Merge branch 'gallium-docs'
[mesa.git] / src / mesa / drivers / dri / radeon / radeon_mipmap_tree.c
1 /*
2 * Copyright (C) 2009 Maciej Cencora.
3 * Copyright (C) 2008 Nicolai Haehnle.
4 *
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining
8 * a copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sublicense, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial
17 * portions of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
20 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
22 * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
23 * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
24 * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
25 * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 */
28
29 #include "radeon_mipmap_tree.h"
30
31 #include <errno.h>
32 #include <unistd.h>
33
34 #include "main/simple_list.h"
35 #include "main/texcompress.h"
36 #include "main/teximage.h"
37 #include "main/texobj.h"
38 #include "radeon_texture.h"
39
40 static unsigned get_aligned_compressed_row_stride(
41 gl_format format,
42 unsigned width,
43 unsigned minStride)
44 {
45 const unsigned blockSize = _mesa_get_format_bytes(format);
46 unsigned blockWidth, blockHeight, numXBlocks;
47
48 _mesa_get_format_block_size(format, &blockWidth, &blockHeight);
49 numXBlocks = (width + blockWidth - 1) / blockWidth;
50
51 while (numXBlocks * blockSize < minStride)
52 {
53 ++numXBlocks;
54 }
55
56 return numXBlocks * blockSize;
57 }
58
59 static unsigned get_compressed_image_size(
60 gl_format format,
61 unsigned rowStride,
62 unsigned height)
63 {
64 unsigned blockWidth, blockHeight;
65
66 _mesa_get_format_block_size(format, &blockWidth, &blockHeight);
67
68 return rowStride * ((height + blockHeight - 1) / blockHeight);
69 }
70
71 static int find_next_power_of_two(GLuint value)
72 {
73 int i, tmp;
74
75 i = 0;
76 tmp = value - 1;
77 while (tmp) {
78 tmp >>= 1;
79 i++;
80 }
81 return (1 << i);
82 }
83
84 /**
85 * Compute sizes and fill in offset and blit information for the given
86 * image (determined by \p face and \p level).
87 *
88 * \param curOffset points to the offset at which the image is to be stored
89 * and is updated by this function according to the size of the image.
90 */
91 static void compute_tex_image_offset(radeonContextPtr rmesa, radeon_mipmap_tree *mt,
92 GLuint face, GLuint level, GLuint* curOffset)
93 {
94 radeon_mipmap_level *lvl = &mt->levels[level];
95 uint32_t row_align;
96 GLuint height;
97
98 height = find_next_power_of_two(lvl->height);
99
100 /* Find image size in bytes */
101 if (_mesa_is_format_compressed(mt->mesaFormat)) {
102 lvl->rowstride = get_aligned_compressed_row_stride(mt->mesaFormat, lvl->width, rmesa->texture_compressed_row_align);
103 lvl->size = get_compressed_image_size(mt->mesaFormat, lvl->rowstride, height);
104 } else if (mt->target == GL_TEXTURE_RECTANGLE_NV) {
105 row_align = rmesa->texture_rect_row_align - 1;
106 lvl->rowstride = (_mesa_format_row_stride(mt->mesaFormat, lvl->width) + row_align) & ~row_align;
107 lvl->size = lvl->rowstride * height;
108 } else if (mt->tilebits & RADEON_TXO_MICRO_TILE) {
109 /* tile pattern is 16 bytes x2. mipmaps stay 32 byte aligned,
110 * though the actual offset may be different (if texture is less than
111 * 32 bytes width) to the untiled case */
112 lvl->rowstride = (_mesa_format_row_stride(mt->mesaFormat, lvl->width) * 2 + 31) & ~31;
113 lvl->size = lvl->rowstride * ((height + 1) / 2) * lvl->depth;
114 } else {
115 row_align = rmesa->texture_row_align - 1;
116 lvl->rowstride = (_mesa_format_row_stride(mt->mesaFormat, lvl->width) + row_align) & ~row_align;
117 lvl->size = lvl->rowstride * height * lvl->depth;
118 }
119 assert(lvl->size > 0);
120
121 /* All images are aligned to a 32-byte offset */
122 *curOffset = (*curOffset + 0x1f) & ~0x1f;
123 lvl->faces[face].offset = *curOffset;
124 *curOffset += lvl->size;
125
126 if (RADEON_DEBUG & RADEON_TEXTURE)
127 fprintf(stderr,
128 "level %d, face %d: rs:%d %dx%d at %d\n",
129 level, face, lvl->rowstride, lvl->width, height, lvl->faces[face].offset);
130 }
131
132 static GLuint minify(GLuint size, GLuint levels)
133 {
134 size = size >> levels;
135 if (size < 1)
136 size = 1;
137 return size;
138 }
139
140
141 static void calculate_miptree_layout_r100(radeonContextPtr rmesa, radeon_mipmap_tree *mt)
142 {
143 GLuint curOffset, i, face, level;
144
145 assert(mt->numLevels <= rmesa->glCtx->Const.MaxTextureLevels);
146
147 curOffset = 0;
148 for(face = 0; face < mt->faces; face++) {
149
150 for(i = 0, level = mt->baseLevel; i < mt->numLevels; i++, level++) {
151 mt->levels[level].valid = 1;
152 mt->levels[level].width = minify(mt->width0, i);
153 mt->levels[level].height = minify(mt->height0, i);
154 mt->levels[level].depth = minify(mt->depth0, i);
155 compute_tex_image_offset(rmesa, mt, face, level, &curOffset);
156 }
157 }
158
159 /* Note the required size in memory */
160 mt->totalsize = (curOffset + RADEON_OFFSET_MASK) & ~RADEON_OFFSET_MASK;
161 }
162
163 static void calculate_miptree_layout_r300(radeonContextPtr rmesa, radeon_mipmap_tree *mt)
164 {
165 GLuint curOffset, i, level;
166
167 assert(mt->numLevels <= rmesa->glCtx->Const.MaxTextureLevels);
168
169 curOffset = 0;
170 for(i = 0, level = mt->baseLevel; i < mt->numLevels; i++, level++) {
171 GLuint face;
172
173 mt->levels[level].valid = 1;
174 mt->levels[level].width = minify(mt->width0, i);
175 mt->levels[level].height = minify(mt->height0, i);
176 mt->levels[level].depth = minify(mt->depth0, i);
177
178 for(face = 0; face < mt->faces; face++)
179 compute_tex_image_offset(rmesa, mt, face, level, &curOffset);
180 }
181
182 /* Note the required size in memory */
183 mt->totalsize = (curOffset + RADEON_OFFSET_MASK) & ~RADEON_OFFSET_MASK;
184 }
185
186 /**
187 * Create a new mipmap tree, calculate its layout and allocate memory.
188 */
189 static radeon_mipmap_tree* radeon_miptree_create(radeonContextPtr rmesa,
190 GLenum target, gl_format mesaFormat, GLuint baseLevel, GLuint numLevels,
191 GLuint width0, GLuint height0, GLuint depth0, GLuint tilebits)
192 {
193 radeon_mipmap_tree *mt = CALLOC_STRUCT(_radeon_mipmap_tree);
194
195 mt->mesaFormat = mesaFormat;
196 mt->refcount = 1;
197 mt->target = target;
198 mt->faces = (target == GL_TEXTURE_CUBE_MAP) ? 6 : 1;
199 mt->baseLevel = baseLevel;
200 mt->numLevels = numLevels;
201 mt->width0 = width0;
202 mt->height0 = height0;
203 mt->depth0 = depth0;
204 mt->tilebits = tilebits;
205
206 if (rmesa->radeonScreen->chip_family >= CHIP_FAMILY_R300)
207 calculate_miptree_layout_r300(rmesa, mt);
208 else
209 calculate_miptree_layout_r100(rmesa, mt);
210
211 mt->bo = radeon_bo_open(rmesa->radeonScreen->bom,
212 0, mt->totalsize, 1024,
213 RADEON_GEM_DOMAIN_VRAM,
214 0);
215
216 return mt;
217 }
218
219 void radeon_miptree_reference(radeon_mipmap_tree *mt, radeon_mipmap_tree **ptr)
220 {
221 assert(!*ptr);
222
223 mt->refcount++;
224 assert(mt->refcount > 0);
225
226 *ptr = mt;
227 }
228
229 void radeon_miptree_unreference(radeon_mipmap_tree **ptr)
230 {
231 radeon_mipmap_tree *mt = *ptr;
232 if (!mt)
233 return;
234
235 assert(mt->refcount > 0);
236
237 mt->refcount--;
238 if (!mt->refcount) {
239 radeon_bo_unref(mt->bo);
240 free(mt);
241 }
242
243 *ptr = 0;
244 }
245
246 /**
247 * Calculate min and max LOD for the given texture object.
248 * @param[in] tObj texture object whose LOD values to calculate
249 * @param[out] pminLod minimal LOD
250 * @param[out] pmaxLod maximal LOD
251 */
252 static void calculate_min_max_lod(struct gl_texture_object *tObj,
253 unsigned *pminLod, unsigned *pmaxLod)
254 {
255 int minLod, maxLod;
256 /* Yes, this looks overly complicated, but it's all needed.
257 */
258 switch (tObj->Target) {
259 case GL_TEXTURE_1D:
260 case GL_TEXTURE_2D:
261 case GL_TEXTURE_3D:
262 case GL_TEXTURE_CUBE_MAP:
263 if (tObj->MinFilter == GL_NEAREST || tObj->MinFilter == GL_LINEAR) {
264 /* GL_NEAREST and GL_LINEAR only care about GL_TEXTURE_BASE_LEVEL.
265 */
266 minLod = maxLod = tObj->BaseLevel;
267 } else {
268 minLod = tObj->BaseLevel + (GLint)(tObj->MinLod);
269 minLod = MAX2(minLod, tObj->BaseLevel);
270 minLod = MIN2(minLod, tObj->MaxLevel);
271 maxLod = tObj->BaseLevel + (GLint)(tObj->MaxLod + 0.5);
272 maxLod = MIN2(maxLod, tObj->MaxLevel);
273 maxLod = MIN2(maxLod, tObj->Image[0][minLod]->MaxLog2 + minLod);
274 maxLod = MAX2(maxLod, minLod); /* need at least one level */
275 }
276 break;
277 case GL_TEXTURE_RECTANGLE_NV:
278 case GL_TEXTURE_4D_SGIS:
279 minLod = maxLod = 0;
280 break;
281 default:
282 return;
283 }
284
285 /* save these values */
286 *pminLod = minLod;
287 *pmaxLod = maxLod;
288 }
289
290 /**
291 * Checks whether the given miptree can hold the given texture image at the
292 * given face and level.
293 */
294 GLboolean radeon_miptree_matches_image(radeon_mipmap_tree *mt,
295 struct gl_texture_image *texImage, GLuint face, GLuint level)
296 {
297 radeon_mipmap_level *lvl;
298
299 if (face >= mt->faces)
300 return GL_FALSE;
301
302 if (texImage->TexFormat != mt->mesaFormat)
303 return GL_FALSE;
304
305 lvl = &mt->levels[level];
306 if (!lvl->valid ||
307 lvl->width != texImage->Width ||
308 lvl->height != texImage->Height ||
309 lvl->depth != texImage->Depth)
310 return GL_FALSE;
311
312 return GL_TRUE;
313 }
314
315 /**
316 * Checks whether the given miptree has the right format to store the given texture object.
317 */
318 static GLboolean radeon_miptree_matches_texture(radeon_mipmap_tree *mt, struct gl_texture_object *texObj)
319 {
320 struct gl_texture_image *firstImage;
321 unsigned numLevels;
322 radeon_mipmap_level *mtBaseLevel;
323
324 if (texObj->BaseLevel < mt->baseLevel)
325 return GL_FALSE;
326
327 mtBaseLevel = &mt->levels[texObj->BaseLevel - mt->baseLevel];
328 firstImage = texObj->Image[0][texObj->BaseLevel];
329 numLevels = MIN2(texObj->MaxLevel - texObj->BaseLevel + 1, firstImage->MaxLog2 + 1);
330
331 if (RADEON_DEBUG & RADEON_TEXTURE) {
332 fprintf(stderr, "Checking if miptree %p matches texObj %p\n", mt, texObj);
333 fprintf(stderr, "target %d vs %d\n", mt->target, texObj->Target);
334 fprintf(stderr, "format %d vs %d\n", mt->mesaFormat, firstImage->TexFormat);
335 fprintf(stderr, "numLevels %d vs %d\n", mt->numLevels, numLevels);
336 fprintf(stderr, "width0 %d vs %d\n", mtBaseLevel->width, firstImage->Width);
337 fprintf(stderr, "height0 %d vs %d\n", mtBaseLevel->height, firstImage->Height);
338 fprintf(stderr, "depth0 %d vs %d\n", mtBaseLevel->depth, firstImage->Depth);
339 if (mt->target == texObj->Target &&
340 mt->mesaFormat == firstImage->TexFormat &&
341 mt->numLevels >= numLevels &&
342 mtBaseLevel->width == firstImage->Width &&
343 mtBaseLevel->height == firstImage->Height &&
344 mtBaseLevel->depth == firstImage->Depth) {
345 fprintf(stderr, "MATCHED\n");
346 } else {
347 fprintf(stderr, "NOT MATCHED\n");
348 }
349 }
350
351 return (mt->target == texObj->Target &&
352 mt->mesaFormat == firstImage->TexFormat &&
353 mt->numLevels >= numLevels &&
354 mtBaseLevel->width == firstImage->Width &&
355 mtBaseLevel->height == firstImage->Height &&
356 mtBaseLevel->depth == firstImage->Depth);
357 }
358
359 /**
360 * Try to allocate a mipmap tree for the given texture object.
361 * @param[in] rmesa radeon context
362 * @param[in] t radeon texture object
363 */
364 void radeon_try_alloc_miptree(radeonContextPtr rmesa, radeonTexObj *t)
365 {
366 struct gl_texture_object *texObj = &t->base;
367 struct gl_texture_image *texImg = texObj->Image[0][texObj->BaseLevel];
368 GLuint numLevels;
369
370 assert(!t->mt);
371
372 if (!texImg)
373 return;
374
375 numLevels = MIN2(texObj->MaxLevel - texObj->BaseLevel + 1, texImg->MaxLog2 + 1);
376
377 t->mt = radeon_miptree_create(rmesa, t->base.Target,
378 texImg->TexFormat, texObj->BaseLevel,
379 numLevels, texImg->Width, texImg->Height,
380 texImg->Depth, t->tile_bits);
381 }
382
383 /* Although we use the image_offset[] array to store relative offsets
384 * to cube faces, Mesa doesn't know anything about this and expects
385 * each cube face to be treated as a separate image.
386 *
387 * These functions present that view to mesa:
388 */
389 void
390 radeon_miptree_depth_offsets(radeon_mipmap_tree *mt, GLuint level, GLuint *offsets)
391 {
392 if (mt->target != GL_TEXTURE_3D || mt->faces == 1) {
393 offsets[0] = 0;
394 } else {
395 int i;
396 for (i = 0; i < 6; i++) {
397 offsets[i] = mt->levels[level].faces[i].offset;
398 }
399 }
400 }
401
402 GLuint
403 radeon_miptree_image_offset(radeon_mipmap_tree *mt,
404 GLuint face, GLuint level)
405 {
406 if (mt->target == GL_TEXTURE_CUBE_MAP_ARB)
407 return (mt->levels[level].faces[face].offset);
408 else
409 return mt->levels[level].faces[0].offset;
410 }
411
412 /**
413 * Ensure that the given image is stored in the given miptree from now on.
414 */
415 static void migrate_image_to_miptree(radeon_mipmap_tree *mt,
416 radeon_texture_image *image,
417 int face, int level)
418 {
419 radeon_mipmap_level *dstlvl = &mt->levels[level];
420 unsigned char *dest;
421
422 assert(image->mt != mt);
423 assert(dstlvl->valid);
424 assert(dstlvl->width == image->base.Width);
425 assert(dstlvl->height == image->base.Height);
426 assert(dstlvl->depth == image->base.Depth);
427
428 radeon_bo_map(mt->bo, GL_TRUE);
429 dest = mt->bo->ptr + dstlvl->faces[face].offset;
430
431 if (image->mt) {
432 /* Format etc. should match, so we really just need a memcpy().
433 * In fact, that memcpy() could be done by the hardware in many
434 * cases, provided that we have a proper memory manager.
435 */
436 assert(mt->mesaFormat == image->base.TexFormat);
437
438 radeon_mipmap_level *srclvl = &image->mt->levels[image->mtlevel];
439
440 /* TODO: bring back these assertions once the FBOs are fixed */
441 #if 0
442 assert(image->mtlevel == level);
443 assert(srclvl->size == dstlvl->size);
444 assert(srclvl->rowstride == dstlvl->rowstride);
445 #endif
446
447 radeon_bo_map(image->mt->bo, GL_FALSE);
448
449 memcpy(dest,
450 image->mt->bo->ptr + srclvl->faces[face].offset,
451 dstlvl->size);
452 radeon_bo_unmap(image->mt->bo);
453
454 radeon_miptree_unreference(&image->mt);
455 } else if (image->base.Data) {
456 /* This condition should be removed, it's here to workaround
457 * a segfault when mapping textures during software fallbacks.
458 */
459 const uint32_t srcrowstride = _mesa_format_row_stride(image->base.TexFormat, image->base.Width);
460 uint32_t rows = image->base.Height * image->base.Depth;
461
462 if (_mesa_is_format_compressed(image->base.TexFormat)) {
463 uint32_t blockWidth, blockHeight;
464 _mesa_get_format_block_size(image->base.TexFormat, &blockWidth, &blockHeight);
465 rows = (rows + blockHeight - 1) / blockHeight;
466 }
467
468 copy_rows(dest, dstlvl->rowstride, image->base.Data, srcrowstride,
469 rows, srcrowstride);
470
471 _mesa_free_texmemory(image->base.Data);
472 image->base.Data = 0;
473 }
474
475 radeon_bo_unmap(mt->bo);
476
477 radeon_miptree_reference(mt, &image->mt);
478 image->mtface = face;
479 image->mtlevel = level;
480 }
481
482 /**
483 * Filter matching miptrees, and select one with the most of data.
484 * @param[in] texObj radeon texture object
485 * @param[in] firstLevel first texture level to check
486 * @param[in] lastLevel last texture level to check
487 */
488 static radeon_mipmap_tree * get_biggest_matching_miptree(radeonTexObj *texObj,
489 unsigned firstLevel,
490 unsigned lastLevel)
491 {
492 const unsigned numLevels = lastLevel - firstLevel + 1;
493 unsigned *mtSizes = calloc(numLevels, sizeof(unsigned));
494 radeon_mipmap_tree **mts = calloc(numLevels, sizeof(radeon_mipmap_tree *));
495 unsigned mtCount = 0;
496 unsigned maxMtIndex = 0;
497 radeon_mipmap_tree *tmp;
498
499 for (unsigned level = firstLevel; level <= lastLevel; ++level) {
500 radeon_texture_image *img = get_radeon_texture_image(texObj->base.Image[0][level]);
501 unsigned found = 0;
502 // TODO: why this hack??
503 if (!img)
504 break;
505
506 if (!img->mt)
507 continue;
508
509 for (int i = 0; i < mtCount; ++i) {
510 if (mts[i] == img->mt) {
511 found = 1;
512 mtSizes[i] += img->mt->levels[img->mtlevel].size;
513 break;
514 }
515 }
516
517 if (!found && radeon_miptree_matches_texture(img->mt, &texObj->base)) {
518 mtSizes[mtCount] = img->mt->levels[img->mtlevel].size;
519 mts[mtCount] = img->mt;
520 mtCount++;
521 }
522 }
523
524 if (mtCount == 0) {
525 return NULL;
526 }
527
528 for (int i = 1; i < mtCount; ++i) {
529 if (mtSizes[i] > mtSizes[maxMtIndex]) {
530 maxMtIndex = i;
531 }
532 }
533
534 tmp = mts[maxMtIndex];
535 free(mtSizes);
536 free(mts);
537
538 return tmp;
539 }
540
541 /**
542 * Validate texture mipmap tree.
543 * If individual images are stored in different mipmap trees
544 * use the mipmap tree that has the most of the correct data.
545 */
546 int radeon_validate_texture_miptree(GLcontext * ctx, struct gl_texture_object *texObj)
547 {
548 radeonContextPtr rmesa = RADEON_CONTEXT(ctx);
549 radeonTexObj *t = radeon_tex_obj(texObj);
550
551 if (t->validated || t->image_override) {
552 return GL_TRUE;
553 }
554
555 if (texObj->Image[0][texObj->BaseLevel]->Border > 0)
556 return GL_FALSE;
557
558 _mesa_test_texobj_completeness(rmesa->glCtx, texObj);
559 if (!texObj->_Complete) {
560 return GL_FALSE;
561 }
562
563 calculate_min_max_lod(&t->base, &t->minLod, &t->maxLod);
564
565 if (RADEON_DEBUG & RADEON_TEXTURE)
566 fprintf(stderr, "%s: Validating texture %p now, minLod = %d, maxLod = %d\n",
567 __FUNCTION__, texObj ,t->minLod, t->maxLod);
568
569 radeon_mipmap_tree *dst_miptree;
570 dst_miptree = get_biggest_matching_miptree(t, t->minLod, t->maxLod);
571
572 if (!dst_miptree) {
573 radeon_miptree_unreference(&t->mt);
574 radeon_try_alloc_miptree(rmesa, t);
575 dst_miptree = t->mt;
576 if (RADEON_DEBUG & RADEON_TEXTURE) {
577 fprintf(stderr, "%s: No matching miptree found, allocated new one %p\n", __FUNCTION__, t->mt);
578 }
579 } else if (RADEON_DEBUG & RADEON_TEXTURE) {
580 fprintf(stderr, "%s: Using miptree %p\n", __FUNCTION__, t->mt);
581 }
582
583 const unsigned faces = texObj->Target == GL_TEXTURE_CUBE_MAP ? 6 : 1;
584 unsigned face, level;
585 radeon_texture_image *img;
586 /* Validate only the levels that will actually be used during rendering */
587 for (face = 0; face < faces; ++face) {
588 for (level = t->minLod; level <= t->maxLod; ++level) {
589 img = get_radeon_texture_image(texObj->Image[face][level]);
590
591 if (RADEON_DEBUG & RADEON_TEXTURE) {
592 fprintf(stderr, "Checking image level %d, face %d, mt %p ... ", level, face, img->mt);
593 }
594
595 if (img->mt != dst_miptree) {
596 if (RADEON_DEBUG & RADEON_TEXTURE) {
597 fprintf(stderr, "MIGRATING\n");
598 }
599 struct radeon_bo *src_bo = (img->mt) ? img->mt->bo : img->bo;
600 if (src_bo && radeon_bo_is_referenced_by_cs(src_bo, rmesa->cmdbuf.cs)) {
601 radeon_firevertices(rmesa);
602 }
603 migrate_image_to_miptree(dst_miptree, img, face, level);
604 } else if (RADEON_DEBUG & RADEON_TEXTURE) {
605 fprintf(stderr, "OK\n");
606 }
607 }
608 }
609
610 t->validated = GL_TRUE;
611
612 return GL_TRUE;
613 }
614
615 uint32_t get_base_teximage_offset(radeonTexObj *texObj)
616 {
617 if (!texObj->mt) {
618 return 0;
619 } else {
620 return radeon_miptree_image_offset(texObj->mt, 0, texObj->minLod);
621 }
622 }