vk_fsr.cpp 14 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396
  1. // Copyright 2021 yuzu Emulator Project
  2. // Licensed under GPLv2 or any later version
  3. // Refer to the license.txt file included.
  4. #include "common/common_types.h"
  5. #include "common/div_ceil.h"
  6. #include "video_core/host_shaders/vulkan_fidelityfx_fsr_easu_comp_spv.h"
  7. #include "video_core/host_shaders/vulkan_fidelityfx_fsr_rcas_comp_spv.h"
  8. #include "video_core/renderer_vulkan/vk_fsr.h"
  9. #include "video_core/renderer_vulkan/vk_scheduler.h"
  10. #include "video_core/renderer_vulkan/vk_shader_util.h"
  11. #include "video_core/vulkan_common/vulkan_device.h"
  12. namespace Vulkan {
  13. FSR::FSR(const Device& device_, MemoryAllocator& memory_allocator_, size_t image_count_,
  14. VkExtent2D output_size_)
  15. : device{device_}, memory_allocator{memory_allocator_}, image_count{image_count_},
  16. output_size{output_size_} {
  17. CreateImages();
  18. CreateSampler();
  19. CreateShaders();
  20. CreateDescriptorPool();
  21. CreateDescriptorSetLayout();
  22. CreateDescriptorSets();
  23. CreatePipelineLayout();
  24. CreatePipeline();
  25. }
  26. VkImageView FSR::Draw(VKScheduler& scheduler, size_t image_index, VkImageView image_view,
  27. const Common::Rectangle<int>& crop_rect) {
  28. UpdateDescriptorSet(image_index, image_view);
  29. scheduler.Record([this, image_index, crop_rect](vk::CommandBuffer cmdbuf) {
  30. const VkImageMemoryBarrier base_barrier{
  31. .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
  32. .pNext = nullptr,
  33. .srcAccessMask = 0,
  34. .dstAccessMask = 0,
  35. .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
  36. .newLayout = VK_IMAGE_LAYOUT_GENERAL,
  37. .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
  38. .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
  39. .image = {},
  40. .subresourceRange =
  41. {
  42. .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
  43. .baseMipLevel = 0,
  44. .levelCount = 1,
  45. .baseArrayLayer = 0,
  46. .layerCount = 1,
  47. },
  48. };
  49. // TODO: Support clear color
  50. cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *easu_pipeline);
  51. cmdbuf.PushConstants(*pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT,
  52. VkExtent2D{
  53. .width = static_cast<u32>(crop_rect.GetWidth()),
  54. .height = static_cast<u32>(crop_rect.GetHeight()),
  55. });
  56. {
  57. VkImageMemoryBarrier fsr_write_barrier = base_barrier;
  58. fsr_write_barrier.image = *images[image_index],
  59. fsr_write_barrier.oldLayout = VK_IMAGE_LAYOUT_UNDEFINED;
  60. cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
  61. VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, fsr_write_barrier);
  62. }
  63. cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline_layout, 0,
  64. descriptor_sets[image_index * 2], {});
  65. cmdbuf.Dispatch(Common::DivCeil(output_size.width, 16u),
  66. Common::DivCeil(output_size.height, 16u), 1);
  67. cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *rcas_pipeline);
  68. cmdbuf.PushConstants(*pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, output_size);
  69. {
  70. std::array<VkImageMemoryBarrier, 2> barriers;
  71. auto& fsr_read_barrier = barriers[0];
  72. auto& blit_write_barrier = barriers[1];
  73. fsr_read_barrier = base_barrier;
  74. fsr_read_barrier.image = *images[image_index];
  75. fsr_read_barrier.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT;
  76. fsr_read_barrier.dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
  77. blit_write_barrier = base_barrier;
  78. blit_write_barrier.image = *images[image_count + image_index];
  79. blit_write_barrier.oldLayout = VK_IMAGE_LAYOUT_UNDEFINED;
  80. blit_write_barrier.newLayout = VK_IMAGE_LAYOUT_GENERAL;
  81. cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
  82. VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, {}, {}, barriers);
  83. }
  84. cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline_layout, 0,
  85. descriptor_sets[image_index * 2 + 1], {});
  86. cmdbuf.Dispatch(Common::DivCeil(output_size.width, 16u),
  87. Common::DivCeil(output_size.height, 16u), 1);
  88. {
  89. std::array<VkImageMemoryBarrier, 1> barriers;
  90. auto& blit_read_barrier = barriers[0];
  91. blit_read_barrier = base_barrier;
  92. blit_read_barrier.image = *images[image_count + image_index];
  93. blit_read_barrier.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT;
  94. blit_read_barrier.dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
  95. cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
  96. VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, {}, {}, barriers);
  97. }
  98. });
  99. return *image_views[image_count + image_index];
  100. }
  101. void FSR::CreateDescriptorPool() {
  102. const std::array<VkDescriptorPoolSize, 2> pool_sizes{{
  103. {
  104. .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
  105. .descriptorCount = static_cast<u32>(image_count * 2),
  106. },
  107. {
  108. .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
  109. .descriptorCount = static_cast<u32>(image_count * 2),
  110. },
  111. }};
  112. const VkDescriptorPoolCreateInfo ci{
  113. .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
  114. .pNext = nullptr,
  115. .flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT,
  116. .maxSets = static_cast<u32>(image_count * 2),
  117. .poolSizeCount = static_cast<u32>(pool_sizes.size()),
  118. .pPoolSizes = pool_sizes.data(),
  119. };
  120. descriptor_pool = device.GetLogical().CreateDescriptorPool(ci);
  121. }
  122. void FSR::CreateDescriptorSetLayout() {
  123. const std::array<VkDescriptorSetLayoutBinding, 2> layout_bindings{{
  124. {
  125. .binding = 0,
  126. .descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
  127. .descriptorCount = 1,
  128. .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
  129. .pImmutableSamplers = sampler.address(),
  130. },
  131. {
  132. .binding = 1,
  133. .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
  134. .descriptorCount = 1,
  135. .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
  136. .pImmutableSamplers = sampler.address(),
  137. },
  138. }};
  139. const VkDescriptorSetLayoutCreateInfo ci{
  140. .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
  141. .pNext = nullptr,
  142. .flags = 0,
  143. .bindingCount = static_cast<u32>(layout_bindings.size()),
  144. .pBindings = layout_bindings.data(),
  145. };
  146. descriptor_set_layout = device.GetLogical().CreateDescriptorSetLayout(ci);
  147. }
  148. void FSR::CreateDescriptorSets() {
  149. const u32 sets = static_cast<u32>(image_count * 2);
  150. const std::vector layouts(sets, *descriptor_set_layout);
  151. const VkDescriptorSetAllocateInfo ai{
  152. .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO,
  153. .pNext = nullptr,
  154. .descriptorPool = *descriptor_pool,
  155. .descriptorSetCount = sets,
  156. .pSetLayouts = layouts.data(),
  157. };
  158. descriptor_sets = descriptor_pool.Allocate(ai);
  159. }
  160. void FSR::CreateImages() {
  161. images.resize(image_count * 2);
  162. image_views.resize(image_count * 2);
  163. buffer_commits.resize(image_count * 2);
  164. for (size_t i = 0; i < image_count * 2; ++i) {
  165. images[i] = device.GetLogical().CreateImage(VkImageCreateInfo{
  166. .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
  167. .pNext = nullptr,
  168. .flags = 0,
  169. .imageType = VK_IMAGE_TYPE_2D,
  170. .format = VK_FORMAT_R16G16B16A16_SFLOAT,
  171. .extent =
  172. {
  173. .width = output_size.width,
  174. .height = output_size.height,
  175. .depth = 1,
  176. },
  177. .mipLevels = 1,
  178. .arrayLayers = 1,
  179. .samples = VK_SAMPLE_COUNT_1_BIT,
  180. .tiling = VK_IMAGE_TILING_OPTIMAL,
  181. .usage = VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_STORAGE_BIT |
  182. VK_IMAGE_USAGE_SAMPLED_BIT,
  183. .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
  184. .queueFamilyIndexCount = 0,
  185. .pQueueFamilyIndices = nullptr,
  186. .initialLayout = VK_IMAGE_LAYOUT_UNDEFINED,
  187. });
  188. buffer_commits[i] = memory_allocator.Commit(images[i], MemoryUsage::DeviceLocal);
  189. image_views[i] = device.GetLogical().CreateImageView(VkImageViewCreateInfo{
  190. .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
  191. .pNext = nullptr,
  192. .flags = 0,
  193. .image = *images[i],
  194. .viewType = VK_IMAGE_VIEW_TYPE_2D,
  195. .format = VK_FORMAT_R16G16B16A16_SFLOAT,
  196. .components =
  197. {
  198. .r = VK_COMPONENT_SWIZZLE_IDENTITY,
  199. .g = VK_COMPONENT_SWIZZLE_IDENTITY,
  200. .b = VK_COMPONENT_SWIZZLE_IDENTITY,
  201. .a = VK_COMPONENT_SWIZZLE_IDENTITY,
  202. },
  203. .subresourceRange =
  204. {
  205. .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
  206. .baseMipLevel = 0,
  207. .levelCount = 1,
  208. .baseArrayLayer = 0,
  209. .layerCount = 1,
  210. },
  211. });
  212. }
  213. }
  214. void FSR::CreatePipelineLayout() {
  215. VkPushConstantRange push_const{
  216. .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
  217. .offset = 0,
  218. .size = sizeof(std::array<u32, 2>),
  219. };
  220. VkPipelineLayoutCreateInfo ci{
  221. .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
  222. .pNext = nullptr,
  223. .flags = 0,
  224. .setLayoutCount = 1,
  225. .pSetLayouts = descriptor_set_layout.address(),
  226. .pushConstantRangeCount = 1,
  227. .pPushConstantRanges = &push_const,
  228. };
  229. pipeline_layout = device.GetLogical().CreatePipelineLayout(ci);
  230. }
  231. void FSR::UpdateDescriptorSet(std::size_t image_index, VkImageView image_view) const {
  232. const auto fsr_image_view = *image_views[image_index];
  233. const auto blit_image_view = *image_views[image_count + image_index];
  234. const VkDescriptorImageInfo image_info{
  235. .sampler = VK_NULL_HANDLE,
  236. .imageView = image_view,
  237. .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
  238. };
  239. const VkDescriptorImageInfo fsr_image_info{
  240. .sampler = VK_NULL_HANDLE,
  241. .imageView = fsr_image_view,
  242. .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
  243. };
  244. const VkDescriptorImageInfo blit_image_info{
  245. .sampler = VK_NULL_HANDLE,
  246. .imageView = blit_image_view,
  247. .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
  248. };
  249. VkWriteDescriptorSet sampler_write{
  250. .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
  251. .pNext = nullptr,
  252. .dstSet = descriptor_sets[image_index * 2],
  253. .dstBinding = 0,
  254. .dstArrayElement = 0,
  255. .descriptorCount = 1,
  256. .descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
  257. .pImageInfo = &image_info,
  258. .pBufferInfo = nullptr,
  259. .pTexelBufferView = nullptr,
  260. };
  261. VkWriteDescriptorSet output_write{
  262. .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
  263. .pNext = nullptr,
  264. .dstSet = descriptor_sets[image_index * 2],
  265. .dstBinding = 1,
  266. .dstArrayElement = 0,
  267. .descriptorCount = 1,
  268. .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
  269. .pImageInfo = &fsr_image_info,
  270. .pBufferInfo = nullptr,
  271. .pTexelBufferView = nullptr,
  272. };
  273. device.GetLogical().UpdateDescriptorSets(std::array{sampler_write, output_write}, {});
  274. sampler_write.dstSet = descriptor_sets[image_index * 2 + 1];
  275. sampler_write.pImageInfo = &fsr_image_info;
  276. output_write.dstSet = descriptor_sets[image_index * 2 + 1];
  277. output_write.pImageInfo = &blit_image_info;
  278. device.GetLogical().UpdateDescriptorSets(std::array{sampler_write, output_write}, {});
  279. }
  280. void FSR::CreateSampler() {
  281. const VkSamplerCreateInfo ci{
  282. .sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO,
  283. .pNext = nullptr,
  284. .flags = 0,
  285. .magFilter = VK_FILTER_LINEAR,
  286. .minFilter = VK_FILTER_LINEAR,
  287. .mipmapMode = VK_SAMPLER_MIPMAP_MODE_LINEAR,
  288. .addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE,
  289. .addressModeV = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE,
  290. .addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE,
  291. .mipLodBias = 0.0f,
  292. .anisotropyEnable = VK_FALSE,
  293. .maxAnisotropy = 0.0f,
  294. .compareEnable = VK_FALSE,
  295. .compareOp = VK_COMPARE_OP_NEVER,
  296. .minLod = 0.0f,
  297. .maxLod = 0.0f,
  298. .borderColor = VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK,
  299. .unnormalizedCoordinates = VK_FALSE,
  300. };
  301. sampler = device.GetLogical().CreateSampler(ci);
  302. }
  303. void FSR::CreateShaders() {
  304. easu_shader = BuildShader(device, VULKAN_FIDELITYFX_FSR_EASU_COMP_SPV);
  305. rcas_shader = BuildShader(device, VULKAN_FIDELITYFX_FSR_RCAS_COMP_SPV);
  306. }
  307. void FSR::CreatePipeline() {
  308. VkPipelineShaderStageCreateInfo shader_stage_easu{
  309. .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
  310. .pNext = nullptr,
  311. .flags = 0,
  312. .stage = VK_SHADER_STAGE_COMPUTE_BIT,
  313. .module = *easu_shader,
  314. .pName = "main",
  315. .pSpecializationInfo = nullptr,
  316. };
  317. VkPipelineShaderStageCreateInfo shader_stage_rcas{
  318. .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
  319. .pNext = nullptr,
  320. .flags = 0,
  321. .stage = VK_SHADER_STAGE_COMPUTE_BIT,
  322. .module = *rcas_shader,
  323. .pName = "main",
  324. .pSpecializationInfo = nullptr,
  325. };
  326. VkComputePipelineCreateInfo pipeline_ci_easu{
  327. .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
  328. .pNext = nullptr,
  329. .flags = 0,
  330. .stage = shader_stage_easu,
  331. .layout = *pipeline_layout,
  332. .basePipelineHandle = VK_NULL_HANDLE,
  333. .basePipelineIndex = 0,
  334. };
  335. VkComputePipelineCreateInfo pipeline_ci_rcas{
  336. .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
  337. .pNext = nullptr,
  338. .flags = 0,
  339. .stage = shader_stage_rcas,
  340. .layout = *pipeline_layout,
  341. .basePipelineHandle = VK_NULL_HANDLE,
  342. .basePipelineIndex = 0,
  343. };
  344. easu_pipeline = device.GetLogical().CreateComputePipeline(pipeline_ci_easu);
  345. rcas_pipeline = device.GetLogical().CreateComputePipeline(pipeline_ci_rcas);
  346. }
  347. } // namespace Vulkan