projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radeonsi: determine the rasterization primitive type accurately (v2)
[mesa.git] / src / gallium / drivers / radeonsi / si_pipe.h
1 /*
2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3 * Copyright 2018 Advanced Micro Devices, Inc.
4 * All Rights Reserved.
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
10 * license, and/or sell copies of the Software, and to permit persons to whom
11 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
21 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
22 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
23 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 */
25 #ifndef SI_PIPE_H
26 #define SI_PIPE_H
27
28 #include "si_shader.h"
29 #include "si_state.h"
30
31 #include "util/u_dynarray.h"
32 #include "util/u_idalloc.h"
33 #include "util/u_threaded_context.h"
34
35 #ifdef PIPE_ARCH_BIG_ENDIAN
36 #define SI_BIG_ENDIAN 1
37 #else
38 #define SI_BIG_ENDIAN 0
39 #endif
40
41 #define ATI_VENDOR_ID 0x1002
42 #define SI_PRIM_DISCARD_DEBUG 0
43 #define SI_NOT_QUERY 0xffffffff
44
45 /* The base vertex and primitive restart can be any number, but we must pick
46 * one which will mean "unknown" for the purpose of state tracking and
47 * the number shouldn't be a commonly-used one. */
48 #define SI_BASE_VERTEX_UNKNOWN INT_MIN
49 #define SI_RESTART_INDEX_UNKNOWN INT_MIN
50 #define SI_INSTANCE_COUNT_UNKNOWN INT_MIN
51 #define SI_NUM_SMOOTH_AA_SAMPLES 8
52 #define SI_MAX_POINT_SIZE 2048
53 #define SI_GS_PER_ES 128
54 /* Alignment for optimal CP DMA performance. */
55 #define SI_CPDMA_ALIGNMENT 32
56
57 /* Tunables for compute-based clear_buffer and copy_buffer: */
58 #define SI_COMPUTE_CLEAR_DW_PER_THREAD 4
59 #define SI_COMPUTE_COPY_DW_PER_THREAD 4
60 #define SI_COMPUTE_DST_CACHE_POLICY L2_STREAM
61
62 /* Pipeline & streamout query controls. */
63 #define SI_CONTEXT_START_PIPELINE_STATS (1 << 0)
64 #define SI_CONTEXT_STOP_PIPELINE_STATS (1 << 1)
65 #define SI_CONTEXT_FLUSH_FOR_RENDER_COND (1 << 2)
66 /* Instruction cache. */
67 #define SI_CONTEXT_INV_ICACHE (1 << 3)
68 /* Scalar cache. (GFX6-9: scalar L1; GFX10: scalar L0)
69 * GFX10: This also invalidates the L1 shader array cache. */
70 #define SI_CONTEXT_INV_SCACHE (1 << 4)
71 /* Vector cache. (GFX6-9: vector L1; GFX10: vector L0)
72 * GFX10: This also invalidates the L1 shader array cache. */
73 #define SI_CONTEXT_INV_VCACHE (1 << 5)
74 /* L2 cache + L2 metadata cache writeback & invalidate.
75 * GFX6-8: Used by shaders only. GFX9-10: Used by everything. */
76 #define SI_CONTEXT_INV_L2 (1 << 6)
77 /* L2 writeback (write dirty L2 lines to memory for non-L2 clients).
78 * Only used for coherency with non-L2 clients like CB, DB, CP on GFX6-8.
79 * GFX6-7 will do complete invalidation, because the writeback is unsupported. */
80 #define SI_CONTEXT_WB_L2 (1 << 7)
81 /* Writeback & invalidate the L2 metadata cache only. It can only be coupled with
82 * a CB or DB flush. */
83 #define SI_CONTEXT_INV_L2_METADATA (1 << 8)
84 /* Framebuffer caches. */
85 #define SI_CONTEXT_FLUSH_AND_INV_DB (1 << 9)
86 #define SI_CONTEXT_FLUSH_AND_INV_DB_META (1 << 10)
87 #define SI_CONTEXT_FLUSH_AND_INV_CB (1 << 11)
88 /* Engine synchronization. */
89 #define SI_CONTEXT_VS_PARTIAL_FLUSH (1 << 12)
90 #define SI_CONTEXT_PS_PARTIAL_FLUSH (1 << 13)
91 #define SI_CONTEXT_CS_PARTIAL_FLUSH (1 << 14)
92 #define SI_CONTEXT_VGT_FLUSH (1 << 15)
93 #define SI_CONTEXT_VGT_STREAMOUT_SYNC (1 << 16)
94
95 #define SI_PREFETCH_VBO_DESCRIPTORS (1 << 0)
96 #define SI_PREFETCH_LS (1 << 1)
97 #define SI_PREFETCH_HS (1 << 2)
98 #define SI_PREFETCH_ES (1 << 3)
99 #define SI_PREFETCH_GS (1 << 4)
100 #define SI_PREFETCH_VS (1 << 5)
101 #define SI_PREFETCH_PS (1 << 6)
102
103 #define SI_MAX_BORDER_COLORS 4096
104 #define SI_MAX_VIEWPORTS 16
105 #define SIX_BITS 0x3F
106 #define SI_MAP_BUFFER_ALIGNMENT 64
107 #define SI_MAX_VARIABLE_THREADS_PER_BLOCK 1024
108
109 #define SI_RESOURCE_FLAG_TRANSFER (PIPE_RESOURCE_FLAG_DRV_PRIV << 0)
110 #define SI_RESOURCE_FLAG_FLUSHED_DEPTH (PIPE_RESOURCE_FLAG_DRV_PRIV << 1)
111 #define SI_RESOURCE_FLAG_FORCE_MSAA_TILING (PIPE_RESOURCE_FLAG_DRV_PRIV << 2)
112 #define SI_RESOURCE_FLAG_DISABLE_DCC (PIPE_RESOURCE_FLAG_DRV_PRIV << 3)
113 #define SI_RESOURCE_FLAG_UNMAPPABLE (PIPE_RESOURCE_FLAG_DRV_PRIV << 4)
114 #define SI_RESOURCE_FLAG_READ_ONLY (PIPE_RESOURCE_FLAG_DRV_PRIV << 5)
115 #define SI_RESOURCE_FLAG_32BIT (PIPE_RESOURCE_FLAG_DRV_PRIV << 6)
116 #define SI_RESOURCE_FLAG_CLEAR (PIPE_RESOURCE_FLAG_DRV_PRIV << 7)
117 /* For const_uploader, upload data via GTT and copy to VRAM on context flush via SDMA. */
118 #define SI_RESOURCE_FLAG_UPLOAD_FLUSH_EXPLICIT_VIA_SDMA (PIPE_RESOURCE_FLAG_DRV_PRIV << 8)
119
120 enum si_clear_code
121 {
122 DCC_CLEAR_COLOR_0000 = 0x00000000,
123 DCC_CLEAR_COLOR_0001 = 0x40404040,
124 DCC_CLEAR_COLOR_1110 = 0x80808080,
125 DCC_CLEAR_COLOR_1111 = 0xC0C0C0C0,
126 DCC_CLEAR_COLOR_REG = 0x20202020,
127 DCC_UNCOMPRESSED = 0xFFFFFFFF,
128 };
129
130 #define SI_IMAGE_ACCESS_AS_BUFFER (1 << 7)
131
132 /* Debug flags. */
133 enum {
134 /* Shader logging options: */
135 DBG_VS = PIPE_SHADER_VERTEX,
136 DBG_PS = PIPE_SHADER_FRAGMENT,
137 DBG_GS = PIPE_SHADER_GEOMETRY,
138 DBG_TCS = PIPE_SHADER_TESS_CTRL,
139 DBG_TES = PIPE_SHADER_TESS_EVAL,
140 DBG_CS = PIPE_SHADER_COMPUTE,
141 DBG_NO_IR,
142 DBG_NO_TGSI,
143 DBG_NO_ASM,
144 DBG_PREOPT_IR,
145
146 /* Shader compiler options the shader cache should be aware of: */
147 DBG_FS_CORRECT_DERIVS_AFTER_KILL,
148 DBG_UNSAFE_MATH,
149 DBG_SI_SCHED,
150 DBG_GISEL,
151
152 /* Shader compiler options (with no effect on the shader cache): */
153 DBG_CHECK_IR,
154 DBG_MONOLITHIC_SHADERS,
155 DBG_NO_OPT_VARIANT,
156
157 /* Information logging options: */
158 DBG_INFO,
159 DBG_TEX,
160 DBG_COMPUTE,
161 DBG_VM,
162
163 /* Driver options: */
164 DBG_FORCE_DMA,
165 DBG_NO_ASYNC_DMA,
166 DBG_NO_WC,
167 DBG_CHECK_VM,
168 DBG_RESERVE_VMID,
169 DBG_ZERO_VRAM,
170
171 /* 3D engine options: */
172 DBG_ALWAYS_PD,
173 DBG_PD,
174 DBG_NO_PD,
175 DBG_SWITCH_ON_EOP,
176 DBG_NO_OUT_OF_ORDER,
177 DBG_NO_DPBB,
178 DBG_NO_DFSM,
179 DBG_DPBB,
180 DBG_DFSM,
181 DBG_NO_HYPERZ,
182 DBG_NO_RB_PLUS,
183 DBG_NO_2D_TILING,
184 DBG_NO_TILING,
185 DBG_NO_DCC,
186 DBG_NO_DCC_CLEAR,
187 DBG_NO_DCC_FB,
188 DBG_NO_DCC_MSAA,
189 DBG_NO_FMASK,
190
191 /* Tests: */
192 DBG_TEST_DMA,
193 DBG_TEST_VMFAULT_CP,
194 DBG_TEST_VMFAULT_SDMA,
195 DBG_TEST_VMFAULT_SHADER,
196 DBG_TEST_DMA_PERF,
197 DBG_TEST_GDS,
198 DBG_TEST_GDS_MM,
199 DBG_TEST_GDS_OA_MM,
200 };
201
202 #define DBG_ALL_SHADERS (((1 << (DBG_CS + 1)) - 1))
203 #define DBG(name) (1ull << DBG_##name)
204
205 enum si_cache_policy {
206 L2_BYPASS,
207 L2_STREAM, /* same as SLC=1 */
208 L2_LRU, /* same as SLC=0 */
209 };
210
211 enum si_coherency {
212 SI_COHERENCY_NONE, /* no cache flushes needed */
213 SI_COHERENCY_SHADER,
214 SI_COHERENCY_CB_META,
215 SI_COHERENCY_CP,
216 };
217
218 struct si_compute;
219 struct si_shader_context;
220 struct hash_table;
221 struct u_suballocator;
222
223 /* Only 32-bit buffer allocations are supported, gallium doesn't support more
224 * at the moment.
225 */
226 struct si_resource {
227 struct threaded_resource b;
228
229 /* Winsys objects. */
230 struct pb_buffer *buf;
231 uint64_t gpu_address;
232 /* Memory usage if the buffer placement is optimal. */
233 uint64_t vram_usage;
234 uint64_t gart_usage;
235
236 /* Resource properties. */
237 uint64_t bo_size;
238 unsigned bo_alignment;
239 enum radeon_bo_domain domains;
240 enum radeon_bo_flag flags;
241 unsigned bind_history;
242 int max_forced_staging_uploads;
243
244 /* The buffer range which is initialized (with a write transfer,
245 * streamout, DMA, or as a random access target). The rest of
246 * the buffer is considered invalid and can be mapped unsynchronized.
247 *
248 * This allows unsychronized mapping of a buffer range which hasn't
249 * been used yet. It's for applications which forget to use
250 * the unsynchronized map flag and expect the driver to figure it out.
251 */
252 struct util_range valid_buffer_range;
253
254 /* For buffers only. This indicates that a write operation has been
255 * performed by TC L2, but the cache hasn't been flushed.
256 * Any hw block which doesn't use or bypasses TC L2 should check this
257 * flag and flush the cache before using the buffer.
258 *
259 * For example, TC L2 must be flushed if a buffer which has been
260 * modified by a shader store instruction is about to be used as
261 * an index buffer. The reason is that VGT DMA index fetching doesn't
262 * use TC L2.
263 */
264 bool TC_L2_dirty;
265
266 /* Whether this resource is referenced by bindless handles. */
267 bool texture_handle_allocated;
268 bool image_handle_allocated;
269
270 /* Whether the resource has been exported via resource_get_handle. */
271 unsigned external_usage; /* PIPE_HANDLE_USAGE_* */
272 };
273
274 struct si_transfer {
275 struct threaded_transfer b;
276 struct si_resource *staging;
277 unsigned offset;
278 };
279
280 struct si_texture {
281 struct si_resource buffer;
282
283 struct radeon_surf surface;
284 uint64_t size;
285 struct si_texture *flushed_depth_texture;
286
287 /* One texture allocation can contain these buffers:
288 * - image (pixel data)
289 * - FMASK buffer (MSAA compression)
290 * - CMASK buffer (MSAA compression and/or legacy fast color clear)
291 * - HTILE buffer (Z/S compression and fast Z/S clear)
292 * - DCC buffer (color compression and new fast color clear)
293 * - displayable DCC buffer (if the DCC buffer is not displayable)
294 * - DCC retile mapping buffer (if the DCC buffer is not displayable)
295 */
296 uint64_t fmask_offset;
297 uint64_t cmask_offset;
298 uint64_t cmask_base_address_reg;
299 struct si_resource *cmask_buffer;
300 uint64_t dcc_offset; /* 0 = disabled */
301 uint64_t display_dcc_offset;
302 uint64_t dcc_retile_map_offset;
303 unsigned cb_color_info; /* fast clear enable bit */
304 unsigned color_clear_value[2];
305 unsigned last_msaa_resolve_target_micro_mode;
306 unsigned num_level0_transfers;
307
308 /* Depth buffer compression and fast clear. */
309 uint64_t htile_offset;
310 float depth_clear_value;
311 uint16_t dirty_level_mask; /* each bit says if that mipmap is compressed */
312 uint16_t stencil_dirty_level_mask; /* each bit says if that mipmap is compressed */
313 enum pipe_format db_render_format:16;
314 uint8_t stencil_clear_value;
315 bool tc_compatible_htile:1;
316 bool htile_stencil_disabled:1;
317 bool depth_cleared:1; /* if it was cleared at least once */
318 bool stencil_cleared:1; /* if it was cleared at least once */
319 bool upgraded_depth:1; /* upgraded from unorm to Z32_FLOAT */
320 bool is_depth:1;
321 bool db_compatible:1;
322 bool can_sample_z:1;
323 bool can_sample_s:1;
324
325 /* We need to track DCC dirtiness, because st/dri usually calls
326 * flush_resource twice per frame (not a bug) and we don't wanna
327 * decompress DCC twice. Also, the dirty tracking must be done even
328 * if DCC isn't used, because it's required by the DCC usage analysis
329 * for a possible future enablement.
330 */
331 bool separate_dcc_dirty:1;
332 /* Statistics gathering for the DCC enablement heuristic. */
333 bool dcc_gather_statistics:1;
334 /* Counter that should be non-zero if the texture is bound to a
335 * framebuffer.
336 */
337 unsigned framebuffers_bound;
338 /* Whether the texture is a displayable back buffer and needs DCC
339 * decompression, which is expensive. Therefore, it's enabled only
340 * if statistics suggest that it will pay off and it's allocated
341 * separately. It can't be bound as a sampler by apps. Limited to
342 * target == 2D and last_level == 0. If enabled, dcc_offset contains
343 * the absolute GPUVM address, not the relative one.
344 */
345 struct si_resource *dcc_separate_buffer;
346 /* When DCC is temporarily disabled, the separate buffer is here. */
347 struct si_resource *last_dcc_separate_buffer;
348 /* Estimate of how much this color buffer is written to in units of
349 * full-screen draws: ps_invocations / (width * height)
350 * Shader kills, late Z, and blending with trivial discards make it
351 * inaccurate (we need to count CB updates, not PS invocations).
352 */
353 unsigned ps_draw_ratio;
354 /* The number of clears since the last DCC usage analysis. */
355 unsigned num_slow_clears;
356 };
357
358 struct si_surface {
359 struct pipe_surface base;
360
361 /* These can vary with block-compressed textures. */
362 uint16_t width0;
363 uint16_t height0;
364
365 bool color_initialized:1;
366 bool depth_initialized:1;
367
368 /* Misc. color flags. */
369 bool color_is_int8:1;
370 bool color_is_int10:1;
371 bool dcc_incompatible:1;
372
373 /* Color registers. */
374 unsigned cb_color_info;
375 unsigned cb_color_view;
376 unsigned cb_color_attrib;
377 unsigned cb_color_attrib2; /* GFX9 and later */
378 unsigned cb_color_attrib3; /* GFX10 and later */
379 unsigned cb_dcc_control; /* GFX8 and later */
380 unsigned spi_shader_col_format:8; /* no blending, no alpha-to-coverage. */
381 unsigned spi_shader_col_format_alpha:8; /* alpha-to-coverage */
382 unsigned spi_shader_col_format_blend:8; /* blending without alpha. */
383 unsigned spi_shader_col_format_blend_alpha:8; /* blending with alpha. */
384
385 /* DB registers. */
386 uint64_t db_depth_base; /* DB_Z_READ/WRITE_BASE */
387 uint64_t db_stencil_base;
388 uint64_t db_htile_data_base;
389 unsigned db_depth_info;
390 unsigned db_z_info;
391 unsigned db_z_info2; /* GFX9 only */
392 unsigned db_depth_view;
393 unsigned db_depth_size;
394 unsigned db_depth_slice;
395 unsigned db_stencil_info;
396 unsigned db_stencil_info2; /* GFX9 only */
397 unsigned db_htile_surface;
398 };
399
400 struct si_mmio_counter {
401 unsigned busy;
402 unsigned idle;
403 };
404
405 union si_mmio_counters {
406 struct {
407 /* For global GPU load including SDMA. */
408 struct si_mmio_counter gpu;
409
410 /* GRBM_STATUS */
411 struct si_mmio_counter spi;
412 struct si_mmio_counter gui;
413 struct si_mmio_counter ta;
414 struct si_mmio_counter gds;
415 struct si_mmio_counter vgt;
416 struct si_mmio_counter ia;
417 struct si_mmio_counter sx;
418 struct si_mmio_counter wd;
419 struct si_mmio_counter bci;
420 struct si_mmio_counter sc;
421 struct si_mmio_counter pa;
422 struct si_mmio_counter db;
423 struct si_mmio_counter cp;
424 struct si_mmio_counter cb;
425
426 /* SRBM_STATUS2 */
427 struct si_mmio_counter sdma;
428
429 /* CP_STAT */
430 struct si_mmio_counter pfp;
431 struct si_mmio_counter meq;
432 struct si_mmio_counter me;
433 struct si_mmio_counter surf_sync;
434 struct si_mmio_counter cp_dma;
435 struct si_mmio_counter scratch_ram;
436 } named;
437 unsigned array[0];
438 };
439
440 struct si_memory_object {
441 struct pipe_memory_object b;
442 struct pb_buffer *buf;
443 uint32_t stride;
444 };
445
446 /* Saved CS data for debugging features. */
447 struct radeon_saved_cs {
448 uint32_t *ib;
449 unsigned num_dw;
450
451 struct radeon_bo_list_item *bo_list;
452 unsigned bo_count;
453 };
454
455 struct si_screen {
456 struct pipe_screen b;
457 struct radeon_winsys *ws;
458 struct disk_cache *disk_shader_cache;
459
460 struct radeon_info info;
461 uint64_t debug_flags;
462 char renderer_string[183];
463
464 void (*make_texture_descriptor)(
465 struct si_screen *screen,
466 struct si_texture *tex,
467 bool sampler,
468 enum pipe_texture_target target,
469 enum pipe_format pipe_format,
470 const unsigned char state_swizzle[4],
471 unsigned first_level, unsigned last_level,
472 unsigned first_layer, unsigned last_layer,
473 unsigned width, unsigned height, unsigned depth,
474 uint32_t *state,
475 uint32_t *fmask_state);
476
477 unsigned pa_sc_raster_config;
478 unsigned pa_sc_raster_config_1;
479 unsigned se_tile_repeat;
480 unsigned gs_table_depth;
481 unsigned tess_offchip_block_dw_size;
482 unsigned tess_offchip_ring_size;
483 unsigned tess_factor_ring_size;
484 unsigned vgt_hs_offchip_param;
485 unsigned eqaa_force_coverage_samples;
486 unsigned eqaa_force_z_samples;
487 unsigned eqaa_force_color_samples;
488 bool has_clear_state;
489 bool has_distributed_tess;
490 bool has_draw_indirect_multi;
491 bool has_out_of_order_rast;
492 bool assume_no_z_fights;
493 bool commutative_blend_add;
494 bool has_gfx9_scissor_bug;
495 bool has_msaa_sample_loc_bug;
496 bool has_ls_vgpr_init_bug;
497 bool has_dcc_constant_encode;
498 bool dpbb_allowed;
499 bool dfsm_allowed;
500 bool llvm_has_working_vgpr_indexing;
501
502 struct {
503 #define OPT_BOOL(name, dflt, description) bool name:1;
504 #include "si_debug_options.h"
505 } options;
506
507 /* Whether shaders are monolithic (1-part) or separate (3-part). */
508 bool use_monolithic_shaders;
509 bool record_llvm_ir;
510 bool has_rbplus; /* if RB+ registers exist */
511 bool rbplus_allowed; /* if RB+ is allowed */
512 bool dcc_msaa_allowed;
513 bool cpdma_prefetch_writes_memory;
514
515 struct slab_parent_pool pool_transfers;
516
517 /* Texture filter settings. */
518 int force_aniso; /* -1 = disabled */
519
520 /* Auxiliary context. Mainly used to initialize resources.
521 * It must be locked prior to using and flushed before unlocking. */
522 struct pipe_context *aux_context;
523 mtx_t aux_context_lock;
524
525 /* This must be in the screen, because UE4 uses one context for
526 * compilation and another one for rendering.
527 */
528 unsigned num_compilations;
529 /* Along with ST_DEBUG=precompile, this should show if applications
530 * are loading shaders on demand. This is a monotonic counter.
531 */
532 unsigned num_shaders_created;
533 unsigned num_shader_cache_hits;
534
535 /* GPU load thread. */
536 mtx_t gpu_load_mutex;
537 thrd_t gpu_load_thread;
538 union si_mmio_counters mmio_counters;
539 volatile unsigned gpu_load_stop_thread; /* bool */
540
541 /* Performance counters. */
542 struct si_perfcounters *perfcounters;
543
544 /* If pipe_screen wants to recompute and re-emit the framebuffer,
545 * sampler, and image states of all contexts, it should atomically
546 * increment this.
547 *
548 * Each context will compare this with its own last known value of
549 * the counter before drawing and re-emit the states accordingly.
550 */
551 unsigned dirty_tex_counter;
552 unsigned dirty_buf_counter;
553
554 /* Atomically increment this counter when an existing texture's
555 * metadata is enabled or disabled in a way that requires changing
556 * contexts' compressed texture binding masks.
557 */
558 unsigned compressed_colortex_counter;
559
560 struct {
561 /* Context flags to set so that all writes from earlier jobs
562 * in the CP are seen by L2 clients.
563 */
564 unsigned cp_to_L2;
565
566 /* Context flags to set so that all writes from earlier jobs
567 * that end in L2 are seen by CP.
568 */
569 unsigned L2_to_cp;
570 } barrier_flags;
571
572 mtx_t shader_parts_mutex;
573 struct si_shader_part *vs_prologs;
574 struct si_shader_part *tcs_epilogs;
575 struct si_shader_part *gs_prologs;
576 struct si_shader_part *ps_prologs;
577 struct si_shader_part *ps_epilogs;
578
579 /* Shader cache in memory.
580 *
581 * Design & limitations:
582 * - The shader cache is per screen (= per process), never saved to
583 * disk, and skips redundant shader compilations from TGSI to bytecode.
584 * - It can only be used with one-variant-per-shader support, in which
585 * case only the main (typically middle) part of shaders is cached.
586 * - Only VS, TCS, TES, PS are cached, out of which only the hw VS
587 * variants of VS and TES are cached, so LS and ES aren't.
588 * - GS and CS aren't cached, but it's certainly possible to cache
589 * those as well.
590 */
591 mtx_t shader_cache_mutex;
592 struct hash_table *shader_cache;
593
594 /* Shader compiler queue for multithreaded compilation. */
595 struct util_queue shader_compiler_queue;
596 /* Use at most 3 normal compiler threads on quadcore and better.
597 * Hyperthreaded CPUs report the number of threads, but we want
598 * the number of cores. We only need this many threads for shader-db. */
599 struct ac_llvm_compiler compiler[24]; /* used by the queue only */
600
601 struct util_queue shader_compiler_queue_low_priority;
602 /* Use at most 2 low priority threads on quadcore and better.
603 * We want to minimize the impact on multithreaded Mesa. */
604 struct ac_llvm_compiler compiler_lowp[10];
605 };
606
607 struct si_blend_color {
608 struct pipe_blend_color state;
609 bool any_nonzeros;
610 };
611
612 struct si_sampler_view {
613 struct pipe_sampler_view base;
614 /* [0..7] = image descriptor
615 * [4..7] = buffer descriptor */
616 uint32_t state[8];
617 uint32_t fmask_state[8];
618 const struct legacy_surf_level *base_level_info;
619 ubyte base_level;
620 ubyte block_width;
621 bool is_stencil_sampler;
622 bool is_integer;
623 bool dcc_incompatible;
624 };
625
626 #define SI_SAMPLER_STATE_MAGIC 0x34f1c35a
627
628 struct si_sampler_state {
629 #ifndef NDEBUG
630 unsigned magic;
631 #endif
632 uint32_t val[4];
633 uint32_t integer_val[4];
634 uint32_t upgraded_depth_val[4];
635 };
636
637 struct si_cs_shader_state {
638 struct si_compute *program;
639 struct si_compute *emitted_program;
640 unsigned offset;
641 bool initialized;
642 bool uses_scratch;
643 };
644
645 struct si_samplers {
646 struct pipe_sampler_view *views[SI_NUM_SAMPLERS];
647 struct si_sampler_state *sampler_states[SI_NUM_SAMPLERS];
648
649 /* The i-th bit is set if that element is enabled (non-NULL resource). */
650 unsigned enabled_mask;
651 uint32_t needs_depth_decompress_mask;
652 uint32_t needs_color_decompress_mask;
653 };
654
655 struct si_images {
656 struct pipe_image_view views[SI_NUM_IMAGES];
657 uint32_t needs_color_decompress_mask;
658 unsigned enabled_mask;
659 };
660
661 struct si_framebuffer {
662 struct pipe_framebuffer_state state;
663 unsigned colorbuf_enabled_4bit;
664 unsigned spi_shader_col_format;
665 unsigned spi_shader_col_format_alpha;
666 unsigned spi_shader_col_format_blend;
667 unsigned spi_shader_col_format_blend_alpha;
668 ubyte nr_samples:5; /* at most 16xAA */
669 ubyte log_samples:3; /* at most 4 = 16xAA */
670 ubyte nr_color_samples; /* at most 8xAA */
671 ubyte compressed_cb_mask;
672 ubyte uncompressed_cb_mask;
673 ubyte color_is_int8;
674 ubyte color_is_int10;
675 ubyte dirty_cbufs;
676 ubyte dcc_overwrite_combiner_watermark;
677 bool dirty_zsbuf;
678 bool any_dst_linear;
679 bool CB_has_shader_readable_metadata;
680 bool DB_has_shader_readable_metadata;
681 bool all_DCC_pipe_aligned;
682 };
683
684 enum si_quant_mode {
685 /* This is the list we want to support. */
686 SI_QUANT_MODE_16_8_FIXED_POINT_1_256TH,
687 SI_QUANT_MODE_14_10_FIXED_POINT_1_1024TH,
688 SI_QUANT_MODE_12_12_FIXED_POINT_1_4096TH,
689 };
690
691 struct si_signed_scissor {
692 int minx;
693 int miny;
694 int maxx;
695 int maxy;
696 enum si_quant_mode quant_mode;
697 };
698
699 struct si_viewports {
700 struct pipe_viewport_state states[SI_MAX_VIEWPORTS];
701 struct si_signed_scissor as_scissor[SI_MAX_VIEWPORTS];
702 bool y_inverted;
703 };
704
705 struct si_clip_state {
706 struct pipe_clip_state state;
707 bool any_nonzeros;
708 };
709
710 struct si_streamout_target {
711 struct pipe_stream_output_target b;
712
713 /* The buffer where BUFFER_FILLED_SIZE is stored. */
714 struct si_resource *buf_filled_size;
715 unsigned buf_filled_size_offset;
716 bool buf_filled_size_valid;
717
718 unsigned stride_in_dw;
719 };
720
721 struct si_streamout {
722 bool begin_emitted;
723
724 unsigned enabled_mask;
725 unsigned num_targets;
726 struct si_streamout_target *targets[PIPE_MAX_SO_BUFFERS];
727
728 unsigned append_bitmask;
729 bool suspended;
730
731 /* External state which comes from the vertex shader,
732 * it must be set explicitly when binding a shader. */
733 uint16_t *stride_in_dw;
734 unsigned enabled_stream_buffers_mask; /* stream0 buffers0-3 in 4 LSB */
735
736 /* The state of VGT_STRMOUT_BUFFER_(CONFIG|EN). */
737 unsigned hw_enabled_mask;
738
739 /* The state of VGT_STRMOUT_(CONFIG|EN). */
740 bool streamout_enabled;
741 bool prims_gen_query_enabled;
742 int num_prims_gen_queries;
743 };
744
745 /* A shader state consists of the shader selector, which is a constant state
746 * object shared by multiple contexts and shouldn't be modified, and
747 * the current shader variant selected for this context.
748 */
749 struct si_shader_ctx_state {
750 struct si_shader_selector *cso;
751 struct si_shader *current;
752 };
753
754 #define SI_NUM_VGT_PARAM_KEY_BITS 12
755 #define SI_NUM_VGT_PARAM_STATES (1 << SI_NUM_VGT_PARAM_KEY_BITS)
756
757 /* The IA_MULTI_VGT_PARAM key used to index the table of precomputed values.
758 * Some fields are set by state-change calls, most are set by draw_vbo.
759 */
760 union si_vgt_param_key {
761 struct {
762 #ifdef PIPE_ARCH_LITTLE_ENDIAN
763 unsigned prim:4;
764 unsigned uses_instancing:1;
765 unsigned multi_instances_smaller_than_primgroup:1;
766 unsigned primitive_restart:1;
767 unsigned count_from_stream_output:1;
768 unsigned line_stipple_enabled:1;
769 unsigned uses_tess:1;
770 unsigned tess_uses_prim_id:1;
771 unsigned uses_gs:1;
772 unsigned _pad:32 - SI_NUM_VGT_PARAM_KEY_BITS;
773 #else /* PIPE_ARCH_BIG_ENDIAN */
774 unsigned _pad:32 - SI_NUM_VGT_PARAM_KEY_BITS;
775 unsigned uses_gs:1;
776 unsigned tess_uses_prim_id:1;
777 unsigned uses_tess:1;
778 unsigned line_stipple_enabled:1;
779 unsigned count_from_stream_output:1;
780 unsigned primitive_restart:1;
781 unsigned multi_instances_smaller_than_primgroup:1;
782 unsigned uses_instancing:1;
783 unsigned prim:4;
784 #endif
785 } u;
786 uint32_t index;
787 };
788
789 #define SI_NUM_VGT_STAGES_KEY_BITS 4
790 #define SI_NUM_VGT_STAGES_STATES (1 << SI_NUM_VGT_STAGES_KEY_BITS)
791
792 /* The VGT_SHADER_STAGES key used to index the table of precomputed values.
793 * Some fields are set by state-change calls, most are set by draw_vbo.
794 */
795 union si_vgt_stages_key {
796 struct {
797 #ifdef PIPE_ARCH_LITTLE_ENDIAN
798 unsigned tess:1;
799 unsigned gs:1;
800 unsigned ngg:1; /* gfx10+ */
801 unsigned streamout:1; /* only used with NGG */
802 unsigned _pad:32 - SI_NUM_VGT_STAGES_KEY_BITS;
803 #else /* PIPE_ARCH_BIG_ENDIAN */
804 unsigned _pad:32 - SI_NUM_VGT_STAGES_KEY_BITS;
805 unsigned streamout:1;
806 unsigned ngg:1;
807 unsigned gs:1;
808 unsigned tess:1;
809 #endif
810 } u;
811 uint32_t index;
812 };
813
814 struct si_texture_handle
815 {
816 unsigned desc_slot;
817 bool desc_dirty;
818 struct pipe_sampler_view *view;
819 struct si_sampler_state sstate;
820 };
821
822 struct si_image_handle
823 {
824 unsigned desc_slot;
825 bool desc_dirty;
826 struct pipe_image_view view;
827 };
828
829 struct si_saved_cs {
830 struct pipe_reference reference;
831 struct si_context *ctx;
832 struct radeon_saved_cs gfx;
833 struct radeon_saved_cs compute;
834 struct si_resource *trace_buf;
835 unsigned trace_id;
836
837 unsigned gfx_last_dw;
838 unsigned compute_last_dw;
839 bool flushed;
840 int64_t time_flush;
841 };
842
843 struct si_sdma_upload {
844 struct si_resource *dst;
845 struct si_resource *src;
846 unsigned src_offset;
847 unsigned dst_offset;
848 unsigned size;
849 };
850
851 struct si_context {
852 struct pipe_context b; /* base class */
853
854 enum radeon_family family;
855 enum chip_class chip_class;
856
857 struct radeon_winsys *ws;
858 struct radeon_winsys_ctx *ctx;
859 struct radeon_cmdbuf *gfx_cs; /* compute IB if graphics is disabled */
860 struct radeon_cmdbuf *dma_cs;
861 struct pipe_fence_handle *last_gfx_fence;
862 struct pipe_fence_handle *last_sdma_fence;
863 struct si_resource *eop_bug_scratch;
864 struct u_upload_mgr *cached_gtt_allocator;
865 struct threaded_context *tc;
866 struct u_suballocator *allocator_zeroed_memory;
867 struct slab_child_pool pool_transfers;
868 struct slab_child_pool pool_transfers_unsync; /* for threaded_context */
869 struct pipe_device_reset_callback device_reset_callback;
870 struct u_log_context *log;
871 void *query_result_shader;
872 void *sh_query_result_shader;
873
874 void (*emit_cache_flush)(struct si_context *ctx);
875
876 struct blitter_context *blitter;
877 void *custom_dsa_flush;
878 void *custom_blend_resolve;
879 void *custom_blend_fmask_decompress;
880 void *custom_blend_eliminate_fastclear;
881 void *custom_blend_dcc_decompress;
882 void *vs_blit_pos;
883 void *vs_blit_pos_layered;
884 void *vs_blit_color;
885 void *vs_blit_color_layered;
886 void *vs_blit_texcoord;
887 void *cs_clear_buffer;
888 void *cs_copy_buffer;
889 void *cs_copy_image;
890 void *cs_copy_image_1d_array;
891 void *cs_clear_render_target;
892 void *cs_clear_render_target_1d_array;
893 void *cs_dcc_retile;
894 struct si_screen *screen;
895 struct pipe_debug_callback debug;
896 struct ac_llvm_compiler compiler; /* only non-threaded compilation */
897 struct si_shader_ctx_state fixed_func_tcs_shader;
898 /* Offset 0: EOP flush number; Offset 4: GDS prim restart counter */
899 struct si_resource *wait_mem_scratch;
900 unsigned wait_mem_number;
901 uint16_t prefetch_L2_mask;
902
903 bool has_graphics;
904 bool gfx_flush_in_progress:1;
905 bool gfx_last_ib_is_busy:1;
906 bool compute_is_busy:1;
907
908 unsigned num_gfx_cs_flushes;
909 unsigned initial_gfx_cs_size;
910 unsigned last_dirty_tex_counter;
911 unsigned last_dirty_buf_counter;
912 unsigned last_compressed_colortex_counter;
913 unsigned last_num_draw_calls;
914 unsigned flags; /* flush flags */
915 /* Current unaccounted memory usage. */
916 uint64_t vram;
917 uint64_t gtt;
918
919 /* Compute-based primitive discard. */
920 unsigned prim_discard_vertex_count_threshold;
921 struct pb_buffer *gds;
922 struct pb_buffer *gds_oa;
923 struct radeon_cmdbuf *prim_discard_compute_cs;
924 unsigned compute_gds_offset;
925 struct si_shader *compute_ib_last_shader;
926 uint32_t compute_rewind_va;
927 unsigned compute_num_prims_in_batch;
928 bool preserve_prim_restart_gds_at_flush;
929 /* index_ring is divided into 2 halves for doublebuffering. */
930 struct si_resource *index_ring;
931 unsigned index_ring_base; /* offset of a per-IB portion */
932 unsigned index_ring_offset; /* offset within a per-IB portion */
933 unsigned index_ring_size_per_ib; /* max available size per IB */
934 bool prim_discard_compute_ib_initialized;
935 /* For tracking the last execution barrier - it can be either
936 * a WRITE_DATA packet or a fence. */
937 uint32_t *last_pkt3_write_data;
938 struct si_resource *barrier_buf;
939 unsigned barrier_buf_offset;
940 struct pipe_fence_handle *last_ib_barrier_fence;
941 struct si_resource *last_ib_barrier_buf;
942 unsigned last_ib_barrier_buf_offset;
943
944 /* Atoms (direct states). */
945 union si_state_atoms atoms;
946 unsigned dirty_atoms; /* mask */
947 /* PM4 states (precomputed immutable states) */
948 unsigned dirty_states;
949 union si_state queued;
950 union si_state emitted;
951
952 /* Atom declarations. */
953 struct si_framebuffer framebuffer;
954 unsigned sample_locs_num_samples;
955 uint16_t sample_mask;
956 unsigned last_cb_target_mask;
957 struct si_blend_color blend_color;
958 struct si_clip_state clip_state;
959 struct si_shader_data shader_pointers;
960 struct si_stencil_ref stencil_ref;
961 struct pipe_scissor_state scissors[SI_MAX_VIEWPORTS];
962 struct si_streamout streamout;
963 struct si_viewports viewports;
964 unsigned num_window_rectangles;
965 bool window_rectangles_include;
966 struct pipe_scissor_state window_rectangles[4];
967
968 /* Precomputed states. */
969 struct si_pm4_state *init_config;
970 struct si_pm4_state *init_config_gs_rings;
971 bool init_config_has_vgt_flush;
972 struct si_pm4_state *vgt_shader_config[SI_NUM_VGT_STAGES_STATES];
973
974 /* shaders */
975 struct si_shader_ctx_state ps_shader;
976 struct si_shader_ctx_state gs_shader;
977 struct si_shader_ctx_state vs_shader;
978 struct si_shader_ctx_state tcs_shader;
979 struct si_shader_ctx_state tes_shader;
980 struct si_shader_ctx_state cs_prim_discard_state;
981 struct si_cs_shader_state cs_shader_state;
982
983 /* shader information */
984 struct si_vertex_elements *vertex_elements;
985 unsigned sprite_coord_enable;
986 unsigned cs_max_waves_per_sh;
987 bool flatshade;
988 bool do_update_shaders;
989
990 /* vertex buffer descriptors */
991 uint32_t *vb_descriptors_gpu_list;
992 struct si_resource *vb_descriptors_buffer;
993 unsigned vb_descriptors_offset;
994
995 /* shader descriptors */
996 struct si_descriptors descriptors[SI_NUM_DESCS];
997 unsigned descriptors_dirty;
998 unsigned shader_pointers_dirty;
999 unsigned shader_needs_decompress_mask;
1000 struct si_buffer_resources rw_buffers;
1001 struct si_buffer_resources const_and_shader_buffers[SI_NUM_SHADERS];
1002 struct si_samplers samplers[SI_NUM_SHADERS];
1003 struct si_images images[SI_NUM_SHADERS];
1004 bool bo_list_add_all_resident_resources;
1005 bool bo_list_add_all_gfx_resources;
1006 bool bo_list_add_all_compute_resources;
1007
1008 /* other shader resources */
1009 struct pipe_constant_buffer null_const_buf; /* used for set_constant_buffer(NULL) on GFX7 */
1010 struct pipe_resource *esgs_ring;
1011 struct pipe_resource *gsvs_ring;
1012 struct pipe_resource *tess_rings;
1013 union pipe_color_union *border_color_table; /* in CPU memory, any endian */
1014 struct si_resource *border_color_buffer;
1015 union pipe_color_union *border_color_map; /* in VRAM (slow access), little endian */
1016 unsigned border_color_count;
1017 unsigned num_vs_blit_sgprs;
1018 uint32_t vs_blit_sh_data[SI_VS_BLIT_SGPRS_POS_TEXCOORD];
1019 uint32_t cs_user_data[4];
1020
1021 /* Vertex and index buffers. */
1022 bool vertex_buffers_dirty;
1023 bool vertex_buffer_pointer_dirty;
1024 struct pipe_vertex_buffer vertex_buffer[SI_NUM_VERTEX_BUFFERS];
1025 uint16_t vertex_buffer_unaligned; /* bitmask of not dword-aligned buffers */
1026
1027 /* MSAA config state. */
1028 int ps_iter_samples;
1029 bool ps_uses_fbfetch;
1030 bool smoothing_enabled;
1031
1032 /* DB render state. */
1033 unsigned ps_db_shader_control;
1034 unsigned dbcb_copy_sample;
1035 bool dbcb_depth_copy_enabled:1;
1036 bool dbcb_stencil_copy_enabled:1;
1037 bool db_flush_depth_inplace:1;
1038 bool db_flush_stencil_inplace:1;
1039 bool db_depth_clear:1;
1040 bool db_depth_disable_expclear:1;
1041 bool db_stencil_clear:1;
1042 bool db_stencil_disable_expclear:1;
1043 bool occlusion_queries_disabled:1;
1044 bool generate_mipmap_for_depth:1;
1045
1046 /* Emitted draw state. */
1047 bool gs_tri_strip_adj_fix:1;
1048 bool ls_vgpr_fix:1;
1049 bool prim_discard_cs_instancing:1;
1050 bool ngg:1;
1051 int last_index_size;
1052 int last_base_vertex;
1053 int last_start_instance;
1054 int last_instance_count;
1055 int last_drawid;
1056 int last_sh_base_reg;
1057 int last_primitive_restart_en;
1058 int last_restart_index;
1059 int last_prim;
1060 int last_multi_vgt_param;
1061 int last_rast_prim;
1062 int last_flatshade_first;
1063 unsigned last_sc_line_stipple;
1064 unsigned current_vs_state;
1065 unsigned last_vs_state;
1066 enum pipe_prim_type current_rast_prim; /* primitive type after TES, GS */
1067
1068 /* Scratch buffer */
1069 struct si_resource *scratch_buffer;
1070 unsigned scratch_waves;
1071 unsigned spi_tmpring_size;
1072
1073 struct si_resource *compute_scratch_buffer;
1074
1075 /* Emitted derived tessellation state. */
1076 /* Local shader (VS), or HS if LS-HS are merged. */
1077 struct si_shader *last_ls;
1078 struct si_shader_selector *last_tcs;
1079 int last_num_tcs_input_cp;
1080 int last_tes_sh_base;
1081 bool last_tess_uses_primid;
1082 unsigned last_num_patches;
1083 int last_ls_hs_config;
1084
1085 /* Debug state. */
1086 bool is_debug;
1087 struct si_saved_cs *current_saved_cs;
1088 uint64_t dmesg_timestamp;
1089 unsigned apitrace_call_number;
1090
1091 /* Other state */
1092 bool need_check_render_feedback;
1093 bool decompression_enabled;
1094 bool dpbb_force_off;
1095 bool vs_writes_viewport_index;
1096 bool vs_disables_clipping_viewport;
1097
1098 /* Precomputed IA_MULTI_VGT_PARAM */
1099 union si_vgt_param_key ia_multi_vgt_param_key;
1100 unsigned ia_multi_vgt_param[SI_NUM_VGT_PARAM_STATES];
1101
1102 /* Bindless descriptors. */
1103 struct si_descriptors bindless_descriptors;
1104 struct util_idalloc bindless_used_slots;
1105 unsigned num_bindless_descriptors;
1106 bool bindless_descriptors_dirty;
1107 bool graphics_bindless_pointer_dirty;
1108 bool compute_bindless_pointer_dirty;
1109
1110 /* Allocated bindless handles */
1111 struct hash_table *tex_handles;
1112 struct hash_table *img_handles;
1113
1114 /* Resident bindless handles */
1115 struct util_dynarray resident_tex_handles;
1116 struct util_dynarray resident_img_handles;
1117
1118 /* Resident bindless handles which need decompression */
1119 struct util_dynarray resident_tex_needs_color_decompress;
1120 struct util_dynarray resident_img_needs_color_decompress;
1121 struct util_dynarray resident_tex_needs_depth_decompress;
1122
1123 /* Bindless state */
1124 bool uses_bindless_samplers;
1125 bool uses_bindless_images;
1126
1127 /* MSAA sample locations.
1128 * The first index is the sample index.
1129 * The second index is the coordinate: X, Y. */
1130 struct {
1131 float x1[1][2];
1132 float x2[2][2];
1133 float x4[4][2];
1134 float x8[8][2];
1135 float x16[16][2];
1136 } sample_positions;
1137 struct pipe_resource *sample_pos_buffer;
1138
1139 /* Misc stats. */
1140 unsigned num_draw_calls;
1141 unsigned num_decompress_calls;
1142 unsigned num_mrt_draw_calls;
1143 unsigned num_prim_restart_calls;
1144 unsigned num_spill_draw_calls;
1145 unsigned num_compute_calls;
1146 unsigned num_spill_compute_calls;
1147 unsigned num_dma_calls;
1148 unsigned num_cp_dma_calls;
1149 unsigned num_vs_flushes;
1150 unsigned num_ps_flushes;
1151 unsigned num_cs_flushes;
1152 unsigned num_cb_cache_flushes;
1153 unsigned num_db_cache_flushes;
1154 unsigned num_L2_invalidates;
1155 unsigned num_L2_writebacks;
1156 unsigned num_resident_handles;
1157 uint64_t num_alloc_tex_transfer_bytes;
1158 unsigned last_tex_ps_draw_ratio; /* for query */
1159 unsigned compute_num_verts_accepted;
1160 unsigned compute_num_verts_rejected;
1161 unsigned compute_num_verts_ineligible; /* due to low vertex count */
1162 unsigned context_roll;
1163
1164 /* Queries. */
1165 /* Maintain the list of active queries for pausing between IBs. */
1166 int num_occlusion_queries;
1167 int num_perfect_occlusion_queries;
1168 int num_pipeline_stat_queries;
1169 struct list_head active_queries;
1170 unsigned num_cs_dw_queries_suspend;
1171
1172 /* Render condition. */
1173 struct pipe_query *render_cond;
1174 unsigned render_cond_mode;
1175 bool render_cond_invert;
1176 bool render_cond_force_off; /* for u_blitter */
1177
1178 /* For uploading data via GTT and copy to VRAM on context flush via SDMA. */
1179 bool sdma_uploads_in_progress;
1180 struct si_sdma_upload *sdma_uploads;
1181 unsigned num_sdma_uploads;
1182 unsigned max_sdma_uploads;
1183
1184 /* Shader-based queries. */
1185 struct list_head shader_query_buffers;
1186 unsigned num_active_shader_queries;
1187
1188 /* Statistics gathering for the DCC enablement heuristic. It can't be
1189 * in si_texture because si_texture can be shared by multiple
1190 * contexts. This is for back buffers only. We shouldn't get too many
1191 * of those.
1192 *
1193 * X11 DRI3 rotates among a finite set of back buffers. They should
1194 * all fit in this array. If they don't, separate DCC might never be
1195 * enabled by DCC stat gathering.
1196 */
1197 struct {
1198 struct si_texture *tex;
1199 /* Query queue: 0 = usually active, 1 = waiting, 2 = readback. */
1200 struct pipe_query *ps_stats[3];
1201 /* If all slots are used and another slot is needed,
1202 * the least recently used slot is evicted based on this. */
1203 int64_t last_use_timestamp;
1204 bool query_active;
1205 } dcc_stats[5];
1206
1207 /* Copy one resource to another using async DMA. */
1208 void (*dma_copy)(struct pipe_context *ctx,
1209 struct pipe_resource *dst,
1210 unsigned dst_level,
1211 unsigned dst_x, unsigned dst_y, unsigned dst_z,
1212 struct pipe_resource *src,
1213 unsigned src_level,
1214 const struct pipe_box *src_box);
1215
1216 struct si_tracked_regs tracked_regs;
1217 };
1218
1219 /* cik_sdma.c */
1220 void cik_init_sdma_functions(struct si_context *sctx);
1221
1222 /* si_blit.c */
1223 enum si_blitter_op /* bitmask */
1224 {
1225 SI_SAVE_TEXTURES = 1,
1226 SI_SAVE_FRAMEBUFFER = 2,
1227 SI_SAVE_FRAGMENT_STATE = 4,
1228 SI_DISABLE_RENDER_COND = 8,
1229 };
1230
1231 void si_blitter_begin(struct si_context *sctx, enum si_blitter_op op);
1232 void si_blitter_end(struct si_context *sctx);
1233 void si_init_blit_functions(struct si_context *sctx);
1234 void si_decompress_textures(struct si_context *sctx, unsigned shader_mask);
1235 void si_resource_copy_region(struct pipe_context *ctx,
1236 struct pipe_resource *dst,
1237 unsigned dst_level,
1238 unsigned dstx, unsigned dsty, unsigned dstz,
1239 struct pipe_resource *src,
1240 unsigned src_level,
1241 const struct pipe_box *src_box);
1242 void si_decompress_dcc(struct si_context *sctx, struct si_texture *tex);
1243
1244 /* si_buffer.c */
1245 bool si_rings_is_buffer_referenced(struct si_context *sctx,
1246 struct pb_buffer *buf,
1247 enum radeon_bo_usage usage);
1248 void *si_buffer_map_sync_with_rings(struct si_context *sctx,
1249 struct si_resource *resource,
1250 unsigned usage);
1251 void si_init_resource_fields(struct si_screen *sscreen,
1252 struct si_resource *res,
1253 uint64_t size, unsigned alignment);
1254 bool si_alloc_resource(struct si_screen *sscreen,
1255 struct si_resource *res);
1256 struct pipe_resource *pipe_aligned_buffer_create(struct pipe_screen *screen,
1257 unsigned flags, unsigned usage,
1258 unsigned size, unsigned alignment);
1259 struct si_resource *si_aligned_buffer_create(struct pipe_screen *screen,
1260 unsigned flags, unsigned usage,
1261 unsigned size, unsigned alignment);
1262 void si_replace_buffer_storage(struct pipe_context *ctx,
1263 struct pipe_resource *dst,
1264 struct pipe_resource *src);
1265 void si_init_screen_buffer_functions(struct si_screen *sscreen);
1266 void si_init_buffer_functions(struct si_context *sctx);
1267
1268 /* si_clear.c */
1269 enum pipe_format si_simplify_cb_format(enum pipe_format format);
1270 bool vi_alpha_is_on_msb(struct si_screen *sscreen, enum pipe_format format);
1271 void vi_dcc_clear_level(struct si_context *sctx,
1272 struct si_texture *tex,
1273 unsigned level, unsigned clear_value);
1274 void si_init_clear_functions(struct si_context *sctx);
1275
1276 /* si_compute_blit.c */
1277 unsigned si_get_flush_flags(struct si_context *sctx, enum si_coherency coher,
1278 enum si_cache_policy cache_policy);
1279 void si_clear_buffer(struct si_context *sctx, struct pipe_resource *dst,
1280 uint64_t offset, uint64_t size, uint32_t *clear_value,
1281 uint32_t clear_value_size, enum si_coherency coher,
1282 bool force_cpdma);
1283 void si_copy_buffer(struct si_context *sctx,
1284 struct pipe_resource *dst, struct pipe_resource *src,
1285 uint64_t dst_offset, uint64_t src_offset, unsigned size);
1286 void si_compute_copy_image(struct si_context *sctx,
1287 struct pipe_resource *dst,
1288 unsigned dst_level,
1289 struct pipe_resource *src,
1290 unsigned src_level,
1291 unsigned dstx, unsigned dsty, unsigned dstz,
1292 const struct pipe_box *src_box);
1293 void si_compute_clear_render_target(struct pipe_context *ctx,
1294 struct pipe_surface *dstsurf,
1295 const union pipe_color_union *color,
1296 unsigned dstx, unsigned dsty,
1297 unsigned width, unsigned height,
1298 bool render_condition_enabled);
1299 void si_retile_dcc(struct si_context *sctx, struct si_texture *tex);
1300 void si_init_compute_blit_functions(struct si_context *sctx);
1301
1302 /* si_cp_dma.c */
1303 #define SI_CPDMA_SKIP_CHECK_CS_SPACE (1 << 0) /* don't call need_cs_space */
1304 #define SI_CPDMA_SKIP_SYNC_AFTER (1 << 1) /* don't wait for DMA after the copy */
1305 #define SI_CPDMA_SKIP_SYNC_BEFORE (1 << 2) /* don't wait for DMA before the copy (RAW hazards) */
1306 #define SI_CPDMA_SKIP_GFX_SYNC (1 << 3) /* don't flush caches and don't wait for PS/CS */
1307 #define SI_CPDMA_SKIP_BO_LIST_UPDATE (1 << 4) /* don't update the BO list */
1308 #define SI_CPDMA_SKIP_ALL (SI_CPDMA_SKIP_CHECK_CS_SPACE | \
1309 SI_CPDMA_SKIP_SYNC_AFTER | \
1310 SI_CPDMA_SKIP_SYNC_BEFORE | \
1311 SI_CPDMA_SKIP_GFX_SYNC | \
1312 SI_CPDMA_SKIP_BO_LIST_UPDATE)
1313
1314 void si_cp_dma_wait_for_idle(struct si_context *sctx);
1315 void si_cp_dma_clear_buffer(struct si_context *sctx, struct radeon_cmdbuf *cs,
1316 struct pipe_resource *dst, uint64_t offset,
1317 uint64_t size, unsigned value, unsigned user_flags,
1318 enum si_coherency coher, enum si_cache_policy cache_policy);
1319 void si_cp_dma_copy_buffer(struct si_context *sctx,
1320 struct pipe_resource *dst, struct pipe_resource *src,
1321 uint64_t dst_offset, uint64_t src_offset, unsigned size,
1322 unsigned user_flags, enum si_coherency coher,
1323 enum si_cache_policy cache_policy);
1324 void cik_prefetch_TC_L2_async(struct si_context *sctx, struct pipe_resource *buf,
1325 uint64_t offset, unsigned size);
1326 void cik_emit_prefetch_L2(struct si_context *sctx, bool vertex_stage_only);
1327 void si_test_gds(struct si_context *sctx);
1328 void si_cp_write_data(struct si_context *sctx, struct si_resource *buf,
1329 unsigned offset, unsigned size, unsigned dst_sel,
1330 unsigned engine, const void *data);
1331 void si_cp_copy_data(struct si_context *sctx, struct radeon_cmdbuf *cs,
1332 unsigned dst_sel, struct si_resource *dst, unsigned dst_offset,
1333 unsigned src_sel, struct si_resource *src, unsigned src_offset);
1334
1335 /* si_debug.c */
1336 void si_save_cs(struct radeon_winsys *ws, struct radeon_cmdbuf *cs,
1337 struct radeon_saved_cs *saved, bool get_buffer_list);
1338 void si_clear_saved_cs(struct radeon_saved_cs *saved);
1339 void si_destroy_saved_cs(struct si_saved_cs *scs);
1340 void si_auto_log_cs(void *data, struct u_log_context *log);
1341 void si_log_hw_flush(struct si_context *sctx);
1342 void si_log_draw_state(struct si_context *sctx, struct u_log_context *log);
1343 void si_log_compute_state(struct si_context *sctx, struct u_log_context *log);
1344 void si_init_debug_functions(struct si_context *sctx);
1345 void si_check_vm_faults(struct si_context *sctx,
1346 struct radeon_saved_cs *saved, enum ring_type ring);
1347 bool si_replace_shader(unsigned num, struct si_shader_binary *binary);
1348
1349 /* si_dma.c */
1350 void si_init_dma_functions(struct si_context *sctx);
1351
1352 /* si_dma_cs.c */
1353 void si_dma_emit_timestamp(struct si_context *sctx, struct si_resource *dst,
1354 uint64_t offset);
1355 void si_sdma_clear_buffer(struct si_context *sctx, struct pipe_resource *dst,
1356 uint64_t offset, uint64_t size, unsigned clear_value);
1357 void si_need_dma_space(struct si_context *ctx, unsigned num_dw,
1358 struct si_resource *dst, struct si_resource *src);
1359 void si_flush_dma_cs(struct si_context *ctx, unsigned flags,
1360 struct pipe_fence_handle **fence);
1361 void si_screen_clear_buffer(struct si_screen *sscreen, struct pipe_resource *dst,
1362 uint64_t offset, uint64_t size, unsigned value);
1363
1364 /* si_fence.c */
1365 void si_cp_release_mem(struct si_context *ctx, struct radeon_cmdbuf *cs,
1366 unsigned event, unsigned event_flags,
1367 unsigned dst_sel, unsigned int_sel, unsigned data_sel,
1368 struct si_resource *buf, uint64_t va,
1369 uint32_t new_fence, unsigned query_type);
1370 unsigned si_cp_write_fence_dwords(struct si_screen *screen);
1371 void si_cp_wait_mem(struct si_context *ctx, struct radeon_cmdbuf *cs,
1372 uint64_t va, uint32_t ref, uint32_t mask, unsigned flags);
1373 void si_init_fence_functions(struct si_context *ctx);
1374 void si_init_screen_fence_functions(struct si_screen *screen);
1375 struct pipe_fence_handle *si_create_fence(struct pipe_context *ctx,
1376 struct tc_unflushed_batch_token *tc_token);
1377
1378 /* si_get.c */
1379 void si_init_screen_get_functions(struct si_screen *sscreen);
1380
1381 /* si_gfx_cs.c */
1382 void si_flush_gfx_cs(struct si_context *ctx, unsigned flags,
1383 struct pipe_fence_handle **fence);
1384 void si_allocate_gds(struct si_context *ctx);
1385 void si_begin_new_gfx_cs(struct si_context *ctx);
1386 void si_need_gfx_cs_space(struct si_context *ctx);
1387 void si_unref_sdma_uploads(struct si_context *sctx);
1388
1389 /* si_gpu_load.c */
1390 void si_gpu_load_kill_thread(struct si_screen *sscreen);
1391 uint64_t si_begin_counter(struct si_screen *sscreen, unsigned type);
1392 unsigned si_end_counter(struct si_screen *sscreen, unsigned type,
1393 uint64_t begin);
1394
1395 /* si_compute.c */
1396 void si_emit_initial_compute_regs(struct si_context *sctx, struct radeon_cmdbuf *cs);
1397 unsigned si_get_compute_resource_limits(struct si_screen *sscreen,
1398 unsigned waves_per_threadgroup,
1399 unsigned max_waves_per_sh,
1400 unsigned threadgroups_per_cu);
1401 void si_init_compute_functions(struct si_context *sctx);
1402
1403 /* si_compute_prim_discard.c */
1404 enum si_prim_discard_outcome {
1405 SI_PRIM_DISCARD_ENABLED,
1406 SI_PRIM_DISCARD_DISABLED,
1407 SI_PRIM_DISCARD_DRAW_SPLIT,
1408 };
1409
1410 void si_build_prim_discard_compute_shader(struct si_shader_context *ctx);
1411 enum si_prim_discard_outcome
1412 si_prepare_prim_discard_or_split_draw(struct si_context *sctx,
1413 const struct pipe_draw_info *info,
1414 bool primitive_restart);
1415 void si_compute_signal_gfx(struct si_context *sctx);
1416 void si_dispatch_prim_discard_cs_and_draw(struct si_context *sctx,
1417 const struct pipe_draw_info *info,
1418 unsigned index_size,
1419 unsigned base_vertex,
1420 uint64_t input_indexbuf_va,
1421 unsigned input_indexbuf_max_elements);
1422 void si_initialize_prim_discard_tunables(struct si_context *sctx);
1423
1424 /* si_perfcounters.c */
1425 void si_init_perfcounters(struct si_screen *screen);
1426 void si_destroy_perfcounters(struct si_screen *screen);
1427
1428 /* si_pipe.c */
1429 bool si_check_device_reset(struct si_context *sctx);
1430
1431 /* si_query.c */
1432 void si_init_screen_query_functions(struct si_screen *sscreen);
1433 void si_init_query_functions(struct si_context *sctx);
1434 void si_suspend_queries(struct si_context *sctx);
1435 void si_resume_queries(struct si_context *sctx);
1436
1437 /* si_shaderlib_tgsi.c */
1438 void *si_get_blitter_vs(struct si_context *sctx, enum blitter_attrib_type type,
1439 unsigned num_layers);
1440 void *si_create_fixed_func_tcs(struct si_context *sctx);
1441 void *si_create_dma_compute_shader(struct pipe_context *ctx,
1442 unsigned num_dwords_per_thread,
1443 bool dst_stream_cache_policy, bool is_copy);
1444 void *si_create_copy_image_compute_shader(struct pipe_context *ctx);
1445 void *si_create_copy_image_compute_shader_1d_array(struct pipe_context *ctx);
1446 void *si_clear_render_target_shader(struct pipe_context *ctx);
1447 void *si_clear_render_target_shader_1d_array(struct pipe_context *ctx);
1448 void *si_create_dcc_retile_cs(struct pipe_context *ctx);
1449 void *si_create_query_result_cs(struct si_context *sctx);
1450 void *gfx10_create_sh_query_result_cs(struct si_context *sctx);
1451
1452 /* gfx10_query.c */
1453 void gfx10_init_query(struct si_context *sctx);
1454 void gfx10_destroy_query(struct si_context *sctx);
1455
1456 /* si_test_dma.c */
1457 void si_test_dma(struct si_screen *sscreen);
1458
1459 /* si_test_clearbuffer.c */
1460 void si_test_dma_perf(struct si_screen *sscreen);
1461
1462 /* si_uvd.c */
1463 struct pipe_video_codec *si_uvd_create_decoder(struct pipe_context *context,
1464 const struct pipe_video_codec *templ);
1465
1466 struct pipe_video_buffer *si_video_buffer_create(struct pipe_context *pipe,
1467 const struct pipe_video_buffer *tmpl);
1468
1469 /* si_viewport.c */
1470 void si_update_vs_viewport_state(struct si_context *ctx);
1471 void si_init_viewport_functions(struct si_context *ctx);
1472
1473 /* si_texture.c */
1474 bool si_prepare_for_dma_blit(struct si_context *sctx,
1475 struct si_texture *dst,
1476 unsigned dst_level, unsigned dstx,
1477 unsigned dsty, unsigned dstz,
1478 struct si_texture *src,
1479 unsigned src_level,
1480 const struct pipe_box *src_box);
1481 void si_eliminate_fast_color_clear(struct si_context *sctx,
1482 struct si_texture *tex);
1483 void si_texture_discard_cmask(struct si_screen *sscreen,
1484 struct si_texture *tex);
1485 bool si_init_flushed_depth_texture(struct pipe_context *ctx,
1486 struct pipe_resource *texture);
1487 void si_print_texture_info(struct si_screen *sscreen,
1488 struct si_texture *tex, struct u_log_context *log);
1489 struct pipe_resource *si_texture_create(struct pipe_screen *screen,
1490 const struct pipe_resource *templ);
1491 bool vi_dcc_formats_compatible(struct si_screen *sscreen,
1492 enum pipe_format format1,
1493 enum pipe_format format2);
1494 bool vi_dcc_formats_are_incompatible(struct pipe_resource *tex,
1495 unsigned level,
1496 enum pipe_format view_format);
1497 void vi_disable_dcc_if_incompatible_format(struct si_context *sctx,
1498 struct pipe_resource *tex,
1499 unsigned level,
1500 enum pipe_format view_format);
1501 struct pipe_surface *si_create_surface_custom(struct pipe_context *pipe,
1502 struct pipe_resource *texture,
1503 const struct pipe_surface *templ,
1504 unsigned width0, unsigned height0,
1505 unsigned width, unsigned height);
1506 unsigned si_translate_colorswap(enum pipe_format format, bool do_endian_swap);
1507 void vi_separate_dcc_try_enable(struct si_context *sctx,
1508 struct si_texture *tex);
1509 void vi_separate_dcc_start_query(struct si_context *sctx,
1510 struct si_texture *tex);
1511 void vi_separate_dcc_stop_query(struct si_context *sctx,
1512 struct si_texture *tex);
1513 void vi_separate_dcc_process_and_reset_stats(struct pipe_context *ctx,
1514 struct si_texture *tex);
1515 bool si_texture_disable_dcc(struct si_context *sctx,
1516 struct si_texture *tex);
1517 void si_init_screen_texture_functions(struct si_screen *sscreen);
1518 void si_init_context_texture_functions(struct si_context *sctx);
1519
1520
1521 /*
1522 * common helpers
1523 */
1524
1525 static inline struct si_resource *si_resource(struct pipe_resource *r)
1526 {
1527 return (struct si_resource*)r;
1528 }
1529
1530 static inline void
1531 si_resource_reference(struct si_resource **ptr, struct si_resource *res)
1532 {
1533 pipe_resource_reference((struct pipe_resource **)ptr,
1534 (struct pipe_resource *)res);
1535 }
1536
1537 static inline void
1538 si_texture_reference(struct si_texture **ptr, struct si_texture *res)
1539 {
1540 pipe_resource_reference((struct pipe_resource **)ptr, &res->buffer.b.b);
1541 }
1542
1543 static inline bool
1544 vi_dcc_enabled(struct si_texture *tex, unsigned level)
1545 {
1546 return tex->dcc_offset && level < tex->surface.num_dcc_levels;
1547 }
1548
1549 static inline unsigned
1550 si_tile_mode_index(struct si_texture *tex, unsigned level, bool stencil)
1551 {
1552 if (stencil)
1553 return tex->surface.u.legacy.stencil_tiling_index[level];
1554 else
1555 return tex->surface.u.legacy.tiling_index[level];
1556 }
1557
1558 static inline unsigned
1559 si_get_minimum_num_gfx_cs_dwords(struct si_context *sctx)
1560 {
1561 /* Don't count the needed CS space exactly and just use an upper bound.
1562 *
1563 * Also reserve space for stopping queries at the end of IB, because
1564 * the number of active queries is unlimited in theory.
1565 */
1566 return 2048 + sctx->num_cs_dw_queries_suspend;
1567 }
1568
1569 static inline void
1570 si_context_add_resource_size(struct si_context *sctx, struct pipe_resource *r)
1571 {
1572 if (r) {
1573 /* Add memory usage for need_gfx_cs_space */
1574 sctx->vram += si_resource(r)->vram_usage;
1575 sctx->gtt += si_resource(r)->gart_usage;
1576 }
1577 }
1578
1579 static inline void
1580 si_invalidate_draw_sh_constants(struct si_context *sctx)
1581 {
1582 sctx->last_base_vertex = SI_BASE_VERTEX_UNKNOWN;
1583 sctx->last_instance_count = SI_INSTANCE_COUNT_UNKNOWN;
1584 }
1585
1586 static inline unsigned
1587 si_get_atom_bit(struct si_context *sctx, struct si_atom *atom)
1588 {
1589 return 1 << (atom - sctx->atoms.array);
1590 }
1591
1592 static inline void
1593 si_set_atom_dirty(struct si_context *sctx, struct si_atom *atom, bool dirty)
1594 {
1595 unsigned bit = si_get_atom_bit(sctx, atom);
1596
1597 if (dirty)
1598 sctx->dirty_atoms |= bit;
1599 else
1600 sctx->dirty_atoms &= ~bit;
1601 }
1602
1603 static inline bool
1604 si_is_atom_dirty(struct si_context *sctx, struct si_atom *atom)
1605 {
1606 return (sctx->dirty_atoms & si_get_atom_bit(sctx, atom)) != 0;
1607 }
1608
1609 static inline void
1610 si_mark_atom_dirty(struct si_context *sctx, struct si_atom *atom)
1611 {
1612 si_set_atom_dirty(sctx, atom, true);
1613 }
1614
1615 static inline struct si_shader_ctx_state *si_get_vs(struct si_context *sctx)
1616 {
1617 if (sctx->gs_shader.cso)
1618 return &sctx->gs_shader;
1619 if (sctx->tes_shader.cso)
1620 return &sctx->tes_shader;
1621
1622 return &sctx->vs_shader;
1623 }
1624
1625 static inline struct tgsi_shader_info *si_get_vs_info(struct si_context *sctx)
1626 {
1627 struct si_shader_ctx_state *vs = si_get_vs(sctx);
1628
1629 return vs->cso ? &vs->cso->info : NULL;
1630 }
1631
1632 static inline struct si_shader* si_get_vs_state(struct si_context *sctx)
1633 {
1634 if (sctx->gs_shader.cso)
1635 return sctx->gs_shader.cso->gs_copy_shader;
1636
1637 struct si_shader_ctx_state *vs = si_get_vs(sctx);
1638 return vs->current ? vs->current : NULL;
1639 }
1640
1641 static inline bool si_can_dump_shader(struct si_screen *sscreen,
1642 unsigned processor)
1643 {
1644 return sscreen->debug_flags & (1 << processor);
1645 }
1646
1647 static inline bool si_get_strmout_en(struct si_context *sctx)
1648 {
1649 return sctx->streamout.streamout_enabled ||
1650 sctx->streamout.prims_gen_query_enabled;
1651 }
1652
1653 static inline unsigned
1654 si_optimal_tcc_alignment(struct si_context *sctx, unsigned upload_size)
1655 {
1656 unsigned alignment, tcc_cache_line_size;
1657
1658 /* If the upload size is less than the cache line size (e.g. 16, 32),
1659 * the whole thing will fit into a cache line if we align it to its size.
1660 * The idea is that multiple small uploads can share a cache line.
1661 * If the upload size is greater, align it to the cache line size.
1662 */
1663 alignment = util_next_power_of_two(upload_size);
1664 tcc_cache_line_size = sctx->screen->info.tcc_cache_line_size;
1665 return MIN2(alignment, tcc_cache_line_size);
1666 }
1667
1668 static inline void
1669 si_saved_cs_reference(struct si_saved_cs **dst, struct si_saved_cs *src)
1670 {
1671 if (pipe_reference(&(*dst)->reference, &src->reference))
1672 si_destroy_saved_cs(*dst);
1673
1674 *dst = src;
1675 }
1676
1677 static inline void
1678 si_make_CB_shader_coherent(struct si_context *sctx, unsigned num_samples,
1679 bool shaders_read_metadata, bool dcc_pipe_aligned)
1680 {
1681 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_CB |
1682 SI_CONTEXT_INV_VCACHE;
1683
1684 if (sctx->chip_class >= GFX9) {
1685 /* Single-sample color is coherent with shaders on GFX9, but
1686 * L2 metadata must be flushed if shaders read metadata.
1687 * (DCC, CMASK).
1688 */
1689 if (num_samples >= 2 ||
1690 (shaders_read_metadata && !dcc_pipe_aligned))
1691 sctx->flags |= SI_CONTEXT_INV_L2;
1692 else if (shaders_read_metadata)
1693 sctx->flags |= SI_CONTEXT_INV_L2_METADATA;
1694 } else {
1695 /* GFX6-GFX8 */
1696 sctx->flags |= SI_CONTEXT_INV_L2;
1697 }
1698 }
1699
1700 static inline void
1701 si_make_DB_shader_coherent(struct si_context *sctx, unsigned num_samples,
1702 bool include_stencil, bool shaders_read_metadata)
1703 {
1704 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_DB |
1705 SI_CONTEXT_INV_VCACHE;
1706
1707 if (sctx->chip_class >= GFX9) {
1708 /* Single-sample depth (not stencil) is coherent with shaders
1709 * on GFX9, but L2 metadata must be flushed if shaders read
1710 * metadata.
1711 */
1712 if (num_samples >= 2 || include_stencil)
1713 sctx->flags |= SI_CONTEXT_INV_L2;
1714 else if (shaders_read_metadata)
1715 sctx->flags |= SI_CONTEXT_INV_L2_METADATA;
1716 } else {
1717 /* GFX6-GFX8 */
1718 sctx->flags |= SI_CONTEXT_INV_L2;
1719 }
1720 }
1721
1722 static inline bool
1723 si_can_sample_zs(struct si_texture *tex, bool stencil_sampler)
1724 {
1725 return (stencil_sampler && tex->can_sample_s) ||
1726 (!stencil_sampler && tex->can_sample_z);
1727 }
1728
1729 static inline bool
1730 si_htile_enabled(struct si_texture *tex, unsigned level, unsigned zs_mask)
1731 {
1732 if (zs_mask == PIPE_MASK_S && tex->htile_stencil_disabled)
1733 return false;
1734
1735 return tex->htile_offset && level == 0;
1736 }
1737
1738 static inline bool
1739 vi_tc_compat_htile_enabled(struct si_texture *tex, unsigned level, unsigned zs_mask)
1740 {
1741 assert(!tex->tc_compatible_htile || tex->htile_offset);
1742 return tex->tc_compatible_htile && si_htile_enabled(tex, level, zs_mask);
1743 }
1744
1745 static inline unsigned si_get_ps_iter_samples(struct si_context *sctx)
1746 {
1747 if (sctx->ps_uses_fbfetch)
1748 return sctx->framebuffer.nr_color_samples;
1749
1750 return MIN2(sctx->ps_iter_samples, sctx->framebuffer.nr_color_samples);
1751 }
1752
1753 static inline unsigned si_get_total_colormask(struct si_context *sctx)
1754 {
1755 if (sctx->queued.named.rasterizer->rasterizer_discard)
1756 return 0;
1757
1758 struct si_shader_selector *ps = sctx->ps_shader.cso;
1759 if (!ps)
1760 return 0;
1761
1762 unsigned colormask = sctx->framebuffer.colorbuf_enabled_4bit &
1763 sctx->queued.named.blend->cb_target_mask;
1764
1765 if (!ps->info.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS])
1766 colormask &= ps->colors_written_4bit;
1767 else if (!ps->colors_written_4bit)
1768 colormask = 0; /* color0 writes all cbufs, but it's not written */
1769
1770 return colormask;
1771 }
1772
1773 #define UTIL_ALL_PRIM_LINE_MODES ((1 << PIPE_PRIM_LINES) | \
1774 (1 << PIPE_PRIM_LINE_LOOP) | \
1775 (1 << PIPE_PRIM_LINE_STRIP) | \
1776 (1 << PIPE_PRIM_LINES_ADJACENCY) | \
1777 (1 << PIPE_PRIM_LINE_STRIP_ADJACENCY))
1778
1779 static inline bool util_prim_is_lines(unsigned prim)
1780 {
1781 return ((1 << prim) & UTIL_ALL_PRIM_LINE_MODES) != 0;
1782 }
1783
1784 static inline bool util_prim_is_points_or_lines(unsigned prim)
1785 {
1786 return ((1 << prim) & (UTIL_ALL_PRIM_LINE_MODES |
1787 (1 << PIPE_PRIM_POINTS))) != 0;
1788 }
1789
1790 static inline bool util_rast_prim_is_triangles(unsigned prim)
1791 {
1792 return ((1 << prim) & ((1 << PIPE_PRIM_TRIANGLES) |
1793 (1 << PIPE_PRIM_TRIANGLE_STRIP) |
1794 (1 << PIPE_PRIM_TRIANGLE_FAN) |
1795 (1 << PIPE_PRIM_QUADS) |
1796 (1 << PIPE_PRIM_QUAD_STRIP) |
1797 (1 << PIPE_PRIM_POLYGON) |
1798 (1 << PIPE_PRIM_TRIANGLES_ADJACENCY) |
1799 (1 << PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY)));
1800 }
1801
1802 /**
1803 * Return true if there is enough memory in VRAM and GTT for the buffers
1804 * added so far.
1805 *
1806 * \param vram VRAM memory size not added to the buffer list yet
1807 * \param gtt GTT memory size not added to the buffer list yet
1808 */
1809 static inline bool
1810 radeon_cs_memory_below_limit(struct si_screen *screen,
1811 struct radeon_cmdbuf *cs,
1812 uint64_t vram, uint64_t gtt)
1813 {
1814 vram += cs->used_vram;
1815 gtt += cs->used_gart;
1816
1817 /* Anything that goes above the VRAM size should go to GTT. */
1818 if (vram > screen->info.vram_size)
1819 gtt += vram - screen->info.vram_size;
1820
1821 /* Now we just need to check if we have enough GTT. */
1822 return gtt < screen->info.gart_size * 0.7;
1823 }
1824
1825 /**
1826 * Add a buffer to the buffer list for the given command stream (CS).
1827 *
1828 * All buffers used by a CS must be added to the list. This tells the kernel
1829 * driver which buffers are used by GPU commands. Other buffers can
1830 * be swapped out (not accessible) during execution.
1831 *
1832 * The buffer list becomes empty after every context flush and must be
1833 * rebuilt.
1834 */
1835 static inline void radeon_add_to_buffer_list(struct si_context *sctx,
1836 struct radeon_cmdbuf *cs,
1837 struct si_resource *bo,
1838 enum radeon_bo_usage usage,
1839 enum radeon_bo_priority priority)
1840 {
1841 assert(usage);
1842 sctx->ws->cs_add_buffer(
1843 cs, bo->buf,
1844 (enum radeon_bo_usage)(usage | RADEON_USAGE_SYNCHRONIZED),
1845 bo->domains, priority);
1846 }
1847
1848 /**
1849 * Same as above, but also checks memory usage and flushes the context
1850 * accordingly.
1851 *
1852 * When this SHOULD NOT be used:
1853 *
1854 * - if si_context_add_resource_size has been called for the buffer
1855 * followed by *_need_cs_space for checking the memory usage
1856 *
1857 * - if si_need_dma_space has been called for the buffer
1858 *
1859 * - when emitting state packets and draw packets (because preceding packets
1860 * can't be re-emitted at that point)
1861 *
1862 * - if shader resource "enabled_mask" is not up-to-date or there is
1863 * a different constraint disallowing a context flush
1864 */
1865 static inline void
1866 radeon_add_to_gfx_buffer_list_check_mem(struct si_context *sctx,
1867 struct si_resource *bo,
1868 enum radeon_bo_usage usage,
1869 enum radeon_bo_priority priority,
1870 bool check_mem)
1871 {
1872 if (check_mem &&
1873 !radeon_cs_memory_below_limit(sctx->screen, sctx->gfx_cs,
1874 sctx->vram + bo->vram_usage,
1875 sctx->gtt + bo->gart_usage))
1876 si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
1877
1878 radeon_add_to_buffer_list(sctx, sctx->gfx_cs, bo, usage, priority);
1879 }
1880
1881 static inline bool si_compute_prim_discard_enabled(struct si_context *sctx)
1882 {
1883 return sctx->prim_discard_vertex_count_threshold != UINT_MAX;
1884 }
1885
1886 #define PRINT_ERR(fmt, args...) \
1887 fprintf(stderr, "EE %s:%d %s - " fmt, __FILE__, __LINE__, __func__, ##args)
1888
1889 #endif