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
build_color_shaders(struct nir_shader ** out_vs,struct nir_shader ** out_fs,uint32_t frag_output)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_system_value(&vs_b, nir_intrinsic_load_instance_id, 0);
85 nir_ssa_def *base_instance = nir_load_system_value(&vs_b, nir_intrinsic_load_base_instance, 0);
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
create_pipeline(struct radv_device * device,struct radv_render_pass * render_pass,uint32_t samples,struct nir_shader * vs_nir,struct nir_shader * fs_nir,const VkPipelineVertexInputStateCreateInfo * vi_state,const VkPipelineDepthStencilStateCreateInfo * ds_state,const VkPipelineColorBlendStateCreateInfo * cb_state,const VkPipelineLayout layout,const struct radv_graphics_pipeline_create_info * extra,const VkAllocationCallbacks * alloc,VkPipeline * pipeline)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
create_color_renderpass(struct radv_device * device,VkFormat vk_format,uint32_t samples,VkRenderPass * pass)198 create_color_renderpass(struct radv_device *device,
199 VkFormat vk_format,
200 uint32_t samples,
201 VkRenderPass *pass)
202 {
203 return radv_CreateRenderPass(radv_device_to_handle(device),
204 &(VkRenderPassCreateInfo) {
205 .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
206 .attachmentCount = 1,
207 .pAttachments = &(VkAttachmentDescription) {
208 .format = vk_format,
209 .samples = samples,
210 .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
211 .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
212 .initialLayout = VK_IMAGE_LAYOUT_GENERAL,
213 .finalLayout = VK_IMAGE_LAYOUT_GENERAL,
214 },
215 .subpassCount = 1,
216 .pSubpasses = &(VkSubpassDescription) {
217 .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
218 .inputAttachmentCount = 0,
219 .colorAttachmentCount = 1,
220 .pColorAttachments = &(VkAttachmentReference) {
221 .attachment = 0,
222 .layout = VK_IMAGE_LAYOUT_GENERAL,
223 },
224 .pResolveAttachments = NULL,
225 .pDepthStencilAttachment = &(VkAttachmentReference) {
226 .attachment = VK_ATTACHMENT_UNUSED,
227 .layout = VK_IMAGE_LAYOUT_GENERAL,
228 },
229 .preserveAttachmentCount = 1,
230 .pPreserveAttachments = (uint32_t[]) { 0 },
231 },
232 .dependencyCount = 0,
233 }, &device->meta_state.alloc, pass);
234 }
235
236 static VkResult
create_color_pipeline(struct radv_device * device,uint32_t samples,uint32_t frag_output,VkPipeline * pipeline,VkRenderPass pass)237 create_color_pipeline(struct radv_device *device,
238 uint32_t samples,
239 uint32_t frag_output,
240 VkPipeline *pipeline,
241 VkRenderPass pass)
242 {
243 struct nir_shader *vs_nir;
244 struct nir_shader *fs_nir;
245 VkResult result;
246 build_color_shaders(&vs_nir, &fs_nir, frag_output);
247
248 const VkPipelineVertexInputStateCreateInfo vi_state = {
249 .sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
250 .vertexBindingDescriptionCount = 0,
251 .vertexAttributeDescriptionCount = 0,
252 };
253
254 const VkPipelineDepthStencilStateCreateInfo ds_state = {
255 .sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
256 .depthTestEnable = false,
257 .depthWriteEnable = false,
258 .depthBoundsTestEnable = false,
259 .stencilTestEnable = false,
260 };
261
262 VkPipelineColorBlendAttachmentState blend_attachment_state[MAX_RTS] = { 0 };
263 blend_attachment_state[frag_output] = (VkPipelineColorBlendAttachmentState) {
264 .blendEnable = false,
265 .colorWriteMask = VK_COLOR_COMPONENT_A_BIT |
266 VK_COLOR_COMPONENT_R_BIT |
267 VK_COLOR_COMPONENT_G_BIT |
268 VK_COLOR_COMPONENT_B_BIT,
269 };
270
271 const VkPipelineColorBlendStateCreateInfo cb_state = {
272 .sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
273 .logicOpEnable = false,
274 .attachmentCount = MAX_RTS,
275 .pAttachments = blend_attachment_state
276 };
277
278
279 struct radv_graphics_pipeline_create_info extra = {
280 .use_rectlist = true,
281 };
282 result = create_pipeline(device, radv_render_pass_from_handle(pass),
283 samples, vs_nir, fs_nir, &vi_state, &ds_state, &cb_state,
284 device->meta_state.clear_color_p_layout,
285 &extra, &device->meta_state.alloc, pipeline);
286
287 return result;
288 }
289
290 void
radv_device_finish_meta_clear_state(struct radv_device * device)291 radv_device_finish_meta_clear_state(struct radv_device *device)
292 {
293 struct radv_meta_state *state = &device->meta_state;
294
295 for (uint32_t i = 0; i < ARRAY_SIZE(state->clear); ++i) {
296 for (uint32_t j = 0; j < ARRAY_SIZE(state->clear[i].color_pipelines); ++j) {
297 radv_DestroyPipeline(radv_device_to_handle(device),
298 state->clear[i].color_pipelines[j],
299 &state->alloc);
300 radv_DestroyRenderPass(radv_device_to_handle(device),
301 state->clear[i].render_pass[j],
302 &state->alloc);
303 }
304
305 for (uint32_t j = 0; j < NUM_DEPTH_CLEAR_PIPELINES; j++) {
306 radv_DestroyPipeline(radv_device_to_handle(device),
307 state->clear[i].depth_only_pipeline[j],
308 &state->alloc);
309 radv_DestroyPipeline(radv_device_to_handle(device),
310 state->clear[i].stencil_only_pipeline[j],
311 &state->alloc);
312 radv_DestroyPipeline(radv_device_to_handle(device),
313 state->clear[i].depthstencil_pipeline[j],
314 &state->alloc);
315 }
316 radv_DestroyRenderPass(radv_device_to_handle(device),
317 state->clear[i].depthstencil_rp,
318 &state->alloc);
319 }
320 radv_DestroyPipelineLayout(radv_device_to_handle(device),
321 state->clear_color_p_layout,
322 &state->alloc);
323 radv_DestroyPipelineLayout(radv_device_to_handle(device),
324 state->clear_depth_p_layout,
325 &state->alloc);
326 }
327
328 static void
emit_color_clear(struct radv_cmd_buffer * cmd_buffer,const VkClearAttachment * clear_att,const VkClearRect * clear_rect,uint32_t view_mask)329 emit_color_clear(struct radv_cmd_buffer *cmd_buffer,
330 const VkClearAttachment *clear_att,
331 const VkClearRect *clear_rect,
332 uint32_t view_mask)
333 {
334 struct radv_device *device = cmd_buffer->device;
335 const struct radv_subpass *subpass = cmd_buffer->state.subpass;
336 const struct radv_framebuffer *fb = cmd_buffer->state.framebuffer;
337 const uint32_t subpass_att = clear_att->colorAttachment;
338 const uint32_t pass_att = subpass->color_attachments[subpass_att].attachment;
339 const struct radv_image_view *iview = fb->attachments[pass_att].attachment;
340 const uint32_t samples = iview->image->info.samples;
341 const uint32_t samples_log2 = ffs(samples) - 1;
342 unsigned fs_key = radv_format_meta_fs_key(iview->vk_format);
343 VkClearColorValue clear_value = clear_att->clearValue.color;
344 VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
345 VkPipeline pipeline;
346
347 if (fs_key == -1) {
348 radv_finishme("color clears incomplete");
349 return;
350 }
351
352 pipeline = device->meta_state.clear[samples_log2].color_pipelines[fs_key];
353 if (!pipeline) {
354 radv_finishme("color clears incomplete");
355 return;
356 }
357 assert(samples_log2 < ARRAY_SIZE(device->meta_state.clear));
358 assert(pipeline);
359 assert(clear_att->aspectMask == VK_IMAGE_ASPECT_COLOR_BIT);
360 assert(clear_att->colorAttachment < subpass->color_count);
361
362 radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
363 device->meta_state.clear_color_p_layout,
364 VK_SHADER_STAGE_FRAGMENT_BIT, 0, 16,
365 &clear_value);
366
367 struct radv_subpass clear_subpass = {
368 .color_count = 1,
369 .color_attachments = (VkAttachmentReference[]) {
370 subpass->color_attachments[clear_att->colorAttachment]
371 },
372 .depth_stencil_attachment = (VkAttachmentReference) { VK_ATTACHMENT_UNUSED, VK_IMAGE_LAYOUT_UNDEFINED }
373 };
374
375 radv_cmd_buffer_set_subpass(cmd_buffer, &clear_subpass, false);
376
377 radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS,
378 pipeline);
379
380 radv_CmdSetViewport(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1, &(VkViewport) {
381 .x = clear_rect->rect.offset.x,
382 .y = clear_rect->rect.offset.y,
383 .width = clear_rect->rect.extent.width,
384 .height = clear_rect->rect.extent.height,
385 .minDepth = 0.0f,
386 .maxDepth = 1.0f
387 });
388
389 radv_CmdSetScissor(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1, &clear_rect->rect);
390
391 if (view_mask) {
392 unsigned i;
393 for_each_bit(i, view_mask)
394 radv_CmdDraw(cmd_buffer_h, 3, 1, 0, i);
395 } else {
396 radv_CmdDraw(cmd_buffer_h, 3, clear_rect->layerCount, 0, clear_rect->baseArrayLayer);
397 }
398
399 radv_cmd_buffer_set_subpass(cmd_buffer, subpass, false);
400 }
401
402
403 static void
build_depthstencil_shader(struct nir_shader ** out_vs,struct nir_shader ** out_fs)404 build_depthstencil_shader(struct nir_shader **out_vs, struct nir_shader **out_fs)
405 {
406 nir_builder vs_b, fs_b;
407
408 nir_builder_init_simple_shader(&vs_b, NULL, MESA_SHADER_VERTEX, NULL);
409 nir_builder_init_simple_shader(&fs_b, NULL, MESA_SHADER_FRAGMENT, NULL);
410
411 vs_b.shader->info.name = ralloc_strdup(vs_b.shader, "meta_clear_depthstencil_vs");
412 fs_b.shader->info.name = ralloc_strdup(fs_b.shader, "meta_clear_depthstencil_fs");
413 const struct glsl_type *position_out_type = glsl_vec4_type();
414
415 nir_variable *vs_out_pos =
416 nir_variable_create(vs_b.shader, nir_var_shader_out, position_out_type,
417 "gl_Position");
418 vs_out_pos->data.location = VARYING_SLOT_POS;
419
420 nir_intrinsic_instr *in_color_load = nir_intrinsic_instr_create(vs_b.shader, nir_intrinsic_load_push_constant);
421 nir_intrinsic_set_base(in_color_load, 0);
422 nir_intrinsic_set_range(in_color_load, 4);
423 in_color_load->src[0] = nir_src_for_ssa(nir_imm_int(&vs_b, 0));
424 in_color_load->num_components = 1;
425 nir_ssa_dest_init(&in_color_load->instr, &in_color_load->dest, 1, 32, "depth value");
426 nir_builder_instr_insert(&vs_b, &in_color_load->instr);
427
428 nir_ssa_def *outvec = radv_meta_gen_rect_vertices_comp2(&vs_b, &in_color_load->dest.ssa);
429 nir_store_var(&vs_b, vs_out_pos, outvec, 0xf);
430
431 const struct glsl_type *layer_type = glsl_int_type();
432 nir_variable *vs_out_layer =
433 nir_variable_create(vs_b.shader, nir_var_shader_out, layer_type,
434 "v_layer");
435 vs_out_layer->data.location = VARYING_SLOT_LAYER;
436 vs_out_layer->data.interpolation = INTERP_MODE_FLAT;
437 nir_ssa_def *inst_id = nir_load_system_value(&vs_b, nir_intrinsic_load_instance_id, 0);
438 nir_ssa_def *base_instance = nir_load_system_value(&vs_b, nir_intrinsic_load_base_instance, 0);
439
440 nir_ssa_def *layer_id = nir_iadd(&vs_b, inst_id, base_instance);
441 nir_store_var(&vs_b, vs_out_layer, layer_id, 0x1);
442
443 *out_vs = vs_b.shader;
444 *out_fs = fs_b.shader;
445 }
446
447 static VkResult
create_depthstencil_renderpass(struct radv_device * device,uint32_t samples,VkRenderPass * render_pass)448 create_depthstencil_renderpass(struct radv_device *device,
449 uint32_t samples,
450 VkRenderPass *render_pass)
451 {
452 return radv_CreateRenderPass(radv_device_to_handle(device),
453 &(VkRenderPassCreateInfo) {
454 .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
455 .attachmentCount = 1,
456 .pAttachments = &(VkAttachmentDescription) {
457 .format = VK_FORMAT_D32_SFLOAT_S8_UINT,
458 .samples = samples,
459 .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
460 .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
461 .initialLayout = VK_IMAGE_LAYOUT_GENERAL,
462 .finalLayout = VK_IMAGE_LAYOUT_GENERAL,
463 },
464 .subpassCount = 1,
465 .pSubpasses = &(VkSubpassDescription) {
466 .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
467 .inputAttachmentCount = 0,
468 .colorAttachmentCount = 0,
469 .pColorAttachments = NULL,
470 .pResolveAttachments = NULL,
471 .pDepthStencilAttachment = &(VkAttachmentReference) {
472 .attachment = 0,
473 .layout = VK_IMAGE_LAYOUT_GENERAL,
474 },
475 .preserveAttachmentCount = 1,
476 .pPreserveAttachments = (uint32_t[]) { 0 },
477 },
478 .dependencyCount = 0,
479 }, &device->meta_state.alloc, render_pass);
480 }
481
482 static VkResult
create_depthstencil_pipeline(struct radv_device * device,VkImageAspectFlags aspects,uint32_t samples,int index,VkPipeline * pipeline,VkRenderPass render_pass)483 create_depthstencil_pipeline(struct radv_device *device,
484 VkImageAspectFlags aspects,
485 uint32_t samples,
486 int index,
487 VkPipeline *pipeline,
488 VkRenderPass render_pass)
489 {
490 struct nir_shader *vs_nir, *fs_nir;
491 VkResult result;
492 build_depthstencil_shader(&vs_nir, &fs_nir);
493
494 const VkPipelineVertexInputStateCreateInfo vi_state = {
495 .sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
496 .vertexBindingDescriptionCount = 0,
497 .vertexAttributeDescriptionCount = 0,
498 };
499
500 const VkPipelineDepthStencilStateCreateInfo ds_state = {
501 .sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO,
502 .depthTestEnable = (aspects & VK_IMAGE_ASPECT_DEPTH_BIT),
503 .depthCompareOp = VK_COMPARE_OP_ALWAYS,
504 .depthWriteEnable = (aspects & VK_IMAGE_ASPECT_DEPTH_BIT),
505 .depthBoundsTestEnable = false,
506 .stencilTestEnable = (aspects & VK_IMAGE_ASPECT_STENCIL_BIT),
507 .front = {
508 .passOp = VK_STENCIL_OP_REPLACE,
509 .compareOp = VK_COMPARE_OP_ALWAYS,
510 .writeMask = UINT32_MAX,
511 .reference = 0, /* dynamic */
512 },
513 .back = { 0 /* dont care */ },
514 };
515
516 const VkPipelineColorBlendStateCreateInfo cb_state = {
517 .sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
518 .logicOpEnable = false,
519 .attachmentCount = 0,
520 .pAttachments = NULL,
521 };
522
523 struct radv_graphics_pipeline_create_info extra = {
524 .use_rectlist = true,
525 };
526
527 if (aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
528 extra.db_depth_clear = index == DEPTH_CLEAR_SLOW ? false : true;
529 extra.db_depth_disable_expclear = index == DEPTH_CLEAR_FAST_NO_EXPCLEAR ? true : false;
530 }
531 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
532 extra.db_stencil_clear = index == DEPTH_CLEAR_SLOW ? false : true;
533 extra.db_stencil_disable_expclear = index == DEPTH_CLEAR_FAST_NO_EXPCLEAR ? true : false;
534 }
535 result = create_pipeline(device, radv_render_pass_from_handle(render_pass),
536 samples, vs_nir, fs_nir, &vi_state, &ds_state, &cb_state,
537 device->meta_state.clear_depth_p_layout,
538 &extra, &device->meta_state.alloc, pipeline);
539 return result;
540 }
541
depth_view_can_fast_clear(struct radv_cmd_buffer * cmd_buffer,const struct radv_image_view * iview,VkImageAspectFlags aspects,VkImageLayout layout,const VkClearRect * clear_rect,VkClearDepthStencilValue clear_value)542 static bool depth_view_can_fast_clear(struct radv_cmd_buffer *cmd_buffer,
543 const struct radv_image_view *iview,
544 VkImageAspectFlags aspects,
545 VkImageLayout layout,
546 const VkClearRect *clear_rect,
547 VkClearDepthStencilValue clear_value)
548 {
549 uint32_t queue_mask = radv_image_queue_family_mask(iview->image,
550 cmd_buffer->queue_family_index,
551 cmd_buffer->queue_family_index);
552 if (clear_rect->rect.offset.x || clear_rect->rect.offset.y ||
553 clear_rect->rect.extent.width != iview->extent.width ||
554 clear_rect->rect.extent.height != iview->extent.height)
555 return false;
556 if (iview->image->tc_compatible_htile &&
557 (((aspects & VK_IMAGE_ASPECT_DEPTH_BIT) && clear_value.depth != 0.0 &&
558 clear_value.depth != 1.0) ||
559 ((aspects & VK_IMAGE_ASPECT_STENCIL_BIT) && clear_value.stencil != 0)))
560 return false;
561 if (iview->image->surface.htile_size &&
562 iview->base_mip == 0 &&
563 iview->base_layer == 0 &&
564 radv_layout_is_htile_compressed(iview->image, layout, queue_mask) &&
565 !radv_image_extent_compare(iview->image, &iview->extent))
566 return true;
567 return false;
568 }
569
570 static VkPipeline
pick_depthstencil_pipeline(struct radv_cmd_buffer * cmd_buffer,struct radv_meta_state * meta_state,const struct radv_image_view * iview,int samples_log2,VkImageAspectFlags aspects,VkImageLayout layout,const VkClearRect * clear_rect,VkClearDepthStencilValue clear_value)571 pick_depthstencil_pipeline(struct radv_cmd_buffer *cmd_buffer,
572 struct radv_meta_state *meta_state,
573 const struct radv_image_view *iview,
574 int samples_log2,
575 VkImageAspectFlags aspects,
576 VkImageLayout layout,
577 const VkClearRect *clear_rect,
578 VkClearDepthStencilValue clear_value)
579 {
580 bool fast = depth_view_can_fast_clear(cmd_buffer, iview, aspects, layout, clear_rect, clear_value);
581 int index = DEPTH_CLEAR_SLOW;
582
583 if (fast) {
584 /* we don't know the previous clear values, so we always have
585 * the NO_EXPCLEAR path */
586 index = DEPTH_CLEAR_FAST_NO_EXPCLEAR;
587 }
588
589 switch (aspects) {
590 case VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT:
591 return meta_state->clear[samples_log2].depthstencil_pipeline[index];
592 case VK_IMAGE_ASPECT_DEPTH_BIT:
593 return meta_state->clear[samples_log2].depth_only_pipeline[index];
594 case VK_IMAGE_ASPECT_STENCIL_BIT:
595 return meta_state->clear[samples_log2].stencil_only_pipeline[index];
596 }
597 unreachable("expected depth or stencil aspect");
598 }
599
600 static void
emit_depthstencil_clear(struct radv_cmd_buffer * cmd_buffer,const VkClearAttachment * clear_att,const VkClearRect * clear_rect)601 emit_depthstencil_clear(struct radv_cmd_buffer *cmd_buffer,
602 const VkClearAttachment *clear_att,
603 const VkClearRect *clear_rect)
604 {
605 struct radv_device *device = cmd_buffer->device;
606 struct radv_meta_state *meta_state = &device->meta_state;
607 const struct radv_subpass *subpass = cmd_buffer->state.subpass;
608 const struct radv_framebuffer *fb = cmd_buffer->state.framebuffer;
609 const uint32_t pass_att = subpass->depth_stencil_attachment.attachment;
610 VkClearDepthStencilValue clear_value = clear_att->clearValue.depthStencil;
611 VkImageAspectFlags aspects = clear_att->aspectMask;
612 const struct radv_image_view *iview = fb->attachments[pass_att].attachment;
613 const uint32_t samples = iview->image->info.samples;
614 const uint32_t samples_log2 = ffs(samples) - 1;
615 VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
616
617 assert(pass_att != VK_ATTACHMENT_UNUSED);
618
619 if (!(aspects & VK_IMAGE_ASPECT_DEPTH_BIT))
620 clear_value.depth = 1.0f;
621
622 radv_CmdPushConstants(radv_cmd_buffer_to_handle(cmd_buffer),
623 device->meta_state.clear_depth_p_layout,
624 VK_SHADER_STAGE_VERTEX_BIT, 0, 4,
625 &clear_value.depth);
626
627 uint32_t prev_reference = cmd_buffer->state.dynamic.stencil_reference.front;
628 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
629 radv_CmdSetStencilReference(cmd_buffer_h, VK_STENCIL_FACE_FRONT_BIT,
630 clear_value.stencil);
631 }
632
633 VkPipeline pipeline = pick_depthstencil_pipeline(cmd_buffer,
634 meta_state,
635 iview,
636 samples_log2,
637 aspects,
638 subpass->depth_stencil_attachment.layout,
639 clear_rect,
640 clear_value);
641
642 radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS,
643 pipeline);
644
645 if (depth_view_can_fast_clear(cmd_buffer, iview, aspects,
646 subpass->depth_stencil_attachment.layout,
647 clear_rect, clear_value))
648 radv_set_depth_clear_regs(cmd_buffer, iview->image, clear_value, aspects);
649
650 radv_CmdSetViewport(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1, &(VkViewport) {
651 .x = clear_rect->rect.offset.x,
652 .y = clear_rect->rect.offset.y,
653 .width = clear_rect->rect.extent.width,
654 .height = clear_rect->rect.extent.height,
655 .minDepth = 0.0f,
656 .maxDepth = 1.0f
657 });
658
659 radv_CmdSetScissor(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1, &clear_rect->rect);
660
661 radv_CmdDraw(cmd_buffer_h, 3, clear_rect->layerCount, 0, clear_rect->baseArrayLayer);
662
663 if (aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
664 radv_CmdSetStencilReference(cmd_buffer_h, VK_STENCIL_FACE_FRONT_BIT,
665 prev_reference);
666 }
667 }
668
669 static bool
emit_fast_htile_clear(struct radv_cmd_buffer * cmd_buffer,const VkClearAttachment * clear_att,const VkClearRect * clear_rect,enum radv_cmd_flush_bits * pre_flush,enum radv_cmd_flush_bits * post_flush)670 emit_fast_htile_clear(struct radv_cmd_buffer *cmd_buffer,
671 const VkClearAttachment *clear_att,
672 const VkClearRect *clear_rect,
673 enum radv_cmd_flush_bits *pre_flush,
674 enum radv_cmd_flush_bits *post_flush)
675 {
676 const struct radv_subpass *subpass = cmd_buffer->state.subpass;
677 const uint32_t pass_att = subpass->depth_stencil_attachment.attachment;
678 VkImageLayout image_layout = subpass->depth_stencil_attachment.layout;
679 const struct radv_framebuffer *fb = cmd_buffer->state.framebuffer;
680 const struct radv_image_view *iview = fb->attachments[pass_att].attachment;
681 VkClearDepthStencilValue clear_value = clear_att->clearValue.depthStencil;
682 VkImageAspectFlags aspects = clear_att->aspectMask;
683 uint32_t clear_word, flush_bits;
684
685 if (!iview->image->surface.htile_size)
686 return false;
687
688 if (cmd_buffer->device->instance->debug_flags & RADV_DEBUG_NO_FAST_CLEARS)
689 return false;
690
691 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)))
692 goto fail;
693
694 /* don't fast clear 3D */
695 if (iview->image->type == VK_IMAGE_TYPE_3D)
696 goto fail;
697
698 /* all layers are bound */
699 if (iview->base_layer > 0)
700 goto fail;
701 if (iview->image->info.array_size != iview->layer_count)
702 goto fail;
703
704 if (!radv_image_extent_compare(iview->image, &iview->extent))
705 goto fail;
706
707 if (clear_rect->rect.offset.x || clear_rect->rect.offset.y ||
708 clear_rect->rect.extent.width != iview->image->info.width ||
709 clear_rect->rect.extent.height != iview->image->info.height)
710 goto fail;
711
712 if (clear_rect->baseArrayLayer != 0)
713 goto fail;
714 if (clear_rect->layerCount != iview->image->info.array_size)
715 goto fail;
716
717 if ((clear_value.depth != 0.0 && clear_value.depth != 1.0) || !(aspects & VK_IMAGE_ASPECT_DEPTH_BIT))
718 goto fail;
719
720 if (vk_format_aspects(iview->image->vk_format) & VK_IMAGE_ASPECT_STENCIL_BIT) {
721 if (clear_value.stencil != 0 || !(aspects & VK_IMAGE_ASPECT_STENCIL_BIT))
722 goto fail;
723 clear_word = clear_value.depth ? 0xfffc0000 : 0;
724 } else
725 clear_word = clear_value.depth ? 0xfffffff0 : 0;
726
727 if (pre_flush) {
728 cmd_buffer->state.flush_bits |= (RADV_CMD_FLAG_FLUSH_AND_INV_DB |
729 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META) & ~ *pre_flush;
730 *pre_flush |= cmd_buffer->state.flush_bits;
731 } else
732 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_DB |
733 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META;
734
735 flush_bits = radv_fill_buffer(cmd_buffer, iview->image->bo,
736 iview->image->offset + iview->image->htile_offset,
737 iview->image->surface.htile_size, clear_word);
738
739 radv_set_depth_clear_regs(cmd_buffer, iview->image, clear_value, aspects);
740 if (post_flush) {
741 *post_flush |= flush_bits;
742 } else {
743 cmd_buffer->state.flush_bits |= flush_bits;
744 }
745
746 return true;
747 fail:
748 return false;
749 }
750
751 static VkFormat pipeline_formats[] = {
752 VK_FORMAT_R8G8B8A8_UNORM,
753 VK_FORMAT_R8G8B8A8_UINT,
754 VK_FORMAT_R8G8B8A8_SINT,
755 VK_FORMAT_A2R10G10B10_UINT_PACK32,
756 VK_FORMAT_A2R10G10B10_SINT_PACK32,
757 VK_FORMAT_R16G16B16A16_UNORM,
758 VK_FORMAT_R16G16B16A16_SNORM,
759 VK_FORMAT_R16G16B16A16_UINT,
760 VK_FORMAT_R16G16B16A16_SINT,
761 VK_FORMAT_R32_SFLOAT,
762 VK_FORMAT_R32G32_SFLOAT,
763 VK_FORMAT_R32G32B32A32_SFLOAT
764 };
765
766 VkResult
radv_device_init_meta_clear_state(struct radv_device * device)767 radv_device_init_meta_clear_state(struct radv_device *device)
768 {
769 VkResult res;
770 struct radv_meta_state *state = &device->meta_state;
771
772 VkPipelineLayoutCreateInfo pl_color_create_info = {
773 .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
774 .setLayoutCount = 0,
775 .pushConstantRangeCount = 1,
776 .pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_FRAGMENT_BIT, 0, 16},
777 };
778
779 res = radv_CreatePipelineLayout(radv_device_to_handle(device),
780 &pl_color_create_info,
781 &device->meta_state.alloc,
782 &device->meta_state.clear_color_p_layout);
783 if (res != VK_SUCCESS)
784 goto fail;
785
786 VkPipelineLayoutCreateInfo pl_depth_create_info = {
787 .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
788 .setLayoutCount = 0,
789 .pushConstantRangeCount = 1,
790 .pPushConstantRanges = &(VkPushConstantRange){VK_SHADER_STAGE_VERTEX_BIT, 0, 4},
791 };
792
793 res = radv_CreatePipelineLayout(radv_device_to_handle(device),
794 &pl_depth_create_info,
795 &device->meta_state.alloc,
796 &device->meta_state.clear_depth_p_layout);
797 if (res != VK_SUCCESS)
798 goto fail;
799
800 for (uint32_t i = 0; i < ARRAY_SIZE(state->clear); ++i) {
801 uint32_t samples = 1 << i;
802 for (uint32_t j = 0; j < ARRAY_SIZE(pipeline_formats); ++j) {
803 VkFormat format = pipeline_formats[j];
804 unsigned fs_key = radv_format_meta_fs_key(format);
805 assert(!state->clear[i].color_pipelines[fs_key]);
806
807 res = create_color_renderpass(device, format, samples,
808 &state->clear[i].render_pass[fs_key]);
809 if (res != VK_SUCCESS)
810 goto fail;
811
812 res = create_color_pipeline(device, samples, 0, &state->clear[i].color_pipelines[fs_key],
813 state->clear[i].render_pass[fs_key]);
814 if (res != VK_SUCCESS)
815 goto fail;
816
817 }
818
819 res = create_depthstencil_renderpass(device,
820 samples,
821 &state->clear[i].depthstencil_rp);
822 if (res != VK_SUCCESS)
823 goto fail;
824
825 for (uint32_t j = 0; j < NUM_DEPTH_CLEAR_PIPELINES; j++) {
826 res = create_depthstencil_pipeline(device,
827 VK_IMAGE_ASPECT_DEPTH_BIT,
828 samples,
829 j,
830 &state->clear[i].depth_only_pipeline[j],
831 state->clear[i].depthstencil_rp);
832 if (res != VK_SUCCESS)
833 goto fail;
834
835 res = create_depthstencil_pipeline(device,
836 VK_IMAGE_ASPECT_STENCIL_BIT,
837 samples,
838 j,
839 &state->clear[i].stencil_only_pipeline[j],
840 state->clear[i].depthstencil_rp);
841 if (res != VK_SUCCESS)
842 goto fail;
843
844 res = create_depthstencil_pipeline(device,
845 VK_IMAGE_ASPECT_DEPTH_BIT |
846 VK_IMAGE_ASPECT_STENCIL_BIT,
847 samples,
848 j,
849 &state->clear[i].depthstencil_pipeline[j],
850 state->clear[i].depthstencil_rp);
851 if (res != VK_SUCCESS)
852 goto fail;
853 }
854 }
855 return VK_SUCCESS;
856
857 fail:
858 radv_device_finish_meta_clear_state(device);
859 return res;
860 }
861
vi_get_fast_clear_parameters(VkFormat format,const VkClearColorValue * clear_value,uint32_t * reset_value,bool * can_avoid_fast_clear_elim)862 static void vi_get_fast_clear_parameters(VkFormat format,
863 const VkClearColorValue *clear_value,
864 uint32_t* reset_value,
865 bool *can_avoid_fast_clear_elim)
866 {
867 bool values[4] = {};
868 int extra_channel;
869 bool main_value = false;
870 bool extra_value = false;
871 int i;
872 *can_avoid_fast_clear_elim = false;
873
874 *reset_value = 0x20202020U;
875
876 const struct vk_format_description *desc = vk_format_description(format);
877 if (format == VK_FORMAT_B10G11R11_UFLOAT_PACK32 ||
878 format == VK_FORMAT_R5G6B5_UNORM_PACK16 ||
879 format == VK_FORMAT_B5G6R5_UNORM_PACK16)
880 extra_channel = -1;
881 else if (desc->layout == VK_FORMAT_LAYOUT_PLAIN) {
882 if (radv_translate_colorswap(format, false) <= 1)
883 extra_channel = desc->nr_channels - 1;
884 else
885 extra_channel = 0;
886 } else
887 return;
888
889 for (i = 0; i < 4; i++) {
890 int index = desc->swizzle[i] - VK_SWIZZLE_X;
891 if (desc->swizzle[i] < VK_SWIZZLE_X ||
892 desc->swizzle[i] > VK_SWIZZLE_W)
893 continue;
894
895 if (desc->channel[i].pure_integer &&
896 desc->channel[i].type == VK_FORMAT_TYPE_SIGNED) {
897 /* Use the maximum value for clamping the clear color. */
898 int max = u_bit_consecutive(0, desc->channel[i].size - 1);
899
900 values[i] = clear_value->int32[i] != 0;
901 if (clear_value->int32[i] != 0 && MIN2(clear_value->int32[i], max) != max)
902 return;
903 } else if (desc->channel[i].pure_integer &&
904 desc->channel[i].type == VK_FORMAT_TYPE_UNSIGNED) {
905 /* Use the maximum value for clamping the clear color. */
906 unsigned max = u_bit_consecutive(0, desc->channel[i].size);
907
908 values[i] = clear_value->uint32[i] != 0U;
909 if (clear_value->uint32[i] != 0U && MIN2(clear_value->uint32[i], max) != max)
910 return;
911 } else {
912 values[i] = clear_value->float32[i] != 0.0F;
913 if (clear_value->float32[i] != 0.0F && clear_value->float32[i] != 1.0F)
914 return;
915 }
916
917 if (index == extra_channel)
918 extra_value = values[i];
919 else
920 main_value = values[i];
921 }
922
923 for (int i = 0; i < 4; ++i)
924 if (values[i] != main_value &&
925 desc->swizzle[i] - VK_SWIZZLE_X != extra_channel &&
926 desc->swizzle[i] >= VK_SWIZZLE_X &&
927 desc->swizzle[i] <= VK_SWIZZLE_W)
928 return;
929
930 *can_avoid_fast_clear_elim = true;
931 if (main_value)
932 *reset_value |= 0x80808080U;
933
934 if (extra_value)
935 *reset_value |= 0x40404040U;
936 return;
937 }
938
939 static bool
emit_fast_color_clear(struct radv_cmd_buffer * cmd_buffer,const VkClearAttachment * clear_att,const VkClearRect * clear_rect,enum radv_cmd_flush_bits * pre_flush,enum radv_cmd_flush_bits * post_flush,uint32_t view_mask)940 emit_fast_color_clear(struct radv_cmd_buffer *cmd_buffer,
941 const VkClearAttachment *clear_att,
942 const VkClearRect *clear_rect,
943 enum radv_cmd_flush_bits *pre_flush,
944 enum radv_cmd_flush_bits *post_flush,
945 uint32_t view_mask)
946 {
947 const struct radv_subpass *subpass = cmd_buffer->state.subpass;
948 const uint32_t subpass_att = clear_att->colorAttachment;
949 const uint32_t pass_att = subpass->color_attachments[subpass_att].attachment;
950 VkImageLayout image_layout = subpass->color_attachments[subpass_att].layout;
951 const struct radv_framebuffer *fb = cmd_buffer->state.framebuffer;
952 const struct radv_image_view *iview = fb->attachments[pass_att].attachment;
953 VkClearColorValue clear_value = clear_att->clearValue.color;
954 uint32_t clear_color[2], flush_bits;
955 bool ret;
956
957 if (!iview->image->cmask.size && !iview->image->surface.dcc_size)
958 return false;
959
960 if (cmd_buffer->device->instance->debug_flags & RADV_DEBUG_NO_FAST_CLEARS)
961 return false;
962
963 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)))
964 goto fail;
965
966 /* don't fast clear 3D */
967 if (iview->image->type == VK_IMAGE_TYPE_3D)
968 goto fail;
969
970 /* all layers are bound */
971 if (iview->base_layer > 0)
972 goto fail;
973 if (iview->image->info.array_size != iview->layer_count)
974 goto fail;
975
976 if (iview->image->info.levels > 1)
977 goto fail;
978
979 if (iview->image->surface.is_linear)
980 goto fail;
981 if (!radv_image_extent_compare(iview->image, &iview->extent))
982 goto fail;
983
984 if (clear_rect->rect.offset.x || clear_rect->rect.offset.y ||
985 clear_rect->rect.extent.width != iview->image->info.width ||
986 clear_rect->rect.extent.height != iview->image->info.height)
987 goto fail;
988
989 if (view_mask && (iview->image->info.array_size >= 32 ||
990 (1u << iview->image->info.array_size) - 1u != view_mask))
991 goto fail;
992 if (!view_mask && clear_rect->baseArrayLayer != 0)
993 goto fail;
994 if (!view_mask && clear_rect->layerCount != iview->image->info.array_size)
995 goto fail;
996
997 /* RB+ doesn't work with CMASK fast clear on Stoney. */
998 if (!iview->image->surface.dcc_size &&
999 cmd_buffer->device->physical_device->rad_info.family == CHIP_STONEY)
1000 goto fail;
1001
1002 /* DCC */
1003 ret = radv_format_pack_clear_color(iview->image->vk_format,
1004 clear_color, &clear_value);
1005 if (ret == false)
1006 goto fail;
1007
1008 if (pre_flush) {
1009 cmd_buffer->state.flush_bits |= (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
1010 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META) & ~ *pre_flush;
1011 *pre_flush |= cmd_buffer->state.flush_bits;
1012 } else
1013 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB |
1014 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META;
1015 /* clear cmask buffer */
1016 if (iview->image->surface.dcc_size) {
1017 uint32_t reset_value;
1018 bool can_avoid_fast_clear_elim;
1019 vi_get_fast_clear_parameters(iview->image->vk_format,
1020 &clear_value, &reset_value,
1021 &can_avoid_fast_clear_elim);
1022
1023 flush_bits = radv_fill_buffer(cmd_buffer, iview->image->bo,
1024 iview->image->offset + iview->image->dcc_offset,
1025 iview->image->surface.dcc_size, reset_value);
1026 radv_set_dcc_need_cmask_elim_pred(cmd_buffer, iview->image,
1027 !can_avoid_fast_clear_elim);
1028 } else {
1029 flush_bits = radv_fill_buffer(cmd_buffer, iview->image->bo,
1030 iview->image->offset + iview->image->cmask.offset,
1031 iview->image->cmask.size, 0);
1032 }
1033
1034 if (post_flush) {
1035 *post_flush |= flush_bits;
1036 } else {
1037 cmd_buffer->state.flush_bits |= flush_bits;
1038 }
1039
1040 radv_set_color_clear_regs(cmd_buffer, iview->image, subpass_att, clear_color);
1041
1042 return true;
1043 fail:
1044 return false;
1045 }
1046
1047 /**
1048 * The parameters mean that same as those in vkCmdClearAttachments.
1049 */
1050 static void
emit_clear(struct radv_cmd_buffer * cmd_buffer,const VkClearAttachment * clear_att,const VkClearRect * clear_rect,enum radv_cmd_flush_bits * pre_flush,enum radv_cmd_flush_bits * post_flush,uint32_t view_mask)1051 emit_clear(struct radv_cmd_buffer *cmd_buffer,
1052 const VkClearAttachment *clear_att,
1053 const VkClearRect *clear_rect,
1054 enum radv_cmd_flush_bits *pre_flush,
1055 enum radv_cmd_flush_bits *post_flush,
1056 uint32_t view_mask)
1057 {
1058 if (clear_att->aspectMask & VK_IMAGE_ASPECT_COLOR_BIT) {
1059 if (!emit_fast_color_clear(cmd_buffer, clear_att, clear_rect,
1060 pre_flush, post_flush, view_mask))
1061 emit_color_clear(cmd_buffer, clear_att, clear_rect, view_mask);
1062 } else {
1063 assert(clear_att->aspectMask & (VK_IMAGE_ASPECT_DEPTH_BIT |
1064 VK_IMAGE_ASPECT_STENCIL_BIT));
1065 if (!emit_fast_htile_clear(cmd_buffer, clear_att, clear_rect,
1066 pre_flush, post_flush))
1067 emit_depthstencil_clear(cmd_buffer, clear_att, clear_rect);
1068 }
1069 }
1070
1071 static inline bool
radv_attachment_needs_clear(struct radv_cmd_state * cmd_state,uint32_t a)1072 radv_attachment_needs_clear(struct radv_cmd_state *cmd_state, uint32_t a)
1073 {
1074 uint32_t view_mask = cmd_state->subpass->view_mask;
1075 return (a != VK_ATTACHMENT_UNUSED &&
1076 cmd_state->attachments[a].pending_clear_aspects &&
1077 (!view_mask || (view_mask & ~cmd_state->attachments[a].cleared_views)));
1078 }
1079
1080 static bool
radv_subpass_needs_clear(struct radv_cmd_buffer * cmd_buffer)1081 radv_subpass_needs_clear(struct radv_cmd_buffer *cmd_buffer)
1082 {
1083 struct radv_cmd_state *cmd_state = &cmd_buffer->state;
1084 uint32_t a;
1085
1086 if (!cmd_state->subpass)
1087 return false;
1088
1089 for (uint32_t i = 0; i < cmd_state->subpass->color_count; ++i) {
1090 a = cmd_state->subpass->color_attachments[i].attachment;
1091 if (radv_attachment_needs_clear(cmd_state, a))
1092 return true;
1093 }
1094
1095 a = cmd_state->subpass->depth_stencil_attachment.attachment;
1096 return radv_attachment_needs_clear(cmd_state, a);
1097 }
1098
1099 static void
radv_subpass_clear_attachment(struct radv_cmd_buffer * cmd_buffer,struct radv_attachment_state * attachment,const VkClearAttachment * clear_att,enum radv_cmd_flush_bits * pre_flush,enum radv_cmd_flush_bits * post_flush)1100 radv_subpass_clear_attachment(struct radv_cmd_buffer *cmd_buffer,
1101 struct radv_attachment_state *attachment,
1102 const VkClearAttachment *clear_att,
1103 enum radv_cmd_flush_bits *pre_flush,
1104 enum radv_cmd_flush_bits *post_flush)
1105 {
1106 struct radv_cmd_state *cmd_state = &cmd_buffer->state;
1107 uint32_t view_mask = cmd_state->subpass->view_mask;
1108
1109 VkClearRect clear_rect = {
1110 .rect = cmd_state->render_area,
1111 .baseArrayLayer = 0,
1112 .layerCount = cmd_state->framebuffer->layers,
1113 };
1114
1115 emit_clear(cmd_buffer, clear_att, &clear_rect, pre_flush, post_flush,
1116 view_mask & ~attachment->cleared_views);
1117 if (view_mask)
1118 attachment->cleared_views |= view_mask;
1119 else
1120 attachment->pending_clear_aspects = 0;
1121 }
1122
1123 /**
1124 * Emit any pending attachment clears for the current subpass.
1125 *
1126 * @see radv_attachment_state::pending_clear_aspects
1127 */
1128 void
radv_cmd_buffer_clear_subpass(struct radv_cmd_buffer * cmd_buffer)1129 radv_cmd_buffer_clear_subpass(struct radv_cmd_buffer *cmd_buffer)
1130 {
1131 struct radv_cmd_state *cmd_state = &cmd_buffer->state;
1132 struct radv_meta_saved_state saved_state;
1133 enum radv_cmd_flush_bits pre_flush = 0;
1134 enum radv_cmd_flush_bits post_flush = 0;
1135
1136 if (!radv_subpass_needs_clear(cmd_buffer))
1137 return;
1138
1139 radv_meta_save(&saved_state, cmd_buffer,
1140 RADV_META_SAVE_GRAPHICS_PIPELINE |
1141 RADV_META_SAVE_CONSTANTS);
1142
1143 for (uint32_t i = 0; i < cmd_state->subpass->color_count; ++i) {
1144 uint32_t a = cmd_state->subpass->color_attachments[i].attachment;
1145
1146 if (!radv_attachment_needs_clear(cmd_state, a))
1147 continue;
1148
1149 assert(cmd_state->attachments[a].pending_clear_aspects ==
1150 VK_IMAGE_ASPECT_COLOR_BIT);
1151
1152 VkClearAttachment clear_att = {
1153 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
1154 .colorAttachment = i, /* Use attachment index relative to subpass */
1155 .clearValue = cmd_state->attachments[a].clear_value,
1156 };
1157
1158 radv_subpass_clear_attachment(cmd_buffer,
1159 &cmd_state->attachments[a],
1160 &clear_att, &pre_flush,
1161 &post_flush);
1162 }
1163
1164 uint32_t ds = cmd_state->subpass->depth_stencil_attachment.attachment;
1165 if (radv_attachment_needs_clear(cmd_state, ds)) {
1166 VkClearAttachment clear_att = {
1167 .aspectMask = cmd_state->attachments[ds].pending_clear_aspects,
1168 .clearValue = cmd_state->attachments[ds].clear_value,
1169 };
1170
1171 radv_subpass_clear_attachment(cmd_buffer,
1172 &cmd_state->attachments[ds],
1173 &clear_att, &pre_flush,
1174 &post_flush);
1175 }
1176
1177 radv_meta_restore(&saved_state, cmd_buffer);
1178 cmd_buffer->state.flush_bits |= post_flush;
1179 }
1180
1181 static void
radv_clear_image_layer(struct radv_cmd_buffer * cmd_buffer,struct radv_image * image,VkImageLayout image_layout,const VkImageSubresourceRange * range,VkFormat format,int level,int layer,const VkClearValue * clear_val)1182 radv_clear_image_layer(struct radv_cmd_buffer *cmd_buffer,
1183 struct radv_image *image,
1184 VkImageLayout image_layout,
1185 const VkImageSubresourceRange *range,
1186 VkFormat format, int level, int layer,
1187 const VkClearValue *clear_val)
1188 {
1189 VkDevice device_h = radv_device_to_handle(cmd_buffer->device);
1190 struct radv_image_view iview;
1191 uint32_t width = radv_minify(image->info.width, range->baseMipLevel + level);
1192 uint32_t height = radv_minify(image->info.height, range->baseMipLevel + level);
1193
1194 radv_image_view_init(&iview, cmd_buffer->device,
1195 &(VkImageViewCreateInfo) {
1196 .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
1197 .image = radv_image_to_handle(image),
1198 .viewType = radv_meta_get_view_type(image),
1199 .format = format,
1200 .subresourceRange = {
1201 .aspectMask = range->aspectMask,
1202 .baseMipLevel = range->baseMipLevel + level,
1203 .levelCount = 1,
1204 .baseArrayLayer = range->baseArrayLayer + layer,
1205 .layerCount = 1
1206 },
1207 });
1208
1209 VkFramebuffer fb;
1210 radv_CreateFramebuffer(device_h,
1211 &(VkFramebufferCreateInfo) {
1212 .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
1213 .attachmentCount = 1,
1214 .pAttachments = (VkImageView[]) {
1215 radv_image_view_to_handle(&iview),
1216 },
1217 .width = width,
1218 .height = height,
1219 .layers = 1
1220 },
1221 &cmd_buffer->pool->alloc,
1222 &fb);
1223
1224 VkAttachmentDescription att_desc = {
1225 .format = iview.vk_format,
1226 .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
1227 .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
1228 .stencilLoadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
1229 .stencilStoreOp = VK_ATTACHMENT_STORE_OP_STORE,
1230 .initialLayout = image_layout,
1231 .finalLayout = image_layout,
1232 };
1233
1234 VkSubpassDescription subpass_desc = {
1235 .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
1236 .inputAttachmentCount = 0,
1237 .colorAttachmentCount = 0,
1238 .pColorAttachments = NULL,
1239 .pResolveAttachments = NULL,
1240 .pDepthStencilAttachment = NULL,
1241 .preserveAttachmentCount = 0,
1242 .pPreserveAttachments = NULL,
1243 };
1244
1245 const VkAttachmentReference att_ref = {
1246 .attachment = 0,
1247 .layout = image_layout,
1248 };
1249
1250 if (range->aspectMask & VK_IMAGE_ASPECT_COLOR_BIT) {
1251 subpass_desc.colorAttachmentCount = 1;
1252 subpass_desc.pColorAttachments = &att_ref;
1253 } else {
1254 subpass_desc.pDepthStencilAttachment = &att_ref;
1255 }
1256
1257 VkRenderPass pass;
1258 radv_CreateRenderPass(device_h,
1259 &(VkRenderPassCreateInfo) {
1260 .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
1261 .attachmentCount = 1,
1262 .pAttachments = &att_desc,
1263 .subpassCount = 1,
1264 .pSubpasses = &subpass_desc,
1265 },
1266 &cmd_buffer->pool->alloc,
1267 &pass);
1268
1269 radv_CmdBeginRenderPass(radv_cmd_buffer_to_handle(cmd_buffer),
1270 &(VkRenderPassBeginInfo) {
1271 .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
1272 .renderArea = {
1273 .offset = { 0, 0, },
1274 .extent = {
1275 .width = width,
1276 .height = height,
1277 },
1278 },
1279 .renderPass = pass,
1280 .framebuffer = fb,
1281 .clearValueCount = 0,
1282 .pClearValues = NULL,
1283 },
1284 VK_SUBPASS_CONTENTS_INLINE);
1285
1286 VkClearAttachment clear_att = {
1287 .aspectMask = range->aspectMask,
1288 .colorAttachment = 0,
1289 .clearValue = *clear_val,
1290 };
1291
1292 VkClearRect clear_rect = {
1293 .rect = {
1294 .offset = { 0, 0 },
1295 .extent = { width, height },
1296 },
1297 .baseArrayLayer = range->baseArrayLayer,
1298 .layerCount = 1, /* FINISHME: clear multi-layer framebuffer */
1299 };
1300
1301 emit_clear(cmd_buffer, &clear_att, &clear_rect, NULL, NULL, 0);
1302
1303 radv_CmdEndRenderPass(radv_cmd_buffer_to_handle(cmd_buffer));
1304 radv_DestroyRenderPass(device_h, pass,
1305 &cmd_buffer->pool->alloc);
1306 radv_DestroyFramebuffer(device_h, fb,
1307 &cmd_buffer->pool->alloc);
1308 }
1309 static void
radv_cmd_clear_image(struct radv_cmd_buffer * cmd_buffer,struct radv_image * image,VkImageLayout image_layout,const VkClearValue * clear_value,uint32_t range_count,const VkImageSubresourceRange * ranges,bool cs)1310 radv_cmd_clear_image(struct radv_cmd_buffer *cmd_buffer,
1311 struct radv_image *image,
1312 VkImageLayout image_layout,
1313 const VkClearValue *clear_value,
1314 uint32_t range_count,
1315 const VkImageSubresourceRange *ranges,
1316 bool cs)
1317 {
1318 VkFormat format = image->vk_format;
1319 VkClearValue internal_clear_value = *clear_value;
1320
1321 if (format == VK_FORMAT_E5B9G9R9_UFLOAT_PACK32) {
1322 uint32_t value;
1323 format = VK_FORMAT_R32_UINT;
1324 value = float3_to_rgb9e5(clear_value->color.float32);
1325 internal_clear_value.color.uint32[0] = value;
1326 }
1327
1328 if (format == VK_FORMAT_R4G4_UNORM_PACK8) {
1329 uint8_t r, g;
1330 format = VK_FORMAT_R8_UINT;
1331 r = float_to_ubyte(clear_value->color.float32[0]) >> 4;
1332 g = float_to_ubyte(clear_value->color.float32[1]) >> 4;
1333 internal_clear_value.color.uint32[0] = (r << 4) | (g & 0xf);
1334 }
1335
1336 for (uint32_t r = 0; r < range_count; r++) {
1337 const VkImageSubresourceRange *range = &ranges[r];
1338 for (uint32_t l = 0; l < radv_get_levelCount(image, range); ++l) {
1339 const uint32_t layer_count = image->type == VK_IMAGE_TYPE_3D ?
1340 radv_minify(image->info.depth, range->baseMipLevel + l) :
1341 radv_get_layerCount(image, range);
1342 for (uint32_t s = 0; s < layer_count; ++s) {
1343
1344 if (cs) {
1345 struct radv_meta_blit2d_surf surf;
1346 surf.format = format;
1347 surf.image = image;
1348 surf.level = range->baseMipLevel + l;
1349 surf.layer = range->baseArrayLayer + s;
1350 surf.aspect_mask = range->aspectMask;
1351 radv_meta_clear_image_cs(cmd_buffer, &surf,
1352 &internal_clear_value.color);
1353 } else {
1354 radv_clear_image_layer(cmd_buffer, image, image_layout,
1355 range, format, l, s, &internal_clear_value);
1356 }
1357 }
1358 }
1359 }
1360 }
1361
radv_CmdClearColorImage(VkCommandBuffer commandBuffer,VkImage image_h,VkImageLayout imageLayout,const VkClearColorValue * pColor,uint32_t rangeCount,const VkImageSubresourceRange * pRanges)1362 void radv_CmdClearColorImage(
1363 VkCommandBuffer commandBuffer,
1364 VkImage image_h,
1365 VkImageLayout imageLayout,
1366 const VkClearColorValue* pColor,
1367 uint32_t rangeCount,
1368 const VkImageSubresourceRange* pRanges)
1369 {
1370 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
1371 RADV_FROM_HANDLE(radv_image, image, image_h);
1372 struct radv_meta_saved_state saved_state;
1373 bool cs = cmd_buffer->queue_family_index == RADV_QUEUE_COMPUTE;
1374
1375 if (cs) {
1376 radv_meta_save(&saved_state, cmd_buffer,
1377 RADV_META_SAVE_COMPUTE_PIPELINE |
1378 RADV_META_SAVE_CONSTANTS |
1379 RADV_META_SAVE_DESCRIPTORS);
1380 } else {
1381 radv_meta_save(&saved_state, cmd_buffer,
1382 RADV_META_SAVE_GRAPHICS_PIPELINE |
1383 RADV_META_SAVE_CONSTANTS);
1384 }
1385
1386 radv_cmd_clear_image(cmd_buffer, image, imageLayout,
1387 (const VkClearValue *) pColor,
1388 rangeCount, pRanges, cs);
1389
1390 radv_meta_restore(&saved_state, cmd_buffer);
1391 }
1392
radv_CmdClearDepthStencilImage(VkCommandBuffer commandBuffer,VkImage image_h,VkImageLayout imageLayout,const VkClearDepthStencilValue * pDepthStencil,uint32_t rangeCount,const VkImageSubresourceRange * pRanges)1393 void radv_CmdClearDepthStencilImage(
1394 VkCommandBuffer commandBuffer,
1395 VkImage image_h,
1396 VkImageLayout imageLayout,
1397 const VkClearDepthStencilValue* pDepthStencil,
1398 uint32_t rangeCount,
1399 const VkImageSubresourceRange* pRanges)
1400 {
1401 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
1402 RADV_FROM_HANDLE(radv_image, image, image_h);
1403 struct radv_meta_saved_state saved_state;
1404
1405 radv_meta_save(&saved_state, cmd_buffer,
1406 RADV_META_SAVE_GRAPHICS_PIPELINE |
1407 RADV_META_SAVE_CONSTANTS);
1408
1409 radv_cmd_clear_image(cmd_buffer, image, imageLayout,
1410 (const VkClearValue *) pDepthStencil,
1411 rangeCount, pRanges, false);
1412
1413 radv_meta_restore(&saved_state, cmd_buffer);
1414 }
1415
radv_CmdClearAttachments(VkCommandBuffer commandBuffer,uint32_t attachmentCount,const VkClearAttachment * pAttachments,uint32_t rectCount,const VkClearRect * pRects)1416 void radv_CmdClearAttachments(
1417 VkCommandBuffer commandBuffer,
1418 uint32_t attachmentCount,
1419 const VkClearAttachment* pAttachments,
1420 uint32_t rectCount,
1421 const VkClearRect* pRects)
1422 {
1423 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, commandBuffer);
1424 struct radv_meta_saved_state saved_state;
1425 enum radv_cmd_flush_bits pre_flush = 0;
1426 enum radv_cmd_flush_bits post_flush = 0;
1427
1428 if (!cmd_buffer->state.subpass)
1429 return;
1430
1431 radv_meta_save(&saved_state, cmd_buffer,
1432 RADV_META_SAVE_GRAPHICS_PIPELINE |
1433 RADV_META_SAVE_CONSTANTS);
1434
1435 /* FINISHME: We can do better than this dumb loop. It thrashes too much
1436 * state.
1437 */
1438 for (uint32_t a = 0; a < attachmentCount; ++a) {
1439 for (uint32_t r = 0; r < rectCount; ++r) {
1440 emit_clear(cmd_buffer, &pAttachments[a], &pRects[r], &pre_flush, &post_flush,
1441 cmd_buffer->state.subpass->view_mask);
1442 }
1443 }
1444
1445 radv_meta_restore(&saved_state, cmd_buffer);
1446 cmd_buffer->state.flush_bits |= post_flush;
1447 }
1448