1/*
2 * Copyright © 2014-2017 Broadcom
3 * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 * IN THE SOFTWARE.
23 */
24
25#include "pipe/p_defines.h"
26#include "util/u_blit.h"
27#include "util/u_memory.h"
28#include "util/u_format.h"
29#include "util/u_inlines.h"
30#include "util/u_surface.h"
31#include "util/u_transfer_helper.h"
32#include "util/u_upload_mgr.h"
33#include "util/u_format_zs.h"
34#include "util/u_drm.h"
35
36#include "drm-uapi/drm_fourcc.h"
37#include "v3d_screen.h"
38#include "v3d_context.h"
39#include "v3d_resource.h"
40#include "v3d_tiling.h"
41#include "broadcom/cle/v3d_packet_v33_pack.h"
42
43static void
44v3d_debug_resource_layout(struct v3d_resource *rsc, const char *caller)
45{
46        if (!(V3D_DEBUG & V3D_DEBUG_SURFACE))
47                return;
48
49        struct pipe_resource *prsc = &rsc->base;
50
51        if (prsc->target == PIPE_BUFFER) {
52                fprintf(stderr,
53                        "rsc %s %p (format %s), %dx%d buffer @0x%08x-0x%08x\n",
54                        caller, rsc,
55                        util_format_short_name(prsc->format),
56                        prsc->width0, prsc->height0,
57                        rsc->bo->offset,
58                        rsc->bo->offset + rsc->bo->size - 1);
59                return;
60        }
61
62        static const char *const tiling_descriptions[] = {
63                [VC5_TILING_RASTER] = "R",
64                [VC5_TILING_LINEARTILE] = "LT",
65                [VC5_TILING_UBLINEAR_1_COLUMN] = "UB1",
66                [VC5_TILING_UBLINEAR_2_COLUMN] = "UB2",
67                [VC5_TILING_UIF_NO_XOR] = "UIF",
68                [VC5_TILING_UIF_XOR] = "UIF^",
69        };
70
71        for (int i = 0; i <= prsc->last_level; i++) {
72                struct v3d_resource_slice *slice = &rsc->slices[i];
73
74                int level_width = slice->stride / rsc->cpp;
75                int level_height = slice->padded_height;
76                int level_depth =
77                        u_minify(util_next_power_of_two(prsc->depth0), i);
78
79                fprintf(stderr,
80                        "rsc %s %p (format %s), %dx%d: "
81                        "level %d (%s) %dx%dx%d -> %dx%dx%d, stride %d@0x%08x\n",
82                        caller, rsc,
83                        util_format_short_name(prsc->format),
84                        prsc->width0, prsc->height0,
85                        i, tiling_descriptions[slice->tiling],
86                        u_minify(prsc->width0, i),
87                        u_minify(prsc->height0, i),
88                        u_minify(prsc->depth0, i),
89                        level_width,
90                        level_height,
91                        level_depth,
92                        slice->stride,
93                        rsc->bo->offset + slice->offset);
94        }
95}
96
97static bool
98v3d_resource_bo_alloc(struct v3d_resource *rsc)
99{
100        struct pipe_resource *prsc = &rsc->base;
101        struct pipe_screen *pscreen = prsc->screen;
102        struct v3d_bo *bo;
103
104        bo = v3d_bo_alloc(v3d_screen(pscreen), rsc->size, "resource");
105        if (bo) {
106                v3d_bo_unreference(&rsc->bo);
107                rsc->bo = bo;
108                v3d_debug_resource_layout(rsc, "alloc");
109                return true;
110        } else {
111                return false;
112        }
113}
114
115static void
116v3d_resource_transfer_unmap(struct pipe_context *pctx,
117                            struct pipe_transfer *ptrans)
118{
119        struct v3d_context *v3d = v3d_context(pctx);
120        struct v3d_transfer *trans = v3d_transfer(ptrans);
121
122        if (trans->map) {
123                struct v3d_resource *rsc = v3d_resource(ptrans->resource);
124                struct v3d_resource_slice *slice = &rsc->slices[ptrans->level];
125
126                if (ptrans->usage & PIPE_TRANSFER_WRITE) {
127                        for (int z = 0; z < ptrans->box.depth; z++) {
128                                void *dst = rsc->bo->map +
129                                        v3d_layer_offset(&rsc->base,
130                                                         ptrans->level,
131                                                         ptrans->box.z + z);
132                                v3d_store_tiled_image(dst,
133                                                      slice->stride,
134                                                      (trans->map +
135                                                       ptrans->stride *
136                                                       ptrans->box.height * z),
137                                                      ptrans->stride,
138                                                      slice->tiling, rsc->cpp,
139                                                      slice->padded_height,
140                                                      &ptrans->box);
141                        }
142                }
143                free(trans->map);
144        }
145
146        pipe_resource_reference(&ptrans->resource, NULL);
147        slab_free(&v3d->transfer_pool, ptrans);
148}
149
150static void
151v3d_map_usage_prep(struct pipe_context *pctx,
152                   struct pipe_resource *prsc,
153                   unsigned usage)
154{
155        struct v3d_context *v3d = v3d_context(pctx);
156        struct v3d_resource *rsc = v3d_resource(prsc);
157
158        if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
159                if (v3d_resource_bo_alloc(rsc)) {
160                        /* If it might be bound as one of our vertex buffers
161                         * or UBOs, make sure we re-emit vertex buffer state
162                         * or uniforms.
163                         */
164                        if (prsc->bind & PIPE_BIND_VERTEX_BUFFER)
165                                v3d->dirty |= VC5_DIRTY_VTXBUF;
166                        if (prsc->bind & PIPE_BIND_CONSTANT_BUFFER)
167                                v3d->dirty |= VC5_DIRTY_CONSTBUF;
168                } else {
169                        /* If we failed to reallocate, flush users so that we
170                         * don't violate any syncing requirements.
171                         */
172                        v3d_flush_jobs_reading_resource(v3d, prsc);
173                }
174        } else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
175                /* If we're writing and the buffer is being used by the CL, we
176                 * have to flush the CL first.  If we're only reading, we need
177                 * to flush if the CL has written our buffer.
178                 */
179                if (usage & PIPE_TRANSFER_WRITE)
180                        v3d_flush_jobs_reading_resource(v3d, prsc);
181                else
182                        v3d_flush_jobs_writing_resource(v3d, prsc);
183        }
184
185        if (usage & PIPE_TRANSFER_WRITE) {
186                rsc->writes++;
187                rsc->initialized_buffers = ~0;
188        }
189}
190
191static void *
192v3d_resource_transfer_map(struct pipe_context *pctx,
193                          struct pipe_resource *prsc,
194                          unsigned level, unsigned usage,
195                          const struct pipe_box *box,
196                          struct pipe_transfer **pptrans)
197{
198        struct v3d_context *v3d = v3d_context(pctx);
199        struct v3d_resource *rsc = v3d_resource(prsc);
200        struct v3d_transfer *trans;
201        struct pipe_transfer *ptrans;
202        enum pipe_format format = prsc->format;
203        char *buf;
204
205        /* MSAA maps should have been handled by u_transfer_helper. */
206        assert(prsc->nr_samples <= 1);
207
208        /* Upgrade DISCARD_RANGE to WHOLE_RESOURCE if the whole resource is
209         * being mapped.
210         */
211        if ((usage & PIPE_TRANSFER_DISCARD_RANGE) &&
212            !(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
213            !(prsc->flags & PIPE_RESOURCE_FLAG_MAP_PERSISTENT) &&
214            prsc->last_level == 0 &&
215            prsc->width0 == box->width &&
216            prsc->height0 == box->height &&
217            prsc->depth0 == box->depth &&
218            prsc->array_size == 1 &&
219            rsc->bo->private) {
220                usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE;
221        }
222
223        v3d_map_usage_prep(pctx, prsc, usage);
224
225        trans = slab_alloc(&v3d->transfer_pool);
226        if (!trans)
227                return NULL;
228
229        /* XXX: Handle DONTBLOCK, DISCARD_RANGE, PERSISTENT, COHERENT. */
230
231        /* slab_alloc_st() doesn't zero: */
232        memset(trans, 0, sizeof(*trans));
233        ptrans = &trans->base;
234
235        pipe_resource_reference(&ptrans->resource, prsc);
236        ptrans->level = level;
237        ptrans->usage = usage;
238        ptrans->box = *box;
239
240        /* Note that the current kernel implementation is synchronous, so no
241         * need to do syncing stuff here yet.
242         */
243
244        if (usage & PIPE_TRANSFER_UNSYNCHRONIZED)
245                buf = v3d_bo_map_unsynchronized(rsc->bo);
246        else
247                buf = v3d_bo_map(rsc->bo);
248        if (!buf) {
249                fprintf(stderr, "Failed to map bo\n");
250                goto fail;
251        }
252
253        *pptrans = ptrans;
254
255        /* Our load/store routines work on entire compressed blocks. */
256        ptrans->box.x /= util_format_get_blockwidth(format);
257        ptrans->box.y /= util_format_get_blockheight(format);
258        ptrans->box.width = DIV_ROUND_UP(ptrans->box.width,
259                                         util_format_get_blockwidth(format));
260        ptrans->box.height = DIV_ROUND_UP(ptrans->box.height,
261                                          util_format_get_blockheight(format));
262
263        struct v3d_resource_slice *slice = &rsc->slices[level];
264        if (rsc->tiled) {
265                /* No direct mappings of tiled, since we need to manually
266                 * tile/untile.
267                 */
268                if (usage & PIPE_TRANSFER_MAP_DIRECTLY)
269                        return NULL;
270
271                ptrans->stride = ptrans->box.width * rsc->cpp;
272                ptrans->layer_stride = ptrans->stride * ptrans->box.height;
273
274                trans->map = malloc(ptrans->layer_stride * ptrans->box.depth);
275
276                if (usage & PIPE_TRANSFER_READ) {
277                        for (int z = 0; z < ptrans->box.depth; z++) {
278                                void *src = rsc->bo->map +
279                                        v3d_layer_offset(&rsc->base,
280                                                         ptrans->level,
281                                                         ptrans->box.z + z);
282                                v3d_load_tiled_image((trans->map +
283                                                      ptrans->stride *
284                                                      ptrans->box.height * z),
285                                                     ptrans->stride,
286                                                     src,
287                                                     slice->stride,
288                                                     slice->tiling, rsc->cpp,
289                                                     slice->padded_height,
290                                             &ptrans->box);
291                        }
292                }
293                return trans->map;
294        } else {
295                ptrans->stride = slice->stride;
296                ptrans->layer_stride = rsc->cube_map_stride;
297
298                return buf + slice->offset +
299                        ptrans->box.y * ptrans->stride +
300                        ptrans->box.x * rsc->cpp +
301                        ptrans->box.z * rsc->cube_map_stride;
302        }
303
304
305fail:
306        v3d_resource_transfer_unmap(pctx, ptrans);
307        return NULL;
308}
309
310static void
311v3d_texture_subdata(struct pipe_context *pctx,
312                    struct pipe_resource *prsc,
313                    unsigned level,
314                    unsigned usage,
315                    const struct pipe_box *box,
316                    const void *data,
317                    unsigned stride,
318                    unsigned layer_stride)
319{
320        struct v3d_resource *rsc = v3d_resource(prsc);
321        struct v3d_resource_slice *slice = &rsc->slices[level];
322
323        /* For a direct mapping, we can just take the u_transfer path. */
324        if (!rsc->tiled) {
325                return u_default_texture_subdata(pctx, prsc, level, usage, box,
326                                                 data, stride, layer_stride);
327        }
328
329        /* Otherwise, map and store the texture data directly into the tiled
330         * texture.  Note that gallium's texture_subdata may be called with
331         * obvious usage flags missing!
332         */
333        v3d_map_usage_prep(pctx, prsc, usage | (PIPE_TRANSFER_WRITE |
334                                                PIPE_TRANSFER_DISCARD_RANGE));
335
336        void *buf;
337        if (usage & PIPE_TRANSFER_UNSYNCHRONIZED)
338                buf = v3d_bo_map_unsynchronized(rsc->bo);
339        else
340                buf = v3d_bo_map(rsc->bo);
341
342        for (int i = 0; i < box->depth; i++) {
343                v3d_store_tiled_image(buf +
344                                      v3d_layer_offset(&rsc->base,
345                                                       level,
346                                                       box->z + i),
347                                      slice->stride,
348                                      (void *)data + layer_stride * i,
349                                      stride,
350                                      slice->tiling, rsc->cpp, slice->padded_height,
351                                      box);
352        }
353}
354
355static void
356v3d_resource_destroy(struct pipe_screen *pscreen,
357                     struct pipe_resource *prsc)
358{
359        struct v3d_screen *screen = v3d_screen(pscreen);
360        struct v3d_resource *rsc = v3d_resource(prsc);
361
362        if (rsc->scanout)
363                renderonly_scanout_destroy(rsc->scanout, screen->ro);
364
365        v3d_bo_unreference(&rsc->bo);
366        free(rsc);
367}
368
369static boolean
370v3d_resource_get_handle(struct pipe_screen *pscreen,
371                        struct pipe_context *pctx,
372                        struct pipe_resource *prsc,
373                        struct winsys_handle *whandle,
374                        unsigned usage)
375{
376        struct v3d_screen *screen = v3d_screen(pscreen);
377        struct v3d_resource *rsc = v3d_resource(prsc);
378        struct v3d_bo *bo = rsc->bo;
379
380        whandle->stride = rsc->slices[0].stride;
381        whandle->offset = 0;
382
383        /* If we're passing some reference to our BO out to some other part of
384         * the system, then we can't do any optimizations about only us being
385         * the ones seeing it (like BO caching).
386         */
387        bo->private = false;
388
389        if (rsc->tiled) {
390                /* A shared tiled buffer should always be allocated as UIF,
391                 * not UBLINEAR or LT.
392                 */
393                assert(rsc->slices[0].tiling == VC5_TILING_UIF_XOR ||
394                       rsc->slices[0].tiling == VC5_TILING_UIF_NO_XOR);
395                whandle->modifier = DRM_FORMAT_MOD_BROADCOM_UIF;
396        } else {
397                whandle->modifier = DRM_FORMAT_MOD_LINEAR;
398        }
399
400        switch (whandle->type) {
401        case WINSYS_HANDLE_TYPE_SHARED:
402                return v3d_bo_flink(bo, &whandle->handle);
403        case WINSYS_HANDLE_TYPE_KMS:
404                if (screen->ro) {
405                        assert(rsc->scanout);
406                        bool ok = renderonly_get_handle(rsc->scanout, whandle);
407                        whandle->stride = rsc->slices[0].stride;
408                        return ok;
409                }
410                whandle->handle = bo->handle;
411                return TRUE;
412        case WINSYS_HANDLE_TYPE_FD:
413                whandle->handle = v3d_bo_get_dmabuf(bo);
414                return whandle->handle != -1;
415        }
416
417        return FALSE;
418}
419
420#define PAGE_UB_ROWS (VC5_UIFCFG_PAGE_SIZE / VC5_UIFBLOCK_ROW_SIZE)
421#define PAGE_UB_ROWS_TIMES_1_5 ((PAGE_UB_ROWS * 3) >> 1)
422#define PAGE_CACHE_UB_ROWS (VC5_PAGE_CACHE_SIZE / VC5_UIFBLOCK_ROW_SIZE)
423#define PAGE_CACHE_MINUS_1_5_UB_ROWS (PAGE_CACHE_UB_ROWS - PAGE_UB_ROWS_TIMES_1_5)
424
425/**
426 * Computes the HW's UIFblock padding for a given height/cpp.
427 *
428 * The goal of the padding is to keep pages of the same color (bank number) at
429 * least half a page away from each other vertically when crossing between
430 * between columns of UIF blocks.
431 */
432static uint32_t
433v3d_get_ub_pad(struct v3d_resource *rsc, uint32_t height)
434{
435        uint32_t utile_h = v3d_utile_height(rsc->cpp);
436        uint32_t uif_block_h = utile_h * 2;
437        uint32_t height_ub = height / uif_block_h;
438
439        uint32_t height_offset_in_pc = height_ub % PAGE_CACHE_UB_ROWS;
440
441        /* For the perfectly-aligned-for-UIF-XOR case, don't add any pad. */
442        if (height_offset_in_pc == 0)
443                return 0;
444
445        /* Try padding up to where we're offset by at least half a page. */
446        if (height_offset_in_pc < PAGE_UB_ROWS_TIMES_1_5) {
447                /* If we fit entirely in the page cache, don't pad. */
448                if (height_ub < PAGE_CACHE_UB_ROWS)
449                        return 0;
450                else
451                        return PAGE_UB_ROWS_TIMES_1_5 - height_offset_in_pc;
452        }
453
454        /* If we're close to being aligned to page cache size, then round up
455         * and rely on XOR.
456         */
457        if (height_offset_in_pc > PAGE_CACHE_MINUS_1_5_UB_ROWS)
458                return PAGE_CACHE_UB_ROWS - height_offset_in_pc;
459
460        /* Otherwise, we're far enough away (top and bottom) to not need any
461         * padding.
462         */
463        return 0;
464}
465
466static void
467v3d_setup_slices(struct v3d_resource *rsc, uint32_t winsys_stride,
468                 bool uif_top)
469{
470        struct pipe_resource *prsc = &rsc->base;
471        uint32_t width = prsc->width0;
472        uint32_t height = prsc->height0;
473        uint32_t depth = prsc->depth0;
474        /* Note that power-of-two padding is based on level 1.  These are not
475         * equivalent to just util_next_power_of_two(dimension), because at a
476         * level 0 dimension of 9, the level 1 power-of-two padded value is 4,
477         * not 8.
478         */
479        uint32_t pot_width = 2 * util_next_power_of_two(u_minify(width, 1));
480        uint32_t pot_height = 2 * util_next_power_of_two(u_minify(height, 1));
481        uint32_t pot_depth = 2 * util_next_power_of_two(u_minify(depth, 1));
482        uint32_t offset = 0;
483        uint32_t utile_w = v3d_utile_width(rsc->cpp);
484        uint32_t utile_h = v3d_utile_height(rsc->cpp);
485        uint32_t uif_block_w = utile_w * 2;
486        uint32_t uif_block_h = utile_h * 2;
487        uint32_t block_width = util_format_get_blockwidth(prsc->format);
488        uint32_t block_height = util_format_get_blockheight(prsc->format);
489        bool msaa = prsc->nr_samples > 1;
490
491        /* MSAA textures/renderbuffers are always laid out as single-level
492         * UIF.
493         */
494        uif_top |= msaa;
495
496        /* Check some easy mistakes to make in a resource_create() call that
497         * will break our setup.
498         */
499        assert(prsc->array_size != 0);
500        assert(prsc->depth0 != 0);
501
502        for (int i = prsc->last_level; i >= 0; i--) {
503                struct v3d_resource_slice *slice = &rsc->slices[i];
504
505                uint32_t level_width, level_height, level_depth;
506                if (i < 2) {
507                        level_width = u_minify(width, i);
508                        level_height = u_minify(height, i);
509                } else {
510                        level_width = u_minify(pot_width, i);
511                        level_height = u_minify(pot_height, i);
512                }
513                if (i < 1)
514                        level_depth = u_minify(depth, i);
515                else
516                        level_depth = u_minify(pot_depth, i);
517
518                if (msaa) {
519                        level_width *= 2;
520                        level_height *= 2;
521                }
522
523                level_width = DIV_ROUND_UP(level_width, block_width);
524                level_height = DIV_ROUND_UP(level_height, block_height);
525
526                if (!rsc->tiled) {
527                        slice->tiling = VC5_TILING_RASTER;
528                        if (prsc->target == PIPE_TEXTURE_1D)
529                                level_width = align(level_width, 64 / rsc->cpp);
530                } else {
531                        if ((i != 0 || !uif_top) &&
532                            (level_width <= utile_w ||
533                             level_height <= utile_h)) {
534                                slice->tiling = VC5_TILING_LINEARTILE;
535                                level_width = align(level_width, utile_w);
536                                level_height = align(level_height, utile_h);
537                        } else if ((i != 0 || !uif_top) &&
538                                   level_width <= uif_block_w) {
539                                slice->tiling = VC5_TILING_UBLINEAR_1_COLUMN;
540                                level_width = align(level_width, uif_block_w);
541                                level_height = align(level_height, uif_block_h);
542                        } else if ((i != 0 || !uif_top) &&
543                                   level_width <= 2 * uif_block_w) {
544                                slice->tiling = VC5_TILING_UBLINEAR_2_COLUMN;
545                                level_width = align(level_width, 2 * uif_block_w);
546                                level_height = align(level_height, uif_block_h);
547                        } else {
548                                /* We align the width to a 4-block column of
549                                 * UIF blocks, but we only align height to UIF
550                                 * blocks.
551                                 */
552                                level_width = align(level_width,
553                                                    4 * uif_block_w);
554                                level_height = align(level_height,
555                                                     uif_block_h);
556
557                                slice->ub_pad = v3d_get_ub_pad(rsc,
558                                                               level_height);
559                                level_height += slice->ub_pad * uif_block_h;
560
561                                /* If the padding set us to to be aligned to
562                                 * the page cache size, then the HW will use
563                                 * the XOR bit on odd columns to get us
564                                 * perfectly misaligned
565                                 */
566                                if ((level_height / uif_block_h) %
567                                    (VC5_PAGE_CACHE_SIZE /
568                                     VC5_UIFBLOCK_ROW_SIZE) == 0) {
569                                        slice->tiling = VC5_TILING_UIF_XOR;
570                                } else {
571                                        slice->tiling = VC5_TILING_UIF_NO_XOR;
572                                }
573                        }
574                }
575
576                slice->offset = offset;
577                if (winsys_stride)
578                        slice->stride = winsys_stride;
579                else
580                        slice->stride = level_width * rsc->cpp;
581                slice->padded_height = level_height;
582                slice->size = level_height * slice->stride;
583
584                uint32_t slice_total_size = slice->size * level_depth;
585
586                /* The HW aligns level 1's base to a page if any of level 1 or
587                 * below could be UIF XOR.  The lower levels then inherit the
588                 * alignment for as long as necesary, thanks to being power of
589                 * two aligned.
590                 */
591                if (i == 1 &&
592                    level_width > 4 * uif_block_w &&
593                    level_height > PAGE_CACHE_MINUS_1_5_UB_ROWS * uif_block_h) {
594                        slice_total_size = align(slice_total_size,
595                                                 VC5_UIFCFG_PAGE_SIZE);
596                }
597
598                offset += slice_total_size;
599
600        }
601        rsc->size = offset;
602
603        /* UIF/UBLINEAR levels need to be aligned to UIF-blocks, and LT only
604         * needs to be aligned to utile boundaries.  Since tiles are laid out
605         * from small to big in memory, we need to align the later UIF slices
606         * to UIF blocks, if they were preceded by non-UIF-block-aligned LT
607         * slices.
608         *
609         * We additionally align to 4k, which improves UIF XOR performance.
610         */
611        uint32_t page_align_offset = (align(rsc->slices[0].offset, 4096) -
612                                      rsc->slices[0].offset);
613        if (page_align_offset) {
614                rsc->size += page_align_offset;
615                for (int i = 0; i <= prsc->last_level; i++)
616                        rsc->slices[i].offset += page_align_offset;
617        }
618
619        /* Arrays and cube textures have a stride which is the distance from
620         * one full mipmap tree to the next (64b aligned).  For 3D textures,
621         * we need to program the stride between slices of miplevel 0.
622         */
623        if (prsc->target != PIPE_TEXTURE_3D) {
624                rsc->cube_map_stride = align(rsc->slices[0].offset +
625                                             rsc->slices[0].size, 64);
626                rsc->size += rsc->cube_map_stride * (prsc->array_size - 1);
627        } else {
628                rsc->cube_map_stride = rsc->slices[0].size;
629        }
630}
631
632uint32_t
633v3d_layer_offset(struct pipe_resource *prsc, uint32_t level, uint32_t layer)
634{
635        struct v3d_resource *rsc = v3d_resource(prsc);
636        struct v3d_resource_slice *slice = &rsc->slices[level];
637
638        if (prsc->target == PIPE_TEXTURE_3D)
639                return slice->offset + layer * slice->size;
640        else
641                return slice->offset + layer * rsc->cube_map_stride;
642}
643
644static struct v3d_resource *
645v3d_resource_setup(struct pipe_screen *pscreen,
646                   const struct pipe_resource *tmpl)
647{
648        struct v3d_screen *screen = v3d_screen(pscreen);
649        struct v3d_resource *rsc = CALLOC_STRUCT(v3d_resource);
650        if (!rsc)
651                return NULL;
652        struct pipe_resource *prsc = &rsc->base;
653
654        *prsc = *tmpl;
655
656        pipe_reference_init(&prsc->reference, 1);
657        prsc->screen = pscreen;
658
659        if (prsc->nr_samples <= 1 ||
660            screen->devinfo.ver >= 40 ||
661            util_format_is_depth_or_stencil(prsc->format)) {
662                rsc->cpp = util_format_get_blocksize(prsc->format);
663                if (screen->devinfo.ver < 40 && prsc->nr_samples > 1)
664                        rsc->cpp *= prsc->nr_samples;
665        } else {
666                assert(v3d_rt_format_supported(&screen->devinfo, prsc->format));
667                uint32_t output_image_format =
668                        v3d_get_rt_format(&screen->devinfo, prsc->format);
669                uint32_t internal_type;
670                uint32_t internal_bpp;
671                v3d_get_internal_type_bpp_for_output_format(&screen->devinfo,
672                                                            output_image_format,
673                                                            &internal_type,
674                                                            &internal_bpp);
675                switch (internal_bpp) {
676                case V3D_INTERNAL_BPP_32:
677                        rsc->cpp = 4;
678                        break;
679                case V3D_INTERNAL_BPP_64:
680                        rsc->cpp = 8;
681                        break;
682                case V3D_INTERNAL_BPP_128:
683                        rsc->cpp = 16;
684                        break;
685                }
686        }
687
688        assert(rsc->cpp);
689
690        return rsc;
691}
692
693static struct pipe_resource *
694v3d_resource_create_with_modifiers(struct pipe_screen *pscreen,
695                                   const struct pipe_resource *tmpl,
696                                   const uint64_t *modifiers,
697                                   int count)
698{
699        struct v3d_screen *screen = v3d_screen(pscreen);
700
701        bool linear_ok = drm_find_modifier(DRM_FORMAT_MOD_LINEAR, modifiers, count);
702        struct v3d_resource *rsc = v3d_resource_setup(pscreen, tmpl);
703        struct pipe_resource *prsc = &rsc->base;
704        /* Use a tiled layout if we can, for better 3D performance. */
705        bool should_tile = true;
706
707        /* VBOs/PBOs are untiled (and 1 height). */
708        if (tmpl->target == PIPE_BUFFER)
709                should_tile = false;
710
711        /* Cursors are always linear, and the user can request linear as well.
712         */
713        if (tmpl->bind & (PIPE_BIND_LINEAR | PIPE_BIND_CURSOR))
714                should_tile = false;
715
716        /* 1D and 1D_ARRAY textures are always raster-order. */
717        if (tmpl->target == PIPE_TEXTURE_1D ||
718            tmpl->target == PIPE_TEXTURE_1D_ARRAY)
719                should_tile = false;
720
721        /* Scanout BOs for simulator need to be linear for interaction with
722         * i965.
723         */
724        if (using_v3d_simulator &&
725            tmpl->bind & (PIPE_BIND_SHARED | PIPE_BIND_SCANOUT))
726                should_tile = false;
727
728        /* If using the old-school SCANOUT flag, we don't know what the screen
729         * might support other than linear. Just force linear.
730         */
731        if (tmpl->bind & PIPE_BIND_SCANOUT)
732                should_tile = false;
733
734        /* No user-specified modifier; determine our own. */
735        if (count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID) {
736                linear_ok = true;
737                rsc->tiled = should_tile;
738        } else if (should_tile &&
739                   drm_find_modifier(DRM_FORMAT_MOD_BROADCOM_UIF,
740                                 modifiers, count)) {
741                rsc->tiled = true;
742        } else if (linear_ok) {
743                rsc->tiled = false;
744        } else {
745                fprintf(stderr, "Unsupported modifier requested\n");
746                goto fail;
747        }
748
749        rsc->internal_format = prsc->format;
750
751        v3d_setup_slices(rsc, 0, tmpl->bind & PIPE_BIND_SHARED);
752
753        /* If we're in a renderonly setup, use the other device to perform our
754         * allocation and just import it to v3d.  The other device may be
755         * using CMA, and V3D can import from CMA but doesn't do CMA
756         * allocations on its own.
757         *
758         * We always allocate this way for SHARED, because get_handle will
759         * need a resource on the display fd.
760         */
761        if (screen->ro && (tmpl->bind & (PIPE_BIND_SCANOUT |
762                                         PIPE_BIND_SHARED))) {
763                struct winsys_handle handle;
764                struct pipe_resource scanout_tmpl = {
765                        .target = prsc->target,
766                        .format = PIPE_FORMAT_RGBA8888_UNORM,
767                        .width0 = 1024, /* one page */
768                        .height0 = align(rsc->size, 4096) / 4096,
769                        .depth0 = 1,
770                        .array_size = 1,
771                };
772
773                rsc->scanout =
774                        renderonly_scanout_for_resource(&scanout_tmpl,
775                                                        screen->ro,
776                                                        &handle);
777
778                if (!rsc->scanout) {
779                        fprintf(stderr, "Failed to create scanout resource\n");
780                        return NULL;
781                }
782                assert(handle.type == WINSYS_HANDLE_TYPE_FD);
783                rsc->bo = v3d_bo_open_dmabuf(screen, handle.handle);
784                close(handle.handle);
785
786                if (!rsc->bo)
787                        goto fail;
788
789                v3d_debug_resource_layout(rsc, "renderonly");
790
791                return prsc;
792        } else {
793                if (!v3d_resource_bo_alloc(rsc))
794                        goto fail;
795        }
796
797        return prsc;
798fail:
799        v3d_resource_destroy(pscreen, prsc);
800        return NULL;
801}
802
803struct pipe_resource *
804v3d_resource_create(struct pipe_screen *pscreen,
805                    const struct pipe_resource *tmpl)
806{
807        const uint64_t mod = DRM_FORMAT_MOD_INVALID;
808        return v3d_resource_create_with_modifiers(pscreen, tmpl, &mod, 1);
809}
810
811static struct pipe_resource *
812v3d_resource_from_handle(struct pipe_screen *pscreen,
813                         const struct pipe_resource *tmpl,
814                         struct winsys_handle *whandle,
815                         unsigned usage)
816{
817        struct v3d_screen *screen = v3d_screen(pscreen);
818        struct v3d_resource *rsc = v3d_resource_setup(pscreen, tmpl);
819        struct pipe_resource *prsc = &rsc->base;
820        struct v3d_resource_slice *slice = &rsc->slices[0];
821
822        if (!rsc)
823                return NULL;
824
825        switch (whandle->modifier) {
826        case DRM_FORMAT_MOD_LINEAR:
827                rsc->tiled = false;
828                break;
829        case DRM_FORMAT_MOD_BROADCOM_UIF:
830                rsc->tiled = true;
831                break;
832        case DRM_FORMAT_MOD_INVALID:
833                rsc->tiled = screen->ro == NULL;
834                break;
835        default:
836                fprintf(stderr,
837                        "Attempt to import unsupported modifier 0x%llx\n",
838                        (long long)whandle->modifier);
839                goto fail;
840        }
841
842        if (whandle->offset != 0) {
843                fprintf(stderr,
844                        "Attempt to import unsupported winsys offset %u\n",
845                        whandle->offset);
846                goto fail;
847        }
848
849        switch (whandle->type) {
850        case WINSYS_HANDLE_TYPE_SHARED:
851                rsc->bo = v3d_bo_open_name(screen, whandle->handle);
852                break;
853        case WINSYS_HANDLE_TYPE_FD:
854                rsc->bo = v3d_bo_open_dmabuf(screen, whandle->handle);
855                break;
856        default:
857                fprintf(stderr,
858                        "Attempt to import unsupported handle type %d\n",
859                        whandle->type);
860                goto fail;
861        }
862
863        if (!rsc->bo)
864                goto fail;
865
866        rsc->internal_format = prsc->format;
867
868        v3d_setup_slices(rsc, whandle->stride, true);
869        v3d_debug_resource_layout(rsc, "import");
870
871        if (screen->ro) {
872                /* Make sure that renderonly has a handle to our buffer in the
873                 * display's fd, so that a later renderonly_get_handle()
874                 * returns correct handles or GEM names.
875                 */
876                rsc->scanout =
877                        renderonly_create_gpu_import_for_resource(prsc,
878                                                                  screen->ro,
879                                                                  NULL);
880                if (!rsc->scanout) {
881                        fprintf(stderr, "Failed to create scanout resource.\n");
882                        goto fail;
883                }
884        }
885
886        if (whandle->stride != slice->stride) {
887                static bool warned = false;
888                if (!warned) {
889                        warned = true;
890                        fprintf(stderr,
891                                "Attempting to import %dx%d %s with "
892                                "unsupported stride %d instead of %d\n",
893                                prsc->width0, prsc->height0,
894                                util_format_short_name(prsc->format),
895                                whandle->stride,
896                                slice->stride);
897                }
898                goto fail;
899        }
900
901        return prsc;
902
903fail:
904        v3d_resource_destroy(pscreen, prsc);
905        return NULL;
906}
907
908void
909v3d_update_shadow_texture(struct pipe_context *pctx,
910                          struct pipe_sampler_view *pview)
911{
912        struct v3d_context *v3d = v3d_context(pctx);
913        struct v3d_sampler_view *view = v3d_sampler_view(pview);
914        struct v3d_resource *shadow = v3d_resource(view->texture);
915        struct v3d_resource *orig = v3d_resource(pview->texture);
916
917        assert(view->texture != pview->texture);
918
919        if (shadow->writes == orig->writes && orig->bo->private)
920                return;
921
922        perf_debug("Updating %dx%d@%d shadow for linear texture\n",
923                   orig->base.width0, orig->base.height0,
924                   pview->u.tex.first_level);
925
926        for (int i = 0; i <= shadow->base.last_level; i++) {
927                unsigned width = u_minify(shadow->base.width0, i);
928                unsigned height = u_minify(shadow->base.height0, i);
929                struct pipe_blit_info info = {
930                        .dst = {
931                                .resource = &shadow->base,
932                                .level = i,
933                                .box = {
934                                        .x = 0,
935                                        .y = 0,
936                                        .z = 0,
937                                        .width = width,
938                                        .height = height,
939                                        .depth = 1,
940                                },
941                                .format = shadow->base.format,
942                        },
943                        .src = {
944                                .resource = &orig->base,
945                                .level = pview->u.tex.first_level + i,
946                                .box = {
947                                        .x = 0,
948                                        .y = 0,
949                                        .z = 0,
950                                        .width = width,
951                                        .height = height,
952                                        .depth = 1,
953                                },
954                                .format = orig->base.format,
955                        },
956                        .mask = util_format_get_mask(orig->base.format),
957                };
958                pctx->blit(pctx, &info);
959        }
960
961        shadow->writes = orig->writes;
962}
963
964static struct pipe_surface *
965v3d_create_surface(struct pipe_context *pctx,
966                   struct pipe_resource *ptex,
967                   const struct pipe_surface *surf_tmpl)
968{
969        struct v3d_context *v3d = v3d_context(pctx);
970        struct v3d_screen *screen = v3d->screen;
971        struct v3d_surface *surface = CALLOC_STRUCT(v3d_surface);
972        struct v3d_resource *rsc = v3d_resource(ptex);
973
974        if (!surface)
975                return NULL;
976
977        assert(surf_tmpl->u.tex.first_layer == surf_tmpl->u.tex.last_layer);
978
979        struct pipe_surface *psurf = &surface->base;
980        unsigned level = surf_tmpl->u.tex.level;
981        struct v3d_resource_slice *slice = &rsc->slices[level];
982
983        pipe_reference_init(&psurf->reference, 1);
984        pipe_resource_reference(&psurf->texture, ptex);
985
986        psurf->context = pctx;
987        psurf->format = surf_tmpl->format;
988        psurf->width = u_minify(ptex->width0, level);
989        psurf->height = u_minify(ptex->height0, level);
990        psurf->u.tex.level = level;
991        psurf->u.tex.first_layer = surf_tmpl->u.tex.first_layer;
992        psurf->u.tex.last_layer = surf_tmpl->u.tex.last_layer;
993
994        surface->offset = v3d_layer_offset(ptex, level,
995                                           psurf->u.tex.first_layer);
996        surface->tiling = slice->tiling;
997
998        surface->format = v3d_get_rt_format(&screen->devinfo, psurf->format);
999
1000        const struct util_format_description *desc =
1001                util_format_description(psurf->format);
1002
1003        surface->swap_rb = (desc->swizzle[0] == PIPE_SWIZZLE_Z &&
1004                            psurf->format != PIPE_FORMAT_B5G6R5_UNORM);
1005
1006        if (util_format_is_depth_or_stencil(psurf->format)) {
1007                switch (psurf->format) {
1008                case PIPE_FORMAT_Z16_UNORM:
1009                        surface->internal_type = V3D_INTERNAL_TYPE_DEPTH_16;
1010                        break;
1011                case PIPE_FORMAT_Z32_FLOAT:
1012                case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
1013                        surface->internal_type = V3D_INTERNAL_TYPE_DEPTH_32F;
1014                        break;
1015                default:
1016                        surface->internal_type = V3D_INTERNAL_TYPE_DEPTH_24;
1017                }
1018        } else {
1019                uint32_t bpp, type;
1020                v3d_get_internal_type_bpp_for_output_format(&screen->devinfo,
1021                                                            surface->format,
1022                                                            &type, &bpp);
1023                surface->internal_type = type;
1024                surface->internal_bpp = bpp;
1025        }
1026
1027        if (surface->tiling == VC5_TILING_UIF_NO_XOR ||
1028            surface->tiling == VC5_TILING_UIF_XOR) {
1029                surface->padded_height_of_output_image_in_uif_blocks =
1030                        (slice->padded_height /
1031                         (2 * v3d_utile_height(rsc->cpp)));
1032        }
1033
1034        if (rsc->separate_stencil) {
1035                surface->separate_stencil =
1036                        v3d_create_surface(pctx, &rsc->separate_stencil->base,
1037                                           surf_tmpl);
1038        }
1039
1040        return &surface->base;
1041}
1042
1043static void
1044v3d_surface_destroy(struct pipe_context *pctx, struct pipe_surface *psurf)
1045{
1046        struct v3d_surface *surf = v3d_surface(psurf);
1047
1048        if (surf->separate_stencil)
1049                pipe_surface_reference(&surf->separate_stencil, NULL);
1050
1051        pipe_resource_reference(&psurf->texture, NULL);
1052        FREE(psurf);
1053}
1054
1055static void
1056v3d_flush_resource(struct pipe_context *pctx, struct pipe_resource *resource)
1057{
1058        /* All calls to flush_resource are followed by a flush of the context,
1059         * so there's nothing to do.
1060         */
1061}
1062
1063static enum pipe_format
1064v3d_resource_get_internal_format(struct pipe_resource *prsc)
1065{
1066        return v3d_resource(prsc)->internal_format;
1067}
1068
1069static void
1070v3d_resource_set_stencil(struct pipe_resource *prsc,
1071                         struct pipe_resource *stencil)
1072{
1073        v3d_resource(prsc)->separate_stencil = v3d_resource(stencil);
1074}
1075
1076static struct pipe_resource *
1077v3d_resource_get_stencil(struct pipe_resource *prsc)
1078{
1079        struct v3d_resource *rsc = v3d_resource(prsc);
1080
1081        return &rsc->separate_stencil->base;
1082}
1083
1084static const struct u_transfer_vtbl transfer_vtbl = {
1085        .resource_create          = v3d_resource_create,
1086        .resource_destroy         = v3d_resource_destroy,
1087        .transfer_map             = v3d_resource_transfer_map,
1088        .transfer_unmap           = v3d_resource_transfer_unmap,
1089        .transfer_flush_region    = u_default_transfer_flush_region,
1090        .get_internal_format      = v3d_resource_get_internal_format,
1091        .set_stencil              = v3d_resource_set_stencil,
1092        .get_stencil              = v3d_resource_get_stencil,
1093};
1094
1095void
1096v3d_resource_screen_init(struct pipe_screen *pscreen)
1097{
1098        pscreen->resource_create_with_modifiers =
1099                v3d_resource_create_with_modifiers;
1100        pscreen->resource_create = u_transfer_helper_resource_create;
1101        pscreen->resource_from_handle = v3d_resource_from_handle;
1102        pscreen->resource_get_handle = v3d_resource_get_handle;
1103        pscreen->resource_destroy = u_transfer_helper_resource_destroy;
1104        pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl,
1105                                                            true, false,
1106                                                            true, true);
1107}
1108
1109void
1110v3d_resource_context_init(struct pipe_context *pctx)
1111{
1112        pctx->transfer_map = u_transfer_helper_transfer_map;
1113        pctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
1114        pctx->transfer_unmap = u_transfer_helper_transfer_unmap;
1115        pctx->buffer_subdata = u_default_buffer_subdata;
1116        pctx->texture_subdata = v3d_texture_subdata;
1117        pctx->create_surface = v3d_create_surface;
1118        pctx->surface_destroy = v3d_surface_destroy;
1119        pctx->resource_copy_region = util_resource_copy_region;
1120        pctx->blit = v3d_blit;
1121        pctx->generate_mipmap = v3d_generate_mipmap;
1122        pctx->flush_resource = v3d_flush_resource;
1123}
1124