src/mesa/drivers/dri/r200/r200_texmem.c

   1 /* $XFree86: xc/lib/GL/mesa/src/drv/r200/r200_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 "r200_context.h"
  47 #include "r200_ioctl.h"
  48 #include "r200_tex.h"
  49 #include "radeon_reg.h"
  50
  51 #include <unistd.h>  /* for usleep() */
  52
  53
  54 /**
  55  * Destroy any device-dependent state associated with the texture.  This may
  56  * include NULLing out hardware state that points to the texture.
  57  */
  58 void
  59 r200DestroyTexObj( r200ContextPtr rmesa, r200TexObjPtr t )
  60 {
  61    if ( R200_DEBUG & DEBUG_TEXTURE ) {
  62       fprintf( stderr, "%s( %p, %p )\n", __FUNCTION__,
  63                (void *)t, (void *)t->base.tObj );
  64    }
  65
  66    if ( rmesa != NULL ) {
  67       unsigned   i;
  68
  69
  70       for ( i = 0 ; i < rmesa->glCtx->Const.MaxTextureUnits ; i++ ) {
  71          if ( t == rmesa->state.texture.unit[i].texobj ) {
  72             rmesa->state.texture.unit[i].texobj = NULL;
  73             rmesa->hw.tex[i].dirty = GL_FALSE;
  74             rmesa->hw.cube[i].dirty = GL_FALSE;
  75          }
  76       }
  77    }
  78 }
  79
  80
  81 /* ------------------------------------------------------------
  82  * Texture image conversions
  83  */
  84
  85
  86 static void r200UploadGARTClientSubImage( r200ContextPtr rmesa,
  87                                           r200TexObjPtr t,
  88                                           struct gl_texture_image *texImage,
  89                                           GLint hwlevel,
  90                                           GLint x, GLint y,
  91                                           GLint width, GLint height )
  92 {
  93    const struct gl_texture_format *texFormat = texImage->TexFormat;
  94    GLuint srcPitch, dstPitch;
  95    int blit_format;
  96    int srcOffset;
  97
  98    /*
  99     * XXX it appears that we always upload the full image, not a subimage.
 100     * I.e. x==0, y==0, width=texWidth, height=texWidth.  If this is ever
 101     * changed, the src pitch will have to change.
 102     */
 103    switch ( texFormat->TexelBytes ) {
 104    case 1:
 105       blit_format = R200_CP_COLOR_FORMAT_CI8;
 106       srcPitch = t->image[0][0].width * texFormat->TexelBytes;
 107       dstPitch = t->image[0][0].width * texFormat->TexelBytes;
 108       break;
 109    case 2:
 110       blit_format = R200_CP_COLOR_FORMAT_RGB565;
 111       srcPitch = t->image[0][0].width * texFormat->TexelBytes;
 112       dstPitch = t->image[0][0].width * texFormat->TexelBytes;
 113       break;
 114    case 4:
 115       blit_format = R200_CP_COLOR_FORMAT_ARGB8888;
 116       srcPitch = t->image[0][0].width * texFormat->TexelBytes;
 117       dstPitch = t->image[0][0].width * texFormat->TexelBytes;
 118       break;
 119    default:
 120       return;
 121    }
 122
 123    t->image[0][hwlevel].data = texImage->Data;
 124    srcOffset = r200GartOffsetFromVirtual( rmesa, texImage->Data );
 125
 126    assert( srcOffset != ~0 );
 127
 128    /* Don't currently need to cope with small pitches?
 129     */
 130    width = texImage->Width;
 131    height = texImage->Height;
 132
 133    r200EmitWait( rmesa, RADEON_WAIT_3D );
 134
 135    r200EmitBlit( rmesa, blit_format,
 136                  srcPitch,
 137                  srcOffset,
 138                  dstPitch,
 139                  t->bufAddr,
 140                  x,
 141                  y,
 142                  t->image[0][hwlevel].x + x,
 143                  t->image[0][hwlevel].y + y,
 144                  width,
 145                  height );
 146
 147    r200EmitWait( rmesa, RADEON_WAIT_2D );
 148 }
 149
 150 static void r200UploadRectSubImage( r200ContextPtr rmesa,
 151                                     r200TexObjPtr t,
 152                                     struct gl_texture_image *texImage,
 153                                     GLint x, GLint y,
 154                                     GLint width, GLint height )
 155 {
 156    const struct gl_texture_format *texFormat = texImage->TexFormat;
 157    int blit_format, dstPitch, done;
 158
 159    switch ( texFormat->TexelBytes ) {
 160    case 1:
 161       blit_format = R200_CP_COLOR_FORMAT_CI8;
 162       break;
 163    case 2:
 164       blit_format = R200_CP_COLOR_FORMAT_RGB565;
 165       break;
 166    case 4:
 167       blit_format = R200_CP_COLOR_FORMAT_ARGB8888;
 168       break;
 169    default:
 170       return;
 171    }
 172
 173    t->image[0][0].data = texImage->Data;
 174
 175    /* Currently don't need to cope with small pitches.
 176     */
 177    width = texImage->Width;
 178    height = texImage->Height;
 179    dstPitch = t->pp_txpitch + 32;
 180
 181    if (rmesa->prefer_gart_client_texturing && texImage->IsClientData) {
 182       /* In this case, could also use GART texturing.  This is
 183        * currently disabled, but has been tested & works.
 184        */
 185       t->pp_txoffset = r200GartOffsetFromVirtual( rmesa, texImage->Data );
 186       t->pp_txpitch = texImage->RowStride * texFormat->TexelBytes - 32;
 187
 188       if (R200_DEBUG & DEBUG_TEXTURE)
 189          fprintf(stderr,
 190                  "Using GART texturing for rectangular client texture\n");
 191
 192       /* Release FB memory allocated for this image:
 193        */
 194       /* FIXME This may not be correct as driSwapOutTextureObject sets
 195        * FIXME dirty_images.  It may be fine, though.
 196        */
 197       if ( t->base.memBlock ) {
 198          driSwapOutTextureObject( (driTextureObject *) t );
 199       }
 200    }
 201    else if (texImage->IsClientData) {
 202       /* Data already in GART memory, with usable pitch.
 203        */
 204       GLuint srcPitch;
 205       srcPitch = texImage->RowStride * texFormat->TexelBytes;
 206       r200EmitBlit( rmesa,
 207                     blit_format,
 208                     srcPitch,
 209                     r200GartOffsetFromVirtual( rmesa, texImage->Data ),
 210                     dstPitch, t->bufAddr,
 211                     0, 0,
 212                     0, 0,
 213                     width, height );
 214    }
 215    else {
 216       /* Data not in GART memory, or bad pitch.
 217        */
 218       for (done = 0; done < height ; ) {
 219          struct r200_dma_region region;
 220          int lines = MIN2( height - done, RADEON_BUFFER_SIZE / dstPitch );
 221          int src_pitch;
 222          char *tex;
 223
 224          src_pitch = texImage->RowStride * texFormat->TexelBytes;
 225
 226          tex = (char *)texImage->Data + done * src_pitch;
 227
 228          memset(&region, 0, sizeof(region));
 229          r200AllocDmaRegion( rmesa, &region, lines * dstPitch, 1024 );
 230
 231          /* Copy texdata to dma:
 232           */
 233          if (0)
 234             fprintf(stderr, "%s: src_pitch %d dst_pitch %d\n",
 235                     __FUNCTION__, src_pitch, dstPitch);
 236
 237          if (src_pitch == dstPitch) {
 238             memcpy( region.address + region.start, tex, lines * src_pitch );
 239          }
 240          else {
 241             char *buf = region.address + region.start;
 242             int i;
 243             for (i = 0 ; i < lines ; i++) {
 244                memcpy( buf, tex, src_pitch );
 245                buf += dstPitch;
 246                tex += src_pitch;
 247             }
 248          }
 249
 250          r200EmitWait( rmesa, RADEON_WAIT_3D );
 251
 252          /* Blit to framebuffer
 253           */
 254          r200EmitBlit( rmesa,
 255                        blit_format,
 256                        dstPitch, GET_START( &region ),
 257                        dstPitch | (t->tile_bits >> 16),
 258                        t->bufAddr,
 259                        0, 0,
 260                        0, done,
 261                        width, lines );
 262
 263          r200EmitWait( rmesa, RADEON_WAIT_2D );
 264
 265          r200ReleaseDmaRegion( rmesa, &region, __FUNCTION__ );
 266          done += lines;
 267       }
 268    }
 269 }
 270
 271
 272 /**
 273  * Upload the texture image associated with texture \a t at the specified
 274  * level at the address relative to \a start.
 275  */
 276 static void uploadSubImage( r200ContextPtr rmesa, r200TexObjPtr t,
 277                             GLint hwlevel,
 278                             GLint x, GLint y, GLint width, GLint height,
 279                             GLuint face )
 280 {
 281    struct gl_texture_image *texImage = NULL;
 282    GLuint offset;
 283    GLint imageWidth, imageHeight;
 284    GLint ret;
 285    drm_radeon_texture_t tex;
 286    drm_radeon_tex_image_t tmp;
 287    const int level = hwlevel + t->base.firstLevel;
 288
 289    if ( R200_DEBUG & DEBUG_TEXTURE ) {
 290       fprintf( stderr, "%s( %p, %p ) level/width/height/face = %d/%d/%d/%u\n",
 291                __FUNCTION__, (void *)t, (void *)t->base.tObj,
 292                level, width, height, face );
 293    }
 294
 295    ASSERT(face < 6);
 296
 297    /* Ensure we have a valid texture to upload */
 298    if ( ( hwlevel < 0 ) || ( hwlevel >= RADEON_MAX_TEXTURE_LEVELS ) ) {
 299       _mesa_problem(NULL, "bad texture level in %s", __FUNCTION__);
 300       return;
 301    }
 302
 303    texImage = t->base.tObj->Image[face][level];
 304
 305    if ( !texImage ) {
 306       if ( R200_DEBUG & DEBUG_TEXTURE )
 307          fprintf( stderr, "%s: texImage %d is NULL!\n", __FUNCTION__, level );
 308       return;
 309    }
 310    if ( !texImage->Data ) {
 311       if ( R200_DEBUG & DEBUG_TEXTURE )
 312          fprintf( stderr, "%s: image data is NULL!\n", __FUNCTION__ );
 313       return;
 314    }
 315
 316
 317    if (t->base.tObj->Target == GL_TEXTURE_RECTANGLE_NV) {
 318       assert(level == 0);
 319       assert(hwlevel == 0);
 320       if ( R200_DEBUG & DEBUG_TEXTURE )
 321          fprintf( stderr, "%s: image data is rectangular\n", __FUNCTION__);
 322       r200UploadRectSubImage( rmesa, t, texImage, x, y, width, height );
 323       return;
 324    }
 325    else if (texImage->IsClientData) {
 326       if ( R200_DEBUG & DEBUG_TEXTURE )
 327          fprintf( stderr, "%s: image data is in GART client storage\n",
 328                   __FUNCTION__);
 329       r200UploadGARTClientSubImage( rmesa, t, texImage, hwlevel,
 330                                    x, y, width, height );
 331       return;
 332    }
 333    else if ( R200_DEBUG & DEBUG_TEXTURE )
 334       fprintf( stderr, "%s: image data is in normal memory\n",
 335                __FUNCTION__);
 336
 337
 338    imageWidth = texImage->Width;
 339    imageHeight = texImage->Height;
 340
 341    offset = t->bufAddr + t->base.totalSize / 6 * face;
 342
 343    if ( R200_DEBUG & (DEBUG_TEXTURE|DEBUG_IOCTL) ) {
 344       GLint imageX = 0;
 345       GLint imageY = 0;
 346       GLint blitX = t->image[face][hwlevel].x;
 347       GLint blitY = t->image[face][hwlevel].y;
 348       GLint blitWidth = t->image[face][hwlevel].width;
 349       GLint blitHeight = t->image[face][hwlevel].height;
 350       fprintf( stderr, "   upload image: %d,%d at %d,%d\n",
 351                imageWidth, imageHeight, imageX, imageY );
 352       fprintf( stderr, "   upload  blit: %d,%d at %d,%d\n",
 353                blitWidth, blitHeight, blitX, blitY );
 354       fprintf( stderr, "       blit ofs: 0x%07x level: %d/%d\n",
 355                (GLuint)offset, hwlevel, level );
 356    }
 357
 358    t->image[face][hwlevel].data = texImage->Data;
 359
 360    /* Init the DRM_RADEON_TEXTURE command / drm_radeon_texture_t struct.
 361     * NOTE: we're always use a 1KB-wide blit and I8 texture format.
 362     * We used to use 1, 2 and 4-byte texels and used to use the texture
 363     * width to dictate the blit width - but that won't work for compressed
 364     * textures. (Brian)
 365     * NOTE: can't do that with texture tiling. (sroland)
 366     */
 367    tex.offset = offset;
 368    tex.image = &tmp;
 369    /* copy (x,y,width,height,data) */
 370    memcpy( &tmp, &t->image[face][hwlevel], sizeof(tmp) );
 371
 372    if (texImage->TexFormat->TexelBytes) {
 373       /* use multi-byte upload scheme */
 374       tex.height = imageHeight;
 375       tex.width = imageWidth;
 376       tex.format = t->pp_txformat & R200_TXFORMAT_FORMAT_MASK;
 377       tex.pitch = MAX2((texImage->Width * texImage->TexFormat->TexelBytes) / 64, 1);
 378       tex.offset += tmp.x & ~1023;
 379       tmp.x = tmp.x % 1024;
 380       if (t->tile_bits & R200_TXO_MICRO_TILE) {
 381          /* need something like "tiled coordinates" ? */
 382          tmp.y = tmp.x / (tex.pitch * 128) * 2;
 383          tmp.x = tmp.x % (tex.pitch * 128) / 2 / texImage->TexFormat->TexelBytes;
 384          tex.pitch |= RADEON_DST_TILE_MICRO >> 22;
 385       }
 386       else {
 387          tmp.x = tmp.x >> (texImage->TexFormat->TexelBytes >> 1);
 388       }
 389       if ((t->tile_bits & R200_TXO_MACRO_TILE) &&
 390          (texImage->Width * texImage->TexFormat->TexelBytes >= 256) &&
 391          ((!(t->tile_bits & R200_TXO_MICRO_TILE) && (texImage->Height >= 8)) ||
 392             (texImage->Height >= 16))) {
 393          /* weird: R200 disables macro tiling if mip width is smaller than 256 bytes,
 394             OR if height is smaller than 8 automatically, but if micro tiling is active
 395             the limit is height 16 instead ? */
 396          tex.pitch |= RADEON_DST_TILE_MACRO >> 22;
 397       }
 398    }
 399    else {
 400       /* In case of for instance 8x8 texture (2x2 dxt blocks), padding after the first two blocks is
 401          needed (only with dxt1 since 2 dxt3/dxt5 blocks already use 32 Byte). */
 402       /* set tex.height to 1/4 since 1 "macropixel" (dxt-block) has 4 real pixels. Needed
 403          so the kernel module reads the right amount of data. */
 404       tex.format = R200_TXFORMAT_I8; /* any 1-byte texel format */
 405       tex.pitch = (BLIT_WIDTH_BYTES / 64);
 406       tex.height = (imageHeight + 3) / 4;
 407       tex.width = (imageWidth + 3) / 4;
 408       switch (t->pp_txformat & R200_TXFORMAT_FORMAT_MASK) {
 409       case R200_TXFORMAT_DXT1:
 410            tex.width *= 8;
 411            break;
 412       case R200_TXFORMAT_DXT23:
 413       case R200_TXFORMAT_DXT45:
 414            tex.width *= 16;
 415            break;
 416       default:
 417           fprintf(stderr, "unknown compressed tex format in uploadSubImage\n");
 418       }
 419    }
 420
 421    LOCK_HARDWARE( rmesa );
 422    do {
 423       ret = drmCommandWriteRead( rmesa->dri.fd, DRM_RADEON_TEXTURE,
 424                                  &tex, sizeof(drm_radeon_texture_t) );
 425       if (ret) {
 426          if (R200_DEBUG & DEBUG_IOCTL)
 427             fprintf(stderr, "DRM_RADEON_TEXTURE:  again!\n");
 428          usleep(1);
 429       }
 430    } while ( ret && errno == EAGAIN );
 431
 432    UNLOCK_HARDWARE( rmesa );
 433
 434    if ( ret ) {
 435       fprintf( stderr, "DRM_RADEON_TEXTURE: return = %d\n", ret );
 436       fprintf( stderr, "   offset=0x%08x\n",
 437                offset );
 438       fprintf( stderr, "   image width=%d height=%d\n",
 439                imageWidth, imageHeight );
 440       fprintf( stderr, "    blit width=%d height=%d data=%p\n",
 441                t->image[face][hwlevel].width, t->image[face][hwlevel].height,
 442                t->image[face][hwlevel].data );
 443       exit( 1 );
 444    }
 445 }
 446
 447
 448 /**
 449  * Upload the texture images associated with texture \a t.  This might
 450  * require the allocation of texture memory.
 451  *
 452  * \param rmesa Context pointer
 453  * \param t Texture to be uploaded
 454  * \param face Cube map face to be uploaded.  Zero for non-cube maps.
 455  */
 456
 457 int r200UploadTexImages( r200ContextPtr rmesa, r200TexObjPtr t, GLuint face )
 458 {
 459    const int numLevels = t->base.lastLevel - t->base.firstLevel + 1;
 460
 461    if ( R200_DEBUG & (DEBUG_TEXTURE|DEBUG_IOCTL) ) {
 462       fprintf( stderr, "%s( %p, %p ) sz=%d lvls=%d-%d\n", __FUNCTION__,
 463                (void *)rmesa->glCtx, (void *)t->base.tObj, t->base.totalSize,
 464                t->base.firstLevel, t->base.lastLevel );
 465    }
 466
 467    if ( !t || t->base.totalSize == 0 )
 468       return 0;
 469
 470    if (R200_DEBUG & DEBUG_SYNC) {
 471       fprintf(stderr, "%s: Syncing\n", __FUNCTION__ );
 472       r200Finish( rmesa->glCtx );
 473    }
 474
 475    LOCK_HARDWARE( rmesa );
 476
 477    if ( t->base.memBlock == NULL ) {
 478       int heap;
 479
 480       heap = driAllocateTexture( rmesa->texture_heaps, rmesa->nr_heaps,
 481                                  (driTextureObject *) t );
 482       if ( heap == -1 ) {
 483          UNLOCK_HARDWARE( rmesa );
 484          return -1;
 485       }
 486
 487       /* Set the base offset of the texture image */
 488       t->bufAddr = rmesa->r200Screen->texOffset[heap]
 489            + t->base.memBlock->ofs;
 490       t->pp_txoffset = t->bufAddr;
 491
 492       if (!(t->base.tObj->Image[0][0]->IsClientData)) {
 493          /* hope it's safe to add that here... */
 494          t->pp_txoffset |= t->tile_bits;
 495       }
 496
 497       /* Mark this texobj as dirty on all units:
 498        */
 499       t->dirty_state = TEX_ALL;
 500    }
 501
 502    /* Let the world know we've used this memory recently.
 503     */
 504    driUpdateTextureLRU( (driTextureObject *) t );
 505    UNLOCK_HARDWARE( rmesa );
 506
 507    /* Upload any images that are new */
 508    if (t->base.dirty_images[face]) {
 509       int i;
 510       for ( i = 0 ; i < numLevels ; i++ ) {
 511          if ( (t->base.dirty_images[face] & (1 << (i+t->base.firstLevel))) != 0 ) {
 512             uploadSubImage( rmesa, t, i, 0, 0, t->image[face][i].width,
 513                             t->image[face][i].height, face );
 514          }
 515       }
 516       t->base.dirty_images[face] = 0;
 517    }
 518
 519
 520    if (R200_DEBUG & DEBUG_SYNC) {
 521       fprintf(stderr, "%s: Syncing\n", __FUNCTION__ );
 522       r200Finish( rmesa->glCtx );
 523    }
 524
 525    return 0;
 526 }