2 * Copyright © 2015 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 #include "radv_debug.h"
25 #include "radv_meta.h"
26 #include "radv_private.h"
27 #include "nir/nir_builder.h"
29 #include "util/format_rgb9e5.h"
30 #include "vk_format.h"
34 DEPTH_CLEAR_FAST_EXPCLEAR
,
35 DEPTH_CLEAR_FAST_NO_EXPCLEAR
39 build_color_shaders(struct nir_shader
**out_vs
,
40 struct nir_shader
**out_fs
,
46 nir_builder_init_simple_shader(&vs_b
, NULL
, MESA_SHADER_VERTEX
, NULL
);
47 nir_builder_init_simple_shader(&fs_b
, NULL
, MESA_SHADER_FRAGMENT
, NULL
);
49 vs_b
.shader
->info
.name
= ralloc_strdup(vs_b
.shader
, "meta_clear_color_vs");
50 fs_b
.shader
->info
.name
= ralloc_strdup(fs_b
.shader
, "meta_clear_color_fs");
52 const struct glsl_type
*position_type
= glsl_vec4_type();
53 const struct glsl_type
*color_type
= glsl_vec4_type();
55 nir_variable
*vs_out_pos
=
56 nir_variable_create(vs_b
.shader
, nir_var_shader_out
, position_type
,
58 vs_out_pos
->data
.location
= VARYING_SLOT_POS
;
60 nir_intrinsic_instr
*in_color_load
= nir_intrinsic_instr_create(fs_b
.shader
, nir_intrinsic_load_push_constant
);
61 nir_intrinsic_set_base(in_color_load
, 0);
62 nir_intrinsic_set_range(in_color_load
, 16);
63 in_color_load
->src
[0] = nir_src_for_ssa(nir_imm_int(&fs_b
, 0));
64 in_color_load
->num_components
= 4;
65 nir_ssa_dest_init(&in_color_load
->instr
, &in_color_load
->dest
, 4, 32, "clear color");
66 nir_builder_instr_insert(&fs_b
, &in_color_load
->instr
);
68 nir_variable
*fs_out_color
=
69 nir_variable_create(fs_b
.shader
, nir_var_shader_out
, color_type
,
71 fs_out_color
->data
.location
= FRAG_RESULT_DATA0
+ frag_output
;
73 nir_store_var(&fs_b
, fs_out_color
, &in_color_load
->dest
.ssa
, 0xf);
75 nir_ssa_def
*outvec
= radv_meta_gen_rect_vertices(&vs_b
);
76 nir_store_var(&vs_b
, vs_out_pos
, outvec
, 0xf);
78 const struct glsl_type
*layer_type
= glsl_int_type();
79 nir_variable
*vs_out_layer
=
80 nir_variable_create(vs_b
.shader
, nir_var_shader_out
, layer_type
,
82 vs_out_layer
->data
.location
= VARYING_SLOT_LAYER
;
83 vs_out_layer
->data
.interpolation
= INTERP_MODE_FLAT
;
84 nir_ssa_def
*inst_id
= nir_load_instance_id(&vs_b
);
85 nir_ssa_def
*base_instance
= nir_load_base_instance(&vs_b
);
87 nir_ssa_def
*layer_id
= nir_iadd(&vs_b
, inst_id
, base_instance
);
88 nir_store_var(&vs_b
, vs_out_layer
, layer_id
, 0x1);
90 *out_vs
= vs_b
.shader
;
91 *out_fs
= fs_b
.shader
;
95 create_pipeline(struct radv_device
*device
,
96 struct radv_render_pass
*render_pass
,
98 struct nir_shader
*vs_nir
,
99 struct nir_shader
*fs_nir
,
100 const VkPipelineVertexInputStateCreateInfo
*vi_state
,
101 const VkPipelineDepthStencilStateCreateInfo
*ds_state
,
102 const VkPipelineColorBlendStateCreateInfo
*cb_state
,
103 const VkPipelineLayout layout
,
104 const struct radv_graphics_pipeline_create_info
*extra
,
105 const VkAllocationCallbacks
*alloc
,
106 VkPipeline
*pipeline
)
108 VkDevice device_h
= radv_device_to_handle(device
);
111 struct radv_shader_module vs_m
= { .nir
= vs_nir
};
112 struct radv_shader_module fs_m
= { .nir
= fs_nir
};
114 result
= radv_graphics_pipeline_create(device_h
,
115 radv_pipeline_cache_to_handle(&device
->meta_state
.cache
),
116 &(VkGraphicsPipelineCreateInfo
) {
117 .sType
= VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO
,
118 .stageCount
= fs_nir
? 2 : 1,
119 .pStages
= (VkPipelineShaderStageCreateInfo
[]) {
121 .sType
= VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO
,
122 .stage
= VK_SHADER_STAGE_VERTEX_BIT
,
123 .module
= radv_shader_module_to_handle(&vs_m
),
127 .sType
= VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO
,
128 .stage
= VK_SHADER_STAGE_FRAGMENT_BIT
,
129 .module
= radv_shader_module_to_handle(&fs_m
),
133 .pVertexInputState
= vi_state
,
134 .pInputAssemblyState
= &(VkPipelineInputAssemblyStateCreateInfo
) {
135 .sType
= VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO
,
136 .topology
= VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
,
137 .primitiveRestartEnable
= false,
139 .pViewportState
= &(VkPipelineViewportStateCreateInfo
) {
140 .sType
= VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO
,
144 .pRasterizationState
= &(VkPipelineRasterizationStateCreateInfo
) {
145 .sType
= VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO
,
146 .rasterizerDiscardEnable
= false,
147 .polygonMode
= VK_POLYGON_MODE_FILL
,
148 .cullMode
= VK_CULL_MODE_NONE
,
149 .frontFace
= VK_FRONT_FACE_COUNTER_CLOCKWISE
,
150 .depthBiasEnable
= false,
152 .pMultisampleState
= &(VkPipelineMultisampleStateCreateInfo
) {
153 .sType
= VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO
,
154 .rasterizationSamples
= samples
,
155 .sampleShadingEnable
= false,
157 .alphaToCoverageEnable
= false,
158 .alphaToOneEnable
= false,
160 .pDepthStencilState
= ds_state
,
161 .pColorBlendState
= cb_state
,
162 .pDynamicState
= &(VkPipelineDynamicStateCreateInfo
) {
163 /* The meta clear pipeline declares all state as dynamic.
164 * As a consequence, vkCmdBindPipeline writes no dynamic state
165 * to the cmd buffer. Therefore, at the end of the meta clear,
166 * we need only restore dynamic state was vkCmdSet.
168 .sType
= VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO
,
169 .dynamicStateCount
= 8,
170 .pDynamicStates
= (VkDynamicState
[]) {
171 /* Everything except stencil write mask */
172 VK_DYNAMIC_STATE_VIEWPORT
,
173 VK_DYNAMIC_STATE_SCISSOR
,
174 VK_DYNAMIC_STATE_LINE_WIDTH
,
175 VK_DYNAMIC_STATE_DEPTH_BIAS
,
176 VK_DYNAMIC_STATE_BLEND_CONSTANTS
,
177 VK_DYNAMIC_STATE_DEPTH_BOUNDS
,
178 VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK
,
179 VK_DYNAMIC_STATE_STENCIL_REFERENCE
,
184 .renderPass
= radv_render_pass_to_handle(render_pass
),
198 create_color_renderpass(struct radv_device
*device
,
203 mtx_lock(&device
->meta_state
.mtx
);
205 mtx_unlock (&device
->meta_state
.mtx
);
209 VkResult result
= radv_CreateRenderPass(radv_device_to_handle(device
),
210 &(VkRenderPassCreateInfo
) {
211 .sType
= VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO
,
212 .attachmentCount
= 1,
213 .pAttachments
= &(VkAttachmentDescription
) {
216 .loadOp
= VK_ATTACHMENT_LOAD_OP_LOAD
,
217 .storeOp
= VK_ATTACHMENT_STORE_OP_STORE
,
218 .initialLayout
= VK_IMAGE_LAYOUT_GENERAL
,
219 .finalLayout
= VK_IMAGE_LAYOUT_GENERAL
,
222 .pSubpasses
= &(VkSubpassDescription
) {
223 .pipelineBindPoint
= VK_PIPELINE_BIND_POINT_GRAPHICS
,
224 .inputAttachmentCount
= 0,
225 .colorAttachmentCount
= 1,
226 .pColorAttachments
= &(VkAttachmentReference
) {
228 .layout
= VK_IMAGE_LAYOUT_GENERAL
,
230 .pResolveAttachments
= NULL
,
231 .pDepthStencilAttachment
= &(VkAttachmentReference
) {
232 .attachment
= VK_ATTACHMENT_UNUSED
,
233 .layout
= VK_IMAGE_LAYOUT_GENERAL
,
235 .preserveAttachmentCount
= 0,
236 .pPreserveAttachments
= NULL
,
238 .dependencyCount
= 0,
239 }, &device
->meta_state
.alloc
, pass
);
240 mtx_unlock(&device
->meta_state
.mtx
);
245 create_color_pipeline(struct radv_device
*device
,
247 uint32_t frag_output
,
248 VkPipeline
*pipeline
,
251 struct nir_shader
*vs_nir
;
252 struct nir_shader
*fs_nir
;
255 mtx_lock(&device
->meta_state
.mtx
);
257 mtx_unlock(&device
->meta_state
.mtx
);
261 build_color_shaders(&vs_nir
, &fs_nir
, frag_output
);
263 const VkPipelineVertexInputStateCreateInfo vi_state
= {
264 .sType
= VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO
,
265 .vertexBindingDescriptionCount
= 0,
266 .vertexAttributeDescriptionCount
= 0,
269 const VkPipelineDepthStencilStateCreateInfo ds_state
= {
270 .sType
= VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO
,
271 .depthTestEnable
= false,
272 .depthWriteEnable
= false,
273 .depthBoundsTestEnable
= false,
274 .stencilTestEnable
= false,
277 VkPipelineColorBlendAttachmentState blend_attachment_state
[MAX_RTS
] = { 0 };
278 blend_attachment_state
[frag_output
] = (VkPipelineColorBlendAttachmentState
) {
279 .blendEnable
= false,
280 .colorWriteMask
= VK_COLOR_COMPONENT_A_BIT
|
281 VK_COLOR_COMPONENT_R_BIT
|
282 VK_COLOR_COMPONENT_G_BIT
|
283 VK_COLOR_COMPONENT_B_BIT
,
286 const VkPipelineColorBlendStateCreateInfo cb_state
= {
287 .sType
= VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO
,
288 .logicOpEnable
= false,
289 .attachmentCount
= MAX_RTS
,
290 .pAttachments
= blend_attachment_state
294 struct radv_graphics_pipeline_create_info extra
= {
295 .use_rectlist
= true,
297 result
= create_pipeline(device
, radv_render_pass_from_handle(pass
),
298 samples
, vs_nir
, fs_nir
, &vi_state
, &ds_state
, &cb_state
,
299 device
->meta_state
.clear_color_p_layout
,
300 &extra
, &device
->meta_state
.alloc
, pipeline
);
302 mtx_unlock(&device
->meta_state
.mtx
);
307 finish_meta_clear_htile_mask_state(struct radv_device
*device
)
309 struct radv_meta_state
*state
= &device
->meta_state
;
311 radv_DestroyPipeline(radv_device_to_handle(device
),
312 state
->clear_htile_mask_pipeline
,
314 radv_DestroyPipelineLayout(radv_device_to_handle(device
),
315 state
->clear_htile_mask_p_layout
,
317 radv_DestroyDescriptorSetLayout(radv_device_to_handle(device
),
318 state
->clear_htile_mask_ds_layout
,
323 radv_device_finish_meta_clear_state(struct radv_device
*device
)
325 struct radv_meta_state
*state
= &device
->meta_state
;
327 for (uint32_t i
= 0; i
< ARRAY_SIZE(state
->clear
); ++i
) {
328 for (uint32_t j
= 0; j
< ARRAY_SIZE(state
->clear
[i
].color_pipelines
); ++j
) {
329 radv_DestroyPipeline(radv_device_to_handle(device
),
330 state
->clear
[i
].color_pipelines
[j
],
332 radv_DestroyRenderPass(radv_device_to_handle(device
),
333 state
->clear
[i
].render_pass
[j
],
337 for (uint32_t j
= 0; j
< NUM_DEPTH_CLEAR_PIPELINES
; j
++) {
338 radv_DestroyPipeline(radv_device_to_handle(device
),
339 state
->clear
[i
].depth_only_pipeline
[j
],
341 radv_DestroyPipeline(radv_device_to_handle(device
),
342 state
->clear
[i
].stencil_only_pipeline
[j
],
344 radv_DestroyPipeline(radv_device_to_handle(device
),
345 state
->clear
[i
].depthstencil_pipeline
[j
],
348 radv_DestroyRenderPass(radv_device_to_handle(device
),
349 state
->clear
[i
].depthstencil_rp
,
352 radv_DestroyPipelineLayout(radv_device_to_handle(device
),
353 state
->clear_color_p_layout
,
355 radv_DestroyPipelineLayout(radv_device_to_handle(device
),
356 state
->clear_depth_p_layout
,
359 finish_meta_clear_htile_mask_state(device
);
363 emit_color_clear(struct radv_cmd_buffer
*cmd_buffer
,
364 const VkClearAttachment
*clear_att
,
365 const VkClearRect
*clear_rect
,
368 struct radv_device
*device
= cmd_buffer
->device
;
369 const struct radv_subpass
*subpass
= cmd_buffer
->state
.subpass
;
370 const uint32_t subpass_att
= clear_att
->colorAttachment
;
371 const uint32_t pass_att
= subpass
->color_attachments
[subpass_att
].attachment
;
372 const struct radv_image_view
*iview
= cmd_buffer
->state
.attachments
?
373 cmd_buffer
->state
.attachments
[pass_att
].iview
: NULL
;
374 uint32_t samples
, samples_log2
;
377 VkClearColorValue clear_value
= clear_att
->clearValue
.color
;
378 VkCommandBuffer cmd_buffer_h
= radv_cmd_buffer_to_handle(cmd_buffer
);
381 /* When a framebuffer is bound to the current command buffer, get the
382 * number of samples from it. Otherwise, get the number of samples from
383 * the render pass because it's likely a secondary command buffer.
386 samples
= iview
->image
->info
.samples
;
387 format
= iview
->vk_format
;
389 samples
= cmd_buffer
->state
.pass
->attachments
[pass_att
].samples
;
390 format
= cmd_buffer
->state
.pass
->attachments
[pass_att
].format
;
393 samples_log2
= ffs(samples
) - 1;
394 fs_key
= radv_format_meta_fs_key(format
);
397 radv_finishme("color clears incomplete");
401 if (device
->meta_state
.clear
[samples_log2
].render_pass
[fs_key
] == VK_NULL_HANDLE
) {
402 VkResult ret
= create_color_renderpass(device
, radv_fs_key_format_exemplars
[fs_key
],
404 &device
->meta_state
.clear
[samples_log2
].render_pass
[fs_key
]);
405 if (ret
!= VK_SUCCESS
) {
406 cmd_buffer
->record_result
= ret
;
411 if (device
->meta_state
.clear
[samples_log2
].color_pipelines
[fs_key
] == VK_NULL_HANDLE
) {
412 VkResult ret
= create_color_pipeline(device
, samples
, 0,
413 &device
->meta_state
.clear
[samples_log2
].color_pipelines
[fs_key
],
414 device
->meta_state
.clear
[samples_log2
].render_pass
[fs_key
]);
415 if (ret
!= VK_SUCCESS
) {
416 cmd_buffer
->record_result
= ret
;
421 pipeline
= device
->meta_state
.clear
[samples_log2
].color_pipelines
[fs_key
];
423 radv_finishme("color clears incomplete");
426 assert(samples_log2
< ARRAY_SIZE(device
->meta_state
.clear
));
428 assert(clear_att
->aspectMask
== VK_IMAGE_ASPECT_COLOR_BIT
);
429 assert(clear_att
->colorAttachment
< subpass
->color_count
);
431 radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer
),
432 device
->meta_state
.clear_color_p_layout
,
433 VK_SHADER_STAGE_FRAGMENT_BIT
, 0, 16,
436 struct radv_subpass clear_subpass
= {
438 .color_attachments
= (struct radv_subpass_attachment
[]) {
439 subpass
->color_attachments
[clear_att
->colorAttachment
]
441 .depth_stencil_attachment
= NULL
,
444 radv_cmd_buffer_set_subpass(cmd_buffer
, &clear_subpass
);
446 radv_CmdBindPipeline(cmd_buffer_h
, VK_PIPELINE_BIND_POINT_GRAPHICS
,
449 radv_CmdSetViewport(radv_cmd_buffer_to_handle(cmd_buffer
), 0, 1, &(VkViewport
) {
450 .x
= clear_rect
->rect
.offset
.x
,
451 .y
= clear_rect
->rect
.offset
.y
,
452 .width
= clear_rect
->rect
.extent
.width
,
453 .height
= clear_rect
->rect
.extent
.height
,
458 radv_CmdSetScissor(radv_cmd_buffer_to_handle(cmd_buffer
), 0, 1, &clear_rect
->rect
);
462 for_each_bit(i
, view_mask
)
463 radv_CmdDraw(cmd_buffer_h
, 3, 1, 0, i
);
465 radv_CmdDraw(cmd_buffer_h
, 3, clear_rect
->layerCount
, 0, clear_rect
->baseArrayLayer
);
468 radv_cmd_buffer_set_subpass(cmd_buffer
, subpass
);
473 build_depthstencil_shader(struct nir_shader
**out_vs
, struct nir_shader
**out_fs
)
475 nir_builder vs_b
, fs_b
;
477 nir_builder_init_simple_shader(&vs_b
, NULL
, MESA_SHADER_VERTEX
, NULL
);
478 nir_builder_init_simple_shader(&fs_b
, NULL
, MESA_SHADER_FRAGMENT
, NULL
);
480 vs_b
.shader
->info
.name
= ralloc_strdup(vs_b
.shader
, "meta_clear_depthstencil_vs");
481 fs_b
.shader
->info
.name
= ralloc_strdup(fs_b
.shader
, "meta_clear_depthstencil_fs");
482 const struct glsl_type
*position_out_type
= glsl_vec4_type();
484 nir_variable
*vs_out_pos
=
485 nir_variable_create(vs_b
.shader
, nir_var_shader_out
, position_out_type
,
487 vs_out_pos
->data
.location
= VARYING_SLOT_POS
;
489 nir_intrinsic_instr
*in_color_load
= nir_intrinsic_instr_create(fs_b
.shader
, nir_intrinsic_load_push_constant
);
490 nir_intrinsic_set_base(in_color_load
, 0);
491 nir_intrinsic_set_range(in_color_load
, 4);
492 in_color_load
->src
[0] = nir_src_for_ssa(nir_imm_int(&fs_b
, 0));
493 in_color_load
->num_components
= 1;
494 nir_ssa_dest_init(&in_color_load
->instr
, &in_color_load
->dest
, 1, 32, "depth value");
495 nir_builder_instr_insert(&fs_b
, &in_color_load
->instr
);
497 nir_variable
*fs_out_depth
=
498 nir_variable_create(fs_b
.shader
, nir_var_shader_out
,
499 glsl_int_type(), "f_depth");
500 fs_out_depth
->data
.location
= FRAG_RESULT_DEPTH
;
501 nir_store_var(&fs_b
, fs_out_depth
, &in_color_load
->dest
.ssa
, 0x1);
503 nir_ssa_def
*outvec
= radv_meta_gen_rect_vertices(&vs_b
);
504 nir_store_var(&vs_b
, vs_out_pos
, outvec
, 0xf);
506 const struct glsl_type
*layer_type
= glsl_int_type();
507 nir_variable
*vs_out_layer
=
508 nir_variable_create(vs_b
.shader
, nir_var_shader_out
, layer_type
,
510 vs_out_layer
->data
.location
= VARYING_SLOT_LAYER
;
511 vs_out_layer
->data
.interpolation
= INTERP_MODE_FLAT
;
512 nir_ssa_def
*inst_id
= nir_load_instance_id(&vs_b
);
513 nir_ssa_def
*base_instance
= nir_load_base_instance(&vs_b
);
515 nir_ssa_def
*layer_id
= nir_iadd(&vs_b
, inst_id
, base_instance
);
516 nir_store_var(&vs_b
, vs_out_layer
, layer_id
, 0x1);
518 *out_vs
= vs_b
.shader
;
519 *out_fs
= fs_b
.shader
;
523 create_depthstencil_renderpass(struct radv_device
*device
,
525 VkRenderPass
*render_pass
)
527 mtx_lock(&device
->meta_state
.mtx
);
529 mtx_unlock(&device
->meta_state
.mtx
);
533 VkResult result
= radv_CreateRenderPass(radv_device_to_handle(device
),
534 &(VkRenderPassCreateInfo
) {
535 .sType
= VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO
,
536 .attachmentCount
= 1,
537 .pAttachments
= &(VkAttachmentDescription
) {
538 .format
= VK_FORMAT_D32_SFLOAT_S8_UINT
,
540 .loadOp
= VK_ATTACHMENT_LOAD_OP_LOAD
,
541 .storeOp
= VK_ATTACHMENT_STORE_OP_STORE
,
542 .initialLayout
= VK_IMAGE_LAYOUT_GENERAL
,
543 .finalLayout
= VK_IMAGE_LAYOUT_GENERAL
,
546 .pSubpasses
= &(VkSubpassDescription
) {
547 .pipelineBindPoint
= VK_PIPELINE_BIND_POINT_GRAPHICS
,
548 .inputAttachmentCount
= 0,
549 .colorAttachmentCount
= 0,
550 .pColorAttachments
= NULL
,
551 .pResolveAttachments
= NULL
,
552 .pDepthStencilAttachment
= &(VkAttachmentReference
) {
554 .layout
= VK_IMAGE_LAYOUT_GENERAL
,
556 .preserveAttachmentCount
= 0,
557 .pPreserveAttachments
= NULL
,
559 .dependencyCount
= 0,
560 }, &device
->meta_state
.alloc
, render_pass
);
561 mtx_unlock(&device
->meta_state
.mtx
);
566 create_depthstencil_pipeline(struct radv_device
*device
,
567 VkImageAspectFlags aspects
,
570 VkPipeline
*pipeline
,
571 VkRenderPass render_pass
)
573 struct nir_shader
*vs_nir
, *fs_nir
;
576 mtx_lock(&device
->meta_state
.mtx
);
578 mtx_unlock(&device
->meta_state
.mtx
);
582 build_depthstencil_shader(&vs_nir
, &fs_nir
);
584 const VkPipelineVertexInputStateCreateInfo vi_state
= {
585 .sType
= VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO
,
586 .vertexBindingDescriptionCount
= 0,
587 .vertexAttributeDescriptionCount
= 0,
590 const VkPipelineDepthStencilStateCreateInfo ds_state
= {
591 .sType
= VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO
,
592 .depthTestEnable
= (aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
),
593 .depthCompareOp
= VK_COMPARE_OP_ALWAYS
,
594 .depthWriteEnable
= (aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
),
595 .depthBoundsTestEnable
= false,
596 .stencilTestEnable
= (aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
),
598 .passOp
= VK_STENCIL_OP_REPLACE
,
599 .compareOp
= VK_COMPARE_OP_ALWAYS
,
600 .writeMask
= UINT32_MAX
,
601 .reference
= 0, /* dynamic */
603 .back
= { 0 /* dont care */ },
606 const VkPipelineColorBlendStateCreateInfo cb_state
= {
607 .sType
= VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO
,
608 .logicOpEnable
= false,
609 .attachmentCount
= 0,
610 .pAttachments
= NULL
,
613 struct radv_graphics_pipeline_create_info extra
= {
614 .use_rectlist
= true,
617 if (aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
618 extra
.db_depth_clear
= index
== DEPTH_CLEAR_SLOW
? false : true;
619 extra
.db_depth_disable_expclear
= index
== DEPTH_CLEAR_FAST_NO_EXPCLEAR
? true : false;
621 if (aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
622 extra
.db_stencil_clear
= index
== DEPTH_CLEAR_SLOW
? false : true;
623 extra
.db_stencil_disable_expclear
= index
== DEPTH_CLEAR_FAST_NO_EXPCLEAR
? true : false;
625 result
= create_pipeline(device
, radv_render_pass_from_handle(render_pass
),
626 samples
, vs_nir
, fs_nir
, &vi_state
, &ds_state
, &cb_state
,
627 device
->meta_state
.clear_depth_p_layout
,
628 &extra
, &device
->meta_state
.alloc
, pipeline
);
630 mtx_unlock(&device
->meta_state
.mtx
);
634 static bool depth_view_can_fast_clear(struct radv_cmd_buffer
*cmd_buffer
,
635 const struct radv_image_view
*iview
,
636 VkImageAspectFlags aspects
,
637 VkImageLayout layout
,
639 const VkClearRect
*clear_rect
,
640 VkClearDepthStencilValue clear_value
)
645 uint32_t queue_mask
= radv_image_queue_family_mask(iview
->image
,
646 cmd_buffer
->queue_family_index
,
647 cmd_buffer
->queue_family_index
);
648 if (clear_rect
->rect
.offset
.x
|| clear_rect
->rect
.offset
.y
||
649 clear_rect
->rect
.extent
.width
!= iview
->extent
.width
||
650 clear_rect
->rect
.extent
.height
!= iview
->extent
.height
)
652 if (radv_image_is_tc_compat_htile(iview
->image
) &&
653 (((aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) && clear_value
.depth
!= 0.0 &&
654 clear_value
.depth
!= 1.0) ||
655 ((aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) && clear_value
.stencil
!= 0)))
657 if (radv_image_has_htile(iview
->image
) &&
658 iview
->base_mip
== 0 &&
659 iview
->base_layer
== 0 &&
660 iview
->layer_count
== iview
->image
->info
.array_size
&&
661 radv_layout_is_htile_compressed(iview
->image
, layout
, in_render_loop
, queue_mask
) &&
662 radv_image_extent_compare(iview
->image
, &iview
->extent
))
668 pick_depthstencil_pipeline(struct radv_cmd_buffer
*cmd_buffer
,
669 struct radv_meta_state
*meta_state
,
670 const struct radv_image_view
*iview
,
672 VkImageAspectFlags aspects
,
673 VkImageLayout layout
,
675 const VkClearRect
*clear_rect
,
676 VkClearDepthStencilValue clear_value
)
678 bool fast
= depth_view_can_fast_clear(cmd_buffer
, iview
, aspects
, layout
,
679 in_render_loop
, clear_rect
, clear_value
);
680 int index
= DEPTH_CLEAR_SLOW
;
681 VkPipeline
*pipeline
;
684 /* we don't know the previous clear values, so we always have
685 * the NO_EXPCLEAR path */
686 index
= DEPTH_CLEAR_FAST_NO_EXPCLEAR
;
690 case VK_IMAGE_ASPECT_DEPTH_BIT
| VK_IMAGE_ASPECT_STENCIL_BIT
:
691 pipeline
= &meta_state
->clear
[samples_log2
].depthstencil_pipeline
[index
];
693 case VK_IMAGE_ASPECT_DEPTH_BIT
:
694 pipeline
= &meta_state
->clear
[samples_log2
].depth_only_pipeline
[index
];
696 case VK_IMAGE_ASPECT_STENCIL_BIT
:
697 pipeline
= &meta_state
->clear
[samples_log2
].stencil_only_pipeline
[index
];
700 unreachable("expected depth or stencil aspect");
703 if (cmd_buffer
->device
->meta_state
.clear
[samples_log2
].depthstencil_rp
== VK_NULL_HANDLE
) {
704 VkResult ret
= create_depthstencil_renderpass(cmd_buffer
->device
, 1u << samples_log2
,
705 &cmd_buffer
->device
->meta_state
.clear
[samples_log2
].depthstencil_rp
);
706 if (ret
!= VK_SUCCESS
) {
707 cmd_buffer
->record_result
= ret
;
708 return VK_NULL_HANDLE
;
712 if (*pipeline
== VK_NULL_HANDLE
) {
713 VkResult ret
= create_depthstencil_pipeline(cmd_buffer
->device
, aspects
, 1u << samples_log2
, index
,
714 pipeline
, cmd_buffer
->device
->meta_state
.clear
[samples_log2
].depthstencil_rp
);
715 if (ret
!= VK_SUCCESS
) {
716 cmd_buffer
->record_result
= ret
;
717 return VK_NULL_HANDLE
;
724 emit_depthstencil_clear(struct radv_cmd_buffer
*cmd_buffer
,
725 const VkClearAttachment
*clear_att
,
726 const VkClearRect
*clear_rect
,
727 struct radv_subpass_attachment
*ds_att
,
730 struct radv_device
*device
= cmd_buffer
->device
;
731 struct radv_meta_state
*meta_state
= &device
->meta_state
;
732 const struct radv_subpass
*subpass
= cmd_buffer
->state
.subpass
;
733 const uint32_t pass_att
= ds_att
->attachment
;
734 VkClearDepthStencilValue clear_value
= clear_att
->clearValue
.depthStencil
;
735 VkImageAspectFlags aspects
= clear_att
->aspectMask
;
736 const struct radv_image_view
*iview
= cmd_buffer
->state
.attachments
?
737 cmd_buffer
->state
.attachments
[pass_att
].iview
: NULL
;
738 uint32_t samples
, samples_log2
;
739 VkCommandBuffer cmd_buffer_h
= radv_cmd_buffer_to_handle(cmd_buffer
);
741 /* When a framebuffer is bound to the current command buffer, get the
742 * number of samples from it. Otherwise, get the number of samples from
743 * the render pass because it's likely a secondary command buffer.
746 samples
= iview
->image
->info
.samples
;
748 samples
= cmd_buffer
->state
.pass
->attachments
[pass_att
].samples
;
751 samples_log2
= ffs(samples
) - 1;
753 assert(pass_att
!= VK_ATTACHMENT_UNUSED
);
755 if (!(aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
))
756 clear_value
.depth
= 1.0f
;
758 radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer
),
759 device
->meta_state
.clear_depth_p_layout
,
760 VK_SHADER_STAGE_FRAGMENT_BIT
, 0, 4,
763 uint32_t prev_reference
= cmd_buffer
->state
.dynamic
.stencil_reference
.front
;
764 if (aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
765 radv_CmdSetStencilReference(cmd_buffer_h
, VK_STENCIL_FACE_FRONT_BIT
,
766 clear_value
.stencil
);
769 VkPipeline pipeline
= pick_depthstencil_pipeline(cmd_buffer
,
775 ds_att
->in_render_loop
,
781 struct radv_subpass clear_subpass
= {
783 .color_attachments
= NULL
,
784 .depth_stencil_attachment
= ds_att
,
787 radv_cmd_buffer_set_subpass(cmd_buffer
, &clear_subpass
);
789 radv_CmdBindPipeline(cmd_buffer_h
, VK_PIPELINE_BIND_POINT_GRAPHICS
,
792 if (depth_view_can_fast_clear(cmd_buffer
, iview
, aspects
,
793 ds_att
->layout
, ds_att
->in_render_loop
,
794 clear_rect
, clear_value
))
795 radv_update_ds_clear_metadata(cmd_buffer
, iview
,
796 clear_value
, aspects
);
798 radv_CmdSetViewport(radv_cmd_buffer_to_handle(cmd_buffer
), 0, 1, &(VkViewport
) {
799 .x
= clear_rect
->rect
.offset
.x
,
800 .y
= clear_rect
->rect
.offset
.y
,
801 .width
= clear_rect
->rect
.extent
.width
,
802 .height
= clear_rect
->rect
.extent
.height
,
807 radv_CmdSetScissor(radv_cmd_buffer_to_handle(cmd_buffer
), 0, 1, &clear_rect
->rect
);
811 for_each_bit(i
, view_mask
)
812 radv_CmdDraw(cmd_buffer_h
, 3, 1, 0, i
);
814 radv_CmdDraw(cmd_buffer_h
, 3, clear_rect
->layerCount
, 0, clear_rect
->baseArrayLayer
);
817 if (aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
818 radv_CmdSetStencilReference(cmd_buffer_h
, VK_STENCIL_FACE_FRONT_BIT
,
822 radv_cmd_buffer_set_subpass(cmd_buffer
, subpass
);
826 clear_htile_mask(struct radv_cmd_buffer
*cmd_buffer
,
827 struct radeon_winsys_bo
*bo
, uint64_t offset
, uint64_t size
,
828 uint32_t htile_value
, uint32_t htile_mask
)
830 struct radv_device
*device
= cmd_buffer
->device
;
831 struct radv_meta_state
*state
= &device
->meta_state
;
832 uint64_t block_count
= round_up_u64(size
, 1024);
833 struct radv_meta_saved_state saved_state
;
835 radv_meta_save(&saved_state
, cmd_buffer
,
836 RADV_META_SAVE_COMPUTE_PIPELINE
|
837 RADV_META_SAVE_CONSTANTS
|
838 RADV_META_SAVE_DESCRIPTORS
);
840 struct radv_buffer dst_buffer
= {
846 radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer
),
847 VK_PIPELINE_BIND_POINT_COMPUTE
,
848 state
->clear_htile_mask_pipeline
);
850 radv_meta_push_descriptor_set(cmd_buffer
, VK_PIPELINE_BIND_POINT_COMPUTE
,
851 state
->clear_htile_mask_p_layout
,
853 1, /* descriptorWriteCount */
854 (VkWriteDescriptorSet
[]) {
856 .sType
= VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET
,
858 .dstArrayElement
= 0,
859 .descriptorCount
= 1,
860 .descriptorType
= VK_DESCRIPTOR_TYPE_STORAGE_BUFFER
,
861 .pBufferInfo
= &(VkDescriptorBufferInfo
) {
862 .buffer
= radv_buffer_to_handle(&dst_buffer
),
869 const unsigned constants
[2] = {
870 htile_value
& htile_mask
,
874 radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer
),
875 state
->clear_htile_mask_p_layout
,
876 VK_SHADER_STAGE_COMPUTE_BIT
, 0, 8,
879 radv_CmdDispatch(radv_cmd_buffer_to_handle(cmd_buffer
), block_count
, 1, 1);
881 radv_meta_restore(&saved_state
, cmd_buffer
);
883 return RADV_CMD_FLAG_CS_PARTIAL_FLUSH
|
884 RADV_CMD_FLAG_INV_VCACHE
|
889 radv_get_htile_fast_clear_value(const struct radv_image
*image
,
890 VkClearDepthStencilValue value
)
892 uint32_t clear_value
;
894 if (!image
->planes
[0].surface
.has_stencil
) {
895 clear_value
= value
.depth
? 0xfffffff0 : 0;
897 clear_value
= value
.depth
? 0xfffc0000 : 0;
904 radv_get_htile_mask(const struct radv_image
*image
, VkImageAspectFlags aspects
)
908 if (!image
->planes
[0].surface
.has_stencil
) {
909 /* All the HTILE buffer is used when there is no stencil. */
912 if (aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
)
914 if (aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
)
922 radv_is_fast_clear_depth_allowed(VkClearDepthStencilValue value
)
924 return value
.depth
== 1.0f
|| value
.depth
== 0.0f
;
928 radv_is_fast_clear_stencil_allowed(VkClearDepthStencilValue value
)
930 return value
.stencil
== 0;
934 * Determine if the given image can be fast cleared.
937 radv_image_can_fast_clear(struct radv_device
*device
, struct radv_image
*image
)
939 if (device
->instance
->debug_flags
& RADV_DEBUG_NO_FAST_CLEARS
)
942 if (vk_format_is_color(image
->vk_format
)) {
943 if (!radv_image_has_cmask(image
) && !radv_image_has_dcc(image
))
946 /* RB+ doesn't work with CMASK fast clear on Stoney. */
947 if (!radv_image_has_dcc(image
) &&
948 device
->physical_device
->rad_info
.family
== CHIP_STONEY
)
951 if (!radv_image_has_htile(image
))
955 /* Do not fast clears 3D images. */
956 if (image
->type
== VK_IMAGE_TYPE_3D
)
963 * Determine if the given image view can be fast cleared.
966 radv_image_view_can_fast_clear(struct radv_device
*device
,
967 const struct radv_image_view
*iview
)
969 struct radv_image
*image
;
973 image
= iview
->image
;
975 /* Only fast clear if the image itself can be fast cleared. */
976 if (!radv_image_can_fast_clear(device
, image
))
979 /* Only fast clear if all layers are bound. */
980 if (iview
->base_layer
> 0 ||
981 iview
->layer_count
!= image
->info
.array_size
)
984 /* Only fast clear if the view covers the whole image. */
985 if (!radv_image_extent_compare(image
, &iview
->extent
))
992 radv_can_fast_clear_depth(struct radv_cmd_buffer
*cmd_buffer
,
993 const struct radv_image_view
*iview
,
994 VkImageLayout image_layout
,
996 VkImageAspectFlags aspects
,
997 const VkClearRect
*clear_rect
,
998 const VkClearDepthStencilValue clear_value
,
1001 if (!radv_image_view_can_fast_clear(cmd_buffer
->device
, iview
))
1004 if (!radv_layout_is_htile_compressed(iview
->image
, image_layout
, in_render_loop
,
1005 radv_image_queue_family_mask(iview
->image
,
1006 cmd_buffer
->queue_family_index
,
1007 cmd_buffer
->queue_family_index
)))
1010 if (clear_rect
->rect
.offset
.x
|| clear_rect
->rect
.offset
.y
||
1011 clear_rect
->rect
.extent
.width
!= iview
->image
->info
.width
||
1012 clear_rect
->rect
.extent
.height
!= iview
->image
->info
.height
)
1015 if (view_mask
&& (iview
->image
->info
.array_size
>= 32 ||
1016 (1u << iview
->image
->info
.array_size
) - 1u != view_mask
))
1018 if (!view_mask
&& clear_rect
->baseArrayLayer
!= 0)
1020 if (!view_mask
&& clear_rect
->layerCount
!= iview
->image
->info
.array_size
)
1023 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
!= GFX9
&&
1024 (!(aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) ||
1025 ((vk_format_aspects(iview
->image
->vk_format
) & VK_IMAGE_ASPECT_STENCIL_BIT
) &&
1026 !(aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
))))
1029 if (((aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) &&
1030 !radv_is_fast_clear_depth_allowed(clear_value
)) ||
1031 ((aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) &&
1032 !radv_is_fast_clear_stencil_allowed(clear_value
)))
1039 radv_fast_clear_depth(struct radv_cmd_buffer
*cmd_buffer
,
1040 const struct radv_image_view
*iview
,
1041 const VkClearAttachment
*clear_att
,
1042 enum radv_cmd_flush_bits
*pre_flush
,
1043 enum radv_cmd_flush_bits
*post_flush
)
1045 VkClearDepthStencilValue clear_value
= clear_att
->clearValue
.depthStencil
;
1046 VkImageAspectFlags aspects
= clear_att
->aspectMask
;
1047 uint32_t clear_word
, flush_bits
;
1048 uint32_t htile_mask
;
1050 clear_word
= radv_get_htile_fast_clear_value(iview
->image
, clear_value
);
1051 htile_mask
= radv_get_htile_mask(iview
->image
, aspects
);
1054 cmd_buffer
->state
.flush_bits
|= (RADV_CMD_FLAG_FLUSH_AND_INV_DB
|
1055 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META
) & ~ *pre_flush
;
1056 *pre_flush
|= cmd_buffer
->state
.flush_bits
;
1059 if (htile_mask
== UINT_MAX
) {
1060 /* Clear the whole HTILE buffer. */
1061 flush_bits
= radv_fill_buffer(cmd_buffer
, iview
->image
->bo
,
1062 iview
->image
->offset
+ iview
->image
->htile_offset
,
1063 iview
->image
->planes
[0].surface
.htile_size
, clear_word
);
1065 /* Only clear depth or stencil bytes in the HTILE buffer. */
1066 /* TODO: Implement that path for GFX10. */
1067 assert(cmd_buffer
->device
->physical_device
->rad_info
.chip_class
== GFX9
);
1068 flush_bits
= clear_htile_mask(cmd_buffer
, iview
->image
->bo
,
1069 iview
->image
->offset
+ iview
->image
->htile_offset
,
1070 iview
->image
->planes
[0].surface
.htile_size
, clear_word
,
1074 radv_update_ds_clear_metadata(cmd_buffer
, iview
, clear_value
, aspects
);
1076 *post_flush
|= flush_bits
;
1081 build_clear_htile_mask_shader()
1085 nir_builder_init_simple_shader(&b
, NULL
, MESA_SHADER_COMPUTE
, NULL
);
1086 b
.shader
->info
.name
= ralloc_strdup(b
.shader
, "meta_clear_htile_mask");
1087 b
.shader
->info
.cs
.local_size
[0] = 64;
1088 b
.shader
->info
.cs
.local_size
[1] = 1;
1089 b
.shader
->info
.cs
.local_size
[2] = 1;
1091 nir_ssa_def
*invoc_id
= nir_load_local_invocation_id(&b
);
1092 nir_ssa_def
*wg_id
= nir_load_work_group_id(&b
);
1093 nir_ssa_def
*block_size
= nir_imm_ivec4(&b
,
1094 b
.shader
->info
.cs
.local_size
[0],
1095 b
.shader
->info
.cs
.local_size
[1],
1096 b
.shader
->info
.cs
.local_size
[2], 0);
1098 nir_ssa_def
*global_id
= nir_iadd(&b
, nir_imul(&b
, wg_id
, block_size
), invoc_id
);
1100 nir_ssa_def
*offset
= nir_imul(&b
, global_id
, nir_imm_int(&b
, 16));
1101 offset
= nir_channel(&b
, offset
, 0);
1103 nir_intrinsic_instr
*buf
=
1104 nir_intrinsic_instr_create(b
.shader
,
1105 nir_intrinsic_vulkan_resource_index
);
1107 buf
->src
[0] = nir_src_for_ssa(nir_imm_int(&b
, 0));
1108 buf
->num_components
= 1;
1109 nir_intrinsic_set_desc_set(buf
, 0);
1110 nir_intrinsic_set_binding(buf
, 0);
1111 nir_ssa_dest_init(&buf
->instr
, &buf
->dest
, buf
->num_components
, 32, NULL
);
1112 nir_builder_instr_insert(&b
, &buf
->instr
);
1114 nir_intrinsic_instr
*constants
=
1115 nir_intrinsic_instr_create(b
.shader
,
1116 nir_intrinsic_load_push_constant
);
1117 nir_intrinsic_set_base(constants
, 0);
1118 nir_intrinsic_set_range(constants
, 8);
1119 constants
->src
[0] = nir_src_for_ssa(nir_imm_int(&b
, 0));
1120 constants
->num_components
= 2;
1121 nir_ssa_dest_init(&constants
->instr
, &constants
->dest
, 2, 32, "constants");
1122 nir_builder_instr_insert(&b
, &constants
->instr
);
1124 nir_intrinsic_instr
*load
=
1125 nir_intrinsic_instr_create(b
.shader
, nir_intrinsic_load_ssbo
);
1126 load
->src
[0] = nir_src_for_ssa(&buf
->dest
.ssa
);
1127 load
->src
[1] = nir_src_for_ssa(offset
);
1128 nir_ssa_dest_init(&load
->instr
, &load
->dest
, 4, 32, NULL
);
1129 load
->num_components
= 4;
1130 nir_builder_instr_insert(&b
, &load
->instr
);
1132 /* data = (data & ~htile_mask) | (htile_value & htile_mask) */
1134 nir_iand(&b
, &load
->dest
.ssa
,
1135 nir_channel(&b
, &constants
->dest
.ssa
, 1));
1136 data
= nir_ior(&b
, data
, nir_channel(&b
, &constants
->dest
.ssa
, 0));
1138 nir_intrinsic_instr
*store
=
1139 nir_intrinsic_instr_create(b
.shader
, nir_intrinsic_store_ssbo
);
1140 store
->src
[0] = nir_src_for_ssa(data
);
1141 store
->src
[1] = nir_src_for_ssa(&buf
->dest
.ssa
);
1142 store
->src
[2] = nir_src_for_ssa(offset
);
1143 nir_intrinsic_set_write_mask(store
, 0xf);
1144 nir_intrinsic_set_access(store
, ACCESS_NON_READABLE
);
1145 store
->num_components
= 4;
1146 nir_builder_instr_insert(&b
, &store
->instr
);
1152 init_meta_clear_htile_mask_state(struct radv_device
*device
)
1154 struct radv_meta_state
*state
= &device
->meta_state
;
1155 struct radv_shader_module cs
= { .nir
= NULL
};
1158 cs
.nir
= build_clear_htile_mask_shader();
1160 VkDescriptorSetLayoutCreateInfo ds_layout_info
= {
1161 .sType
= VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO
,
1162 .flags
= VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR
,
1164 .pBindings
= (VkDescriptorSetLayoutBinding
[]) {
1167 .descriptorType
= VK_DESCRIPTOR_TYPE_STORAGE_BUFFER
,
1168 .descriptorCount
= 1,
1169 .stageFlags
= VK_SHADER_STAGE_COMPUTE_BIT
,
1170 .pImmutableSamplers
= NULL
1175 result
= radv_CreateDescriptorSetLayout(radv_device_to_handle(device
),
1176 &ds_layout_info
, &state
->alloc
,
1177 &state
->clear_htile_mask_ds_layout
);
1178 if (result
!= VK_SUCCESS
)
1181 VkPipelineLayoutCreateInfo p_layout_info
= {
1182 .sType
= VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO
,
1183 .setLayoutCount
= 1,
1184 .pSetLayouts
= &state
->clear_htile_mask_ds_layout
,
1185 .pushConstantRangeCount
= 1,
1186 .pPushConstantRanges
= &(VkPushConstantRange
){
1187 VK_SHADER_STAGE_COMPUTE_BIT
, 0, 8,
1191 result
= radv_CreatePipelineLayout(radv_device_to_handle(device
),
1192 &p_layout_info
, &state
->alloc
,
1193 &state
->clear_htile_mask_p_layout
);
1194 if (result
!= VK_SUCCESS
)
1197 VkPipelineShaderStageCreateInfo shader_stage
= {
1198 .sType
= VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO
,
1199 .stage
= VK_SHADER_STAGE_COMPUTE_BIT
,
1200 .module
= radv_shader_module_to_handle(&cs
),
1202 .pSpecializationInfo
= NULL
,
1205 VkComputePipelineCreateInfo pipeline_info
= {
1206 .sType
= VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO
,
1207 .stage
= shader_stage
,
1209 .layout
= state
->clear_htile_mask_p_layout
,
1212 result
= radv_CreateComputePipelines(radv_device_to_handle(device
),
1213 radv_pipeline_cache_to_handle(&state
->cache
),
1214 1, &pipeline_info
, NULL
,
1215 &state
->clear_htile_mask_pipeline
);
1217 ralloc_free(cs
.nir
);
1220 ralloc_free(cs
.nir
);
1225 radv_device_init_meta_clear_state(struct radv_device
*device
, bool on_demand
)
1228 struct radv_meta_state
*state
= &device
->meta_state
;
1230 VkPipelineLayoutCreateInfo pl_color_create_info
= {
1231 .sType
= VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO
,
1232 .setLayoutCount
= 0,
1233 .pushConstantRangeCount
= 1,
1234 .pPushConstantRanges
= &(VkPushConstantRange
){VK_SHADER_STAGE_FRAGMENT_BIT
, 0, 16},
1237 res
= radv_CreatePipelineLayout(radv_device_to_handle(device
),
1238 &pl_color_create_info
,
1239 &device
->meta_state
.alloc
,
1240 &device
->meta_state
.clear_color_p_layout
);
1241 if (res
!= VK_SUCCESS
)
1244 VkPipelineLayoutCreateInfo pl_depth_create_info
= {
1245 .sType
= VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO
,
1246 .setLayoutCount
= 0,
1247 .pushConstantRangeCount
= 1,
1248 .pPushConstantRanges
= &(VkPushConstantRange
){VK_SHADER_STAGE_FRAGMENT_BIT
, 0, 4},
1251 res
= radv_CreatePipelineLayout(radv_device_to_handle(device
),
1252 &pl_depth_create_info
,
1253 &device
->meta_state
.alloc
,
1254 &device
->meta_state
.clear_depth_p_layout
);
1255 if (res
!= VK_SUCCESS
)
1258 res
= init_meta_clear_htile_mask_state(device
);
1259 if (res
!= VK_SUCCESS
)
1265 for (uint32_t i
= 0; i
< ARRAY_SIZE(state
->clear
); ++i
) {
1266 uint32_t samples
= 1 << i
;
1267 for (uint32_t j
= 0; j
< NUM_META_FS_KEYS
; ++j
) {
1268 VkFormat format
= radv_fs_key_format_exemplars
[j
];
1269 unsigned fs_key
= radv_format_meta_fs_key(format
);
1270 assert(!state
->clear
[i
].color_pipelines
[fs_key
]);
1272 res
= create_color_renderpass(device
, format
, samples
,
1273 &state
->clear
[i
].render_pass
[fs_key
]);
1274 if (res
!= VK_SUCCESS
)
1277 res
= create_color_pipeline(device
, samples
, 0, &state
->clear
[i
].color_pipelines
[fs_key
],
1278 state
->clear
[i
].render_pass
[fs_key
]);
1279 if (res
!= VK_SUCCESS
)
1284 res
= create_depthstencil_renderpass(device
,
1286 &state
->clear
[i
].depthstencil_rp
);
1287 if (res
!= VK_SUCCESS
)
1290 for (uint32_t j
= 0; j
< NUM_DEPTH_CLEAR_PIPELINES
; j
++) {
1291 res
= create_depthstencil_pipeline(device
,
1292 VK_IMAGE_ASPECT_DEPTH_BIT
,
1295 &state
->clear
[i
].depth_only_pipeline
[j
],
1296 state
->clear
[i
].depthstencil_rp
);
1297 if (res
!= VK_SUCCESS
)
1300 res
= create_depthstencil_pipeline(device
,
1301 VK_IMAGE_ASPECT_STENCIL_BIT
,
1304 &state
->clear
[i
].stencil_only_pipeline
[j
],
1305 state
->clear
[i
].depthstencil_rp
);
1306 if (res
!= VK_SUCCESS
)
1309 res
= create_depthstencil_pipeline(device
,
1310 VK_IMAGE_ASPECT_DEPTH_BIT
|
1311 VK_IMAGE_ASPECT_STENCIL_BIT
,
1314 &state
->clear
[i
].depthstencil_pipeline
[j
],
1315 state
->clear
[i
].depthstencil_rp
);
1316 if (res
!= VK_SUCCESS
)
1323 radv_device_finish_meta_clear_state(device
);
1328 radv_get_cmask_fast_clear_value(const struct radv_image
*image
)
1330 uint32_t value
= 0; /* Default value when no DCC. */
1332 /* The fast-clear value is different for images that have both DCC and
1335 if (radv_image_has_dcc(image
)) {
1336 /* DCC fast clear with MSAA should clear CMASK to 0xC. */
1337 return image
->info
.samples
> 1 ? 0xcccccccc : 0xffffffff;
1344 radv_clear_cmask(struct radv_cmd_buffer
*cmd_buffer
,
1345 struct radv_image
*image
,
1346 const VkImageSubresourceRange
*range
, uint32_t value
)
1348 uint64_t offset
= image
->offset
+ image
->cmask_offset
;
1351 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1352 /* TODO: clear layers. */
1353 size
= image
->planes
[0].surface
.cmask_size
;
1355 unsigned cmask_slice_size
=
1356 image
->planes
[0].surface
.cmask_slice_size
;
1358 offset
+= cmask_slice_size
* range
->baseArrayLayer
;
1359 size
= cmask_slice_size
* radv_get_layerCount(image
, range
);
1362 return radv_fill_buffer(cmd_buffer
, image
->bo
, offset
, size
, value
);
1367 radv_clear_fmask(struct radv_cmd_buffer
*cmd_buffer
,
1368 struct radv_image
*image
,
1369 const VkImageSubresourceRange
*range
, uint32_t value
)
1371 uint64_t offset
= image
->offset
+ image
->fmask_offset
;
1374 /* MSAA images do not support mipmap levels. */
1375 assert(range
->baseMipLevel
== 0 &&
1376 radv_get_levelCount(image
, range
) == 1);
1378 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1379 /* TODO: clear layers. */
1380 size
= image
->planes
[0].surface
.fmask_size
;
1382 unsigned fmask_slice_size
=
1383 image
->planes
[0].surface
.u
.legacy
.fmask
.slice_size
;
1386 offset
+= fmask_slice_size
* range
->baseArrayLayer
;
1387 size
= fmask_slice_size
* radv_get_layerCount(image
, range
);
1390 return radv_fill_buffer(cmd_buffer
, image
->bo
, offset
, size
, value
);
1394 radv_clear_dcc(struct radv_cmd_buffer
*cmd_buffer
,
1395 struct radv_image
*image
,
1396 const VkImageSubresourceRange
*range
, uint32_t value
)
1398 uint32_t level_count
= radv_get_levelCount(image
, range
);
1399 uint32_t flush_bits
= 0;
1401 /* Mark the image as being compressed. */
1402 radv_update_dcc_metadata(cmd_buffer
, image
, range
, true);
1404 for (uint32_t l
= 0; l
< level_count
; l
++) {
1405 uint64_t offset
= image
->offset
+ image
->dcc_offset
;
1406 uint32_t level
= range
->baseMipLevel
+ l
;
1409 if (cmd_buffer
->device
->physical_device
->rad_info
.chip_class
>= GFX9
) {
1410 /* Mipmap levels aren't implemented. */
1412 size
= image
->planes
[0].surface
.dcc_size
;
1414 const struct legacy_surf_level
*surf_level
=
1415 &image
->planes
[0].surface
.u
.legacy
.level
[level
];
1417 /* If dcc_fast_clear_size is 0 (which might happens for
1418 * mipmaps) the fill buffer operation below is a no-op.
1419 * This can only happen during initialization as the
1420 * fast clear path fallbacks to slow clears if one
1421 * level can't be fast cleared.
1423 offset
+= surf_level
->dcc_offset
+
1424 surf_level
->dcc_slice_fast_clear_size
* range
->baseArrayLayer
;
1425 size
= surf_level
->dcc_slice_fast_clear_size
* radv_get_layerCount(image
, range
);
1428 flush_bits
|= radv_fill_buffer(cmd_buffer
, image
->bo
, offset
,
1436 radv_clear_htile(struct radv_cmd_buffer
*cmd_buffer
, struct radv_image
*image
,
1437 const VkImageSubresourceRange
*range
, uint32_t value
)
1439 unsigned layer_count
= radv_get_layerCount(image
, range
);
1440 uint64_t size
= image
->planes
[0].surface
.htile_slice_size
* layer_count
;
1441 uint64_t offset
= image
->offset
+ image
->htile_offset
+
1442 image
->planes
[0].surface
.htile_slice_size
* range
->baseArrayLayer
;
1444 return radv_fill_buffer(cmd_buffer
, image
->bo
, offset
, size
, value
);
1448 RADV_DCC_CLEAR_REG
= 0x20202020U
,
1449 RADV_DCC_CLEAR_MAIN_1
= 0x80808080U
,
1450 RADV_DCC_CLEAR_SECONDARY_1
= 0x40404040U
1453 static void vi_get_fast_clear_parameters(struct radv_device
*device
,
1454 VkFormat image_format
,
1455 VkFormat view_format
,
1456 const VkClearColorValue
*clear_value
,
1457 uint32_t* reset_value
,
1458 bool *can_avoid_fast_clear_elim
)
1460 bool values
[4] = {};
1462 bool main_value
= false;
1463 bool extra_value
= false;
1465 *can_avoid_fast_clear_elim
= false;
1467 *reset_value
= RADV_DCC_CLEAR_REG
;
1469 const struct vk_format_description
*desc
= vk_format_description(view_format
);
1470 if (view_format
== VK_FORMAT_B10G11R11_UFLOAT_PACK32
||
1471 view_format
== VK_FORMAT_R5G6B5_UNORM_PACK16
||
1472 view_format
== VK_FORMAT_B5G6R5_UNORM_PACK16
)
1474 else if (desc
->layout
== VK_FORMAT_LAYOUT_PLAIN
) {
1475 if (vi_alpha_is_on_msb(device
, view_format
))
1476 extra_channel
= desc
->nr_channels
- 1;
1482 bool image_alpha_is_on_msb
= vi_alpha_is_on_msb(device
, image_format
);
1483 bool view_alpha_is_on_msb
= vi_alpha_is_on_msb(device
, view_format
);
1485 for (i
= 0; i
< 4; i
++) {
1486 int index
= desc
->swizzle
[i
] - VK_SWIZZLE_X
;
1487 if (desc
->swizzle
[i
] < VK_SWIZZLE_X
||
1488 desc
->swizzle
[i
] > VK_SWIZZLE_W
)
1491 if (desc
->channel
[i
].pure_integer
&&
1492 desc
->channel
[i
].type
== VK_FORMAT_TYPE_SIGNED
) {
1493 /* Use the maximum value for clamping the clear color. */
1494 int max
= u_bit_consecutive(0, desc
->channel
[i
].size
- 1);
1496 values
[i
] = clear_value
->int32
[i
] != 0;
1497 if (clear_value
->int32
[i
] != 0 && MIN2(clear_value
->int32
[i
], max
) != max
)
1499 } else if (desc
->channel
[i
].pure_integer
&&
1500 desc
->channel
[i
].type
== VK_FORMAT_TYPE_UNSIGNED
) {
1501 /* Use the maximum value for clamping the clear color. */
1502 unsigned max
= u_bit_consecutive(0, desc
->channel
[i
].size
);
1504 values
[i
] = clear_value
->uint32
[i
] != 0U;
1505 if (clear_value
->uint32
[i
] != 0U && MIN2(clear_value
->uint32
[i
], max
) != max
)
1508 values
[i
] = clear_value
->float32
[i
] != 0.0F
;
1509 if (clear_value
->float32
[i
] != 0.0F
&& clear_value
->float32
[i
] != 1.0F
)
1513 if (index
== extra_channel
)
1514 extra_value
= values
[i
];
1516 main_value
= values
[i
];
1519 /* If alpha isn't present, make it the same as color, and vice versa. */
1521 extra_value
= main_value
;
1522 else if (!main_value
)
1523 main_value
= extra_value
;
1525 if (main_value
!= extra_value
) {
1526 assert(image_alpha_is_on_msb
== view_alpha_is_on_msb
);
1527 return; /* require ELIMINATE_FAST_CLEAR */
1530 for (int i
= 0; i
< 4; ++i
)
1531 if (values
[i
] != main_value
&&
1532 desc
->swizzle
[i
] - VK_SWIZZLE_X
!= extra_channel
&&
1533 desc
->swizzle
[i
] >= VK_SWIZZLE_X
&&
1534 desc
->swizzle
[i
] <= VK_SWIZZLE_W
)
1537 *can_avoid_fast_clear_elim
= true;
1540 *reset_value
|= RADV_DCC_CLEAR_MAIN_1
;
1543 *reset_value
|= RADV_DCC_CLEAR_SECONDARY_1
;
1548 radv_can_fast_clear_color(struct radv_cmd_buffer
*cmd_buffer
,
1549 const struct radv_image_view
*iview
,
1550 VkImageLayout image_layout
,
1551 bool in_render_loop
,
1552 const VkClearRect
*clear_rect
,
1553 VkClearColorValue clear_value
,
1556 uint32_t clear_color
[2];
1558 if (!radv_image_view_can_fast_clear(cmd_buffer
->device
, iview
))
1561 if (!radv_layout_can_fast_clear(iview
->image
, image_layout
, in_render_loop
,
1562 radv_image_queue_family_mask(iview
->image
,
1563 cmd_buffer
->queue_family_index
,
1564 cmd_buffer
->queue_family_index
)))
1567 if (clear_rect
->rect
.offset
.x
|| clear_rect
->rect
.offset
.y
||
1568 clear_rect
->rect
.extent
.width
!= iview
->image
->info
.width
||
1569 clear_rect
->rect
.extent
.height
!= iview
->image
->info
.height
)
1572 if (view_mask
&& (iview
->image
->info
.array_size
>= 32 ||
1573 (1u << iview
->image
->info
.array_size
) - 1u != view_mask
))
1575 if (!view_mask
&& clear_rect
->baseArrayLayer
!= 0)
1577 if (!view_mask
&& clear_rect
->layerCount
!= iview
->image
->info
.array_size
)
1581 if (!radv_format_pack_clear_color(iview
->vk_format
,
1582 clear_color
, &clear_value
))
1585 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
)) {
1586 bool can_avoid_fast_clear_elim
;
1587 uint32_t reset_value
;
1589 vi_get_fast_clear_parameters(cmd_buffer
->device
,
1590 iview
->image
->vk_format
,
1592 &clear_value
, &reset_value
,
1593 &can_avoid_fast_clear_elim
);
1595 if (iview
->image
->info
.samples
> 1) {
1596 /* DCC fast clear with MSAA should clear CMASK. */
1597 /* FIXME: This doesn't work for now. There is a
1598 * hardware bug with fast clears and DCC for MSAA
1599 * textures. AMDVLK has a workaround but it doesn't
1600 * seem to work here. Note that we might emit useless
1601 * CB flushes but that shouldn't matter.
1603 if (!can_avoid_fast_clear_elim
)
1607 if (iview
->image
->info
.levels
> 1 &&
1608 cmd_buffer
->device
->physical_device
->rad_info
.chip_class
== GFX8
) {
1609 for (uint32_t l
= 0; l
< iview
->level_count
; l
++) {
1610 uint32_t level
= iview
->base_mip
+ l
;
1611 struct legacy_surf_level
*surf_level
=
1612 &iview
->image
->planes
[0].surface
.u
.legacy
.level
[level
];
1614 /* Do not fast clears if one level can't be
1617 if (!surf_level
->dcc_fast_clear_size
)
1628 radv_fast_clear_color(struct radv_cmd_buffer
*cmd_buffer
,
1629 const struct radv_image_view
*iview
,
1630 const VkClearAttachment
*clear_att
,
1631 uint32_t subpass_att
,
1632 enum radv_cmd_flush_bits
*pre_flush
,
1633 enum radv_cmd_flush_bits
*post_flush
)
1635 VkClearColorValue clear_value
= clear_att
->clearValue
.color
;
1636 uint32_t clear_color
[2], flush_bits
= 0;
1637 uint32_t cmask_clear_value
;
1638 VkImageSubresourceRange range
= {
1639 .aspectMask
= iview
->aspect_mask
,
1640 .baseMipLevel
= iview
->base_mip
,
1641 .levelCount
= iview
->level_count
,
1642 .baseArrayLayer
= iview
->base_layer
,
1643 .layerCount
= iview
->layer_count
,
1647 cmd_buffer
->state
.flush_bits
|= (RADV_CMD_FLAG_FLUSH_AND_INV_CB
|
1648 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META
) & ~ *pre_flush
;
1649 *pre_flush
|= cmd_buffer
->state
.flush_bits
;
1653 radv_format_pack_clear_color(iview
->vk_format
, clear_color
, &clear_value
);
1655 cmask_clear_value
= radv_get_cmask_fast_clear_value(iview
->image
);
1657 /* clear cmask buffer */
1658 if (radv_dcc_enabled(iview
->image
, iview
->base_mip
)) {
1659 uint32_t reset_value
;
1660 bool can_avoid_fast_clear_elim
;
1661 bool need_decompress_pass
= false;
1663 vi_get_fast_clear_parameters(cmd_buffer
->device
,
1664 iview
->image
->vk_format
,
1666 &clear_value
, &reset_value
,
1667 &can_avoid_fast_clear_elim
);
1669 if (radv_image_has_cmask(iview
->image
)) {
1670 flush_bits
= radv_clear_cmask(cmd_buffer
, iview
->image
,
1671 &range
, cmask_clear_value
);
1673 need_decompress_pass
= true;
1676 if (!can_avoid_fast_clear_elim
)
1677 need_decompress_pass
= true;
1679 flush_bits
|= radv_clear_dcc(cmd_buffer
, iview
->image
, &range
,
1682 radv_update_fce_metadata(cmd_buffer
, iview
->image
, &range
,
1683 need_decompress_pass
);
1685 flush_bits
= radv_clear_cmask(cmd_buffer
, iview
->image
,
1686 &range
, cmask_clear_value
);
1690 *post_flush
|= flush_bits
;
1693 radv_update_color_clear_metadata(cmd_buffer
, iview
, subpass_att
,
1698 * The parameters mean that same as those in vkCmdClearAttachments.
1701 emit_clear(struct radv_cmd_buffer
*cmd_buffer
,
1702 const VkClearAttachment
*clear_att
,
1703 const VkClearRect
*clear_rect
,
1704 enum radv_cmd_flush_bits
*pre_flush
,
1705 enum radv_cmd_flush_bits
*post_flush
,
1707 bool ds_resolve_clear
)
1709 const struct radv_framebuffer
*fb
= cmd_buffer
->state
.framebuffer
;
1710 const struct radv_subpass
*subpass
= cmd_buffer
->state
.subpass
;
1711 VkImageAspectFlags aspects
= clear_att
->aspectMask
;
1713 if (aspects
& VK_IMAGE_ASPECT_COLOR_BIT
) {
1714 const uint32_t subpass_att
= clear_att
->colorAttachment
;
1715 assert(subpass_att
< subpass
->color_count
);
1716 const uint32_t pass_att
= subpass
->color_attachments
[subpass_att
].attachment
;
1717 if (pass_att
== VK_ATTACHMENT_UNUSED
)
1720 VkImageLayout image_layout
= subpass
->color_attachments
[subpass_att
].layout
;
1721 bool in_render_loop
= subpass
->color_attachments
[subpass_att
].in_render_loop
;
1722 const struct radv_image_view
*iview
= fb
? cmd_buffer
->state
.attachments
[pass_att
].iview
: NULL
;
1723 VkClearColorValue clear_value
= clear_att
->clearValue
.color
;
1725 if (radv_can_fast_clear_color(cmd_buffer
, iview
, image_layout
, in_render_loop
,
1726 clear_rect
, clear_value
, view_mask
)) {
1727 radv_fast_clear_color(cmd_buffer
, iview
, clear_att
,
1728 subpass_att
, pre_flush
,
1731 emit_color_clear(cmd_buffer
, clear_att
, clear_rect
, view_mask
);
1734 struct radv_subpass_attachment
*ds_att
= subpass
->depth_stencil_attachment
;
1736 if (ds_resolve_clear
)
1737 ds_att
= subpass
->ds_resolve_attachment
;
1739 if (!ds_att
|| ds_att
->attachment
== VK_ATTACHMENT_UNUSED
)
1742 VkImageLayout image_layout
= ds_att
->layout
;
1743 bool in_render_loop
= ds_att
->in_render_loop
;
1744 const struct radv_image_view
*iview
= fb
? cmd_buffer
->state
.attachments
[ds_att
->attachment
].iview
: NULL
;
1745 VkClearDepthStencilValue clear_value
= clear_att
->clearValue
.depthStencil
;
1747 assert(aspects
& (VK_IMAGE_ASPECT_DEPTH_BIT
|
1748 VK_IMAGE_ASPECT_STENCIL_BIT
));
1750 if (radv_can_fast_clear_depth(cmd_buffer
, iview
, image_layout
,
1751 in_render_loop
, aspects
, clear_rect
,
1752 clear_value
, view_mask
)) {
1753 radv_fast_clear_depth(cmd_buffer
, iview
, clear_att
,
1754 pre_flush
, post_flush
);
1756 emit_depthstencil_clear(cmd_buffer
, clear_att
, clear_rect
,
1763 radv_attachment_needs_clear(struct radv_cmd_state
*cmd_state
, uint32_t a
)
1765 uint32_t view_mask
= cmd_state
->subpass
->view_mask
;
1766 return (a
!= VK_ATTACHMENT_UNUSED
&&
1767 cmd_state
->attachments
[a
].pending_clear_aspects
&&
1768 (!view_mask
|| (view_mask
& ~cmd_state
->attachments
[a
].cleared_views
)));
1772 radv_subpass_needs_clear(struct radv_cmd_buffer
*cmd_buffer
)
1774 struct radv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
1777 if (!cmd_state
->subpass
)
1780 for (uint32_t i
= 0; i
< cmd_state
->subpass
->color_count
; ++i
) {
1781 a
= cmd_state
->subpass
->color_attachments
[i
].attachment
;
1782 if (radv_attachment_needs_clear(cmd_state
, a
))
1786 if (cmd_state
->subpass
->depth_stencil_attachment
) {
1787 a
= cmd_state
->subpass
->depth_stencil_attachment
->attachment
;
1788 if (radv_attachment_needs_clear(cmd_state
, a
))
1792 if (!cmd_state
->subpass
->ds_resolve_attachment
)
1795 a
= cmd_state
->subpass
->ds_resolve_attachment
->attachment
;
1796 return radv_attachment_needs_clear(cmd_state
, a
);
1800 radv_subpass_clear_attachment(struct radv_cmd_buffer
*cmd_buffer
,
1801 struct radv_attachment_state
*attachment
,
1802 const VkClearAttachment
*clear_att
,
1803 enum radv_cmd_flush_bits
*pre_flush
,
1804 enum radv_cmd_flush_bits
*post_flush
,
1805 bool ds_resolve_clear
)
1807 struct radv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
1808 uint32_t view_mask
= cmd_state
->subpass
->view_mask
;
1810 VkClearRect clear_rect
= {
1811 .rect
= cmd_state
->render_area
,
1812 .baseArrayLayer
= 0,
1813 .layerCount
= cmd_state
->framebuffer
->layers
,
1816 emit_clear(cmd_buffer
, clear_att
, &clear_rect
, pre_flush
, post_flush
,
1817 view_mask
& ~attachment
->cleared_views
, ds_resolve_clear
);
1819 attachment
->cleared_views
|= view_mask
;
1821 attachment
->pending_clear_aspects
= 0;
1825 * Emit any pending attachment clears for the current subpass.
1827 * @see radv_attachment_state::pending_clear_aspects
1830 radv_cmd_buffer_clear_subpass(struct radv_cmd_buffer
*cmd_buffer
)
1832 struct radv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
1833 struct radv_meta_saved_state saved_state
;
1834 enum radv_cmd_flush_bits pre_flush
= 0;
1835 enum radv_cmd_flush_bits post_flush
= 0;
1837 if (!radv_subpass_needs_clear(cmd_buffer
))
1840 radv_meta_save(&saved_state
, cmd_buffer
,
1841 RADV_META_SAVE_GRAPHICS_PIPELINE
|
1842 RADV_META_SAVE_CONSTANTS
);
1844 for (uint32_t i
= 0; i
< cmd_state
->subpass
->color_count
; ++i
) {
1845 uint32_t a
= cmd_state
->subpass
->color_attachments
[i
].attachment
;
1847 if (!radv_attachment_needs_clear(cmd_state
, a
))
1850 assert(cmd_state
->attachments
[a
].pending_clear_aspects
==
1851 VK_IMAGE_ASPECT_COLOR_BIT
);
1853 VkClearAttachment clear_att
= {
1854 .aspectMask
= VK_IMAGE_ASPECT_COLOR_BIT
,
1855 .colorAttachment
= i
, /* Use attachment index relative to subpass */
1856 .clearValue
= cmd_state
->attachments
[a
].clear_value
,
1859 radv_subpass_clear_attachment(cmd_buffer
,
1860 &cmd_state
->attachments
[a
],
1861 &clear_att
, &pre_flush
,
1862 &post_flush
, false);
1865 if (cmd_state
->subpass
->depth_stencil_attachment
) {
1866 uint32_t ds
= cmd_state
->subpass
->depth_stencil_attachment
->attachment
;
1867 if (radv_attachment_needs_clear(cmd_state
, ds
)) {
1868 VkClearAttachment clear_att
= {
1869 .aspectMask
= cmd_state
->attachments
[ds
].pending_clear_aspects
,
1870 .clearValue
= cmd_state
->attachments
[ds
].clear_value
,
1873 radv_subpass_clear_attachment(cmd_buffer
,
1874 &cmd_state
->attachments
[ds
],
1875 &clear_att
, &pre_flush
,
1876 &post_flush
, false);
1880 if (cmd_state
->subpass
->ds_resolve_attachment
) {
1881 uint32_t ds_resolve
= cmd_state
->subpass
->ds_resolve_attachment
->attachment
;
1882 if (radv_attachment_needs_clear(cmd_state
, ds_resolve
)) {
1883 VkClearAttachment clear_att
= {
1884 .aspectMask
= cmd_state
->attachments
[ds_resolve
].pending_clear_aspects
,
1885 .clearValue
= cmd_state
->attachments
[ds_resolve
].clear_value
,
1888 radv_subpass_clear_attachment(cmd_buffer
,
1889 &cmd_state
->attachments
[ds_resolve
],
1890 &clear_att
, &pre_flush
,
1895 radv_meta_restore(&saved_state
, cmd_buffer
);
1896 cmd_buffer
->state
.flush_bits
|= post_flush
;
1900 radv_clear_image_layer(struct radv_cmd_buffer
*cmd_buffer
,
1901 struct radv_image
*image
,
1902 VkImageLayout image_layout
,
1903 const VkImageSubresourceRange
*range
,
1904 VkFormat format
, int level
, int layer
,
1905 const VkClearValue
*clear_val
)
1907 VkDevice device_h
= radv_device_to_handle(cmd_buffer
->device
);
1908 struct radv_image_view iview
;
1909 uint32_t width
= radv_minify(image
->info
.width
, range
->baseMipLevel
+ level
);
1910 uint32_t height
= radv_minify(image
->info
.height
, range
->baseMipLevel
+ level
);
1912 radv_image_view_init(&iview
, cmd_buffer
->device
,
1913 &(VkImageViewCreateInfo
) {
1914 .sType
= VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO
,
1915 .image
= radv_image_to_handle(image
),
1916 .viewType
= radv_meta_get_view_type(image
),
1918 .subresourceRange
= {
1919 .aspectMask
= range
->aspectMask
,
1920 .baseMipLevel
= range
->baseMipLevel
+ level
,
1922 .baseArrayLayer
= range
->baseArrayLayer
+ layer
,
1928 radv_CreateFramebuffer(device_h
,
1929 &(VkFramebufferCreateInfo
) {
1930 .sType
= VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO
,
1931 .attachmentCount
= 1,
1932 .pAttachments
= (VkImageView
[]) {
1933 radv_image_view_to_handle(&iview
),
1939 &cmd_buffer
->pool
->alloc
,
1942 VkAttachmentDescription att_desc
= {
1943 .format
= iview
.vk_format
,
1944 .loadOp
= VK_ATTACHMENT_LOAD_OP_LOAD
,
1945 .storeOp
= VK_ATTACHMENT_STORE_OP_STORE
,
1946 .stencilLoadOp
= VK_ATTACHMENT_LOAD_OP_LOAD
,
1947 .stencilStoreOp
= VK_ATTACHMENT_STORE_OP_STORE
,
1948 .initialLayout
= image_layout
,
1949 .finalLayout
= image_layout
,
1952 VkSubpassDescription subpass_desc
= {
1953 .pipelineBindPoint
= VK_PIPELINE_BIND_POINT_GRAPHICS
,
1954 .inputAttachmentCount
= 0,
1955 .colorAttachmentCount
= 0,
1956 .pColorAttachments
= NULL
,
1957 .pResolveAttachments
= NULL
,
1958 .pDepthStencilAttachment
= NULL
,
1959 .preserveAttachmentCount
= 0,
1960 .pPreserveAttachments
= NULL
,
1963 const VkAttachmentReference att_ref
= {
1965 .layout
= image_layout
,
1968 if (range
->aspectMask
& VK_IMAGE_ASPECT_COLOR_BIT
) {
1969 subpass_desc
.colorAttachmentCount
= 1;
1970 subpass_desc
.pColorAttachments
= &att_ref
;
1972 subpass_desc
.pDepthStencilAttachment
= &att_ref
;
1976 radv_CreateRenderPass(device_h
,
1977 &(VkRenderPassCreateInfo
) {
1978 .sType
= VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO
,
1979 .attachmentCount
= 1,
1980 .pAttachments
= &att_desc
,
1982 .pSubpasses
= &subpass_desc
,
1984 &cmd_buffer
->pool
->alloc
,
1987 radv_CmdBeginRenderPass(radv_cmd_buffer_to_handle(cmd_buffer
),
1988 &(VkRenderPassBeginInfo
) {
1989 .sType
= VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO
,
1991 .offset
= { 0, 0, },
1999 .clearValueCount
= 0,
2000 .pClearValues
= NULL
,
2002 VK_SUBPASS_CONTENTS_INLINE
);
2004 VkClearAttachment clear_att
= {
2005 .aspectMask
= range
->aspectMask
,
2006 .colorAttachment
= 0,
2007 .clearValue
= *clear_val
,
2010 VkClearRect clear_rect
= {
2013 .extent
= { width
, height
},
2015 .baseArrayLayer
= range
->baseArrayLayer
,
2016 .layerCount
= 1, /* FINISHME: clear multi-layer framebuffer */
2019 emit_clear(cmd_buffer
, &clear_att
, &clear_rect
, NULL
, NULL
, 0, false);
2021 radv_CmdEndRenderPass(radv_cmd_buffer_to_handle(cmd_buffer
));
2022 radv_DestroyRenderPass(device_h
, pass
,
2023 &cmd_buffer
->pool
->alloc
);
2024 radv_DestroyFramebuffer(device_h
, fb
,
2025 &cmd_buffer
->pool
->alloc
);
2029 * Return TRUE if a fast color or depth clear has been performed.
2032 radv_fast_clear_range(struct radv_cmd_buffer
*cmd_buffer
,
2033 struct radv_image
*image
,
2035 VkImageLayout image_layout
,
2036 bool in_render_loop
,
2037 const VkImageSubresourceRange
*range
,
2038 const VkClearValue
*clear_val
)
2040 struct radv_image_view iview
;
2042 radv_image_view_init(&iview
, cmd_buffer
->device
,
2043 &(VkImageViewCreateInfo
) {
2044 .sType
= VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO
,
2045 .image
= radv_image_to_handle(image
),
2046 .viewType
= radv_meta_get_view_type(image
),
2047 .format
= image
->vk_format
,
2048 .subresourceRange
= {
2049 .aspectMask
= range
->aspectMask
,
2050 .baseMipLevel
= range
->baseMipLevel
,
2051 .levelCount
= range
->levelCount
,
2052 .baseArrayLayer
= range
->baseArrayLayer
,
2053 .layerCount
= range
->layerCount
,
2057 VkClearRect clear_rect
= {
2061 radv_minify(image
->info
.width
, range
->baseMipLevel
),
2062 radv_minify(image
->info
.height
, range
->baseMipLevel
),
2065 .baseArrayLayer
= range
->baseArrayLayer
,
2066 .layerCount
= range
->layerCount
,
2069 VkClearAttachment clear_att
= {
2070 .aspectMask
= range
->aspectMask
,
2071 .colorAttachment
= 0,
2072 .clearValue
= *clear_val
,
2075 if (vk_format_is_color(format
)) {
2076 if (radv_can_fast_clear_color(cmd_buffer
, &iview
, image_layout
,
2077 in_render_loop
, &clear_rect
,
2078 clear_att
.clearValue
.color
, 0)) {
2079 radv_fast_clear_color(cmd_buffer
, &iview
, &clear_att
,
2080 clear_att
.colorAttachment
,
2085 if (radv_can_fast_clear_depth(cmd_buffer
, &iview
, image_layout
,
2086 in_render_loop
,range
->aspectMask
,
2087 &clear_rect
, clear_att
.clearValue
.depthStencil
,
2089 radv_fast_clear_depth(cmd_buffer
, &iview
, &clear_att
,
2099 radv_cmd_clear_image(struct radv_cmd_buffer
*cmd_buffer
,
2100 struct radv_image
*image
,
2101 VkImageLayout image_layout
,
2102 const VkClearValue
*clear_value
,
2103 uint32_t range_count
,
2104 const VkImageSubresourceRange
*ranges
,
2107 VkFormat format
= image
->vk_format
;
2108 VkClearValue internal_clear_value
= *clear_value
;
2110 if (format
== VK_FORMAT_E5B9G9R9_UFLOAT_PACK32
) {
2112 format
= VK_FORMAT_R32_UINT
;
2113 value
= float3_to_rgb9e5(clear_value
->color
.float32
);
2114 internal_clear_value
.color
.uint32
[0] = value
;
2117 if (format
== VK_FORMAT_R4G4_UNORM_PACK8
) {
2119 format
= VK_FORMAT_R8_UINT
;
2120 r
= float_to_ubyte(clear_value
->color
.float32
[0]) >> 4;
2121 g
= float_to_ubyte(clear_value
->color
.float32
[1]) >> 4;
2122 internal_clear_value
.color
.uint32
[0] = (r
<< 4) | (g
& 0xf);
2125 if (format
== VK_FORMAT_R32G32B32_UINT
||
2126 format
== VK_FORMAT_R32G32B32_SINT
||
2127 format
== VK_FORMAT_R32G32B32_SFLOAT
)
2130 for (uint32_t r
= 0; r
< range_count
; r
++) {
2131 const VkImageSubresourceRange
*range
= &ranges
[r
];
2133 /* Try to perform a fast clear first, otherwise fallback to
2137 radv_fast_clear_range(cmd_buffer
, image
, format
,
2138 image_layout
, false, range
,
2139 &internal_clear_value
)) {
2143 for (uint32_t l
= 0; l
< radv_get_levelCount(image
, range
); ++l
) {
2144 const uint32_t layer_count
= image
->type
== VK_IMAGE_TYPE_3D
?
2145 radv_minify(image
->info
.depth
, range
->baseMipLevel
+ l
) :
2146 radv_get_layerCount(image
, range
);
2147 for (uint32_t s
= 0; s
< layer_count
; ++s
) {
2150 struct radv_meta_blit2d_surf surf
;
2151 surf
.format
= format
;
2153 surf
.level
= range
->baseMipLevel
+ l
;
2154 surf
.layer
= range
->baseArrayLayer
+ s
;
2155 surf
.aspect_mask
= range
->aspectMask
;
2156 radv_meta_clear_image_cs(cmd_buffer
, &surf
,
2157 &internal_clear_value
.color
);
2159 radv_clear_image_layer(cmd_buffer
, image
, image_layout
,
2160 range
, format
, l
, s
, &internal_clear_value
);
2167 void radv_CmdClearColorImage(
2168 VkCommandBuffer commandBuffer
,
2170 VkImageLayout imageLayout
,
2171 const VkClearColorValue
* pColor
,
2172 uint32_t rangeCount
,
2173 const VkImageSubresourceRange
* pRanges
)
2175 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2176 RADV_FROM_HANDLE(radv_image
, image
, image_h
);
2177 struct radv_meta_saved_state saved_state
;
2178 bool cs
= cmd_buffer
->queue_family_index
== RADV_QUEUE_COMPUTE
;
2181 radv_meta_save(&saved_state
, cmd_buffer
,
2182 RADV_META_SAVE_COMPUTE_PIPELINE
|
2183 RADV_META_SAVE_CONSTANTS
|
2184 RADV_META_SAVE_DESCRIPTORS
);
2186 radv_meta_save(&saved_state
, cmd_buffer
,
2187 RADV_META_SAVE_GRAPHICS_PIPELINE
|
2188 RADV_META_SAVE_CONSTANTS
);
2191 radv_cmd_clear_image(cmd_buffer
, image
, imageLayout
,
2192 (const VkClearValue
*) pColor
,
2193 rangeCount
, pRanges
, cs
);
2195 radv_meta_restore(&saved_state
, cmd_buffer
);
2198 void radv_CmdClearDepthStencilImage(
2199 VkCommandBuffer commandBuffer
,
2201 VkImageLayout imageLayout
,
2202 const VkClearDepthStencilValue
* pDepthStencil
,
2203 uint32_t rangeCount
,
2204 const VkImageSubresourceRange
* pRanges
)
2206 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2207 RADV_FROM_HANDLE(radv_image
, image
, image_h
);
2208 struct radv_meta_saved_state saved_state
;
2210 radv_meta_save(&saved_state
, cmd_buffer
,
2211 RADV_META_SAVE_GRAPHICS_PIPELINE
|
2212 RADV_META_SAVE_CONSTANTS
);
2214 radv_cmd_clear_image(cmd_buffer
, image
, imageLayout
,
2215 (const VkClearValue
*) pDepthStencil
,
2216 rangeCount
, pRanges
, false);
2218 radv_meta_restore(&saved_state
, cmd_buffer
);
2221 void radv_CmdClearAttachments(
2222 VkCommandBuffer commandBuffer
,
2223 uint32_t attachmentCount
,
2224 const VkClearAttachment
* pAttachments
,
2226 const VkClearRect
* pRects
)
2228 RADV_FROM_HANDLE(radv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2229 struct radv_meta_saved_state saved_state
;
2230 enum radv_cmd_flush_bits pre_flush
= 0;
2231 enum radv_cmd_flush_bits post_flush
= 0;
2233 if (!cmd_buffer
->state
.subpass
)
2236 radv_meta_save(&saved_state
, cmd_buffer
,
2237 RADV_META_SAVE_GRAPHICS_PIPELINE
|
2238 RADV_META_SAVE_CONSTANTS
);
2240 /* FINISHME: We can do better than this dumb loop. It thrashes too much
2243 for (uint32_t a
= 0; a
< attachmentCount
; ++a
) {
2244 for (uint32_t r
= 0; r
< rectCount
; ++r
) {
2245 emit_clear(cmd_buffer
, &pAttachments
[a
], &pRects
[r
], &pre_flush
, &post_flush
,
2246 cmd_buffer
->state
.subpass
->view_mask
, false);
2250 radv_meta_restore(&saved_state
, cmd_buffer
);
2251 cmd_buffer
->state
.flush_bits
|= post_flush
;