projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac/surface/gfx6: don't overallocate mipmapped HTILE
[mesa.git] / src / amd / common / ac_surface.c
1 /*
2 * Copyright © 2011 Red Hat All Rights Reserved.
3 * Copyright © 2017 Advanced Micro Devices, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining
7 * a 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 SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
15 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
16 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
17 * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
18 * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 *
23 * The above copyright notice and this permission notice (including the
24 * next paragraph) shall be included in all copies or substantial portions
25 * of the Software.
26 */
27
28 #include "ac_surface.h"
29 #include "amd_family.h"
30 #include "addrlib/amdgpu_asic_addr.h"
31 #include "ac_gpu_info.h"
32 #include "util/macros.h"
33 #include "util/u_atomic.h"
34 #include "util/u_math.h"
35
36 #include <errno.h>
37 #include <stdio.h>
38 #include <stdlib.h>
39 #include <amdgpu.h>
40 #include <amdgpu_drm.h>
41
42 #include "addrlib/addrinterface.h"
43
44 #ifndef CIASICIDGFXENGINE_SOUTHERNISLAND
45 #define CIASICIDGFXENGINE_SOUTHERNISLAND 0x0000000A
46 #endif
47
48 #ifndef CIASICIDGFXENGINE_ARCTICISLAND
49 #define CIASICIDGFXENGINE_ARCTICISLAND 0x0000000D
50 #endif
51
52 static unsigned get_first(unsigned x, unsigned y)
53 {
54 return x;
55 }
56
57 static void addrlib_family_rev_id(enum radeon_family family,
58 unsigned *addrlib_family,
59 unsigned *addrlib_revid)
60 {
61 switch (family) {
62 case CHIP_TAHITI:
63 *addrlib_family = FAMILY_SI;
64 *addrlib_revid = get_first(AMDGPU_TAHITI_RANGE);
65 break;
66 case CHIP_PITCAIRN:
67 *addrlib_family = FAMILY_SI;
68 *addrlib_revid = get_first(AMDGPU_PITCAIRN_RANGE);
69 break;
70 case CHIP_VERDE:
71 *addrlib_family = FAMILY_SI;
72 *addrlib_revid = get_first(AMDGPU_CAPEVERDE_RANGE);
73 break;
74 case CHIP_OLAND:
75 *addrlib_family = FAMILY_SI;
76 *addrlib_revid = get_first(AMDGPU_OLAND_RANGE);
77 break;
78 case CHIP_HAINAN:
79 *addrlib_family = FAMILY_SI;
80 *addrlib_revid = get_first(AMDGPU_HAINAN_RANGE);
81 break;
82 case CHIP_BONAIRE:
83 *addrlib_family = FAMILY_CI;
84 *addrlib_revid = get_first(AMDGPU_BONAIRE_RANGE);
85 break;
86 case CHIP_KAVERI:
87 *addrlib_family = FAMILY_KV;
88 *addrlib_revid = get_first(AMDGPU_SPECTRE_RANGE);
89 break;
90 case CHIP_KABINI:
91 *addrlib_family = FAMILY_KV;
92 *addrlib_revid = get_first(AMDGPU_KALINDI_RANGE);
93 break;
94 case CHIP_HAWAII:
95 *addrlib_family = FAMILY_CI;
96 *addrlib_revid = get_first(AMDGPU_HAWAII_RANGE);
97 break;
98 case CHIP_MULLINS:
99 *addrlib_family = FAMILY_KV;
100 *addrlib_revid = get_first(AMDGPU_GODAVARI_RANGE);
101 break;
102 case CHIP_TONGA:
103 *addrlib_family = FAMILY_VI;
104 *addrlib_revid = get_first(AMDGPU_TONGA_RANGE);
105 break;
106 case CHIP_ICELAND:
107 *addrlib_family = FAMILY_VI;
108 *addrlib_revid = get_first(AMDGPU_ICELAND_RANGE);
109 break;
110 case CHIP_CARRIZO:
111 *addrlib_family = FAMILY_CZ;
112 *addrlib_revid = get_first(AMDGPU_CARRIZO_RANGE);
113 break;
114 case CHIP_STONEY:
115 *addrlib_family = FAMILY_CZ;
116 *addrlib_revid = get_first(AMDGPU_STONEY_RANGE);
117 break;
118 case CHIP_FIJI:
119 *addrlib_family = FAMILY_VI;
120 *addrlib_revid = get_first(AMDGPU_FIJI_RANGE);
121 break;
122 case CHIP_POLARIS10:
123 *addrlib_family = FAMILY_VI;
124 *addrlib_revid = get_first(AMDGPU_POLARIS10_RANGE);
125 break;
126 case CHIP_POLARIS11:
127 *addrlib_family = FAMILY_VI;
128 *addrlib_revid = get_first(AMDGPU_POLARIS11_RANGE);
129 break;
130 case CHIP_POLARIS12:
131 *addrlib_family = FAMILY_VI;
132 *addrlib_revid = get_first(AMDGPU_POLARIS12_RANGE);
133 break;
134 case CHIP_VEGAM:
135 *addrlib_family = FAMILY_VI;
136 *addrlib_revid = get_first(AMDGPU_VEGAM_RANGE);
137 break;
138 case CHIP_VEGA10:
139 *addrlib_family = FAMILY_AI;
140 *addrlib_revid = get_first(AMDGPU_VEGA10_RANGE);
141 break;
142 case CHIP_VEGA12:
143 *addrlib_family = FAMILY_AI;
144 *addrlib_revid = get_first(AMDGPU_VEGA12_RANGE);
145 break;
146 case CHIP_RAVEN:
147 *addrlib_family = FAMILY_RV;
148 *addrlib_revid = get_first(AMDGPU_RAVEN_RANGE);
149 break;
150 default:
151 fprintf(stderr, "amdgpu: Unknown family.\n");
152 }
153 }
154
155 static void *ADDR_API allocSysMem(const ADDR_ALLOCSYSMEM_INPUT * pInput)
156 {
157 return malloc(pInput->sizeInBytes);
158 }
159
160 static ADDR_E_RETURNCODE ADDR_API freeSysMem(const ADDR_FREESYSMEM_INPUT * pInput)
161 {
162 free(pInput->pVirtAddr);
163 return ADDR_OK;
164 }
165
166 ADDR_HANDLE amdgpu_addr_create(const struct radeon_info *info,
167 const struct amdgpu_gpu_info *amdinfo,
168 uint64_t *max_alignment)
169 {
170 ADDR_CREATE_INPUT addrCreateInput = {0};
171 ADDR_CREATE_OUTPUT addrCreateOutput = {0};
172 ADDR_REGISTER_VALUE regValue = {0};
173 ADDR_CREATE_FLAGS createFlags = {{0}};
174 ADDR_GET_MAX_ALINGMENTS_OUTPUT addrGetMaxAlignmentsOutput = {0};
175 ADDR_E_RETURNCODE addrRet;
176
177 addrCreateInput.size = sizeof(ADDR_CREATE_INPUT);
178 addrCreateOutput.size = sizeof(ADDR_CREATE_OUTPUT);
179
180 regValue.gbAddrConfig = amdinfo->gb_addr_cfg;
181 createFlags.value = 0;
182
183 addrlib_family_rev_id(info->family, &addrCreateInput.chipFamily, &addrCreateInput.chipRevision);
184 if (addrCreateInput.chipFamily == FAMILY_UNKNOWN)
185 return NULL;
186
187 if (addrCreateInput.chipFamily >= FAMILY_AI) {
188 addrCreateInput.chipEngine = CIASICIDGFXENGINE_ARCTICISLAND;
189 regValue.blockVarSizeLog2 = 0;
190 } else {
191 regValue.noOfBanks = amdinfo->mc_arb_ramcfg & 0x3;
192 regValue.noOfRanks = (amdinfo->mc_arb_ramcfg & 0x4) >> 2;
193
194 regValue.backendDisables = amdinfo->enabled_rb_pipes_mask;
195 regValue.pTileConfig = amdinfo->gb_tile_mode;
196 regValue.noOfEntries = ARRAY_SIZE(amdinfo->gb_tile_mode);
197 if (addrCreateInput.chipFamily == FAMILY_SI) {
198 regValue.pMacroTileConfig = NULL;
199 regValue.noOfMacroEntries = 0;
200 } else {
201 regValue.pMacroTileConfig = amdinfo->gb_macro_tile_mode;
202 regValue.noOfMacroEntries = ARRAY_SIZE(amdinfo->gb_macro_tile_mode);
203 }
204
205 createFlags.useTileIndex = 1;
206 createFlags.useHtileSliceAlign = 1;
207
208 addrCreateInput.chipEngine = CIASICIDGFXENGINE_SOUTHERNISLAND;
209 }
210
211 addrCreateInput.callbacks.allocSysMem = allocSysMem;
212 addrCreateInput.callbacks.freeSysMem = freeSysMem;
213 addrCreateInput.callbacks.debugPrint = 0;
214 addrCreateInput.createFlags = createFlags;
215 addrCreateInput.regValue = regValue;
216
217 addrRet = AddrCreate(&addrCreateInput, &addrCreateOutput);
218 if (addrRet != ADDR_OK)
219 return NULL;
220
221 if (max_alignment) {
222 addrRet = AddrGetMaxAlignments(addrCreateOutput.hLib, &addrGetMaxAlignmentsOutput);
223 if (addrRet == ADDR_OK){
224 *max_alignment = addrGetMaxAlignmentsOutput.baseAlign;
225 }
226 }
227 return addrCreateOutput.hLib;
228 }
229
230 static int surf_config_sanity(const struct ac_surf_config *config,
231 unsigned flags)
232 {
233 /* FMASK is allocated together with the color surface and can't be
234 * allocated separately.
235 */
236 assert(!(flags & RADEON_SURF_FMASK));
237 if (flags & RADEON_SURF_FMASK)
238 return -EINVAL;
239
240 /* all dimension must be at least 1 ! */
241 if (!config->info.width || !config->info.height || !config->info.depth ||
242 !config->info.array_size || !config->info.levels)
243 return -EINVAL;
244
245 switch (config->info.samples) {
246 case 0:
247 case 1:
248 case 2:
249 case 4:
250 case 8:
251 break;
252 case 16:
253 if (flags & RADEON_SURF_Z_OR_SBUFFER)
254 return -EINVAL;
255 break;
256 default:
257 return -EINVAL;
258 }
259
260 if (!(flags & RADEON_SURF_Z_OR_SBUFFER)) {
261 switch (config->info.color_samples) {
262 case 0:
263 case 1:
264 case 2:
265 case 4:
266 case 8:
267 break;
268 default:
269 return -EINVAL;
270 }
271 }
272
273 if (config->is_3d && config->info.array_size > 1)
274 return -EINVAL;
275 if (config->is_cube && config->info.depth > 1)
276 return -EINVAL;
277
278 return 0;
279 }
280
281 static int gfx6_compute_level(ADDR_HANDLE addrlib,
282 const struct ac_surf_config *config,
283 struct radeon_surf *surf, bool is_stencil,
284 unsigned level, bool compressed,
285 ADDR_COMPUTE_SURFACE_INFO_INPUT *AddrSurfInfoIn,
286 ADDR_COMPUTE_SURFACE_INFO_OUTPUT *AddrSurfInfoOut,
287 ADDR_COMPUTE_DCCINFO_INPUT *AddrDccIn,
288 ADDR_COMPUTE_DCCINFO_OUTPUT *AddrDccOut,
289 ADDR_COMPUTE_HTILE_INFO_INPUT *AddrHtileIn,
290 ADDR_COMPUTE_HTILE_INFO_OUTPUT *AddrHtileOut)
291 {
292 struct legacy_surf_level *surf_level;
293 ADDR_E_RETURNCODE ret;
294
295 AddrSurfInfoIn->mipLevel = level;
296 AddrSurfInfoIn->width = u_minify(config->info.width, level);
297 AddrSurfInfoIn->height = u_minify(config->info.height, level);
298
299 /* Make GFX6 linear surfaces compatible with GFX9 for hybrid graphics,
300 * because GFX9 needs linear alignment of 256 bytes.
301 */
302 if (config->info.levels == 1 &&
303 AddrSurfInfoIn->tileMode == ADDR_TM_LINEAR_ALIGNED &&
304 AddrSurfInfoIn->bpp &&
305 util_is_power_of_two_or_zero(AddrSurfInfoIn->bpp)) {
306 unsigned alignment = 256 / (AddrSurfInfoIn->bpp / 8);
307
308 AddrSurfInfoIn->width = align(AddrSurfInfoIn->width, alignment);
309 }
310
311 if (config->is_3d)
312 AddrSurfInfoIn->numSlices = u_minify(config->info.depth, level);
313 else if (config->is_cube)
314 AddrSurfInfoIn->numSlices = 6;
315 else
316 AddrSurfInfoIn->numSlices = config->info.array_size;
317
318 if (level > 0) {
319 /* Set the base level pitch. This is needed for calculation
320 * of non-zero levels. */
321 if (is_stencil)
322 AddrSurfInfoIn->basePitch = surf->u.legacy.stencil_level[0].nblk_x;
323 else
324 AddrSurfInfoIn->basePitch = surf->u.legacy.level[0].nblk_x;
325
326 /* Convert blocks to pixels for compressed formats. */
327 if (compressed)
328 AddrSurfInfoIn->basePitch *= surf->blk_w;
329 }
330
331 ret = AddrComputeSurfaceInfo(addrlib,
332 AddrSurfInfoIn,
333 AddrSurfInfoOut);
334 if (ret != ADDR_OK) {
335 return ret;
336 }
337
338 surf_level = is_stencil ? &surf->u.legacy.stencil_level[level] : &surf->u.legacy.level[level];
339 surf_level->offset = align64(surf->surf_size, AddrSurfInfoOut->baseAlign);
340 surf_level->slice_size_dw = AddrSurfInfoOut->sliceSize / 4;
341 surf_level->nblk_x = AddrSurfInfoOut->pitch;
342 surf_level->nblk_y = AddrSurfInfoOut->height;
343
344 switch (AddrSurfInfoOut->tileMode) {
345 case ADDR_TM_LINEAR_ALIGNED:
346 surf_level->mode = RADEON_SURF_MODE_LINEAR_ALIGNED;
347 break;
348 case ADDR_TM_1D_TILED_THIN1:
349 surf_level->mode = RADEON_SURF_MODE_1D;
350 break;
351 case ADDR_TM_2D_TILED_THIN1:
352 surf_level->mode = RADEON_SURF_MODE_2D;
353 break;
354 default:
355 assert(0);
356 }
357
358 if (is_stencil)
359 surf->u.legacy.stencil_tiling_index[level] = AddrSurfInfoOut->tileIndex;
360 else
361 surf->u.legacy.tiling_index[level] = AddrSurfInfoOut->tileIndex;
362
363 surf->surf_size = surf_level->offset + AddrSurfInfoOut->surfSize;
364
365 /* Clear DCC fields at the beginning. */
366 surf_level->dcc_offset = 0;
367
368 /* The previous level's flag tells us if we can use DCC for this level. */
369 if (AddrSurfInfoIn->flags.dccCompatible &&
370 (level == 0 || AddrDccOut->subLvlCompressible)) {
371 bool prev_level_clearable = level == 0 ||
372 AddrDccOut->dccRamSizeAligned;
373
374 AddrDccIn->colorSurfSize = AddrSurfInfoOut->surfSize;
375 AddrDccIn->tileMode = AddrSurfInfoOut->tileMode;
376 AddrDccIn->tileInfo = *AddrSurfInfoOut->pTileInfo;
377 AddrDccIn->tileIndex = AddrSurfInfoOut->tileIndex;
378 AddrDccIn->macroModeIndex = AddrSurfInfoOut->macroModeIndex;
379
380 ret = AddrComputeDccInfo(addrlib,
381 AddrDccIn,
382 AddrDccOut);
383
384 if (ret == ADDR_OK) {
385 surf_level->dcc_offset = surf->dcc_size;
386 surf->num_dcc_levels = level + 1;
387 surf->dcc_size = surf_level->dcc_offset + AddrDccOut->dccRamSize;
388 surf->dcc_alignment = MAX2(surf->dcc_alignment, AddrDccOut->dccRamBaseAlign);
389
390 /* If the DCC size of a subresource (1 mip level or 1 slice)
391 * is not aligned, the DCC memory layout is not contiguous for
392 * that subresource, which means we can't use fast clear.
393 *
394 * We only do fast clears for whole mipmap levels. If we did
395 * per-slice fast clears, the same restriction would apply.
396 * (i.e. only compute the slice size and see if it's aligned)
397 *
398 * The last level can be non-contiguous and still be clearable
399 * if it's interleaved with the next level that doesn't exist.
400 */
401 if (AddrDccOut->dccRamSizeAligned ||
402 (prev_level_clearable && level == config->info.levels - 1))
403 surf_level->dcc_fast_clear_size = AddrDccOut->dccFastClearSize;
404 else
405 surf_level->dcc_fast_clear_size = 0;
406 }
407 }
408
409 /* TC-compatible HTILE. */
410 if (!is_stencil &&
411 AddrSurfInfoIn->flags.depth &&
412 surf_level->mode == RADEON_SURF_MODE_2D &&
413 level == 0) {
414 AddrHtileIn->flags.tcCompatible = AddrSurfInfoIn->flags.tcCompatible;
415 AddrHtileIn->pitch = AddrSurfInfoOut->pitch;
416 AddrHtileIn->height = AddrSurfInfoOut->height;
417 AddrHtileIn->numSlices = AddrSurfInfoOut->depth;
418 AddrHtileIn->blockWidth = ADDR_HTILE_BLOCKSIZE_8;
419 AddrHtileIn->blockHeight = ADDR_HTILE_BLOCKSIZE_8;
420 AddrHtileIn->pTileInfo = AddrSurfInfoOut->pTileInfo;
421 AddrHtileIn->tileIndex = AddrSurfInfoOut->tileIndex;
422 AddrHtileIn->macroModeIndex = AddrSurfInfoOut->macroModeIndex;
423
424 ret = AddrComputeHtileInfo(addrlib,
425 AddrHtileIn,
426 AddrHtileOut);
427
428 if (ret == ADDR_OK) {
429 surf->htile_size = AddrHtileOut->htileBytes;
430 surf->htile_slice_size = AddrHtileOut->sliceSize;
431 surf->htile_alignment = AddrHtileOut->baseAlign;
432 }
433 }
434
435 return 0;
436 }
437
438 #define G_009910_MICRO_TILE_MODE(x) (((x) >> 0) & 0x03)
439 #define G_009910_MICRO_TILE_MODE_NEW(x) (((x) >> 22) & 0x07)
440
441 static void gfx6_set_micro_tile_mode(struct radeon_surf *surf,
442 const struct radeon_info *info)
443 {
444 uint32_t tile_mode = info->si_tile_mode_array[surf->u.legacy.tiling_index[0]];
445
446 if (info->chip_class >= CIK)
447 surf->micro_tile_mode = G_009910_MICRO_TILE_MODE_NEW(tile_mode);
448 else
449 surf->micro_tile_mode = G_009910_MICRO_TILE_MODE(tile_mode);
450 }
451
452 static unsigned cik_get_macro_tile_index(struct radeon_surf *surf)
453 {
454 unsigned index, tileb;
455
456 tileb = 8 * 8 * surf->bpe;
457 tileb = MIN2(surf->u.legacy.tile_split, tileb);
458
459 for (index = 0; tileb > 64; index++)
460 tileb >>= 1;
461
462 assert(index < 16);
463 return index;
464 }
465
466 static bool get_display_flag(const struct ac_surf_config *config,
467 const struct radeon_surf *surf)
468 {
469 unsigned num_channels = config->info.num_channels;
470 unsigned bpe = surf->bpe;
471
472 if (surf->flags & RADEON_SURF_SCANOUT &&
473 config->info.samples <= 1 &&
474 surf->blk_w <= 2 && surf->blk_h == 1) {
475 /* subsampled */
476 if (surf->blk_w == 2 && surf->blk_h == 1)
477 return true;
478
479 if (/* RGBA8 or RGBA16F */
480 (bpe >= 4 && bpe <= 8 && num_channels == 4) ||
481 /* R5G6B5 or R5G5B5A1 */
482 (bpe == 2 && num_channels >= 3) ||
483 /* C8 palette */
484 (bpe == 1 && num_channels == 1))
485 return true;
486 }
487 return false;
488 }
489
490 /**
491 * This must be called after the first level is computed.
492 *
493 * Copy surface-global settings like pipe/bank config from level 0 surface
494 * computation, and compute tile swizzle.
495 */
496 static int gfx6_surface_settings(ADDR_HANDLE addrlib,
497 const struct radeon_info *info,
498 const struct ac_surf_config *config,
499 ADDR_COMPUTE_SURFACE_INFO_OUTPUT* csio,
500 struct radeon_surf *surf)
501 {
502 surf->surf_alignment = csio->baseAlign;
503 surf->u.legacy.pipe_config = csio->pTileInfo->pipeConfig - 1;
504 gfx6_set_micro_tile_mode(surf, info);
505
506 /* For 2D modes only. */
507 if (csio->tileMode >= ADDR_TM_2D_TILED_THIN1) {
508 surf->u.legacy.bankw = csio->pTileInfo->bankWidth;
509 surf->u.legacy.bankh = csio->pTileInfo->bankHeight;
510 surf->u.legacy.mtilea = csio->pTileInfo->macroAspectRatio;
511 surf->u.legacy.tile_split = csio->pTileInfo->tileSplitBytes;
512 surf->u.legacy.num_banks = csio->pTileInfo->banks;
513 surf->u.legacy.macro_tile_index = csio->macroModeIndex;
514 } else {
515 surf->u.legacy.macro_tile_index = 0;
516 }
517
518 /* Compute tile swizzle. */
519 /* TODO: fix tile swizzle with mipmapping for SI */
520 if ((info->chip_class >= CIK || config->info.levels == 1) &&
521 config->info.surf_index &&
522 surf->u.legacy.level[0].mode == RADEON_SURF_MODE_2D &&
523 !(surf->flags & (RADEON_SURF_Z_OR_SBUFFER | RADEON_SURF_SHAREABLE)) &&
524 !get_display_flag(config, surf)) {
525 ADDR_COMPUTE_BASE_SWIZZLE_INPUT AddrBaseSwizzleIn = {0};
526 ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT AddrBaseSwizzleOut = {0};
527
528 AddrBaseSwizzleIn.size = sizeof(ADDR_COMPUTE_BASE_SWIZZLE_INPUT);
529 AddrBaseSwizzleOut.size = sizeof(ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT);
530
531 AddrBaseSwizzleIn.surfIndex = p_atomic_inc_return(config->info.surf_index) - 1;
532 AddrBaseSwizzleIn.tileIndex = csio->tileIndex;
533 AddrBaseSwizzleIn.macroModeIndex = csio->macroModeIndex;
534 AddrBaseSwizzleIn.pTileInfo = csio->pTileInfo;
535 AddrBaseSwizzleIn.tileMode = csio->tileMode;
536
537 int r = AddrComputeBaseSwizzle(addrlib, &AddrBaseSwizzleIn,
538 &AddrBaseSwizzleOut);
539 if (r != ADDR_OK)
540 return r;
541
542 assert(AddrBaseSwizzleOut.tileSwizzle <=
543 u_bit_consecutive(0, sizeof(surf->tile_swizzle) * 8));
544 surf->tile_swizzle = AddrBaseSwizzleOut.tileSwizzle;
545 }
546 return 0;
547 }
548
549 /**
550 * Fill in the tiling information in \p surf based on the given surface config.
551 *
552 * The following fields of \p surf must be initialized by the caller:
553 * blk_w, blk_h, bpe, flags.
554 */
555 static int gfx6_compute_surface(ADDR_HANDLE addrlib,
556 const struct radeon_info *info,
557 const struct ac_surf_config *config,
558 enum radeon_surf_mode mode,
559 struct radeon_surf *surf)
560 {
561 unsigned level;
562 bool compressed;
563 ADDR_COMPUTE_SURFACE_INFO_INPUT AddrSurfInfoIn = {0};
564 ADDR_COMPUTE_SURFACE_INFO_OUTPUT AddrSurfInfoOut = {0};
565 ADDR_COMPUTE_DCCINFO_INPUT AddrDccIn = {0};
566 ADDR_COMPUTE_DCCINFO_OUTPUT AddrDccOut = {0};
567 ADDR_COMPUTE_HTILE_INFO_INPUT AddrHtileIn = {0};
568 ADDR_COMPUTE_HTILE_INFO_OUTPUT AddrHtileOut = {0};
569 ADDR_TILEINFO AddrTileInfoIn = {0};
570 ADDR_TILEINFO AddrTileInfoOut = {0};
571 int r;
572
573 AddrSurfInfoIn.size = sizeof(ADDR_COMPUTE_SURFACE_INFO_INPUT);
574 AddrSurfInfoOut.size = sizeof(ADDR_COMPUTE_SURFACE_INFO_OUTPUT);
575 AddrDccIn.size = sizeof(ADDR_COMPUTE_DCCINFO_INPUT);
576 AddrDccOut.size = sizeof(ADDR_COMPUTE_DCCINFO_OUTPUT);
577 AddrHtileIn.size = sizeof(ADDR_COMPUTE_HTILE_INFO_INPUT);
578 AddrHtileOut.size = sizeof(ADDR_COMPUTE_HTILE_INFO_OUTPUT);
579 AddrSurfInfoOut.pTileInfo = &AddrTileInfoOut;
580
581 compressed = surf->blk_w == 4 && surf->blk_h == 4;
582
583 /* MSAA requires 2D tiling. */
584 if (config->info.samples > 1)
585 mode = RADEON_SURF_MODE_2D;
586
587 /* DB doesn't support linear layouts. */
588 if (surf->flags & (RADEON_SURF_Z_OR_SBUFFER) &&
589 mode < RADEON_SURF_MODE_1D)
590 mode = RADEON_SURF_MODE_1D;
591
592 /* Set the requested tiling mode. */
593 switch (mode) {
594 case RADEON_SURF_MODE_LINEAR_ALIGNED:
595 AddrSurfInfoIn.tileMode = ADDR_TM_LINEAR_ALIGNED;
596 break;
597 case RADEON_SURF_MODE_1D:
598 AddrSurfInfoIn.tileMode = ADDR_TM_1D_TILED_THIN1;
599 break;
600 case RADEON_SURF_MODE_2D:
601 AddrSurfInfoIn.tileMode = ADDR_TM_2D_TILED_THIN1;
602 break;
603 default:
604 assert(0);
605 }
606
607 /* The format must be set correctly for the allocation of compressed
608 * textures to work. In other cases, setting the bpp is sufficient.
609 */
610 if (compressed) {
611 switch (surf->bpe) {
612 case 8:
613 AddrSurfInfoIn.format = ADDR_FMT_BC1;
614 break;
615 case 16:
616 AddrSurfInfoIn.format = ADDR_FMT_BC3;
617 break;
618 default:
619 assert(0);
620 }
621 }
622 else {
623 AddrDccIn.bpp = AddrSurfInfoIn.bpp = surf->bpe * 8;
624 }
625
626 AddrDccIn.numSamples = AddrSurfInfoIn.numSamples =
627 MAX2(1, config->info.samples);
628 AddrSurfInfoIn.tileIndex = -1;
629
630 if (!(surf->flags & RADEON_SURF_Z_OR_SBUFFER)) {
631 AddrDccIn.numSamples = AddrSurfInfoIn.numFrags =
632 MAX2(1, config->info.color_samples);
633 }
634
635 /* Set the micro tile type. */
636 if (surf->flags & RADEON_SURF_SCANOUT)
637 AddrSurfInfoIn.tileType = ADDR_DISPLAYABLE;
638 else if (surf->flags & RADEON_SURF_Z_OR_SBUFFER)
639 AddrSurfInfoIn.tileType = ADDR_DEPTH_SAMPLE_ORDER;
640 else
641 AddrSurfInfoIn.tileType = ADDR_NON_DISPLAYABLE;
642
643 AddrSurfInfoIn.flags.color = !(surf->flags & RADEON_SURF_Z_OR_SBUFFER);
644 AddrSurfInfoIn.flags.depth = (surf->flags & RADEON_SURF_ZBUFFER) != 0;
645 AddrSurfInfoIn.flags.cube = config->is_cube;
646 AddrSurfInfoIn.flags.display = get_display_flag(config, surf);
647 AddrSurfInfoIn.flags.pow2Pad = config->info.levels > 1;
648 AddrSurfInfoIn.flags.tcCompatible = (surf->flags & RADEON_SURF_TC_COMPATIBLE_HTILE) != 0;
649
650 /* Only degrade the tile mode for space if TC-compatible HTILE hasn't been
651 * requested, because TC-compatible HTILE requires 2D tiling.
652 */
653 AddrSurfInfoIn.flags.opt4Space = !AddrSurfInfoIn.flags.tcCompatible &&
654 !AddrSurfInfoIn.flags.fmask &&
655 config->info.samples <= 1 &&
656 (surf->flags & RADEON_SURF_OPTIMIZE_FOR_SPACE);
657
658 /* DCC notes:
659 * - If we add MSAA support, keep in mind that CB can't decompress 8bpp
660 * with samples >= 4.
661 * - Mipmapped array textures have low performance (discovered by a closed
662 * driver team).
663 */
664 AddrSurfInfoIn.flags.dccCompatible =
665 info->chip_class >= VI &&
666 !(surf->flags & RADEON_SURF_Z_OR_SBUFFER) &&
667 !(surf->flags & RADEON_SURF_DISABLE_DCC) &&
668 !compressed &&
669 ((config->info.array_size == 1 && config->info.depth == 1) ||
670 config->info.levels == 1);
671
672 AddrSurfInfoIn.flags.noStencil = (surf->flags & RADEON_SURF_SBUFFER) == 0;
673 AddrSurfInfoIn.flags.compressZ = AddrSurfInfoIn.flags.depth;
674
675 /* On CI/VI, the DB uses the same pitch and tile mode (except tilesplit)
676 * for Z and stencil. This can cause a number of problems which we work
677 * around here:
678 *
679 * - a depth part that is incompatible with mipmapped texturing
680 * - at least on Stoney, entirely incompatible Z/S aspects (e.g.
681 * incorrect tiling applied to the stencil part, stencil buffer
682 * memory accesses that go out of bounds) even without mipmapping
683 *
684 * Some piglit tests that are prone to different types of related
685 * failures:
686 * ./bin/ext_framebuffer_multisample-upsample 2 stencil
687 * ./bin/framebuffer-blit-levels {draw,read} stencil
688 * ./bin/ext_framebuffer_multisample-unaligned-blit N {depth,stencil} {msaa,upsample,downsample}
689 * ./bin/fbo-depth-array fs-writes-{depth,stencil} / {depth,stencil}-{clear,layered-clear,draw}
690 * ./bin/depthstencil-render-miplevels 1024 d=s=z24_s8
691 */
692 int stencil_tile_idx = -1;
693
694 if (AddrSurfInfoIn.flags.depth && !AddrSurfInfoIn.flags.noStencil &&
695 (config->info.levels > 1 || info->family == CHIP_STONEY)) {
696 /* Compute stencilTileIdx that is compatible with the (depth)
697 * tileIdx. This degrades the depth surface if necessary to
698 * ensure that a matching stencilTileIdx exists. */
699 AddrSurfInfoIn.flags.matchStencilTileCfg = 1;
700
701 /* Keep the depth mip-tail compatible with texturing. */
702 AddrSurfInfoIn.flags.noStencil = 1;
703 }
704
705 /* Set preferred macrotile parameters. This is usually required
706 * for shared resources. This is for 2D tiling only. */
707 if (AddrSurfInfoIn.tileMode >= ADDR_TM_2D_TILED_THIN1 &&
708 surf->u.legacy.bankw && surf->u.legacy.bankh &&
709 surf->u.legacy.mtilea && surf->u.legacy.tile_split) {
710 /* If any of these parameters are incorrect, the calculation
711 * will fail. */
712 AddrTileInfoIn.banks = surf->u.legacy.num_banks;
713 AddrTileInfoIn.bankWidth = surf->u.legacy.bankw;
714 AddrTileInfoIn.bankHeight = surf->u.legacy.bankh;
715 AddrTileInfoIn.macroAspectRatio = surf->u.legacy.mtilea;
716 AddrTileInfoIn.tileSplitBytes = surf->u.legacy.tile_split;
717 AddrTileInfoIn.pipeConfig = surf->u.legacy.pipe_config + 1; /* +1 compared to GB_TILE_MODE */
718 AddrSurfInfoIn.flags.opt4Space = 0;
719 AddrSurfInfoIn.pTileInfo = &AddrTileInfoIn;
720
721 /* If AddrSurfInfoIn.pTileInfo is set, Addrlib doesn't set
722 * the tile index, because we are expected to know it if
723 * we know the other parameters.
724 *
725 * This is something that can easily be fixed in Addrlib.
726 * For now, just figure it out here.
727 * Note that only 2D_TILE_THIN1 is handled here.
728 */
729 assert(!(surf->flags & RADEON_SURF_Z_OR_SBUFFER));
730 assert(AddrSurfInfoIn.tileMode == ADDR_TM_2D_TILED_THIN1);
731
732 if (info->chip_class == SI) {
733 if (AddrSurfInfoIn.tileType == ADDR_DISPLAYABLE) {
734 if (surf->bpe == 2)
735 AddrSurfInfoIn.tileIndex = 11; /* 16bpp */
736 else
737 AddrSurfInfoIn.tileIndex = 12; /* 32bpp */
738 } else {
739 if (surf->bpe == 1)
740 AddrSurfInfoIn.tileIndex = 14; /* 8bpp */
741 else if (surf->bpe == 2)
742 AddrSurfInfoIn.tileIndex = 15; /* 16bpp */
743 else if (surf->bpe == 4)
744 AddrSurfInfoIn.tileIndex = 16; /* 32bpp */
745 else
746 AddrSurfInfoIn.tileIndex = 17; /* 64bpp (and 128bpp) */
747 }
748 } else {
749 /* CIK - VI */
750 if (AddrSurfInfoIn.tileType == ADDR_DISPLAYABLE)
751 AddrSurfInfoIn.tileIndex = 10; /* 2D displayable */
752 else
753 AddrSurfInfoIn.tileIndex = 14; /* 2D non-displayable */
754
755 /* Addrlib doesn't set this if tileIndex is forced like above. */
756 AddrSurfInfoOut.macroModeIndex = cik_get_macro_tile_index(surf);
757 }
758 }
759
760 surf->has_stencil = !!(surf->flags & RADEON_SURF_SBUFFER);
761 surf->num_dcc_levels = 0;
762 surf->surf_size = 0;
763 surf->dcc_size = 0;
764 surf->dcc_alignment = 1;
765 surf->htile_size = 0;
766 surf->htile_slice_size = 0;
767 surf->htile_alignment = 1;
768
769 const bool only_stencil = (surf->flags & RADEON_SURF_SBUFFER) &&
770 !(surf->flags & RADEON_SURF_ZBUFFER);
771
772 /* Calculate texture layout information. */
773 if (!only_stencil) {
774 for (level = 0; level < config->info.levels; level++) {
775 r = gfx6_compute_level(addrlib, config, surf, false, level, compressed,
776 &AddrSurfInfoIn, &AddrSurfInfoOut,
777 &AddrDccIn, &AddrDccOut, &AddrHtileIn, &AddrHtileOut);
778 if (r)
779 return r;
780
781 if (level > 0)
782 continue;
783
784 /* Check that we actually got a TC-compatible HTILE if
785 * we requested it (only for level 0, since we're not
786 * supporting HTILE on higher mip levels anyway). */
787 assert(AddrSurfInfoOut.tcCompatible ||
788 !AddrSurfInfoIn.flags.tcCompatible ||
789 AddrSurfInfoIn.flags.matchStencilTileCfg);
790
791 if (AddrSurfInfoIn.flags.matchStencilTileCfg) {
792 if (!AddrSurfInfoOut.tcCompatible) {
793 AddrSurfInfoIn.flags.tcCompatible = 0;
794 surf->flags &= ~RADEON_SURF_TC_COMPATIBLE_HTILE;
795 }
796
797 AddrSurfInfoIn.flags.matchStencilTileCfg = 0;
798 AddrSurfInfoIn.tileIndex = AddrSurfInfoOut.tileIndex;
799 stencil_tile_idx = AddrSurfInfoOut.stencilTileIdx;
800
801 assert(stencil_tile_idx >= 0);
802 }
803
804 r = gfx6_surface_settings(addrlib, info, config,
805 &AddrSurfInfoOut, surf);
806 if (r)
807 return r;
808 }
809 }
810
811 /* Calculate texture layout information for stencil. */
812 if (surf->flags & RADEON_SURF_SBUFFER) {
813 AddrSurfInfoIn.tileIndex = stencil_tile_idx;
814 AddrSurfInfoIn.bpp = 8;
815 AddrSurfInfoIn.flags.depth = 0;
816 AddrSurfInfoIn.flags.stencil = 1;
817 AddrSurfInfoIn.flags.tcCompatible = 0;
818 /* This will be ignored if AddrSurfInfoIn.pTileInfo is NULL. */
819 AddrTileInfoIn.tileSplitBytes = surf->u.legacy.stencil_tile_split;
820
821 for (level = 0; level < config->info.levels; level++) {
822 r = gfx6_compute_level(addrlib, config, surf, true, level, compressed,
823 &AddrSurfInfoIn, &AddrSurfInfoOut,
824 &AddrDccIn, &AddrDccOut,
825 NULL, NULL);
826 if (r)
827 return r;
828
829 /* DB uses the depth pitch for both stencil and depth. */
830 if (!only_stencil) {
831 if (surf->u.legacy.stencil_level[level].nblk_x !=
832 surf->u.legacy.level[level].nblk_x)
833 surf->u.legacy.stencil_adjusted = true;
834 } else {
835 surf->u.legacy.level[level].nblk_x =
836 surf->u.legacy.stencil_level[level].nblk_x;
837 }
838
839 if (level == 0) {
840 if (only_stencil) {
841 r = gfx6_surface_settings(addrlib, info, config,
842 &AddrSurfInfoOut, surf);
843 if (r)
844 return r;
845 }
846
847 /* For 2D modes only. */
848 if (AddrSurfInfoOut.tileMode >= ADDR_TM_2D_TILED_THIN1) {
849 surf->u.legacy.stencil_tile_split =
850 AddrSurfInfoOut.pTileInfo->tileSplitBytes;
851 }
852 }
853 }
854 }
855
856 /* Compute FMASK. */
857 if (config->info.samples >= 2 && AddrSurfInfoIn.flags.color) {
858 ADDR_COMPUTE_FMASK_INFO_INPUT fin = {0};
859 ADDR_COMPUTE_FMASK_INFO_OUTPUT fout = {0};
860 ADDR_TILEINFO fmask_tile_info = {};
861
862 fin.size = sizeof(fin);
863 fout.size = sizeof(fout);
864
865 fin.tileMode = AddrSurfInfoOut.tileMode;
866 fin.pitch = AddrSurfInfoOut.pitch;
867 fin.height = config->info.height;
868 fin.numSlices = AddrSurfInfoIn.numSlices;
869 fin.numSamples = AddrSurfInfoIn.numSamples;
870 fin.numFrags = AddrSurfInfoIn.numFrags;
871 fin.tileIndex = -1;
872 fout.pTileInfo = &fmask_tile_info;
873
874 r = AddrComputeFmaskInfo(addrlib, &fin, &fout);
875 if (r)
876 return r;
877
878 surf->fmask_size = fout.fmaskBytes;
879 surf->fmask_alignment = fout.baseAlign;
880 surf->fmask_tile_swizzle = 0;
881
882 surf->u.legacy.fmask.slice_tile_max =
883 (fout.pitch * fout.height) / 64;
884 if (surf->u.legacy.fmask.slice_tile_max)
885 surf->u.legacy.fmask.slice_tile_max -= 1;
886
887 surf->u.legacy.fmask.tiling_index = fout.tileIndex;
888 surf->u.legacy.fmask.bankh = fout.pTileInfo->bankHeight;
889 surf->u.legacy.fmask.pitch_in_pixels = fout.pitch;
890
891 /* Compute tile swizzle for FMASK. */
892 if (config->info.fmask_surf_index &&
893 !(surf->flags & RADEON_SURF_SHAREABLE)) {
894 ADDR_COMPUTE_BASE_SWIZZLE_INPUT xin = {0};
895 ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT xout = {0};
896
897 xin.size = sizeof(ADDR_COMPUTE_BASE_SWIZZLE_INPUT);
898 xout.size = sizeof(ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT);
899
900 /* This counter starts from 1 instead of 0. */
901 xin.surfIndex = p_atomic_inc_return(config->info.fmask_surf_index);
902 xin.tileIndex = fout.tileIndex;
903 xin.macroModeIndex = fout.macroModeIndex;
904 xin.pTileInfo = fout.pTileInfo;
905 xin.tileMode = fin.tileMode;
906
907 int r = AddrComputeBaseSwizzle(addrlib, &xin, &xout);
908 if (r != ADDR_OK)
909 return r;
910
911 assert(xout.tileSwizzle <=
912 u_bit_consecutive(0, sizeof(surf->tile_swizzle) * 8));
913 surf->fmask_tile_swizzle = xout.tileSwizzle;
914 }
915 }
916
917 /* Recalculate the whole DCC miptree size including disabled levels.
918 * This is what addrlib does, but calling addrlib would be a lot more
919 * complicated.
920 */
921 if (surf->dcc_size && config->info.levels > 1) {
922 /* The smallest miplevels that are never compressed by DCC
923 * still read the DCC buffer via TC if the base level uses DCC,
924 * and for some reason the DCC buffer needs to be larger if
925 * the miptree uses non-zero tile_swizzle. Otherwise there are
926 * VM faults.
927 *
928 * "dcc_alignment * 4" was determined by trial and error.
929 */
930 surf->dcc_size = align64(surf->surf_size >> 8,
931 surf->dcc_alignment * 4);
932 }
933
934 /* Make sure HTILE covers the whole miptree, because the shader reads
935 * TC-compatible HTILE even for levels where it's disabled by DB.
936 */
937 if (surf->htile_size && config->info.levels > 1 &&
938 surf->flags & RADEON_SURF_TC_COMPATIBLE_HTILE) {
939 /* MSAA can't occur with levels > 1, so ignore the sample count. */
940 const unsigned total_pixels = surf->surf_size / surf->bpe;
941 const unsigned htile_block_size = 8 * 8;
942 const unsigned htile_element_size = 4;
943
944 surf->htile_size = (total_pixels / htile_block_size) *
945 htile_element_size;
946 surf->htile_size = align(surf->htile_size, surf->htile_alignment);
947 }
948
949 surf->is_linear = surf->u.legacy.level[0].mode == RADEON_SURF_MODE_LINEAR_ALIGNED;
950 surf->is_displayable = surf->is_linear ||
951 surf->micro_tile_mode == RADEON_MICRO_MODE_DISPLAY ||
952 surf->micro_tile_mode == RADEON_MICRO_MODE_ROTATED;
953 return 0;
954 }
955
956 /* This is only called when expecting a tiled layout. */
957 static int
958 gfx9_get_preferred_swizzle_mode(ADDR_HANDLE addrlib,
959 ADDR2_COMPUTE_SURFACE_INFO_INPUT *in,
960 bool is_fmask, unsigned flags,
961 AddrSwizzleMode *swizzle_mode)
962 {
963 ADDR_E_RETURNCODE ret;
964 ADDR2_GET_PREFERRED_SURF_SETTING_INPUT sin = {0};
965 ADDR2_GET_PREFERRED_SURF_SETTING_OUTPUT sout = {0};
966
967 sin.size = sizeof(ADDR2_GET_PREFERRED_SURF_SETTING_INPUT);
968 sout.size = sizeof(ADDR2_GET_PREFERRED_SURF_SETTING_OUTPUT);
969
970 sin.flags = in->flags;
971 sin.resourceType = in->resourceType;
972 sin.format = in->format;
973 sin.resourceLoction = ADDR_RSRC_LOC_INVIS;
974 /* TODO: We could allow some of these: */
975 sin.forbiddenBlock.micro = 1; /* don't allow the 256B swizzle modes */
976 sin.forbiddenBlock.var = 1; /* don't allow the variable-sized swizzle modes */
977 sin.forbiddenBlock.linear = 1; /* don't allow linear swizzle modes */
978 sin.bpp = in->bpp;
979 sin.width = in->width;
980 sin.height = in->height;
981 sin.numSlices = in->numSlices;
982 sin.numMipLevels = in->numMipLevels;
983 sin.numSamples = in->numSamples;
984 sin.numFrags = in->numFrags;
985
986 if (flags & RADEON_SURF_SCANOUT) {
987 sin.preferredSwSet.sw_D = 1;
988 /* Raven only allows S for displayable surfaces with < 64 bpp, so
989 * allow it as fallback */
990 sin.preferredSwSet.sw_S = 1;
991 } else if (in->flags.depth || in->flags.stencil || is_fmask)
992 sin.preferredSwSet.sw_Z = 1;
993 else
994 sin.preferredSwSet.sw_S = 1;
995
996 if (is_fmask) {
997 sin.flags.display = 0;
998 sin.flags.color = 0;
999 sin.flags.fmask = 1;
1000 }
1001
1002 ret = Addr2GetPreferredSurfaceSetting(addrlib, &sin, &sout);
1003 if (ret != ADDR_OK)
1004 return ret;
1005
1006 *swizzle_mode = sout.swizzleMode;
1007 return 0;
1008 }
1009
1010 static int gfx9_compute_miptree(ADDR_HANDLE addrlib,
1011 const struct ac_surf_config *config,
1012 struct radeon_surf *surf, bool compressed,
1013 ADDR2_COMPUTE_SURFACE_INFO_INPUT *in)
1014 {
1015 ADDR2_MIP_INFO mip_info[RADEON_SURF_MAX_LEVELS] = {};
1016 ADDR2_COMPUTE_SURFACE_INFO_OUTPUT out = {0};
1017 ADDR_E_RETURNCODE ret;
1018
1019 out.size = sizeof(ADDR2_COMPUTE_SURFACE_INFO_OUTPUT);
1020 out.pMipInfo = mip_info;
1021
1022 ret = Addr2ComputeSurfaceInfo(addrlib, in, &out);
1023 if (ret != ADDR_OK)
1024 return ret;
1025
1026 if (in->flags.stencil) {
1027 surf->u.gfx9.stencil.swizzle_mode = in->swizzleMode;
1028 surf->u.gfx9.stencil.epitch = out.epitchIsHeight ? out.mipChainHeight - 1 :
1029 out.mipChainPitch - 1;
1030 surf->surf_alignment = MAX2(surf->surf_alignment, out.baseAlign);
1031 surf->u.gfx9.stencil_offset = align(surf->surf_size, out.baseAlign);
1032 surf->surf_size = surf->u.gfx9.stencil_offset + out.surfSize;
1033 return 0;
1034 }
1035
1036 surf->u.gfx9.surf.swizzle_mode = in->swizzleMode;
1037 surf->u.gfx9.surf.epitch = out.epitchIsHeight ? out.mipChainHeight - 1 :
1038 out.mipChainPitch - 1;
1039
1040 /* CMASK fast clear uses these even if FMASK isn't allocated.
1041 * FMASK only supports the Z swizzle modes, whose numbers are multiples of 4.
1042 */
1043 surf->u.gfx9.fmask.swizzle_mode = surf->u.gfx9.surf.swizzle_mode & ~0x3;
1044 surf->u.gfx9.fmask.epitch = surf->u.gfx9.surf.epitch;
1045
1046 surf->u.gfx9.surf_slice_size = out.sliceSize;
1047 surf->u.gfx9.surf_pitch = out.pitch;
1048 surf->u.gfx9.surf_height = out.height;
1049 surf->surf_size = out.surfSize;
1050 surf->surf_alignment = out.baseAlign;
1051
1052 if (in->swizzleMode == ADDR_SW_LINEAR) {
1053 for (unsigned i = 0; i < in->numMipLevels; i++)
1054 surf->u.gfx9.offset[i] = mip_info[i].offset;
1055 }
1056
1057 if (in->flags.depth) {
1058 assert(in->swizzleMode != ADDR_SW_LINEAR);
1059
1060 /* HTILE */
1061 ADDR2_COMPUTE_HTILE_INFO_INPUT hin = {0};
1062 ADDR2_COMPUTE_HTILE_INFO_OUTPUT hout = {0};
1063
1064 hin.size = sizeof(ADDR2_COMPUTE_HTILE_INFO_INPUT);
1065 hout.size = sizeof(ADDR2_COMPUTE_HTILE_INFO_OUTPUT);
1066
1067 hin.hTileFlags.pipeAligned = !in->flags.metaPipeUnaligned;
1068 hin.hTileFlags.rbAligned = !in->flags.metaRbUnaligned;
1069 hin.depthFlags = in->flags;
1070 hin.swizzleMode = in->swizzleMode;
1071 hin.unalignedWidth = in->width;
1072 hin.unalignedHeight = in->height;
1073 hin.numSlices = in->numSlices;
1074 hin.numMipLevels = in->numMipLevels;
1075
1076 ret = Addr2ComputeHtileInfo(addrlib, &hin, &hout);
1077 if (ret != ADDR_OK)
1078 return ret;
1079
1080 surf->u.gfx9.htile.rb_aligned = hin.hTileFlags.rbAligned;
1081 surf->u.gfx9.htile.pipe_aligned = hin.hTileFlags.pipeAligned;
1082 surf->htile_size = hout.htileBytes;
1083 surf->htile_slice_size = hout.sliceSize;
1084 surf->htile_alignment = hout.baseAlign;
1085 } else {
1086 /* Compute tile swizzle for the color surface.
1087 * All *_X and *_T modes can use the swizzle.
1088 */
1089 if (config->info.surf_index &&
1090 in->swizzleMode >= ADDR_SW_64KB_Z_T &&
1091 !out.mipChainInTail &&
1092 !(surf->flags & RADEON_SURF_SHAREABLE) &&
1093 !in->flags.display) {
1094 ADDR2_COMPUTE_PIPEBANKXOR_INPUT xin = {0};
1095 ADDR2_COMPUTE_PIPEBANKXOR_OUTPUT xout = {0};
1096
1097 xin.size = sizeof(ADDR2_COMPUTE_PIPEBANKXOR_INPUT);
1098 xout.size = sizeof(ADDR2_COMPUTE_PIPEBANKXOR_OUTPUT);
1099
1100 xin.surfIndex = p_atomic_inc_return(config->info.surf_index) - 1;
1101 xin.flags = in->flags;
1102 xin.swizzleMode = in->swizzleMode;
1103 xin.resourceType = in->resourceType;
1104 xin.format = in->format;
1105 xin.numSamples = in->numSamples;
1106 xin.numFrags = in->numFrags;
1107
1108 ret = Addr2ComputePipeBankXor(addrlib, &xin, &xout);
1109 if (ret != ADDR_OK)
1110 return ret;
1111
1112 assert(xout.pipeBankXor <=
1113 u_bit_consecutive(0, sizeof(surf->tile_swizzle) * 8));
1114 surf->tile_swizzle = xout.pipeBankXor;
1115 }
1116
1117 /* DCC */
1118 if (!(surf->flags & RADEON_SURF_DISABLE_DCC) &&
1119 !compressed &&
1120 in->swizzleMode != ADDR_SW_LINEAR) {
1121 ADDR2_COMPUTE_DCCINFO_INPUT din = {0};
1122 ADDR2_COMPUTE_DCCINFO_OUTPUT dout = {0};
1123 ADDR2_META_MIP_INFO meta_mip_info[RADEON_SURF_MAX_LEVELS] = {};
1124
1125 din.size = sizeof(ADDR2_COMPUTE_DCCINFO_INPUT);
1126 dout.size = sizeof(ADDR2_COMPUTE_DCCINFO_OUTPUT);
1127 dout.pMipInfo = meta_mip_info;
1128
1129 din.dccKeyFlags.pipeAligned = !in->flags.metaPipeUnaligned;
1130 din.dccKeyFlags.rbAligned = !in->flags.metaRbUnaligned;
1131 din.colorFlags = in->flags;
1132 din.resourceType = in->resourceType;
1133 din.swizzleMode = in->swizzleMode;
1134 din.bpp = in->bpp;
1135 din.unalignedWidth = in->width;
1136 din.unalignedHeight = in->height;
1137 din.numSlices = in->numSlices;
1138 din.numFrags = in->numFrags;
1139 din.numMipLevels = in->numMipLevels;
1140 din.dataSurfaceSize = out.surfSize;
1141
1142 ret = Addr2ComputeDccInfo(addrlib, &din, &dout);
1143 if (ret != ADDR_OK)
1144 return ret;
1145
1146 surf->u.gfx9.dcc.rb_aligned = din.dccKeyFlags.rbAligned;
1147 surf->u.gfx9.dcc.pipe_aligned = din.dccKeyFlags.pipeAligned;
1148 surf->u.gfx9.dcc_pitch_max = dout.pitch - 1;
1149 surf->dcc_size = dout.dccRamSize;
1150 surf->dcc_alignment = dout.dccRamBaseAlign;
1151 surf->num_dcc_levels = in->numMipLevels;
1152
1153 /* Disable DCC for levels that are in the mip tail.
1154 *
1155 * There are two issues that this is intended to
1156 * address:
1157 *
1158 * 1. Multiple mip levels may share a cache line. This
1159 * can lead to corruption when switching between
1160 * rendering to different mip levels because the
1161 * RBs don't maintain coherency.
1162 *
1163 * 2. Texturing with metadata after rendering sometimes
1164 * fails with corruption, probably for a similar
1165 * reason.
1166 *
1167 * Working around these issues for all levels in the
1168 * mip tail may be overly conservative, but it's what
1169 * Vulkan does.
1170 *
1171 * Alternative solutions that also work but are worse:
1172 * - Disable DCC entirely.
1173 * - Flush TC L2 after rendering.
1174 */
1175 for (unsigned i = 0; i < in->numMipLevels; i++) {
1176 if (meta_mip_info[i].inMiptail) {
1177 surf->num_dcc_levels = i;
1178 break;
1179 }
1180 }
1181
1182 if (!surf->num_dcc_levels)
1183 surf->dcc_size = 0;
1184 }
1185
1186 /* FMASK */
1187 if (in->numSamples > 1) {
1188 ADDR2_COMPUTE_FMASK_INFO_INPUT fin = {0};
1189 ADDR2_COMPUTE_FMASK_INFO_OUTPUT fout = {0};
1190
1191 fin.size = sizeof(ADDR2_COMPUTE_FMASK_INFO_INPUT);
1192 fout.size = sizeof(ADDR2_COMPUTE_FMASK_INFO_OUTPUT);
1193
1194 ret = gfx9_get_preferred_swizzle_mode(addrlib, in,
1195 true, surf->flags,
1196 &fin.swizzleMode);
1197 if (ret != ADDR_OK)
1198 return ret;
1199
1200 fin.unalignedWidth = in->width;
1201 fin.unalignedHeight = in->height;
1202 fin.numSlices = in->numSlices;
1203 fin.numSamples = in->numSamples;
1204 fin.numFrags = in->numFrags;
1205
1206 ret = Addr2ComputeFmaskInfo(addrlib, &fin, &fout);
1207 if (ret != ADDR_OK)
1208 return ret;
1209
1210 surf->u.gfx9.fmask.swizzle_mode = fin.swizzleMode;
1211 surf->u.gfx9.fmask.epitch = fout.pitch - 1;
1212 surf->fmask_size = fout.fmaskBytes;
1213 surf->fmask_alignment = fout.baseAlign;
1214
1215 /* Compute tile swizzle for the FMASK surface. */
1216 if (config->info.fmask_surf_index &&
1217 fin.swizzleMode >= ADDR_SW_64KB_Z_T &&
1218 !(surf->flags & RADEON_SURF_SHAREABLE)) {
1219 ADDR2_COMPUTE_PIPEBANKXOR_INPUT xin = {0};
1220 ADDR2_COMPUTE_PIPEBANKXOR_OUTPUT xout = {0};
1221
1222 xin.size = sizeof(ADDR2_COMPUTE_PIPEBANKXOR_INPUT);
1223 xout.size = sizeof(ADDR2_COMPUTE_PIPEBANKXOR_OUTPUT);
1224
1225 /* This counter starts from 1 instead of 0. */
1226 xin.surfIndex = p_atomic_inc_return(config->info.fmask_surf_index);
1227 xin.flags = in->flags;
1228 xin.swizzleMode = in->swizzleMode;
1229 xin.resourceType = in->resourceType;
1230 xin.format = in->format;
1231 xin.numSamples = in->numSamples;
1232 xin.numFrags = in->numFrags;
1233
1234 ret = Addr2ComputePipeBankXor(addrlib, &xin, &xout);
1235 if (ret != ADDR_OK)
1236 return ret;
1237
1238 assert(xout.pipeBankXor <=
1239 u_bit_consecutive(0, sizeof(surf->fmask_tile_swizzle) * 8));
1240 surf->fmask_tile_swizzle = xout.pipeBankXor;
1241 }
1242 }
1243
1244 /* CMASK */
1245 if (in->swizzleMode != ADDR_SW_LINEAR) {
1246 ADDR2_COMPUTE_CMASK_INFO_INPUT cin = {0};
1247 ADDR2_COMPUTE_CMASK_INFO_OUTPUT cout = {0};
1248
1249 cin.size = sizeof(ADDR2_COMPUTE_CMASK_INFO_INPUT);
1250 cout.size = sizeof(ADDR2_COMPUTE_CMASK_INFO_OUTPUT);
1251
1252 if (in->numSamples > 1) {
1253 /* FMASK is always aligned. */
1254 cin.cMaskFlags.pipeAligned = 1;
1255 cin.cMaskFlags.rbAligned = 1;
1256 } else {
1257 cin.cMaskFlags.pipeAligned = !in->flags.metaPipeUnaligned;
1258 cin.cMaskFlags.rbAligned = !in->flags.metaRbUnaligned;
1259 }
1260 cin.colorFlags = in->flags;
1261 cin.resourceType = in->resourceType;
1262 cin.unalignedWidth = in->width;
1263 cin.unalignedHeight = in->height;
1264 cin.numSlices = in->numSlices;
1265
1266 if (in->numSamples > 1)
1267 cin.swizzleMode = surf->u.gfx9.fmask.swizzle_mode;
1268 else
1269 cin.swizzleMode = in->swizzleMode;
1270
1271 ret = Addr2ComputeCmaskInfo(addrlib, &cin, &cout);
1272 if (ret != ADDR_OK)
1273 return ret;
1274
1275 surf->u.gfx9.cmask.rb_aligned = cin.cMaskFlags.rbAligned;
1276 surf->u.gfx9.cmask.pipe_aligned = cin.cMaskFlags.pipeAligned;
1277 surf->u.gfx9.cmask_size = cout.cmaskBytes;
1278 surf->u.gfx9.cmask_alignment = cout.baseAlign;
1279 }
1280 }
1281
1282 return 0;
1283 }
1284
1285 static int gfx9_compute_surface(ADDR_HANDLE addrlib,
1286 const struct radeon_info *info,
1287 const struct ac_surf_config *config,
1288 enum radeon_surf_mode mode,
1289 struct radeon_surf *surf)
1290 {
1291 bool compressed;
1292 ADDR2_COMPUTE_SURFACE_INFO_INPUT AddrSurfInfoIn = {0};
1293 int r;
1294
1295 AddrSurfInfoIn.size = sizeof(ADDR2_COMPUTE_SURFACE_INFO_INPUT);
1296
1297 compressed = surf->blk_w == 4 && surf->blk_h == 4;
1298
1299 /* The format must be set correctly for the allocation of compressed
1300 * textures to work. In other cases, setting the bpp is sufficient. */
1301 if (compressed) {
1302 switch (surf->bpe) {
1303 case 8:
1304 AddrSurfInfoIn.format = ADDR_FMT_BC1;
1305 break;
1306 case 16:
1307 AddrSurfInfoIn.format = ADDR_FMT_BC3;
1308 break;
1309 default:
1310 assert(0);
1311 }
1312 } else {
1313 switch (surf->bpe) {
1314 case 1:
1315 assert(!(surf->flags & RADEON_SURF_ZBUFFER));
1316 AddrSurfInfoIn.format = ADDR_FMT_8;
1317 break;
1318 case 2:
1319 assert(surf->flags & RADEON_SURF_ZBUFFER ||
1320 !(surf->flags & RADEON_SURF_SBUFFER));
1321 AddrSurfInfoIn.format = ADDR_FMT_16;
1322 break;
1323 case 4:
1324 assert(surf->flags & RADEON_SURF_ZBUFFER ||
1325 !(surf->flags & RADEON_SURF_SBUFFER));
1326 AddrSurfInfoIn.format = ADDR_FMT_32;
1327 break;
1328 case 8:
1329 assert(!(surf->flags & RADEON_SURF_Z_OR_SBUFFER));
1330 AddrSurfInfoIn.format = ADDR_FMT_32_32;
1331 break;
1332 case 12:
1333 assert(!(surf->flags & RADEON_SURF_Z_OR_SBUFFER));
1334 AddrSurfInfoIn.format = ADDR_FMT_32_32_32;
1335 break;
1336 case 16:
1337 assert(!(surf->flags & RADEON_SURF_Z_OR_SBUFFER));
1338 AddrSurfInfoIn.format = ADDR_FMT_32_32_32_32;
1339 break;
1340 default:
1341 assert(0);
1342 }
1343 AddrSurfInfoIn.bpp = surf->bpe * 8;
1344 }
1345
1346 AddrSurfInfoIn.flags.color = !(surf->flags & RADEON_SURF_Z_OR_SBUFFER);
1347 AddrSurfInfoIn.flags.depth = (surf->flags & RADEON_SURF_ZBUFFER) != 0;
1348 AddrSurfInfoIn.flags.display = get_display_flag(config, surf);
1349 /* flags.texture currently refers to TC-compatible HTILE */
1350 AddrSurfInfoIn.flags.texture = AddrSurfInfoIn.flags.color ||
1351 surf->flags & RADEON_SURF_TC_COMPATIBLE_HTILE;
1352 AddrSurfInfoIn.flags.opt4space = 1;
1353
1354 AddrSurfInfoIn.numMipLevels = config->info.levels;
1355 AddrSurfInfoIn.numSamples = MAX2(1, config->info.samples);
1356 AddrSurfInfoIn.numFrags = AddrSurfInfoIn.numSamples;
1357
1358 if (!(surf->flags & RADEON_SURF_Z_OR_SBUFFER))
1359 AddrSurfInfoIn.numFrags = MAX2(1, config->info.color_samples);
1360
1361 /* GFX9 doesn't support 1D depth textures, so allocate all 1D textures
1362 * as 2D to avoid having shader variants for 1D vs 2D, so all shaders
1363 * must sample 1D textures as 2D. */
1364 if (config->is_3d)
1365 AddrSurfInfoIn.resourceType = ADDR_RSRC_TEX_3D;
1366 else
1367 AddrSurfInfoIn.resourceType = ADDR_RSRC_TEX_2D;
1368
1369 AddrSurfInfoIn.width = config->info.width;
1370 AddrSurfInfoIn.height = config->info.height;
1371
1372 if (config->is_3d)
1373 AddrSurfInfoIn.numSlices = config->info.depth;
1374 else if (config->is_cube)
1375 AddrSurfInfoIn.numSlices = 6;
1376 else
1377 AddrSurfInfoIn.numSlices = config->info.array_size;
1378
1379 /* This is propagated to HTILE/DCC/CMASK. */
1380 AddrSurfInfoIn.flags.metaPipeUnaligned = 0;
1381 AddrSurfInfoIn.flags.metaRbUnaligned = 0;
1382
1383 switch (mode) {
1384 case RADEON_SURF_MODE_LINEAR_ALIGNED:
1385 assert(config->info.samples <= 1);
1386 assert(!(surf->flags & RADEON_SURF_Z_OR_SBUFFER));
1387 AddrSurfInfoIn.swizzleMode = ADDR_SW_LINEAR;
1388 break;
1389
1390 case RADEON_SURF_MODE_1D:
1391 case RADEON_SURF_MODE_2D:
1392 if (surf->flags & RADEON_SURF_IMPORTED) {
1393 AddrSurfInfoIn.swizzleMode = surf->u.gfx9.surf.swizzle_mode;
1394 break;
1395 }
1396
1397 r = gfx9_get_preferred_swizzle_mode(addrlib, &AddrSurfInfoIn,
1398 false, surf->flags,
1399 &AddrSurfInfoIn.swizzleMode);
1400 if (r)
1401 return r;
1402 break;
1403
1404 default:
1405 assert(0);
1406 }
1407
1408 surf->u.gfx9.resource_type = AddrSurfInfoIn.resourceType;
1409 surf->has_stencil = !!(surf->flags & RADEON_SURF_SBUFFER);
1410
1411 surf->num_dcc_levels = 0;
1412 surf->surf_size = 0;
1413 surf->fmask_size = 0;
1414 surf->dcc_size = 0;
1415 surf->htile_size = 0;
1416 surf->htile_slice_size = 0;
1417 surf->u.gfx9.surf_offset = 0;
1418 surf->u.gfx9.stencil_offset = 0;
1419 surf->u.gfx9.cmask_size = 0;
1420
1421 /* Calculate texture layout information. */
1422 r = gfx9_compute_miptree(addrlib, config, surf, compressed,
1423 &AddrSurfInfoIn);
1424 if (r)
1425 return r;
1426
1427 /* Calculate texture layout information for stencil. */
1428 if (surf->flags & RADEON_SURF_SBUFFER) {
1429 AddrSurfInfoIn.flags.stencil = 1;
1430 AddrSurfInfoIn.bpp = 8;
1431 AddrSurfInfoIn.format = ADDR_FMT_8;
1432
1433 if (!AddrSurfInfoIn.flags.depth) {
1434 r = gfx9_get_preferred_swizzle_mode(addrlib, &AddrSurfInfoIn,
1435 false, surf->flags,
1436 &AddrSurfInfoIn.swizzleMode);
1437 if (r)
1438 return r;
1439 } else
1440 AddrSurfInfoIn.flags.depth = 0;
1441
1442 r = gfx9_compute_miptree(addrlib, config, surf, compressed,
1443 &AddrSurfInfoIn);
1444 if (r)
1445 return r;
1446 }
1447
1448 surf->is_linear = surf->u.gfx9.surf.swizzle_mode == ADDR_SW_LINEAR;
1449
1450 /* Query whether the surface is displayable. */
1451 bool displayable = false;
1452 r = Addr2IsValidDisplaySwizzleMode(addrlib, surf->u.gfx9.surf.swizzle_mode,
1453 surf->bpe * 8, &displayable);
1454 if (r)
1455 return r;
1456 surf->is_displayable = displayable;
1457
1458 switch (surf->u.gfx9.surf.swizzle_mode) {
1459 /* S = standard. */
1460 case ADDR_SW_256B_S:
1461 case ADDR_SW_4KB_S:
1462 case ADDR_SW_64KB_S:
1463 case ADDR_SW_VAR_S:
1464 case ADDR_SW_64KB_S_T:
1465 case ADDR_SW_4KB_S_X:
1466 case ADDR_SW_64KB_S_X:
1467 case ADDR_SW_VAR_S_X:
1468 surf->micro_tile_mode = RADEON_MICRO_MODE_THIN;
1469 break;
1470
1471 /* D = display. */
1472 case ADDR_SW_LINEAR:
1473 case ADDR_SW_256B_D:
1474 case ADDR_SW_4KB_D:
1475 case ADDR_SW_64KB_D:
1476 case ADDR_SW_VAR_D:
1477 case ADDR_SW_64KB_D_T:
1478 case ADDR_SW_4KB_D_X:
1479 case ADDR_SW_64KB_D_X:
1480 case ADDR_SW_VAR_D_X:
1481 surf->micro_tile_mode = RADEON_MICRO_MODE_DISPLAY;
1482 break;
1483
1484 /* R = rotated. */
1485 case ADDR_SW_256B_R:
1486 case ADDR_SW_4KB_R:
1487 case ADDR_SW_64KB_R:
1488 case ADDR_SW_VAR_R:
1489 case ADDR_SW_64KB_R_T:
1490 case ADDR_SW_4KB_R_X:
1491 case ADDR_SW_64KB_R_X:
1492 case ADDR_SW_VAR_R_X:
1493 surf->micro_tile_mode = RADEON_MICRO_MODE_ROTATED;
1494 break;
1495
1496 /* Z = depth. */
1497 case ADDR_SW_4KB_Z:
1498 case ADDR_SW_64KB_Z:
1499 case ADDR_SW_VAR_Z:
1500 case ADDR_SW_64KB_Z_T:
1501 case ADDR_SW_4KB_Z_X:
1502 case ADDR_SW_64KB_Z_X:
1503 case ADDR_SW_VAR_Z_X:
1504 surf->micro_tile_mode = RADEON_MICRO_MODE_DEPTH;
1505 break;
1506
1507 default:
1508 assert(0);
1509 }
1510
1511 /* Temporary workaround to prevent VM faults and hangs. */
1512 if (info->family == CHIP_VEGA12)
1513 surf->fmask_size *= 8;
1514
1515 return 0;
1516 }
1517
1518 int ac_compute_surface(ADDR_HANDLE addrlib, const struct radeon_info *info,
1519 const struct ac_surf_config *config,
1520 enum radeon_surf_mode mode,
1521 struct radeon_surf *surf)
1522 {
1523 int r;
1524
1525 r = surf_config_sanity(config, surf->flags);
1526 if (r)
1527 return r;
1528
1529 if (info->chip_class >= GFX9)
1530 return gfx9_compute_surface(addrlib, info, config, mode, surf);
1531 else
1532 return gfx6_compute_surface(addrlib, info, config, mode, surf);
1533 }