src/gallium/drivers/freedreno/freedreno_context.h

   1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
   2
   3 /*
   4  * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  *
  25  * Authors:
  26  *    Rob Clark <robclark@freedesktop.org>
  27  */
  28
  29 #ifndef FREEDRENO_CONTEXT_H_
  30 #define FREEDRENO_CONTEXT_H_
  31
  32 #include "pipe/p_context.h"
  33 #include "indices/u_primconvert.h"
  34 #include "util/u_blitter.h"
  35 #include "util/list.h"
  36 #include "util/slab.h"
  37 #include "util/u_string.h"
  38
  39 #include "freedreno_batch.h"
  40 #include "freedreno_screen.h"
  41 #include "freedreno_gmem.h"
  42 #include "freedreno_util.h"
  43
  44 #define BORDER_COLOR_UPLOAD_SIZE (2 * PIPE_MAX_SAMPLERS * BORDERCOLOR_SIZE)
  45
  46 struct fd_vertex_stateobj;
  47
  48 struct fd_texture_stateobj {
  49         struct pipe_sampler_view *textures[PIPE_MAX_SAMPLERS];
  50         unsigned num_textures;
  51         unsigned valid_textures;
  52         struct pipe_sampler_state *samplers[PIPE_MAX_SAMPLERS];
  53         unsigned num_samplers;
  54         unsigned valid_samplers;
  55         /* number of samples per sampler, 2 bits per sampler: */
  56         uint32_t samples;
  57 };
  58
  59 struct fd_program_stateobj {
  60         void *vp, *fp;
  61
  62         /* rest only used by fd2.. split out: */
  63         uint8_t num_exports;
  64         /* Indexed by semantic name or TGSI_SEMANTIC_COUNT + semantic index
  65          * for TGSI_SEMANTIC_GENERIC.  Special vs exports (position and point-
  66          * size) are not included in this
  67          */
  68         uint8_t export_linkage[63];
  69 };
  70
  71 struct fd_constbuf_stateobj {
  72         struct pipe_constant_buffer cb[PIPE_MAX_CONSTANT_BUFFERS];
  73         uint32_t enabled_mask;
  74 };
  75
  76 struct fd_shaderbuf_stateobj {
  77         struct pipe_shader_buffer sb[PIPE_MAX_SHADER_BUFFERS];
  78         uint32_t enabled_mask;
  79 };
  80
  81 struct fd_shaderimg_stateobj {
  82         struct pipe_image_view si[PIPE_MAX_SHADER_IMAGES];
  83         uint32_t enabled_mask;
  84 };
  85
  86 struct fd_vertexbuf_stateobj {
  87         struct pipe_vertex_buffer vb[PIPE_MAX_ATTRIBS];
  88         unsigned count;
  89         uint32_t enabled_mask;
  90 };
  91
  92 struct fd_vertex_stateobj {
  93         struct pipe_vertex_element pipe[PIPE_MAX_ATTRIBS];
  94         unsigned num_elements;
  95 };
  96
  97 struct fd_streamout_stateobj {
  98         struct pipe_stream_output_target *targets[PIPE_MAX_SO_BUFFERS];
  99         unsigned num_targets;
 100         /* Track offset from vtxcnt for streamout data.  This counter
 101          * is just incremented by # of vertices on each draw until
 102          * reset or new streamout buffer bound.
 103          *
 104          * When we eventually have GS, the CPU won't actually know the
 105          * number of vertices per draw, so I think we'll have to do
 106          * something more clever.
 107          */
 108         unsigned offsets[PIPE_MAX_SO_BUFFERS];
 109 };
 110
 111 #define MAX_GLOBAL_BUFFERS 16
 112 struct fd_global_bindings_stateobj {
 113         struct pipe_resource *buf[MAX_GLOBAL_BUFFERS];
 114         uint32_t enabled_mask;
 115 };
 116
 117 /* group together the vertex and vertexbuf state.. for ease of passing
 118  * around, and because various internal operations (gmem<->mem, etc)
 119  * need their own vertex state:
 120  */
 121 struct fd_vertex_state {
 122         struct fd_vertex_stateobj *vtx;
 123         struct fd_vertexbuf_stateobj vertexbuf;
 124 };
 125
 126 /* global 3d pipeline dirty state: */
 127 enum fd_dirty_3d_state {
 128         FD_DIRTY_BLEND       = BIT(0),
 129         FD_DIRTY_RASTERIZER  = BIT(1),
 130         FD_DIRTY_ZSA         = BIT(2),
 131         FD_DIRTY_BLEND_COLOR = BIT(3),
 132         FD_DIRTY_STENCIL_REF = BIT(4),
 133         FD_DIRTY_SAMPLE_MASK = BIT(5),
 134         FD_DIRTY_FRAMEBUFFER = BIT(6),
 135         FD_DIRTY_STIPPLE     = BIT(7),
 136         FD_DIRTY_VIEWPORT    = BIT(8),
 137         FD_DIRTY_VTXSTATE    = BIT(9),
 138         FD_DIRTY_VTXBUF      = BIT(10),
 139
 140         FD_DIRTY_SCISSOR     = BIT(12),
 141         FD_DIRTY_STREAMOUT   = BIT(13),
 142         FD_DIRTY_UCP         = BIT(14),
 143         FD_DIRTY_BLEND_DUAL  = BIT(15),
 144
 145         /* These are a bit redundent with fd_dirty_shader_state, and possibly
 146          * should be removed.  (But OTOH kinda convenient in some places)
 147          */
 148         FD_DIRTY_PROG        = BIT(16),
 149         FD_DIRTY_CONST       = BIT(17),
 150         FD_DIRTY_TEX         = BIT(18),
 151
 152         /* only used by a2xx.. possibly can be removed.. */
 153         FD_DIRTY_TEXSTATE    = BIT(19),
 154 };
 155
 156 /* per shader-stage dirty state: */
 157 enum fd_dirty_shader_state {
 158         FD_DIRTY_SHADER_PROG  = BIT(0),
 159         FD_DIRTY_SHADER_CONST = BIT(1),
 160         FD_DIRTY_SHADER_TEX   = BIT(2),
 161         FD_DIRTY_SHADER_SSBO  = BIT(3),
 162         FD_DIRTY_SHADER_IMAGE = BIT(4),
 163 };
 164
 165 struct fd_context {
 166         struct pipe_context base;
 167
 168         struct fd_device *dev;
 169         struct fd_screen *screen;
 170         struct fd_pipe *pipe;
 171
 172         struct util_queue flush_queue;
 173
 174         struct blitter_context *blitter;
 175         void *clear_rs_state;
 176         struct primconvert_context *primconvert;
 177
 178         /* slab for pipe_transfer allocations: */
 179         struct slab_child_pool transfer_pool;
 180
 181         /**
 182          * query related state:
 183          */
 184         /*@{*/
 185         /* slabs for fd_hw_sample and fd_hw_sample_period allocations: */
 186         struct slab_mempool sample_pool;
 187         struct slab_mempool sample_period_pool;
 188
 189         /* sample-providers for hw queries: */
 190         const struct fd_hw_sample_provider *hw_sample_providers[MAX_HW_SAMPLE_PROVIDERS];
 191
 192         /* list of active queries: */
 193         struct list_head hw_active_queries;
 194
 195         /* sample-providers for accumulating hw queries: */
 196         const struct fd_acc_sample_provider *acc_sample_providers[MAX_HW_SAMPLE_PROVIDERS];
 197
 198         /* list of active accumulating queries: */
 199         struct list_head acc_active_queries;
 200         /*@}*/
 201
 202         /* table with PIPE_PRIM_MAX entries mapping PIPE_PRIM_x to
 203          * DI_PT_x value to use for draw initiator.  There are some
 204          * slight differences between generation:
 205          */
 206         const uint8_t *primtypes;
 207         uint32_t primtype_mask;
 208
 209         /* shaders used by clear, and gmem->mem blits: */
 210         struct fd_program_stateobj solid_prog; // TODO move to screen?
 211
 212         /* shaders used by mem->gmem blits: */
 213         struct fd_program_stateobj blit_prog[MAX_RENDER_TARGETS]; // TODO move to screen?
 214         struct fd_program_stateobj blit_z, blit_zs;
 215
 216         /* Stats/counters:
 217          */
 218         struct {
 219                 uint64_t prims_emitted;
 220                 uint64_t prims_generated;
 221                 uint64_t draw_calls;
 222                 uint64_t batch_total, batch_sysmem, batch_gmem, batch_nondraw, batch_restore;
 223                 uint64_t staging_uploads, shadow_uploads;
 224                 uint64_t vs_regs, fs_regs;
 225         } stats;
 226
 227         /* Current batch.. the rule here is that you can deref ctx->batch
 228          * in codepaths from pipe_context entrypoints.  But not in code-
 229          * paths from fd_batch_flush() (basically, the stuff that gets
 230          * called from GMEM code), since in those code-paths the batch
 231          * you care about is not necessarily the same as ctx->batch.
 232          */
 233         struct fd_batch *batch;
 234
 235         /* Are we in process of shadowing a resource? Used to detect recursion
 236          * in transfer_map, and skip unneeded synchronization.
 237          */
 238         bool in_shadow : 1;
 239
 240         /* Ie. in blit situation where we no longer care about previous framebuffer
 241          * contents.  Main point is to eliminate blits from fd_try_shadow_resource().
 242          * For example, in case of texture upload + gen-mipmaps.
 243          */
 244         bool in_blit : 1;
 245
 246         struct pipe_scissor_state scissor;
 247
 248         /* we don't have a disable/enable bit for scissor, so instead we keep
 249          * a disabled-scissor state which matches the entire bound framebuffer
 250          * and use that when scissor is not enabled.
 251          */
 252         struct pipe_scissor_state disabled_scissor;
 253
 254         /* Current gmem/tiling configuration.. gets updated on render_tiles()
 255          * if out of date with current maximal-scissor/cpp:
 256          *
 257          * (NOTE: this is kind of related to the batch, but moving it there
 258          * means we'd always have to recalc tiles ever batch)
 259          */
 260         struct fd_gmem_stateobj gmem;
 261         struct fd_vsc_pipe      vsc_pipe[16];
 262         struct fd_tile          tile[512];
 263
 264         /* which state objects need to be re-emit'd: */
 265         enum fd_dirty_3d_state dirty;
 266
 267         /* per shader-stage dirty status: */
 268         enum fd_dirty_shader_state dirty_shader[PIPE_SHADER_TYPES];
 269
 270         void *compute;
 271         struct pipe_blend_state *blend;
 272         struct pipe_rasterizer_state *rasterizer;
 273         struct pipe_depth_stencil_alpha_state *zsa;
 274
 275         struct fd_texture_stateobj tex[PIPE_SHADER_TYPES];
 276
 277         struct fd_program_stateobj prog;
 278
 279         struct fd_vertex_state vtx;
 280
 281         struct pipe_blend_color blend_color;
 282         struct pipe_stencil_ref stencil_ref;
 283         unsigned sample_mask;
 284         struct pipe_poly_stipple stipple;
 285         struct pipe_viewport_state viewport;
 286         struct fd_constbuf_stateobj constbuf[PIPE_SHADER_TYPES];
 287         struct fd_shaderbuf_stateobj shaderbuf[PIPE_SHADER_TYPES];
 288         struct fd_shaderimg_stateobj shaderimg[PIPE_SHADER_TYPES];
 289         struct fd_streamout_stateobj streamout;
 290         struct fd_global_bindings_stateobj global_bindings;
 291         struct pipe_clip_state ucp;
 292
 293         struct pipe_query *cond_query;
 294         bool cond_cond; /* inverted rendering condition */
 295         uint cond_mode;
 296
 297         struct pipe_debug_callback debug;
 298
 299         /* GMEM/tile handling fxns: */
 300         void (*emit_tile_init)(struct fd_batch *batch);
 301         void (*emit_tile_prep)(struct fd_batch *batch, struct fd_tile *tile);
 302         void (*emit_tile_mem2gmem)(struct fd_batch *batch, struct fd_tile *tile);
 303         void (*emit_tile_renderprep)(struct fd_batch *batch, struct fd_tile *tile);
 304         void (*emit_tile_gmem2mem)(struct fd_batch *batch, struct fd_tile *tile);
 305         void (*emit_tile_fini)(struct fd_batch *batch);   /* optional */
 306
 307         /* optional, for GMEM bypass: */
 308         void (*emit_sysmem_prep)(struct fd_batch *batch);
 309         void (*emit_sysmem_fini)(struct fd_batch *batch);
 310
 311         /* draw: */
 312         bool (*draw_vbo)(struct fd_context *ctx, const struct pipe_draw_info *info,
 313                         unsigned index_offset);
 314         bool (*clear)(struct fd_context *ctx, unsigned buffers,
 315                         const union pipe_color_union *color, double depth, unsigned stencil);
 316
 317         /* compute: */
 318         void (*launch_grid)(struct fd_context *ctx, const struct pipe_grid_info *info);
 319
 320         /* constant emit:  (note currently not used/needed for a2xx) */
 321         void (*emit_const)(struct fd_ringbuffer *ring, enum shader_t type,
 322                         uint32_t regid, uint32_t offset, uint32_t sizedwords,
 323                         const uint32_t *dwords, struct pipe_resource *prsc);
 324         /* emit bo addresses as constant: */
 325         void (*emit_const_bo)(struct fd_ringbuffer *ring, enum shader_t type, boolean write,
 326                         uint32_t regid, uint32_t num, struct pipe_resource **prscs, uint32_t *offsets);
 327
 328         /* indirect-branch emit: */
 329         void (*emit_ib)(struct fd_ringbuffer *ring, struct fd_ringbuffer *target);
 330
 331         /* query: */
 332         struct fd_query * (*create_query)(struct fd_context *ctx, unsigned query_type);
 333         void (*query_prepare)(struct fd_batch *batch, uint32_t num_tiles);
 334         void (*query_prepare_tile)(struct fd_batch *batch, uint32_t n,
 335                         struct fd_ringbuffer *ring);
 336         void (*query_set_stage)(struct fd_batch *batch, enum fd_render_stage stage);
 337
 338         /* blitter: */
 339         void (*blit)(struct fd_context *ctx, const struct pipe_blit_info *info);
 340
 341         /* simple gpu "memcpy": */
 342         void (*mem_to_mem)(struct fd_ringbuffer *ring, struct pipe_resource *dst,
 343                         unsigned dst_off, struct pipe_resource *src, unsigned src_off,
 344                         unsigned sizedwords);
 345
 346         /*
 347          * Common pre-cooked VBO state (used for a3xx and later):
 348          */
 349
 350         /* for clear/gmem->mem vertices, and mem->gmem */
 351         struct pipe_resource *solid_vbuf;
 352
 353         /* for mem->gmem tex coords: */
 354         struct pipe_resource *blit_texcoord_vbuf;
 355
 356         /* vertex state for solid_vbuf:
 357          *    - solid_vbuf / 12 / R32G32B32_FLOAT
 358          */
 359         struct fd_vertex_state solid_vbuf_state;
 360
 361         /* vertex state for blit_prog:
 362          *    - blit_texcoord_vbuf / 8 / R32G32_FLOAT
 363          *    - solid_vbuf / 12 / R32G32B32_FLOAT
 364          */
 365         struct fd_vertex_state blit_vbuf_state;
 366 };
 367
 368 static inline struct fd_context *
 369 fd_context(struct pipe_context *pctx)
 370 {
 371         return (struct fd_context *)pctx;
 372 }
 373
 374 static inline void
 375 fd_context_assert_locked(struct fd_context *ctx)
 376 {
 377         pipe_mutex_assert_locked(ctx->screen->lock);
 378 }
 379
 380 static inline void
 381 fd_context_lock(struct fd_context *ctx)
 382 {
 383         mtx_lock(&ctx->screen->lock);
 384 }
 385
 386 static inline void
 387 fd_context_unlock(struct fd_context *ctx)
 388 {
 389         mtx_unlock(&ctx->screen->lock);
 390 }
 391
 392 /* mark all state dirty: */
 393 static inline void
 394 fd_context_all_dirty(struct fd_context *ctx)
 395 {
 396         ctx->dirty = ~0;
 397         for (unsigned i = 0; i < PIPE_SHADER_TYPES; i++)
 398                 ctx->dirty_shader[i] = ~0;
 399 }
 400
 401 static inline void
 402 fd_context_all_clean(struct fd_context *ctx)
 403 {
 404         ctx->dirty = 0;
 405         for (unsigned i = 0; i < PIPE_SHADER_TYPES; i++) {
 406                 /* don't mark compute state as clean, since it is not emitted
 407                  * during normal draw call.  The places that call _all_dirty(),
 408                  * it is safe to mark compute state dirty as well, but the
 409                  * inverse is not true.
 410                  */
 411                 if (i == PIPE_SHADER_COMPUTE)
 412                         continue;
 413                 ctx->dirty_shader[i] = 0;
 414         }
 415 }
 416
 417 static inline struct pipe_scissor_state *
 418 fd_context_get_scissor(struct fd_context *ctx)
 419 {
 420         if (ctx->rasterizer && ctx->rasterizer->scissor)
 421                 return &ctx->scissor;
 422         return &ctx->disabled_scissor;
 423 }
 424
 425 static inline bool
 426 fd_supported_prim(struct fd_context *ctx, unsigned prim)
 427 {
 428         return (1 << prim) & ctx->primtype_mask;
 429 }
 430
 431 static inline struct fd_batch *
 432 fd_context_batch(struct fd_context *ctx)
 433 {
 434         return ctx->batch;
 435 }
 436
 437 static inline void
 438 fd_batch_set_stage(struct fd_batch *batch, enum fd_render_stage stage)
 439 {
 440         struct fd_context *ctx = batch->ctx;
 441
 442         /* special case: internal blits (like mipmap level generation)
 443          * go through normal draw path (via util_blitter_blit()).. but
 444          * we need to ignore the FD_STAGE_DRAW which will be set, so we
 445          * don't enable queries which should be paused during internal
 446          * blits:
 447          */
 448         if ((batch->stage == FD_STAGE_BLIT) &&
 449                         (stage != FD_STAGE_NULL))
 450                 return;
 451
 452         if (ctx->query_set_stage)
 453                 ctx->query_set_stage(batch, stage);
 454
 455         batch->stage = stage;
 456 }
 457
 458 void fd_context_setup_common_vbos(struct fd_context *ctx);
 459 void fd_context_cleanup_common_vbos(struct fd_context *ctx);
 460
 461 struct pipe_context * fd_context_init(struct fd_context *ctx,
 462                 struct pipe_screen *pscreen, const uint8_t *primtypes,
 463                 void *priv, unsigned flags);
 464
 465 void fd_context_destroy(struct pipe_context *pctx);
 466
 467 #endif /* FREEDRENO_CONTEXT_H_ */