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