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 /* XXX: Thread safety?
  53  */
  54 GLubyte *
  55 intel_region_map(struct intel_context *intel, struct intel_region *region)
  56 {
  57    DBG("%s\n", __FUNCTION__);
  58    if (!region->map_refcount++) {
  59       if (region->pbo)
  60          intel_region_cow(intel, region);
  61
  62       dri_bo_map(region->buffer, GL_TRUE);
  63       region->map = region->buffer->virtual;
  64    }
  65
  66    return region->map;
  67 }
  68
  69 void
  70 intel_region_unmap(struct intel_context *intel, struct intel_region *region)
  71 {
  72    DBG("%s\n", __FUNCTION__);
  73    if (!--region->map_refcount) {
  74       dri_bo_unmap(region->buffer);
  75       region->map = NULL;
  76    }
  77 }
  78
  79 struct intel_region *
  80 intel_region_alloc(struct intel_context *intel,
  81                    GLuint cpp, GLuint pitch, GLuint height)
  82 {
  83    struct intel_region *region = calloc(sizeof(*region), 1);
  84
  85    DBG("%s\n", __FUNCTION__);
  86
  87    region->cpp = cpp;
  88    region->pitch = pitch;
  89    region->height = height;     /* needed? */
  90    region->refcount = 1;
  91
  92    region->buffer = dri_bo_alloc(intel->bufmgr, "region",
  93                                  pitch * cpp * height, 64,
  94                                  DRM_BO_FLAG_MEM_LOCAL |
  95                                  DRM_BO_FLAG_CACHED |
  96                                  DRM_BO_FLAG_CACHED_MAPPED);
  97    return region;
  98 }
  99
 100 void
 101 intel_region_reference(struct intel_region **dst, struct intel_region *src)
 102 {
 103    assert(*dst == NULL);
 104    if (src) {
 105       src->refcount++;
 106       *dst = src;
 107    }
 108 }
 109
 110 void
 111 intel_region_release(struct intel_region **region)
 112 {
 113    if (!*region)
 114       return;
 115
 116    DBG("%s %d\n", __FUNCTION__, (*region)->refcount - 1);
 117
 118    ASSERT((*region)->refcount > 0);
 119    (*region)->refcount--;
 120
 121    if ((*region)->refcount == 0) {
 122       assert((*region)->map_refcount == 0);
 123
 124       if ((*region)->pbo)
 125          (*region)->pbo->region = NULL;
 126       (*region)->pbo = NULL;
 127       dri_bo_unreference((*region)->buffer);
 128       free(*region);
 129    }
 130    *region = NULL;
 131 }
 132
 133 /*
 134  * XXX Move this into core Mesa?
 135  */
 136 void
 137 _mesa_copy_rect(GLubyte * dst,
 138                 GLuint cpp,
 139                 GLuint dst_pitch,
 140                 GLuint dst_x,
 141                 GLuint dst_y,
 142                 GLuint width,
 143                 GLuint height,
 144                 const GLubyte * src,
 145                 GLuint src_pitch, GLuint src_x, GLuint src_y)
 146 {
 147    GLuint i;
 148
 149    dst_pitch *= cpp;
 150    src_pitch *= cpp;
 151    dst += dst_x * cpp;
 152    src += src_x * cpp;
 153    dst += dst_y * dst_pitch;
 154    src += src_y * dst_pitch;
 155    width *= cpp;
 156
 157    if (width == dst_pitch && width == src_pitch)
 158       memcpy(dst, src, height * width);
 159    else {
 160       for (i = 0; i < height; i++) {
 161          memcpy(dst, src, width);
 162          dst += dst_pitch;
 163          src += src_pitch;
 164       }
 165    }
 166 }
 167
 168
 169 /* Upload data to a rectangular sub-region.  Lots of choices how to do this:
 170  *
 171  * - memcpy by span to current destination
 172  * - upload data as new buffer and blit
 173  *
 174  * Currently always memcpy.
 175  */
 176 void
 177 intel_region_data(struct intel_context *intel,
 178                   struct intel_region *dst,
 179                   GLuint dst_offset,
 180                   GLuint dstx, GLuint dsty,
 181                   const void *src, GLuint src_pitch,
 182                   GLuint srcx, GLuint srcy, GLuint width, GLuint height)
 183 {
 184    GLboolean locked = GL_FALSE;
 185
 186    DBG("%s\n", __FUNCTION__);
 187
 188    if (intel == NULL)
 189       return;
 190
 191    if (dst->pbo) {
 192       if (dstx == 0 &&
 193           dsty == 0 && width == dst->pitch && height == dst->height)
 194          intel_region_release_pbo(intel, dst);
 195       else
 196          intel_region_cow(intel, dst);
 197    }
 198
 199    if (!intel->locked) {
 200       LOCK_HARDWARE(intel);
 201       locked = GL_TRUE;
 202    }
 203
 204    _mesa_copy_rect(intel_region_map(intel, dst) + dst_offset,
 205                    dst->cpp,
 206                    dst->pitch,
 207                    dstx, dsty, width, height, src, src_pitch, srcx, srcy);
 208
 209    intel_region_unmap(intel, dst);
 210
 211    if (locked)
 212       UNLOCK_HARDWARE(intel);
 213
 214 }
 215
 216 /* Copy rectangular sub-regions. Need better logic about when to
 217  * push buffers into AGP - will currently do so whenever possible.
 218  */
 219 void
 220 intel_region_copy(struct intel_context *intel,
 221                   struct intel_region *dst,
 222                   GLuint dst_offset,
 223                   GLuint dstx, GLuint dsty,
 224                   struct intel_region *src,
 225                   GLuint src_offset,
 226                   GLuint srcx, GLuint srcy, GLuint width, GLuint height)
 227 {
 228    DBG("%s\n", __FUNCTION__);
 229
 230    if (intel == NULL)
 231       return;
 232
 233    if (dst->pbo) {
 234       if (dstx == 0 &&
 235           dsty == 0 && width == dst->pitch && height == dst->height)
 236          intel_region_release_pbo(intel, dst);
 237       else
 238          intel_region_cow(intel, dst);
 239    }
 240
 241    assert(src->cpp == dst->cpp);
 242
 243    intelEmitCopyBlit(intel,
 244                      dst->cpp,
 245                      src->pitch, src->buffer, src_offset, src->tiled,
 246                      dst->pitch, dst->buffer, dst_offset, dst->tiled,
 247                      srcx, srcy, dstx, dsty, width, height,
 248                      GL_COPY);
 249 }
 250
 251 /* Fill a rectangular sub-region.  Need better logic about when to
 252  * push buffers into AGP - will currently do so whenever possible.
 253  */
 254 void
 255 intel_region_fill(struct intel_context *intel,
 256                   struct intel_region *dst,
 257                   GLuint dst_offset,
 258                   GLuint dstx, GLuint dsty,
 259                   GLuint width, GLuint height, GLuint color)
 260 {
 261    DBG("%s\n", __FUNCTION__);
 262
 263    if (intel == NULL)
 264       return;
 265
 266    if (dst->pbo) {
 267       if (dstx == 0 &&
 268           dsty == 0 && width == dst->pitch && height == dst->height)
 269          intel_region_release_pbo(intel, dst);
 270       else
 271          intel_region_cow(intel, dst);
 272    }
 273
 274    intelEmitFillBlit(intel,
 275                      dst->cpp,
 276                      dst->pitch, dst->buffer, dst_offset, dst->tiled,
 277                      dstx, dsty, width, height, color);
 278 }
 279
 280 /* Attach to a pbo, discarding our data.  Effectively zero-copy upload
 281  * the pbo's data.
 282  */
 283 void
 284 intel_region_attach_pbo(struct intel_context *intel,
 285                         struct intel_region *region,
 286                         struct intel_buffer_object *pbo)
 287 {
 288    if (region->pbo == pbo)
 289       return;
 290
 291    /* If there is already a pbo attached, break the cow tie now.
 292     * Don't call intel_region_release_pbo() as that would
 293     * unnecessarily allocate a new buffer we would have to immediately
 294     * discard.
 295     */
 296    if (region->pbo) {
 297       region->pbo->region = NULL;
 298       region->pbo = NULL;
 299    }
 300
 301    if (region->buffer) {
 302       dri_bo_unreference(region->buffer);
 303       region->buffer = NULL;
 304    }
 305
 306    region->pbo = pbo;
 307    region->pbo->region = region;
 308    dri_bo_reference(pbo->buffer);
 309    region->buffer = pbo->buffer;
 310 }
 311
 312
 313 /* Break the COW tie to the pbo and allocate a new buffer.
 314  * The pbo gets to keep the data.
 315  */
 316 void
 317 intel_region_release_pbo(struct intel_context *intel,
 318                          struct intel_region *region)
 319 {
 320    assert(region->buffer == region->pbo->buffer);
 321    region->pbo->region = NULL;
 322    region->pbo = NULL;
 323    dri_bo_unreference(region->buffer);
 324    region->buffer = NULL;
 325
 326    region->buffer = dri_bo_alloc(intel->bufmgr, "region",
 327                                  region->pitch * region->cpp * region->height,
 328                                  64,
 329                                  DRM_BO_FLAG_MEM_LOCAL |
 330                                  DRM_BO_FLAG_CACHED |
 331                                  DRM_BO_FLAG_CACHED_MAPPED);
 332 }
 333
 334 /* Break the COW tie to the pbo.  Both the pbo and the region end up
 335  * with a copy of the data.
 336  */
 337 void
 338 intel_region_cow(struct intel_context *intel, struct intel_region *region)
 339 {
 340    struct intel_buffer_object *pbo = region->pbo;
 341    GLboolean was_locked = intel->locked;
 342
 343    if (intel == NULL)
 344       return;
 345
 346    intel_region_release_pbo(intel, region);
 347
 348    assert(region->cpp * region->pitch * region->height == pbo->Base.Size);
 349
 350    DBG("%s (%d bytes)\n", __FUNCTION__, pbo->Base.Size);
 351
 352    /* Now blit from the texture buffer to the new buffer:
 353     */
 354
 355    intel_batchbuffer_flush(intel->batch);
 356
 357    was_locked = intel->locked;
 358    if (intel->locked)
 359       LOCK_HARDWARE(intel);
 360
 361    intelEmitCopyBlit(intel,
 362                      region->cpp,
 363                      region->pitch, region->buffer, 0, region->tiled,
 364                      region->pitch, pbo->buffer, 0, region->tiled,
 365                      0, 0, 0, 0,
 366                      region->pitch, region->height,
 367                      GL_COPY);
 368
 369    intel_batchbuffer_flush(intel->batch);
 370
 371    if (was_locked)
 372       UNLOCK_HARDWARE(intel);
 373 }
 374
 375 dri_bo *
 376 intel_region_buffer(struct intel_context *intel,
 377                     struct intel_region *region, GLuint flag)
 378 {
 379    if (region->pbo) {
 380       if (flag == INTEL_WRITE_PART)
 381          intel_region_cow(intel, region);
 382       else if (flag == INTEL_WRITE_FULL)
 383          intel_region_release_pbo(intel, region);
 384    }
 385
 386    return region->buffer;
 387 }
 388
 389 static struct intel_region *
 390 intel_recreate_static(struct intel_context *intel,
 391                       const char *name,
 392                       struct intel_region *region,
 393                       intelRegion *region_desc,
 394                       GLuint mem_type)
 395 {
 396    intelScreenPrivate *intelScreen = intel->intelScreen;
 397
 398    if (region == NULL) {
 399       region = calloc(sizeof(*region), 1);
 400       region->refcount = 1;
 401    }
 402
 403    region->cpp = intel->ctx.Visual.rgbBits / 8;
 404    region->pitch = intelScreen->pitch;
 405    region->height = intelScreen->height;     /* needed? */
 406    region->tiled = region_desc->tiled;
 407
 408    if (intel->ttm) {
 409       assert(region_desc->bo_handle != -1);
 410       region->buffer = intel_ttm_bo_create_from_handle(intel->bufmgr,
 411                                                        name,
 412                                                        region_desc->bo_handle);
 413    } else {
 414       region->buffer = dri_bo_alloc_static(intel->bufmgr,
 415                                            name,
 416                                            region_desc->offset,
 417                                            intelScreen->pitch *
 418                                            intelScreen->height,
 419                                            region_desc->map,
 420                                            DRM_BO_FLAG_MEM_TT);
 421    }
 422
 423    assert(region->buffer != NULL);
 424
 425    return region;
 426 }
 427
 428 /**
 429  * Create intel_region structs to describe the static front, back, and depth
 430  * buffers created by the xserver.
 431  *
 432  * Although FBO's mean we now no longer use these as render targets in
 433  * all circumstances, they won't go away until the back and depth
 434  * buffers become private, and the front buffer will remain even then.
 435  *
 436  * Note that these don't allocate video memory, just describe
 437  * allocations alread made by the X server.
 438  */
 439 void
 440 intel_recreate_static_regions(struct intel_context *intel)
 441 {
 442    intelScreenPrivate *intelScreen = intel->intelScreen;
 443
 444    intel->front_region =
 445       intel_recreate_static(intel, "front",
 446                             intel->front_region,
 447                             &intelScreen->front,
 448                             DRM_BO_FLAG_MEM_TT);
 449
 450    intel->back_region =
 451       intel_recreate_static(intel, "back",
 452                             intel->back_region,
 453                             &intelScreen->back,
 454                             DRM_BO_FLAG_MEM_TT);
 455
 456 #ifdef I915
 457    if (intelScreen->third.handle) {
 458       intel->third_region =
 459          intel_recreate_static(intel, "third",
 460                                intel->third_region,
 461                                &intelScreen->third,
 462                                DRM_BO_FLAG_MEM_TT);
 463    }
 464 #endif /* I915 */
 465
 466    /* Still assumes front.cpp == depth.cpp.  We can kill this when we move to
 467     * private buffers.
 468     */
 469    intel->depth_region =
 470       intel_recreate_static(intel, "depth",
 471                             intel->depth_region,
 472                             &intelScreen->depth,
 473                             DRM_BO_FLAG_MEM_TT);
 474 }