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 <sys/ioctl.h>
  43 #include <errno.h>
  44
  45 #include "main/hash.h"
  46 #include "intel_context.h"
  47 #include "intel_regions.h"
  48 #include "intel_blit.h"
  49 #include "intel_buffer_objects.h"
  50 #include "intel_bufmgr.h"
  51 #include "intel_batchbuffer.h"
  52
  53 #define FILE_DEBUG_FLAG DEBUG_REGION
  54
  55 /* This should be set to the maximum backtrace size desired.
  56  * Set it to 0 to disable backtrace debugging.
  57  */
  58 #define DEBUG_BACKTRACE_SIZE 0
  59
  60 #if DEBUG_BACKTRACE_SIZE == 0
  61 /* Use the standard debug output */
  62 #define _DBG(...) DBG(__VA_ARGS__)
  63 #else
  64 /* Use backtracing debug output */
  65 #define _DBG(...) {debug_backtrace(); DBG(__VA_ARGS__);}
  66
  67 /* Backtracing debug support */
  68 #include <execinfo.h>
  69
  70 static void
  71 debug_backtrace(void)
  72 {
  73    void *trace[DEBUG_BACKTRACE_SIZE];
  74    char **strings = NULL;
  75    int traceSize;
  76    register int i;
  77
  78    traceSize = backtrace(trace, DEBUG_BACKTRACE_SIZE);
  79    strings = backtrace_symbols(trace, traceSize);
  80    if (strings == NULL) {
  81       DBG("no backtrace:");
  82       return;
  83    }
  84
  85    /* Spit out all the strings with a colon separator.  Ignore
  86     * the first, since we don't really care about the call
  87     * to debug_backtrace() itself.  Skip until the final "/" in
  88     * the trace to avoid really long lines.
  89     */
  90    for (i = 1; i < traceSize; i++) {
  91       char *p = strings[i], *slash = strings[i];
  92       while (*p) {
  93          if (*p++ == '/') {
  94             slash = p;
  95          }
  96       }
  97
  98       DBG("%s:", slash);
  99    }
 100
 101    /* Free up the memory, and we're done */
 102    free(strings);
 103 }
 104
 105 #endif
 106
 107
 108
 109 /* XXX: Thread safety?
 110  */
 111 GLubyte *
 112 intel_region_map(struct intel_context *intel, struct intel_region *region)
 113 {
 114    intel_flush(&intel->ctx);
 115
 116    _DBG("%s %p\n", __FUNCTION__, region);
 117    if (!region->map_refcount++) {
 118       if (region->pbo)
 119          intel_region_cow(intel, region);
 120
 121       if (region->tiling != I915_TILING_NONE)
 122          drm_intel_gem_bo_map_gtt(region->buffer);
 123       else
 124          drm_intel_bo_map(region->buffer, GL_TRUE);
 125       region->map = region->buffer->virtual;
 126    }
 127
 128    return region->map;
 129 }
 130
 131 void
 132 intel_region_unmap(struct intel_context *intel, struct intel_region *region)
 133 {
 134    _DBG("%s %p\n", __FUNCTION__, region);
 135    if (!--region->map_refcount) {
 136       if (region->tiling != I915_TILING_NONE)
 137          drm_intel_gem_bo_unmap_gtt(region->buffer);
 138       else
 139          drm_intel_bo_unmap(region->buffer);
 140       region->map = NULL;
 141    }
 142 }
 143
 144 static struct intel_region *
 145 intel_region_alloc_internal(struct intel_screen *screen,
 146                             GLuint cpp,
 147                             GLuint width, GLuint height, GLuint pitch,
 148                             uint32_t tiling, drm_intel_bo *buffer)
 149 {
 150    struct intel_region *region;
 151
 152    if (buffer == NULL) {
 153       _DBG("%s <-- NULL\n", __FUNCTION__);
 154       return NULL;
 155    }
 156
 157    region = calloc(sizeof(*region), 1);
 158    if (region == NULL)
 159       return region;
 160
 161    region->cpp = cpp;
 162    region->width = width;
 163    region->height = height;
 164    region->pitch = pitch;
 165    region->refcount = 1;
 166    region->buffer = buffer;
 167    region->tiling = tiling;
 168    region->screen = screen;
 169
 170    _DBG("%s <-- %p\n", __FUNCTION__, region);
 171    return region;
 172 }
 173
 174 struct intel_region *
 175 intel_region_alloc(struct intel_screen *screen,
 176                    uint32_t tiling,
 177                    GLuint cpp, GLuint width, GLuint height,
 178                    GLboolean expect_accelerated_upload)
 179 {
 180    drm_intel_bo *buffer;
 181    unsigned long flags = 0;
 182    unsigned long aligned_pitch;
 183
 184    if (expect_accelerated_upload)
 185       flags |= BO_ALLOC_FOR_RENDER;
 186
 187    buffer = drm_intel_bo_alloc_tiled(screen->bufmgr, "region",
 188                                      width, height, cpp,
 189                                      &tiling, &aligned_pitch, flags);
 190
 191    return intel_region_alloc_internal(screen, cpp, width, height,
 192                                       aligned_pitch / cpp, tiling, buffer);
 193 }
 194
 195 struct intel_region *
 196 intel_region_alloc_for_handle(struct intel_screen *screen,
 197                               GLuint cpp,
 198                               GLuint width, GLuint height, GLuint pitch,
 199                               GLuint handle, const char *name)
 200 {
 201    struct intel_region *region, *dummy;
 202    drm_intel_bo *buffer;
 203    int ret;
 204    uint32_t bit_6_swizzle, tiling;
 205
 206    region = _mesa_HashLookup(screen->named_regions, handle);
 207    if (region != NULL) {
 208       dummy = NULL;
 209       if (region->width != width || region->height != height ||
 210           region->cpp != cpp || region->pitch != pitch) {
 211          fprintf(stderr,
 212                  "Region for name %d already exists but is not compatible\n",
 213                  handle);
 214          return NULL;
 215       }
 216       intel_region_reference(&dummy, region);
 217       return dummy;
 218    }
 219
 220    buffer = intel_bo_gem_create_from_name(screen->bufmgr, name, handle);
 221    if (buffer == NULL)
 222       return NULL;
 223    ret = drm_intel_bo_get_tiling(buffer, &tiling, &bit_6_swizzle);
 224    if (ret != 0) {
 225       fprintf(stderr, "Couldn't get tiling of buffer %d (%s): %s\n",
 226               handle, name, strerror(-ret));
 227       drm_intel_bo_unreference(buffer);
 228       return NULL;
 229    }
 230
 231    region = intel_region_alloc_internal(screen, cpp,
 232                                         width, height, pitch, tiling, buffer);
 233    if (region == NULL) {
 234       drm_intel_bo_unreference(buffer);
 235       return NULL;
 236    }
 237
 238    region->name = handle;
 239    _mesa_HashInsert(screen->named_regions, handle, region);
 240
 241    return region;
 242 }
 243
 244 void
 245 intel_region_reference(struct intel_region **dst, struct intel_region *src)
 246 {
 247    if (src)
 248       _DBG("%s %p %d\n", __FUNCTION__, src, src->refcount);
 249
 250    assert(*dst == NULL);
 251    if (src) {
 252       src->refcount++;
 253       *dst = src;
 254    }
 255 }
 256
 257 void
 258 intel_region_release(struct intel_region **region_handle)
 259 {
 260    struct intel_region *region = *region_handle;
 261
 262    if (region == NULL) {
 263       _DBG("%s NULL\n", __FUNCTION__);
 264       return;
 265    }
 266
 267    _DBG("%s %p %d\n", __FUNCTION__, region, region->refcount - 1);
 268
 269    ASSERT(region->refcount > 0);
 270    region->refcount--;
 271
 272    if (region->refcount == 0) {
 273       assert(region->map_refcount == 0);
 274
 275       if (region->pbo)
 276          region->pbo->region = NULL;
 277       region->pbo = NULL;
 278       drm_intel_bo_unreference(region->buffer);
 279
 280       if (region->name > 0)
 281          _mesa_HashRemove(region->screen->named_regions, region->name);
 282
 283       free(region);
 284    }
 285    *region_handle = NULL;
 286 }
 287
 288 /*
 289  * XXX Move this into core Mesa?
 290  */
 291 void
 292 _mesa_copy_rect(GLubyte * dst,
 293                 GLuint cpp,
 294                 GLuint dst_pitch,
 295                 GLuint dst_x,
 296                 GLuint dst_y,
 297                 GLuint width,
 298                 GLuint height,
 299                 const GLubyte * src,
 300                 GLuint src_pitch, GLuint src_x, GLuint src_y)
 301 {
 302    GLuint i;
 303
 304    dst_pitch *= cpp;
 305    src_pitch *= cpp;
 306    dst += dst_x * cpp;
 307    src += src_x * cpp;
 308    dst += dst_y * dst_pitch;
 309    src += src_y * src_pitch;
 310    width *= cpp;
 311
 312    if (width == dst_pitch && width == src_pitch)
 313       memcpy(dst, src, height * width);
 314    else {
 315       for (i = 0; i < height; i++) {
 316          memcpy(dst, src, width);
 317          dst += dst_pitch;
 318          src += src_pitch;
 319       }
 320    }
 321 }
 322
 323
 324 /* Upload data to a rectangular sub-region.  Lots of choices how to do this:
 325  *
 326  * - memcpy by span to current destination
 327  * - upload data as new buffer and blit
 328  *
 329  * Currently always memcpy.
 330  */
 331 void
 332 intel_region_data(struct intel_context *intel,
 333                   struct intel_region *dst,
 334                   GLuint dst_offset,
 335                   GLuint dstx, GLuint dsty,
 336                   const void *src, GLuint src_pitch,
 337                   GLuint srcx, GLuint srcy, GLuint width, GLuint height)
 338 {
 339    _DBG("%s\n", __FUNCTION__);
 340
 341    if (intel == NULL)
 342       return;
 343
 344    if (dst->pbo) {
 345       if (dstx == 0 &&
 346           dsty == 0 && width == dst->pitch && height == dst->height)
 347          intel_region_release_pbo(intel, dst);
 348       else
 349          intel_region_cow(intel, dst);
 350    }
 351
 352    intel_prepare_render(intel);
 353
 354    _mesa_copy_rect(intel_region_map(intel, dst) + dst_offset,
 355                    dst->cpp,
 356                    dst->pitch,
 357                    dstx, dsty, width, height, src, src_pitch, srcx, srcy);
 358
 359    intel_region_unmap(intel, dst);
 360 }
 361
 362 /* Copy rectangular sub-regions. Need better logic about when to
 363  * push buffers into AGP - will currently do so whenever possible.
 364  */
 365 GLboolean
 366 intel_region_copy(struct intel_context *intel,
 367                   struct intel_region *dst,
 368                   GLuint dst_offset,
 369                   GLuint dstx, GLuint dsty,
 370                   struct intel_region *src,
 371                   GLuint src_offset,
 372                   GLuint srcx, GLuint srcy, GLuint width, GLuint height,
 373                   GLboolean flip,
 374                   GLenum logicop)
 375 {
 376    uint32_t src_pitch = src->pitch;
 377
 378    _DBG("%s\n", __FUNCTION__);
 379
 380    if (intel == NULL)
 381       return GL_FALSE;
 382
 383    if (dst->pbo) {
 384       if (dstx == 0 &&
 385           dsty == 0 && width == dst->pitch && height == dst->height)
 386          intel_region_release_pbo(intel, dst);
 387       else
 388          intel_region_cow(intel, dst);
 389    }
 390
 391    assert(src->cpp == dst->cpp);
 392
 393    if (flip)
 394       src_pitch = -src_pitch;
 395
 396    return intelEmitCopyBlit(intel,
 397                             dst->cpp,
 398                             src_pitch, src->buffer, src_offset, src->tiling,
 399                             dst->pitch, dst->buffer, dst_offset, dst->tiling,
 400                             srcx, srcy, dstx, dsty, width, height,
 401                             logicop);
 402 }
 403
 404 /* Attach to a pbo, discarding our data.  Effectively zero-copy upload
 405  * the pbo's data.
 406  */
 407 void
 408 intel_region_attach_pbo(struct intel_context *intel,
 409                         struct intel_region *region,
 410                         struct intel_buffer_object *pbo)
 411 {
 412    drm_intel_bo *buffer;
 413
 414    if (region->pbo == pbo)
 415       return;
 416
 417    _DBG("%s %p %p\n", __FUNCTION__, region, pbo);
 418
 419    /* If there is already a pbo attached, break the cow tie now.
 420     * Don't call intel_region_release_pbo() as that would
 421     * unnecessarily allocate a new buffer we would have to immediately
 422     * discard.
 423     */
 424    if (region->pbo) {
 425       region->pbo->region = NULL;
 426       region->pbo = NULL;
 427    }
 428
 429    if (region->buffer) {
 430       drm_intel_bo_unreference(region->buffer);
 431       region->buffer = NULL;
 432    }
 433
 434    /* make sure pbo has a buffer of its own */
 435    buffer = intel_bufferobj_buffer(intel, pbo, INTEL_WRITE_FULL);
 436
 437    region->pbo = pbo;
 438    region->pbo->region = region;
 439    drm_intel_bo_reference(buffer);
 440    region->buffer = buffer;
 441    region->tiling = I915_TILING_NONE;
 442 }
 443
 444
 445 /* Break the COW tie to the pbo and allocate a new buffer.
 446  * The pbo gets to keep the data.
 447  */
 448 void
 449 intel_region_release_pbo(struct intel_context *intel,
 450                          struct intel_region *region)
 451 {
 452    _DBG("%s %p\n", __FUNCTION__, region);
 453    assert(region->buffer == region->pbo->buffer);
 454    region->pbo->region = NULL;
 455    region->pbo = NULL;
 456    drm_intel_bo_unreference(region->buffer);
 457    region->buffer = NULL;
 458
 459    region->buffer = drm_intel_bo_alloc(intel->bufmgr, "region",
 460                                        region->pitch * region->cpp *
 461                                        region->height,
 462                                        64);
 463 }
 464
 465 /* Break the COW tie to the pbo.  Both the pbo and the region end up
 466  * with a copy of the data.
 467  */
 468 void
 469 intel_region_cow(struct intel_context *intel, struct intel_region *region)
 470 {
 471    struct intel_buffer_object *pbo = region->pbo;
 472    GLboolean ok;
 473
 474    intel_region_release_pbo(intel, region);
 475
 476    assert(region->cpp * region->pitch * region->height == pbo->Base.Size);
 477
 478    _DBG("%s %p (%d bytes)\n", __FUNCTION__, region, pbo->Base.Size);
 479
 480    /* Now blit from the texture buffer to the new buffer:
 481     */
 482
 483    intel_prepare_render(intel);
 484    ok = intelEmitCopyBlit(intel,
 485                           region->cpp,
 486                           region->pitch, pbo->buffer, 0, region->tiling,
 487                           region->pitch, region->buffer, 0, region->tiling,
 488                           0, 0, 0, 0,
 489                           region->pitch, region->height,
 490                           GL_COPY);
 491    assert(ok);
 492 }
 493
 494 drm_intel_bo *
 495 intel_region_buffer(struct intel_context *intel,
 496                     struct intel_region *region, GLuint flag)
 497 {
 498    if (region->pbo) {
 499       if (flag == INTEL_WRITE_PART)
 500          intel_region_cow(intel, region);
 501       else if (flag == INTEL_WRITE_FULL)
 502          intel_region_release_pbo(intel, region);
 503    }
 504
 505    return region->buffer;
 506 }