2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include "radv_private.h"
29 #include "radv_radeon_winsys.h"
30 #include "radv_shader.h"
34 #include "vk_format.h"
35 #include "radv_debug.h"
36 #include "radv_meta.h"
41 RADV_PREFETCH_VBO_DESCRIPTORS
= (1 << 0),
42 RADV_PREFETCH_VS
= (1 << 1),
43 RADV_PREFETCH_TCS
= (1 << 2),
44 RADV_PREFETCH_TES
= (1 << 3),
45 RADV_PREFETCH_GS
= (1 << 4),
46 RADV_PREFETCH_PS
= (1 << 5),
47 RADV_PREFETCH_SHADERS
= (RADV_PREFETCH_VS
|
54 static void radv_handle_image_transition(struct radv_cmd_buffer
*cmd_buffer
,
55 struct radv_image
*image
,
56 VkImageLayout src_layout
,
57 VkImageLayout dst_layout
,
60 const VkImageSubresourceRange
*range
,
61 VkImageAspectFlags pending_clears
);
63 const struct radv_dynamic_state default_dynamic_state
= {
76 .blend_constants
= { 0.0f
, 0.0f
, 0.0f
, 0.0f
},
81 .stencil_compare_mask
= {
85 .stencil_write_mask
= {
89 .stencil_reference
= {
96 radv_bind_dynamic_state(struct radv_cmd_buffer
*cmd_buffer
,
97 const struct radv_dynamic_state
*src
)
99 struct radv_dynamic_state
*dest
= &cmd_buffer
->state
.dynamic
;
100 uint32_t copy_mask
= src
->mask
;
101 uint32_t dest_mask
= 0;
103 /* Make sure to copy the number of viewports/scissors because they can
104 * only be specified at pipeline creation time.
106 dest
->viewport
.count
= src
->viewport
.count
;
107 dest
->scissor
.count
= src
->scissor
.count
;
108 dest
->discard_rectangle
.count
= src
->discard_rectangle
.count
;
110 if (copy_mask
& RADV_DYNAMIC_VIEWPORT
) {
111 if (memcmp(&dest
->viewport
.viewports
, &src
->viewport
.viewports
,
112 src
->viewport
.count
* sizeof(VkViewport
))) {
113 typed_memcpy(dest
->viewport
.viewports
,
114 src
->viewport
.viewports
,
115 src
->viewport
.count
);
116 dest_mask
|= RADV_DYNAMIC_VIEWPORT
;
120 if (copy_mask
& RADV_DYNAMIC_SCISSOR
) {
121 if (memcmp(&dest
->scissor
.scissors
, &src
->scissor
.scissors
,
122 src
->scissor
.count
* sizeof(VkRect2D
))) {
123 typed_memcpy(dest
->scissor
.scissors
,
124 src
->scissor
.scissors
, src
->scissor
.count
);
125 dest_mask
|= RADV_DYNAMIC_SCISSOR
;
129 if (copy_mask
& RADV_DYNAMIC_LINE_WIDTH
) {
130 if (dest
->line_width
!= src
->line_width
) {
131 dest
->line_width
= src
->line_width
;
132 dest_mask
|= RADV_DYNAMIC_LINE_WIDTH
;
136 if (copy_mask
& RADV_DYNAMIC_DEPTH_BIAS
) {
137 if (memcmp(&dest
->depth_bias
, &src
->depth_bias
,
138 sizeof(src
->depth_bias
))) {
139 dest
->depth_bias
= src
->depth_bias
;
140 dest_mask
|= RADV_DYNAMIC_DEPTH_BIAS
;
144 if (copy_mask
& RADV_DYNAMIC_BLEND_CONSTANTS
) {
145 if (memcmp(&dest
->blend_constants
, &src
->blend_constants
,
146 sizeof(src
->blend_constants
))) {
147 typed_memcpy(dest
->blend_constants
,
148 src
->blend_constants
, 4);
149 dest_mask
|= RADV_DYNAMIC_BLEND_CONSTANTS
;
153 if (copy_mask
& RADV_DYNAMIC_DEPTH_BOUNDS
) {
154 if (memcmp(&dest
->depth_bounds
, &src
->depth_bounds
,
155 sizeof(src
->depth_bounds
))) {
156 dest
->depth_bounds
= src
->depth_bounds
;
157 dest_mask
|= RADV_DYNAMIC_DEPTH_BOUNDS
;
161 if (copy_mask
& RADV_DYNAMIC_STENCIL_COMPARE_MASK
) {
162 if (memcmp(&dest
->stencil_compare_mask
,
163 &src
->stencil_compare_mask
,
164 sizeof(src
->stencil_compare_mask
))) {
165 dest
->stencil_compare_mask
= src
->stencil_compare_mask
;
166 dest_mask
|= RADV_DYNAMIC_STENCIL_COMPARE_MASK
;
170 if (copy_mask
& RADV_DYNAMIC_STENCIL_WRITE_MASK
) {
171 if (memcmp(&dest
->stencil_write_mask
, &src
->stencil_write_mask
,
172 sizeof(src
->stencil_write_mask
))) {
173 dest
->stencil_write_mask
= src
->stencil_write_mask
;
174 dest_mask
|= RADV_DYNAMIC_STENCIL_WRITE_MASK
;
178 if (copy_mask
& RADV_DYNAMIC_STENCIL_REFERENCE
) {
179 if (memcmp(&dest
->stencil_reference
, &src
->stencil_reference
,
180 sizeof(src
->stencil_reference
))) {
181 dest
->stencil_reference
= src
->stencil_reference
;
182 dest_mask
|= RADV_DYNAMIC_STENCIL_REFERENCE
;
186 if (copy_mask
& RADV_DYNAMIC_DISCARD_RECTANGLE
) {
187 if (memcmp(&dest
->discard_rectangle
.rectangles
, &src
->discard_rectangle
.rectangles
,
188 src
->discard_rectangle
.count
* sizeof(VkRect2D
))) {
189 typed_memcpy(dest
->discard_rectangle
.rectangles
,
190 src
->discard_rectangle
.rectangles
,
191 src
->discard_rectangle
.count
);
192 dest_mask
|= RADV_DYNAMIC_DISCARD_RECTANGLE
;
196 cmd_buffer
->state
.dirty
|= dest_mask
;
199 bool radv_cmd_buffer_uses_mec(struct radv_cmd_buffer
*cmd_buffer
)
201 return cmd_buffer
->queue_family_index
== RADV_QUEUE_COMPUTE
&&
202 cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= CIK
;
205 enum ring_type
radv_queue_family_to_ring(int f
) {
207 case RADV_QUEUE_GENERAL
:
209 case RADV_QUEUE_COMPUTE
:
211 case RADV_QUEUE_TRANSFER
:
214 unreachable("Unknown queue family");
218 static VkResult
radv_create_cmd_buffer(
219 struct radv_device
* device
,
220 struct radv_cmd_pool
* pool
,
221 VkCommandBufferLevel level
,
222 VkCommandBuffer
* pCommandBuffer
)
224 struct radv_cmd_buffer
*cmd_buffer
;
226 cmd_buffer
= vk_zalloc(&pool
->alloc
, sizeof(*cmd_buffer
), 8,
227 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
228 if (cmd_buffer
== NULL
)
229 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
231 cmd_buffer
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
232 cmd_buffer
->device
= device
;
233 cmd_buffer
->pool
= pool
;
234 cmd_buffer
->level
= level
;
237 list_addtail(&cmd_buffer
->pool_link
, &pool
->cmd_buffers
);
238 cmd_buffer
->queue_family_index
= pool
->queue_family_index
;
241 /* Init the pool_link so we can safefly call list_del when we destroy
244 list_inithead(&cmd_buffer
->pool_link
);
245 cmd_buffer
->queue_family_index
= RADV_QUEUE_GENERAL
;
248 ring
= radv_queue_family_to_ring(cmd_buffer
->queue_family_index
);
250 cmd_buffer
->cs
= device
->ws
->cs_create(device
->ws
, ring
);
251 if (!cmd_buffer
->cs
) {
252 vk_free(&cmd_buffer
->pool
->alloc
, cmd_buffer
);
253 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
256 *pCommandBuffer
= radv_cmd_buffer_to_handle(cmd_buffer
);
258 list_inithead(&cmd_buffer
->upload
.list
);
264 radv_cmd_buffer_destroy(struct radv_cmd_buffer
*cmd_buffer
)
266 list_del(&cmd_buffer
->pool_link
);
268 list_for_each_entry_safe(struct radv_cmd_buffer_upload
, up
,
269 &cmd_buffer
->upload
.list
, list
) {
270 cmd_buffer
->device
->ws
->buffer_destroy(up
->upload_bo
);
275 if (cmd_buffer
->upload
.upload_bo
)
276 cmd_buffer
->device
->ws
->buffer_destroy(cmd_buffer
->upload
.upload_bo
);
277 cmd_buffer
->device
->ws
->cs_destroy(cmd_buffer
->cs
);
279 for (unsigned i
= 0; i
< VK_PIPELINE_BIND_POINT_RANGE_SIZE
; i
++)
280 free(cmd_buffer
->descriptors
[i
].push_set
.set
.mapped_ptr
);
282 vk_free(&cmd_buffer
->pool
->alloc
, cmd_buffer
);
286 radv_reset_cmd_buffer(struct radv_cmd_buffer
*cmd_buffer
)
289 cmd_buffer
->device
->ws
->cs_reset(cmd_buffer
->cs
);
291 list_for_each_entry_safe(struct radv_cmd_buffer_upload
, up
,
292 &cmd_buffer
->upload
.list
, list
) {
293 cmd_buffer
->device
->ws
->buffer_destroy(up
->upload_bo
);
298 cmd_buffer
->push_constant_stages
= 0;
299 cmd_buffer
->scratch_size_needed
= 0;
300 cmd_buffer
->compute_scratch_size_needed
= 0;
301 cmd_buffer
->esgs_ring_size_needed
= 0;
302 cmd_buffer
->gsvs_ring_size_needed
= 0;
303 cmd_buffer
->tess_rings_needed
= false;
304 cmd_buffer
->sample_positions_needed
= false;
306 if (cmd_buffer
->upload
.upload_bo
)
307 radv_cs_add_buffer(cmd_buffer
->device
->ws
, cmd_buffer
->cs
,
308 cmd_buffer
->upload
.upload_bo
, 8);
309 cmd_buffer
->upload
.offset
= 0;
311 cmd_buffer
->record_result
= VK_SUCCESS
;
313 cmd_buffer
->ring_offsets_idx
= -1;
315 for (unsigned i
= 0; i
< VK_PIPELINE_BIND_POINT_RANGE_SIZE
; i
++) {
316 cmd_buffer
->descriptors
[i
].dirty
= 0;
317 cmd_buffer
->descriptors
[i
].valid
= 0;
318 cmd_buffer
->descriptors
[i
].push_dirty
= false;
321 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
323 radv_cmd_buffer_upload_alloc(cmd_buffer
, 8, 0,
324 &cmd_buffer
->gfx9_fence_offset
,
326 cmd_buffer
->gfx9_fence_bo
= cmd_buffer
->upload
.upload_bo
;
329 cmd_buffer
->status
= RADV_CMD_BUFFER_STATUS_INITIAL
;
331 return cmd_buffer
->record_result
;
335 radv_cmd_buffer_resize_upload_buf(struct radv_cmd_buffer
*cmd_buffer
,
339 struct radeon_winsys_bo
*bo
;
340 struct radv_cmd_buffer_upload
*upload
;
341 struct radv_device
*device
= cmd_buffer
->device
;
343 new_size
= MAX2(min_needed
, 16 * 1024);
344 new_size
= MAX2(new_size
, 2 * cmd_buffer
->upload
.size
);
346 bo
= device
->ws
->buffer_create(device
->ws
,
349 RADEON_FLAG_CPU_ACCESS
|
350 RADEON_FLAG_NO_INTERPROCESS_SHARING
);
353 cmd_buffer
->record_result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
357 radv_cs_add_buffer(device
->ws
, cmd_buffer
->cs
, bo
, 8);
358 if (cmd_buffer
->upload
.upload_bo
) {
359 upload
= malloc(sizeof(*upload
));
362 cmd_buffer
->record_result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
363 device
->ws
->buffer_destroy(bo
);
367 memcpy(upload
, &cmd_buffer
->upload
, sizeof(*upload
));
368 list_add(&upload
->list
, &cmd_buffer
->upload
.list
);
371 cmd_buffer
->upload
.upload_bo
= bo
;
372 cmd_buffer
->upload
.size
= new_size
;
373 cmd_buffer
->upload
.offset
= 0;
374 cmd_buffer
->upload
.map
= device
->ws
->buffer_map(cmd_buffer
->upload
.upload_bo
);
376 if (!cmd_buffer
->upload
.map
) {
377 cmd_buffer
->record_result
= VK_ERROR_OUT_OF_DEVICE_MEMORY
;
385 radv_cmd_buffer_upload_alloc(struct radv_cmd_buffer
*cmd_buffer
,
388 unsigned *out_offset
,
391 uint64_t offset
= align(cmd_buffer
->upload
.offset
, alignment
);
392 if (offset
+ size
> cmd_buffer
->upload
.size
) {
393 if (!radv_cmd_buffer_resize_upload_buf(cmd_buffer
, size
))
398 *out_offset
= offset
;
399 *ptr
= cmd_buffer
->upload
.map
+ offset
;
401 cmd_buffer
->upload
.offset
= offset
+ size
;
406 radv_cmd_buffer_upload_data(struct radv_cmd_buffer
*cmd_buffer
,
407 unsigned size
, unsigned alignment
,
408 const void *data
, unsigned *out_offset
)
412 if (!radv_cmd_buffer_upload_alloc(cmd_buffer
, size
, alignment
,
413 out_offset
, (void **)&ptr
))
417 memcpy(ptr
, data
, size
);
423 radv_emit_write_data_packet(struct radeon_winsys_cs
*cs
, uint64_t va
,
424 unsigned count
, const uint32_t *data
)
426 radeon_emit(cs
, PKT3(PKT3_WRITE_DATA
, 2 + count
, 0));
427 radeon_emit(cs
, S_370_DST_SEL(V_370_MEM_ASYNC
) |
428 S_370_WR_CONFIRM(1) |
429 S_370_ENGINE_SEL(V_370_ME
));
431 radeon_emit(cs
, va
>> 32);
432 radeon_emit_array(cs
, data
, count
);
435 void radv_cmd_buffer_trace_emit(struct radv_cmd_buffer
*cmd_buffer
)
437 struct radv_device
*device
= cmd_buffer
->device
;
438 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
441 va
= radv_buffer_get_va(device
->trace_bo
);
442 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
)
445 MAYBE_UNUSED
unsigned cdw_max
= radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, 7);
447 ++cmd_buffer
->state
.trace_id
;
448 radv_cs_add_buffer(device
->ws
, cs
, device
->trace_bo
, 8);
449 radv_emit_write_data_packet(cs
, va
, 1, &cmd_buffer
->state
.trace_id
);
450 radeon_emit(cs
, PKT3(PKT3_NOP
, 0, 0));
451 radeon_emit(cs
, AC_ENCODE_TRACE_POINT(cmd_buffer
->state
.trace_id
));
455 radv_cmd_buffer_after_draw(struct radv_cmd_buffer
*cmd_buffer
,
456 enum radv_cmd_flush_bits flags
)
458 if (cmd_buffer
->device
->instance
->debug_flags
& RADV_DEBUG_SYNC_SHADERS
) {
459 uint32_t *ptr
= NULL
;
462 assert(flags
& (RADV_CMD_FLAG_PS_PARTIAL_FLUSH
|
463 RADV_CMD_FLAG_CS_PARTIAL_FLUSH
));
465 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
== GFX9
) {
466 va
= radv_buffer_get_va(cmd_buffer
->gfx9_fence_bo
) +
467 cmd_buffer
->gfx9_fence_offset
;
468 ptr
= &cmd_buffer
->gfx9_fence_idx
;
471 /* Force wait for graphics or compute engines to be idle. */
472 si_cs_emit_cache_flush(cmd_buffer
->cs
,
473 cmd_buffer
->device
->physical_device
->rad_info
.chip_class
,
475 radv_cmd_buffer_uses_mec(cmd_buffer
),
479 if (unlikely(cmd_buffer
->device
->trace_bo
))
480 radv_cmd_buffer_trace_emit(cmd_buffer
);
484 radv_save_pipeline(struct radv_cmd_buffer
*cmd_buffer
,
485 struct radv_pipeline
*pipeline
, enum ring_type ring
)
487 struct radv_device
*device
= cmd_buffer
->device
;
488 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
492 va
= radv_buffer_get_va(device
->trace_bo
);
502 assert(!"invalid ring type");
505 MAYBE_UNUSED
unsigned cdw_max
= radeon_check_space(device
->ws
,
508 data
[0] = (uintptr_t)pipeline
;
509 data
[1] = (uintptr_t)pipeline
>> 32;
511 radv_cs_add_buffer(device
->ws
, cs
, device
->trace_bo
, 8);
512 radv_emit_write_data_packet(cs
, va
, 2, data
);
515 void radv_set_descriptor_set(struct radv_cmd_buffer
*cmd_buffer
,
516 VkPipelineBindPoint bind_point
,
517 struct radv_descriptor_set
*set
,
520 struct radv_descriptor_state
*descriptors_state
=
521 radv_get_descriptors_state(cmd_buffer
, bind_point
);
523 descriptors_state
->sets
[idx
] = set
;
525 descriptors_state
->valid
|= (1u << idx
);
527 descriptors_state
->valid
&= ~(1u << idx
);
528 descriptors_state
->dirty
|= (1u << idx
);
532 radv_save_descriptors(struct radv_cmd_buffer
*cmd_buffer
,
533 VkPipelineBindPoint bind_point
)
535 struct radv_descriptor_state
*descriptors_state
=
536 radv_get_descriptors_state(cmd_buffer
, bind_point
);
537 struct radv_device
*device
= cmd_buffer
->device
;
538 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
539 uint32_t data
[MAX_SETS
* 2] = {};
542 va
= radv_buffer_get_va(device
->trace_bo
) + 24;
544 MAYBE_UNUSED
unsigned cdw_max
= radeon_check_space(device
->ws
,
545 cmd_buffer
->cs
, 4 + MAX_SETS
* 2);
547 for_each_bit(i
, descriptors_state
->valid
) {
548 struct radv_descriptor_set
*set
= descriptors_state
->sets
[i
];
549 data
[i
* 2] = (uintptr_t)set
;
550 data
[i
* 2 + 1] = (uintptr_t)set
>> 32;
553 radv_cs_add_buffer(device
->ws
, cs
, device
->trace_bo
, 8);
554 radv_emit_write_data_packet(cs
, va
, MAX_SETS
* 2, data
);
557 struct radv_userdata_info
*
558 radv_lookup_user_sgpr(struct radv_pipeline
*pipeline
,
559 gl_shader_stage stage
,
562 if (stage
== MESA_SHADER_VERTEX
) {
563 if (pipeline
->shaders
[MESA_SHADER_VERTEX
])
564 return &pipeline
->shaders
[MESA_SHADER_VERTEX
]->info
.user_sgprs_locs
.shader_data
[idx
];
565 if (pipeline
->shaders
[MESA_SHADER_TESS_CTRL
])
566 return &pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.user_sgprs_locs
.shader_data
[idx
];
567 if (pipeline
->shaders
[MESA_SHADER_GEOMETRY
])
568 return &pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
.user_sgprs_locs
.shader_data
[idx
];
569 } else if (stage
== MESA_SHADER_TESS_EVAL
) {
570 if (pipeline
->shaders
[MESA_SHADER_TESS_EVAL
])
571 return &pipeline
->shaders
[MESA_SHADER_TESS_EVAL
]->info
.user_sgprs_locs
.shader_data
[idx
];
572 if (pipeline
->shaders
[MESA_SHADER_GEOMETRY
])
573 return &pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
.user_sgprs_locs
.shader_data
[idx
];
575 return &pipeline
->shaders
[stage
]->info
.user_sgprs_locs
.shader_data
[idx
];
579 radv_emit_userdata_address(struct radv_cmd_buffer
*cmd_buffer
,
580 struct radv_pipeline
*pipeline
,
581 gl_shader_stage stage
,
582 int idx
, uint64_t va
)
584 struct radv_userdata_info
*loc
= radv_lookup_user_sgpr(pipeline
, stage
, idx
);
585 uint32_t base_reg
= pipeline
->user_data_0
[stage
];
586 if (loc
->sgpr_idx
== -1)
588 assert(loc
->num_sgprs
== 2);
589 assert(!loc
->indirect
);
590 radeon_set_sh_reg_seq(cmd_buffer
->cs
, base_reg
+ loc
->sgpr_idx
* 4, 2);
591 radeon_emit(cmd_buffer
->cs
, va
);
592 radeon_emit(cmd_buffer
->cs
, va
>> 32);
596 radv_update_multisample_state(struct radv_cmd_buffer
*cmd_buffer
,
597 struct radv_pipeline
*pipeline
)
599 int num_samples
= pipeline
->graphics
.ms
.num_samples
;
600 struct radv_multisample_state
*ms
= &pipeline
->graphics
.ms
;
601 struct radv_pipeline
*old_pipeline
= cmd_buffer
->state
.emitted_pipeline
;
603 if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.needs_sample_positions
)
604 cmd_buffer
->sample_positions_needed
= true;
606 if (old_pipeline
&& num_samples
== old_pipeline
->graphics
.ms
.num_samples
)
609 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028BDC_PA_SC_LINE_CNTL
, 2);
610 radeon_emit(cmd_buffer
->cs
, ms
->pa_sc_line_cntl
);
611 radeon_emit(cmd_buffer
->cs
, ms
->pa_sc_aa_config
);
613 radeon_set_context_reg(cmd_buffer
->cs
, R_028A48_PA_SC_MODE_CNTL_0
, ms
->pa_sc_mode_cntl_0
);
615 radv_cayman_emit_msaa_sample_locs(cmd_buffer
->cs
, num_samples
);
617 /* GFX9: Flush DFSM when the AA mode changes. */
618 if (cmd_buffer
->device
->dfsm_allowed
) {
619 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
620 radeon_emit(cmd_buffer
->cs
, EVENT_TYPE(V_028A90_FLUSH_DFSM
) | EVENT_INDEX(0));
625 radv_emit_shader_prefetch(struct radv_cmd_buffer
*cmd_buffer
,
626 struct radv_shader_variant
*shader
)
633 va
= radv_buffer_get_va(shader
->bo
) + shader
->bo_offset
;
635 si_cp_dma_prefetch(cmd_buffer
, va
, shader
->code_size
);
639 radv_emit_prefetch_L2(struct radv_cmd_buffer
*cmd_buffer
,
640 struct radv_pipeline
*pipeline
,
641 bool vertex_stage_only
)
643 struct radv_cmd_state
*state
= &cmd_buffer
->state
;
644 uint32_t mask
= state
->prefetch_L2_mask
;
646 if (vertex_stage_only
) {
647 /* Fast prefetch path for starting draws as soon as possible.
649 mask
= state
->prefetch_L2_mask
& (RADV_PREFETCH_VS
|
650 RADV_PREFETCH_VBO_DESCRIPTORS
);
653 if (mask
& RADV_PREFETCH_VS
)
654 radv_emit_shader_prefetch(cmd_buffer
,
655 pipeline
->shaders
[MESA_SHADER_VERTEX
]);
657 if (mask
& RADV_PREFETCH_VBO_DESCRIPTORS
)
658 si_cp_dma_prefetch(cmd_buffer
, state
->vb_va
, state
->vb_size
);
660 if (mask
& RADV_PREFETCH_TCS
)
661 radv_emit_shader_prefetch(cmd_buffer
,
662 pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]);
664 if (mask
& RADV_PREFETCH_TES
)
665 radv_emit_shader_prefetch(cmd_buffer
,
666 pipeline
->shaders
[MESA_SHADER_TESS_EVAL
]);
668 if (mask
& RADV_PREFETCH_GS
) {
669 radv_emit_shader_prefetch(cmd_buffer
,
670 pipeline
->shaders
[MESA_SHADER_GEOMETRY
]);
671 radv_emit_shader_prefetch(cmd_buffer
, pipeline
->gs_copy_shader
);
674 if (mask
& RADV_PREFETCH_PS
)
675 radv_emit_shader_prefetch(cmd_buffer
,
676 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]);
678 state
->prefetch_L2_mask
&= ~mask
;
682 radv_emit_rbplus_state(struct radv_cmd_buffer
*cmd_buffer
)
684 if (!cmd_buffer
->device
->physical_device
->rbplus_allowed
)
687 struct radv_pipeline
*pipeline
= cmd_buffer
->state
.pipeline
;
688 struct radv_framebuffer
*framebuffer
= cmd_buffer
->state
.framebuffer
;
689 const struct radv_subpass
*subpass
= cmd_buffer
->state
.subpass
;
691 unsigned sx_ps_downconvert
= 0;
692 unsigned sx_blend_opt_epsilon
= 0;
693 unsigned sx_blend_opt_control
= 0;
695 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
696 if (subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
)
699 int idx
= subpass
->color_attachments
[i
].attachment
;
700 struct radv_color_buffer_info
*cb
= &framebuffer
->attachments
[idx
].cb
;
702 unsigned format
= G_028C70_FORMAT(cb
->cb_color_info
);
703 unsigned swap
= G_028C70_COMP_SWAP(cb
->cb_color_info
);
704 uint32_t spi_format
= (pipeline
->graphics
.col_format
>> (i
* 4)) & 0xf;
705 uint32_t colormask
= (pipeline
->graphics
.cb_target_mask
>> (i
* 4)) & 0xf;
707 bool has_alpha
, has_rgb
;
709 /* Set if RGB and A are present. */
710 has_alpha
= !G_028C74_FORCE_DST_ALPHA_1(cb
->cb_color_attrib
);
712 if (format
== V_028C70_COLOR_8
||
713 format
== V_028C70_COLOR_16
||
714 format
== V_028C70_COLOR_32
)
715 has_rgb
= !has_alpha
;
719 /* Check the colormask and export format. */
720 if (!(colormask
& 0x7))
722 if (!(colormask
& 0x8))
725 if (spi_format
== V_028714_SPI_SHADER_ZERO
) {
730 /* Disable value checking for disabled channels. */
732 sx_blend_opt_control
|= S_02875C_MRT0_COLOR_OPT_DISABLE(1) << (i
* 4);
734 sx_blend_opt_control
|= S_02875C_MRT0_ALPHA_OPT_DISABLE(1) << (i
* 4);
736 /* Enable down-conversion for 32bpp and smaller formats. */
738 case V_028C70_COLOR_8
:
739 case V_028C70_COLOR_8_8
:
740 case V_028C70_COLOR_8_8_8_8
:
741 /* For 1 and 2-channel formats, use the superset thereof. */
742 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
||
743 spi_format
== V_028714_SPI_SHADER_UINT16_ABGR
||
744 spi_format
== V_028714_SPI_SHADER_SINT16_ABGR
) {
745 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_8_8_8_8
<< (i
* 4);
746 sx_blend_opt_epsilon
|= V_028758_8BIT_FORMAT
<< (i
* 4);
750 case V_028C70_COLOR_5_6_5
:
751 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
752 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_5_6_5
<< (i
* 4);
753 sx_blend_opt_epsilon
|= V_028758_6BIT_FORMAT
<< (i
* 4);
757 case V_028C70_COLOR_1_5_5_5
:
758 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
759 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_1_5_5_5
<< (i
* 4);
760 sx_blend_opt_epsilon
|= V_028758_5BIT_FORMAT
<< (i
* 4);
764 case V_028C70_COLOR_4_4_4_4
:
765 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
766 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_4_4_4_4
<< (i
* 4);
767 sx_blend_opt_epsilon
|= V_028758_4BIT_FORMAT
<< (i
* 4);
771 case V_028C70_COLOR_32
:
772 if (swap
== V_028C70_SWAP_STD
&&
773 spi_format
== V_028714_SPI_SHADER_32_R
)
774 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_32_R
<< (i
* 4);
775 else if (swap
== V_028C70_SWAP_ALT_REV
&&
776 spi_format
== V_028714_SPI_SHADER_32_AR
)
777 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_32_A
<< (i
* 4);
780 case V_028C70_COLOR_16
:
781 case V_028C70_COLOR_16_16
:
782 /* For 1-channel formats, use the superset thereof. */
783 if (spi_format
== V_028714_SPI_SHADER_UNORM16_ABGR
||
784 spi_format
== V_028714_SPI_SHADER_SNORM16_ABGR
||
785 spi_format
== V_028714_SPI_SHADER_UINT16_ABGR
||
786 spi_format
== V_028714_SPI_SHADER_SINT16_ABGR
) {
787 if (swap
== V_028C70_SWAP_STD
||
788 swap
== V_028C70_SWAP_STD_REV
)
789 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_16_16_GR
<< (i
* 4);
791 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_16_16_AR
<< (i
* 4);
795 case V_028C70_COLOR_10_11_11
:
796 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
797 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_10_11_11
<< (i
* 4);
798 sx_blend_opt_epsilon
|= V_028758_11BIT_FORMAT
<< (i
* 4);
802 case V_028C70_COLOR_2_10_10_10
:
803 if (spi_format
== V_028714_SPI_SHADER_FP16_ABGR
) {
804 sx_ps_downconvert
|= V_028754_SX_RT_EXPORT_2_10_10_10
<< (i
* 4);
805 sx_blend_opt_epsilon
|= V_028758_10BIT_FORMAT
<< (i
* 4);
811 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028754_SX_PS_DOWNCONVERT
, 3);
812 radeon_emit(cmd_buffer
->cs
, sx_ps_downconvert
);
813 radeon_emit(cmd_buffer
->cs
, sx_blend_opt_epsilon
);
814 radeon_emit(cmd_buffer
->cs
, sx_blend_opt_control
);
818 radv_emit_graphics_pipeline(struct radv_cmd_buffer
*cmd_buffer
)
820 struct radv_pipeline
*pipeline
= cmd_buffer
->state
.pipeline
;
822 if (!pipeline
|| cmd_buffer
->state
.emitted_pipeline
== pipeline
)
825 radv_update_multisample_state(cmd_buffer
, pipeline
);
827 cmd_buffer
->scratch_size_needed
=
828 MAX2(cmd_buffer
->scratch_size_needed
,
829 pipeline
->max_waves
* pipeline
->scratch_bytes_per_wave
);
831 if (!cmd_buffer
->state
.emitted_pipeline
||
832 cmd_buffer
->state
.emitted_pipeline
->graphics
.can_use_guardband
!=
833 pipeline
->graphics
.can_use_guardband
)
834 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_DYNAMIC_SCISSOR
;
836 radeon_emit_array(cmd_buffer
->cs
, pipeline
->cs
.buf
, pipeline
->cs
.cdw
);
838 for (unsigned i
= 0; i
< MESA_SHADER_COMPUTE
; i
++) {
839 if (!pipeline
->shaders
[i
])
842 radv_cs_add_buffer(cmd_buffer
->device
->ws
, cmd_buffer
->cs
,
843 pipeline
->shaders
[i
]->bo
, 8);
846 if (radv_pipeline_has_gs(pipeline
))
847 radv_cs_add_buffer(cmd_buffer
->device
->ws
, cmd_buffer
->cs
,
848 pipeline
->gs_copy_shader
->bo
, 8);
850 if (unlikely(cmd_buffer
->device
->trace_bo
))
851 radv_save_pipeline(cmd_buffer
, pipeline
, RING_GFX
);
853 cmd_buffer
->state
.emitted_pipeline
= pipeline
;
855 cmd_buffer
->state
.dirty
&= ~RADV_CMD_DIRTY_PIPELINE
;
859 radv_emit_viewport(struct radv_cmd_buffer
*cmd_buffer
)
861 si_write_viewport(cmd_buffer
->cs
, 0, cmd_buffer
->state
.dynamic
.viewport
.count
,
862 cmd_buffer
->state
.dynamic
.viewport
.viewports
);
866 radv_emit_scissor(struct radv_cmd_buffer
*cmd_buffer
)
868 uint32_t count
= cmd_buffer
->state
.dynamic
.scissor
.count
;
870 /* Vega10/Raven scissor bug workaround. This must be done before VPORT
871 * scissor registers are changed. There is also a more efficient but
872 * more involved alternative workaround.
874 if (cmd_buffer
->device
->physical_device
->has_scissor_bug
) {
875 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_PS_PARTIAL_FLUSH
;
876 si_emit_cache_flush(cmd_buffer
);
878 si_write_scissors(cmd_buffer
->cs
, 0, count
,
879 cmd_buffer
->state
.dynamic
.scissor
.scissors
,
880 cmd_buffer
->state
.dynamic
.viewport
.viewports
,
881 cmd_buffer
->state
.emitted_pipeline
->graphics
.can_use_guardband
);
885 radv_emit_discard_rectangle(struct radv_cmd_buffer
*cmd_buffer
)
887 if (!cmd_buffer
->state
.dynamic
.discard_rectangle
.count
)
890 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028210_PA_SC_CLIPRECT_0_TL
,
891 cmd_buffer
->state
.dynamic
.discard_rectangle
.count
* 2);
892 for (unsigned i
= 0; i
< cmd_buffer
->state
.dynamic
.discard_rectangle
.count
; ++i
) {
893 VkRect2D rect
= cmd_buffer
->state
.dynamic
.discard_rectangle
.rectangles
[i
];
894 radeon_emit(cmd_buffer
->cs
, S_028210_TL_X(rect
.offset
.x
) | S_028210_TL_Y(rect
.offset
.y
));
895 radeon_emit(cmd_buffer
->cs
, S_028214_BR_X(rect
.offset
.x
+ rect
.extent
.width
) |
896 S_028214_BR_Y(rect
.offset
.y
+ rect
.extent
.height
));
901 radv_emit_line_width(struct radv_cmd_buffer
*cmd_buffer
)
903 unsigned width
= cmd_buffer
->state
.dynamic
.line_width
* 8;
905 radeon_set_context_reg(cmd_buffer
->cs
, R_028A08_PA_SU_LINE_CNTL
,
906 S_028A08_WIDTH(CLAMP(width
, 0, 0xFFF)));
910 radv_emit_blend_constants(struct radv_cmd_buffer
*cmd_buffer
)
912 struct radv_dynamic_state
*d
= &cmd_buffer
->state
.dynamic
;
914 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028414_CB_BLEND_RED
, 4);
915 radeon_emit_array(cmd_buffer
->cs
, (uint32_t *)d
->blend_constants
, 4);
919 radv_emit_stencil(struct radv_cmd_buffer
*cmd_buffer
)
921 struct radv_dynamic_state
*d
= &cmd_buffer
->state
.dynamic
;
923 radeon_set_context_reg_seq(cmd_buffer
->cs
,
924 R_028430_DB_STENCILREFMASK
, 2);
925 radeon_emit(cmd_buffer
->cs
,
926 S_028430_STENCILTESTVAL(d
->stencil_reference
.front
) |
927 S_028430_STENCILMASK(d
->stencil_compare_mask
.front
) |
928 S_028430_STENCILWRITEMASK(d
->stencil_write_mask
.front
) |
929 S_028430_STENCILOPVAL(1));
930 radeon_emit(cmd_buffer
->cs
,
931 S_028434_STENCILTESTVAL_BF(d
->stencil_reference
.back
) |
932 S_028434_STENCILMASK_BF(d
->stencil_compare_mask
.back
) |
933 S_028434_STENCILWRITEMASK_BF(d
->stencil_write_mask
.back
) |
934 S_028434_STENCILOPVAL_BF(1));
938 radv_emit_depth_bounds(struct radv_cmd_buffer
*cmd_buffer
)
940 struct radv_dynamic_state
*d
= &cmd_buffer
->state
.dynamic
;
942 radeon_set_context_reg(cmd_buffer
->cs
, R_028020_DB_DEPTH_BOUNDS_MIN
,
943 fui(d
->depth_bounds
.min
));
944 radeon_set_context_reg(cmd_buffer
->cs
, R_028024_DB_DEPTH_BOUNDS_MAX
,
945 fui(d
->depth_bounds
.max
));
949 radv_emit_depth_bias(struct radv_cmd_buffer
*cmd_buffer
)
951 struct radv_dynamic_state
*d
= &cmd_buffer
->state
.dynamic
;
952 unsigned slope
= fui(d
->depth_bias
.slope
* 16.0f
);
953 unsigned bias
= fui(d
->depth_bias
.bias
* cmd_buffer
->state
.offset_scale
);
956 radeon_set_context_reg_seq(cmd_buffer
->cs
,
957 R_028B7C_PA_SU_POLY_OFFSET_CLAMP
, 5);
958 radeon_emit(cmd_buffer
->cs
, fui(d
->depth_bias
.clamp
)); /* CLAMP */
959 radeon_emit(cmd_buffer
->cs
, slope
); /* FRONT SCALE */
960 radeon_emit(cmd_buffer
->cs
, bias
); /* FRONT OFFSET */
961 radeon_emit(cmd_buffer
->cs
, slope
); /* BACK SCALE */
962 radeon_emit(cmd_buffer
->cs
, bias
); /* BACK OFFSET */
966 radv_emit_fb_color_state(struct radv_cmd_buffer
*cmd_buffer
,
968 struct radv_attachment_info
*att
,
969 struct radv_image
*image
,
970 VkImageLayout layout
)
972 bool is_vi
= cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= VI
;
973 struct radv_color_buffer_info
*cb
= &att
->cb
;
974 uint32_t cb_color_info
= cb
->cb_color_info
;
976 if (!radv_layout_dcc_compressed(image
, layout
,
977 radv_image_queue_family_mask(image
,
978 cmd_buffer
->queue_family_index
,
979 cmd_buffer
->queue_family_index
))) {
980 cb_color_info
&= C_028C70_DCC_ENABLE
;
983 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
984 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028C60_CB_COLOR0_BASE
+ index
* 0x3c, 11);
985 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_base
);
986 radeon_emit(cmd_buffer
->cs
, S_028C64_BASE_256B(cb
->cb_color_base
>> 32));
987 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_attrib2
);
988 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_view
);
989 radeon_emit(cmd_buffer
->cs
, cb_color_info
);
990 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_attrib
);
991 radeon_emit(cmd_buffer
->cs
, cb
->cb_dcc_control
);
992 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_cmask
);
993 radeon_emit(cmd_buffer
->cs
, S_028C80_BASE_256B(cb
->cb_color_cmask
>> 32));
994 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_fmask
);
995 radeon_emit(cmd_buffer
->cs
, S_028C88_BASE_256B(cb
->cb_color_fmask
>> 32));
997 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028C94_CB_COLOR0_DCC_BASE
+ index
* 0x3c, 2);
998 radeon_emit(cmd_buffer
->cs
, cb
->cb_dcc_base
);
999 radeon_emit(cmd_buffer
->cs
, S_028C98_BASE_256B(cb
->cb_dcc_base
>> 32));
1001 radeon_set_context_reg(cmd_buffer
->cs
, R_0287A0_CB_MRT0_EPITCH
+ index
* 4,
1002 S_0287A0_EPITCH(att
->attachment
->image
->surface
.u
.gfx9
.surf
.epitch
));
1004 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028C60_CB_COLOR0_BASE
+ index
* 0x3c, 11);
1005 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_base
);
1006 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_pitch
);
1007 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_slice
);
1008 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_view
);
1009 radeon_emit(cmd_buffer
->cs
, cb_color_info
);
1010 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_attrib
);
1011 radeon_emit(cmd_buffer
->cs
, cb
->cb_dcc_control
);
1012 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_cmask
);
1013 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_cmask_slice
);
1014 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_fmask
);
1015 radeon_emit(cmd_buffer
->cs
, cb
->cb_color_fmask_slice
);
1017 if (is_vi
) { /* DCC BASE */
1018 radeon_set_context_reg(cmd_buffer
->cs
, R_028C94_CB_COLOR0_DCC_BASE
+ index
* 0x3c, cb
->cb_dcc_base
);
1024 radv_emit_fb_ds_state(struct radv_cmd_buffer
*cmd_buffer
,
1025 struct radv_ds_buffer_info
*ds
,
1026 struct radv_image
*image
,
1027 VkImageLayout layout
)
1029 uint32_t db_z_info
= ds
->db_z_info
;
1030 uint32_t db_stencil_info
= ds
->db_stencil_info
;
1032 if (!radv_layout_has_htile(image
, layout
,
1033 radv_image_queue_family_mask(image
,
1034 cmd_buffer
->queue_family_index
,
1035 cmd_buffer
->queue_family_index
))) {
1036 db_z_info
&= C_028040_TILE_SURFACE_ENABLE
;
1037 db_stencil_info
|= S_028044_TILE_STENCIL_DISABLE(1);
1040 radeon_set_context_reg(cmd_buffer
->cs
, R_028008_DB_DEPTH_VIEW
, ds
->db_depth_view
);
1041 radeon_set_context_reg(cmd_buffer
->cs
, R_028ABC_DB_HTILE_SURFACE
, ds
->db_htile_surface
);
1044 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1045 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028014_DB_HTILE_DATA_BASE
, 3);
1046 radeon_emit(cmd_buffer
->cs
, ds
->db_htile_data_base
);
1047 radeon_emit(cmd_buffer
->cs
, S_028018_BASE_HI(ds
->db_htile_data_base
>> 32));
1048 radeon_emit(cmd_buffer
->cs
, ds
->db_depth_size
);
1050 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028038_DB_Z_INFO
, 10);
1051 radeon_emit(cmd_buffer
->cs
, db_z_info
); /* DB_Z_INFO */
1052 radeon_emit(cmd_buffer
->cs
, db_stencil_info
); /* DB_STENCIL_INFO */
1053 radeon_emit(cmd_buffer
->cs
, ds
->db_z_read_base
); /* DB_Z_READ_BASE */
1054 radeon_emit(cmd_buffer
->cs
, S_028044_BASE_HI(ds
->db_z_read_base
>> 32)); /* DB_Z_READ_BASE_HI */
1055 radeon_emit(cmd_buffer
->cs
, ds
->db_stencil_read_base
); /* DB_STENCIL_READ_BASE */
1056 radeon_emit(cmd_buffer
->cs
, S_02804C_BASE_HI(ds
->db_stencil_read_base
>> 32)); /* DB_STENCIL_READ_BASE_HI */
1057 radeon_emit(cmd_buffer
->cs
, ds
->db_z_write_base
); /* DB_Z_WRITE_BASE */
1058 radeon_emit(cmd_buffer
->cs
, S_028054_BASE_HI(ds
->db_z_write_base
>> 32)); /* DB_Z_WRITE_BASE_HI */
1059 radeon_emit(cmd_buffer
->cs
, ds
->db_stencil_write_base
); /* DB_STENCIL_WRITE_BASE */
1060 radeon_emit(cmd_buffer
->cs
, S_02805C_BASE_HI(ds
->db_stencil_write_base
>> 32)); /* DB_STENCIL_WRITE_BASE_HI */
1062 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028068_DB_Z_INFO2
, 2);
1063 radeon_emit(cmd_buffer
->cs
, ds
->db_z_info2
);
1064 radeon_emit(cmd_buffer
->cs
, ds
->db_stencil_info2
);
1066 radeon_set_context_reg(cmd_buffer
->cs
, R_028014_DB_HTILE_DATA_BASE
, ds
->db_htile_data_base
);
1068 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_02803C_DB_DEPTH_INFO
, 9);
1069 radeon_emit(cmd_buffer
->cs
, ds
->db_depth_info
); /* R_02803C_DB_DEPTH_INFO */
1070 radeon_emit(cmd_buffer
->cs
, db_z_info
); /* R_028040_DB_Z_INFO */
1071 radeon_emit(cmd_buffer
->cs
, db_stencil_info
); /* R_028044_DB_STENCIL_INFO */
1072 radeon_emit(cmd_buffer
->cs
, ds
->db_z_read_base
); /* R_028048_DB_Z_READ_BASE */
1073 radeon_emit(cmd_buffer
->cs
, ds
->db_stencil_read_base
); /* R_02804C_DB_STENCIL_READ_BASE */
1074 radeon_emit(cmd_buffer
->cs
, ds
->db_z_write_base
); /* R_028050_DB_Z_WRITE_BASE */
1075 radeon_emit(cmd_buffer
->cs
, ds
->db_stencil_write_base
); /* R_028054_DB_STENCIL_WRITE_BASE */
1076 radeon_emit(cmd_buffer
->cs
, ds
->db_depth_size
); /* R_028058_DB_DEPTH_SIZE */
1077 radeon_emit(cmd_buffer
->cs
, ds
->db_depth_slice
); /* R_02805C_DB_DEPTH_SLICE */
1081 radeon_set_context_reg(cmd_buffer
->cs
, R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL
,
1082 ds
->pa_su_poly_offset_db_fmt_cntl
);
1086 radv_set_depth_clear_regs(struct radv_cmd_buffer
*cmd_buffer
,
1087 struct radv_image
*image
,
1088 VkClearDepthStencilValue ds_clear_value
,
1089 VkImageAspectFlags aspects
)
1091 uint64_t va
= radv_buffer_get_va(image
->bo
);
1092 va
+= image
->offset
+ image
->clear_value_offset
;
1093 unsigned reg_offset
= 0, reg_count
= 0;
1095 assert(radv_image_has_htile(image
));
1097 if (aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
1103 if (aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
)
1106 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_WRITE_DATA
, 2 + reg_count
, 0));
1107 radeon_emit(cmd_buffer
->cs
, S_370_DST_SEL(V_370_MEM_ASYNC
) |
1108 S_370_WR_CONFIRM(1) |
1109 S_370_ENGINE_SEL(V_370_PFP
));
1110 radeon_emit(cmd_buffer
->cs
, va
);
1111 radeon_emit(cmd_buffer
->cs
, va
>> 32);
1112 if (aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
)
1113 radeon_emit(cmd_buffer
->cs
, ds_clear_value
.stencil
);
1114 if (aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
)
1115 radeon_emit(cmd_buffer
->cs
, fui(ds_clear_value
.depth
));
1117 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028028_DB_STENCIL_CLEAR
+ 4 * reg_offset
, reg_count
);
1118 if (aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
)
1119 radeon_emit(cmd_buffer
->cs
, ds_clear_value
.stencil
); /* R_028028_DB_STENCIL_CLEAR */
1120 if (aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
)
1121 radeon_emit(cmd_buffer
->cs
, fui(ds_clear_value
.depth
)); /* R_02802C_DB_DEPTH_CLEAR */
1125 radv_load_depth_clear_regs(struct radv_cmd_buffer
*cmd_buffer
,
1126 struct radv_image
*image
)
1128 VkImageAspectFlags aspects
= vk_format_aspects(image
->vk_format
);
1129 uint64_t va
= radv_buffer_get_va(image
->bo
);
1130 va
+= image
->offset
+ image
->clear_value_offset
;
1131 unsigned reg_offset
= 0, reg_count
= 0;
1133 if (!radv_image_has_htile(image
))
1136 if (aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
1142 if (aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
)
1145 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_COPY_DATA
, 4, 0));
1146 radeon_emit(cmd_buffer
->cs
, COPY_DATA_SRC_SEL(COPY_DATA_MEM
) |
1147 COPY_DATA_DST_SEL(COPY_DATA_REG
) |
1148 (reg_count
== 2 ? COPY_DATA_COUNT_SEL
: 0));
1149 radeon_emit(cmd_buffer
->cs
, va
);
1150 radeon_emit(cmd_buffer
->cs
, va
>> 32);
1151 radeon_emit(cmd_buffer
->cs
, (R_028028_DB_STENCIL_CLEAR
+ 4 * reg_offset
) >> 2);
1152 radeon_emit(cmd_buffer
->cs
, 0);
1154 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, 0));
1155 radeon_emit(cmd_buffer
->cs
, 0);
1159 *with DCC some colors don't require CMASK elimiation before being
1160 * used as a texture. This sets a predicate value to determine if the
1161 * cmask eliminate is required.
1164 radv_set_dcc_need_cmask_elim_pred(struct radv_cmd_buffer
*cmd_buffer
,
1165 struct radv_image
*image
,
1168 uint64_t pred_val
= value
;
1169 uint64_t va
= radv_buffer_get_va(image
->bo
);
1170 va
+= image
->offset
+ image
->dcc_pred_offset
;
1172 assert(radv_image_has_dcc(image
));
1174 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_WRITE_DATA
, 4, 0));
1175 radeon_emit(cmd_buffer
->cs
, S_370_DST_SEL(V_370_MEM_ASYNC
) |
1176 S_370_WR_CONFIRM(1) |
1177 S_370_ENGINE_SEL(V_370_PFP
));
1178 radeon_emit(cmd_buffer
->cs
, va
);
1179 radeon_emit(cmd_buffer
->cs
, va
>> 32);
1180 radeon_emit(cmd_buffer
->cs
, pred_val
);
1181 radeon_emit(cmd_buffer
->cs
, pred_val
>> 32);
1185 radv_set_color_clear_regs(struct radv_cmd_buffer
*cmd_buffer
,
1186 struct radv_image
*image
,
1188 uint32_t color_values
[2])
1190 uint64_t va
= radv_buffer_get_va(image
->bo
);
1191 va
+= image
->offset
+ image
->clear_value_offset
;
1193 assert(radv_image_has_cmask(image
) || radv_image_has_dcc(image
));
1195 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_WRITE_DATA
, 4, 0));
1196 radeon_emit(cmd_buffer
->cs
, S_370_DST_SEL(V_370_MEM_ASYNC
) |
1197 S_370_WR_CONFIRM(1) |
1198 S_370_ENGINE_SEL(V_370_PFP
));
1199 radeon_emit(cmd_buffer
->cs
, va
);
1200 radeon_emit(cmd_buffer
->cs
, va
>> 32);
1201 radeon_emit(cmd_buffer
->cs
, color_values
[0]);
1202 radeon_emit(cmd_buffer
->cs
, color_values
[1]);
1204 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028C8C_CB_COLOR0_CLEAR_WORD0
+ idx
* 0x3c, 2);
1205 radeon_emit(cmd_buffer
->cs
, color_values
[0]);
1206 radeon_emit(cmd_buffer
->cs
, color_values
[1]);
1210 radv_load_color_clear_regs(struct radv_cmd_buffer
*cmd_buffer
,
1211 struct radv_image
*image
,
1214 uint64_t va
= radv_buffer_get_va(image
->bo
);
1215 va
+= image
->offset
+ image
->clear_value_offset
;
1217 if (!radv_image_has_cmask(image
) && !radv_image_has_dcc(image
))
1220 uint32_t reg
= R_028C8C_CB_COLOR0_CLEAR_WORD0
+ idx
* 0x3c;
1222 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_COPY_DATA
, 4, cmd_buffer
->state
.predicating
));
1223 radeon_emit(cmd_buffer
->cs
, COPY_DATA_SRC_SEL(COPY_DATA_MEM
) |
1224 COPY_DATA_DST_SEL(COPY_DATA_REG
) |
1225 COPY_DATA_COUNT_SEL
);
1226 radeon_emit(cmd_buffer
->cs
, va
);
1227 radeon_emit(cmd_buffer
->cs
, va
>> 32);
1228 radeon_emit(cmd_buffer
->cs
, reg
>> 2);
1229 radeon_emit(cmd_buffer
->cs
, 0);
1231 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_PFP_SYNC_ME
, 0, cmd_buffer
->state
.predicating
));
1232 radeon_emit(cmd_buffer
->cs
, 0);
1236 radv_emit_framebuffer_state(struct radv_cmd_buffer
*cmd_buffer
)
1239 struct radv_framebuffer
*framebuffer
= cmd_buffer
->state
.framebuffer
;
1240 const struct radv_subpass
*subpass
= cmd_buffer
->state
.subpass
;
1242 /* this may happen for inherited secondary recording */
1246 for (i
= 0; i
< 8; ++i
) {
1247 if (i
>= subpass
->color_count
|| subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
) {
1248 radeon_set_context_reg(cmd_buffer
->cs
, R_028C70_CB_COLOR0_INFO
+ i
* 0x3C,
1249 S_028C70_FORMAT(V_028C70_COLOR_INVALID
));
1253 int idx
= subpass
->color_attachments
[i
].attachment
;
1254 struct radv_attachment_info
*att
= &framebuffer
->attachments
[idx
];
1255 struct radv_image
*image
= att
->attachment
->image
;
1256 VkImageLayout layout
= subpass
->color_attachments
[i
].layout
;
1258 radv_cs_add_buffer(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, att
->attachment
->bo
, 8);
1260 assert(att
->attachment
->aspect_mask
& VK_IMAGE_ASPECT_COLOR_BIT
);
1261 radv_emit_fb_color_state(cmd_buffer
, i
, att
, image
, layout
);
1263 radv_load_color_clear_regs(cmd_buffer
, image
, i
);
1266 if(subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
1267 int idx
= subpass
->depth_stencil_attachment
.attachment
;
1268 VkImageLayout layout
= subpass
->depth_stencil_attachment
.layout
;
1269 struct radv_attachment_info
*att
= &framebuffer
->attachments
[idx
];
1270 struct radv_image
*image
= att
->attachment
->image
;
1271 radv_cs_add_buffer(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, att
->attachment
->bo
, 8);
1272 MAYBE_UNUSED
uint32_t queue_mask
= radv_image_queue_family_mask(image
,
1273 cmd_buffer
->queue_family_index
,
1274 cmd_buffer
->queue_family_index
);
1275 /* We currently don't support writing decompressed HTILE */
1276 assert(radv_layout_has_htile(image
, layout
, queue_mask
) ==
1277 radv_layout_is_htile_compressed(image
, layout
, queue_mask
));
1279 radv_emit_fb_ds_state(cmd_buffer
, &att
->ds
, image
, layout
);
1281 if (att
->ds
.offset_scale
!= cmd_buffer
->state
.offset_scale
) {
1282 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS
;
1283 cmd_buffer
->state
.offset_scale
= att
->ds
.offset_scale
;
1285 radv_load_depth_clear_regs(cmd_buffer
, image
);
1287 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1288 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028038_DB_Z_INFO
, 2);
1290 radeon_set_context_reg_seq(cmd_buffer
->cs
, R_028040_DB_Z_INFO
, 2);
1292 radeon_emit(cmd_buffer
->cs
, S_028040_FORMAT(V_028040_Z_INVALID
)); /* DB_Z_INFO */
1293 radeon_emit(cmd_buffer
->cs
, S_028044_FORMAT(V_028044_STENCIL_INVALID
)); /* DB_STENCIL_INFO */
1295 radeon_set_context_reg(cmd_buffer
->cs
, R_028208_PA_SC_WINDOW_SCISSOR_BR
,
1296 S_028208_BR_X(framebuffer
->width
) |
1297 S_028208_BR_Y(framebuffer
->height
));
1299 if (cmd_buffer
->device
->dfsm_allowed
) {
1300 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_EVENT_WRITE
, 0, 0));
1301 radeon_emit(cmd_buffer
->cs
, EVENT_TYPE(V_028A90_BREAK_BATCH
) | EVENT_INDEX(0));
1304 cmd_buffer
->state
.dirty
&= ~RADV_CMD_DIRTY_FRAMEBUFFER
;
1308 radv_emit_index_buffer(struct radv_cmd_buffer
*cmd_buffer
)
1310 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
1311 struct radv_cmd_state
*state
= &cmd_buffer
->state
;
1313 if (state
->index_type
!= state
->last_index_type
) {
1314 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1315 radeon_set_uconfig_reg_idx(cs
, R_03090C_VGT_INDEX_TYPE
,
1316 2, state
->index_type
);
1318 radeon_emit(cs
, PKT3(PKT3_INDEX_TYPE
, 0, 0));
1319 radeon_emit(cs
, state
->index_type
);
1322 state
->last_index_type
= state
->index_type
;
1325 radeon_emit(cs
, PKT3(PKT3_INDEX_BASE
, 1, 0));
1326 radeon_emit(cs
, state
->index_va
);
1327 radeon_emit(cs
, state
->index_va
>> 32);
1329 radeon_emit(cs
, PKT3(PKT3_INDEX_BUFFER_SIZE
, 0, 0));
1330 radeon_emit(cs
, state
->max_index_count
);
1332 cmd_buffer
->state
.dirty
&= ~RADV_CMD_DIRTY_INDEX_BUFFER
;
1335 void radv_set_db_count_control(struct radv_cmd_buffer
*cmd_buffer
)
1337 bool has_perfect_queries
= cmd_buffer
->state
.perfect_occlusion_queries_enabled
;
1338 struct radv_pipeline
*pipeline
= cmd_buffer
->state
.pipeline
;
1339 uint32_t pa_sc_mode_cntl_1
=
1340 pipeline
? pipeline
->graphics
.ms
.pa_sc_mode_cntl_1
: 0;
1341 uint32_t db_count_control
;
1343 if(!cmd_buffer
->state
.active_occlusion_queries
) {
1344 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1345 if (G_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(pa_sc_mode_cntl_1
) &&
1346 pipeline
->graphics
.disable_out_of_order_rast_for_occlusion
&&
1347 has_perfect_queries
) {
1348 /* Re-enable out-of-order rasterization if the
1349 * bound pipeline supports it and if it's has
1350 * been disabled before starting any perfect
1351 * occlusion queries.
1353 radeon_set_context_reg(cmd_buffer
->cs
,
1354 R_028A4C_PA_SC_MODE_CNTL_1
,
1357 db_count_control
= 0;
1359 db_count_control
= S_028004_ZPASS_INCREMENT_DISABLE(1);
1362 const struct radv_subpass
*subpass
= cmd_buffer
->state
.subpass
;
1363 uint32_t sample_rate
= subpass
? util_logbase2(subpass
->max_sample_count
) : 0;
1365 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1367 S_028004_PERFECT_ZPASS_COUNTS(has_perfect_queries
) |
1368 S_028004_SAMPLE_RATE(sample_rate
) |
1369 S_028004_ZPASS_ENABLE(1) |
1370 S_028004_SLICE_EVEN_ENABLE(1) |
1371 S_028004_SLICE_ODD_ENABLE(1);
1373 if (G_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(pa_sc_mode_cntl_1
) &&
1374 pipeline
->graphics
.disable_out_of_order_rast_for_occlusion
&&
1375 has_perfect_queries
) {
1376 /* If the bound pipeline has enabled
1377 * out-of-order rasterization, we should
1378 * disable it before starting any perfect
1379 * occlusion queries.
1381 pa_sc_mode_cntl_1
&= C_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE
;
1383 radeon_set_context_reg(cmd_buffer
->cs
,
1384 R_028A4C_PA_SC_MODE_CNTL_1
,
1388 db_count_control
= S_028004_PERFECT_ZPASS_COUNTS(1) |
1389 S_028004_SAMPLE_RATE(sample_rate
);
1393 radeon_set_context_reg(cmd_buffer
->cs
, R_028004_DB_COUNT_CONTROL
, db_count_control
);
1397 radv_cmd_buffer_flush_dynamic_state(struct radv_cmd_buffer
*cmd_buffer
)
1399 uint32_t states
= cmd_buffer
->state
.dirty
& cmd_buffer
->state
.emitted_pipeline
->graphics
.needed_dynamic_state
;
1401 if (states
& (RADV_CMD_DIRTY_DYNAMIC_VIEWPORT
))
1402 radv_emit_viewport(cmd_buffer
);
1404 if (states
& (RADV_CMD_DIRTY_DYNAMIC_SCISSOR
| RADV_CMD_DIRTY_DYNAMIC_VIEWPORT
))
1405 radv_emit_scissor(cmd_buffer
);
1407 if (states
& RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH
)
1408 radv_emit_line_width(cmd_buffer
);
1410 if (states
& RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS
)
1411 radv_emit_blend_constants(cmd_buffer
);
1413 if (states
& (RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE
|
1414 RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK
|
1415 RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK
))
1416 radv_emit_stencil(cmd_buffer
);
1418 if (states
& RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS
)
1419 radv_emit_depth_bounds(cmd_buffer
);
1421 if (states
& RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS
)
1422 radv_emit_depth_bias(cmd_buffer
);
1424 if (states
& RADV_CMD_DIRTY_DYNAMIC_DISCARD_RECTANGLE
)
1425 radv_emit_discard_rectangle(cmd_buffer
);
1427 cmd_buffer
->state
.dirty
&= ~states
;
1431 emit_stage_descriptor_set_userdata(struct radv_cmd_buffer
*cmd_buffer
,
1432 struct radv_pipeline
*pipeline
,
1435 gl_shader_stage stage
)
1437 struct radv_userdata_info
*desc_set_loc
= &pipeline
->shaders
[stage
]->info
.user_sgprs_locs
.descriptor_sets
[idx
];
1438 uint32_t base_reg
= pipeline
->user_data_0
[stage
];
1440 if (desc_set_loc
->sgpr_idx
== -1 || desc_set_loc
->indirect
)
1443 assert(!desc_set_loc
->indirect
);
1444 assert(desc_set_loc
->num_sgprs
== 2);
1445 radeon_set_sh_reg_seq(cmd_buffer
->cs
,
1446 base_reg
+ desc_set_loc
->sgpr_idx
* 4, 2);
1447 radeon_emit(cmd_buffer
->cs
, va
);
1448 radeon_emit(cmd_buffer
->cs
, va
>> 32);
1452 radv_emit_descriptor_set_userdata(struct radv_cmd_buffer
*cmd_buffer
,
1453 VkShaderStageFlags stages
,
1454 struct radv_descriptor_set
*set
,
1457 if (cmd_buffer
->state
.pipeline
) {
1458 radv_foreach_stage(stage
, stages
) {
1459 if (cmd_buffer
->state
.pipeline
->shaders
[stage
])
1460 emit_stage_descriptor_set_userdata(cmd_buffer
, cmd_buffer
->state
.pipeline
,
1466 if (cmd_buffer
->state
.compute_pipeline
&& (stages
& VK_SHADER_STAGE_COMPUTE_BIT
))
1467 emit_stage_descriptor_set_userdata(cmd_buffer
, cmd_buffer
->state
.compute_pipeline
,
1469 MESA_SHADER_COMPUTE
);
1473 radv_flush_push_descriptors(struct radv_cmd_buffer
*cmd_buffer
,
1474 VkPipelineBindPoint bind_point
)
1476 struct radv_descriptor_state
*descriptors_state
=
1477 radv_get_descriptors_state(cmd_buffer
, bind_point
);
1478 struct radv_descriptor_set
*set
= &descriptors_state
->push_set
.set
;
1481 if (!radv_cmd_buffer_upload_data(cmd_buffer
, set
->size
, 32,
1486 set
->va
= radv_buffer_get_va(cmd_buffer
->upload
.upload_bo
);
1487 set
->va
+= bo_offset
;
1491 radv_flush_indirect_descriptor_sets(struct radv_cmd_buffer
*cmd_buffer
,
1492 VkPipelineBindPoint bind_point
)
1494 struct radv_descriptor_state
*descriptors_state
=
1495 radv_get_descriptors_state(cmd_buffer
, bind_point
);
1496 uint32_t size
= MAX_SETS
* 2 * 4;
1500 if (!radv_cmd_buffer_upload_alloc(cmd_buffer
, size
,
1501 256, &offset
, &ptr
))
1504 for (unsigned i
= 0; i
< MAX_SETS
; i
++) {
1505 uint32_t *uptr
= ((uint32_t *)ptr
) + i
* 2;
1506 uint64_t set_va
= 0;
1507 struct radv_descriptor_set
*set
= descriptors_state
->sets
[i
];
1508 if (descriptors_state
->valid
& (1u << i
))
1510 uptr
[0] = set_va
& 0xffffffff;
1511 uptr
[1] = set_va
>> 32;
1514 uint64_t va
= radv_buffer_get_va(cmd_buffer
->upload
.upload_bo
);
1517 if (cmd_buffer
->state
.pipeline
) {
1518 if (cmd_buffer
->state
.pipeline
->shaders
[MESA_SHADER_VERTEX
])
1519 radv_emit_userdata_address(cmd_buffer
, cmd_buffer
->state
.pipeline
, MESA_SHADER_VERTEX
,
1520 AC_UD_INDIRECT_DESCRIPTOR_SETS
, va
);
1522 if (cmd_buffer
->state
.pipeline
->shaders
[MESA_SHADER_FRAGMENT
])
1523 radv_emit_userdata_address(cmd_buffer
, cmd_buffer
->state
.pipeline
, MESA_SHADER_FRAGMENT
,
1524 AC_UD_INDIRECT_DESCRIPTOR_SETS
, va
);
1526 if (radv_pipeline_has_gs(cmd_buffer
->state
.pipeline
))
1527 radv_emit_userdata_address(cmd_buffer
, cmd_buffer
->state
.pipeline
, MESA_SHADER_GEOMETRY
,
1528 AC_UD_INDIRECT_DESCRIPTOR_SETS
, va
);
1530 if (radv_pipeline_has_tess(cmd_buffer
->state
.pipeline
))
1531 radv_emit_userdata_address(cmd_buffer
, cmd_buffer
->state
.pipeline
, MESA_SHADER_TESS_CTRL
,
1532 AC_UD_INDIRECT_DESCRIPTOR_SETS
, va
);
1534 if (radv_pipeline_has_tess(cmd_buffer
->state
.pipeline
))
1535 radv_emit_userdata_address(cmd_buffer
, cmd_buffer
->state
.pipeline
, MESA_SHADER_TESS_EVAL
,
1536 AC_UD_INDIRECT_DESCRIPTOR_SETS
, va
);
1539 if (cmd_buffer
->state
.compute_pipeline
)
1540 radv_emit_userdata_address(cmd_buffer
, cmd_buffer
->state
.compute_pipeline
, MESA_SHADER_COMPUTE
,
1541 AC_UD_INDIRECT_DESCRIPTOR_SETS
, va
);
1545 radv_flush_descriptors(struct radv_cmd_buffer
*cmd_buffer
,
1546 VkShaderStageFlags stages
)
1548 VkPipelineBindPoint bind_point
= stages
& VK_SHADER_STAGE_COMPUTE_BIT
?
1549 VK_PIPELINE_BIND_POINT_COMPUTE
:
1550 VK_PIPELINE_BIND_POINT_GRAPHICS
;
1551 struct radv_descriptor_state
*descriptors_state
=
1552 radv_get_descriptors_state(cmd_buffer
, bind_point
);
1555 if (!descriptors_state
->dirty
)
1558 if (descriptors_state
->push_dirty
)
1559 radv_flush_push_descriptors(cmd_buffer
, bind_point
);
1561 if ((cmd_buffer
->state
.pipeline
&& cmd_buffer
->state
.pipeline
->need_indirect_descriptor_sets
) ||
1562 (cmd_buffer
->state
.compute_pipeline
&& cmd_buffer
->state
.compute_pipeline
->need_indirect_descriptor_sets
)) {
1563 radv_flush_indirect_descriptor_sets(cmd_buffer
, bind_point
);
1566 MAYBE_UNUSED
unsigned cdw_max
= radeon_check_space(cmd_buffer
->device
->ws
,
1568 MAX_SETS
* MESA_SHADER_STAGES
* 4);
1570 for_each_bit(i
, descriptors_state
->dirty
) {
1571 struct radv_descriptor_set
*set
= descriptors_state
->sets
[i
];
1572 if (!(descriptors_state
->valid
& (1u << i
)))
1575 radv_emit_descriptor_set_userdata(cmd_buffer
, stages
, set
, i
);
1577 descriptors_state
->dirty
= 0;
1578 descriptors_state
->push_dirty
= false;
1580 if (unlikely(cmd_buffer
->device
->trace_bo
))
1581 radv_save_descriptors(cmd_buffer
, bind_point
);
1583 assert(cmd_buffer
->cs
->cdw
<= cdw_max
);
1587 radv_flush_constants(struct radv_cmd_buffer
*cmd_buffer
,
1588 VkShaderStageFlags stages
)
1590 struct radv_pipeline
*pipeline
= stages
& VK_SHADER_STAGE_COMPUTE_BIT
1591 ? cmd_buffer
->state
.compute_pipeline
1592 : cmd_buffer
->state
.pipeline
;
1593 struct radv_pipeline_layout
*layout
= pipeline
->layout
;
1598 stages
&= cmd_buffer
->push_constant_stages
;
1600 (!layout
->push_constant_size
&& !layout
->dynamic_offset_count
))
1603 if (!radv_cmd_buffer_upload_alloc(cmd_buffer
, layout
->push_constant_size
+
1604 16 * layout
->dynamic_offset_count
,
1605 256, &offset
, &ptr
))
1608 memcpy(ptr
, cmd_buffer
->push_constants
, layout
->push_constant_size
);
1609 memcpy((char*)ptr
+ layout
->push_constant_size
, cmd_buffer
->dynamic_buffers
,
1610 16 * layout
->dynamic_offset_count
);
1612 va
= radv_buffer_get_va(cmd_buffer
->upload
.upload_bo
);
1615 MAYBE_UNUSED
unsigned cdw_max
= radeon_check_space(cmd_buffer
->device
->ws
,
1616 cmd_buffer
->cs
, MESA_SHADER_STAGES
* 4);
1618 radv_foreach_stage(stage
, stages
) {
1619 if (pipeline
->shaders
[stage
]) {
1620 radv_emit_userdata_address(cmd_buffer
, pipeline
, stage
,
1621 AC_UD_PUSH_CONSTANTS
, va
);
1625 cmd_buffer
->push_constant_stages
&= ~stages
;
1626 assert(cmd_buffer
->cs
->cdw
<= cdw_max
);
1630 radv_flush_vertex_descriptors(struct radv_cmd_buffer
*cmd_buffer
,
1631 bool pipeline_is_dirty
)
1633 if ((pipeline_is_dirty
||
1634 (cmd_buffer
->state
.dirty
& RADV_CMD_DIRTY_VERTEX_BUFFER
)) &&
1635 cmd_buffer
->state
.pipeline
->vertex_elements
.count
&&
1636 radv_get_vertex_shader(cmd_buffer
->state
.pipeline
)->info
.info
.vs
.has_vertex_buffers
) {
1637 struct radv_vertex_elements_info
*velems
= &cmd_buffer
->state
.pipeline
->vertex_elements
;
1641 uint32_t count
= velems
->count
;
1644 /* allocate some descriptor state for vertex buffers */
1645 if (!radv_cmd_buffer_upload_alloc(cmd_buffer
, count
* 16, 256,
1646 &vb_offset
, &vb_ptr
))
1649 for (i
= 0; i
< count
; i
++) {
1650 uint32_t *desc
= &((uint32_t *)vb_ptr
)[i
* 4];
1652 int vb
= velems
->binding
[i
];
1653 struct radv_buffer
*buffer
= cmd_buffer
->vertex_bindings
[vb
].buffer
;
1654 uint32_t stride
= cmd_buffer
->state
.pipeline
->binding_stride
[vb
];
1656 va
= radv_buffer_get_va(buffer
->bo
);
1658 offset
= cmd_buffer
->vertex_bindings
[vb
].offset
+ velems
->offset
[i
];
1659 va
+= offset
+ buffer
->offset
;
1661 desc
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32) | S_008F04_STRIDE(stride
);
1662 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
<= CIK
&& stride
)
1663 desc
[2] = (buffer
->size
- offset
- velems
->format_size
[i
]) / stride
+ 1;
1665 desc
[2] = buffer
->size
- offset
;
1666 desc
[3] = velems
->rsrc_word3
[i
];
1669 va
= radv_buffer_get_va(cmd_buffer
->upload
.upload_bo
);
1672 radv_emit_userdata_address(cmd_buffer
, cmd_buffer
->state
.pipeline
, MESA_SHADER_VERTEX
,
1673 AC_UD_VS_VERTEX_BUFFERS
, va
);
1675 cmd_buffer
->state
.vb_va
= va
;
1676 cmd_buffer
->state
.vb_size
= count
* 16;
1677 cmd_buffer
->state
.prefetch_L2_mask
|= RADV_PREFETCH_VBO_DESCRIPTORS
;
1679 cmd_buffer
->state
.dirty
&= ~RADV_CMD_DIRTY_VERTEX_BUFFER
;
1683 radv_upload_graphics_shader_descriptors(struct radv_cmd_buffer
*cmd_buffer
, bool pipeline_is_dirty
)
1685 radv_flush_vertex_descriptors(cmd_buffer
, pipeline_is_dirty
);
1686 radv_flush_descriptors(cmd_buffer
, VK_SHADER_STAGE_ALL_GRAPHICS
);
1687 radv_flush_constants(cmd_buffer
, VK_SHADER_STAGE_ALL_GRAPHICS
);
1691 radv_emit_draw_registers(struct radv_cmd_buffer
*cmd_buffer
, bool indexed_draw
,
1692 bool instanced_draw
, bool indirect_draw
,
1693 uint32_t draw_vertex_count
)
1695 struct radeon_info
*info
= &cmd_buffer
->device
->physical_device
->rad_info
;
1696 struct radv_cmd_state
*state
= &cmd_buffer
->state
;
1697 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
1698 uint32_t ia_multi_vgt_param
;
1699 int32_t primitive_reset_en
;
1702 ia_multi_vgt_param
=
1703 si_get_ia_multi_vgt_param(cmd_buffer
, instanced_draw
,
1704 indirect_draw
, draw_vertex_count
);
1706 if (state
->last_ia_multi_vgt_param
!= ia_multi_vgt_param
) {
1707 if (info
->chip_class
>= GFX9
) {
1708 radeon_set_uconfig_reg_idx(cs
,
1709 R_030960_IA_MULTI_VGT_PARAM
,
1710 4, ia_multi_vgt_param
);
1711 } else if (info
->chip_class
>= CIK
) {
1712 radeon_set_context_reg_idx(cs
,
1713 R_028AA8_IA_MULTI_VGT_PARAM
,
1714 1, ia_multi_vgt_param
);
1716 radeon_set_context_reg(cs
, R_028AA8_IA_MULTI_VGT_PARAM
,
1717 ia_multi_vgt_param
);
1719 state
->last_ia_multi_vgt_param
= ia_multi_vgt_param
;
1722 /* Primitive restart. */
1723 primitive_reset_en
=
1724 indexed_draw
&& state
->pipeline
->graphics
.prim_restart_enable
;
1726 if (primitive_reset_en
!= state
->last_primitive_reset_en
) {
1727 state
->last_primitive_reset_en
= primitive_reset_en
;
1728 if (info
->chip_class
>= GFX9
) {
1729 radeon_set_uconfig_reg(cs
,
1730 R_03092C_VGT_MULTI_PRIM_IB_RESET_EN
,
1731 primitive_reset_en
);
1733 radeon_set_context_reg(cs
,
1734 R_028A94_VGT_MULTI_PRIM_IB_RESET_EN
,
1735 primitive_reset_en
);
1739 if (primitive_reset_en
) {
1740 uint32_t primitive_reset_index
=
1741 state
->index_type
? 0xffffffffu
: 0xffffu
;
1743 if (primitive_reset_index
!= state
->last_primitive_reset_index
) {
1744 radeon_set_context_reg(cs
,
1745 R_02840C_VGT_MULTI_PRIM_IB_RESET_INDX
,
1746 primitive_reset_index
);
1747 state
->last_primitive_reset_index
= primitive_reset_index
;
1752 static void radv_stage_flush(struct radv_cmd_buffer
*cmd_buffer
,
1753 VkPipelineStageFlags src_stage_mask
)
1755 if (src_stage_mask
& (VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT
|
1756 VK_PIPELINE_STAGE_TRANSFER_BIT
|
1757 VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT
|
1758 VK_PIPELINE_STAGE_ALL_COMMANDS_BIT
)) {
1759 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_CS_PARTIAL_FLUSH
;
1762 if (src_stage_mask
& (VK_PIPELINE_STAGE_TESSELLATION_CONTROL_SHADER_BIT
|
1763 VK_PIPELINE_STAGE_TESSELLATION_EVALUATION_SHADER_BIT
|
1764 VK_PIPELINE_STAGE_GEOMETRY_SHADER_BIT
|
1765 VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT
|
1766 VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT
|
1767 VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT
|
1768 VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT
|
1769 VK_PIPELINE_STAGE_TRANSFER_BIT
|
1770 VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT
|
1771 VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT
|
1772 VK_PIPELINE_STAGE_ALL_COMMANDS_BIT
)) {
1773 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_PS_PARTIAL_FLUSH
;
1774 } else if (src_stage_mask
& (VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT
|
1775 VK_PIPELINE_STAGE_VERTEX_INPUT_BIT
|
1776 VK_PIPELINE_STAGE_VERTEX_SHADER_BIT
)) {
1777 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_VS_PARTIAL_FLUSH
;
1781 static enum radv_cmd_flush_bits
1782 radv_src_access_flush(struct radv_cmd_buffer
*cmd_buffer
,
1783 VkAccessFlags src_flags
)
1785 enum radv_cmd_flush_bits flush_bits
= 0;
1787 for_each_bit(b
, src_flags
) {
1788 switch ((VkAccessFlagBits
)(1 << b
)) {
1789 case VK_ACCESS_SHADER_WRITE_BIT
:
1790 flush_bits
|= RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2
;
1792 case VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT
:
1793 flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
1794 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
;
1796 case VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT
:
1797 flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
1798 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
;
1800 case VK_ACCESS_TRANSFER_WRITE_BIT
:
1801 flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
1802 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
|
1803 RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
1804 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
|
1805 RADV_CMD_FLAG_INV_GLOBAL_L2
;
1814 static enum radv_cmd_flush_bits
1815 radv_dst_access_flush(struct radv_cmd_buffer
*cmd_buffer
,
1816 VkAccessFlags dst_flags
,
1817 struct radv_image
*image
)
1819 enum radv_cmd_flush_bits flush_bits
= 0;
1821 for_each_bit(b
, dst_flags
) {
1822 switch ((VkAccessFlagBits
)(1 << b
)) {
1823 case VK_ACCESS_INDIRECT_COMMAND_READ_BIT
:
1824 case VK_ACCESS_INDEX_READ_BIT
:
1826 case VK_ACCESS_UNIFORM_READ_BIT
:
1827 flush_bits
|= RADV_CMD_FLAG_INV_VMEM_L1
| RADV_CMD_FLAG_INV_SMEM_L1
;
1829 case VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT
:
1830 case VK_ACCESS_SHADER_READ_BIT
:
1831 case VK_ACCESS_TRANSFER_READ_BIT
:
1832 case VK_ACCESS_INPUT_ATTACHMENT_READ_BIT
:
1833 flush_bits
|= RADV_CMD_FLAG_INV_VMEM_L1
|
1834 RADV_CMD_FLAG_INV_GLOBAL_L2
;
1836 case VK_ACCESS_COLOR_ATTACHMENT_READ_BIT
:
1837 /* TODO: change to image && when the image gets passed
1838 * through from the subpass. */
1839 if (!image
|| (image
->usage
& VK_IMAGE_USAGE_STORAGE_BIT
))
1840 flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
1841 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
;
1843 case VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT
:
1844 if (!image
|| (image
->usage
& VK_IMAGE_USAGE_STORAGE_BIT
))
1845 flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
1846 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
;
1855 static void radv_subpass_barrier(struct radv_cmd_buffer
*cmd_buffer
, const struct radv_subpass_barrier
*barrier
)
1857 cmd_buffer
->state
.flush_bits
|= radv_src_access_flush(cmd_buffer
, barrier
->src_access_mask
);
1858 radv_stage_flush(cmd_buffer
, barrier
->src_stage_mask
);
1859 cmd_buffer
->state
.flush_bits
|= radv_dst_access_flush(cmd_buffer
, barrier
->dst_access_mask
,
1863 static void radv_handle_subpass_image_transition(struct radv_cmd_buffer
*cmd_buffer
,
1864 VkAttachmentReference att
)
1866 unsigned idx
= att
.attachment
;
1867 struct radv_image_view
*view
= cmd_buffer
->state
.framebuffer
->attachments
[idx
].attachment
;
1868 VkImageSubresourceRange range
;
1869 range
.aspectMask
= 0;
1870 range
.baseMipLevel
= view
->base_mip
;
1871 range
.levelCount
= 1;
1872 range
.baseArrayLayer
= view
->base_layer
;
1873 range
.layerCount
= cmd_buffer
->state
.framebuffer
->layers
;
1875 radv_handle_image_transition(cmd_buffer
,
1877 cmd_buffer
->state
.attachments
[idx
].current_layout
,
1878 att
.layout
, 0, 0, &range
,
1879 cmd_buffer
->state
.attachments
[idx
].pending_clear_aspects
);
1881 cmd_buffer
->state
.attachments
[idx
].current_layout
= att
.layout
;
1887 radv_cmd_buffer_set_subpass(struct radv_cmd_buffer
*cmd_buffer
,
1888 const struct radv_subpass
*subpass
, bool transitions
)
1891 radv_subpass_barrier(cmd_buffer
, &subpass
->start_barrier
);
1893 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
1894 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
)
1895 radv_handle_subpass_image_transition(cmd_buffer
,
1896 subpass
->color_attachments
[i
]);
1899 for (unsigned i
= 0; i
< subpass
->input_count
; ++i
) {
1900 radv_handle_subpass_image_transition(cmd_buffer
,
1901 subpass
->input_attachments
[i
]);
1904 if (subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
1905 radv_handle_subpass_image_transition(cmd_buffer
,
1906 subpass
->depth_stencil_attachment
);
1910 cmd_buffer
->state
.subpass
= subpass
;
1912 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_FRAMEBUFFER
;
1916 radv_cmd_state_setup_attachments(struct radv_cmd_buffer
*cmd_buffer
,
1917 struct radv_render_pass
*pass
,
1918 const VkRenderPassBeginInfo
*info
)
1920 struct radv_cmd_state
*state
= &cmd_buffer
->state
;
1922 if (pass
->attachment_count
== 0) {
1923 state
->attachments
= NULL
;
1927 state
->attachments
= vk_alloc(&cmd_buffer
->pool
->alloc
,
1928 pass
->attachment_count
*
1929 sizeof(state
->attachments
[0]),
1930 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1931 if (state
->attachments
== NULL
) {
1932 cmd_buffer
->record_result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
1933 return cmd_buffer
->record_result
;
1936 for (uint32_t i
= 0; i
< pass
->attachment_count
; ++i
) {
1937 struct radv_render_pass_attachment
*att
= &pass
->attachments
[i
];
1938 VkImageAspectFlags att_aspects
= vk_format_aspects(att
->format
);
1939 VkImageAspectFlags clear_aspects
= 0;
1941 if (att_aspects
== VK_IMAGE_ASPECT_COLOR_BIT
) {
1942 /* color attachment */
1943 if (att
->load_op
== VK_ATTACHMENT_LOAD_OP_CLEAR
) {
1944 clear_aspects
|= VK_IMAGE_ASPECT_COLOR_BIT
;
1947 /* depthstencil attachment */
1948 if ((att_aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) &&
1949 att
->load_op
== VK_ATTACHMENT_LOAD_OP_CLEAR
) {
1950 clear_aspects
|= VK_IMAGE_ASPECT_DEPTH_BIT
;
1951 if ((att_aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) &&
1952 att
->stencil_load_op
== VK_ATTACHMENT_LOAD_OP_DONT_CARE
)
1953 clear_aspects
|= VK_IMAGE_ASPECT_STENCIL_BIT
;
1955 if ((att_aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) &&
1956 att
->stencil_load_op
== VK_ATTACHMENT_LOAD_OP_CLEAR
) {
1957 clear_aspects
|= VK_IMAGE_ASPECT_STENCIL_BIT
;
1961 state
->attachments
[i
].pending_clear_aspects
= clear_aspects
;
1962 state
->attachments
[i
].cleared_views
= 0;
1963 if (clear_aspects
&& info
) {
1964 assert(info
->clearValueCount
> i
);
1965 state
->attachments
[i
].clear_value
= info
->pClearValues
[i
];
1968 state
->attachments
[i
].current_layout
= att
->initial_layout
;
1974 VkResult
radv_AllocateCommandBuffers(
1976 const VkCommandBufferAllocateInfo
*pAllocateInfo
,
1977 VkCommandBuffer
*pCommandBuffers
)
1979 RADV_FROM_HANDLE(radv_device
, device
, _device
);
1980 RADV_FROM_HANDLE(radv_cmd_pool
, pool
, pAllocateInfo
->commandPool
);
1982 VkResult result
= VK_SUCCESS
;
1985 for (i
= 0; i
< pAllocateInfo
->commandBufferCount
; i
++) {
1987 if (!list_empty(&pool
->free_cmd_buffers
)) {
1988 struct radv_cmd_buffer
*cmd_buffer
= list_first_entry(&pool
->free_cmd_buffers
, struct radv_cmd_buffer
, pool_link
);
1990 list_del(&cmd_buffer
->pool_link
);
1991 list_addtail(&cmd_buffer
->pool_link
, &pool
->cmd_buffers
);
1993 result
= radv_reset_cmd_buffer(cmd_buffer
);
1994 cmd_buffer
->_loader_data
.loaderMagic
= ICD_LOADER_MAGIC
;
1995 cmd_buffer
->level
= pAllocateInfo
->level
;
1997 pCommandBuffers
[i
] = radv_cmd_buffer_to_handle(cmd_buffer
);
1999 result
= radv_create_cmd_buffer(device
, pool
, pAllocateInfo
->level
,
2000 &pCommandBuffers
[i
]);
2002 if (result
!= VK_SUCCESS
)
2006 if (result
!= VK_SUCCESS
) {
2007 radv_FreeCommandBuffers(_device
, pAllocateInfo
->commandPool
,
2008 i
, pCommandBuffers
);
2010 /* From the Vulkan 1.0.66 spec:
2012 * "vkAllocateCommandBuffers can be used to create multiple
2013 * command buffers. If the creation of any of those command
2014 * buffers fails, the implementation must destroy all
2015 * successfully created command buffer objects from this
2016 * command, set all entries of the pCommandBuffers array to
2017 * NULL and return the error."
2019 memset(pCommandBuffers
, 0,
2020 sizeof(*pCommandBuffers
) * pAllocateInfo
->commandBufferCount
);
2026 void radv_FreeCommandBuffers(
2028 VkCommandPool commandPool
,
2029 uint32_t commandBufferCount
,
2030 const VkCommandBuffer
*pCommandBuffers
)
2032 for (uint32_t i
= 0; i
< commandBufferCount
; i
++) {
2033 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, pCommandBuffers
[i
]);
2036 if (cmd_buffer
->pool
) {
2037 list_del(&cmd_buffer
->pool_link
);
2038 list_addtail(&cmd_buffer
->pool_link
, &cmd_buffer
->pool
->free_cmd_buffers
);
2040 radv_cmd_buffer_destroy(cmd_buffer
);
2046 VkResult
radv_ResetCommandBuffer(
2047 VkCommandBuffer commandBuffer
,
2048 VkCommandBufferResetFlags flags
)
2050 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2051 return radv_reset_cmd_buffer(cmd_buffer
);
2054 static void emit_gfx_buffer_state(struct radv_cmd_buffer
*cmd_buffer
)
2056 struct radv_device
*device
= cmd_buffer
->device
;
2057 if (device
->gfx_init
) {
2058 uint64_t va
= radv_buffer_get_va(device
->gfx_init
);
2059 radv_cs_add_buffer(device
->ws
, cmd_buffer
->cs
, device
->gfx_init
, 8);
2060 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_INDIRECT_BUFFER_CIK
, 2, 0));
2061 radeon_emit(cmd_buffer
->cs
, va
);
2062 radeon_emit(cmd_buffer
->cs
, va
>> 32);
2063 radeon_emit(cmd_buffer
->cs
, device
->gfx_init_size_dw
& 0xffff);
2065 si_init_config(cmd_buffer
);
2068 VkResult
radv_BeginCommandBuffer(
2069 VkCommandBuffer commandBuffer
,
2070 const VkCommandBufferBeginInfo
*pBeginInfo
)
2072 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2073 VkResult result
= VK_SUCCESS
;
2075 if (cmd_buffer
->status
!= RADV_CMD_BUFFER_STATUS_INITIAL
) {
2076 /* If the command buffer has already been resetted with
2077 * vkResetCommandBuffer, no need to do it again.
2079 result
= radv_reset_cmd_buffer(cmd_buffer
);
2080 if (result
!= VK_SUCCESS
)
2084 memset(&cmd_buffer
->state
, 0, sizeof(cmd_buffer
->state
));
2085 cmd_buffer
->state
.last_primitive_reset_en
= -1;
2086 cmd_buffer
->state
.last_index_type
= -1;
2087 cmd_buffer
->state
.last_num_instances
= -1;
2088 cmd_buffer
->state
.last_vertex_offset
= -1;
2089 cmd_buffer
->state
.last_first_instance
= -1;
2090 cmd_buffer
->usage_flags
= pBeginInfo
->flags
;
2092 /* setup initial configuration into command buffer */
2093 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
) {
2094 switch (cmd_buffer
->queue_family_index
) {
2095 case RADV_QUEUE_GENERAL
:
2096 emit_gfx_buffer_state(cmd_buffer
);
2098 case RADV_QUEUE_COMPUTE
:
2099 si_init_compute(cmd_buffer
);
2101 case RADV_QUEUE_TRANSFER
:
2107 if (pBeginInfo
->flags
& VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT
) {
2108 assert(pBeginInfo
->pInheritanceInfo
);
2109 cmd_buffer
->state
.framebuffer
= radv_framebuffer_from_handle(pBeginInfo
->pInheritanceInfo
->framebuffer
);
2110 cmd_buffer
->state
.pass
= radv_render_pass_from_handle(pBeginInfo
->pInheritanceInfo
->renderPass
);
2112 struct radv_subpass
*subpass
=
2113 &cmd_buffer
->state
.pass
->subpasses
[pBeginInfo
->pInheritanceInfo
->subpass
];
2115 result
= radv_cmd_state_setup_attachments(cmd_buffer
, cmd_buffer
->state
.pass
, NULL
);
2116 if (result
!= VK_SUCCESS
)
2119 radv_cmd_buffer_set_subpass(cmd_buffer
, subpass
, false);
2122 if (unlikely(cmd_buffer
->device
->trace_bo
))
2123 radv_cmd_buffer_trace_emit(cmd_buffer
);
2125 cmd_buffer
->status
= RADV_CMD_BUFFER_STATUS_RECORDING
;
2130 void radv_CmdBindVertexBuffers(
2131 VkCommandBuffer commandBuffer
,
2132 uint32_t firstBinding
,
2133 uint32_t bindingCount
,
2134 const VkBuffer
* pBuffers
,
2135 const VkDeviceSize
* pOffsets
)
2137 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2138 struct radv_vertex_binding
*vb
= cmd_buffer
->vertex_bindings
;
2139 bool changed
= false;
2141 /* We have to defer setting up vertex buffer since we need the buffer
2142 * stride from the pipeline. */
2144 assert(firstBinding
+ bindingCount
<= MAX_VBS
);
2145 for (uint32_t i
= 0; i
< bindingCount
; i
++) {
2146 uint32_t idx
= firstBinding
+ i
;
2149 (vb
[idx
].buffer
!= radv_buffer_from_handle(pBuffers
[i
]) ||
2150 vb
[idx
].offset
!= pOffsets
[i
])) {
2154 vb
[idx
].buffer
= radv_buffer_from_handle(pBuffers
[i
]);
2155 vb
[idx
].offset
= pOffsets
[i
];
2157 radv_cs_add_buffer(cmd_buffer
->device
->ws
, cmd_buffer
->cs
,
2158 vb
[idx
].buffer
->bo
, 8);
2162 /* No state changes. */
2166 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_VERTEX_BUFFER
;
2169 void radv_CmdBindIndexBuffer(
2170 VkCommandBuffer commandBuffer
,
2172 VkDeviceSize offset
,
2173 VkIndexType indexType
)
2175 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2176 RADV_FROM_HANDLE(radv_buffer
, index_buffer
, buffer
);
2178 if (cmd_buffer
->state
.index_buffer
== index_buffer
&&
2179 cmd_buffer
->state
.index_offset
== offset
&&
2180 cmd_buffer
->state
.index_type
== indexType
) {
2181 /* No state changes. */
2185 cmd_buffer
->state
.index_buffer
= index_buffer
;
2186 cmd_buffer
->state
.index_offset
= offset
;
2187 cmd_buffer
->state
.index_type
= indexType
; /* vk matches hw */
2188 cmd_buffer
->state
.index_va
= radv_buffer_get_va(index_buffer
->bo
);
2189 cmd_buffer
->state
.index_va
+= index_buffer
->offset
+ offset
;
2191 int index_size_shift
= cmd_buffer
->state
.index_type
? 2 : 1;
2192 cmd_buffer
->state
.max_index_count
= (index_buffer
->size
- offset
) >> index_size_shift
;
2193 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_INDEX_BUFFER
;
2194 radv_cs_add_buffer(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, index_buffer
->bo
, 8);
2199 radv_bind_descriptor_set(struct radv_cmd_buffer
*cmd_buffer
,
2200 VkPipelineBindPoint bind_point
,
2201 struct radv_descriptor_set
*set
, unsigned idx
)
2203 struct radeon_winsys
*ws
= cmd_buffer
->device
->ws
;
2205 radv_set_descriptor_set(cmd_buffer
, bind_point
, set
, idx
);
2209 assert(!(set
->layout
->flags
& VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR
));
2211 if (!cmd_buffer
->device
->use_global_bo_list
) {
2212 for (unsigned j
= 0; j
< set
->layout
->buffer_count
; ++j
)
2213 if (set
->descriptors
[j
])
2214 radv_cs_add_buffer(ws
, cmd_buffer
->cs
, set
->descriptors
[j
], 7);
2218 radv_cs_add_buffer(ws
, cmd_buffer
->cs
, set
->bo
, 8);
2221 void radv_CmdBindDescriptorSets(
2222 VkCommandBuffer commandBuffer
,
2223 VkPipelineBindPoint pipelineBindPoint
,
2224 VkPipelineLayout _layout
,
2226 uint32_t descriptorSetCount
,
2227 const VkDescriptorSet
* pDescriptorSets
,
2228 uint32_t dynamicOffsetCount
,
2229 const uint32_t* pDynamicOffsets
)
2231 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2232 RADV_FROM_HANDLE(radv_pipeline_layout
, layout
, _layout
);
2233 unsigned dyn_idx
= 0;
2235 const bool no_dynamic_bounds
= cmd_buffer
->device
->instance
->debug_flags
& RADV_DEBUG_NO_DYNAMIC_BOUNDS
;
2237 for (unsigned i
= 0; i
< descriptorSetCount
; ++i
) {
2238 unsigned idx
= i
+ firstSet
;
2239 RADV_FROM_HANDLE(radv_descriptor_set
, set
, pDescriptorSets
[i
]);
2240 radv_bind_descriptor_set(cmd_buffer
, pipelineBindPoint
, set
, idx
);
2242 for(unsigned j
= 0; j
< set
->layout
->dynamic_offset_count
; ++j
, ++dyn_idx
) {
2243 unsigned idx
= j
+ layout
->set
[i
+ firstSet
].dynamic_offset_start
;
2244 uint32_t *dst
= cmd_buffer
->dynamic_buffers
+ idx
* 4;
2245 assert(dyn_idx
< dynamicOffsetCount
);
2247 struct radv_descriptor_range
*range
= set
->dynamic_descriptors
+ j
;
2248 uint64_t va
= range
->va
+ pDynamicOffsets
[dyn_idx
];
2250 dst
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32);
2251 dst
[2] = no_dynamic_bounds
? 0xffffffffu
: range
->size
;
2252 dst
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
2253 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
2254 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
2255 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
2256 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
2257 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
2258 cmd_buffer
->push_constant_stages
|=
2259 set
->layout
->dynamic_shader_stages
;
2264 static bool radv_init_push_descriptor_set(struct radv_cmd_buffer
*cmd_buffer
,
2265 struct radv_descriptor_set
*set
,
2266 struct radv_descriptor_set_layout
*layout
,
2267 VkPipelineBindPoint bind_point
)
2269 struct radv_descriptor_state
*descriptors_state
=
2270 radv_get_descriptors_state(cmd_buffer
, bind_point
);
2271 set
->size
= layout
->size
;
2272 set
->layout
= layout
;
2274 if (descriptors_state
->push_set
.capacity
< set
->size
) {
2275 size_t new_size
= MAX2(set
->size
, 1024);
2276 new_size
= MAX2(new_size
, 2 * descriptors_state
->push_set
.capacity
);
2277 new_size
= MIN2(new_size
, 96 * MAX_PUSH_DESCRIPTORS
);
2279 free(set
->mapped_ptr
);
2280 set
->mapped_ptr
= malloc(new_size
);
2282 if (!set
->mapped_ptr
) {
2283 descriptors_state
->push_set
.capacity
= 0;
2284 cmd_buffer
->record_result
= VK_ERROR_OUT_OF_HOST_MEMORY
;
2288 descriptors_state
->push_set
.capacity
= new_size
;
2294 void radv_meta_push_descriptor_set(
2295 struct radv_cmd_buffer
* cmd_buffer
,
2296 VkPipelineBindPoint pipelineBindPoint
,
2297 VkPipelineLayout _layout
,
2299 uint32_t descriptorWriteCount
,
2300 const VkWriteDescriptorSet
* pDescriptorWrites
)
2302 RADV_FROM_HANDLE(radv_pipeline_layout
, layout
, _layout
);
2303 struct radv_descriptor_set
*push_set
= &cmd_buffer
->meta_push_descriptors
;
2307 assert(layout
->set
[set
].layout
->flags
& VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR
);
2309 push_set
->size
= layout
->set
[set
].layout
->size
;
2310 push_set
->layout
= layout
->set
[set
].layout
;
2312 if (!radv_cmd_buffer_upload_alloc(cmd_buffer
, push_set
->size
, 32,
2314 (void**) &push_set
->mapped_ptr
))
2317 push_set
->va
= radv_buffer_get_va(cmd_buffer
->upload
.upload_bo
);
2318 push_set
->va
+= bo_offset
;
2320 radv_update_descriptor_sets(cmd_buffer
->device
, cmd_buffer
,
2321 radv_descriptor_set_to_handle(push_set
),
2322 descriptorWriteCount
, pDescriptorWrites
, 0, NULL
);
2324 radv_set_descriptor_set(cmd_buffer
, pipelineBindPoint
, push_set
, set
);
2327 void radv_CmdPushDescriptorSetKHR(
2328 VkCommandBuffer commandBuffer
,
2329 VkPipelineBindPoint pipelineBindPoint
,
2330 VkPipelineLayout _layout
,
2332 uint32_t descriptorWriteCount
,
2333 const VkWriteDescriptorSet
* pDescriptorWrites
)
2335 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2336 RADV_FROM_HANDLE(radv_pipeline_layout
, layout
, _layout
);
2337 struct radv_descriptor_state
*descriptors_state
=
2338 radv_get_descriptors_state(cmd_buffer
, pipelineBindPoint
);
2339 struct radv_descriptor_set
*push_set
= &descriptors_state
->push_set
.set
;
2341 assert(layout
->set
[set
].layout
->flags
& VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR
);
2343 if (!radv_init_push_descriptor_set(cmd_buffer
, push_set
,
2344 layout
->set
[set
].layout
,
2348 radv_update_descriptor_sets(cmd_buffer
->device
, cmd_buffer
,
2349 radv_descriptor_set_to_handle(push_set
),
2350 descriptorWriteCount
, pDescriptorWrites
, 0, NULL
);
2352 radv_set_descriptor_set(cmd_buffer
, pipelineBindPoint
, push_set
, set
);
2353 descriptors_state
->push_dirty
= true;
2356 void radv_CmdPushDescriptorSetWithTemplateKHR(
2357 VkCommandBuffer commandBuffer
,
2358 VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate
,
2359 VkPipelineLayout _layout
,
2363 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2364 RADV_FROM_HANDLE(radv_pipeline_layout
, layout
, _layout
);
2365 RADV_FROM_HANDLE(radv_descriptor_update_template
, templ
, descriptorUpdateTemplate
);
2366 struct radv_descriptor_state
*descriptors_state
=
2367 radv_get_descriptors_state(cmd_buffer
, templ
->bind_point
);
2368 struct radv_descriptor_set
*push_set
= &descriptors_state
->push_set
.set
;
2370 assert(layout
->set
[set
].layout
->flags
& VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR
);
2372 if (!radv_init_push_descriptor_set(cmd_buffer
, push_set
,
2373 layout
->set
[set
].layout
,
2377 radv_update_descriptor_set_with_template(cmd_buffer
->device
, cmd_buffer
, push_set
,
2378 descriptorUpdateTemplate
, pData
);
2380 radv_set_descriptor_set(cmd_buffer
, templ
->bind_point
, push_set
, set
);
2381 descriptors_state
->push_dirty
= true;
2384 void radv_CmdPushConstants(VkCommandBuffer commandBuffer
,
2385 VkPipelineLayout layout
,
2386 VkShaderStageFlags stageFlags
,
2389 const void* pValues
)
2391 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2392 memcpy(cmd_buffer
->push_constants
+ offset
, pValues
, size
);
2393 cmd_buffer
->push_constant_stages
|= stageFlags
;
2396 VkResult
radv_EndCommandBuffer(
2397 VkCommandBuffer commandBuffer
)
2399 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2401 if (cmd_buffer
->queue_family_index
!= RADV_QUEUE_TRANSFER
) {
2402 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
== SI
)
2403 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_CS_PARTIAL_FLUSH
| RADV_CMD_FLAG_PS_PARTIAL_FLUSH
| RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2
;
2404 si_emit_cache_flush(cmd_buffer
);
2407 vk_free(&cmd_buffer
->pool
->alloc
, cmd_buffer
->state
.attachments
);
2409 if (!cmd_buffer
->device
->ws
->cs_finalize(cmd_buffer
->cs
))
2410 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2412 cmd_buffer
->status
= RADV_CMD_BUFFER_STATUS_EXECUTABLE
;
2414 return cmd_buffer
->record_result
;
2418 radv_emit_compute_pipeline(struct radv_cmd_buffer
*cmd_buffer
)
2420 struct radv_pipeline
*pipeline
= cmd_buffer
->state
.compute_pipeline
;
2422 if (!pipeline
|| pipeline
== cmd_buffer
->state
.emitted_compute_pipeline
)
2425 cmd_buffer
->state
.emitted_compute_pipeline
= pipeline
;
2427 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
, pipeline
->cs
.cdw
);
2428 radeon_emit_array(cmd_buffer
->cs
, pipeline
->cs
.buf
, pipeline
->cs
.cdw
);
2430 cmd_buffer
->compute_scratch_size_needed
=
2431 MAX2(cmd_buffer
->compute_scratch_size_needed
,
2432 pipeline
->max_waves
* pipeline
->scratch_bytes_per_wave
);
2434 radv_cs_add_buffer(cmd_buffer
->device
->ws
, cmd_buffer
->cs
,
2435 pipeline
->shaders
[MESA_SHADER_COMPUTE
]->bo
, 8);
2437 if (unlikely(cmd_buffer
->device
->trace_bo
))
2438 radv_save_pipeline(cmd_buffer
, pipeline
, RING_COMPUTE
);
2441 static void radv_mark_descriptor_sets_dirty(struct radv_cmd_buffer
*cmd_buffer
,
2442 VkPipelineBindPoint bind_point
)
2444 struct radv_descriptor_state
*descriptors_state
=
2445 radv_get_descriptors_state(cmd_buffer
, bind_point
);
2447 descriptors_state
->dirty
|= descriptors_state
->valid
;
2450 void radv_CmdBindPipeline(
2451 VkCommandBuffer commandBuffer
,
2452 VkPipelineBindPoint pipelineBindPoint
,
2453 VkPipeline _pipeline
)
2455 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2456 RADV_FROM_HANDLE(radv_pipeline
, pipeline
, _pipeline
);
2458 switch (pipelineBindPoint
) {
2459 case VK_PIPELINE_BIND_POINT_COMPUTE
:
2460 if (cmd_buffer
->state
.compute_pipeline
== pipeline
)
2462 radv_mark_descriptor_sets_dirty(cmd_buffer
, pipelineBindPoint
);
2464 cmd_buffer
->state
.compute_pipeline
= pipeline
;
2465 cmd_buffer
->push_constant_stages
|= VK_SHADER_STAGE_COMPUTE_BIT
;
2467 case VK_PIPELINE_BIND_POINT_GRAPHICS
:
2468 if (cmd_buffer
->state
.pipeline
== pipeline
)
2470 radv_mark_descriptor_sets_dirty(cmd_buffer
, pipelineBindPoint
);
2472 cmd_buffer
->state
.pipeline
= pipeline
;
2476 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_PIPELINE
;
2477 cmd_buffer
->push_constant_stages
|= pipeline
->active_stages
;
2479 /* the new vertex shader might not have the same user regs */
2480 cmd_buffer
->state
.last_first_instance
= -1;
2481 cmd_buffer
->state
.last_vertex_offset
= -1;
2483 /* Prefetch all pipeline shaders at first draw time. */
2484 cmd_buffer
->state
.prefetch_L2_mask
|= RADV_PREFETCH_SHADERS
;
2486 radv_bind_dynamic_state(cmd_buffer
, &pipeline
->dynamic_state
);
2488 if (pipeline
->graphics
.esgs_ring_size
> cmd_buffer
->esgs_ring_size_needed
)
2489 cmd_buffer
->esgs_ring_size_needed
= pipeline
->graphics
.esgs_ring_size
;
2490 if (pipeline
->graphics
.gsvs_ring_size
> cmd_buffer
->gsvs_ring_size_needed
)
2491 cmd_buffer
->gsvs_ring_size_needed
= pipeline
->graphics
.gsvs_ring_size
;
2493 if (radv_pipeline_has_tess(pipeline
))
2494 cmd_buffer
->tess_rings_needed
= true;
2496 if (radv_pipeline_has_gs(pipeline
)) {
2497 struct radv_userdata_info
*loc
= radv_lookup_user_sgpr(cmd_buffer
->state
.pipeline
, MESA_SHADER_GEOMETRY
,
2498 AC_UD_SCRATCH_RING_OFFSETS
);
2499 if (cmd_buffer
->ring_offsets_idx
== -1)
2500 cmd_buffer
->ring_offsets_idx
= loc
->sgpr_idx
;
2501 else if (loc
->sgpr_idx
!= -1)
2502 assert(loc
->sgpr_idx
== cmd_buffer
->ring_offsets_idx
);
2506 assert(!"invalid bind point");
2511 void radv_CmdSetViewport(
2512 VkCommandBuffer commandBuffer
,
2513 uint32_t firstViewport
,
2514 uint32_t viewportCount
,
2515 const VkViewport
* pViewports
)
2517 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2518 struct radv_cmd_state
*state
= &cmd_buffer
->state
;
2519 MAYBE_UNUSED
const uint32_t total_count
= firstViewport
+ viewportCount
;
2521 assert(firstViewport
< MAX_VIEWPORTS
);
2522 assert(total_count
>= 1 && total_count
<= MAX_VIEWPORTS
);
2524 if (cmd_buffer
->device
->physical_device
->has_scissor_bug
) {
2525 /* Try to skip unnecessary PS partial flushes when the viewports
2528 if (!(state
->dirty
& (RADV_CMD_DIRTY_DYNAMIC_VIEWPORT
|
2529 RADV_CMD_DIRTY_DYNAMIC_SCISSOR
)) &&
2530 !memcmp(state
->dynamic
.viewport
.viewports
+ firstViewport
,
2531 pViewports
, viewportCount
* sizeof(*pViewports
))) {
2536 memcpy(state
->dynamic
.viewport
.viewports
+ firstViewport
, pViewports
,
2537 viewportCount
* sizeof(*pViewports
));
2539 state
->dirty
|= RADV_CMD_DIRTY_DYNAMIC_VIEWPORT
;
2542 void radv_CmdSetScissor(
2543 VkCommandBuffer commandBuffer
,
2544 uint32_t firstScissor
,
2545 uint32_t scissorCount
,
2546 const VkRect2D
* pScissors
)
2548 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2549 struct radv_cmd_state
*state
= &cmd_buffer
->state
;
2550 MAYBE_UNUSED
const uint32_t total_count
= firstScissor
+ scissorCount
;
2552 assert(firstScissor
< MAX_SCISSORS
);
2553 assert(total_count
>= 1 && total_count
<= MAX_SCISSORS
);
2555 if (cmd_buffer
->device
->physical_device
->has_scissor_bug
) {
2556 /* Try to skip unnecessary PS partial flushes when the scissors
2559 if (!(state
->dirty
& (RADV_CMD_DIRTY_DYNAMIC_VIEWPORT
|
2560 RADV_CMD_DIRTY_DYNAMIC_SCISSOR
)) &&
2561 !memcmp(state
->dynamic
.scissor
.scissors
+ firstScissor
,
2562 pScissors
, scissorCount
* sizeof(*pScissors
))) {
2567 memcpy(state
->dynamic
.scissor
.scissors
+ firstScissor
, pScissors
,
2568 scissorCount
* sizeof(*pScissors
));
2570 state
->dirty
|= RADV_CMD_DIRTY_DYNAMIC_SCISSOR
;
2573 void radv_CmdSetLineWidth(
2574 VkCommandBuffer commandBuffer
,
2577 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2578 cmd_buffer
->state
.dynamic
.line_width
= lineWidth
;
2579 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH
;
2582 void radv_CmdSetDepthBias(
2583 VkCommandBuffer commandBuffer
,
2584 float depthBiasConstantFactor
,
2585 float depthBiasClamp
,
2586 float depthBiasSlopeFactor
)
2588 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2590 cmd_buffer
->state
.dynamic
.depth_bias
.bias
= depthBiasConstantFactor
;
2591 cmd_buffer
->state
.dynamic
.depth_bias
.clamp
= depthBiasClamp
;
2592 cmd_buffer
->state
.dynamic
.depth_bias
.slope
= depthBiasSlopeFactor
;
2594 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS
;
2597 void radv_CmdSetBlendConstants(
2598 VkCommandBuffer commandBuffer
,
2599 const float blendConstants
[4])
2601 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2603 memcpy(cmd_buffer
->state
.dynamic
.blend_constants
,
2604 blendConstants
, sizeof(float) * 4);
2606 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS
;
2609 void radv_CmdSetDepthBounds(
2610 VkCommandBuffer commandBuffer
,
2611 float minDepthBounds
,
2612 float maxDepthBounds
)
2614 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2616 cmd_buffer
->state
.dynamic
.depth_bounds
.min
= minDepthBounds
;
2617 cmd_buffer
->state
.dynamic
.depth_bounds
.max
= maxDepthBounds
;
2619 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS
;
2622 void radv_CmdSetStencilCompareMask(
2623 VkCommandBuffer commandBuffer
,
2624 VkStencilFaceFlags faceMask
,
2625 uint32_t compareMask
)
2627 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2629 if (faceMask
& VK_STENCIL_FACE_FRONT_BIT
)
2630 cmd_buffer
->state
.dynamic
.stencil_compare_mask
.front
= compareMask
;
2631 if (faceMask
& VK_STENCIL_FACE_BACK_BIT
)
2632 cmd_buffer
->state
.dynamic
.stencil_compare_mask
.back
= compareMask
;
2634 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK
;
2637 void radv_CmdSetStencilWriteMask(
2638 VkCommandBuffer commandBuffer
,
2639 VkStencilFaceFlags faceMask
,
2642 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2644 if (faceMask
& VK_STENCIL_FACE_FRONT_BIT
)
2645 cmd_buffer
->state
.dynamic
.stencil_write_mask
.front
= writeMask
;
2646 if (faceMask
& VK_STENCIL_FACE_BACK_BIT
)
2647 cmd_buffer
->state
.dynamic
.stencil_write_mask
.back
= writeMask
;
2649 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK
;
2652 void radv_CmdSetStencilReference(
2653 VkCommandBuffer commandBuffer
,
2654 VkStencilFaceFlags faceMask
,
2657 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2659 if (faceMask
& VK_STENCIL_FACE_FRONT_BIT
)
2660 cmd_buffer
->state
.dynamic
.stencil_reference
.front
= reference
;
2661 if (faceMask
& VK_STENCIL_FACE_BACK_BIT
)
2662 cmd_buffer
->state
.dynamic
.stencil_reference
.back
= reference
;
2664 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE
;
2667 void radv_CmdSetDiscardRectangleEXT(
2668 VkCommandBuffer commandBuffer
,
2669 uint32_t firstDiscardRectangle
,
2670 uint32_t discardRectangleCount
,
2671 const VkRect2D
* pDiscardRectangles
)
2673 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2674 struct radv_cmd_state
*state
= &cmd_buffer
->state
;
2675 MAYBE_UNUSED
const uint32_t total_count
= firstDiscardRectangle
+ discardRectangleCount
;
2677 assert(firstDiscardRectangle
< MAX_DISCARD_RECTANGLES
);
2678 assert(total_count
>= 1 && total_count
<= MAX_DISCARD_RECTANGLES
);
2680 typed_memcpy(&state
->dynamic
.discard_rectangle
.rectangles
[firstDiscardRectangle
],
2681 pDiscardRectangles
, discardRectangleCount
);
2683 state
->dirty
|= RADV_CMD_DIRTY_DYNAMIC_DISCARD_RECTANGLE
;
2686 void radv_CmdExecuteCommands(
2687 VkCommandBuffer commandBuffer
,
2688 uint32_t commandBufferCount
,
2689 const VkCommandBuffer
* pCmdBuffers
)
2691 RADV_FROM_HANDLE(radv_cmd_buffer
, primary
, commandBuffer
);
2693 assert(commandBufferCount
> 0);
2695 /* Emit pending flushes on primary prior to executing secondary */
2696 si_emit_cache_flush(primary
);
2698 for (uint32_t i
= 0; i
< commandBufferCount
; i
++) {
2699 RADV_FROM_HANDLE(radv_cmd_buffer
, secondary
, pCmdBuffers
[i
]);
2701 primary
->scratch_size_needed
= MAX2(primary
->scratch_size_needed
,
2702 secondary
->scratch_size_needed
);
2703 primary
->compute_scratch_size_needed
= MAX2(primary
->compute_scratch_size_needed
,
2704 secondary
->compute_scratch_size_needed
);
2706 if (secondary
->esgs_ring_size_needed
> primary
->esgs_ring_size_needed
)
2707 primary
->esgs_ring_size_needed
= secondary
->esgs_ring_size_needed
;
2708 if (secondary
->gsvs_ring_size_needed
> primary
->gsvs_ring_size_needed
)
2709 primary
->gsvs_ring_size_needed
= secondary
->gsvs_ring_size_needed
;
2710 if (secondary
->tess_rings_needed
)
2711 primary
->tess_rings_needed
= true;
2712 if (secondary
->sample_positions_needed
)
2713 primary
->sample_positions_needed
= true;
2715 if (secondary
->ring_offsets_idx
!= -1) {
2716 if (primary
->ring_offsets_idx
== -1)
2717 primary
->ring_offsets_idx
= secondary
->ring_offsets_idx
;
2719 assert(secondary
->ring_offsets_idx
== primary
->ring_offsets_idx
);
2721 primary
->device
->ws
->cs_execute_secondary(primary
->cs
, secondary
->cs
);
2724 /* When the secondary command buffer is compute only we don't
2725 * need to re-emit the current graphics pipeline.
2727 if (secondary
->state
.emitted_pipeline
) {
2728 primary
->state
.emitted_pipeline
=
2729 secondary
->state
.emitted_pipeline
;
2732 /* When the secondary command buffer is graphics only we don't
2733 * need to re-emit the current compute pipeline.
2735 if (secondary
->state
.emitted_compute_pipeline
) {
2736 primary
->state
.emitted_compute_pipeline
=
2737 secondary
->state
.emitted_compute_pipeline
;
2740 /* Only re-emit the draw packets when needed. */
2741 if (secondary
->state
.last_primitive_reset_en
!= -1) {
2742 primary
->state
.last_primitive_reset_en
=
2743 secondary
->state
.last_primitive_reset_en
;
2746 if (secondary
->state
.last_primitive_reset_index
) {
2747 primary
->state
.last_primitive_reset_index
=
2748 secondary
->state
.last_primitive_reset_index
;
2751 if (secondary
->state
.last_ia_multi_vgt_param
) {
2752 primary
->state
.last_ia_multi_vgt_param
=
2753 secondary
->state
.last_ia_multi_vgt_param
;
2756 primary
->state
.last_first_instance
= secondary
->state
.last_first_instance
;
2757 primary
->state
.last_num_instances
= secondary
->state
.last_num_instances
;
2758 primary
->state
.last_vertex_offset
= secondary
->state
.last_vertex_offset
;
2760 if (secondary
->state
.last_index_type
!= -1) {
2761 primary
->state
.last_index_type
=
2762 secondary
->state
.last_index_type
;
2766 /* After executing commands from secondary buffers we have to dirty
2769 primary
->state
.dirty
|= RADV_CMD_DIRTY_PIPELINE
|
2770 RADV_CMD_DIRTY_INDEX_BUFFER
|
2771 RADV_CMD_DIRTY_DYNAMIC_ALL
;
2772 radv_mark_descriptor_sets_dirty(primary
, VK_PIPELINE_BIND_POINT_GRAPHICS
);
2773 radv_mark_descriptor_sets_dirty(primary
, VK_PIPELINE_BIND_POINT_COMPUTE
);
2776 VkResult
radv_CreateCommandPool(
2778 const VkCommandPoolCreateInfo
* pCreateInfo
,
2779 const VkAllocationCallbacks
* pAllocator
,
2780 VkCommandPool
* pCmdPool
)
2782 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2783 struct radv_cmd_pool
*pool
;
2785 pool
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*pool
), 8,
2786 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2788 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2791 pool
->alloc
= *pAllocator
;
2793 pool
->alloc
= device
->alloc
;
2795 list_inithead(&pool
->cmd_buffers
);
2796 list_inithead(&pool
->free_cmd_buffers
);
2798 pool
->queue_family_index
= pCreateInfo
->queueFamilyIndex
;
2800 *pCmdPool
= radv_cmd_pool_to_handle(pool
);
2806 void radv_DestroyCommandPool(
2808 VkCommandPool commandPool
,
2809 const VkAllocationCallbacks
* pAllocator
)
2811 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2812 RADV_FROM_HANDLE(radv_cmd_pool
, pool
, commandPool
);
2817 list_for_each_entry_safe(struct radv_cmd_buffer
, cmd_buffer
,
2818 &pool
->cmd_buffers
, pool_link
) {
2819 radv_cmd_buffer_destroy(cmd_buffer
);
2822 list_for_each_entry_safe(struct radv_cmd_buffer
, cmd_buffer
,
2823 &pool
->free_cmd_buffers
, pool_link
) {
2824 radv_cmd_buffer_destroy(cmd_buffer
);
2827 vk_free2(&device
->alloc
, pAllocator
, pool
);
2830 VkResult
radv_ResetCommandPool(
2832 VkCommandPool commandPool
,
2833 VkCommandPoolResetFlags flags
)
2835 RADV_FROM_HANDLE(radv_cmd_pool
, pool
, commandPool
);
2838 list_for_each_entry(struct radv_cmd_buffer
, cmd_buffer
,
2839 &pool
->cmd_buffers
, pool_link
) {
2840 result
= radv_reset_cmd_buffer(cmd_buffer
);
2841 if (result
!= VK_SUCCESS
)
2848 void radv_TrimCommandPool(
2850 VkCommandPool commandPool
,
2851 VkCommandPoolTrimFlagsKHR flags
)
2853 RADV_FROM_HANDLE(radv_cmd_pool
, pool
, commandPool
);
2858 list_for_each_entry_safe(struct radv_cmd_buffer
, cmd_buffer
,
2859 &pool
->free_cmd_buffers
, pool_link
) {
2860 radv_cmd_buffer_destroy(cmd_buffer
);
2864 void radv_CmdBeginRenderPass(
2865 VkCommandBuffer commandBuffer
,
2866 const VkRenderPassBeginInfo
* pRenderPassBegin
,
2867 VkSubpassContents contents
)
2869 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2870 RADV_FROM_HANDLE(radv_render_pass
, pass
, pRenderPassBegin
->renderPass
);
2871 RADV_FROM_HANDLE(radv_framebuffer
, framebuffer
, pRenderPassBegin
->framebuffer
);
2873 MAYBE_UNUSED
unsigned cdw_max
= radeon_check_space(cmd_buffer
->device
->ws
,
2874 cmd_buffer
->cs
, 2048);
2875 MAYBE_UNUSED VkResult result
;
2877 cmd_buffer
->state
.framebuffer
= framebuffer
;
2878 cmd_buffer
->state
.pass
= pass
;
2879 cmd_buffer
->state
.render_area
= pRenderPassBegin
->renderArea
;
2881 result
= radv_cmd_state_setup_attachments(cmd_buffer
, pass
, pRenderPassBegin
);
2882 if (result
!= VK_SUCCESS
)
2885 radv_cmd_buffer_set_subpass(cmd_buffer
, pass
->subpasses
, true);
2886 assert(cmd_buffer
->cs
->cdw
<= cdw_max
);
2888 radv_cmd_buffer_clear_subpass(cmd_buffer
);
2891 void radv_CmdNextSubpass(
2892 VkCommandBuffer commandBuffer
,
2893 VkSubpassContents contents
)
2895 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2897 radv_cmd_buffer_resolve_subpass(cmd_buffer
);
2899 radeon_check_space(cmd_buffer
->device
->ws
, cmd_buffer
->cs
,
2902 radv_cmd_buffer_set_subpass(cmd_buffer
, cmd_buffer
->state
.subpass
+ 1, true);
2903 radv_cmd_buffer_clear_subpass(cmd_buffer
);
2906 static void radv_emit_view_index(struct radv_cmd_buffer
*cmd_buffer
, unsigned index
)
2908 struct radv_pipeline
*pipeline
= cmd_buffer
->state
.pipeline
;
2909 for (unsigned stage
= 0; stage
< MESA_SHADER_STAGES
; ++stage
) {
2910 if (!pipeline
->shaders
[stage
])
2912 struct radv_userdata_info
*loc
= radv_lookup_user_sgpr(pipeline
, stage
, AC_UD_VIEW_INDEX
);
2913 if (loc
->sgpr_idx
== -1)
2915 uint32_t base_reg
= pipeline
->user_data_0
[stage
];
2916 radeon_set_sh_reg(cmd_buffer
->cs
, base_reg
+ loc
->sgpr_idx
* 4, index
);
2919 if (pipeline
->gs_copy_shader
) {
2920 struct radv_userdata_info
*loc
= &pipeline
->gs_copy_shader
->info
.user_sgprs_locs
.shader_data
[AC_UD_VIEW_INDEX
];
2921 if (loc
->sgpr_idx
!= -1) {
2922 uint32_t base_reg
= R_00B130_SPI_SHADER_USER_DATA_VS_0
;
2923 radeon_set_sh_reg(cmd_buffer
->cs
, base_reg
+ loc
->sgpr_idx
* 4, index
);
2929 radv_cs_emit_draw_packet(struct radv_cmd_buffer
*cmd_buffer
,
2930 uint32_t vertex_count
)
2932 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_DRAW_INDEX_AUTO
, 1, cmd_buffer
->state
.predicating
));
2933 radeon_emit(cmd_buffer
->cs
, vertex_count
);
2934 radeon_emit(cmd_buffer
->cs
, V_0287F0_DI_SRC_SEL_AUTO_INDEX
|
2935 S_0287F0_USE_OPAQUE(0));
2939 radv_cs_emit_draw_indexed_packet(struct radv_cmd_buffer
*cmd_buffer
,
2941 uint32_t index_count
)
2943 radeon_emit(cmd_buffer
->cs
, PKT3(PKT3_DRAW_INDEX_2
, 4, false));
2944 radeon_emit(cmd_buffer
->cs
, cmd_buffer
->state
.max_index_count
);
2945 radeon_emit(cmd_buffer
->cs
, index_va
);
2946 radeon_emit(cmd_buffer
->cs
, index_va
>> 32);
2947 radeon_emit(cmd_buffer
->cs
, index_count
);
2948 radeon_emit(cmd_buffer
->cs
, V_0287F0_DI_SRC_SEL_DMA
);
2952 radv_cs_emit_indirect_draw_packet(struct radv_cmd_buffer
*cmd_buffer
,
2954 uint32_t draw_count
,
2958 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
2959 unsigned di_src_sel
= indexed
? V_0287F0_DI_SRC_SEL_DMA
2960 : V_0287F0_DI_SRC_SEL_AUTO_INDEX
;
2961 bool draw_id_enable
= radv_get_vertex_shader(cmd_buffer
->state
.pipeline
)->info
.info
.vs
.needs_draw_id
;
2962 uint32_t base_reg
= cmd_buffer
->state
.pipeline
->graphics
.vtx_base_sgpr
;
2965 /* just reset draw state for vertex data */
2966 cmd_buffer
->state
.last_first_instance
= -1;
2967 cmd_buffer
->state
.last_num_instances
= -1;
2968 cmd_buffer
->state
.last_vertex_offset
= -1;
2970 if (draw_count
== 1 && !count_va
&& !draw_id_enable
) {
2971 radeon_emit(cs
, PKT3(indexed
? PKT3_DRAW_INDEX_INDIRECT
:
2972 PKT3_DRAW_INDIRECT
, 3, false));
2974 radeon_emit(cs
, (base_reg
- SI_SH_REG_OFFSET
) >> 2);
2975 radeon_emit(cs
, ((base_reg
+ 4) - SI_SH_REG_OFFSET
) >> 2);
2976 radeon_emit(cs
, di_src_sel
);
2978 radeon_emit(cs
, PKT3(indexed
? PKT3_DRAW_INDEX_INDIRECT_MULTI
:
2979 PKT3_DRAW_INDIRECT_MULTI
,
2982 radeon_emit(cs
, (base_reg
- SI_SH_REG_OFFSET
) >> 2);
2983 radeon_emit(cs
, ((base_reg
+ 4) - SI_SH_REG_OFFSET
) >> 2);
2984 radeon_emit(cs
, (((base_reg
+ 8) - SI_SH_REG_OFFSET
) >> 2) |
2985 S_2C3_DRAW_INDEX_ENABLE(draw_id_enable
) |
2986 S_2C3_COUNT_INDIRECT_ENABLE(!!count_va
));
2987 radeon_emit(cs
, draw_count
); /* count */
2988 radeon_emit(cs
, count_va
); /* count_addr */
2989 radeon_emit(cs
, count_va
>> 32);
2990 radeon_emit(cs
, stride
); /* stride */
2991 radeon_emit(cs
, di_src_sel
);
2995 struct radv_draw_info
{
2997 * Number of vertices.
3002 * Index of the first vertex.
3004 int32_t vertex_offset
;
3007 * First instance id.
3009 uint32_t first_instance
;
3012 * Number of instances.
3014 uint32_t instance_count
;
3017 * First index (indexed draws only).
3019 uint32_t first_index
;
3022 * Whether it's an indexed draw.
3027 * Indirect draw parameters resource.
3029 struct radv_buffer
*indirect
;
3030 uint64_t indirect_offset
;
3034 * Draw count parameters resource.
3036 struct radv_buffer
*count_buffer
;
3037 uint64_t count_buffer_offset
;
3041 radv_emit_draw_packets(struct radv_cmd_buffer
*cmd_buffer
,
3042 const struct radv_draw_info
*info
)
3044 struct radv_cmd_state
*state
= &cmd_buffer
->state
;
3045 struct radeon_winsys
*ws
= cmd_buffer
->device
->ws
;
3046 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
3048 if (info
->indirect
) {
3049 uint64_t va
= radv_buffer_get_va(info
->indirect
->bo
);
3050 uint64_t count_va
= 0;
3052 va
+= info
->indirect
->offset
+ info
->indirect_offset
;
3054 radv_cs_add_buffer(ws
, cs
, info
->indirect
->bo
, 8);
3056 radeon_emit(cs
, PKT3(PKT3_SET_BASE
, 2, 0));
3058 radeon_emit(cs
, va
);
3059 radeon_emit(cs
, va
>> 32);
3061 if (info
->count_buffer
) {
3062 count_va
= radv_buffer_get_va(info
->count_buffer
->bo
);
3063 count_va
+= info
->count_buffer
->offset
+
3064 info
->count_buffer_offset
;
3066 radv_cs_add_buffer(ws
, cs
, info
->count_buffer
->bo
, 8);
3069 if (!state
->subpass
->view_mask
) {
3070 radv_cs_emit_indirect_draw_packet(cmd_buffer
,
3077 for_each_bit(i
, state
->subpass
->view_mask
) {
3078 radv_emit_view_index(cmd_buffer
, i
);
3080 radv_cs_emit_indirect_draw_packet(cmd_buffer
,
3088 assert(state
->pipeline
->graphics
.vtx_base_sgpr
);
3090 if (info
->vertex_offset
!= state
->last_vertex_offset
||
3091 info
->first_instance
!= state
->last_first_instance
) {
3092 radeon_set_sh_reg_seq(cs
, state
->pipeline
->graphics
.vtx_base_sgpr
,
3093 state
->pipeline
->graphics
.vtx_emit_num
);
3095 radeon_emit(cs
, info
->vertex_offset
);
3096 radeon_emit(cs
, info
->first_instance
);
3097 if (state
->pipeline
->graphics
.vtx_emit_num
== 3)
3099 state
->last_first_instance
= info
->first_instance
;
3100 state
->last_vertex_offset
= info
->vertex_offset
;
3103 if (state
->last_num_instances
!= info
->instance_count
) {
3104 radeon_emit(cs
, PKT3(PKT3_NUM_INSTANCES
, 0, false));
3105 radeon_emit(cs
, info
->instance_count
);
3106 state
->last_num_instances
= info
->instance_count
;
3109 if (info
->indexed
) {
3110 int index_size
= state
->index_type
? 4 : 2;
3113 index_va
= state
->index_va
;
3114 index_va
+= info
->first_index
* index_size
;
3116 if (!state
->subpass
->view_mask
) {
3117 radv_cs_emit_draw_indexed_packet(cmd_buffer
,
3122 for_each_bit(i
, state
->subpass
->view_mask
) {
3123 radv_emit_view_index(cmd_buffer
, i
);
3125 radv_cs_emit_draw_indexed_packet(cmd_buffer
,
3131 if (!state
->subpass
->view_mask
) {
3132 radv_cs_emit_draw_packet(cmd_buffer
, info
->count
);
3135 for_each_bit(i
, state
->subpass
->view_mask
) {
3136 radv_emit_view_index(cmd_buffer
, i
);
3138 radv_cs_emit_draw_packet(cmd_buffer
,
3147 radv_emit_all_graphics_states(struct radv_cmd_buffer
*cmd_buffer
,
3148 const struct radv_draw_info
*info
)
3150 if ((cmd_buffer
->state
.dirty
& RADV_CMD_DIRTY_FRAMEBUFFER
) ||
3151 cmd_buffer
->state
.emitted_pipeline
!= cmd_buffer
->state
.pipeline
)
3152 radv_emit_rbplus_state(cmd_buffer
);
3154 if (cmd_buffer
->state
.dirty
& RADV_CMD_DIRTY_PIPELINE
)
3155 radv_emit_graphics_pipeline(cmd_buffer
);
3157 if (cmd_buffer
->state
.dirty
& RADV_CMD_DIRTY_FRAMEBUFFER
)
3158 radv_emit_framebuffer_state(cmd_buffer
);
3160 if (info
->indexed
) {
3161 if (cmd_buffer
->state
.dirty
& RADV_CMD_DIRTY_INDEX_BUFFER
)
3162 radv_emit_index_buffer(cmd_buffer
);
3164 /* On CI and later, non-indexed draws overwrite VGT_INDEX_TYPE,
3165 * so the state must be re-emitted before the next indexed
3168 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
3169 cmd_buffer
->state
.last_index_type
= -1;
3170 cmd_buffer
->state
.dirty
|= RADV_CMD_DIRTY_INDEX_BUFFER
;
3174 radv_cmd_buffer_flush_dynamic_state(cmd_buffer
);
3176 radv_emit_draw_registers(cmd_buffer
, info
->indexed
,
3177 info
->instance_count
> 1, info
->indirect
,
3178 info
->indirect
? 0 : info
->count
);
3182 radv_draw(struct radv_cmd_buffer
*cmd_buffer
,
3183 const struct radv_draw_info
*info
)
3186 cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= CIK
;
3187 bool pipeline_is_dirty
=
3188 (cmd_buffer
->state
.dirty
& RADV_CMD_DIRTY_PIPELINE
) &&
3189 cmd_buffer
->state
.pipeline
&&
3190 cmd_buffer
->state
.pipeline
!= cmd_buffer
->state
.emitted_pipeline
;
3192 MAYBE_UNUSED
unsigned cdw_max
=
3193 radeon_check_space(cmd_buffer
->device
->ws
,
3194 cmd_buffer
->cs
, 4096);
3196 /* Use optimal packet order based on whether we need to sync the
3199 if (cmd_buffer
->state
.flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
3200 RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
3201 RADV_CMD_FLAG_PS_PARTIAL_FLUSH
|
3202 RADV_CMD_FLAG_CS_PARTIAL_FLUSH
)) {
3203 /* If we have to wait for idle, set all states first, so that
3204 * all SET packets are processed in parallel with previous draw
3205 * calls. Then upload descriptors, set shader pointers, and
3206 * draw, and prefetch at the end. This ensures that the time
3207 * the CUs are idle is very short. (there are only SET_SH
3208 * packets between the wait and the draw)
3210 radv_emit_all_graphics_states(cmd_buffer
, info
);
3211 si_emit_cache_flush(cmd_buffer
);
3212 /* <-- CUs are idle here --> */
3214 radv_upload_graphics_shader_descriptors(cmd_buffer
, pipeline_is_dirty
);
3216 radv_emit_draw_packets(cmd_buffer
, info
);
3217 /* <-- CUs are busy here --> */
3219 /* Start prefetches after the draw has been started. Both will
3220 * run in parallel, but starting the draw first is more
3223 if (has_prefetch
&& cmd_buffer
->state
.prefetch_L2_mask
) {
3224 radv_emit_prefetch_L2(cmd_buffer
,
3225 cmd_buffer
->state
.pipeline
, false);
3228 /* If we don't wait for idle, start prefetches first, then set
3229 * states, and draw at the end.
3231 si_emit_cache_flush(cmd_buffer
);
3233 if (has_prefetch
&& cmd_buffer
->state
.prefetch_L2_mask
) {
3234 /* Only prefetch the vertex shader and VBO descriptors
3235 * in order to start the draw as soon as possible.
3237 radv_emit_prefetch_L2(cmd_buffer
,
3238 cmd_buffer
->state
.pipeline
, true);
3241 radv_upload_graphics_shader_descriptors(cmd_buffer
, pipeline_is_dirty
);
3243 radv_emit_all_graphics_states(cmd_buffer
, info
);
3244 radv_emit_draw_packets(cmd_buffer
, info
);
3246 /* Prefetch the remaining shaders after the draw has been
3249 if (has_prefetch
&& cmd_buffer
->state
.prefetch_L2_mask
) {
3250 radv_emit_prefetch_L2(cmd_buffer
,
3251 cmd_buffer
->state
.pipeline
, false);
3255 assert(cmd_buffer
->cs
->cdw
<= cdw_max
);
3256 radv_cmd_buffer_after_draw(cmd_buffer
, RADV_CMD_FLAG_PS_PARTIAL_FLUSH
);
3260 VkCommandBuffer commandBuffer
,
3261 uint32_t vertexCount
,
3262 uint32_t instanceCount
,
3263 uint32_t firstVertex
,
3264 uint32_t firstInstance
)
3266 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3267 struct radv_draw_info info
= {};
3269 info
.count
= vertexCount
;
3270 info
.instance_count
= instanceCount
;
3271 info
.first_instance
= firstInstance
;
3272 info
.vertex_offset
= firstVertex
;
3274 radv_draw(cmd_buffer
, &info
);
3277 void radv_CmdDrawIndexed(
3278 VkCommandBuffer commandBuffer
,
3279 uint32_t indexCount
,
3280 uint32_t instanceCount
,
3281 uint32_t firstIndex
,
3282 int32_t vertexOffset
,
3283 uint32_t firstInstance
)
3285 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3286 struct radv_draw_info info
= {};
3288 info
.indexed
= true;
3289 info
.count
= indexCount
;
3290 info
.instance_count
= instanceCount
;
3291 info
.first_index
= firstIndex
;
3292 info
.vertex_offset
= vertexOffset
;
3293 info
.first_instance
= firstInstance
;
3295 radv_draw(cmd_buffer
, &info
);
3298 void radv_CmdDrawIndirect(
3299 VkCommandBuffer commandBuffer
,
3301 VkDeviceSize offset
,
3305 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3306 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3307 struct radv_draw_info info
= {};
3309 info
.count
= drawCount
;
3310 info
.indirect
= buffer
;
3311 info
.indirect_offset
= offset
;
3312 info
.stride
= stride
;
3314 radv_draw(cmd_buffer
, &info
);
3317 void radv_CmdDrawIndexedIndirect(
3318 VkCommandBuffer commandBuffer
,
3320 VkDeviceSize offset
,
3324 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3325 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3326 struct radv_draw_info info
= {};
3328 info
.indexed
= true;
3329 info
.count
= drawCount
;
3330 info
.indirect
= buffer
;
3331 info
.indirect_offset
= offset
;
3332 info
.stride
= stride
;
3334 radv_draw(cmd_buffer
, &info
);
3337 void radv_CmdDrawIndirectCountAMD(
3338 VkCommandBuffer commandBuffer
,
3340 VkDeviceSize offset
,
3341 VkBuffer _countBuffer
,
3342 VkDeviceSize countBufferOffset
,
3343 uint32_t maxDrawCount
,
3346 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3347 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3348 RADV_FROM_HANDLE(radv_buffer
, count_buffer
, _countBuffer
);
3349 struct radv_draw_info info
= {};
3351 info
.count
= maxDrawCount
;
3352 info
.indirect
= buffer
;
3353 info
.indirect_offset
= offset
;
3354 info
.count_buffer
= count_buffer
;
3355 info
.count_buffer_offset
= countBufferOffset
;
3356 info
.stride
= stride
;
3358 radv_draw(cmd_buffer
, &info
);
3361 void radv_CmdDrawIndexedIndirectCountAMD(
3362 VkCommandBuffer commandBuffer
,
3364 VkDeviceSize offset
,
3365 VkBuffer _countBuffer
,
3366 VkDeviceSize countBufferOffset
,
3367 uint32_t maxDrawCount
,
3370 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3371 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3372 RADV_FROM_HANDLE(radv_buffer
, count_buffer
, _countBuffer
);
3373 struct radv_draw_info info
= {};
3375 info
.indexed
= true;
3376 info
.count
= maxDrawCount
;
3377 info
.indirect
= buffer
;
3378 info
.indirect_offset
= offset
;
3379 info
.count_buffer
= count_buffer
;
3380 info
.count_buffer_offset
= countBufferOffset
;
3381 info
.stride
= stride
;
3383 radv_draw(cmd_buffer
, &info
);
3386 struct radv_dispatch_info
{
3388 * Determine the layout of the grid (in block units) to be used.
3393 * A starting offset for the grid. If unaligned is set, the offset
3394 * must still be aligned.
3396 uint32_t offsets
[3];
3398 * Whether it's an unaligned compute dispatch.
3403 * Indirect compute parameters resource.
3405 struct radv_buffer
*indirect
;
3406 uint64_t indirect_offset
;
3410 radv_emit_dispatch_packets(struct radv_cmd_buffer
*cmd_buffer
,
3411 const struct radv_dispatch_info
*info
)
3413 struct radv_pipeline
*pipeline
= cmd_buffer
->state
.compute_pipeline
;
3414 struct radv_shader_variant
*compute_shader
= pipeline
->shaders
[MESA_SHADER_COMPUTE
];
3415 unsigned dispatch_initiator
= cmd_buffer
->device
->dispatch_initiator
;
3416 struct radeon_winsys
*ws
= cmd_buffer
->device
->ws
;
3417 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
3418 struct radv_userdata_info
*loc
;
3420 loc
= radv_lookup_user_sgpr(pipeline
, MESA_SHADER_COMPUTE
,
3421 AC_UD_CS_GRID_SIZE
);
3423 MAYBE_UNUSED
unsigned cdw_max
= radeon_check_space(ws
, cs
, 25);
3425 if (info
->indirect
) {
3426 uint64_t va
= radv_buffer_get_va(info
->indirect
->bo
);
3428 va
+= info
->indirect
->offset
+ info
->indirect_offset
;
3430 radv_cs_add_buffer(ws
, cs
, info
->indirect
->bo
, 8);
3432 if (loc
->sgpr_idx
!= -1) {
3433 for (unsigned i
= 0; i
< 3; ++i
) {
3434 radeon_emit(cs
, PKT3(PKT3_COPY_DATA
, 4, 0));
3435 radeon_emit(cs
, COPY_DATA_SRC_SEL(COPY_DATA_MEM
) |
3436 COPY_DATA_DST_SEL(COPY_DATA_REG
));
3437 radeon_emit(cs
, (va
+ 4 * i
));
3438 radeon_emit(cs
, (va
+ 4 * i
) >> 32);
3439 radeon_emit(cs
, ((R_00B900_COMPUTE_USER_DATA_0
3440 + loc
->sgpr_idx
* 4) >> 2) + i
);
3445 if (radv_cmd_buffer_uses_mec(cmd_buffer
)) {
3446 radeon_emit(cs
, PKT3(PKT3_DISPATCH_INDIRECT
, 2, 0) |
3447 PKT3_SHADER_TYPE_S(1));
3448 radeon_emit(cs
, va
);
3449 radeon_emit(cs
, va
>> 32);
3450 radeon_emit(cs
, dispatch_initiator
);
3452 radeon_emit(cs
, PKT3(PKT3_SET_BASE
, 2, 0) |
3453 PKT3_SHADER_TYPE_S(1));
3455 radeon_emit(cs
, va
);
3456 radeon_emit(cs
, va
>> 32);
3458 radeon_emit(cs
, PKT3(PKT3_DISPATCH_INDIRECT
, 1, 0) |
3459 PKT3_SHADER_TYPE_S(1));
3461 radeon_emit(cs
, dispatch_initiator
);
3464 unsigned blocks
[3] = { info
->blocks
[0], info
->blocks
[1], info
->blocks
[2] };
3465 unsigned offsets
[3] = { info
->offsets
[0], info
->offsets
[1], info
->offsets
[2] };
3467 if (info
->unaligned
) {
3468 unsigned *cs_block_size
= compute_shader
->info
.cs
.block_size
;
3469 unsigned remainder
[3];
3471 /* If aligned, these should be an entire block size,
3474 remainder
[0] = blocks
[0] + cs_block_size
[0] -
3475 align_u32_npot(blocks
[0], cs_block_size
[0]);
3476 remainder
[1] = blocks
[1] + cs_block_size
[1] -
3477 align_u32_npot(blocks
[1], cs_block_size
[1]);
3478 remainder
[2] = blocks
[2] + cs_block_size
[2] -
3479 align_u32_npot(blocks
[2], cs_block_size
[2]);
3481 blocks
[0] = round_up_u32(blocks
[0], cs_block_size
[0]);
3482 blocks
[1] = round_up_u32(blocks
[1], cs_block_size
[1]);
3483 blocks
[2] = round_up_u32(blocks
[2], cs_block_size
[2]);
3485 for(unsigned i
= 0; i
< 3; ++i
) {
3486 assert(offsets
[i
] % cs_block_size
[i
] == 0);
3487 offsets
[i
] /= cs_block_size
[i
];
3490 radeon_set_sh_reg_seq(cs
, R_00B81C_COMPUTE_NUM_THREAD_X
, 3);
3492 S_00B81C_NUM_THREAD_FULL(cs_block_size
[0]) |
3493 S_00B81C_NUM_THREAD_PARTIAL(remainder
[0]));
3495 S_00B81C_NUM_THREAD_FULL(cs_block_size
[1]) |
3496 S_00B81C_NUM_THREAD_PARTIAL(remainder
[1]));
3498 S_00B81C_NUM_THREAD_FULL(cs_block_size
[2]) |
3499 S_00B81C_NUM_THREAD_PARTIAL(remainder
[2]));
3501 dispatch_initiator
|= S_00B800_PARTIAL_TG_EN(1);
3504 if (loc
->sgpr_idx
!= -1) {
3505 assert(!loc
->indirect
);
3506 assert(loc
->num_sgprs
== 3);
3508 radeon_set_sh_reg_seq(cs
, R_00B900_COMPUTE_USER_DATA_0
+
3509 loc
->sgpr_idx
* 4, 3);
3510 radeon_emit(cs
, blocks
[0]);
3511 radeon_emit(cs
, blocks
[1]);
3512 radeon_emit(cs
, blocks
[2]);
3515 if (offsets
[0] || offsets
[1] || offsets
[2]) {
3516 radeon_set_sh_reg_seq(cs
, R_00B810_COMPUTE_START_X
, 3);
3517 radeon_emit(cs
, offsets
[0]);
3518 radeon_emit(cs
, offsets
[1]);
3519 radeon_emit(cs
, offsets
[2]);
3521 /* The blocks in the packet are not counts but end values. */
3522 for (unsigned i
= 0; i
< 3; ++i
)
3523 blocks
[i
] += offsets
[i
];
3525 dispatch_initiator
|= S_00B800_FORCE_START_AT_000(1);
3528 radeon_emit(cs
, PKT3(PKT3_DISPATCH_DIRECT
, 3, 0) |
3529 PKT3_SHADER_TYPE_S(1));
3530 radeon_emit(cs
, blocks
[0]);
3531 radeon_emit(cs
, blocks
[1]);
3532 radeon_emit(cs
, blocks
[2]);
3533 radeon_emit(cs
, dispatch_initiator
);
3536 assert(cmd_buffer
->cs
->cdw
<= cdw_max
);
3540 radv_upload_compute_shader_descriptors(struct radv_cmd_buffer
*cmd_buffer
)
3542 radv_flush_descriptors(cmd_buffer
, VK_SHADER_STAGE_COMPUTE_BIT
);
3543 radv_flush_constants(cmd_buffer
, VK_SHADER_STAGE_COMPUTE_BIT
);
3547 radv_dispatch(struct radv_cmd_buffer
*cmd_buffer
,
3548 const struct radv_dispatch_info
*info
)
3550 struct radv_pipeline
*pipeline
= cmd_buffer
->state
.compute_pipeline
;
3552 cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= CIK
;
3553 bool pipeline_is_dirty
= pipeline
&&
3554 pipeline
!= cmd_buffer
->state
.emitted_compute_pipeline
;
3556 if (cmd_buffer
->state
.flush_bits
& (RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
3557 RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
3558 RADV_CMD_FLAG_PS_PARTIAL_FLUSH
|
3559 RADV_CMD_FLAG_CS_PARTIAL_FLUSH
)) {
3560 /* If we have to wait for idle, set all states first, so that
3561 * all SET packets are processed in parallel with previous draw
3562 * calls. Then upload descriptors, set shader pointers, and
3563 * dispatch, and prefetch at the end. This ensures that the
3564 * time the CUs are idle is very short. (there are only SET_SH
3565 * packets between the wait and the draw)
3567 radv_emit_compute_pipeline(cmd_buffer
);
3568 si_emit_cache_flush(cmd_buffer
);
3569 /* <-- CUs are idle here --> */
3571 radv_upload_compute_shader_descriptors(cmd_buffer
);
3573 radv_emit_dispatch_packets(cmd_buffer
, info
);
3574 /* <-- CUs are busy here --> */
3576 /* Start prefetches after the dispatch has been started. Both
3577 * will run in parallel, but starting the dispatch first is
3580 if (has_prefetch
&& pipeline_is_dirty
) {
3581 radv_emit_shader_prefetch(cmd_buffer
,
3582 pipeline
->shaders
[MESA_SHADER_COMPUTE
]);
3585 /* If we don't wait for idle, start prefetches first, then set
3586 * states, and dispatch at the end.
3588 si_emit_cache_flush(cmd_buffer
);
3590 if (has_prefetch
&& pipeline_is_dirty
) {
3591 radv_emit_shader_prefetch(cmd_buffer
,
3592 pipeline
->shaders
[MESA_SHADER_COMPUTE
]);
3595 radv_upload_compute_shader_descriptors(cmd_buffer
);
3597 radv_emit_compute_pipeline(cmd_buffer
);
3598 radv_emit_dispatch_packets(cmd_buffer
, info
);
3601 radv_cmd_buffer_after_draw(cmd_buffer
, RADV_CMD_FLAG_CS_PARTIAL_FLUSH
);
3604 void radv_CmdDispatchBase(
3605 VkCommandBuffer commandBuffer
,
3613 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3614 struct radv_dispatch_info info
= {};
3620 info
.offsets
[0] = base_x
;
3621 info
.offsets
[1] = base_y
;
3622 info
.offsets
[2] = base_z
;
3623 radv_dispatch(cmd_buffer
, &info
);
3626 void radv_CmdDispatch(
3627 VkCommandBuffer commandBuffer
,
3632 radv_CmdDispatchBase(commandBuffer
, 0, 0, 0, x
, y
, z
);
3635 void radv_CmdDispatchIndirect(
3636 VkCommandBuffer commandBuffer
,
3638 VkDeviceSize offset
)
3640 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3641 RADV_FROM_HANDLE(radv_buffer
, buffer
, _buffer
);
3642 struct radv_dispatch_info info
= {};
3644 info
.indirect
= buffer
;
3645 info
.indirect_offset
= offset
;
3647 radv_dispatch(cmd_buffer
, &info
);
3650 void radv_unaligned_dispatch(
3651 struct radv_cmd_buffer
*cmd_buffer
,
3656 struct radv_dispatch_info info
= {};
3663 radv_dispatch(cmd_buffer
, &info
);
3666 void radv_CmdEndRenderPass(
3667 VkCommandBuffer commandBuffer
)
3669 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3671 radv_subpass_barrier(cmd_buffer
, &cmd_buffer
->state
.pass
->end_barrier
);
3673 radv_cmd_buffer_resolve_subpass(cmd_buffer
);
3675 for (unsigned i
= 0; i
< cmd_buffer
->state
.framebuffer
->attachment_count
; ++i
) {
3676 VkImageLayout layout
= cmd_buffer
->state
.pass
->attachments
[i
].final_layout
;
3677 radv_handle_subpass_image_transition(cmd_buffer
,
3678 (VkAttachmentReference
){i
, layout
});
3681 vk_free(&cmd_buffer
->pool
->alloc
, cmd_buffer
->state
.attachments
);
3683 cmd_buffer
->state
.pass
= NULL
;
3684 cmd_buffer
->state
.subpass
= NULL
;
3685 cmd_buffer
->state
.attachments
= NULL
;
3686 cmd_buffer
->state
.framebuffer
= NULL
;
3690 * For HTILE we have the following interesting clear words:
3691 * 0xfffff30f: Uncompressed, full depth range, for depth+stencil HTILE
3692 * 0xfffc000f: Uncompressed, full depth range, for depth only HTILE.
3693 * 0xfffffff0: Clear depth to 1.0
3694 * 0x00000000: Clear depth to 0.0
3696 static void radv_initialize_htile(struct radv_cmd_buffer
*cmd_buffer
,
3697 struct radv_image
*image
,
3698 const VkImageSubresourceRange
*range
,
3699 uint32_t clear_word
)
3701 assert(range
->baseMipLevel
== 0);
3702 assert(range
->levelCount
== 1 || range
->levelCount
== VK_REMAINING_ARRAY_LAYERS
);
3703 unsigned layer_count
= radv_get_layerCount(image
, range
);
3704 uint64_t size
= image
->surface
.htile_slice_size
* layer_count
;
3705 uint64_t offset
= image
->offset
+ image
->htile_offset
+
3706 image
->surface
.htile_slice_size
* range
->baseArrayLayer
;
3707 struct radv_cmd_state
*state
= &cmd_buffer
->state
;
3709 state
->flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
3710 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
;
3712 state
->flush_bits
|= radv_fill_buffer(cmd_buffer
, image
->bo
, offset
,
3715 state
->flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
;
3718 static void radv_handle_depth_image_transition(struct radv_cmd_buffer
*cmd_buffer
,
3719 struct radv_image
*image
,
3720 VkImageLayout src_layout
,
3721 VkImageLayout dst_layout
,
3722 unsigned src_queue_mask
,
3723 unsigned dst_queue_mask
,
3724 const VkImageSubresourceRange
*range
,
3725 VkImageAspectFlags pending_clears
)
3727 if (!radv_image_has_htile(image
))
3730 if (dst_layout
== VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
&&
3731 (pending_clears
& vk_format_aspects(image
->vk_format
)) == vk_format_aspects(image
->vk_format
) &&
3732 cmd_buffer
->state
.render_area
.offset
.x
== 0 && cmd_buffer
->state
.render_area
.offset
.y
== 0 &&
3733 cmd_buffer
->state
.render_area
.extent
.width
== image
->info
.width
&&
3734 cmd_buffer
->state
.render_area
.extent
.height
== image
->info
.height
) {
3735 /* The clear will initialize htile. */
3737 } else if (src_layout
== VK_IMAGE_LAYOUT_UNDEFINED
&&
3738 radv_layout_has_htile(image
, dst_layout
, dst_queue_mask
)) {
3739 /* TODO: merge with the clear if applicable */
3740 radv_initialize_htile(cmd_buffer
, image
, range
, 0);
3741 } else if (!radv_layout_is_htile_compressed(image
, src_layout
, src_queue_mask
) &&
3742 radv_layout_is_htile_compressed(image
, dst_layout
, dst_queue_mask
)) {
3743 uint32_t clear_value
= vk_format_is_stencil(image
->vk_format
) ? 0xfffff30f : 0xfffc000f;
3744 radv_initialize_htile(cmd_buffer
, image
, range
, clear_value
);
3745 } else if (radv_layout_is_htile_compressed(image
, src_layout
, src_queue_mask
) &&
3746 !radv_layout_is_htile_compressed(image
, dst_layout
, dst_queue_mask
)) {
3747 VkImageSubresourceRange local_range
= *range
;
3748 local_range
.aspectMask
= VK_IMAGE_ASPECT_DEPTH_BIT
;
3749 local_range
.baseMipLevel
= 0;
3750 local_range
.levelCount
= 1;
3752 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
3753 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
;
3755 radv_decompress_depth_image_inplace(cmd_buffer
, image
, &local_range
);
3757 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
3758 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
;
3762 static void radv_initialise_cmask(struct radv_cmd_buffer
*cmd_buffer
,
3763 struct radv_image
*image
, uint32_t value
)
3765 struct radv_cmd_state
*state
= &cmd_buffer
->state
;
3767 state
->flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
3768 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
;
3770 state
->flush_bits
|= radv_clear_cmask(cmd_buffer
, image
, value
);
3772 state
->flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
;
3775 void radv_initialize_dcc(struct radv_cmd_buffer
*cmd_buffer
,
3776 struct radv_image
*image
, uint32_t value
)
3778 struct radv_cmd_state
*state
= &cmd_buffer
->state
;
3780 state
->flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
3781 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
;
3783 state
->flush_bits
|= radv_clear_dcc(cmd_buffer
, image
, value
);
3785 state
->flush_bits
|= RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
3786 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
;
3790 * Initialize DCC/FMASK/CMASK metadata for a color image.
3792 static void radv_init_color_image_metadata(struct radv_cmd_buffer
*cmd_buffer
,
3793 struct radv_image
*image
,
3794 VkImageLayout src_layout
,
3795 VkImageLayout dst_layout
,
3796 unsigned src_queue_mask
,
3797 unsigned dst_queue_mask
)
3799 if (radv_image_has_cmask(image
)) {
3800 uint32_t value
= 0xffffffffu
; /* Fully expanded mode. */
3802 /* TODO: clarify this. */
3803 if (radv_image_has_fmask(image
)) {
3804 value
= 0xccccccccu
;
3807 radv_initialise_cmask(cmd_buffer
, image
, value
);
3810 if (radv_image_has_dcc(image
)) {
3811 uint32_t value
= 0xffffffffu
; /* Fully expanded mode. */
3813 if (radv_layout_dcc_compressed(image
, dst_layout
,
3815 value
= 0x20202020u
;
3818 radv_initialize_dcc(cmd_buffer
, image
, value
);
3823 * Handle color image transitions for DCC/FMASK/CMASK.
3825 static void radv_handle_color_image_transition(struct radv_cmd_buffer
*cmd_buffer
,
3826 struct radv_image
*image
,
3827 VkImageLayout src_layout
,
3828 VkImageLayout dst_layout
,
3829 unsigned src_queue_mask
,
3830 unsigned dst_queue_mask
,
3831 const VkImageSubresourceRange
*range
)
3833 if (src_layout
== VK_IMAGE_LAYOUT_UNDEFINED
) {
3834 radv_init_color_image_metadata(cmd_buffer
, image
,
3835 src_layout
, dst_layout
,
3836 src_queue_mask
, dst_queue_mask
);
3840 if (radv_image_has_dcc(image
)) {
3841 if (src_layout
== VK_IMAGE_LAYOUT_PREINITIALIZED
) {
3842 radv_initialize_dcc(cmd_buffer
, image
, 0xffffffffu
);
3843 } else if (radv_layout_dcc_compressed(image
, src_layout
, src_queue_mask
) &&
3844 !radv_layout_dcc_compressed(image
, dst_layout
, dst_queue_mask
)) {
3845 radv_decompress_dcc(cmd_buffer
, image
, range
);
3846 } else if (radv_layout_can_fast_clear(image
, src_layout
, src_queue_mask
) &&
3847 !radv_layout_can_fast_clear(image
, dst_layout
, dst_queue_mask
)) {
3848 radv_fast_clear_flush_image_inplace(cmd_buffer
, image
, range
);
3850 } else if (radv_image_has_cmask(image
) || radv_image_has_fmask(image
)) {
3851 if (radv_layout_can_fast_clear(image
, src_layout
, src_queue_mask
) &&
3852 !radv_layout_can_fast_clear(image
, dst_layout
, dst_queue_mask
)) {
3853 radv_fast_clear_flush_image_inplace(cmd_buffer
, image
, range
);
3858 static void radv_handle_image_transition(struct radv_cmd_buffer
*cmd_buffer
,
3859 struct radv_image
*image
,
3860 VkImageLayout src_layout
,
3861 VkImageLayout dst_layout
,
3862 uint32_t src_family
,
3863 uint32_t dst_family
,
3864 const VkImageSubresourceRange
*range
,
3865 VkImageAspectFlags pending_clears
)
3867 if (image
->exclusive
&& src_family
!= dst_family
) {
3868 /* This is an acquire or a release operation and there will be
3869 * a corresponding release/acquire. Do the transition in the
3870 * most flexible queue. */
3872 assert(src_family
== cmd_buffer
->queue_family_index
||
3873 dst_family
== cmd_buffer
->queue_family_index
);
3875 if (cmd_buffer
->queue_family_index
== RADV_QUEUE_TRANSFER
)
3878 if (cmd_buffer
->queue_family_index
== RADV_QUEUE_COMPUTE
&&
3879 (src_family
== RADV_QUEUE_GENERAL
||
3880 dst_family
== RADV_QUEUE_GENERAL
))
3884 unsigned src_queue_mask
=
3885 radv_image_queue_family_mask(image
, src_family
,
3886 cmd_buffer
->queue_family_index
);
3887 unsigned dst_queue_mask
=
3888 radv_image_queue_family_mask(image
, dst_family
,
3889 cmd_buffer
->queue_family_index
);
3891 if (vk_format_is_depth(image
->vk_format
)) {
3892 radv_handle_depth_image_transition(cmd_buffer
, image
,
3893 src_layout
, dst_layout
,
3894 src_queue_mask
, dst_queue_mask
,
3895 range
, pending_clears
);
3897 radv_handle_color_image_transition(cmd_buffer
, image
,
3898 src_layout
, dst_layout
,
3899 src_queue_mask
, dst_queue_mask
,
3904 void radv_CmdPipelineBarrier(
3905 VkCommandBuffer commandBuffer
,
3906 VkPipelineStageFlags srcStageMask
,
3907 VkPipelineStageFlags destStageMask
,
3909 uint32_t memoryBarrierCount
,
3910 const VkMemoryBarrier
* pMemoryBarriers
,
3911 uint32_t bufferMemoryBarrierCount
,
3912 const VkBufferMemoryBarrier
* pBufferMemoryBarriers
,
3913 uint32_t imageMemoryBarrierCount
,
3914 const VkImageMemoryBarrier
* pImageMemoryBarriers
)
3916 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3917 enum radv_cmd_flush_bits src_flush_bits
= 0;
3918 enum radv_cmd_flush_bits dst_flush_bits
= 0;
3920 for (uint32_t i
= 0; i
< memoryBarrierCount
; i
++) {
3921 src_flush_bits
|= radv_src_access_flush(cmd_buffer
, pMemoryBarriers
[i
].srcAccessMask
);
3922 dst_flush_bits
|= radv_dst_access_flush(cmd_buffer
, pMemoryBarriers
[i
].dstAccessMask
,
3926 for (uint32_t i
= 0; i
< bufferMemoryBarrierCount
; i
++) {
3927 src_flush_bits
|= radv_src_access_flush(cmd_buffer
, pBufferMemoryBarriers
[i
].srcAccessMask
);
3928 dst_flush_bits
|= radv_dst_access_flush(cmd_buffer
, pBufferMemoryBarriers
[i
].dstAccessMask
,
3932 for (uint32_t i
= 0; i
< imageMemoryBarrierCount
; i
++) {
3933 RADV_FROM_HANDLE(radv_image
, image
, pImageMemoryBarriers
[i
].image
);
3934 src_flush_bits
|= radv_src_access_flush(cmd_buffer
, pImageMemoryBarriers
[i
].srcAccessMask
);
3935 dst_flush_bits
|= radv_dst_access_flush(cmd_buffer
, pImageMemoryBarriers
[i
].dstAccessMask
,
3939 radv_stage_flush(cmd_buffer
, srcStageMask
);
3940 cmd_buffer
->state
.flush_bits
|= src_flush_bits
;
3942 for (uint32_t i
= 0; i
< imageMemoryBarrierCount
; i
++) {
3943 RADV_FROM_HANDLE(radv_image
, image
, pImageMemoryBarriers
[i
].image
);
3944 radv_handle_image_transition(cmd_buffer
, image
,
3945 pImageMemoryBarriers
[i
].oldLayout
,
3946 pImageMemoryBarriers
[i
].newLayout
,
3947 pImageMemoryBarriers
[i
].srcQueueFamilyIndex
,
3948 pImageMemoryBarriers
[i
].dstQueueFamilyIndex
,
3949 &pImageMemoryBarriers
[i
].subresourceRange
,
3953 cmd_buffer
->state
.flush_bits
|= dst_flush_bits
;
3957 static void write_event(struct radv_cmd_buffer
*cmd_buffer
,
3958 struct radv_event
*event
,
3959 VkPipelineStageFlags stageMask
,
3962 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
3963 uint64_t va
= radv_buffer_get_va(event
->bo
);
3965 radv_cs_add_buffer(cmd_buffer
->device
->ws
, cs
, event
->bo
, 8);
3967 MAYBE_UNUSED
unsigned cdw_max
= radeon_check_space(cmd_buffer
->device
->ws
, cs
, 18);
3969 /* TODO: this is overkill. Probably should figure something out from
3970 * the stage mask. */
3972 si_cs_emit_write_event_eop(cs
,
3973 cmd_buffer
->state
.predicating
,
3974 cmd_buffer
->device
->physical_device
->rad_info
.chip_class
,
3975 radv_cmd_buffer_uses_mec(cmd_buffer
),
3976 V_028A90_BOTTOM_OF_PIPE_TS
, 0,
3979 assert(cmd_buffer
->cs
->cdw
<= cdw_max
);
3982 void radv_CmdSetEvent(VkCommandBuffer commandBuffer
,
3984 VkPipelineStageFlags stageMask
)
3986 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3987 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3989 write_event(cmd_buffer
, event
, stageMask
, 1);
3992 void radv_CmdResetEvent(VkCommandBuffer commandBuffer
,
3994 VkPipelineStageFlags stageMask
)
3996 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3997 RADV_FROM_HANDLE(radv_event
, event
, _event
);
3999 write_event(cmd_buffer
, event
, stageMask
, 0);
4002 void radv_CmdWaitEvents(VkCommandBuffer commandBuffer
,
4003 uint32_t eventCount
,
4004 const VkEvent
* pEvents
,
4005 VkPipelineStageFlags srcStageMask
,
4006 VkPipelineStageFlags dstStageMask
,
4007 uint32_t memoryBarrierCount
,
4008 const VkMemoryBarrier
* pMemoryBarriers
,
4009 uint32_t bufferMemoryBarrierCount
,
4010 const VkBufferMemoryBarrier
* pBufferMemoryBarriers
,
4011 uint32_t imageMemoryBarrierCount
,
4012 const VkImageMemoryBarrier
* pImageMemoryBarriers
)
4014 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
4015 struct radeon_winsys_cs
*cs
= cmd_buffer
->cs
;
4017 for (unsigned i
= 0; i
< eventCount
; ++i
) {
4018 RADV_FROM_HANDLE(radv_event
, event
, pEvents
[i
]);
4019 uint64_t va
= radv_buffer_get_va(event
->bo
);
4021 radv_cs_add_buffer(cmd_buffer
->device
->ws
, cs
, event
->bo
, 8);
4023 MAYBE_UNUSED
unsigned cdw_max
= radeon_check_space(cmd_buffer
->device
->ws
, cs
, 7);
4025 si_emit_wait_fence(cs
, false, va
, 1, 0xffffffff);
4026 assert(cmd_buffer
->cs
->cdw
<= cdw_max
);
4030 for (uint32_t i
= 0; i
< imageMemoryBarrierCount
; i
++) {
4031 RADV_FROM_HANDLE(radv_image
, image
, pImageMemoryBarriers
[i
].image
);
4033 radv_handle_image_transition(cmd_buffer
, image
,
4034 pImageMemoryBarriers
[i
].oldLayout
,
4035 pImageMemoryBarriers
[i
].newLayout
,
4036 pImageMemoryBarriers
[i
].srcQueueFamilyIndex
,
4037 pImageMemoryBarriers
[i
].dstQueueFamilyIndex
,
4038 &pImageMemoryBarriers
[i
].subresourceRange
,
4042 /* TODO: figure out how to do memory barriers without waiting */
4043 cmd_buffer
->state
.flush_bits
|= RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER
|
4044 RADV_CMD_FLAG_INV_GLOBAL_L2
|
4045 RADV_CMD_FLAG_INV_VMEM_L1
|
4046 RADV_CMD_FLAG_INV_SMEM_L1
;
4050 void radv_CmdSetDeviceMask(VkCommandBuffer commandBuffer
,
4051 uint32_t deviceMask
)