1 /*
2 * Copyright © 2014-2017 Broadcom
3 * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
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
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 * IN THE SOFTWARE.
23 */
24
25 #include "pipe/p_defines.h"
26 #include "util/u_memory.h"
27 #include "util/format/u_format.h"
28 #include "util/u_inlines.h"
29 #include "util/u_surface.h"
30 #include "util/u_transfer_helper.h"
31 #include "util/u_upload_mgr.h"
32 #include "util/format/u_format_zs.h"
33 #include "util/u_drm.h"
34
35 #include "drm-uapi/drm_fourcc.h"
36 #include "v3d_screen.h"
37 #include "v3d_context.h"
38 #include "v3d_resource.h"
39 #include "v3d_tiling.h"
40 #include "broadcom/cle/v3d_packet_v33_pack.h"
41
42 static void
v3d_debug_resource_layout(struct v3d_resource * rsc,const char * caller)43 v3d_debug_resource_layout(struct v3d_resource *rsc, const char *caller)
44 {
45 if (!(V3D_DEBUG & V3D_DEBUG_SURFACE))
46 return;
47
48 struct pipe_resource *prsc = &rsc->base;
49
50 if (prsc->target == PIPE_BUFFER) {
51 fprintf(stderr,
52 "rsc %s %p (format %s), %dx%d buffer @0x%08x-0x%08x\n",
53 caller, rsc,
54 util_format_short_name(prsc->format),
55 prsc->width0, prsc->height0,
56 rsc->bo->offset,
57 rsc->bo->offset + rsc->bo->size - 1);
58 return;
59 }
60
61 static const char *const tiling_descriptions[] = {
62 [VC5_TILING_RASTER] = "R",
63 [VC5_TILING_LINEARTILE] = "LT",
64 [VC5_TILING_UBLINEAR_1_COLUMN] = "UB1",
65 [VC5_TILING_UBLINEAR_2_COLUMN] = "UB2",
66 [VC5_TILING_UIF_NO_XOR] = "UIF",
67 [VC5_TILING_UIF_XOR] = "UIF^",
68 };
69
70 for (int i = 0; i <= prsc->last_level; i++) {
71 struct v3d_resource_slice *slice = &rsc->slices[i];
72
73 int level_width = slice->stride / rsc->cpp;
74 int level_height = slice->padded_height;
75 int level_depth =
76 u_minify(util_next_power_of_two(prsc->depth0), i);
77
78 fprintf(stderr,
79 "rsc %s %p (format %s), %dx%d: "
80 "level %d (%s) %dx%dx%d -> %dx%dx%d, stride %d@0x%08x\n",
81 caller, rsc,
82 util_format_short_name(prsc->format),
83 prsc->width0, prsc->height0,
84 i, tiling_descriptions[slice->tiling],
85 u_minify(prsc->width0, i),
86 u_minify(prsc->height0, i),
87 u_minify(prsc->depth0, i),
88 level_width,
89 level_height,
90 level_depth,
91 slice->stride,
92 rsc->bo->offset + slice->offset);
93 }
94 }
95
96 static bool
v3d_resource_bo_alloc(struct v3d_resource * rsc)97 v3d_resource_bo_alloc(struct v3d_resource *rsc)
98 {
99 struct pipe_resource *prsc = &rsc->base;
100 struct pipe_screen *pscreen = prsc->screen;
101 struct v3d_bo *bo;
102
103 bo = v3d_bo_alloc(v3d_screen(pscreen), rsc->size, "resource");
104 if (bo) {
105 v3d_bo_unreference(&rsc->bo);
106 rsc->bo = bo;
107 v3d_debug_resource_layout(rsc, "alloc");
108 return true;
109 } else {
110 return false;
111 }
112 }
113
114 static void
v3d_resource_transfer_unmap(struct pipe_context * pctx,struct pipe_transfer * ptrans)115 v3d_resource_transfer_unmap(struct pipe_context *pctx,
116 struct pipe_transfer *ptrans)
117 {
118 struct v3d_context *v3d = v3d_context(pctx);
119 struct v3d_transfer *trans = v3d_transfer(ptrans);
120
121 if (trans->map) {
122 struct v3d_resource *rsc = v3d_resource(ptrans->resource);
123 struct v3d_resource_slice *slice = &rsc->slices[ptrans->level];
124
125 if (ptrans->usage & PIPE_MAP_WRITE) {
126 for (int z = 0; z < ptrans->box.depth; z++) {
127 void *dst = rsc->bo->map +
128 v3d_layer_offset(&rsc->base,
129 ptrans->level,
130 ptrans->box.z + z);
131 v3d_store_tiled_image(dst,
132 slice->stride,
133 (trans->map +
134 ptrans->stride *
135 ptrans->box.height * z),
136 ptrans->stride,
137 slice->tiling, rsc->cpp,
138 slice->padded_height,
139 &ptrans->box);
140 }
141 }
142 free(trans->map);
143 }
144
145 pipe_resource_reference(&ptrans->resource, NULL);
146 slab_free(&v3d->transfer_pool, ptrans);
147 }
148
149 static void
rebind_sampler_views(struct v3d_context * v3d,struct v3d_resource * rsc)150 rebind_sampler_views(struct v3d_context *v3d,
151 struct v3d_resource *rsc)
152 {
153 for (int st = 0; st < PIPE_SHADER_TYPES; st++) {
154 struct v3d_texture_stateobj *tex = v3d->tex + st;
155
156 for (unsigned i = 0; i < tex->num_textures; i++) {
157 struct pipe_sampler_view *psview = tex->textures[i];
158
159 if (psview->texture != &rsc->base)
160 continue;
161
162 struct v3d_sampler_view *sview =
163 v3d_sampler_view(psview);
164
165 v3d_create_texture_shader_state_bo(v3d, sview);
166
167 v3d_flag_dirty_sampler_state(v3d, st);
168 }
169 }
170 }
171
172 static void
v3d_map_usage_prep(struct pipe_context * pctx,struct pipe_resource * prsc,unsigned usage)173 v3d_map_usage_prep(struct pipe_context *pctx,
174 struct pipe_resource *prsc,
175 unsigned usage)
176 {
177 struct v3d_context *v3d = v3d_context(pctx);
178 struct v3d_resource *rsc = v3d_resource(prsc);
179
180 if (usage & PIPE_MAP_DISCARD_WHOLE_RESOURCE) {
181 if (v3d_resource_bo_alloc(rsc)) {
182 /* If it might be bound as one of our vertex buffers
183 * or UBOs, make sure we re-emit vertex buffer state
184 * or uniforms.
185 */
186 if (prsc->bind & PIPE_BIND_VERTEX_BUFFER)
187 v3d->dirty |= VC5_DIRTY_VTXBUF;
188 if (prsc->bind & PIPE_BIND_CONSTANT_BUFFER)
189 v3d->dirty |= VC5_DIRTY_CONSTBUF;
190 if (prsc->bind & PIPE_BIND_SAMPLER_VIEW)
191 rebind_sampler_views(v3d, rsc);
192 } else {
193 /* If we failed to reallocate, flush users so that we
194 * don't violate any syncing requirements.
195 */
196 v3d_flush_jobs_reading_resource(v3d, prsc,
197 V3D_FLUSH_DEFAULT,
198 false);
199 }
200 } else if (!(usage & PIPE_MAP_UNSYNCHRONIZED)) {
201 /* If we're writing and the buffer is being used by the CL, we
202 * have to flush the CL first. If we're only reading, we need
203 * to flush if the CL has written our buffer.
204 */
205 if (usage & PIPE_MAP_WRITE) {
206 v3d_flush_jobs_reading_resource(v3d, prsc,
207 V3D_FLUSH_ALWAYS,
208 false);
209 } else {
210 v3d_flush_jobs_writing_resource(v3d, prsc,
211 V3D_FLUSH_ALWAYS,
212 false);
213 }
214 }
215
216 if (usage & PIPE_MAP_WRITE) {
217 rsc->writes++;
218 rsc->initialized_buffers = ~0;
219 }
220 }
221
222 static void *
v3d_resource_transfer_map(struct pipe_context * pctx,struct pipe_resource * prsc,unsigned level,unsigned usage,const struct pipe_box * box,struct pipe_transfer ** pptrans)223 v3d_resource_transfer_map(struct pipe_context *pctx,
224 struct pipe_resource *prsc,
225 unsigned level, unsigned usage,
226 const struct pipe_box *box,
227 struct pipe_transfer **pptrans)
228 {
229 struct v3d_context *v3d = v3d_context(pctx);
230 struct v3d_resource *rsc = v3d_resource(prsc);
231 struct v3d_transfer *trans;
232 struct pipe_transfer *ptrans;
233 enum pipe_format format = prsc->format;
234 char *buf;
235
236 /* MSAA maps should have been handled by u_transfer_helper. */
237 assert(prsc->nr_samples <= 1);
238
239 /* Upgrade DISCARD_RANGE to WHOLE_RESOURCE if the whole resource is
240 * being mapped.
241 */
242 if ((usage & PIPE_MAP_DISCARD_RANGE) &&
243 !(usage & PIPE_MAP_UNSYNCHRONIZED) &&
244 !(prsc->flags & PIPE_RESOURCE_FLAG_MAP_PERSISTENT) &&
245 prsc->last_level == 0 &&
246 prsc->width0 == box->width &&
247 prsc->height0 == box->height &&
248 prsc->depth0 == box->depth &&
249 prsc->array_size == 1 &&
250 rsc->bo->private) {
251 usage |= PIPE_MAP_DISCARD_WHOLE_RESOURCE;
252 }
253
254 v3d_map_usage_prep(pctx, prsc, usage);
255
256 trans = slab_alloc(&v3d->transfer_pool);
257 if (!trans)
258 return NULL;
259
260 /* XXX: Handle DONTBLOCK, DISCARD_RANGE, PERSISTENT, COHERENT. */
261
262 /* slab_alloc_st() doesn't zero: */
263 memset(trans, 0, sizeof(*trans));
264 ptrans = &trans->base;
265
266 pipe_resource_reference(&ptrans->resource, prsc);
267 ptrans->level = level;
268 ptrans->usage = usage;
269 ptrans->box = *box;
270
271 /* Note that the current kernel implementation is synchronous, so no
272 * need to do syncing stuff here yet.
273 */
274
275 if (usage & PIPE_MAP_UNSYNCHRONIZED)
276 buf = v3d_bo_map_unsynchronized(rsc->bo);
277 else
278 buf = v3d_bo_map(rsc->bo);
279 if (!buf) {
280 fprintf(stderr, "Failed to map bo\n");
281 goto fail;
282 }
283
284 *pptrans = ptrans;
285
286 /* Our load/store routines work on entire compressed blocks. */
287 ptrans->box.x /= util_format_get_blockwidth(format);
288 ptrans->box.y /= util_format_get_blockheight(format);
289 ptrans->box.width = DIV_ROUND_UP(ptrans->box.width,
290 util_format_get_blockwidth(format));
291 ptrans->box.height = DIV_ROUND_UP(ptrans->box.height,
292 util_format_get_blockheight(format));
293
294 struct v3d_resource_slice *slice = &rsc->slices[level];
295 if (rsc->tiled) {
296 /* No direct mappings of tiled, since we need to manually
297 * tile/untile.
298 */
299 if (usage & PIPE_MAP_DIRECTLY)
300 return NULL;
301
302 ptrans->stride = ptrans->box.width * rsc->cpp;
303 ptrans->layer_stride = ptrans->stride * ptrans->box.height;
304
305 trans->map = malloc(ptrans->layer_stride * ptrans->box.depth);
306
307 if (usage & PIPE_MAP_READ) {
308 for (int z = 0; z < ptrans->box.depth; z++) {
309 void *src = rsc->bo->map +
310 v3d_layer_offset(&rsc->base,
311 ptrans->level,
312 ptrans->box.z + z);
313 v3d_load_tiled_image((trans->map +
314 ptrans->stride *
315 ptrans->box.height * z),
316 ptrans->stride,
317 src,
318 slice->stride,
319 slice->tiling, rsc->cpp,
320 slice->padded_height,
321 &ptrans->box);
322 }
323 }
324 return trans->map;
325 } else {
326 ptrans->stride = slice->stride;
327 ptrans->layer_stride = rsc->cube_map_stride;
328
329 return buf + slice->offset +
330 ptrans->box.y * ptrans->stride +
331 ptrans->box.x * rsc->cpp +
332 ptrans->box.z * rsc->cube_map_stride;
333 }
334
335
336 fail:
337 v3d_resource_transfer_unmap(pctx, ptrans);
338 return NULL;
339 }
340
341 static void
v3d_texture_subdata(struct pipe_context * pctx,struct pipe_resource * prsc,unsigned level,unsigned usage,const struct pipe_box * box,const void * data,unsigned stride,unsigned layer_stride)342 v3d_texture_subdata(struct pipe_context *pctx,
343 struct pipe_resource *prsc,
344 unsigned level,
345 unsigned usage,
346 const struct pipe_box *box,
347 const void *data,
348 unsigned stride,
349 unsigned layer_stride)
350 {
351 struct v3d_resource *rsc = v3d_resource(prsc);
352 struct v3d_resource_slice *slice = &rsc->slices[level];
353
354 /* For a direct mapping, we can just take the u_transfer path. */
355 if (!rsc->tiled) {
356 return u_default_texture_subdata(pctx, prsc, level, usage, box,
357 data, stride, layer_stride);
358 }
359
360 /* Otherwise, map and store the texture data directly into the tiled
361 * texture. Note that gallium's texture_subdata may be called with
362 * obvious usage flags missing!
363 */
364 v3d_map_usage_prep(pctx, prsc, usage | (PIPE_MAP_WRITE |
365 PIPE_MAP_DISCARD_RANGE));
366
367 void *buf;
368 if (usage & PIPE_MAP_UNSYNCHRONIZED)
369 buf = v3d_bo_map_unsynchronized(rsc->bo);
370 else
371 buf = v3d_bo_map(rsc->bo);
372
373 for (int i = 0; i < box->depth; i++) {
374 v3d_store_tiled_image(buf +
375 v3d_layer_offset(&rsc->base,
376 level,
377 box->z + i),
378 slice->stride,
379 (void *)data + layer_stride * i,
380 stride,
381 slice->tiling, rsc->cpp, slice->padded_height,
382 box);
383 }
384 }
385
386 static void
v3d_resource_destroy(struct pipe_screen * pscreen,struct pipe_resource * prsc)387 v3d_resource_destroy(struct pipe_screen *pscreen,
388 struct pipe_resource *prsc)
389 {
390 struct v3d_screen *screen = v3d_screen(pscreen);
391 struct v3d_resource *rsc = v3d_resource(prsc);
392
393 if (rsc->scanout)
394 renderonly_scanout_destroy(rsc->scanout, screen->ro);
395
396 v3d_bo_unreference(&rsc->bo);
397 free(rsc);
398 }
399
400 static bool
v3d_resource_get_handle(struct pipe_screen * pscreen,struct pipe_context * pctx,struct pipe_resource * prsc,struct winsys_handle * whandle,unsigned usage)401 v3d_resource_get_handle(struct pipe_screen *pscreen,
402 struct pipe_context *pctx,
403 struct pipe_resource *prsc,
404 struct winsys_handle *whandle,
405 unsigned usage)
406 {
407 struct v3d_screen *screen = v3d_screen(pscreen);
408 struct v3d_resource *rsc = v3d_resource(prsc);
409 struct v3d_bo *bo = rsc->bo;
410
411 whandle->stride = rsc->slices[0].stride;
412 whandle->offset = 0;
413
414 /* If we're passing some reference to our BO out to some other part of
415 * the system, then we can't do any optimizations about only us being
416 * the ones seeing it (like BO caching).
417 */
418 bo->private = false;
419
420 if (rsc->tiled) {
421 /* A shared tiled buffer should always be allocated as UIF,
422 * not UBLINEAR or LT.
423 */
424 assert(rsc->slices[0].tiling == VC5_TILING_UIF_XOR ||
425 rsc->slices[0].tiling == VC5_TILING_UIF_NO_XOR);
426 whandle->modifier = DRM_FORMAT_MOD_BROADCOM_UIF;
427 } else {
428 whandle->modifier = DRM_FORMAT_MOD_LINEAR;
429 }
430
431 switch (whandle->type) {
432 case WINSYS_HANDLE_TYPE_SHARED:
433 return v3d_bo_flink(bo, &whandle->handle);
434 case WINSYS_HANDLE_TYPE_KMS:
435 if (screen->ro) {
436 assert(rsc->scanout);
437 bool ok = renderonly_get_handle(rsc->scanout, whandle);
438 whandle->stride = rsc->slices[0].stride;
439 return ok;
440 }
441 whandle->handle = bo->handle;
442 return true;
443 case WINSYS_HANDLE_TYPE_FD:
444 whandle->handle = v3d_bo_get_dmabuf(bo);
445 return whandle->handle != -1;
446 }
447
448 return false;
449 }
450
451 #define PAGE_UB_ROWS (VC5_UIFCFG_PAGE_SIZE / VC5_UIFBLOCK_ROW_SIZE)
452 #define PAGE_UB_ROWS_TIMES_1_5 ((PAGE_UB_ROWS * 3) >> 1)
453 #define PAGE_CACHE_UB_ROWS (VC5_PAGE_CACHE_SIZE / VC5_UIFBLOCK_ROW_SIZE)
454 #define PAGE_CACHE_MINUS_1_5_UB_ROWS (PAGE_CACHE_UB_ROWS - PAGE_UB_ROWS_TIMES_1_5)
455
456 /**
457 * Computes the HW's UIFblock padding for a given height/cpp.
458 *
459 * The goal of the padding is to keep pages of the same color (bank number) at
460 * least half a page away from each other vertically when crossing between
461 * between columns of UIF blocks.
462 */
463 static uint32_t
v3d_get_ub_pad(struct v3d_resource * rsc,uint32_t height)464 v3d_get_ub_pad(struct v3d_resource *rsc, uint32_t height)
465 {
466 uint32_t utile_h = v3d_utile_height(rsc->cpp);
467 uint32_t uif_block_h = utile_h * 2;
468 uint32_t height_ub = height / uif_block_h;
469
470 uint32_t height_offset_in_pc = height_ub % PAGE_CACHE_UB_ROWS;
471
472 /* For the perfectly-aligned-for-UIF-XOR case, don't add any pad. */
473 if (height_offset_in_pc == 0)
474 return 0;
475
476 /* Try padding up to where we're offset by at least half a page. */
477 if (height_offset_in_pc < PAGE_UB_ROWS_TIMES_1_5) {
478 /* If we fit entirely in the page cache, don't pad. */
479 if (height_ub < PAGE_CACHE_UB_ROWS)
480 return 0;
481 else
482 return PAGE_UB_ROWS_TIMES_1_5 - height_offset_in_pc;
483 }
484
485 /* If we're close to being aligned to page cache size, then round up
486 * and rely on XOR.
487 */
488 if (height_offset_in_pc > PAGE_CACHE_MINUS_1_5_UB_ROWS)
489 return PAGE_CACHE_UB_ROWS - height_offset_in_pc;
490
491 /* Otherwise, we're far enough away (top and bottom) to not need any
492 * padding.
493 */
494 return 0;
495 }
496
497 static void
v3d_setup_slices(struct v3d_resource * rsc,uint32_t winsys_stride,bool uif_top)498 v3d_setup_slices(struct v3d_resource *rsc, uint32_t winsys_stride,
499 bool uif_top)
500 {
501 struct pipe_resource *prsc = &rsc->base;
502 uint32_t width = prsc->width0;
503 uint32_t height = prsc->height0;
504 uint32_t depth = prsc->depth0;
505 /* Note that power-of-two padding is based on level 1. These are not
506 * equivalent to just util_next_power_of_two(dimension), because at a
507 * level 0 dimension of 9, the level 1 power-of-two padded value is 4,
508 * not 8.
509 */
510 uint32_t pot_width = 2 * util_next_power_of_two(u_minify(width, 1));
511 uint32_t pot_height = 2 * util_next_power_of_two(u_minify(height, 1));
512 uint32_t pot_depth = 2 * util_next_power_of_two(u_minify(depth, 1));
513 uint32_t offset = 0;
514 uint32_t utile_w = v3d_utile_width(rsc->cpp);
515 uint32_t utile_h = v3d_utile_height(rsc->cpp);
516 uint32_t uif_block_w = utile_w * 2;
517 uint32_t uif_block_h = utile_h * 2;
518 uint32_t block_width = util_format_get_blockwidth(prsc->format);
519 uint32_t block_height = util_format_get_blockheight(prsc->format);
520 bool msaa = prsc->nr_samples > 1;
521
522 /* MSAA textures/renderbuffers are always laid out as single-level
523 * UIF.
524 */
525 uif_top |= msaa;
526
527 /* Check some easy mistakes to make in a resource_create() call that
528 * will break our setup.
529 */
530 assert(prsc->array_size != 0);
531 assert(prsc->depth0 != 0);
532
533 for (int i = prsc->last_level; i >= 0; i--) {
534 struct v3d_resource_slice *slice = &rsc->slices[i];
535
536 uint32_t level_width, level_height, level_depth;
537 if (i < 2) {
538 level_width = u_minify(width, i);
539 level_height = u_minify(height, i);
540 } else {
541 level_width = u_minify(pot_width, i);
542 level_height = u_minify(pot_height, i);
543 }
544 if (i < 1)
545 level_depth = u_minify(depth, i);
546 else
547 level_depth = u_minify(pot_depth, i);
548
549 if (msaa) {
550 level_width *= 2;
551 level_height *= 2;
552 }
553
554 level_width = DIV_ROUND_UP(level_width, block_width);
555 level_height = DIV_ROUND_UP(level_height, block_height);
556
557 if (!rsc->tiled) {
558 slice->tiling = VC5_TILING_RASTER;
559 if (prsc->target == PIPE_TEXTURE_1D)
560 level_width = align(level_width, 64 / rsc->cpp);
561 } else {
562 if ((i != 0 || !uif_top) &&
563 (level_width <= utile_w ||
564 level_height <= utile_h)) {
565 slice->tiling = VC5_TILING_LINEARTILE;
566 level_width = align(level_width, utile_w);
567 level_height = align(level_height, utile_h);
568 } else if ((i != 0 || !uif_top) &&
569 level_width <= uif_block_w) {
570 slice->tiling = VC5_TILING_UBLINEAR_1_COLUMN;
571 level_width = align(level_width, uif_block_w);
572 level_height = align(level_height, uif_block_h);
573 } else if ((i != 0 || !uif_top) &&
574 level_width <= 2 * uif_block_w) {
575 slice->tiling = VC5_TILING_UBLINEAR_2_COLUMN;
576 level_width = align(level_width, 2 * uif_block_w);
577 level_height = align(level_height, uif_block_h);
578 } else {
579 /* We align the width to a 4-block column of
580 * UIF blocks, but we only align height to UIF
581 * blocks.
582 */
583 level_width = align(level_width,
584 4 * uif_block_w);
585 level_height = align(level_height,
586 uif_block_h);
587
588 slice->ub_pad = v3d_get_ub_pad(rsc,
589 level_height);
590 level_height += slice->ub_pad * uif_block_h;
591
592 /* If the padding set us to to be aligned to
593 * the page cache size, then the HW will use
594 * the XOR bit on odd columns to get us
595 * perfectly misaligned
596 */
597 if ((level_height / uif_block_h) %
598 (VC5_PAGE_CACHE_SIZE /
599 VC5_UIFBLOCK_ROW_SIZE) == 0) {
600 slice->tiling = VC5_TILING_UIF_XOR;
601 } else {
602 slice->tiling = VC5_TILING_UIF_NO_XOR;
603 }
604 }
605 }
606
607 slice->offset = offset;
608 if (winsys_stride)
609 slice->stride = winsys_stride;
610 else
611 slice->stride = level_width * rsc->cpp;
612 slice->padded_height = level_height;
613 slice->size = level_height * slice->stride;
614
615 uint32_t slice_total_size = slice->size * level_depth;
616
617 /* The HW aligns level 1's base to a page if any of level 1 or
618 * below could be UIF XOR. The lower levels then inherit the
619 * alignment for as long as necesary, thanks to being power of
620 * two aligned.
621 */
622 if (i == 1 &&
623 level_width > 4 * uif_block_w &&
624 level_height > PAGE_CACHE_MINUS_1_5_UB_ROWS * uif_block_h) {
625 slice_total_size = align(slice_total_size,
626 VC5_UIFCFG_PAGE_SIZE);
627 }
628
629 offset += slice_total_size;
630
631 }
632 rsc->size = offset;
633
634 /* UIF/UBLINEAR levels need to be aligned to UIF-blocks, and LT only
635 * needs to be aligned to utile boundaries. Since tiles are laid out
636 * from small to big in memory, we need to align the later UIF slices
637 * to UIF blocks, if they were preceded by non-UIF-block-aligned LT
638 * slices.
639 *
640 * We additionally align to 4k, which improves UIF XOR performance.
641 */
642 uint32_t page_align_offset = (align(rsc->slices[0].offset, 4096) -
643 rsc->slices[0].offset);
644 if (page_align_offset) {
645 rsc->size += page_align_offset;
646 for (int i = 0; i <= prsc->last_level; i++)
647 rsc->slices[i].offset += page_align_offset;
648 }
649
650 /* Arrays and cube textures have a stride which is the distance from
651 * one full mipmap tree to the next (64b aligned). For 3D textures,
652 * we need to program the stride between slices of miplevel 0.
653 */
654 if (prsc->target != PIPE_TEXTURE_3D) {
655 rsc->cube_map_stride = align(rsc->slices[0].offset +
656 rsc->slices[0].size, 64);
657 rsc->size += rsc->cube_map_stride * (prsc->array_size - 1);
658 } else {
659 rsc->cube_map_stride = rsc->slices[0].size;
660 }
661 }
662
663 uint32_t
v3d_layer_offset(struct pipe_resource * prsc,uint32_t level,uint32_t layer)664 v3d_layer_offset(struct pipe_resource *prsc, uint32_t level, uint32_t layer)
665 {
666 struct v3d_resource *rsc = v3d_resource(prsc);
667 struct v3d_resource_slice *slice = &rsc->slices[level];
668
669 if (prsc->target == PIPE_TEXTURE_3D)
670 return slice->offset + layer * slice->size;
671 else
672 return slice->offset + layer * rsc->cube_map_stride;
673 }
674
675 static struct v3d_resource *
v3d_resource_setup(struct pipe_screen * pscreen,const struct pipe_resource * tmpl)676 v3d_resource_setup(struct pipe_screen *pscreen,
677 const struct pipe_resource *tmpl)
678 {
679 struct v3d_screen *screen = v3d_screen(pscreen);
680 struct v3d_resource *rsc = CALLOC_STRUCT(v3d_resource);
681 if (!rsc)
682 return NULL;
683 struct pipe_resource *prsc = &rsc->base;
684
685 *prsc = *tmpl;
686
687 pipe_reference_init(&prsc->reference, 1);
688 prsc->screen = pscreen;
689
690 if (prsc->nr_samples <= 1 ||
691 screen->devinfo.ver >= 40 ||
692 util_format_is_depth_or_stencil(prsc->format)) {
693 rsc->cpp = util_format_get_blocksize(prsc->format);
694 if (screen->devinfo.ver < 40 && prsc->nr_samples > 1)
695 rsc->cpp *= prsc->nr_samples;
696 } else {
697 assert(v3d_rt_format_supported(&screen->devinfo, prsc->format));
698 uint32_t output_image_format =
699 v3d_get_rt_format(&screen->devinfo, prsc->format);
700 uint32_t internal_type;
701 uint32_t internal_bpp;
702 v3d_get_internal_type_bpp_for_output_format(&screen->devinfo,
703 output_image_format,
704 &internal_type,
705 &internal_bpp);
706 switch (internal_bpp) {
707 case V3D_INTERNAL_BPP_32:
708 rsc->cpp = 4;
709 break;
710 case V3D_INTERNAL_BPP_64:
711 rsc->cpp = 8;
712 break;
713 case V3D_INTERNAL_BPP_128:
714 rsc->cpp = 16;
715 break;
716 }
717 }
718
719 assert(rsc->cpp);
720
721 return rsc;
722 }
723
724 static struct pipe_resource *
v3d_resource_create_with_modifiers(struct pipe_screen * pscreen,const struct pipe_resource * tmpl,const uint64_t * modifiers,int count)725 v3d_resource_create_with_modifiers(struct pipe_screen *pscreen,
726 const struct pipe_resource *tmpl,
727 const uint64_t *modifiers,
728 int count)
729 {
730 struct v3d_screen *screen = v3d_screen(pscreen);
731
732 bool linear_ok = drm_find_modifier(DRM_FORMAT_MOD_LINEAR, modifiers, count);
733 struct v3d_resource *rsc = v3d_resource_setup(pscreen, tmpl);
734 struct pipe_resource *prsc = &rsc->base;
735 /* Use a tiled layout if we can, for better 3D performance. */
736 bool should_tile = true;
737
738 /* VBOs/PBOs are untiled (and 1 height). */
739 if (tmpl->target == PIPE_BUFFER)
740 should_tile = false;
741
742 /* Cursors are always linear, and the user can request linear as well.
743 */
744 if (tmpl->bind & (PIPE_BIND_LINEAR | PIPE_BIND_CURSOR))
745 should_tile = false;
746
747 /* 1D and 1D_ARRAY textures are always raster-order. */
748 if (tmpl->target == PIPE_TEXTURE_1D ||
749 tmpl->target == PIPE_TEXTURE_1D_ARRAY)
750 should_tile = false;
751
752 /* Scanout BOs for simulator need to be linear for interaction with
753 * i965.
754 */
755 if (using_v3d_simulator &&
756 tmpl->bind & (PIPE_BIND_SHARED | PIPE_BIND_SCANOUT))
757 should_tile = false;
758
759 /* If using the old-school SCANOUT flag, we don't know what the screen
760 * might support other than linear. Just force linear.
761 */
762 if (tmpl->bind & PIPE_BIND_SCANOUT)
763 should_tile = false;
764
765 /* No user-specified modifier; determine our own. */
766 if (count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID) {
767 linear_ok = true;
768 rsc->tiled = should_tile;
769 } else if (should_tile &&
770 drm_find_modifier(DRM_FORMAT_MOD_BROADCOM_UIF,
771 modifiers, count)) {
772 rsc->tiled = true;
773 } else if (linear_ok) {
774 rsc->tiled = false;
775 } else {
776 fprintf(stderr, "Unsupported modifier requested\n");
777 goto fail;
778 }
779
780 rsc->internal_format = prsc->format;
781
782 v3d_setup_slices(rsc, 0, tmpl->bind & PIPE_BIND_SHARED);
783
784 /* If we're in a renderonly setup, use the other device to perform our
785 * allocation and just import it to v3d. The other device may be
786 * using CMA, and V3D can import from CMA but doesn't do CMA
787 * allocations on its own.
788 *
789 * We always allocate this way for SHARED, because get_handle will
790 * need a resource on the display fd.
791 */
792 if (screen->ro && (tmpl->bind & (PIPE_BIND_SCANOUT |
793 PIPE_BIND_SHARED))) {
794 struct winsys_handle handle;
795 struct pipe_resource scanout_tmpl = {
796 .target = prsc->target,
797 .format = PIPE_FORMAT_RGBA8888_UNORM,
798 .width0 = 1024, /* one page */
799 .height0 = align(rsc->size, 4096) / 4096,
800 .depth0 = 1,
801 .array_size = 1,
802 };
803
804 rsc->scanout =
805 renderonly_scanout_for_resource(&scanout_tmpl,
806 screen->ro,
807 &handle);
808
809 if (!rsc->scanout) {
810 fprintf(stderr, "Failed to create scanout resource\n");
811 return NULL;
812 }
813 assert(handle.type == WINSYS_HANDLE_TYPE_FD);
814 rsc->bo = v3d_bo_open_dmabuf(screen, handle.handle);
815 close(handle.handle);
816
817 if (!rsc->bo)
818 goto fail;
819
820 v3d_debug_resource_layout(rsc, "renderonly");
821
822 return prsc;
823 } else {
824 if (!v3d_resource_bo_alloc(rsc))
825 goto fail;
826 }
827
828 return prsc;
829 fail:
830 v3d_resource_destroy(pscreen, prsc);
831 return NULL;
832 }
833
834 struct pipe_resource *
v3d_resource_create(struct pipe_screen * pscreen,const struct pipe_resource * tmpl)835 v3d_resource_create(struct pipe_screen *pscreen,
836 const struct pipe_resource *tmpl)
837 {
838 const uint64_t mod = DRM_FORMAT_MOD_INVALID;
839 return v3d_resource_create_with_modifiers(pscreen, tmpl, &mod, 1);
840 }
841
842 static struct pipe_resource *
v3d_resource_from_handle(struct pipe_screen * pscreen,const struct pipe_resource * tmpl,struct winsys_handle * whandle,unsigned usage)843 v3d_resource_from_handle(struct pipe_screen *pscreen,
844 const struct pipe_resource *tmpl,
845 struct winsys_handle *whandle,
846 unsigned usage)
847 {
848 struct v3d_screen *screen = v3d_screen(pscreen);
849 struct v3d_resource *rsc = v3d_resource_setup(pscreen, tmpl);
850 struct pipe_resource *prsc = &rsc->base;
851 struct v3d_resource_slice *slice = &rsc->slices[0];
852
853 if (!rsc)
854 return NULL;
855
856 switch (whandle->modifier) {
857 case DRM_FORMAT_MOD_LINEAR:
858 rsc->tiled = false;
859 break;
860 case DRM_FORMAT_MOD_BROADCOM_UIF:
861 rsc->tiled = true;
862 break;
863 case DRM_FORMAT_MOD_INVALID:
864 rsc->tiled = screen->ro == NULL;
865 break;
866 default:
867 fprintf(stderr,
868 "Attempt to import unsupported modifier 0x%llx\n",
869 (long long)whandle->modifier);
870 goto fail;
871 }
872
873 switch (whandle->type) {
874 case WINSYS_HANDLE_TYPE_SHARED:
875 rsc->bo = v3d_bo_open_name(screen, whandle->handle);
876 break;
877 case WINSYS_HANDLE_TYPE_FD:
878 rsc->bo = v3d_bo_open_dmabuf(screen, whandle->handle);
879 break;
880 default:
881 fprintf(stderr,
882 "Attempt to import unsupported handle type %d\n",
883 whandle->type);
884 goto fail;
885 }
886
887 if (!rsc->bo)
888 goto fail;
889
890 rsc->internal_format = prsc->format;
891
892 v3d_setup_slices(rsc, whandle->stride, true);
893 v3d_debug_resource_layout(rsc, "import");
894
895 if (whandle->offset != 0) {
896 if (rsc->tiled) {
897 fprintf(stderr,
898 "Attempt to import unsupported winsys offset %u\n",
899 whandle->offset);
900 goto fail;
901 }
902 rsc->slices[0].offset += whandle->offset;
903
904 if (rsc->slices[0].offset + rsc->slices[0].size >
905 rsc->bo->size) {
906 fprintf(stderr, "Attempt to import "
907 "with overflowing offset (%d + %d > %d)\n",
908 whandle->offset,
909 rsc->slices[0].size,
910 rsc->bo->size);
911 goto fail;
912 }
913 }
914
915 if (screen->ro) {
916 /* Make sure that renderonly has a handle to our buffer in the
917 * display's fd, so that a later renderonly_get_handle()
918 * returns correct handles or GEM names.
919 */
920 rsc->scanout =
921 renderonly_create_gpu_import_for_resource(prsc,
922 screen->ro,
923 NULL);
924 if (!rsc->scanout) {
925 fprintf(stderr, "Failed to create scanout resource.\n");
926 goto fail;
927 }
928 }
929
930 if (rsc->tiled && whandle->stride != slice->stride) {
931 static bool warned = false;
932 if (!warned) {
933 warned = true;
934 fprintf(stderr,
935 "Attempting to import %dx%d %s with "
936 "unsupported stride %d instead of %d\n",
937 prsc->width0, prsc->height0,
938 util_format_short_name(prsc->format),
939 whandle->stride,
940 slice->stride);
941 }
942 goto fail;
943 } else if (!rsc->tiled) {
944 slice->stride = whandle->stride;
945 }
946
947 return prsc;
948
949 fail:
950 v3d_resource_destroy(pscreen, prsc);
951 return NULL;
952 }
953
954 void
v3d_update_shadow_texture(struct pipe_context * pctx,struct pipe_sampler_view * pview)955 v3d_update_shadow_texture(struct pipe_context *pctx,
956 struct pipe_sampler_view *pview)
957 {
958 struct v3d_context *v3d = v3d_context(pctx);
959 struct v3d_sampler_view *view = v3d_sampler_view(pview);
960 struct v3d_resource *shadow = v3d_resource(view->texture);
961 struct v3d_resource *orig = v3d_resource(pview->texture);
962
963 assert(view->texture != pview->texture);
964
965 if (shadow->writes == orig->writes && orig->bo->private)
966 return;
967
968 perf_debug("Updating %dx%d@%d shadow for linear texture\n",
969 orig->base.width0, orig->base.height0,
970 pview->u.tex.first_level);
971
972 for (int i = 0; i <= shadow->base.last_level; i++) {
973 unsigned width = u_minify(shadow->base.width0, i);
974 unsigned height = u_minify(shadow->base.height0, i);
975 struct pipe_blit_info info = {
976 .dst = {
977 .resource = &shadow->base,
978 .level = i,
979 .box = {
980 .x = 0,
981 .y = 0,
982 .z = 0,
983 .width = width,
984 .height = height,
985 .depth = 1,
986 },
987 .format = shadow->base.format,
988 },
989 .src = {
990 .resource = &orig->base,
991 .level = pview->u.tex.first_level + i,
992 .box = {
993 .x = 0,
994 .y = 0,
995 .z = 0,
996 .width = width,
997 .height = height,
998 .depth = 1,
999 },
1000 .format = orig->base.format,
1001 },
1002 .mask = util_format_get_mask(orig->base.format),
1003 };
1004 pctx->blit(pctx, &info);
1005 }
1006
1007 shadow->writes = orig->writes;
1008 }
1009
1010 static struct pipe_surface *
v3d_create_surface(struct pipe_context * pctx,struct pipe_resource * ptex,const struct pipe_surface * surf_tmpl)1011 v3d_create_surface(struct pipe_context *pctx,
1012 struct pipe_resource *ptex,
1013 const struct pipe_surface *surf_tmpl)
1014 {
1015 struct v3d_context *v3d = v3d_context(pctx);
1016 struct v3d_screen *screen = v3d->screen;
1017 struct v3d_surface *surface = CALLOC_STRUCT(v3d_surface);
1018 struct v3d_resource *rsc = v3d_resource(ptex);
1019
1020 if (!surface)
1021 return NULL;
1022
1023 struct pipe_surface *psurf = &surface->base;
1024 unsigned level = surf_tmpl->u.tex.level;
1025 struct v3d_resource_slice *slice = &rsc->slices[level];
1026
1027 pipe_reference_init(&psurf->reference, 1);
1028 pipe_resource_reference(&psurf->texture, ptex);
1029
1030 psurf->context = pctx;
1031 psurf->format = surf_tmpl->format;
1032 psurf->width = u_minify(ptex->width0, level);
1033 psurf->height = u_minify(ptex->height0, level);
1034 psurf->u.tex.level = level;
1035 psurf->u.tex.first_layer = surf_tmpl->u.tex.first_layer;
1036 psurf->u.tex.last_layer = surf_tmpl->u.tex.last_layer;
1037
1038 surface->offset = v3d_layer_offset(ptex, level,
1039 psurf->u.tex.first_layer);
1040 surface->tiling = slice->tiling;
1041
1042 surface->format = v3d_get_rt_format(&screen->devinfo, psurf->format);
1043
1044 const struct util_format_description *desc =
1045 util_format_description(psurf->format);
1046
1047 surface->swap_rb = (desc->swizzle[0] == PIPE_SWIZZLE_Z &&
1048 psurf->format != PIPE_FORMAT_B5G6R5_UNORM);
1049
1050 if (util_format_is_depth_or_stencil(psurf->format)) {
1051 switch (psurf->format) {
1052 case PIPE_FORMAT_Z16_UNORM:
1053 surface->internal_type = V3D_INTERNAL_TYPE_DEPTH_16;
1054 break;
1055 case PIPE_FORMAT_Z32_FLOAT:
1056 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
1057 surface->internal_type = V3D_INTERNAL_TYPE_DEPTH_32F;
1058 break;
1059 default:
1060 surface->internal_type = V3D_INTERNAL_TYPE_DEPTH_24;
1061 }
1062 } else {
1063 uint32_t bpp, type;
1064 v3d_get_internal_type_bpp_for_output_format(&screen->devinfo,
1065 surface->format,
1066 &type, &bpp);
1067 surface->internal_type = type;
1068 surface->internal_bpp = bpp;
1069 }
1070
1071 if (surface->tiling == VC5_TILING_UIF_NO_XOR ||
1072 surface->tiling == VC5_TILING_UIF_XOR) {
1073 surface->padded_height_of_output_image_in_uif_blocks =
1074 (slice->padded_height /
1075 (2 * v3d_utile_height(rsc->cpp)));
1076 }
1077
1078 if (rsc->separate_stencil) {
1079 surface->separate_stencil =
1080 v3d_create_surface(pctx, &rsc->separate_stencil->base,
1081 surf_tmpl);
1082 }
1083
1084 return &surface->base;
1085 }
1086
1087 static void
v3d_surface_destroy(struct pipe_context * pctx,struct pipe_surface * psurf)1088 v3d_surface_destroy(struct pipe_context *pctx, struct pipe_surface *psurf)
1089 {
1090 struct v3d_surface *surf = v3d_surface(psurf);
1091
1092 if (surf->separate_stencil)
1093 pipe_surface_reference(&surf->separate_stencil, NULL);
1094
1095 pipe_resource_reference(&psurf->texture, NULL);
1096 FREE(psurf);
1097 }
1098
1099 static void
v3d_flush_resource(struct pipe_context * pctx,struct pipe_resource * resource)1100 v3d_flush_resource(struct pipe_context *pctx, struct pipe_resource *resource)
1101 {
1102 /* All calls to flush_resource are followed by a flush of the context,
1103 * so there's nothing to do.
1104 */
1105 }
1106
1107 static enum pipe_format
v3d_resource_get_internal_format(struct pipe_resource * prsc)1108 v3d_resource_get_internal_format(struct pipe_resource *prsc)
1109 {
1110 return v3d_resource(prsc)->internal_format;
1111 }
1112
1113 static void
v3d_resource_set_stencil(struct pipe_resource * prsc,struct pipe_resource * stencil)1114 v3d_resource_set_stencil(struct pipe_resource *prsc,
1115 struct pipe_resource *stencil)
1116 {
1117 v3d_resource(prsc)->separate_stencil = v3d_resource(stencil);
1118 }
1119
1120 static struct pipe_resource *
v3d_resource_get_stencil(struct pipe_resource * prsc)1121 v3d_resource_get_stencil(struct pipe_resource *prsc)
1122 {
1123 struct v3d_resource *rsc = v3d_resource(prsc);
1124
1125 return &rsc->separate_stencil->base;
1126 }
1127
1128 static const struct u_transfer_vtbl transfer_vtbl = {
1129 .resource_create = v3d_resource_create,
1130 .resource_destroy = v3d_resource_destroy,
1131 .transfer_map = v3d_resource_transfer_map,
1132 .transfer_unmap = v3d_resource_transfer_unmap,
1133 .transfer_flush_region = u_default_transfer_flush_region,
1134 .get_internal_format = v3d_resource_get_internal_format,
1135 .set_stencil = v3d_resource_set_stencil,
1136 .get_stencil = v3d_resource_get_stencil,
1137 };
1138
1139 void
v3d_resource_screen_init(struct pipe_screen * pscreen)1140 v3d_resource_screen_init(struct pipe_screen *pscreen)
1141 {
1142 pscreen->resource_create_with_modifiers =
1143 v3d_resource_create_with_modifiers;
1144 pscreen->resource_create = u_transfer_helper_resource_create;
1145 pscreen->resource_from_handle = v3d_resource_from_handle;
1146 pscreen->resource_get_handle = v3d_resource_get_handle;
1147 pscreen->resource_destroy = u_transfer_helper_resource_destroy;
1148 pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl,
1149 true, false,
1150 true, true);
1151 }
1152
1153 void
v3d_resource_context_init(struct pipe_context * pctx)1154 v3d_resource_context_init(struct pipe_context *pctx)
1155 {
1156 pctx->transfer_map = u_transfer_helper_transfer_map;
1157 pctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
1158 pctx->transfer_unmap = u_transfer_helper_transfer_unmap;
1159 pctx->buffer_subdata = u_default_buffer_subdata;
1160 pctx->texture_subdata = v3d_texture_subdata;
1161 pctx->create_surface = v3d_create_surface;
1162 pctx->surface_destroy = v3d_surface_destroy;
1163 pctx->resource_copy_region = util_resource_copy_region;
1164 pctx->blit = v3d_blit;
1165 pctx->generate_mipmap = v3d_generate_mipmap;
1166 pctx->flush_resource = v3d_flush_resource;
1167 }
1168