src/gallium/drivers/radeon/radeon_winsys.h

   1 /*
   2  * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
   3  * Copyright 2010 Marek Olšák <maraeo@gmail.com>
   4  * Copyright 2018 Advanced Micro Devices, Inc.
   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 "Software"),
   9  * to deal in the Software without restriction, including without limitation
  10  * on the rights to use, copy, modify, merge, publish, distribute, sub
  11  * license, and/or sell copies of the Software, and to permit persons to whom
  12  * the Software is furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the next
  15  * paragraph) shall be included in all copies or substantial portions of the
  16  * Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  20  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  21  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  22  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  23  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  24  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
  25
  26 #ifndef RADEON_WINSYS_H
  27 #define RADEON_WINSYS_H
  28
  29 /* The public winsys interface header for the radeon driver. */
  30
  31 /* Whether the next IB can start immediately and not wait for draws and
  32  * dispatches from the current IB to finish. */
  33 #define RADEON_FLUSH_START_NEXT_GFX_IB_NOW      (1u << 31)
  34
  35 #define RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW \
  36         (PIPE_FLUSH_ASYNC | RADEON_FLUSH_START_NEXT_GFX_IB_NOW)
  37
  38 #include "pipebuffer/pb_buffer.h"
  39
  40 #include "amd/common/ac_gpu_info.h"
  41 #include "amd/common/ac_surface.h"
  42
  43 /* Tiling flags. */
  44 enum radeon_bo_layout {
  45     RADEON_LAYOUT_LINEAR = 0,
  46     RADEON_LAYOUT_TILED,
  47     RADEON_LAYOUT_SQUARETILED,
  48
  49     RADEON_LAYOUT_UNKNOWN
  50 };
  51
  52 enum radeon_bo_domain { /* bitfield */
  53     RADEON_DOMAIN_GTT  = 2,
  54     RADEON_DOMAIN_VRAM = 4,
  55     RADEON_DOMAIN_VRAM_GTT = RADEON_DOMAIN_VRAM | RADEON_DOMAIN_GTT,
  56     RADEON_DOMAIN_GDS = 8,
  57     RADEON_DOMAIN_OA = 16,
  58 };
  59
  60 enum radeon_bo_flag { /* bitfield */
  61     RADEON_FLAG_GTT_WC =        (1 << 0),
  62     RADEON_FLAG_NO_CPU_ACCESS = (1 << 1),
  63     RADEON_FLAG_NO_SUBALLOC =   (1 << 2),
  64     RADEON_FLAG_SPARSE =        (1 << 3),
  65     RADEON_FLAG_NO_INTERPROCESS_SHARING = (1 << 4),
  66     RADEON_FLAG_READ_ONLY =     (1 << 5),
  67     RADEON_FLAG_32BIT =    (1 << 6),
  68 };
  69
  70 enum radeon_dependency_flag {
  71     /* Add the dependency to the parallel compute IB only. */
  72     RADEON_DEPENDENCY_PARALLEL_COMPUTE_ONLY = 1 << 0,
  73
  74     /* Instead of waiting for a job to finish execution, the dependency will
  75      * be signaled when the job starts execution.
  76      */
  77     RADEON_DEPENDENCY_START_FENCE = 1 << 1,
  78 };
  79
  80 enum radeon_bo_usage { /* bitfield */
  81     RADEON_USAGE_READ = 2,
  82     RADEON_USAGE_WRITE = 4,
  83     RADEON_USAGE_READWRITE = RADEON_USAGE_READ | RADEON_USAGE_WRITE,
  84
  85     /* The winsys ensures that the CS submission will be scheduled after
  86      * previously flushed CSs referencing this BO in a conflicting way.
  87      */
  88     RADEON_USAGE_SYNCHRONIZED = 8
  89 };
  90
  91 enum radeon_transfer_flags {
  92    /* Indicates that the caller will unmap the buffer.
  93     *
  94     * Not unmapping buffers is an important performance optimization for
  95     * OpenGL (avoids kernel overhead for frequently mapped buffers).
  96     */
  97    RADEON_TRANSFER_TEMPORARY = (PIPE_TRANSFER_DRV_PRV << 0),
  98 };
  99
 100 #define RADEON_SPARSE_PAGE_SIZE (64 * 1024)
 101
 102 enum ring_type {
 103     RING_GFX = 0,
 104     RING_COMPUTE,
 105     RING_DMA,
 106     RING_UVD,
 107     RING_VCE,
 108     RING_UVD_ENC,
 109     RING_VCN_DEC,
 110     RING_VCN_ENC,
 111     RING_VCN_JPEG,
 112     RING_LAST,
 113 };
 114
 115 enum radeon_value_id {
 116     RADEON_REQUESTED_VRAM_MEMORY,
 117     RADEON_REQUESTED_GTT_MEMORY,
 118     RADEON_MAPPED_VRAM,
 119     RADEON_MAPPED_GTT,
 120     RADEON_BUFFER_WAIT_TIME_NS,
 121     RADEON_NUM_MAPPED_BUFFERS,
 122     RADEON_TIMESTAMP,
 123     RADEON_NUM_GFX_IBS,
 124     RADEON_NUM_SDMA_IBS,
 125     RADEON_GFX_BO_LIST_COUNTER, /* number of BOs submitted in gfx IBs */
 126     RADEON_GFX_IB_SIZE_COUNTER,
 127     RADEON_NUM_BYTES_MOVED,
 128     RADEON_NUM_EVICTIONS,
 129     RADEON_NUM_VRAM_CPU_PAGE_FAULTS,
 130     RADEON_VRAM_USAGE,
 131     RADEON_VRAM_VIS_USAGE,
 132     RADEON_GTT_USAGE,
 133     RADEON_GPU_TEMPERATURE, /* DRM 2.42.0 */
 134     RADEON_CURRENT_SCLK,
 135     RADEON_CURRENT_MCLK,
 136     RADEON_CS_THREAD_TIME,
 137 };
 138
 139 enum radeon_bo_priority {
 140     /* Each group of two has the same priority. */
 141     RADEON_PRIO_FENCE = 0,
 142     RADEON_PRIO_TRACE,
 143
 144     RADEON_PRIO_SO_FILLED_SIZE = 2,
 145     RADEON_PRIO_QUERY,
 146
 147     RADEON_PRIO_IB1 = 4, /* main IB submitted to the kernel */
 148     RADEON_PRIO_IB2, /* IB executed with INDIRECT_BUFFER */
 149
 150     RADEON_PRIO_DRAW_INDIRECT = 6,
 151     RADEON_PRIO_INDEX_BUFFER,
 152
 153     RADEON_PRIO_CP_DMA = 8,
 154     RADEON_PRIO_BORDER_COLORS,
 155
 156     RADEON_PRIO_CONST_BUFFER = 10,
 157     RADEON_PRIO_DESCRIPTORS,
 158
 159     RADEON_PRIO_SAMPLER_BUFFER = 12,
 160     RADEON_PRIO_VERTEX_BUFFER,
 161
 162     RADEON_PRIO_SHADER_RW_BUFFER = 14,
 163     RADEON_PRIO_COMPUTE_GLOBAL,
 164
 165     RADEON_PRIO_SAMPLER_TEXTURE = 16,
 166     RADEON_PRIO_SHADER_RW_IMAGE,
 167
 168     RADEON_PRIO_SAMPLER_TEXTURE_MSAA = 18,
 169     RADEON_PRIO_COLOR_BUFFER,
 170
 171     RADEON_PRIO_DEPTH_BUFFER = 20,
 172
 173     RADEON_PRIO_COLOR_BUFFER_MSAA = 22,
 174
 175     RADEON_PRIO_DEPTH_BUFFER_MSAA = 24,
 176
 177     RADEON_PRIO_SEPARATE_META = 26,
 178     RADEON_PRIO_SHADER_BINARY, /* the hw can't hide instruction cache misses */
 179
 180     RADEON_PRIO_SHADER_RINGS = 28,
 181
 182     RADEON_PRIO_SCRATCH_BUFFER = 30,
 183     /* 31 is the maximum value */
 184 };
 185
 186 struct winsys_handle;
 187 struct radeon_winsys_ctx;
 188
 189 struct radeon_cmdbuf_chunk {
 190     unsigned cdw;  /* Number of used dwords. */
 191     unsigned max_dw; /* Maximum number of dwords. */
 192     uint32_t *buf; /* The base pointer of the chunk. */
 193 };
 194
 195 struct radeon_cmdbuf {
 196     struct radeon_cmdbuf_chunk current;
 197     struct radeon_cmdbuf_chunk *prev;
 198     unsigned                      num_prev; /* Number of previous chunks. */
 199     unsigned                      max_prev; /* Space in array pointed to by prev. */
 200     unsigned                      prev_dw; /* Total number of dwords in previous chunks. */
 201
 202     /* Memory usage of the buffer list. These are always 0 for preamble IBs. */
 203     uint64_t                      used_vram;
 204     uint64_t                      used_gart;
 205     uint64_t                      gpu_address;
 206 };
 207
 208 /* Tiling info for display code, DRI sharing, and other data. */
 209 struct radeon_bo_metadata {
 210     /* Tiling flags describing the texture layout for display code
 211      * and DRI sharing.
 212      */
 213     union {
 214         struct {
 215             enum radeon_bo_layout   microtile;
 216             enum radeon_bo_layout   macrotile;
 217             unsigned                pipe_config;
 218             unsigned                bankw;
 219             unsigned                bankh;
 220             unsigned                tile_split;
 221             unsigned                mtilea;
 222             unsigned                num_banks;
 223             unsigned                stride;
 224             bool                    scanout;
 225         } legacy;
 226
 227         struct {
 228             /* surface flags */
 229             unsigned swizzle_mode:5;
 230
 231             /* DCC flags */
 232             /* [31:8]: max offset = 4GB - 256; 0 = DCC disabled */
 233             unsigned dcc_offset_256B:24;
 234             unsigned dcc_pitch_max:14;   /* (mip chain pitch - 1) for DCN */
 235             unsigned dcc_independent_64B:1;
 236         } gfx9;
 237     } u;
 238
 239     /* Additional metadata associated with the buffer, in bytes.
 240      * The maximum size is 64 * 4. This is opaque for the winsys & kernel.
 241      * Supported by amdgpu only.
 242      */
 243     uint32_t                size_metadata;
 244     uint32_t                metadata[64];
 245 };
 246
 247 enum radeon_feature_id {
 248     RADEON_FID_R300_HYPERZ_ACCESS,     /* ZMask + HiZ */
 249     RADEON_FID_R300_CMASK_ACCESS,
 250 };
 251
 252 struct radeon_bo_list_item {
 253     uint64_t bo_size;
 254     uint64_t vm_address;
 255     uint32_t priority_usage; /* mask of (1 << RADEON_PRIO_*) */
 256 };
 257
 258 struct radeon_winsys {
 259     /**
 260      * The screen object this winsys was created for
 261      */
 262     struct pipe_screen *screen;
 263
 264     /**
 265      * Decrement the winsys reference count.
 266      *
 267      * \param ws  The winsys this function is called for.
 268      * \return    True if the winsys and screen should be destroyed.
 269      */
 270     bool (*unref)(struct radeon_winsys *ws);
 271
 272     /**
 273      * Destroy this winsys.
 274      *
 275      * \param ws        The winsys this function is called from.
 276      */
 277     void (*destroy)(struct radeon_winsys *ws);
 278
 279     /**
 280      * Query an info structure from winsys.
 281      *
 282      * \param ws        The winsys this function is called from.
 283      * \param info      Return structure
 284      */
 285     void (*query_info)(struct radeon_winsys *ws,
 286                        struct radeon_info *info);
 287
 288     /**
 289      * A hint for the winsys that it should pin its execution threads to
 290      * a group of cores sharing a specific L3 cache if the CPU has multiple
 291      * L3 caches. This is needed for good multithreading performance on
 292      * AMD Zen CPUs.
 293      */
 294     void (*pin_threads_to_L3_cache)(struct radeon_winsys *ws, unsigned cache);
 295
 296     /**************************************************************************
 297      * Buffer management. Buffer attributes are mostly fixed over its lifetime.
 298      *
 299      * Remember that gallium gets to choose the interface it needs, and the
 300      * window systems must then implement that interface (rather than the
 301      * other way around...).
 302      *************************************************************************/
 303
 304     /**
 305      * Create a buffer object.
 306      *
 307      * \param ws        The winsys this function is called from.
 308      * \param size      The size to allocate.
 309      * \param alignment An alignment of the buffer in memory.
 310      * \param use_reusable_pool Whether the cache buffer manager should be used.
 311      * \param domain    A bitmask of the RADEON_DOMAIN_* flags.
 312      * \return          The created buffer object.
 313      */
 314     struct pb_buffer *(*buffer_create)(struct radeon_winsys *ws,
 315                                        uint64_t size,
 316                                        unsigned alignment,
 317                                        enum radeon_bo_domain domain,
 318                                        enum radeon_bo_flag flags);
 319
 320     /**
 321      * Map the entire data store of a buffer object into the client's address
 322      * space.
 323      *
 324      * Callers are expected to unmap buffers again if and only if the
 325      * RADEON_TRANSFER_TEMPORARY flag is set in \p usage.
 326      *
 327      * \param buf       A winsys buffer object to map.
 328      * \param cs        A command stream to flush if the buffer is referenced by it.
 329      * \param usage     A bitmask of the PIPE_TRANSFER_* and RADEON_TRANSFER_* flags.
 330      * \return          The pointer at the beginning of the buffer.
 331      */
 332     void *(*buffer_map)(struct pb_buffer *buf,
 333                         struct radeon_cmdbuf *cs,
 334                         enum pipe_transfer_usage usage);
 335
 336     /**
 337      * Unmap a buffer object from the client's address space.
 338      *
 339      * \param buf       A winsys buffer object to unmap.
 340      */
 341     void (*buffer_unmap)(struct pb_buffer *buf);
 342
 343     /**
 344      * Wait for the buffer and return true if the buffer is not used
 345      * by the device.
 346      *
 347      * The timeout of 0 will only return the status.
 348      * The timeout of PIPE_TIMEOUT_INFINITE will always wait until the buffer
 349      * is idle.
 350      */
 351     bool (*buffer_wait)(struct pb_buffer *buf, uint64_t timeout,
 352                         enum radeon_bo_usage usage);
 353
 354     /**
 355      * Return buffer metadata.
 356      * (tiling info for display code, DRI sharing, and other data)
 357      *
 358      * \param buf       A winsys buffer object to get the flags from.
 359      * \param md        Metadata
 360      */
 361     void (*buffer_get_metadata)(struct pb_buffer *buf,
 362                                 struct radeon_bo_metadata *md);
 363
 364     /**
 365      * Set buffer metadata.
 366      * (tiling info for display code, DRI sharing, and other data)
 367      *
 368      * \param buf       A winsys buffer object to set the flags for.
 369      * \param md        Metadata
 370      */
 371     void (*buffer_set_metadata)(struct pb_buffer *buf,
 372                                 struct radeon_bo_metadata *md);
 373
 374     /**
 375      * Get a winsys buffer from a winsys handle. The internal structure
 376      * of the handle is platform-specific and only a winsys should access it.
 377      *
 378      * \param ws        The winsys this function is called from.
 379      * \param whandle   A winsys handle pointer as was received from a state
 380      *                  tracker.
 381      * \param stride    The returned buffer stride in bytes.
 382      */
 383     struct pb_buffer *(*buffer_from_handle)(struct radeon_winsys *ws,
 384                                             struct winsys_handle *whandle,
 385                                             unsigned vm_alignment,
 386                                             unsigned *stride, unsigned *offset);
 387
 388     /**
 389      * Get a winsys buffer from a user pointer. The resulting buffer can't
 390      * be exported. Both pointer and size must be page aligned.
 391      *
 392      * \param ws        The winsys this function is called from.
 393      * \param pointer   User pointer to turn into a buffer object.
 394      * \param Size      Size in bytes for the new buffer.
 395      */
 396     struct pb_buffer *(*buffer_from_ptr)(struct radeon_winsys *ws,
 397                                          void *pointer, uint64_t size);
 398
 399     /**
 400      * Whether the buffer was created from a user pointer.
 401      *
 402      * \param buf       A winsys buffer object
 403      * \return          whether \p buf was created via buffer_from_ptr
 404      */
 405     bool (*buffer_is_user_ptr)(struct pb_buffer *buf);
 406
 407     /** Whether the buffer was suballocated. */
 408     bool (*buffer_is_suballocated)(struct pb_buffer *buf);
 409
 410     /**
 411      * Get a winsys handle from a winsys buffer. The internal structure
 412      * of the handle is platform-specific and only a winsys should access it.
 413      *
 414      * \param ws        The winsys instance for which the handle is to be valid
 415      * \param buf       A winsys buffer object to get the handle from.
 416      * \param whandle   A winsys handle pointer.
 417      * \param stride    A stride of the buffer in bytes, for texturing.
 418      * \return          true on success.
 419      */
 420     bool (*buffer_get_handle)(struct radeon_winsys *ws,
 421                               struct pb_buffer *buf,
 422                               unsigned stride, unsigned offset,
 423                               unsigned slice_size,
 424                               struct winsys_handle *whandle);
 425
 426     /**
 427      * Change the commitment of a (64KB-page aligned) region of the given
 428      * sparse buffer.
 429      *
 430      * \warning There is no automatic synchronization with command submission.
 431      *
 432      * \note Only implemented by the amdgpu winsys.
 433      *
 434      * \return false on out of memory or other failure, true on success.
 435      */
 436     bool (*buffer_commit)(struct pb_buffer *buf,
 437                           uint64_t offset, uint64_t size,
 438                           bool commit);
 439
 440     /**
 441      * Return the virtual address of a buffer.
 442      *
 443      * When virtual memory is not in use, this is the offset relative to the
 444      * relocation base (non-zero for sub-allocated buffers).
 445      *
 446      * \param buf       A winsys buffer object
 447      * \return          virtual address
 448      */
 449     uint64_t (*buffer_get_virtual_address)(struct pb_buffer *buf);
 450
 451     /**
 452      * Return the offset of this buffer relative to the relocation base.
 453      * This is only non-zero for sub-allocated buffers.
 454      *
 455      * This is only supported in the radeon winsys, since amdgpu uses virtual
 456      * addresses in submissions even for the video engines.
 457      *
 458      * \param buf      A winsys buffer object
 459      * \return         the offset for relocations
 460      */
 461     unsigned (*buffer_get_reloc_offset)(struct pb_buffer *buf);
 462
 463     /**
 464      * Query the initial placement of the buffer from the kernel driver.
 465      */
 466     enum radeon_bo_domain (*buffer_get_initial_domain)(struct pb_buffer *buf);
 467
 468     /**************************************************************************
 469      * Command submission.
 470      *
 471      * Each pipe context should create its own command stream and submit
 472      * commands independently of other contexts.
 473      *************************************************************************/
 474
 475     /**
 476      * Create a command submission context.
 477      * Various command streams can be submitted to the same context.
 478      */
 479     struct radeon_winsys_ctx *(*ctx_create)(struct radeon_winsys *ws);
 480
 481     /**
 482      * Destroy a context.
 483      */
 484     void (*ctx_destroy)(struct radeon_winsys_ctx *ctx);
 485
 486     /**
 487      * Query a GPU reset status.
 488      */
 489     enum pipe_reset_status (*ctx_query_reset_status)(struct radeon_winsys_ctx *ctx);
 490
 491     /**
 492      * Create a command stream.
 493      *
 494      * \param ctx       The submission context
 495      * \param ring_type The ring type (GFX, DMA, UVD)
 496      * \param flush     Flush callback function associated with the command stream.
 497      * \param user      User pointer that will be passed to the flush callback.
 498      */
 499     struct radeon_cmdbuf *(*cs_create)(struct radeon_winsys_ctx *ctx,
 500                                        enum ring_type ring_type,
 501                                        void (*flush)(void *ctx, unsigned flags,
 502                                                      struct pipe_fence_handle **fence),
 503                                        void *flush_ctx,
 504                                        bool stop_exec_on_failure);
 505
 506     /**
 507      * Add a parallel compute IB to a gfx IB. It will share the buffer list
 508      * and fence dependencies with the gfx IB. The gfx flush call will submit
 509      * both IBs at the same time.
 510      *
 511      * The compute IB doesn't have an output fence, so the primary IB has
 512      * to use a wait packet for synchronization.
 513      *
 514      * The returned IB is only a stream for writing packets to the new
 515      * IB. Calling other winsys functions with it is not allowed, not even
 516      * "cs_destroy". Use the gfx IB instead.
 517      *
 518      * \param cs              Gfx IB
 519      */
 520     struct radeon_cmdbuf *(*cs_add_parallel_compute_ib)(struct radeon_cmdbuf *cs,
 521                                                         bool uses_gds_ordered_append);
 522
 523     /**
 524      * Destroy a command stream.
 525      *
 526      * \param cs        A command stream to destroy.
 527      */
 528     void (*cs_destroy)(struct radeon_cmdbuf *cs);
 529
 530     /**
 531      * Add a buffer. Each buffer used by a CS must be added using this function.
 532      *
 533      * \param cs      Command stream
 534      * \param buf     Buffer
 535      * \param usage   Whether the buffer is used for read and/or write.
 536      * \param domain  Bitmask of the RADEON_DOMAIN_* flags.
 537      * \param priority  A higher number means a greater chance of being
 538      *                  placed in the requested domain. 15 is the maximum.
 539      * \return Buffer index.
 540      */
 541     unsigned (*cs_add_buffer)(struct radeon_cmdbuf *cs,
 542                              struct pb_buffer *buf,
 543                              enum radeon_bo_usage usage,
 544                              enum radeon_bo_domain domain,
 545                              enum radeon_bo_priority priority);
 546
 547     /**
 548      * Return the index of an already-added buffer.
 549      *
 550      * Not supported on amdgpu. Drivers with GPUVM should not care about
 551      * buffer indices.
 552      *
 553      * \param cs        Command stream
 554      * \param buf       Buffer
 555      * \return          The buffer index, or -1 if the buffer has not been added.
 556      */
 557     int (*cs_lookup_buffer)(struct radeon_cmdbuf *cs,
 558                             struct pb_buffer *buf);
 559
 560     /**
 561      * Return true if there is enough memory in VRAM and GTT for the buffers
 562      * added so far. If the validation fails, all buffers which have
 563      * been added since the last call of cs_validate will be removed and
 564      * the CS will be flushed (provided there are still any buffers).
 565      *
 566      * \param cs        A command stream to validate.
 567      */
 568     bool (*cs_validate)(struct radeon_cmdbuf *cs);
 569
 570     /**
 571      * Check whether the given number of dwords is available in the IB.
 572      * Optionally chain a new chunk of the IB if necessary and supported.
 573      *
 574      * \param cs        A command stream.
 575      * \param dw        Number of CS dwords requested by the caller.
 576      * \param force_chaining  Chain the IB into a new buffer now to discard
 577      *                        the CP prefetch cache (to emulate PKT3_REWIND)
 578      * \return true if there is enough space
 579      */
 580     bool (*cs_check_space)(struct radeon_cmdbuf *cs, unsigned dw,
 581                            bool force_chaining);
 582
 583     /**
 584      * Return the buffer list.
 585      *
 586      * This is the buffer list as passed to the kernel, i.e. it only contains
 587      * the parent buffers of sub-allocated buffers.
 588      *
 589      * \param cs    Command stream
 590      * \param list  Returned buffer list. Set to NULL to query the count only.
 591      * \return      The buffer count.
 592      */
 593     unsigned (*cs_get_buffer_list)(struct radeon_cmdbuf *cs,
 594                                    struct radeon_bo_list_item *list);
 595
 596     /**
 597      * Flush a command stream.
 598      *
 599      * \param cs          A command stream to flush.
 600      * \param flags,      PIPE_FLUSH_* flags.
 601      * \param fence       Pointer to a fence. If non-NULL, a fence is inserted
 602      *                    after the CS and is returned through this parameter.
 603      * \return Negative POSIX error code or 0 for success.
 604      *         Asynchronous submissions never return an error.
 605      */
 606     int (*cs_flush)(struct radeon_cmdbuf *cs,
 607                     unsigned flags,
 608                     struct pipe_fence_handle **fence);
 609
 610     /**
 611      * Create a fence before the CS is flushed.
 612      * The user must flush manually to complete the initializaton of the fence.
 613      *
 614      * The fence must not be used for anything except \ref cs_add_fence_dependency
 615      * before the flush.
 616      */
 617     struct pipe_fence_handle *(*cs_get_next_fence)(struct radeon_cmdbuf *cs);
 618
 619     /**
 620      * Return true if a buffer is referenced by a command stream.
 621      *
 622      * \param cs        A command stream.
 623      * \param buf       A winsys buffer.
 624      */
 625     bool (*cs_is_buffer_referenced)(struct radeon_cmdbuf *cs,
 626                                     struct pb_buffer *buf,
 627                                     enum radeon_bo_usage usage);
 628
 629     /**
 630      * Request access to a feature for a command stream.
 631      *
 632      * \param cs        A command stream.
 633      * \param fid       Feature ID, one of RADEON_FID_*
 634      * \param enable    Whether to enable or disable the feature.
 635      */
 636     bool (*cs_request_feature)(struct radeon_cmdbuf *cs,
 637                                enum radeon_feature_id fid,
 638                                bool enable);
 639      /**
 640       * Make sure all asynchronous flush of the cs have completed
 641       *
 642       * \param cs        A command stream.
 643       */
 644     void (*cs_sync_flush)(struct radeon_cmdbuf *cs);
 645
 646     /**
 647      * Add a fence dependency to the CS, so that the CS will wait for
 648      * the fence before execution.
 649      *
 650      * \param dependency_flags  Bitmask of RADEON_DEPENDENCY_*
 651      */
 652     void (*cs_add_fence_dependency)(struct radeon_cmdbuf *cs,
 653                                     struct pipe_fence_handle *fence,
 654                                     unsigned dependency_flags);
 655
 656     /**
 657      * Signal a syncobj when the CS finishes execution.
 658      */
 659     void (*cs_add_syncobj_signal)(struct radeon_cmdbuf *cs,
 660                                   struct pipe_fence_handle *fence);
 661
 662     /**
 663      * Wait for the fence and return true if the fence has been signalled.
 664      * The timeout of 0 will only return the status.
 665      * The timeout of PIPE_TIMEOUT_INFINITE will always wait until the fence
 666      * is signalled.
 667      */
 668     bool (*fence_wait)(struct radeon_winsys *ws,
 669                        struct pipe_fence_handle *fence,
 670                        uint64_t timeout);
 671
 672     /**
 673      * Reference counting for fences.
 674      */
 675     void (*fence_reference)(struct pipe_fence_handle **dst,
 676                             struct pipe_fence_handle *src);
 677
 678     /**
 679      * Create a new fence object corresponding to the given syncobj fd.
 680      */
 681     struct pipe_fence_handle *(*fence_import_syncobj)(struct radeon_winsys *ws,
 682                                                       int fd);
 683
 684     /**
 685      * Create a new fence object corresponding to the given sync_file.
 686      */
 687     struct pipe_fence_handle *(*fence_import_sync_file)(struct radeon_winsys *ws,
 688                                                         int fd);
 689
 690     /**
 691      * Return a sync_file FD corresponding to the given fence object.
 692      */
 693     int (*fence_export_sync_file)(struct radeon_winsys *ws,
 694                                   struct pipe_fence_handle *fence);
 695
 696     /**
 697      * Return a sync file FD that is already signalled.
 698      */
 699     int (*export_signalled_sync_file)(struct radeon_winsys *ws);
 700
 701     /**
 702      * Initialize surface
 703      *
 704      * \param ws        The winsys this function is called from.
 705      * \param tex       Input texture description
 706      * \param flags     Bitmask of RADEON_SURF_* flags
 707      * \param bpe       Bytes per pixel, it can be different for Z buffers.
 708      * \param mode      Preferred tile mode. (linear, 1D, or 2D)
 709      * \param surf      Output structure
 710      */
 711     int (*surface_init)(struct radeon_winsys *ws,
 712                         const struct pipe_resource *tex,
 713                         unsigned flags, unsigned bpe,
 714                         enum radeon_surf_mode mode,
 715                         struct radeon_surf *surf);
 716
 717     uint64_t (*query_value)(struct radeon_winsys *ws,
 718                             enum radeon_value_id value);
 719
 720     bool (*read_registers)(struct radeon_winsys *ws, unsigned reg_offset,
 721                            unsigned num_registers, uint32_t *out);
 722 };
 723
 724 static inline bool radeon_emitted(struct radeon_cmdbuf *cs, unsigned num_dw)
 725 {
 726     return cs && (cs->prev_dw + cs->current.cdw > num_dw);
 727 }
 728
 729 static inline void radeon_emit(struct radeon_cmdbuf *cs, uint32_t value)
 730 {
 731     cs->current.buf[cs->current.cdw++] = value;
 732 }
 733
 734 static inline void radeon_emit_array(struct radeon_cmdbuf *cs,
 735                                      const uint32_t *values, unsigned count)
 736 {
 737     memcpy(cs->current.buf + cs->current.cdw, values, count * 4);
 738     cs->current.cdw += count;
 739 }
 740
 741 enum radeon_heap {
 742     RADEON_HEAP_VRAM_NO_CPU_ACCESS,
 743     RADEON_HEAP_VRAM_READ_ONLY,
 744     RADEON_HEAP_VRAM_READ_ONLY_32BIT,
 745     RADEON_HEAP_VRAM_32BIT,
 746     RADEON_HEAP_VRAM,
 747     RADEON_HEAP_GTT_WC,
 748     RADEON_HEAP_GTT_WC_READ_ONLY,
 749     RADEON_HEAP_GTT_WC_READ_ONLY_32BIT,
 750     RADEON_HEAP_GTT_WC_32BIT,
 751     RADEON_HEAP_GTT,
 752     RADEON_MAX_SLAB_HEAPS,
 753     RADEON_MAX_CACHED_HEAPS = RADEON_MAX_SLAB_HEAPS,
 754 };
 755
 756 static inline enum radeon_bo_domain radeon_domain_from_heap(enum radeon_heap heap)
 757 {
 758     switch (heap) {
 759     case RADEON_HEAP_VRAM_NO_CPU_ACCESS:
 760     case RADEON_HEAP_VRAM_READ_ONLY:
 761     case RADEON_HEAP_VRAM_READ_ONLY_32BIT:
 762     case RADEON_HEAP_VRAM_32BIT:
 763     case RADEON_HEAP_VRAM:
 764         return RADEON_DOMAIN_VRAM;
 765     case RADEON_HEAP_GTT_WC:
 766     case RADEON_HEAP_GTT_WC_READ_ONLY:
 767     case RADEON_HEAP_GTT_WC_READ_ONLY_32BIT:
 768     case RADEON_HEAP_GTT_WC_32BIT:
 769     case RADEON_HEAP_GTT:
 770         return RADEON_DOMAIN_GTT;
 771     default:
 772         assert(0);
 773         return (enum radeon_bo_domain)0;
 774     }
 775 }
 776
 777 static inline unsigned radeon_flags_from_heap(enum radeon_heap heap)
 778 {
 779     unsigned flags = RADEON_FLAG_NO_INTERPROCESS_SHARING |
 780                      (heap != RADEON_HEAP_GTT ? RADEON_FLAG_GTT_WC : 0);
 781
 782     switch (heap) {
 783     case RADEON_HEAP_VRAM_NO_CPU_ACCESS:
 784         return flags |
 785                RADEON_FLAG_NO_CPU_ACCESS;
 786
 787     case RADEON_HEAP_VRAM_READ_ONLY:
 788     case RADEON_HEAP_GTT_WC_READ_ONLY:
 789         return flags |
 790                RADEON_FLAG_READ_ONLY;
 791
 792     case RADEON_HEAP_VRAM_READ_ONLY_32BIT:
 793     case RADEON_HEAP_GTT_WC_READ_ONLY_32BIT:
 794         return flags |
 795                RADEON_FLAG_READ_ONLY |
 796                RADEON_FLAG_32BIT;
 797
 798     case RADEON_HEAP_VRAM_32BIT:
 799     case RADEON_HEAP_GTT_WC_32BIT:
 800         return flags |
 801                RADEON_FLAG_32BIT;
 802
 803     case RADEON_HEAP_VRAM:
 804     case RADEON_HEAP_GTT_WC:
 805     case RADEON_HEAP_GTT:
 806     default:
 807         return flags;
 808     }
 809 }
 810
 811 /* Return the heap index for winsys allocators, or -1 on failure. */
 812 static inline int radeon_get_heap_index(enum radeon_bo_domain domain,
 813                                         enum radeon_bo_flag flags)
 814 {
 815     /* VRAM implies WC (write combining) */
 816     assert(!(domain & RADEON_DOMAIN_VRAM) || flags & RADEON_FLAG_GTT_WC);
 817     /* NO_CPU_ACCESS implies VRAM only. */
 818     assert(!(flags & RADEON_FLAG_NO_CPU_ACCESS) || domain == RADEON_DOMAIN_VRAM);
 819
 820     /* Resources with interprocess sharing don't use any winsys allocators. */
 821     if (!(flags & RADEON_FLAG_NO_INTERPROCESS_SHARING))
 822         return -1;
 823
 824     /* Unsupported flags: NO_SUBALLOC, SPARSE. */
 825     if (flags & ~(RADEON_FLAG_GTT_WC |
 826                   RADEON_FLAG_NO_CPU_ACCESS |
 827                   RADEON_FLAG_NO_INTERPROCESS_SHARING |
 828                   RADEON_FLAG_READ_ONLY |
 829                   RADEON_FLAG_32BIT))
 830         return -1;
 831
 832     switch (domain) {
 833     case RADEON_DOMAIN_VRAM:
 834         switch (flags & (RADEON_FLAG_NO_CPU_ACCESS |
 835                          RADEON_FLAG_READ_ONLY |
 836                          RADEON_FLAG_32BIT)) {
 837         case RADEON_FLAG_NO_CPU_ACCESS | RADEON_FLAG_READ_ONLY | RADEON_FLAG_32BIT:
 838         case RADEON_FLAG_NO_CPU_ACCESS | RADEON_FLAG_READ_ONLY:
 839             assert(!"NO_CPU_ACCESS | READ_ONLY doesn't make sense");
 840             return -1;
 841         case RADEON_FLAG_NO_CPU_ACCESS | RADEON_FLAG_32BIT:
 842             assert(!"NO_CPU_ACCESS with 32BIT is disallowed");
 843             return -1;
 844         case RADEON_FLAG_NO_CPU_ACCESS:
 845             return RADEON_HEAP_VRAM_NO_CPU_ACCESS;
 846         case RADEON_FLAG_READ_ONLY | RADEON_FLAG_32BIT:
 847             return RADEON_HEAP_VRAM_READ_ONLY_32BIT;
 848         case RADEON_FLAG_READ_ONLY:
 849             return RADEON_HEAP_VRAM_READ_ONLY;
 850         case RADEON_FLAG_32BIT:
 851             return RADEON_HEAP_VRAM_32BIT;
 852         case 0:
 853             return RADEON_HEAP_VRAM;
 854         }
 855         break;
 856     case RADEON_DOMAIN_GTT:
 857         switch (flags & (RADEON_FLAG_GTT_WC |
 858                          RADEON_FLAG_READ_ONLY |
 859                          RADEON_FLAG_32BIT)) {
 860         case RADEON_FLAG_GTT_WC | RADEON_FLAG_READ_ONLY | RADEON_FLAG_32BIT:
 861             return RADEON_HEAP_GTT_WC_READ_ONLY_32BIT;
 862         case RADEON_FLAG_GTT_WC | RADEON_FLAG_READ_ONLY:
 863             return RADEON_HEAP_GTT_WC_READ_ONLY;
 864         case RADEON_FLAG_GTT_WC | RADEON_FLAG_32BIT:
 865             return RADEON_HEAP_GTT_WC_32BIT;
 866         case RADEON_FLAG_GTT_WC:
 867             return RADEON_HEAP_GTT_WC;
 868         case RADEON_FLAG_READ_ONLY | RADEON_FLAG_32BIT:
 869         case RADEON_FLAG_READ_ONLY:
 870             assert(!"READ_ONLY without WC is disallowed");
 871             return -1;
 872         case RADEON_FLAG_32BIT:
 873             assert(!"32BIT without WC is disallowed");
 874             return -1;
 875         case 0:
 876             return RADEON_HEAP_GTT;
 877         }
 878         break;
 879     default:
 880         break;
 881     }
 882     return -1;
 883 }
 884
 885 #endif