projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac: add rbplus_allowed to ac_gpu_info
[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_SI_SCHED,
149 DBG_GISEL,
150 DBG_W32_GE,
151 DBG_W32_PS,
152 DBG_W32_CS,
153 DBG_W64_GE,
154 DBG_W64_PS,
155 DBG_W64_CS,
156
157 /* Shader compiler options (with no effect on the shader cache): */
158 DBG_CHECK_IR,
159 DBG_MONOLITHIC_SHADERS,
160 DBG_NO_OPT_VARIANT,
161
162 /* Information logging options: */
163 DBG_INFO,
164 DBG_TEX,
165 DBG_COMPUTE,
166 DBG_VM,
167
168 /* Driver options: */
169 DBG_FORCE_DMA,
170 DBG_NO_ASYNC_DMA,
171 DBG_NO_WC,
172 DBG_CHECK_VM,
173 DBG_RESERVE_VMID,
174 DBG_ZERO_VRAM,
175
176 /* 3D engine options: */
177 DBG_NO_GFX,
178 DBG_ALWAYS_PD,
179 DBG_PD,
180 DBG_NO_PD,
181 DBG_SWITCH_ON_EOP,
182 DBG_NO_OUT_OF_ORDER,
183 DBG_NO_DPBB,
184 DBG_NO_DFSM,
185 DBG_DPBB,
186 DBG_DFSM,
187 DBG_NO_HYPERZ,
188 DBG_NO_RB_PLUS,
189 DBG_NO_2D_TILING,
190 DBG_NO_TILING,
191 DBG_NO_DCC,
192 DBG_NO_DCC_CLEAR,
193 DBG_NO_DCC_FB,
194 DBG_NO_DCC_MSAA,
195 DBG_NO_FMASK,
196
197 /* Tests: */
198 DBG_TEST_DMA,
199 DBG_TEST_VMFAULT_CP,
200 DBG_TEST_VMFAULT_SDMA,
201 DBG_TEST_VMFAULT_SHADER,
202 DBG_TEST_DMA_PERF,
203 DBG_TEST_GDS,
204 DBG_TEST_GDS_MM,
205 DBG_TEST_GDS_OA_MM,
206 };
207
208 #define DBG_ALL_SHADERS (((1 << (DBG_CS + 1)) - 1))
209 #define DBG(name) (1ull << DBG_##name)
210
211 enum si_cache_policy {
212 L2_BYPASS,
213 L2_STREAM, /* same as SLC=1 */
214 L2_LRU, /* same as SLC=0 */
215 };
216
217 enum si_coherency {
218 SI_COHERENCY_NONE, /* no cache flushes needed */
219 SI_COHERENCY_SHADER,
220 SI_COHERENCY_CB_META,
221 SI_COHERENCY_CP,
222 };
223
224 struct si_compute;
225 struct si_shader_context;
226 struct hash_table;
227 struct u_suballocator;
228
229 /* Only 32-bit buffer allocations are supported, gallium doesn't support more
230 * at the moment.
231 */
232 struct si_resource {
233 struct threaded_resource b;
234
235 /* Winsys objects. */
236 struct pb_buffer *buf;
237 uint64_t gpu_address;
238 /* Memory usage if the buffer placement is optimal. */
239 uint64_t vram_usage;
240 uint64_t gart_usage;
241
242 /* Resource properties. */
243 uint64_t bo_size;
244 unsigned bo_alignment;
245 enum radeon_bo_domain domains;
246 enum radeon_bo_flag flags;
247 unsigned bind_history;
248 int max_forced_staging_uploads;
249
250 /* The buffer range which is initialized (with a write transfer,
251 * streamout, DMA, or as a random access target). The rest of
252 * the buffer is considered invalid and can be mapped unsynchronized.
253 *
254 * This allows unsychronized mapping of a buffer range which hasn't
255 * been used yet. It's for applications which forget to use
256 * the unsynchronized map flag and expect the driver to figure it out.
257 */
258 struct util_range valid_buffer_range;
259
260 /* For buffers only. This indicates that a write operation has been
261 * performed by TC L2, but the cache hasn't been flushed.
262 * Any hw block which doesn't use or bypasses TC L2 should check this
263 * flag and flush the cache before using the buffer.
264 *
265 * For example, TC L2 must be flushed if a buffer which has been
266 * modified by a shader store instruction is about to be used as
267 * an index buffer. The reason is that VGT DMA index fetching doesn't
268 * use TC L2.
269 */
270 bool TC_L2_dirty;
271
272 /* Whether this resource is referenced by bindless handles. */
273 bool texture_handle_allocated;
274 bool image_handle_allocated;
275
276 /* Whether the resource has been exported via resource_get_handle. */
277 unsigned external_usage; /* PIPE_HANDLE_USAGE_* */
278 };
279
280 struct si_transfer {
281 struct threaded_transfer b;
282 struct si_resource *staging;
283 unsigned offset;
284 };
285
286 struct si_texture {
287 struct si_resource buffer;
288
289 struct radeon_surf surface;
290 uint64_t size;
291 struct si_texture *flushed_depth_texture;
292
293 /* One texture allocation can contain these buffers:
294 * - image (pixel data)
295 * - FMASK buffer (MSAA compression)
296 * - CMASK buffer (MSAA compression and/or legacy fast color clear)
297 * - HTILE buffer (Z/S compression and fast Z/S clear)
298 * - DCC buffer (color compression and new fast color clear)
299 * - displayable DCC buffer (if the DCC buffer is not displayable)
300 * - DCC retile mapping buffer (if the DCC buffer is not displayable)
301 */
302 uint64_t fmask_offset;
303 uint64_t cmask_offset;
304 uint64_t cmask_base_address_reg;
305 struct si_resource *cmask_buffer;
306 uint64_t dcc_offset; /* 0 = disabled */
307 uint64_t display_dcc_offset;
308 uint64_t dcc_retile_map_offset;
309 unsigned cb_color_info; /* fast clear enable bit */
310 unsigned color_clear_value[2];
311 unsigned last_msaa_resolve_target_micro_mode;
312 unsigned num_level0_transfers;
313
314 /* Depth buffer compression and fast clear. */
315 uint64_t htile_offset;
316 float depth_clear_value;
317 uint16_t dirty_level_mask; /* each bit says if that mipmap is compressed */
318 uint16_t stencil_dirty_level_mask; /* each bit says if that mipmap is compressed */
319 enum pipe_format db_render_format:16;
320 uint8_t stencil_clear_value;
321 bool tc_compatible_htile:1;
322 bool htile_stencil_disabled:1;
323 bool depth_cleared:1; /* if it was cleared at least once */
324 bool stencil_cleared:1; /* if it was cleared at least once */
325 bool upgraded_depth:1; /* upgraded from unorm to Z32_FLOAT */
326 bool is_depth:1;
327 bool db_compatible:1;
328 bool can_sample_z:1;
329 bool can_sample_s:1;
330
331 /* We need to track DCC dirtiness, because st/dri usually calls
332 * flush_resource twice per frame (not a bug) and we don't wanna
333 * decompress DCC twice. Also, the dirty tracking must be done even
334 * if DCC isn't used, because it's required by the DCC usage analysis
335 * for a possible future enablement.
336 */
337 bool separate_dcc_dirty:1;
338 /* Statistics gathering for the DCC enablement heuristic. */
339 bool dcc_gather_statistics:1;
340 /* Counter that should be non-zero if the texture is bound to a
341 * framebuffer.
342 */
343 unsigned framebuffers_bound;
344 /* Whether the texture is a displayable back buffer and needs DCC
345 * decompression, which is expensive. Therefore, it's enabled only
346 * if statistics suggest that it will pay off and it's allocated
347 * separately. It can't be bound as a sampler by apps. Limited to
348 * target == 2D and last_level == 0. If enabled, dcc_offset contains
349 * the absolute GPUVM address, not the relative one.
350 */
351 struct si_resource *dcc_separate_buffer;
352 /* When DCC is temporarily disabled, the separate buffer is here. */
353 struct si_resource *last_dcc_separate_buffer;
354 /* Estimate of how much this color buffer is written to in units of
355 * full-screen draws: ps_invocations / (width * height)
356 * Shader kills, late Z, and blending with trivial discards make it
357 * inaccurate (we need to count CB updates, not PS invocations).
358 */
359 unsigned ps_draw_ratio;
360 /* The number of clears since the last DCC usage analysis. */
361 unsigned num_slow_clears;
362 };
363
364 struct si_surface {
365 struct pipe_surface base;
366
367 /* These can vary with block-compressed textures. */
368 uint16_t width0;
369 uint16_t height0;
370
371 bool color_initialized:1;
372 bool depth_initialized:1;
373
374 /* Misc. color flags. */
375 bool color_is_int8:1;
376 bool color_is_int10:1;
377 bool dcc_incompatible:1;
378
379 /* Color registers. */
380 unsigned cb_color_info;
381 unsigned cb_color_view;
382 unsigned cb_color_attrib;
383 unsigned cb_color_attrib2; /* GFX9 and later */
384 unsigned cb_color_attrib3; /* GFX10 and later */
385 unsigned cb_dcc_control; /* GFX8 and later */
386 unsigned spi_shader_col_format:8; /* no blending, no alpha-to-coverage. */
387 unsigned spi_shader_col_format_alpha:8; /* alpha-to-coverage */
388 unsigned spi_shader_col_format_blend:8; /* blending without alpha. */
389 unsigned spi_shader_col_format_blend_alpha:8; /* blending with alpha. */
390
391 /* DB registers. */
392 uint64_t db_depth_base; /* DB_Z_READ/WRITE_BASE */
393 uint64_t db_stencil_base;
394 uint64_t db_htile_data_base;
395 unsigned db_depth_info;
396 unsigned db_z_info;
397 unsigned db_z_info2; /* GFX9 only */
398 unsigned db_depth_view;
399 unsigned db_depth_size;
400 unsigned db_depth_slice;
401 unsigned db_stencil_info;
402 unsigned db_stencil_info2; /* GFX9 only */
403 unsigned db_htile_surface;
404 };
405
406 struct si_mmio_counter {
407 unsigned busy;
408 unsigned idle;
409 };
410
411 union si_mmio_counters {
412 struct {
413 /* For global GPU load including SDMA. */
414 struct si_mmio_counter gpu;
415
416 /* GRBM_STATUS */
417 struct si_mmio_counter spi;
418 struct si_mmio_counter gui;
419 struct si_mmio_counter ta;
420 struct si_mmio_counter gds;
421 struct si_mmio_counter vgt;
422 struct si_mmio_counter ia;
423 struct si_mmio_counter sx;
424 struct si_mmio_counter wd;
425 struct si_mmio_counter bci;
426 struct si_mmio_counter sc;
427 struct si_mmio_counter pa;
428 struct si_mmio_counter db;
429 struct si_mmio_counter cp;
430 struct si_mmio_counter cb;
431
432 /* SRBM_STATUS2 */
433 struct si_mmio_counter sdma;
434
435 /* CP_STAT */
436 struct si_mmio_counter pfp;
437 struct si_mmio_counter meq;
438 struct si_mmio_counter me;
439 struct si_mmio_counter surf_sync;
440 struct si_mmio_counter cp_dma;
441 struct si_mmio_counter scratch_ram;
442 } named;
443 unsigned array[0];
444 };
445
446 struct si_memory_object {
447 struct pipe_memory_object b;
448 struct pb_buffer *buf;
449 uint32_t stride;
450 };
451
452 /* Saved CS data for debugging features. */
453 struct radeon_saved_cs {
454 uint32_t *ib;
455 unsigned num_dw;
456
457 struct radeon_bo_list_item *bo_list;
458 unsigned bo_count;
459 };
460
461 struct si_screen {
462 struct pipe_screen b;
463 struct radeon_winsys *ws;
464 struct disk_cache *disk_shader_cache;
465
466 struct radeon_info info;
467 uint64_t debug_flags;
468 char renderer_string[183];
469
470 void (*make_texture_descriptor)(
471 struct si_screen *screen,
472 struct si_texture *tex,
473 bool sampler,
474 enum pipe_texture_target target,
475 enum pipe_format pipe_format,
476 const unsigned char state_swizzle[4],
477 unsigned first_level, unsigned last_level,
478 unsigned first_layer, unsigned last_layer,
479 unsigned width, unsigned height, unsigned depth,
480 uint32_t *state,
481 uint32_t *fmask_state);
482
483 unsigned pa_sc_raster_config;
484 unsigned pa_sc_raster_config_1;
485 unsigned se_tile_repeat;
486 unsigned gs_table_depth;
487 unsigned tess_offchip_block_dw_size;
488 unsigned tess_offchip_ring_size;
489 unsigned tess_factor_ring_size;
490 unsigned vgt_hs_offchip_param;
491 unsigned eqaa_force_coverage_samples;
492 unsigned eqaa_force_z_samples;
493 unsigned eqaa_force_color_samples;
494 bool has_draw_indirect_multi;
495 bool has_out_of_order_rast;
496 bool assume_no_z_fights;
497 bool commutative_blend_add;
498 bool has_msaa_sample_loc_bug;
499 bool has_ls_vgpr_init_bug;
500 bool dpbb_allowed;
501 bool dfsm_allowed;
502 bool llvm_has_working_vgpr_indexing;
503 bool use_ngg;
504 bool use_ngg_streamout;
505
506 struct {
507 #define OPT_BOOL(name, dflt, description) bool name:1;
508 #include "si_debug_options.h"
509 } options;
510
511 /* Whether shaders are monolithic (1-part) or separate (3-part). */
512 bool use_monolithic_shaders;
513 bool record_llvm_ir;
514 bool dcc_msaa_allowed;
515
516 struct slab_parent_pool pool_transfers;
517
518 /* Texture filter settings. */
519 int force_aniso; /* -1 = disabled */
520
521 /* Auxiliary context. Mainly used to initialize resources.
522 * It must be locked prior to using and flushed before unlocking. */
523 struct pipe_context *aux_context;
524 mtx_t aux_context_lock;
525
526 /* This must be in the screen, because UE4 uses one context for
527 * compilation and another one for rendering.
528 */
529 unsigned num_compilations;
530 /* Along with ST_DEBUG=precompile, this should show if applications
531 * are loading shaders on demand. This is a monotonic counter.
532 */
533 unsigned num_shaders_created;
534 unsigned num_shader_cache_hits;
535
536 /* GPU load thread. */
537 mtx_t gpu_load_mutex;
538 thrd_t gpu_load_thread;
539 union si_mmio_counters mmio_counters;
540 volatile unsigned gpu_load_stop_thread; /* bool */
541
542 /* Performance counters. */
543 struct si_perfcounters *perfcounters;
544
545 /* If pipe_screen wants to recompute and re-emit the framebuffer,
546 * sampler, and image states of all contexts, it should atomically
547 * increment this.
548 *
549 * Each context will compare this with its own last known value of
550 * the counter before drawing and re-emit the states accordingly.
551 */
552 unsigned dirty_tex_counter;
553 unsigned dirty_buf_counter;
554
555 /* Atomically increment this counter when an existing texture's
556 * metadata is enabled or disabled in a way that requires changing
557 * contexts' compressed texture binding masks.
558 */
559 unsigned compressed_colortex_counter;
560
561 struct {
562 /* Context flags to set so that all writes from earlier jobs
563 * in the CP are seen by L2 clients.
564 */
565 unsigned cp_to_L2;
566
567 /* Context flags to set so that all writes from earlier jobs
568 * that end in L2 are seen by CP.
569 */
570 unsigned L2_to_cp;
571 } barrier_flags;
572
573 mtx_t shader_parts_mutex;
574 struct si_shader_part *vs_prologs;
575 struct si_shader_part *tcs_epilogs;
576 struct si_shader_part *gs_prologs;
577 struct si_shader_part *ps_prologs;
578 struct si_shader_part *ps_epilogs;
579
580 /* Shader cache in memory.
581 *
582 * Design & limitations:
583 * - The shader cache is per screen (= per process), never saved to
584 * disk, and skips redundant shader compilations from TGSI to bytecode.
585 * - It can only be used with one-variant-per-shader support, in which
586 * case only the main (typically middle) part of shaders is cached.
587 * - Only VS, TCS, TES, PS are cached, out of which only the hw VS
588 * variants of VS and TES are cached, so LS and ES aren't.
589 * - GS and CS aren't cached, but it's certainly possible to cache
590 * those as well.
591 */
592 mtx_t shader_cache_mutex;
593 struct hash_table *shader_cache;
594
595 /* Shader compiler queue for multithreaded compilation. */
596 struct util_queue shader_compiler_queue;
597 /* Use at most 3 normal compiler threads on quadcore and better.
598 * Hyperthreaded CPUs report the number of threads, but we want
599 * the number of cores. We only need this many threads for shader-db. */
600 struct ac_llvm_compiler compiler[24]; /* used by the queue only */
601
602 struct util_queue shader_compiler_queue_low_priority;
603 /* Use at most 2 low priority threads on quadcore and better.
604 * We want to minimize the impact on multithreaded Mesa. */
605 struct ac_llvm_compiler compiler_lowp[10];
606
607 unsigned compute_wave_size;
608 unsigned ps_wave_size;
609 unsigned ge_wave_size;
610 };
611
612 struct si_blend_color {
613 struct pipe_blend_color state;
614 bool any_nonzeros;
615 };
616
617 struct si_sampler_view {
618 struct pipe_sampler_view base;
619 /* [0..7] = image descriptor
620 * [4..7] = buffer descriptor */
621 uint32_t state[8];
622 uint32_t fmask_state[8];
623 const struct legacy_surf_level *base_level_info;
624 ubyte base_level;
625 ubyte block_width;
626 bool is_stencil_sampler;
627 bool is_integer;
628 bool dcc_incompatible;
629 };
630
631 #define SI_SAMPLER_STATE_MAGIC 0x34f1c35a
632
633 struct si_sampler_state {
634 #ifndef NDEBUG
635 unsigned magic;
636 #endif
637 uint32_t val[4];
638 uint32_t integer_val[4];
639 uint32_t upgraded_depth_val[4];
640 };
641
642 struct si_cs_shader_state {
643 struct si_compute *program;
644 struct si_compute *emitted_program;
645 unsigned offset;
646 bool initialized;
647 bool uses_scratch;
648 };
649
650 struct si_samplers {
651 struct pipe_sampler_view *views[SI_NUM_SAMPLERS];
652 struct si_sampler_state *sampler_states[SI_NUM_SAMPLERS];
653
654 /* The i-th bit is set if that element is enabled (non-NULL resource). */
655 unsigned enabled_mask;
656 uint32_t needs_depth_decompress_mask;
657 uint32_t needs_color_decompress_mask;
658 };
659
660 struct si_images {
661 struct pipe_image_view views[SI_NUM_IMAGES];
662 uint32_t needs_color_decompress_mask;
663 unsigned enabled_mask;
664 };
665
666 struct si_framebuffer {
667 struct pipe_framebuffer_state state;
668 unsigned colorbuf_enabled_4bit;
669 unsigned spi_shader_col_format;
670 unsigned spi_shader_col_format_alpha;
671 unsigned spi_shader_col_format_blend;
672 unsigned spi_shader_col_format_blend_alpha;
673 ubyte nr_samples:5; /* at most 16xAA */
674 ubyte log_samples:3; /* at most 4 = 16xAA */
675 ubyte nr_color_samples; /* at most 8xAA */
676 ubyte compressed_cb_mask;
677 ubyte uncompressed_cb_mask;
678 ubyte color_is_int8;
679 ubyte color_is_int10;
680 ubyte dirty_cbufs;
681 ubyte dcc_overwrite_combiner_watermark;
682 ubyte min_bytes_per_pixel;
683 bool dirty_zsbuf;
684 bool any_dst_linear;
685 bool CB_has_shader_readable_metadata;
686 bool DB_has_shader_readable_metadata;
687 bool all_DCC_pipe_aligned;
688 };
689
690 enum si_quant_mode {
691 /* This is the list we want to support. */
692 SI_QUANT_MODE_16_8_FIXED_POINT_1_256TH,
693 SI_QUANT_MODE_14_10_FIXED_POINT_1_1024TH,
694 SI_QUANT_MODE_12_12_FIXED_POINT_1_4096TH,
695 };
696
697 struct si_signed_scissor {
698 int minx;
699 int miny;
700 int maxx;
701 int maxy;
702 enum si_quant_mode quant_mode;
703 };
704
705 struct si_viewports {
706 struct pipe_viewport_state states[SI_MAX_VIEWPORTS];
707 struct si_signed_scissor as_scissor[SI_MAX_VIEWPORTS];
708 bool y_inverted;
709 };
710
711 struct si_clip_state {
712 struct pipe_clip_state state;
713 bool any_nonzeros;
714 };
715
716 struct si_streamout_target {
717 struct pipe_stream_output_target b;
718
719 /* The buffer where BUFFER_FILLED_SIZE is stored. */
720 struct si_resource *buf_filled_size;
721 unsigned buf_filled_size_offset;
722 bool buf_filled_size_valid;
723
724 unsigned stride_in_dw;
725 };
726
727 struct si_streamout {
728 bool begin_emitted;
729
730 unsigned enabled_mask;
731 unsigned num_targets;
732 struct si_streamout_target *targets[PIPE_MAX_SO_BUFFERS];
733
734 unsigned append_bitmask;
735 bool suspended;
736
737 /* External state which comes from the vertex shader,
738 * it must be set explicitly when binding a shader. */
739 uint16_t *stride_in_dw;
740 unsigned enabled_stream_buffers_mask; /* stream0 buffers0-3 in 4 LSB */
741
742 /* The state of VGT_STRMOUT_BUFFER_(CONFIG|EN). */
743 unsigned hw_enabled_mask;
744
745 /* The state of VGT_STRMOUT_(CONFIG|EN). */
746 bool streamout_enabled;
747 bool prims_gen_query_enabled;
748 int num_prims_gen_queries;
749 };
750
751 /* A shader state consists of the shader selector, which is a constant state
752 * object shared by multiple contexts and shouldn't be modified, and
753 * the current shader variant selected for this context.
754 */
755 struct si_shader_ctx_state {
756 struct si_shader_selector *cso;
757 struct si_shader *current;
758 };
759
760 #define SI_NUM_VGT_PARAM_KEY_BITS 12
761 #define SI_NUM_VGT_PARAM_STATES (1 << SI_NUM_VGT_PARAM_KEY_BITS)
762
763 /* The IA_MULTI_VGT_PARAM key used to index the table of precomputed values.
764 * Some fields are set by state-change calls, most are set by draw_vbo.
765 */
766 union si_vgt_param_key {
767 struct {
768 #ifdef PIPE_ARCH_LITTLE_ENDIAN
769 unsigned prim:4;
770 unsigned uses_instancing:1;
771 unsigned multi_instances_smaller_than_primgroup:1;
772 unsigned primitive_restart:1;
773 unsigned count_from_stream_output:1;
774 unsigned line_stipple_enabled:1;
775 unsigned uses_tess:1;
776 unsigned tess_uses_prim_id:1;
777 unsigned uses_gs:1;
778 unsigned _pad:32 - SI_NUM_VGT_PARAM_KEY_BITS;
779 #else /* PIPE_ARCH_BIG_ENDIAN */
780 unsigned _pad:32 - SI_NUM_VGT_PARAM_KEY_BITS;
781 unsigned uses_gs:1;
782 unsigned tess_uses_prim_id:1;
783 unsigned uses_tess:1;
784 unsigned line_stipple_enabled:1;
785 unsigned count_from_stream_output:1;
786 unsigned primitive_restart:1;
787 unsigned multi_instances_smaller_than_primgroup:1;
788 unsigned uses_instancing:1;
789 unsigned prim:4;
790 #endif
791 } u;
792 uint32_t index;
793 };
794
795 #define SI_NUM_VGT_STAGES_KEY_BITS 4
796 #define SI_NUM_VGT_STAGES_STATES (1 << SI_NUM_VGT_STAGES_KEY_BITS)
797
798 /* The VGT_SHADER_STAGES key used to index the table of precomputed values.
799 * Some fields are set by state-change calls, most are set by draw_vbo.
800 */
801 union si_vgt_stages_key {
802 struct {
803 #ifdef PIPE_ARCH_LITTLE_ENDIAN
804 unsigned tess:1;
805 unsigned gs:1;
806 unsigned ngg:1; /* gfx10+ */
807 unsigned streamout:1; /* only used with NGG */
808 unsigned _pad:32 - SI_NUM_VGT_STAGES_KEY_BITS;
809 #else /* PIPE_ARCH_BIG_ENDIAN */
810 unsigned _pad:32 - SI_NUM_VGT_STAGES_KEY_BITS;
811 unsigned streamout:1;
812 unsigned ngg:1;
813 unsigned gs:1;
814 unsigned tess:1;
815 #endif
816 } u;
817 uint32_t index;
818 };
819
820 struct si_texture_handle
821 {
822 unsigned desc_slot;
823 bool desc_dirty;
824 struct pipe_sampler_view *view;
825 struct si_sampler_state sstate;
826 };
827
828 struct si_image_handle
829 {
830 unsigned desc_slot;
831 bool desc_dirty;
832 struct pipe_image_view view;
833 };
834
835 struct si_saved_cs {
836 struct pipe_reference reference;
837 struct si_context *ctx;
838 struct radeon_saved_cs gfx;
839 struct radeon_saved_cs compute;
840 struct si_resource *trace_buf;
841 unsigned trace_id;
842
843 unsigned gfx_last_dw;
844 unsigned compute_last_dw;
845 bool flushed;
846 int64_t time_flush;
847 };
848
849 struct si_sdma_upload {
850 struct si_resource *dst;
851 struct si_resource *src;
852 unsigned src_offset;
853 unsigned dst_offset;
854 unsigned size;
855 };
856
857 struct si_context {
858 struct pipe_context b; /* base class */
859
860 enum radeon_family family;
861 enum chip_class chip_class;
862
863 struct radeon_winsys *ws;
864 struct radeon_winsys_ctx *ctx;
865 struct radeon_cmdbuf *gfx_cs; /* compute IB if graphics is disabled */
866 struct radeon_cmdbuf *dma_cs;
867 struct pipe_fence_handle *last_gfx_fence;
868 struct pipe_fence_handle *last_sdma_fence;
869 struct si_resource *eop_bug_scratch;
870 struct u_upload_mgr *cached_gtt_allocator;
871 struct threaded_context *tc;
872 struct u_suballocator *allocator_zeroed_memory;
873 struct slab_child_pool pool_transfers;
874 struct slab_child_pool pool_transfers_unsync; /* for threaded_context */
875 struct pipe_device_reset_callback device_reset_callback;
876 struct u_log_context *log;
877 void *query_result_shader;
878 void *sh_query_result_shader;
879
880 void (*emit_cache_flush)(struct si_context *ctx);
881
882 struct blitter_context *blitter;
883 void *noop_blend;
884 void *noop_dsa;
885 void *discard_rasterizer_state;
886 void *custom_dsa_flush;
887 void *custom_blend_resolve;
888 void *custom_blend_fmask_decompress;
889 void *custom_blend_eliminate_fastclear;
890 void *custom_blend_dcc_decompress;
891 void *vs_blit_pos;
892 void *vs_blit_pos_layered;
893 void *vs_blit_color;
894 void *vs_blit_color_layered;
895 void *vs_blit_texcoord;
896 void *cs_clear_buffer;
897 void *cs_copy_buffer;
898 void *cs_copy_image;
899 void *cs_copy_image_1d_array;
900 void *cs_clear_render_target;
901 void *cs_clear_render_target_1d_array;
902 void *cs_dcc_retile;
903 struct si_screen *screen;
904 struct pipe_debug_callback debug;
905 struct ac_llvm_compiler compiler; /* only non-threaded compilation */
906 struct si_shader_ctx_state fixed_func_tcs_shader;
907 /* Offset 0: EOP flush number; Offset 4: GDS prim restart counter */
908 struct si_resource *wait_mem_scratch;
909 unsigned wait_mem_number;
910 uint16_t prefetch_L2_mask;
911
912 bool has_graphics;
913 bool gfx_flush_in_progress:1;
914 bool gfx_last_ib_is_busy:1;
915 bool compute_is_busy:1;
916
917 unsigned num_gfx_cs_flushes;
918 unsigned initial_gfx_cs_size;
919 unsigned last_dirty_tex_counter;
920 unsigned last_dirty_buf_counter;
921 unsigned last_compressed_colortex_counter;
922 unsigned last_num_draw_calls;
923 unsigned flags; /* flush flags */
924 /* Current unaccounted memory usage. */
925 uint64_t vram;
926 uint64_t gtt;
927
928 /* Compute-based primitive discard. */
929 unsigned prim_discard_vertex_count_threshold;
930 struct pb_buffer *gds;
931 struct pb_buffer *gds_oa;
932 struct radeon_cmdbuf *prim_discard_compute_cs;
933 unsigned compute_gds_offset;
934 struct si_shader *compute_ib_last_shader;
935 uint32_t compute_rewind_va;
936 unsigned compute_num_prims_in_batch;
937 bool preserve_prim_restart_gds_at_flush;
938 /* index_ring is divided into 2 halves for doublebuffering. */
939 struct si_resource *index_ring;
940 unsigned index_ring_base; /* offset of a per-IB portion */
941 unsigned index_ring_offset; /* offset within a per-IB portion */
942 unsigned index_ring_size_per_ib; /* max available size per IB */
943 bool prim_discard_compute_ib_initialized;
944 /* For tracking the last execution barrier - it can be either
945 * a WRITE_DATA packet or a fence. */
946 uint32_t *last_pkt3_write_data;
947 struct si_resource *barrier_buf;
948 unsigned barrier_buf_offset;
949 struct pipe_fence_handle *last_ib_barrier_fence;
950 struct si_resource *last_ib_barrier_buf;
951 unsigned last_ib_barrier_buf_offset;
952
953 /* Atoms (direct states). */
954 union si_state_atoms atoms;
955 unsigned dirty_atoms; /* mask */
956 /* PM4 states (precomputed immutable states) */
957 unsigned dirty_states;
958 union si_state queued;
959 union si_state emitted;
960
961 /* Atom declarations. */
962 struct si_framebuffer framebuffer;
963 unsigned sample_locs_num_samples;
964 uint16_t sample_mask;
965 unsigned last_cb_target_mask;
966 struct si_blend_color blend_color;
967 struct si_clip_state clip_state;
968 struct si_shader_data shader_pointers;
969 struct si_stencil_ref stencil_ref;
970 struct pipe_scissor_state scissors[SI_MAX_VIEWPORTS];
971 struct si_streamout streamout;
972 struct si_viewports viewports;
973 unsigned num_window_rectangles;
974 bool window_rectangles_include;
975 struct pipe_scissor_state window_rectangles[4];
976
977 /* Precomputed states. */
978 struct si_pm4_state *init_config;
979 struct si_pm4_state *init_config_gs_rings;
980 bool init_config_has_vgt_flush;
981 struct si_pm4_state *vgt_shader_config[SI_NUM_VGT_STAGES_STATES];
982
983 /* shaders */
984 struct si_shader_ctx_state ps_shader;
985 struct si_shader_ctx_state gs_shader;
986 struct si_shader_ctx_state vs_shader;
987 struct si_shader_ctx_state tcs_shader;
988 struct si_shader_ctx_state tes_shader;
989 struct si_shader_ctx_state cs_prim_discard_state;
990 struct si_cs_shader_state cs_shader_state;
991
992 /* shader information */
993 struct si_vertex_elements *vertex_elements;
994 unsigned sprite_coord_enable;
995 unsigned cs_max_waves_per_sh;
996 bool flatshade;
997 bool do_update_shaders;
998
999 /* vertex buffer descriptors */
1000 uint32_t *vb_descriptors_gpu_list;
1001 struct si_resource *vb_descriptors_buffer;
1002 unsigned vb_descriptors_offset;
1003
1004 /* shader descriptors */
1005 struct si_descriptors descriptors[SI_NUM_DESCS];
1006 unsigned descriptors_dirty;
1007 unsigned shader_pointers_dirty;
1008 unsigned shader_needs_decompress_mask;
1009 struct si_buffer_resources rw_buffers;
1010 struct si_buffer_resources const_and_shader_buffers[SI_NUM_SHADERS];
1011 struct si_samplers samplers[SI_NUM_SHADERS];
1012 struct si_images images[SI_NUM_SHADERS];
1013 bool bo_list_add_all_resident_resources;
1014 bool bo_list_add_all_gfx_resources;
1015 bool bo_list_add_all_compute_resources;
1016
1017 /* other shader resources */
1018 struct pipe_constant_buffer null_const_buf; /* used for set_constant_buffer(NULL) on GFX7 */
1019 struct pipe_resource *esgs_ring;
1020 struct pipe_resource *gsvs_ring;
1021 struct pipe_resource *tess_rings;
1022 union pipe_color_union *border_color_table; /* in CPU memory, any endian */
1023 struct si_resource *border_color_buffer;
1024 union pipe_color_union *border_color_map; /* in VRAM (slow access), little endian */
1025 unsigned border_color_count;
1026 unsigned num_vs_blit_sgprs;
1027 uint32_t vs_blit_sh_data[SI_VS_BLIT_SGPRS_POS_TEXCOORD];
1028 uint32_t cs_user_data[4];
1029
1030 /* Vertex and index buffers. */
1031 bool vertex_buffers_dirty;
1032 bool vertex_buffer_pointer_dirty;
1033 struct pipe_vertex_buffer vertex_buffer[SI_NUM_VERTEX_BUFFERS];
1034 uint16_t vertex_buffer_unaligned; /* bitmask of not dword-aligned buffers */
1035
1036 /* MSAA config state. */
1037 int ps_iter_samples;
1038 bool ps_uses_fbfetch;
1039 bool smoothing_enabled;
1040
1041 /* DB render state. */
1042 unsigned ps_db_shader_control;
1043 unsigned dbcb_copy_sample;
1044 bool dbcb_depth_copy_enabled:1;
1045 bool dbcb_stencil_copy_enabled:1;
1046 bool db_flush_depth_inplace:1;
1047 bool db_flush_stencil_inplace:1;
1048 bool db_depth_clear:1;
1049 bool db_depth_disable_expclear:1;
1050 bool db_stencil_clear:1;
1051 bool db_stencil_disable_expclear:1;
1052 bool occlusion_queries_disabled:1;
1053 bool generate_mipmap_for_depth:1;
1054
1055 /* Emitted draw state. */
1056 bool gs_tri_strip_adj_fix:1;
1057 bool ls_vgpr_fix:1;
1058 bool prim_discard_cs_instancing:1;
1059 bool ngg:1;
1060 int last_index_size;
1061 int last_base_vertex;
1062 int last_start_instance;
1063 int last_instance_count;
1064 int last_drawid;
1065 int last_sh_base_reg;
1066 int last_primitive_restart_en;
1067 int last_restart_index;
1068 int last_prim;
1069 int last_multi_vgt_param;
1070 int last_rast_prim;
1071 int last_flatshade_first;
1072 int last_binning_enabled;
1073 unsigned last_sc_line_stipple;
1074 unsigned current_vs_state;
1075 unsigned last_vs_state;
1076 enum pipe_prim_type current_rast_prim; /* primitive type after TES, GS */
1077
1078 /* Scratch buffer */
1079 struct si_resource *scratch_buffer;
1080 unsigned scratch_waves;
1081 unsigned spi_tmpring_size;
1082
1083 struct si_resource *compute_scratch_buffer;
1084
1085 /* Emitted derived tessellation state. */
1086 /* Local shader (VS), or HS if LS-HS are merged. */
1087 struct si_shader *last_ls;
1088 struct si_shader_selector *last_tcs;
1089 int last_num_tcs_input_cp;
1090 int last_tes_sh_base;
1091 bool last_tess_uses_primid;
1092 unsigned last_num_patches;
1093 int last_ls_hs_config;
1094
1095 /* Debug state. */
1096 bool is_debug;
1097 struct si_saved_cs *current_saved_cs;
1098 uint64_t dmesg_timestamp;
1099 unsigned apitrace_call_number;
1100
1101 /* Other state */
1102 bool need_check_render_feedback;
1103 bool decompression_enabled;
1104 bool dpbb_force_off;
1105 bool vs_writes_viewport_index;
1106 bool vs_disables_clipping_viewport;
1107
1108 /* Precomputed IA_MULTI_VGT_PARAM */
1109 union si_vgt_param_key ia_multi_vgt_param_key;
1110 unsigned ia_multi_vgt_param[SI_NUM_VGT_PARAM_STATES];
1111
1112 /* Bindless descriptors. */
1113 struct si_descriptors bindless_descriptors;
1114 struct util_idalloc bindless_used_slots;
1115 unsigned num_bindless_descriptors;
1116 bool bindless_descriptors_dirty;
1117 bool graphics_bindless_pointer_dirty;
1118 bool compute_bindless_pointer_dirty;
1119
1120 /* Allocated bindless handles */
1121 struct hash_table *tex_handles;
1122 struct hash_table *img_handles;
1123
1124 /* Resident bindless handles */
1125 struct util_dynarray resident_tex_handles;
1126 struct util_dynarray resident_img_handles;
1127
1128 /* Resident bindless handles which need decompression */
1129 struct util_dynarray resident_tex_needs_color_decompress;
1130 struct util_dynarray resident_img_needs_color_decompress;
1131 struct util_dynarray resident_tex_needs_depth_decompress;
1132
1133 /* Bindless state */
1134 bool uses_bindless_samplers;
1135 bool uses_bindless_images;
1136
1137 /* MSAA sample locations.
1138 * The first index is the sample index.
1139 * The second index is the coordinate: X, Y. */
1140 struct {
1141 float x1[1][2];
1142 float x2[2][2];
1143 float x4[4][2];
1144 float x8[8][2];
1145 float x16[16][2];
1146 } sample_positions;
1147 struct pipe_resource *sample_pos_buffer;
1148
1149 /* Misc stats. */
1150 unsigned num_draw_calls;
1151 unsigned num_decompress_calls;
1152 unsigned num_mrt_draw_calls;
1153 unsigned num_prim_restart_calls;
1154 unsigned num_spill_draw_calls;
1155 unsigned num_compute_calls;
1156 unsigned num_spill_compute_calls;
1157 unsigned num_dma_calls;
1158 unsigned num_cp_dma_calls;
1159 unsigned num_vs_flushes;
1160 unsigned num_ps_flushes;
1161 unsigned num_cs_flushes;
1162 unsigned num_cb_cache_flushes;
1163 unsigned num_db_cache_flushes;
1164 unsigned num_L2_invalidates;
1165 unsigned num_L2_writebacks;
1166 unsigned num_resident_handles;
1167 uint64_t num_alloc_tex_transfer_bytes;
1168 unsigned last_tex_ps_draw_ratio; /* for query */
1169 unsigned compute_num_verts_accepted;
1170 unsigned compute_num_verts_rejected;
1171 unsigned compute_num_verts_ineligible; /* due to low vertex count */
1172 unsigned context_roll;
1173
1174 /* Queries. */
1175 /* Maintain the list of active queries for pausing between IBs. */
1176 int num_occlusion_queries;
1177 int num_perfect_occlusion_queries;
1178 int num_pipeline_stat_queries;
1179 struct list_head active_queries;
1180 unsigned num_cs_dw_queries_suspend;
1181
1182 /* Render condition. */
1183 struct pipe_query *render_cond;
1184 unsigned render_cond_mode;
1185 bool render_cond_invert;
1186 bool render_cond_force_off; /* for u_blitter */
1187
1188 /* For uploading data via GTT and copy to VRAM on context flush via SDMA. */
1189 bool sdma_uploads_in_progress;
1190 struct si_sdma_upload *sdma_uploads;
1191 unsigned num_sdma_uploads;
1192 unsigned max_sdma_uploads;
1193
1194 /* Shader-based queries. */
1195 struct list_head shader_query_buffers;
1196 unsigned num_active_shader_queries;
1197
1198 /* Statistics gathering for the DCC enablement heuristic. It can't be
1199 * in si_texture because si_texture can be shared by multiple
1200 * contexts. This is for back buffers only. We shouldn't get too many
1201 * of those.
1202 *
1203 * X11 DRI3 rotates among a finite set of back buffers. They should
1204 * all fit in this array. If they don't, separate DCC might never be
1205 * enabled by DCC stat gathering.
1206 */
1207 struct {
1208 struct si_texture *tex;
1209 /* Query queue: 0 = usually active, 1 = waiting, 2 = readback. */
1210 struct pipe_query *ps_stats[3];
1211 /* If all slots are used and another slot is needed,
1212 * the least recently used slot is evicted based on this. */
1213 int64_t last_use_timestamp;
1214 bool query_active;
1215 } dcc_stats[5];
1216
1217 /* Copy one resource to another using async DMA. */
1218 void (*dma_copy)(struct pipe_context *ctx,
1219 struct pipe_resource *dst,
1220 unsigned dst_level,
1221 unsigned dst_x, unsigned dst_y, unsigned dst_z,
1222 struct pipe_resource *src,
1223 unsigned src_level,
1224 const struct pipe_box *src_box);
1225
1226 struct si_tracked_regs tracked_regs;
1227 };
1228
1229 /* cik_sdma.c */
1230 void cik_init_sdma_functions(struct si_context *sctx);
1231
1232 /* si_blit.c */
1233 enum si_blitter_op /* bitmask */
1234 {
1235 SI_SAVE_TEXTURES = 1,
1236 SI_SAVE_FRAMEBUFFER = 2,
1237 SI_SAVE_FRAGMENT_STATE = 4,
1238 SI_DISABLE_RENDER_COND = 8,
1239 };
1240
1241 void si_blitter_begin(struct si_context *sctx, enum si_blitter_op op);
1242 void si_blitter_end(struct si_context *sctx);
1243 void si_init_blit_functions(struct si_context *sctx);
1244 void si_decompress_textures(struct si_context *sctx, unsigned shader_mask);
1245 void si_resource_copy_region(struct pipe_context *ctx,
1246 struct pipe_resource *dst,
1247 unsigned dst_level,
1248 unsigned dstx, unsigned dsty, unsigned dstz,
1249 struct pipe_resource *src,
1250 unsigned src_level,
1251 const struct pipe_box *src_box);
1252 void si_decompress_dcc(struct si_context *sctx, struct si_texture *tex);
1253
1254 /* si_buffer.c */
1255 bool si_rings_is_buffer_referenced(struct si_context *sctx,
1256 struct pb_buffer *buf,
1257 enum radeon_bo_usage usage);
1258 void *si_buffer_map_sync_with_rings(struct si_context *sctx,
1259 struct si_resource *resource,
1260 unsigned usage);
1261 void si_init_resource_fields(struct si_screen *sscreen,
1262 struct si_resource *res,
1263 uint64_t size, unsigned alignment);
1264 bool si_alloc_resource(struct si_screen *sscreen,
1265 struct si_resource *res);
1266 struct pipe_resource *pipe_aligned_buffer_create(struct pipe_screen *screen,
1267 unsigned flags, unsigned usage,
1268 unsigned size, unsigned alignment);
1269 struct si_resource *si_aligned_buffer_create(struct pipe_screen *screen,
1270 unsigned flags, unsigned usage,
1271 unsigned size, unsigned alignment);
1272 void si_replace_buffer_storage(struct pipe_context *ctx,
1273 struct pipe_resource *dst,
1274 struct pipe_resource *src);
1275 void si_init_screen_buffer_functions(struct si_screen *sscreen);
1276 void si_init_buffer_functions(struct si_context *sctx);
1277
1278 /* si_clear.c */
1279 enum pipe_format si_simplify_cb_format(enum pipe_format format);
1280 bool vi_alpha_is_on_msb(struct si_screen *sscreen, enum pipe_format format);
1281 bool vi_dcc_clear_level(struct si_context *sctx,
1282 struct si_texture *tex,
1283 unsigned level, unsigned clear_value);
1284 void si_init_clear_functions(struct si_context *sctx);
1285
1286 /* si_compute_blit.c */
1287 unsigned si_get_flush_flags(struct si_context *sctx, enum si_coherency coher,
1288 enum si_cache_policy cache_policy);
1289 void si_clear_buffer(struct si_context *sctx, struct pipe_resource *dst,
1290 uint64_t offset, uint64_t size, uint32_t *clear_value,
1291 uint32_t clear_value_size, enum si_coherency coher,
1292 bool force_cpdma);
1293 void si_copy_buffer(struct si_context *sctx,
1294 struct pipe_resource *dst, struct pipe_resource *src,
1295 uint64_t dst_offset, uint64_t src_offset, unsigned size);
1296 void si_compute_copy_image(struct si_context *sctx,
1297 struct pipe_resource *dst,
1298 unsigned dst_level,
1299 struct pipe_resource *src,
1300 unsigned src_level,
1301 unsigned dstx, unsigned dsty, unsigned dstz,
1302 const struct pipe_box *src_box);
1303 void si_compute_clear_render_target(struct pipe_context *ctx,
1304 struct pipe_surface *dstsurf,
1305 const union pipe_color_union *color,
1306 unsigned dstx, unsigned dsty,
1307 unsigned width, unsigned height,
1308 bool render_condition_enabled);
1309 void si_retile_dcc(struct si_context *sctx, struct si_texture *tex);
1310 void si_init_compute_blit_functions(struct si_context *sctx);
1311
1312 /* si_cp_dma.c */
1313 #define SI_CPDMA_SKIP_CHECK_CS_SPACE (1 << 0) /* don't call need_cs_space */
1314 #define SI_CPDMA_SKIP_SYNC_AFTER (1 << 1) /* don't wait for DMA after the copy */
1315 #define SI_CPDMA_SKIP_SYNC_BEFORE (1 << 2) /* don't wait for DMA before the copy (RAW hazards) */
1316 #define SI_CPDMA_SKIP_GFX_SYNC (1 << 3) /* don't flush caches and don't wait for PS/CS */
1317 #define SI_CPDMA_SKIP_BO_LIST_UPDATE (1 << 4) /* don't update the BO list */
1318 #define SI_CPDMA_SKIP_ALL (SI_CPDMA_SKIP_CHECK_CS_SPACE | \
1319 SI_CPDMA_SKIP_SYNC_AFTER | \
1320 SI_CPDMA_SKIP_SYNC_BEFORE | \
1321 SI_CPDMA_SKIP_GFX_SYNC | \
1322 SI_CPDMA_SKIP_BO_LIST_UPDATE)
1323
1324 void si_cp_dma_wait_for_idle(struct si_context *sctx);
1325 void si_cp_dma_clear_buffer(struct si_context *sctx, struct radeon_cmdbuf *cs,
1326 struct pipe_resource *dst, uint64_t offset,
1327 uint64_t size, unsigned value, unsigned user_flags,
1328 enum si_coherency coher, enum si_cache_policy cache_policy);
1329 void si_cp_dma_copy_buffer(struct si_context *sctx,
1330 struct pipe_resource *dst, struct pipe_resource *src,
1331 uint64_t dst_offset, uint64_t src_offset, unsigned size,
1332 unsigned user_flags, enum si_coherency coher,
1333 enum si_cache_policy cache_policy);
1334 void cik_prefetch_TC_L2_async(struct si_context *sctx, struct pipe_resource *buf,
1335 uint64_t offset, unsigned size);
1336 void cik_emit_prefetch_L2(struct si_context *sctx, bool vertex_stage_only);
1337 void si_test_gds(struct si_context *sctx);
1338 void si_cp_write_data(struct si_context *sctx, struct si_resource *buf,
1339 unsigned offset, unsigned size, unsigned dst_sel,
1340 unsigned engine, const void *data);
1341 void si_cp_copy_data(struct si_context *sctx, struct radeon_cmdbuf *cs,
1342 unsigned dst_sel, struct si_resource *dst, unsigned dst_offset,
1343 unsigned src_sel, struct si_resource *src, unsigned src_offset);
1344
1345 /* si_debug.c */
1346 void si_save_cs(struct radeon_winsys *ws, struct radeon_cmdbuf *cs,
1347 struct radeon_saved_cs *saved, bool get_buffer_list);
1348 void si_clear_saved_cs(struct radeon_saved_cs *saved);
1349 void si_destroy_saved_cs(struct si_saved_cs *scs);
1350 void si_auto_log_cs(void *data, struct u_log_context *log);
1351 void si_log_hw_flush(struct si_context *sctx);
1352 void si_log_draw_state(struct si_context *sctx, struct u_log_context *log);
1353 void si_log_compute_state(struct si_context *sctx, struct u_log_context *log);
1354 void si_init_debug_functions(struct si_context *sctx);
1355 void si_check_vm_faults(struct si_context *sctx,
1356 struct radeon_saved_cs *saved, enum ring_type ring);
1357 bool si_replace_shader(unsigned num, struct si_shader_binary *binary);
1358
1359 /* si_dma.c */
1360 void si_init_dma_functions(struct si_context *sctx);
1361
1362 /* si_dma_cs.c */
1363 void si_dma_emit_timestamp(struct si_context *sctx, struct si_resource *dst,
1364 uint64_t offset);
1365 void si_sdma_clear_buffer(struct si_context *sctx, struct pipe_resource *dst,
1366 uint64_t offset, uint64_t size, unsigned clear_value);
1367 void si_need_dma_space(struct si_context *ctx, unsigned num_dw,
1368 struct si_resource *dst, struct si_resource *src);
1369 void si_flush_dma_cs(struct si_context *ctx, unsigned flags,
1370 struct pipe_fence_handle **fence);
1371 void si_screen_clear_buffer(struct si_screen *sscreen, struct pipe_resource *dst,
1372 uint64_t offset, uint64_t size, unsigned value);
1373
1374 /* si_fence.c */
1375 void si_cp_release_mem(struct si_context *ctx, struct radeon_cmdbuf *cs,
1376 unsigned event, unsigned event_flags,
1377 unsigned dst_sel, unsigned int_sel, unsigned data_sel,
1378 struct si_resource *buf, uint64_t va,
1379 uint32_t new_fence, unsigned query_type);
1380 unsigned si_cp_write_fence_dwords(struct si_screen *screen);
1381 void si_cp_wait_mem(struct si_context *ctx, struct radeon_cmdbuf *cs,
1382 uint64_t va, uint32_t ref, uint32_t mask, unsigned flags);
1383 void si_init_fence_functions(struct si_context *ctx);
1384 void si_init_screen_fence_functions(struct si_screen *screen);
1385 struct pipe_fence_handle *si_create_fence(struct pipe_context *ctx,
1386 struct tc_unflushed_batch_token *tc_token);
1387
1388 /* si_get.c */
1389 void si_init_screen_get_functions(struct si_screen *sscreen);
1390
1391 /* si_gfx_cs.c */
1392 void si_flush_gfx_cs(struct si_context *ctx, unsigned flags,
1393 struct pipe_fence_handle **fence);
1394 void si_allocate_gds(struct si_context *ctx);
1395 void si_begin_new_gfx_cs(struct si_context *ctx);
1396 void si_need_gfx_cs_space(struct si_context *ctx);
1397 void si_unref_sdma_uploads(struct si_context *sctx);
1398
1399 /* si_gpu_load.c */
1400 void si_gpu_load_kill_thread(struct si_screen *sscreen);
1401 uint64_t si_begin_counter(struct si_screen *sscreen, unsigned type);
1402 unsigned si_end_counter(struct si_screen *sscreen, unsigned type,
1403 uint64_t begin);
1404
1405 /* si_compute.c */
1406 void si_emit_initial_compute_regs(struct si_context *sctx, struct radeon_cmdbuf *cs);
1407 void si_init_compute_functions(struct si_context *sctx);
1408
1409 /* si_compute_prim_discard.c */
1410 enum si_prim_discard_outcome {
1411 SI_PRIM_DISCARD_ENABLED,
1412 SI_PRIM_DISCARD_DISABLED,
1413 SI_PRIM_DISCARD_DRAW_SPLIT,
1414 };
1415
1416 void si_build_prim_discard_compute_shader(struct si_shader_context *ctx);
1417 enum si_prim_discard_outcome
1418 si_prepare_prim_discard_or_split_draw(struct si_context *sctx,
1419 const struct pipe_draw_info *info,
1420 bool primitive_restart);
1421 void si_compute_signal_gfx(struct si_context *sctx);
1422 void si_dispatch_prim_discard_cs_and_draw(struct si_context *sctx,
1423 const struct pipe_draw_info *info,
1424 unsigned index_size,
1425 unsigned base_vertex,
1426 uint64_t input_indexbuf_va,
1427 unsigned input_indexbuf_max_elements);
1428 void si_initialize_prim_discard_tunables(struct si_context *sctx);
1429
1430 /* si_perfcounters.c */
1431 void si_init_perfcounters(struct si_screen *screen);
1432 void si_destroy_perfcounters(struct si_screen *screen);
1433
1434 /* si_pipe.c */
1435 bool si_check_device_reset(struct si_context *sctx);
1436
1437 /* si_query.c */
1438 void si_init_screen_query_functions(struct si_screen *sscreen);
1439 void si_init_query_functions(struct si_context *sctx);
1440 void si_suspend_queries(struct si_context *sctx);
1441 void si_resume_queries(struct si_context *sctx);
1442
1443 /* si_shaderlib_tgsi.c */
1444 void *si_get_blitter_vs(struct si_context *sctx, enum blitter_attrib_type type,
1445 unsigned num_layers);
1446 void *si_create_fixed_func_tcs(struct si_context *sctx);
1447 void *si_create_dma_compute_shader(struct pipe_context *ctx,
1448 unsigned num_dwords_per_thread,
1449 bool dst_stream_cache_policy, bool is_copy);
1450 void *si_create_copy_image_compute_shader(struct pipe_context *ctx);
1451 void *si_create_copy_image_compute_shader_1d_array(struct pipe_context *ctx);
1452 void *si_clear_render_target_shader(struct pipe_context *ctx);
1453 void *si_clear_render_target_shader_1d_array(struct pipe_context *ctx);
1454 void *si_create_dcc_retile_cs(struct pipe_context *ctx);
1455 void *si_create_query_result_cs(struct si_context *sctx);
1456 void *gfx10_create_sh_query_result_cs(struct si_context *sctx);
1457
1458 /* gfx10_query.c */
1459 void gfx10_init_query(struct si_context *sctx);
1460 void gfx10_destroy_query(struct si_context *sctx);
1461
1462 /* si_test_dma.c */
1463 void si_test_dma(struct si_screen *sscreen);
1464
1465 /* si_test_clearbuffer.c */
1466 void si_test_dma_perf(struct si_screen *sscreen);
1467
1468 /* si_uvd.c */
1469 struct pipe_video_codec *si_uvd_create_decoder(struct pipe_context *context,
1470 const struct pipe_video_codec *templ);
1471
1472 struct pipe_video_buffer *si_video_buffer_create(struct pipe_context *pipe,
1473 const struct pipe_video_buffer *tmpl);
1474
1475 /* si_viewport.c */
1476 void si_update_vs_viewport_state(struct si_context *ctx);
1477 void si_init_viewport_functions(struct si_context *ctx);
1478
1479 /* si_texture.c */
1480 bool si_prepare_for_dma_blit(struct si_context *sctx,
1481 struct si_texture *dst,
1482 unsigned dst_level, unsigned dstx,
1483 unsigned dsty, unsigned dstz,
1484 struct si_texture *src,
1485 unsigned src_level,
1486 const struct pipe_box *src_box);
1487 void si_eliminate_fast_color_clear(struct si_context *sctx,
1488 struct si_texture *tex);
1489 void si_texture_discard_cmask(struct si_screen *sscreen,
1490 struct si_texture *tex);
1491 bool si_init_flushed_depth_texture(struct pipe_context *ctx,
1492 struct pipe_resource *texture);
1493 void si_print_texture_info(struct si_screen *sscreen,
1494 struct si_texture *tex, struct u_log_context *log);
1495 struct pipe_resource *si_texture_create(struct pipe_screen *screen,
1496 const struct pipe_resource *templ);
1497 bool vi_dcc_formats_compatible(struct si_screen *sscreen,
1498 enum pipe_format format1,
1499 enum pipe_format format2);
1500 bool vi_dcc_formats_are_incompatible(struct pipe_resource *tex,
1501 unsigned level,
1502 enum pipe_format view_format);
1503 void vi_disable_dcc_if_incompatible_format(struct si_context *sctx,
1504 struct pipe_resource *tex,
1505 unsigned level,
1506 enum pipe_format view_format);
1507 struct pipe_surface *si_create_surface_custom(struct pipe_context *pipe,
1508 struct pipe_resource *texture,
1509 const struct pipe_surface *templ,
1510 unsigned width0, unsigned height0,
1511 unsigned width, unsigned height);
1512 unsigned si_translate_colorswap(enum pipe_format format, bool do_endian_swap);
1513 void vi_separate_dcc_try_enable(struct si_context *sctx,
1514 struct si_texture *tex);
1515 void vi_separate_dcc_start_query(struct si_context *sctx,
1516 struct si_texture *tex);
1517 void vi_separate_dcc_stop_query(struct si_context *sctx,
1518 struct si_texture *tex);
1519 void vi_separate_dcc_process_and_reset_stats(struct pipe_context *ctx,
1520 struct si_texture *tex);
1521 bool si_texture_disable_dcc(struct si_context *sctx,
1522 struct si_texture *tex);
1523 void si_init_screen_texture_functions(struct si_screen *sscreen);
1524 void si_init_context_texture_functions(struct si_context *sctx);
1525
1526
1527 /*
1528 * common helpers
1529 */
1530
1531 static inline struct si_resource *si_resource(struct pipe_resource *r)
1532 {
1533 return (struct si_resource*)r;
1534 }
1535
1536 static inline void
1537 si_resource_reference(struct si_resource **ptr, struct si_resource *res)
1538 {
1539 pipe_resource_reference((struct pipe_resource **)ptr,
1540 (struct pipe_resource *)res);
1541 }
1542
1543 static inline void
1544 si_texture_reference(struct si_texture **ptr, struct si_texture *res)
1545 {
1546 pipe_resource_reference((struct pipe_resource **)ptr, &res->buffer.b.b);
1547 }
1548
1549 static inline bool
1550 vi_dcc_enabled(struct si_texture *tex, unsigned level)
1551 {
1552 return tex->dcc_offset && level < tex->surface.num_dcc_levels;
1553 }
1554
1555 static inline unsigned
1556 si_tile_mode_index(struct si_texture *tex, unsigned level, bool stencil)
1557 {
1558 if (stencil)
1559 return tex->surface.u.legacy.stencil_tiling_index[level];
1560 else
1561 return tex->surface.u.legacy.tiling_index[level];
1562 }
1563
1564 static inline unsigned
1565 si_get_minimum_num_gfx_cs_dwords(struct si_context *sctx)
1566 {
1567 /* Don't count the needed CS space exactly and just use an upper bound.
1568 *
1569 * Also reserve space for stopping queries at the end of IB, because
1570 * the number of active queries is unlimited in theory.
1571 */
1572 return 2048 + sctx->num_cs_dw_queries_suspend;
1573 }
1574
1575 static inline void
1576 si_context_add_resource_size(struct si_context *sctx, struct pipe_resource *r)
1577 {
1578 if (r) {
1579 /* Add memory usage for need_gfx_cs_space */
1580 sctx->vram += si_resource(r)->vram_usage;
1581 sctx->gtt += si_resource(r)->gart_usage;
1582 }
1583 }
1584
1585 static inline void
1586 si_invalidate_draw_sh_constants(struct si_context *sctx)
1587 {
1588 sctx->last_base_vertex = SI_BASE_VERTEX_UNKNOWN;
1589 sctx->last_instance_count = SI_INSTANCE_COUNT_UNKNOWN;
1590 }
1591
1592 static inline unsigned
1593 si_get_atom_bit(struct si_context *sctx, struct si_atom *atom)
1594 {
1595 return 1 << (atom - sctx->atoms.array);
1596 }
1597
1598 static inline void
1599 si_set_atom_dirty(struct si_context *sctx, struct si_atom *atom, bool dirty)
1600 {
1601 unsigned bit = si_get_atom_bit(sctx, atom);
1602
1603 if (dirty)
1604 sctx->dirty_atoms |= bit;
1605 else
1606 sctx->dirty_atoms &= ~bit;
1607 }
1608
1609 static inline bool
1610 si_is_atom_dirty(struct si_context *sctx, struct si_atom *atom)
1611 {
1612 return (sctx->dirty_atoms & si_get_atom_bit(sctx, atom)) != 0;
1613 }
1614
1615 static inline void
1616 si_mark_atom_dirty(struct si_context *sctx, struct si_atom *atom)
1617 {
1618 si_set_atom_dirty(sctx, atom, true);
1619 }
1620
1621 static inline struct si_shader_ctx_state *si_get_vs(struct si_context *sctx)
1622 {
1623 if (sctx->gs_shader.cso)
1624 return &sctx->gs_shader;
1625 if (sctx->tes_shader.cso)
1626 return &sctx->tes_shader;
1627
1628 return &sctx->vs_shader;
1629 }
1630
1631 static inline struct tgsi_shader_info *si_get_vs_info(struct si_context *sctx)
1632 {
1633 struct si_shader_ctx_state *vs = si_get_vs(sctx);
1634
1635 return vs->cso ? &vs->cso->info : NULL;
1636 }
1637
1638 static inline struct si_shader* si_get_vs_state(struct si_context *sctx)
1639 {
1640 if (sctx->gs_shader.cso &&
1641 sctx->gs_shader.current &&
1642 !sctx->gs_shader.current->key.as_ngg)
1643 return sctx->gs_shader.cso->gs_copy_shader;
1644
1645 struct si_shader_ctx_state *vs = si_get_vs(sctx);
1646 return vs->current ? vs->current : NULL;
1647 }
1648
1649 static inline bool si_can_dump_shader(struct si_screen *sscreen,
1650 unsigned processor)
1651 {
1652 return sscreen->debug_flags & (1 << processor);
1653 }
1654
1655 static inline bool si_get_strmout_en(struct si_context *sctx)
1656 {
1657 return sctx->streamout.streamout_enabled ||
1658 sctx->streamout.prims_gen_query_enabled;
1659 }
1660
1661 static inline unsigned
1662 si_optimal_tcc_alignment(struct si_context *sctx, unsigned upload_size)
1663 {
1664 unsigned alignment, tcc_cache_line_size;
1665
1666 /* If the upload size is less than the cache line size (e.g. 16, 32),
1667 * the whole thing will fit into a cache line if we align it to its size.
1668 * The idea is that multiple small uploads can share a cache line.
1669 * If the upload size is greater, align it to the cache line size.
1670 */
1671 alignment = util_next_power_of_two(upload_size);
1672 tcc_cache_line_size = sctx->screen->info.tcc_cache_line_size;
1673 return MIN2(alignment, tcc_cache_line_size);
1674 }
1675
1676 static inline void
1677 si_saved_cs_reference(struct si_saved_cs **dst, struct si_saved_cs *src)
1678 {
1679 if (pipe_reference(&(*dst)->reference, &src->reference))
1680 si_destroy_saved_cs(*dst);
1681
1682 *dst = src;
1683 }
1684
1685 static inline void
1686 si_make_CB_shader_coherent(struct si_context *sctx, unsigned num_samples,
1687 bool shaders_read_metadata, bool dcc_pipe_aligned)
1688 {
1689 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_CB |
1690 SI_CONTEXT_INV_VCACHE;
1691
1692 if (sctx->chip_class >= GFX10) {
1693 if (shaders_read_metadata)
1694 sctx->flags |= SI_CONTEXT_INV_L2_METADATA;
1695 } else if (sctx->chip_class == GFX9) {
1696 /* Single-sample color is coherent with shaders on GFX9, but
1697 * L2 metadata must be flushed if shaders read metadata.
1698 * (DCC, CMASK).
1699 */
1700 if (num_samples >= 2 ||
1701 (shaders_read_metadata && !dcc_pipe_aligned))
1702 sctx->flags |= SI_CONTEXT_INV_L2;
1703 else if (shaders_read_metadata)
1704 sctx->flags |= SI_CONTEXT_INV_L2_METADATA;
1705 } else {
1706 /* GFX6-GFX8 */
1707 sctx->flags |= SI_CONTEXT_INV_L2;
1708 }
1709 }
1710
1711 static inline void
1712 si_make_DB_shader_coherent(struct si_context *sctx, unsigned num_samples,
1713 bool include_stencil, bool shaders_read_metadata)
1714 {
1715 sctx->flags |= SI_CONTEXT_FLUSH_AND_INV_DB |
1716 SI_CONTEXT_INV_VCACHE;
1717
1718 if (sctx->chip_class >= GFX10) {
1719 if (shaders_read_metadata)
1720 sctx->flags |= SI_CONTEXT_INV_L2_METADATA;
1721 } else if (sctx->chip_class == GFX9) {
1722 /* Single-sample depth (not stencil) is coherent with shaders
1723 * on GFX9, but L2 metadata must be flushed if shaders read
1724 * metadata.
1725 */
1726 if (num_samples >= 2 || include_stencil)
1727 sctx->flags |= SI_CONTEXT_INV_L2;
1728 else if (shaders_read_metadata)
1729 sctx->flags |= SI_CONTEXT_INV_L2_METADATA;
1730 } else {
1731 /* GFX6-GFX8 */
1732 sctx->flags |= SI_CONTEXT_INV_L2;
1733 }
1734 }
1735
1736 static inline bool
1737 si_can_sample_zs(struct si_texture *tex, bool stencil_sampler)
1738 {
1739 return (stencil_sampler && tex->can_sample_s) ||
1740 (!stencil_sampler && tex->can_sample_z);
1741 }
1742
1743 static inline bool
1744 si_htile_enabled(struct si_texture *tex, unsigned level, unsigned zs_mask)
1745 {
1746 if (zs_mask == PIPE_MASK_S && tex->htile_stencil_disabled)
1747 return false;
1748
1749 return tex->htile_offset && level == 0;
1750 }
1751
1752 static inline bool
1753 vi_tc_compat_htile_enabled(struct si_texture *tex, unsigned level, unsigned zs_mask)
1754 {
1755 assert(!tex->tc_compatible_htile || tex->htile_offset);
1756 return tex->tc_compatible_htile && si_htile_enabled(tex, level, zs_mask);
1757 }
1758
1759 static inline unsigned si_get_ps_iter_samples(struct si_context *sctx)
1760 {
1761 if (sctx->ps_uses_fbfetch)
1762 return sctx->framebuffer.nr_color_samples;
1763
1764 return MIN2(sctx->ps_iter_samples, sctx->framebuffer.nr_color_samples);
1765 }
1766
1767 static inline unsigned si_get_total_colormask(struct si_context *sctx)
1768 {
1769 if (sctx->queued.named.rasterizer->rasterizer_discard)
1770 return 0;
1771
1772 struct si_shader_selector *ps = sctx->ps_shader.cso;
1773 if (!ps)
1774 return 0;
1775
1776 unsigned colormask = sctx->framebuffer.colorbuf_enabled_4bit &
1777 sctx->queued.named.blend->cb_target_mask;
1778
1779 if (!ps->info.properties[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS])
1780 colormask &= ps->colors_written_4bit;
1781 else if (!ps->colors_written_4bit)
1782 colormask = 0; /* color0 writes all cbufs, but it's not written */
1783
1784 return colormask;
1785 }
1786
1787 #define UTIL_ALL_PRIM_LINE_MODES ((1 << PIPE_PRIM_LINES) | \
1788 (1 << PIPE_PRIM_LINE_LOOP) | \
1789 (1 << PIPE_PRIM_LINE_STRIP) | \
1790 (1 << PIPE_PRIM_LINES_ADJACENCY) | \
1791 (1 << PIPE_PRIM_LINE_STRIP_ADJACENCY))
1792
1793 static inline bool util_prim_is_lines(unsigned prim)
1794 {
1795 return ((1 << prim) & UTIL_ALL_PRIM_LINE_MODES) != 0;
1796 }
1797
1798 static inline bool util_prim_is_points_or_lines(unsigned prim)
1799 {
1800 return ((1 << prim) & (UTIL_ALL_PRIM_LINE_MODES |
1801 (1 << PIPE_PRIM_POINTS))) != 0;
1802 }
1803
1804 static inline bool util_rast_prim_is_triangles(unsigned prim)
1805 {
1806 return ((1 << prim) & ((1 << PIPE_PRIM_TRIANGLES) |
1807 (1 << PIPE_PRIM_TRIANGLE_STRIP) |
1808 (1 << PIPE_PRIM_TRIANGLE_FAN) |
1809 (1 << PIPE_PRIM_QUADS) |
1810 (1 << PIPE_PRIM_QUAD_STRIP) |
1811 (1 << PIPE_PRIM_POLYGON) |
1812 (1 << PIPE_PRIM_TRIANGLES_ADJACENCY) |
1813 (1 << PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY)));
1814 }
1815
1816 /**
1817 * Return true if there is enough memory in VRAM and GTT for the buffers
1818 * added so far.
1819 *
1820 * \param vram VRAM memory size not added to the buffer list yet
1821 * \param gtt GTT memory size not added to the buffer list yet
1822 */
1823 static inline bool
1824 radeon_cs_memory_below_limit(struct si_screen *screen,
1825 struct radeon_cmdbuf *cs,
1826 uint64_t vram, uint64_t gtt)
1827 {
1828 vram += cs->used_vram;
1829 gtt += cs->used_gart;
1830
1831 /* Anything that goes above the VRAM size should go to GTT. */
1832 if (vram > screen->info.vram_size)
1833 gtt += vram - screen->info.vram_size;
1834
1835 /* Now we just need to check if we have enough GTT. */
1836 return gtt < screen->info.gart_size * 0.7;
1837 }
1838
1839 /**
1840 * Add a buffer to the buffer list for the given command stream (CS).
1841 *
1842 * All buffers used by a CS must be added to the list. This tells the kernel
1843 * driver which buffers are used by GPU commands. Other buffers can
1844 * be swapped out (not accessible) during execution.
1845 *
1846 * The buffer list becomes empty after every context flush and must be
1847 * rebuilt.
1848 */
1849 static inline void radeon_add_to_buffer_list(struct si_context *sctx,
1850 struct radeon_cmdbuf *cs,
1851 struct si_resource *bo,
1852 enum radeon_bo_usage usage,
1853 enum radeon_bo_priority priority)
1854 {
1855 assert(usage);
1856 sctx->ws->cs_add_buffer(
1857 cs, bo->buf,
1858 (enum radeon_bo_usage)(usage | RADEON_USAGE_SYNCHRONIZED),
1859 bo->domains, priority);
1860 }
1861
1862 /**
1863 * Same as above, but also checks memory usage and flushes the context
1864 * accordingly.
1865 *
1866 * When this SHOULD NOT be used:
1867 *
1868 * - if si_context_add_resource_size has been called for the buffer
1869 * followed by *_need_cs_space for checking the memory usage
1870 *
1871 * - if si_need_dma_space has been called for the buffer
1872 *
1873 * - when emitting state packets and draw packets (because preceding packets
1874 * can't be re-emitted at that point)
1875 *
1876 * - if shader resource "enabled_mask" is not up-to-date or there is
1877 * a different constraint disallowing a context flush
1878 */
1879 static inline void
1880 radeon_add_to_gfx_buffer_list_check_mem(struct si_context *sctx,
1881 struct si_resource *bo,
1882 enum radeon_bo_usage usage,
1883 enum radeon_bo_priority priority,
1884 bool check_mem)
1885 {
1886 if (check_mem &&
1887 !radeon_cs_memory_below_limit(sctx->screen, sctx->gfx_cs,
1888 sctx->vram + bo->vram_usage,
1889 sctx->gtt + bo->gart_usage))
1890 si_flush_gfx_cs(sctx, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW, NULL);
1891
1892 radeon_add_to_buffer_list(sctx, sctx->gfx_cs, bo, usage, priority);
1893 }
1894
1895 static inline bool si_compute_prim_discard_enabled(struct si_context *sctx)
1896 {
1897 return sctx->prim_discard_vertex_count_threshold != UINT_MAX;
1898 }
1899
1900 static inline unsigned si_get_wave_size(struct si_screen *sscreen,
1901 enum pipe_shader_type shader_type,
1902 bool ngg, bool es)
1903 {
1904 if (shader_type == PIPE_SHADER_COMPUTE)
1905 return sscreen->compute_wave_size;
1906 else if (shader_type == PIPE_SHADER_FRAGMENT)
1907 return sscreen->ps_wave_size;
1908 else if ((shader_type == PIPE_SHADER_TESS_EVAL && es && !ngg) ||
1909 (shader_type == PIPE_SHADER_GEOMETRY && !ngg)) /* legacy GS only supports Wave64 */
1910 return 64;
1911 else
1912 return sscreen->ge_wave_size;
1913 }
1914
1915 static inline unsigned si_get_shader_wave_size(struct si_shader *shader)
1916 {
1917 return si_get_wave_size(shader->selector->screen, shader->selector->type,
1918 shader->key.as_ngg, shader->key.as_es);
1919 }
1920
1921 #define PRINT_ERR(fmt, args...) \
1922 fprintf(stderr, "EE %s:%d %s - " fmt, __FILE__, __LINE__, __func__, ##args)
1923
1924 #endif