1 /*
2 * Copyright (C) 2019-2022 Collabora, Ltd.
3 * Copyright (C) 2018-2019 Alyssa Rosenzweig
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 *
24 */
25
26 #include "util/log.h"
27 #include "util/macros.h"
28 #include "util/u_math.h"
29 #include "pan_texture.h"
30
31 /*
32 * List of supported modifiers, in descending order of preference. AFBC is
33 * faster than u-interleaved tiling which is faster than linear. Within AFBC,
34 * enabling the YUV-like transform is typically a win where possible.
35 * AFRC is only used if explicitely asked for (only for RGB formats).
36 */
37 uint64_t pan_best_modifiers[PAN_MODIFIER_COUNT] = {
38 DRM_FORMAT_MOD_ARM_AFBC(AFBC_FORMAT_MOD_BLOCK_SIZE_32x8 |
39 AFBC_FORMAT_MOD_SPARSE | AFBC_FORMAT_MOD_SPLIT),
40 DRM_FORMAT_MOD_ARM_AFBC(AFBC_FORMAT_MOD_BLOCK_SIZE_32x8 |
41 AFBC_FORMAT_MOD_SPARSE | AFBC_FORMAT_MOD_SPLIT |
42 AFBC_FORMAT_MOD_YTR),
43
44 DRM_FORMAT_MOD_ARM_AFBC(AFBC_FORMAT_MOD_BLOCK_SIZE_16x16 |
45 AFBC_FORMAT_MOD_TILED | AFBC_FORMAT_MOD_SC |
46 AFBC_FORMAT_MOD_SPARSE | AFBC_FORMAT_MOD_YTR),
47
48 DRM_FORMAT_MOD_ARM_AFBC(AFBC_FORMAT_MOD_BLOCK_SIZE_16x16 |
49 AFBC_FORMAT_MOD_TILED | AFBC_FORMAT_MOD_SC |
50 AFBC_FORMAT_MOD_SPARSE),
51
52 DRM_FORMAT_MOD_ARM_AFBC(AFBC_FORMAT_MOD_BLOCK_SIZE_16x16 |
53 AFBC_FORMAT_MOD_SPARSE | AFBC_FORMAT_MOD_YTR),
54
55 DRM_FORMAT_MOD_ARM_AFBC(AFBC_FORMAT_MOD_BLOCK_SIZE_16x16 |
56 AFBC_FORMAT_MOD_SPARSE),
57
58 DRM_FORMAT_MOD_ARM_16X16_BLOCK_U_INTERLEAVED,
59 DRM_FORMAT_MOD_LINEAR,
60
61 DRM_FORMAT_MOD_ARM_AFRC(
62 AFRC_FORMAT_MOD_CU_SIZE_P0(AFRC_FORMAT_MOD_CU_SIZE_16)),
63 DRM_FORMAT_MOD_ARM_AFRC(
64 AFRC_FORMAT_MOD_CU_SIZE_P0(AFRC_FORMAT_MOD_CU_SIZE_24)),
65 DRM_FORMAT_MOD_ARM_AFRC(
66 AFRC_FORMAT_MOD_CU_SIZE_P0(AFRC_FORMAT_MOD_CU_SIZE_32)),
67 DRM_FORMAT_MOD_ARM_AFRC(
68 AFRC_FORMAT_MOD_CU_SIZE_P0(AFRC_FORMAT_MOD_CU_SIZE_16) |
69 AFRC_FORMAT_MOD_LAYOUT_SCAN),
70 DRM_FORMAT_MOD_ARM_AFRC(
71 AFRC_FORMAT_MOD_CU_SIZE_P0(AFRC_FORMAT_MOD_CU_SIZE_24) |
72 AFRC_FORMAT_MOD_LAYOUT_SCAN),
73 DRM_FORMAT_MOD_ARM_AFRC(
74 AFRC_FORMAT_MOD_CU_SIZE_P0(AFRC_FORMAT_MOD_CU_SIZE_32) |
75 AFRC_FORMAT_MOD_LAYOUT_SCAN),
76 };
77
78 /* Table of AFBC superblock sizes */
79 static const struct pan_block_size afbc_superblock_sizes[] = {
80 [AFBC_FORMAT_MOD_BLOCK_SIZE_16x16] = {16, 16},
81 [AFBC_FORMAT_MOD_BLOCK_SIZE_32x8] = {32, 8},
82 [AFBC_FORMAT_MOD_BLOCK_SIZE_64x4] = {64, 4},
83 };
84
85 /*
86 * Given an AFBC modifier, return the superblock size.
87 *
88 * We do not yet have any use cases for multiplanar YCBCr formats with different
89 * superblock sizes on the luma and chroma planes. These formats are unsupported
90 * for now.
91 */
92 struct pan_block_size
panfrost_afbc_superblock_size(uint64_t modifier)93 panfrost_afbc_superblock_size(uint64_t modifier)
94 {
95 unsigned index = (modifier & AFBC_FORMAT_MOD_BLOCK_SIZE_MASK);
96
97 assert(drm_is_afbc(modifier));
98 assert(index < ARRAY_SIZE(afbc_superblock_sizes));
99
100 return afbc_superblock_sizes[index];
101 }
102
103 /*
104 * Given an AFBC modifier, return the render size.
105 */
106 struct pan_block_size
panfrost_afbc_renderblock_size(uint64_t modifier)107 panfrost_afbc_renderblock_size(uint64_t modifier)
108 {
109 unsigned index = (modifier & AFBC_FORMAT_MOD_BLOCK_SIZE_MASK);
110
111 assert(drm_is_afbc(modifier));
112 assert(index < ARRAY_SIZE(afbc_superblock_sizes));
113
114 struct pan_block_size blk_size = afbc_superblock_sizes[index];
115
116 /* The GPU needs to render 16x16 tiles. For wide tiles, that means we
117 * have to extend the render region to have a height of 16 pixels.
118 */
119 blk_size.height = ALIGN_POT(blk_size.height, 16);
120 return blk_size;
121 }
122
123 /*
124 * Given an AFBC modifier, return the width of the superblock.
125 */
126 unsigned
panfrost_afbc_superblock_width(uint64_t modifier)127 panfrost_afbc_superblock_width(uint64_t modifier)
128 {
129 return panfrost_afbc_superblock_size(modifier).width;
130 }
131
132 /*
133 * Given an AFBC modifier, return the height of the superblock.
134 */
135 unsigned
panfrost_afbc_superblock_height(uint64_t modifier)136 panfrost_afbc_superblock_height(uint64_t modifier)
137 {
138 return panfrost_afbc_superblock_size(modifier).height;
139 }
140
141 /*
142 * Given an AFBC modifier, return if "wide blocks" are used. Wide blocks are
143 * defined as superblocks wider than 16 pixels, the minimum (and default) super
144 * block width.
145 */
146 bool
panfrost_afbc_is_wide(uint64_t modifier)147 panfrost_afbc_is_wide(uint64_t modifier)
148 {
149 return panfrost_afbc_superblock_width(modifier) > 16;
150 }
151
152 /*
153 * Given an AFBC modifier, return the subblock size (subdivision of a
154 * superblock). This is always 4x4 for now as we only support one AFBC
155 * superblock layout.
156 */
157 struct pan_block_size
panfrost_afbc_subblock_size(uint64_t modifier)158 panfrost_afbc_subblock_size(uint64_t modifier)
159 {
160 return (struct pan_block_size){4, 4};
161 }
162
163 /*
164 * Given an AFRC modifier, return whether the layout is optimized for scan
165 * order (vs rotation order).
166 */
167 bool
panfrost_afrc_is_scan(uint64_t modifier)168 panfrost_afrc_is_scan(uint64_t modifier)
169 {
170 return modifier & AFRC_FORMAT_MOD_LAYOUT_SCAN;
171 }
172
173 struct pan_block_size
panfrost_afrc_clump_size(enum pipe_format format,bool scan)174 panfrost_afrc_clump_size(enum pipe_format format, bool scan)
175 {
176 struct pan_afrc_format_info finfo = panfrost_afrc_get_format_info(format);
177
178 switch (finfo.num_comps) {
179 case 1:
180 return scan ? (struct pan_block_size){16, 4}
181 : (struct pan_block_size){8, 8};
182 case 2:
183 return (struct pan_block_size){8, 4};
184 case 3:
185 case 4:
186 return (struct pan_block_size){4, 4};
187 default:
188 assert(0);
189 return (struct pan_block_size){0, 0};
190 }
191 }
192
193 static struct pan_block_size
panfrost_afrc_layout_size(uint64_t modifier)194 panfrost_afrc_layout_size(uint64_t modifier)
195 {
196 if (panfrost_afrc_is_scan(modifier))
197 return (struct pan_block_size){16, 4};
198 else
199 return (struct pan_block_size){8, 8};
200 }
201
202 struct pan_block_size
panfrost_afrc_tile_size(enum pipe_format format,uint64_t modifier)203 panfrost_afrc_tile_size(enum pipe_format format, uint64_t modifier)
204 {
205 bool scan = panfrost_afrc_is_scan(modifier);
206 struct pan_block_size clump_sz = panfrost_afrc_clump_size(format, scan);
207 struct pan_block_size layout_sz = panfrost_afrc_layout_size(modifier);
208
209 return (struct pan_block_size){clump_sz.width * layout_sz.width,
210 clump_sz.height * layout_sz.height};
211 }
212
213 unsigned
panfrost_afrc_block_size_from_modifier(uint64_t modifier)214 panfrost_afrc_block_size_from_modifier(uint64_t modifier)
215 {
216 switch (modifier & AFRC_FORMAT_MOD_CU_SIZE_MASK) {
217 case AFRC_FORMAT_MOD_CU_SIZE_16:
218 return 16;
219 case AFRC_FORMAT_MOD_CU_SIZE_24:
220 return 24;
221 case AFRC_FORMAT_MOD_CU_SIZE_32:
222 return 32;
223 default:
224 unreachable("invalid coding unit size flag in modifier");
225 };
226 }
227
228 static unsigned
panfrost_afrc_buffer_alignment_from_modifier(uint64_t modifier)229 panfrost_afrc_buffer_alignment_from_modifier(uint64_t modifier)
230 {
231 switch (modifier & AFRC_FORMAT_MOD_CU_SIZE_MASK) {
232 case AFRC_FORMAT_MOD_CU_SIZE_16:
233 return 1024;
234 case AFRC_FORMAT_MOD_CU_SIZE_24:
235 return 512;
236 case AFRC_FORMAT_MOD_CU_SIZE_32:
237 return 2048;
238 default:
239 unreachable("invalid coding unit size flag in modifier");
240 };
241 }
242
243 /*
244 * Determine the number of bytes between rows of paging tiles in an AFRC image
245 */
246 uint32_t
pan_afrc_row_stride(enum pipe_format format,uint64_t modifier,uint32_t width)247 pan_afrc_row_stride(enum pipe_format format, uint64_t modifier, uint32_t width)
248 {
249 struct pan_block_size tile_size = panfrost_afrc_tile_size(format, modifier);
250 unsigned block_size = panfrost_afrc_block_size_from_modifier(modifier);
251
252 return (width / tile_size.width) * block_size * AFRC_CLUMPS_PER_TILE;
253 }
254
255 /*
256 * Given a format, determine the tile size used for u-interleaving. For formats
257 * that are already block compressed, this is 4x4. For all other formats, this
258 * is 16x16, hence the modifier name.
259 */
260 static inline struct pan_block_size
panfrost_u_interleaved_tile_size(enum pipe_format format)261 panfrost_u_interleaved_tile_size(enum pipe_format format)
262 {
263 if (util_format_is_compressed(format))
264 return (struct pan_block_size){4, 4};
265 else
266 return (struct pan_block_size){16, 16};
267 }
268
269 /*
270 * Determine the block size used for interleaving. For u-interleaving, this is
271 * the tile size. For AFBC, this is the superblock size. For AFRC, this is the
272 * paging tile size. For linear textures, this is trivially 1x1.
273 */
274 struct pan_block_size
panfrost_block_size(uint64_t modifier,enum pipe_format format)275 panfrost_block_size(uint64_t modifier, enum pipe_format format)
276 {
277 if (modifier == DRM_FORMAT_MOD_ARM_16X16_BLOCK_U_INTERLEAVED)
278 return panfrost_u_interleaved_tile_size(format);
279 else if (drm_is_afbc(modifier))
280 return panfrost_afbc_superblock_size(modifier);
281 else if (drm_is_afrc(modifier))
282 return panfrost_afrc_tile_size(format, modifier);
283 else
284 return (struct pan_block_size){1, 1};
285 }
286
287 /* For non-AFBC and non-wide AFBC, the render block size matches
288 * the block size, but for wide AFBC, the GPU wants the block height
289 * to be 16 pixels high.
290 */
291 struct pan_block_size
panfrost_renderblock_size(uint64_t modifier,enum pipe_format format)292 panfrost_renderblock_size(uint64_t modifier, enum pipe_format format)
293 {
294 if (!drm_is_afbc(modifier))
295 return panfrost_block_size(modifier, format);
296
297 return panfrost_afbc_renderblock_size(modifier);
298 }
299
300 /*
301 * Determine the tile size used by AFBC. This tiles superblocks themselves.
302 * Current GPUs support either 8x8 tiling or no tiling (1x1)
303 */
304 static inline unsigned
pan_afbc_tile_size(uint64_t modifier)305 pan_afbc_tile_size(uint64_t modifier)
306 {
307 return (modifier & AFBC_FORMAT_MOD_TILED) ? 8 : 1;
308 }
309
310 /*
311 * Determine the number of bytes between header rows for an AFBC image. For an
312 * image with linear headers, this is simply the number of header blocks
313 * (=superblocks) per row times the numbers of bytes per header block. For an
314 * image with tiled headers, this is multipled by the number of rows of
315 * header blocks are in a tile together.
316 */
317 uint32_t
pan_afbc_row_stride(uint64_t modifier,uint32_t width)318 pan_afbc_row_stride(uint64_t modifier, uint32_t width)
319 {
320 unsigned block_width = panfrost_afbc_superblock_width(modifier);
321
322 return (width / block_width) * pan_afbc_tile_size(modifier) *
323 AFBC_HEADER_BYTES_PER_TILE;
324 }
325
326 /*
327 * Determine the number of header blocks between header rows. This is equal to
328 * the number of bytes between header rows divided by the bytes per blocks of a
329 * header tile. This is also divided by the tile size to give a "line stride" in
330 * blocks, rather than a real row stride. This is required by Bifrost.
331 */
332 uint32_t
pan_afbc_stride_blocks(uint64_t modifier,uint32_t row_stride_bytes)333 pan_afbc_stride_blocks(uint64_t modifier, uint32_t row_stride_bytes)
334 {
335 return row_stride_bytes /
336 (AFBC_HEADER_BYTES_PER_TILE * pan_afbc_tile_size(modifier));
337 }
338
339 /*
340 * Determine the required alignment for the slice offset of an image. For
341 * now, this is always aligned on 64-byte boundaries. */
342 uint32_t
pan_slice_align(uint64_t modifier)343 pan_slice_align(uint64_t modifier)
344 {
345 return 64;
346 }
347
348 /*
349 * Determine the required alignment for the body offset of an AFBC image. For
350 * now, this depends only on whether tiling is in use. These minimum alignments
351 * are required on all current GPUs.
352 */
353 uint32_t
pan_afbc_body_align(unsigned arch,uint64_t modifier)354 pan_afbc_body_align(unsigned arch, uint64_t modifier)
355 {
356 if (modifier & AFBC_FORMAT_MOD_TILED)
357 return 4096;
358
359 if (arch >= 6)
360 return 128;
361
362 return 64;
363 }
364
365 static inline unsigned
format_minimum_alignment(unsigned arch,enum pipe_format format,uint64_t mod)366 format_minimum_alignment(unsigned arch, enum pipe_format format, uint64_t mod)
367 {
368 if (drm_is_afbc(mod))
369 return 16;
370
371 if (drm_is_afrc(mod))
372 return panfrost_afrc_buffer_alignment_from_modifier(mod);
373
374 if (arch < 7)
375 return 64;
376
377 switch (format) {
378 /* For v7+, NV12/NV21/I420 have a looser alignment requirement of 16 bytes */
379 case PIPE_FORMAT_R8_G8B8_420_UNORM:
380 case PIPE_FORMAT_G8_B8R8_420_UNORM:
381 case PIPE_FORMAT_R8_G8_B8_420_UNORM:
382 case PIPE_FORMAT_R8_B8_G8_420_UNORM:
383 case PIPE_FORMAT_R8_G8B8_422_UNORM:
384 case PIPE_FORMAT_R8_B8G8_422_UNORM:
385 return 16;
386 /* the 10 bit formats have even looser alignment */
387 case PIPE_FORMAT_R10_G10B10_420_UNORM:
388 case PIPE_FORMAT_R10_G10B10_422_UNORM:
389 return 1;
390 default:
391 return 64;
392 }
393 }
394
395 /*
396 * Computes sizes for checksumming, which is 8 bytes per 16x16 tile.
397 * Checksumming is believed to be a CRC variant (CRC64 based on the size?).
398 * This feature is also known as "transaction elimination".
399 * CRC values are prefetched by 32x32 regions so size needs to be aligned.
400 */
401
402 #define CHECKSUM_TILE_WIDTH 16
403 #define CHECKSUM_TILE_HEIGHT 16
404 #define CHECKSUM_REGION_SIZE 32
405 #define CHECKSUM_X_TILE_PER_REGION (CHECKSUM_REGION_SIZE / CHECKSUM_TILE_WIDTH)
406 #define CHECKSUM_Y_TILE_PER_REGION (CHECKSUM_REGION_SIZE / CHECKSUM_TILE_HEIGHT)
407 #define CHECKSUM_BYTES_PER_TILE 8
408
409 unsigned
panfrost_compute_checksum_size(struct pan_image_slice_layout * slice,unsigned width,unsigned height)410 panfrost_compute_checksum_size(struct pan_image_slice_layout *slice,
411 unsigned width, unsigned height)
412 {
413 unsigned tile_count_x =
414 CHECKSUM_X_TILE_PER_REGION * DIV_ROUND_UP(width, CHECKSUM_REGION_SIZE);
415 unsigned tile_count_y =
416 CHECKSUM_Y_TILE_PER_REGION * DIV_ROUND_UP(height, CHECKSUM_REGION_SIZE);
417
418 slice->crc.stride = tile_count_x * CHECKSUM_BYTES_PER_TILE;
419
420 return slice->crc.stride * tile_count_y;
421 }
422
423 unsigned
panfrost_get_layer_stride(const struct pan_image_layout * layout,unsigned level)424 panfrost_get_layer_stride(const struct pan_image_layout *layout, unsigned level)
425 {
426 if (layout->dim != MALI_TEXTURE_DIMENSION_3D)
427 return layout->array_stride;
428 else if (drm_is_afbc(layout->modifier))
429 return layout->slices[level].afbc.surface_stride;
430 else
431 return layout->slices[level].surface_stride;
432 }
433
434 unsigned
panfrost_get_legacy_stride(const struct pan_image_layout * layout,unsigned level)435 panfrost_get_legacy_stride(const struct pan_image_layout *layout,
436 unsigned level)
437 {
438 unsigned row_stride = layout->slices[level].row_stride;
439 struct pan_block_size block_size =
440 panfrost_renderblock_size(layout->modifier, layout->format);
441
442 if (drm_is_afbc(layout->modifier)) {
443 unsigned width = u_minify(layout->width, level);
444 unsigned alignment =
445 block_size.width * pan_afbc_tile_size(layout->modifier);
446
447 width = ALIGN_POT(width, alignment);
448 return width * util_format_get_blocksize(layout->format);
449 } else if (drm_is_afrc(layout->modifier)) {
450 struct pan_block_size tile_size =
451 panfrost_afrc_tile_size(layout->format, layout->modifier);
452
453 return row_stride / tile_size.height;
454 } else {
455 return row_stride / block_size.height;
456 }
457 }
458
459 unsigned
panfrost_from_legacy_stride(unsigned legacy_stride,enum pipe_format format,uint64_t modifier)460 panfrost_from_legacy_stride(unsigned legacy_stride, enum pipe_format format,
461 uint64_t modifier)
462 {
463 struct pan_block_size block_size =
464 panfrost_renderblock_size(modifier, format);
465
466 if (drm_is_afbc(modifier)) {
467 unsigned width = legacy_stride / util_format_get_blocksize(format);
468
469 return pan_afbc_row_stride(modifier, width);
470 } else if (drm_is_afrc(modifier)) {
471 struct pan_block_size tile_size =
472 panfrost_afrc_tile_size(format, modifier);
473
474 return legacy_stride * tile_size.height;
475 } else {
476 return legacy_stride * block_size.height;
477 }
478 }
479
480 /* Computes the offset into a texture at a particular level/face. Add to
481 * the base address of a texture to get the address to that level/face */
482
483 unsigned
panfrost_texture_offset(const struct pan_image_layout * layout,unsigned level,unsigned array_idx,unsigned surface_idx)484 panfrost_texture_offset(const struct pan_image_layout *layout, unsigned level,
485 unsigned array_idx, unsigned surface_idx)
486 {
487 return layout->slices[level].offset + (array_idx * layout->array_stride) +
488 (surface_idx * layout->slices[level].surface_stride);
489 }
490
491 bool
pan_image_layout_init(unsigned arch,struct pan_image_layout * layout,const struct pan_image_explicit_layout * explicit_layout)492 pan_image_layout_init(unsigned arch, struct pan_image_layout *layout,
493 const struct pan_image_explicit_layout *explicit_layout)
494 {
495 /* Explicit stride only work with non-mipmap, non-array, single-sample
496 * 2D image without CRC.
497 */
498 if (explicit_layout &&
499 (layout->depth > 1 || layout->nr_samples > 1 || layout->array_size > 1 ||
500 layout->dim != MALI_TEXTURE_DIMENSION_2D || layout->nr_slices > 1 ||
501 layout->crc))
502 return false;
503
504 bool afbc = drm_is_afbc(layout->modifier);
505 bool afrc = drm_is_afrc(layout->modifier);
506 int align_req =
507 format_minimum_alignment(arch, layout->format, layout->modifier);
508
509 /* Mandate alignment */
510 if (explicit_layout) {
511 bool rejected = false;
512
513 int align_mask = align_req - 1;
514
515 if (arch >= 7) {
516 rejected = ((explicit_layout->offset & align_mask) ||
517 (explicit_layout->row_stride & align_mask));
518 } else {
519 rejected = (explicit_layout->offset & align_mask);
520 }
521
522 if (rejected) {
523 mesa_loge(
524 "panfrost: rejecting image due to unsupported offset or stride "
525 "alignment.\n");
526 return false;
527 }
528 }
529
530 unsigned fmt_blocksize = util_format_get_blocksize(layout->format);
531
532 /* MSAA is implemented as a 3D texture with z corresponding to the
533 * sample #, horrifyingly enough */
534
535 assert(layout->depth == 1 || layout->nr_samples == 1);
536
537 bool linear = layout->modifier == DRM_FORMAT_MOD_LINEAR;
538 bool is_3d = layout->dim == MALI_TEXTURE_DIMENSION_3D;
539
540 uint64_t offset = explicit_layout ? explicit_layout->offset : 0;
541 struct pan_block_size renderblk_size =
542 panfrost_renderblock_size(layout->modifier, layout->format);
543 struct pan_block_size block_size =
544 panfrost_block_size(layout->modifier, layout->format);
545
546 unsigned width = layout->width;
547 unsigned height = layout->height;
548 unsigned depth = layout->depth;
549
550 unsigned align_w = renderblk_size.width;
551 unsigned align_h = renderblk_size.height;
552
553 /* For tiled AFBC, align to tiles of superblocks (this can be large) */
554 if (afbc) {
555 align_w *= pan_afbc_tile_size(layout->modifier);
556 align_h *= pan_afbc_tile_size(layout->modifier);
557 }
558
559 for (unsigned l = 0; l < layout->nr_slices; ++l) {
560 struct pan_image_slice_layout *slice = &layout->slices[l];
561
562 unsigned effective_width =
563 ALIGN_POT(util_format_get_nblocksx(layout->format, width), align_w);
564 unsigned effective_height =
565 ALIGN_POT(util_format_get_nblocksy(layout->format, height), align_h);
566 unsigned row_stride;
567
568 /* Align levels to cache-line as a performance improvement for
569 * linear/tiled and as a requirement for AFBC */
570
571 offset = ALIGN_POT(offset, pan_slice_align(layout->modifier));
572
573 slice->offset = offset;
574
575 if (afrc) {
576 row_stride = pan_afrc_row_stride(layout->format, layout->modifier,
577 effective_width);
578 } else {
579 row_stride = fmt_blocksize * effective_width * block_size.height;
580 }
581
582 /* On v7+ row_stride and offset alignment requirement are equal */
583 if (arch >= 7) {
584 row_stride = ALIGN_POT(row_stride, align_req);
585 }
586
587 if (explicit_layout && !afbc && !afrc) {
588 /* Make sure the explicit stride is valid */
589 if (explicit_layout->row_stride < row_stride) {
590 mesa_loge("panfrost: rejecting image due to invalid row stride.\n");
591 return false;
592 }
593
594 row_stride = explicit_layout->row_stride;
595 } else if (linear) {
596 /* Keep lines alignment on 64 byte for performance */
597 row_stride = ALIGN_POT(row_stride, 64);
598 }
599
600 uint64_t slice_one_size =
601 (uint64_t)row_stride * (effective_height / block_size.height);
602
603 /* Compute AFBC sizes if necessary */
604 if (afbc) {
605 slice->row_stride =
606 pan_afbc_row_stride(layout->modifier, effective_width);
607 slice->afbc.stride = effective_width / block_size.width;
608 slice->afbc.nr_blocks =
609 slice->afbc.stride * (effective_height / block_size.height);
610 slice->afbc.header_size =
611 ALIGN_POT(slice->afbc.nr_blocks * AFBC_HEADER_BYTES_PER_TILE,
612 pan_afbc_body_align(arch, layout->modifier));
613
614 if (explicit_layout &&
615 explicit_layout->row_stride < slice->row_stride) {
616 mesa_loge("panfrost: rejecting image due to invalid row stride.\n");
617 return false;
618 }
619
620 /* AFBC body size */
621 slice->afbc.body_size = slice_one_size;
622
623 /* 3D AFBC resources have all headers placed at the
624 * beginning instead of having them split per depth
625 * level
626 */
627 if (is_3d) {
628 slice->afbc.surface_stride = slice->afbc.header_size;
629 slice->afbc.header_size *= depth;
630 slice->afbc.body_size *= depth;
631 offset += slice->afbc.header_size;
632 } else {
633 slice_one_size += slice->afbc.header_size;
634 slice->afbc.surface_stride = slice_one_size;
635 }
636 } else {
637 slice->row_stride = row_stride;
638 }
639
640 uint64_t slice_full_size = slice_one_size * depth * layout->nr_samples;
641
642 slice->surface_stride = slice_one_size;
643
644 /* Compute AFBC sizes if necessary */
645
646 offset += slice_full_size;
647 slice->size = slice_full_size;
648
649 /* Add a checksum region if necessary */
650 if (layout->crc) {
651 slice->crc.size = panfrost_compute_checksum_size(slice, width, height);
652
653 slice->crc.offset = offset;
654 offset += slice->crc.size;
655 slice->size += slice->crc.size;
656 }
657
658 width = u_minify(width, 1);
659 height = u_minify(height, 1);
660 depth = u_minify(depth, 1);
661 }
662
663 /* Arrays and cubemaps have the entire miptree duplicated */
664 layout->array_stride = ALIGN_POT(offset, 64);
665 if (explicit_layout)
666 layout->data_size = offset;
667 else
668 layout->data_size = ALIGN_POT(
669 (uint64_t)layout->array_stride * (uint64_t)layout->array_size, 4096);
670
671 return true;
672 }
673
674 void
pan_iview_get_surface(const struct pan_image_view * iview,unsigned level,unsigned layer,unsigned sample,struct pan_surface * surf)675 pan_iview_get_surface(const struct pan_image_view *iview, unsigned level,
676 unsigned layer, unsigned sample, struct pan_surface *surf)
677 {
678 const struct util_format_description *fdesc =
679 util_format_description(iview->format);
680
681
682 /* In case of multiplanar depth/stencil, the stencil is always on
683 * plane 1. Combined depth/stencil only has one plane, so depth
684 * will be on plane 0 in either case.
685 */
686 const struct pan_image *image = util_format_has_stencil(fdesc)
687 ? pan_image_view_get_s_plane(iview)
688 : pan_image_view_get_plane(iview, 0);
689
690 level += iview->first_level;
691 assert(level < image->layout.nr_slices);
692
693 layer += iview->first_layer;
694
695 bool is_3d = image->layout.dim == MALI_TEXTURE_DIMENSION_3D;
696 const struct pan_image_slice_layout *slice = &image->layout.slices[level];
697 uint64_t base = image->data.base + image->data.offset;
698
699 if (drm_is_afbc(image->layout.modifier)) {
700 assert(!sample);
701
702 if (is_3d) {
703 ASSERTED unsigned depth = u_minify(image->layout.depth, level);
704 assert(layer < depth);
705 surf->afbc.header =
706 base + slice->offset + (layer * slice->afbc.surface_stride);
707 surf->afbc.body = base + slice->offset + slice->afbc.header_size +
708 (slice->surface_stride * layer);
709 } else {
710 assert(layer < image->layout.array_size);
711 surf->afbc.header =
712 base + panfrost_texture_offset(&image->layout, level, layer, 0);
713 surf->afbc.body = surf->afbc.header + slice->afbc.header_size;
714 }
715 } else {
716 unsigned array_idx = is_3d ? 0 : layer;
717 unsigned surface_idx = is_3d ? layer : sample;
718
719 surf->data = base + panfrost_texture_offset(&image->layout, level,
720 array_idx, surface_idx);
721 }
722 }
723