projects / mesa.git / blob
? search:


e169303f4a37c17eef0061b0e977914ebb1cb42d
[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,
618 whandle->handle, whandle->stride);
619 break;
620 case WINSYS_HANDLE_TYPE_FD:
621 rsc->bo = vc4_bo_open_dmabuf(screen,
622 whandle->handle, whandle->stride);
623 break;
624 default:
625 fprintf(stderr,
626 "Attempt to import unsupported handle type %d\n",
627 whandle->type);
628 }
629
630 if (!rsc->bo)
631 goto fail;
632
633 struct drm_vc4_get_tiling get_tiling = {
634 .handle = rsc->bo->handle,
635 };
636 int ret = vc4_ioctl(screen->fd, DRM_IOCTL_VC4_GET_TILING, &get_tiling);
637
638 if (ret != 0) {
639 whandle->modifier = DRM_FORMAT_MOD_LINEAR;
640 } else if (whandle->modifier == DRM_FORMAT_MOD_INVALID) {
641 whandle->modifier = get_tiling.modifier;
642 } else if (whandle->modifier != get_tiling.modifier) {
643 fprintf(stderr,
644 "Modifier 0x%llx vs. tiling (0x%llx) mismatch\n",
645 (long long)whandle->modifier, get_tiling.modifier);
646 goto fail;
647 }
648
649 switch (whandle->modifier) {
650 case DRM_FORMAT_MOD_LINEAR:
651 rsc->tiled = false;
652 break;
653 case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED:
654 rsc->tiled = true;
655 break;
656 default:
657 fprintf(stderr,
658 "Attempt to import unsupported modifier 0x%llx\n",
659 (long long)whandle->modifier);
660 goto fail;
661 }
662
663 rsc->vc4_format = get_resource_texture_format(prsc);
664 vc4_setup_slices(rsc, "import");
665
666 if (whandle->offset != 0) {
667 if (rsc->tiled) {
668 fprintf(stderr,
669 "Attempt to import unsupported "
670 "winsys offset %u\n",
671 whandle->offset);
672 goto fail;
673 }
674
675 rsc->slices[0].offset += whandle->offset;
676
677 if (rsc->slices[0].offset + rsc->slices[0].size >
678 rsc->bo->size) {
679 fprintf(stderr, "Attempt to import "
680 "with overflowing offset (%d + %d > %d)\n",
681 whandle->offset,
682 rsc->slices[0].size,
683 rsc->bo->size);
684 goto fail;
685 }
686 }
687
688 if (screen->ro) {
689 /* Make sure that renderonly has a handle to our buffer in the
690 * display's fd, so that a later renderonly_get_handle()
691 * returns correct handles or GEM names.
692 */
693 rsc->scanout =
694 renderonly_create_gpu_import_for_resource(prsc,
695 screen->ro,
696 NULL);
697 if (!rsc->scanout)
698 goto fail;
699 }
700
701 if (rsc->tiled && whandle->stride != slice->stride) {
702 static bool warned = false;
703 if (!warned) {
704 warned = true;
705 fprintf(stderr,
706 "Attempting to import %dx%d %s with "
707 "unsupported stride %d instead of %d\n",
708 prsc->width0, prsc->height0,
709 util_format_short_name(prsc->format),
710 whandle->stride,
711 slice->stride);
712 }
713 goto fail;
714 } else if (!rsc->tiled) {
715 slice->stride = whandle->stride;
716 }
717
718 return prsc;
719
720 fail:
721 vc4_resource_destroy(pscreen, prsc);
722 return NULL;
723 }
724
725 static struct pipe_surface *
726 vc4_create_surface(struct pipe_context *pctx,
727 struct pipe_resource *ptex,
728 const struct pipe_surface *surf_tmpl)
729 {
730 struct vc4_surface *surface = CALLOC_STRUCT(vc4_surface);
731 struct vc4_resource *rsc = vc4_resource(ptex);
732
733 if (!surface)
734 return NULL;
735
736 assert(surf_tmpl->u.tex.first_layer == surf_tmpl->u.tex.last_layer);
737
738 struct pipe_surface *psurf = &surface->base;
739 unsigned level = surf_tmpl->u.tex.level;
740
741 pipe_reference_init(&psurf->reference, 1);
742 pipe_resource_reference(&psurf->texture, ptex);
743
744 psurf->context = pctx;
745 psurf->format = surf_tmpl->format;
746 psurf->width = u_minify(ptex->width0, level);
747 psurf->height = u_minify(ptex->height0, level);
748 psurf->u.tex.level = level;
749 psurf->u.tex.first_layer = surf_tmpl->u.tex.first_layer;
750 psurf->u.tex.last_layer = surf_tmpl->u.tex.last_layer;
751 surface->offset = (rsc->slices[level].offset +
752 psurf->u.tex.first_layer * rsc->cube_map_stride);
753 surface->tiling = rsc->slices[level].tiling;
754
755 return &surface->base;
756 }
757
758 static void
759 vc4_surface_destroy(struct pipe_context *pctx, struct pipe_surface *psurf)
760 {
761 pipe_resource_reference(&psurf->texture, NULL);
762 FREE(psurf);
763 }
764
765 static void
766 vc4_dump_surface_non_msaa(struct pipe_surface *psurf)
767 {
768 struct pipe_resource *prsc = psurf->texture;
769 struct vc4_resource *rsc = vc4_resource(prsc);
770 uint32_t *map = vc4_bo_map(rsc->bo);
771 uint32_t stride = rsc->slices[0].stride / 4;
772 uint32_t width = psurf->width;
773 uint32_t height = psurf->height;
774 uint32_t chunk_w = width / 79;
775 uint32_t chunk_h = height / 40;
776 uint32_t found_colors[10];
777 uint32_t num_found_colors = 0;
778
779 if (rsc->vc4_format != VC4_TEXTURE_TYPE_RGBA32R) {
780 fprintf(stderr, "%s: Unsupported format %s\n",
781 __func__, util_format_short_name(psurf->format));
782 return;
783 }
784
785 for (int by = 0; by < height; by += chunk_h) {
786 for (int bx = 0; bx < width; bx += chunk_w) {
787 int all_found_color = -1; /* nothing found */
788
789 for (int y = by; y < MIN2(height, by + chunk_h); y++) {
790 for (int x = bx; x < MIN2(width, bx + chunk_w); x++) {
791 uint32_t pix = map[y * stride + x];
792
793 int i;
794 for (i = 0; i < num_found_colors; i++) {
795 if (pix == found_colors[i])
796 break;
797 }
798 if (i == num_found_colors &&
799 num_found_colors <
800 ARRAY_SIZE(found_colors)) {
801 found_colors[num_found_colors++] = pix;
802 }
803
804 if (i < num_found_colors) {
805 if (all_found_color == -1)
806 all_found_color = i;
807 else if (i != all_found_color)
808 all_found_color = ARRAY_SIZE(found_colors);
809 }
810 }
811 }
812 /* If all pixels for this chunk have a consistent
813 * value, then print a character for it. Either a
814 * fixed name (particularly common for piglit tests),
815 * or a runtime-generated number.
816 */
817 if (all_found_color >= 0 &&
818 all_found_color < ARRAY_SIZE(found_colors)) {
819 static const struct {
820 uint32_t val;
821 const char *c;
822 } named_colors[] = {
823 { 0xff000000, "█" },
824 { 0x00000000, "█" },
825 { 0xffff0000, "r" },
826 { 0xff00ff00, "g" },
827 { 0xff0000ff, "b" },
828 { 0xffffffff, "w" },
829 };
830 int i;
831 for (i = 0; i < ARRAY_SIZE(named_colors); i++) {
832 if (named_colors[i].val ==
833 found_colors[all_found_color]) {
834 fprintf(stderr, "%s",
835 named_colors[i].c);
836 break;
837 }
838 }
839 /* For unnamed colors, print a number and the
840 * numbers will have values printed at the
841 * end.
842 */
843 if (i == ARRAY_SIZE(named_colors)) {
844 fprintf(stderr, "%c",
845 '0' + all_found_color);
846 }
847 } else {
848 /* If there's no consistent color, print this.
849 */
850 fprintf(stderr, ".");
851 }
852 }
853 fprintf(stderr, "\n");
854 }
855
856 for (int i = 0; i < num_found_colors; i++) {
857 fprintf(stderr, "color %d: 0x%08x\n", i, found_colors[i]);
858 }
859 }
860
861 static uint32_t
862 vc4_surface_msaa_get_sample(struct pipe_surface *psurf,
863 uint32_t x, uint32_t y, uint32_t sample)
864 {
865 struct pipe_resource *prsc = psurf->texture;
866 struct vc4_resource *rsc = vc4_resource(prsc);
867 uint32_t tile_w = 32, tile_h = 32;
868 uint32_t tiles_w = DIV_ROUND_UP(psurf->width, 32);
869
870 uint32_t tile_x = x / tile_w;
871 uint32_t tile_y = y / tile_h;
872 uint32_t *tile = (vc4_bo_map(rsc->bo) +
873 VC4_TILE_BUFFER_SIZE * (tile_y * tiles_w + tile_x));
874 uint32_t subtile_x = x % tile_w;
875 uint32_t subtile_y = y % tile_h;
876
877 uint32_t quad_samples = VC4_MAX_SAMPLES * 4;
878 uint32_t tile_stride = quad_samples * tile_w / 2;
879
880 return *((uint32_t *)tile +
881 (subtile_y >> 1) * tile_stride +
882 (subtile_x >> 1) * quad_samples +
883 ((subtile_y & 1) << 1) +
884 (subtile_x & 1) +
885 sample);
886 }
887
888 static void
889 vc4_dump_surface_msaa_char(struct pipe_surface *psurf,
890 uint32_t start_x, uint32_t start_y,
891 uint32_t w, uint32_t h)
892 {
893 bool all_same_color = true;
894 uint32_t all_pix = 0;
895
896 for (int y = start_y; y < start_y + h; y++) {
897 for (int x = start_x; x < start_x + w; x++) {
898 for (int s = 0; s < VC4_MAX_SAMPLES; s++) {
899 uint32_t pix = vc4_surface_msaa_get_sample(psurf,
900 x, y,
901 s);
902 if (x == start_x && y == start_y)
903 all_pix = pix;
904 else if (all_pix != pix)
905 all_same_color = false;
906 }
907 }
908 }
909 if (all_same_color) {
910 static const struct {
911 uint32_t val;
912 const char *c;
913 } named_colors[] = {
914 { 0xff000000, "█" },
915 { 0x00000000, "█" },
916 { 0xffff0000, "r" },
917 { 0xff00ff00, "g" },
918 { 0xff0000ff, "b" },
919 { 0xffffffff, "w" },
920 };
921 int i;
922 for (i = 0; i < ARRAY_SIZE(named_colors); i++) {
923 if (named_colors[i].val == all_pix) {
924 fprintf(stderr, "%s",
925 named_colors[i].c);
926 return;
927 }
928 }
929 fprintf(stderr, "x");
930 } else {
931 fprintf(stderr, ".");
932 }
933 }
934
935 static void
936 vc4_dump_surface_msaa(struct pipe_surface *psurf)
937 {
938 uint32_t tile_w = 32, tile_h = 32;
939 uint32_t tiles_w = DIV_ROUND_UP(psurf->width, tile_w);
940 uint32_t tiles_h = DIV_ROUND_UP(psurf->height, tile_h);
941 uint32_t char_w = 140, char_h = 60;
942 uint32_t char_w_per_tile = char_w / tiles_w - 1;
943 uint32_t char_h_per_tile = char_h / tiles_h - 1;
944
945 fprintf(stderr, "Surface: %dx%d (%dx MSAA)\n",
946 psurf->width, psurf->height, psurf->texture->nr_samples);
947
948 for (int x = 0; x < (char_w_per_tile + 1) * tiles_w; x++)
949 fprintf(stderr, "-");
950 fprintf(stderr, "\n");
951
952 for (int ty = 0; ty < psurf->height; ty += tile_h) {
953 for (int y = 0; y < char_h_per_tile; y++) {
954
955 for (int tx = 0; tx < psurf->width; tx += tile_w) {
956 for (int x = 0; x < char_w_per_tile; x++) {
957 uint32_t bx1 = (x * tile_w /
958 char_w_per_tile);
959 uint32_t bx2 = ((x + 1) * tile_w /
960 char_w_per_tile);
961 uint32_t by1 = (y * tile_h /
962 char_h_per_tile);
963 uint32_t by2 = ((y + 1) * tile_h /
964 char_h_per_tile);
965
966 vc4_dump_surface_msaa_char(psurf,
967 tx + bx1,
968 ty + by1,
969 bx2 - bx1,
970 by2 - by1);
971 }
972 fprintf(stderr, "|");
973 }
974 fprintf(stderr, "\n");
975 }
976
977 for (int x = 0; x < (char_w_per_tile + 1) * tiles_w; x++)
978 fprintf(stderr, "-");
979 fprintf(stderr, "\n");
980 }
981 }
982
983 /** Debug routine to dump the contents of an 8888 surface to the console */
984 void
985 vc4_dump_surface(struct pipe_surface *psurf)
986 {
987 if (!psurf)
988 return;
989
990 if (psurf->texture->nr_samples > 1)
991 vc4_dump_surface_msaa(psurf);
992 else
993 vc4_dump_surface_non_msaa(psurf);
994 }
995
996 static void
997 vc4_flush_resource(struct pipe_context *pctx, struct pipe_resource *resource)
998 {
999 /* All calls to flush_resource are followed by a flush of the context,
1000 * so there's nothing to do.
1001 */
1002 }
1003
1004 void
1005 vc4_update_shadow_baselevel_texture(struct pipe_context *pctx,
1006 struct pipe_sampler_view *pview)
1007 {
1008 struct vc4_sampler_view *view = vc4_sampler_view(pview);
1009 struct vc4_resource *shadow = vc4_resource(view->texture);
1010 struct vc4_resource *orig = vc4_resource(pview->texture);
1011
1012 assert(view->texture != pview->texture);
1013
1014 if (shadow->writes == orig->writes && orig->bo->private)
1015 return;
1016
1017 perf_debug("Updating %dx%d@%d shadow texture due to %s\n",
1018 orig->base.width0, orig->base.height0,
1019 pview->u.tex.first_level,
1020 pview->u.tex.first_level ? "base level" : "raster layout");
1021
1022 for (int i = 0; i <= shadow->base.last_level; i++) {
1023 unsigned width = u_minify(shadow->base.width0, i);
1024 unsigned height = u_minify(shadow->base.height0, i);
1025 struct pipe_blit_info info = {
1026 .dst = {
1027 .resource = &shadow->base,
1028 .level = i,
1029 .box = {
1030 .x = 0,
1031 .y = 0,
1032 .z = 0,
1033 .width = width,
1034 .height = height,
1035 .depth = 1,
1036 },
1037 .format = shadow->base.format,
1038 },
1039 .src = {
1040 .resource = &orig->base,
1041 .level = pview->u.tex.first_level + i,
1042 .box = {
1043 .x = 0,
1044 .y = 0,
1045 .z = 0,
1046 .width = width,
1047 .height = height,
1048 .depth = 1,
1049 },
1050 .format = orig->base.format,
1051 },
1052 .mask = ~0,
1053 };
1054 pctx->blit(pctx, &info);
1055 }
1056
1057 shadow->writes = orig->writes;
1058 }
1059
1060 /**
1061 * Converts a 4-byte index buffer to 2 bytes.
1062 *
1063 * Since GLES2 only has support for 1 and 2-byte indices, the hardware doesn't
1064 * include 4-byte index support, and we have to shrink it down.
1065 *
1066 * There's no fallback support for when indices end up being larger than 2^16,
1067 * though it will at least assertion fail. Also, if the original index data
1068 * was in user memory, it would be nice to not have uploaded it to a VBO
1069 * before translating.
1070 */
1071 struct pipe_resource *
1072 vc4_get_shadow_index_buffer(struct pipe_context *pctx,
1073 const struct pipe_draw_info *info,
1074 uint32_t offset,
1075 uint32_t count,
1076 uint32_t *shadow_offset)
1077 {
1078 struct vc4_context *vc4 = vc4_context(pctx);
1079 struct vc4_resource *orig = vc4_resource(info->index.resource);
1080 perf_debug("Fallback conversion for %d uint indices\n", count);
1081
1082 void *data;
1083 struct pipe_resource *shadow_rsc = NULL;
1084 u_upload_alloc(vc4->uploader, 0, count * 2, 4,
1085 shadow_offset, &shadow_rsc, &data);
1086 uint16_t *dst = data;
1087
1088 struct pipe_transfer *src_transfer = NULL;
1089 const uint32_t *src;
1090 if (info->has_user_indices) {
1091 src = info->index.user;
1092 } else {
1093 src = pipe_buffer_map_range(pctx, &orig->base,
1094 offset,
1095 count * 4,
1096 PIPE_TRANSFER_READ, &src_transfer);
1097 }
1098
1099 for (int i = 0; i < count; i++) {
1100 uint32_t src_index = src[i];
1101 assert(src_index <= 0xffff);
1102 dst[i] = src_index;
1103 }
1104
1105 if (src_transfer)
1106 pctx->transfer_unmap(pctx, src_transfer);
1107
1108 return shadow_rsc;
1109 }
1110
1111 static const struct u_transfer_vtbl transfer_vtbl = {
1112 .resource_create = vc4_resource_create,
1113 .resource_destroy = vc4_resource_destroy,
1114 .transfer_map = vc4_resource_transfer_map,
1115 .transfer_unmap = vc4_resource_transfer_unmap,
1116 .transfer_flush_region = u_default_transfer_flush_region,
1117 };
1118
1119 void
1120 vc4_resource_screen_init(struct pipe_screen *pscreen)
1121 {
1122 struct vc4_screen *screen = vc4_screen(pscreen);
1123
1124 pscreen->resource_create = vc4_resource_create;
1125 pscreen->resource_create_with_modifiers =
1126 vc4_resource_create_with_modifiers;
1127 pscreen->resource_from_handle = vc4_resource_from_handle;
1128 pscreen->resource_destroy = u_resource_destroy_vtbl;
1129 pscreen->resource_get_handle = vc4_resource_get_handle;
1130 pscreen->resource_destroy = vc4_resource_destroy;
1131 pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl,
1132 false, false, true);
1133
1134 /* Test if the kernel has GET_TILING; it will return -EINVAL if the
1135 * ioctl does not exist, but -ENOENT if we pass an impossible handle.
1136 * 0 cannot be a valid GEM object, so use that.
1137 */
1138 struct drm_vc4_get_tiling get_tiling = {
1139 .handle = 0x0,
1140 };
1141 int ret = vc4_ioctl(screen->fd, DRM_IOCTL_VC4_GET_TILING, &get_tiling);
1142 if (ret == -1 && errno == ENOENT)
1143 screen->has_tiling_ioctl = true;
1144 }
1145
1146 void
1147 vc4_resource_context_init(struct pipe_context *pctx)
1148 {
1149 pctx->transfer_map = u_transfer_helper_transfer_map;
1150 pctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
1151 pctx->transfer_unmap = u_transfer_helper_transfer_unmap;
1152 pctx->buffer_subdata = u_default_buffer_subdata;
1153 pctx->texture_subdata = vc4_texture_subdata;
1154 pctx->create_surface = vc4_create_surface;
1155 pctx->surface_destroy = vc4_surface_destroy;
1156 pctx->resource_copy_region = util_resource_copy_region;
1157 pctx->blit = vc4_blit;
1158 pctx->flush_resource = vc4_flush_resource;
1159 }