src/gallium/drivers/radeonsi/si_state.h

   1 /*
   2  * Copyright 2012 Advanced Micro Devices, Inc.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * on the rights to use, copy, modify, merge, publish, distribute, sub
   9  * license, and/or sell copies of the Software, and to permit persons to whom
  10  * the Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  */
  24
  25 #ifndef SI_STATE_H
  26 #define SI_STATE_H
  27
  28 #include "si_pm4.h"
  29
  30 #include "pipebuffer/pb_slab.h"
  31 #include "util/u_blitter.h"
  32
  33 #define SI_NUM_GRAPHICS_SHADERS (PIPE_SHADER_TESS_EVAL+1)
  34 #define SI_NUM_SHADERS (PIPE_SHADER_COMPUTE+1)
  35
  36 #define SI_NUM_VERTEX_BUFFERS           SI_MAX_ATTRIBS
  37 #define SI_NUM_SAMPLERS                 32 /* OpenGL textures units per shader */
  38 #define SI_NUM_CONST_BUFFERS            16
  39 #define SI_NUM_IMAGES                   16
  40 #define SI_NUM_SHADER_BUFFERS           16
  41
  42 struct si_screen;
  43 struct si_shader;
  44 struct si_shader_selector;
  45 struct r600_texture;
  46 struct si_qbo_state;
  47
  48 /* State atoms are callbacks which write a sequence of packets into a GPU
  49  * command buffer (AKA indirect buffer, AKA IB, AKA command stream, AKA CS).
  50  */
  51 struct si_atom {
  52         void (*emit)(struct si_context *ctx);
  53 };
  54
  55 struct si_state_blend {
  56         struct si_pm4_state     pm4;
  57         uint32_t                cb_target_mask;
  58         /* Set 0xf or 0x0 (4 bits) per render target if the following is
  59          * true. ANDed with spi_shader_col_format.
  60          */
  61         unsigned                cb_target_enabled_4bit;
  62         unsigned                blend_enable_4bit;
  63         unsigned                need_src_alpha_4bit;
  64         unsigned                commutative_4bit;
  65         bool                    alpha_to_coverage:1;
  66         bool                    alpha_to_one:1;
  67         bool                    dual_src_blend:1;
  68         bool                    logicop_enable:1;
  69 };
  70
  71 struct si_state_rasterizer {
  72         struct si_pm4_state     pm4;
  73         /* poly offset states for 16-bit, 24-bit, and 32-bit zbuffers */
  74         struct si_pm4_state     *pm4_poly_offset;
  75         unsigned                pa_sc_line_stipple;
  76         unsigned                pa_cl_clip_cntl;
  77         float                   line_width;
  78         float                   max_point_size;
  79         unsigned                sprite_coord_enable:8;
  80         unsigned                clip_plane_enable:8;
  81         unsigned                flatshade:1;
  82         unsigned                two_side:1;
  83         unsigned                multisample_enable:1;
  84         unsigned                force_persample_interp:1;
  85         unsigned                line_stipple_enable:1;
  86         unsigned                poly_stipple_enable:1;
  87         unsigned                line_smooth:1;
  88         unsigned                poly_smooth:1;
  89         unsigned                uses_poly_offset:1;
  90         unsigned                clamp_fragment_color:1;
  91         unsigned                clamp_vertex_color:1;
  92         unsigned                rasterizer_discard:1;
  93         unsigned                scissor_enable:1;
  94         unsigned                clip_halfz:1;
  95 };
  96
  97 struct si_dsa_stencil_ref_part {
  98         uint8_t                 valuemask[2];
  99         uint8_t                 writemask[2];
 100 };
 101
 102 struct si_dsa_order_invariance {
 103         /** Whether the final result in Z/S buffers is guaranteed to be
 104          * invariant under changes to the order in which fragments arrive. */
 105         bool zs:1;
 106
 107         /** Whether the set of fragments that pass the combined Z/S test is
 108          * guaranteed to be invariant under changes to the order in which
 109          * fragments arrive. */
 110         bool pass_set:1;
 111
 112         /** Whether the last fragment that passes the combined Z/S test at each
 113          * sample is guaranteed to be invariant under changes to the order in
 114          * which fragments arrive. */
 115         bool pass_last:1;
 116 };
 117
 118 struct si_state_dsa {
 119         struct si_pm4_state             pm4;
 120         struct si_dsa_stencil_ref_part  stencil_ref;
 121
 122         /* 0 = without stencil buffer, 1 = when both Z and S buffers are present */
 123         struct si_dsa_order_invariance  order_invariance[2];
 124
 125         ubyte                           alpha_func:3;
 126         bool                            depth_enabled:1;
 127         bool                            depth_write_enabled:1;
 128         bool                            stencil_enabled:1;
 129         bool                            stencil_write_enabled:1;
 130         bool                            db_can_write:1;
 131
 132 };
 133
 134 struct si_stencil_ref {
 135         struct pipe_stencil_ref         state;
 136         struct si_dsa_stencil_ref_part  dsa_part;
 137 };
 138
 139 struct si_vertex_elements
 140 {
 141         uint32_t                        instance_divisors[SI_MAX_ATTRIBS];
 142         uint32_t                        rsrc_word3[SI_MAX_ATTRIBS];
 143         uint16_t                        src_offset[SI_MAX_ATTRIBS];
 144         uint8_t                         fix_fetch[SI_MAX_ATTRIBS];
 145         uint8_t                         format_size[SI_MAX_ATTRIBS];
 146         uint8_t                         vertex_buffer_index[SI_MAX_ATTRIBS];
 147
 148         uint8_t                         count;
 149         bool                            uses_instance_divisors;
 150
 151         uint16_t                        first_vb_use_mask;
 152         /* Vertex buffer descriptor list size aligned for optimal prefetch. */
 153         uint16_t                        desc_list_byte_size;
 154         uint16_t                        instance_divisor_is_one; /* bitmask of inputs */
 155         uint16_t                        instance_divisor_is_fetched;  /* bitmask of inputs */
 156 };
 157
 158 union si_state {
 159         struct {
 160                 struct si_state_blend           *blend;
 161                 struct si_state_rasterizer      *rasterizer;
 162                 struct si_state_dsa             *dsa;
 163                 struct si_pm4_state             *poly_offset;
 164                 struct si_pm4_state             *ls;
 165                 struct si_pm4_state             *hs;
 166                 struct si_pm4_state             *es;
 167                 struct si_pm4_state             *gs;
 168                 struct si_pm4_state             *vgt_shader_config;
 169                 struct si_pm4_state             *vs;
 170                 struct si_pm4_state             *ps;
 171         } named;
 172         struct si_pm4_state     *array[0];
 173 };
 174
 175 #define SI_NUM_STATES (sizeof(union si_state) / sizeof(struct si_pm4_state *))
 176
 177 union si_state_atoms {
 178         struct {
 179                 /* The order matters. */
 180                 struct si_atom render_cond;
 181                 struct si_atom streamout_begin;
 182                 struct si_atom streamout_enable; /* must be after streamout_begin */
 183                 struct si_atom framebuffer;
 184                 struct si_atom msaa_sample_locs;
 185                 struct si_atom db_render_state;
 186                 struct si_atom dpbb_state;
 187                 struct si_atom msaa_config;
 188                 struct si_atom sample_mask;
 189                 struct si_atom cb_render_state;
 190                 struct si_atom blend_color;
 191                 struct si_atom clip_regs;
 192                 struct si_atom clip_state;
 193                 struct si_atom shader_pointers;
 194                 struct si_atom scissors;
 195                 struct si_atom viewports;
 196                 struct si_atom stencil_ref;
 197                 struct si_atom spi_map;
 198                 struct si_atom scratch_state;
 199         } s;
 200         struct si_atom array[0];
 201 };
 202
 203 #define SI_NUM_ATOMS (sizeof(union si_state_atoms)/sizeof(struct si_atom*))
 204
 205 struct si_shader_data {
 206         uint32_t                sh_base[SI_NUM_SHADERS];
 207 };
 208
 209 /* The list of registers whose emitted values are remembered by si_context. */
 210 enum si_tracked_reg {
 211         SI_TRACKED_DB_RENDER_CONTROL, /* 2 consecutive registers */
 212         SI_TRACKED_DB_COUNT_CONTROL,
 213
 214         SI_TRACKED_DB_RENDER_OVERRIDE2,
 215         SI_TRACKED_DB_SHADER_CONTROL,
 216
 217         SI_TRACKED_CB_TARGET_MASK,
 218         SI_TRACKED_CB_DCC_CONTROL,
 219
 220         SI_TRACKED_SX_PS_DOWNCONVERT, /* 3 consecutive registers */
 221         SI_TRACKED_SX_BLEND_OPT_EPSILON,
 222         SI_TRACKED_SX_BLEND_OPT_CONTROL,
 223
 224         SI_TRACKED_PA_SC_LINE_CNTL, /* 2 consecutive registers */
 225         SI_TRACKED_PA_SC_AA_CONFIG,
 226
 227         SI_TRACKED_DB_EQAA,
 228         SI_TRACKED_PA_SC_MODE_CNTL_1,
 229
 230         SI_TRACKED_PA_SU_SMALL_PRIM_FILTER_CNTL,
 231
 232         SI_TRACKED_PA_CL_VS_OUT_CNTL,
 233         SI_TRACKED_PA_CL_CLIP_CNTL,
 234
 235         SI_NUM_TRACKED_REGS,
 236 };
 237
 238 struct si_tracked_regs {
 239         uint32_t                reg_saved;
 240         uint32_t                reg_value[SI_NUM_TRACKED_REGS];
 241 };
 242
 243 /* Private read-write buffer slots. */
 244 enum {
 245         SI_ES_RING_ESGS,
 246         SI_GS_RING_ESGS,
 247
 248         SI_RING_GSVS,
 249
 250         SI_VS_STREAMOUT_BUF0,
 251         SI_VS_STREAMOUT_BUF1,
 252         SI_VS_STREAMOUT_BUF2,
 253         SI_VS_STREAMOUT_BUF3,
 254
 255         SI_HS_CONST_DEFAULT_TESS_LEVELS,
 256         SI_VS_CONST_INSTANCE_DIVISORS,
 257         SI_VS_CONST_CLIP_PLANES,
 258         SI_PS_CONST_POLY_STIPPLE,
 259         SI_PS_CONST_SAMPLE_POSITIONS,
 260
 261         /* Image descriptor of color buffer 0 for KHR_blend_equation_advanced. */
 262         SI_PS_IMAGE_COLORBUF0,
 263         SI_PS_IMAGE_COLORBUF0_HI,
 264         SI_PS_IMAGE_COLORBUF0_FMASK,
 265         SI_PS_IMAGE_COLORBUF0_FMASK_HI,
 266
 267         SI_NUM_RW_BUFFERS,
 268 };
 269
 270 /* Indices into sctx->descriptors, laid out so that gfx and compute pipelines
 271  * are contiguous:
 272  *
 273  *  0 - rw buffers
 274  *  1 - vertex const and shader buffers
 275  *  2 - vertex samplers and images
 276  *  3 - fragment const and shader buffer
 277  *   ...
 278  *  11 - compute const and shader buffers
 279  *  12 - compute samplers and images
 280  */
 281 enum {
 282         SI_SHADER_DESCS_CONST_AND_SHADER_BUFFERS,
 283         SI_SHADER_DESCS_SAMPLERS_AND_IMAGES,
 284         SI_NUM_SHADER_DESCS,
 285 };
 286
 287 #define SI_DESCS_RW_BUFFERS            0
 288 #define SI_DESCS_FIRST_SHADER          1
 289 #define SI_DESCS_FIRST_COMPUTE         (SI_DESCS_FIRST_SHADER + \
 290                                         PIPE_SHADER_COMPUTE * SI_NUM_SHADER_DESCS)
 291 #define SI_NUM_DESCS                   (SI_DESCS_FIRST_SHADER + \
 292                                         SI_NUM_SHADERS * SI_NUM_SHADER_DESCS)
 293
 294 #define SI_DESCS_SHADER_MASK(name) \
 295         u_bit_consecutive(SI_DESCS_FIRST_SHADER + \
 296                           PIPE_SHADER_##name * SI_NUM_SHADER_DESCS, \
 297                           SI_NUM_SHADER_DESCS)
 298
 299 /* This represents descriptors in memory, such as buffer resources,
 300  * image resources, and sampler states.
 301  */
 302 struct si_descriptors {
 303         /* The list of descriptors in malloc'd memory. */
 304         uint32_t *list;
 305         /* The list in mapped GPU memory. */
 306         uint32_t *gpu_list;
 307
 308         /* The buffer where the descriptors have been uploaded. */
 309         struct r600_resource *buffer;
 310         uint64_t gpu_address;
 311
 312         /* The maximum number of descriptors. */
 313         uint32_t num_elements;
 314
 315         /* Slots that are used by currently-bound shaders.
 316          * It determines which slots are uploaded.
 317          */
 318         uint32_t first_active_slot;
 319         uint32_t num_active_slots;
 320
 321         /* The SH register offset relative to USER_DATA*_0 where the pointer
 322          * to the descriptor array will be stored. */
 323         short shader_userdata_offset;
 324         /* The size of one descriptor. */
 325         ubyte element_dw_size;
 326         /* If there is only one slot enabled, bind it directly instead of
 327          * uploading descriptors. -1 if disabled. */
 328         signed char slot_index_to_bind_directly;
 329 };
 330
 331 struct si_buffer_resources {
 332         struct pipe_resource            **buffers; /* this has num_buffers elements */
 333
 334         enum radeon_bo_usage            shader_usage:4; /* READ, WRITE, or READWRITE */
 335         enum radeon_bo_usage            shader_usage_constbuf:4;
 336         enum radeon_bo_priority         priority:6;
 337         enum radeon_bo_priority         priority_constbuf:6;
 338
 339         /* The i-th bit is set if that element is enabled (non-NULL resource). */
 340         unsigned                        enabled_mask;
 341 };
 342
 343 #define si_pm4_block_idx(member) \
 344         (offsetof(union si_state, named.member) / sizeof(struct si_pm4_state *))
 345
 346 #define si_pm4_state_changed(sctx, member) \
 347         ((sctx)->queued.named.member != (sctx)->emitted.named.member)
 348
 349 #define si_pm4_state_enabled_and_changed(sctx, member) \
 350         ((sctx)->queued.named.member && si_pm4_state_changed(sctx, member))
 351
 352 #define si_pm4_bind_state(sctx, member, value) \
 353         do { \
 354                 (sctx)->queued.named.member = (value); \
 355                 (sctx)->dirty_states |= 1 << si_pm4_block_idx(member); \
 356         } while(0)
 357
 358 #define si_pm4_delete_state(sctx, member, value) \
 359         do { \
 360                 if ((sctx)->queued.named.member == (value)) { \
 361                         (sctx)->queued.named.member = NULL; \
 362                 } \
 363                 si_pm4_free_state(sctx, (struct si_pm4_state *)(value), \
 364                                   si_pm4_block_idx(member)); \
 365         } while(0)
 366
 367 /* si_descriptors.c */
 368 void si_set_mutable_tex_desc_fields(struct si_screen *sscreen,
 369                                     struct r600_texture *tex,
 370                                     const struct legacy_surf_level *base_level_info,
 371                                     unsigned base_level, unsigned first_level,
 372                                     unsigned block_width, bool is_stencil,
 373                                     uint32_t *state);
 374 void si_update_ps_colorbuf0_slot(struct si_context *sctx);
 375 void si_get_pipe_constant_buffer(struct si_context *sctx, uint shader,
 376                                  uint slot, struct pipe_constant_buffer *cbuf);
 377 void si_get_shader_buffers(struct si_context *sctx,
 378                            enum pipe_shader_type shader,
 379                            uint start_slot, uint count,
 380                            struct pipe_shader_buffer *sbuf);
 381 void si_set_ring_buffer(struct si_context *sctx, uint slot,
 382                         struct pipe_resource *buffer,
 383                         unsigned stride, unsigned num_records,
 384                         bool add_tid, bool swizzle,
 385                         unsigned element_size, unsigned index_stride, uint64_t offset);
 386 void si_init_all_descriptors(struct si_context *sctx);
 387 bool si_upload_vertex_buffer_descriptors(struct si_context *sctx);
 388 bool si_upload_graphics_shader_descriptors(struct si_context *sctx);
 389 bool si_upload_compute_shader_descriptors(struct si_context *sctx);
 390 void si_release_all_descriptors(struct si_context *sctx);
 391 void si_all_descriptors_begin_new_cs(struct si_context *sctx);
 392 void si_all_resident_buffers_begin_new_cs(struct si_context *sctx);
 393 void si_upload_const_buffer(struct si_context *sctx, struct r600_resource **rbuffer,
 394                             const uint8_t *ptr, unsigned size, uint32_t *const_offset);
 395 void si_update_all_texture_descriptors(struct si_context *sctx);
 396 void si_shader_change_notify(struct si_context *sctx);
 397 void si_update_needs_color_decompress_masks(struct si_context *sctx);
 398 void si_emit_graphics_shader_pointers(struct si_context *sctx);
 399 void si_emit_compute_shader_pointers(struct si_context *sctx);
 400 void si_set_rw_buffer(struct si_context *sctx,
 401                       uint slot, const struct pipe_constant_buffer *input);
 402 void si_set_active_descriptors(struct si_context *sctx, unsigned desc_idx,
 403                                uint64_t new_active_mask);
 404 void si_set_active_descriptors_for_shader(struct si_context *sctx,
 405                                           struct si_shader_selector *sel);
 406 bool si_bindless_descriptor_can_reclaim_slab(void *priv,
 407                                              struct pb_slab_entry *entry);
 408 struct pb_slab *si_bindless_descriptor_slab_alloc(void *priv, unsigned heap,
 409                                                   unsigned entry_size,
 410                                                   unsigned group_index);
 411 void si_bindless_descriptor_slab_free(void *priv, struct pb_slab *pslab);
 412 void si_rebind_buffer(struct si_context *sctx, struct pipe_resource *buf,
 413                       uint64_t old_va);
 414 /* si_state.c */
 415 void si_init_state_functions(struct si_context *sctx);
 416 void si_init_screen_state_functions(struct si_screen *sscreen);
 417 void
 418 si_make_buffer_descriptor(struct si_screen *screen, struct r600_resource *buf,
 419                           enum pipe_format format,
 420                           unsigned offset, unsigned size,
 421                           uint32_t *state);
 422 void
 423 si_make_texture_descriptor(struct si_screen *screen,
 424                            struct r600_texture *tex,
 425                            bool sampler,
 426                            enum pipe_texture_target target,
 427                            enum pipe_format pipe_format,
 428                            const unsigned char state_swizzle[4],
 429                            unsigned first_level, unsigned last_level,
 430                            unsigned first_layer, unsigned last_layer,
 431                            unsigned width, unsigned height, unsigned depth,
 432                            uint32_t *state,
 433                            uint32_t *fmask_state);
 434 struct pipe_sampler_view *
 435 si_create_sampler_view_custom(struct pipe_context *ctx,
 436                               struct pipe_resource *texture,
 437                               const struct pipe_sampler_view *state,
 438                               unsigned width0, unsigned height0,
 439                               unsigned force_level);
 440 void si_update_fb_dirtiness_after_rendering(struct si_context *sctx);
 441 void si_update_ps_iter_samples(struct si_context *sctx);
 442 void si_save_qbo_state(struct si_context *sctx, struct si_qbo_state *st);
 443 void si_set_occlusion_query_state(struct si_context *sctx,
 444                                   bool old_perfect_enable);
 445
 446 /* si_state_binning.c */
 447 void si_emit_dpbb_state(struct si_context *sctx);
 448
 449 /* si_state_shaders.c */
 450 bool si_update_shaders(struct si_context *sctx);
 451 void si_init_shader_functions(struct si_context *sctx);
 452 bool si_init_shader_cache(struct si_screen *sscreen);
 453 void si_destroy_shader_cache(struct si_screen *sscreen);
 454 void si_get_active_slot_masks(const struct tgsi_shader_info *info,
 455                               uint32_t *const_and_shader_buffers,
 456                               uint64_t *samplers_and_images);
 457 void *si_get_blit_vs(struct si_context *sctx, enum blitter_attrib_type type,
 458                      unsigned num_layers);
 459
 460 /* si_state_draw.c */
 461 void si_init_ia_multi_vgt_param_table(struct si_context *sctx);
 462 void si_emit_cache_flush(struct si_context *sctx);
 463 void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *dinfo);
 464 void si_draw_rectangle(struct blitter_context *blitter,
 465                        void *vertex_elements_cso,
 466                        blitter_get_vs_func get_vs,
 467                        int x1, int y1, int x2, int y2,
 468                        float depth, unsigned num_instances,
 469                        enum blitter_attrib_type type,
 470                        const union blitter_attrib *attrib);
 471 void si_trace_emit(struct si_context *sctx);
 472
 473 /* si_state_msaa.c */
 474 void si_init_msaa_functions(struct si_context *sctx);
 475 void si_emit_sample_locations(struct radeon_winsys_cs *cs, int nr_samples);
 476
 477 /* si_state_streamout.c */
 478 void si_streamout_buffers_dirty(struct si_context *sctx);
 479 void si_emit_streamout_end(struct si_context *sctx);
 480 void si_update_prims_generated_query_state(struct si_context *sctx,
 481                                            unsigned type, int diff);
 482 void si_init_streamout_functions(struct si_context *sctx);
 483
 484
 485 static inline unsigned si_get_constbuf_slot(unsigned slot)
 486 {
 487         /* Constant buffers are in slots [16..31], ascending */
 488         return SI_NUM_SHADER_BUFFERS + slot;
 489 }
 490
 491 static inline unsigned si_get_shaderbuf_slot(unsigned slot)
 492 {
 493         /* shader buffers are in slots [15..0], descending */
 494         return SI_NUM_SHADER_BUFFERS - 1 - slot;
 495 }
 496
 497 static inline unsigned si_get_sampler_slot(unsigned slot)
 498 {
 499         /* samplers are in slots [8..39], ascending */
 500         return SI_NUM_IMAGES / 2 + slot;
 501 }
 502
 503 static inline unsigned si_get_image_slot(unsigned slot)
 504 {
 505         /* images are in slots [15..0] (sampler slots [7..0]), descending */
 506         return SI_NUM_IMAGES - 1 - slot;
 507 }
 508
 509 #endif