2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3 * Copyright 2018 Advanced Micro Devices, Inc.
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:
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
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.
28 #include "si_shader.h"
31 #include "util/u_dynarray.h"
32 #include "util/u_idalloc.h"
33 #include "util/u_threaded_context.h"
35 #ifdef PIPE_ARCH_BIG_ENDIAN
36 #define SI_BIG_ENDIAN 1
38 #define SI_BIG_ENDIAN 0
41 #define ATI_VENDOR_ID 0x1002
43 #define SI_NOT_QUERY 0xffffffff
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
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
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 /* SMEM L1, other names: KCACHE, constant cache, DCACHE, data cache */
69 #define SI_CONTEXT_INV_SMEM_L1 (1 << 4)
70 /* VMEM L1 can optionally be bypassed (GLC=1). Other names: TC L1 */
71 #define SI_CONTEXT_INV_VMEM_L1 (1 << 5)
72 /* Used by everything except CB/DB, can be bypassed (SLC=1). Other names: TC L2 */
73 #define SI_CONTEXT_INV_GLOBAL_L2 (1 << 6)
74 /* Write dirty L2 lines back to memory (shader and CP DMA stores), but don't
75 * invalidate L2. GFX6-GFX7 can't do it, so they will do complete invalidation. */
76 #define SI_CONTEXT_WRITEBACK_GLOBAL_L2 (1 << 7)
77 /* Writeback & invalidate the L2 metadata cache. It can only be coupled with
78 * a CB or DB flush. */
79 #define SI_CONTEXT_INV_L2_METADATA (1 << 8)
80 /* Framebuffer caches. */
81 #define SI_CONTEXT_FLUSH_AND_INV_DB (1 << 9)
82 #define SI_CONTEXT_FLUSH_AND_INV_DB_META (1 << 10)
83 #define SI_CONTEXT_FLUSH_AND_INV_CB (1 << 11)
84 /* Engine synchronization. */
85 #define SI_CONTEXT_VS_PARTIAL_FLUSH (1 << 12)
86 #define SI_CONTEXT_PS_PARTIAL_FLUSH (1 << 13)
87 #define SI_CONTEXT_CS_PARTIAL_FLUSH (1 << 14)
88 #define SI_CONTEXT_VGT_FLUSH (1 << 15)
89 #define SI_CONTEXT_VGT_STREAMOUT_SYNC (1 << 16)
91 #define SI_PREFETCH_VBO_DESCRIPTORS (1 << 0)
92 #define SI_PREFETCH_LS (1 << 1)
93 #define SI_PREFETCH_HS (1 << 2)
94 #define SI_PREFETCH_ES (1 << 3)
95 #define SI_PREFETCH_GS (1 << 4)
96 #define SI_PREFETCH_VS (1 << 5)
97 #define SI_PREFETCH_PS (1 << 6)
99 #define SI_MAX_BORDER_COLORS 4096
100 #define SI_MAX_VIEWPORTS 16
101 #define SIX_BITS 0x3F
102 #define SI_MAP_BUFFER_ALIGNMENT 64
103 #define SI_MAX_VARIABLE_THREADS_PER_BLOCK 1024
105 #define SI_RESOURCE_FLAG_TRANSFER (PIPE_RESOURCE_FLAG_DRV_PRIV << 0)
106 #define SI_RESOURCE_FLAG_FLUSHED_DEPTH (PIPE_RESOURCE_FLAG_DRV_PRIV << 1)
107 #define SI_RESOURCE_FLAG_FORCE_MSAA_TILING (PIPE_RESOURCE_FLAG_DRV_PRIV << 2)
108 #define SI_RESOURCE_FLAG_DISABLE_DCC (PIPE_RESOURCE_FLAG_DRV_PRIV << 3)
109 #define SI_RESOURCE_FLAG_UNMAPPABLE (PIPE_RESOURCE_FLAG_DRV_PRIV << 4)
110 #define SI_RESOURCE_FLAG_READ_ONLY (PIPE_RESOURCE_FLAG_DRV_PRIV << 5)
111 #define SI_RESOURCE_FLAG_32BIT (PIPE_RESOURCE_FLAG_DRV_PRIV << 6)
112 #define SI_RESOURCE_FLAG_CLEAR (PIPE_RESOURCE_FLAG_DRV_PRIV << 7)
113 /* For const_uploader, upload data via GTT and copy to VRAM on context flush via SDMA. */
114 #define SI_RESOURCE_FLAG_UPLOAD_FLUSH_EXPLICIT_VIA_SDMA (PIPE_RESOURCE_FLAG_DRV_PRIV << 8)
118 DCC_CLEAR_COLOR_0000
= 0x00000000,
119 DCC_CLEAR_COLOR_0001
= 0x40404040,
120 DCC_CLEAR_COLOR_1110
= 0x80808080,
121 DCC_CLEAR_COLOR_1111
= 0xC0C0C0C0,
122 DCC_CLEAR_COLOR_REG
= 0x20202020,
123 DCC_UNCOMPRESSED
= 0xFFFFFFFF,
126 #define SI_IMAGE_ACCESS_AS_BUFFER (1 << 7)
130 /* Shader logging options: */
131 DBG_VS
= PIPE_SHADER_VERTEX
,
132 DBG_PS
= PIPE_SHADER_FRAGMENT
,
133 DBG_GS
= PIPE_SHADER_GEOMETRY
,
134 DBG_TCS
= PIPE_SHADER_TESS_CTRL
,
135 DBG_TES
= PIPE_SHADER_TESS_EVAL
,
136 DBG_CS
= PIPE_SHADER_COMPUTE
,
142 /* Shader compiler options the shader cache should be aware of: */
143 DBG_FS_CORRECT_DERIVS_AFTER_KILL
,
148 /* Shader compiler options (with no effect on the shader cache): */
150 DBG_MONOLITHIC_SHADERS
,
153 /* Information logging options: */
159 /* Driver options: */
167 /* 3D engine options: */
187 DBG_TEST_VMFAULT_SDMA
,
188 DBG_TEST_VMFAULT_SHADER
,
195 #define DBG_ALL_SHADERS (((1 << (DBG_CS + 1)) - 1))
196 #define DBG(name) (1ull << DBG_##name)
198 enum si_cache_policy
{
200 L2_STREAM
, /* same as SLC=1 */
201 L2_LRU
, /* same as SLC=0 */
205 SI_COHERENCY_NONE
, /* no cache flushes needed */
207 SI_COHERENCY_CB_META
,
213 struct u_suballocator
;
215 /* Only 32-bit buffer allocations are supported, gallium doesn't support more
219 struct threaded_resource b
;
221 /* Winsys objects. */
222 struct pb_buffer
*buf
;
223 uint64_t gpu_address
;
224 /* Memory usage if the buffer placement is optimal. */
228 /* Resource properties. */
230 unsigned bo_alignment
;
231 enum radeon_bo_domain domains
;
232 enum radeon_bo_flag flags
;
233 unsigned bind_history
;
234 int max_forced_staging_uploads
;
236 /* The buffer range which is initialized (with a write transfer,
237 * streamout, DMA, or as a random access target). The rest of
238 * the buffer is considered invalid and can be mapped unsynchronized.
240 * This allows unsychronized mapping of a buffer range which hasn't
241 * been used yet. It's for applications which forget to use
242 * the unsynchronized map flag and expect the driver to figure it out.
244 struct util_range valid_buffer_range
;
246 /* For buffers only. This indicates that a write operation has been
247 * performed by TC L2, but the cache hasn't been flushed.
248 * Any hw block which doesn't use or bypasses TC L2 should check this
249 * flag and flush the cache before using the buffer.
251 * For example, TC L2 must be flushed if a buffer which has been
252 * modified by a shader store instruction is about to be used as
253 * an index buffer. The reason is that VGT DMA index fetching doesn't
258 /* Whether this resource is referenced by bindless handles. */
259 bool texture_handle_allocated
;
260 bool image_handle_allocated
;
262 /* Whether the resource has been exported via resource_get_handle. */
263 unsigned external_usage
; /* PIPE_HANDLE_USAGE_* */
267 struct threaded_transfer b
;
268 struct si_resource
*staging
;
273 struct si_resource buffer
;
275 struct radeon_surf surface
;
277 struct si_texture
*flushed_depth_texture
;
279 /* One texture allocation can contain these buffers:
280 * - image (pixel data)
281 * - FMASK buffer (MSAA compression)
282 * - CMASK buffer (MSAA compression and/or legacy fast color clear)
283 * - HTILE buffer (Z/S compression and fast Z/S clear)
284 * - DCC buffer (color compression and new fast color clear)
285 * - displayable DCC buffer (if the DCC buffer is not displayable)
286 * - DCC retile mapping buffer (if the DCC buffer is not displayable)
288 uint64_t fmask_offset
;
289 uint64_t cmask_offset
;
290 uint64_t cmask_base_address_reg
;
291 struct si_resource
*cmask_buffer
;
292 uint64_t dcc_offset
; /* 0 = disabled */
293 uint64_t display_dcc_offset
;
294 uint64_t dcc_retile_map_offset
;
295 unsigned cb_color_info
; /* fast clear enable bit */
296 unsigned color_clear_value
[2];
297 unsigned last_msaa_resolve_target_micro_mode
;
298 unsigned num_level0_transfers
;
300 /* Depth buffer compression and fast clear. */
301 uint64_t htile_offset
;
302 float depth_clear_value
;
303 uint16_t dirty_level_mask
; /* each bit says if that mipmap is compressed */
304 uint16_t stencil_dirty_level_mask
; /* each bit says if that mipmap is compressed */
305 enum pipe_format db_render_format
:16;
306 uint8_t stencil_clear_value
;
307 bool tc_compatible_htile
:1;
308 bool depth_cleared
:1; /* if it was cleared at least once */
309 bool stencil_cleared
:1; /* if it was cleared at least once */
310 bool upgraded_depth
:1; /* upgraded from unorm to Z32_FLOAT */
312 bool db_compatible
:1;
316 /* We need to track DCC dirtiness, because st/dri usually calls
317 * flush_resource twice per frame (not a bug) and we don't wanna
318 * decompress DCC twice. Also, the dirty tracking must be done even
319 * if DCC isn't used, because it's required by the DCC usage analysis
320 * for a possible future enablement.
322 bool separate_dcc_dirty
:1;
323 /* Statistics gathering for the DCC enablement heuristic. */
324 bool dcc_gather_statistics
:1;
325 /* Counter that should be non-zero if the texture is bound to a
328 unsigned framebuffers_bound
;
329 /* Whether the texture is a displayable back buffer and needs DCC
330 * decompression, which is expensive. Therefore, it's enabled only
331 * if statistics suggest that it will pay off and it's allocated
332 * separately. It can't be bound as a sampler by apps. Limited to
333 * target == 2D and last_level == 0. If enabled, dcc_offset contains
334 * the absolute GPUVM address, not the relative one.
336 struct si_resource
*dcc_separate_buffer
;
337 /* When DCC is temporarily disabled, the separate buffer is here. */
338 struct si_resource
*last_dcc_separate_buffer
;
339 /* Estimate of how much this color buffer is written to in units of
340 * full-screen draws: ps_invocations / (width * height)
341 * Shader kills, late Z, and blending with trivial discards make it
342 * inaccurate (we need to count CB updates, not PS invocations).
344 unsigned ps_draw_ratio
;
345 /* The number of clears since the last DCC usage analysis. */
346 unsigned num_slow_clears
;
350 struct pipe_surface base
;
352 /* These can vary with block-compressed textures. */
356 bool color_initialized
:1;
357 bool depth_initialized
:1;
359 /* Misc. color flags. */
360 bool color_is_int8
:1;
361 bool color_is_int10
:1;
362 bool dcc_incompatible
:1;
364 /* Color registers. */
365 unsigned cb_color_info
;
366 unsigned cb_color_view
;
367 unsigned cb_color_attrib
;
368 unsigned cb_color_attrib2
; /* GFX9 and later */
369 unsigned cb_dcc_control
; /* GFX8 and later */
370 unsigned spi_shader_col_format
:8; /* no blending, no alpha-to-coverage. */
371 unsigned spi_shader_col_format_alpha
:8; /* alpha-to-coverage */
372 unsigned spi_shader_col_format_blend
:8; /* blending without alpha. */
373 unsigned spi_shader_col_format_blend_alpha
:8; /* blending with alpha. */
376 uint64_t db_depth_base
; /* DB_Z_READ/WRITE_BASE */
377 uint64_t db_stencil_base
;
378 uint64_t db_htile_data_base
;
379 unsigned db_depth_info
;
381 unsigned db_z_info2
; /* GFX9+ */
382 unsigned db_depth_view
;
383 unsigned db_depth_size
;
384 unsigned db_depth_slice
;
385 unsigned db_stencil_info
;
386 unsigned db_stencil_info2
; /* GFX9+ */
387 unsigned db_htile_surface
;
390 struct si_mmio_counter
{
395 union si_mmio_counters
{
397 /* For global GPU load including SDMA. */
398 struct si_mmio_counter gpu
;
401 struct si_mmio_counter spi
;
402 struct si_mmio_counter gui
;
403 struct si_mmio_counter ta
;
404 struct si_mmio_counter gds
;
405 struct si_mmio_counter vgt
;
406 struct si_mmio_counter ia
;
407 struct si_mmio_counter sx
;
408 struct si_mmio_counter wd
;
409 struct si_mmio_counter bci
;
410 struct si_mmio_counter sc
;
411 struct si_mmio_counter pa
;
412 struct si_mmio_counter db
;
413 struct si_mmio_counter cp
;
414 struct si_mmio_counter cb
;
417 struct si_mmio_counter sdma
;
420 struct si_mmio_counter pfp
;
421 struct si_mmio_counter meq
;
422 struct si_mmio_counter me
;
423 struct si_mmio_counter surf_sync
;
424 struct si_mmio_counter cp_dma
;
425 struct si_mmio_counter scratch_ram
;
430 struct si_memory_object
{
431 struct pipe_memory_object b
;
432 struct pb_buffer
*buf
;
436 /* Saved CS data for debugging features. */
437 struct radeon_saved_cs
{
441 struct radeon_bo_list_item
*bo_list
;
446 struct pipe_screen b
;
447 struct radeon_winsys
*ws
;
448 struct disk_cache
*disk_shader_cache
;
450 struct radeon_info info
;
451 uint64_t debug_flags
;
452 char renderer_string
[183];
454 unsigned pa_sc_raster_config
;
455 unsigned pa_sc_raster_config_1
;
456 unsigned se_tile_repeat
;
457 unsigned gs_table_depth
;
458 unsigned tess_offchip_block_dw_size
;
459 unsigned tess_offchip_ring_size
;
460 unsigned tess_factor_ring_size
;
461 unsigned vgt_hs_offchip_param
;
462 unsigned eqaa_force_coverage_samples
;
463 unsigned eqaa_force_z_samples
;
464 unsigned eqaa_force_color_samples
;
465 bool has_clear_state
;
466 bool has_distributed_tess
;
467 bool has_draw_indirect_multi
;
468 bool has_out_of_order_rast
;
469 bool assume_no_z_fights
;
470 bool commutative_blend_add
;
471 bool has_gfx9_scissor_bug
;
472 bool has_msaa_sample_loc_bug
;
473 bool has_ls_vgpr_init_bug
;
474 bool has_dcc_constant_encode
;
477 bool llvm_has_working_vgpr_indexing
;
480 #define OPT_BOOL(name, dflt, description) bool name:1;
481 #include "si_debug_options.h"
484 /* Whether shaders are monolithic (1-part) or separate (3-part). */
485 bool use_monolithic_shaders
;
487 bool has_rbplus
; /* if RB+ registers exist */
488 bool rbplus_allowed
; /* if RB+ is allowed */
489 bool dcc_msaa_allowed
;
490 bool cpdma_prefetch_writes_memory
;
492 struct slab_parent_pool pool_transfers
;
494 /* Texture filter settings. */
495 int force_aniso
; /* -1 = disabled */
497 /* Auxiliary context. Mainly used to initialize resources.
498 * It must be locked prior to using and flushed before unlocking. */
499 struct pipe_context
*aux_context
;
500 mtx_t aux_context_lock
;
502 /* This must be in the screen, because UE4 uses one context for
503 * compilation and another one for rendering.
505 unsigned num_compilations
;
506 /* Along with ST_DEBUG=precompile, this should show if applications
507 * are loading shaders on demand. This is a monotonic counter.
509 unsigned num_shaders_created
;
510 unsigned num_shader_cache_hits
;
512 /* GPU load thread. */
513 mtx_t gpu_load_mutex
;
514 thrd_t gpu_load_thread
;
515 union si_mmio_counters mmio_counters
;
516 volatile unsigned gpu_load_stop_thread
; /* bool */
518 /* Performance counters. */
519 struct si_perfcounters
*perfcounters
;
521 /* If pipe_screen wants to recompute and re-emit the framebuffer,
522 * sampler, and image states of all contexts, it should atomically
525 * Each context will compare this with its own last known value of
526 * the counter before drawing and re-emit the states accordingly.
528 unsigned dirty_tex_counter
;
530 /* Atomically increment this counter when an existing texture's
531 * metadata is enabled or disabled in a way that requires changing
532 * contexts' compressed texture binding masks.
534 unsigned compressed_colortex_counter
;
537 /* Context flags to set so that all writes from earlier jobs
538 * in the CP are seen by L2 clients.
542 /* Context flags to set so that all writes from earlier jobs
543 * that end in L2 are seen by CP.
548 mtx_t shader_parts_mutex
;
549 struct si_shader_part
*vs_prologs
;
550 struct si_shader_part
*tcs_epilogs
;
551 struct si_shader_part
*gs_prologs
;
552 struct si_shader_part
*ps_prologs
;
553 struct si_shader_part
*ps_epilogs
;
555 /* Shader cache in memory.
557 * Design & limitations:
558 * - The shader cache is per screen (= per process), never saved to
559 * disk, and skips redundant shader compilations from TGSI to bytecode.
560 * - It can only be used with one-variant-per-shader support, in which
561 * case only the main (typically middle) part of shaders is cached.
562 * - Only VS, TCS, TES, PS are cached, out of which only the hw VS
563 * variants of VS and TES are cached, so LS and ES aren't.
564 * - GS and CS aren't cached, but it's certainly possible to cache
567 mtx_t shader_cache_mutex
;
568 struct hash_table
*shader_cache
;
570 /* Shader compiler queue for multithreaded compilation. */
571 struct util_queue shader_compiler_queue
;
572 /* Use at most 3 normal compiler threads on quadcore and better.
573 * Hyperthreaded CPUs report the number of threads, but we want
574 * the number of cores. We only need this many threads for shader-db. */
575 struct ac_llvm_compiler compiler
[24]; /* used by the queue only */
577 struct util_queue shader_compiler_queue_low_priority
;
578 /* Use at most 2 low priority threads on quadcore and better.
579 * We want to minimize the impact on multithreaded Mesa. */
580 struct ac_llvm_compiler compiler_lowp
[10];
583 struct si_blend_color
{
584 struct pipe_blend_color state
;
588 struct si_sampler_view
{
589 struct pipe_sampler_view base
;
590 /* [0..7] = image descriptor
591 * [4..7] = buffer descriptor */
593 uint32_t fmask_state
[8];
594 const struct legacy_surf_level
*base_level_info
;
597 bool is_stencil_sampler
;
599 bool dcc_incompatible
;
602 #define SI_SAMPLER_STATE_MAGIC 0x34f1c35a
604 struct si_sampler_state
{
609 uint32_t integer_val
[4];
610 uint32_t upgraded_depth_val
[4];
613 struct si_cs_shader_state
{
614 struct si_compute
*program
;
615 struct si_compute
*emitted_program
;
622 struct pipe_sampler_view
*views
[SI_NUM_SAMPLERS
];
623 struct si_sampler_state
*sampler_states
[SI_NUM_SAMPLERS
];
625 /* The i-th bit is set if that element is enabled (non-NULL resource). */
626 unsigned enabled_mask
;
627 uint32_t needs_depth_decompress_mask
;
628 uint32_t needs_color_decompress_mask
;
632 struct pipe_image_view views
[SI_NUM_IMAGES
];
633 uint32_t needs_color_decompress_mask
;
634 unsigned enabled_mask
;
637 struct si_framebuffer
{
638 struct pipe_framebuffer_state state
;
639 unsigned colorbuf_enabled_4bit
;
640 unsigned spi_shader_col_format
;
641 unsigned spi_shader_col_format_alpha
;
642 unsigned spi_shader_col_format_blend
;
643 unsigned spi_shader_col_format_blend_alpha
;
644 ubyte nr_samples
:5; /* at most 16xAA */
645 ubyte log_samples
:3; /* at most 4 = 16xAA */
646 ubyte nr_color_samples
; /* at most 8xAA */
647 ubyte compressed_cb_mask
;
648 ubyte uncompressed_cb_mask
;
650 ubyte color_is_int10
;
652 ubyte dcc_overwrite_combiner_watermark
;
655 bool CB_has_shader_readable_metadata
;
656 bool DB_has_shader_readable_metadata
;
657 bool all_DCC_pipe_aligned
;
661 /* This is the list we want to support. */
662 SI_QUANT_MODE_16_8_FIXED_POINT_1_256TH
,
663 SI_QUANT_MODE_14_10_FIXED_POINT_1_1024TH
,
664 SI_QUANT_MODE_12_12_FIXED_POINT_1_4096TH
,
667 struct si_signed_scissor
{
672 enum si_quant_mode quant_mode
;
675 struct si_viewports
{
676 struct pipe_viewport_state states
[SI_MAX_VIEWPORTS
];
677 struct si_signed_scissor as_scissor
[SI_MAX_VIEWPORTS
];
680 struct si_clip_state
{
681 struct pipe_clip_state state
;
685 struct si_streamout_target
{
686 struct pipe_stream_output_target b
;
688 /* The buffer where BUFFER_FILLED_SIZE is stored. */
689 struct si_resource
*buf_filled_size
;
690 unsigned buf_filled_size_offset
;
691 bool buf_filled_size_valid
;
693 unsigned stride_in_dw
;
696 struct si_streamout
{
699 unsigned enabled_mask
;
700 unsigned num_targets
;
701 struct si_streamout_target
*targets
[PIPE_MAX_SO_BUFFERS
];
703 unsigned append_bitmask
;
706 /* External state which comes from the vertex shader,
707 * it must be set explicitly when binding a shader. */
708 uint16_t *stride_in_dw
;
709 unsigned enabled_stream_buffers_mask
; /* stream0 buffers0-3 in 4 LSB */
711 /* The state of VGT_STRMOUT_BUFFER_(CONFIG|EN). */
712 unsigned hw_enabled_mask
;
714 /* The state of VGT_STRMOUT_(CONFIG|EN). */
715 bool streamout_enabled
;
716 bool prims_gen_query_enabled
;
717 int num_prims_gen_queries
;
720 /* A shader state consists of the shader selector, which is a constant state
721 * object shared by multiple contexts and shouldn't be modified, and
722 * the current shader variant selected for this context.
724 struct si_shader_ctx_state
{
725 struct si_shader_selector
*cso
;
726 struct si_shader
*current
;
729 #define SI_NUM_VGT_PARAM_KEY_BITS 12
730 #define SI_NUM_VGT_PARAM_STATES (1 << SI_NUM_VGT_PARAM_KEY_BITS)
732 /* The IA_MULTI_VGT_PARAM key used to index the table of precomputed values.
733 * Some fields are set by state-change calls, most are set by draw_vbo.
735 union si_vgt_param_key
{
737 #ifdef PIPE_ARCH_LITTLE_ENDIAN
739 unsigned uses_instancing
:1;
740 unsigned multi_instances_smaller_than_primgroup
:1;
741 unsigned primitive_restart
:1;
742 unsigned count_from_stream_output
:1;
743 unsigned line_stipple_enabled
:1;
744 unsigned uses_tess
:1;
745 unsigned tess_uses_prim_id
:1;
747 unsigned _pad
:32 - SI_NUM_VGT_PARAM_KEY_BITS
;
748 #else /* PIPE_ARCH_BIG_ENDIAN */
749 unsigned _pad
:32 - SI_NUM_VGT_PARAM_KEY_BITS
;
751 unsigned tess_uses_prim_id
:1;
752 unsigned uses_tess
:1;
753 unsigned line_stipple_enabled
:1;
754 unsigned count_from_stream_output
:1;
755 unsigned primitive_restart
:1;
756 unsigned multi_instances_smaller_than_primgroup
:1;
757 unsigned uses_instancing
:1;
764 struct si_texture_handle
768 struct pipe_sampler_view
*view
;
769 struct si_sampler_state sstate
;
772 struct si_image_handle
776 struct pipe_image_view view
;
780 struct pipe_reference reference
;
781 struct si_context
*ctx
;
782 struct radeon_saved_cs gfx
;
783 struct si_resource
*trace_buf
;
786 unsigned gfx_last_dw
;
791 struct si_sdma_upload
{
792 struct si_resource
*dst
;
793 struct si_resource
*src
;
800 struct pipe_context b
; /* base class */
802 enum radeon_family family
;
803 enum chip_class chip_class
;
805 struct radeon_winsys
*ws
;
806 struct radeon_winsys_ctx
*ctx
;
807 struct radeon_cmdbuf
*gfx_cs
; /* compute IB if graphics is disabled */
808 struct radeon_cmdbuf
*dma_cs
;
809 struct pipe_fence_handle
*last_gfx_fence
;
810 struct pipe_fence_handle
*last_sdma_fence
;
811 struct si_resource
*eop_bug_scratch
;
812 struct u_upload_mgr
*cached_gtt_allocator
;
813 struct threaded_context
*tc
;
814 struct u_suballocator
*allocator_zeroed_memory
;
815 struct slab_child_pool pool_transfers
;
816 struct slab_child_pool pool_transfers_unsync
; /* for threaded_context */
817 struct pipe_device_reset_callback device_reset_callback
;
818 struct u_log_context
*log
;
819 void *query_result_shader
;
820 struct blitter_context
*blitter
;
821 void *custom_dsa_flush
;
822 void *custom_blend_resolve
;
823 void *custom_blend_fmask_decompress
;
824 void *custom_blend_eliminate_fastclear
;
825 void *custom_blend_dcc_decompress
;
827 void *vs_blit_pos_layered
;
829 void *vs_blit_color_layered
;
830 void *vs_blit_texcoord
;
831 void *cs_clear_buffer
;
832 void *cs_copy_buffer
;
834 void *cs_copy_image_1d_array
;
835 void *cs_clear_render_target
;
836 void *cs_clear_render_target_1d_array
;
838 struct si_screen
*screen
;
839 struct pipe_debug_callback debug
;
840 struct ac_llvm_compiler compiler
; /* only non-threaded compilation */
841 struct si_shader_ctx_state fixed_func_tcs_shader
;
842 struct si_resource
*wait_mem_scratch
;
843 unsigned wait_mem_number
;
844 uint16_t prefetch_L2_mask
;
847 bool gfx_flush_in_progress
:1;
848 bool gfx_last_ib_is_busy
:1;
849 bool compute_is_busy
:1;
851 unsigned num_gfx_cs_flushes
;
852 unsigned initial_gfx_cs_size
;
853 unsigned gpu_reset_counter
;
854 unsigned last_dirty_tex_counter
;
855 unsigned last_compressed_colortex_counter
;
856 unsigned last_num_draw_calls
;
857 unsigned flags
; /* flush flags */
858 /* Current unaccounted memory usage. */
862 /* Atoms (direct states). */
863 union si_state_atoms atoms
;
864 unsigned dirty_atoms
; /* mask */
865 /* PM4 states (precomputed immutable states) */
866 unsigned dirty_states
;
867 union si_state queued
;
868 union si_state emitted
;
870 /* Atom declarations. */
871 struct si_framebuffer framebuffer
;
872 unsigned sample_locs_num_samples
;
873 uint16_t sample_mask
;
874 unsigned last_cb_target_mask
;
875 struct si_blend_color blend_color
;
876 struct si_clip_state clip_state
;
877 struct si_shader_data shader_pointers
;
878 struct si_stencil_ref stencil_ref
;
879 struct pipe_scissor_state scissors
[SI_MAX_VIEWPORTS
];
880 struct si_streamout streamout
;
881 struct si_viewports viewports
;
882 unsigned num_window_rectangles
;
883 bool window_rectangles_include
;
884 struct pipe_scissor_state window_rectangles
[4];
886 /* Precomputed states. */
887 struct si_pm4_state
*init_config
;
888 struct si_pm4_state
*init_config_gs_rings
;
889 bool init_config_has_vgt_flush
;
890 struct si_pm4_state
*vgt_shader_config
[4];
893 struct si_shader_ctx_state ps_shader
;
894 struct si_shader_ctx_state gs_shader
;
895 struct si_shader_ctx_state vs_shader
;
896 struct si_shader_ctx_state tcs_shader
;
897 struct si_shader_ctx_state tes_shader
;
898 struct si_cs_shader_state cs_shader_state
;
900 /* shader information */
901 struct si_vertex_elements
*vertex_elements
;
902 unsigned sprite_coord_enable
;
903 unsigned cs_max_waves_per_sh
;
905 bool do_update_shaders
;
907 /* vertex buffer descriptors */
908 uint32_t *vb_descriptors_gpu_list
;
909 struct si_resource
*vb_descriptors_buffer
;
910 unsigned vb_descriptors_offset
;
912 /* shader descriptors */
913 struct si_descriptors descriptors
[SI_NUM_DESCS
];
914 unsigned descriptors_dirty
;
915 unsigned shader_pointers_dirty
;
916 unsigned shader_needs_decompress_mask
;
917 struct si_buffer_resources rw_buffers
;
918 struct si_buffer_resources const_and_shader_buffers
[SI_NUM_SHADERS
];
919 struct si_samplers samplers
[SI_NUM_SHADERS
];
920 struct si_images images
[SI_NUM_SHADERS
];
921 bool bo_list_add_all_resident_resources
;
922 bool bo_list_add_all_gfx_resources
;
923 bool bo_list_add_all_compute_resources
;
925 /* other shader resources */
926 struct pipe_constant_buffer null_const_buf
; /* used for set_constant_buffer(NULL) on GFX7 */
927 struct pipe_resource
*esgs_ring
;
928 struct pipe_resource
*gsvs_ring
;
929 struct pipe_resource
*tess_rings
;
930 union pipe_color_union
*border_color_table
; /* in CPU memory, any endian */
931 struct si_resource
*border_color_buffer
;
932 union pipe_color_union
*border_color_map
; /* in VRAM (slow access), little endian */
933 unsigned border_color_count
;
934 unsigned num_vs_blit_sgprs
;
935 uint32_t vs_blit_sh_data
[SI_VS_BLIT_SGPRS_POS_TEXCOORD
];
936 uint32_t cs_user_data
[4];
938 /* Vertex and index buffers. */
939 bool vertex_buffers_dirty
;
940 bool vertex_buffer_pointer_dirty
;
941 struct pipe_vertex_buffer vertex_buffer
[SI_NUM_VERTEX_BUFFERS
];
942 uint16_t vertex_buffer_unaligned
; /* bitmask of not dword-aligned buffers */
944 /* MSAA config state. */
946 bool ps_uses_fbfetch
;
947 bool smoothing_enabled
;
949 /* DB render state. */
950 unsigned ps_db_shader_control
;
951 unsigned dbcb_copy_sample
;
952 bool dbcb_depth_copy_enabled
:1;
953 bool dbcb_stencil_copy_enabled
:1;
954 bool db_flush_depth_inplace
:1;
955 bool db_flush_stencil_inplace
:1;
956 bool db_depth_clear
:1;
957 bool db_depth_disable_expclear
:1;
958 bool db_stencil_clear
:1;
959 bool db_stencil_disable_expclear
:1;
960 bool occlusion_queries_disabled
:1;
961 bool generate_mipmap_for_depth
:1;
963 /* Emitted draw state. */
964 bool gs_tri_strip_adj_fix
:1;
967 int last_base_vertex
;
968 int last_start_instance
;
969 int last_instance_count
;
971 int last_sh_base_reg
;
972 int last_primitive_restart_en
;
973 int last_restart_index
;
975 int last_multi_vgt_param
;
977 unsigned last_sc_line_stipple
;
978 unsigned current_vs_state
;
979 unsigned last_vs_state
;
980 enum pipe_prim_type current_rast_prim
; /* primitive type after TES, GS */
983 struct si_resource
*scratch_buffer
;
984 unsigned scratch_waves
;
985 unsigned spi_tmpring_size
;
987 struct si_resource
*compute_scratch_buffer
;
989 /* Emitted derived tessellation state. */
990 /* Local shader (VS), or HS if LS-HS are merged. */
991 struct si_shader
*last_ls
;
992 struct si_shader_selector
*last_tcs
;
993 int last_num_tcs_input_cp
;
994 int last_tes_sh_base
;
995 bool last_tess_uses_primid
;
996 unsigned last_num_patches
;
997 int last_ls_hs_config
;
1001 struct si_saved_cs
*current_saved_cs
;
1002 uint64_t dmesg_timestamp
;
1003 unsigned apitrace_call_number
;
1006 bool need_check_render_feedback
;
1007 bool decompression_enabled
;
1008 bool dpbb_force_off
;
1009 bool vs_writes_viewport_index
;
1010 bool vs_disables_clipping_viewport
;
1012 /* Precomputed IA_MULTI_VGT_PARAM */
1013 union si_vgt_param_key ia_multi_vgt_param_key
;
1014 unsigned ia_multi_vgt_param
[SI_NUM_VGT_PARAM_STATES
];
1016 /* Bindless descriptors. */
1017 struct si_descriptors bindless_descriptors
;
1018 struct util_idalloc bindless_used_slots
;
1019 unsigned num_bindless_descriptors
;
1020 bool bindless_descriptors_dirty
;
1021 bool graphics_bindless_pointer_dirty
;
1022 bool compute_bindless_pointer_dirty
;
1024 /* Allocated bindless handles */
1025 struct hash_table
*tex_handles
;
1026 struct hash_table
*img_handles
;
1028 /* Resident bindless handles */
1029 struct util_dynarray resident_tex_handles
;
1030 struct util_dynarray resident_img_handles
;
1032 /* Resident bindless handles which need decompression */
1033 struct util_dynarray resident_tex_needs_color_decompress
;
1034 struct util_dynarray resident_img_needs_color_decompress
;
1035 struct util_dynarray resident_tex_needs_depth_decompress
;
1037 /* Bindless state */
1038 bool uses_bindless_samplers
;
1039 bool uses_bindless_images
;
1041 /* MSAA sample locations.
1042 * The first index is the sample index.
1043 * The second index is the coordinate: X, Y. */
1051 struct pipe_resource
*sample_pos_buffer
;
1054 unsigned num_draw_calls
;
1055 unsigned num_decompress_calls
;
1056 unsigned num_mrt_draw_calls
;
1057 unsigned num_prim_restart_calls
;
1058 unsigned num_spill_draw_calls
;
1059 unsigned num_compute_calls
;
1060 unsigned num_spill_compute_calls
;
1061 unsigned num_dma_calls
;
1062 unsigned num_cp_dma_calls
;
1063 unsigned num_vs_flushes
;
1064 unsigned num_ps_flushes
;
1065 unsigned num_cs_flushes
;
1066 unsigned num_cb_cache_flushes
;
1067 unsigned num_db_cache_flushes
;
1068 unsigned num_L2_invalidates
;
1069 unsigned num_L2_writebacks
;
1070 unsigned num_resident_handles
;
1071 uint64_t num_alloc_tex_transfer_bytes
;
1072 unsigned last_tex_ps_draw_ratio
; /* for query */
1073 unsigned context_roll
;
1076 /* Maintain the list of active queries for pausing between IBs. */
1077 int num_occlusion_queries
;
1078 int num_perfect_occlusion_queries
;
1079 struct list_head active_queries
;
1080 unsigned num_cs_dw_queries_suspend
;
1082 /* Render condition. */
1083 struct pipe_query
*render_cond
;
1084 unsigned render_cond_mode
;
1085 bool render_cond_invert
;
1086 bool render_cond_force_off
; /* for u_blitter */
1088 /* For uploading data via GTT and copy to VRAM on context flush via SDMA. */
1089 bool sdma_uploads_in_progress
;
1090 struct si_sdma_upload
*sdma_uploads
;
1091 unsigned num_sdma_uploads
;
1092 unsigned max_sdma_uploads
;
1094 /* Statistics gathering for the DCC enablement heuristic. It can't be
1095 * in si_texture because si_texture can be shared by multiple
1096 * contexts. This is for back buffers only. We shouldn't get too many
1099 * X11 DRI3 rotates among a finite set of back buffers. They should
1100 * all fit in this array. If they don't, separate DCC might never be
1101 * enabled by DCC stat gathering.
1104 struct si_texture
*tex
;
1105 /* Query queue: 0 = usually active, 1 = waiting, 2 = readback. */
1106 struct pipe_query
*ps_stats
[3];
1107 /* If all slots are used and another slot is needed,
1108 * the least recently used slot is evicted based on this. */
1109 int64_t last_use_timestamp
;
1113 /* Copy one resource to another using async DMA. */
1114 void (*dma_copy
)(struct pipe_context
*ctx
,
1115 struct pipe_resource
*dst
,
1117 unsigned dst_x
, unsigned dst_y
, unsigned dst_z
,
1118 struct pipe_resource
*src
,
1120 const struct pipe_box
*src_box
);
1122 struct si_tracked_regs tracked_regs
;
1126 void cik_init_sdma_functions(struct si_context
*sctx
);
1129 enum si_blitter_op
/* bitmask */
1131 SI_SAVE_TEXTURES
= 1,
1132 SI_SAVE_FRAMEBUFFER
= 2,
1133 SI_SAVE_FRAGMENT_STATE
= 4,
1134 SI_DISABLE_RENDER_COND
= 8,
1137 void si_blitter_begin(struct si_context
*sctx
, enum si_blitter_op op
);
1138 void si_blitter_end(struct si_context
*sctx
);
1139 void si_init_blit_functions(struct si_context
*sctx
);
1140 void si_decompress_textures(struct si_context
*sctx
, unsigned shader_mask
);
1141 void si_resource_copy_region(struct pipe_context
*ctx
,
1142 struct pipe_resource
*dst
,
1144 unsigned dstx
, unsigned dsty
, unsigned dstz
,
1145 struct pipe_resource
*src
,
1147 const struct pipe_box
*src_box
);
1148 void si_decompress_dcc(struct si_context
*sctx
, struct si_texture
*tex
);
1149 void si_blit_decompress_depth(struct pipe_context
*ctx
,
1150 struct si_texture
*texture
,
1151 struct si_texture
*staging
,
1152 unsigned first_level
, unsigned last_level
,
1153 unsigned first_layer
, unsigned last_layer
,
1154 unsigned first_sample
, unsigned last_sample
);
1157 bool si_rings_is_buffer_referenced(struct si_context
*sctx
,
1158 struct pb_buffer
*buf
,
1159 enum radeon_bo_usage usage
);
1160 void *si_buffer_map_sync_with_rings(struct si_context
*sctx
,
1161 struct si_resource
*resource
,
1163 void si_init_resource_fields(struct si_screen
*sscreen
,
1164 struct si_resource
*res
,
1165 uint64_t size
, unsigned alignment
);
1166 bool si_alloc_resource(struct si_screen
*sscreen
,
1167 struct si_resource
*res
);
1168 struct pipe_resource
*pipe_aligned_buffer_create(struct pipe_screen
*screen
,
1169 unsigned flags
, unsigned usage
,
1170 unsigned size
, unsigned alignment
);
1171 struct si_resource
*si_aligned_buffer_create(struct pipe_screen
*screen
,
1172 unsigned flags
, unsigned usage
,
1173 unsigned size
, unsigned alignment
);
1174 void si_replace_buffer_storage(struct pipe_context
*ctx
,
1175 struct pipe_resource
*dst
,
1176 struct pipe_resource
*src
);
1177 void si_init_screen_buffer_functions(struct si_screen
*sscreen
);
1178 void si_init_buffer_functions(struct si_context
*sctx
);
1181 enum pipe_format
si_simplify_cb_format(enum pipe_format format
);
1182 bool vi_alpha_is_on_msb(enum pipe_format format
);
1183 void vi_dcc_clear_level(struct si_context
*sctx
,
1184 struct si_texture
*tex
,
1185 unsigned level
, unsigned clear_value
);
1186 void si_init_clear_functions(struct si_context
*sctx
);
1188 /* si_compute_blit.c */
1189 unsigned si_get_flush_flags(struct si_context
*sctx
, enum si_coherency coher
,
1190 enum si_cache_policy cache_policy
);
1191 void si_clear_buffer(struct si_context
*sctx
, struct pipe_resource
*dst
,
1192 uint64_t offset
, uint64_t size
, uint32_t *clear_value
,
1193 uint32_t clear_value_size
, enum si_coherency coher
,
1195 void si_copy_buffer(struct si_context
*sctx
,
1196 struct pipe_resource
*dst
, struct pipe_resource
*src
,
1197 uint64_t dst_offset
, uint64_t src_offset
, unsigned size
);
1198 void si_compute_copy_image(struct si_context
*sctx
,
1199 struct pipe_resource
*dst
,
1201 struct pipe_resource
*src
,
1203 unsigned dstx
, unsigned dsty
, unsigned dstz
,
1204 const struct pipe_box
*src_box
);
1205 void si_compute_clear_render_target(struct pipe_context
*ctx
,
1206 struct pipe_surface
*dstsurf
,
1207 const union pipe_color_union
*color
,
1208 unsigned dstx
, unsigned dsty
,
1209 unsigned width
, unsigned height
,
1210 bool render_condition_enabled
);
1211 void si_retile_dcc(struct si_context
*sctx
, struct si_texture
*tex
);
1212 void si_init_compute_blit_functions(struct si_context
*sctx
);
1215 #define SI_CPDMA_SKIP_CHECK_CS_SPACE (1 << 0) /* don't call need_cs_space */
1216 #define SI_CPDMA_SKIP_SYNC_AFTER (1 << 1) /* don't wait for DMA after the copy */
1217 #define SI_CPDMA_SKIP_SYNC_BEFORE (1 << 2) /* don't wait for DMA before the copy (RAW hazards) */
1218 #define SI_CPDMA_SKIP_GFX_SYNC (1 << 3) /* don't flush caches and don't wait for PS/CS */
1219 #define SI_CPDMA_SKIP_BO_LIST_UPDATE (1 << 4) /* don't update the BO list */
1220 #define SI_CPDMA_SKIP_ALL (SI_CPDMA_SKIP_CHECK_CS_SPACE | \
1221 SI_CPDMA_SKIP_SYNC_AFTER | \
1222 SI_CPDMA_SKIP_SYNC_BEFORE | \
1223 SI_CPDMA_SKIP_GFX_SYNC | \
1224 SI_CPDMA_SKIP_BO_LIST_UPDATE)
1226 void si_cp_dma_wait_for_idle(struct si_context
*sctx
);
1227 void si_cp_dma_clear_buffer(struct si_context
*sctx
, struct radeon_cmdbuf
*cs
,
1228 struct pipe_resource
*dst
, uint64_t offset
,
1229 uint64_t size
, unsigned value
, unsigned user_flags
,
1230 enum si_coherency coher
, enum si_cache_policy cache_policy
);
1231 void si_cp_dma_copy_buffer(struct si_context
*sctx
,
1232 struct pipe_resource
*dst
, struct pipe_resource
*src
,
1233 uint64_t dst_offset
, uint64_t src_offset
, unsigned size
,
1234 unsigned user_flags
, enum si_coherency coher
,
1235 enum si_cache_policy cache_policy
);
1236 void cik_prefetch_TC_L2_async(struct si_context
*sctx
, struct pipe_resource
*buf
,
1237 uint64_t offset
, unsigned size
);
1238 void cik_emit_prefetch_L2(struct si_context
*sctx
, bool vertex_stage_only
);
1239 void si_test_gds(struct si_context
*sctx
);
1240 void si_cp_write_data(struct si_context
*sctx
, struct si_resource
*buf
,
1241 unsigned offset
, unsigned size
, unsigned dst_sel
,
1242 unsigned engine
, const void *data
);
1243 void si_cp_copy_data(struct si_context
*sctx
,
1244 unsigned dst_sel
, struct si_resource
*dst
, unsigned dst_offset
,
1245 unsigned src_sel
, struct si_resource
*src
, unsigned src_offset
);
1248 void si_save_cs(struct radeon_winsys
*ws
, struct radeon_cmdbuf
*cs
,
1249 struct radeon_saved_cs
*saved
, bool get_buffer_list
);
1250 void si_clear_saved_cs(struct radeon_saved_cs
*saved
);
1251 void si_destroy_saved_cs(struct si_saved_cs
*scs
);
1252 void si_auto_log_cs(void *data
, struct u_log_context
*log
);
1253 void si_log_hw_flush(struct si_context
*sctx
);
1254 void si_log_draw_state(struct si_context
*sctx
, struct u_log_context
*log
);
1255 void si_log_compute_state(struct si_context
*sctx
, struct u_log_context
*log
);
1256 void si_init_debug_functions(struct si_context
*sctx
);
1257 void si_check_vm_faults(struct si_context
*sctx
,
1258 struct radeon_saved_cs
*saved
, enum ring_type ring
);
1259 bool si_replace_shader(unsigned num
, struct ac_shader_binary
*binary
);
1262 void si_init_dma_functions(struct si_context
*sctx
);
1265 void si_dma_emit_timestamp(struct si_context
*sctx
, struct si_resource
*dst
,
1267 void si_sdma_clear_buffer(struct si_context
*sctx
, struct pipe_resource
*dst
,
1268 uint64_t offset
, uint64_t size
, unsigned clear_value
);
1269 void si_need_dma_space(struct si_context
*ctx
, unsigned num_dw
,
1270 struct si_resource
*dst
, struct si_resource
*src
);
1271 void si_flush_dma_cs(struct si_context
*ctx
, unsigned flags
,
1272 struct pipe_fence_handle
**fence
);
1273 void si_screen_clear_buffer(struct si_screen
*sscreen
, struct pipe_resource
*dst
,
1274 uint64_t offset
, uint64_t size
, unsigned value
);
1277 void si_cp_release_mem(struct si_context
*ctx
,
1278 unsigned event
, unsigned event_flags
,
1279 unsigned dst_sel
, unsigned int_sel
, unsigned data_sel
,
1280 struct si_resource
*buf
, uint64_t va
,
1281 uint32_t new_fence
, unsigned query_type
);
1282 unsigned si_cp_write_fence_dwords(struct si_screen
*screen
);
1283 void si_cp_wait_mem(struct si_context
*ctx
, struct radeon_cmdbuf
*cs
,
1284 uint64_t va
, uint32_t ref
, uint32_t mask
, unsigned flags
);
1285 void si_init_fence_functions(struct si_context
*ctx
);
1286 void si_init_screen_fence_functions(struct si_screen
*screen
);
1287 struct pipe_fence_handle
*si_create_fence(struct pipe_context
*ctx
,
1288 struct tc_unflushed_batch_token
*tc_token
);
1291 void si_init_screen_get_functions(struct si_screen
*sscreen
);
1294 void si_flush_gfx_cs(struct si_context
*ctx
, unsigned flags
,
1295 struct pipe_fence_handle
**fence
);
1296 void si_begin_new_gfx_cs(struct si_context
*ctx
);
1297 void si_need_gfx_cs_space(struct si_context
*ctx
);
1298 void si_unref_sdma_uploads(struct si_context
*sctx
);
1301 void si_gpu_load_kill_thread(struct si_screen
*sscreen
);
1302 uint64_t si_begin_counter(struct si_screen
*sscreen
, unsigned type
);
1303 unsigned si_end_counter(struct si_screen
*sscreen
, unsigned type
,
1307 void si_emit_initial_compute_regs(struct si_context
*sctx
, struct radeon_cmdbuf
*cs
);
1308 unsigned si_get_compute_resource_limits(struct si_screen
*sscreen
,
1309 unsigned waves_per_threadgroup
,
1310 unsigned max_waves_per_sh
,
1311 unsigned threadgroups_per_cu
);
1312 void si_init_compute_functions(struct si_context
*sctx
);
1314 /* si_perfcounters.c */
1315 void si_init_perfcounters(struct si_screen
*screen
);
1316 void si_destroy_perfcounters(struct si_screen
*screen
);
1319 bool si_check_device_reset(struct si_context
*sctx
);
1322 void si_init_screen_query_functions(struct si_screen
*sscreen
);
1323 void si_init_query_functions(struct si_context
*sctx
);
1324 void si_suspend_queries(struct si_context
*sctx
);
1325 void si_resume_queries(struct si_context
*sctx
);
1327 /* si_shaderlib_tgsi.c */
1328 void *si_get_blitter_vs(struct si_context
*sctx
, enum blitter_attrib_type type
,
1329 unsigned num_layers
);
1330 void *si_create_fixed_func_tcs(struct si_context
*sctx
);
1331 void *si_create_dma_compute_shader(struct pipe_context
*ctx
,
1332 unsigned num_dwords_per_thread
,
1333 bool dst_stream_cache_policy
, bool is_copy
);
1334 void *si_create_copy_image_compute_shader(struct pipe_context
*ctx
);
1335 void *si_create_copy_image_compute_shader_1d_array(struct pipe_context
*ctx
);
1336 void *si_clear_render_target_shader(struct pipe_context
*ctx
);
1337 void *si_clear_render_target_shader_1d_array(struct pipe_context
*ctx
);
1338 void *si_create_dcc_retile_cs(struct pipe_context
*ctx
);
1339 void *si_create_query_result_cs(struct si_context
*sctx
);
1342 void si_test_dma(struct si_screen
*sscreen
);
1344 /* si_test_clearbuffer.c */
1345 void si_test_dma_perf(struct si_screen
*sscreen
);
1348 struct pipe_video_codec
*si_uvd_create_decoder(struct pipe_context
*context
,
1349 const struct pipe_video_codec
*templ
);
1351 struct pipe_video_buffer
*si_video_buffer_create(struct pipe_context
*pipe
,
1352 const struct pipe_video_buffer
*tmpl
);
1355 void si_update_vs_viewport_state(struct si_context
*ctx
);
1356 void si_init_viewport_functions(struct si_context
*ctx
);
1359 bool si_prepare_for_dma_blit(struct si_context
*sctx
,
1360 struct si_texture
*dst
,
1361 unsigned dst_level
, unsigned dstx
,
1362 unsigned dsty
, unsigned dstz
,
1363 struct si_texture
*src
,
1365 const struct pipe_box
*src_box
);
1366 void si_eliminate_fast_color_clear(struct si_context
*sctx
,
1367 struct si_texture
*tex
);
1368 void si_texture_discard_cmask(struct si_screen
*sscreen
,
1369 struct si_texture
*tex
);
1370 bool si_init_flushed_depth_texture(struct pipe_context
*ctx
,
1371 struct pipe_resource
*texture
,
1372 struct si_texture
**staging
);
1373 void si_print_texture_info(struct si_screen
*sscreen
,
1374 struct si_texture
*tex
, struct u_log_context
*log
);
1375 struct pipe_resource
*si_texture_create(struct pipe_screen
*screen
,
1376 const struct pipe_resource
*templ
);
1377 bool vi_dcc_formats_compatible(enum pipe_format format1
,
1378 enum pipe_format format2
);
1379 bool vi_dcc_formats_are_incompatible(struct pipe_resource
*tex
,
1381 enum pipe_format view_format
);
1382 void vi_disable_dcc_if_incompatible_format(struct si_context
*sctx
,
1383 struct pipe_resource
*tex
,
1385 enum pipe_format view_format
);
1386 struct pipe_surface
*si_create_surface_custom(struct pipe_context
*pipe
,
1387 struct pipe_resource
*texture
,
1388 const struct pipe_surface
*templ
,
1389 unsigned width0
, unsigned height0
,
1390 unsigned width
, unsigned height
);
1391 unsigned si_translate_colorswap(enum pipe_format format
, bool do_endian_swap
);
1392 void vi_separate_dcc_try_enable(struct si_context
*sctx
,
1393 struct si_texture
*tex
);
1394 void vi_separate_dcc_start_query(struct si_context
*sctx
,
1395 struct si_texture
*tex
);
1396 void vi_separate_dcc_stop_query(struct si_context
*sctx
,
1397 struct si_texture
*tex
);
1398 void vi_separate_dcc_process_and_reset_stats(struct pipe_context
*ctx
,
1399 struct si_texture
*tex
);
1400 bool si_texture_disable_dcc(struct si_context
*sctx
,
1401 struct si_texture
*tex
);
1402 void si_init_screen_texture_functions(struct si_screen
*sscreen
);
1403 void si_init_context_texture_functions(struct si_context
*sctx
);
1410 static inline struct si_resource
*si_resource(struct pipe_resource
*r
)
1412 return (struct si_resource
*)r
;
1416 si_resource_reference(struct si_resource
**ptr
, struct si_resource
*res
)
1418 pipe_resource_reference((struct pipe_resource
**)ptr
,
1419 (struct pipe_resource
*)res
);
1423 si_texture_reference(struct si_texture
**ptr
, struct si_texture
*res
)
1425 pipe_resource_reference((struct pipe_resource
**)ptr
, &res
->buffer
.b
.b
);
1429 vi_dcc_enabled(struct si_texture
*tex
, unsigned level
)
1431 return tex
->dcc_offset
&& level
< tex
->surface
.num_dcc_levels
;
1434 static inline unsigned
1435 si_tile_mode_index(struct si_texture
*tex
, unsigned level
, bool stencil
)
1438 return tex
->surface
.u
.legacy
.stencil_tiling_index
[level
];
1440 return tex
->surface
.u
.legacy
.tiling_index
[level
];
1443 static inline unsigned
1444 si_get_minimum_num_gfx_cs_dwords(struct si_context
*sctx
)
1446 /* Don't count the needed CS space exactly and just use an upper bound.
1448 * Also reserve space for stopping queries at the end of IB, because
1449 * the number of active queries is unlimited in theory.
1451 return 2048 + sctx
->num_cs_dw_queries_suspend
;
1455 si_context_add_resource_size(struct si_context
*sctx
, struct pipe_resource
*r
)
1458 /* Add memory usage for need_gfx_cs_space */
1459 sctx
->vram
+= si_resource(r
)->vram_usage
;
1460 sctx
->gtt
+= si_resource(r
)->gart_usage
;
1465 si_invalidate_draw_sh_constants(struct si_context
*sctx
)
1467 sctx
->last_base_vertex
= SI_BASE_VERTEX_UNKNOWN
;
1468 sctx
->last_instance_count
= SI_INSTANCE_COUNT_UNKNOWN
;
1471 static inline unsigned
1472 si_get_atom_bit(struct si_context
*sctx
, struct si_atom
*atom
)
1474 return 1 << (atom
- sctx
->atoms
.array
);
1478 si_set_atom_dirty(struct si_context
*sctx
, struct si_atom
*atom
, bool dirty
)
1480 unsigned bit
= si_get_atom_bit(sctx
, atom
);
1483 sctx
->dirty_atoms
|= bit
;
1485 sctx
->dirty_atoms
&= ~bit
;
1489 si_is_atom_dirty(struct si_context
*sctx
, struct si_atom
*atom
)
1491 return (sctx
->dirty_atoms
& si_get_atom_bit(sctx
, atom
)) != 0;
1495 si_mark_atom_dirty(struct si_context
*sctx
, struct si_atom
*atom
)
1497 si_set_atom_dirty(sctx
, atom
, true);
1500 static inline struct si_shader_ctx_state
*si_get_vs(struct si_context
*sctx
)
1502 if (sctx
->gs_shader
.cso
)
1503 return &sctx
->gs_shader
;
1504 if (sctx
->tes_shader
.cso
)
1505 return &sctx
->tes_shader
;
1507 return &sctx
->vs_shader
;
1510 static inline struct tgsi_shader_info
*si_get_vs_info(struct si_context
*sctx
)
1512 struct si_shader_ctx_state
*vs
= si_get_vs(sctx
);
1514 return vs
->cso
? &vs
->cso
->info
: NULL
;
1517 static inline struct si_shader
* si_get_vs_state(struct si_context
*sctx
)
1519 if (sctx
->gs_shader
.cso
)
1520 return sctx
->gs_shader
.cso
->gs_copy_shader
;
1522 struct si_shader_ctx_state
*vs
= si_get_vs(sctx
);
1523 return vs
->current
? vs
->current
: NULL
;
1526 static inline bool si_can_dump_shader(struct si_screen
*sscreen
,
1529 return sscreen
->debug_flags
& (1 << processor
);
1532 static inline bool si_get_strmout_en(struct si_context
*sctx
)
1534 return sctx
->streamout
.streamout_enabled
||
1535 sctx
->streamout
.prims_gen_query_enabled
;
1538 static inline unsigned
1539 si_optimal_tcc_alignment(struct si_context
*sctx
, unsigned upload_size
)
1541 unsigned alignment
, tcc_cache_line_size
;
1543 /* If the upload size is less than the cache line size (e.g. 16, 32),
1544 * the whole thing will fit into a cache line if we align it to its size.
1545 * The idea is that multiple small uploads can share a cache line.
1546 * If the upload size is greater, align it to the cache line size.
1548 alignment
= util_next_power_of_two(upload_size
);
1549 tcc_cache_line_size
= sctx
->screen
->info
.tcc_cache_line_size
;
1550 return MIN2(alignment
, tcc_cache_line_size
);
1554 si_saved_cs_reference(struct si_saved_cs
**dst
, struct si_saved_cs
*src
)
1556 if (pipe_reference(&(*dst
)->reference
, &src
->reference
))
1557 si_destroy_saved_cs(*dst
);
1563 si_make_CB_shader_coherent(struct si_context
*sctx
, unsigned num_samples
,
1564 bool shaders_read_metadata
, bool dcc_pipe_aligned
)
1566 sctx
->flags
|= SI_CONTEXT_FLUSH_AND_INV_CB
|
1567 SI_CONTEXT_INV_VMEM_L1
;
1569 if (sctx
->chip_class
>= GFX9
) {
1570 /* Single-sample color is coherent with shaders on GFX9, but
1571 * L2 metadata must be flushed if shaders read metadata.
1574 if (num_samples
>= 2 ||
1575 (shaders_read_metadata
&& !dcc_pipe_aligned
))
1576 sctx
->flags
|= SI_CONTEXT_INV_GLOBAL_L2
;
1577 else if (shaders_read_metadata
)
1578 sctx
->flags
|= SI_CONTEXT_INV_L2_METADATA
;
1581 sctx
->flags
|= SI_CONTEXT_INV_GLOBAL_L2
;
1586 si_make_DB_shader_coherent(struct si_context
*sctx
, unsigned num_samples
,
1587 bool include_stencil
, bool shaders_read_metadata
)
1589 sctx
->flags
|= SI_CONTEXT_FLUSH_AND_INV_DB
|
1590 SI_CONTEXT_INV_VMEM_L1
;
1592 if (sctx
->chip_class
>= GFX9
) {
1593 /* Single-sample depth (not stencil) is coherent with shaders
1594 * on GFX9, but L2 metadata must be flushed if shaders read
1597 if (num_samples
>= 2 || include_stencil
)
1598 sctx
->flags
|= SI_CONTEXT_INV_GLOBAL_L2
;
1599 else if (shaders_read_metadata
)
1600 sctx
->flags
|= SI_CONTEXT_INV_L2_METADATA
;
1603 sctx
->flags
|= SI_CONTEXT_INV_GLOBAL_L2
;
1608 si_can_sample_zs(struct si_texture
*tex
, bool stencil_sampler
)
1610 return (stencil_sampler
&& tex
->can_sample_s
) ||
1611 (!stencil_sampler
&& tex
->can_sample_z
);
1615 si_htile_enabled(struct si_texture
*tex
, unsigned level
)
1617 return tex
->htile_offset
&& level
== 0;
1621 vi_tc_compat_htile_enabled(struct si_texture
*tex
, unsigned level
)
1623 assert(!tex
->tc_compatible_htile
|| tex
->htile_offset
);
1624 return tex
->tc_compatible_htile
&& level
== 0;
1627 static inline unsigned si_get_ps_iter_samples(struct si_context
*sctx
)
1629 if (sctx
->ps_uses_fbfetch
)
1630 return sctx
->framebuffer
.nr_color_samples
;
1632 return MIN2(sctx
->ps_iter_samples
, sctx
->framebuffer
.nr_color_samples
);
1635 static inline unsigned si_get_total_colormask(struct si_context
*sctx
)
1637 if (sctx
->queued
.named
.rasterizer
->rasterizer_discard
)
1640 struct si_shader_selector
*ps
= sctx
->ps_shader
.cso
;
1644 unsigned colormask
= sctx
->framebuffer
.colorbuf_enabled_4bit
&
1645 sctx
->queued
.named
.blend
->cb_target_mask
;
1647 if (!ps
->info
.properties
[TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS
])
1648 colormask
&= ps
->colors_written_4bit
;
1649 else if (!ps
->colors_written_4bit
)
1650 colormask
= 0; /* color0 writes all cbufs, but it's not written */
1655 #define UTIL_ALL_PRIM_LINE_MODES ((1 << PIPE_PRIM_LINES) | \
1656 (1 << PIPE_PRIM_LINE_LOOP) | \
1657 (1 << PIPE_PRIM_LINE_STRIP) | \
1658 (1 << PIPE_PRIM_LINES_ADJACENCY) | \
1659 (1 << PIPE_PRIM_LINE_STRIP_ADJACENCY))
1661 static inline bool util_prim_is_lines(unsigned prim
)
1663 return ((1 << prim
) & UTIL_ALL_PRIM_LINE_MODES
) != 0;
1666 static inline bool util_prim_is_points_or_lines(unsigned prim
)
1668 return ((1 << prim
) & (UTIL_ALL_PRIM_LINE_MODES
|
1669 (1 << PIPE_PRIM_POINTS
))) != 0;
1673 * Return true if there is enough memory in VRAM and GTT for the buffers
1676 * \param vram VRAM memory size not added to the buffer list yet
1677 * \param gtt GTT memory size not added to the buffer list yet
1680 radeon_cs_memory_below_limit(struct si_screen
*screen
,
1681 struct radeon_cmdbuf
*cs
,
1682 uint64_t vram
, uint64_t gtt
)
1684 vram
+= cs
->used_vram
;
1685 gtt
+= cs
->used_gart
;
1687 /* Anything that goes above the VRAM size should go to GTT. */
1688 if (vram
> screen
->info
.vram_size
)
1689 gtt
+= vram
- screen
->info
.vram_size
;
1691 /* Now we just need to check if we have enough GTT. */
1692 return gtt
< screen
->info
.gart_size
* 0.7;
1696 * Add a buffer to the buffer list for the given command stream (CS).
1698 * All buffers used by a CS must be added to the list. This tells the kernel
1699 * driver which buffers are used by GPU commands. Other buffers can
1700 * be swapped out (not accessible) during execution.
1702 * The buffer list becomes empty after every context flush and must be
1705 static inline void radeon_add_to_buffer_list(struct si_context
*sctx
,
1706 struct radeon_cmdbuf
*cs
,
1707 struct si_resource
*bo
,
1708 enum radeon_bo_usage usage
,
1709 enum radeon_bo_priority priority
)
1712 sctx
->ws
->cs_add_buffer(
1714 (enum radeon_bo_usage
)(usage
| RADEON_USAGE_SYNCHRONIZED
),
1715 bo
->domains
, priority
);
1719 * Same as above, but also checks memory usage and flushes the context
1722 * When this SHOULD NOT be used:
1724 * - if si_context_add_resource_size has been called for the buffer
1725 * followed by *_need_cs_space for checking the memory usage
1727 * - if si_need_dma_space has been called for the buffer
1729 * - when emitting state packets and draw packets (because preceding packets
1730 * can't be re-emitted at that point)
1732 * - if shader resource "enabled_mask" is not up-to-date or there is
1733 * a different constraint disallowing a context flush
1736 radeon_add_to_gfx_buffer_list_check_mem(struct si_context
*sctx
,
1737 struct si_resource
*bo
,
1738 enum radeon_bo_usage usage
,
1739 enum radeon_bo_priority priority
,
1743 !radeon_cs_memory_below_limit(sctx
->screen
, sctx
->gfx_cs
,
1744 sctx
->vram
+ bo
->vram_usage
,
1745 sctx
->gtt
+ bo
->gart_usage
))
1746 si_flush_gfx_cs(sctx
, RADEON_FLUSH_ASYNC_START_NEXT_GFX_IB_NOW
, NULL
);
1748 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
, bo
, usage
, priority
);
1751 #define PRINT_ERR(fmt, args...) \
1752 fprintf(stderr, "EE %s:%d %s - " fmt, __FILE__, __LINE__, __func__, ##args)