src/gallium/drivers/panfrost/pan_screen.h

   1 /**************************************************************************
   2  *
   3  * Copyright 2018-2019 Alyssa Rosenzweig
   4  * Copyright 2018-2019 Collabora, Ltd.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 #ifndef PAN_SCREEN_H
  30 #define PAN_SCREEN_H
  31
  32 #include <xf86drm.h>
  33 #include "pipe/p_screen.h"
  34 #include "pipe/p_defines.h"
  35 #include "renderonly/renderonly.h"
  36 #include "util/u_dynarray.h"
  37 #include "util/bitset.h"
  38
  39 #include <panfrost-misc.h>
  40 #include "pan_allocate.h"
  41
  42 struct panfrost_batch;
  43 struct panfrost_context;
  44 struct panfrost_resource;
  45 struct panfrost_screen;
  46
  47 /* Driver limits */
  48 #define PAN_MAX_CONST_BUFFERS 16
  49
  50 /* Flags for allocated memory */
  51
  52 /* This memory region is executable */
  53 #define PAN_ALLOCATE_EXECUTE            (1 << 0)
  54
  55 /* This memory region should be lazily allocated and grow-on-page-fault. Must
  56  * be used in conjunction with INVISIBLE */
  57 #define PAN_ALLOCATE_GROWABLE           (1 << 1)
  58
  59 /* This memory region should not be mapped to the CPU */
  60 #define PAN_ALLOCATE_INVISIBLE          (1 << 2)
  61
  62 /* This memory region will be used for varyings and needs to have the cache
  63  * bits twiddled accordingly */
  64 #define PAN_ALLOCATE_COHERENT_LOCAL     (1 << 3)
  65
  66 /* This region may not be used immediately and will not mmap on allocate
  67  * (semantically distinct from INVISIBLE, which cannot never be mmaped) */
  68 #define PAN_ALLOCATE_DELAY_MMAP         (1 << 4)
  69
  70 /* Transient slab size. This is a balance between fragmentation against cache
  71  * locality and ease of bookkeeping */
  72
  73 #define TRANSIENT_SLAB_PAGES (32) /* 128kb */
  74 #define TRANSIENT_SLAB_SIZE (4096 * TRANSIENT_SLAB_PAGES)
  75
  76 /* Maximum number of transient slabs so we don't need dynamic arrays. Most
  77  * interesting Mali boards are 4GB RAM max, so if the entire RAM was filled
  78  * with transient slabs, you could never exceed (4GB / TRANSIENT_SLAB_SIZE)
  79  * allocations anyway. By capping, we can use a fixed-size bitset for tracking
  80  * free slabs, eliminating quite a bit of complexity. We can pack the free
  81  * state of 8 slabs into a single byte, so for 128kb transient slabs the bitset
  82  * occupies a cheap 4kb of memory */
  83
  84 #define MAX_TRANSIENT_SLABS (1024*1024 / TRANSIENT_SLAB_PAGES)
  85
  86 /* How many power-of-two levels in the BO cache do we want? 2^12
  87  * minimum chosen as it is the page size that all allocations are
  88  * rounded to */
  89
  90 #define MIN_BO_CACHE_BUCKET (12) /* 2^12 = 4KB */
  91 #define MAX_BO_CACHE_BUCKET (22) /* 2^22 = 4MB */
  92
  93 /* Fencepost problem, hence the off-by-one */
  94 #define NR_BO_CACHE_BUCKETS (MAX_BO_CACHE_BUCKET - MIN_BO_CACHE_BUCKET + 1)
  95
  96 struct panfrost_screen {
  97         struct pipe_screen base;
  98         int fd;
  99
 100         /* Properties of the GPU in use */
 101         unsigned gpu_id;
 102         bool require_sfbd;
 103
 104         drmVersionPtr kernel_version;
 105
 106         struct renderonly *ro;
 107
 108         pthread_mutex_t transient_lock;
 109
 110         /* Transient memory management is based on borrowing fixed-size slabs
 111          * off the screen (loaning them out to the batch). Dynamic array
 112          * container of panfrost_bo */
 113
 114         struct util_dynarray transient_bo;
 115
 116         /* Set of free transient BOs */
 117         BITSET_DECLARE(free_transient, MAX_TRANSIENT_SLABS);
 118
 119         pthread_mutex_t bo_cache_lock;
 120
 121         /* The BO cache is a set of buckets with power-of-two sizes ranging
 122          * from 2^12 (4096, the page size) to 2^(12 + MAX_BO_CACHE_BUCKETS).
 123          * Each bucket is a linked list of free panfrost_bo objects. */
 124
 125         struct list_head bo_cache[NR_BO_CACHE_BUCKETS];
 126 };
 127
 128 static inline struct panfrost_screen *
 129 pan_screen(struct pipe_screen *p)
 130 {
 131         return (struct panfrost_screen *)p;
 132 }
 133
 134 /* Get a transient BO off the screen given a
 135  * particular index */
 136
 137 static inline struct panfrost_bo *
 138 pan_bo_for_index(struct panfrost_screen *screen, unsigned index)
 139 {
 140         return *(util_dynarray_element(&screen->transient_bo,
 141                                 struct panfrost_bo *, index));
 142 }
 143
 144 void
 145 panfrost_drm_allocate_slab(struct panfrost_screen *screen,
 146                            struct panfrost_memory *mem,
 147                            size_t pages,
 148                            bool same_va,
 149                            int extra_flags,
 150                            int commit_count,
 151                            int extent);
 152 void
 153 panfrost_drm_free_slab(struct panfrost_screen *screen,
 154                        struct panfrost_memory *mem);
 155 struct panfrost_bo *
 156 panfrost_drm_create_bo(struct panfrost_screen *screen, size_t size,
 157                        uint32_t flags);
 158 void
 159 panfrost_drm_mmap_bo(struct panfrost_screen *screen, struct panfrost_bo *bo);
 160 void
 161 panfrost_drm_release_bo(struct panfrost_screen *screen, struct panfrost_bo *bo, bool cacheable);
 162 struct panfrost_bo *
 163 panfrost_drm_import_bo(struct panfrost_screen *screen, int fd);
 164 int
 165 panfrost_drm_export_bo(struct panfrost_screen *screen, const struct panfrost_bo *bo);
 166 int
 167 panfrost_drm_submit_vs_fs_batch(struct panfrost_batch *batch, bool has_draws);
 168 void
 169 panfrost_drm_force_flush_fragment(struct panfrost_context *ctx,
 170                                   struct pipe_fence_handle **fence);
 171 unsigned
 172 panfrost_drm_query_gpu_version(struct panfrost_screen *screen);
 173 int
 174 panfrost_drm_init_context(struct panfrost_context *ctx);
 175 void
 176 panfrost_drm_fence_reference(struct pipe_screen *screen,
 177                              struct pipe_fence_handle **ptr,
 178                              struct pipe_fence_handle *fence);
 179 boolean
 180 panfrost_drm_fence_finish(struct pipe_screen *pscreen,
 181                           struct pipe_context *ctx,
 182                           struct pipe_fence_handle *fence,
 183                           uint64_t timeout);
 184 struct panfrost_bo *
 185 panfrost_bo_cache_fetch(
 186                 struct panfrost_screen *screen,
 187                 size_t size, uint32_t flags);
 188
 189 bool
 190 panfrost_bo_cache_put(
 191                 struct panfrost_screen *screen,
 192                 struct panfrost_bo *bo);
 193
 194 void
 195 panfrost_bo_cache_evict_all(
 196                 struct panfrost_screen *screen);
 197
 198 #endif /* PAN_SCREEN_H */