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)
62 /* Framebuffer caches. */
63 #define SI_CONTEXT_FLUSH_AND_INV_DB (R600_CONTEXT_PRIVATE_FLAG << 7)
64 #define SI_CONTEXT_FLUSH_AND_INV_CB (R600_CONTEXT_PRIVATE_FLAG << 8)
65 /* Engine synchronization. */
66 #define SI_CONTEXT_VS_PARTIAL_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 9)
67 #define SI_CONTEXT_PS_PARTIAL_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 10)
68 #define SI_CONTEXT_CS_PARTIAL_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 11)
69 #define SI_CONTEXT_VGT_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 12)
70 #define SI_CONTEXT_VGT_STREAMOUT_SYNC (R600_CONTEXT_PRIVATE_FLAG << 13)
72 #define SI_PREFETCH_VBO_DESCRIPTORS (1 << 0)
73 #define SI_PREFETCH_LS (1 << 1)
74 #define SI_PREFETCH_HS (1 << 2)
75 #define SI_PREFETCH_ES (1 << 3)
76 #define SI_PREFETCH_GS (1 << 4)
77 #define SI_PREFETCH_VS (1 << 5)
78 #define SI_PREFETCH_PS (1 << 6)
80 #define SI_MAX_BORDER_COLORS 4096
85 struct u_suballocator
;
88 struct r600_common_screen b
;
89 unsigned gs_table_depth
;
90 unsigned tess_offchip_block_dw_size
;
92 bool has_distributed_tess
;
93 bool has_draw_indirect_multi
;
95 bool has_msaa_sample_loc_bug
;
96 bool llvm_has_working_vgpr_indexing
;
98 /* Whether shaders are monolithic (1-part) or separate (3-part). */
99 bool use_monolithic_shaders
;
102 mtx_t shader_parts_mutex
;
103 struct si_shader_part
*vs_prologs
;
104 struct si_shader_part
*tcs_epilogs
;
105 struct si_shader_part
*gs_prologs
;
106 struct si_shader_part
*ps_prologs
;
107 struct si_shader_part
*ps_epilogs
;
109 /* Shader cache in memory.
111 * Design & limitations:
112 * - The shader cache is per screen (= per process), never saved to
113 * disk, and skips redundant shader compilations from TGSI to bytecode.
114 * - It can only be used with one-variant-per-shader support, in which
115 * case only the main (typically middle) part of shaders is cached.
116 * - Only VS, TCS, TES, PS are cached, out of which only the hw VS
117 * variants of VS and TES are cached, so LS and ES aren't.
118 * - GS and CS aren't cached, but it's certainly possible to cache
121 mtx_t shader_cache_mutex
;
122 struct hash_table
*shader_cache
;
124 /* Shader compiler queue for multithreaded compilation. */
125 struct util_queue shader_compiler_queue
;
126 /* Use at most 3 normal compiler threads on quadcore and better.
127 * Hyperthreaded CPUs report the number of threads, but we want
128 * the number of cores. */
129 LLVMTargetMachineRef tm
[3]; /* used by the queue only */
131 struct util_queue shader_compiler_queue_low_priority
;
132 /* Use at most 2 low priority threads on quadcore and better.
133 * We want to minimize the impact on multithreaded Mesa. */
134 LLVMTargetMachineRef tm_low_priority
[2]; /* at most 2 threads */
137 struct si_blend_color
{
138 struct r600_atom atom
;
139 struct pipe_blend_color state
;
143 struct si_sampler_view
{
144 struct pipe_sampler_view base
;
145 /* [0..7] = image descriptor
146 * [4..7] = buffer descriptor */
148 uint32_t fmask_state
[8];
149 const struct legacy_surf_level
*base_level_info
;
152 bool is_stencil_sampler
;
153 bool dcc_incompatible
;
156 #define SI_SAMPLER_STATE_MAGIC 0x34f1c35a
158 struct si_sampler_state
{
165 struct si_cs_shader_state
{
166 struct si_compute
*program
;
167 struct si_compute
*emitted_program
;
173 struct si_textures_info
{
174 struct si_sampler_views views
;
175 uint32_t needs_depth_decompress_mask
;
176 uint32_t needs_color_decompress_mask
;
179 struct si_images_info
{
180 struct pipe_image_view views
[SI_NUM_IMAGES
];
181 uint32_t needs_color_decompress_mask
;
182 unsigned enabled_mask
;
185 struct si_framebuffer
{
186 struct r600_atom atom
;
187 struct pipe_framebuffer_state state
;
188 unsigned colorbuf_enabled_4bit
;
189 unsigned spi_shader_col_format
;
190 unsigned spi_shader_col_format_alpha
;
191 unsigned spi_shader_col_format_blend
;
192 unsigned spi_shader_col_format_blend_alpha
;
193 ubyte nr_samples
:5; /* at most 16xAA */
194 ubyte log_samples
:3; /* at most 4 = 16xAA */
195 ubyte compressed_cb_mask
;
197 ubyte color_is_int10
;
203 struct si_clip_state
{
204 struct r600_atom atom
;
205 struct pipe_clip_state state
;
209 struct si_sample_locs
{
210 struct r600_atom atom
;
214 struct si_sample_mask
{
215 struct r600_atom atom
;
216 uint16_t sample_mask
;
219 /* A shader state consists of the shader selector, which is a constant state
220 * object shared by multiple contexts and shouldn't be modified, and
221 * the current shader variant selected for this context.
223 struct si_shader_ctx_state
{
224 struct si_shader_selector
*cso
;
225 struct si_shader
*current
;
228 #define SI_NUM_VGT_PARAM_KEY_BITS 12
229 #define SI_NUM_VGT_PARAM_STATES (1 << SI_NUM_VGT_PARAM_KEY_BITS)
231 /* The IA_MULTI_VGT_PARAM key used to index the table of precomputed values.
232 * Some fields are set by state-change calls, most are set by draw_vbo.
234 union si_vgt_param_key
{
237 unsigned uses_instancing
:1;
238 unsigned multi_instances_smaller_than_primgroup
:1;
239 unsigned primitive_restart
:1;
240 unsigned count_from_stream_output
:1;
241 unsigned line_stipple_enabled
:1;
242 unsigned uses_tess
:1;
243 unsigned tess_uses_prim_id
:1;
245 unsigned _pad
:32 - SI_NUM_VGT_PARAM_KEY_BITS
;
250 struct si_texture_handle
254 struct pipe_sampler_view
*view
;
255 struct si_sampler_state sstate
;
258 struct si_image_handle
262 struct pipe_image_view view
;
266 struct pipe_reference reference
;
267 struct si_context
*ctx
;
268 struct radeon_saved_cs gfx
;
269 struct r600_resource
*trace_buf
;
272 unsigned gfx_last_dw
;
277 struct r600_common_context b
;
278 struct blitter_context
*blitter
;
279 void *custom_dsa_flush
;
280 void *custom_blend_resolve
;
281 void *custom_blend_fmask_decompress
;
282 void *custom_blend_eliminate_fastclear
;
283 void *custom_blend_dcc_decompress
;
284 struct si_screen
*screen
;
285 LLVMTargetMachineRef tm
; /* only non-threaded compilation */
286 struct si_shader_ctx_state fixed_func_tcs_shader
;
287 struct r600_resource
*wait_mem_scratch
;
288 unsigned wait_mem_number
;
289 uint16_t prefetch_L2_mask
;
291 bool gfx_flush_in_progress
:1;
292 bool compute_is_busy
:1;
294 /* Atoms (direct states). */
295 union si_state_atoms atoms
;
296 unsigned dirty_atoms
; /* mask */
297 /* PM4 states (precomputed immutable states) */
298 unsigned dirty_states
;
299 union si_state queued
;
300 union si_state emitted
;
302 /* Atom declarations. */
303 struct si_framebuffer framebuffer
;
304 struct si_sample_locs msaa_sample_locs
;
305 struct r600_atom db_render_state
;
306 struct r600_atom msaa_config
;
307 struct si_sample_mask sample_mask
;
308 struct r600_atom cb_render_state
;
309 unsigned last_cb_target_mask
;
310 struct si_blend_color blend_color
;
311 struct r600_atom clip_regs
;
312 struct si_clip_state clip_state
;
313 struct si_shader_data shader_pointers
;
314 struct si_stencil_ref stencil_ref
;
315 struct r600_atom spi_map
;
317 /* Precomputed states. */
318 struct si_pm4_state
*init_config
;
319 struct si_pm4_state
*init_config_gs_rings
;
320 bool init_config_has_vgt_flush
;
321 struct si_pm4_state
*vgt_shader_config
[4];
324 struct si_shader_ctx_state ps_shader
;
325 struct si_shader_ctx_state gs_shader
;
326 struct si_shader_ctx_state vs_shader
;
327 struct si_shader_ctx_state tcs_shader
;
328 struct si_shader_ctx_state tes_shader
;
329 struct si_cs_shader_state cs_shader_state
;
331 /* shader information */
332 struct si_vertex_elements
*vertex_elements
;
333 unsigned sprite_coord_enable
;
335 bool do_update_shaders
;
337 /* shader descriptors */
338 struct si_descriptors vertex_buffers
;
339 struct si_descriptors descriptors
[SI_NUM_DESCS
];
340 unsigned descriptors_dirty
;
341 unsigned shader_pointers_dirty
;
342 unsigned shader_needs_decompress_mask
;
343 struct si_buffer_resources rw_buffers
;
344 struct si_buffer_resources const_and_shader_buffers
[SI_NUM_SHADERS
];
345 struct si_textures_info samplers
[SI_NUM_SHADERS
];
346 struct si_images_info images
[SI_NUM_SHADERS
];
348 /* other shader resources */
349 struct pipe_constant_buffer null_const_buf
; /* used for set_constant_buffer(NULL) on CIK */
350 struct pipe_resource
*esgs_ring
;
351 struct pipe_resource
*gsvs_ring
;
352 struct pipe_resource
*tf_ring
;
353 struct pipe_resource
*tess_offchip_ring
;
354 union pipe_color_union
*border_color_table
; /* in CPU memory, any endian */
355 struct r600_resource
*border_color_buffer
;
356 union pipe_color_union
*border_color_map
; /* in VRAM (slow access), little endian */
357 unsigned border_color_count
;
359 /* Vertex and index buffers. */
360 bool vertex_buffers_dirty
;
361 bool vertex_buffer_pointer_dirty
;
362 struct pipe_vertex_buffer vertex_buffer
[SI_NUM_VERTEX_BUFFERS
];
364 /* MSAA config state. */
366 bool smoothing_enabled
;
368 /* DB render state. */
369 unsigned ps_db_shader_control
;
370 unsigned dbcb_copy_sample
;
371 bool dbcb_depth_copy_enabled
:1;
372 bool dbcb_stencil_copy_enabled
:1;
373 bool db_flush_depth_inplace
:1;
374 bool db_flush_stencil_inplace
:1;
375 bool db_depth_clear
:1;
376 bool db_depth_disable_expclear
:1;
377 bool db_stencil_clear
:1;
378 bool db_stencil_disable_expclear
:1;
379 bool occlusion_queries_disabled
:1;
380 bool generate_mipmap_for_depth
:1;
382 /* Emitted draw state. */
383 bool gs_tri_strip_adj_fix
:1;
385 int last_base_vertex
;
386 int last_start_instance
;
388 int last_sh_base_reg
;
389 int last_primitive_restart_en
;
390 int last_restart_index
;
391 int last_gs_out_prim
;
393 int last_multi_vgt_param
;
395 unsigned last_sc_line_stipple
;
396 unsigned current_vs_state
;
397 unsigned last_vs_state
;
398 enum pipe_prim_type current_rast_prim
; /* primitive type after TES, GS */
401 struct r600_atom scratch_state
;
402 struct r600_resource
*scratch_buffer
;
403 unsigned scratch_waves
;
404 unsigned spi_tmpring_size
;
406 struct r600_resource
*compute_scratch_buffer
;
408 /* Emitted derived tessellation state. */
409 /* Local shader (VS), or HS if LS-HS are merged. */
410 struct si_shader
*last_ls
;
411 struct si_shader_selector
*last_tcs
;
412 int last_num_tcs_input_cp
;
413 int last_tes_sh_base
;
414 bool last_tess_uses_primid
;
415 unsigned last_num_patches
;
419 struct si_saved_cs
*current_saved_cs
;
420 uint64_t dmesg_timestamp
;
421 unsigned apitrace_call_number
;
424 bool need_check_render_feedback
;
425 bool decompression_enabled
;
427 /* Precomputed IA_MULTI_VGT_PARAM */
428 union si_vgt_param_key ia_multi_vgt_param_key
;
429 unsigned ia_multi_vgt_param
[SI_NUM_VGT_PARAM_STATES
];
431 /* Bindless descriptors. */
432 struct si_descriptors bindless_descriptors
;
433 struct util_idalloc bindless_used_slots
;
434 unsigned num_bindless_descriptors
;
435 bool bindless_descriptors_dirty
;
436 bool graphics_bindless_pointer_dirty
;
437 bool compute_bindless_pointer_dirty
;
439 /* Allocated bindless handles */
440 struct hash_table
*tex_handles
;
441 struct hash_table
*img_handles
;
443 /* Resident bindless handles */
444 struct util_dynarray resident_tex_handles
;
445 struct util_dynarray resident_img_handles
;
447 /* Resident bindless handles which need decompression */
448 struct util_dynarray resident_tex_needs_color_decompress
;
449 struct util_dynarray resident_img_needs_color_decompress
;
450 struct util_dynarray resident_tex_needs_depth_decompress
;
453 bool uses_bindless_samplers
;
454 bool uses_bindless_images
;
458 void cik_init_sdma_functions(struct si_context
*sctx
);
461 void si_init_blit_functions(struct si_context
*sctx
);
462 void si_decompress_graphics_textures(struct si_context
*sctx
);
463 void si_decompress_compute_textures(struct si_context
*sctx
);
464 void si_resource_copy_region(struct pipe_context
*ctx
,
465 struct pipe_resource
*dst
,
467 unsigned dstx
, unsigned dsty
, unsigned dstz
,
468 struct pipe_resource
*src
,
470 const struct pipe_box
*src_box
);
473 #define SI_CPDMA_SKIP_CHECK_CS_SPACE (1 << 0) /* don't call need_cs_space */
474 #define SI_CPDMA_SKIP_SYNC_AFTER (1 << 1) /* don't wait for DMA after the copy */
475 #define SI_CPDMA_SKIP_SYNC_BEFORE (1 << 2) /* don't wait for DMA before the copy (RAW hazards) */
476 #define SI_CPDMA_SKIP_GFX_SYNC (1 << 3) /* don't flush caches and don't wait for PS/CS */
477 #define SI_CPDMA_SKIP_BO_LIST_UPDATE (1 << 4) /* don't update the BO list */
478 #define SI_CPDMA_SKIP_ALL (SI_CPDMA_SKIP_CHECK_CS_SPACE | \
479 SI_CPDMA_SKIP_SYNC_AFTER | \
480 SI_CPDMA_SKIP_SYNC_BEFORE | \
481 SI_CPDMA_SKIP_GFX_SYNC | \
482 SI_CPDMA_SKIP_BO_LIST_UPDATE)
484 void si_copy_buffer(struct si_context
*sctx
,
485 struct pipe_resource
*dst
, struct pipe_resource
*src
,
486 uint64_t dst_offset
, uint64_t src_offset
, unsigned size
,
487 unsigned user_flags
);
488 void cik_prefetch_TC_L2_async(struct si_context
*sctx
, struct pipe_resource
*buf
,
489 uint64_t offset
, unsigned size
);
490 void cik_emit_prefetch_L2(struct si_context
*sctx
);
491 void si_init_cp_dma_functions(struct si_context
*sctx
);
494 void si_auto_log_cs(void *data
, struct u_log_context
*log
);
495 void si_log_hw_flush(struct si_context
*sctx
);
496 void si_log_draw_state(struct si_context
*sctx
, struct u_log_context
*log
);
497 void si_log_compute_state(struct si_context
*sctx
, struct u_log_context
*log
);
498 void si_init_debug_functions(struct si_context
*sctx
);
499 void si_check_vm_faults(struct r600_common_context
*ctx
,
500 struct radeon_saved_cs
*saved
, enum ring_type ring
);
501 bool si_replace_shader(unsigned num
, struct ac_shader_binary
*binary
);
504 void si_init_dma_functions(struct si_context
*sctx
);
506 /* si_hw_context.c */
507 void si_destroy_saved_cs(struct si_saved_cs
*scs
);
508 void si_context_gfx_flush(void *context
, unsigned flags
,
509 struct pipe_fence_handle
**fence
);
510 void si_begin_new_cs(struct si_context
*ctx
);
511 void si_need_cs_space(struct si_context
*ctx
);
514 void si_init_compute_functions(struct si_context
*sctx
);
516 /* si_perfcounters.c */
517 void si_init_perfcounters(struct si_screen
*screen
);
520 struct pipe_video_codec
*si_uvd_create_decoder(struct pipe_context
*context
,
521 const struct pipe_video_codec
*templ
);
523 struct pipe_video_buffer
*si_video_buffer_create(struct pipe_context
*pipe
,
524 const struct pipe_video_buffer
*tmpl
);
531 si_invalidate_draw_sh_constants(struct si_context
*sctx
)
533 sctx
->last_base_vertex
= SI_BASE_VERTEX_UNKNOWN
;
537 si_set_atom_dirty(struct si_context
*sctx
,
538 struct r600_atom
*atom
, bool dirty
)
540 unsigned bit
= 1 << atom
->id
;
543 sctx
->dirty_atoms
|= bit
;
545 sctx
->dirty_atoms
&= ~bit
;
549 si_is_atom_dirty(struct si_context
*sctx
,
550 struct r600_atom
*atom
)
552 unsigned bit
= 1 << atom
->id
;
554 return sctx
->dirty_atoms
& bit
;
558 si_mark_atom_dirty(struct si_context
*sctx
,
559 struct r600_atom
*atom
)
561 si_set_atom_dirty(sctx
, atom
, true);
564 static inline struct si_shader_ctx_state
*si_get_vs(struct si_context
*sctx
)
566 if (sctx
->gs_shader
.cso
)
567 return &sctx
->gs_shader
;
568 if (sctx
->tes_shader
.cso
)
569 return &sctx
->tes_shader
;
571 return &sctx
->vs_shader
;
574 static inline struct tgsi_shader_info
*si_get_vs_info(struct si_context
*sctx
)
576 struct si_shader_ctx_state
*vs
= si_get_vs(sctx
);
578 return vs
->cso
? &vs
->cso
->info
: NULL
;
581 static inline struct si_shader
* si_get_vs_state(struct si_context
*sctx
)
583 if (sctx
->gs_shader
.cso
)
584 return sctx
->gs_shader
.cso
->gs_copy_shader
;
586 struct si_shader_ctx_state
*vs
= si_get_vs(sctx
);
587 return vs
->current
? vs
->current
: NULL
;
590 static inline unsigned
591 si_optimal_tcc_alignment(struct si_context
*sctx
, unsigned upload_size
)
593 unsigned alignment
, tcc_cache_line_size
;
595 /* If the upload size is less than the cache line size (e.g. 16, 32),
596 * the whole thing will fit into a cache line if we align it to its size.
597 * The idea is that multiple small uploads can share a cache line.
598 * If the upload size is greater, align it to the cache line size.
600 alignment
= util_next_power_of_two(upload_size
);
601 tcc_cache_line_size
= sctx
->screen
->b
.info
.tcc_cache_line_size
;
602 return MIN2(alignment
, tcc_cache_line_size
);
606 si_saved_cs_reference(struct si_saved_cs
**dst
, struct si_saved_cs
*src
)
608 if (pipe_reference(&(*dst
)->reference
, &src
->reference
))
609 si_destroy_saved_cs(*dst
);