src/gallium/drivers/freedreno/freedreno_context.h

   1 /*
   2  * Copyright (C) 2012 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_CONTEXT_H_
  28 #define FREEDRENO_CONTEXT_H_
  29
  30 #include "pipe/p_context.h"
  31 #include "indices/u_primconvert.h"
  32 #include "util/u_blitter.h"
  33 #include "util/list.h"
  34 #include "util/slab.h"
  35 #include "util/u_string.h"
  36
  37 #include "freedreno_batch.h"
  38 #include "freedreno_screen.h"
  39 #include "freedreno_gmem.h"
  40 #include "freedreno_util.h"
  41
  42 #define BORDER_COLOR_UPLOAD_SIZE (2 * PIPE_MAX_SAMPLERS * BORDERCOLOR_SIZE)
  43
  44 struct fd_vertex_stateobj;
  45
  46 struct fd_texture_stateobj {
  47         struct pipe_sampler_view *textures[PIPE_MAX_SAMPLERS];
  48         unsigned num_textures;
  49         unsigned valid_textures;
  50         struct pipe_sampler_state *samplers[PIPE_MAX_SAMPLERS];
  51         unsigned num_samplers;
  52         unsigned valid_samplers;
  53         /* number of samples per sampler, 2 bits per sampler: */
  54         uint32_t samples;
  55 };
  56
  57 struct fd_program_stateobj {
  58         void *vs, *hs, *ds, *gs, *fs;
  59 };
  60
  61 struct fd_constbuf_stateobj {
  62         struct pipe_constant_buffer cb[PIPE_MAX_CONSTANT_BUFFERS];
  63         uint32_t enabled_mask;
  64 };
  65
  66 struct fd_shaderbuf_stateobj {
  67         struct pipe_shader_buffer sb[PIPE_MAX_SHADER_BUFFERS];
  68         uint32_t enabled_mask;
  69         uint32_t writable_mask;
  70 };
  71
  72 struct fd_shaderimg_stateobj {
  73         struct pipe_image_view si[PIPE_MAX_SHADER_IMAGES];
  74         uint32_t enabled_mask;
  75 };
  76
  77 struct fd_vertexbuf_stateobj {
  78         struct pipe_vertex_buffer vb[PIPE_MAX_ATTRIBS];
  79         unsigned count;
  80         uint32_t enabled_mask;
  81 };
  82
  83 struct fd_vertex_stateobj {
  84         struct pipe_vertex_element pipe[PIPE_MAX_ATTRIBS];
  85         unsigned num_elements;
  86 };
  87
  88 struct fd_streamout_stateobj {
  89         struct pipe_stream_output_target *targets[PIPE_MAX_SO_BUFFERS];
  90         /* Bitmask of stream that should be reset. */
  91         unsigned reset;
  92
  93         unsigned num_targets;
  94         /* Track offset from vtxcnt for streamout data.  This counter
  95          * is just incremented by # of vertices on each draw until
  96          * reset or new streamout buffer bound.
  97          *
  98          * When we eventually have GS, the CPU won't actually know the
  99          * number of vertices per draw, so I think we'll have to do
 100          * something more clever.
 101          */
 102         unsigned offsets[PIPE_MAX_SO_BUFFERS];
 103 };
 104
 105 #define MAX_GLOBAL_BUFFERS 16
 106 struct fd_global_bindings_stateobj {
 107         struct pipe_resource *buf[MAX_GLOBAL_BUFFERS];
 108         uint32_t enabled_mask;
 109 };
 110
 111 /* group together the vertex and vertexbuf state.. for ease of passing
 112  * around, and because various internal operations (gmem<->mem, etc)
 113  * need their own vertex state:
 114  */
 115 struct fd_vertex_state {
 116         struct fd_vertex_stateobj *vtx;
 117         struct fd_vertexbuf_stateobj vertexbuf;
 118 };
 119
 120 /* global 3d pipeline dirty state: */
 121 enum fd_dirty_3d_state {
 122         FD_DIRTY_BLEND       = BIT(0),
 123         FD_DIRTY_RASTERIZER  = BIT(1),
 124         FD_DIRTY_ZSA         = BIT(2),
 125         FD_DIRTY_BLEND_COLOR = BIT(3),
 126         FD_DIRTY_STENCIL_REF = BIT(4),
 127         FD_DIRTY_SAMPLE_MASK = BIT(5),
 128         FD_DIRTY_FRAMEBUFFER = BIT(6),
 129         FD_DIRTY_STIPPLE     = BIT(7),
 130         FD_DIRTY_VIEWPORT    = BIT(8),
 131         FD_DIRTY_VTXSTATE    = BIT(9),
 132         FD_DIRTY_VTXBUF      = BIT(10),
 133         FD_DIRTY_MIN_SAMPLES = BIT(11),
 134
 135         FD_DIRTY_SCISSOR     = BIT(12),
 136         FD_DIRTY_STREAMOUT   = BIT(13),
 137         FD_DIRTY_UCP         = BIT(14),
 138         FD_DIRTY_BLEND_DUAL  = BIT(15),
 139
 140         /* These are a bit redundent with fd_dirty_shader_state, and possibly
 141          * should be removed.  (But OTOH kinda convenient in some places)
 142          */
 143         FD_DIRTY_PROG        = BIT(16),
 144         FD_DIRTY_CONST       = BIT(17),
 145         FD_DIRTY_TEX         = BIT(18),
 146         FD_DIRTY_IMAGE       = BIT(19),
 147         FD_DIRTY_SSBO        = BIT(20),
 148
 149         /* only used by a2xx.. possibly can be removed.. */
 150         FD_DIRTY_TEXSTATE    = BIT(21),
 151 };
 152
 153 /* per shader-stage dirty state: */
 154 enum fd_dirty_shader_state {
 155         FD_DIRTY_SHADER_PROG  = BIT(0),
 156         FD_DIRTY_SHADER_CONST = BIT(1),
 157         FD_DIRTY_SHADER_TEX   = BIT(2),
 158         FD_DIRTY_SHADER_SSBO  = BIT(3),
 159         FD_DIRTY_SHADER_IMAGE = BIT(4),
 160 };
 161
 162 struct fd_context {
 163         struct pipe_context base;
 164
 165         struct list_head node;   /* node in screen->context_list */
 166
 167         /* We currently need to serialize emitting GMEM batches, because of
 168          * VSC state access in the context.
 169          *
 170          * In practice this lock should not be contended, since pipe_context
 171          * use should be single threaded.  But it is needed to protect the
 172          * case, with batch reordering where a ctxB batch triggers flushing
 173          * a ctxA batch
 174          */
 175         mtx_t gmem_lock;
 176
 177         struct fd_device *dev;
 178         struct fd_screen *screen;
 179         struct fd_pipe *pipe;
 180
 181         struct blitter_context *blitter;
 182         void *clear_rs_state;
 183         struct primconvert_context *primconvert;
 184
 185         /* slab for pipe_transfer allocations: */
 186         struct slab_child_pool transfer_pool;
 187
 188         /**
 189          * query related state:
 190          */
 191         /*@{*/
 192         /* slabs for fd_hw_sample and fd_hw_sample_period allocations: */
 193         struct slab_mempool sample_pool;
 194         struct slab_mempool sample_period_pool;
 195
 196         /* sample-providers for hw queries: */
 197         const struct fd_hw_sample_provider *hw_sample_providers[MAX_HW_SAMPLE_PROVIDERS];
 198
 199         /* list of active queries: */
 200         struct list_head hw_active_queries;
 201
 202         /* sample-providers for accumulating hw queries: */
 203         const struct fd_acc_sample_provider *acc_sample_providers[MAX_HW_SAMPLE_PROVIDERS];
 204
 205         /* list of active accumulating queries: */
 206         struct list_head acc_active_queries;
 207         /*@}*/
 208
 209         /* Whether we need to walk the acc_active_queries next fd_set_stage() to
 210          * update active queries (even if stage doesn't change).
 211          */
 212         bool update_active_queries;
 213
 214         /* Current state of pctx->set_active_query_state() (i.e. "should drawing
 215          * be counted against non-perfcounter queries")
 216          */
 217         bool active_queries;
 218
 219         /* table with PIPE_PRIM_MAX entries mapping PIPE_PRIM_x to
 220          * DI_PT_x value to use for draw initiator.  There are some
 221          * slight differences between generation:
 222          */
 223         const uint8_t *primtypes;
 224         uint32_t primtype_mask;
 225
 226         /* shaders used by clear, and gmem->mem blits: */
 227         struct fd_program_stateobj solid_prog; // TODO move to screen?
 228
 229         /* shaders used by mem->gmem blits: */
 230         struct fd_program_stateobj blit_prog[MAX_RENDER_TARGETS]; // TODO move to screen?
 231         struct fd_program_stateobj blit_z, blit_zs;
 232
 233         /* Stats/counters:
 234          */
 235         struct {
 236                 uint64_t prims_emitted;
 237                 uint64_t prims_generated;
 238                 uint64_t draw_calls;
 239                 uint64_t batch_total, batch_sysmem, batch_gmem, batch_nondraw, batch_restore;
 240                 uint64_t staging_uploads, shadow_uploads;
 241                 uint64_t vs_regs, hs_regs, ds_regs, gs_regs, fs_regs;
 242         } stats;
 243
 244         /* Current batch.. the rule here is that you can deref ctx->batch
 245          * in codepaths from pipe_context entrypoints.  But not in code-
 246          * paths from fd_batch_flush() (basically, the stuff that gets
 247          * called from GMEM code), since in those code-paths the batch
 248          * you care about is not necessarily the same as ctx->batch.
 249          */
 250         struct fd_batch *batch;
 251
 252         /* NULL if there has been rendering since last flush.  Otherwise
 253          * keeps a reference to the last fence so we can re-use it rather
 254          * than having to flush no-op batch.
 255          */
 256         struct pipe_fence_handle *last_fence;
 257
 258         /* track last known reset status globally and per-context to
 259          * determine if more resets occurred since then.  If global reset
 260          * count increases, it means some other context crashed.  If
 261          * per-context reset count increases, it means we crashed the
 262          * gpu.
 263          */
 264         uint32_t context_reset_count, global_reset_count;
 265
 266         /* Are we in process of shadowing a resource? Used to detect recursion
 267          * in transfer_map, and skip unneeded synchronization.
 268          */
 269         bool in_shadow : 1;
 270
 271         /* Ie. in blit situation where we no longer care about previous framebuffer
 272          * contents.  Main point is to eliminate blits from fd_try_shadow_resource().
 273          * For example, in case of texture upload + gen-mipmaps.
 274          */
 275         bool in_discard_blit : 1;
 276
 277         struct pipe_scissor_state scissor;
 278
 279         /* we don't have a disable/enable bit for scissor, so instead we keep
 280          * a disabled-scissor state which matches the entire bound framebuffer
 281          * and use that when scissor is not enabled.
 282          */
 283         struct pipe_scissor_state disabled_scissor;
 284
 285         /* Per vsc pipe bo's (a2xx-a5xx): */
 286         struct fd_bo *vsc_pipe_bo[32];
 287
 288         /* which state objects need to be re-emit'd: */
 289         enum fd_dirty_3d_state dirty;
 290
 291         /* per shader-stage dirty status: */
 292         enum fd_dirty_shader_state dirty_shader[PIPE_SHADER_TYPES];
 293
 294         void *compute;
 295         struct pipe_blend_state *blend;
 296         struct pipe_rasterizer_state *rasterizer;
 297         struct pipe_depth_stencil_alpha_state *zsa;
 298
 299         struct fd_texture_stateobj tex[PIPE_SHADER_TYPES];
 300
 301         struct fd_program_stateobj prog;
 302
 303         struct fd_vertex_state vtx;
 304
 305         struct pipe_blend_color blend_color;
 306         struct pipe_stencil_ref stencil_ref;
 307         unsigned sample_mask;
 308         unsigned min_samples;
 309         /* local context fb state, for when ctx->batch is null: */
 310         struct pipe_framebuffer_state framebuffer;
 311         struct pipe_poly_stipple stipple;
 312         struct pipe_viewport_state viewport;
 313         struct pipe_scissor_state viewport_scissor;
 314         struct fd_constbuf_stateobj constbuf[PIPE_SHADER_TYPES];
 315         struct fd_shaderbuf_stateobj shaderbuf[PIPE_SHADER_TYPES];
 316         struct fd_shaderimg_stateobj shaderimg[PIPE_SHADER_TYPES];
 317         struct fd_streamout_stateobj streamout;
 318         struct fd_global_bindings_stateobj global_bindings;
 319         struct pipe_clip_state ucp;
 320
 321         struct pipe_query *cond_query;
 322         bool cond_cond; /* inverted rendering condition */
 323         uint cond_mode;
 324
 325         struct pipe_debug_callback debug;
 326
 327         /* GMEM/tile handling fxns: */
 328         void (*emit_tile_init)(struct fd_batch *batch);
 329         void (*emit_tile_prep)(struct fd_batch *batch, const struct fd_tile *tile);
 330         void (*emit_tile_mem2gmem)(struct fd_batch *batch, const struct fd_tile *tile);
 331         void (*emit_tile_renderprep)(struct fd_batch *batch, const struct fd_tile *tile);
 332         void (*emit_tile)(struct fd_batch *batch, const struct fd_tile *tile);
 333         void (*emit_tile_gmem2mem)(struct fd_batch *batch, const struct fd_tile *tile);
 334         void (*emit_tile_fini)(struct fd_batch *batch);   /* optional */
 335
 336         /* optional, for GMEM bypass: */
 337         void (*emit_sysmem_prep)(struct fd_batch *batch);
 338         void (*emit_sysmem_fini)(struct fd_batch *batch);
 339
 340         /* draw: */
 341         bool (*draw_vbo)(struct fd_context *ctx, const struct pipe_draw_info *info,
 342                         unsigned index_offset);
 343         bool (*clear)(struct fd_context *ctx, unsigned buffers,
 344                         const union pipe_color_union *color, double depth, unsigned stencil);
 345
 346         /* compute: */
 347         void (*launch_grid)(struct fd_context *ctx, const struct pipe_grid_info *info);
 348
 349         /* query: */
 350         struct fd_query * (*create_query)(struct fd_context *ctx, unsigned query_type, unsigned index);
 351         void (*query_prepare)(struct fd_batch *batch, uint32_t num_tiles);
 352         void (*query_prepare_tile)(struct fd_batch *batch, uint32_t n,
 353                         struct fd_ringbuffer *ring);
 354         void (*query_set_stage)(struct fd_batch *batch, enum fd_render_stage stage);
 355
 356         /* blitter: */
 357         bool (*blit)(struct fd_context *ctx, const struct pipe_blit_info *info);
 358
 359         /* handling for barriers: */
 360         void (*framebuffer_barrier)(struct fd_context *ctx);
 361
 362         /* logger: */
 363         void (*record_timestamp)(struct fd_ringbuffer *ring, struct fd_bo *bo, unsigned offset);
 364         uint64_t (*ts_to_ns)(uint64_t ts);
 365
 366         struct list_head log_chunks;  /* list of flushed log chunks in fifo order */
 367         unsigned frame_nr;            /* frame counter (for fd_log) */
 368         FILE *log_out;
 369
 370         /*
 371          * Common pre-cooked VBO state (used for a3xx and later):
 372          */
 373
 374         /* for clear/gmem->mem vertices, and mem->gmem */
 375         struct pipe_resource *solid_vbuf;
 376
 377         /* for mem->gmem tex coords: */
 378         struct pipe_resource *blit_texcoord_vbuf;
 379
 380         /* vertex state for solid_vbuf:
 381          *    - solid_vbuf / 12 / R32G32B32_FLOAT
 382          */
 383         struct fd_vertex_state solid_vbuf_state;
 384
 385         /* vertex state for blit_prog:
 386          *    - blit_texcoord_vbuf / 8 / R32G32_FLOAT
 387          *    - solid_vbuf / 12 / R32G32B32_FLOAT
 388          */
 389         struct fd_vertex_state blit_vbuf_state;
 390
 391         /*
 392          * Info about state of previous draw, for state that comes from
 393          * pipe_draw_info (ie. not part of a CSO).  This allows us to
 394          * skip some register emit when the state doesn't change from
 395          * draw-to-draw
 396          */
 397         struct {
 398                 bool dirty;               /* last draw state unknown */
 399                 bool primitive_restart;
 400                 uint32_t index_start;
 401                 uint32_t instance_start;
 402                 uint32_t restart_index;
 403         } last;
 404 };
 405
 406 static inline struct fd_context *
 407 fd_context(struct pipe_context *pctx)
 408 {
 409         return (struct fd_context *)pctx;
 410 }
 411
 412 static inline void
 413 fd_context_assert_locked(struct fd_context *ctx)
 414 {
 415         pipe_mutex_assert_locked(ctx->screen->lock);
 416 }
 417
 418 static inline void
 419 fd_context_lock(struct fd_context *ctx)
 420 {
 421         mtx_lock(&ctx->screen->lock);
 422 }
 423
 424 static inline void
 425 fd_context_unlock(struct fd_context *ctx)
 426 {
 427         mtx_unlock(&ctx->screen->lock);
 428 }
 429
 430 /* mark all state dirty: */
 431 static inline void
 432 fd_context_all_dirty(struct fd_context *ctx)
 433 {
 434         ctx->last.dirty = true;
 435         ctx->dirty = ~0;
 436         for (unsigned i = 0; i < PIPE_SHADER_TYPES; i++)
 437                 ctx->dirty_shader[i] = ~0;
 438 }
 439
 440 static inline void
 441 fd_context_all_clean(struct fd_context *ctx)
 442 {
 443         ctx->dirty = 0;
 444         for (unsigned i = 0; i < PIPE_SHADER_TYPES; i++) {
 445                 /* don't mark compute state as clean, since it is not emitted
 446                  * during normal draw call.  The places that call _all_dirty(),
 447                  * it is safe to mark compute state dirty as well, but the
 448                  * inverse is not true.
 449                  */
 450                 if (i == PIPE_SHADER_COMPUTE)
 451                         continue;
 452                 ctx->dirty_shader[i] = 0;
 453         }
 454 }
 455
 456 static inline struct pipe_scissor_state *
 457 fd_context_get_scissor(struct fd_context *ctx)
 458 {
 459         if (ctx->rasterizer && ctx->rasterizer->scissor)
 460                 return &ctx->scissor;
 461         return &ctx->disabled_scissor;
 462 }
 463
 464 static inline bool
 465 fd_supported_prim(struct fd_context *ctx, unsigned prim)
 466 {
 467         return (1 << prim) & ctx->primtype_mask;
 468 }
 469
 470 static inline struct fd_batch *
 471 fd_context_batch(struct fd_context *ctx)
 472 {
 473         if (unlikely(!ctx->batch)) {
 474                 struct fd_batch *batch =
 475                         fd_batch_from_fb(&ctx->screen->batch_cache, ctx, &ctx->framebuffer);
 476                 util_copy_framebuffer_state(&batch->framebuffer, &ctx->framebuffer);
 477                 ctx->batch = batch;
 478                 fd_context_all_dirty(ctx);
 479         }
 480         return ctx->batch;
 481 }
 482
 483 static inline void
 484 fd_batch_set_stage(struct fd_batch *batch, enum fd_render_stage stage)
 485 {
 486         struct fd_context *ctx = batch->ctx;
 487
 488         if (ctx->query_set_stage)
 489                 ctx->query_set_stage(batch, stage);
 490
 491         batch->stage = stage;
 492 }
 493
 494 void fd_context_setup_common_vbos(struct fd_context *ctx);
 495 void fd_context_cleanup_common_vbos(struct fd_context *ctx);
 496
 497 struct pipe_context * fd_context_init(struct fd_context *ctx,
 498                 struct pipe_screen *pscreen, const uint8_t *primtypes,
 499                 void *priv, unsigned flags);
 500
 501 void fd_context_destroy(struct pipe_context *pctx);
 502
 503 #endif /* FREEDRENO_CONTEXT_H_ */