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    intelFlush(&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          dri_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          dri_bo_unmap(region->buffer);
 140       region->map = NULL;
 141    }
 142 }
 143
 144 static struct intel_region *
 145 intel_region_alloc_internal(struct intel_context *intel,
 146                             GLuint cpp,
 147                             GLuint width, GLuint height, GLuint pitch,
 148                             dri_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    region->cpp = cpp;
 159    region->width = width;
 160    region->height = height;
 161    region->pitch = pitch;
 162    region->refcount = 1;
 163    region->buffer = buffer;
 164
 165    /* Default to no tiling */
 166    region->tiling = I915_TILING_NONE;
 167
 168    _DBG("%s <-- %p\n", __FUNCTION__, region);
 169    return region;
 170 }
 171
 172 struct intel_region *
 173 intel_region_alloc(struct intel_context *intel,
 174                    uint32_t tiling,
 175                    GLuint cpp, GLuint width, GLuint height,
 176                    GLboolean expect_accelerated_upload)
 177 {
 178    dri_bo *buffer;
 179    struct intel_region *region;
 180    unsigned long flags = 0;
 181    unsigned long aligned_pitch;
 182
 183    if (expect_accelerated_upload)
 184       flags |= BO_ALLOC_FOR_RENDER;
 185
 186    buffer = drm_intel_bo_alloc_tiled(intel->bufmgr, "region",
 187                                      width, height, cpp,
 188                                      &tiling, &aligned_pitch, flags);
 189
 190    region = intel_region_alloc_internal(intel, cpp, width, height,
 191                                         aligned_pitch / cpp, buffer);
 192    region->tiling = tiling;
 193
 194    return region;
 195 }
 196
 197 struct intel_region *
 198 intel_region_alloc_for_handle(struct intel_context *intel,
 199                               GLuint cpp,
 200                               GLuint width, GLuint height, GLuint pitch,
 201                               GLuint handle, const char *name)
 202 {
 203    struct intel_region *region, *dummy;
 204    dri_bo *buffer;
 205    int ret;
 206    uint32_t bit_6_swizzle;
 207
 208    region = _mesa_HashLookup(intel->intelScreen->named_regions, handle);
 209    if (region != NULL) {
 210       dummy = NULL;
 211       if (region->width != width || region->height != height ||
 212           region->cpp != cpp || region->pitch != pitch) {
 213          fprintf(stderr,
 214                  "Region for name %d already exists but is not compatible\n",
 215                  handle);
 216          return NULL;
 217       }
 218       intel_region_reference(&dummy, region);
 219       return dummy;
 220    }
 221
 222    buffer = intel_bo_gem_create_from_name(intel->bufmgr, name, handle);
 223
 224    region = intel_region_alloc_internal(intel, cpp,
 225                                         width, height, pitch, buffer);
 226    if (region == NULL)
 227       return region;
 228
 229    ret = dri_bo_get_tiling(region->buffer, &region->tiling,
 230                            &bit_6_swizzle);
 231    if (ret != 0) {
 232       fprintf(stderr, "Couldn't get tiling of buffer %d (%s): %s\n",
 233               handle, name, strerror(-ret));
 234       intel_region_release(&region);
 235       return NULL;
 236    }
 237
 238    region->name = handle;
 239    region->screen = intel->intelScreen;
 240    _mesa_HashInsert(intel->intelScreen->named_regions, handle, region);
 241
 242    return region;
 243 }
 244
 245 void
 246 intel_region_reference(struct intel_region **dst, struct intel_region *src)
 247 {
 248    if (src)
 249       _DBG("%s %p %d\n", __FUNCTION__, src, src->refcount);
 250
 251    assert(*dst == NULL);
 252    if (src) {
 253       src->refcount++;
 254       *dst = src;
 255    }
 256 }
 257
 258 void
 259 intel_region_release(struct intel_region **region_handle)
 260 {
 261    struct intel_region *region = *region_handle;
 262
 263    if (region == NULL) {
 264       _DBG("%s NULL\n", __FUNCTION__);
 265       return;
 266    }
 267
 268    _DBG("%s %p %d\n", __FUNCTION__, region, region->refcount - 1);
 269
 270    ASSERT(region->refcount > 0);
 271    region->refcount--;
 272
 273    if (region->refcount == 0) {
 274       assert(region->map_refcount == 0);
 275
 276       if (region->pbo)
 277          region->pbo->region = NULL;
 278       region->pbo = NULL;
 279       dri_bo_unreference(region->buffer);
 280
 281       if (region->name > 0)
 282          _mesa_HashRemove(region->screen->named_regions, region->name);
 283
 284       free(region);
 285    }
 286    *region_handle = NULL;
 287 }
 288
 289 /*
 290  * XXX Move this into core Mesa?
 291  */
 292 void
 293 _mesa_copy_rect(GLubyte * dst,
 294                 GLuint cpp,
 295                 GLuint dst_pitch,
 296                 GLuint dst_x,
 297                 GLuint dst_y,
 298                 GLuint width,
 299                 GLuint height,
 300                 const GLubyte * src,
 301                 GLuint src_pitch, GLuint src_x, GLuint src_y)
 302 {
 303    GLuint i;
 304
 305    dst_pitch *= cpp;
 306    src_pitch *= cpp;
 307    dst += dst_x * cpp;
 308    src += src_x * cpp;
 309    dst += dst_y * dst_pitch;
 310    src += src_y * src_pitch;
 311    width *= cpp;
 312
 313    if (width == dst_pitch && width == src_pitch)
 314       memcpy(dst, src, height * width);
 315    else {
 316       for (i = 0; i < height; i++) {
 317          memcpy(dst, src, width);
 318          dst += dst_pitch;
 319          src += src_pitch;
 320       }
 321    }
 322 }
 323
 324
 325 /* Upload data to a rectangular sub-region.  Lots of choices how to do this:
 326  *
 327  * - memcpy by span to current destination
 328  * - upload data as new buffer and blit
 329  *
 330  * Currently always memcpy.
 331  */
 332 void
 333 intel_region_data(struct intel_context *intel,
 334                   struct intel_region *dst,
 335                   GLuint dst_offset,
 336                   GLuint dstx, GLuint dsty,
 337                   const void *src, GLuint src_pitch,
 338                   GLuint srcx, GLuint srcy, GLuint width, GLuint height)
 339 {
 340    _DBG("%s\n", __FUNCTION__);
 341
 342    if (intel == NULL)
 343       return;
 344
 345    if (dst->pbo) {
 346       if (dstx == 0 &&
 347           dsty == 0 && width == dst->pitch && height == dst->height)
 348          intel_region_release_pbo(intel, dst);
 349       else
 350          intel_region_cow(intel, dst);
 351    }
 352
 353    intel_prepare_render(intel);
 354
 355    _mesa_copy_rect(intel_region_map(intel, dst) + dst_offset,
 356                    dst->cpp,
 357                    dst->pitch,
 358                    dstx, dsty, width, height, src, src_pitch, srcx, srcy);
 359
 360    intel_region_unmap(intel, dst);
 361 }
 362
 363 /* Copy rectangular sub-regions. Need better logic about when to
 364  * push buffers into AGP - will currently do so whenever possible.
 365  */
 366 GLboolean
 367 intel_region_copy(struct intel_context *intel,
 368                   struct intel_region *dst,
 369                   GLuint dst_offset,
 370                   GLuint dstx, GLuint dsty,
 371                   struct intel_region *src,
 372                   GLuint src_offset,
 373                   GLuint srcx, GLuint srcy, GLuint width, GLuint height,
 374                   GLboolean flip,
 375                   GLenum logicop)
 376 {
 377    uint32_t src_pitch = src->pitch;
 378
 379    _DBG("%s\n", __FUNCTION__);
 380
 381    if (intel == NULL)
 382       return GL_FALSE;
 383
 384    if (dst->pbo) {
 385       if (dstx == 0 &&
 386           dsty == 0 && width == dst->pitch && height == dst->height)
 387          intel_region_release_pbo(intel, dst);
 388       else
 389          intel_region_cow(intel, dst);
 390    }
 391
 392    assert(src->cpp == dst->cpp);
 393
 394    if (flip)
 395       src_pitch = -src_pitch;
 396
 397    return intelEmitCopyBlit(intel,
 398                             dst->cpp,
 399                             src_pitch, src->buffer, src_offset, src->tiling,
 400                             dst->pitch, dst->buffer, dst_offset, dst->tiling,
 401                             srcx, srcy, dstx, dsty, width, height,
 402                             logicop);
 403 }
 404
 405 /* Attach to a pbo, discarding our data.  Effectively zero-copy upload
 406  * the pbo's data.
 407  */
 408 void
 409 intel_region_attach_pbo(struct intel_context *intel,
 410                         struct intel_region *region,
 411                         struct intel_buffer_object *pbo)
 412 {
 413    dri_bo *buffer;
 414
 415    if (region->pbo == pbo)
 416       return;
 417
 418    _DBG("%s %p %p\n", __FUNCTION__, region, pbo);
 419
 420    /* If there is already a pbo attached, break the cow tie now.
 421     * Don't call intel_region_release_pbo() as that would
 422     * unnecessarily allocate a new buffer we would have to immediately
 423     * discard.
 424     */
 425    if (region->pbo) {
 426       region->pbo->region = NULL;
 427       region->pbo = NULL;
 428    }
 429
 430    if (region->buffer) {
 431       dri_bo_unreference(region->buffer);
 432       region->buffer = NULL;
 433    }
 434
 435    /* make sure pbo has a buffer of its own */
 436    buffer = intel_bufferobj_buffer(intel, pbo, INTEL_WRITE_FULL);
 437
 438    region->pbo = pbo;
 439    region->pbo->region = region;
 440    dri_bo_reference(buffer);
 441    region->buffer = buffer;
 442    region->tiling = I915_TILING_NONE;
 443 }
 444
 445
 446 /* Break the COW tie to the pbo and allocate a new buffer.
 447  * The pbo gets to keep the data.
 448  */
 449 void
 450 intel_region_release_pbo(struct intel_context *intel,
 451                          struct intel_region *region)
 452 {
 453    _DBG("%s %p\n", __FUNCTION__, region);
 454    assert(region->buffer == region->pbo->buffer);
 455    region->pbo->region = NULL;
 456    region->pbo = NULL;
 457    dri_bo_unreference(region->buffer);
 458    region->buffer = NULL;
 459
 460    region->buffer = dri_bo_alloc(intel->bufmgr, "region",
 461                                  region->pitch * region->cpp * 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 dri_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 }