xref: /aosp_15_r20/external/mesa3d/src/gallium/drivers/panfrost/pan_resource.c (revision 6104692788411f58d303aa86923a9ff6ecaded22)
1 /*
2  * Copyright (C) 2008 VMware, Inc.
3  * Copyright (C) 2012 Rob Clark <[email protected]>
4  * Copyright (C) 2014-2017 Broadcom
5  * Copyright (C) 2018-2019 Alyssa Rosenzweig
6  * Copyright (C) 2019 Collabora, Ltd.
7  * Copyright (C) 2023 Amazon.com, Inc. or its affiliates
8  *
9  * Permission is hereby granted, free of charge, to any person obtaining a
10  * copy of this software and associated documentation files (the "Software"),
11  * to deal in the Software without restriction, including without limitation
12  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
13  * and/or sell copies of the Software, and to permit persons to whom the
14  * Software is furnished to do so, subject to the following conditions:
15  *
16  * The above copyright notice and this permission notice (including the next
17  * paragraph) shall be included in all copies or substantial portions of the
18  * Software.
19  *
20  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
21  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
22  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
23  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
24  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
25  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26  * SOFTWARE.
27  *
28  * Authors (Collabora):
29  *   Tomeu Vizoso <[email protected]>
30  *   Alyssa Rosenzweig <[email protected]>
31  *
32  */
33 
34 #include <fcntl.h>
35 #include <xf86drm.h>
36 #include "drm-uapi/drm_fourcc.h"
37 
38 #include "frontend/winsys_handle.h"
39 #include "util/format/u_format.h"
40 #include "util/u_debug_image.h"
41 #include "util/u_drm.h"
42 #include "util/u_gen_mipmap.h"
43 #include "util/u_memory.h"
44 #include "util/u_resource.h"
45 #include "util/u_surface.h"
46 #include "util/u_transfer.h"
47 #include "util/u_transfer_helper.h"
48 
49 #include "decode.h"
50 #include "pan_bo.h"
51 #include "pan_context.h"
52 #include "pan_resource.h"
53 #include "pan_screen.h"
54 #include "pan_tiling.h"
55 #include "pan_util.h"
56 
57 static void
panfrost_clear_depth_stencil(struct pipe_context * pipe,struct pipe_surface * dst,unsigned clear_flags,double depth,unsigned stencil,unsigned dstx,unsigned dsty,unsigned width,unsigned height,bool render_condition_enabled)58 panfrost_clear_depth_stencil(struct pipe_context *pipe,
59                              struct pipe_surface *dst, unsigned clear_flags,
60                              double depth, unsigned stencil, unsigned dstx,
61                              unsigned dsty, unsigned width, unsigned height,
62                              bool render_condition_enabled)
63 {
64    struct panfrost_context *ctx = pan_context(pipe);
65 
66    if (render_condition_enabled && !panfrost_render_condition_check(ctx))
67       return;
68 
69    panfrost_blitter_save(
70       ctx, render_condition_enabled ? PAN_RENDER_COND : PAN_RENDER_BASE);
71    util_blitter_clear_depth_stencil(ctx->blitter, dst, clear_flags, depth,
72                                     stencil, dstx, dsty, width, height);
73 }
74 
75 static void
panfrost_clear_render_target(struct pipe_context * pipe,struct pipe_surface * dst,const union pipe_color_union * color,unsigned dstx,unsigned dsty,unsigned width,unsigned height,bool render_condition_enabled)76 panfrost_clear_render_target(struct pipe_context *pipe,
77                              struct pipe_surface *dst,
78                              const union pipe_color_union *color, unsigned dstx,
79                              unsigned dsty, unsigned width, unsigned height,
80                              bool render_condition_enabled)
81 {
82    struct panfrost_context *ctx = pan_context(pipe);
83 
84    if (render_condition_enabled && !panfrost_render_condition_check(ctx))
85       return;
86 
87    panfrost_blitter_save(
88       ctx, render_condition_enabled ? PAN_RENDER_COND : PAN_RENDER_BASE);
89    util_blitter_clear_render_target(ctx->blitter, dst, color, dstx, dsty, width,
90                                     height);
91 }
92 
93 static struct pipe_resource *
panfrost_resource_from_handle(struct pipe_screen * pscreen,const struct pipe_resource * templat,struct winsys_handle * whandle,unsigned usage)94 panfrost_resource_from_handle(struct pipe_screen *pscreen,
95                               const struct pipe_resource *templat,
96                               struct winsys_handle *whandle, unsigned usage)
97 {
98    struct panfrost_device *dev = pan_device(pscreen);
99    struct panfrost_resource *rsc;
100    struct pipe_resource *prsc;
101 
102    assert(whandle->type == WINSYS_HANDLE_TYPE_FD);
103 
104    rsc = CALLOC_STRUCT(panfrost_resource);
105    if (!rsc)
106       return NULL;
107 
108    prsc = &rsc->base;
109 
110    *prsc = *templat;
111 
112    pipe_reference_init(&prsc->reference, 1);
113    prsc->screen = pscreen;
114 
115    uint64_t mod = whandle->modifier == DRM_FORMAT_MOD_INVALID
116                      ? DRM_FORMAT_MOD_LINEAR
117                      : whandle->modifier;
118    enum mali_texture_dimension dim =
119       panfrost_translate_texture_dimension(templat->target);
120    struct pan_image_explicit_layout explicit_layout = {
121       .offset = whandle->offset,
122       .row_stride =
123          panfrost_from_legacy_stride(whandle->stride, templat->format, mod),
124    };
125 
126    rsc->image.layout = (struct pan_image_layout){
127       .modifier = mod,
128       .format = templat->format,
129       .dim = dim,
130       .width = prsc->width0,
131       .height = prsc->height0,
132       .depth = prsc->depth0,
133       .array_size = prsc->array_size,
134       .nr_samples = MAX2(prsc->nr_samples, 1),
135       .nr_slices = 1,
136    };
137 
138    bool valid =
139       pan_image_layout_init(dev->arch, &rsc->image.layout, &explicit_layout);
140 
141    if (!valid) {
142       FREE(rsc);
143       return NULL;
144    }
145 
146    rsc->bo = panfrost_bo_import(dev, whandle->handle);
147    /* Sometimes an import can fail e.g. on an invalid buffer fd, out of
148     * memory space to mmap it etc.
149     */
150    if (!rsc->bo) {
151       FREE(rsc);
152       return NULL;
153    }
154 
155    rsc->image.data.base = rsc->bo->ptr.gpu;
156    rsc->modifier_constant = true;
157 
158    BITSET_SET(rsc->valid.data, 0);
159    panfrost_resource_set_damage_region(pscreen, &rsc->base, 0, NULL);
160 
161    if (dev->ro) {
162       rsc->scanout =
163          renderonly_create_gpu_import_for_resource(prsc, dev->ro, NULL);
164       /* failure is expected in some cases.. */
165    }
166 
167    return prsc;
168 }
169 
170 static bool
panfrost_resource_get_handle(struct pipe_screen * pscreen,struct pipe_context * ctx,struct pipe_resource * pt,struct winsys_handle * handle,unsigned usage)171 panfrost_resource_get_handle(struct pipe_screen *pscreen,
172                              struct pipe_context *ctx, struct pipe_resource *pt,
173                              struct winsys_handle *handle, unsigned usage)
174 {
175    struct panfrost_device *dev = pan_device(pscreen);
176    struct panfrost_resource *rsrc;
177    struct renderonly_scanout *scanout;
178    struct pipe_resource *plane_res =
179       util_resource_at_index(pt, handle->plane);
180 
181    if (!plane_res)
182       return false;
183 
184    rsrc = pan_resource(plane_res);
185    scanout = rsrc->scanout;
186 
187    handle->modifier = rsrc->image.layout.modifier;
188    rsrc->modifier_constant = true;
189 
190    if (handle->type == WINSYS_HANDLE_TYPE_KMS && dev->ro) {
191       return renderonly_get_handle(scanout, handle);
192    } else if (handle->type == WINSYS_HANDLE_TYPE_KMS) {
193       handle->handle = panfrost_bo_handle(rsrc->bo);
194    } else if (handle->type == WINSYS_HANDLE_TYPE_FD) {
195       int fd = panfrost_bo_export(rsrc->bo);
196 
197       if (fd < 0)
198          return false;
199 
200       handle->handle = fd;
201    } else {
202       /* Other handle types not supported */
203       return false;
204    }
205 
206    handle->stride = panfrost_get_legacy_stride(&rsrc->image.layout, 0);
207    handle->offset = rsrc->image.layout.slices[0].offset;
208    return true;
209 }
210 
211 static bool
panfrost_resource_get_param(struct pipe_screen * pscreen,struct pipe_context * pctx,struct pipe_resource * prsc,unsigned plane,unsigned layer,unsigned level,enum pipe_resource_param param,unsigned usage,uint64_t * value)212 panfrost_resource_get_param(struct pipe_screen *pscreen,
213                             struct pipe_context *pctx,
214                             struct pipe_resource *prsc, unsigned plane,
215                             unsigned layer, unsigned level,
216                             enum pipe_resource_param param, unsigned usage,
217                             uint64_t *value)
218 {
219    struct pipe_resource *plane_res = util_resource_at_index(prsc, plane);
220    struct panfrost_resource *rsrc = pan_resource(plane_res);
221 
222    switch (param) {
223    case PIPE_RESOURCE_PARAM_STRIDE:
224       *value = panfrost_get_legacy_stride(&rsrc->image.layout, level);
225       return true;
226    case PIPE_RESOURCE_PARAM_OFFSET:
227       *value = rsrc->image.layout.slices[level].offset;
228       return true;
229    case PIPE_RESOURCE_PARAM_MODIFIER:
230       *value = rsrc->image.layout.modifier;
231       return true;
232    case PIPE_RESOURCE_PARAM_NPLANES:
233       *value = util_resource_num(prsc);
234       return true;
235    default:
236       return false;
237    }
238 }
239 
240 static void
panfrost_flush_resource(struct pipe_context * pctx,struct pipe_resource * prsc)241 panfrost_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
242 {
243    /* TODO */
244 }
245 
246 static struct pipe_surface *
panfrost_create_surface(struct pipe_context * pipe,struct pipe_resource * pt,const struct pipe_surface * surf_tmpl)247 panfrost_create_surface(struct pipe_context *pipe, struct pipe_resource *pt,
248                         const struct pipe_surface *surf_tmpl)
249 {
250    struct pipe_surface *ps = NULL;
251 
252    ps = CALLOC_STRUCT(pipe_surface);
253 
254    if (ps) {
255       pipe_reference_init(&ps->reference, 1);
256       pipe_resource_reference(&ps->texture, pt);
257       ps->context = pipe;
258       ps->format = surf_tmpl->format;
259 
260       if (pt->target != PIPE_BUFFER) {
261          assert(surf_tmpl->u.tex.level <= pt->last_level);
262          ps->width = u_minify(pt->width0, surf_tmpl->u.tex.level);
263          ps->height = u_minify(pt->height0, surf_tmpl->u.tex.level);
264          ps->nr_samples = surf_tmpl->nr_samples;
265          ps->u.tex.level = surf_tmpl->u.tex.level;
266          ps->u.tex.first_layer = surf_tmpl->u.tex.first_layer;
267          ps->u.tex.last_layer = surf_tmpl->u.tex.last_layer;
268       } else {
269          /* setting width as number of elements should get us correct
270           * renderbuffer width */
271          ps->width =
272             surf_tmpl->u.buf.last_element - surf_tmpl->u.buf.first_element + 1;
273          ps->height = pt->height0;
274          ps->u.buf.first_element = surf_tmpl->u.buf.first_element;
275          ps->u.buf.last_element = surf_tmpl->u.buf.last_element;
276          assert(ps->u.buf.first_element <= ps->u.buf.last_element);
277          assert(ps->u.buf.last_element < ps->width);
278       }
279    }
280 
281    return ps;
282 }
283 
284 static void
panfrost_surface_destroy(struct pipe_context * pipe,struct pipe_surface * surf)285 panfrost_surface_destroy(struct pipe_context *pipe, struct pipe_surface *surf)
286 {
287    assert(surf->texture);
288    pipe_resource_reference(&surf->texture, NULL);
289    free(surf);
290 }
291 
292 static inline bool
panfrost_is_2d(const struct panfrost_resource * pres)293 panfrost_is_2d(const struct panfrost_resource *pres)
294 {
295    return (pres->base.target == PIPE_TEXTURE_2D) ||
296           (pres->base.target == PIPE_TEXTURE_RECT);
297 }
298 
299 /* Based on the usage, determine if it makes sense to use u-inteleaved tiling.
300  * We only have routines to tile 2D textures of sane bpps. On the hardware
301  * level, not all usages are valid for tiling. Finally, if the app is hinting
302  * that the contents frequently change, tiling will be a loss.
303  *
304  * On platforms where it is supported, AFBC is even better. */
305 
306 static bool
panfrost_should_afbc(struct panfrost_device * dev,const struct panfrost_resource * pres,enum pipe_format fmt)307 panfrost_should_afbc(struct panfrost_device *dev,
308                      const struct panfrost_resource *pres, enum pipe_format fmt)
309 {
310    /* AFBC resources may be rendered to, textured from, or shared across
311     * processes, but may not be used as e.g buffers */
312    const unsigned valid_binding =
313       PIPE_BIND_DEPTH_STENCIL | PIPE_BIND_RENDER_TARGET | PIPE_BIND_BLENDABLE |
314       PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT |
315       PIPE_BIND_SHARED;
316 
317    if (pres->base.bind & ~valid_binding)
318       return false;
319 
320    /* AFBC support is optional */
321    if (!dev->has_afbc)
322       return false;
323 
324    /* AFBC<-->staging is expensive */
325    if (pres->base.usage == PIPE_USAGE_STREAM)
326       return false;
327 
328    /* If constant (non-data-dependent) format is requested, don't AFBC: */
329    if (pres->base.bind & PIPE_BIND_CONST_BW)
330       return false;
331 
332    /* Only a small selection of formats are AFBC'able */
333    if (!panfrost_format_supports_afbc(dev->arch, fmt))
334       return false;
335 
336    /* AFBC does not support layered (GLES3 style) multisampling. Use
337     * EXT_multisampled_render_to_texture instead */
338    if (pres->base.nr_samples > 1)
339       return false;
340 
341    switch (pres->base.target) {
342    case PIPE_TEXTURE_2D:
343    case PIPE_TEXTURE_RECT:
344    case PIPE_TEXTURE_2D_ARRAY:
345    case PIPE_TEXTURE_CUBE:
346    case PIPE_TEXTURE_CUBE_ARRAY:
347       break;
348 
349    case PIPE_TEXTURE_3D:
350       /* 3D AFBC is only supported on Bifrost v7+. It's supposed to
351        * be supported on Midgard but it doesn't seem to work */
352       if (dev->arch != 7)
353          return false;
354 
355       break;
356 
357    default:
358       return false;
359    }
360 
361    /* For one tile, AFBC is a loss compared to u-interleaved */
362    if (pres->base.width0 <= 16 && pres->base.height0 <= 16)
363       return false;
364 
365    /* Otherwise, we'd prefer AFBC as it is dramatically more efficient
366     * than linear or usually even u-interleaved */
367    return true;
368 }
369 
370 /*
371  * For a resource we want to use AFBC with, should we use AFBC with tiled
372  * headers? On GPUs that support it, this is believed to be beneficial for
373  * images that are at least 128x128.
374  */
375 static bool
panfrost_should_tile_afbc(const struct panfrost_device * dev,const struct panfrost_resource * pres)376 panfrost_should_tile_afbc(const struct panfrost_device *dev,
377                           const struct panfrost_resource *pres)
378 {
379    return panfrost_afbc_can_tile(dev->arch) && pres->base.width0 >= 128 &&
380           pres->base.height0 >= 128 && !(dev->debug & PAN_DBG_FORCE_PACK);
381 }
382 
383 bool
panfrost_should_pack_afbc(struct panfrost_device * dev,const struct panfrost_resource * prsrc)384 panfrost_should_pack_afbc(struct panfrost_device *dev,
385                           const struct panfrost_resource *prsrc)
386 {
387    const unsigned valid_binding = PIPE_BIND_DEPTH_STENCIL |
388                                   PIPE_BIND_RENDER_TARGET |
389                                   PIPE_BIND_SAMPLER_VIEW;
390 
391    return panfrost_afbc_can_pack(prsrc->base.format) && panfrost_is_2d(prsrc) &&
392           drm_is_afbc(prsrc->image.layout.modifier) &&
393           (prsrc->image.layout.modifier & AFBC_FORMAT_MOD_SPARSE) &&
394           (prsrc->base.bind & ~valid_binding) == 0 &&
395           !prsrc->modifier_constant && prsrc->base.width0 >= 32 &&
396           prsrc->base.height0 >= 32;
397 }
398 
399 static bool
panfrost_should_tile(struct panfrost_device * dev,const struct panfrost_resource * pres,enum pipe_format fmt)400 panfrost_should_tile(struct panfrost_device *dev,
401                      const struct panfrost_resource *pres, enum pipe_format fmt)
402 {
403    const unsigned valid_binding =
404       PIPE_BIND_DEPTH_STENCIL | PIPE_BIND_RENDER_TARGET | PIPE_BIND_BLENDABLE |
405       PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT |
406       PIPE_BIND_SHARED;
407 
408    /* The purpose of tiling is improving locality in both X- and
409     * Y-directions. If there is only a single pixel in either direction,
410     * tiling does not make sense; using a linear layout instead is optimal
411     * for both memory usage and performance.
412     */
413    if (MIN2(pres->base.width0, pres->base.height0) < 2)
414       return false;
415 
416    bool can_tile = (pres->base.target != PIPE_BUFFER) &&
417                    ((pres->base.bind & ~valid_binding) == 0);
418 
419    return can_tile && (pres->base.usage != PIPE_USAGE_STREAM);
420 }
421 
422 static bool
panfrost_should_afrc(struct panfrost_device * dev,const struct panfrost_resource * pres,enum pipe_format fmt)423 panfrost_should_afrc(struct panfrost_device *dev,
424                      const struct panfrost_resource *pres, enum pipe_format fmt)
425 {
426    const unsigned valid_binding = PIPE_BIND_RENDER_TARGET |
427                                   PIPE_BIND_BLENDABLE | PIPE_BIND_SAMPLER_VIEW |
428                                   PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SHARED;
429 
430    if (pres->base.bind & ~valid_binding)
431       return false;
432 
433    /* AFRC support is optional */
434    if (!dev->has_afrc)
435       return false;
436 
437    /* AFRC<-->staging is expensive */
438    if (pres->base.usage == PIPE_USAGE_STREAM)
439       return false;
440 
441    /* Only a small selection of formats are AFRC'able */
442    if (!panfrost_format_supports_afrc(fmt))
443       return false;
444 
445    /* AFRC does not support layered (GLES3 style) multisampling. Use
446     * EXT_multisampled_render_to_texture instead */
447    if (pres->base.nr_samples > 1)
448       return false;
449 
450    switch (pres->base.target) {
451    case PIPE_TEXTURE_2D:
452    case PIPE_TEXTURE_RECT:
453    case PIPE_TEXTURE_2D_ARRAY:
454    case PIPE_TEXTURE_CUBE:
455    case PIPE_TEXTURE_CUBE_ARRAY:
456    case PIPE_TEXTURE_3D:
457       break;
458 
459    default:
460       return false;
461    }
462 
463    return true;
464 }
465 
466 static uint64_t
panfrost_best_modifier(struct pipe_screen * pscreen,const struct panfrost_resource * pres,enum pipe_format fmt)467 panfrost_best_modifier(struct pipe_screen *pscreen,
468                        const struct panfrost_resource *pres,
469                        enum pipe_format fmt)
470 {
471    struct panfrost_screen *screen = pan_screen(pscreen);
472    struct panfrost_device *dev = pan_device(pscreen);
473 
474    /* Force linear textures when debugging tiling/compression */
475    if (unlikely(dev->debug & PAN_DBG_LINEAR))
476       return DRM_FORMAT_MOD_LINEAR;
477 
478    int afrc_rate = screen->force_afrc_rate;
479    if (afrc_rate < 0)
480       afrc_rate = pres->base.compression_rate;
481    if (afrc_rate > PIPE_COMPRESSION_FIXED_RATE_NONE &&
482        panfrost_should_afrc(dev, pres, fmt)) {
483       /* It's not really possible to decide on a global AFRC-rate,
484        * because the set of valid AFRC rates varies from format to
485        * format. So instead, treat this as a minimum rate, and search
486        * for the next valid one.
487        */
488       for (int i = afrc_rate; i < 12; ++i) {
489          if (panfrost_afrc_get_modifiers(fmt, i, 0, NULL)) {
490             afrc_rate = i;
491             break;
492          }
493       }
494    }
495 
496    if (afrc_rate != PIPE_COMPRESSION_FIXED_RATE_NONE &&
497        panfrost_should_afrc(dev, pres, fmt)) {
498       uint64_t mod;
499       unsigned num_mods = 0;
500 
501       STATIC_ASSERT(PIPE_COMPRESSION_FIXED_RATE_DEFAULT == PAN_AFRC_RATE_DEFAULT);
502       num_mods = panfrost_afrc_get_modifiers(fmt, afrc_rate, 1, &mod);
503       if (num_mods > 0) {
504          return mod;
505       }
506    }
507 
508    if (panfrost_should_afbc(dev, pres, fmt)) {
509       uint64_t afbc = AFBC_FORMAT_MOD_BLOCK_SIZE_16x16 | AFBC_FORMAT_MOD_SPARSE;
510 
511       if (panfrost_afbc_can_ytr(pres->base.format))
512          afbc |= AFBC_FORMAT_MOD_YTR;
513 
514       if (panfrost_should_tile_afbc(dev, pres))
515          afbc |= AFBC_FORMAT_MOD_TILED | AFBC_FORMAT_MOD_SC;
516 
517       return DRM_FORMAT_MOD_ARM_AFBC(afbc);
518    } else if (panfrost_should_tile(dev, pres, fmt))
519       return DRM_FORMAT_MOD_ARM_16X16_BLOCK_U_INTERLEAVED;
520    else
521       return DRM_FORMAT_MOD_LINEAR;
522 }
523 
524 static bool
panfrost_should_checksum(const struct panfrost_device * dev,const struct panfrost_resource * pres)525 panfrost_should_checksum(const struct panfrost_device *dev,
526                          const struct panfrost_resource *pres)
527 {
528    /* Checksumming is disabled by default due to fundamental unsoundness */
529    if (!(dev->debug & PAN_DBG_CRC))
530       return false;
531 
532    /* When checksumming is enabled, the tile data must fit in the
533     * size of the writeback buffer, so don't checksum formats
534     * that use too much space. */
535 
536    unsigned bytes_per_pixel_max = (dev->arch == 6) ? 6 : 4;
537 
538    unsigned bytes_per_pixel = MAX2(pres->base.nr_samples, 1) *
539                               util_format_get_blocksize(pres->base.format);
540 
541    return pres->base.bind & PIPE_BIND_RENDER_TARGET && panfrost_is_2d(pres) &&
542           bytes_per_pixel <= bytes_per_pixel_max && pres->base.last_level == 0;
543 }
544 
545 static void
panfrost_resource_setup(struct pipe_screen * screen,struct panfrost_resource * pres,uint64_t modifier,enum pipe_format fmt)546 panfrost_resource_setup(struct pipe_screen *screen,
547                         struct panfrost_resource *pres, uint64_t modifier,
548                         enum pipe_format fmt)
549 {
550    struct panfrost_device *dev = pan_device(screen);
551    uint64_t chosen_mod = modifier != DRM_FORMAT_MOD_INVALID
552                             ? modifier
553                             : panfrost_best_modifier(screen, pres, fmt);
554    enum mali_texture_dimension dim =
555       panfrost_translate_texture_dimension(pres->base.target);
556 
557    /* We can only switch tiled->linear if the resource isn't already
558     * linear and if we control the modifier */
559    pres->modifier_constant = !(chosen_mod != DRM_FORMAT_MOD_LINEAR &&
560                                modifier == DRM_FORMAT_MOD_INVALID);
561 
562    /* Z32_S8X24 variants are actually stored in 2 planes (one per
563     * component), we have to adjust the format on the first plane.
564     */
565    if (fmt == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT)
566       fmt = PIPE_FORMAT_Z32_FLOAT;
567 
568    pres->image.layout = (struct pan_image_layout){
569       .modifier = chosen_mod,
570       .format = fmt,
571       .dim = dim,
572       .width = pres->base.width0,
573       .height = pres->base.height0,
574       .depth = pres->base.depth0,
575       .array_size = pres->base.array_size,
576       .nr_samples = MAX2(pres->base.nr_samples, 1),
577       .nr_slices = pres->base.last_level + 1,
578       .crc = panfrost_should_checksum(dev, pres),
579    };
580 
581    /* Update the compression rate with the correct value as we
582     * want the real bitrate and not DEFAULT */
583    pres->base.compression_rate = panfrost_afrc_get_rate(fmt, chosen_mod);
584 
585    ASSERTED bool valid =
586       pan_image_layout_init(dev->arch, &pres->image.layout, NULL);
587    assert(valid);
588 }
589 
590 static void
panfrost_resource_init_afbc_headers(struct panfrost_resource * pres)591 panfrost_resource_init_afbc_headers(struct panfrost_resource *pres)
592 {
593    panfrost_bo_mmap(pres->bo);
594 
595    unsigned nr_samples = MAX2(pres->base.nr_samples, 1);
596 
597    for (unsigned i = 0; i < pres->base.array_size; ++i) {
598       for (unsigned l = 0; l <= pres->base.last_level; ++l) {
599          struct pan_image_slice_layout *slice = &pres->image.layout.slices[l];
600 
601          for (unsigned s = 0; s < nr_samples; ++s) {
602             void *ptr = pres->bo->ptr.cpu +
603                         (i * pres->image.layout.array_stride) + slice->offset +
604                         (s * slice->afbc.surface_stride);
605 
606             /* Zero-ed AFBC headers seem to encode a plain
607              * black. Let's use this pattern to keep the
608              * initialization simple.
609              */
610             memset(ptr, 0, slice->afbc.header_size);
611          }
612       }
613    }
614 }
615 
616 void
panfrost_resource_set_damage_region(struct pipe_screen * screen,struct pipe_resource * res,unsigned int nrects,const struct pipe_box * rects)617 panfrost_resource_set_damage_region(struct pipe_screen *screen,
618                                     struct pipe_resource *res,
619                                     unsigned int nrects,
620                                     const struct pipe_box *rects)
621 {
622    struct panfrost_device *dev = pan_device(screen);
623    struct panfrost_resource *pres = pan_resource(res);
624    struct pipe_scissor_state *damage_extent = &pres->damage.extent;
625    unsigned int i;
626 
627    /* Partial updates are implemented with a tile enable map only on v5.
628     * Later architectures have a more efficient method of implementing
629     * partial updates (frame shaders), while earlier architectures lack
630     * tile enable maps altogether.
631     */
632    if (dev->arch == 5 && nrects > 1) {
633       if (!pres->damage.tile_map.data) {
634          pres->damage.tile_map.stride =
635             ALIGN_POT(DIV_ROUND_UP(res->width0, 32 * 8), 64);
636          pres->damage.tile_map.size =
637             pres->damage.tile_map.stride * DIV_ROUND_UP(res->height0, 32);
638          pres->damage.tile_map.data = malloc(pres->damage.tile_map.size);
639       }
640 
641       memset(pres->damage.tile_map.data, 0, pres->damage.tile_map.size);
642       pres->damage.tile_map.enable = true;
643    } else {
644       pres->damage.tile_map.enable = false;
645    }
646 
647    /* Track the damage extent: the quad including all damage regions. Will
648     * be used restrict the rendering area */
649 
650    damage_extent->minx = 0xffff;
651    damage_extent->miny = 0xffff;
652 
653    unsigned enable_count = 0;
654 
655    for (i = 0; i < nrects; i++) {
656       int x = rects[i].x, w = rects[i].width, h = rects[i].height;
657       int y = res->height0 - (rects[i].y + h);
658 
659       damage_extent->minx = MIN2(damage_extent->minx, x);
660       damage_extent->miny = MIN2(damage_extent->miny, y);
661       damage_extent->maxx = MAX2(damage_extent->maxx, MIN2(x + w, res->width0));
662       damage_extent->maxy =
663          MAX2(damage_extent->maxy, MIN2(y + h, res->height0));
664 
665       if (!pres->damage.tile_map.enable)
666          continue;
667 
668       unsigned t_x_start = x / 32;
669       unsigned t_x_end = (x + w - 1) / 32;
670       unsigned t_y_start = y / 32;
671       unsigned t_y_end = (y + h - 1) / 32;
672 
673       for (unsigned t_y = t_y_start; t_y <= t_y_end; t_y++) {
674          for (unsigned t_x = t_x_start; t_x <= t_x_end; t_x++) {
675             unsigned b = (t_y * pres->damage.tile_map.stride * 8) + t_x;
676 
677             if (BITSET_TEST(pres->damage.tile_map.data, b))
678                continue;
679 
680             BITSET_SET(pres->damage.tile_map.data, b);
681             enable_count++;
682          }
683       }
684    }
685 
686    if (nrects == 0) {
687       damage_extent->minx = 0;
688       damage_extent->miny = 0;
689       damage_extent->maxx = res->width0;
690       damage_extent->maxy = res->height0;
691    }
692 
693    if (pres->damage.tile_map.enable) {
694       unsigned t_x_start = damage_extent->minx / 32;
695       unsigned t_x_end = damage_extent->maxx / 32;
696       unsigned t_y_start = damage_extent->miny / 32;
697       unsigned t_y_end = damage_extent->maxy / 32;
698       unsigned tile_count =
699          (t_x_end - t_x_start + 1) * (t_y_end - t_y_start + 1);
700 
701       /* Don't bother passing a tile-enable-map if the amount of
702        * tiles to reload is to close to the total number of tiles.
703        */
704       if (tile_count - enable_count < 10)
705          pres->damage.tile_map.enable = false;
706    }
707 }
708 
709 struct pipe_resource *
panfrost_resource_create_with_modifier(struct pipe_screen * screen,const struct pipe_resource * template,uint64_t modifier)710 panfrost_resource_create_with_modifier(struct pipe_screen *screen,
711                                        const struct pipe_resource *template,
712                                        uint64_t modifier)
713 {
714    struct panfrost_device *dev = pan_device(screen);
715 
716    struct panfrost_resource *so = CALLOC_STRUCT(panfrost_resource);
717 
718    if (!so)
719       return NULL;
720 
721    so->base = *template;
722    so->base.screen = screen;
723 
724    pipe_reference_init(&so->base.reference, 1);
725 
726    util_range_init(&so->valid_buffer_range);
727 
728    if (template->bind & PAN_BIND_SHARED_MASK) {
729       /* For compatibility with older consumers that may not be
730        * modifiers aware, treat INVALID as LINEAR for shared
731        * resources.
732        */
733       if (modifier == DRM_FORMAT_MOD_INVALID)
734          modifier = DRM_FORMAT_MOD_LINEAR;
735 
736       /* At any rate, we can't change the modifier later for shared
737        * resources, since we have no way to propagate the modifier
738        * change.
739        */
740       so->modifier_constant = true;
741    }
742 
743    panfrost_resource_setup(screen, so, modifier, template->format);
744 
745    /* Guess a label based on the bind */
746    unsigned bind = template->bind;
747    const char *label = (bind & PIPE_BIND_INDEX_BUFFER)     ? "Index buffer"
748                        : (bind & PIPE_BIND_SCANOUT)        ? "Scanout"
749                        : (bind & PIPE_BIND_DISPLAY_TARGET) ? "Display target"
750                        : (bind & PIPE_BIND_SHARED)         ? "Shared resource"
751                        : (bind & PIPE_BIND_RENDER_TARGET)  ? "Render target"
752                        : (bind & PIPE_BIND_DEPTH_STENCIL)
753                           ? "Depth/stencil buffer"
754                        : (bind & PIPE_BIND_SAMPLER_VIEW)    ? "Texture"
755                        : (bind & PIPE_BIND_VERTEX_BUFFER)   ? "Vertex buffer"
756                        : (bind & PIPE_BIND_CONSTANT_BUFFER) ? "Constant buffer"
757                        : (bind & PIPE_BIND_GLOBAL)          ? "Global memory"
758                        : (bind & PIPE_BIND_SHADER_BUFFER)   ? "Shader buffer"
759                        : (bind & PIPE_BIND_SHADER_IMAGE)    ? "Shader image"
760                                                             : "Other resource";
761 
762    if (dev->ro && (template->bind & PIPE_BIND_SCANOUT)) {
763       struct winsys_handle handle;
764       struct pan_block_size blocksize =
765          panfrost_block_size(modifier, template->format);
766 
767       /* Block-based texture formats are only used for texture
768        * compression (not framebuffer compression!), which doesn't
769        * make sense to share across processes.
770        */
771       assert(util_format_get_blockwidth(template->format) == 1);
772       assert(util_format_get_blockheight(template->format) == 1);
773 
774       /* Present a resource with similar dimensions that, if allocated
775        * as a linear image, is big enough to fit the resource in the
776        * actual layout. For linear images, this is a no-op. For 16x16
777        * tiling, this aligns the dimensions to 16x16.
778        *
779        * For AFBC, this aligns the width to the superblock width (as
780        * expected) and adds extra rows to account for the header. This
781        * is a bit of a lie, but it's the best we can do with dumb
782        * buffers, which are extremely not meant for AFBC. And yet this
783        * has to work anyway...
784        *
785        * Moral of the story: if you're reading this comment, that
786        * means you're working on WSI and so it's already too late for
787        * you. I'm sorry.
788        */
789       unsigned width = ALIGN_POT(template->width0, blocksize.width);
790       unsigned stride = ALIGN_POT(template->width0, blocksize.width) *
791                         util_format_get_blocksize(template->format);
792       unsigned size = so->image.layout.data_size;
793       unsigned effective_rows = DIV_ROUND_UP(size, stride);
794 
795       struct pipe_resource scanout_tmpl = {
796          .target = so->base.target,
797          .format = template->format,
798          .width0 = width,
799          .height0 = effective_rows,
800          .depth0 = 1,
801          .array_size = 1,
802       };
803 
804       so->scanout =
805          renderonly_scanout_for_resource(&scanout_tmpl, dev->ro, &handle);
806 
807       if (!so->scanout) {
808          fprintf(stderr, "Failed to create scanout resource\n");
809          FREE(so);
810          return NULL;
811       }
812       assert(handle.type == WINSYS_HANDLE_TYPE_FD);
813       so->bo = panfrost_bo_import(dev, handle.handle);
814       close(handle.handle);
815 
816       if (!so->bo) {
817          FREE(so);
818          return NULL;
819       }
820 
821       so->image.data.base = so->bo->ptr.gpu;
822    } else {
823       /* We create a BO immediately but don't bother mapping, since we don't
824        * care to map e.g. FBOs which the CPU probably won't touch */
825       uint32_t flags = PAN_BO_DELAY_MMAP;
826 
827       /* If the resource is never exported, we can make the BO private. */
828       if (template->bind & PIPE_BIND_SHARED)
829          flags |= PAN_BO_SHAREABLE;
830 
831       so->bo =
832          panfrost_bo_create(dev, so->image.layout.data_size, flags, label);
833 
834       if (!so->bo) {
835          FREE(so);
836          return NULL;
837       }
838 
839       so->image.data.base = so->bo->ptr.gpu;
840 
841       so->constant_stencil = true;
842    }
843 
844    if (drm_is_afbc(so->image.layout.modifier))
845       panfrost_resource_init_afbc_headers(so);
846 
847    panfrost_resource_set_damage_region(screen, &so->base, 0, NULL);
848 
849    if (template->bind & PIPE_BIND_INDEX_BUFFER)
850       so->index_cache = CALLOC_STRUCT(panfrost_minmax_cache);
851 
852    return (struct pipe_resource *)so;
853 }
854 
855 /* Default is to create a resource as don't care */
856 
857 static struct pipe_resource *
panfrost_resource_create(struct pipe_screen * screen,const struct pipe_resource * template)858 panfrost_resource_create(struct pipe_screen *screen,
859                          const struct pipe_resource *template)
860 {
861    return panfrost_resource_create_with_modifier(screen, template,
862                                                  DRM_FORMAT_MOD_INVALID);
863 }
864 
865 /* If no modifier is specified, we'll choose. Otherwise, the order of
866  * preference is compressed, tiled, linear. */
867 
868 static struct pipe_resource *
panfrost_resource_create_with_modifiers(struct pipe_screen * screen,const struct pipe_resource * template,const uint64_t * modifiers,int count)869 panfrost_resource_create_with_modifiers(struct pipe_screen *screen,
870                                         const struct pipe_resource *template,
871                                         const uint64_t *modifiers, int count)
872 {
873    for (unsigned i = 0; i < PAN_MODIFIER_COUNT; ++i) {
874       if (drm_find_modifier(pan_best_modifiers[i], modifiers, count)) {
875          return panfrost_resource_create_with_modifier(screen, template,
876                                                        pan_best_modifiers[i]);
877       }
878    }
879 
880    /* If we didn't find one, app specified invalid */
881    assert(count == 1 && modifiers[0] == DRM_FORMAT_MOD_INVALID);
882    return panfrost_resource_create(screen, template);
883 }
884 
885 static void
panfrost_resource_destroy(struct pipe_screen * screen,struct pipe_resource * pt)886 panfrost_resource_destroy(struct pipe_screen *screen, struct pipe_resource *pt)
887 {
888    struct panfrost_device *dev = pan_device(screen);
889    struct panfrost_resource *rsrc = (struct panfrost_resource *)pt;
890 
891    if (rsrc->scanout)
892       renderonly_scanout_destroy(rsrc->scanout, dev->ro);
893 
894    if (rsrc->bo)
895       panfrost_bo_unreference(rsrc->bo);
896 
897    free(rsrc->index_cache);
898    free(rsrc->damage.tile_map.data);
899 
900    util_range_destroy(&rsrc->valid_buffer_range);
901    free(rsrc);
902 }
903 
904 /* Most of the time we can do CPU-side transfers, but sometimes we need to use
905  * the 3D pipe for this. Let's wrap u_blitter to blit to/from staging textures.
906  * Code adapted from freedreno */
907 
908 static struct panfrost_resource *
pan_alloc_staging(struct panfrost_context * ctx,struct panfrost_resource * rsc,unsigned level,const struct pipe_box * box)909 pan_alloc_staging(struct panfrost_context *ctx, struct panfrost_resource *rsc,
910                   unsigned level, const struct pipe_box *box)
911 {
912    struct pipe_context *pctx = &ctx->base;
913    struct pipe_resource tmpl = rsc->base;
914 
915    tmpl.width0 = box->width;
916    tmpl.height0 = box->height;
917    /* for array textures, box->depth is the array_size, otherwise
918     * for 3d textures, it is the depth:
919     */
920    if (tmpl.array_size > 1) {
921       if (tmpl.target == PIPE_TEXTURE_CUBE)
922          tmpl.target = PIPE_TEXTURE_2D_ARRAY;
923       tmpl.array_size = box->depth;
924       tmpl.depth0 = 1;
925    } else {
926       tmpl.array_size = 1;
927       tmpl.depth0 = box->depth;
928    }
929    tmpl.last_level = 0;
930    tmpl.bind |= PIPE_BIND_LINEAR;
931    tmpl.bind &= ~PAN_BIND_SHARED_MASK;
932    tmpl.compression_rate = PIPE_COMPRESSION_FIXED_RATE_NONE;
933 
934    struct pipe_resource *pstaging =
935       pctx->screen->resource_create(pctx->screen, &tmpl);
936    if (!pstaging)
937       return NULL;
938 
939    return pan_resource(pstaging);
940 }
941 
942 static void
pan_blit_from_staging(struct pipe_context * pctx,struct panfrost_transfer * trans)943 pan_blit_from_staging(struct pipe_context *pctx,
944                       struct panfrost_transfer *trans)
945 {
946    struct pipe_resource *dst = trans->base.resource;
947    struct pipe_blit_info blit = {0};
948 
949    blit.dst.resource = dst;
950    blit.dst.format = dst->format;
951    blit.dst.level = trans->base.level;
952    blit.dst.box = trans->base.box;
953    blit.src.resource = trans->staging.rsrc;
954    blit.src.format = trans->staging.rsrc->format;
955    blit.src.level = 0;
956    blit.src.box = trans->staging.box;
957    blit.mask = util_format_get_mask(blit.src.format);
958    blit.filter = PIPE_TEX_FILTER_NEAREST;
959 
960    panfrost_blit_no_afbc_legalization(pctx, &blit);
961 }
962 
963 static void
pan_blit_to_staging(struct pipe_context * pctx,struct panfrost_transfer * trans)964 pan_blit_to_staging(struct pipe_context *pctx, struct panfrost_transfer *trans)
965 {
966    struct pipe_resource *src = trans->base.resource;
967    struct pipe_blit_info blit = {0};
968 
969    blit.src.resource = src;
970    blit.src.format = src->format;
971    blit.src.level = trans->base.level;
972    blit.src.box = trans->base.box;
973    blit.dst.resource = trans->staging.rsrc;
974    blit.dst.format = trans->staging.rsrc->format;
975    blit.dst.level = 0;
976    blit.dst.box = trans->staging.box;
977    blit.mask = util_format_get_mask(blit.dst.format);
978    blit.filter = PIPE_TEX_FILTER_NEAREST;
979 
980    panfrost_blit_no_afbc_legalization(pctx, &blit);
981 }
982 
983 static void
panfrost_load_tiled_images(struct panfrost_transfer * transfer,struct panfrost_resource * rsrc)984 panfrost_load_tiled_images(struct panfrost_transfer *transfer,
985                            struct panfrost_resource *rsrc)
986 {
987    struct pipe_transfer *ptrans = &transfer->base;
988    unsigned level = ptrans->level;
989 
990    /* If the requested level of the image is uninitialized, it's not
991     * necessary to copy it. Leave the result unintiialized too.
992     */
993    if (!BITSET_TEST(rsrc->valid.data, level))
994       return;
995 
996    struct panfrost_bo *bo = rsrc->bo;
997    unsigned stride = panfrost_get_layer_stride(&rsrc->image.layout, level);
998 
999    /* Otherwise, load each layer separately, required to load from 3D and
1000     * array textures.
1001     */
1002    for (unsigned z = 0; z < ptrans->box.depth; ++z) {
1003       void *dst = transfer->map + (ptrans->layer_stride * z);
1004       uint8_t *map = bo->ptr.cpu + rsrc->image.layout.slices[level].offset +
1005                      (z + ptrans->box.z) * stride;
1006 
1007       panfrost_load_tiled_image(dst, map, ptrans->box.x, ptrans->box.y,
1008                                 ptrans->box.width, ptrans->box.height,
1009                                 ptrans->stride,
1010                                 rsrc->image.layout.slices[level].row_stride,
1011                                 rsrc->image.layout.format);
1012    }
1013 }
1014 
1015 #if MESA_DEBUG
1016 
1017 static unsigned
get_superblock_size(uint32_t * hdr,unsigned uncompressed_size)1018 get_superblock_size(uint32_t *hdr, unsigned uncompressed_size)
1019 {
1020    /* AFBC superblock layout 0 */
1021    unsigned body_base_ptr_len = 32;
1022    unsigned nr_subblocks = 16;
1023    unsigned sz_len = 6; /* bits */
1024    unsigned mask = (1 << sz_len) - 1;
1025    unsigned size = 0;
1026 
1027    /* Sum up all of the subblock sizes */
1028    for (int i = 0; i < nr_subblocks; i++) {
1029       unsigned bitoffset = body_base_ptr_len + (i * sz_len);
1030       unsigned start = bitoffset / 32;
1031       unsigned end = (bitoffset + (sz_len - 1)) / 32;
1032       unsigned offset = bitoffset % 32;
1033       unsigned subblock_size;
1034 
1035       if (start != end)
1036          subblock_size = (hdr[start] >> offset) | (hdr[end] << (32 - offset));
1037       else
1038          subblock_size = hdr[start] >> offset;
1039       subblock_size = (subblock_size == 1) ? uncompressed_size : subblock_size;
1040       size += subblock_size & mask;
1041 
1042       if (i == 0 && size == 0)
1043          return 0;
1044    }
1045 
1046    return size;
1047 }
1048 
1049 static void
dump_block(struct panfrost_resource * rsrc,uint32_t idx)1050 dump_block(struct panfrost_resource *rsrc, uint32_t idx)
1051 {
1052    panfrost_bo_wait(rsrc->bo, INT64_MAX, false);
1053 
1054    uint8_t *ptr = rsrc->bo->ptr.cpu;
1055    uint32_t *header = (uint32_t *)(ptr + (idx * AFBC_HEADER_BYTES_PER_TILE));
1056    uint32_t body_base_ptr = header[0];
1057    uint32_t *body = (uint32_t *)(ptr + body_base_ptr);
1058    struct pan_block_size block_sz =
1059       panfrost_afbc_subblock_size(rsrc->image.layout.modifier);
1060    unsigned pixel_sz = util_format_get_blocksize(rsrc->base.format);
1061    unsigned uncompressed_size = pixel_sz * block_sz.width * block_sz.height;
1062    unsigned size = get_superblock_size(header, uncompressed_size);
1063 
1064    fprintf(stderr, "  Header: %08x %08x %08x %08x (size: %u bytes)\n",
1065            header[0], header[1], header[2], header[3], size);
1066    if (size > 0) {
1067       fprintf(stderr, "  Body:   %08x %08x %08x %08x\n", body[0], body[1],
1068               body[2], body[3]);
1069    } else {
1070       uint8_t *comp = (uint8_t *)(header + 2);
1071       fprintf(stderr, "  Color:  0x%02x%02x%02x%02x\n", comp[0], comp[1],
1072               comp[2], comp[3]);
1073    }
1074    fprintf(stderr, "\n");
1075 }
1076 
1077 void
pan_dump_resource(struct panfrost_context * ctx,struct panfrost_resource * rsc)1078 pan_dump_resource(struct panfrost_context *ctx, struct panfrost_resource *rsc)
1079 {
1080    struct pipe_context *pctx = &ctx->base;
1081    struct pipe_resource tmpl = rsc->base;
1082    struct pipe_resource *plinear = NULL;
1083    struct panfrost_resource *linear = rsc;
1084    struct pipe_blit_info blit = {0};
1085    struct pipe_box box;
1086    char buffer[1024];
1087 
1088    if (rsc->image.layout.modifier != DRM_FORMAT_MOD_LINEAR) {
1089       tmpl.bind |= PIPE_BIND_LINEAR;
1090       tmpl.bind &= ~PAN_BIND_SHARED_MASK;
1091 
1092       plinear = pctx->screen->resource_create(pctx->screen, &tmpl);
1093       u_box_2d(0, 0, rsc->base.width0, rsc->base.height0, &box);
1094 
1095       blit.src.resource = &rsc->base;
1096       blit.src.format = rsc->base.format;
1097       blit.src.level = 0;
1098       blit.src.box = box;
1099       blit.dst.resource = plinear;
1100       blit.dst.format = rsc->base.format;
1101       blit.dst.level = 0;
1102       blit.dst.box = box;
1103       blit.mask = util_format_get_mask(blit.dst.format);
1104       blit.filter = PIPE_TEX_FILTER_NEAREST;
1105 
1106       panfrost_blit(pctx, &blit);
1107 
1108       linear = pan_resource(plinear);
1109    }
1110 
1111    panfrost_flush_writer(ctx, linear, "dump image");
1112    panfrost_bo_wait(linear->bo, INT64_MAX, false);
1113    panfrost_bo_mmap(linear->bo);
1114 
1115    static unsigned frame_count = 0;
1116    frame_count++;
1117    snprintf(buffer, sizeof(buffer), "dump_image.%04d", frame_count);
1118 
1119    debug_dump_image(buffer, rsc->base.format, 0 /* UNUSED */, rsc->base.width0,
1120                     rsc->base.height0,
1121                     linear->image.layout.slices[0].row_stride,
1122                     linear->bo->ptr.cpu);
1123 
1124    if (plinear)
1125       pipe_resource_reference(&plinear, NULL);
1126 }
1127 
1128 #endif
1129 
1130 /* Get scan-order index from (x, y) position when blocks are
1131  * arranged in z-order in 8x8 tiles */
1132 static unsigned
get_morton_index(unsigned x,unsigned y,unsigned stride)1133 get_morton_index(unsigned x, unsigned y, unsigned stride)
1134 {
1135    unsigned i = ((x << 0) & 1) | ((y << 1) & 2) | ((x << 1) & 4) |
1136                 ((y << 2) & 8) | ((x << 2) & 16) | ((y << 3) & 32);
1137    return (((y & ~7) * stride) + ((x & ~7) << 3)) + i;
1138 }
1139 
1140 static void
panfrost_store_tiled_images(struct panfrost_transfer * transfer,struct panfrost_resource * rsrc)1141 panfrost_store_tiled_images(struct panfrost_transfer *transfer,
1142                             struct panfrost_resource *rsrc)
1143 {
1144    struct panfrost_bo *bo = rsrc->bo;
1145    struct pipe_transfer *ptrans = &transfer->base;
1146    unsigned level = ptrans->level;
1147    unsigned stride = panfrost_get_layer_stride(&rsrc->image.layout, level);
1148 
1149    /* Otherwise, store each layer separately, required to store to 3D and
1150     * array textures.
1151     */
1152    for (unsigned z = 0; z < ptrans->box.depth; ++z) {
1153       void *src = transfer->map + (ptrans->layer_stride * z);
1154       uint8_t *map = bo->ptr.cpu + rsrc->image.layout.slices[level].offset +
1155                      (z + ptrans->box.z) * stride;
1156 
1157       panfrost_store_tiled_image(map, src, ptrans->box.x, ptrans->box.y,
1158                                  ptrans->box.width, ptrans->box.height,
1159                                  rsrc->image.layout.slices[level].row_stride,
1160                                  ptrans->stride, rsrc->image.layout.format);
1161    }
1162 }
1163 
1164 static bool
panfrost_box_covers_resource(const struct pipe_resource * resource,const struct pipe_box * box)1165 panfrost_box_covers_resource(const struct pipe_resource *resource,
1166                              const struct pipe_box *box)
1167 {
1168    return resource->last_level == 0 &&
1169           util_texrange_covers_whole_level(resource, 0, box->x, box->y, box->z,
1170                                            box->width, box->height, box->depth);
1171 }
1172 
1173 static bool
panfrost_can_discard(struct pipe_resource * resource,const struct pipe_box * box,unsigned usage)1174 panfrost_can_discard(struct pipe_resource *resource, const struct pipe_box *box,
1175                      unsigned usage)
1176 {
1177    struct panfrost_resource *rsrc = pan_resource(resource);
1178 
1179    return ((usage & PIPE_MAP_DISCARD_RANGE) &&
1180            !(usage & PIPE_MAP_UNSYNCHRONIZED) &&
1181            !(resource->flags & PIPE_RESOURCE_FLAG_MAP_PERSISTENT) &&
1182            panfrost_box_covers_resource(resource, box) &&
1183            !(rsrc->bo->flags & PAN_BO_SHARED));
1184 }
1185 
1186 static void *
panfrost_ptr_map(struct pipe_context * pctx,struct pipe_resource * resource,unsigned level,unsigned usage,const struct pipe_box * box,struct pipe_transfer ** out_transfer)1187 panfrost_ptr_map(struct pipe_context *pctx, struct pipe_resource *resource,
1188                  unsigned level,
1189                  unsigned usage, /* a combination of PIPE_MAP_x */
1190                  const struct pipe_box *box,
1191                  struct pipe_transfer **out_transfer)
1192 {
1193    struct panfrost_context *ctx = pan_context(pctx);
1194    struct panfrost_device *dev = pan_device(pctx->screen);
1195    struct panfrost_resource *rsrc = pan_resource(resource);
1196    enum pipe_format format = rsrc->image.layout.format;
1197    int bytes_per_block = util_format_get_blocksize(format);
1198    struct panfrost_bo *bo = rsrc->bo;
1199 
1200    /* Can't map tiled/compressed directly */
1201    if ((usage & PIPE_MAP_DIRECTLY) &&
1202        rsrc->image.layout.modifier != DRM_FORMAT_MOD_LINEAR)
1203       return NULL;
1204 
1205    struct panfrost_transfer *transfer = rzalloc(pctx, struct panfrost_transfer);
1206    transfer->base.level = level;
1207    transfer->base.usage = usage;
1208    transfer->base.box = *box;
1209 
1210    pipe_resource_reference(&transfer->base.resource, resource);
1211    *out_transfer = &transfer->base;
1212 
1213    if (usage & PIPE_MAP_WRITE)
1214       rsrc->constant_stencil = false;
1215 
1216    /* We don't have s/w routines for AFBC/AFRC, so use a staging texture */
1217    if (drm_is_afbc(rsrc->image.layout.modifier) ||
1218        drm_is_afrc(rsrc->image.layout.modifier)) {
1219       struct panfrost_resource *staging =
1220          pan_alloc_staging(ctx, rsrc, level, box);
1221       assert(staging);
1222 
1223       /* Staging resources have one LOD: level 0. Query the strides
1224        * on this LOD.
1225        */
1226       transfer->base.stride = staging->image.layout.slices[0].row_stride;
1227       transfer->base.layer_stride =
1228          panfrost_get_layer_stride(&staging->image.layout, 0);
1229 
1230       transfer->staging.rsrc = &staging->base;
1231 
1232       transfer->staging.box = *box;
1233       transfer->staging.box.x = 0;
1234       transfer->staging.box.y = 0;
1235       transfer->staging.box.z = 0;
1236 
1237       assert(transfer->staging.rsrc != NULL);
1238 
1239       bool valid = BITSET_TEST(rsrc->valid.data, level);
1240 
1241       if ((usage & PIPE_MAP_READ) &&
1242           (valid || panfrost_any_batch_writes_rsrc(ctx, rsrc))) {
1243          pan_blit_to_staging(pctx, transfer);
1244          panfrost_flush_writer(ctx, staging, "AFBC/AFRC tex read staging blit");
1245          panfrost_bo_wait(staging->bo, INT64_MAX, false);
1246       }
1247 
1248       panfrost_bo_mmap(staging->bo);
1249       return staging->bo->ptr.cpu;
1250    }
1251 
1252    bool already_mapped = bo->ptr.cpu != NULL;
1253 
1254    /* If we haven't already mmaped, now's the time */
1255    panfrost_bo_mmap(bo);
1256 
1257    if (dev->debug & (PAN_DBG_TRACE | PAN_DBG_SYNC)) {
1258       pandecode_inject_mmap(dev->decode_ctx, bo->ptr.gpu, bo->ptr.cpu,
1259                             panfrost_bo_size(bo), NULL);
1260    }
1261 
1262    /* Upgrade writes to uninitialized ranges to UNSYNCHRONIZED */
1263    if ((usage & PIPE_MAP_WRITE) && resource->target == PIPE_BUFFER &&
1264        !util_ranges_intersect(&rsrc->valid_buffer_range, box->x,
1265                               box->x + box->width)) {
1266 
1267       usage |= PIPE_MAP_UNSYNCHRONIZED;
1268    }
1269 
1270    /* Upgrade DISCARD_RANGE to WHOLE_RESOURCE if the whole resource is
1271     * being mapped.
1272     */
1273    if (panfrost_can_discard(resource, box, usage)) {
1274       usage |= PIPE_MAP_DISCARD_WHOLE_RESOURCE;
1275    }
1276 
1277    bool create_new_bo = usage & PIPE_MAP_DISCARD_WHOLE_RESOURCE;
1278    bool copy_resource = false;
1279 
1280    if (!(usage & PIPE_MAP_UNSYNCHRONIZED) &&
1281        !(resource->flags & PIPE_RESOURCE_FLAG_MAP_PERSISTENT) &&
1282        (usage & PIPE_MAP_WRITE) && panfrost_any_batch_reads_rsrc(ctx, rsrc)) {
1283       /* When a resource to be modified is already being used by a
1284        * pending batch, it is often faster to copy the whole BO than
1285        * to flush and split the frame in two.
1286        */
1287 
1288       panfrost_flush_writer(ctx, rsrc, "Shadow resource creation");
1289       panfrost_bo_wait(bo, INT64_MAX, false);
1290 
1291       create_new_bo = true;
1292       copy_resource = !(usage & PIPE_MAP_DISCARD_WHOLE_RESOURCE);
1293    }
1294 
1295    /* Shadowing with separate stencil may require additional accounting.
1296     * Bail in these exotic cases.
1297     */
1298    if (rsrc->separate_stencil) {
1299       create_new_bo = false;
1300       copy_resource = false;
1301    }
1302 
1303    if (create_new_bo &&
1304        (!(resource->flags & PIPE_RESOURCE_FLAG_MAP_PERSISTENT) ||
1305         !already_mapped)) {
1306       /* Make sure we re-emit any descriptors using this resource */
1307       panfrost_dirty_state_all(ctx);
1308 
1309       /* If the BO is used by one of the pending batches or if it's
1310        * not ready yet (still accessed by one of the already flushed
1311        * batches), we try to allocate a new one to avoid waiting.
1312        */
1313       if (panfrost_any_batch_reads_rsrc(ctx, rsrc) ||
1314           !panfrost_bo_wait(bo, 0, true)) {
1315          /* We want the BO to be MMAPed. */
1316          uint32_t flags = bo->flags & ~PAN_BO_DELAY_MMAP;
1317          struct panfrost_bo *newbo = NULL;
1318 
1319          /* When the BO has been imported/exported, we can't
1320           * replace it by another one, otherwise the
1321           * importer/exporter wouldn't see the change we're
1322           * doing to it.
1323           */
1324          if (!(bo->flags & PAN_BO_SHARED)) {
1325             newbo =
1326                panfrost_bo_create(dev, panfrost_bo_size(bo), flags, bo->label);
1327          }
1328 
1329          if (newbo) {
1330             if (copy_resource) {
1331                memcpy(newbo->ptr.cpu, rsrc->bo->ptr.cpu, panfrost_bo_size(bo));
1332             }
1333 
1334             /* Swap the pointers, dropping a reference to
1335              * the old BO which is no long referenced from
1336              * the resource.
1337              */
1338             panfrost_bo_unreference(rsrc->bo);
1339             rsrc->bo = newbo;
1340             rsrc->image.data.base = newbo->ptr.gpu;
1341 
1342             if (!copy_resource && drm_is_afbc(rsrc->image.layout.modifier))
1343                panfrost_resource_init_afbc_headers(rsrc);
1344 
1345             bo = newbo;
1346          } else {
1347             /* Allocation failed or was impossible, let's
1348              * fall back on a flush+wait.
1349              */
1350             panfrost_flush_batches_accessing_rsrc(
1351                ctx, rsrc, "Resource access with high memory pressure");
1352             panfrost_bo_wait(bo, INT64_MAX, true);
1353          }
1354       }
1355    } else if (!(usage & PIPE_MAP_UNSYNCHRONIZED)) {
1356       if (usage & PIPE_MAP_WRITE) {
1357          panfrost_flush_batches_accessing_rsrc(ctx, rsrc, "Synchronized write");
1358          panfrost_bo_wait(bo, INT64_MAX, true);
1359       } else if (usage & PIPE_MAP_READ) {
1360          panfrost_flush_writer(ctx, rsrc, "Synchronized read");
1361          panfrost_bo_wait(bo, INT64_MAX, false);
1362       }
1363    }
1364 
1365    /* For access to compressed textures, we want the (x, y, w, h)
1366     * region-of-interest in blocks, not pixels. Then we compute the stride
1367     * between rows of blocks as the width in blocks times the width per
1368     * block, etc.
1369     */
1370    struct pipe_box box_blocks;
1371    u_box_pixels_to_blocks(&box_blocks, box, format);
1372 
1373    if (rsrc->image.layout.modifier ==
1374        DRM_FORMAT_MOD_ARM_16X16_BLOCK_U_INTERLEAVED) {
1375       transfer->base.stride = box_blocks.width * bytes_per_block;
1376       transfer->base.layer_stride = transfer->base.stride * box_blocks.height;
1377       transfer->map =
1378          ralloc_size(transfer, transfer->base.layer_stride * box->depth);
1379 
1380       if (usage & PIPE_MAP_READ)
1381          panfrost_load_tiled_images(transfer, rsrc);
1382 
1383       return transfer->map;
1384    } else {
1385       assert(rsrc->image.layout.modifier == DRM_FORMAT_MOD_LINEAR);
1386 
1387       /* Direct, persistent writes create holes in time for
1388        * caching... I don't know if this is actually possible but we
1389        * should still get it right */
1390 
1391       unsigned dpw = PIPE_MAP_DIRECTLY | PIPE_MAP_WRITE | PIPE_MAP_PERSISTENT;
1392 
1393       if ((usage & dpw) == dpw && rsrc->index_cache)
1394          return NULL;
1395 
1396       transfer->base.stride = rsrc->image.layout.slices[level].row_stride;
1397       transfer->base.layer_stride =
1398          panfrost_get_layer_stride(&rsrc->image.layout, level);
1399 
1400       /* By mapping direct-write, we're implicitly already
1401        * initialized (maybe), so be conservative */
1402 
1403       if (usage & PIPE_MAP_WRITE) {
1404          BITSET_SET(rsrc->valid.data, level);
1405          panfrost_minmax_cache_invalidate(
1406             rsrc->index_cache, transfer->base.box.x, transfer->base.box.width);
1407       }
1408 
1409       return bo->ptr.cpu + rsrc->image.layout.slices[level].offset +
1410              box->z * transfer->base.layer_stride +
1411              box_blocks.y * rsrc->image.layout.slices[level].row_stride +
1412              box_blocks.x * bytes_per_block;
1413    }
1414 }
1415 
1416 void
pan_resource_modifier_convert(struct panfrost_context * ctx,struct panfrost_resource * rsrc,uint64_t modifier,bool copy_resource,const char * reason)1417 pan_resource_modifier_convert(struct panfrost_context *ctx,
1418                               struct panfrost_resource *rsrc, uint64_t modifier,
1419                               bool copy_resource, const char *reason)
1420 {
1421    assert(!rsrc->modifier_constant);
1422 
1423    struct pipe_resource *tmp_prsrc = panfrost_resource_create_with_modifier(
1424       ctx->base.screen, &rsrc->base, modifier);
1425    struct panfrost_resource *tmp_rsrc = pan_resource(tmp_prsrc);
1426 
1427    if (copy_resource) {
1428       struct pipe_blit_info blit = {
1429          .dst.resource = &tmp_rsrc->base,
1430          .dst.format = tmp_rsrc->base.format,
1431          .src.resource = &rsrc->base,
1432          .src.format = rsrc->base.format,
1433          .mask = util_format_get_mask(tmp_rsrc->base.format),
1434          .filter = PIPE_TEX_FILTER_NEAREST,
1435       };
1436 
1437       /* data_valid is not valid until flushed */
1438       panfrost_flush_writer(ctx, rsrc, "AFBC/AFRC decompressing blit");
1439 
1440       for (int i = 0; i <= rsrc->base.last_level; i++) {
1441          if (BITSET_TEST(rsrc->valid.data, i)) {
1442             blit.dst.level = blit.src.level = i;
1443 
1444             u_box_3d(0, 0, 0, u_minify(rsrc->base.width0, i),
1445                      u_minify(rsrc->base.height0, i),
1446                      util_num_layers(&rsrc->base, i), &blit.dst.box);
1447             blit.src.box = blit.dst.box;
1448 
1449             panfrost_blit_no_afbc_legalization(&ctx->base, &blit);
1450          }
1451       }
1452 
1453       /* we lose track of tmp_rsrc after this point, and the BO migration
1454        * (from tmp_rsrc to rsrc) doesn't transfer the last_writer to rsrc
1455        */
1456       panfrost_flush_writer(ctx, tmp_rsrc, "AFBC/AFRC decompressing blit");
1457    }
1458 
1459    panfrost_bo_unreference(rsrc->bo);
1460 
1461    rsrc->bo = tmp_rsrc->bo;
1462    rsrc->image.data.base = rsrc->bo->ptr.gpu;
1463    panfrost_bo_reference(rsrc->bo);
1464 
1465    panfrost_resource_setup(ctx->base.screen, rsrc, modifier,
1466                            tmp_rsrc->base.format);
1467    /* panfrost_resource_setup will force the modifier to stay constant when
1468     * called with a specific modifier. We don't want that here, we want to
1469     * be able to convert back to another modifier if needed */
1470    rsrc->modifier_constant = false;
1471    pipe_resource_reference(&tmp_prsrc, NULL);
1472 }
1473 
1474 /* Validate that an AFBC/AFRC resource may be used as a particular format. If it
1475  * may not, decompress it on the fly. Failure to do so can produce wrong results
1476  * or invalid data faults when sampling or rendering to AFBC */
1477 
1478 void
pan_legalize_format(struct panfrost_context * ctx,struct panfrost_resource * rsrc,enum pipe_format format,bool write,bool discard)1479 pan_legalize_format(struct panfrost_context *ctx,
1480                     struct panfrost_resource *rsrc, enum pipe_format format,
1481                     bool write, bool discard)
1482 {
1483    struct panfrost_device *dev = pan_device(ctx->base.screen);
1484    enum pipe_format old_format = rsrc->base.format;
1485    enum pipe_format new_format = format;
1486    bool compatible = true;
1487 
1488    if (!drm_is_afbc(rsrc->image.layout.modifier) &&
1489        !drm_is_afrc(rsrc->image.layout.modifier))
1490       return;
1491 
1492    if (drm_is_afbc(rsrc->image.layout.modifier)) {
1493       compatible = (panfrost_afbc_format(dev->arch, old_format) ==
1494                     panfrost_afbc_format(dev->arch, new_format));
1495    } else if (drm_is_afrc(rsrc->image.layout.modifier)) {
1496       struct pan_afrc_format_info old_info =
1497          panfrost_afrc_get_format_info(old_format);
1498       struct pan_afrc_format_info new_info =
1499          panfrost_afrc_get_format_info(new_format);
1500       compatible = !memcmp(&old_info, &new_info, sizeof(old_info));
1501    }
1502 
1503    if (!compatible) {
1504       pan_resource_modifier_convert(
1505          ctx, rsrc, DRM_FORMAT_MOD_ARM_16X16_BLOCK_U_INTERLEAVED, !discard,
1506          drm_is_afbc(rsrc->image.layout.modifier)
1507             ? "Reinterpreting AFBC surface as incompatible format"
1508             : "Reinterpreting AFRC surface as incompatible format");
1509       return;
1510    }
1511 
1512    /* Can't write to AFBC-P resources */
1513    if (write && drm_is_afbc(rsrc->image.layout.modifier) &&
1514       (rsrc->image.layout.modifier & AFBC_FORMAT_MOD_SPARSE) == 0) {
1515       pan_resource_modifier_convert(
1516          ctx, rsrc, rsrc->image.layout.modifier | AFBC_FORMAT_MOD_SPARSE,
1517          !discard, "Legalizing resource to allow writing");
1518    }
1519 }
1520 
1521 static bool
panfrost_should_linear_convert(struct panfrost_context * ctx,struct panfrost_resource * prsrc,struct pipe_transfer * transfer)1522 panfrost_should_linear_convert(struct panfrost_context *ctx,
1523                                struct panfrost_resource *prsrc,
1524                                struct pipe_transfer *transfer)
1525 {
1526    if (prsrc->modifier_constant)
1527       return false;
1528 
1529    /* Overwriting the entire resource indicates streaming, for which
1530     * linear layout is most efficient due to the lack of expensive
1531     * conversion.
1532     *
1533     * For now we just switch to linear after a number of complete
1534     * overwrites to keep things simple, but we could do better.
1535     *
1536     * This mechanism is only implemented for 2D resources. This suffices
1537     * for video players, its intended use case.
1538     */
1539 
1540    bool entire_overwrite = panfrost_is_2d(prsrc) &&
1541                            prsrc->base.last_level == 0 &&
1542                            transfer->box.width == prsrc->base.width0 &&
1543                            transfer->box.height == prsrc->base.height0 &&
1544                            transfer->box.x == 0 && transfer->box.y == 0;
1545 
1546    if (entire_overwrite)
1547       ++prsrc->modifier_updates;
1548 
1549    if (prsrc->modifier_updates >= LAYOUT_CONVERT_THRESHOLD) {
1550       perf_debug(ctx, "Transitioning to linear due to streaming usage");
1551       return true;
1552    } else {
1553       return false;
1554    }
1555 }
1556 
1557 struct panfrost_bo *
panfrost_get_afbc_superblock_sizes(struct panfrost_context * ctx,struct panfrost_resource * rsrc,unsigned first_level,unsigned last_level,unsigned * out_offsets)1558 panfrost_get_afbc_superblock_sizes(struct panfrost_context *ctx,
1559                                    struct panfrost_resource *rsrc,
1560                                    unsigned first_level, unsigned last_level,
1561                                    unsigned *out_offsets)
1562 {
1563    struct panfrost_screen *screen = pan_screen(ctx->base.screen);
1564    struct panfrost_device *dev = pan_device(ctx->base.screen);
1565    struct panfrost_batch *batch;
1566    struct panfrost_bo *bo;
1567    unsigned metadata_size = 0;
1568 
1569    for (int level = first_level; level <= last_level; ++level) {
1570       struct pan_image_slice_layout *slice = &rsrc->image.layout.slices[level];
1571       unsigned sz = slice->afbc.nr_blocks * sizeof(struct pan_afbc_block_info);
1572       out_offsets[level - first_level] = metadata_size;
1573       metadata_size += sz;
1574    }
1575 
1576    panfrost_flush_batches_accessing_rsrc(ctx, rsrc, "AFBC before size flush");
1577    batch = panfrost_get_fresh_batch_for_fbo(ctx, "AFBC superblock sizes");
1578    bo = panfrost_bo_create(dev, metadata_size, 0, "AFBC superblock sizes");
1579    assert(bo);
1580 
1581    for (int level = first_level; level <= last_level; ++level) {
1582       unsigned offset = out_offsets[level - first_level];
1583       screen->vtbl.afbc_size(batch, rsrc, bo, offset, level);
1584    }
1585 
1586    panfrost_flush_batches_accessing_rsrc(ctx, rsrc, "AFBC after size flush");
1587 
1588    return bo;
1589 }
1590 
1591 void
panfrost_pack_afbc(struct panfrost_context * ctx,struct panfrost_resource * prsrc)1592 panfrost_pack_afbc(struct panfrost_context *ctx,
1593                    struct panfrost_resource *prsrc)
1594 {
1595    struct panfrost_screen *screen = pan_screen(ctx->base.screen);
1596    struct panfrost_device *dev = pan_device(ctx->base.screen);
1597    struct panfrost_bo *metadata_bo;
1598    unsigned metadata_offsets[PIPE_MAX_TEXTURE_LEVELS];
1599 
1600    uint64_t src_modifier = prsrc->image.layout.modifier;
1601    uint64_t dst_modifier =
1602       src_modifier & ~(AFBC_FORMAT_MOD_TILED | AFBC_FORMAT_MOD_SPARSE);
1603    bool is_tiled = src_modifier & AFBC_FORMAT_MOD_TILED;
1604    unsigned last_level = prsrc->base.last_level;
1605    struct pan_image_slice_layout slice_infos[PIPE_MAX_TEXTURE_LEVELS] = {0};
1606    unsigned total_size = 0;
1607 
1608    /* It doesn't make sense to pack everything if we need to unpack right
1609     * away to upload data to another level */
1610    for (int i = 0; i <= last_level; i++) {
1611       if (!BITSET_TEST(prsrc->valid.data, i))
1612          return;
1613    }
1614 
1615    metadata_bo = panfrost_get_afbc_superblock_sizes(ctx, prsrc, 0, last_level,
1616                                                     metadata_offsets);
1617    panfrost_bo_wait(metadata_bo, INT64_MAX, false);
1618 
1619    for (unsigned level = 0; level <= last_level; ++level) {
1620       struct pan_image_slice_layout *src_slice =
1621          &prsrc->image.layout.slices[level];
1622       struct pan_image_slice_layout *dst_slice = &slice_infos[level];
1623 
1624       unsigned width = u_minify(prsrc->base.width0, level);
1625       unsigned height = u_minify(prsrc->base.height0, level);
1626       unsigned src_stride =
1627          pan_afbc_stride_blocks(src_modifier, src_slice->row_stride);
1628       unsigned dst_stride =
1629          DIV_ROUND_UP(width, panfrost_afbc_superblock_width(dst_modifier));
1630       unsigned dst_height =
1631          DIV_ROUND_UP(height, panfrost_afbc_superblock_height(dst_modifier));
1632 
1633       uint32_t offset = 0;
1634       struct pan_afbc_block_info *meta =
1635          metadata_bo->ptr.cpu + metadata_offsets[level];
1636 
1637       for (unsigned y = 0, i = 0; y < dst_height; ++y) {
1638          for (unsigned x = 0; x < dst_stride; ++x, ++i) {
1639             unsigned idx = is_tiled ? get_morton_index(x, y, src_stride) : i;
1640             uint32_t size = meta[idx].size;
1641             meta[idx].offset = offset; /* write the start offset */
1642             offset += size;
1643          }
1644       }
1645 
1646       total_size = ALIGN_POT(total_size, pan_slice_align(dst_modifier));
1647       {
1648          dst_slice->afbc.stride = dst_stride;
1649          dst_slice->afbc.nr_blocks = dst_stride * dst_height;
1650          dst_slice->afbc.header_size =
1651             ALIGN_POT(dst_stride * dst_height * AFBC_HEADER_BYTES_PER_TILE,
1652                       pan_afbc_body_align(dst_modifier));
1653          dst_slice->afbc.body_size = offset;
1654          dst_slice->afbc.surface_stride = dst_slice->afbc.header_size + offset;
1655 
1656          dst_slice->offset = total_size;
1657          dst_slice->row_stride = dst_stride * AFBC_HEADER_BYTES_PER_TILE;
1658          dst_slice->surface_stride = dst_slice->afbc.surface_stride;
1659          dst_slice->size = dst_slice->afbc.surface_stride;
1660       }
1661       total_size += dst_slice->afbc.surface_stride;
1662    }
1663 
1664    unsigned new_size = ALIGN_POT(total_size, 4096); // FIXME
1665    unsigned old_size = panfrost_bo_size(prsrc->bo);
1666    unsigned ratio = 100 * new_size / old_size;
1667 
1668    if (ratio > screen->max_afbc_packing_ratio)
1669       return;
1670 
1671    perf_debug(ctx, "%i%%: %i KB -> %i KB\n", ratio, old_size / 1024,
1672               new_size / 1024);
1673 
1674    struct panfrost_bo *dst =
1675       panfrost_bo_create(dev, new_size, 0, "AFBC compact texture");
1676    assert(dst);
1677    struct panfrost_batch *batch =
1678       panfrost_get_fresh_batch_for_fbo(ctx, "AFBC compaction");
1679 
1680    for (unsigned level = 0; level <= last_level; ++level) {
1681       struct pan_image_slice_layout *slice = &slice_infos[level];
1682       screen->vtbl.afbc_pack(batch, prsrc, dst, slice, metadata_bo,
1683                              metadata_offsets[level], level);
1684       prsrc->image.layout.slices[level] = *slice;
1685    }
1686 
1687    panfrost_flush_batches_accessing_rsrc(ctx, prsrc, "AFBC compaction flush");
1688 
1689    prsrc->image.layout.modifier = dst_modifier;
1690    panfrost_bo_unreference(prsrc->bo);
1691    prsrc->bo = dst;
1692    prsrc->image.data.base = dst->ptr.gpu;
1693    panfrost_bo_unreference(metadata_bo);
1694 }
1695 
1696 static void
panfrost_ptr_unmap(struct pipe_context * pctx,struct pipe_transfer * transfer)1697 panfrost_ptr_unmap(struct pipe_context *pctx, struct pipe_transfer *transfer)
1698 {
1699    /* Gallium expects writeback here, so we tile */
1700 
1701    struct panfrost_context *ctx = pan_context(pctx);
1702    struct pipe_screen *screen = ctx->base.screen;
1703    struct panfrost_transfer *trans = pan_transfer(transfer);
1704    struct panfrost_resource *prsrc =
1705       (struct panfrost_resource *)transfer->resource;
1706    struct panfrost_device *dev = pan_device(pctx->screen);
1707 
1708    if (transfer->usage & PIPE_MAP_WRITE)
1709       prsrc->valid.crc = false;
1710 
1711    /* AFBC/AFRC will use a staging resource. `initialized` will be set when
1712     * the fragment job is created; this is deferred to prevent useless surface
1713     * reloads that can cascade into DATA_INVALID_FAULTs due to reading
1714     * malformed AFBC/AFRC data if uninitialized */
1715 
1716    if (trans->staging.rsrc) {
1717       if (transfer->usage & PIPE_MAP_WRITE) {
1718          if (panfrost_should_linear_convert(ctx, prsrc, transfer)) {
1719 
1720             panfrost_bo_unreference(prsrc->bo);
1721 
1722             panfrost_resource_setup(screen, prsrc, DRM_FORMAT_MOD_LINEAR,
1723                                     prsrc->image.layout.format);
1724 
1725             prsrc->bo = pan_resource(trans->staging.rsrc)->bo;
1726             prsrc->image.data.base = prsrc->bo->ptr.gpu;
1727             panfrost_bo_reference(prsrc->bo);
1728          } else {
1729             bool discard = panfrost_can_discard(&prsrc->base, &transfer->box,
1730                                                 transfer->usage);
1731             pan_legalize_format(ctx, prsrc, prsrc->image.layout.format, true,
1732                                 discard);
1733             pan_blit_from_staging(pctx, trans);
1734             panfrost_flush_batches_accessing_rsrc(
1735                ctx, pan_resource(trans->staging.rsrc),
1736                "AFBC write staging blit");
1737 
1738             if (pan_screen(pctx->screen)->force_afbc_packing) {
1739                if (panfrost_should_pack_afbc(dev, prsrc))
1740                   panfrost_pack_afbc(ctx, prsrc);
1741             }
1742          }
1743       }
1744 
1745       pipe_resource_reference(&trans->staging.rsrc, NULL);
1746    }
1747 
1748    /* Tiling will occur in software from a staging cpu buffer */
1749    if (trans->map) {
1750       struct panfrost_bo *bo = prsrc->bo;
1751 
1752       if (transfer->usage & PIPE_MAP_WRITE) {
1753          BITSET_SET(prsrc->valid.data, transfer->level);
1754 
1755          if (prsrc->image.layout.modifier ==
1756              DRM_FORMAT_MOD_ARM_16X16_BLOCK_U_INTERLEAVED) {
1757             if (panfrost_should_linear_convert(ctx, prsrc, transfer)) {
1758                panfrost_resource_setup(screen, prsrc, DRM_FORMAT_MOD_LINEAR,
1759                                        prsrc->image.layout.format);
1760                if (prsrc->image.layout.data_size > panfrost_bo_size(bo)) {
1761                   const char *label = bo->label;
1762                   panfrost_bo_unreference(bo);
1763                   bo = prsrc->bo = panfrost_bo_create(
1764                      dev, prsrc->image.layout.data_size, 0, label);
1765                   prsrc->image.data.base = prsrc->bo->ptr.gpu;
1766                   assert(bo);
1767                }
1768 
1769                util_copy_rect(
1770                   bo->ptr.cpu + prsrc->image.layout.slices[0].offset,
1771                   prsrc->base.format, prsrc->image.layout.slices[0].row_stride,
1772                   0, 0, transfer->box.width, transfer->box.height, trans->map,
1773                   transfer->stride, 0, 0);
1774             } else {
1775                panfrost_store_tiled_images(trans, prsrc);
1776             }
1777          }
1778       }
1779    }
1780 
1781    util_range_add(&prsrc->base, &prsrc->valid_buffer_range, transfer->box.x,
1782                   transfer->box.x + transfer->box.width);
1783 
1784    if (transfer->usage & PIPE_MAP_WRITE) {
1785       panfrost_minmax_cache_invalidate(prsrc->index_cache, transfer->box.x,
1786                                        transfer->box.width);
1787    }
1788 
1789    /* Derefence the resource */
1790    pipe_resource_reference(&transfer->resource, NULL);
1791 
1792    /* Transfer itself is RALLOCed at the moment */
1793    ralloc_free(transfer);
1794 }
1795 
1796 static void
panfrost_ptr_flush_region(struct pipe_context * pctx,struct pipe_transfer * transfer,const struct pipe_box * box)1797 panfrost_ptr_flush_region(struct pipe_context *pctx,
1798                           struct pipe_transfer *transfer,
1799                           const struct pipe_box *box)
1800 {
1801    struct panfrost_resource *rsc = pan_resource(transfer->resource);
1802 
1803    if (transfer->resource->target == PIPE_BUFFER) {
1804       util_range_add(&rsc->base, &rsc->valid_buffer_range,
1805                      transfer->box.x + box->x,
1806                      transfer->box.x + box->x + box->width);
1807    } else {
1808       BITSET_SET(rsc->valid.data, transfer->level);
1809    }
1810 }
1811 
1812 static void
panfrost_invalidate_resource(struct pipe_context * pctx,struct pipe_resource * prsrc)1813 panfrost_invalidate_resource(struct pipe_context *pctx,
1814                              struct pipe_resource *prsrc)
1815 {
1816    struct panfrost_context *ctx = pan_context(pctx);
1817    struct panfrost_batch *batch = panfrost_get_batch_for_fbo(ctx);
1818    struct panfrost_resource *rsrc = pan_resource(prsrc);
1819 
1820    rsrc->constant_stencil = true;
1821 
1822    /* Handle the glInvalidateFramebuffer case */
1823    if (batch->key.zsbuf && batch->key.zsbuf->texture == prsrc)
1824       batch->resolve &= ~PIPE_CLEAR_DEPTHSTENCIL;
1825 
1826    for (unsigned i = 0; i < batch->key.nr_cbufs; ++i) {
1827       struct pipe_surface *surf = batch->key.cbufs[i];
1828 
1829       if (surf && surf->texture == prsrc)
1830          batch->resolve &= ~(PIPE_CLEAR_COLOR0 << i);
1831    }
1832 }
1833 
1834 static enum pipe_format
panfrost_resource_get_internal_format(struct pipe_resource * rsrc)1835 panfrost_resource_get_internal_format(struct pipe_resource *rsrc)
1836 {
1837    struct panfrost_resource *prsrc = (struct panfrost_resource *)rsrc;
1838    return prsrc->image.layout.format;
1839 }
1840 
1841 void
panfrost_set_image_view_planes(struct pan_image_view * iview,struct pipe_resource * texture)1842 panfrost_set_image_view_planes(struct pan_image_view *iview,
1843                                struct pipe_resource *texture)
1844 {
1845    struct panfrost_resource *prsrc_plane = (struct panfrost_resource *)texture;
1846 
1847    for (int i = 0; i < MAX_IMAGE_PLANES && prsrc_plane; i++) {
1848       iview->planes[i] = &prsrc_plane->image;
1849       prsrc_plane = (struct panfrost_resource *)prsrc_plane->base.next;
1850    }
1851 }
1852 
1853 static bool
panfrost_generate_mipmap(struct pipe_context * pctx,struct pipe_resource * prsrc,enum pipe_format format,unsigned base_level,unsigned last_level,unsigned first_layer,unsigned last_layer)1854 panfrost_generate_mipmap(struct pipe_context *pctx, struct pipe_resource *prsrc,
1855                          enum pipe_format format, unsigned base_level,
1856                          unsigned last_level, unsigned first_layer,
1857                          unsigned last_layer)
1858 {
1859    struct panfrost_resource *rsrc = pan_resource(prsrc);
1860 
1861    perf_debug(pan_context(pctx), "Unoptimized mipmap generation");
1862 
1863    /* Generating a mipmap invalidates the written levels, so make that
1864     * explicit so we don't try to wallpaper them back and end up with
1865     * u_blitter recursion */
1866 
1867    assert(rsrc->bo);
1868    for (unsigned l = base_level + 1; l <= last_level; ++l)
1869       BITSET_CLEAR(rsrc->valid.data, l);
1870 
1871    /* Beyond that, we just delegate the hard stuff. */
1872 
1873    bool blit_res =
1874       util_gen_mipmap(pctx, prsrc, format, base_level, last_level, first_layer,
1875                       last_layer, PIPE_TEX_FILTER_LINEAR);
1876 
1877    return blit_res;
1878 }
1879 
1880 static void
panfrost_resource_set_stencil(struct pipe_resource * prsrc,struct pipe_resource * stencil)1881 panfrost_resource_set_stencil(struct pipe_resource *prsrc,
1882                               struct pipe_resource *stencil)
1883 {
1884    pan_resource(prsrc)->separate_stencil = pan_resource(stencil);
1885 }
1886 
1887 static struct pipe_resource *
panfrost_resource_get_stencil(struct pipe_resource * prsrc)1888 panfrost_resource_get_stencil(struct pipe_resource *prsrc)
1889 {
1890    if (!pan_resource(prsrc)->separate_stencil)
1891       return NULL;
1892 
1893    return &pan_resource(prsrc)->separate_stencil->base;
1894 }
1895 
1896 static const struct u_transfer_vtbl transfer_vtbl = {
1897    .resource_create = panfrost_resource_create,
1898    .resource_destroy = panfrost_resource_destroy,
1899    .transfer_map = panfrost_ptr_map,
1900    .transfer_unmap = panfrost_ptr_unmap,
1901    .transfer_flush_region = panfrost_ptr_flush_region,
1902    .get_internal_format = panfrost_resource_get_internal_format,
1903    .set_stencil = panfrost_resource_set_stencil,
1904    .get_stencil = panfrost_resource_get_stencil,
1905 };
1906 
1907 void
panfrost_resource_screen_init(struct pipe_screen * pscreen)1908 panfrost_resource_screen_init(struct pipe_screen *pscreen)
1909 {
1910    pscreen->resource_create_with_modifiers =
1911       panfrost_resource_create_with_modifiers;
1912    pscreen->resource_create = u_transfer_helper_resource_create;
1913    pscreen->resource_destroy = u_transfer_helper_resource_destroy;
1914    pscreen->resource_from_handle = panfrost_resource_from_handle;
1915    pscreen->resource_get_handle = panfrost_resource_get_handle;
1916    pscreen->resource_get_param = panfrost_resource_get_param;
1917    pscreen->transfer_helper = u_transfer_helper_create(
1918       &transfer_vtbl,
1919       U_TRANSFER_HELPER_SEPARATE_Z32S8 | U_TRANSFER_HELPER_MSAA_MAP);
1920 }
1921 void
panfrost_resource_screen_destroy(struct pipe_screen * pscreen)1922 panfrost_resource_screen_destroy(struct pipe_screen *pscreen)
1923 {
1924    u_transfer_helper_destroy(pscreen->transfer_helper);
1925 }
1926 
1927 void
panfrost_resource_context_init(struct pipe_context * pctx)1928 panfrost_resource_context_init(struct pipe_context *pctx)
1929 {
1930    pctx->buffer_map = u_transfer_helper_transfer_map;
1931    pctx->buffer_unmap = u_transfer_helper_transfer_unmap;
1932    pctx->texture_map = u_transfer_helper_transfer_map;
1933    pctx->texture_unmap = u_transfer_helper_transfer_unmap;
1934    pctx->create_surface = panfrost_create_surface;
1935    pctx->surface_destroy = panfrost_surface_destroy;
1936    pctx->resource_copy_region = util_resource_copy_region;
1937    pctx->blit = panfrost_blit;
1938    pctx->generate_mipmap = panfrost_generate_mipmap;
1939    pctx->flush_resource = panfrost_flush_resource;
1940    pctx->invalidate_resource = panfrost_invalidate_resource;
1941    pctx->transfer_flush_region = u_transfer_helper_transfer_flush_region;
1942    pctx->buffer_subdata = u_default_buffer_subdata;
1943    pctx->texture_subdata = u_default_texture_subdata;
1944    pctx->clear_buffer = u_default_clear_buffer;
1945    pctx->clear_render_target = panfrost_clear_render_target;
1946    pctx->clear_depth_stencil = panfrost_clear_depth_stencil;
1947 }
1948