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"
33 #ifdef PIPE_ARCH_BIG_ENDIAN
34 #define SI_BIG_ENDIAN 1
36 #define SI_BIG_ENDIAN 0
39 /* The base vertex and primitive restart can be any number, but we must pick
40 * one which will mean "unknown" for the purpose of state tracking and
41 * the number shouldn't be a commonly-used one. */
42 #define SI_BASE_VERTEX_UNKNOWN INT_MIN
43 #define SI_RESTART_INDEX_UNKNOWN INT_MIN
44 #define SI_NUM_SMOOTH_AA_SAMPLES 8
45 #define SI_GS_PER_ES 128
46 /* Alignment for optimal CP DMA performance. */
47 #define SI_CPDMA_ALIGNMENT 32
49 /* Instruction cache. */
50 #define SI_CONTEXT_INV_ICACHE (R600_CONTEXT_PRIVATE_FLAG << 0)
51 /* SMEM L1, other names: KCACHE, constant cache, DCACHE, data cache */
52 #define SI_CONTEXT_INV_SMEM_L1 (R600_CONTEXT_PRIVATE_FLAG << 1)
53 /* VMEM L1 can optionally be bypassed (GLC=1). Other names: TC L1 */
54 #define SI_CONTEXT_INV_VMEM_L1 (R600_CONTEXT_PRIVATE_FLAG << 2)
55 /* Used by everything except CB/DB, can be bypassed (SLC=1). Other names: TC L2 */
56 #define SI_CONTEXT_INV_GLOBAL_L2 (R600_CONTEXT_PRIVATE_FLAG << 3)
57 /* Write dirty L2 lines back to memory (shader and CP DMA stores), but don't
58 * invalidate L2. SI-CIK can't do it, so they will do complete invalidation. */
59 #define SI_CONTEXT_WRITEBACK_GLOBAL_L2 (R600_CONTEXT_PRIVATE_FLAG << 4)
61 /* Framebuffer caches. */
62 #define SI_CONTEXT_FLUSH_AND_INV_DB (R600_CONTEXT_PRIVATE_FLAG << 7)
63 #define SI_CONTEXT_FLUSH_AND_INV_CB (R600_CONTEXT_PRIVATE_FLAG << 8)
64 /* Engine synchronization. */
65 #define SI_CONTEXT_VS_PARTIAL_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 9)
66 #define SI_CONTEXT_PS_PARTIAL_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 10)
67 #define SI_CONTEXT_CS_PARTIAL_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 11)
68 #define SI_CONTEXT_VGT_FLUSH (R600_CONTEXT_PRIVATE_FLAG << 12)
69 #define SI_CONTEXT_VGT_STREAMOUT_SYNC (R600_CONTEXT_PRIVATE_FLAG << 13)
71 #define SI_MAX_BORDER_COLORS 4096
76 struct u_suballocator
;
79 struct r600_common_screen b
;
80 unsigned gs_table_depth
;
81 unsigned tess_offchip_block_dw_size
;
82 bool has_distributed_tess
;
83 bool has_draw_indirect_multi
;
85 bool has_msaa_sample_loc_bug
;
86 bool llvm_has_working_vgpr_indexing
;
88 /* Whether shaders are monolithic (1-part) or separate (3-part). */
89 bool use_monolithic_shaders
;
92 mtx_t shader_parts_mutex
;
93 struct si_shader_part
*vs_prologs
;
94 struct si_shader_part
*tcs_epilogs
;
95 struct si_shader_part
*gs_prologs
;
96 struct si_shader_part
*ps_prologs
;
97 struct si_shader_part
*ps_epilogs
;
99 /* Shader cache in memory.
101 * Design & limitations:
102 * - The shader cache is per screen (= per process), never saved to
103 * disk, and skips redundant shader compilations from TGSI to bytecode.
104 * - It can only be used with one-variant-per-shader support, in which
105 * case only the main (typically middle) part of shaders is cached.
106 * - Only VS, TCS, TES, PS are cached, out of which only the hw VS
107 * variants of VS and TES are cached, so LS and ES aren't.
108 * - GS and CS aren't cached, but it's certainly possible to cache
111 mtx_t shader_cache_mutex
;
112 struct hash_table
*shader_cache
;
114 /* Shader compiler queue for multithreaded compilation. */
115 struct util_queue shader_compiler_queue
;
116 LLVMTargetMachineRef tm
[4]; /* used by the queue only */
118 struct util_queue shader_compiler_queue_low_priority
;
119 LLVMTargetMachineRef tm_low_priority
[4];
122 struct si_blend_color
{
123 struct r600_atom atom
;
124 struct pipe_blend_color state
;
127 struct si_sampler_view
{
128 struct pipe_sampler_view base
;
129 /* [0..7] = image descriptor
130 * [4..7] = buffer descriptor */
132 uint32_t fmask_state
[8];
133 const struct legacy_surf_level
*base_level_info
;
136 bool is_stencil_sampler
;
137 bool dcc_incompatible
;
140 #define SI_SAMPLER_STATE_MAGIC 0x34f1c35a
142 struct si_sampler_state
{
149 struct si_cs_shader_state
{
150 struct si_compute
*program
;
151 struct si_compute
*emitted_program
;
157 struct si_textures_info
{
158 struct si_sampler_views views
;
159 uint32_t needs_depth_decompress_mask
;
160 uint32_t needs_color_decompress_mask
;
163 struct si_images_info
{
164 struct pipe_image_view views
[SI_NUM_IMAGES
];
165 uint32_t needs_color_decompress_mask
;
166 unsigned enabled_mask
;
169 struct si_framebuffer
{
170 struct r600_atom atom
;
171 struct pipe_framebuffer_state state
;
172 unsigned colorbuf_enabled_4bit
;
173 unsigned spi_shader_col_format
;
174 unsigned spi_shader_col_format_alpha
;
175 unsigned spi_shader_col_format_blend
;
176 unsigned spi_shader_col_format_blend_alpha
;
177 ubyte nr_samples
:5; /* at most 16xAA */
178 ubyte log_samples
:3; /* at most 4 = 16xAA */
179 ubyte compressed_cb_mask
;
181 ubyte color_is_int10
;
185 bool do_update_surf_dirtiness
;
188 struct si_clip_state
{
189 struct r600_atom atom
;
190 struct pipe_clip_state state
;
193 struct si_sample_locs
{
194 struct r600_atom atom
;
198 struct si_sample_mask
{
199 struct r600_atom atom
;
200 uint16_t sample_mask
;
203 /* A shader state consists of the shader selector, which is a constant state
204 * object shared by multiple contexts and shouldn't be modified, and
205 * the current shader variant selected for this context.
207 struct si_shader_ctx_state
{
208 struct si_shader_selector
*cso
;
209 struct si_shader
*current
;
212 #define SI_NUM_VGT_PARAM_KEY_BITS 12
213 #define SI_NUM_VGT_PARAM_STATES (1 << SI_NUM_VGT_PARAM_KEY_BITS)
215 /* The IA_MULTI_VGT_PARAM key used to index the table of precomputed values.
216 * Some fields are set by state-change calls, most are set by draw_vbo.
218 union si_vgt_param_key
{
221 unsigned uses_instancing
:1;
222 unsigned multi_instances_smaller_than_primgroup
:1;
223 unsigned primitive_restart
:1;
224 unsigned count_from_stream_output
:1;
225 unsigned line_stipple_enabled
:1;
226 unsigned uses_tess
:1;
227 unsigned tess_uses_prim_id
:1;
229 unsigned _pad
:32 - SI_NUM_VGT_PARAM_KEY_BITS
;
234 struct si_bindless_descriptor
236 struct pb_slab_entry entry
;
237 struct r600_resource
*buffer
;
239 uint32_t desc_list
[16];
243 struct si_texture_handle
245 struct si_bindless_descriptor
*desc
;
246 struct pipe_sampler_view
*view
;
247 struct si_sampler_state sstate
;
250 struct si_image_handle
252 struct si_bindless_descriptor
*desc
;
253 struct pipe_image_view view
;
257 struct r600_common_context b
;
258 struct blitter_context
*blitter
;
259 void *custom_dsa_flush
;
260 void *custom_blend_resolve
;
261 void *custom_blend_fmask_decompress
;
262 void *custom_blend_eliminate_fastclear
;
263 void *custom_blend_dcc_decompress
;
264 struct si_screen
*screen
;
265 LLVMTargetMachineRef tm
; /* only non-threaded compilation */
266 struct si_shader_ctx_state fixed_func_tcs_shader
;
267 struct r600_resource
*wait_mem_scratch
;
268 unsigned wait_mem_number
;
270 struct radeon_winsys_cs
*ce_ib
;
271 struct radeon_winsys_cs
*ce_preamble_ib
;
272 struct r600_resource
*ce_ram_saved_buffer
;
273 struct u_suballocator
*ce_suballocator
;
274 unsigned ce_ram_saved_offset
;
275 uint16_t total_ce_ram_allocated
;
276 bool ce_need_synchronization
:1;
278 bool gfx_flush_in_progress
:1;
279 bool compute_is_busy
:1;
281 /* Atoms (direct states). */
282 union si_state_atoms atoms
;
283 unsigned dirty_atoms
; /* mask */
284 /* PM4 states (precomputed immutable states) */
285 unsigned dirty_states
;
286 union si_state queued
;
287 union si_state emitted
;
289 /* Atom declarations. */
290 struct r600_atom prefetch_L2
;
291 struct si_framebuffer framebuffer
;
292 struct si_sample_locs msaa_sample_locs
;
293 struct r600_atom db_render_state
;
294 struct r600_atom msaa_config
;
295 struct si_sample_mask sample_mask
;
296 struct r600_atom cb_render_state
;
297 unsigned last_cb_target_mask
;
298 struct si_blend_color blend_color
;
299 struct r600_atom clip_regs
;
300 struct si_clip_state clip_state
;
301 struct si_shader_data shader_userdata
;
302 struct si_stencil_ref stencil_ref
;
303 struct r600_atom spi_map
;
305 /* Precomputed states. */
306 struct si_pm4_state
*init_config
;
307 struct si_pm4_state
*init_config_gs_rings
;
308 bool init_config_has_vgt_flush
;
309 struct si_pm4_state
*vgt_shader_config
[4];
312 struct si_shader_ctx_state ps_shader
;
313 struct si_shader_ctx_state gs_shader
;
314 struct si_shader_ctx_state vs_shader
;
315 struct si_shader_ctx_state tcs_shader
;
316 struct si_shader_ctx_state tes_shader
;
317 struct si_cs_shader_state cs_shader_state
;
319 /* shader information */
320 struct si_vertex_elements
*vertex_elements
;
321 unsigned sprite_coord_enable
;
323 bool do_update_shaders
;
325 /* shader descriptors */
326 struct si_descriptors vertex_buffers
;
327 struct si_descriptors descriptors
[SI_NUM_DESCS
];
328 unsigned descriptors_dirty
;
329 unsigned shader_pointers_dirty
;
330 unsigned shader_needs_decompress_mask
;
331 struct si_buffer_resources rw_buffers
;
332 struct si_buffer_resources const_and_shader_buffers
[SI_NUM_SHADERS
];
333 struct si_textures_info samplers
[SI_NUM_SHADERS
];
334 struct si_images_info images
[SI_NUM_SHADERS
];
336 /* other shader resources */
337 struct pipe_constant_buffer null_const_buf
; /* used for set_constant_buffer(NULL) on CIK */
338 struct pipe_resource
*esgs_ring
;
339 struct pipe_resource
*gsvs_ring
;
340 struct pipe_resource
*tf_ring
;
341 struct pipe_resource
*tess_offchip_ring
;
342 union pipe_color_union
*border_color_table
; /* in CPU memory, any endian */
343 struct r600_resource
*border_color_buffer
;
344 union pipe_color_union
*border_color_map
; /* in VRAM (slow access), little endian */
345 unsigned border_color_count
;
347 /* Vertex and index buffers. */
348 bool vertex_buffers_dirty
;
349 bool vertex_buffer_pointer_dirty
;
350 struct pipe_vertex_buffer vertex_buffer
[SI_NUM_VERTEX_BUFFERS
];
352 /* MSAA config state. */
354 bool smoothing_enabled
;
356 /* DB render state. */
357 unsigned ps_db_shader_control
;
358 unsigned dbcb_copy_sample
;
359 bool dbcb_depth_copy_enabled
:1;
360 bool dbcb_stencil_copy_enabled
:1;
361 bool db_flush_depth_inplace
:1;
362 bool db_flush_stencil_inplace
:1;
363 bool db_depth_clear
:1;
364 bool db_depth_disable_expclear
:1;
365 bool db_stencil_clear
:1;
366 bool db_stencil_disable_expclear
:1;
367 bool occlusion_queries_disabled
:1;
368 bool generate_mipmap_for_depth
:1;
370 /* Emitted draw state. */
371 bool gs_tri_strip_adj_fix
:1;
373 int last_base_vertex
;
374 int last_start_instance
;
376 int last_sh_base_reg
;
377 int last_primitive_restart_en
;
378 int last_restart_index
;
379 int last_gs_out_prim
;
381 int last_multi_vgt_param
;
383 unsigned last_sc_line_stipple
;
384 unsigned current_vs_state
;
385 unsigned last_vs_state
;
386 enum pipe_prim_type current_rast_prim
; /* primitive type after TES, GS */
389 struct r600_atom scratch_state
;
390 struct r600_resource
*scratch_buffer
;
391 unsigned scratch_waves
;
392 unsigned spi_tmpring_size
;
394 struct r600_resource
*compute_scratch_buffer
;
396 /* Emitted derived tessellation state. */
397 /* Local shader (VS), or HS if LS-HS are merged. */
398 struct si_shader
*last_ls
;
399 struct si_shader_selector
*last_tcs
;
400 int last_num_tcs_input_cp
;
401 int last_tes_sh_base
;
402 bool last_tess_uses_primid
;
403 unsigned last_num_patches
;
407 struct radeon_saved_cs last_gfx
;
408 struct r600_resource
*last_trace_buf
;
409 struct r600_resource
*trace_buf
;
411 uint64_t dmesg_timestamp
;
412 unsigned apitrace_call_number
;
415 bool need_check_render_feedback
;
416 bool decompression_enabled
;
418 /* Precomputed IA_MULTI_VGT_PARAM */
419 union si_vgt_param_key ia_multi_vgt_param_key
;
420 unsigned ia_multi_vgt_param
[SI_NUM_VGT_PARAM_STATES
];
422 /* Slab allocator for bindless descriptors. */
423 struct pb_slabs bindless_descriptor_slabs
;
425 /* Bindless descriptors. */
426 struct util_dynarray bindless_descriptors
;
427 bool bindless_descriptors_dirty
;
429 /* Allocated bindless handles */
430 struct hash_table
*tex_handles
;
431 struct hash_table
*img_handles
;
433 /* Resident bindless handles */
434 struct util_dynarray resident_tex_handles
;
435 struct util_dynarray resident_img_handles
;
437 /* Resident bindless handles which need decompression */
438 struct util_dynarray resident_tex_needs_color_decompress
;
439 struct util_dynarray resident_img_needs_color_decompress
;
440 struct util_dynarray resident_tex_needs_depth_decompress
;
443 bool uses_bindless_samplers
;
444 bool uses_bindless_images
;
448 void cik_init_sdma_functions(struct si_context
*sctx
);
451 void si_init_blit_functions(struct si_context
*sctx
);
452 void si_decompress_graphics_textures(struct si_context
*sctx
);
453 void si_decompress_compute_textures(struct si_context
*sctx
);
454 void si_resource_copy_region(struct pipe_context
*ctx
,
455 struct pipe_resource
*dst
,
457 unsigned dstx
, unsigned dsty
, unsigned dstz
,
458 struct pipe_resource
*src
,
460 const struct pipe_box
*src_box
);
463 #define SI_CPDMA_SKIP_CHECK_CS_SPACE (1 << 0) /* don't call need_cs_space */
464 #define SI_CPDMA_SKIP_SYNC_AFTER (1 << 1) /* don't wait for DMA after the copy */
465 #define SI_CPDMA_SKIP_SYNC_BEFORE (1 << 2) /* don't wait for DMA before the copy (RAW hazards) */
466 #define SI_CPDMA_SKIP_GFX_SYNC (1 << 3) /* don't flush caches and don't wait for PS/CS */
467 #define SI_CPDMA_SKIP_BO_LIST_UPDATE (1 << 4) /* don't update the BO list */
468 #define SI_CPDMA_SKIP_ALL (SI_CPDMA_SKIP_CHECK_CS_SPACE | \
469 SI_CPDMA_SKIP_SYNC_AFTER | \
470 SI_CPDMA_SKIP_SYNC_BEFORE | \
471 SI_CPDMA_SKIP_GFX_SYNC | \
472 SI_CPDMA_SKIP_BO_LIST_UPDATE)
474 void si_copy_buffer(struct si_context
*sctx
,
475 struct pipe_resource
*dst
, struct pipe_resource
*src
,
476 uint64_t dst_offset
, uint64_t src_offset
, unsigned size
,
477 unsigned user_flags
);
478 void cik_prefetch_TC_L2_async(struct si_context
*sctx
, struct pipe_resource
*buf
,
479 uint64_t offset
, unsigned size
);
480 void si_init_cp_dma_functions(struct si_context
*sctx
);
483 void si_init_debug_functions(struct si_context
*sctx
);
484 void si_check_vm_faults(struct r600_common_context
*ctx
,
485 struct radeon_saved_cs
*saved
, enum ring_type ring
);
486 bool si_replace_shader(unsigned num
, struct ac_shader_binary
*binary
);
489 void si_init_dma_functions(struct si_context
*sctx
);
491 /* si_hw_context.c */
492 void si_context_gfx_flush(void *context
, unsigned flags
,
493 struct pipe_fence_handle
**fence
);
494 void si_begin_new_cs(struct si_context
*ctx
);
495 void si_need_cs_space(struct si_context
*ctx
);
498 void si_init_compute_functions(struct si_context
*sctx
);
500 /* si_perfcounters.c */
501 void si_init_perfcounters(struct si_screen
*screen
);
504 struct pipe_video_codec
*si_uvd_create_decoder(struct pipe_context
*context
,
505 const struct pipe_video_codec
*templ
);
507 struct pipe_video_buffer
*si_video_buffer_create(struct pipe_context
*pipe
,
508 const struct pipe_video_buffer
*tmpl
);
515 si_invalidate_draw_sh_constants(struct si_context
*sctx
)
517 sctx
->last_base_vertex
= SI_BASE_VERTEX_UNKNOWN
;
521 si_set_atom_dirty(struct si_context
*sctx
,
522 struct r600_atom
*atom
, bool dirty
)
524 unsigned bit
= 1 << atom
->id
;
527 sctx
->dirty_atoms
|= bit
;
529 sctx
->dirty_atoms
&= ~bit
;
533 si_is_atom_dirty(struct si_context
*sctx
,
534 struct r600_atom
*atom
)
536 unsigned bit
= 1 << atom
->id
;
538 return sctx
->dirty_atoms
& bit
;
542 si_mark_atom_dirty(struct si_context
*sctx
,
543 struct r600_atom
*atom
)
545 si_set_atom_dirty(sctx
, atom
, true);
548 static inline struct si_shader_ctx_state
*si_get_vs(struct si_context
*sctx
)
550 if (sctx
->gs_shader
.cso
)
551 return &sctx
->gs_shader
;
552 if (sctx
->tes_shader
.cso
)
553 return &sctx
->tes_shader
;
555 return &sctx
->vs_shader
;
558 static inline struct tgsi_shader_info
*si_get_vs_info(struct si_context
*sctx
)
560 struct si_shader_ctx_state
*vs
= si_get_vs(sctx
);
562 return vs
->cso
? &vs
->cso
->info
: NULL
;
565 static inline struct si_shader
* si_get_vs_state(struct si_context
*sctx
)
567 if (sctx
->gs_shader
.cso
)
568 return sctx
->gs_shader
.cso
->gs_copy_shader
;
570 struct si_shader_ctx_state
*vs
= si_get_vs(sctx
);
571 return vs
->current
? vs
->current
: NULL
;
574 static inline unsigned
575 si_optimal_tcc_alignment(struct si_context
*sctx
, unsigned upload_size
)
577 unsigned alignment
, tcc_cache_line_size
;
579 /* If the upload size is less than the cache line size (e.g. 16, 32),
580 * the whole thing will fit into a cache line if we align it to its size.
581 * The idea is that multiple small uploads can share a cache line.
582 * If the upload size is greater, align it to the cache line size.
584 alignment
= util_next_power_of_two(upload_size
);
585 tcc_cache_line_size
= sctx
->screen
->b
.info
.tcc_cache_line_size
;
586 return MIN2(alignment
, tcc_cache_line_size
);