src_offset,
 235                   GLuint srcx, GLuint srcy, GLuint width, GLuint height)
 236 {
 237    DBG("%s\n", __FUNCTION__);
 238
 239    if (intel == NULL)
 240       return;
 241
 242    if (dst->pbo) {
 243       if (dstx == 0 &&
 244           dsty == 0 && width == dst->pitch && height == dst->height)
 245          intel_region_release_pbo(intel, dst);
 246       else
 247          intel_region_cow(intel, dst);
 248    }
 249
 250    assert(src->cpp == dst->cpp);
 251
 252    intelEmitCopyBlit(intel,
 253                      dst->cpp,
 254                      src->pitch, src->buffer, src_offset, src->tiled,
 255                      dst->pitch, dst->buffer, dst_offset, dst->tiled,
 256                      srcx, srcy, dstx, dsty, width, height,
 257                      GL_COPY);
 258 }
 259
 260 /* Fill a rectangular sub-region.  Need better logic about when to
 261  * push buffers into AGP - will currently do so whenever possible.
 262  */
 263 void
 264 intel_region_fill(struct intel_context *intel,
 265                   struct intel_region *dst,
 266                   GLuint dst_offset,
 267                   GLuint dstx, GLuint dsty,
 268                   GLuint width, GLuint height, GLuint color)
 269 {
 270    DBG("%s\n", __FUNCTION__);
 271
 272    if (intel == NULL)
 273       return;
 274
 275    if (dst->pbo) {
 276       if (dstx == 0 &&
 277           dsty == 0 && width == dst->pitch && height == dst->height)
 278          intel_region_release_pbo(intel, dst);
 279       else
 280          intel_region_cow(intel, dst);
 281    }
 282
 283    intelEmitFillBlit(intel,
 284                      dst->cpp,
 285                      dst->pitch, dst->buffer, dst_offset, dst->tiled,
 286                      dstx, dsty, width, height, color);
 287 }
 288
 289 /* Attach to a pbo, discarding our data.  Effectively zero-copy upload
 290  * the pbo's data.
 291  */
 292 void
 293 intel_region_attach_pbo(struct intel_context *intel,
 294                         struct intel_region *region,
 295                         struct intel_buffer_object *pbo)
 296 {
 297    if (region->pbo == pbo)
 298       return;
 299
 300    /* If there is already a pbo attached, break the cow tie now.
 301     * Don't call intel_region_release_pbo() as that would
 302     * unnecessarily allocate a new buffer we would have to immediately
 303     * discard.
 304     */
 305    if (region->pbo) {
 306       region->pbo->region = NULL;
 307       region->pbo = NULL;
 308    }
 309
 310    if (region->buffer) {
 311       dri_bo_unreference(region->buffer);
 312       region->buffer = NULL;
 313    }
 314
 315    region->pbo = pbo;
 316    region->pbo->region = region;
 317    dri_bo_reference(pbo->buffer);
 318    region->buffer = pbo->buffer;
 319 }
 320
 321
 322 /* Break the COW tie to the pbo and allocate a new buffer.
 323  * The pbo gets to keep the data.
 324  */
 325 void
 326 intel_region_release_pbo(struct intel_context *intel,
 327                          struct intel_region *region)
 328 {
 329    assert(region->buffer == region->pbo->buffer);
 330    region->pbo->region = NULL;
 331    region->pbo = NULL;
 332    dri_bo_unreference(region->buffer);
 333    region->buffer = NULL;
 334
 335    region->buffer = dri_bo_alloc(intel->bufmgr, "region",
 336                                  region->pitch * region->cpp * region->height,
 337                                  64, DRM_BO_FLAG_MEM_TT);
 338 }
 339
 340 /* Break the COW tie to the pbo.  Both the pbo and the region end up
 341  * with a copy of the data.
 342  */
 343 void
 344 intel_region_cow(struct intel_context *intel, struct intel_region *region)
 345 {
 346    struct intel_buffer_object *pbo = region->pbo;
 347    GLboolean was_locked = intel->locked;
 348
 349    if (intel == NULL)
 350       return;
 351
 352    intel_region_release_pbo(intel, region);
 353
 354    assert(region->cpp * region->pitch * region->height == pbo->Base.Size);
 355
 356    DBG("%s (%d bytes)\n", __FUNCTION__, pbo->Base.Size);
 357
 358    /* Now blit from the texture buffer to the new buffer:
 359     */
 360
 361    intel_batchbuffer_flush(intel->batch);
 362
 363    was_locked = intel->locked;
 364    if (intel->locked)
 365       LOCK_HARDWARE(intel);
 366
 367    intelEmitCopyBlit(intel,
 368                      region->cpp,
 369                      region->pitch, region->buffer, 0, region->tiled,
 370                      region->pitch, pbo->buffer, 0, region->tiled,
 371                      0, 0, 0, 0,
 372                      region->pitch, region->height,
 373                      GL_COPY);
 374
 375    intel_batchbuffer_flush(intel->batch);
 376
 377    if (was_locked)
 378       UNLOCK_HARDWARE(intel);
 379 }
 380
 381 dri_bo *
 382 intel_region_buffer(struct intel_context *intel,
 383                     struct intel_region *region, GLuint flag)
 384 {
 385    if (region->pbo) {
 386       if (flag == INTEL_WRITE_PART)
 387          intel_region_cow(intel, region);
 388       else if (flag == INTEL_WRITE_FULL)
 389          intel_region_release_pbo(intel, region);
 390    }
 391
 392    return region->buffer;
 393 }
 394
 395 static struct intel_region *
 396 intel_recreate_static(struct intel_context *intel,
 397                       const char *name,
 398                       struct intel_region *region,
 399                       intelRegion *region_desc,
 400                       GLuint mem_type)
 401 {
 402    intelScreenPrivate *intelScreen = intel->intelScreen;
 403
 404    if (region == NULL) {
 405       region = calloc(sizeof(*region), 1);
 406       region->refcount = 1;
 407    }
 408
 409    region->cpp = intelScreen->cpp;
 410    region->pitch = region_desc->pitch / intelScreen->cpp;
 411    region->height = intelScreen->height;     /* needed? */
 412    region->tiled = region_desc->tiled;
 413
 414    if (intel->ttm) {
 415       assert(region_desc->bo_handle != -1);
 416       region->buffer = intel_ttm_bo_create_from_handle(intel->bufmgr,
 417                                                        name,
 418                                                        region_desc->bo_handle);
 419    } else {
 420       region->buffer = dri_bo_alloc_static(intel->bufmgr,
 421                                            name,
 422                                            region_desc->offset,
 423                                            region_desc->pitch *
 424                                            intelScreen->height,
 425                                            region_desc->map,
 426                                            DRM_BO_FLAG_MEM_TT);
 427    }
 428
 429    assert(region->buffer != NULL);
 430
 431    return region;
 432 }
 433
 434 /**
 435  * Create intel_region structs to describe the static front, back, and depth
 436  * buffers created by the xserver.
 437  *
 438  * Although FBO's mean we now no longer use these as render targets in
 439  * all circumstances, they won't go away until the back and depth
 440  * buffers become private, and the front buffer will remain even then.
 441  *
 442  * Note that these don't allocate video memory, just describe
 443  * allocations alread made by the X server.
 444  */
 445 void
 446 intel_recreate_static_regions(struct intel_context *intel)
 447 {
 448    intelScreenPrivate *intelScreen = intel->intelScreen;
 449
 450    intel->front_region =
 451       intel_recreate_static(intel, "front",
 452                             intel->front_region,
 453                             &intelScreen->front,
 454                             DRM_BO_FLAG_MEM_TT);
 455
 456    intel->back_region =
 457       intel_recreate_static(intel, "back",
 458                             intel->back_region,
 459                             &intelScreen->back,
 460                             DRM_BO_FLAG_MEM_TT);
 461
 462    if (intelScreen->third.handle) {
 463       intel->third_region =
 464          intel_recreate_static(intel, "third",
 465                                intel->third_region,
 466                                &intelScreen->third,
 467                                DRM_BO_FLAG_MEM_TT);
 468    }
 469
 470    /* Still assumes front.cpp == depth.cpp.  We can kill this when we move to
 471     * private buffers.
 472     */
 473    intel->depth_region =
 474       intel_recreate_static(intel, "depth",
 475                             intel->depth_region,
 476                             &intelScreen->depth,
 477                             DRM_BO_FLAG_MEM_TT);
 478 }