broadcom/vc5: Don't allocate simulator BOs at offset 0.
[mesa.git] / src / gallium / drivers / vc5 / vc5_resource.c
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_blit.h"
27 #include "util/u_memory.h"
28 #include "util/u_format.h"
29 #include "util/u_inlines.h"
30 #include "util/u_surface.h"
31 #include "util/u_transfer_helper.h"
32 #include "util/u_upload_mgr.h"
33 #include "util/u_format_zs.h"
34
35 #include "drm_fourcc.h"
36 #include "vc5_screen.h"
37 #include "vc5_context.h"
38 #include "vc5_resource.h"
39 #include "vc5_tiling.h"
40 #include "broadcom/cle/v3d_packet_v33_pack.h"
41
42 static void
43 vc5_debug_resource_layout(struct vc5_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 vc5_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
97 vc5_resource_bo_alloc(struct vc5_resource *rsc)
98 {
99 struct pipe_resource *prsc = &rsc->base;
100 struct pipe_screen *pscreen = prsc->screen;
101 struct vc5_bo *bo;
102
103 bo = vc5_bo_alloc(vc5_screen(pscreen), rsc->size, "resource");
104 if (bo) {
105 vc5_bo_unreference(&rsc->bo);
106 rsc->bo = bo;
107 vc5_debug_resource_layout(rsc, "alloc");
108 return true;
109 } else {
110 return false;
111 }
112 }
113
114 static void
115 vc5_resource_transfer_unmap(struct pipe_context *pctx,
116 struct pipe_transfer *ptrans)
117 {
118 struct vc5_context *vc5 = vc5_context(pctx);
119 struct vc5_transfer *trans = vc5_transfer(ptrans);
120
121 if (trans->map) {
122 struct vc5_resource *rsc = vc5_resource(ptrans->resource);
123 struct vc5_resource_slice *slice = &rsc->slices[ptrans->level];
124
125 if (ptrans->usage & PIPE_TRANSFER_WRITE) {
126 for (int z = 0; z < ptrans->box.depth; z++) {
127 void *dst = rsc->bo->map +
128 vc5_layer_offset(&rsc->base,
129 ptrans->level,
130 ptrans->box.z + z);
131 vc5_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(&vc5->transfer_pool, ptrans);
147 }
148
149 static void *
150 vc5_resource_transfer_map(struct pipe_context *pctx,
151 struct pipe_resource *prsc,
152 unsigned level, unsigned usage,
153 const struct pipe_box *box,
154 struct pipe_transfer **pptrans)
155 {
156 struct vc5_context *vc5 = vc5_context(pctx);
157 struct vc5_resource *rsc = vc5_resource(prsc);
158 struct vc5_transfer *trans;
159 struct pipe_transfer *ptrans;
160 enum pipe_format format = prsc->format;
161 char *buf;
162
163 /* MSAA maps should have been handled by u_transfer_helper. */
164 assert(prsc->nr_samples <= 1);
165
166 /* Upgrade DISCARD_RANGE to WHOLE_RESOURCE if the whole resource is
167 * being mapped.
168 */
169 if ((usage & PIPE_TRANSFER_DISCARD_RANGE) &&
170 !(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
171 !(prsc->flags & PIPE_RESOURCE_FLAG_MAP_COHERENT) &&
172 prsc->last_level == 0 &&
173 prsc->width0 == box->width &&
174 prsc->height0 == box->height &&
175 prsc->depth0 == box->depth &&
176 prsc->array_size == 1 &&
177 rsc->bo->private) {
178 usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE;
179 }
180
181 if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
182 if (vc5_resource_bo_alloc(rsc)) {
183 /* If it might be bound as one of our vertex buffers
184 * or UBOs, make sure we re-emit vertex buffer state
185 * or uniforms.
186 */
187 if (prsc->bind & PIPE_BIND_VERTEX_BUFFER)
188 vc5->dirty |= VC5_DIRTY_VTXBUF;
189 if (prsc->bind & PIPE_BIND_CONSTANT_BUFFER)
190 vc5->dirty |= VC5_DIRTY_CONSTBUF;
191 } else {
192 /* If we failed to reallocate, flush users so that we
193 * don't violate any syncing requirements.
194 */
195 vc5_flush_jobs_reading_resource(vc5, prsc);
196 }
197 } else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
198 /* If we're writing and the buffer is being used by the CL, we
199 * have to flush the CL first. If we're only reading, we need
200 * to flush if the CL has written our buffer.
201 */
202 if (usage & PIPE_TRANSFER_WRITE)
203 vc5_flush_jobs_reading_resource(vc5, prsc);
204 else
205 vc5_flush_jobs_writing_resource(vc5, prsc);
206 }
207
208 if (usage & PIPE_TRANSFER_WRITE) {
209 rsc->writes++;
210 rsc->initialized_buffers = ~0;
211 }
212
213 trans = slab_alloc(&vc5->transfer_pool);
214 if (!trans)
215 return NULL;
216
217 /* XXX: Handle DONTBLOCK, DISCARD_RANGE, PERSISTENT, COHERENT. */
218
219 /* slab_alloc_st() doesn't zero: */
220 memset(trans, 0, sizeof(*trans));
221 ptrans = &trans->base;
222
223 pipe_resource_reference(&ptrans->resource, prsc);
224 ptrans->level = level;
225 ptrans->usage = usage;
226 ptrans->box = *box;
227
228 /* Note that the current kernel implementation is synchronous, so no
229 * need to do syncing stuff here yet.
230 */
231
232 if (usage & PIPE_TRANSFER_UNSYNCHRONIZED)
233 buf = vc5_bo_map_unsynchronized(rsc->bo);
234 else
235 buf = vc5_bo_map(rsc->bo);
236 if (!buf) {
237 fprintf(stderr, "Failed to map bo\n");
238 goto fail;
239 }
240
241 *pptrans = ptrans;
242
243 /* Our load/store routines work on entire compressed blocks. */
244 ptrans->box.x /= util_format_get_blockwidth(format);
245 ptrans->box.y /= util_format_get_blockheight(format);
246 ptrans->box.width = DIV_ROUND_UP(ptrans->box.width,
247 util_format_get_blockwidth(format));
248 ptrans->box.height = DIV_ROUND_UP(ptrans->box.height,
249 util_format_get_blockheight(format));
250
251 struct vc5_resource_slice *slice = &rsc->slices[level];
252 if (rsc->tiled) {
253 /* No direct mappings of tiled, since we need to manually
254 * tile/untile.
255 */
256 if (usage & PIPE_TRANSFER_MAP_DIRECTLY)
257 return NULL;
258
259 ptrans->stride = ptrans->box.width * rsc->cpp;
260 ptrans->layer_stride = ptrans->stride * ptrans->box.height;
261
262 trans->map = malloc(ptrans->layer_stride * ptrans->box.depth);
263
264 if (usage & PIPE_TRANSFER_READ) {
265 for (int z = 0; z < ptrans->box.depth; z++) {
266 void *src = rsc->bo->map +
267 vc5_layer_offset(&rsc->base,
268 ptrans->level,
269 ptrans->box.z + z);
270 vc5_load_tiled_image((trans->map +
271 ptrans->stride *
272 ptrans->box.height * z),
273 ptrans->stride,
274 src,
275 slice->stride,
276 slice->tiling, rsc->cpp,
277 slice->padded_height,
278 &ptrans->box);
279 }
280 }
281 return trans->map;
282 } else {
283 ptrans->stride = slice->stride;
284 ptrans->layer_stride = ptrans->stride;
285
286 return buf + slice->offset +
287 ptrans->box.y * ptrans->stride +
288 ptrans->box.x * rsc->cpp +
289 ptrans->box.z * rsc->cube_map_stride;
290 }
291
292
293 fail:
294 vc5_resource_transfer_unmap(pctx, ptrans);
295 return NULL;
296 }
297
298 static void
299 vc5_resource_destroy(struct pipe_screen *pscreen,
300 struct pipe_resource *prsc)
301 {
302 struct vc5_resource *rsc = vc5_resource(prsc);
303
304 vc5_bo_unreference(&rsc->bo);
305 free(rsc);
306 }
307
308 static boolean
309 vc5_resource_get_handle(struct pipe_screen *pscreen,
310 struct pipe_context *pctx,
311 struct pipe_resource *prsc,
312 struct winsys_handle *whandle,
313 unsigned usage)
314 {
315 struct vc5_resource *rsc = vc5_resource(prsc);
316 struct vc5_bo *bo = rsc->bo;
317
318 whandle->stride = rsc->slices[0].stride;
319
320 /* If we're passing some reference to our BO out to some other part of
321 * the system, then we can't do any optimizations about only us being
322 * the ones seeing it (like BO caching).
323 */
324 bo->private = false;
325
326 switch (whandle->type) {
327 case DRM_API_HANDLE_TYPE_SHARED:
328 return vc5_bo_flink(bo, &whandle->handle);
329 case DRM_API_HANDLE_TYPE_KMS:
330 whandle->handle = bo->handle;
331 return TRUE;
332 case DRM_API_HANDLE_TYPE_FD:
333 whandle->handle = vc5_bo_get_dmabuf(bo);
334 return whandle->handle != -1;
335 }
336
337 return FALSE;
338 }
339
340 #define PAGE_UB_ROWS (VC5_UIFCFG_PAGE_SIZE / VC5_UIFBLOCK_ROW_SIZE)
341 #define PAGE_UB_ROWS_TIMES_1_5 ((PAGE_UB_ROWS * 3) >> 1)
342 #define PAGE_CACHE_UB_ROWS (VC5_PAGE_CACHE_SIZE / VC5_UIFBLOCK_ROW_SIZE)
343 #define PAGE_CACHE_MINUS_1_5_UB_ROWS (PAGE_CACHE_UB_ROWS - PAGE_UB_ROWS_TIMES_1_5)
344
345 /**
346 * Computes the HW's UIFblock padding for a given height/cpp.
347 *
348 * The goal of the padding is to keep pages of the same color (bank number) at
349 * least half a page away from each other vertically when crossing between
350 * between columns of UIF blocks.
351 */
352 static uint32_t
353 vc5_get_ub_pad(struct vc5_resource *rsc, uint32_t height)
354 {
355 uint32_t utile_h = vc5_utile_height(rsc->cpp);
356 uint32_t uif_block_h = utile_h * 2;
357 uint32_t height_ub = height / uif_block_h;
358
359 uint32_t height_offset_in_pc = height_ub % PAGE_CACHE_UB_ROWS;
360
361 /* For the perfectly-aligned-for-UIF-XOR case, don't add any pad. */
362 if (height_offset_in_pc == 0)
363 return 0;
364
365 /* Try padding up to where we're offset by at least half a page. */
366 if (height_offset_in_pc < PAGE_UB_ROWS_TIMES_1_5) {
367 /* If we fit entirely in the page cache, don't pad. */
368 if (height_ub < PAGE_CACHE_UB_ROWS)
369 return 0;
370 else
371 return PAGE_UB_ROWS_TIMES_1_5 - height_offset_in_pc;
372 }
373
374 /* If we're close to being aligned to page cache size, then round up
375 * and rely on XOR.
376 */
377 if (height_offset_in_pc > PAGE_CACHE_MINUS_1_5_UB_ROWS)
378 return PAGE_CACHE_UB_ROWS - height_offset_in_pc;
379
380 /* Otherwise, we're far enough away (top and bottom) to not need any
381 * padding.
382 */
383 return 0;
384 }
385
386 static void
387 vc5_setup_slices(struct vc5_resource *rsc)
388 {
389 struct pipe_resource *prsc = &rsc->base;
390 uint32_t width = prsc->width0;
391 uint32_t height = prsc->height0;
392 uint32_t depth = prsc->depth0;
393 /* Note that power-of-two padding is based on level 1. These are not
394 * equivalent to just util_next_power_of_two(dimension), because at a
395 * level 0 dimension of 9, the level 1 power-of-two padded value is 4,
396 * not 8.
397 */
398 uint32_t pot_width = 2 * util_next_power_of_two(u_minify(width, 1));
399 uint32_t pot_height = 2 * util_next_power_of_two(u_minify(height, 1));
400 uint32_t pot_depth = 2 * util_next_power_of_two(u_minify(depth, 1));
401 uint32_t offset = 0;
402 uint32_t utile_w = vc5_utile_width(rsc->cpp);
403 uint32_t utile_h = vc5_utile_height(rsc->cpp);
404 uint32_t uif_block_w = utile_w * 2;
405 uint32_t uif_block_h = utile_h * 2;
406 uint32_t block_width = util_format_get_blockwidth(prsc->format);
407 uint32_t block_height = util_format_get_blockheight(prsc->format);
408 bool msaa = prsc->nr_samples > 1;
409 /* MSAA textures/renderbuffers are always laid out as single-level
410 * UIF.
411 */
412 bool uif_top = msaa;
413
414 for (int i = prsc->last_level; i >= 0; i--) {
415 struct vc5_resource_slice *slice = &rsc->slices[i];
416
417 uint32_t level_width, level_height, level_depth;
418 if (i < 2) {
419 level_width = u_minify(width, i);
420 level_height = u_minify(height, i);
421 } else {
422 level_width = u_minify(pot_width, i);
423 level_height = u_minify(pot_height, i);
424 }
425 if (i < 1)
426 level_depth = u_minify(depth, i);
427 else
428 level_depth = u_minify(pot_depth, i);
429
430 if (msaa) {
431 level_width *= 2;
432 level_height *= 2;
433 }
434
435 level_width = DIV_ROUND_UP(level_width, block_width);
436 level_height = DIV_ROUND_UP(level_height, block_height);
437
438 if (!rsc->tiled) {
439 slice->tiling = VC5_TILING_RASTER;
440 if (prsc->target == PIPE_TEXTURE_1D)
441 level_width = align(level_width, 64 / rsc->cpp);
442 } else {
443 if ((i != 0 || !uif_top) &&
444 (level_width <= utile_w ||
445 level_height <= utile_h)) {
446 slice->tiling = VC5_TILING_LINEARTILE;
447 level_width = align(level_width, utile_w);
448 level_height = align(level_height, utile_h);
449 } else if ((i != 0 || !uif_top) &&
450 level_width <= uif_block_w) {
451 slice->tiling = VC5_TILING_UBLINEAR_1_COLUMN;
452 level_width = align(level_width, uif_block_w);
453 level_height = align(level_height, uif_block_h);
454 } else if ((i != 0 || !uif_top) &&
455 level_width <= 2 * uif_block_w) {
456 slice->tiling = VC5_TILING_UBLINEAR_2_COLUMN;
457 level_width = align(level_width, 2 * uif_block_w);
458 level_height = align(level_height, uif_block_h);
459 } else {
460 /* We align the width to a 4-block column of
461 * UIF blocks, but we only align height to UIF
462 * blocks.
463 */
464 level_width = align(level_width,
465 4 * uif_block_w);
466 level_height = align(level_height,
467 uif_block_h);
468
469 slice->ub_pad = vc5_get_ub_pad(rsc,
470 level_height);
471 level_height += slice->ub_pad * uif_block_h;
472
473 /* If the padding set us to to be aligned to
474 * the page cache size, then the HW will use
475 * the XOR bit on odd columns to get us
476 * perfectly misaligned
477 */
478 if ((level_height / uif_block_h) %
479 (VC5_PAGE_CACHE_SIZE /
480 VC5_UIFBLOCK_ROW_SIZE) == 0) {
481 slice->tiling = VC5_TILING_UIF_XOR;
482 } else {
483 slice->tiling = VC5_TILING_UIF_NO_XOR;
484 }
485 }
486 }
487
488 slice->offset = offset;
489 slice->stride = level_width * rsc->cpp;
490 slice->padded_height = level_height;
491 slice->size = level_height * slice->stride;
492
493 uint32_t slice_total_size = slice->size * level_depth;
494
495 /* The HW aligns level 1's base to a page if any of level 1 or
496 * below could be UIF XOR. The lower levels then inherit the
497 * alignment for as long as necesary, thanks to being power of
498 * two aligned.
499 */
500 if (i == 1 &&
501 level_width > 4 * uif_block_w &&
502 level_height > PAGE_CACHE_MINUS_1_5_UB_ROWS * uif_block_h) {
503 slice_total_size = align(slice_total_size,
504 VC5_UIFCFG_PAGE_SIZE);
505 }
506
507 offset += slice_total_size;
508
509 }
510 rsc->size = offset;
511
512 /* UIF/UBLINEAR levels need to be aligned to UIF-blocks, and LT only
513 * needs to be aligned to utile boundaries. Since tiles are laid out
514 * from small to big in memory, we need to align the later UIF slices
515 * to UIF blocks, if they were preceded by non-UIF-block-aligned LT
516 * slices.
517 *
518 * We additionally align to 4k, which improves UIF XOR performance.
519 */
520 uint32_t page_align_offset = (align(rsc->slices[0].offset, 4096) -
521 rsc->slices[0].offset);
522 if (page_align_offset) {
523 rsc->size += page_align_offset;
524 for (int i = 0; i <= prsc->last_level; i++)
525 rsc->slices[i].offset += page_align_offset;
526 }
527
528 /* Arrays and cube textures have a stride which is the distance from
529 * one full mipmap tree to the next (64b aligned). For 3D textures,
530 * we need to program the stride between slices of miplevel 0.
531 */
532 if (prsc->target != PIPE_TEXTURE_3D) {
533 rsc->cube_map_stride = align(rsc->slices[0].offset +
534 rsc->slices[0].size, 64);
535 rsc->size += rsc->cube_map_stride * (prsc->array_size - 1);
536 } else {
537 rsc->cube_map_stride = rsc->slices[0].size;
538 }
539 }
540
541 uint32_t
542 vc5_layer_offset(struct pipe_resource *prsc, uint32_t level, uint32_t layer)
543 {
544 struct vc5_resource *rsc = vc5_resource(prsc);
545 struct vc5_resource_slice *slice = &rsc->slices[level];
546
547 if (prsc->target == PIPE_TEXTURE_3D)
548 return slice->offset + layer * slice->size;
549 else
550 return slice->offset + layer * rsc->cube_map_stride;
551 }
552
553 static struct vc5_resource *
554 vc5_resource_setup(struct pipe_screen *pscreen,
555 const struct pipe_resource *tmpl)
556 {
557 struct vc5_screen *screen = vc5_screen(pscreen);
558 struct vc5_resource *rsc = CALLOC_STRUCT(vc5_resource);
559 if (!rsc)
560 return NULL;
561 struct pipe_resource *prsc = &rsc->base;
562
563 *prsc = *tmpl;
564
565 pipe_reference_init(&prsc->reference, 1);
566 prsc->screen = pscreen;
567
568 if (prsc->nr_samples <= 1 ||
569 util_format_is_depth_or_stencil(prsc->format)) {
570 rsc->cpp = util_format_get_blocksize(prsc->format) *
571 MAX2(prsc->nr_samples, 1);
572 } else {
573 assert(vc5_rt_format_supported(&screen->devinfo, prsc->format));
574 uint32_t output_image_format =
575 vc5_get_rt_format(&screen->devinfo, prsc->format);
576 uint32_t internal_type;
577 uint32_t internal_bpp;
578 vc5_get_internal_type_bpp_for_output_format(&screen->devinfo,
579 output_image_format,
580 &internal_type,
581 &internal_bpp);
582 switch (internal_bpp) {
583 case V3D_INTERNAL_BPP_32:
584 rsc->cpp = 4;
585 break;
586 case V3D_INTERNAL_BPP_64:
587 rsc->cpp = 8;
588 break;
589 case V3D_INTERNAL_BPP_128:
590 rsc->cpp = 16;
591 break;
592 }
593 }
594
595 assert(rsc->cpp);
596
597 return rsc;
598 }
599
600 static bool
601 find_modifier(uint64_t needle, const uint64_t *haystack, int count)
602 {
603 int i;
604
605 for (i = 0; i < count; i++) {
606 if (haystack[i] == needle)
607 return true;
608 }
609
610 return false;
611 }
612
613 static struct pipe_resource *
614 vc5_resource_create_with_modifiers(struct pipe_screen *pscreen,
615 const struct pipe_resource *tmpl,
616 const uint64_t *modifiers,
617 int count)
618 {
619 bool linear_ok = find_modifier(DRM_FORMAT_MOD_LINEAR, modifiers, count);
620 struct vc5_resource *rsc = vc5_resource_setup(pscreen, tmpl);
621 struct pipe_resource *prsc = &rsc->base;
622 /* Use a tiled layout if we can, for better 3D performance. */
623 bool should_tile = true;
624
625 /* VBOs/PBOs are untiled (and 1 height). */
626 if (tmpl->target == PIPE_BUFFER)
627 should_tile = false;
628
629 /* Cursors are always linear, and the user can request linear as well.
630 */
631 if (tmpl->bind & (PIPE_BIND_LINEAR | PIPE_BIND_CURSOR))
632 should_tile = false;
633
634 /* 1D and 1D_ARRAY textures are always raster-order. */
635 if (tmpl->target == PIPE_TEXTURE_1D ||
636 tmpl->target == PIPE_TEXTURE_1D_ARRAY)
637 should_tile = false;
638
639 /* Scanout BOs for simulator need to be linear for interaction with
640 * i965.
641 */
642 if (using_vc5_simulator &&
643 tmpl->bind & (PIPE_BIND_SHARED | PIPE_BIND_SCANOUT))
644 should_tile = false;
645
646 /* No user-specified modifier; determine our own. */
647 if (count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID) {
648 linear_ok = true;
649 rsc->tiled = should_tile;
650 } else if (should_tile &&
651 find_modifier(DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED,
652 modifiers, count)) {
653 rsc->tiled = true;
654 } else if (linear_ok) {
655 rsc->tiled = false;
656 } else {
657 fprintf(stderr, "Unsupported modifier requested\n");
658 return NULL;
659 }
660
661 rsc->internal_format = prsc->format;
662
663 vc5_setup_slices(rsc);
664 if (!vc5_resource_bo_alloc(rsc))
665 goto fail;
666
667 return prsc;
668 fail:
669 vc5_resource_destroy(pscreen, prsc);
670 return NULL;
671 }
672
673 struct pipe_resource *
674 vc5_resource_create(struct pipe_screen *pscreen,
675 const struct pipe_resource *tmpl)
676 {
677 const uint64_t mod = DRM_FORMAT_MOD_INVALID;
678 return vc5_resource_create_with_modifiers(pscreen, tmpl, &mod, 1);
679 }
680
681 static struct pipe_resource *
682 vc5_resource_from_handle(struct pipe_screen *pscreen,
683 const struct pipe_resource *tmpl,
684 struct winsys_handle *whandle,
685 unsigned usage)
686 {
687 struct vc5_screen *screen = vc5_screen(pscreen);
688 struct vc5_resource *rsc = vc5_resource_setup(pscreen, tmpl);
689 struct pipe_resource *prsc = &rsc->base;
690 struct vc5_resource_slice *slice = &rsc->slices[0];
691
692 if (!rsc)
693 return NULL;
694
695 switch (whandle->modifier) {
696 case DRM_FORMAT_MOD_LINEAR:
697 case DRM_FORMAT_MOD_INVALID:
698 rsc->tiled = false;
699 break;
700 /* XXX: UIF */
701 default:
702 fprintf(stderr,
703 "Attempt to import unsupported modifier 0x%llx\n",
704 (long long)whandle->modifier);
705 goto fail;
706 }
707
708 if (whandle->offset != 0) {
709 fprintf(stderr,
710 "Attempt to import unsupported winsys offset %u\n",
711 whandle->offset);
712 goto fail;
713 }
714
715 switch (whandle->type) {
716 case DRM_API_HANDLE_TYPE_SHARED:
717 rsc->bo = vc5_bo_open_name(screen,
718 whandle->handle, whandle->stride);
719 break;
720 case DRM_API_HANDLE_TYPE_FD:
721 rsc->bo = vc5_bo_open_dmabuf(screen,
722 whandle->handle, whandle->stride);
723 break;
724 default:
725 fprintf(stderr,
726 "Attempt to import unsupported handle type %d\n",
727 whandle->type);
728 goto fail;
729 }
730
731 if (!rsc->bo)
732 goto fail;
733
734 vc5_setup_slices(rsc);
735 vc5_debug_resource_layout(rsc, "import");
736
737 if (whandle->stride != slice->stride) {
738 static bool warned = false;
739 if (!warned) {
740 warned = true;
741 fprintf(stderr,
742 "Attempting to import %dx%d %s with "
743 "unsupported stride %d instead of %d\n",
744 prsc->width0, prsc->height0,
745 util_format_short_name(prsc->format),
746 whandle->stride,
747 slice->stride);
748 }
749 goto fail;
750 }
751
752 return prsc;
753
754 fail:
755 vc5_resource_destroy(pscreen, prsc);
756 return NULL;
757 }
758
759 static struct pipe_surface *
760 vc5_create_surface(struct pipe_context *pctx,
761 struct pipe_resource *ptex,
762 const struct pipe_surface *surf_tmpl)
763 {
764 struct vc5_context *vc5 = vc5_context(pctx);
765 struct vc5_screen *screen = vc5->screen;
766 struct vc5_surface *surface = CALLOC_STRUCT(vc5_surface);
767 struct vc5_resource *rsc = vc5_resource(ptex);
768
769 if (!surface)
770 return NULL;
771
772 assert(surf_tmpl->u.tex.first_layer == surf_tmpl->u.tex.last_layer);
773
774 struct pipe_surface *psurf = &surface->base;
775 unsigned level = surf_tmpl->u.tex.level;
776 struct vc5_resource_slice *slice = &rsc->slices[level];
777
778 pipe_reference_init(&psurf->reference, 1);
779 pipe_resource_reference(&psurf->texture, ptex);
780
781 psurf->context = pctx;
782 psurf->format = surf_tmpl->format;
783 psurf->width = u_minify(ptex->width0, level);
784 psurf->height = u_minify(ptex->height0, level);
785 psurf->u.tex.level = level;
786 psurf->u.tex.first_layer = surf_tmpl->u.tex.first_layer;
787 psurf->u.tex.last_layer = surf_tmpl->u.tex.last_layer;
788
789 surface->offset = vc5_layer_offset(ptex, level,
790 psurf->u.tex.first_layer);
791 surface->tiling = slice->tiling;
792
793 surface->format = vc5_get_rt_format(&screen->devinfo, psurf->format);
794
795 if (util_format_is_depth_or_stencil(psurf->format)) {
796 switch (psurf->format) {
797 case PIPE_FORMAT_Z16_UNORM:
798 surface->internal_type = V3D_INTERNAL_TYPE_DEPTH_16;
799 break;
800 case PIPE_FORMAT_Z32_FLOAT:
801 case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
802 surface->internal_type = V3D_INTERNAL_TYPE_DEPTH_32F;
803 break;
804 default:
805 surface->internal_type = V3D_INTERNAL_TYPE_DEPTH_24;
806 }
807 } else {
808 uint32_t bpp, type;
809 vc5_get_internal_type_bpp_for_output_format(&screen->devinfo,
810 surface->format,
811 &type, &bpp);
812 surface->internal_type = type;
813 surface->internal_bpp = bpp;
814 }
815
816 if (surface->tiling == VC5_TILING_UIF_NO_XOR ||
817 surface->tiling == VC5_TILING_UIF_XOR) {
818 surface->padded_height_of_output_image_in_uif_blocks =
819 (slice->padded_height /
820 (2 * vc5_utile_height(rsc->cpp)));
821 }
822
823 if (rsc->separate_stencil) {
824 surface->separate_stencil =
825 vc5_create_surface(pctx, &rsc->separate_stencil->base,
826 surf_tmpl);
827 }
828
829 return &surface->base;
830 }
831
832 static void
833 vc5_surface_destroy(struct pipe_context *pctx, struct pipe_surface *psurf)
834 {
835 struct vc5_surface *surf = vc5_surface(psurf);
836
837 if (surf->separate_stencil)
838 pipe_surface_reference(&surf->separate_stencil, NULL);
839
840 pipe_resource_reference(&psurf->texture, NULL);
841 FREE(psurf);
842 }
843
844 static void
845 vc5_flush_resource(struct pipe_context *pctx, struct pipe_resource *resource)
846 {
847 /* All calls to flush_resource are followed by a flush of the context,
848 * so there's nothing to do.
849 */
850 }
851
852 static enum pipe_format
853 vc5_resource_get_internal_format(struct pipe_resource *prsc)
854 {
855 return vc5_resource(prsc)->internal_format;
856 }
857
858 static void
859 vc5_resource_set_stencil(struct pipe_resource *prsc,
860 struct pipe_resource *stencil)
861 {
862 vc5_resource(prsc)->separate_stencil = vc5_resource(stencil);
863 }
864
865 static struct pipe_resource *
866 vc5_resource_get_stencil(struct pipe_resource *prsc)
867 {
868 struct vc5_resource *rsc = vc5_resource(prsc);
869
870 return &rsc->separate_stencil->base;
871 }
872
873 static const struct u_transfer_vtbl transfer_vtbl = {
874 .resource_create = vc5_resource_create,
875 .resource_destroy = vc5_resource_destroy,
876 .transfer_map = vc5_resource_transfer_map,
877 .transfer_unmap = vc5_resource_transfer_unmap,
878 .transfer_flush_region = u_default_transfer_flush_region,
879 .get_internal_format = vc5_resource_get_internal_format,
880 .set_stencil = vc5_resource_set_stencil,
881 .get_stencil = vc5_resource_get_stencil,
882 };
883
884 void
885 vc5_resource_screen_init(struct pipe_screen *pscreen)
886 {
887 pscreen->resource_create_with_modifiers =
888 vc5_resource_create_with_modifiers;
889 pscreen->resource_create = u_transfer_helper_resource_create;
890 pscreen->resource_from_handle = vc5_resource_from_handle;
891 pscreen->resource_get_handle = vc5_resource_get_handle;
892 pscreen->resource_destroy = u_transfer_helper_resource_destroy;
893 pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl,
894 true, true, true);
895 }
896
897 void
898 vc5_resource_context_init(struct pipe_context *pctx)
899 {
900 pctx->transfer_map = u_transfer_helper_transfer_map;
901 pctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
902 pctx->transfer_unmap = u_transfer_helper_transfer_unmap;
903 pctx->buffer_subdata = u_default_buffer_subdata;
904 pctx->texture_subdata = u_default_texture_subdata;
905 pctx->create_surface = vc5_create_surface;
906 pctx->surface_destroy = vc5_surface_destroy;
907 pctx->resource_copy_region = util_resource_copy_region;
908 pctx->blit = vc5_blit;
909 pctx->flush_resource = vc5_flush_resource;
910 }