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_memory.h"
27#include "util/format/u_format.h"
28#include "util/u_inlines.h"
29#include "util/u_surface.h"
30#include "util/u_transfer_helper.h"
31#include "util/u_upload_mgr.h"
32#include "util/u_drm.h"
33
34#include "drm-uapi/drm_fourcc.h"
35#include "drm-uapi/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_MAP_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_MAP_DISCARD_RANGE) &&
116            !(usage & PIPE_MAP_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_MAP_DISCARD_WHOLE_RESOURCE;
125        }
126
127        if (usage & PIPE_MAP_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_MAP_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_MAP_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_MAP_WRITE) {
152                rsc->writes++;
153                rsc->initialized_buffers = ~0;
154        }
155
156        trans = slab_zalloc(&vc4->transfer_pool);
157        if (!trans)
158                return NULL;
159
160        /* XXX: Handle DONTBLOCK, DISCARD_RANGE, PERSISTENT, COHERENT. */
161
162        ptrans = &trans->base;
163
164        pipe_resource_reference(&ptrans->resource, prsc);
165        ptrans->level = level;
166        ptrans->usage = usage;
167        ptrans->box = *box;
168
169        if (usage & PIPE_MAP_UNSYNCHRONIZED)
170                buf = vc4_bo_map_unsynchronized(rsc->bo);
171        else
172                buf = vc4_bo_map(rsc->bo);
173        if (!buf) {
174                fprintf(stderr, "Failed to map bo\n");
175                goto fail;
176        }
177
178        *pptrans = ptrans;
179
180        struct vc4_resource_slice *slice = &rsc->slices[level];
181        if (rsc->tiled) {
182                /* No direct mappings of tiled, since we need to manually
183                 * tile/untile.
184                 */
185                if (usage & PIPE_MAP_DIRECTLY)
186                        return NULL;
187
188                /* Our load/store routines work on entire compressed blocks. */
189                u_box_pixels_to_blocks(&ptrans->box, &ptrans->box, format);
190
191                ptrans->stride = ptrans->box.width * rsc->cpp;
192                ptrans->layer_stride = ptrans->stride * ptrans->box.height;
193
194                trans->map = malloc(ptrans->layer_stride * ptrans->box.depth);
195
196                if (usage & PIPE_MAP_READ) {
197                        vc4_load_tiled_image(trans->map, ptrans->stride,
198                                             buf + slice->offset +
199                                             ptrans->box.z * rsc->cube_map_stride,
200                                             slice->stride,
201                                             slice->tiling, rsc->cpp,
202                                             &ptrans->box);
203                }
204                return trans->map;
205        } else {
206                ptrans->stride = slice->stride;
207                ptrans->layer_stride = ptrans->stride;
208
209                return buf + slice->offset +
210                        ptrans->box.y / util_format_get_blockheight(format) * ptrans->stride +
211                        ptrans->box.x / util_format_get_blockwidth(format) * rsc->cpp +
212                        ptrans->box.z * rsc->cube_map_stride;
213        }
214
215
216fail:
217        vc4_resource_transfer_unmap(pctx, ptrans);
218        return NULL;
219}
220
221static void
222vc4_texture_subdata(struct pipe_context *pctx,
223                    struct pipe_resource *prsc,
224                    unsigned level,
225                    unsigned usage,
226                    const struct pipe_box *box,
227                    const void *data,
228                    unsigned stride,
229                    unsigned layer_stride)
230{
231        struct vc4_resource *rsc = vc4_resource(prsc);
232        struct vc4_resource_slice *slice = &rsc->slices[level];
233
234        /* For a direct mapping, we can just take the u_transfer path. */
235        if (!rsc->tiled ||
236            box->depth != 1 ||
237            (usage & PIPE_MAP_DISCARD_WHOLE_RESOURCE)) {
238                return u_default_texture_subdata(pctx, prsc, level, usage, box,
239                                                 data, stride, layer_stride);
240        }
241
242        /* Otherwise, map and store the texture data directly into the tiled
243         * texture.
244         */
245        void *buf;
246        if (usage & PIPE_MAP_UNSYNCHRONIZED)
247                buf = vc4_bo_map_unsynchronized(rsc->bo);
248        else
249                buf = vc4_bo_map(rsc->bo);
250
251        vc4_store_tiled_image(buf + slice->offset +
252                              box->z * rsc->cube_map_stride,
253                              slice->stride,
254                              (void *)data, stride,
255                              slice->tiling, rsc->cpp,
256                              box);
257}
258
259static void
260vc4_resource_destroy(struct pipe_screen *pscreen,
261                     struct pipe_resource *prsc)
262{
263        struct vc4_screen *screen = vc4_screen(pscreen);
264        struct vc4_resource *rsc = vc4_resource(prsc);
265        vc4_bo_unreference(&rsc->bo);
266
267        if (rsc->scanout)
268                renderonly_scanout_destroy(rsc->scanout, screen->ro);
269
270        free(rsc);
271}
272
273static uint64_t
274vc4_resource_modifier(struct vc4_resource *rsc)
275{
276        if (rsc->tiled)
277                return DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED;
278        else
279                return DRM_FORMAT_MOD_LINEAR;
280}
281
282static bool
283vc4_resource_get_handle(struct pipe_screen *pscreen,
284                        struct pipe_context *pctx,
285                        struct pipe_resource *prsc,
286                        struct winsys_handle *whandle,
287                        unsigned usage)
288{
289        struct vc4_screen *screen = vc4_screen(pscreen);
290        struct vc4_resource *rsc = vc4_resource(prsc);
291
292        whandle->stride = rsc->slices[0].stride;
293        whandle->offset = 0;
294        whandle->modifier = vc4_resource_modifier(rsc);
295
296        /* If we're passing some reference to our BO out to some other part of
297         * the system, then we can't do any optimizations about only us being
298         * the ones seeing it (like BO caching or shadow update avoidance).
299         */
300        rsc->bo->private = false;
301
302        switch (whandle->type) {
303        case WINSYS_HANDLE_TYPE_SHARED:
304                if (screen->ro) {
305                        /* This could probably be supported, assuming that a
306                         * control node was used for pl111.
307                         */
308                        fprintf(stderr, "flink unsupported with pl111\n");
309                        return false;
310                }
311
312                return vc4_bo_flink(rsc->bo, &whandle->handle);
313        case WINSYS_HANDLE_TYPE_KMS:
314                if (screen->ro) {
315                        return renderonly_get_handle(rsc->scanout, whandle);
316                }
317                whandle->handle = rsc->bo->handle;
318                return true;
319        case WINSYS_HANDLE_TYPE_FD:
320                /* FDs are cross-device, so we can export directly from vc4.
321                 */
322                whandle->handle = vc4_bo_get_dmabuf(rsc->bo);
323                return whandle->handle != -1;
324        }
325
326        return false;
327}
328
329static bool
330vc4_resource_get_param(struct pipe_screen *pscreen,
331                       struct pipe_context *pctx, struct pipe_resource *prsc,
332                       unsigned plane, unsigned layer, unsigned level,
333                       enum pipe_resource_param param,
334                       unsigned usage, uint64_t *value)
335{
336        struct vc4_resource *rsc = vc4_resource(prsc);
337
338        switch (param) {
339        case PIPE_RESOURCE_PARAM_STRIDE:
340                *value = rsc->slices[level].stride;
341                return true;
342        case PIPE_RESOURCE_PARAM_OFFSET:
343                *value = 0;
344                return true;
345        case PIPE_RESOURCE_PARAM_MODIFIER:
346                *value = vc4_resource_modifier(rsc);
347                return true;
348        default:
349                return false;
350        }
351}
352
353static void
354vc4_setup_slices(struct vc4_resource *rsc, const char *caller)
355{
356        struct pipe_resource *prsc = &rsc->base;
357        uint32_t width = prsc->width0;
358        uint32_t height = prsc->height0;
359        if (prsc->format == PIPE_FORMAT_ETC1_RGB8) {
360                width = (width + 3) >> 2;
361                height = (height + 3) >> 2;
362        }
363
364        uint32_t pot_width = util_next_power_of_two(width);
365        uint32_t pot_height = util_next_power_of_two(height);
366        uint32_t offset = 0;
367        uint32_t utile_w = vc4_utile_width(rsc->cpp);
368        uint32_t utile_h = vc4_utile_height(rsc->cpp);
369
370        for (int i = prsc->last_level; i >= 0; i--) {
371                struct vc4_resource_slice *slice = &rsc->slices[i];
372
373                uint32_t level_width, level_height;
374                if (i == 0) {
375                        level_width = width;
376                        level_height = height;
377                } else {
378                        level_width = u_minify(pot_width, i);
379                        level_height = u_minify(pot_height, i);
380                }
381
382                if (!rsc->tiled) {
383                        slice->tiling = VC4_TILING_FORMAT_LINEAR;
384                        if (prsc->nr_samples > 1) {
385                                /* MSAA (4x) surfaces are stored as raw tile buffer contents. */
386                                level_width = align(level_width, 32);
387                                level_height = align(level_height, 32);
388                        } else {
389                                level_width = align(level_width, utile_w);
390                        }
391                } else {
392                        if (vc4_size_is_lt(level_width, level_height,
393                                           rsc->cpp)) {
394                                slice->tiling = VC4_TILING_FORMAT_LT;
395                                level_width = align(level_width, utile_w);
396                                level_height = align(level_height, utile_h);
397                        } else {
398                                slice->tiling = VC4_TILING_FORMAT_T;
399                                level_width = align(level_width,
400                                                    4 * 2 * utile_w);
401                                level_height = align(level_height,
402                                                     4 * 2 * utile_h);
403                        }
404                }
405
406                slice->offset = offset;
407                slice->stride = (level_width * rsc->cpp *
408                                 MAX2(prsc->nr_samples, 1));
409                slice->size = level_height * slice->stride;
410
411                offset += slice->size;
412
413                if (vc4_debug & VC4_DEBUG_SURFACE) {
414                        static const char tiling_chars[] = {
415                                [VC4_TILING_FORMAT_LINEAR] = 'R',
416                                [VC4_TILING_FORMAT_LT] = 'L',
417                                [VC4_TILING_FORMAT_T] = 'T'
418                        };
419                        fprintf(stderr,
420                                "rsc %s %p (format %s: vc4 %d), %dx%d: "
421                                "level %d (%c) -> %dx%d, stride %d@0x%08x\n",
422                                caller, rsc,
423                                util_format_short_name(prsc->format),
424                                rsc->vc4_format,
425                                prsc->width0, prsc->height0,
426                                i, tiling_chars[slice->tiling],
427                                level_width, level_height,
428                                slice->stride, slice->offset);
429                }
430        }
431
432        /* The texture base pointer that has to point to level 0 doesn't have
433         * intra-page bits, so we have to align it, and thus shift up all the
434         * smaller slices.
435         */
436        uint32_t page_align_offset = (align(rsc->slices[0].offset, 4096) -
437                                      rsc->slices[0].offset);
438        if (page_align_offset) {
439                for (int i = 0; i <= prsc->last_level; i++)
440                        rsc->slices[i].offset += page_align_offset;
441        }
442
443        /* Cube map faces appear as whole miptrees at a page-aligned offset
444         * from the first face's miptree.
445         */
446        if (prsc->target == PIPE_TEXTURE_CUBE) {
447                rsc->cube_map_stride = align(rsc->slices[0].offset +
448                                             rsc->slices[0].size, 4096);
449        }
450}
451
452static struct vc4_resource *
453vc4_resource_setup(struct pipe_screen *pscreen,
454                   const struct pipe_resource *tmpl)
455{
456        struct vc4_resource *rsc = CALLOC_STRUCT(vc4_resource);
457        if (!rsc)
458                return NULL;
459        struct pipe_resource *prsc = &rsc->base;
460
461        *prsc = *tmpl;
462
463        pipe_reference_init(&prsc->reference, 1);
464        prsc->screen = pscreen;
465
466        if (prsc->nr_samples <= 1)
467                rsc->cpp = util_format_get_blocksize(tmpl->format);
468        else
469                rsc->cpp = sizeof(uint32_t);
470
471        assert(rsc->cpp);
472
473        return rsc;
474}
475
476static enum vc4_texture_data_type
477get_resource_texture_format(struct pipe_resource *prsc)
478{
479        struct vc4_resource *rsc = vc4_resource(prsc);
480        uint8_t format = vc4_get_tex_format(prsc->format);
481
482        if (!rsc->tiled) {
483                if (prsc->nr_samples > 1) {
484                        return ~0;
485                } else {
486                        if (format == VC4_TEXTURE_TYPE_RGBA8888)
487                                return VC4_TEXTURE_TYPE_RGBA32R;
488                        else
489                                return ~0;
490                }
491        }
492
493        return format;
494}
495
496static struct pipe_resource *
497vc4_resource_create_with_modifiers(struct pipe_screen *pscreen,
498                                   const struct pipe_resource *tmpl,
499                                   const uint64_t *modifiers,
500                                   int count)
501{
502        struct vc4_screen *screen = vc4_screen(pscreen);
503        struct vc4_resource *rsc = vc4_resource_setup(pscreen, tmpl);
504        struct pipe_resource *prsc = &rsc->base;
505        bool linear_ok = drm_find_modifier(DRM_FORMAT_MOD_LINEAR, modifiers, count);
506        /* Use a tiled layout if we can, for better 3D performance. */
507        bool should_tile = true;
508
509        /* VBOs/PBOs are untiled (and 1 height). */
510        if (tmpl->target == PIPE_BUFFER)
511                should_tile = false;
512
513        /* MSAA buffers are linear. */
514        if (tmpl->nr_samples > 1)
515                should_tile = false;
516
517        /* No tiling when we're sharing with another device (pl111). */
518        if (screen->ro && (tmpl->bind & PIPE_BIND_SCANOUT))
519                should_tile = false;
520
521        /* Cursors are always linear, and the user can request linear as well.
522         */
523        if (tmpl->bind & (PIPE_BIND_LINEAR | PIPE_BIND_CURSOR))
524                should_tile = false;
525
526        /* No shared objects with LT format -- the kernel only has T-format
527         * metadata.  LT objects are small enough it's not worth the trouble to
528         * give them metadata to tile.
529         */
530        if ((tmpl->bind & (PIPE_BIND_SHARED | PIPE_BIND_SCANOUT)) &&
531            vc4_size_is_lt(prsc->width0, prsc->height0, rsc->cpp))
532                should_tile = false;
533
534        /* If we're sharing or scanning out, we need the ioctl present to
535         * inform the kernel or the other side.
536         */
537        if ((tmpl->bind & (PIPE_BIND_SHARED |
538                           PIPE_BIND_SCANOUT)) && !screen->has_tiling_ioctl)
539                should_tile = false;
540
541        /* No user-specified modifier; determine our own. */
542        if (count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID) {
543                linear_ok = true;
544                rsc->tiled = should_tile;
545        } else if (should_tile &&
546                   drm_find_modifier(DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED,
547                                 modifiers, count)) {
548                rsc->tiled = true;
549        } else if (linear_ok) {
550                rsc->tiled = false;
551        } else {
552                fprintf(stderr, "Unsupported modifier requested\n");
553                return NULL;
554        }
555
556        if (tmpl->target != PIPE_BUFFER)
557                rsc->vc4_format = get_resource_texture_format(prsc);
558
559        vc4_setup_slices(rsc, "create");
560        if (!vc4_resource_bo_alloc(rsc))
561                goto fail;
562
563        if (screen->has_tiling_ioctl) {
564                uint64_t modifier;
565                if (rsc->tiled)
566                        modifier = DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED;
567                else
568                        modifier = DRM_FORMAT_MOD_LINEAR;
569                struct drm_vc4_set_tiling set_tiling = {
570                        .handle = rsc->bo->handle,
571                        .modifier = modifier,
572                };
573                int ret = vc4_ioctl(screen->fd, DRM_IOCTL_VC4_SET_TILING,
574                                    &set_tiling);
575                if (ret != 0)
576                        goto fail;
577        }
578
579        /* Set up the "scanout resource" (the dmabuf export of our buffer to
580         * the KMS handle) if the buffer might ever have
581         * resource_get_handle(WINSYS_HANDLE_TYPE_KMS) called on it.
582         * create_with_modifiers() doesn't give us usage flags, so we have to
583         * assume that all calls with modifiers are scanout-possible.
584         */
585        if (screen->ro &&
586            ((tmpl->bind & PIPE_BIND_SCANOUT) ||
587             !(count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID))) {
588                rsc->scanout =
589                        renderonly_scanout_for_resource(prsc, screen->ro, NULL);
590                if (!rsc->scanout)
591                        goto fail;
592        }
593
594        vc4_bo_label(screen, rsc->bo, "%sresource %dx%d@%d/%d",
595                     (tmpl->bind & PIPE_BIND_SCANOUT) ? "scanout " : "",
596                     tmpl->width0, tmpl->height0,
597                     rsc->cpp * 8, prsc->last_level);
598
599        return prsc;
600fail:
601        vc4_resource_destroy(pscreen, prsc);
602        return NULL;
603}
604
605struct pipe_resource *
606vc4_resource_create(struct pipe_screen *pscreen,
607                    const struct pipe_resource *tmpl)
608{
609        const uint64_t mod = DRM_FORMAT_MOD_INVALID;
610        return vc4_resource_create_with_modifiers(pscreen, tmpl, &mod, 1);
611}
612
613static struct pipe_resource *
614vc4_resource_from_handle(struct pipe_screen *pscreen,
615                         const struct pipe_resource *tmpl,
616                         struct winsys_handle *whandle,
617                         unsigned usage)
618{
619        struct vc4_screen *screen = vc4_screen(pscreen);
620        struct vc4_resource *rsc = vc4_resource_setup(pscreen, tmpl);
621        struct pipe_resource *prsc = &rsc->base;
622        struct vc4_resource_slice *slice = &rsc->slices[0];
623
624        if (!rsc)
625                return NULL;
626
627        switch (whandle->type) {
628        case WINSYS_HANDLE_TYPE_SHARED:
629                rsc->bo = vc4_bo_open_name(screen, whandle->handle);
630                break;
631        case WINSYS_HANDLE_TYPE_FD:
632                rsc->bo = vc4_bo_open_dmabuf(screen, whandle->handle);
633                break;
634        default:
635                fprintf(stderr,
636                        "Attempt to import unsupported handle type %d\n",
637                        whandle->type);
638        }
639
640        if (!rsc->bo)
641                goto fail;
642
643        struct drm_vc4_get_tiling get_tiling = {
644                .handle = rsc->bo->handle,
645        };
646        int ret = vc4_ioctl(screen->fd, DRM_IOCTL_VC4_GET_TILING, &get_tiling);
647
648        if (ret != 0) {
649                whandle->modifier = DRM_FORMAT_MOD_LINEAR;
650        } else if (whandle->modifier == DRM_FORMAT_MOD_INVALID) {
651                whandle->modifier = get_tiling.modifier;
652        } else if (whandle->modifier != get_tiling.modifier) {
653                fprintf(stderr,
654                        "Modifier 0x%llx vs. tiling (0x%llx) mismatch\n",
655                        (long long)whandle->modifier, get_tiling.modifier);
656                goto fail;
657        }
658
659        switch (whandle->modifier) {
660        case DRM_FORMAT_MOD_LINEAR:
661                rsc->tiled = false;
662                break;
663        case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED:
664                rsc->tiled = true;
665                break;
666        default:
667                fprintf(stderr,
668                        "Attempt to import unsupported modifier 0x%llx\n",
669                        (long long)whandle->modifier);
670                goto fail;
671        }
672
673        rsc->vc4_format = get_resource_texture_format(prsc);
674        vc4_setup_slices(rsc, "import");
675
676        if (whandle->offset != 0) {
677                if (rsc->tiled) {
678                        fprintf(stderr,
679                                "Attempt to import unsupported "
680                                "winsys offset %u\n",
681                                whandle->offset);
682                        goto fail;
683                }
684
685                rsc->slices[0].offset += whandle->offset;
686
687                if (rsc->slices[0].offset + rsc->slices[0].size >
688                    rsc->bo->size) {
689                        fprintf(stderr, "Attempt to import "
690                                "with overflowing offset (%d + %d > %d)\n",
691                                whandle->offset,
692                                rsc->slices[0].size,
693                                rsc->bo->size);
694                        goto fail;
695                }
696        }
697
698        if (screen->ro) {
699                /* Make sure that renderonly has a handle to our buffer in the
700                 * display's fd, so that a later renderonly_get_handle()
701                 * returns correct handles or GEM names.
702                 */
703                rsc->scanout =
704                        renderonly_create_gpu_import_for_resource(prsc,
705                                                                  screen->ro,
706                                                                  NULL);
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] = { 0 };
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_context *vc4 = vc4_context(pctx);
1017        struct vc4_sampler_view *view = vc4_sampler_view(pview);
1018        struct vc4_resource *shadow = vc4_resource(view->texture);
1019        struct vc4_resource *orig = vc4_resource(pview->texture);
1020
1021        assert(view->texture != pview->texture);
1022
1023        if (shadow->writes == orig->writes && orig->bo->private)
1024                return;
1025
1026        perf_debug("Updating %dx%d@%d shadow texture due to %s\n",
1027                   orig->base.width0, orig->base.height0,
1028                   pview->u.tex.first_level,
1029                   pview->u.tex.first_level ? "base level" : "raster layout");
1030
1031        for (int i = 0; i <= shadow->base.last_level; i++) {
1032                unsigned width = u_minify(shadow->base.width0, i);
1033                unsigned height = u_minify(shadow->base.height0, i);
1034                struct pipe_blit_info info = {
1035                        .dst = {
1036                                .resource = &shadow->base,
1037                                .level = i,
1038                                .box = {
1039                                        .x = 0,
1040                                        .y = 0,
1041                                        .z = 0,
1042                                        .width = width,
1043                                        .height = height,
1044                                        .depth = 1,
1045                                },
1046                                .format = shadow->base.format,
1047                        },
1048                        .src = {
1049                                .resource = &orig->base,
1050                                .level = pview->u.tex.first_level + i,
1051                                .box = {
1052                                        .x = 0,
1053                                        .y = 0,
1054                                        .z = 0,
1055                                        .width = width,
1056                                        .height = height,
1057                                        .depth = 1,
1058                                },
1059                                .format = orig->base.format,
1060                        },
1061                        .mask = ~0,
1062                };
1063                pctx->blit(pctx, &info);
1064        }
1065
1066        shadow->writes = orig->writes;
1067}
1068
1069/**
1070 * Converts a 4-byte index buffer to 2 bytes.
1071 *
1072 * Since GLES2 only has support for 1 and 2-byte indices, the hardware doesn't
1073 * include 4-byte index support, and we have to shrink it down.
1074 *
1075 * There's no fallback support for when indices end up being larger than 2^16,
1076 * though it will at least assertion fail.  Also, if the original index data
1077 * was in user memory, it would be nice to not have uploaded it to a VBO
1078 * before translating.
1079 */
1080struct pipe_resource *
1081vc4_get_shadow_index_buffer(struct pipe_context *pctx,
1082                            const struct pipe_draw_info *info,
1083                            uint32_t offset,
1084                            uint32_t count,
1085                            uint32_t *shadow_offset)
1086{
1087        struct vc4_context *vc4 = vc4_context(pctx);
1088        struct vc4_resource *orig = vc4_resource(info->index.resource);
1089        perf_debug("Fallback conversion for %d uint indices\n", count);
1090
1091        void *data;
1092        struct pipe_resource *shadow_rsc = NULL;
1093        u_upload_alloc(vc4->uploader, 0, count * 2, 4,
1094                       shadow_offset, &shadow_rsc, &data);
1095        uint16_t *dst = data;
1096
1097        struct pipe_transfer *src_transfer = NULL;
1098        const uint32_t *src;
1099        if (info->has_user_indices) {
1100                src = (uint32_t*)((char*)info->index.user + offset);
1101        } else {
1102                src = pipe_buffer_map_range(pctx, &orig->base,
1103                                            offset,
1104                                            count * 4,
1105                                            PIPE_MAP_READ, &src_transfer);
1106        }
1107
1108        for (int i = 0; i < count; i++) {
1109                uint32_t src_index = src[i];
1110                assert(src_index <= 0xffff);
1111                dst[i] = src_index;
1112        }
1113
1114        if (src_transfer)
1115                pctx->buffer_unmap(pctx, src_transfer);
1116
1117        return shadow_rsc;
1118}
1119
1120static const struct u_transfer_vtbl transfer_vtbl = {
1121        .resource_create          = vc4_resource_create,
1122        .resource_destroy         = vc4_resource_destroy,
1123        .transfer_map             = vc4_resource_transfer_map,
1124        .transfer_unmap           = vc4_resource_transfer_unmap,
1125        .transfer_flush_region    = u_default_transfer_flush_region,
1126};
1127
1128void
1129vc4_resource_screen_init(struct pipe_screen *pscreen)
1130{
1131        struct vc4_screen *screen = vc4_screen(pscreen);
1132
1133        pscreen->resource_create = vc4_resource_create;
1134        pscreen->resource_create_with_modifiers =
1135                vc4_resource_create_with_modifiers;
1136        pscreen->resource_from_handle = vc4_resource_from_handle;
1137        pscreen->resource_get_handle = vc4_resource_get_handle;
1138        pscreen->resource_get_param = vc4_resource_get_param;
1139        pscreen->resource_destroy = vc4_resource_destroy;
1140        pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl,
1141                                                            false, false,
1142                                                            false, true,
1143                                                            false);
1144
1145        /* Test if the kernel has GET_TILING; it will return -EINVAL if the
1146         * ioctl does not exist, but -ENOENT if we pass an impossible handle.
1147         * 0 cannot be a valid GEM object, so use that.
1148         */
1149        struct drm_vc4_get_tiling get_tiling = {
1150                .handle = 0x0,
1151        };
1152        int ret = vc4_ioctl(screen->fd, DRM_IOCTL_VC4_GET_TILING, &get_tiling);
1153        if (ret == -1 && errno == ENOENT)
1154                screen->has_tiling_ioctl = true;
1155}
1156
1157void
1158vc4_resource_context_init(struct pipe_context *pctx)
1159{
1160        pctx->buffer_map = u_transfer_helper_transfer_map;
1161        pctx->texture_map = u_transfer_helper_transfer_map;
1162        pctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
1163        pctx->buffer_unmap = u_transfer_helper_transfer_unmap;
1164        pctx->texture_unmap = u_transfer_helper_transfer_unmap;
1165        pctx->buffer_subdata = u_default_buffer_subdata;
1166        pctx->texture_subdata = vc4_texture_subdata;
1167        pctx->create_surface = vc4_create_surface;
1168        pctx->surface_destroy = vc4_surface_destroy;
1169        pctx->resource_copy_region = util_resource_copy_region;
1170        pctx->blit = vc4_blit;
1171        pctx->flush_resource = vc4_flush_resource;
1172}
1173