v3d_resource.c revision 01e04c3f
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
35#include "drm_fourcc.h"
36#include "v3d_screen.h"
37#include "v3d_context.h"
38#include "v3d_resource.h"
39#include "v3d_tiling.h"
40#include "broadcom/cle/v3d_packet_v33_pack.h"
41
42static void
43v3d_debug_resource_layout(struct v3d_resource *rsc, const char *caller)
44{
45        if (!(V3D_DEBUG & V3D_DEBUG_SURFACE))
46                return;
47
48        struct pipe_resource *prsc = &rsc->base;
49
50        if (prsc->target == PIPE_BUFFER) {
51                fprintf(stderr,
52                        "rsc %s %p (format %s), %dx%d buffer @0x%08x-0x%08x\n",
53                        caller, rsc,
54                        util_format_short_name(prsc->format),
55                        prsc->width0, prsc->height0,
56                        rsc->bo->offset,
57                        rsc->bo->offset + rsc->bo->size - 1);
58                return;
59        }
60
61        static const char *const tiling_descriptions[] = {
62                [VC5_TILING_RASTER] = "R",
63                [VC5_TILING_LINEARTILE] = "LT",
64                [VC5_TILING_UBLINEAR_1_COLUMN] = "UB1",
65                [VC5_TILING_UBLINEAR_2_COLUMN] = "UB2",
66                [VC5_TILING_UIF_NO_XOR] = "UIF",
67                [VC5_TILING_UIF_XOR] = "UIF^",
68        };
69
70        for (int i = 0; i <= prsc->last_level; i++) {
71                struct v3d_resource_slice *slice = &rsc->slices[i];
72
73                int level_width = slice->stride / rsc->cpp;
74                int level_height = slice->padded_height;
75                int level_depth =
76                        u_minify(util_next_power_of_two(prsc->depth0), i);
77
78                fprintf(stderr,
79                        "rsc %s %p (format %s), %dx%d: "
80                        "level %d (%s) %dx%dx%d -> %dx%dx%d, stride %d@0x%08x\n",
81                        caller, rsc,
82                        util_format_short_name(prsc->format),
83                        prsc->width0, prsc->height0,
84                        i, tiling_descriptions[slice->tiling],
85                        u_minify(prsc->width0, i),
86                        u_minify(prsc->height0, i),
87                        u_minify(prsc->depth0, i),
88                        level_width,
89                        level_height,
90                        level_depth,
91                        slice->stride,
92                        rsc->bo->offset + slice->offset);
93        }
94}
95
96static bool
97v3d_resource_bo_alloc(struct v3d_resource *rsc)
98{
99        struct pipe_resource *prsc = &rsc->base;
100        struct pipe_screen *pscreen = prsc->screen;
101        struct v3d_bo *bo;
102
103        bo = v3d_bo_alloc(v3d_screen(pscreen), rsc->size, "resource");
104        if (bo) {
105                v3d_bo_unreference(&rsc->bo);
106                rsc->bo = bo;
107                v3d_debug_resource_layout(rsc, "alloc");
108                return true;
109        } else {
110                return false;
111        }
112}
113
114static void
115v3d_resource_transfer_unmap(struct pipe_context *pctx,
116                            struct pipe_transfer *ptrans)
117{
118        struct v3d_context *v3d = v3d_context(pctx);
119        struct v3d_transfer *trans = v3d_transfer(ptrans);
120
121        if (trans->map) {
122                struct v3d_resource *rsc = v3d_resource(ptrans->resource);
123                struct v3d_resource_slice *slice = &rsc->slices[ptrans->level];
124
125                if (ptrans->usage & PIPE_TRANSFER_WRITE) {
126                        for (int z = 0; z < ptrans->box.depth; z++) {
127                                void *dst = rsc->bo->map +
128                                        v3d_layer_offset(&rsc->base,
129                                                         ptrans->level,
130                                                         ptrans->box.z + z);
131                                v3d_store_tiled_image(dst,
132                                                      slice->stride,
133                                                      (trans->map +
134                                                       ptrans->stride *
135                                                       ptrans->box.height * z),
136                                                      ptrans->stride,
137                                                      slice->tiling, rsc->cpp,
138                                                      slice->padded_height,
139                                                      &ptrans->box);
140                        }
141                }
142                free(trans->map);
143        }
144
145        pipe_resource_reference(&ptrans->resource, NULL);
146        slab_free(&v3d->transfer_pool, ptrans);
147}
148
149static void *
150v3d_resource_transfer_map(struct pipe_context *pctx,
151                          struct pipe_resource *prsc,
152                          unsigned level, unsigned usage,
153                          const struct pipe_box *box,
154                          struct pipe_transfer **pptrans)
155{
156        struct v3d_context *v3d = v3d_context(pctx);
157        struct v3d_resource *rsc = v3d_resource(prsc);
158        struct v3d_transfer *trans;
159        struct pipe_transfer *ptrans;
160        enum pipe_format format = prsc->format;
161        char *buf;
162
163        /* MSAA maps should have been handled by u_transfer_helper. */
164        assert(prsc->nr_samples <= 1);
165
166        /* Upgrade DISCARD_RANGE to WHOLE_RESOURCE if the whole resource is
167         * being mapped.
168         */
169        if ((usage & PIPE_TRANSFER_DISCARD_RANGE) &&
170            !(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
171            !(prsc->flags & PIPE_RESOURCE_FLAG_MAP_PERSISTENT) &&
172            prsc->last_level == 0 &&
173            prsc->width0 == box->width &&
174            prsc->height0 == box->height &&
175            prsc->depth0 == box->depth &&
176            prsc->array_size == 1 &&
177            rsc->bo->private) {
178                usage |= PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE;
179        }
180
181        if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
182                if (v3d_resource_bo_alloc(rsc)) {
183                        /* If it might be bound as one of our vertex buffers
184                         * or UBOs, make sure we re-emit vertex buffer state
185                         * or uniforms.
186                         */
187                        if (prsc->bind & PIPE_BIND_VERTEX_BUFFER)
188                                v3d->dirty |= VC5_DIRTY_VTXBUF;
189                        if (prsc->bind & PIPE_BIND_CONSTANT_BUFFER)
190                                v3d->dirty |= VC5_DIRTY_CONSTBUF;
191                } else {
192                        /* If we failed to reallocate, flush users so that we
193                         * don't violate any syncing requirements.
194                         */
195                        v3d_flush_jobs_reading_resource(v3d, prsc);
196                }
197        } else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
198                /* If we're writing and the buffer is being used by the CL, we
199                 * have to flush the CL first.  If we're only reading, we need
200                 * to flush if the CL has written our buffer.
201                 */
202                if (usage & PIPE_TRANSFER_WRITE)
203                        v3d_flush_jobs_reading_resource(v3d, prsc);
204                else
205                        v3d_flush_jobs_writing_resource(v3d, prsc);
206        }
207
208        if (usage & PIPE_TRANSFER_WRITE) {
209                rsc->writes++;
210                rsc->initialized_buffers = ~0;
211        }
212
213        trans = slab_alloc(&v3d->transfer_pool);
214        if (!trans)
215                return NULL;
216
217        /* XXX: Handle DONTBLOCK, DISCARD_RANGE, PERSISTENT, COHERENT. */
218
219        /* slab_alloc_st() doesn't zero: */
220        memset(trans, 0, sizeof(*trans));
221        ptrans = &trans->base;
222
223        pipe_resource_reference(&ptrans->resource, prsc);
224        ptrans->level = level;
225        ptrans->usage = usage;
226        ptrans->box = *box;
227
228        /* Note that the current kernel implementation is synchronous, so no
229         * need to do syncing stuff here yet.
230         */
231
232        if (usage & PIPE_TRANSFER_UNSYNCHRONIZED)
233                buf = v3d_bo_map_unsynchronized(rsc->bo);
234        else
235                buf = v3d_bo_map(rsc->bo);
236        if (!buf) {
237                fprintf(stderr, "Failed to map bo\n");
238                goto fail;
239        }
240
241        *pptrans = ptrans;
242
243        /* Our load/store routines work on entire compressed blocks. */
244        ptrans->box.x /= util_format_get_blockwidth(format);
245        ptrans->box.y /= util_format_get_blockheight(format);
246        ptrans->box.width = DIV_ROUND_UP(ptrans->box.width,
247                                         util_format_get_blockwidth(format));
248        ptrans->box.height = DIV_ROUND_UP(ptrans->box.height,
249                                          util_format_get_blockheight(format));
250
251        struct v3d_resource_slice *slice = &rsc->slices[level];
252        if (rsc->tiled) {
253                /* No direct mappings of tiled, since we need to manually
254                 * tile/untile.
255                 */
256                if (usage & PIPE_TRANSFER_MAP_DIRECTLY)
257                        return NULL;
258
259                ptrans->stride = ptrans->box.width * rsc->cpp;
260                ptrans->layer_stride = ptrans->stride * ptrans->box.height;
261
262                trans->map = malloc(ptrans->layer_stride * ptrans->box.depth);
263
264                if (usage & PIPE_TRANSFER_READ) {
265                        for (int z = 0; z < ptrans->box.depth; z++) {
266                                void *src = rsc->bo->map +
267                                        v3d_layer_offset(&rsc->base,
268                                                         ptrans->level,
269                                                         ptrans->box.z + z);
270                                v3d_load_tiled_image((trans->map +
271                                                      ptrans->stride *
272                                                      ptrans->box.height * z),
273                                                     ptrans->stride,
274                                                     src,
275                                                     slice->stride,
276                                                     slice->tiling, rsc->cpp,
277                                                     slice->padded_height,
278                                             &ptrans->box);
279                        }
280                }
281                return trans->map;
282        } else {
283                ptrans->stride = slice->stride;
284                ptrans->layer_stride = rsc->cube_map_stride;
285
286                return buf + slice->offset +
287                        ptrans->box.y * ptrans->stride +
288                        ptrans->box.x * rsc->cpp +
289                        ptrans->box.z * rsc->cube_map_stride;
290        }
291
292
293fail:
294        v3d_resource_transfer_unmap(pctx, ptrans);
295        return NULL;
296}
297
298static void
299v3d_resource_destroy(struct pipe_screen *pscreen,
300                     struct pipe_resource *prsc)
301{
302        struct v3d_resource *rsc = v3d_resource(prsc);
303
304        v3d_bo_unreference(&rsc->bo);
305        free(rsc);
306}
307
308static boolean
309v3d_resource_get_handle(struct pipe_screen *pscreen,
310                        struct pipe_context *pctx,
311                        struct pipe_resource *prsc,
312                        struct winsys_handle *whandle,
313                        unsigned usage)
314{
315        struct v3d_resource *rsc = v3d_resource(prsc);
316        struct v3d_bo *bo = rsc->bo;
317
318        whandle->stride = rsc->slices[0].stride;
319        whandle->offset = 0;
320
321        /* If we're passing some reference to our BO out to some other part of
322         * the system, then we can't do any optimizations about only us being
323         * the ones seeing it (like BO caching).
324         */
325        bo->private = false;
326
327        if (rsc->tiled) {
328                /* A shared tiled buffer should always be allocated as UIF,
329                 * not UBLINEAR or LT.
330                 */
331                assert(rsc->slices[0].tiling == VC5_TILING_UIF_XOR ||
332                       rsc->slices[0].tiling == VC5_TILING_UIF_NO_XOR);
333                whandle->modifier = DRM_FORMAT_MOD_BROADCOM_UIF;
334        } else {
335                whandle->modifier = DRM_FORMAT_MOD_LINEAR;
336        }
337
338        switch (whandle->type) {
339        case WINSYS_HANDLE_TYPE_SHARED:
340                return v3d_bo_flink(bo, &whandle->handle);
341        case WINSYS_HANDLE_TYPE_KMS:
342                whandle->handle = bo->handle;
343                return TRUE;
344        case WINSYS_HANDLE_TYPE_FD:
345                whandle->handle = v3d_bo_get_dmabuf(bo);
346                return whandle->handle != -1;
347        }
348
349        return FALSE;
350}
351
352#define PAGE_UB_ROWS (VC5_UIFCFG_PAGE_SIZE / VC5_UIFBLOCK_ROW_SIZE)
353#define PAGE_UB_ROWS_TIMES_1_5 ((PAGE_UB_ROWS * 3) >> 1)
354#define PAGE_CACHE_UB_ROWS (VC5_PAGE_CACHE_SIZE / VC5_UIFBLOCK_ROW_SIZE)
355#define PAGE_CACHE_MINUS_1_5_UB_ROWS (PAGE_CACHE_UB_ROWS - PAGE_UB_ROWS_TIMES_1_5)
356
357/**
358 * Computes the HW's UIFblock padding for a given height/cpp.
359 *
360 * The goal of the padding is to keep pages of the same color (bank number) at
361 * least half a page away from each other vertically when crossing between
362 * between columns of UIF blocks.
363 */
364static uint32_t
365v3d_get_ub_pad(struct v3d_resource *rsc, uint32_t height)
366{
367        uint32_t utile_h = v3d_utile_height(rsc->cpp);
368        uint32_t uif_block_h = utile_h * 2;
369        uint32_t height_ub = height / uif_block_h;
370
371        uint32_t height_offset_in_pc = height_ub % PAGE_CACHE_UB_ROWS;
372
373        /* For the perfectly-aligned-for-UIF-XOR case, don't add any pad. */
374        if (height_offset_in_pc == 0)
375                return 0;
376
377        /* Try padding up to where we're offset by at least half a page. */
378        if (height_offset_in_pc < PAGE_UB_ROWS_TIMES_1_5) {
379                /* If we fit entirely in the page cache, don't pad. */
380                if (height_ub < PAGE_CACHE_UB_ROWS)
381                        return 0;
382                else
383                        return PAGE_UB_ROWS_TIMES_1_5 - height_offset_in_pc;
384        }
385
386        /* If we're close to being aligned to page cache size, then round up
387         * and rely on XOR.
388         */
389        if (height_offset_in_pc > PAGE_CACHE_MINUS_1_5_UB_ROWS)
390                return PAGE_CACHE_UB_ROWS - height_offset_in_pc;
391
392        /* Otherwise, we're far enough away (top and bottom) to not need any
393         * padding.
394         */
395        return 0;
396}
397
398static void
399v3d_setup_slices(struct v3d_resource *rsc)
400{
401        struct pipe_resource *prsc = &rsc->base;
402        uint32_t width = prsc->width0;
403        uint32_t height = prsc->height0;
404        uint32_t depth = prsc->depth0;
405        /* Note that power-of-two padding is based on level 1.  These are not
406         * equivalent to just util_next_power_of_two(dimension), because at a
407         * level 0 dimension of 9, the level 1 power-of-two padded value is 4,
408         * not 8.
409         */
410        uint32_t pot_width = 2 * util_next_power_of_two(u_minify(width, 1));
411        uint32_t pot_height = 2 * util_next_power_of_two(u_minify(height, 1));
412        uint32_t pot_depth = 2 * util_next_power_of_two(u_minify(depth, 1));
413        uint32_t offset = 0;
414        uint32_t utile_w = v3d_utile_width(rsc->cpp);
415        uint32_t utile_h = v3d_utile_height(rsc->cpp);
416        uint32_t uif_block_w = utile_w * 2;
417        uint32_t uif_block_h = utile_h * 2;
418        uint32_t block_width = util_format_get_blockwidth(prsc->format);
419        uint32_t block_height = util_format_get_blockheight(prsc->format);
420        bool msaa = prsc->nr_samples > 1;
421        /* MSAA textures/renderbuffers are always laid out as single-level
422         * UIF.
423         */
424        bool uif_top = msaa;
425
426        for (int i = prsc->last_level; i >= 0; i--) {
427                struct v3d_resource_slice *slice = &rsc->slices[i];
428
429                uint32_t level_width, level_height, level_depth;
430                if (i < 2) {
431                        level_width = u_minify(width, i);
432                        level_height = u_minify(height, i);
433                } else {
434                        level_width = u_minify(pot_width, i);
435                        level_height = u_minify(pot_height, i);
436                }
437                if (i < 1)
438                        level_depth = u_minify(depth, i);
439                else
440                        level_depth = u_minify(pot_depth, i);
441
442                if (msaa) {
443                        level_width *= 2;
444                        level_height *= 2;
445                }
446
447                level_width = DIV_ROUND_UP(level_width, block_width);
448                level_height = DIV_ROUND_UP(level_height, block_height);
449
450                if (!rsc->tiled) {
451                        slice->tiling = VC5_TILING_RASTER;
452                        if (prsc->target == PIPE_TEXTURE_1D)
453                                level_width = align(level_width, 64 / rsc->cpp);
454                } else {
455                        if ((i != 0 || !uif_top) &&
456                            (level_width <= utile_w ||
457                             level_height <= utile_h)) {
458                                slice->tiling = VC5_TILING_LINEARTILE;
459                                level_width = align(level_width, utile_w);
460                                level_height = align(level_height, utile_h);
461                        } else if ((i != 0 || !uif_top) &&
462                                   level_width <= uif_block_w) {
463                                slice->tiling = VC5_TILING_UBLINEAR_1_COLUMN;
464                                level_width = align(level_width, uif_block_w);
465                                level_height = align(level_height, uif_block_h);
466                        } else if ((i != 0 || !uif_top) &&
467                                   level_width <= 2 * uif_block_w) {
468                                slice->tiling = VC5_TILING_UBLINEAR_2_COLUMN;
469                                level_width = align(level_width, 2 * uif_block_w);
470                                level_height = align(level_height, uif_block_h);
471                        } else {
472                                /* We align the width to a 4-block column of
473                                 * UIF blocks, but we only align height to UIF
474                                 * blocks.
475                                 */
476                                level_width = align(level_width,
477                                                    4 * uif_block_w);
478                                level_height = align(level_height,
479                                                     uif_block_h);
480
481                                slice->ub_pad = v3d_get_ub_pad(rsc,
482                                                               level_height);
483                                level_height += slice->ub_pad * uif_block_h;
484
485                                /* If the padding set us to to be aligned to
486                                 * the page cache size, then the HW will use
487                                 * the XOR bit on odd columns to get us
488                                 * perfectly misaligned
489                                 */
490                                if ((level_height / uif_block_h) %
491                                    (VC5_PAGE_CACHE_SIZE /
492                                     VC5_UIFBLOCK_ROW_SIZE) == 0) {
493                                        slice->tiling = VC5_TILING_UIF_XOR;
494                                } else {
495                                        slice->tiling = VC5_TILING_UIF_NO_XOR;
496                                }
497                        }
498                }
499
500                slice->offset = offset;
501                slice->stride = level_width * rsc->cpp;
502                slice->padded_height = level_height;
503                slice->size = level_height * slice->stride;
504
505                uint32_t slice_total_size = slice->size * level_depth;
506
507                /* The HW aligns level 1's base to a page if any of level 1 or
508                 * below could be UIF XOR.  The lower levels then inherit the
509                 * alignment for as long as necesary, thanks to being power of
510                 * two aligned.
511                 */
512                if (i == 1 &&
513                    level_width > 4 * uif_block_w &&
514                    level_height > PAGE_CACHE_MINUS_1_5_UB_ROWS * uif_block_h) {
515                        slice_total_size = align(slice_total_size,
516                                                 VC5_UIFCFG_PAGE_SIZE);
517                }
518
519                offset += slice_total_size;
520
521        }
522        rsc->size = offset;
523
524        /* UIF/UBLINEAR levels need to be aligned to UIF-blocks, and LT only
525         * needs to be aligned to utile boundaries.  Since tiles are laid out
526         * from small to big in memory, we need to align the later UIF slices
527         * to UIF blocks, if they were preceded by non-UIF-block-aligned LT
528         * slices.
529         *
530         * We additionally align to 4k, which improves UIF XOR performance.
531         */
532        uint32_t page_align_offset = (align(rsc->slices[0].offset, 4096) -
533                                      rsc->slices[0].offset);
534        if (page_align_offset) {
535                rsc->size += page_align_offset;
536                for (int i = 0; i <= prsc->last_level; i++)
537                        rsc->slices[i].offset += page_align_offset;
538        }
539
540        /* Arrays and cube textures have a stride which is the distance from
541         * one full mipmap tree to the next (64b aligned).  For 3D textures,
542         * we need to program the stride between slices of miplevel 0.
543         */
544        if (prsc->target != PIPE_TEXTURE_3D) {
545                rsc->cube_map_stride = align(rsc->slices[0].offset +
546                                             rsc->slices[0].size, 64);
547                rsc->size += rsc->cube_map_stride * (prsc->array_size - 1);
548        } else {
549                rsc->cube_map_stride = rsc->slices[0].size;
550        }
551}
552
553uint32_t
554v3d_layer_offset(struct pipe_resource *prsc, uint32_t level, uint32_t layer)
555{
556        struct v3d_resource *rsc = v3d_resource(prsc);
557        struct v3d_resource_slice *slice = &rsc->slices[level];
558
559        if (prsc->target == PIPE_TEXTURE_3D)
560                return slice->offset + layer * slice->size;
561        else
562                return slice->offset + layer * rsc->cube_map_stride;
563}
564
565static struct v3d_resource *
566v3d_resource_setup(struct pipe_screen *pscreen,
567                   const struct pipe_resource *tmpl)
568{
569        struct v3d_screen *screen = v3d_screen(pscreen);
570        struct v3d_resource *rsc = CALLOC_STRUCT(v3d_resource);
571        if (!rsc)
572                return NULL;
573        struct pipe_resource *prsc = &rsc->base;
574
575        *prsc = *tmpl;
576
577        pipe_reference_init(&prsc->reference, 1);
578        prsc->screen = pscreen;
579
580        if (prsc->nr_samples <= 1 ||
581            screen->devinfo.ver >= 40 ||
582            util_format_is_depth_or_stencil(prsc->format)) {
583                rsc->cpp = util_format_get_blocksize(prsc->format);
584                if (screen->devinfo.ver < 40 && prsc->nr_samples > 1)
585                        rsc->cpp *= prsc->nr_samples;
586        } else {
587                assert(v3d_rt_format_supported(&screen->devinfo, prsc->format));
588                uint32_t output_image_format =
589                        v3d_get_rt_format(&screen->devinfo, prsc->format);
590                uint32_t internal_type;
591                uint32_t internal_bpp;
592                v3d_get_internal_type_bpp_for_output_format(&screen->devinfo,
593                                                            output_image_format,
594                                                            &internal_type,
595                                                            &internal_bpp);
596                switch (internal_bpp) {
597                case V3D_INTERNAL_BPP_32:
598                        rsc->cpp = 4;
599                        break;
600                case V3D_INTERNAL_BPP_64:
601                        rsc->cpp = 8;
602                        break;
603                case V3D_INTERNAL_BPP_128:
604                        rsc->cpp = 16;
605                        break;
606                }
607        }
608
609        assert(rsc->cpp);
610
611        return rsc;
612}
613
614static bool
615find_modifier(uint64_t needle, const uint64_t *haystack, int count)
616{
617        int i;
618
619        for (i = 0; i < count; i++) {
620                if (haystack[i] == needle)
621                        return true;
622        }
623
624        return false;
625}
626
627static struct pipe_resource *
628v3d_resource_create_with_modifiers(struct pipe_screen *pscreen,
629                                   const struct pipe_resource *tmpl,
630                                   const uint64_t *modifiers,
631                                   int count)
632{
633        bool linear_ok = find_modifier(DRM_FORMAT_MOD_LINEAR, modifiers, count);
634        struct v3d_resource *rsc = v3d_resource_setup(pscreen, tmpl);
635        struct pipe_resource *prsc = &rsc->base;
636        /* Use a tiled layout if we can, for better 3D performance. */
637        bool should_tile = true;
638
639        /* VBOs/PBOs are untiled (and 1 height). */
640        if (tmpl->target == PIPE_BUFFER)
641                should_tile = false;
642
643        /* Cursors are always linear, and the user can request linear as well.
644         */
645        if (tmpl->bind & (PIPE_BIND_LINEAR | PIPE_BIND_CURSOR))
646                should_tile = false;
647
648        /* 1D and 1D_ARRAY textures are always raster-order. */
649        if (tmpl->target == PIPE_TEXTURE_1D ||
650            tmpl->target == PIPE_TEXTURE_1D_ARRAY)
651                should_tile = false;
652
653        /* Scanout BOs for simulator need to be linear for interaction with
654         * i965.
655         */
656        if (using_v3d_simulator &&
657            tmpl->bind & (PIPE_BIND_SHARED | PIPE_BIND_SCANOUT))
658                should_tile = false;
659
660        /* No user-specified modifier; determine our own. */
661        if (count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID) {
662                linear_ok = true;
663                rsc->tiled = should_tile;
664        } else if (should_tile &&
665                   find_modifier(DRM_FORMAT_MOD_BROADCOM_UIF,
666                                 modifiers, count)) {
667                rsc->tiled = true;
668        } else if (linear_ok) {
669                rsc->tiled = false;
670        } else {
671                fprintf(stderr, "Unsupported modifier requested\n");
672                goto fail;
673        }
674
675        rsc->internal_format = prsc->format;
676
677        v3d_setup_slices(rsc);
678        if (!v3d_resource_bo_alloc(rsc))
679                goto fail;
680
681        return prsc;
682fail:
683        v3d_resource_destroy(pscreen, prsc);
684        return NULL;
685}
686
687struct pipe_resource *
688v3d_resource_create(struct pipe_screen *pscreen,
689                    const struct pipe_resource *tmpl)
690{
691        const uint64_t mod = DRM_FORMAT_MOD_INVALID;
692        return v3d_resource_create_with_modifiers(pscreen, tmpl, &mod, 1);
693}
694
695static struct pipe_resource *
696v3d_resource_from_handle(struct pipe_screen *pscreen,
697                         const struct pipe_resource *tmpl,
698                         struct winsys_handle *whandle,
699                         unsigned usage)
700{
701        struct v3d_screen *screen = v3d_screen(pscreen);
702        struct v3d_resource *rsc = v3d_resource_setup(pscreen, tmpl);
703        struct pipe_resource *prsc = &rsc->base;
704        struct v3d_resource_slice *slice = &rsc->slices[0];
705
706        if (!rsc)
707                return NULL;
708
709        switch (whandle->modifier) {
710        case DRM_FORMAT_MOD_LINEAR:
711                rsc->tiled = false;
712                break;
713        case DRM_FORMAT_MOD_BROADCOM_UIF:
714        case DRM_FORMAT_MOD_INVALID:
715                rsc->tiled = true;
716                break;
717        default:
718                fprintf(stderr,
719                        "Attempt to import unsupported modifier 0x%llx\n",
720                        (long long)whandle->modifier);
721                goto fail;
722        }
723
724        if (whandle->offset != 0) {
725                fprintf(stderr,
726                        "Attempt to import unsupported winsys offset %u\n",
727                        whandle->offset);
728                goto fail;
729        }
730
731        switch (whandle->type) {
732        case WINSYS_HANDLE_TYPE_SHARED:
733                rsc->bo = v3d_bo_open_name(screen,
734                                           whandle->handle, whandle->stride);
735                break;
736        case WINSYS_HANDLE_TYPE_FD:
737                rsc->bo = v3d_bo_open_dmabuf(screen,
738                                             whandle->handle, whandle->stride);
739                break;
740        default:
741                fprintf(stderr,
742                        "Attempt to import unsupported handle type %d\n",
743                        whandle->type);
744                goto fail;
745        }
746
747        if (!rsc->bo)
748                goto fail;
749
750        rsc->internal_format = prsc->format;
751
752        v3d_setup_slices(rsc);
753        v3d_debug_resource_layout(rsc, "import");
754
755        if (whandle->stride != slice->stride) {
756                static bool warned = false;
757                if (!warned) {
758                        warned = true;
759                        fprintf(stderr,
760                                "Attempting to import %dx%d %s with "
761                                "unsupported stride %d instead of %d\n",
762                                prsc->width0, prsc->height0,
763                                util_format_short_name(prsc->format),
764                                whandle->stride,
765                                slice->stride);
766                }
767                goto fail;
768        }
769
770        return prsc;
771
772fail:
773        v3d_resource_destroy(pscreen, prsc);
774        return NULL;
775}
776
777static struct pipe_surface *
778v3d_create_surface(struct pipe_context *pctx,
779                   struct pipe_resource *ptex,
780                   const struct pipe_surface *surf_tmpl)
781{
782        struct v3d_context *v3d = v3d_context(pctx);
783        struct v3d_screen *screen = v3d->screen;
784        struct v3d_surface *surface = CALLOC_STRUCT(v3d_surface);
785        struct v3d_resource *rsc = v3d_resource(ptex);
786
787        if (!surface)
788                return NULL;
789
790        assert(surf_tmpl->u.tex.first_layer == surf_tmpl->u.tex.last_layer);
791
792        struct pipe_surface *psurf = &surface->base;
793        unsigned level = surf_tmpl->u.tex.level;
794        struct v3d_resource_slice *slice = &rsc->slices[level];
795
796        pipe_reference_init(&psurf->reference, 1);
797        pipe_resource_reference(&psurf->texture, ptex);
798
799        psurf->context = pctx;
800        psurf->format = surf_tmpl->format;
801        psurf->width = u_minify(ptex->width0, level);
802        psurf->height = u_minify(ptex->height0, level);
803        psurf->u.tex.level = level;
804        psurf->u.tex.first_layer = surf_tmpl->u.tex.first_layer;
805        psurf->u.tex.last_layer = surf_tmpl->u.tex.last_layer;
806
807        surface->offset = v3d_layer_offset(ptex, level,
808                                           psurf->u.tex.first_layer);
809        surface->tiling = slice->tiling;
810
811        surface->format = v3d_get_rt_format(&screen->devinfo, psurf->format);
812
813        if (util_format_is_depth_or_stencil(psurf->format)) {
814                switch (psurf->format) {
815                case PIPE_FORMAT_Z16_UNORM:
816                        surface->internal_type = V3D_INTERNAL_TYPE_DEPTH_16;
817                        break;
818                case PIPE_FORMAT_Z32_FLOAT:
819                case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
820                        surface->internal_type = V3D_INTERNAL_TYPE_DEPTH_32F;
821                        break;
822                default:
823                        surface->internal_type = V3D_INTERNAL_TYPE_DEPTH_24;
824                }
825        } else {
826                uint32_t bpp, type;
827                v3d_get_internal_type_bpp_for_output_format(&screen->devinfo,
828                                                            surface->format,
829                                                            &type, &bpp);
830                surface->internal_type = type;
831                surface->internal_bpp = bpp;
832        }
833
834        if (surface->tiling == VC5_TILING_UIF_NO_XOR ||
835            surface->tiling == VC5_TILING_UIF_XOR) {
836                surface->padded_height_of_output_image_in_uif_blocks =
837                        (slice->padded_height /
838                         (2 * v3d_utile_height(rsc->cpp)));
839        }
840
841        if (rsc->separate_stencil) {
842                surface->separate_stencil =
843                        v3d_create_surface(pctx, &rsc->separate_stencil->base,
844                                           surf_tmpl);
845        }
846
847        return &surface->base;
848}
849
850static void
851v3d_surface_destroy(struct pipe_context *pctx, struct pipe_surface *psurf)
852{
853        struct v3d_surface *surf = v3d_surface(psurf);
854
855        if (surf->separate_stencil)
856                pipe_surface_reference(&surf->separate_stencil, NULL);
857
858        pipe_resource_reference(&psurf->texture, NULL);
859        FREE(psurf);
860}
861
862static void
863v3d_flush_resource(struct pipe_context *pctx, struct pipe_resource *resource)
864{
865        /* All calls to flush_resource are followed by a flush of the context,
866         * so there's nothing to do.
867         */
868}
869
870static enum pipe_format
871v3d_resource_get_internal_format(struct pipe_resource *prsc)
872{
873        return v3d_resource(prsc)->internal_format;
874}
875
876static void
877v3d_resource_set_stencil(struct pipe_resource *prsc,
878                         struct pipe_resource *stencil)
879{
880        v3d_resource(prsc)->separate_stencil = v3d_resource(stencil);
881}
882
883static struct pipe_resource *
884v3d_resource_get_stencil(struct pipe_resource *prsc)
885{
886        struct v3d_resource *rsc = v3d_resource(prsc);
887
888        return &rsc->separate_stencil->base;
889}
890
891static const struct u_transfer_vtbl transfer_vtbl = {
892        .resource_create          = v3d_resource_create,
893        .resource_destroy         = v3d_resource_destroy,
894        .transfer_map             = v3d_resource_transfer_map,
895        .transfer_unmap           = v3d_resource_transfer_unmap,
896        .transfer_flush_region    = u_default_transfer_flush_region,
897        .get_internal_format      = v3d_resource_get_internal_format,
898        .set_stencil              = v3d_resource_set_stencil,
899        .get_stencil              = v3d_resource_get_stencil,
900};
901
902void
903v3d_resource_screen_init(struct pipe_screen *pscreen)
904{
905        pscreen->resource_create_with_modifiers =
906                v3d_resource_create_with_modifiers;
907        pscreen->resource_create = u_transfer_helper_resource_create;
908        pscreen->resource_from_handle = v3d_resource_from_handle;
909        pscreen->resource_get_handle = v3d_resource_get_handle;
910        pscreen->resource_destroy = u_transfer_helper_resource_destroy;
911        pscreen->transfer_helper = u_transfer_helper_create(&transfer_vtbl,
912                                                            true, false,
913                                                            true, true);
914}
915
916void
917v3d_resource_context_init(struct pipe_context *pctx)
918{
919        pctx->transfer_map = u_transfer_helper_transfer_map;
920        pctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
921        pctx->transfer_unmap = u_transfer_helper_transfer_unmap;
922        pctx->buffer_subdata = u_default_buffer_subdata;
923        pctx->texture_subdata = u_default_texture_subdata;
924        pctx->create_surface = v3d_create_surface;
925        pctx->surface_destroy = v3d_surface_destroy;
926        pctx->resource_copy_region = util_resource_copy_region;
927        pctx->blit = v3d_blit;
928        pctx->flush_resource = v3d_flush_resource;
929}
930