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