src/gallium/drivers/freedreno/freedreno_batch.h

   1 /*
   2  * Copyright (C) 2016 Rob Clark <robclark@freedesktop.org>
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors:
  24  *    Rob Clark <robclark@freedesktop.org>
  25  */
  26
  27 #ifndef FREEDRENO_BATCH_H_
  28 #define FREEDRENO_BATCH_H_
  29
  30 #include "util/u_inlines.h"
  31 #include "util/u_queue.h"
  32 #include "util/list.h"
  33
  34 #include "freedreno_util.h"
  35
  36 #ifdef DEBUG
  37 #  define BATCH_DEBUG (fd_mesa_debug & FD_DBG_MSGS)
  38 #else
  39 #  define BATCH_DEBUG 0
  40 #endif
  41
  42 struct fd_context;
  43 struct fd_resource;
  44 enum fd_resource_status;
  45
  46 /* Bitmask of stages in rendering that a particular query query is
  47  * active.  Queries will be automatically started/stopped (generating
  48  * additional fd_hw_sample_period's) on entrance/exit from stages that
  49  * are applicable to the query.
  50  *
  51  * NOTE: set the stage to NULL at end of IB to ensure no query is still
  52  * active.  Things aren't going to work out the way you want if a query
  53  * is active across IB's (or between tile IB and draw IB)
  54  */
  55 enum fd_render_stage {
  56         FD_STAGE_NULL     = 0x00,
  57         FD_STAGE_DRAW     = 0x01,
  58         FD_STAGE_CLEAR    = 0x02,
  59         /* used for driver internal draws (ie. util_blitter_blit()): */
  60         FD_STAGE_BLIT     = 0x04,
  61         FD_STAGE_ALL      = 0xff,
  62 };
  63
  64 #define MAX_HW_SAMPLE_PROVIDERS 7
  65 struct fd_hw_sample_provider;
  66 struct fd_hw_sample;
  67
  68 /* A batch tracks everything about a cmdstream batch/submit, including the
  69  * ringbuffers used for binning, draw, and gmem cmds, list of associated
  70  * fd_resource-s, etc.
  71  */
  72 struct fd_batch {
  73         struct pipe_reference reference;
  74         unsigned seqno;
  75         unsigned idx;       /* index into cache->batches[] */
  76
  77         int in_fence_fd;
  78         bool needs_out_fence_fd;
  79         struct pipe_fence_handle *fence;
  80
  81         struct fd_context *ctx;
  82
  83         /* do we need to mem2gmem before rendering.  We don't, if for example,
  84          * there was a glClear() that invalidated the entire previous buffer
  85          * contents.  Keep track of which buffer(s) are cleared, or needs
  86          * restore.  Masks of PIPE_CLEAR_*
  87          *
  88          * The 'cleared' bits will be set for buffers which are *entirely*
  89          * cleared, and 'partial_cleared' bits will be set if you must
  90          * check cleared_scissor.
  91          *
  92          * The 'invalidated' bits are set for cleared buffers, and buffers
  93          * where the contents are undefined, ie. what we don't need to restore
  94          * to gmem.
  95          */
  96         enum {
  97                 /* align bitmask values w/ PIPE_CLEAR_*.. since that is convenient.. */
  98                 FD_BUFFER_COLOR   = PIPE_CLEAR_COLOR,
  99                 FD_BUFFER_DEPTH   = PIPE_CLEAR_DEPTH,
 100                 FD_BUFFER_STENCIL = PIPE_CLEAR_STENCIL,
 101                 FD_BUFFER_ALL     = FD_BUFFER_COLOR | FD_BUFFER_DEPTH | FD_BUFFER_STENCIL,
 102         } invalidated, cleared, fast_cleared, restore, resolve;
 103
 104         /* is this a non-draw batch (ie compute/blit which has no pfb state)? */
 105         bool nondraw : 1;
 106         bool needs_flush : 1;
 107         bool flushed : 1;
 108         bool blit : 1;
 109         bool back_blit : 1;      /* only blit so far is resource shadowing back-blit */
 110         bool tessellation : 1;      /* tessellation used in batch */
 111
 112         /* Keep track if WAIT_FOR_IDLE is needed for registers we need
 113          * to update via RMW:
 114          */
 115         bool needs_wfi : 1;
 116
 117         /* To decide whether to render to system memory, keep track of the
 118          * number of draws, and whether any of them require multisample,
 119          * depth_test (or depth write), stencil_test, blending, and
 120          * color_logic_Op (since those functions are disabled when by-
 121          * passing GMEM.
 122          */
 123         enum {
 124                 FD_GMEM_CLEARS_DEPTH_STENCIL = 0x01,
 125                 FD_GMEM_DEPTH_ENABLED        = 0x02,
 126                 FD_GMEM_STENCIL_ENABLED      = 0x04,
 127
 128                 FD_GMEM_BLEND_ENABLED        = 0x10,
 129                 FD_GMEM_LOGICOP_ENABLED      = 0x20,
 130                 FD_GMEM_FB_READ              = 0x40,
 131         } gmem_reason;
 132
 133         /* At submit time, once we've decided that this batch will use GMEM
 134          * rendering, the appropriate gmem state is looked up:
 135          */
 136         const struct fd_gmem_stateobj *gmem_state;
 137
 138         unsigned num_draws;      /* number of draws in current batch */
 139         unsigned num_vertices;   /* number of vertices in current batch */
 140
 141         /* Currently only used on a6xx, to calculate vsc prim/draw stream
 142          * sizes:
 143          */
 144         unsigned num_bins_per_pipe;
 145         unsigned prim_strm_bits;
 146         unsigned draw_strm_bits;
 147
 148         /* Track the maximal bounds of the scissor of all the draws within a
 149          * batch.  Used at the tile rendering step (fd_gmem_render_tiles(),
 150          * mem2gmem/gmem2mem) to avoid needlessly moving data in/out of gmem.
 151          */
 152         struct pipe_scissor_state max_scissor;
 153
 154         /* Keep track of DRAW initiators that need to be patched up depending
 155          * on whether we using binning or not:
 156          */
 157         struct util_dynarray draw_patches;
 158
 159         /* texture state that needs patching for fb_read: */
 160         struct util_dynarray fb_read_patches;
 161
 162         /* Keep track of writes to RB_RENDER_CONTROL which need to be patched
 163          * once we know whether or not to use GMEM, and GMEM tile pitch.
 164          *
 165          * (only for a3xx.. but having gen specific subclasses of fd_batch
 166          * seemed overkill for now)
 167          */
 168         struct util_dynarray rbrc_patches;
 169
 170         /* Keep track of GMEM related values that need to be patched up once we
 171          * know the gmem layout:
 172          */
 173         struct util_dynarray gmem_patches;
 174
 175         /* Keep track of pointer to start of MEM exports for a20x binning shaders
 176          *
 177          * this is so the end of the shader can be cut off at the right point
 178          * depending on the GMEM configuration
 179          */
 180         struct util_dynarray shader_patches;
 181
 182         struct pipe_framebuffer_state framebuffer;
 183
 184         struct fd_submit *submit;
 185
 186         /** draw pass cmdstream: */
 187         struct fd_ringbuffer *draw;
 188         /** binning pass cmdstream: */
 189         struct fd_ringbuffer *binning;
 190         /** tiling/gmem (IB0) cmdstream: */
 191         struct fd_ringbuffer *gmem;
 192
 193         /** preemble cmdstream (executed once before first tile): */
 194         struct fd_ringbuffer *prologue;
 195
 196         /** epilogue cmdstream (executed after each tile): */
 197         struct fd_ringbuffer *epilogue;
 198
 199         struct fd_ringbuffer *tile_setup;
 200         struct fd_ringbuffer *tile_fini;
 201
 202         union pipe_color_union clear_color[MAX_RENDER_TARGETS];
 203         double clear_depth;
 204         unsigned clear_stencil;
 205
 206         /**
 207          * hw query related state:
 208          */
 209         /*@{*/
 210         /* next sample offset.. incremented for each sample in the batch/
 211          * submit, reset to zero on next submit.
 212          */
 213         uint32_t next_sample_offset;
 214
 215         /* cached samples (in case multiple queries need to reference
 216          * the same sample snapshot)
 217          */
 218         struct fd_hw_sample *sample_cache[MAX_HW_SAMPLE_PROVIDERS];
 219
 220         /* which sample providers were active in the current batch: */
 221         uint32_t active_providers;
 222
 223         /* tracking for current stage, to know when to start/stop
 224          * any active queries:
 225          */
 226         enum fd_render_stage stage;
 227
 228         /* list of samples in current batch: */
 229         struct util_dynarray samples;
 230
 231         /* current query result bo and tile stride: */
 232         struct pipe_resource *query_buf;
 233         uint32_t query_tile_stride;
 234         /*@}*/
 235
 236
 237         /* Set of resources used by currently-unsubmitted batch (read or
 238          * write).. does not hold a reference to the resource.
 239          */
 240         struct set *resources;
 241
 242         /** key in batch-cache (if not null): */
 243         const void *key;
 244         uint32_t hash;
 245
 246         /** set of dependent batches.. holds refs to dependent batches: */
 247         uint32_t dependents_mask;
 248
 249         /* Buffer for tessellation engine input
 250          */
 251         struct fd_bo *tessfactor_bo;
 252         uint32_t tessfactor_size;
 253
 254         /* Buffer for passing parameters between TCS and TES
 255          */
 256         struct fd_bo *tessparam_bo;
 257         uint32_t tessparam_size;
 258
 259         struct fd_ringbuffer *tess_addrs_constobj;
 260
 261         struct list_head log_chunks;  /* list of unflushed log chunks in fifo order */
 262 };
 263
 264 struct fd_batch * fd_batch_create(struct fd_context *ctx, bool nondraw);
 265
 266 void fd_batch_reset(struct fd_batch *batch);
 267 void fd_batch_flush(struct fd_batch *batch);
 268 void fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep);
 269 void fd_batch_resource_write(struct fd_batch *batch, struct fd_resource *rsc);
 270 void fd_batch_resource_read_slowpath(struct fd_batch *batch, struct fd_resource *rsc);
 271 void fd_batch_check_size(struct fd_batch *batch);
 272
 273 /* not called directly: */
 274 void __fd_batch_describe(char* buf, const struct fd_batch *batch);
 275 void __fd_batch_destroy(struct fd_batch *batch);
 276
 277 /*
 278  * NOTE the rule is, you need to hold the screen->lock when destroying
 279  * a batch..  so either use fd_batch_reference() (which grabs the lock
 280  * for you) if you don't hold the lock, or fd_batch_reference_locked()
 281  * if you do hold the lock.
 282  *
 283  * WARNING the _locked() version can briefly drop the lock.  Without
 284  * recursive mutexes, I'm not sure there is much else we can do (since
 285  * __fd_batch_destroy() needs to unref resources)
 286  *
 287  * WARNING you must acquire the screen->lock and use the _locked()
 288  * version in case that the batch being ref'd can disappear under
 289  * you.
 290  */
 291
 292 /* fwd-decl prototypes to untangle header dependency :-/ */
 293 static inline void fd_context_assert_locked(struct fd_context *ctx);
 294 static inline void fd_context_lock(struct fd_context *ctx);
 295 static inline void fd_context_unlock(struct fd_context *ctx);
 296
 297 static inline void
 298 fd_batch_reference_locked(struct fd_batch **ptr, struct fd_batch *batch)
 299 {
 300         struct fd_batch *old_batch = *ptr;
 301
 302         /* only need lock if a reference is dropped: */
 303         if (old_batch)
 304                 fd_context_assert_locked(old_batch->ctx);
 305
 306         if (pipe_reference_described(&(*ptr)->reference, &batch->reference,
 307                         (debug_reference_descriptor)__fd_batch_describe))
 308                 __fd_batch_destroy(old_batch);
 309
 310         *ptr = batch;
 311 }
 312
 313 static inline void
 314 fd_batch_reference(struct fd_batch **ptr, struct fd_batch *batch)
 315 {
 316         struct fd_batch *old_batch = *ptr;
 317         struct fd_context *ctx = old_batch ? old_batch->ctx : NULL;
 318
 319         if (ctx)
 320                 fd_context_lock(ctx);
 321
 322         fd_batch_reference_locked(ptr, batch);
 323
 324         if (ctx)
 325                 fd_context_unlock(ctx);
 326 }
 327
 328 #include "freedreno_context.h"
 329
 330 static inline void
 331 fd_reset_wfi(struct fd_batch *batch)
 332 {
 333         batch->needs_wfi = true;
 334 }
 335
 336 void fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring);
 337
 338 /* emit a CP_EVENT_WRITE:
 339  */
 340 static inline void
 341 fd_event_write(struct fd_batch *batch, struct fd_ringbuffer *ring,
 342                 enum vgt_event_type evt)
 343 {
 344         OUT_PKT3(ring, CP_EVENT_WRITE, 1);
 345         OUT_RING(ring, evt);
 346         fd_reset_wfi(batch);
 347 }
 348
 349 /* Get per-tile epilogue */
 350 static inline struct fd_ringbuffer *
 351 fd_batch_get_epilogue(struct fd_batch *batch)
 352 {
 353         if (batch->epilogue == NULL)
 354                 batch->epilogue = fd_submit_new_ringbuffer(batch->submit, 0x1000, 0);
 355
 356         return batch->epilogue;
 357 }
 358
 359 struct fd_ringbuffer * fd_batch_get_prologue(struct fd_batch *batch);
 360
 361 #endif /* FREEDRENO_BATCH_H_ */