3e6633dec03fa9c127690f1794e5842c2163edc6
[mesa.git] / src / amd / vulkan / radv_meta_resolve.c
1 /*
2 * Copyright © 2016 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 <assert.h>
25 #include <stdbool.h>
26
27 #include "radv_meta.h"
28 #include "radv_private.h"
29 #include "nir/nir_builder.h"
30 #include "sid.h"
31
32 /* emit 0, 0, 0, 1 */
33 static nir_shader *
34 build_nir_fs(void)
35 {
36 const struct glsl_type *vec4 = glsl_vec4_type();
37 nir_builder b;
38 nir_variable *f_color; /* vec4, fragment output color */
39
40 nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
41 b.shader->info->name = ralloc_asprintf(b.shader,
42 "meta_resolve_fs");
43
44 f_color = nir_variable_create(b.shader, nir_var_shader_out, vec4,
45 "f_color");
46 f_color->data.location = FRAG_RESULT_DATA0;
47 nir_store_var(&b, f_color, nir_imm_vec4(&b, 0.0, 0.0, 0.0, 1.0), 0xf);
48
49 return b.shader;
50 }
51
52 static VkResult
53 create_pass(struct radv_device *device)
54 {
55 VkResult result;
56 VkDevice device_h = radv_device_to_handle(device);
57 const VkAllocationCallbacks *alloc = &device->meta_state.alloc;
58 VkAttachmentDescription attachments[2];
59 int i;
60
61 for (i = 0; i < 2; i++) {
62 attachments[i].format = VK_FORMAT_UNDEFINED;
63 attachments[i].samples = 1;
64 attachments[i].loadOp = VK_ATTACHMENT_LOAD_OP_LOAD;
65 attachments[i].storeOp = VK_ATTACHMENT_STORE_OP_STORE;
66 }
67 attachments[0].initialLayout = VK_IMAGE_LAYOUT_GENERAL;
68 attachments[0].finalLayout = VK_IMAGE_LAYOUT_GENERAL;
69 attachments[1].initialLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
70 attachments[1].finalLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
71
72 result = radv_CreateRenderPass(device_h,
73 &(VkRenderPassCreateInfo) {
74 .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
75 .attachmentCount = 2,
76 .pAttachments = attachments,
77 .subpassCount = 1,
78 .pSubpasses = &(VkSubpassDescription) {
79 .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
80 .inputAttachmentCount = 0,
81 .colorAttachmentCount = 2,
82 .pColorAttachments = (VkAttachmentReference[]) {
83 {
84 .attachment = 0,
85 .layout = VK_IMAGE_LAYOUT_GENERAL,
86 },
87 {
88 .attachment = 1,
89 .layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
90 },
91 },
92 .pResolveAttachments = NULL,
93 .pDepthStencilAttachment = &(VkAttachmentReference) {
94 .attachment = VK_ATTACHMENT_UNUSED,
95 },
96 .preserveAttachmentCount = 0,
97 .pPreserveAttachments = NULL,
98 },
99 .dependencyCount = 0,
100 },
101 alloc,
102 &device->meta_state.resolve.pass);
103
104 return result;
105 }
106
107 static VkResult
108 create_pipeline(struct radv_device *device,
109 VkShaderModule vs_module_h)
110 {
111 VkResult result;
112 VkDevice device_h = radv_device_to_handle(device);
113
114 struct radv_shader_module fs_module = {
115 .nir = build_nir_fs(),
116 };
117
118 if (!fs_module.nir) {
119 /* XXX: Need more accurate error */
120 result = VK_ERROR_OUT_OF_HOST_MEMORY;
121 goto cleanup;
122 }
123
124 result = radv_graphics_pipeline_create(device_h,
125 radv_pipeline_cache_to_handle(&device->meta_state.cache),
126 &(VkGraphicsPipelineCreateInfo) {
127 .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
128 .stageCount = 2,
129 .pStages = (VkPipelineShaderStageCreateInfo[]) {
130 {
131 .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
132 .stage = VK_SHADER_STAGE_VERTEX_BIT,
133 .module = vs_module_h,
134 .pName = "main",
135 },
136 {
137 .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
138 .stage = VK_SHADER_STAGE_FRAGMENT_BIT,
139 .module = radv_shader_module_to_handle(&fs_module),
140 .pName = "main",
141 },
142 },
143 .pVertexInputState = &(VkPipelineVertexInputStateCreateInfo) {
144 .sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
145 .vertexBindingDescriptionCount = 0,
146 .vertexAttributeDescriptionCount = 0,
147 },
148 .pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) {
149 .sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
150 .topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
151 .primitiveRestartEnable = false,
152 },
153 .pViewportState = &(VkPipelineViewportStateCreateInfo) {
154 .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
155 .viewportCount = 1,
156 .scissorCount = 1,
157 },
158 .pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) {
159 .sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
160 .depthClampEnable = false,
161 .rasterizerDiscardEnable = false,
162 .polygonMode = VK_POLYGON_MODE_FILL,
163 .cullMode = VK_CULL_MODE_NONE,
164 .frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE,
165 },
166 .pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
167 .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
168 .rasterizationSamples = 1,
169 .sampleShadingEnable = false,
170 .pSampleMask = NULL,
171 .alphaToCoverageEnable = false,
172 .alphaToOneEnable = false,
173 },
174 .pColorBlendState = &(VkPipelineColorBlendStateCreateInfo) {
175 .sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
176 .logicOpEnable = false,
177 .attachmentCount = 2,
178 .pAttachments = (VkPipelineColorBlendAttachmentState []) {
179 {
180 .colorWriteMask = VK_COLOR_COMPONENT_R_BIT |
181 VK_COLOR_COMPONENT_G_BIT |
182 VK_COLOR_COMPONENT_B_BIT |
183 VK_COLOR_COMPONENT_A_BIT,
184 },
185 {
186 .colorWriteMask = 0,
187
188 }
189 },
190 },
191 .pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
192 .sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
193 .dynamicStateCount = 2,
194 .pDynamicStates = (VkDynamicState[]) {
195 VK_DYNAMIC_STATE_VIEWPORT,
196 VK_DYNAMIC_STATE_SCISSOR,
197 },
198 },
199 .renderPass = device->meta_state.resolve.pass,
200 .subpass = 0,
201 },
202 &(struct radv_graphics_pipeline_create_info) {
203 .use_rectlist = true,
204 .custom_blend_mode = V_028808_CB_RESOLVE,
205 },
206 &device->meta_state.alloc,
207 &device->meta_state.resolve.pipeline);
208 if (result != VK_SUCCESS)
209 goto cleanup;
210
211 goto cleanup;
212
213 cleanup:
214 ralloc_free(fs_module.nir);
215 return result;
216 }
217
218 void
219 radv_device_finish_meta_resolve_state(struct radv_device *device)
220 {
221 struct radv_meta_state *state = &device->meta_state;
222 VkDevice device_h = radv_device_to_handle(device);
223 VkRenderPass pass_h = device->meta_state.resolve.pass;
224 const VkAllocationCallbacks *alloc = &device->meta_state.alloc;
225
226 if (pass_h)
227 radv_DestroyRenderPass(device_h, pass_h,
228 &device->meta_state.alloc);
229
230 VkPipeline pipeline_h = state->resolve.pipeline;
231 if (pipeline_h) {
232 radv_DestroyPipeline(device_h, pipeline_h, alloc);
233 }
234 }
235
236 VkResult
237 radv_device_init_meta_resolve_state(struct radv_device *device)
238 {
239 VkResult res = VK_SUCCESS;
240
241 zero(device->meta_state.resolve);
242
243 struct radv_shader_module vs_module = { .nir = radv_meta_build_nir_vs_generate_vertices() };
244 if (!vs_module.nir) {
245 /* XXX: Need more accurate error */
246 res = VK_ERROR_OUT_OF_HOST_MEMORY;
247 goto fail;
248 }
249
250 res = create_pass(device);
251 if (res != VK_SUCCESS)
252 goto fail;
253
254 VkShaderModule vs_module_h = radv_shader_module_to_handle(&vs_module);
255 res = create_pipeline(device, vs_module_h);
256 if (res != VK_SUCCESS)
257 goto fail;
258
259 goto cleanup;
260
261 fail:
262 radv_device_finish_meta_resolve_state(device);
263
264 cleanup:
265 ralloc_free(vs_module.nir);
266
267 return res;
268 }
269
270 static void
271 emit_resolve(struct radv_cmd_buffer *cmd_buffer,
272 const VkOffset2D *dest_offset,
273 const VkExtent2D *resolve_extent)
274 {
275 struct radv_device *device = cmd_buffer->device;
276 VkCommandBuffer cmd_buffer_h = radv_cmd_buffer_to_handle(cmd_buffer);
277
278 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB;
279
280 VkPipeline pipeline_h = device->meta_state.resolve.pipeline;
281 RADV_FROM_HANDLE(radv_pipeline, pipeline, pipeline_h);
282
283 if (cmd_buffer->state.pipeline != pipeline) {
284 radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS,
285 pipeline_h);
286 }
287
288 radv_CmdSetViewport(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1, &(VkViewport) {
289 .x = dest_offset->x,
290 .y = dest_offset->y,
291 .width = resolve_extent->width,
292 .height = resolve_extent->height,
293 .minDepth = 0.0f,
294 .maxDepth = 1.0f
295 });
296
297 radv_CmdSetScissor(radv_cmd_buffer_to_handle(cmd_buffer), 0, 1, &(VkRect2D) {
298 .offset = *dest_offset,
299 .extent = *resolve_extent,
300 });
301
302 radv_CmdDraw(cmd_buffer_h, 3, 1, 0, 0);
303 cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_FLUSH_AND_INV_CB;
304 }
305
306 void radv_CmdResolveImage(
307 VkCommandBuffer cmd_buffer_h,
308 VkImage src_image_h,
309 VkImageLayout src_image_layout,
310 VkImage dest_image_h,
311 VkImageLayout dest_image_layout,
312 uint32_t region_count,
313 const VkImageResolve* regions)
314 {
315 RADV_FROM_HANDLE(radv_cmd_buffer, cmd_buffer, cmd_buffer_h);
316 RADV_FROM_HANDLE(radv_image, src_image, src_image_h);
317 RADV_FROM_HANDLE(radv_image, dest_image, dest_image_h);
318 struct radv_device *device = cmd_buffer->device;
319 struct radv_meta_saved_state saved_state;
320 VkDevice device_h = radv_device_to_handle(device);
321 bool use_compute_resolve = false;
322 bool use_fragment_resolve = false;
323 /* we can use the hw resolve only for single full resolves */
324 if (region_count == 1) {
325 if (regions[0].srcOffset.x ||
326 regions[0].srcOffset.y ||
327 regions[0].srcOffset.z)
328 use_compute_resolve = true;
329 if (regions[0].dstOffset.x ||
330 regions[0].dstOffset.y ||
331 regions[0].dstOffset.z)
332 use_compute_resolve = true;
333
334 if (regions[0].extent.width != src_image->info.width ||
335 regions[0].extent.height != src_image->info.height ||
336 regions[0].extent.depth != src_image->info.depth)
337 use_compute_resolve = true;
338 } else
339 use_compute_resolve = true;
340
341 if (use_fragment_resolve) {
342 radv_meta_resolve_fragment_image(cmd_buffer,
343 src_image,
344 src_image_layout,
345 dest_image,
346 dest_image_layout,
347 region_count, regions);
348 return;
349 }
350
351 if (use_compute_resolve) {
352 radv_meta_resolve_compute_image(cmd_buffer,
353 src_image,
354 src_image_layout,
355 dest_image,
356 dest_image_layout,
357 region_count, regions);
358 return;
359 }
360
361 radv_meta_save_graphics_reset_vport_scissor_novertex(&saved_state, cmd_buffer);
362
363 assert(src_image->info.samples > 1);
364 assert(dest_image->info.samples == 1);
365
366 if (src_image->info.samples >= 16) {
367 /* See commit aa3f9aaf31e9056a255f9e0472ebdfdaa60abe54 for the
368 * glBlitFramebuffer workaround for samples >= 16.
369 */
370 radv_finishme("vkCmdResolveImage: need interpolation workaround when "
371 "samples >= 16");
372 }
373
374 if (src_image->info.array_size > 1)
375 radv_finishme("vkCmdResolveImage: multisample array images");
376
377 if (dest_image->surface.dcc_size) {
378 radv_initialize_dcc(cmd_buffer, dest_image, 0xffffffff);
379 }
380 for (uint32_t r = 0; r < region_count; ++r) {
381 const VkImageResolve *region = &regions[r];
382
383 /* From the Vulkan 1.0 spec:
384 *
385 * - The aspectMask member of srcSubresource and dstSubresource must
386 * only contain VK_IMAGE_ASPECT_COLOR_BIT
387 *
388 * - The layerCount member of srcSubresource and dstSubresource must
389 * match
390 */
391 assert(region->srcSubresource.aspectMask == VK_IMAGE_ASPECT_COLOR_BIT);
392 assert(region->dstSubresource.aspectMask == VK_IMAGE_ASPECT_COLOR_BIT);
393 assert(region->srcSubresource.layerCount ==
394 region->dstSubresource.layerCount);
395
396 const uint32_t src_base_layer =
397 radv_meta_get_iview_layer(src_image, &region->srcSubresource,
398 &region->srcOffset);
399
400 const uint32_t dest_base_layer =
401 radv_meta_get_iview_layer(dest_image, &region->dstSubresource,
402 &region->dstOffset);
403
404 /**
405 * From Vulkan 1.0.6 spec: 18.6 Resolving Multisample Images
406 *
407 * extent is the size in texels of the source image to resolve in width,
408 * height and depth. 1D images use only x and width. 2D images use x, y,
409 * width and height. 3D images use x, y, z, width, height and depth.
410 *
411 * srcOffset and dstOffset select the initial x, y, and z offsets in
412 * texels of the sub-regions of the source and destination image data.
413 * extent is the size in texels of the source image to resolve in width,
414 * height and depth. 1D images use only x and width. 2D images use x, y,
415 * width and height. 3D images use x, y, z, width, height and depth.
416 */
417 const struct VkExtent3D extent =
418 radv_sanitize_image_extent(src_image->type, region->extent);
419 const struct VkOffset3D dstOffset =
420 radv_sanitize_image_offset(dest_image->type, region->dstOffset);
421
422
423 for (uint32_t layer = 0; layer < region->srcSubresource.layerCount;
424 ++layer) {
425
426 struct radv_image_view src_iview;
427 radv_image_view_init(&src_iview, cmd_buffer->device,
428 &(VkImageViewCreateInfo) {
429 .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
430 .image = src_image_h,
431 .viewType = radv_meta_get_view_type(src_image),
432 .format = src_image->vk_format,
433 .subresourceRange = {
434 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
435 .baseMipLevel = region->srcSubresource.mipLevel,
436 .levelCount = 1,
437 .baseArrayLayer = src_base_layer + layer,
438 .layerCount = 1,
439 },
440 },
441 cmd_buffer, VK_IMAGE_USAGE_SAMPLED_BIT);
442
443 struct radv_image_view dest_iview;
444 radv_image_view_init(&dest_iview, cmd_buffer->device,
445 &(VkImageViewCreateInfo) {
446 .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
447 .image = dest_image_h,
448 .viewType = radv_meta_get_view_type(dest_image),
449 .format = dest_image->vk_format,
450 .subresourceRange = {
451 .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
452 .baseMipLevel = region->dstSubresource.mipLevel,
453 .levelCount = 1,
454 .baseArrayLayer = dest_base_layer + layer,
455 .layerCount = 1,
456 },
457 },
458 cmd_buffer, VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT);
459
460 VkFramebuffer fb_h;
461 radv_CreateFramebuffer(device_h,
462 &(VkFramebufferCreateInfo) {
463 .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
464 .attachmentCount = 2,
465 .pAttachments = (VkImageView[]) {
466 radv_image_view_to_handle(&src_iview),
467 radv_image_view_to_handle(&dest_iview),
468 },
469 .width = radv_minify(dest_image->info.width,
470 region->dstSubresource.mipLevel),
471 .height = radv_minify(dest_image->info.height,
472 region->dstSubresource.mipLevel),
473 .layers = 1
474 },
475 &cmd_buffer->pool->alloc,
476 &fb_h);
477
478 radv_CmdBeginRenderPass(cmd_buffer_h,
479 &(VkRenderPassBeginInfo) {
480 .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
481 .renderPass = device->meta_state.resolve.pass,
482 .framebuffer = fb_h,
483 .renderArea = {
484 .offset = {
485 dstOffset.x,
486 dstOffset.y,
487 },
488 .extent = {
489 extent.width,
490 extent.height,
491 }
492 },
493 .clearValueCount = 0,
494 .pClearValues = NULL,
495 },
496 VK_SUBPASS_CONTENTS_INLINE);
497
498 emit_resolve(cmd_buffer,
499 &(VkOffset2D) {
500 .x = dstOffset.x,
501 .y = dstOffset.y,
502 },
503 &(VkExtent2D) {
504 .width = extent.width,
505 .height = extent.height,
506 });
507
508 radv_CmdEndRenderPass(cmd_buffer_h);
509
510 radv_DestroyFramebuffer(device_h, fb_h,
511 &cmd_buffer->pool->alloc);
512 }
513 }
514
515 radv_meta_restore(&saved_state, cmd_buffer);
516 }
517
518 /**
519 * Emit any needed resolves for the current subpass.
520 */
521 void
522 radv_cmd_buffer_resolve_subpass(struct radv_cmd_buffer *cmd_buffer)
523 {
524 struct radv_framebuffer *fb = cmd_buffer->state.framebuffer;
525 const struct radv_subpass *subpass = cmd_buffer->state.subpass;
526 struct radv_meta_saved_state saved_state;
527
528 /* FINISHME(perf): Skip clears for resolve attachments.
529 *
530 * From the Vulkan 1.0 spec:
531 *
532 * If the first use of an attachment in a render pass is as a resolve
533 * attachment, then the loadOp is effectively ignored as the resolve is
534 * guaranteed to overwrite all pixels in the render area.
535 */
536
537 if (!subpass->has_resolve)
538 return;
539
540 radv_meta_save_graphics_reset_vport_scissor(&saved_state, cmd_buffer);
541
542 for (uint32_t i = 0; i < subpass->color_count; ++i) {
543 VkAttachmentReference src_att = subpass->color_attachments[i];
544 VkAttachmentReference dest_att = subpass->resolve_attachments[i];
545 struct radv_image *dst_img = cmd_buffer->state.framebuffer->attachments[dest_att.attachment].attachment->image;
546 if (dest_att.attachment == VK_ATTACHMENT_UNUSED)
547 continue;
548
549 if (dst_img->surface.dcc_size) {
550 radv_initialize_dcc(cmd_buffer, dst_img, 0xffffffff);
551 cmd_buffer->state.attachments[dest_att.attachment].current_layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
552 }
553
554 struct radv_subpass resolve_subpass = {
555 .color_count = 2,
556 .color_attachments = (VkAttachmentReference[]) { src_att, dest_att },
557 .depth_stencil_attachment = { .attachment = VK_ATTACHMENT_UNUSED },
558 };
559
560 radv_cmd_buffer_set_subpass(cmd_buffer, &resolve_subpass, false);
561
562 /* Subpass resolves must respect the render area. We can ignore the
563 * render area here because vkCmdBeginRenderPass set the render area
564 * with 3DSTATE_DRAWING_RECTANGLE.
565 *
566 * XXX(chadv): Does the hardware really respect
567 * 3DSTATE_DRAWING_RECTANGLE when draing a 3DPRIM_RECTLIST?
568 */
569 emit_resolve(cmd_buffer,
570 &(VkOffset2D) { 0, 0 },
571 &(VkExtent2D) { fb->width, fb->height });
572 }
573
574 cmd_buffer->state.subpass = subpass;
575 radv_meta_restore(&saved_state, cmd_buffer);
576 }