vc4_resource.c revision 01e04c3f
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
41static bool
42vc4_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
72static void
73vc4_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
98static void *
99vc4_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
229fail:
230        vc4_resource_transfer_unmap(pctx, ptrans);
231        return NULL;
232}
233
234static void
235vc4_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
272static void
273vc4_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
286static boolean
287vc4_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
336static void
337vc4_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
435static struct vc4_resource *
436vc4_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
459static enum vc4_texture_data_type
460get_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
479static bool
480find_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
492static struct pipe_resource *
493vc4_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        /* Set up the "scanout resource" (the dmabuf export of our buffer to
576         * the KMS handle) if the buffer might ever have
577         * resource_get_handle(WINSYS_HANDLE_TYPE_KMS) called on it.
578         * create_with_modifiers() doesn't give us usage flags, so we have to
579         * assume that all calls with modifiers are scanout-possible.
580         */
581        if (screen->ro &&
582            ((tmpl->bind & PIPE_BIND_SCANOUT) ||
583             !(count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID))) {
584                rsc->scanout =
585                        renderonly_scanout_for_resource(prsc, screen->ro, NULL);
586                if (!rsc->scanout)
587                        goto fail;
588        }
589
590        vc4_bo_label(screen, rsc->bo, "%sresource %dx%d@%d/%d",
591                     (tmpl->bind & PIPE_BIND_SCANOUT) ? "scanout " : "",
592                     tmpl->width0, tmpl->height0,
593                     rsc->cpp * 8, prsc->last_level);
594
595        return prsc;
596fail:
597        vc4_resource_destroy(pscreen, prsc);
598        return NULL;
599}
600
601struct pipe_resource *
602vc4_resource_create(struct pipe_screen *pscreen,
603                    const struct pipe_resource *tmpl)
604{
605        const uint64_t mod = DRM_FORMAT_MOD_INVALID;
606        return vc4_resource_create_with_modifiers(pscreen, tmpl, &mod, 1);
607}
608
609static struct pipe_resource *
610vc4_resource_from_handle(struct pipe_screen *pscreen,
611                         const struct pipe_resource *tmpl,
612                         struct winsys_handle *whandle,
613                         unsigned usage)
614{
615        struct vc4_screen *screen = vc4_screen(pscreen);
616        struct vc4_resource *rsc = vc4_resource_setup(pscreen, tmpl);
617        struct pipe_resource *prsc = &rsc->base;
618        struct vc4_resource_slice *slice = &rsc->slices[0];
619
620        if (!rsc)
621                return NULL;
622
623        switch (whandle->type) {
624        case WINSYS_HANDLE_TYPE_SHARED:
625                rsc->bo = vc4_bo_open_name(screen,
626                                           whandle->handle, whandle->stride);
627                break;
628        case WINSYS_HANDLE_TYPE_FD:
629                rsc->bo = vc4_bo_open_dmabuf(screen,
630                                             whandle->handle, whandle->stride);
631                break;
632        default:
633                fprintf(stderr,
634                        "Attempt to import unsupported handle type %d\n",
635                        whandle->type);
636        }
637
638        if (!rsc->bo)
639                goto fail;
640
641        struct drm_vc4_get_tiling get_tiling = {
642                .handle = rsc->bo->handle,
643        };
644        int ret = vc4_ioctl(screen->fd, DRM_IOCTL_VC4_GET_TILING, &get_tiling);
645
646        if (ret != 0) {
647                whandle->modifier = DRM_FORMAT_MOD_LINEAR;
648        } else if (whandle->modifier == DRM_FORMAT_MOD_INVALID) {
649                whandle->modifier = get_tiling.modifier;
650        } else if (whandle->modifier != get_tiling.modifier) {
651                fprintf(stderr,
652                        "Modifier 0x%llx vs. tiling (0x%llx) mismatch\n",
653                        (long long)whandle->modifier, get_tiling.modifier);
654                goto fail;
655        }
656
657        switch (whandle->modifier) {
658        case DRM_FORMAT_MOD_LINEAR:
659                rsc->tiled = false;
660                break;
661        case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED:
662                rsc->tiled = true;
663                break;
664        default:
665                fprintf(stderr,
666                        "Attempt to import unsupported modifier 0x%llx\n",
667                        (long long)whandle->modifier);
668                goto fail;
669        }
670
671        rsc->vc4_format = get_resource_texture_format(prsc);
672        vc4_setup_slices(rsc, "import");
673
674        if (whandle->offset != 0) {
675                if (rsc->tiled) {
676                        fprintf(stderr,
677                                "Attempt to import unsupported "
678                                "winsys offset %u\n",
679                                whandle->offset);
680                        goto fail;
681                }
682
683                rsc->slices[0].offset += whandle->offset;
684
685                if (rsc->slices[0].offset + rsc->slices[0].size >
686                    rsc->bo->size) {
687                        fprintf(stderr, "Attempt to import "
688                                "with overflowing offset (%d + %d > %d)\n",
689                                whandle->offset,
690                                rsc->slices[0].size,
691                                rsc->bo->size);
692                        goto fail;
693                }
694        }
695
696        if (screen->ro) {
697                /* Make sure that renderonly has a handle to our buffer in the
698                 * display's fd, so that a later renderonly_get_handle()
699                 * returns correct handles or GEM names.
700                 */
701                rsc->scanout =
702                        renderonly_create_gpu_import_for_resource(prsc,
703                                                                  screen->ro,
704                                                                  NULL);
705                if (!rsc->scanout)
706                        goto fail;
707        }
708
709        if (rsc->tiled && whandle->stride != slice->stride) {
710                static bool warned = false;
711                if (!warned) {
712                        warned = true;
713                        fprintf(stderr,
714                                "Attempting to import %dx%d %s with "
715                                "unsupported stride %d instead of %d\n",
716                                prsc->width0, prsc->height0,
717                                util_format_short_name(prsc->format),
718                                whandle->stride,
719                                slice->stride);
720                }
721                goto fail;
722        } else if (!rsc->tiled) {
723                slice->stride = whandle->stride;
724        }
725
726        return prsc;
727
728fail:
729        vc4_resource_destroy(pscreen, prsc);
730        return NULL;
731}
732
733static struct pipe_surface *
734vc4_create_surface(struct pipe_context *pctx,
735                   struct pipe_resource *ptex,
736                   const struct pipe_surface *surf_tmpl)
737{
738        struct vc4_surface *surface = CALLOC_STRUCT(vc4_surface);
739        struct vc4_resource *rsc = vc4_resource(ptex);
740
741        if (!surface)
742                return NULL;
743
744        assert(surf_tmpl->u.tex.first_layer == surf_tmpl->u.tex.last_layer);
745
746        struct pipe_surface *psurf = &surface->base;
747        unsigned level = surf_tmpl->u.tex.level;
748
749        pipe_reference_init(&psurf->reference, 1);
750        pipe_resource_reference(&psurf->texture, ptex);
751
752        psurf->context = pctx;
753        psurf->format = surf_tmpl->format;
754        psurf->width = u_minify(ptex->width0, level);
755        psurf->height = u_minify(ptex->height0, level);
756        psurf->u.tex.level = level;
757        psurf->u.tex.first_layer = surf_tmpl->u.tex.first_layer;
758        psurf->u.tex.last_layer = surf_tmpl->u.tex.last_layer;
759        surface->offset = (rsc->slices[level].offset +
760                           psurf->u.tex.first_layer * rsc->cube_map_stride);
761        surface->tiling = rsc->slices[level].tiling;
762
763        return &surface->base;
764}
765
766static void
767vc4_surface_destroy(struct pipe_context *pctx, struct pipe_surface *psurf)
768{
769        pipe_resource_reference(&psurf->texture, NULL);
770        FREE(psurf);
771}
772
773static void
774vc4_dump_surface_non_msaa(struct pipe_surface *psurf)
775{
776        struct pipe_resource *prsc = psurf->texture;
777        struct vc4_resource *rsc = vc4_resource(prsc);
778        uint32_t *map = vc4_bo_map(rsc->bo);
779        uint32_t stride = rsc->slices[0].stride / 4;
780        uint32_t width = psurf->width;
781        uint32_t height = psurf->height;
782        uint32_t chunk_w = width / 79;
783        uint32_t chunk_h = height / 40;
784        uint32_t found_colors[10];
785        uint32_t num_found_colors = 0;
786
787        if (rsc->vc4_format != VC4_TEXTURE_TYPE_RGBA32R) {
788                fprintf(stderr, "%s: Unsupported format %s\n",
789                        __func__, util_format_short_name(psurf->format));
790                return;
791        }
792
793        for (int by = 0; by < height; by += chunk_h) {
794                for (int bx = 0; bx < width; bx += chunk_w) {
795                        int all_found_color = -1; /* nothing found */
796
797                        for (int y = by; y < MIN2(height, by + chunk_h); y++) {
798                                for (int x = bx; x < MIN2(width, bx + chunk_w); x++) {
799                                        uint32_t pix = map[y * stride + x];
800
801                                        int i;
802                                        for (i = 0; i < num_found_colors; i++) {
803                                                if (pix == found_colors[i])
804                                                        break;
805                                        }
806                                        if (i == num_found_colors &&
807                                            num_found_colors <
808                                            ARRAY_SIZE(found_colors)) {
809                                                found_colors[num_found_colors++] = pix;
810                                        }
811
812                                        if (i < num_found_colors) {
813                                                if (all_found_color == -1)
814                                                        all_found_color = i;
815                                                else if (i != all_found_color)
816                                                        all_found_color = ARRAY_SIZE(found_colors);
817                                        }
818                                }
819                        }
820                        /* If all pixels for this chunk have a consistent
821                         * value, then print a character for it.  Either a
822                         * fixed name (particularly common for piglit tests),
823                         * or a runtime-generated number.
824                         */
825                        if (all_found_color >= 0 &&
826                            all_found_color < ARRAY_SIZE(found_colors)) {
827                                static const struct {
828                                        uint32_t val;
829                                        const char *c;
830                                } named_colors[] = {
831                                        { 0xff000000, "█" },
832                                        { 0x00000000, "█" },
833                                        { 0xffff0000, "r" },
834                                        { 0xff00ff00, "g" },
835                                        { 0xff0000ff, "b" },
836                                        { 0xffffffff, "w" },
837                                };
838                                int i;
839                                for (i = 0; i < ARRAY_SIZE(named_colors); i++) {
840                                        if (named_colors[i].val ==
841                                            found_colors[all_found_color]) {
842                                                fprintf(stderr, "%s",
843                                                        named_colors[i].c);
844                                                break;
845                                        }
846                                }
847                                /* For unnamed colors, print a number and the
848                                 * numbers will have values printed at the
849                                 * end.
850                                 */
851                                if (i == ARRAY_SIZE(named_colors)) {
852                                        fprintf(stderr, "%c",
853                                                '0' + all_found_color);
854                                }
855                        } else {
856                                /* If there's no consistent color, print this.
857                                 */
858                                fprintf(stderr, ".");
859                        }
860                }
861                fprintf(stderr, "\n");
862        }
863
864        for (int i = 0; i < num_found_colors; i++) {
865                fprintf(stderr, "color %d: 0x%08x\n", i, found_colors[i]);
866        }
867}
868
869static uint32_t
870vc4_surface_msaa_get_sample(struct pipe_surface *psurf,
871                            uint32_t x, uint32_t y, uint32_t sample)
872{
873        struct pipe_resource *prsc = psurf->texture;
874        struct vc4_resource *rsc = vc4_resource(prsc);
875        uint32_t tile_w = 32, tile_h = 32;
876        uint32_t tiles_w = DIV_ROUND_UP(psurf->width, 32);
877
878        uint32_t tile_x = x / tile_w;
879        uint32_t tile_y = y / tile_h;
880        uint32_t *tile = (vc4_bo_map(rsc->bo) +
881                          VC4_TILE_BUFFER_SIZE * (tile_y * tiles_w + tile_x));
882        uint32_t subtile_x = x % tile_w;
883        uint32_t subtile_y = y % tile_h;
884
885        uint32_t quad_samples = VC4_MAX_SAMPLES * 4;
886        uint32_t tile_stride = quad_samples * tile_w / 2;
887
888        return *((uint32_t *)tile +
889                 (subtile_y >> 1) * tile_stride +
890                 (subtile_x >> 1) * quad_samples +
891                 ((subtile_y & 1) << 1) +
892                 (subtile_x & 1) +
893                 sample);
894}
895
896static void
897vc4_dump_surface_msaa_char(struct pipe_surface *psurf,
898                           uint32_t start_x, uint32_t start_y,
899                           uint32_t w, uint32_t h)
900{
901        bool all_same_color = true;
902        uint32_t all_pix = 0;
903
904        for (int y = start_y; y < start_y + h; y++) {
905                for (int x = start_x; x < start_x + w; x++) {
906                        for (int s = 0; s < VC4_MAX_SAMPLES; s++) {
907                                uint32_t pix = vc4_surface_msaa_get_sample(psurf,
908                                                                           x, y,
909                                                                           s);
910                                if (x == start_x && y == start_y)
911                                        all_pix = pix;
912                                else if (all_pix != pix)
913                                        all_same_color = false;
914                        }
915                }
916        }
917        if (all_same_color) {
918                static const struct {
919                        uint32_t val;
920                        const char *c;
921                } named_colors[] = {
922                        { 0xff000000, "█" },
923                        { 0x00000000, "█" },
924                        { 0xffff0000, "r" },
925                        { 0xff00ff00, "g" },
926                        { 0xff0000ff, "b" },
927                        { 0xffffffff, "w" },
928                };
929                int i;
930                for (i = 0; i < ARRAY_SIZE(named_colors); i++) {
931                        if (named_colors[i].val == all_pix) {
932                                fprintf(stderr, "%s",
933                                        named_colors[i].c);
934                                return;
935                        }
936                }
937                fprintf(stderr, "x");
938        } else {
939                fprintf(stderr, ".");
940        }
941}
942
943static void
944vc4_dump_surface_msaa(struct pipe_surface *psurf)
945{
946        uint32_t tile_w = 32, tile_h = 32;
947        uint32_t tiles_w = DIV_ROUND_UP(psurf->width, tile_w);
948        uint32_t tiles_h = DIV_ROUND_UP(psurf->height, tile_h);
949        uint32_t char_w = 140, char_h = 60;
950        uint32_t char_w_per_tile = char_w / tiles_w - 1;
951        uint32_t char_h_per_tile = char_h / tiles_h - 1;
952
953        fprintf(stderr, "Surface: %dx%d (%dx MSAA)\n",
954                psurf->width, psurf->height, psurf->texture->nr_samples);
955
956        for (int x = 0; x < (char_w_per_tile + 1) * tiles_w; x++)
957                fprintf(stderr, "-");
958        fprintf(stderr, "\n");
959
960        for (int ty = 0; ty < psurf->height; ty += tile_h) {
961                for (int y = 0; y < char_h_per_tile; y++) {
962
963                        for (int tx = 0; tx < psurf->width; tx += tile_w) {
964                                for (int x = 0; x < char_w_per_tile; x++) {
965                                        uint32_t bx1 = (x * tile_w /
966                                                        char_w_per_tile);
967                                        uint32_t bx2 = ((x + 1) * tile_w /
968                                                        char_w_per_tile);
969                                        uint32_t by1 = (y * tile_h /
970                                                        char_h_per_tile);
971                                        uint32_t by2 = ((y + 1) * tile_h /
972                                                        char_h_per_tile);
973
974                                        vc4_dump_surface_msaa_char(psurf,
975                                                                   tx + bx1,
976                                                                   ty + by1,
977                                                                   bx2 - bx1,
978                                                                   by2 - by1);
979                                }
980                                fprintf(stderr, "|");
981                        }
982                        fprintf(stderr, "\n");
983                }
984
985                for (int x = 0; x < (char_w_per_tile + 1) * tiles_w; x++)
986                        fprintf(stderr, "-");
987                fprintf(stderr, "\n");
988        }
989}
990
991/** Debug routine to dump the contents of an 8888 surface to the console */
992void
993vc4_dump_surface(struct pipe_surface *psurf)
994{
995        if (!psurf)
996                return;
997
998        if (psurf->texture->nr_samples > 1)
999                vc4_dump_surface_msaa(psurf);
1000        else
1001                vc4_dump_surface_non_msaa(psurf);
1002}
1003
1004static void
1005vc4_flush_resource(struct pipe_context *pctx, struct pipe_resource *resource)
1006{
1007        /* All calls to flush_resource are followed by a flush of the context,
1008         * so there's nothing to do.
1009         */
1010}
1011
1012void
1013vc4_update_shadow_baselevel_texture(struct pipe_context *pctx,
1014                                    struct pipe_sampler_view *pview)
1015{
1016        struct vc4_sampler_view *view = vc4_sampler_view(pview);
1017        struct vc4_resource *shadow = vc4_resource(view->texture);
1018        struct vc4_resource *orig = vc4_resource(pview->texture);
1019
1020        assert(view->texture != pview->texture);
1021
1022        if (shadow->writes == orig->writes && orig->bo->private)
1023                return;
1024
1025        perf_debug("Updating %dx%d@%d shadow texture due to %s\n",
1026                   orig->base.width0, orig->base.height0,
1027                   pview->u.tex.first_level,
1028                   pview->u.tex.first_level ? "base level" : "raster layout");
1029
1030        for (int i = 0; i <= shadow->base.last_level; i++) {
1031                unsigned width = u_minify(shadow->base.width0, i);
1032                unsigned height = u_minify(shadow->base.height0, i);
1033                struct pipe_blit_info info = {
1034                        .dst = {
1035                                .resource = &shadow->base,
1036                                .level = i,
1037                                .box = {
1038                                        .x = 0,
1039                                        .y = 0,
1040                                        .z = 0,
1041                                        .width = width,
1042                                        .height = height,
1043                                        .depth = 1,
1044                                },
1045                                .format = shadow->base.format,
1046                        },
1047                        .src = {
1048                                .resource = &orig->base,
1049                                .level = pview->u.tex.first_level + i,
1050                                .box = {
1051                                        .x = 0,
1052                                        .y = 0,
1053                                        .z = 0,
1054                                        .width = width,
1055                                        .height = height,
1056                                        .depth = 1,
1057                                },
1058                                .format = orig->base.format,
1059                        },
1060                        .mask = ~0,
1061                };
1062                pctx->blit(pctx, &info);
1063        }
1064
1065        shadow->writes = orig->writes;
1066}
1067
1068/**
1069 * Converts a 4-byte index buffer to 2 bytes.
1070 *
1071 * Since GLES2 only has support for 1 and 2-byte indices, the hardware doesn't
1072 * include 4-byte index support, and we have to shrink it down.
1073 *
1074 * There's no fallback support for when indices end up being larger than 2^16,
1075 * though it will at least assertion fail.  Also, if the original index data
1076 * was in user memory, it would be nice to not have uploaded it to a VBO
1077 * before translating.
1078 */
1079struct pipe_resource *
1080vc4_get_shadow_index_buffer(struct pipe_context *pctx,
1081                            const struct pipe_draw_info *info,
1082                            uint32_t offset,
1083                            uint32_t count,
1084                            uint32_t *shadow_offset)
1085{
1086        struct vc4_context *vc4 = vc4_context(pctx);
1087        struct vc4_resource *orig = vc4_resource(info->index.resource);
1088        perf_debug("Fallback conversion for %d uint indices\n", count);
1089
1090        void *data;
1091        struct pipe_resource *shadow_rsc = NULL;
1092        u_upload_alloc(vc4->uploader, 0, count * 2, 4,
1093                       shadow_offset, &shadow_rsc, &data);
1094        uint16_t *dst = data;
1095
1096        struct pipe_transfer *src_transfer = NULL;
1097        const uint32_t *src;
1098        if (info->has_user_indices) {
1099                src = info->index.user;
1100        } else {
1101                src = pipe_buffer_map_range(pctx, &orig->base,
1102                                            offset,
1103                                            count * 4,
1104                                            PIPE_TRANSFER_READ, &src_transfer);
1105        }
1106
1107        for (int i = 0; i < count; i++) {
1108                uint32_t src_index = src[i];
1109                assert(src_index <= 0xffff);
1110                dst[i] = src_index;
1111        }
1112
1113        if (src_transfer)
1114                pctx->transfer_unmap(pctx, src_transfer);
1115
1116        return shadow_rsc;
1117}
1118
1119static const struct u_transfer_vtbl transfer_vtbl = {
1120        .resource_create          = vc4_resource_create,
1121        .resource_destroy         = vc4_resource_destroy,
1122        .transfer_map             = vc4_resource_transfer_map,
1123        .transfer_unmap           = vc4_resource_transfer_unmap,
1124        .transfer_flush_region    = u_default_transfer_flush_region,
1125};
1126
1127void
1128vc4_resource_screen_init(struct pipe_screen *pscreen)
1129{
1130        struct vc4_screen *screen = vc4_screen(pscreen);
1131
1132        pscreen->resource_create = vc4_resource_create;
1133        pscreen->resource_create_with_modifiers =
1134                vc4_resource_create_with_modifiers;
1135        pscreen->resource_from_handle = vc4_resource_from_handle;
1136        pscreen->resource_destroy = u_resource_destroy_vtbl;
1137        pscreen->resource_get_handle = vc4_resource_get_handle;
1138        pscreen->resource_destroy = vc4_resource_destroy;
1139        pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl,
1140                                                            false, false,
1141                                                            false, true);
1142
1143        /* Test if the kernel has GET_TILING; it will return -EINVAL if the
1144         * ioctl does not exist, but -ENOENT if we pass an impossible handle.
1145         * 0 cannot be a valid GEM object, so use that.
1146         */
1147        struct drm_vc4_get_tiling get_tiling = {
1148                .handle = 0x0,
1149        };
1150        int ret = vc4_ioctl(screen->fd, DRM_IOCTL_VC4_GET_TILING, &get_tiling);
1151        if (ret == -1 && errno == ENOENT)
1152                screen->has_tiling_ioctl = true;
1153}
1154
1155void
1156vc4_resource_context_init(struct pipe_context *pctx)
1157{
1158        pctx->transfer_map = u_transfer_helper_transfer_map;
1159        pctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
1160        pctx->transfer_unmap = u_transfer_helper_transfer_unmap;
1161        pctx->buffer_subdata = u_default_buffer_subdata;
1162        pctx->texture_subdata = vc4_texture_subdata;
1163        pctx->create_surface = vc4_create_surface;
1164        pctx->surface_destroy = vc4_surface_destroy;
1165        pctx->resource_copy_region = util_resource_copy_region;
1166        pctx->blit = vc4_blit;
1167        pctx->flush_resource = vc4_flush_resource;
1168}
1169