src/mesa/drivers/dri/r300/r300_texmem.c

   1 /* $XFree86: xc/lib/GL/mesa/src/drv/r300/r300_texmem.c,v 1.5 2002/12/17 00:32:56 dawes Exp $ */
   2 /**************************************************************************
   3
   4 Copyright (C) Tungsten Graphics 2002.  All Rights Reserved.
   5 The Weather Channel, Inc. funded Tungsten Graphics to develop the
   6 initial release of the Radeon 8500 driver under the XFree86
   7 license. This notice must be preserved.
   8
   9 Permission is hereby granted, free of charge, to any person obtaining
  10 a copy of this software and associated documentation files (the
  11 "Software"), to deal in the Software without restriction, including
  12 without limitation on the rights to use, copy, modify, merge, publish,
  13 distribute, sub license, and/or sell copies of the Software, and to
  14 permit persons to whom the Software is furnished to do so, subject to
  15 the following conditions:
  16
  17 The above copyright notice and this permission notice (including the
  18 next paragraph) shall be included in all copies or substantial
  19 portions of the Software.
  20
  21 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  22 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  23 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  24 NON-INFRINGEMENT. IN NO EVENT SHALL ATI, VA LINUX SYSTEMS AND/OR THEIR
  25 SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
  26 IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR
  27 IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  28 SOFTWARE.
  29
  30 **************************************************************************/
  31
  32 /*
  33  * Authors:
  34  *   Kevin E. Martin <martin@valinux.com>
  35  *   Gareth Hughes <gareth@valinux.com>
  36  *
  37  */
  38
  39 #include <errno.h>
  40
  41 #include "glheader.h"
  42 #include "imports.h"
  43 #include "context.h"
  44 #include "colormac.h"
  45 #include "macros.h"
  46 #include "simple_list.h"
  47 #include "radeon_reg.h"         /* gets definition for usleep */
  48 #include "r300_context.h"
  49 #include "r300_state.h"
  50 #include "r300_cmdbuf.h"
  51 #include "radeon_ioctl.h"
  52 /*
  53 #include "r300_swtcl.h"
  54 */
  55 #include "r300_tex.h"
  56 #include "r300_ioctl.h"
  57 #include <unistd.h>             /* for usleep() */
  58
  59 #ifdef USER_BUFFERS
  60 #include "radeon_mm.h"
  61 #endif
  62
  63 /**
  64  * Destroy any device-dependent state associated with the texture.  This may
  65  * include NULLing out hardware state that points to the texture.
  66  */
  67 void r300DestroyTexObj(r300ContextPtr rmesa, r300TexObjPtr t)
  68 {
  69         if (RADEON_DEBUG & DEBUG_TEXTURE) {
  70                 fprintf(stderr, "%s( %p, %p )\n", __FUNCTION__,
  71                         (void *)t, (void *)t->base.tObj);
  72         }
  73
  74         if (rmesa != NULL) {
  75                 unsigned i;
  76
  77                 for (i = 0; i < rmesa->radeon.glCtx->Const.MaxTextureUnits; i++) {
  78                         if (t == rmesa->state.texture.unit[i].texobj) {
  79                                 rmesa->state.texture.unit[i].texobj = NULL;
  80                                 /* This code below is meant to shorten state
  81                                    pushed to the hardware by not programming
  82                                    unneeded units.
  83
  84                                    This does not appear to be worthwhile on R300 */
  85 #if 0
  86                                 remove_from_list(&rmesa->hw.tex[i]);
  87                                 make_empty_list(&rmesa->hw.tex[i]);
  88                                 remove_from_list(&rmesa->hw.cube[i]);
  89                                 make_empty_list(&rmesa->hw.cube[i]);
  90 #endif
  91                         }
  92                 }
  93         }
  94 }
  95
  96 /* ------------------------------------------------------------
  97  * Texture image conversions
  98  */
  99
 100 static void r300UploadGARTClientSubImage(r300ContextPtr rmesa,
 101                                          r300TexObjPtr t,
 102                                          struct gl_texture_image *texImage,
 103                                          GLint hwlevel,
 104                                          GLint x, GLint y,
 105                                          GLint width, GLint height)
 106 {
 107         const struct gl_texture_format *texFormat = texImage->TexFormat;
 108         GLuint srcPitch, dstPitch;
 109         int blit_format;
 110         int srcOffset;
 111
 112         /*
 113          * XXX it appears that we always upload the full image, not a subimage.
 114          * I.e. x==0, y==0, width=texWidth, height=texWidth.  If this is ever
 115          * changed, the src pitch will have to change.
 116          */
 117         switch (texFormat->TexelBytes) {
 118         case 1:
 119                 blit_format = R300_CP_COLOR_FORMAT_CI8;
 120                 srcPitch = t->image[0][0].width * texFormat->TexelBytes;
 121                 dstPitch = t->image[0][0].width * texFormat->TexelBytes;
 122                 break;
 123         case 2:
 124                 blit_format = R300_CP_COLOR_FORMAT_RGB565;
 125                 srcPitch = t->image[0][0].width * texFormat->TexelBytes;
 126                 dstPitch = t->image[0][0].width * texFormat->TexelBytes;
 127                 break;
 128         case 4:
 129                 blit_format = R300_CP_COLOR_FORMAT_ARGB8888;
 130                 srcPitch = t->image[0][0].width * texFormat->TexelBytes;
 131                 dstPitch = t->image[0][0].width * texFormat->TexelBytes;
 132                 break;
 133         case 8:
 134         case 16:
 135                 blit_format = R300_CP_COLOR_FORMAT_CI8;
 136                 srcPitch = t->image[0][0].width * texFormat->TexelBytes;
 137                 dstPitch = t->image[0][0].width * texFormat->TexelBytes;
 138                 break;
 139         default:
 140                 return;
 141         }
 142
 143         t->image[0][hwlevel].data = texImage->Data;
 144         srcOffset = r300GartOffsetFromVirtual(rmesa, texImage->Data);
 145
 146         assert(srcOffset != ~0);
 147
 148         /* Don't currently need to cope with small pitches?
 149          */
 150         width = texImage->Width;
 151         height = texImage->Height;
 152
 153         if (texFormat->TexelBytes > 4) {
 154                 width *= texFormat->TexelBytes;
 155         }
 156
 157         r300EmitWait(rmesa, R300_WAIT_3D);
 158
 159         r300EmitBlit(rmesa, blit_format,
 160                      srcPitch,
 161                      srcOffset,
 162                      dstPitch,
 163                      t->bufAddr,
 164                      x,
 165                      y,
 166                      t->image[0][hwlevel].x + x,
 167                      t->image[0][hwlevel].y + y, width, height);
 168
 169         r300EmitWait(rmesa, R300_WAIT_2D);
 170 }
 171
 172 static void r300UploadRectSubImage(r300ContextPtr rmesa,
 173                                    r300TexObjPtr t,
 174                                    struct gl_texture_image *texImage,
 175                                    GLint x, GLint y, GLint width, GLint height)
 176 {
 177         const struct gl_texture_format *texFormat = texImage->TexFormat;
 178         int blit_format, dstPitch, done;
 179
 180         switch (texFormat->TexelBytes) {
 181         case 1:
 182                 blit_format = R300_CP_COLOR_FORMAT_CI8;
 183                 break;
 184         case 2:
 185                 blit_format = R300_CP_COLOR_FORMAT_RGB565;
 186                 break;
 187         case 4:
 188                 blit_format = R300_CP_COLOR_FORMAT_ARGB8888;
 189                 break;
 190         case 8:
 191         case 16:
 192                 blit_format = R300_CP_COLOR_FORMAT_CI8;
 193                 break;
 194         default:
 195                 return;
 196         }
 197
 198         t->image[0][0].data = texImage->Data;
 199
 200         /* Currently don't need to cope with small pitches.
 201          */
 202         width = texImage->Width;
 203         height = texImage->Height;
 204         dstPitch = t->pitch;
 205
 206         if (texFormat->TexelBytes > 4) {
 207                 width *= texFormat->TexelBytes;
 208         }
 209
 210         if (rmesa->prefer_gart_client_texturing && texImage->IsClientData) {
 211                 /* In this case, could also use GART texturing.  This is
 212                  * currently disabled, but has been tested & works.
 213                  */
 214                 t->offset =
 215                     r300GartOffsetFromVirtual(rmesa, texImage->Data);
 216                 t->pitch =
 217                     texImage->RowStride * texFormat->TexelBytes - 32;
 218
 219                 if (RADEON_DEBUG & DEBUG_TEXTURE)
 220                         fprintf(stderr,
 221                                 "Using GART texturing for rectangular client texture\n");
 222
 223                 /* Release FB memory allocated for this image:
 224                  */
 225                 /* FIXME This may not be correct as driSwapOutTextureObject sets
 226                  * FIXME dirty_images.  It may be fine, though.
 227                  */
 228                 if (t->base.memBlock) {
 229                         driSwapOutTextureObject((driTextureObject *) t);
 230                 }
 231         } else if (texImage->IsClientData) {
 232                 /* Data already in GART memory, with usable pitch.
 233                  */
 234                 GLuint srcPitch;
 235                 srcPitch = texImage->RowStride * texFormat->TexelBytes;
 236                 r300EmitBlit(rmesa,
 237                              blit_format,
 238                              srcPitch,
 239                              r300GartOffsetFromVirtual(rmesa, texImage->Data),
 240                              dstPitch, t->bufAddr, 0, 0, 0, 0, width, height);
 241         } else {
 242                 /* Data not in GART memory, or bad pitch.
 243                  */
 244                 for (done = 0; done < height;) {
 245                         struct r300_dma_region region;
 246                         int lines =
 247                             MIN2(height - done, RADEON_BUFFER_SIZE / dstPitch);
 248                         int src_pitch;
 249                         char *tex;
 250
 251                         src_pitch = texImage->RowStride * texFormat->TexelBytes;
 252
 253                         tex = (char *)texImage->Data + done * src_pitch;
 254
 255                         memset(&region, 0, sizeof(region));
 256                         r300AllocDmaRegion(rmesa, &region, lines * dstPitch,
 257                                            1024);
 258
 259                         /* Copy texdata to dma:
 260                          */
 261                         if (0)
 262                                 fprintf(stderr,
 263                                         "%s: src_pitch %d dst_pitch %d\n",
 264                                         __FUNCTION__, src_pitch, dstPitch);
 265
 266                         if (src_pitch == dstPitch) {
 267                                 memcpy(region.address + region.start, tex,
 268                                        lines * src_pitch);
 269                         } else {
 270                                 char *buf = region.address + region.start;
 271                                 int i;
 272                                 for (i = 0; i < lines; i++) {
 273                                         memcpy(buf, tex, src_pitch);
 274                                         buf += dstPitch;
 275                                         tex += src_pitch;
 276                                 }
 277                         }
 278
 279                         r300EmitWait(rmesa, R300_WAIT_3D);
 280
 281                         /* Blit to framebuffer
 282                          */
 283                         r300EmitBlit(rmesa,
 284                                      blit_format,
 285                                      dstPitch, GET_START(&region),
 286                                      dstPitch | (t->tile_bits >> 16),
 287                                      t->bufAddr,
 288                                      0, 0,
 289                                      0, done,
 290                                      width, lines);
 291
 292                         r300EmitWait(rmesa, R300_WAIT_2D);
 293 #ifdef USER_BUFFERS
 294                         radeon_mm_use(rmesa, region.buf->id);
 295 #endif
 296
 297                         r300ReleaseDmaRegion(rmesa, &region, __FUNCTION__);
 298                         done += lines;
 299                 }
 300         }
 301 }
 302
 303 /**
 304  * Upload the texture image associated with texture \a t at the specified
 305  * level at the address relative to \a start.
 306  */
 307 static void uploadSubImage( r300ContextPtr rmesa, r300TexObjPtr t,
 308                             GLint hwlevel,
 309                             GLint x, GLint y, GLint width, GLint height,
 310                             GLuint face )
 311 {
 312    struct gl_texture_image *texImage = NULL;
 313    GLuint offset;
 314    GLint imageWidth, imageHeight;
 315    GLint ret;
 316    drm_radeon_texture_t tex;
 317    drm_radeon_tex_image_t tmp;
 318    const int level = hwlevel + t->base.firstLevel;
 319
 320    if ( RADEON_DEBUG & DEBUG_TEXTURE ) {
 321       fprintf( stderr, "%s( %p, %p ) level/width/height/face = %d/%d/%d/%u\n",
 322                __FUNCTION__, (void *)t, (void *)t->base.tObj,
 323                level, width, height, face );
 324    }
 325
 326    ASSERT(face < 6);
 327
 328    /* Ensure we have a valid texture to upload */
 329    if ( ( hwlevel < 0 ) || ( hwlevel >= RADEON_MAX_TEXTURE_LEVELS ) ) {
 330       _mesa_problem(NULL, "bad texture level in %s", __FUNCTION__);
 331       return;
 332    }
 333
 334    texImage = t->base.tObj->Image[face][level];
 335
 336    if ( !texImage ) {
 337       if ( RADEON_DEBUG & DEBUG_TEXTURE )
 338          fprintf( stderr, "%s: texImage %d is NULL!\n", __FUNCTION__, level );
 339       return;
 340    }
 341    if ( !texImage->Data ) {
 342       if ( RADEON_DEBUG & DEBUG_TEXTURE )
 343          fprintf( stderr, "%s: image data is NULL!\n", __FUNCTION__ );
 344       return;
 345    }
 346
 347
 348    if (t->base.tObj->Target == GL_TEXTURE_RECTANGLE_NV) {
 349       assert(level == 0);
 350       assert(hwlevel == 0);
 351       if ( RADEON_DEBUG & DEBUG_TEXTURE )
 352          fprintf( stderr, "%s: image data is rectangular\n", __FUNCTION__);
 353       r300UploadRectSubImage( rmesa, t, texImage, x, y, width, height );
 354       return;
 355    }
 356    else if (texImage->IsClientData) {
 357       if ( RADEON_DEBUG & DEBUG_TEXTURE )
 358          fprintf( stderr, "%s: image data is in GART client storage\n",
 359                   __FUNCTION__);
 360       r300UploadGARTClientSubImage( rmesa, t, texImage, hwlevel,
 361                                    x, y, width, height );
 362       return;
 363    }
 364    else if ( RADEON_DEBUG & DEBUG_TEXTURE )
 365       fprintf( stderr, "%s: image data is in normal memory\n",
 366                __FUNCTION__);
 367
 368
 369    imageWidth = texImage->Width;
 370    imageHeight = texImage->Height;
 371
 372    offset = t->bufAddr + t->base.totalSize / 6 * face;
 373
 374    if ( RADEON_DEBUG & (DEBUG_TEXTURE|DEBUG_IOCTL) ) {
 375       GLint imageX = 0;
 376       GLint imageY = 0;
 377       GLint blitX = t->image[face][hwlevel].x;
 378       GLint blitY = t->image[face][hwlevel].y;
 379       GLint blitWidth = t->image[face][hwlevel].width;
 380       GLint blitHeight = t->image[face][hwlevel].height;
 381       fprintf( stderr, "   upload image: %d,%d at %d,%d\n",
 382                imageWidth, imageHeight, imageX, imageY );
 383       fprintf( stderr, "   upload  blit: %d,%d at %d,%d\n",
 384                blitWidth, blitHeight, blitX, blitY );
 385       fprintf( stderr, "       blit ofs: 0x%07x level: %d/%d\n",
 386                (GLuint)offset, hwlevel, level );
 387    }
 388
 389    t->image[face][hwlevel].data = texImage->Data;
 390
 391    /* Init the DRM_RADEON_TEXTURE command / drm_radeon_texture_t struct.
 392     * NOTE: we're always use a 1KB-wide blit and I8 texture format.
 393     * We used to use 1, 2 and 4-byte texels and used to use the texture
 394     * width to dictate the blit width - but that won't work for compressed
 395     * textures. (Brian)
 396     * NOTE: can't do that with texture tiling. (sroland)
 397     */
 398    tex.offset = offset;
 399    tex.image = &tmp;
 400    /* copy (x,y,width,height,data) */
 401    memcpy( &tmp, &t->image[face][hwlevel], sizeof(tmp) );
 402
 403    if (texImage->TexFormat->TexelBytes > 4) {
 404       const int log2TexelBytes = (3 + (texImage->TexFormat->TexelBytes >> 4));
 405       tex.format = RADEON_TXFORMAT_I8; /* any 1-byte texel format */
 406       tex.pitch = MAX2((texImage->Width * texImage->TexFormat->TexelBytes) / 64, 1);
 407       tex.height = imageHeight;
 408       tex.width = imageWidth << log2TexelBytes;
 409       tex.offset += (tmp.x << log2TexelBytes) & ~1023;
 410       tmp.x = tmp.x % (1024 >> log2TexelBytes);
 411       tmp.width = tmp.width << log2TexelBytes;
 412    }
 413    else if (texImage->TexFormat->TexelBytes) {
 414       /* use multi-byte upload scheme */
 415       tex.height = imageHeight;
 416       tex.width = imageWidth;
 417       switch(texImage->TexFormat->TexelBytes) {
 418       case 1:
 419         tex.format = RADEON_TXFORMAT_I8;
 420         break;
 421       case 2:
 422         tex.format = RADEON_TXFORMAT_AI88;
 423         break;
 424       case 4:
 425         tex.format = RADEON_TXFORMAT_ARGB8888;
 426         break;
 427       }
 428       tex.pitch = MAX2((texImage->Width * texImage->TexFormat->TexelBytes) / 64, 1);
 429       tex.offset += tmp.x & ~1023;
 430       tmp.x = tmp.x % 1024;
 431 #if 1
 432       if (t->tile_bits & R300_TXO_MICRO_TILE) {
 433          /* need something like "tiled coordinates" ? */
 434          tmp.y = tmp.x / (tex.pitch * 128) * 2;
 435          tmp.x = tmp.x % (tex.pitch * 128) / 2 / texImage->TexFormat->TexelBytes;
 436          tex.pitch |= RADEON_DST_TILE_MICRO >> 22;
 437       }
 438       else
 439 #endif
 440       {
 441          tmp.x = tmp.x >> (texImage->TexFormat->TexelBytes >> 1);
 442       }
 443 #if 1
 444       if ((t->tile_bits & R300_TXO_MACRO_TILE) &&
 445          (texImage->Width * texImage->TexFormat->TexelBytes >= 256) &&
 446          ((!(t->tile_bits & R300_TXO_MICRO_TILE) && (texImage->Height >= 8)) ||
 447             (texImage->Height >= 16))) {
 448          /* weird: R200 disables macro tiling if mip width is smaller than 256 bytes,
 449             OR if height is smaller than 8 automatically, but if micro tiling is active
 450             the limit is height 16 instead ? */
 451          tex.pitch |= RADEON_DST_TILE_MACRO >> 22;
 452       }
 453 #endif
 454    }
 455    else {
 456       /* In case of for instance 8x8 texture (2x2 dxt blocks), padding after the first two blocks is
 457          needed (only with dxt1 since 2 dxt3/dxt5 blocks already use 32 Byte). */
 458       /* set tex.height to 1/4 since 1 "macropixel" (dxt-block) has 4 real pixels. Needed
 459          so the kernel module reads the right amount of data. */
 460       tex.format = RADEON_TXFORMAT_I8; /* any 1-byte texel format */
 461       tex.pitch = (R300_BLIT_WIDTH_BYTES / 64);
 462       tex.height = (imageHeight + 3) / 4;
 463       tex.width = (imageWidth + 3) / 4;
 464       if ((t->format & R300_TX_FORMAT_DXT1) == R300_TX_FORMAT_DXT1)
 465       {
 466            tex.width *= 8;
 467       } else {
 468            tex.width *= 16;
 469       }
 470    }
 471
 472    LOCK_HARDWARE( &rmesa->radeon );
 473    do {
 474       ret = drmCommandWriteRead( rmesa->radeon.dri.fd, DRM_RADEON_TEXTURE,
 475                                  &tex, sizeof(drm_radeon_texture_t) );
 476       if (ret) {
 477          if (RADEON_DEBUG & DEBUG_IOCTL)
 478             fprintf(stderr, "DRM_RADEON_TEXTURE:  again!\n");
 479          usleep(1);
 480       }
 481    } while ( ret == -EAGAIN );
 482
 483    UNLOCK_HARDWARE( &rmesa->radeon );
 484
 485    if ( ret ) {
 486       fprintf( stderr, "DRM_RADEON_TEXTURE: return = %d\n", ret );
 487       fprintf( stderr, "   offset=0x%08x\n",
 488                offset );
 489       fprintf( stderr, "   image width=%d height=%d\n",
 490                imageWidth, imageHeight );
 491       fprintf( stderr, "    blit width=%d height=%d data=%p\n",
 492                t->image[face][hwlevel].width, t->image[face][hwlevel].height,
 493                t->image[face][hwlevel].data );
 494       exit( 1 );
 495    }
 496 }
 497
 498 /**
 499  * Upload the texture images associated with texture \a t.  This might
 500  * require the allocation of texture memory.
 501  *
 502  * \param rmesa Context pointer
 503  * \param t Texture to be uploaded
 504  * \param face Cube map face to be uploaded.  Zero for non-cube maps.
 505  */
 506
 507 int r300UploadTexImages(r300ContextPtr rmesa, r300TexObjPtr t, GLuint face)
 508 {
 509         const int numLevels = t->base.lastLevel - t->base.firstLevel + 1;
 510
 511         if (RADEON_DEBUG & (DEBUG_TEXTURE | DEBUG_IOCTL)) {
 512                 fprintf(stderr, "%s( %p, %p ) sz=%d lvls=%d-%d\n", __FUNCTION__,
 513                         (void *)rmesa->radeon.glCtx, (void *)t->base.tObj,
 514                         t->base.totalSize, t->base.firstLevel,
 515                         t->base.lastLevel);
 516         }
 517
 518         if (!t || t->base.totalSize == 0)
 519                 return 0;
 520
 521         if (RADEON_DEBUG & DEBUG_SYNC) {
 522                 fprintf(stderr, "%s: Syncing\n", __FUNCTION__);
 523                 radeonFinish(rmesa->radeon.glCtx);
 524         }
 525
 526         LOCK_HARDWARE(&rmesa->radeon);
 527
 528         if (t->base.memBlock == NULL) {
 529                 int heap;
 530
 531                 heap = driAllocateTexture(rmesa->texture_heaps, rmesa->nr_heaps,
 532                                           (driTextureObject *) t);
 533                 if (heap == -1) {
 534                         UNLOCK_HARDWARE(&rmesa->radeon);
 535                         return -1;
 536                 }
 537
 538                 /* Set the base offset of the texture image */
 539                 t->bufAddr = rmesa->radeon.radeonScreen->texOffset[heap]
 540                     + t->base.memBlock->ofs;
 541                 t->offset = t->bufAddr;
 542
 543                 if (!(t->base.tObj->Image[0][0]->IsClientData)) {
 544                         /* hope it's safe to add that here... */
 545                         t->offset |= t->tile_bits;
 546                 }
 547
 548                 /* Mark this texobj as dirty on all units:
 549                  */
 550                 t->dirty_state = TEX_ALL;
 551         }
 552
 553         /* Let the world know we've used this memory recently.
 554          */
 555         driUpdateTextureLRU((driTextureObject *) t);
 556         UNLOCK_HARDWARE(&rmesa->radeon);
 557
 558         /* Upload any images that are new */
 559         if (t->base.dirty_images[face]) {
 560                 int i;
 561                 for (i = 0; i < numLevels; i++) {
 562                         if ((t->base.
 563                              dirty_images[face] & (1 <<
 564                                                    (i + t->base.firstLevel))) !=
 565                             0) {
 566                                 uploadSubImage(rmesa, t, i, 0, 0,
 567                                                t->image[face][i].width,
 568                                                t->image[face][i].height, face);
 569                         }
 570                 }
 571                 t->base.dirty_images[face] = 0;
 572         }
 573
 574         if (RADEON_DEBUG & DEBUG_SYNC) {
 575                 fprintf(stderr, "%s: Syncing\n", __FUNCTION__);
 576                 radeonFinish(rmesa->radeon.glCtx);
 577         }
 578
 579         return 0;
 580 }