projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
vc4: Drop the winsys_stride relayout in the simluator
[mesa.git] / src / gallium / drivers / vc4 / vc4_resource.c
1 /*
2 * Copyright © 2014 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
34 #include "drm_fourcc.h"
35 #include "vc4_drm.h"
36 #include "vc4_screen.h"
37 #include "vc4_context.h"
38 #include "vc4_resource.h"
39 #include "vc4_tiling.h"
40
41 static bool
42 vc4_resource_bo_alloc(struct vc4_resource *rsc)
43 {
44 struct pipe_resource *prsc = &rsc->base;
45 struct pipe_screen *pscreen = prsc->screen;
46 struct vc4_bo *bo;
47
48 if (vc4_debug & VC4_DEBUG_SURFACE) {
49 fprintf(stderr, "alloc %p: size %d + offset %d -> %d\n",
50 rsc,
51 rsc->slices[0].size,
52 rsc->slices[0].offset,
53 rsc->slices[0].offset +
54 rsc->slices[0].size +
55 rsc->cube_map_stride * (prsc->array_size - 1));
56 }
57
58 bo = vc4_bo_alloc(vc4_screen(pscreen),
59 rsc->slices[0].offset +
60 rsc->slices[0].size +
61 rsc->cube_map_stride * (prsc->array_size - 1),
62 "resource");
63 if (bo) {
64 vc4_bo_unreference(&rsc->bo);
65 rsc->bo = bo;
66 return true;
67 } else {
68 return false;
69 }
70 }
71
72 static void
73 vc4_resource_transfer_unmap(struct pipe_context *pctx,
74 struct pipe_transfer *ptrans)
75 {
76 struct vc4_context *vc4 = vc4_context(pctx);
77 struct vc4_transfer *trans = vc4_transfer(ptrans);
78
79 if (trans->map) {
80 struct vc4_resource *rsc = vc4_resource(ptrans->resource);
81 struct vc4_resource_slice *slice = &rsc->slices[ptrans->level];
82
83 if (ptrans->usage & PIPE_TRANSFER_WRITE) {
84 vc4_store_tiled_image(rsc->bo->map + slice->offset +
85 ptrans->box.z * rsc->cube_map_stride,
86 slice->stride,
87 trans->map, ptrans->stride,
88 slice->tiling, rsc->cpp,
89 &ptrans->box);
90 }
91 free(trans->map);
92 }
93
94 pipe_resource_reference(&ptrans->resource, NULL);
95 slab_free(&vc4->transfer_pool, ptrans);
96 }
97
98 static void *
99 vc4_resource_transfer_map(struct pipe_context *pctx,
100 struct pipe_resource *prsc,
101 unsigned level, unsigned usage,
102 const struct pipe_box *box,
103 struct pipe_transfer **pptrans)
104 {
105 struct vc4_context *vc4 = vc4_context(pctx);
106 struct vc4_resource *rsc = vc4_resource(prsc);
107 struct vc4_transfer *trans;
108 struct pipe_transfer *ptrans;
109 enum pipe_format format = prsc->format;
110 char *buf;
111
112 /* Upgrade DISCARD_RANGE to WHOLE_RESOURCE if the whole resource is
113 * being mapped.
114 */
115 if ((usage & PIPE_TRANSFER_DISCARD_RANGE) &&
116 !(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
117 !(prsc->flags & PIPE_RESOURCE_FLAG_MAP_PERSISTENT) &&
118 prsc->last_level == 0 &&
119 prsc->width0 == box->width &&
120 prsc->height0 == box->height &&
121 prsc->depth0 == box->depth &&
122 prsc->array_size == 1 &&
123 rsc->bo->private) {
124 usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE;
125 }
126
127 if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
128 if (vc4_resource_bo_alloc(rsc)) {
129 /* If it might be bound as one of our vertex buffers,
130 * make sure we re-emit vertex buffer state.
131 */
132 if (prsc->bind & PIPE_BIND_VERTEX_BUFFER)
133 vc4->dirty |= VC4_DIRTY_VTXBUF;
134 } else {
135 /* If we failed to reallocate, flush users so that we
136 * don't violate any syncing requirements.
137 */
138 vc4_flush_jobs_reading_resource(vc4, prsc);
139 }
140 } else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
141 /* If we're writing and the buffer is being used by the CL, we
142 * have to flush the CL first. If we're only reading, we need
143 * to flush if the CL has written our buffer.
144 */
145 if (usage & PIPE_TRANSFER_WRITE)
146 vc4_flush_jobs_reading_resource(vc4, prsc);
147 else
148 vc4_flush_jobs_writing_resource(vc4, prsc);
149 }
150
151 if (usage & PIPE_TRANSFER_WRITE) {
152 rsc->writes++;
153 rsc->initialized_buffers = ~0;
154 }
155
156 trans = slab_alloc(&vc4->transfer_pool);
157 if (!trans)
158 return NULL;
159
160 /* XXX: Handle DONTBLOCK, DISCARD_RANGE, PERSISTENT, COHERENT. */
161
162 /* slab_alloc_st() doesn't zero: */
163 memset(trans, 0, sizeof(*trans));
164 ptrans = &trans->base;
165
166 pipe_resource_reference(&ptrans->resource, prsc);
167 ptrans->level = level;
168 ptrans->usage = usage;
169 ptrans->box = *box;
170
171 if (usage & PIPE_TRANSFER_UNSYNCHRONIZED)
172 buf = vc4_bo_map_unsynchronized(rsc->bo);
173 else
174 buf = vc4_bo_map(rsc->bo);
175 if (!buf) {
176 fprintf(stderr, "Failed to map bo\n");
177 goto fail;
178 }
179
180 *pptrans = ptrans;
181
182 struct vc4_resource_slice *slice = &rsc->slices[level];
183 if (rsc->tiled) {
184 /* No direct mappings of tiled, since we need to manually
185 * tile/untile.
186 */
187 if (usage & PIPE_TRANSFER_MAP_DIRECTLY)
188 return NULL;
189
190 if (format == PIPE_FORMAT_ETC1_RGB8) {
191 /* ETC1 is arranged as 64-bit blocks, where each block
192 * is 4x4 pixels. Texture tiling operates on the
193 * 64-bit block the way it would an uncompressed
194 * pixels.
195 */
196 assert(!(ptrans->box.x & 3));
197 assert(!(ptrans->box.y & 3));
198 ptrans->box.x >>= 2;
199 ptrans->box.y >>= 2;
200 ptrans->box.width = (ptrans->box.width + 3) >> 2;
201 ptrans->box.height = (ptrans->box.height + 3) >> 2;
202 }
203
204 ptrans->stride = ptrans->box.width * rsc->cpp;
205 ptrans->layer_stride = ptrans->stride * ptrans->box.height;
206
207 trans->map = malloc(ptrans->layer_stride * ptrans->box.depth);
208
209 if (usage & PIPE_TRANSFER_READ) {
210 vc4_load_tiled_image(trans->map, ptrans->stride,
211 buf + slice->offset +
212 ptrans->box.z * rsc->cube_map_stride,
213 slice->stride,
214 slice->tiling, rsc->cpp,
215 &ptrans->box);
216 }
217 return trans->map;
218 } else {
219 ptrans->stride = slice->stride;
220 ptrans->layer_stride = ptrans->stride;
221
222 return buf + slice->offset +
223 ptrans->box.y / util_format_get_blockheight(format) * ptrans->stride +
224 ptrans->box.x / util_format_get_blockwidth(format) * rsc->cpp +
225 ptrans->box.z * rsc->cube_map_stride;
226 }
227
228
229 fail:
230 vc4_resource_transfer_unmap(pctx, ptrans);
231 return NULL;
232 }
233
234 static void
235 vc4_texture_subdata(struct pipe_context *pctx,
236 struct pipe_resource *prsc,
237 unsigned level,
238 unsigned usage,
239 const struct pipe_box *box,
240 const void *data,
241 unsigned stride,
242 unsigned layer_stride)
243 {
244 struct vc4_resource *rsc = vc4_resource(prsc);
245 struct vc4_resource_slice *slice = &rsc->slices[level];
246
247 /* For a direct mapping, we can just take the u_transfer path. */
248 if (!rsc->tiled ||
249 box->depth != 1 ||
250 (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE)) {
251 return u_default_texture_subdata(pctx, prsc, level, usage, box,
252 data, stride, layer_stride);
253 }
254
255 /* Otherwise, map and store the texture data directly into the tiled
256 * texture.
257 */
258 void *buf;
259 if (usage & PIPE_TRANSFER_UNSYNCHRONIZED)
260 buf = vc4_bo_map_unsynchronized(rsc->bo);
261 else
262 buf = vc4_bo_map(rsc->bo);
263
264 vc4_store_tiled_image(buf + slice->offset +
265 box->z * rsc->cube_map_stride,
266 slice->stride,
267 (void *)data, stride,
268 slice->tiling, rsc->cpp,
269 box);
270 }
271
272 static void
273 vc4_resource_destroy(struct pipe_screen *pscreen,
274 struct pipe_resource *prsc)
275 {
276 struct vc4_screen *screen = vc4_screen(pscreen);
277 struct vc4_resource *rsc = vc4_resource(prsc);
278 vc4_bo_unreference(&rsc->bo);
279
280 if (rsc->scanout)
281 renderonly_scanout_destroy(rsc->scanout, screen->ro);
282
283 free(rsc);
284 }
285
286 static boolean
287 vc4_resource_get_handle(struct pipe_screen *pscreen,
288 struct pipe_context *pctx,
289 struct pipe_resource *prsc,
290 struct winsys_handle *whandle,
291 unsigned usage)
292 {
293 struct vc4_screen *screen = vc4_screen(pscreen);
294 struct vc4_resource *rsc = vc4_resource(prsc);
295
296 whandle->stride = rsc->slices[0].stride;
297 whandle->offset = 0;
298
299 /* If we're passing some reference to our BO out to some other part of
300 * the system, then we can't do any optimizations about only us being
301 * the ones seeing it (like BO caching or shadow update avoidance).
302 */
303 rsc->bo->private = false;
304
305 if (rsc->tiled)
306 whandle->modifier = DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED;
307 else
308 whandle->modifier = DRM_FORMAT_MOD_LINEAR;
309
310 switch (whandle->type) {
311 case WINSYS_HANDLE_TYPE_SHARED:
312 if (screen->ro) {
313 /* This could probably be supported, assuming that a
314 * control node was used for pl111.
315 */
316 fprintf(stderr, "flink unsupported with pl111\n");
317 return FALSE;
318 }
319
320 return vc4_bo_flink(rsc->bo, &whandle->handle);
321 case WINSYS_HANDLE_TYPE_KMS:
322 if (screen->ro && renderonly_get_handle(rsc->scanout, whandle))
323 return TRUE;
324 whandle->handle = rsc->bo->handle;
325 return TRUE;
326 case WINSYS_HANDLE_TYPE_FD:
327 /* FDs are cross-device, so we can export directly from vc4.
328 */
329 whandle->handle = vc4_bo_get_dmabuf(rsc->bo);
330 return whandle->handle != -1;
331 }
332
333 return FALSE;
334 }
335
336 static void
337 vc4_setup_slices(struct vc4_resource *rsc, const char *caller)
338 {
339 struct pipe_resource *prsc = &rsc->base;
340 uint32_t width = prsc->width0;
341 uint32_t height = prsc->height0;
342 if (prsc->format == PIPE_FORMAT_ETC1_RGB8) {
343 width = (width + 3) >> 2;
344 height = (height + 3) >> 2;
345 }
346
347 uint32_t pot_width = util_next_power_of_two(width);
348 uint32_t pot_height = util_next_power_of_two(height);
349 uint32_t offset = 0;
350 uint32_t utile_w = vc4_utile_width(rsc->cpp);
351 uint32_t utile_h = vc4_utile_height(rsc->cpp);
352
353 for (int i = prsc->last_level; i >= 0; i--) {
354 struct vc4_resource_slice *slice = &rsc->slices[i];
355
356 uint32_t level_width, level_height;
357 if (i == 0) {
358 level_width = width;
359 level_height = height;
360 } else {
361 level_width = u_minify(pot_width, i);
362 level_height = u_minify(pot_height, i);
363 }
364
365 if (!rsc->tiled) {
366 slice->tiling = VC4_TILING_FORMAT_LINEAR;
367 if (prsc->nr_samples > 1) {
368 /* MSAA (4x) surfaces are stored as raw tile buffer contents. */
369 level_width = align(level_width, 32);
370 level_height = align(level_height, 32);
371 } else {
372 level_width = align(level_width, utile_w);
373 }
374 } else {
375 if (vc4_size_is_lt(level_width, level_height,
376 rsc->cpp)) {
377 slice->tiling = VC4_TILING_FORMAT_LT;
378 level_width = align(level_width, utile_w);
379 level_height = align(level_height, utile_h);
380 } else {
381 slice->tiling = VC4_TILING_FORMAT_T;
382 level_width = align(level_width,
383 4 * 2 * utile_w);
384 level_height = align(level_height,
385 4 * 2 * utile_h);
386 }
387 }
388
389 slice->offset = offset;
390 slice->stride = (level_width * rsc->cpp *
391 MAX2(prsc->nr_samples, 1));
392 slice->size = level_height * slice->stride;
393
394 offset += slice->size;
395
396 if (vc4_debug & VC4_DEBUG_SURFACE) {
397 static const char tiling_chars[] = {
398 [VC4_TILING_FORMAT_LINEAR] = 'R',
399 [VC4_TILING_FORMAT_LT] = 'L',
400 [VC4_TILING_FORMAT_T] = 'T'
401 };
402 fprintf(stderr,
403 "rsc %s %p (format %s: vc4 %d), %dx%d: "
404 "level %d (%c) -> %dx%d, stride %d@0x%08x\n",
405 caller, rsc,
406 util_format_short_name(prsc->format),
407 rsc->vc4_format,
408 prsc->width0, prsc->height0,
409 i, tiling_chars[slice->tiling],
410 level_width, level_height,
411 slice->stride, slice->offset);
412 }
413 }
414
415 /* The texture base pointer that has to point to level 0 doesn't have
416 * intra-page bits, so we have to align it, and thus shift up all the
417 * smaller slices.
418 */
419 uint32_t page_align_offset = (align(rsc->slices[0].offset, 4096) -
420 rsc->slices[0].offset);
421 if (page_align_offset) {
422 for (int i = 0; i <= prsc->last_level; i++)
423 rsc->slices[i].offset += page_align_offset;
424 }
425
426 /* Cube map faces appear as whole miptrees at a page-aligned offset
427 * from the first face's miptree.
428 */
429 if (prsc->target == PIPE_TEXTURE_CUBE) {
430 rsc->cube_map_stride = align(rsc->slices[0].offset +
431 rsc->slices[0].size, 4096);
432 }
433 }
434
435 static struct vc4_resource *
436 vc4_resource_setup(struct pipe_screen *pscreen,
437 const struct pipe_resource *tmpl)
438 {
439 struct vc4_resource *rsc = CALLOC_STRUCT(vc4_resource);
440 if (!rsc)
441 return NULL;
442 struct pipe_resource *prsc = &rsc->base;
443
444 *prsc = *tmpl;
445
446 pipe_reference_init(&prsc->reference, 1);
447 prsc->screen = pscreen;
448
449 if (prsc->nr_samples <= 1)
450 rsc->cpp = util_format_get_blocksize(tmpl->format);
451 else
452 rsc->cpp = sizeof(uint32_t);
453
454 assert(rsc->cpp);
455
456 return rsc;
457 }
458
459 static enum vc4_texture_data_type
460 get_resource_texture_format(struct pipe_resource *prsc)
461 {
462 struct vc4_resource *rsc = vc4_resource(prsc);
463 uint8_t format = vc4_get_tex_format(prsc->format);
464
465 if (!rsc->tiled) {
466 if (prsc->nr_samples > 1) {
467 return ~0;
468 } else {
469 if (format == VC4_TEXTURE_TYPE_RGBA8888)
470 return VC4_TEXTURE_TYPE_RGBA32R;
471 else
472 return ~0;
473 }
474 }
475
476 return format;
477 }
478
479 static bool
480 find_modifier(uint64_t needle, const uint64_t *haystack, int count)
481 {
482 int i;
483
484 for (i = 0; i < count; i++) {
485 if (haystack[i] == needle)
486 return true;
487 }
488
489 return false;
490 }
491
492 static struct pipe_resource *
493 vc4_resource_create_with_modifiers(struct pipe_screen *pscreen,
494 const struct pipe_resource *tmpl,
495 const uint64_t *modifiers,
496 int count)
497 {
498 struct vc4_screen *screen = vc4_screen(pscreen);
499 struct vc4_resource *rsc = vc4_resource_setup(pscreen, tmpl);
500 struct pipe_resource *prsc = &rsc->base;
501 bool linear_ok = find_modifier(DRM_FORMAT_MOD_LINEAR, modifiers, count);
502 /* Use a tiled layout if we can, for better 3D performance. */
503 bool should_tile = true;
504
505 /* VBOs/PBOs are untiled (and 1 height). */
506 if (tmpl->target == PIPE_BUFFER)
507 should_tile = false;
508
509 /* MSAA buffers are linear. */
510 if (tmpl->nr_samples > 1)
511 should_tile = false;
512
513 /* No tiling when we're sharing with another device (pl111). */
514 if (screen->ro && (tmpl->bind & PIPE_BIND_SCANOUT))
515 should_tile = false;
516
517 /* Cursors are always linear, and the user can request linear as well.
518 */
519 if (tmpl->bind & (PIPE_BIND_LINEAR | PIPE_BIND_CURSOR))
520 should_tile = false;
521
522 /* No shared objects with LT format -- the kernel only has T-format
523 * metadata. LT objects are small enough it's not worth the trouble to
524 * give them metadata to tile.
525 */
526 if ((tmpl->bind & (PIPE_BIND_SHARED | PIPE_BIND_SCANOUT)) &&
527 vc4_size_is_lt(prsc->width0, prsc->height0, rsc->cpp))
528 should_tile = false;
529
530 /* If we're sharing or scanning out, we need the ioctl present to
531 * inform the kernel or the other side.
532 */
533 if ((tmpl->bind & (PIPE_BIND_SHARED |
534 PIPE_BIND_SCANOUT)) && !screen->has_tiling_ioctl)
535 should_tile = false;
536
537 /* No user-specified modifier; determine our own. */
538 if (count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID) {
539 linear_ok = true;
540 rsc->tiled = should_tile;
541 } else if (should_tile &&
542 find_modifier(DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED,
543 modifiers, count)) {
544 rsc->tiled = true;
545 } else if (linear_ok) {
546 rsc->tiled = false;
547 } else {
548 fprintf(stderr, "Unsupported modifier requested\n");
549 return NULL;
550 }
551
552 if (tmpl->target != PIPE_BUFFER)
553 rsc->vc4_format = get_resource_texture_format(prsc);
554
555 vc4_setup_slices(rsc, "create");
556 if (!vc4_resource_bo_alloc(rsc))
557 goto fail;
558
559 if (screen->has_tiling_ioctl) {
560 uint64_t modifier;
561 if (rsc->tiled)
562 modifier = DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED;
563 else
564 modifier = DRM_FORMAT_MOD_LINEAR;
565 struct drm_vc4_set_tiling set_tiling = {
566 .handle = rsc->bo->handle,
567 .modifier = modifier,
568 };
569 int ret = vc4_ioctl(screen->fd, DRM_IOCTL_VC4_SET_TILING,
570 &set_tiling);
571 if (ret != 0)
572 goto fail;
573 }
574
575 if (screen->ro && tmpl->bind & PIPE_BIND_SCANOUT) {
576 rsc->scanout =
577 renderonly_scanout_for_resource(prsc, screen->ro, NULL);
578 if (!rsc->scanout)
579 goto fail;
580 }
581
582 vc4_bo_label(screen, rsc->bo, "%sresource %dx%d@%d/%d",
583 (tmpl->bind & PIPE_BIND_SCANOUT) ? "scanout " : "",
584 tmpl->width0, tmpl->height0,
585 rsc->cpp * 8, prsc->last_level);
586
587 return prsc;
588 fail:
589 vc4_resource_destroy(pscreen, prsc);
590 return NULL;
591 }
592
593 struct pipe_resource *
594 vc4_resource_create(struct pipe_screen *pscreen,
595 const struct pipe_resource *tmpl)
596 {
597 const uint64_t mod = DRM_FORMAT_MOD_INVALID;
598 return vc4_resource_create_with_modifiers(pscreen, tmpl, &mod, 1);
599 }
600
601 static struct pipe_resource *
602 vc4_resource_from_handle(struct pipe_screen *pscreen,
603 const struct pipe_resource *tmpl,
604 struct winsys_handle *whandle,
605 unsigned usage)
606 {
607 struct vc4_screen *screen = vc4_screen(pscreen);
608 struct vc4_resource *rsc = vc4_resource_setup(pscreen, tmpl);
609 struct pipe_resource *prsc = &rsc->base;
610 struct vc4_resource_slice *slice = &rsc->slices[0];
611
612 if (!rsc)
613 return NULL;
614
615 switch (whandle->type) {
616 case WINSYS_HANDLE_TYPE_SHARED:
617 rsc->bo = vc4_bo_open_name(screen, whandle->handle);
618 break;
619 case WINSYS_HANDLE_TYPE_FD:
620 rsc->bo = vc4_bo_open_dmabuf(screen, whandle->handle);
621 break;
622 default:
623 fprintf(stderr,
624 "Attempt to import unsupported handle type %d\n",
625 whandle->type);
626 }
627
628 if (!rsc->bo)
629 goto fail;
630
631 struct drm_vc4_get_tiling get_tiling = {
632 .handle = rsc->bo->handle,
633 };
634 int ret = vc4_ioctl(screen->fd, DRM_IOCTL_VC4_GET_TILING, &get_tiling);
635
636 if (ret != 0) {
637 whandle->modifier = DRM_FORMAT_MOD_LINEAR;
638 } else if (whandle->modifier == DRM_FORMAT_MOD_INVALID) {
639 whandle->modifier = get_tiling.modifier;
640 } else if (whandle->modifier != get_tiling.modifier) {
641 fprintf(stderr,
642 "Modifier 0x%llx vs. tiling (0x%llx) mismatch\n",
643 (long long)whandle->modifier, get_tiling.modifier);
644 goto fail;
645 }
646
647 switch (whandle->modifier) {
648 case DRM_FORMAT_MOD_LINEAR:
649 rsc->tiled = false;
650 break;
651 case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED:
652 rsc->tiled = true;
653 break;
654 default:
655 fprintf(stderr,
656 "Attempt to import unsupported modifier 0x%llx\n",
657 (long long)whandle->modifier);
658 goto fail;
659 }
660
661 rsc->vc4_format = get_resource_texture_format(prsc);
662 vc4_setup_slices(rsc, "import");
663
664 if (whandle->offset != 0) {
665 if (rsc->tiled) {
666 fprintf(stderr,
667 "Attempt to import unsupported "
668 "winsys offset %u\n",
669 whandle->offset);
670 goto fail;
671 }
672
673 rsc->slices[0].offset += whandle->offset;
674
675 if (rsc->slices[0].offset + rsc->slices[0].size >
676 rsc->bo->size) {
677 fprintf(stderr, "Attempt to import "
678 "with overflowing offset (%d + %d > %d)\n",
679 whandle->offset,
680 rsc->slices[0].size,
681 rsc->bo->size);
682 goto fail;
683 }
684 }
685
686 if (screen->ro) {
687 /* Make sure that renderonly has a handle to our buffer in the
688 * display's fd, so that a later renderonly_get_handle()
689 * returns correct handles or GEM names.
690 */
691 rsc->scanout =
692 renderonly_create_gpu_import_for_resource(prsc,
693 screen->ro,
694 NULL);
695 if (!rsc->scanout)
696 goto fail;
697 }
698
699 if (rsc->tiled && whandle->stride != slice->stride) {
700 static bool warned = false;
701 if (!warned) {
702 warned = true;
703 fprintf(stderr,
704 "Attempting to import %dx%d %s with "
705 "unsupported stride %d instead of %d\n",
706 prsc->width0, prsc->height0,
707 util_format_short_name(prsc->format),
708 whandle->stride,
709 slice->stride);
710 }
711 goto fail;
712 } else if (!rsc->tiled) {
713 slice->stride = whandle->stride;
714 }
715
716 return prsc;
717
718 fail:
719 vc4_resource_destroy(pscreen, prsc);
720 return NULL;
721 }
722
723 static struct pipe_surface *
724 vc4_create_surface(struct pipe_context *pctx,
725 struct pipe_resource *ptex,
726 const struct pipe_surface *surf_tmpl)
727 {
728 struct vc4_surface *surface = CALLOC_STRUCT(vc4_surface);
729 struct vc4_resource *rsc = vc4_resource(ptex);
730
731 if (!surface)
732 return NULL;
733
734 assert(surf_tmpl->u.tex.first_layer == surf_tmpl->u.tex.last_layer);
735
736 struct pipe_surface *psurf = &surface->base;
737 unsigned level = surf_tmpl->u.tex.level;
738
739 pipe_reference_init(&psurf->reference, 1);
740 pipe_resource_reference(&psurf->texture, ptex);
741
742 psurf->context = pctx;
743 psurf->format = surf_tmpl->format;
744 psurf->width = u_minify(ptex->width0, level);
745 psurf->height = u_minify(ptex->height0, level);
746 psurf->u.tex.level = level;
747 psurf->u.tex.first_layer = surf_tmpl->u.tex.first_layer;
748 psurf->u.tex.last_layer = surf_tmpl->u.tex.last_layer;
749 surface->offset = (rsc->slices[level].offset +
750 psurf->u.tex.first_layer * rsc->cube_map_stride);
751 surface->tiling = rsc->slices[level].tiling;
752
753 return &surface->base;
754 }
755
756 static void
757 vc4_surface_destroy(struct pipe_context *pctx, struct pipe_surface *psurf)
758 {
759 pipe_resource_reference(&psurf->texture, NULL);
760 FREE(psurf);
761 }
762
763 static void
764 vc4_dump_surface_non_msaa(struct pipe_surface *psurf)
765 {
766 struct pipe_resource *prsc = psurf->texture;
767 struct vc4_resource *rsc = vc4_resource(prsc);
768 uint32_t *map = vc4_bo_map(rsc->bo);
769 uint32_t stride = rsc->slices[0].stride / 4;
770 uint32_t width = psurf->width;
771 uint32_t height = psurf->height;
772 uint32_t chunk_w = width / 79;
773 uint32_t chunk_h = height / 40;
774 uint32_t found_colors[10];
775 uint32_t num_found_colors = 0;
776
777 if (rsc->vc4_format != VC4_TEXTURE_TYPE_RGBA32R) {
778 fprintf(stderr, "%s: Unsupported format %s\n",
779 __func__, util_format_short_name(psurf->format));
780 return;
781 }
782
783 for (int by = 0; by < height; by += chunk_h) {
784 for (int bx = 0; bx < width; bx += chunk_w) {
785 int all_found_color = -1; /* nothing found */
786
787 for (int y = by; y < MIN2(height, by + chunk_h); y++) {
788 for (int x = bx; x < MIN2(width, bx + chunk_w); x++) {
789 uint32_t pix = map[y * stride + x];
790
791 int i;
792 for (i = 0; i < num_found_colors; i++) {
793 if (pix == found_colors[i])
794 break;
795 }
796 if (i == num_found_colors &&
797 num_found_colors <
798 ARRAY_SIZE(found_colors)) {
799 found_colors[num_found_colors++] = pix;
800 }
801
802 if (i < num_found_colors) {
803 if (all_found_color == -1)
804 all_found_color = i;
805 else if (i != all_found_color)
806 all_found_color = ARRAY_SIZE(found_colors);
807 }
808 }
809 }
810 /* If all pixels for this chunk have a consistent
811 * value, then print a character for it. Either a
812 * fixed name (particularly common for piglit tests),
813 * or a runtime-generated number.
814 */
815 if (all_found_color >= 0 &&
816 all_found_color < ARRAY_SIZE(found_colors)) {
817 static const struct {
818 uint32_t val;
819 const char *c;
820 } named_colors[] = {
821 { 0xff000000, "█" },
822 { 0x00000000, "█" },
823 { 0xffff0000, "r" },
824 { 0xff00ff00, "g" },
825 { 0xff0000ff, "b" },
826 { 0xffffffff, "w" },
827 };
828 int i;
829 for (i = 0; i < ARRAY_SIZE(named_colors); i++) {
830 if (named_colors[i].val ==
831 found_colors[all_found_color]) {
832 fprintf(stderr, "%s",
833 named_colors[i].c);
834 break;
835 }
836 }
837 /* For unnamed colors, print a number and the
838 * numbers will have values printed at the
839 * end.
840 */
841 if (i == ARRAY_SIZE(named_colors)) {
842 fprintf(stderr, "%c",
843 '0' + all_found_color);
844 }
845 } else {
846 /* If there's no consistent color, print this.
847 */
848 fprintf(stderr, ".");
849 }
850 }
851 fprintf(stderr, "\n");
852 }
853
854 for (int i = 0; i < num_found_colors; i++) {
855 fprintf(stderr, "color %d: 0x%08x\n", i, found_colors[i]);
856 }
857 }
858
859 static uint32_t
860 vc4_surface_msaa_get_sample(struct pipe_surface *psurf,
861 uint32_t x, uint32_t y, uint32_t sample)
862 {
863 struct pipe_resource *prsc = psurf->texture;
864 struct vc4_resource *rsc = vc4_resource(prsc);
865 uint32_t tile_w = 32, tile_h = 32;
866 uint32_t tiles_w = DIV_ROUND_UP(psurf->width, 32);
867
868 uint32_t tile_x = x / tile_w;
869 uint32_t tile_y = y / tile_h;
870 uint32_t *tile = (vc4_bo_map(rsc->bo) +
871 VC4_TILE_BUFFER_SIZE * (tile_y * tiles_w + tile_x));
872 uint32_t subtile_x = x % tile_w;
873 uint32_t subtile_y = y % tile_h;
874
875 uint32_t quad_samples = VC4_MAX_SAMPLES * 4;
876 uint32_t tile_stride = quad_samples * tile_w / 2;
877
878 return *((uint32_t *)tile +
879 (subtile_y >> 1) * tile_stride +
880 (subtile_x >> 1) * quad_samples +
881 ((subtile_y & 1) << 1) +
882 (subtile_x & 1) +
883 sample);
884 }
885
886 static void
887 vc4_dump_surface_msaa_char(struct pipe_surface *psurf,
888 uint32_t start_x, uint32_t start_y,
889 uint32_t w, uint32_t h)
890 {
891 bool all_same_color = true;
892 uint32_t all_pix = 0;
893
894 for (int y = start_y; y < start_y + h; y++) {
895 for (int x = start_x; x < start_x + w; x++) {
896 for (int s = 0; s < VC4_MAX_SAMPLES; s++) {
897 uint32_t pix = vc4_surface_msaa_get_sample(psurf,
898 x, y,
899 s);
900 if (x == start_x && y == start_y)
901 all_pix = pix;
902 else if (all_pix != pix)
903 all_same_color = false;
904 }
905 }
906 }
907 if (all_same_color) {
908 static const struct {
909 uint32_t val;
910 const char *c;
911 } named_colors[] = {
912 { 0xff000000, "█" },
913 { 0x00000000, "█" },
914 { 0xffff0000, "r" },
915 { 0xff00ff00, "g" },
916 { 0xff0000ff, "b" },
917 { 0xffffffff, "w" },
918 };
919 int i;
920 for (i = 0; i < ARRAY_SIZE(named_colors); i++) {
921 if (named_colors[i].val == all_pix) {
922 fprintf(stderr, "%s",
923 named_colors[i].c);
924 return;
925 }
926 }
927 fprintf(stderr, "x");
928 } else {
929 fprintf(stderr, ".");
930 }
931 }
932
933 static void
934 vc4_dump_surface_msaa(struct pipe_surface *psurf)
935 {
936 uint32_t tile_w = 32, tile_h = 32;
937 uint32_t tiles_w = DIV_ROUND_UP(psurf->width, tile_w);
938 uint32_t tiles_h = DIV_ROUND_UP(psurf->height, tile_h);
939 uint32_t char_w = 140, char_h = 60;
940 uint32_t char_w_per_tile = char_w / tiles_w - 1;
941 uint32_t char_h_per_tile = char_h / tiles_h - 1;
942
943 fprintf(stderr, "Surface: %dx%d (%dx MSAA)\n",
944 psurf->width, psurf->height, psurf->texture->nr_samples);
945
946 for (int x = 0; x < (char_w_per_tile + 1) * tiles_w; x++)
947 fprintf(stderr, "-");
948 fprintf(stderr, "\n");
949
950 for (int ty = 0; ty < psurf->height; ty += tile_h) {
951 for (int y = 0; y < char_h_per_tile; y++) {
952
953 for (int tx = 0; tx < psurf->width; tx += tile_w) {
954 for (int x = 0; x < char_w_per_tile; x++) {
955 uint32_t bx1 = (x * tile_w /
956 char_w_per_tile);
957 uint32_t bx2 = ((x + 1) * tile_w /
958 char_w_per_tile);
959 uint32_t by1 = (y * tile_h /
960 char_h_per_tile);
961 uint32_t by2 = ((y + 1) * tile_h /
962 char_h_per_tile);
963
964 vc4_dump_surface_msaa_char(psurf,
965 tx + bx1,
966 ty + by1,
967 bx2 - bx1,
968 by2 - by1);
969 }
970 fprintf(stderr, "|");
971 }
972 fprintf(stderr, "\n");
973 }
974
975 for (int x = 0; x < (char_w_per_tile + 1) * tiles_w; x++)
976 fprintf(stderr, "-");
977 fprintf(stderr, "\n");
978 }
979 }
980
981 /** Debug routine to dump the contents of an 8888 surface to the console */
982 void
983 vc4_dump_surface(struct pipe_surface *psurf)
984 {
985 if (!psurf)
986 return;
987
988 if (psurf->texture->nr_samples > 1)
989 vc4_dump_surface_msaa(psurf);
990 else
991 vc4_dump_surface_non_msaa(psurf);
992 }
993
994 static void
995 vc4_flush_resource(struct pipe_context *pctx, struct pipe_resource *resource)
996 {
997 /* All calls to flush_resource are followed by a flush of the context,
998 * so there's nothing to do.
999 */
1000 }
1001
1002 void
1003 vc4_update_shadow_baselevel_texture(struct pipe_context *pctx,
1004 struct pipe_sampler_view *pview)
1005 {
1006 struct vc4_sampler_view *view = vc4_sampler_view(pview);
1007 struct vc4_resource *shadow = vc4_resource(view->texture);
1008 struct vc4_resource *orig = vc4_resource(pview->texture);
1009
1010 assert(view->texture != pview->texture);
1011
1012 if (shadow->writes == orig->writes && orig->bo->private)
1013 return;
1014
1015 perf_debug("Updating %dx%d@%d shadow texture due to %s\n",
1016 orig->base.width0, orig->base.height0,
1017 pview->u.tex.first_level,
1018 pview->u.tex.first_level ? "base level" : "raster layout");
1019
1020 for (int i = 0; i <= shadow->base.last_level; i++) {
1021 unsigned width = u_minify(shadow->base.width0, i);
1022 unsigned height = u_minify(shadow->base.height0, i);
1023 struct pipe_blit_info info = {
1024 .dst = {
1025 .resource = &shadow->base,
1026 .level = i,
1027 .box = {
1028 .x = 0,
1029 .y = 0,
1030 .z = 0,
1031 .width = width,
1032 .height = height,
1033 .depth = 1,
1034 },
1035 .format = shadow->base.format,
1036 },
1037 .src = {
1038 .resource = &orig->base,
1039 .level = pview->u.tex.first_level + i,
1040 .box = {
1041 .x = 0,
1042 .y = 0,
1043 .z = 0,
1044 .width = width,
1045 .height = height,
1046 .depth = 1,
1047 },
1048 .format = orig->base.format,
1049 },
1050 .mask = ~0,
1051 };
1052 pctx->blit(pctx, &info);
1053 }
1054
1055 shadow->writes = orig->writes;
1056 }
1057
1058 /**
1059 * Converts a 4-byte index buffer to 2 bytes.
1060 *
1061 * Since GLES2 only has support for 1 and 2-byte indices, the hardware doesn't
1062 * include 4-byte index support, and we have to shrink it down.
1063 *
1064 * There's no fallback support for when indices end up being larger than 2^16,
1065 * though it will at least assertion fail. Also, if the original index data
1066 * was in user memory, it would be nice to not have uploaded it to a VBO
1067 * before translating.
1068 */
1069 struct pipe_resource *
1070 vc4_get_shadow_index_buffer(struct pipe_context *pctx,
1071 const struct pipe_draw_info *info,
1072 uint32_t offset,
1073 uint32_t count,
1074 uint32_t *shadow_offset)
1075 {
1076 struct vc4_context *vc4 = vc4_context(pctx);
1077 struct vc4_resource *orig = vc4_resource(info->index.resource);
1078 perf_debug("Fallback conversion for %d uint indices\n", count);
1079
1080 void *data;
1081 struct pipe_resource *shadow_rsc = NULL;
1082 u_upload_alloc(vc4->uploader, 0, count * 2, 4,
1083 shadow_offset, &shadow_rsc, &data);
1084 uint16_t *dst = data;
1085
1086 struct pipe_transfer *src_transfer = NULL;
1087 const uint32_t *src;
1088 if (info->has_user_indices) {
1089 src = info->index.user;
1090 } else {
1091 src = pipe_buffer_map_range(pctx, &orig->base,
1092 offset,
1093 count * 4,
1094 PIPE_TRANSFER_READ, &src_transfer);
1095 }
1096
1097 for (int i = 0; i < count; i++) {
1098 uint32_t src_index = src[i];
1099 assert(src_index <= 0xffff);
1100 dst[i] = src_index;
1101 }
1102
1103 if (src_transfer)
1104 pctx->transfer_unmap(pctx, src_transfer);
1105
1106 return shadow_rsc;
1107 }
1108
1109 static const struct u_transfer_vtbl transfer_vtbl = {
1110 .resource_create = vc4_resource_create,
1111 .resource_destroy = vc4_resource_destroy,
1112 .transfer_map = vc4_resource_transfer_map,
1113 .transfer_unmap = vc4_resource_transfer_unmap,
1114 .transfer_flush_region = u_default_transfer_flush_region,
1115 };
1116
1117 void
1118 vc4_resource_screen_init(struct pipe_screen *pscreen)
1119 {
1120 struct vc4_screen *screen = vc4_screen(pscreen);
1121
1122 pscreen->resource_create = vc4_resource_create;
1123 pscreen->resource_create_with_modifiers =
1124 vc4_resource_create_with_modifiers;
1125 pscreen->resource_from_handle = vc4_resource_from_handle;
1126 pscreen->resource_destroy = u_resource_destroy_vtbl;
1127 pscreen->resource_get_handle = vc4_resource_get_handle;
1128 pscreen->resource_destroy = vc4_resource_destroy;
1129 pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl,
1130 false, false,
1131 false, true);
1132
1133 /* Test if the kernel has GET_TILING; it will return -EINVAL if the
1134 * ioctl does not exist, but -ENOENT if we pass an impossible handle.
1135 * 0 cannot be a valid GEM object, so use that.
1136 */
1137 struct drm_vc4_get_tiling get_tiling = {
1138 .handle = 0x0,
1139 };
1140 int ret = vc4_ioctl(screen->fd, DRM_IOCTL_VC4_GET_TILING, &get_tiling);
1141 if (ret == -1 && errno == ENOENT)
1142 screen->has_tiling_ioctl = true;
1143 }
1144
1145 void
1146 vc4_resource_context_init(struct pipe_context *pctx)
1147 {
1148 pctx->transfer_map = u_transfer_helper_transfer_map;
1149 pctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
1150 pctx->transfer_unmap = u_transfer_helper_transfer_unmap;
1151 pctx->buffer_subdata = u_default_buffer_subdata;
1152 pctx->texture_subdata = vc4_texture_subdata;
1153 pctx->create_surface = vc4_create_surface;
1154 pctx->surface_destroy = vc4_surface_destroy;
1155 pctx->resource_copy_region = util_resource_copy_region;
1156 pctx->blit = vc4_blit;
1157 pctx->flush_resource = vc4_flush_resource;
1158 }