projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radv: do not use gfx fast depth clears for layered depth/stencil images
[mesa.git] / src / amd / vulkan / radv_meta_clear.c
1 /*
2 * Copyright © 2015 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 */
23
24 #include "radv_debug.h"
25 #include "radv_meta.h"
26 #include "radv_private.h"
27 #include "nir/nir_builder.h"
28
29 #include "util/format_rgb9e5.h"
30 #include "vk_format.h"
31
32 enum {
33 DEPTH_CLEAR_SLOW,
34 DEPTH_CLEAR_FAST_EXPCLEAR,
35 DEPTH_CLEAR_FAST_NO_EXPCLEAR
36 };
37
38 static void
39 build_color_shaders(struct nir_shader **out_vs,
40 struct nir_shader **out_fs,
41 uint32_t frag_output)
42 {
43 nir_builder vs_b;
44 nir_builder fs_b;
45
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);
48
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");
51
52 const struct glsl_type *position_type = glsl_vec4_type();
53 const struct glsl_type *color_type = glsl_vec4_type();
54
55 nir_variable *vs_out_pos =
56 nir_variable_create(vs_b.shader, nir_var_shader_out, position_type,
57 "gl_Position");
58 vs_out_pos->data.location = VARYING_SLOT_POS;
59
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);
67
68 nir_variable *fs_out_color =
69 nir_variable_create(fs_b.shader, nir_var_shader_out, color_type,
70 "f_color");
71 fs_out_color->data.location = FRAG_RESULT_DATA0 + frag_output;
72
73 nir_store_var(&fs_b, fs_out_color, &in_color_load->dest.ssa, 0xf);
74
75 nir_ssa_def *outvec = radv_meta_gen_rect_vertices(&vs_b);
76 nir_store_var(&vs_b, vs_out_pos, outvec, 0xf);
77
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,
81 "v_layer");
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);
86
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);
89
90 *out_vs = vs_b.shader;
91 *out_fs = fs_b.shader;
92 }
93
94 static VkResult
95 create_pipeline(struct radv_device *device,
96 struct radv_render_pass *render_pass,
97 uint32_t samples,
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)
107 {
108 VkDevice device_h = radv_device_to_handle(device);
109 VkResult result;
110
111 struct radv_shader_module vs_m = { .nir = vs_nir };
112 struct radv_shader_module fs_m = { .nir = fs_nir };
113
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[]) {
120 {
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),
124 .pName = "main",
125 },
126 {
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),
130 .pName = "main",
131 },
132 },
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,
138 },
139 .pViewportState = &(VkPipelineViewportStateCreateInfo) {
140 .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
141 .viewportCount = 1,
142 .scissorCount = 1,
143 },
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,
151 },
152 .pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
153 .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
154 .rasterizationSamples = samples,
155 .sampleShadingEnable = false,
156 .pSampleMask = NULL,
157 .alphaToCoverageEnable = false,
158 .alphaToOneEnable = false,
159 },
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.
167 */
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,
180 },
181 },
182 .layout = layout,
183 .flags = 0,
184 .renderPass = radv_render_pass_to_handle(render_pass),
185 .subpass = 0,
186 },
187 extra,
188 alloc,
189 pipeline);
190
191 ralloc_free(vs_nir);
192 ralloc_free(fs_nir);
193
194 return result;
195 }
196
197 static VkResult
198 create_color_renderpass(struct radv_device *device,
199 VkFormat vk_format,
200 uint32_t samples,
201 VkRenderPass *pass)
202 {
203 mtx_lock(&device->meta_state.mtx);
204 if (*pass) {
205 mtx_unlock (&device->meta_state.mtx);
206 return VK_SUCCESS;
207 }
208
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) {
214 .format = vk_format,
215 .samples = samples,
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,
220 },
221 .subpassCount = 1,
222 .pSubpasses = &(VkSubpassDescription) {
223 .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
224 .inputAttachmentCount = 0,
225 .colorAttachmentCount = 1,
226 .pColorAttachments = &(VkAttachmentReference) {
227 .attachment = 0,
228 .layout = VK_IMAGE_LAYOUT_GENERAL,
229 },
230 .pResolveAttachments = NULL,
231 .pDepthStencilAttachment = &(VkAttachmentReference) {
232 .attachment = VK_ATTACHMENT_UNUSED,
233 .layout = VK_IMAGE_LAYOUT_GENERAL,
234 },
235 .preserveAttachmentCount = 0,
236 .pPreserveAttachments = NULL,
237 },
238 .dependencyCount = 0,
239 }, &device->meta_state.alloc, pass);
240 mtx_unlock(&device->meta_state.mtx);
241 return result;
242 }
243
244 static VkResult
245 create_color_pipeline(struct radv_device *device,
246 uint32_t samples,
247 uint32_t frag_output,
248 VkPipeline *pipeline,
249 VkRenderPass pass)
250 {
251 struct nir_shader *vs_nir;
252 struct nir_shader *fs_nir;
253 VkResult result;
254
255 mtx_lock(&device->meta_state.mtx);
256 if (*pipeline) {
257 mtx_unlock(&device->meta_state.mtx);
258 return VK_SUCCESS;
259 }
260
261 build_color_shaders(&vs_nir, &fs_nir, frag_output);
262
263 const VkPipelineVertexInputStateCreateInfo vi_state = {
264 .sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
265 .vertexBindingDescriptionCount = 0,
266 .vertexAttributeDescriptionCount = 0,
267 };
268
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,
275 };
276
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,
284 };
285
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
291 };
292
293
294 struct radv_graphics_pipeline_create_info extra = {
295 .use_rectlist = true,
296 };
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);
301
302 mtx_unlock(&device->meta_state.mtx);
303 return result;
304 }
305
306 static void
307 finish_meta_clear_htile_mask_state(struct radv_device *device)
308 {
309 struct radv_meta_state *state = &device->meta_state;
310
311 radv_DestroyPipeline(radv_device_to_handle(device),
312 state->clear_htile_mask_pipeline,
313 &state->alloc);
314 radv_DestroyPipelineLayout(radv_device_to_handle(device),
315 state->clear_htile_mask_p_layout,
316 &state->alloc);
317 radv_DestroyDescriptorSetLayout(radv_device_to_handle(device),
318 state->clear_htile_mask_ds_layout,
319 &state->alloc);
320 }
321
322 void
323 radv_device_finish_meta_clear_state(struct radv_device *device)
324 {
325 struct radv_meta_state *state = &device->meta_state;
326
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],
331 &state->alloc);
332 radv_DestroyRenderPass(radv_device_to_handle(device),
333 state->clear[i].render_pass[j],
334 &state->alloc);
335 }
336
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],
340 &state->alloc);
341 radv_DestroyPipeline(radv_device_to_handle(device),
342 state->clear[i].stencil_only_pipeline[j],
343 &state->alloc);
344 radv_DestroyPipeline(radv_device_to_handle(device),
345 state->clear[i].depthstencil_pipeline[j],
346 &state->alloc);
347 }
348 radv_DestroyRenderPass(radv_device_to_handle(device),
349 state->clear[i].depthstencil_rp,
350 &state->alloc);
351 }
352 radv_DestroyPipelineLayout(radv_device_to_handle(device),
353 state->clear_color_p_layout,
354 &state->alloc);
355 radv_DestroyPipelineLayout(radv_device_to_handle(device),
356 state->clear_depth_p_layout,
357 &state->alloc);
358
359 finish_meta_clear_htile_mask_state(device);
360 }
361
362 static void
363 emit_color_clear(struct radv_cmd_buffer *cmd_buffer,
364 const VkClearAttachment *clear_att,
365 const VkClearRect *clear_rect,
366 uint32_t view_mask)
367 {
368 struct radv_device *device = cmd_buffer->device;
369 const struct radv_subpass *subpass = cmd_buffer->state.subpass;
370 const struct radv_framebuffer *fb = cmd_buffer->state.framebuffer;
371 const uint32_t subpass_att = clear_att->colorAttachment;
372 const uint32_t pass_att = subpass->color_attachments[subpass_att].attachment;
373 const struct radv_image_view *iview = fb ? fb->attachments[pass_att].attachment : NULL;
374 uint32_t samples, samples_log2;
375 VkFormat format;
376 unsigned fs_key;
377 VkClearColorValue clear_value = clear_att->clearValue.color;
378 VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
379 VkPipeline pipeline;
380
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.
384 */
385 if (iview) {
386 samples = iview->image->info.samples;
387 format = iview->vk_format;
388 } else {
389 samples = cmd_buffer->state.pass->attachments[pass_att].samples;
390 format = cmd_buffer->state.pass->attachments[pass_att].format;
391 }
392
393 samples_log2 = ffs(samples) - 1;
394 fs_key = radv_format_meta_fs_key(format);
395
396 if (fs_key == -1) {
397 radv_finishme("color clears incomplete");
398 return;
399 }
400
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],
403 samples,
404 &device->meta_state.clear[samples_log2].render_pass[fs_key]);
405 if (ret != VK_SUCCESS) {
406 cmd_buffer->record_result = ret;
407 return;
408 }
409 }
410
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;
417 return;
418 }
419 }
420
421 pipeline = device->meta_state.clear[samples_log2].color_pipelines[fs_key];
422 if (!pipeline) {
423 radv_finishme("color clears incomplete");
424 return;
425 }
426 assert(samples_log2 < ARRAY_SIZE(device->meta_state.clear));
427 assert(pipeline);
428 assert(clear_att->aspectMask == VK_IMAGE_ASPECT_COLOR_BIT);
429 assert(clear_att->colorAttachment < subpass->color_count);
430
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,
434 &clear_value);
435
436 struct radv_subpass clear_subpass = {
437 .color_count = 1,
438 .color_attachments = (struct radv_subpass_attachment[]) {
439 subpass->color_attachments[clear_att->colorAttachment]
440 },
441 .depth_stencil_attachment = NULL,
442 };
443
444 radv_cmd_buffer_set_subpass(cmd_buffer, &clear_subpass);
445
446 radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS,
447 pipeline);
448
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,
454 .minDepth = 0.0f,
455 .maxDepth = 1.0f
456 });
457
458 radv_CmdSetScissor(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1, &clear_rect->rect);
459
460 if (view_mask) {
461 unsigned i;
462 for_each_bit(i, view_mask)
463 radv_CmdDraw(cmd_buffer_h, 3, 1, 0, i);
464 } else {
465 radv_CmdDraw(cmd_buffer_h, 3, clear_rect->layerCount, 0, clear_rect->baseArrayLayer);
466 }
467
468 radv_cmd_buffer_set_subpass(cmd_buffer, subpass);
469 }
470
471
472 static void
473 build_depthstencil_shader(struct nir_shader **out_vs, struct nir_shader **out_fs)
474 {
475 nir_builder vs_b, fs_b;
476
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);
479
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();
483
484 nir_variable *vs_out_pos =
485 nir_variable_create(vs_b.shader, nir_var_shader_out, position_out_type,
486 "gl_Position");
487 vs_out_pos->data.location = VARYING_SLOT_POS;
488
489 nir_intrinsic_instr *in_color_load = nir_intrinsic_instr_create(vs_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(&vs_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(&vs_b, &in_color_load->instr);
496
497 nir_ssa_def *outvec = radv_meta_gen_rect_vertices_comp2(&vs_b, &in_color_load->dest.ssa);
498 nir_store_var(&vs_b, vs_out_pos, outvec, 0xf);
499
500 const struct glsl_type *layer_type = glsl_int_type();
501 nir_variable *vs_out_layer =
502 nir_variable_create(vs_b.shader, nir_var_shader_out, layer_type,
503 "v_layer");
504 vs_out_layer->data.location = VARYING_SLOT_LAYER;
505 vs_out_layer->data.interpolation = INTERP_MODE_FLAT;
506 nir_ssa_def *inst_id = nir_load_instance_id(&vs_b);
507 nir_ssa_def *base_instance = nir_load_base_instance(&vs_b);
508
509 nir_ssa_def *layer_id = nir_iadd(&vs_b, inst_id, base_instance);
510 nir_store_var(&vs_b, vs_out_layer, layer_id, 0x1);
511
512 *out_vs = vs_b.shader;
513 *out_fs = fs_b.shader;
514 }
515
516 static VkResult
517 create_depthstencil_renderpass(struct radv_device *device,
518 uint32_t samples,
519 VkRenderPass *render_pass)
520 {
521 mtx_lock(&device->meta_state.mtx);
522 if (*render_pass) {
523 mtx_unlock(&device->meta_state.mtx);
524 return VK_SUCCESS;
525 }
526
527 VkResult result = radv_CreateRenderPass(radv_device_to_handle(device),
528 &(VkRenderPassCreateInfo) {
529 .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
530 .attachmentCount = 1,
531 .pAttachments = &(VkAttachmentDescription) {
532 .format = VK_FORMAT_D32_SFLOAT_S8_UINT,
533 .samples = samples,
534 .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
535 .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
536 .initialLayout = VK_IMAGE_LAYOUT_GENERAL,
537 .finalLayout = VK_IMAGE_LAYOUT_GENERAL,
538 },
539 .subpassCount = 1,
540 .pSubpasses = &(VkSubpassDescription) {
541 .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
542 .inputAttachmentCount = 0,
543 .colorAttachmentCount = 0,
544 .pColorAttachments = NULL,
545 .pResolveAttachments = NULL,
546 .pDepthStencilAttachment = &(VkAttachmentReference) {
547 .attachment = 0,
548 .layout = VK_IMAGE_LAYOUT_GENERAL,
549 },
550 .preserveAttachmentCount = 0,
551 .pPreserveAttachments = NULL,
552 },
553 .dependencyCount = 0,
554 }, &device->meta_state.alloc, render_pass);
555 mtx_unlock(&device->meta_state.mtx);
556 return result;
557 }
558
559 static VkResult
560 create_depthstencil_pipeline(struct radv_device *device,
561 VkImageAspectFlags aspects,
562 uint32_t samples,
563 int index,
564 VkPipeline *pipeline,
565 VkRenderPass render_pass)
566 {
567 struct nir_shader *vs_nir, *fs_nir;
568 VkResult result;
569
570 mtx_lock(&device->meta_state.mtx);
571 if (*pipeline) {
572 mtx_unlock(&device->meta_state.mtx);
573 return VK_SUCCESS;
574 }
575
576 build_depthstencil_shader(&vs_nir, &fs_nir);
577
578 const VkPipelineVertexInputStateCreateInfo vi_state = {
579 .sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
580 .vertexBindingDescriptionCount = 0,
581 .vertexAttributeDescriptionCount = 0,
582 };
583
584 const VkPipelineDepthStencilStateCreateInfo ds_state = {
585 .sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
586 .depthTestEnable = (aspects & VK_IMAGE_ASPECT_DEPTH_BIT),
587 .depthCompareOp = VK_COMPARE_OP_ALWAYS,
588 .depthWriteEnable = (aspects & VK_IMAGE_ASPECT_DEPTH_BIT),
589 .depthBoundsTestEnable = false,
590 .stencilTestEnable = (aspects & VK_IMAGE_ASPECT_STENCIL_BIT),
591 .front = {
592 .passOp = VK_STENCIL_OP_REPLACE,
593 .compareOp = VK_COMPARE_OP_ALWAYS,
594 .writeMask = UINT32_MAX,
595 .reference = 0, /* dynamic */
596 },
597 .back = { 0 /* dont care */ },
598 };
599
600 const VkPipelineColorBlendStateCreateInfo cb_state = {
601 .sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
602 .logicOpEnable = false,
603 .attachmentCount = 0,
604 .pAttachments = NULL,
605 };
606
607 struct radv_graphics_pipeline_create_info extra = {
608 .use_rectlist = true,
609 };
610
611 if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
612 extra.db_depth_clear = index == DEPTH_CLEAR_SLOW ? false : true;
613 extra.db_depth_disable_expclear = index == DEPTH_CLEAR_FAST_NO_EXPCLEAR ? true : false;
614 }
615 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
616 extra.db_stencil_clear = index == DEPTH_CLEAR_SLOW ? false : true;
617 extra.db_stencil_disable_expclear = index == DEPTH_CLEAR_FAST_NO_EXPCLEAR ? true : false;
618 }
619 result = create_pipeline(device, radv_render_pass_from_handle(render_pass),
620 samples, vs_nir, fs_nir, &vi_state, &ds_state, &cb_state,
621 device->meta_state.clear_depth_p_layout,
622 &extra, &device->meta_state.alloc, pipeline);
623
624 mtx_unlock(&device->meta_state.mtx);
625 return result;
626 }
627
628 static bool depth_view_can_fast_clear(struct radv_cmd_buffer *cmd_buffer,
629 const struct radv_image_view *iview,
630 VkImageAspectFlags aspects,
631 VkImageLayout layout,
632 const VkClearRect *clear_rect,
633 VkClearDepthStencilValue clear_value)
634 {
635 if (!iview)
636 return false;
637
638 uint32_t queue_mask = radv_image_queue_family_mask(iview->image,
639 cmd_buffer->queue_family_index,
640 cmd_buffer->queue_family_index);
641 if (clear_rect->rect.offset.x || clear_rect->rect.offset.y ||
642 clear_rect->rect.extent.width != iview->extent.width ||
643 clear_rect->rect.extent.height != iview->extent.height)
644 return false;
645 if (radv_image_is_tc_compat_htile(iview->image) &&
646 (((aspects & VK_IMAGE_ASPECT_DEPTH_BIT) && clear_value.depth != 0.0 &&
647 clear_value.depth != 1.0) ||
648 ((aspects & VK_IMAGE_ASPECT_STENCIL_BIT) && clear_value.stencil != 0)))
649 return false;
650 if (radv_image_has_htile(iview->image) &&
651 iview->base_mip == 0 &&
652 iview->base_layer == 0 &&
653 iview->layer_count == iview->image->info.array_size &&
654 radv_layout_is_htile_compressed(iview->image, layout, queue_mask) &&
655 radv_image_extent_compare(iview->image, &iview->extent))
656 return true;
657 return false;
658 }
659
660 static VkPipeline
661 pick_depthstencil_pipeline(struct radv_cmd_buffer *cmd_buffer,
662 struct radv_meta_state *meta_state,
663 const struct radv_image_view *iview,
664 int samples_log2,
665 VkImageAspectFlags aspects,
666 VkImageLayout layout,
667 const VkClearRect *clear_rect,
668 VkClearDepthStencilValue clear_value)
669 {
670 bool fast = depth_view_can_fast_clear(cmd_buffer, iview, aspects, layout, clear_rect, clear_value);
671 int index = DEPTH_CLEAR_SLOW;
672 VkPipeline *pipeline;
673
674 if (fast) {
675 /* we don't know the previous clear values, so we always have
676 * the NO_EXPCLEAR path */
677 index = DEPTH_CLEAR_FAST_NO_EXPCLEAR;
678 }
679
680 switch (aspects) {
681 case VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT:
682 pipeline = &meta_state->clear[samples_log2].depthstencil_pipeline[index];
683 break;
684 case VK_IMAGE_ASPECT_DEPTH_BIT:
685 pipeline = &meta_state->clear[samples_log2].depth_only_pipeline[index];
686 break;
687 case VK_IMAGE_ASPECT_STENCIL_BIT:
688 pipeline = &meta_state->clear[samples_log2].stencil_only_pipeline[index];
689 break;
690 default:
691 unreachable("expected depth or stencil aspect");
692 }
693
694 if (cmd_buffer->device->meta_state.clear[samples_log2].depthstencil_rp == VK_NULL_HANDLE) {
695 VkResult ret = create_depthstencil_renderpass(cmd_buffer->device, 1u << samples_log2,
696 &cmd_buffer->device->meta_state.clear[samples_log2].depthstencil_rp);
697 if (ret != VK_SUCCESS) {
698 cmd_buffer->record_result = ret;
699 return VK_NULL_HANDLE;
700 }
701 }
702
703 if (*pipeline == VK_NULL_HANDLE) {
704 VkResult ret = create_depthstencil_pipeline(cmd_buffer->device, aspects, 1u << samples_log2, index,
705 pipeline, 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;
709 }
710 }
711 return *pipeline;
712 }
713
714 static void
715 emit_depthstencil_clear(struct radv_cmd_buffer *cmd_buffer,
716 const VkClearAttachment *clear_att,
717 const VkClearRect *clear_rect,
718 uint32_t view_mask)
719 {
720 struct radv_device *device = cmd_buffer->device;
721 struct radv_meta_state *meta_state = &device->meta_state;
722 const struct radv_subpass *subpass = cmd_buffer->state.subpass;
723 const struct radv_framebuffer *fb = cmd_buffer->state.framebuffer;
724 const uint32_t pass_att = subpass->depth_stencil_attachment->attachment;
725 VkClearDepthStencilValue clear_value = clear_att->clearValue.depthStencil;
726 VkImageAspectFlags aspects = clear_att->aspectMask;
727 const struct radv_image_view *iview = fb ? fb->attachments[pass_att].attachment : NULL;
728 uint32_t samples, samples_log2;
729 VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
730
731 /* When a framebuffer is bound to the current command buffer, get the
732 * number of samples from it. Otherwise, get the number of samples from
733 * the render pass because it's likely a secondary command buffer.
734 */
735 if (iview) {
736 samples = iview->image->info.samples;
737 } else {
738 samples = cmd_buffer->state.pass->attachments[pass_att].samples;
739 }
740
741 samples_log2 = ffs(samples) - 1;
742
743 assert(pass_att != VK_ATTACHMENT_UNUSED);
744
745 if (!(aspects & VK_IMAGE_ASPECT_DEPTH_BIT))
746 clear_value.depth = 1.0f;
747
748 radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
749 device->meta_state.clear_depth_p_layout,
750 VK_SHADER_STAGE_VERTEX_BIT, 0, 4,
751 &clear_value.depth);
752
753 uint32_t prev_reference = cmd_buffer->state.dynamic.stencil_reference.front;
754 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
755 radv_CmdSetStencilReference(cmd_buffer_h, VK_STENCIL_FACE_FRONT_BIT,
756 clear_value.stencil);
757 }
758
759 VkPipeline pipeline = pick_depthstencil_pipeline(cmd_buffer,
760 meta_state,
761 iview,
762 samples_log2,
763 aspects,
764 subpass->depth_stencil_attachment->layout,
765 clear_rect,
766 clear_value);
767 if (!pipeline)
768 return;
769
770 radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS,
771 pipeline);
772
773 if (depth_view_can_fast_clear(cmd_buffer, iview, aspects,
774 subpass->depth_stencil_attachment->layout,
775 clear_rect, clear_value))
776 radv_update_ds_clear_metadata(cmd_buffer, iview->image,
777 clear_value, aspects);
778
779 radv_CmdSetViewport(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1, &(VkViewport) {
780 .x = clear_rect->rect.offset.x,
781 .y = clear_rect->rect.offset.y,
782 .width = clear_rect->rect.extent.width,
783 .height = clear_rect->rect.extent.height,
784 .minDepth = 0.0f,
785 .maxDepth = 1.0f
786 });
787
788 radv_CmdSetScissor(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1, &clear_rect->rect);
789
790 if (view_mask) {
791 unsigned i;
792 for_each_bit(i, view_mask)
793 radv_CmdDraw(cmd_buffer_h, 3, 1, 0, i);
794 } else {
795 radv_CmdDraw(cmd_buffer_h, 3, clear_rect->layerCount, 0, clear_rect->baseArrayLayer);
796 }
797
798 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
799 radv_CmdSetStencilReference(cmd_buffer_h, VK_STENCIL_FACE_FRONT_BIT,
800 prev_reference);
801 }
802 }
803
804 static uint32_t
805 clear_htile_mask(struct radv_cmd_buffer *cmd_buffer,
806 struct radeon_winsys_bo *bo, uint64_t offset, uint64_t size,
807 uint32_t htile_value, uint32_t htile_mask)
808 {
809 struct radv_device *device = cmd_buffer->device;
810 struct radv_meta_state *state = &device->meta_state;
811 uint64_t block_count = round_up_u64(size, 1024);
812 struct radv_meta_saved_state saved_state;
813
814 radv_meta_save(&saved_state, cmd_buffer,
815 RADV_META_SAVE_COMPUTE_PIPELINE |
816 RADV_META_SAVE_CONSTANTS |
817 RADV_META_SAVE_DESCRIPTORS);
818
819 struct radv_buffer dst_buffer = {
820 .bo = bo,
821 .offset = offset,
822 .size = size
823 };
824
825 radv_CmdBindPipeline(radv_cmd_buffer_to_handle(cmd_buffer),
826 VK_PIPELINE_BIND_POINT_COMPUTE,
827 state->clear_htile_mask_pipeline);
828
829 radv_meta_push_descriptor_set(cmd_buffer, VK_PIPELINE_BIND_POINT_COMPUTE,
830 state->clear_htile_mask_p_layout,
831 0, /* set */
832 1, /* descriptorWriteCount */
833 (VkWriteDescriptorSet[]) {
834 {
835 .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
836 .dstBinding = 0,
837 .dstArrayElement = 0,
838 .descriptorCount = 1,
839 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
840 .pBufferInfo = &(VkDescriptorBufferInfo) {
841 .buffer = radv_buffer_to_handle(&dst_buffer),
842 .offset = 0,
843 .range = size
844 }
845 }
846 });
847
848 const unsigned constants[2] = {
849 htile_value & htile_mask,
850 ~htile_mask,
851 };
852
853 radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
854 state->clear_htile_mask_p_layout,
855 VK_SHADER_STAGE_COMPUTE_BIT, 0, 8,
856 constants);
857
858 radv_CmdDispatch(radv_cmd_buffer_to_handle(cmd_buffer), block_count, 1, 1);
859
860 radv_meta_restore(&saved_state, cmd_buffer);
861
862 return RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
863 RADV_CMD_FLAG_INV_VMEM_L1 |
864 RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2;
865 }
866
867 static uint32_t
868 radv_get_htile_fast_clear_value(const struct radv_image *image,
869 VkClearDepthStencilValue value)
870 {
871 uint32_t clear_value;
872
873 if (!image->planes[0].surface.has_stencil) {
874 clear_value = value.depth ? 0xfffffff0 : 0;
875 } else {
876 clear_value = value.depth ? 0xfffc0000 : 0;
877 }
878
879 return clear_value;
880 }
881
882 static uint32_t
883 radv_get_htile_mask(const struct radv_image *image, VkImageAspectFlags aspects)
884 {
885 uint32_t mask = 0;
886
887 if (!image->planes[0].surface.has_stencil) {
888 /* All the HTILE buffer is used when there is no stencil. */
889 mask = UINT32_MAX;
890 } else {
891 if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
892 mask |= 0xfffffc0f;
893 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT)
894 mask |= 0x000003f0;
895 }
896
897 return mask;
898 }
899
900 static bool
901 radv_is_fast_clear_depth_allowed(VkClearDepthStencilValue value)
902 {
903 return value.depth == 1.0f || value.depth == 0.0f;
904 }
905
906 static bool
907 radv_is_fast_clear_stencil_allowed(VkClearDepthStencilValue value)
908 {
909 return value.stencil == 0;
910 }
911
912 /**
913 * Determine if the given image can be fast cleared.
914 */
915 static bool
916 radv_image_can_fast_clear(struct radv_device *device, struct radv_image *image)
917 {
918 if (device->instance->debug_flags & RADV_DEBUG_NO_FAST_CLEARS)
919 return false;
920
921 if (vk_format_is_color(image->vk_format)) {
922 if (!radv_image_has_cmask(image) && !radv_image_has_dcc(image))
923 return false;
924
925 /* RB+ doesn't work with CMASK fast clear on Stoney. */
926 if (!radv_image_has_dcc(image) &&
927 device->physical_device->rad_info.family == CHIP_STONEY)
928 return false;
929 } else {
930 if (!radv_image_has_htile(image))
931 return false;
932 }
933
934 /* Do not fast clears 3D images. */
935 if (image->type == VK_IMAGE_TYPE_3D)
936 return false;
937
938 return true;
939 }
940
941 /**
942 * Determine if the given image view can be fast cleared.
943 */
944 static bool
945 radv_image_view_can_fast_clear(struct radv_device *device,
946 const struct radv_image_view *iview)
947 {
948 struct radv_image *image;
949
950 if (!iview)
951 return false;
952 image = iview->image;
953
954 /* Only fast clear if the image itself can be fast cleared. */
955 if (!radv_image_can_fast_clear(device, image))
956 return false;
957
958 /* Only fast clear if all layers are bound. */
959 if (iview->base_layer > 0 ||
960 iview->layer_count != image->info.array_size)
961 return false;
962
963 /* Only fast clear if the view covers the whole image. */
964 if (!radv_image_extent_compare(image, &iview->extent))
965 return false;
966
967 return true;
968 }
969
970 static bool
971 radv_can_fast_clear_depth(struct radv_cmd_buffer *cmd_buffer,
972 const struct radv_image_view *iview,
973 VkImageLayout image_layout,
974 VkImageAspectFlags aspects,
975 const VkClearRect *clear_rect,
976 const VkClearDepthStencilValue clear_value,
977 uint32_t view_mask)
978 {
979 if (!radv_image_view_can_fast_clear(cmd_buffer->device, iview))
980 return false;
981
982 if (!radv_layout_is_htile_compressed(iview->image, image_layout, radv_image_queue_family_mask(iview->image, cmd_buffer->queue_family_index, cmd_buffer->queue_family_index)))
983 return false;
984
985 if (clear_rect->rect.offset.x || clear_rect->rect.offset.y ||
986 clear_rect->rect.extent.width != iview->image->info.width ||
987 clear_rect->rect.extent.height != iview->image->info.height)
988 return false;
989
990 if (view_mask && (iview->image->info.array_size >= 32 ||
991 (1u << iview->image->info.array_size) - 1u != view_mask))
992 return false;
993 if (!view_mask && clear_rect->baseArrayLayer != 0)
994 return false;
995 if (!view_mask && clear_rect->layerCount != iview->image->info.array_size)
996 return false;
997
998 if (cmd_buffer->device->physical_device->rad_info.chip_class < GFX9 &&
999 (!(aspects & VK_IMAGE_ASPECT_DEPTH_BIT) ||
1000 ((vk_format_aspects(iview->image->vk_format) & VK_IMAGE_ASPECT_STENCIL_BIT) &&
1001 !(aspects & VK_IMAGE_ASPECT_STENCIL_BIT))))
1002 return false;
1003
1004 if (((aspects & VK_IMAGE_ASPECT_DEPTH_BIT) &&
1005 !radv_is_fast_clear_depth_allowed(clear_value)) ||
1006 ((aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
1007 !radv_is_fast_clear_stencil_allowed(clear_value)))
1008 return false;
1009
1010 return true;
1011 }
1012
1013 static void
1014 radv_fast_clear_depth(struct radv_cmd_buffer *cmd_buffer,
1015 const struct radv_image_view *iview,
1016 const VkClearAttachment *clear_att,
1017 enum radv_cmd_flush_bits *pre_flush,
1018 enum radv_cmd_flush_bits *post_flush)
1019 {
1020 VkClearDepthStencilValue clear_value = clear_att->clearValue.depthStencil;
1021 VkImageAspectFlags aspects = clear_att->aspectMask;
1022 uint32_t clear_word, flush_bits;
1023 uint32_t htile_mask;
1024
1025 clear_word = radv_get_htile_fast_clear_value(iview->image, clear_value);
1026 htile_mask = radv_get_htile_mask(iview->image, aspects);
1027
1028 if (pre_flush) {
1029 cmd_buffer->state.flush_bits |= (RADV_CMD_FLAG_FLUSH_AND_INV_DB |
1030 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META) & ~ *pre_flush;
1031 *pre_flush |= cmd_buffer->state.flush_bits;
1032 }
1033
1034 if (htile_mask == UINT_MAX) {
1035 /* Clear the whole HTILE buffer. */
1036 flush_bits = radv_fill_buffer(cmd_buffer, iview->image->bo,
1037 iview->image->offset + iview->image->htile_offset,
1038 iview->image->planes[0].surface.htile_size, clear_word);
1039 } else {
1040 /* Only clear depth or stencil bytes in the HTILE buffer. */
1041 assert(cmd_buffer->device->physical_device->rad_info.chip_class >= GFX9);
1042 flush_bits = clear_htile_mask(cmd_buffer, iview->image->bo,
1043 iview->image->offset + iview->image->htile_offset,
1044 iview->image->planes[0].surface.htile_size, clear_word,
1045 htile_mask);
1046 }
1047
1048 radv_update_ds_clear_metadata(cmd_buffer, iview->image, clear_value, aspects);
1049 if (post_flush) {
1050 *post_flush |= flush_bits;
1051 }
1052 }
1053
1054 static nir_shader *
1055 build_clear_htile_mask_shader()
1056 {
1057 nir_builder b;
1058
1059 nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_COMPUTE, NULL);
1060 b.shader->info.name = ralloc_strdup(b.shader, "meta_clear_htile_mask");
1061 b.shader->info.cs.local_size[0] = 64;
1062 b.shader->info.cs.local_size[1] = 1;
1063 b.shader->info.cs.local_size[2] = 1;
1064
1065 nir_ssa_def *invoc_id = nir_load_local_invocation_id(&b);
1066 nir_ssa_def *wg_id = nir_load_work_group_id(&b);
1067 nir_ssa_def *block_size = nir_imm_ivec4(&b,
1068 b.shader->info.cs.local_size[0],
1069 b.shader->info.cs.local_size[1],
1070 b.shader->info.cs.local_size[2], 0);
1071
1072 nir_ssa_def *global_id = nir_iadd(&b, nir_imul(&b, wg_id, block_size), invoc_id);
1073
1074 nir_ssa_def *offset = nir_imul(&b, global_id, nir_imm_int(&b, 16));
1075 offset = nir_channel(&b, offset, 0);
1076
1077 nir_intrinsic_instr *buf =
1078 nir_intrinsic_instr_create(b.shader,
1079 nir_intrinsic_vulkan_resource_index);
1080
1081 buf->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
1082 buf->num_components = 1;
1083 nir_intrinsic_set_desc_set(buf, 0);
1084 nir_intrinsic_set_binding(buf, 0);
1085 nir_ssa_dest_init(&buf->instr, &buf->dest, buf->num_components, 32, NULL);
1086 nir_builder_instr_insert(&b, &buf->instr);
1087
1088 nir_intrinsic_instr *constants =
1089 nir_intrinsic_instr_create(b.shader,
1090 nir_intrinsic_load_push_constant);
1091 nir_intrinsic_set_base(constants, 0);
1092 nir_intrinsic_set_range(constants, 8);
1093 constants->src[0] = nir_src_for_ssa(nir_imm_int(&b, 0));
1094 constants->num_components = 2;
1095 nir_ssa_dest_init(&constants->instr, &constants->dest, 2, 32, "constants");
1096 nir_builder_instr_insert(&b, &constants->instr);
1097
1098 nir_intrinsic_instr *load =
1099 nir_intrinsic_instr_create(b.shader, nir_intrinsic_load_ssbo);
1100 load->src[0] = nir_src_for_ssa(&buf->dest.ssa);
1101 load->src[1] = nir_src_for_ssa(offset);
1102 nir_ssa_dest_init(&load->instr, &load->dest, 4, 32, NULL);
1103 load->num_components = 4;
1104 nir_builder_instr_insert(&b, &load->instr);
1105
1106 /* data = (data & ~htile_mask) | (htile_value & htile_mask) */
1107 nir_ssa_def *data =
1108 nir_iand(&b, &load->dest.ssa,
1109 nir_channel(&b, &constants->dest.ssa, 1));
1110 data = nir_ior(&b, data, nir_channel(&b, &constants->dest.ssa, 0));
1111
1112 nir_intrinsic_instr *store =
1113 nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_ssbo);
1114 store->src[0] = nir_src_for_ssa(data);
1115 store->src[1] = nir_src_for_ssa(&buf->dest.ssa);
1116 store->src[2] = nir_src_for_ssa(offset);
1117 nir_intrinsic_set_write_mask(store, 0xf);
1118 nir_intrinsic_set_access(store, ACCESS_NON_READABLE);
1119 store->num_components = 4;
1120 nir_builder_instr_insert(&b, &store->instr);
1121
1122 return b.shader;
1123 }
1124
1125 static VkResult
1126 init_meta_clear_htile_mask_state(struct radv_device *device)
1127 {
1128 struct radv_meta_state *state = &device->meta_state;
1129 struct radv_shader_module cs = { .nir = NULL };
1130 VkResult result;
1131
1132 cs.nir = build_clear_htile_mask_shader();
1133
1134 VkDescriptorSetLayoutCreateInfo ds_layout_info = {
1135 .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
1136 .flags = VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR,
1137 .bindingCount = 1,
1138 .pBindings = (VkDescriptorSetLayoutBinding[]) {
1139 {
1140 .binding = 0,
1141 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
1142 .descriptorCount = 1,
1143 .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
1144 .pImmutableSamplers = NULL
1145 },
1146 }
1147 };
1148
1149 result = radv_CreateDescriptorSetLayout(radv_device_to_handle(device),
1150 &ds_layout_info, &state->alloc,
1151 &state->clear_htile_mask_ds_layout);
1152 if (result != VK_SUCCESS)
1153 goto fail;
1154
1155 VkPipelineLayoutCreateInfo p_layout_info = {
1156 .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
1157 .setLayoutCount = 1,
1158 .pSetLayouts = &state->clear_htile_mask_ds_layout,
1159 .pushConstantRangeCount = 1,
1160 .pPushConstantRanges = &(VkPushConstantRange){
1161 VK_SHADER_STAGE_COMPUTE_BIT, 0, 8,
1162 },
1163 };
1164
1165 result = radv_CreatePipelineLayout(radv_device_to_handle(device),
1166 &p_layout_info, &state->alloc,
1167 &state->clear_htile_mask_p_layout);
1168 if (result != VK_SUCCESS)
1169 goto fail;
1170
1171 VkPipelineShaderStageCreateInfo shader_stage = {
1172 .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
1173 .stage = VK_SHADER_STAGE_COMPUTE_BIT,
1174 .module = radv_shader_module_to_handle(&cs),
1175 .pName = "main",
1176 .pSpecializationInfo = NULL,
1177 };
1178
1179 VkComputePipelineCreateInfo pipeline_info = {
1180 .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
1181 .stage = shader_stage,
1182 .flags = 0,
1183 .layout = state->clear_htile_mask_p_layout,
1184 };
1185
1186 result = radv_CreateComputePipelines(radv_device_to_handle(device),
1187 radv_pipeline_cache_to_handle(&state->cache),
1188 1, &pipeline_info, NULL,
1189 &state->clear_htile_mask_pipeline);
1190
1191 ralloc_free(cs.nir);
1192 return result;
1193 fail:
1194 ralloc_free(cs.nir);
1195 return result;
1196 }
1197
1198 VkResult
1199 radv_device_init_meta_clear_state(struct radv_device *device, bool on_demand)
1200 {
1201 VkResult res;
1202 struct radv_meta_state *state = &device->meta_state;
1203
1204 VkPipelineLayoutCreateInfo pl_color_create_info = {
1205 .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
1206 .setLayoutCount = 0,
1207 .pushConstantRangeCount = 1,
1208 .pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_FRAGMENT_BIT, 0, 16},
1209 };
1210
1211 res = radv_CreatePipelineLayout(radv_device_to_handle(device),
1212 &pl_color_create_info,
1213 &device->meta_state.alloc,
1214 &device->meta_state.clear_color_p_layout);
1215 if (res != VK_SUCCESS)
1216 goto fail;
1217
1218 VkPipelineLayoutCreateInfo pl_depth_create_info = {
1219 .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
1220 .setLayoutCount = 0,
1221 .pushConstantRangeCount = 1,
1222 .pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_VERTEX_BIT, 0, 4},
1223 };
1224
1225 res = radv_CreatePipelineLayout(radv_device_to_handle(device),
1226 &pl_depth_create_info,
1227 &device->meta_state.alloc,
1228 &device->meta_state.clear_depth_p_layout);
1229 if (res != VK_SUCCESS)
1230 goto fail;
1231
1232 res = init_meta_clear_htile_mask_state(device);
1233 if (res != VK_SUCCESS)
1234 goto fail;
1235
1236 if (on_demand)
1237 return VK_SUCCESS;
1238
1239 for (uint32_t i = 0; i < ARRAY_SIZE(state->clear); ++i) {
1240 uint32_t samples = 1 << i;
1241 for (uint32_t j = 0; j < NUM_META_FS_KEYS; ++j) {
1242 VkFormat format = radv_fs_key_format_exemplars[j];
1243 unsigned fs_key = radv_format_meta_fs_key(format);
1244 assert(!state->clear[i].color_pipelines[fs_key]);
1245
1246 res = create_color_renderpass(device, format, samples,
1247 &state->clear[i].render_pass[fs_key]);
1248 if (res != VK_SUCCESS)
1249 goto fail;
1250
1251 res = create_color_pipeline(device, samples, 0, &state->clear[i].color_pipelines[fs_key],
1252 state->clear[i].render_pass[fs_key]);
1253 if (res != VK_SUCCESS)
1254 goto fail;
1255
1256 }
1257
1258 res = create_depthstencil_renderpass(device,
1259 samples,
1260 &state->clear[i].depthstencil_rp);
1261 if (res != VK_SUCCESS)
1262 goto fail;
1263
1264 for (uint32_t j = 0; j < NUM_DEPTH_CLEAR_PIPELINES; j++) {
1265 res = create_depthstencil_pipeline(device,
1266 VK_IMAGE_ASPECT_DEPTH_BIT,
1267 samples,
1268 j,
1269 &state->clear[i].depth_only_pipeline[j],
1270 state->clear[i].depthstencil_rp);
1271 if (res != VK_SUCCESS)
1272 goto fail;
1273
1274 res = create_depthstencil_pipeline(device,
1275 VK_IMAGE_ASPECT_STENCIL_BIT,
1276 samples,
1277 j,
1278 &state->clear[i].stencil_only_pipeline[j],
1279 state->clear[i].depthstencil_rp);
1280 if (res != VK_SUCCESS)
1281 goto fail;
1282
1283 res = create_depthstencil_pipeline(device,
1284 VK_IMAGE_ASPECT_DEPTH_BIT |
1285 VK_IMAGE_ASPECT_STENCIL_BIT,
1286 samples,
1287 j,
1288 &state->clear[i].depthstencil_pipeline[j],
1289 state->clear[i].depthstencil_rp);
1290 if (res != VK_SUCCESS)
1291 goto fail;
1292 }
1293 }
1294 return VK_SUCCESS;
1295
1296 fail:
1297 radv_device_finish_meta_clear_state(device);
1298 return res;
1299 }
1300
1301 static uint32_t
1302 radv_get_cmask_fast_clear_value(const struct radv_image *image)
1303 {
1304 uint32_t value = 0; /* Default value when no DCC. */
1305
1306 /* The fast-clear value is different for images that have both DCC and
1307 * CMASK metadata.
1308 */
1309 if (radv_image_has_dcc(image)) {
1310 /* DCC fast clear with MSAA should clear CMASK to 0xC. */
1311 return image->info.samples > 1 ? 0xcccccccc : 0xffffffff;
1312 }
1313
1314 return value;
1315 }
1316
1317 uint32_t
1318 radv_clear_cmask(struct radv_cmd_buffer *cmd_buffer,
1319 struct radv_image *image, uint32_t value)
1320 {
1321 return radv_fill_buffer(cmd_buffer, image->bo,
1322 image->offset + image->cmask.offset,
1323 image->cmask.size, value);
1324 }
1325
1326
1327 uint32_t
1328 radv_clear_fmask(struct radv_cmd_buffer *cmd_buffer,
1329 struct radv_image *image, uint32_t value)
1330 {
1331 return radv_fill_buffer(cmd_buffer, image->bo,
1332 image->offset + image->fmask.offset,
1333 image->fmask.size, value);
1334 }
1335
1336 uint32_t
1337 radv_clear_dcc(struct radv_cmd_buffer *cmd_buffer,
1338 struct radv_image *image, uint32_t value)
1339 {
1340 /* Mark the image as being compressed. */
1341 radv_update_dcc_metadata(cmd_buffer, image, true);
1342
1343 return radv_fill_buffer(cmd_buffer, image->bo,
1344 image->offset + image->dcc_offset,
1345 image->planes[0].surface.dcc_size, value);
1346 }
1347
1348 uint32_t
1349 radv_clear_htile(struct radv_cmd_buffer *cmd_buffer, struct radv_image *image,
1350 const VkImageSubresourceRange *range, uint32_t value)
1351 {
1352 unsigned layer_count = radv_get_layerCount(image, range);
1353 uint64_t size = image->planes[0].surface.htile_slice_size * layer_count;
1354 uint64_t offset = image->offset + image->htile_offset +
1355 image->planes[0].surface.htile_slice_size * range->baseArrayLayer;
1356
1357 return radv_fill_buffer(cmd_buffer, image->bo, offset, size, value);
1358 }
1359
1360 static void vi_get_fast_clear_parameters(VkFormat format,
1361 const VkClearColorValue *clear_value,
1362 uint32_t* reset_value,
1363 bool *can_avoid_fast_clear_elim)
1364 {
1365 bool values[4] = {};
1366 int extra_channel;
1367 bool main_value = false;
1368 bool extra_value = false;
1369 int i;
1370 *can_avoid_fast_clear_elim = false;
1371
1372 *reset_value = 0x20202020U;
1373
1374 const struct vk_format_description *desc = vk_format_description(format);
1375 if (format == VK_FORMAT_B10G11R11_UFLOAT_PACK32 ||
1376 format == VK_FORMAT_R5G6B5_UNORM_PACK16 ||
1377 format == VK_FORMAT_B5G6R5_UNORM_PACK16)
1378 extra_channel = -1;
1379 else if (desc->layout == VK_FORMAT_LAYOUT_PLAIN) {
1380 if (radv_translate_colorswap(format, false) <= 1)
1381 extra_channel = desc->nr_channels - 1;
1382 else
1383 extra_channel = 0;
1384 } else
1385 return;
1386
1387 for (i = 0; i < 4; i++) {
1388 int index = desc->swizzle[i] - VK_SWIZZLE_X;
1389 if (desc->swizzle[i] < VK_SWIZZLE_X ||
1390 desc->swizzle[i] > VK_SWIZZLE_W)
1391 continue;
1392
1393 if (desc->channel[i].pure_integer &&
1394 desc->channel[i].type == VK_FORMAT_TYPE_SIGNED) {
1395 /* Use the maximum value for clamping the clear color. */
1396 int max = u_bit_consecutive(0, desc->channel[i].size - 1);
1397
1398 values[i] = clear_value->int32[i] != 0;
1399 if (clear_value->int32[i] != 0 && MIN2(clear_value->int32[i], max) != max)
1400 return;
1401 } else if (desc->channel[i].pure_integer &&
1402 desc->channel[i].type == VK_FORMAT_TYPE_UNSIGNED) {
1403 /* Use the maximum value for clamping the clear color. */
1404 unsigned max = u_bit_consecutive(0, desc->channel[i].size);
1405
1406 values[i] = clear_value->uint32[i] != 0U;
1407 if (clear_value->uint32[i] != 0U && MIN2(clear_value->uint32[i], max) != max)
1408 return;
1409 } else {
1410 values[i] = clear_value->float32[i] != 0.0F;
1411 if (clear_value->float32[i] != 0.0F && clear_value->float32[i] != 1.0F)
1412 return;
1413 }
1414
1415 if (index == extra_channel)
1416 extra_value = values[i];
1417 else
1418 main_value = values[i];
1419 }
1420
1421 for (int i = 0; i < 4; ++i)
1422 if (values[i] != main_value &&
1423 desc->swizzle[i] - VK_SWIZZLE_X != extra_channel &&
1424 desc->swizzle[i] >= VK_SWIZZLE_X &&
1425 desc->swizzle[i] <= VK_SWIZZLE_W)
1426 return;
1427
1428 *can_avoid_fast_clear_elim = true;
1429 if (main_value)
1430 *reset_value |= 0x80808080U;
1431
1432 if (extra_value)
1433 *reset_value |= 0x40404040U;
1434 return;
1435 }
1436
1437 static bool
1438 radv_can_fast_clear_color(struct radv_cmd_buffer *cmd_buffer,
1439 const struct radv_image_view *iview,
1440 VkImageLayout image_layout,
1441 const VkClearRect *clear_rect,
1442 VkClearColorValue clear_value,
1443 uint32_t view_mask)
1444 {
1445 uint32_t clear_color[2];
1446
1447 if (!radv_image_view_can_fast_clear(cmd_buffer->device, iview))
1448 return false;
1449
1450 if (!radv_layout_can_fast_clear(iview->image, image_layout, radv_image_queue_family_mask(iview->image, cmd_buffer->queue_family_index, cmd_buffer->queue_family_index)))
1451 return false;
1452
1453 if (clear_rect->rect.offset.x || clear_rect->rect.offset.y ||
1454 clear_rect->rect.extent.width != iview->image->info.width ||
1455 clear_rect->rect.extent.height != iview->image->info.height)
1456 return false;
1457
1458 if (view_mask && (iview->image->info.array_size >= 32 ||
1459 (1u << iview->image->info.array_size) - 1u != view_mask))
1460 return false;
1461 if (!view_mask && clear_rect->baseArrayLayer != 0)
1462 return false;
1463 if (!view_mask && clear_rect->layerCount != iview->image->info.array_size)
1464 return false;
1465
1466 /* DCC */
1467 if (!radv_format_pack_clear_color(iview->vk_format,
1468 clear_color, &clear_value))
1469 return false;
1470
1471 if (radv_image_has_dcc(iview->image)) {
1472 bool can_avoid_fast_clear_elim;
1473 uint32_t reset_value;
1474
1475 vi_get_fast_clear_parameters(iview->vk_format,
1476 &clear_value, &reset_value,
1477 &can_avoid_fast_clear_elim);
1478
1479 if (iview->image->info.samples > 1) {
1480 /* DCC fast clear with MSAA should clear CMASK. */
1481 /* FIXME: This doesn't work for now. There is a
1482 * hardware bug with fast clears and DCC for MSAA
1483 * textures. AMDVLK has a workaround but it doesn't
1484 * seem to work here. Note that we might emit useless
1485 * CB flushes but that shouldn't matter.
1486 */
1487 if (!can_avoid_fast_clear_elim)
1488 return false;
1489 }
1490 }
1491
1492 return true;
1493 }
1494
1495
1496 static void
1497 radv_fast_clear_color(struct radv_cmd_buffer *cmd_buffer,
1498 const struct radv_image_view *iview,
1499 const VkClearAttachment *clear_att,
1500 uint32_t subpass_att,
1501 enum radv_cmd_flush_bits *pre_flush,
1502 enum radv_cmd_flush_bits *post_flush)
1503 {
1504 VkClearColorValue clear_value = clear_att->clearValue.color;
1505 uint32_t clear_color[2], flush_bits = 0;
1506 uint32_t cmask_clear_value;
1507
1508 if (pre_flush) {
1509 cmd_buffer->state.flush_bits |= (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
1510 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META) & ~ *pre_flush;
1511 *pre_flush |= cmd_buffer->state.flush_bits;
1512 }
1513
1514 /* DCC */
1515 radv_format_pack_clear_color(iview->vk_format, clear_color, &clear_value);
1516
1517 cmask_clear_value = radv_get_cmask_fast_clear_value(iview->image);
1518
1519 /* clear cmask buffer */
1520 if (radv_image_has_dcc(iview->image)) {
1521 uint32_t reset_value;
1522 bool can_avoid_fast_clear_elim;
1523 bool need_decompress_pass = false;
1524
1525 vi_get_fast_clear_parameters(iview->vk_format,
1526 &clear_value, &reset_value,
1527 &can_avoid_fast_clear_elim);
1528
1529 if (radv_image_has_cmask(iview->image)) {
1530 flush_bits = radv_clear_cmask(cmd_buffer, iview->image,
1531 cmask_clear_value);
1532
1533 need_decompress_pass = true;
1534 }
1535
1536 if (!can_avoid_fast_clear_elim)
1537 need_decompress_pass = true;
1538
1539 flush_bits |= radv_clear_dcc(cmd_buffer, iview->image, reset_value);
1540
1541 radv_update_fce_metadata(cmd_buffer, iview->image,
1542 need_decompress_pass);
1543 } else {
1544 flush_bits = radv_clear_cmask(cmd_buffer, iview->image,
1545 cmask_clear_value);
1546 }
1547
1548 if (post_flush) {
1549 *post_flush |= flush_bits;
1550 }
1551
1552 radv_update_color_clear_metadata(cmd_buffer, iview->image, subpass_att,
1553 clear_color);
1554 }
1555
1556 /**
1557 * The parameters mean that same as those in vkCmdClearAttachments.
1558 */
1559 static void
1560 emit_clear(struct radv_cmd_buffer *cmd_buffer,
1561 const VkClearAttachment *clear_att,
1562 const VkClearRect *clear_rect,
1563 enum radv_cmd_flush_bits *pre_flush,
1564 enum radv_cmd_flush_bits *post_flush,
1565 uint32_t view_mask)
1566 {
1567 const struct radv_framebuffer *fb = cmd_buffer->state.framebuffer;
1568 const struct radv_subpass *subpass = cmd_buffer->state.subpass;
1569 VkImageAspectFlags aspects = clear_att->aspectMask;
1570
1571 if (aspects & VK_IMAGE_ASPECT_COLOR_BIT) {
1572 const uint32_t subpass_att = clear_att->colorAttachment;
1573 assert(subpass_att < subpass->color_count);
1574 const uint32_t pass_att = subpass->color_attachments[subpass_att].attachment;
1575 if (pass_att == VK_ATTACHMENT_UNUSED)
1576 return;
1577
1578 VkImageLayout image_layout = subpass->color_attachments[subpass_att].layout;
1579 const struct radv_image_view *iview = fb ? fb->attachments[pass_att].attachment : NULL;
1580 VkClearColorValue clear_value = clear_att->clearValue.color;
1581
1582 if (radv_can_fast_clear_color(cmd_buffer, iview, image_layout,
1583 clear_rect, clear_value, view_mask)) {
1584 radv_fast_clear_color(cmd_buffer, iview, clear_att,
1585 subpass_att, pre_flush,
1586 post_flush);
1587 } else {
1588 emit_color_clear(cmd_buffer, clear_att, clear_rect, view_mask);
1589 }
1590 } else {
1591 const uint32_t pass_att = subpass->depth_stencil_attachment->attachment;
1592 if (pass_att == VK_ATTACHMENT_UNUSED)
1593 return;
1594
1595 VkImageLayout image_layout = subpass->depth_stencil_attachment->layout;
1596 const struct radv_image_view *iview = fb ? fb->attachments[pass_att].attachment : NULL;
1597 VkClearDepthStencilValue clear_value = clear_att->clearValue.depthStencil;
1598
1599 assert(aspects & (VK_IMAGE_ASPECT_DEPTH_BIT |
1600 VK_IMAGE_ASPECT_STENCIL_BIT));
1601
1602 if (radv_can_fast_clear_depth(cmd_buffer, iview, image_layout,
1603 aspects, clear_rect, clear_value,
1604 view_mask)) {
1605 radv_fast_clear_depth(cmd_buffer, iview, clear_att,
1606 pre_flush, post_flush);
1607 } else {
1608 emit_depthstencil_clear(cmd_buffer, clear_att, clear_rect,
1609 view_mask);
1610 }
1611 }
1612 }
1613
1614 static inline bool
1615 radv_attachment_needs_clear(struct radv_cmd_state *cmd_state, uint32_t a)
1616 {
1617 uint32_t view_mask = cmd_state->subpass->view_mask;
1618 return (a != VK_ATTACHMENT_UNUSED &&
1619 cmd_state->attachments[a].pending_clear_aspects &&
1620 (!view_mask || (view_mask & ~cmd_state->attachments[a].cleared_views)));
1621 }
1622
1623 static bool
1624 radv_subpass_needs_clear(struct radv_cmd_buffer *cmd_buffer)
1625 {
1626 struct radv_cmd_state *cmd_state = &cmd_buffer->state;
1627 uint32_t a;
1628
1629 if (!cmd_state->subpass)
1630 return false;
1631
1632 for (uint32_t i = 0; i < cmd_state->subpass->color_count; ++i) {
1633 a = cmd_state->subpass->color_attachments[i].attachment;
1634 if (radv_attachment_needs_clear(cmd_state, a))
1635 return true;
1636 }
1637
1638 if (!cmd_state->subpass->depth_stencil_attachment)
1639 return false;
1640
1641 a = cmd_state->subpass->depth_stencil_attachment->attachment;
1642 return radv_attachment_needs_clear(cmd_state, a);
1643 }
1644
1645 static void
1646 radv_subpass_clear_attachment(struct radv_cmd_buffer *cmd_buffer,
1647 struct radv_attachment_state *attachment,
1648 const VkClearAttachment *clear_att,
1649 enum radv_cmd_flush_bits *pre_flush,
1650 enum radv_cmd_flush_bits *post_flush)
1651 {
1652 struct radv_cmd_state *cmd_state = &cmd_buffer->state;
1653 uint32_t view_mask = cmd_state->subpass->view_mask;
1654
1655 VkClearRect clear_rect = {
1656 .rect = cmd_state->render_area,
1657 .baseArrayLayer = 0,
1658 .layerCount = cmd_state->framebuffer->layers,
1659 };
1660
1661 emit_clear(cmd_buffer, clear_att, &clear_rect, pre_flush, post_flush,
1662 view_mask & ~attachment->cleared_views);
1663 if (view_mask)
1664 attachment->cleared_views |= view_mask;
1665 else
1666 attachment->pending_clear_aspects = 0;
1667 }
1668
1669 /**
1670 * Emit any pending attachment clears for the current subpass.
1671 *
1672 * @see radv_attachment_state::pending_clear_aspects
1673 */
1674 void
1675 radv_cmd_buffer_clear_subpass(struct radv_cmd_buffer *cmd_buffer)
1676 {
1677 struct radv_cmd_state *cmd_state = &cmd_buffer->state;
1678 struct radv_meta_saved_state saved_state;
1679 enum radv_cmd_flush_bits pre_flush = 0;
1680 enum radv_cmd_flush_bits post_flush = 0;
1681
1682 if (!radv_subpass_needs_clear(cmd_buffer))
1683 return;
1684
1685 radv_meta_save(&saved_state, cmd_buffer,
1686 RADV_META_SAVE_GRAPHICS_PIPELINE |
1687 RADV_META_SAVE_CONSTANTS);
1688
1689 for (uint32_t i = 0; i < cmd_state->subpass->color_count; ++i) {
1690 uint32_t a = cmd_state->subpass->color_attachments[i].attachment;
1691
1692 if (!radv_attachment_needs_clear(cmd_state, a))
1693 continue;
1694
1695 assert(cmd_state->attachments[a].pending_clear_aspects ==
1696 VK_IMAGE_ASPECT_COLOR_BIT);
1697
1698 VkClearAttachment clear_att = {
1699 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
1700 .colorAttachment = i, /* Use attachment index relative to subpass */
1701 .clearValue = cmd_state->attachments[a].clear_value,
1702 };
1703
1704 radv_subpass_clear_attachment(cmd_buffer,
1705 &cmd_state->attachments[a],
1706 &clear_att, &pre_flush,
1707 &post_flush);
1708 }
1709
1710 if (cmd_state->subpass->depth_stencil_attachment) {
1711 uint32_t ds = cmd_state->subpass->depth_stencil_attachment->attachment;
1712 if (radv_attachment_needs_clear(cmd_state, ds)) {
1713 VkClearAttachment clear_att = {
1714 .aspectMask = cmd_state->attachments[ds].pending_clear_aspects,
1715 .clearValue = cmd_state->attachments[ds].clear_value,
1716 };
1717
1718 radv_subpass_clear_attachment(cmd_buffer,
1719 &cmd_state->attachments[ds],
1720 &clear_att, &pre_flush,
1721 &post_flush);
1722 }
1723 }
1724
1725 radv_meta_restore(&saved_state, cmd_buffer);
1726 cmd_buffer->state.flush_bits |= post_flush;
1727 }
1728
1729 static void
1730 radv_clear_image_layer(struct radv_cmd_buffer *cmd_buffer,
1731 struct radv_image *image,
1732 VkImageLayout image_layout,
1733 const VkImageSubresourceRange *range,
1734 VkFormat format, int level, int layer,
1735 const VkClearValue *clear_val)
1736 {
1737 VkDevice device_h = radv_device_to_handle(cmd_buffer->device);
1738 struct radv_image_view iview;
1739 uint32_t width = radv_minify(image->info.width, range->baseMipLevel + level);
1740 uint32_t height = radv_minify(image->info.height, range->baseMipLevel + level);
1741
1742 radv_image_view_init(&iview, cmd_buffer->device,
1743 &(VkImageViewCreateInfo) {
1744 .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
1745 .image = radv_image_to_handle(image),
1746 .viewType = radv_meta_get_view_type(image),
1747 .format = format,
1748 .subresourceRange = {
1749 .aspectMask = range->aspectMask,
1750 .baseMipLevel = range->baseMipLevel + level,
1751 .levelCount = 1,
1752 .baseArrayLayer = range->baseArrayLayer + layer,
1753 .layerCount = 1
1754 },
1755 });
1756
1757 VkFramebuffer fb;
1758 radv_CreateFramebuffer(device_h,
1759 &(VkFramebufferCreateInfo) {
1760 .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
1761 .attachmentCount = 1,
1762 .pAttachments = (VkImageView[]) {
1763 radv_image_view_to_handle(&iview),
1764 },
1765 .width = width,
1766 .height = height,
1767 .layers = 1
1768 },
1769 &cmd_buffer->pool->alloc,
1770 &fb);
1771
1772 VkAttachmentDescription att_desc = {
1773 .format = iview.vk_format,
1774 .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
1775 .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
1776 .stencilLoadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
1777 .stencilStoreOp = VK_ATTACHMENT_STORE_OP_STORE,
1778 .initialLayout = image_layout,
1779 .finalLayout = image_layout,
1780 };
1781
1782 VkSubpassDescription subpass_desc = {
1783 .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
1784 .inputAttachmentCount = 0,
1785 .colorAttachmentCount = 0,
1786 .pColorAttachments = NULL,
1787 .pResolveAttachments = NULL,
1788 .pDepthStencilAttachment = NULL,
1789 .preserveAttachmentCount = 0,
1790 .pPreserveAttachments = NULL,
1791 };
1792
1793 const VkAttachmentReference att_ref = {
1794 .attachment = 0,
1795 .layout = image_layout,
1796 };
1797
1798 if (range->aspectMask & VK_IMAGE_ASPECT_COLOR_BIT) {
1799 subpass_desc.colorAttachmentCount = 1;
1800 subpass_desc.pColorAttachments = &att_ref;
1801 } else {
1802 subpass_desc.pDepthStencilAttachment = &att_ref;
1803 }
1804
1805 VkRenderPass pass;
1806 radv_CreateRenderPass(device_h,
1807 &(VkRenderPassCreateInfo) {
1808 .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
1809 .attachmentCount = 1,
1810 .pAttachments = &att_desc,
1811 .subpassCount = 1,
1812 .pSubpasses = &subpass_desc,
1813 },
1814 &cmd_buffer->pool->alloc,
1815 &pass);
1816
1817 radv_CmdBeginRenderPass(radv_cmd_buffer_to_handle(cmd_buffer),
1818 &(VkRenderPassBeginInfo) {
1819 .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
1820 .renderArea = {
1821 .offset = { 0, 0, },
1822 .extent = {
1823 .width = width,
1824 .height = height,
1825 },
1826 },
1827 .renderPass = pass,
1828 .framebuffer = fb,
1829 .clearValueCount = 0,
1830 .pClearValues = NULL,
1831 },
1832 VK_SUBPASS_CONTENTS_INLINE);
1833
1834 VkClearAttachment clear_att = {
1835 .aspectMask = range->aspectMask,
1836 .colorAttachment = 0,
1837 .clearValue = *clear_val,
1838 };
1839
1840 VkClearRect clear_rect = {
1841 .rect = {
1842 .offset = { 0, 0 },
1843 .extent = { width, height },
1844 },
1845 .baseArrayLayer = range->baseArrayLayer,
1846 .layerCount = 1, /* FINISHME: clear multi-layer framebuffer */
1847 };
1848
1849 emit_clear(cmd_buffer, &clear_att, &clear_rect, NULL, NULL, 0);
1850
1851 radv_CmdEndRenderPass(radv_cmd_buffer_to_handle(cmd_buffer));
1852 radv_DestroyRenderPass(device_h, pass,
1853 &cmd_buffer->pool->alloc);
1854 radv_DestroyFramebuffer(device_h, fb,
1855 &cmd_buffer->pool->alloc);
1856 }
1857
1858 /**
1859 * Return TRUE if a fast color or depth clear has been performed.
1860 */
1861 static bool
1862 radv_fast_clear_range(struct radv_cmd_buffer *cmd_buffer,
1863 struct radv_image *image,
1864 VkFormat format,
1865 VkImageLayout image_layout,
1866 const VkImageSubresourceRange *range,
1867 const VkClearValue *clear_val)
1868 {
1869 struct radv_image_view iview;
1870
1871 radv_image_view_init(&iview, cmd_buffer->device,
1872 &(VkImageViewCreateInfo) {
1873 .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
1874 .image = radv_image_to_handle(image),
1875 .viewType = radv_meta_get_view_type(image),
1876 .format = image->vk_format,
1877 .subresourceRange = {
1878 .aspectMask = range->aspectMask,
1879 .baseMipLevel = range->baseMipLevel,
1880 .levelCount = range->levelCount,
1881 .baseArrayLayer = range->baseArrayLayer,
1882 .layerCount = range->layerCount,
1883 },
1884 });
1885
1886 VkClearRect clear_rect = {
1887 .rect = {
1888 .offset = { 0, 0 },
1889 .extent = {
1890 radv_minify(image->info.width, range->baseMipLevel),
1891 radv_minify(image->info.height, range->baseMipLevel),
1892 },
1893 },
1894 .baseArrayLayer = range->baseArrayLayer,
1895 .layerCount = range->layerCount,
1896 };
1897
1898 VkClearAttachment clear_att = {
1899 .aspectMask = range->aspectMask,
1900 .colorAttachment = 0,
1901 .clearValue = *clear_val,
1902 };
1903
1904 if (vk_format_is_color(format)) {
1905 if (radv_can_fast_clear_color(cmd_buffer, &iview,
1906 image_layout, &clear_rect,
1907 clear_att.clearValue.color, 0)) {
1908 radv_fast_clear_color(cmd_buffer, &iview, &clear_att,
1909 clear_att.colorAttachment,
1910 NULL, NULL);
1911 return true;
1912 }
1913 } else {
1914 if (radv_can_fast_clear_depth(cmd_buffer, &iview, image_layout,
1915 range->aspectMask, &clear_rect,
1916 clear_att.clearValue.depthStencil, 0)) {
1917 radv_fast_clear_depth(cmd_buffer, &iview, &clear_att,
1918 NULL, NULL);
1919 return true;
1920 }
1921 }
1922
1923 return false;
1924 }
1925
1926 static void
1927 radv_cmd_clear_image(struct radv_cmd_buffer *cmd_buffer,
1928 struct radv_image *image,
1929 VkImageLayout image_layout,
1930 const VkClearValue *clear_value,
1931 uint32_t range_count,
1932 const VkImageSubresourceRange *ranges,
1933 bool cs)
1934 {
1935 VkFormat format = image->vk_format;
1936 VkClearValue internal_clear_value = *clear_value;
1937
1938 if (format == VK_FORMAT_E5B9G9R9_UFLOAT_PACK32) {
1939 uint32_t value;
1940 format = VK_FORMAT_R32_UINT;
1941 value = float3_to_rgb9e5(clear_value->color.float32);
1942 internal_clear_value.color.uint32[0] = value;
1943 }
1944
1945 if (format == VK_FORMAT_R4G4_UNORM_PACK8) {
1946 uint8_t r, g;
1947 format = VK_FORMAT_R8_UINT;
1948 r = float_to_ubyte(clear_value->color.float32[0]) >> 4;
1949 g = float_to_ubyte(clear_value->color.float32[1]) >> 4;
1950 internal_clear_value.color.uint32[0] = (r << 4) | (g & 0xf);
1951 }
1952
1953 if (format == VK_FORMAT_R32G32B32_UINT ||
1954 format == VK_FORMAT_R32G32B32_SINT ||
1955 format == VK_FORMAT_R32G32B32_SFLOAT)
1956 cs = true;
1957
1958 for (uint32_t r = 0; r < range_count; r++) {
1959 const VkImageSubresourceRange *range = &ranges[r];
1960
1961 /* Try to perform a fast clear first, otherwise fallback to
1962 * the legacy path.
1963 */
1964 if (!cs &&
1965 radv_fast_clear_range(cmd_buffer, image, format,
1966 image_layout, range,
1967 &internal_clear_value)) {
1968 continue;
1969 }
1970
1971 for (uint32_t l = 0; l < radv_get_levelCount(image, range); ++l) {
1972 const uint32_t layer_count = image->type == VK_IMAGE_TYPE_3D ?
1973 radv_minify(image->info.depth, range->baseMipLevel + l) :
1974 radv_get_layerCount(image, range);
1975 for (uint32_t s = 0; s < layer_count; ++s) {
1976
1977 if (cs) {
1978 struct radv_meta_blit2d_surf surf;
1979 surf.format = format;
1980 surf.image = image;
1981 surf.level = range->baseMipLevel + l;
1982 surf.layer = range->baseArrayLayer + s;
1983 surf.aspect_mask = range->aspectMask;
1984 radv_meta_clear_image_cs(cmd_buffer, &surf,
1985 &internal_clear_value.color);
1986 } else {
1987 radv_clear_image_layer(cmd_buffer, image, image_layout,
1988 range, format, l, s, &internal_clear_value);
1989 }
1990 }
1991 }
1992 }
1993 }
1994
1995 void radv_CmdClearColorImage(
1996 VkCommandBuffer commandBuffer,
1997 VkImage image_h,
1998 VkImageLayout imageLayout,
1999 const VkClearColorValue* pColor,
2000 uint32_t rangeCount,
2001 const VkImageSubresourceRange* pRanges)
2002 {
2003 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2004 RADV_FROM_HANDLE(radv_image, image, image_h);
2005 struct radv_meta_saved_state saved_state;
2006 bool cs = cmd_buffer->queue_family_index == RADV_QUEUE_COMPUTE;
2007
2008 if (cs) {
2009 radv_meta_save(&saved_state, cmd_buffer,
2010 RADV_META_SAVE_COMPUTE_PIPELINE |
2011 RADV_META_SAVE_CONSTANTS |
2012 RADV_META_SAVE_DESCRIPTORS);
2013 } else {
2014 radv_meta_save(&saved_state, cmd_buffer,
2015 RADV_META_SAVE_GRAPHICS_PIPELINE |
2016 RADV_META_SAVE_CONSTANTS);
2017 }
2018
2019 radv_cmd_clear_image(cmd_buffer, image, imageLayout,
2020 (const VkClearValue *) pColor,
2021 rangeCount, pRanges, cs);
2022
2023 radv_meta_restore(&saved_state, cmd_buffer);
2024 }
2025
2026 void radv_CmdClearDepthStencilImage(
2027 VkCommandBuffer commandBuffer,
2028 VkImage image_h,
2029 VkImageLayout imageLayout,
2030 const VkClearDepthStencilValue* pDepthStencil,
2031 uint32_t rangeCount,
2032 const VkImageSubresourceRange* pRanges)
2033 {
2034 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2035 RADV_FROM_HANDLE(radv_image, image, image_h);
2036 struct radv_meta_saved_state saved_state;
2037
2038 radv_meta_save(&saved_state, cmd_buffer,
2039 RADV_META_SAVE_GRAPHICS_PIPELINE |
2040 RADV_META_SAVE_CONSTANTS);
2041
2042 radv_cmd_clear_image(cmd_buffer, image, imageLayout,
2043 (const VkClearValue *) pDepthStencil,
2044 rangeCount, pRanges, false);
2045
2046 radv_meta_restore(&saved_state, cmd_buffer);
2047 }
2048
2049 void radv_CmdClearAttachments(
2050 VkCommandBuffer commandBuffer,
2051 uint32_t attachmentCount,
2052 const VkClearAttachment* pAttachments,
2053 uint32_t rectCount,
2054 const VkClearRect* pRects)
2055 {
2056 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
2057 struct radv_meta_saved_state saved_state;
2058 enum radv_cmd_flush_bits pre_flush = 0;
2059 enum radv_cmd_flush_bits post_flush = 0;
2060
2061 if (!cmd_buffer->state.subpass)
2062 return;
2063
2064 radv_meta_save(&saved_state, cmd_buffer,
2065 RADV_META_SAVE_GRAPHICS_PIPELINE |
2066 RADV_META_SAVE_CONSTANTS);
2067
2068 /* FINISHME: We can do better than this dumb loop. It thrashes too much
2069 * state.
2070 */
2071 for (uint32_t a = 0; a < attachmentCount; ++a) {
2072 for (uint32_t r = 0; r < rectCount; ++r) {
2073 emit_clear(cmd_buffer, &pAttachments[a], &pRects[r], &pre_flush, &post_flush,
2074 cmd_buffer->state.subpass->view_mask);
2075 }
2076 }
2077
2078 radv_meta_restore(&saved_state, cmd_buffer);
2079 cmd_buffer->state.flush_bits |= post_flush;
2080 }