projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
r600g: Implement scratch buffer state management (v2)
[mesa.git] / src / gallium / drivers / r600 / r600_pipe.h
1 /*
2 * Copyright 2010 Jerome Glisse <glisse@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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 *
23 * Authors:
24 * Jerome Glisse
25 */
26 #ifndef R600_PIPE_H
27 #define R600_PIPE_H
28
29 #include "r600_pipe_common.h"
30 #include "r600_cs.h"
31 #include "r600_public.h"
32 #include "pipe/p_defines.h"
33
34 #include "util/u_suballoc.h"
35 #include "util/list.h"
36 #include "util/u_transfer.h"
37 #include "util/u_memory.h"
38
39 #include "tgsi/tgsi_scan.h"
40
41 #define R600_NUM_ATOMS 56
42
43 #define R600_MAX_IMAGES 8
44 /*
45 * ranges reserved for images on evergreen
46 * first set for the immediate buffers,
47 * second for the actual resources for RESQ.
48 */
49 #define R600_IMAGE_IMMED_RESOURCE_OFFSET 160
50 #define R600_IMAGE_REAL_RESOURCE_OFFSET 168
51
52 /* read caches */
53 #define R600_CONTEXT_INV_VERTEX_CACHE (R600_CONTEXT_PRIVATE_FLAG << 0)
54 #define R600_CONTEXT_INV_TEX_CACHE (R600_CONTEXT_PRIVATE_FLAG << 1)
55 #define R600_CONTEXT_INV_CONST_CACHE (R600_CONTEXT_PRIVATE_FLAG << 2)
56 /* read-write caches */
57 #define R600_CONTEXT_FLUSH_AND_INV (R600_CONTEXT_PRIVATE_FLAG << 3)
58 #define R600_CONTEXT_FLUSH_AND_INV_CB_META (R600_CONTEXT_PRIVATE_FLAG << 4)
59 #define R600_CONTEXT_FLUSH_AND_INV_DB_META (R600_CONTEXT_PRIVATE_FLAG << 5)
60 #define R600_CONTEXT_FLUSH_AND_INV_DB (R600_CONTEXT_PRIVATE_FLAG << 6)
61 #define R600_CONTEXT_FLUSH_AND_INV_CB (R600_CONTEXT_PRIVATE_FLAG << 7)
62 /* engine synchronization */
63 #define R600_CONTEXT_PS_PARTIAL_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 8)
64 #define R600_CONTEXT_WAIT_3D_IDLE (R600_CONTEXT_PRIVATE_FLAG << 9)
65 #define R600_CONTEXT_WAIT_CP_DMA_IDLE (R600_CONTEXT_PRIVATE_FLAG << 10)
66 #define R600_CONTEXT_CS_PARTIAL_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 11)
67
68 /* the number of CS dwords for flushing and drawing */
69 #define R600_MAX_FLUSH_CS_DWORDS 18
70 #define R600_MAX_DRAW_CS_DWORDS 58
71 #define R600_MAX_PFP_SYNC_ME_DWORDS 16
72
73 #define EG_MAX_ATOMIC_BUFFERS 8
74
75 #define R600_MAX_USER_CONST_BUFFERS 15
76 #define R600_MAX_DRIVER_CONST_BUFFERS 3
77 #define R600_MAX_CONST_BUFFERS (R600_MAX_USER_CONST_BUFFERS + R600_MAX_DRIVER_CONST_BUFFERS)
78 #define R600_MAX_HW_CONST_BUFFERS 16
79
80 /* start driver buffers after user buffers */
81 #define R600_BUFFER_INFO_CONST_BUFFER (R600_MAX_USER_CONST_BUFFERS)
82 #define R600_UCP_SIZE (4*4*8)
83 #define R600_CS_BLOCK_GRID_SIZE (8 * 4)
84 #define R600_TCS_DEFAULT_LEVELS_SIZE (6 * 4)
85 #define R600_BUFFER_INFO_OFFSET (R600_UCP_SIZE)
86
87 /*
88 * We only access this buffer through vtx clauses hence it's fine to exist
89 * at index beyond 15.
90 */
91 #define R600_LDS_INFO_CONST_BUFFER (R600_MAX_USER_CONST_BUFFERS + 1)
92 /*
93 * Note GS doesn't use a constant buffer binding, just a resource index,
94 * so it's fine to have it exist at index beyond 15. I.e. it's not actually
95 * a const buffer, just a buffer resource.
96 */
97 #define R600_GS_RING_CONST_BUFFER (R600_MAX_USER_CONST_BUFFERS + 2)
98 /* Currently R600_MAX_CONST_BUFFERS just fits on the hw, which has a limit
99 * of 16 const buffers.
100 * UCP/SAMPLE_POSITIONS are never accessed by same shader stage so they can use the same id.
101 *
102 * In order to support d3d 11 mandated minimum of 15 user const buffers
103 * we'd have to squash all use cases into one driver buffer.
104 */
105 #define R600_MAX_CONST_BUFFER_SIZE (4096 * sizeof(float[4]))
106
107 /* HW stages */
108 #define R600_HW_STAGE_PS 0
109 #define R600_HW_STAGE_VS 1
110 #define R600_HW_STAGE_GS 2
111 #define R600_HW_STAGE_ES 3
112 #define EG_HW_STAGE_LS 4
113 #define EG_HW_STAGE_HS 5
114
115 #define R600_NUM_HW_STAGES 4
116 #define EG_NUM_HW_STAGES 6
117
118 struct r600_context;
119 struct r600_bytecode;
120 union r600_shader_key;
121
122 /* This is an atom containing GPU commands that never change.
123 * This is supposed to be copied directly into the CS. */
124 struct r600_command_buffer {
125 uint32_t *buf;
126 unsigned num_dw;
127 unsigned max_num_dw;
128 unsigned pkt_flags;
129 };
130
131 struct r600_db_state {
132 struct r600_atom atom;
133 struct r600_surface *rsurf;
134 };
135
136 struct r600_db_misc_state {
137 struct r600_atom atom;
138 bool occlusion_queries_disabled;
139 bool flush_depthstencil_through_cb;
140 bool flush_depth_inplace;
141 bool flush_stencil_inplace;
142 bool copy_depth, copy_stencil;
143 unsigned copy_sample;
144 unsigned log_samples;
145 unsigned db_shader_control;
146 bool htile_clear;
147 uint8_t ps_conservative_z;
148 };
149
150 struct r600_cb_misc_state {
151 struct r600_atom atom;
152 unsigned cb_color_control; /* this comes from blend state */
153 unsigned blend_colormask; /* 8*4 bits for 8 RGBA colorbuffers */
154 unsigned nr_cbufs;
155 unsigned bound_cbufs_target_mask;
156 unsigned nr_ps_color_outputs;
157 unsigned ps_color_export_mask;
158 unsigned image_rat_enabled_mask;
159 unsigned buffer_rat_enabled_mask;
160 bool multiwrite;
161 bool dual_src_blend;
162 };
163
164 struct r600_clip_misc_state {
165 struct r600_atom atom;
166 unsigned pa_cl_clip_cntl; /* from rasterizer */
167 unsigned pa_cl_vs_out_cntl; /* from vertex shader */
168 unsigned clip_plane_enable; /* from rasterizer */
169 unsigned cc_dist_mask; /* from vertex shader */
170 unsigned clip_dist_write; /* from vertex shader */
171 unsigned cull_dist_write; /* from vertex shader */
172 boolean clip_disable; /* from vertex shader */
173 boolean vs_out_viewport; /* from vertex shader */
174 };
175
176 struct r600_alphatest_state {
177 struct r600_atom atom;
178 unsigned sx_alpha_test_control; /* this comes from dsa state */
179 unsigned sx_alpha_ref; /* this comes from dsa state */
180 bool bypass;
181 bool cb0_export_16bpc; /* from set_framebuffer_state */
182 };
183
184 struct r600_vgt_state {
185 struct r600_atom atom;
186 uint32_t vgt_multi_prim_ib_reset_en;
187 uint32_t vgt_multi_prim_ib_reset_indx;
188 uint32_t vgt_indx_offset;
189 bool last_draw_was_indirect;
190 };
191
192 struct r600_blend_color {
193 struct r600_atom atom;
194 struct pipe_blend_color state;
195 };
196
197 struct r600_clip_state {
198 struct r600_atom atom;
199 struct pipe_clip_state state;
200 };
201
202 struct r600_cs_shader_state {
203 struct r600_atom atom;
204 unsigned kernel_index;
205 unsigned pc;
206 struct r600_pipe_compute *shader;
207 };
208
209 struct r600_framebuffer {
210 struct r600_atom atom;
211 struct pipe_framebuffer_state state;
212 unsigned compressed_cb_mask;
213 unsigned nr_samples;
214 bool export_16bpc;
215 bool cb0_is_integer;
216 bool is_msaa_resolve;
217 bool dual_src_blend;
218 bool do_update_surf_dirtiness;
219 };
220
221 struct r600_sample_mask {
222 struct r600_atom atom;
223 uint16_t sample_mask; /* there are only 8 bits on EG, 16 bits on Cayman */
224 };
225
226 struct r600_config_state {
227 struct r600_atom atom;
228 unsigned sq_gpr_resource_mgmt_1;
229 unsigned sq_gpr_resource_mgmt_2;
230 unsigned sq_gpr_resource_mgmt_3;
231 bool dyn_gpr_enabled;
232 };
233
234 struct r600_stencil_ref
235 {
236 ubyte ref_value[2];
237 ubyte valuemask[2];
238 ubyte writemask[2];
239 };
240
241 struct r600_stencil_ref_state {
242 struct r600_atom atom;
243 struct r600_stencil_ref state;
244 struct pipe_stencil_ref pipe_state;
245 };
246
247 struct r600_shader_stages_state {
248 struct r600_atom atom;
249 unsigned geom_enable;
250 };
251
252 struct r600_gs_rings_state {
253 struct r600_atom atom;
254 unsigned enable;
255 struct pipe_constant_buffer esgs_ring;
256 struct pipe_constant_buffer gsvs_ring;
257 };
258
259 /* This must start from 16. */
260 /* features */
261 #define DBG_NO_CP_DMA (1 << 30)
262 /* shader backend */
263 #define DBG_NO_SB (1 << 21)
264 #define DBG_SB_CS (1 << 22)
265 #define DBG_SB_DRY_RUN (1 << 23)
266 #define DBG_SB_STAT (1 << 24)
267 #define DBG_SB_DUMP (1 << 25)
268 #define DBG_SB_NO_FALLBACK (1 << 26)
269 #define DBG_SB_DISASM (1 << 27)
270 #define DBG_SB_SAFEMATH (1 << 28)
271
272 struct r600_screen {
273 struct r600_common_screen b;
274 bool has_msaa;
275 bool has_compressed_msaa_texturing;
276 bool has_atomics;
277
278 /*for compute global memory binding, we allocate stuff here, instead of
279 * buffers.
280 * XXX: Not sure if this is the best place for global_pool. Also,
281 * it's not thread safe, so it won't work with multiple contexts. */
282 struct compute_memory_pool *global_pool;
283 };
284
285 struct r600_pipe_sampler_view {
286 struct pipe_sampler_view base;
287 struct list_head list;
288 struct r600_resource *tex_resource;
289 uint32_t tex_resource_words[8];
290 bool skip_mip_address_reloc;
291 bool is_stencil_sampler;
292 };
293
294 struct r600_rasterizer_state {
295 struct r600_command_buffer buffer;
296 boolean flatshade;
297 boolean two_side;
298 unsigned sprite_coord_enable;
299 unsigned clip_plane_enable;
300 unsigned pa_sc_line_stipple;
301 unsigned pa_cl_clip_cntl;
302 unsigned pa_su_sc_mode_cntl;
303 float offset_units;
304 float offset_scale;
305 bool offset_enable;
306 bool offset_units_unscaled;
307 bool scissor_enable;
308 bool multisample_enable;
309 bool clip_halfz;
310 bool rasterizer_discard;
311 };
312
313 struct r600_poly_offset_state {
314 struct r600_atom atom;
315 enum pipe_format zs_format;
316 float offset_units;
317 float offset_scale;
318 bool offset_units_unscaled;
319 };
320
321 struct r600_blend_state {
322 struct r600_command_buffer buffer;
323 struct r600_command_buffer buffer_no_blend;
324 unsigned cb_target_mask;
325 unsigned cb_color_control;
326 unsigned cb_color_control_no_blend;
327 bool dual_src_blend;
328 bool alpha_to_one;
329 };
330
331 struct r600_dsa_state {
332 struct r600_command_buffer buffer;
333 unsigned alpha_ref;
334 ubyte valuemask[2];
335 ubyte writemask[2];
336 unsigned zwritemask;
337 unsigned sx_alpha_test_control;
338 };
339
340 struct r600_pipe_shader;
341
342 struct r600_pipe_shader_selector {
343 struct r600_pipe_shader *current;
344
345 struct tgsi_token *tokens;
346 struct pipe_stream_output_info so;
347 struct tgsi_shader_info info;
348
349 unsigned num_shaders;
350
351 enum pipe_shader_type type;
352
353 /* geometry shader properties */
354 enum pipe_prim_type gs_output_prim;
355 unsigned gs_max_out_vertices;
356 unsigned gs_num_invocations;
357
358 /* TCS/VS */
359 uint64_t lds_patch_outputs_written_mask;
360 uint64_t lds_outputs_written_mask;
361 unsigned nr_ps_max_color_exports;
362 };
363
364 struct r600_pipe_sampler_state {
365 uint32_t tex_sampler_words[3];
366 union pipe_color_union border_color;
367 bool border_color_use;
368 bool seamless_cube_map;
369 };
370
371 /* needed for blitter save */
372 #define NUM_TEX_UNITS 16
373
374 struct r600_seamless_cube_map {
375 struct r600_atom atom;
376 bool enabled;
377 };
378
379 struct r600_samplerview_state {
380 struct r600_atom atom;
381 struct r600_pipe_sampler_view *views[NUM_TEX_UNITS];
382 uint32_t enabled_mask;
383 uint32_t dirty_mask;
384 uint32_t compressed_depthtex_mask; /* which textures are depth */
385 uint32_t compressed_colortex_mask;
386 boolean dirty_buffer_constants;
387 };
388
389 struct r600_sampler_states {
390 struct r600_atom atom;
391 struct r600_pipe_sampler_state *states[NUM_TEX_UNITS];
392 uint32_t enabled_mask;
393 uint32_t dirty_mask;
394 uint32_t has_bordercolor_mask; /* which states contain the border color */
395 };
396
397 struct r600_textures_info {
398 struct r600_samplerview_state views;
399 struct r600_sampler_states states;
400 bool is_array_sampler[NUM_TEX_UNITS];
401 };
402
403 struct r600_shader_driver_constants_info {
404 /* currently 128 bytes for UCP/samplepos + sampler buffer constants */
405 uint32_t *constants;
406 uint32_t alloc_size;
407 bool texture_const_dirty;
408 bool vs_ucp_dirty;
409 bool ps_sample_pos_dirty;
410 bool cs_block_grid_size_dirty;
411 bool tcs_default_levels_dirty;
412 };
413
414 struct r600_constbuf_state
415 {
416 struct r600_atom atom;
417 struct pipe_constant_buffer cb[PIPE_MAX_CONSTANT_BUFFERS];
418 uint32_t enabled_mask;
419 uint32_t dirty_mask;
420 };
421
422 struct r600_vertexbuf_state
423 {
424 struct r600_atom atom;
425 struct pipe_vertex_buffer vb[PIPE_MAX_ATTRIBS];
426 uint32_t enabled_mask; /* non-NULL buffers */
427 uint32_t dirty_mask;
428 };
429
430 /* CSO (constant state object, in other words, immutable state). */
431 struct r600_cso_state
432 {
433 struct r600_atom atom;
434 void *cso; /* e.g. r600_blend_state */
435 struct r600_command_buffer *cb;
436 };
437
438 struct r600_fetch_shader {
439 struct r600_resource *buffer;
440 unsigned offset;
441 };
442
443 struct r600_shader_state {
444 struct r600_atom atom;
445 struct r600_pipe_shader *shader;
446 };
447
448 struct r600_atomic_buffer_state {
449 uint32_t enabled_mask;
450 uint32_t dirty_mask;
451 struct pipe_shader_buffer buffer[EG_MAX_ATOMIC_BUFFERS];
452 };
453
454 struct r600_image_view {
455 struct pipe_image_view base;
456 uint32_t cb_color_base;
457 uint32_t cb_color_pitch;
458 uint32_t cb_color_slice;
459 uint32_t cb_color_view;
460 uint32_t cb_color_info;
461 uint32_t cb_color_attrib;
462 uint32_t cb_color_dim;
463 uint32_t cb_color_fmask;
464 uint32_t cb_color_fmask_slice;
465 uint32_t immed_resource_words[8];
466 uint32_t resource_words[8];
467 bool skip_mip_address_reloc;
468 uint32_t buf_size;
469 };
470
471 struct r600_image_state {
472 struct r600_atom atom;
473 uint32_t enabled_mask;
474 uint32_t dirty_mask;
475 uint32_t compressed_depthtex_mask;
476 uint32_t compressed_colortex_mask;
477 boolean dirty_buffer_constants;
478 struct r600_image_view views[R600_MAX_IMAGES];
479 };
480
481 /* Used to spill shader temps */
482 struct r600_scratch_buffer {
483 struct r600_resource *buffer;
484 boolean dirty;
485 unsigned size;
486 unsigned item_size;
487 };
488
489 struct r600_context {
490 struct r600_common_context b;
491 struct r600_screen *screen;
492 struct blitter_context *blitter;
493 struct u_suballocator *allocator_fetch_shader;
494
495 /* Hardware info. */
496 boolean has_vertex_cache;
497 unsigned default_gprs[EG_NUM_HW_STAGES];
498 unsigned current_gprs[EG_NUM_HW_STAGES];
499 unsigned r6xx_num_clause_temp_gprs;
500
501 /* Miscellaneous state objects. */
502 void *custom_dsa_flush;
503 void *custom_blend_resolve;
504 void *custom_blend_decompress;
505 void *custom_blend_fastclear;
506 /* With rasterizer discard, there doesn't have to be a pixel shader.
507 * In that case, we bind this one: */
508 void *dummy_pixel_shader;
509 /* These dummy CMASK and FMASK buffers are used to get around the R6xx hardware
510 * bug where valid CMASK and FMASK are required to be present to avoid
511 * a hardlock in certain operations but aren't actually used
512 * for anything useful. */
513 struct r600_resource *dummy_fmask;
514 struct r600_resource *dummy_cmask;
515
516 /* State binding slots are here. */
517 struct r600_atom *atoms[R600_NUM_ATOMS];
518 /* Dirty atom bitmask for fast tests */
519 uint64_t dirty_atoms;
520 /* States for CS initialization. */
521 struct r600_command_buffer start_cs_cmd; /* invariant state mostly */
522 /** Compute specific registers initializations. The start_cs_cmd atom
523 * must be emitted before start_compute_cs_cmd. */
524 struct r600_command_buffer start_compute_cs_cmd;
525 /* Register states. */
526 struct r600_alphatest_state alphatest_state;
527 struct r600_cso_state blend_state;
528 struct r600_blend_color blend_color;
529 struct r600_cb_misc_state cb_misc_state;
530 struct r600_clip_misc_state clip_misc_state;
531 struct r600_clip_state clip_state;
532 struct r600_db_misc_state db_misc_state;
533 struct r600_db_state db_state;
534 struct r600_cso_state dsa_state;
535 struct r600_framebuffer framebuffer;
536 struct r600_poly_offset_state poly_offset_state;
537 struct r600_cso_state rasterizer_state;
538 struct r600_sample_mask sample_mask;
539 struct r600_seamless_cube_map seamless_cube_map;
540 struct r600_config_state config_state;
541 struct r600_stencil_ref_state stencil_ref;
542 struct r600_vgt_state vgt_state;
543 struct r600_atomic_buffer_state atomic_buffer_state;
544 /* only have images on fragment shader */
545 struct r600_image_state fragment_images;
546 struct r600_image_state compute_images;
547 struct r600_image_state fragment_buffers;
548 struct r600_image_state compute_buffers;
549 /* Shaders and shader resources. */
550 struct r600_cso_state vertex_fetch_shader;
551 struct r600_shader_state hw_shader_stages[EG_NUM_HW_STAGES];
552 struct r600_cs_shader_state cs_shader_state;
553 struct r600_shader_stages_state shader_stages;
554 struct r600_gs_rings_state gs_rings;
555 struct r600_constbuf_state constbuf_state[PIPE_SHADER_TYPES];
556 struct r600_textures_info samplers[PIPE_SHADER_TYPES];
557
558 struct r600_shader_driver_constants_info driver_consts[PIPE_SHADER_TYPES];
559
560 /** Vertex buffers for fetch shaders */
561 struct r600_vertexbuf_state vertex_buffer_state;
562 /** Vertex buffers for compute shaders */
563 struct r600_vertexbuf_state cs_vertex_buffer_state;
564
565 /* Additional context states. */
566 unsigned compute_cb_target_mask;
567 struct r600_pipe_shader_selector *ps_shader;
568 struct r600_pipe_shader_selector *vs_shader;
569 struct r600_pipe_shader_selector *gs_shader;
570
571 struct r600_pipe_shader_selector *tcs_shader;
572 struct r600_pipe_shader_selector *tes_shader;
573
574 struct r600_pipe_shader_selector *fixed_func_tcs_shader;
575
576 struct r600_rasterizer_state *rasterizer;
577 bool alpha_to_one;
578 bool force_blend_disable;
579 bool gs_tri_strip_adj_fix;
580 boolean dual_src_blend;
581 unsigned zwritemask;
582 int ps_iter_samples;
583
584 /* The list of all texture buffer objects in this context.
585 * This list is walked when a buffer is invalidated/reallocated and
586 * the GPU addresses are updated. */
587 struct list_head texture_buffers;
588
589 /* Last draw state (-1 = unset). */
590 enum pipe_prim_type last_primitive_type; /* Last primitive type used in draw_vbo. */
591 enum pipe_prim_type current_rast_prim; /* primitive type after TES, GS */
592 enum pipe_prim_type last_rast_prim;
593 unsigned last_start_instance;
594
595 void *sb_context;
596 struct r600_isa *isa;
597 float sample_positions[4 * 16];
598 float tess_state[8];
599 uint32_t cs_block_grid_sizes[8]; /* 3 for grid + 1 pad, 3 for block + 1 pad*/
600 struct r600_pipe_shader_selector *last_ls;
601 struct r600_pipe_shader_selector *last_tcs;
602 unsigned last_num_tcs_input_cp;
603 unsigned lds_alloc;
604
605 struct r600_scratch_buffer scratch_buffers[MAX2(R600_NUM_HW_STAGES, EG_NUM_HW_STAGES)];
606
607 /* Debug state. */
608 bool is_debug;
609 struct radeon_saved_cs last_gfx;
610 struct r600_resource *last_trace_buf;
611 struct r600_resource *trace_buf;
612 unsigned trace_id;
613
614 bool cmd_buf_is_compute;
615 struct pipe_resource *append_fence;
616 uint32_t append_fence_id;
617 };
618
619 static inline void r600_emit_command_buffer(struct radeon_winsys_cs *cs,
620 struct r600_command_buffer *cb)
621 {
622 assert(cs->current.cdw + cb->num_dw <= cs->current.max_dw);
623 memcpy(cs->current.buf + cs->current.cdw, cb->buf, 4 * cb->num_dw);
624 cs->current.cdw += cb->num_dw;
625 }
626
627 static inline void r600_set_atom_dirty(struct r600_context *rctx,
628 struct r600_atom *atom,
629 bool dirty)
630 {
631 uint64_t mask;
632
633 assert(atom->id != 0);
634 assert(atom->id < sizeof(mask) * 8);
635 mask = 1ull << atom->id;
636 if (dirty)
637 rctx->dirty_atoms |= mask;
638 else
639 rctx->dirty_atoms &= ~mask;
640 }
641
642 static inline void r600_mark_atom_dirty(struct r600_context *rctx,
643 struct r600_atom *atom)
644 {
645 r600_set_atom_dirty(rctx, atom, true);
646 }
647
648 static inline void r600_emit_atom(struct r600_context *rctx, struct r600_atom *atom)
649 {
650 atom->emit(&rctx->b, atom);
651 r600_set_atom_dirty(rctx, atom, false);
652 }
653
654 static inline void r600_set_cso_state(struct r600_context *rctx,
655 struct r600_cso_state *state, void *cso)
656 {
657 state->cso = cso;
658 r600_set_atom_dirty(rctx, &state->atom, cso != NULL);
659 }
660
661 static inline void r600_set_cso_state_with_cb(struct r600_context *rctx,
662 struct r600_cso_state *state, void *cso,
663 struct r600_command_buffer *cb)
664 {
665 state->cb = cb;
666 state->atom.num_dw = cb ? cb->num_dw : 0;
667 r600_set_cso_state(rctx, state, cso);
668 }
669
670 /* compute_memory_pool.c */
671 struct compute_memory_pool;
672 void compute_memory_pool_delete(struct compute_memory_pool* pool);
673 struct compute_memory_pool* compute_memory_pool_new(
674 struct r600_screen *rscreen);
675
676 /* evergreen_state.c */
677 struct pipe_sampler_view *
678 evergreen_create_sampler_view_custom(struct pipe_context *ctx,
679 struct pipe_resource *texture,
680 const struct pipe_sampler_view *state,
681 unsigned width0, unsigned height0,
682 unsigned force_level);
683 void evergreen_init_common_regs(struct r600_context *ctx,
684 struct r600_command_buffer *cb,
685 enum chip_class ctx_chip_class,
686 enum radeon_family ctx_family,
687 int ctx_drm_minor);
688 void cayman_init_common_regs(struct r600_command_buffer *cb,
689 enum chip_class ctx_chip_class,
690 enum radeon_family ctx_family,
691 int ctx_drm_minor);
692
693 void evergreen_init_state_functions(struct r600_context *rctx);
694 void evergreen_init_atom_start_cs(struct r600_context *rctx);
695 void evergreen_update_ps_state(struct pipe_context *ctx, struct r600_pipe_shader *shader);
696 void evergreen_update_es_state(struct pipe_context *ctx, struct r600_pipe_shader *shader);
697 void evergreen_update_gs_state(struct pipe_context *ctx, struct r600_pipe_shader *shader);
698 void evergreen_update_vs_state(struct pipe_context *ctx, struct r600_pipe_shader *shader);
699 void evergreen_update_ls_state(struct pipe_context *ctx, struct r600_pipe_shader *shader);
700 void evergreen_update_hs_state(struct pipe_context *ctx, struct r600_pipe_shader *shader);
701 void *evergreen_create_db_flush_dsa(struct r600_context *rctx);
702 void *evergreen_create_resolve_blend(struct r600_context *rctx);
703 void *evergreen_create_decompress_blend(struct r600_context *rctx);
704 void *evergreen_create_fastclear_blend(struct r600_context *rctx);
705 boolean evergreen_is_format_supported(struct pipe_screen *screen,
706 enum pipe_format format,
707 enum pipe_texture_target target,
708 unsigned sample_count,
709 unsigned usage);
710 void evergreen_init_color_surface(struct r600_context *rctx,
711 struct r600_surface *surf);
712 void evergreen_init_color_surface_rat(struct r600_context *rctx,
713 struct r600_surface *surf);
714 void evergreen_update_db_shader_control(struct r600_context * rctx);
715 bool evergreen_adjust_gprs(struct r600_context *rctx);
716 void evergreen_setup_scratch_buffers(struct r600_context *rctx);
717 uint32_t evergreen_construct_rat_mask(struct r600_context *rctx, struct r600_cb_misc_state *a,
718 unsigned nr_cbufs);
719 /* r600_blit.c */
720 void r600_init_blit_functions(struct r600_context *rctx);
721 void r600_decompress_depth_textures(struct r600_context *rctx,
722 struct r600_samplerview_state *textures);
723 void r600_decompress_depth_images(struct r600_context *rctx,
724 struct r600_image_state *images);
725 void r600_decompress_color_textures(struct r600_context *rctx,
726 struct r600_samplerview_state *textures);
727 void r600_decompress_color_images(struct r600_context *rctx,
728 struct r600_image_state *images);
729 void r600_resource_copy_region(struct pipe_context *ctx,
730 struct pipe_resource *dst,
731 unsigned dst_level,
732 unsigned dstx, unsigned dsty, unsigned dstz,
733 struct pipe_resource *src,
734 unsigned src_level,
735 const struct pipe_box *src_box);
736
737 /* r600_shader.c */
738 int r600_pipe_shader_create(struct pipe_context *ctx,
739 struct r600_pipe_shader *shader,
740 union r600_shader_key key);
741
742 void r600_pipe_shader_destroy(struct pipe_context *ctx, struct r600_pipe_shader *shader);
743
744 /* r600_state.c */
745 struct pipe_sampler_view *
746 r600_create_sampler_view_custom(struct pipe_context *ctx,
747 struct pipe_resource *texture,
748 const struct pipe_sampler_view *state,
749 unsigned width_first_level, unsigned height_first_level);
750 void r600_init_state_functions(struct r600_context *rctx);
751 void r600_init_atom_start_cs(struct r600_context *rctx);
752 void r600_update_ps_state(struct pipe_context *ctx, struct r600_pipe_shader *shader);
753 void r600_update_es_state(struct pipe_context *ctx, struct r600_pipe_shader *shader);
754 void r600_update_gs_state(struct pipe_context *ctx, struct r600_pipe_shader *shader);
755 void r600_update_vs_state(struct pipe_context *ctx, struct r600_pipe_shader *shader);
756 void *r600_create_db_flush_dsa(struct r600_context *rctx);
757 void *r600_create_resolve_blend(struct r600_context *rctx);
758 void *r700_create_resolve_blend(struct r600_context *rctx);
759 void *r600_create_decompress_blend(struct r600_context *rctx);
760 bool r600_adjust_gprs(struct r600_context *rctx);
761 boolean r600_is_format_supported(struct pipe_screen *screen,
762 enum pipe_format format,
763 enum pipe_texture_target target,
764 unsigned sample_count,
765 unsigned usage);
766 void r600_update_db_shader_control(struct r600_context * rctx);
767 void r600_setup_scratch_buffers(struct r600_context *rctx);
768
769 /* r600_hw_context.c */
770 void r600_context_gfx_flush(void *context, unsigned flags,
771 struct pipe_fence_handle **fence);
772 void r600_begin_new_cs(struct r600_context *ctx);
773 void r600_flush_emit(struct r600_context *ctx);
774 void r600_need_cs_space(struct r600_context *ctx, unsigned num_dw, boolean count_draw_in);
775 void r600_emit_pfp_sync_me(struct r600_context *rctx);
776 void r600_cp_dma_copy_buffer(struct r600_context *rctx,
777 struct pipe_resource *dst, uint64_t dst_offset,
778 struct pipe_resource *src, uint64_t src_offset,
779 unsigned size);
780 void evergreen_cp_dma_clear_buffer(struct r600_context *rctx,
781 struct pipe_resource *dst, uint64_t offset,
782 unsigned size, uint32_t clear_value,
783 enum r600_coherency coher);
784 void r600_dma_copy_buffer(struct r600_context *rctx,
785 struct pipe_resource *dst,
786 struct pipe_resource *src,
787 uint64_t dst_offset,
788 uint64_t src_offset,
789 uint64_t size);
790
791 /*
792 * evergreen_hw_context.c
793 */
794 void evergreen_dma_copy_buffer(struct r600_context *rctx,
795 struct pipe_resource *dst,
796 struct pipe_resource *src,
797 uint64_t dst_offset,
798 uint64_t src_offset,
799 uint64_t size);
800 void evergreen_setup_tess_constants(struct r600_context *rctx,
801 const struct pipe_draw_info *info,
802 unsigned *num_patches);
803 uint32_t evergreen_get_ls_hs_config(struct r600_context *rctx,
804 const struct pipe_draw_info *info,
805 unsigned num_patches);
806 void evergreen_set_ls_hs_config(struct r600_context *rctx,
807 struct radeon_winsys_cs *cs,
808 uint32_t ls_hs_config);
809 void evergreen_set_lds_alloc(struct r600_context *rctx,
810 struct radeon_winsys_cs *cs,
811 uint32_t lds_alloc);
812
813 /* r600_state_common.c */
814 void r600_init_common_state_functions(struct r600_context *rctx);
815 void r600_emit_cso_state(struct r600_context *rctx, struct r600_atom *atom);
816 void r600_emit_alphatest_state(struct r600_context *rctx, struct r600_atom *atom);
817 void r600_emit_blend_color(struct r600_context *rctx, struct r600_atom *atom);
818 void r600_emit_vgt_state(struct r600_context *rctx, struct r600_atom *atom);
819 void r600_emit_clip_misc_state(struct r600_context *rctx, struct r600_atom *atom);
820 void r600_emit_stencil_ref(struct r600_context *rctx, struct r600_atom *atom);
821 void r600_emit_shader(struct r600_context *rctx, struct r600_atom *a);
822 void r600_add_atom(struct r600_context *rctx, struct r600_atom *atom, unsigned id);
823 void r600_init_atom(struct r600_context *rctx, struct r600_atom *atom, unsigned id,
824 void (*emit)(struct r600_context *ctx, struct r600_atom *state),
825 unsigned num_dw);
826 void r600_vertex_buffers_dirty(struct r600_context *rctx);
827 void r600_sampler_views_dirty(struct r600_context *rctx,
828 struct r600_samplerview_state *state);
829 void r600_sampler_states_dirty(struct r600_context *rctx,
830 struct r600_sampler_states *state);
831 void r600_constant_buffers_dirty(struct r600_context *rctx, struct r600_constbuf_state *state);
832 void r600_set_sample_locations_constant_buffer(struct r600_context *rctx);
833 void r600_setup_scratch_area_for_shader(struct r600_context *rctx,
834 struct r600_pipe_shader *shader, struct r600_scratch_buffer *scratch,
835 unsigned ring_base_reg, unsigned item_size_reg, unsigned ring_size_reg);
836 uint32_t r600_translate_stencil_op(int s_op);
837 uint32_t r600_translate_fill(uint32_t func);
838 unsigned r600_tex_wrap(unsigned wrap);
839 unsigned r600_tex_mipfilter(unsigned filter);
840 unsigned r600_tex_compare(unsigned compare);
841 bool sampler_state_needs_border_color(const struct pipe_sampler_state *state);
842 unsigned r600_get_swizzle_combined(const unsigned char *swizzle_format,
843 const unsigned char *swizzle_view,
844 boolean vtx);
845 uint32_t r600_translate_texformat(struct pipe_screen *screen, enum pipe_format format,
846 const unsigned char *swizzle_view,
847 uint32_t *word4_p, uint32_t *yuv_format_p,
848 bool do_endian_swap);
849 uint32_t r600_translate_colorformat(enum chip_class chip, enum pipe_format format,
850 bool do_endian_swap);
851 uint32_t r600_colorformat_endian_swap(uint32_t colorformat, bool do_endian_swap);
852
853 /* r600_uvd.c */
854 struct pipe_video_codec *r600_uvd_create_decoder(struct pipe_context *context,
855 const struct pipe_video_codec *decoder);
856
857 struct pipe_video_buffer *r600_video_buffer_create(struct pipe_context *pipe,
858 const struct pipe_video_buffer *tmpl);
859
860 /*
861 * Helpers for building command buffers
862 */
863
864 #define PKT3_SET_CONFIG_REG 0x68
865 #define PKT3_SET_CONTEXT_REG 0x69
866 #define PKT3_SET_CTL_CONST 0x6F
867 #define PKT3_SET_LOOP_CONST 0x6C
868
869 #define R600_CONFIG_REG_OFFSET 0x08000
870 #define R600_CONTEXT_REG_OFFSET 0x28000
871 #define R600_CTL_CONST_OFFSET 0x3CFF0
872 #define R600_LOOP_CONST_OFFSET 0X0003E200
873 #define EG_LOOP_CONST_OFFSET 0x0003A200
874
875 #define PKT_TYPE_S(x) (((unsigned)(x) & 0x3) << 30)
876 #define PKT_COUNT_S(x) (((unsigned)(x) & 0x3FFF) << 16)
877 #define PKT3_IT_OPCODE_S(x) (((unsigned)(x) & 0xFF) << 8)
878 #define PKT3_PREDICATE(x) (((x) >> 0) & 0x1)
879 #define PKT3(op, count, predicate) (PKT_TYPE_S(3) | PKT_COUNT_S(count) | PKT3_IT_OPCODE_S(op) | PKT3_PREDICATE(predicate))
880
881 #define RADEON_CP_PACKET3_COMPUTE_MODE 0x00000002
882
883 /*Evergreen Compute packet3*/
884 #define PKT3C(op, count, predicate) (PKT_TYPE_S(3) | PKT3_IT_OPCODE_S(op) | PKT_COUNT_S(count) | PKT3_PREDICATE(predicate) | RADEON_CP_PACKET3_COMPUTE_MODE)
885
886 static inline void r600_store_value(struct r600_command_buffer *cb, unsigned value)
887 {
888 cb->buf[cb->num_dw++] = value;
889 }
890
891 static inline void r600_store_array(struct r600_command_buffer *cb, unsigned num, unsigned *ptr)
892 {
893 assert(cb->num_dw+num <= cb->max_num_dw);
894 memcpy(&cb->buf[cb->num_dw], ptr, num * sizeof(ptr[0]));
895 cb->num_dw += num;
896 }
897
898 static inline void r600_store_config_reg_seq(struct r600_command_buffer *cb, unsigned reg, unsigned num)
899 {
900 assert(reg < R600_CONTEXT_REG_OFFSET);
901 assert(cb->num_dw+2+num <= cb->max_num_dw);
902 cb->buf[cb->num_dw++] = PKT3(PKT3_SET_CONFIG_REG, num, 0);
903 cb->buf[cb->num_dw++] = (reg - R600_CONFIG_REG_OFFSET) >> 2;
904 }
905
906 /**
907 * Needs cb->pkt_flags set to RADEON_CP_PACKET3_COMPUTE_MODE for compute
908 * shaders.
909 */
910 static inline void r600_store_context_reg_seq(struct r600_command_buffer *cb, unsigned reg, unsigned num)
911 {
912 assert(reg >= R600_CONTEXT_REG_OFFSET && reg < R600_CTL_CONST_OFFSET);
913 assert(cb->num_dw+2+num <= cb->max_num_dw);
914 cb->buf[cb->num_dw++] = PKT3(PKT3_SET_CONTEXT_REG, num, 0) | cb->pkt_flags;
915 cb->buf[cb->num_dw++] = (reg - R600_CONTEXT_REG_OFFSET) >> 2;
916 }
917
918 /**
919 * Needs cb->pkt_flags set to RADEON_CP_PACKET3_COMPUTE_MODE for compute
920 * shaders.
921 */
922 static inline void r600_store_ctl_const_seq(struct r600_command_buffer *cb, unsigned reg, unsigned num)
923 {
924 assert(reg >= R600_CTL_CONST_OFFSET);
925 assert(cb->num_dw+2+num <= cb->max_num_dw);
926 cb->buf[cb->num_dw++] = PKT3(PKT3_SET_CTL_CONST, num, 0) | cb->pkt_flags;
927 cb->buf[cb->num_dw++] = (reg - R600_CTL_CONST_OFFSET) >> 2;
928 }
929
930 static inline void r600_store_loop_const_seq(struct r600_command_buffer *cb, unsigned reg, unsigned num)
931 {
932 assert(reg >= R600_LOOP_CONST_OFFSET);
933 assert(cb->num_dw+2+num <= cb->max_num_dw);
934 cb->buf[cb->num_dw++] = PKT3(PKT3_SET_LOOP_CONST, num, 0);
935 cb->buf[cb->num_dw++] = (reg - R600_LOOP_CONST_OFFSET) >> 2;
936 }
937
938 /**
939 * Needs cb->pkt_flags set to RADEON_CP_PACKET3_COMPUTE_MODE for compute
940 * shaders.
941 */
942 static inline void eg_store_loop_const_seq(struct r600_command_buffer *cb, unsigned reg, unsigned num)
943 {
944 assert(reg >= EG_LOOP_CONST_OFFSET);
945 assert(cb->num_dw+2+num <= cb->max_num_dw);
946 cb->buf[cb->num_dw++] = PKT3(PKT3_SET_LOOP_CONST, num, 0) | cb->pkt_flags;
947 cb->buf[cb->num_dw++] = (reg - EG_LOOP_CONST_OFFSET) >> 2;
948 }
949
950 static inline void r600_store_config_reg(struct r600_command_buffer *cb, unsigned reg, unsigned value)
951 {
952 r600_store_config_reg_seq(cb, reg, 1);
953 r600_store_value(cb, value);
954 }
955
956 static inline void r600_store_context_reg(struct r600_command_buffer *cb, unsigned reg, unsigned value)
957 {
958 r600_store_context_reg_seq(cb, reg, 1);
959 r600_store_value(cb, value);
960 }
961
962 static inline void r600_store_ctl_const(struct r600_command_buffer *cb, unsigned reg, unsigned value)
963 {
964 r600_store_ctl_const_seq(cb, reg, 1);
965 r600_store_value(cb, value);
966 }
967
968 static inline void r600_store_loop_const(struct r600_command_buffer *cb, unsigned reg, unsigned value)
969 {
970 r600_store_loop_const_seq(cb, reg, 1);
971 r600_store_value(cb, value);
972 }
973
974 static inline void eg_store_loop_const(struct r600_command_buffer *cb, unsigned reg, unsigned value)
975 {
976 eg_store_loop_const_seq(cb, reg, 1);
977 r600_store_value(cb, value);
978 }
979
980 void r600_init_command_buffer(struct r600_command_buffer *cb, unsigned num_dw);
981 void r600_release_command_buffer(struct r600_command_buffer *cb);
982
983 static inline void radeon_compute_set_context_reg_seq(struct radeon_winsys_cs *cs, unsigned reg, unsigned num)
984 {
985 radeon_set_context_reg_seq(cs, reg, num);
986 /* Set the compute bit on the packet header */
987 cs->current.buf[cs->current.cdw - 2] |= RADEON_CP_PACKET3_COMPUTE_MODE;
988 }
989
990 static inline void radeon_set_ctl_const_seq(struct radeon_winsys_cs *cs, unsigned reg, unsigned num)
991 {
992 assert(reg >= R600_CTL_CONST_OFFSET);
993 assert(cs->current.cdw + 2 + num <= cs->current.max_dw);
994 radeon_emit(cs, PKT3(PKT3_SET_CTL_CONST, num, 0));
995 radeon_emit(cs, (reg - R600_CTL_CONST_OFFSET) >> 2);
996 }
997
998 static inline void radeon_compute_set_context_reg(struct radeon_winsys_cs *cs, unsigned reg, unsigned value)
999 {
1000 radeon_compute_set_context_reg_seq(cs, reg, 1);
1001 radeon_emit(cs, value);
1002 }
1003
1004 static inline void radeon_set_context_reg_flag(struct radeon_winsys_cs *cs, unsigned reg, unsigned value, unsigned flag)
1005 {
1006 if (flag & RADEON_CP_PACKET3_COMPUTE_MODE) {
1007 radeon_compute_set_context_reg(cs, reg, value);
1008 } else {
1009 radeon_set_context_reg(cs, reg, value);
1010 }
1011 }
1012
1013 static inline void radeon_set_ctl_const(struct radeon_winsys_cs *cs, unsigned reg, unsigned value)
1014 {
1015 radeon_set_ctl_const_seq(cs, reg, 1);
1016 radeon_emit(cs, value);
1017 }
1018
1019 /*
1020 * common helpers
1021 */
1022
1023 /* 12.4 fixed-point */
1024 static inline unsigned r600_pack_float_12p4(float x)
1025 {
1026 return x <= 0 ? 0 :
1027 x >= 4096 ? 0xffff : x * 16;
1028 }
1029
1030 static inline unsigned r600_get_flush_flags(enum r600_coherency coher)
1031 {
1032 switch (coher) {
1033 default:
1034 case R600_COHERENCY_NONE:
1035 return 0;
1036 case R600_COHERENCY_SHADER:
1037 return R600_CONTEXT_INV_CONST_CACHE |
1038 R600_CONTEXT_INV_VERTEX_CACHE |
1039 R600_CONTEXT_INV_TEX_CACHE |
1040 R600_CONTEXT_STREAMOUT_FLUSH;
1041 case R600_COHERENCY_CB_META:
1042 return R600_CONTEXT_FLUSH_AND_INV_CB |
1043 R600_CONTEXT_FLUSH_AND_INV_CB_META;
1044 }
1045 }
1046
1047 #define V_028A6C_OUTPRIM_TYPE_POINTLIST 0
1048 #define V_028A6C_OUTPRIM_TYPE_LINESTRIP 1
1049 #define V_028A6C_OUTPRIM_TYPE_TRISTRIP 2
1050
1051 unsigned r600_conv_prim_to_gs_out(unsigned mode);
1052
1053 void eg_trace_emit(struct r600_context *rctx);
1054 void eg_dump_debug_state(struct pipe_context *ctx, FILE *f,
1055 unsigned flags);
1056
1057 struct r600_pipe_shader_selector *r600_create_shader_state_tokens(struct pipe_context *ctx,
1058 const struct tgsi_token *tokens,
1059 unsigned pipe_shader_type);
1060 int r600_shader_select(struct pipe_context *ctx,
1061 struct r600_pipe_shader_selector* sel,
1062 bool *dirty);
1063
1064 void r600_delete_shader_selector(struct pipe_context *ctx,
1065 struct r600_pipe_shader_selector *sel);
1066
1067 struct r600_shader_atomic;
1068 bool evergreen_emit_atomic_buffer_setup(struct r600_context *rctx,
1069 struct r600_pipe_shader *cs_shader,
1070 struct r600_shader_atomic *combined_atomics,
1071 uint8_t *atomic_used_mask_p);
1072 void evergreen_emit_atomic_buffer_save(struct r600_context *rctx,
1073 bool is_compute,
1074 struct r600_shader_atomic *combined_atomics,
1075 uint8_t *atomic_used_mask_p);
1076 void r600_update_compressed_resource_state(struct r600_context *rctx, bool compute_only);
1077
1078 void eg_setup_buffer_constants(struct r600_context *rctx, int shader_type);
1079 void r600_update_driver_const_buffers(struct r600_context *rctx, bool compute_only);
1080 #endif