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 struct fd_context;
  37 struct fd_resource;
  38 enum fd_resource_status;
  39
  40 /* Bitmask of stages in rendering that a particular query query is
  41  * active.  Queries will be automatically started/stopped (generating
  42  * additional fd_hw_sample_period's) on entrance/exit from stages that
  43  * are applicable to the query.
  44  *
  45  * NOTE: set the stage to NULL at end of IB to ensure no query is still
  46  * active.  Things aren't going to work out the way you want if a query
  47  * is active across IB's (or between tile IB and draw IB)
  48  */
  49 enum fd_render_stage {
  50         FD_STAGE_NULL     = 0x01,
  51         FD_STAGE_DRAW     = 0x02,
  52         FD_STAGE_CLEAR    = 0x04,
  53         /* used for driver internal draws (ie. util_blitter_blit()): */
  54         FD_STAGE_BLIT     = 0x08,
  55         FD_STAGE_ALL      = 0xff,
  56 };
  57
  58 #define MAX_HW_SAMPLE_PROVIDERS 5
  59 struct fd_hw_sample_provider;
  60 struct fd_hw_sample;
  61
  62 /* A batch tracks everything about a cmdstream batch/submit, including the
  63  * ringbuffers used for binning, draw, and gmem cmds, list of associated
  64  * fd_resource-s, etc.
  65  */
  66 struct fd_batch {
  67         struct pipe_reference reference;
  68         unsigned seqno;
  69         unsigned idx;       /* index into cache->batches[] */
  70
  71         int in_fence_fd;
  72         bool needs_out_fence_fd;
  73         struct pipe_fence_handle *fence;
  74
  75         struct fd_context *ctx;
  76
  77         struct util_queue_fence flush_fence;
  78
  79         /* do we need to mem2gmem before rendering.  We don't, if for example,
  80          * there was a glClear() that invalidated the entire previous buffer
  81          * contents.  Keep track of which buffer(s) are cleared, or needs
  82          * restore.  Masks of PIPE_CLEAR_*
  83          *
  84          * The 'cleared' bits will be set for buffers which are *entirely*
  85          * cleared, and 'partial_cleared' bits will be set if you must
  86          * check cleared_scissor.
  87          *
  88          * The 'invalidated' bits are set for cleared buffers, and buffers
  89          * where the contents are undefined, ie. what we don't need to restore
  90          * to gmem.
  91          */
  92         enum {
  93                 /* align bitmask values w/ PIPE_CLEAR_*.. since that is convenient.. */
  94                 FD_BUFFER_COLOR   = PIPE_CLEAR_COLOR,
  95                 FD_BUFFER_DEPTH   = PIPE_CLEAR_DEPTH,
  96                 FD_BUFFER_STENCIL = PIPE_CLEAR_STENCIL,
  97                 FD_BUFFER_ALL     = FD_BUFFER_COLOR | FD_BUFFER_DEPTH | FD_BUFFER_STENCIL,
  98         } invalidated, cleared, fast_cleared, restore, resolve;
  99
 100         /* is this a non-draw batch (ie compute/blit which has no pfb state)? */
 101         bool nondraw : 1;
 102         bool needs_flush : 1;
 103         bool flushed : 1;
 104         bool blit : 1;
 105         bool back_blit : 1;      /* only blit so far is resource shadowing back-blit */
 106
 107         /* Keep track if WAIT_FOR_IDLE is needed for registers we need
 108          * to update via RMW:
 109          */
 110         bool needs_wfi : 1;
 111
 112         /* To decide whether to render to system memory, keep track of the
 113          * number of draws, and whether any of them require multisample,
 114          * depth_test (or depth write), stencil_test, blending, and
 115          * color_logic_Op (since those functions are disabled when by-
 116          * passing GMEM.
 117          */
 118         enum {
 119                 FD_GMEM_CLEARS_DEPTH_STENCIL = 0x01,
 120                 FD_GMEM_DEPTH_ENABLED        = 0x02,
 121                 FD_GMEM_STENCIL_ENABLED      = 0x04,
 122
 123                 FD_GMEM_BLEND_ENABLED        = 0x10,
 124                 FD_GMEM_LOGICOP_ENABLED      = 0x20,
 125         } gmem_reason;
 126         unsigned num_draws;   /* number of draws in current batch */
 127
 128         /* Track the maximal bounds of the scissor of all the draws within a
 129          * batch.  Used at the tile rendering step (fd_gmem_render_tiles(),
 130          * mem2gmem/gmem2mem) to avoid needlessly moving data in/out of gmem.
 131          */
 132         struct pipe_scissor_state max_scissor;
 133
 134         /* Keep track of DRAW initiators that need to be patched up depending
 135          * on whether we using binning or not:
 136          */
 137         struct util_dynarray draw_patches;
 138
 139         /* Keep track of writes to RB_RENDER_CONTROL which need to be patched
 140          * once we know whether or not to use GMEM, and GMEM tile pitch.
 141          *
 142          * (only for a3xx.. but having gen specific subclasses of fd_batch
 143          * seemed overkill for now)
 144          */
 145         struct util_dynarray rbrc_patches;
 146
 147         struct pipe_framebuffer_state framebuffer;
 148
 149         struct fd_submit *submit;
 150
 151         /** draw pass cmdstream: */
 152         struct fd_ringbuffer *draw;
 153         /** binning pass cmdstream: */
 154         struct fd_ringbuffer *binning;
 155         /** tiling/gmem (IB0) cmdstream: */
 156         struct fd_ringbuffer *gmem;
 157
 158         // TODO maybe more generically split out clear and clear_binning rings?
 159         struct fd_ringbuffer *lrz_clear;
 160         struct fd_ringbuffer *tile_setup;
 161         struct fd_ringbuffer *tile_fini;
 162
 163         union pipe_color_union clear_color[MAX_RENDER_TARGETS];
 164         double clear_depth;
 165         unsigned clear_stencil;
 166
 167         /**
 168          * hw query related state:
 169          */
 170         /*@{*/
 171         /* next sample offset.. incremented for each sample in the batch/
 172          * submit, reset to zero on next submit.
 173          */
 174         uint32_t next_sample_offset;
 175
 176         /* cached samples (in case multiple queries need to reference
 177          * the same sample snapshot)
 178          */
 179         struct fd_hw_sample *sample_cache[MAX_HW_SAMPLE_PROVIDERS];
 180
 181         /* which sample providers were active in the current batch: */
 182         uint32_t active_providers;
 183
 184         /* tracking for current stage, to know when to start/stop
 185          * any active queries:
 186          */
 187         enum fd_render_stage stage;
 188
 189         /* list of samples in current batch: */
 190         struct util_dynarray samples;
 191
 192         /* current query result bo and tile stride: */
 193         struct pipe_resource *query_buf;
 194         uint32_t query_tile_stride;
 195         /*@}*/
 196
 197
 198         /* Set of resources used by currently-unsubmitted batch (read or
 199          * write).. does not hold a reference to the resource.
 200          */
 201         struct set *resources;
 202
 203         /** key in batch-cache (if not null): */
 204         const void *key;
 205         uint32_t hash;
 206
 207         /** set of dependent batches.. holds refs to dependent batches: */
 208         uint32_t dependents_mask;
 209 };
 210
 211 struct fd_batch * fd_batch_create(struct fd_context *ctx, bool nondraw);
 212
 213 void fd_batch_reset(struct fd_batch *batch);
 214 void fd_batch_sync(struct fd_batch *batch);
 215 void fd_batch_flush(struct fd_batch *batch, bool sync, bool force);
 216 void fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep);
 217 void fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write);
 218 void fd_batch_check_size(struct fd_batch *batch);
 219
 220 /* not called directly: */
 221 void __fd_batch_describe(char* buf, const struct fd_batch *batch);
 222 void __fd_batch_destroy(struct fd_batch *batch);
 223
 224 /*
 225  * NOTE the rule is, you need to hold the screen->lock when destroying
 226  * a batch..  so either use fd_batch_reference() (which grabs the lock
 227  * for you) if you don't hold the lock, or fd_batch_reference_locked()
 228  * if you do hold the lock.
 229  *
 230  * WARNING the _locked() version can briefly drop the lock.  Without
 231  * recursive mutexes, I'm not sure there is much else we can do (since
 232  * __fd_batch_destroy() needs to unref resources)
 233  */
 234
 235 /* fwd-decl prototypes to untangle header dependency :-/ */
 236 static inline void fd_context_assert_locked(struct fd_context *ctx);
 237 static inline void fd_context_lock(struct fd_context *ctx);
 238 static inline void fd_context_unlock(struct fd_context *ctx);
 239
 240 static inline void
 241 fd_batch_reference_locked(struct fd_batch **ptr, struct fd_batch *batch)
 242 {
 243         struct fd_batch *old_batch = *ptr;
 244
 245         /* only need lock if a reference is dropped: */
 246         if (old_batch)
 247                 fd_context_assert_locked(old_batch->ctx);
 248
 249         if (pipe_reference_described(&(*ptr)->reference, &batch->reference,
 250                         (debug_reference_descriptor)__fd_batch_describe))
 251                 __fd_batch_destroy(old_batch);
 252
 253         *ptr = batch;
 254 }
 255
 256 static inline void
 257 fd_batch_reference(struct fd_batch **ptr, struct fd_batch *batch)
 258 {
 259         struct fd_batch *old_batch = *ptr;
 260         struct fd_context *ctx = old_batch ? old_batch->ctx : NULL;
 261
 262         if (ctx)
 263                 fd_context_lock(ctx);
 264
 265         fd_batch_reference_locked(ptr, batch);
 266
 267         if (ctx)
 268                 fd_context_unlock(ctx);
 269 }
 270
 271 #include "freedreno_context.h"
 272
 273 static inline void
 274 fd_reset_wfi(struct fd_batch *batch)
 275 {
 276         batch->needs_wfi = true;
 277 }
 278
 279 void fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring);
 280
 281 /* emit a CP_EVENT_WRITE:
 282  */
 283 static inline void
 284 fd_event_write(struct fd_batch *batch, struct fd_ringbuffer *ring,
 285                 enum vgt_event_type evt)
 286 {
 287         OUT_PKT3(ring, CP_EVENT_WRITE, 1);
 288         OUT_RING(ring, evt);
 289         fd_reset_wfi(batch);
 290 }
 291
 292 #endif /* FREEDRENO_BATCH_H_ */