2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
29 #include "si_shader.h"
31 #include "util/u_dynarray.h"
32 #include "util/u_idalloc.h"
34 #ifdef PIPE_ARCH_BIG_ENDIAN
35 #define SI_BIG_ENDIAN 1
37 #define SI_BIG_ENDIAN 0
40 /* The base vertex and primitive restart can be any number, but we must pick
41 * one which will mean "unknown" for the purpose of state tracking and
42 * the number shouldn't be a commonly-used one. */
43 #define SI_BASE_VERTEX_UNKNOWN INT_MIN
44 #define SI_RESTART_INDEX_UNKNOWN INT_MIN
45 #define SI_NUM_SMOOTH_AA_SAMPLES 8
46 #define SI_GS_PER_ES 128
47 /* Alignment for optimal CP DMA performance. */
48 #define SI_CPDMA_ALIGNMENT 32
50 /* Instruction cache. */
51 #define SI_CONTEXT_INV_ICACHE (R600_CONTEXT_PRIVATE_FLAG << 0)
52 /* SMEM L1, other names: KCACHE, constant cache, DCACHE, data cache */
53 #define SI_CONTEXT_INV_SMEM_L1 (R600_CONTEXT_PRIVATE_FLAG << 1)
54 /* VMEM L1 can optionally be bypassed (GLC=1). Other names: TC L1 */
55 #define SI_CONTEXT_INV_VMEM_L1 (R600_CONTEXT_PRIVATE_FLAG << 2)
56 /* Used by everything except CB/DB, can be bypassed (SLC=1). Other names: TC L2 */
57 #define SI_CONTEXT_INV_GLOBAL_L2 (R600_CONTEXT_PRIVATE_FLAG << 3)
58 /* Write dirty L2 lines back to memory (shader and CP DMA stores), but don't
59 * invalidate L2. SI-CIK can't do it, so they will do complete invalidation. */
60 #define SI_CONTEXT_WRITEBACK_GLOBAL_L2 (R600_CONTEXT_PRIVATE_FLAG << 4)
61 /* Writeback & invalidate the L2 metadata cache. It can only be coupled with
62 * a CB or DB flush. */
63 #define SI_CONTEXT_INV_L2_METADATA (R600_CONTEXT_PRIVATE_FLAG << 5)
64 /* Framebuffer caches. */
65 #define SI_CONTEXT_FLUSH_AND_INV_DB (R600_CONTEXT_PRIVATE_FLAG << 6)
66 #define SI_CONTEXT_FLUSH_AND_INV_DB_META (R600_CONTEXT_PRIVATE_FLAG << 7)
67 #define SI_CONTEXT_FLUSH_AND_INV_CB (R600_CONTEXT_PRIVATE_FLAG << 8)
68 /* Engine synchronization. */
69 #define SI_CONTEXT_VS_PARTIAL_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 9)
70 #define SI_CONTEXT_PS_PARTIAL_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 10)
71 #define SI_CONTEXT_CS_PARTIAL_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 11)
72 #define SI_CONTEXT_VGT_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 12)
73 #define SI_CONTEXT_VGT_STREAMOUT_SYNC (R600_CONTEXT_PRIVATE_FLAG << 13)
75 #define SI_PREFETCH_VBO_DESCRIPTORS (1 << 0)
76 #define SI_PREFETCH_LS (1 << 1)
77 #define SI_PREFETCH_HS (1 << 2)
78 #define SI_PREFETCH_ES (1 << 3)
79 #define SI_PREFETCH_GS (1 << 4)
80 #define SI_PREFETCH_VS (1 << 5)
81 #define SI_PREFETCH_PS (1 << 6)
83 #define SI_MAX_BORDER_COLORS 4096
88 struct u_suballocator
;
91 struct r600_common_screen b
;
92 unsigned gs_table_depth
;
93 unsigned tess_offchip_block_dw_size
;
95 bool has_distributed_tess
;
96 bool has_draw_indirect_multi
;
97 bool has_out_of_order_rast
;
98 bool assume_no_z_fights
;
99 bool commutative_blend_add
;
100 bool has_msaa_sample_loc_bug
;
103 bool llvm_has_working_vgpr_indexing
;
105 /* Whether shaders are monolithic (1-part) or separate (3-part). */
106 bool use_monolithic_shaders
;
109 mtx_t shader_parts_mutex
;
110 struct si_shader_part
*vs_prologs
;
111 struct si_shader_part
*tcs_epilogs
;
112 struct si_shader_part
*gs_prologs
;
113 struct si_shader_part
*ps_prologs
;
114 struct si_shader_part
*ps_epilogs
;
116 /* Shader cache in memory.
118 * Design & limitations:
119 * - The shader cache is per screen (= per process), never saved to
120 * disk, and skips redundant shader compilations from TGSI to bytecode.
121 * - It can only be used with one-variant-per-shader support, in which
122 * case only the main (typically middle) part of shaders is cached.
123 * - Only VS, TCS, TES, PS are cached, out of which only the hw VS
124 * variants of VS and TES are cached, so LS and ES aren't.
125 * - GS and CS aren't cached, but it's certainly possible to cache
128 mtx_t shader_cache_mutex
;
129 struct hash_table
*shader_cache
;
131 /* Shader compiler queue for multithreaded compilation. */
132 struct util_queue shader_compiler_queue
;
133 /* Use at most 3 normal compiler threads on quadcore and better.
134 * Hyperthreaded CPUs report the number of threads, but we want
135 * the number of cores. */
136 LLVMTargetMachineRef tm
[3]; /* used by the queue only */
138 struct util_queue shader_compiler_queue_low_priority
;
139 /* Use at most 2 low priority threads on quadcore and better.
140 * We want to minimize the impact on multithreaded Mesa. */
141 LLVMTargetMachineRef tm_low_priority
[2]; /* at most 2 threads */
144 struct si_blend_color
{
145 struct r600_atom atom
;
146 struct pipe_blend_color state
;
150 struct si_sampler_view
{
151 struct pipe_sampler_view base
;
152 /* [0..7] = image descriptor
153 * [4..7] = buffer descriptor */
155 uint32_t fmask_state
[8];
156 const struct legacy_surf_level
*base_level_info
;
159 bool is_stencil_sampler
;
161 bool dcc_incompatible
;
164 #define SI_SAMPLER_STATE_MAGIC 0x34f1c35a
166 struct si_sampler_state
{
171 uint32_t integer_val
[4];
172 uint32_t upgraded_depth_val
[4];
175 struct si_cs_shader_state
{
176 struct si_compute
*program
;
177 struct si_compute
*emitted_program
;
183 struct si_textures_info
{
184 struct si_sampler_views views
;
185 uint32_t needs_depth_decompress_mask
;
186 uint32_t needs_color_decompress_mask
;
189 struct si_images_info
{
190 struct pipe_image_view views
[SI_NUM_IMAGES
];
191 uint32_t needs_color_decompress_mask
;
192 unsigned enabled_mask
;
195 struct si_framebuffer
{
196 struct r600_atom atom
;
197 struct pipe_framebuffer_state state
;
198 unsigned colorbuf_enabled_4bit
;
199 unsigned spi_shader_col_format
;
200 unsigned spi_shader_col_format_alpha
;
201 unsigned spi_shader_col_format_blend
;
202 unsigned spi_shader_col_format_blend_alpha
;
203 ubyte nr_samples
:5; /* at most 16xAA */
204 ubyte log_samples
:3; /* at most 4 = 16xAA */
205 ubyte compressed_cb_mask
;
207 ubyte color_is_int10
;
211 bool CB_has_shader_readable_metadata
;
212 bool DB_has_shader_readable_metadata
;
215 struct si_clip_state
{
216 struct r600_atom atom
;
217 struct pipe_clip_state state
;
221 struct si_sample_locs
{
222 struct r600_atom atom
;
226 struct si_sample_mask
{
227 struct r600_atom atom
;
228 uint16_t sample_mask
;
231 /* A shader state consists of the shader selector, which is a constant state
232 * object shared by multiple contexts and shouldn't be modified, and
233 * the current shader variant selected for this context.
235 struct si_shader_ctx_state
{
236 struct si_shader_selector
*cso
;
237 struct si_shader
*current
;
240 #define SI_NUM_VGT_PARAM_KEY_BITS 12
241 #define SI_NUM_VGT_PARAM_STATES (1 << SI_NUM_VGT_PARAM_KEY_BITS)
243 /* The IA_MULTI_VGT_PARAM key used to index the table of precomputed values.
244 * Some fields are set by state-change calls, most are set by draw_vbo.
246 union si_vgt_param_key
{
249 unsigned uses_instancing
:1;
250 unsigned multi_instances_smaller_than_primgroup
:1;
251 unsigned primitive_restart
:1;
252 unsigned count_from_stream_output
:1;
253 unsigned line_stipple_enabled
:1;
254 unsigned uses_tess
:1;
255 unsigned tess_uses_prim_id
:1;
257 unsigned _pad
:32 - SI_NUM_VGT_PARAM_KEY_BITS
;
262 struct si_texture_handle
266 struct pipe_sampler_view
*view
;
267 struct si_sampler_state sstate
;
270 struct si_image_handle
274 struct pipe_image_view view
;
278 struct pipe_reference reference
;
279 struct si_context
*ctx
;
280 struct radeon_saved_cs gfx
;
281 struct r600_resource
*trace_buf
;
284 unsigned gfx_last_dw
;
289 struct r600_common_context b
;
290 struct blitter_context
*blitter
;
291 void *custom_dsa_flush
;
292 void *custom_blend_resolve
;
293 void *custom_blend_fmask_decompress
;
294 void *custom_blend_eliminate_fastclear
;
295 void *custom_blend_dcc_decompress
;
296 struct si_screen
*screen
;
297 LLVMTargetMachineRef tm
; /* only non-threaded compilation */
298 struct si_shader_ctx_state fixed_func_tcs_shader
;
299 struct r600_resource
*wait_mem_scratch
;
300 unsigned wait_mem_number
;
301 uint16_t prefetch_L2_mask
;
303 bool gfx_flush_in_progress
:1;
304 bool compute_is_busy
:1;
306 /* Atoms (direct states). */
307 union si_state_atoms atoms
;
308 unsigned dirty_atoms
; /* mask */
309 /* PM4 states (precomputed immutable states) */
310 unsigned dirty_states
;
311 union si_state queued
;
312 union si_state emitted
;
314 /* Atom declarations. */
315 struct si_framebuffer framebuffer
;
316 struct si_sample_locs msaa_sample_locs
;
317 struct r600_atom db_render_state
;
318 struct r600_atom dpbb_state
;
319 struct r600_atom msaa_config
;
320 struct si_sample_mask sample_mask
;
321 struct r600_atom cb_render_state
;
322 unsigned last_cb_target_mask
;
323 struct si_blend_color blend_color
;
324 struct r600_atom clip_regs
;
325 struct si_clip_state clip_state
;
326 struct si_shader_data shader_pointers
;
327 struct si_stencil_ref stencil_ref
;
328 struct r600_atom spi_map
;
330 /* Precomputed states. */
331 struct si_pm4_state
*init_config
;
332 struct si_pm4_state
*init_config_gs_rings
;
333 bool init_config_has_vgt_flush
;
334 struct si_pm4_state
*vgt_shader_config
[4];
337 struct si_shader_ctx_state ps_shader
;
338 struct si_shader_ctx_state gs_shader
;
339 struct si_shader_ctx_state vs_shader
;
340 struct si_shader_ctx_state tcs_shader
;
341 struct si_shader_ctx_state tes_shader
;
342 struct si_cs_shader_state cs_shader_state
;
344 /* shader information */
345 struct si_vertex_elements
*vertex_elements
;
346 unsigned sprite_coord_enable
;
348 bool do_update_shaders
;
350 /* shader descriptors */
351 struct si_descriptors vertex_buffers
;
352 struct si_descriptors descriptors
[SI_NUM_DESCS
];
353 unsigned descriptors_dirty
;
354 unsigned shader_pointers_dirty
;
355 unsigned shader_needs_decompress_mask
;
356 struct si_buffer_resources rw_buffers
;
357 struct si_buffer_resources const_and_shader_buffers
[SI_NUM_SHADERS
];
358 struct si_textures_info samplers
[SI_NUM_SHADERS
];
359 struct si_images_info images
[SI_NUM_SHADERS
];
361 /* other shader resources */
362 struct pipe_constant_buffer null_const_buf
; /* used for set_constant_buffer(NULL) on CIK */
363 struct pipe_resource
*esgs_ring
;
364 struct pipe_resource
*gsvs_ring
;
365 struct pipe_resource
*tf_ring
;
366 struct pipe_resource
*tess_offchip_ring
;
367 union pipe_color_union
*border_color_table
; /* in CPU memory, any endian */
368 struct r600_resource
*border_color_buffer
;
369 union pipe_color_union
*border_color_map
; /* in VRAM (slow access), little endian */
370 unsigned border_color_count
;
372 /* Vertex and index buffers. */
373 bool vertex_buffers_dirty
;
374 bool vertex_buffer_pointer_dirty
;
375 struct pipe_vertex_buffer vertex_buffer
[SI_NUM_VERTEX_BUFFERS
];
377 /* MSAA config state. */
379 bool smoothing_enabled
;
381 /* DB render state. */
382 unsigned ps_db_shader_control
;
383 unsigned dbcb_copy_sample
;
384 bool dbcb_depth_copy_enabled
:1;
385 bool dbcb_stencil_copy_enabled
:1;
386 bool db_flush_depth_inplace
:1;
387 bool db_flush_stencil_inplace
:1;
388 bool db_depth_clear
:1;
389 bool db_depth_disable_expclear
:1;
390 bool db_stencil_clear
:1;
391 bool db_stencil_disable_expclear
:1;
392 bool occlusion_queries_disabled
:1;
393 bool generate_mipmap_for_depth
:1;
395 /* Emitted draw state. */
396 bool gs_tri_strip_adj_fix
:1;
399 int last_base_vertex
;
400 int last_start_instance
;
402 int last_sh_base_reg
;
403 int last_primitive_restart_en
;
404 int last_restart_index
;
405 int last_gs_out_prim
;
407 int last_multi_vgt_param
;
409 unsigned last_sc_line_stipple
;
410 unsigned current_vs_state
;
411 unsigned last_vs_state
;
414 struct r600_atom scratch_state
;
415 struct r600_resource
*scratch_buffer
;
416 unsigned scratch_waves
;
417 unsigned spi_tmpring_size
;
419 struct r600_resource
*compute_scratch_buffer
;
421 /* Emitted derived tessellation state. */
422 /* Local shader (VS), or HS if LS-HS are merged. */
423 struct si_shader
*last_ls
;
424 struct si_shader_selector
*last_tcs
;
425 int last_num_tcs_input_cp
;
426 int last_tes_sh_base
;
427 bool last_tess_uses_primid
;
428 unsigned last_num_patches
;
432 struct si_saved_cs
*current_saved_cs
;
433 uint64_t dmesg_timestamp
;
434 unsigned apitrace_call_number
;
437 bool need_check_render_feedback
;
438 bool decompression_enabled
;
440 /* Precomputed IA_MULTI_VGT_PARAM */
441 union si_vgt_param_key ia_multi_vgt_param_key
;
442 unsigned ia_multi_vgt_param
[SI_NUM_VGT_PARAM_STATES
];
444 /* Bindless descriptors. */
445 struct si_descriptors bindless_descriptors
;
446 struct util_idalloc bindless_used_slots
;
447 unsigned num_bindless_descriptors
;
448 bool bindless_descriptors_dirty
;
449 bool graphics_bindless_pointer_dirty
;
450 bool compute_bindless_pointer_dirty
;
452 /* Allocated bindless handles */
453 struct hash_table
*tex_handles
;
454 struct hash_table
*img_handles
;
456 /* Resident bindless handles */
457 struct util_dynarray resident_tex_handles
;
458 struct util_dynarray resident_img_handles
;
460 /* Resident bindless handles which need decompression */
461 struct util_dynarray resident_tex_needs_color_decompress
;
462 struct util_dynarray resident_img_needs_color_decompress
;
463 struct util_dynarray resident_tex_needs_depth_decompress
;
466 bool uses_bindless_samplers
;
467 bool uses_bindless_images
;
471 void cik_init_sdma_functions(struct si_context
*sctx
);
474 void si_init_blit_functions(struct si_context
*sctx
);
475 void si_decompress_graphics_textures(struct si_context
*sctx
);
476 void si_decompress_compute_textures(struct si_context
*sctx
);
477 void si_resource_copy_region(struct pipe_context
*ctx
,
478 struct pipe_resource
*dst
,
480 unsigned dstx
, unsigned dsty
, unsigned dstz
,
481 struct pipe_resource
*src
,
483 const struct pipe_box
*src_box
);
486 #define SI_CPDMA_SKIP_CHECK_CS_SPACE (1 << 0) /* don't call need_cs_space */
487 #define SI_CPDMA_SKIP_SYNC_AFTER (1 << 1) /* don't wait for DMA after the copy */
488 #define SI_CPDMA_SKIP_SYNC_BEFORE (1 << 2) /* don't wait for DMA before the copy (RAW hazards) */
489 #define SI_CPDMA_SKIP_GFX_SYNC (1 << 3) /* don't flush caches and don't wait for PS/CS */
490 #define SI_CPDMA_SKIP_BO_LIST_UPDATE (1 << 4) /* don't update the BO list */
491 #define SI_CPDMA_SKIP_ALL (SI_CPDMA_SKIP_CHECK_CS_SPACE | \
492 SI_CPDMA_SKIP_SYNC_AFTER | \
493 SI_CPDMA_SKIP_SYNC_BEFORE | \
494 SI_CPDMA_SKIP_GFX_SYNC | \
495 SI_CPDMA_SKIP_BO_LIST_UPDATE)
497 void si_copy_buffer(struct si_context
*sctx
,
498 struct pipe_resource
*dst
, struct pipe_resource
*src
,
499 uint64_t dst_offset
, uint64_t src_offset
, unsigned size
,
500 unsigned user_flags
);
501 void cik_prefetch_TC_L2_async(struct si_context
*sctx
, struct pipe_resource
*buf
,
502 uint64_t offset
, unsigned size
);
503 void cik_emit_prefetch_L2(struct si_context
*sctx
);
504 void si_init_cp_dma_functions(struct si_context
*sctx
);
507 void si_auto_log_cs(void *data
, struct u_log_context
*log
);
508 void si_log_hw_flush(struct si_context
*sctx
);
509 void si_log_draw_state(struct si_context
*sctx
, struct u_log_context
*log
);
510 void si_log_compute_state(struct si_context
*sctx
, struct u_log_context
*log
);
511 void si_init_debug_functions(struct si_context
*sctx
);
512 void si_check_vm_faults(struct r600_common_context
*ctx
,
513 struct radeon_saved_cs
*saved
, enum ring_type ring
);
514 bool si_replace_shader(unsigned num
, struct ac_shader_binary
*binary
);
517 void si_init_dma_functions(struct si_context
*sctx
);
519 /* si_hw_context.c */
520 void si_destroy_saved_cs(struct si_saved_cs
*scs
);
521 void si_context_gfx_flush(void *context
, unsigned flags
,
522 struct pipe_fence_handle
**fence
);
523 void si_begin_new_cs(struct si_context
*ctx
);
524 void si_need_cs_space(struct si_context
*ctx
);
527 void si_init_compute_functions(struct si_context
*sctx
);
529 /* si_perfcounters.c */
530 void si_init_perfcounters(struct si_screen
*screen
);
533 struct pipe_video_codec
*si_uvd_create_decoder(struct pipe_context
*context
,
534 const struct pipe_video_codec
*templ
);
536 struct pipe_video_buffer
*si_video_buffer_create(struct pipe_context
*pipe
,
537 const struct pipe_video_buffer
*tmpl
);
544 si_invalidate_draw_sh_constants(struct si_context
*sctx
)
546 sctx
->last_base_vertex
= SI_BASE_VERTEX_UNKNOWN
;
550 si_set_atom_dirty(struct si_context
*sctx
,
551 struct r600_atom
*atom
, bool dirty
)
553 unsigned bit
= 1 << atom
->id
;
556 sctx
->dirty_atoms
|= bit
;
558 sctx
->dirty_atoms
&= ~bit
;
562 si_is_atom_dirty(struct si_context
*sctx
,
563 struct r600_atom
*atom
)
565 unsigned bit
= 1 << atom
->id
;
567 return sctx
->dirty_atoms
& bit
;
571 si_mark_atom_dirty(struct si_context
*sctx
,
572 struct r600_atom
*atom
)
574 si_set_atom_dirty(sctx
, atom
, true);
577 static inline struct si_shader_ctx_state
*si_get_vs(struct si_context
*sctx
)
579 if (sctx
->gs_shader
.cso
)
580 return &sctx
->gs_shader
;
581 if (sctx
->tes_shader
.cso
)
582 return &sctx
->tes_shader
;
584 return &sctx
->vs_shader
;
587 static inline struct tgsi_shader_info
*si_get_vs_info(struct si_context
*sctx
)
589 struct si_shader_ctx_state
*vs
= si_get_vs(sctx
);
591 return vs
->cso
? &vs
->cso
->info
: NULL
;
594 static inline struct si_shader
* si_get_vs_state(struct si_context
*sctx
)
596 if (sctx
->gs_shader
.cso
)
597 return sctx
->gs_shader
.cso
->gs_copy_shader
;
599 struct si_shader_ctx_state
*vs
= si_get_vs(sctx
);
600 return vs
->current
? vs
->current
: NULL
;
603 static inline unsigned
604 si_optimal_tcc_alignment(struct si_context
*sctx
, unsigned upload_size
)
606 unsigned alignment
, tcc_cache_line_size
;
608 /* If the upload size is less than the cache line size (e.g. 16, 32),
609 * the whole thing will fit into a cache line if we align it to its size.
610 * The idea is that multiple small uploads can share a cache line.
611 * If the upload size is greater, align it to the cache line size.
613 alignment
= util_next_power_of_two(upload_size
);
614 tcc_cache_line_size
= sctx
->screen
->b
.info
.tcc_cache_line_size
;
615 return MIN2(alignment
, tcc_cache_line_size
);
619 si_saved_cs_reference(struct si_saved_cs
**dst
, struct si_saved_cs
*src
)
621 if (pipe_reference(&(*dst
)->reference
, &src
->reference
))
622 si_destroy_saved_cs(*dst
);
628 si_make_CB_shader_coherent(struct si_context
*sctx
, unsigned num_samples
,
629 bool shaders_read_metadata
)
631 sctx
->b
.flags
|= SI_CONTEXT_FLUSH_AND_INV_CB
|
632 SI_CONTEXT_INV_VMEM_L1
;
634 if (sctx
->b
.chip_class
>= GFX9
) {
635 /* Single-sample color is coherent with shaders on GFX9, but
636 * L2 metadata must be flushed if shaders read metadata.
639 if (num_samples
>= 2)
640 sctx
->b
.flags
|= SI_CONTEXT_INV_GLOBAL_L2
;
641 else if (shaders_read_metadata
)
642 sctx
->b
.flags
|= SI_CONTEXT_INV_L2_METADATA
;
645 sctx
->b
.flags
|= SI_CONTEXT_INV_GLOBAL_L2
;
650 si_make_DB_shader_coherent(struct si_context
*sctx
, unsigned num_samples
,
651 bool include_stencil
, bool shaders_read_metadata
)
653 sctx
->b
.flags
|= SI_CONTEXT_FLUSH_AND_INV_DB
|
654 SI_CONTEXT_INV_VMEM_L1
;
656 if (sctx
->b
.chip_class
>= GFX9
) {
657 /* Single-sample depth (not stencil) is coherent with shaders
658 * on GFX9, but L2 metadata must be flushed if shaders read
661 if (num_samples
>= 2 || include_stencil
)
662 sctx
->b
.flags
|= SI_CONTEXT_INV_GLOBAL_L2
;
663 else if (shaders_read_metadata
)
664 sctx
->b
.flags
|= SI_CONTEXT_INV_L2_METADATA
;
667 sctx
->b
.flags
|= SI_CONTEXT_INV_GLOBAL_L2
;