vulkan_device.cpp 49 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211
  1. // Copyright 2018 yuzu Emulator Project
  2. // Licensed under GPLv2 or any later version
  3. // Refer to the license.txt file included.
  4. #include <algorithm>
  5. #include <bitset>
  6. #include <chrono>
  7. #include <optional>
  8. #include <string_view>
  9. #include <thread>
  10. #include <unordered_set>
  11. #include <utility>
  12. #include <vector>
  13. #include "common/assert.h"
  14. #include "common/settings.h"
  15. #include "video_core/vulkan_common/nsight_aftermath_tracker.h"
  16. #include "video_core/vulkan_common/vulkan_device.h"
  17. #include "video_core/vulkan_common/vulkan_wrapper.h"
  18. namespace Vulkan {
  19. namespace {
  20. namespace Alternatives {
  21. constexpr std::array DEPTH24_UNORM_STENCIL8_UINT{
  22. VK_FORMAT_D32_SFLOAT_S8_UINT,
  23. VK_FORMAT_D16_UNORM_S8_UINT,
  24. VK_FORMAT_UNDEFINED,
  25. };
  26. constexpr std::array DEPTH16_UNORM_STENCIL8_UINT{
  27. VK_FORMAT_D24_UNORM_S8_UINT,
  28. VK_FORMAT_D32_SFLOAT_S8_UINT,
  29. VK_FORMAT_UNDEFINED,
  30. };
  31. } // namespace Alternatives
  32. enum class NvidiaArchitecture {
  33. AmpereOrNewer,
  34. Turing,
  35. VoltaOrOlder,
  36. };
  37. constexpr std::array REQUIRED_EXTENSIONS{
  38. VK_KHR_MAINTENANCE1_EXTENSION_NAME,
  39. VK_KHR_STORAGE_BUFFER_STORAGE_CLASS_EXTENSION_NAME,
  40. VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
  41. VK_KHR_16BIT_STORAGE_EXTENSION_NAME,
  42. VK_KHR_8BIT_STORAGE_EXTENSION_NAME,
  43. VK_KHR_DRIVER_PROPERTIES_EXTENSION_NAME,
  44. VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME,
  45. VK_KHR_TIMELINE_SEMAPHORE_EXTENSION_NAME,
  46. VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
  47. VK_KHR_SHADER_FLOAT_CONTROLS_EXTENSION_NAME,
  48. VK_KHR_VARIABLE_POINTERS_EXTENSION_NAME,
  49. VK_EXT_VERTEX_ATTRIBUTE_DIVISOR_EXTENSION_NAME,
  50. VK_EXT_SHADER_SUBGROUP_BALLOT_EXTENSION_NAME,
  51. VK_EXT_SHADER_SUBGROUP_VOTE_EXTENSION_NAME,
  52. VK_EXT_ROBUSTNESS_2_EXTENSION_NAME,
  53. VK_EXT_HOST_QUERY_RESET_EXTENSION_NAME,
  54. VK_EXT_SHADER_DEMOTE_TO_HELPER_INVOCATION_EXTENSION_NAME,
  55. #ifdef _WIN32
  56. VK_KHR_EXTERNAL_MEMORY_WIN32_EXTENSION_NAME,
  57. #endif
  58. #ifdef __unix__
  59. VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME,
  60. #endif
  61. };
  62. template <typename T>
  63. void SetNext(void**& next, T& data) {
  64. *next = &data;
  65. next = &data.pNext;
  66. }
  67. constexpr const VkFormat* GetFormatAlternatives(VkFormat format) {
  68. switch (format) {
  69. case VK_FORMAT_D24_UNORM_S8_UINT:
  70. return Alternatives::DEPTH24_UNORM_STENCIL8_UINT.data();
  71. case VK_FORMAT_D16_UNORM_S8_UINT:
  72. return Alternatives::DEPTH16_UNORM_STENCIL8_UINT.data();
  73. default:
  74. return nullptr;
  75. }
  76. }
  77. VkFormatFeatureFlags GetFormatFeatures(VkFormatProperties properties, FormatType format_type) {
  78. switch (format_type) {
  79. case FormatType::Linear:
  80. return properties.linearTilingFeatures;
  81. case FormatType::Optimal:
  82. return properties.optimalTilingFeatures;
  83. case FormatType::Buffer:
  84. return properties.bufferFeatures;
  85. default:
  86. return {};
  87. }
  88. }
  89. std::unordered_map<VkFormat, VkFormatProperties> GetFormatProperties(vk::PhysicalDevice physical) {
  90. static constexpr std::array formats{
  91. VK_FORMAT_A8B8G8R8_UNORM_PACK32,
  92. VK_FORMAT_A8B8G8R8_UINT_PACK32,
  93. VK_FORMAT_A8B8G8R8_SNORM_PACK32,
  94. VK_FORMAT_A8B8G8R8_SINT_PACK32,
  95. VK_FORMAT_A8B8G8R8_SRGB_PACK32,
  96. VK_FORMAT_B5G6R5_UNORM_PACK16,
  97. VK_FORMAT_A2B10G10R10_UNORM_PACK32,
  98. VK_FORMAT_A2B10G10R10_UINT_PACK32,
  99. VK_FORMAT_A1R5G5B5_UNORM_PACK16,
  100. VK_FORMAT_R32G32B32A32_SFLOAT,
  101. VK_FORMAT_R32G32B32A32_SINT,
  102. VK_FORMAT_R32G32B32A32_UINT,
  103. VK_FORMAT_R32G32_SFLOAT,
  104. VK_FORMAT_R32G32_SINT,
  105. VK_FORMAT_R32G32_UINT,
  106. VK_FORMAT_R16G16B16A16_SINT,
  107. VK_FORMAT_R16G16B16A16_UINT,
  108. VK_FORMAT_R16G16B16A16_SNORM,
  109. VK_FORMAT_R16G16B16A16_UNORM,
  110. VK_FORMAT_R16G16_UNORM,
  111. VK_FORMAT_R16G16_SNORM,
  112. VK_FORMAT_R16G16_SFLOAT,
  113. VK_FORMAT_R16G16_SINT,
  114. VK_FORMAT_R16_UNORM,
  115. VK_FORMAT_R16_SNORM,
  116. VK_FORMAT_R16_UINT,
  117. VK_FORMAT_R8G8B8A8_SRGB,
  118. VK_FORMAT_R8G8_UNORM,
  119. VK_FORMAT_R8G8_SNORM,
  120. VK_FORMAT_R8G8_SINT,
  121. VK_FORMAT_R8G8_UINT,
  122. VK_FORMAT_R8_UNORM,
  123. VK_FORMAT_R8_SNORM,
  124. VK_FORMAT_R8_SINT,
  125. VK_FORMAT_R8_UINT,
  126. VK_FORMAT_B10G11R11_UFLOAT_PACK32,
  127. VK_FORMAT_R32_SFLOAT,
  128. VK_FORMAT_R32_UINT,
  129. VK_FORMAT_R32_SINT,
  130. VK_FORMAT_R16_SFLOAT,
  131. VK_FORMAT_R16G16B16A16_SFLOAT,
  132. VK_FORMAT_B8G8R8A8_UNORM,
  133. VK_FORMAT_B8G8R8A8_SRGB,
  134. VK_FORMAT_R4G4B4A4_UNORM_PACK16,
  135. VK_FORMAT_D32_SFLOAT,
  136. VK_FORMAT_D16_UNORM,
  137. VK_FORMAT_D16_UNORM_S8_UINT,
  138. VK_FORMAT_D24_UNORM_S8_UINT,
  139. VK_FORMAT_D32_SFLOAT_S8_UINT,
  140. VK_FORMAT_BC1_RGBA_UNORM_BLOCK,
  141. VK_FORMAT_BC2_UNORM_BLOCK,
  142. VK_FORMAT_BC3_UNORM_BLOCK,
  143. VK_FORMAT_BC4_UNORM_BLOCK,
  144. VK_FORMAT_BC4_SNORM_BLOCK,
  145. VK_FORMAT_BC5_UNORM_BLOCK,
  146. VK_FORMAT_BC5_SNORM_BLOCK,
  147. VK_FORMAT_BC7_UNORM_BLOCK,
  148. VK_FORMAT_BC6H_UFLOAT_BLOCK,
  149. VK_FORMAT_BC6H_SFLOAT_BLOCK,
  150. VK_FORMAT_BC1_RGBA_SRGB_BLOCK,
  151. VK_FORMAT_BC2_SRGB_BLOCK,
  152. VK_FORMAT_BC3_SRGB_BLOCK,
  153. VK_FORMAT_BC7_SRGB_BLOCK,
  154. VK_FORMAT_ASTC_4x4_UNORM_BLOCK,
  155. VK_FORMAT_ASTC_4x4_SRGB_BLOCK,
  156. VK_FORMAT_ASTC_5x4_UNORM_BLOCK,
  157. VK_FORMAT_ASTC_5x4_SRGB_BLOCK,
  158. VK_FORMAT_ASTC_5x5_UNORM_BLOCK,
  159. VK_FORMAT_ASTC_5x5_SRGB_BLOCK,
  160. VK_FORMAT_ASTC_6x5_UNORM_BLOCK,
  161. VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
  162. VK_FORMAT_ASTC_6x6_UNORM_BLOCK,
  163. VK_FORMAT_ASTC_6x6_SRGB_BLOCK,
  164. VK_FORMAT_ASTC_8x5_UNORM_BLOCK,
  165. VK_FORMAT_ASTC_8x5_SRGB_BLOCK,
  166. VK_FORMAT_ASTC_8x6_UNORM_BLOCK,
  167. VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
  168. VK_FORMAT_ASTC_8x8_UNORM_BLOCK,
  169. VK_FORMAT_ASTC_8x8_SRGB_BLOCK,
  170. VK_FORMAT_ASTC_10x5_UNORM_BLOCK,
  171. VK_FORMAT_ASTC_10x5_SRGB_BLOCK,
  172. VK_FORMAT_ASTC_10x6_UNORM_BLOCK,
  173. VK_FORMAT_ASTC_10x6_SRGB_BLOCK,
  174. VK_FORMAT_ASTC_10x8_UNORM_BLOCK,
  175. VK_FORMAT_ASTC_10x8_SRGB_BLOCK,
  176. VK_FORMAT_ASTC_10x10_UNORM_BLOCK,
  177. VK_FORMAT_ASTC_10x10_SRGB_BLOCK,
  178. VK_FORMAT_ASTC_12x10_UNORM_BLOCK,
  179. VK_FORMAT_ASTC_12x10_SRGB_BLOCK,
  180. VK_FORMAT_ASTC_12x12_UNORM_BLOCK,
  181. VK_FORMAT_ASTC_12x12_SRGB_BLOCK,
  182. VK_FORMAT_ASTC_8x6_UNORM_BLOCK,
  183. VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
  184. VK_FORMAT_ASTC_6x5_UNORM_BLOCK,
  185. VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
  186. VK_FORMAT_E5B9G9R9_UFLOAT_PACK32,
  187. };
  188. std::unordered_map<VkFormat, VkFormatProperties> format_properties;
  189. for (const auto format : formats) {
  190. format_properties.emplace(format, physical.GetFormatProperties(format));
  191. }
  192. return format_properties;
  193. }
  194. std::vector<std::string> GetSupportedExtensions(vk::PhysicalDevice physical) {
  195. const std::vector extensions = physical.EnumerateDeviceExtensionProperties();
  196. std::vector<std::string> supported_extensions;
  197. supported_extensions.reserve(extensions.size());
  198. for (const auto& extension : extensions) {
  199. supported_extensions.emplace_back(extension.extensionName);
  200. }
  201. return supported_extensions;
  202. }
  203. NvidiaArchitecture GetNvidiaArchitecture(vk::PhysicalDevice physical,
  204. std::span<const std::string> exts) {
  205. if (std::ranges::find(exts, VK_KHR_FRAGMENT_SHADING_RATE_EXTENSION_NAME) != exts.end()) {
  206. VkPhysicalDeviceFragmentShadingRatePropertiesKHR shading_rate_props{};
  207. shading_rate_props.sType =
  208. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADING_RATE_PROPERTIES_KHR;
  209. VkPhysicalDeviceProperties2KHR physical_properties{};
  210. physical_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR;
  211. physical_properties.pNext = &shading_rate_props;
  212. physical.GetProperties2KHR(physical_properties);
  213. if (shading_rate_props.primitiveFragmentShadingRateWithMultipleViewports) {
  214. // Only Ampere and newer support this feature
  215. return NvidiaArchitecture::AmpereOrNewer;
  216. }
  217. }
  218. if (std::ranges::find(exts, VK_NV_SHADING_RATE_IMAGE_EXTENSION_NAME) != exts.end()) {
  219. return NvidiaArchitecture::Turing;
  220. }
  221. return NvidiaArchitecture::VoltaOrOlder;
  222. }
  223. } // Anonymous namespace
  224. Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR surface,
  225. const vk::InstanceDispatch& dld_)
  226. : instance{instance_}, dld{dld_}, physical{physical_}, properties{physical.GetProperties()},
  227. supported_extensions{GetSupportedExtensions(physical)},
  228. format_properties(GetFormatProperties(physical)) {
  229. CheckSuitability(surface != nullptr);
  230. SetupFamilies(surface);
  231. SetupFeatures();
  232. SetupProperties();
  233. const auto queue_cis = GetDeviceQueueCreateInfos();
  234. const std::vector extensions = LoadExtensions(surface != nullptr);
  235. VkPhysicalDeviceFeatures2 features2{
  236. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2,
  237. .pNext = nullptr,
  238. .features{
  239. .robustBufferAccess = true,
  240. .fullDrawIndexUint32 = false,
  241. .imageCubeArray = true,
  242. .independentBlend = true,
  243. .geometryShader = true,
  244. .tessellationShader = true,
  245. .sampleRateShading = true,
  246. .dualSrcBlend = true,
  247. .logicOp = false,
  248. .multiDrawIndirect = false,
  249. .drawIndirectFirstInstance = false,
  250. .depthClamp = true,
  251. .depthBiasClamp = true,
  252. .fillModeNonSolid = true,
  253. .depthBounds = is_depth_bounds_supported,
  254. .wideLines = true,
  255. .largePoints = true,
  256. .alphaToOne = false,
  257. .multiViewport = true,
  258. .samplerAnisotropy = true,
  259. .textureCompressionETC2 = false,
  260. .textureCompressionASTC_LDR = is_optimal_astc_supported,
  261. .textureCompressionBC = false,
  262. .occlusionQueryPrecise = true,
  263. .pipelineStatisticsQuery = false,
  264. .vertexPipelineStoresAndAtomics = true,
  265. .fragmentStoresAndAtomics = true,
  266. .shaderTessellationAndGeometryPointSize = false,
  267. .shaderImageGatherExtended = true,
  268. .shaderStorageImageExtendedFormats = false,
  269. .shaderStorageImageMultisample = is_shader_storage_image_multisample,
  270. .shaderStorageImageReadWithoutFormat = is_formatless_image_load_supported,
  271. .shaderStorageImageWriteWithoutFormat = true,
  272. .shaderUniformBufferArrayDynamicIndexing = false,
  273. .shaderSampledImageArrayDynamicIndexing = false,
  274. .shaderStorageBufferArrayDynamicIndexing = false,
  275. .shaderStorageImageArrayDynamicIndexing = false,
  276. .shaderClipDistance = true,
  277. .shaderCullDistance = true,
  278. .shaderFloat64 = is_shader_float64_supported,
  279. .shaderInt64 = is_shader_int64_supported,
  280. .shaderInt16 = is_shader_int16_supported,
  281. .shaderResourceResidency = false,
  282. .shaderResourceMinLod = false,
  283. .sparseBinding = false,
  284. .sparseResidencyBuffer = false,
  285. .sparseResidencyImage2D = false,
  286. .sparseResidencyImage3D = false,
  287. .sparseResidency2Samples = false,
  288. .sparseResidency4Samples = false,
  289. .sparseResidency8Samples = false,
  290. .sparseResidency16Samples = false,
  291. .sparseResidencyAliased = false,
  292. .variableMultisampleRate = false,
  293. .inheritedQueries = false,
  294. },
  295. };
  296. const void* first_next = &features2;
  297. void** next = &features2.pNext;
  298. VkPhysicalDeviceTimelineSemaphoreFeaturesKHR timeline_semaphore{
  299. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR,
  300. .pNext = nullptr,
  301. .timelineSemaphore = true,
  302. };
  303. SetNext(next, timeline_semaphore);
  304. VkPhysicalDevice16BitStorageFeaturesKHR bit16_storage{
  305. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR,
  306. .pNext = nullptr,
  307. .storageBuffer16BitAccess = true,
  308. .uniformAndStorageBuffer16BitAccess = true,
  309. .storagePushConstant16 = false,
  310. .storageInputOutput16 = false,
  311. };
  312. SetNext(next, bit16_storage);
  313. VkPhysicalDevice8BitStorageFeaturesKHR bit8_storage{
  314. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR,
  315. .pNext = nullptr,
  316. .storageBuffer8BitAccess = false,
  317. .uniformAndStorageBuffer8BitAccess = true,
  318. .storagePushConstant8 = false,
  319. };
  320. SetNext(next, bit8_storage);
  321. VkPhysicalDeviceRobustness2FeaturesEXT robustness2{
  322. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT,
  323. .pNext = nullptr,
  324. .robustBufferAccess2 = true,
  325. .robustImageAccess2 = true,
  326. .nullDescriptor = true,
  327. };
  328. SetNext(next, robustness2);
  329. VkPhysicalDeviceHostQueryResetFeaturesEXT host_query_reset{
  330. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT,
  331. .pNext = nullptr,
  332. .hostQueryReset = true,
  333. };
  334. SetNext(next, host_query_reset);
  335. VkPhysicalDeviceVariablePointerFeaturesKHR variable_pointers{
  336. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES_KHR,
  337. .pNext = nullptr,
  338. .variablePointersStorageBuffer = VK_TRUE,
  339. .variablePointers = VK_TRUE,
  340. };
  341. SetNext(next, variable_pointers);
  342. VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT demote{
  343. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT,
  344. .pNext = nullptr,
  345. .shaderDemoteToHelperInvocation = true,
  346. };
  347. SetNext(next, demote);
  348. VkPhysicalDeviceFloat16Int8FeaturesKHR float16_int8;
  349. if (is_float16_supported) {
  350. float16_int8 = {
  351. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR,
  352. .pNext = nullptr,
  353. .shaderFloat16 = true,
  354. .shaderInt8 = false,
  355. };
  356. SetNext(next, float16_int8);
  357. } else {
  358. LOG_INFO(Render_Vulkan, "Device doesn't support float16 natively");
  359. }
  360. if (!nv_viewport_swizzle) {
  361. LOG_INFO(Render_Vulkan, "Device doesn't support viewport swizzles");
  362. }
  363. if (!nv_viewport_array2) {
  364. LOG_INFO(Render_Vulkan, "Device doesn't support viewport masks");
  365. }
  366. if (!nv_geometry_shader_passthrough) {
  367. LOG_INFO(Render_Vulkan, "Device doesn't support passthrough geometry shaders");
  368. }
  369. VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR std430_layout;
  370. if (khr_uniform_buffer_standard_layout) {
  371. std430_layout = {
  372. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR,
  373. .pNext = nullptr,
  374. .uniformBufferStandardLayout = true,
  375. };
  376. SetNext(next, std430_layout);
  377. } else {
  378. LOG_INFO(Render_Vulkan, "Device doesn't support packed UBOs");
  379. }
  380. VkPhysicalDeviceIndexTypeUint8FeaturesEXT index_type_uint8;
  381. if (ext_index_type_uint8) {
  382. index_type_uint8 = {
  383. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT,
  384. .pNext = nullptr,
  385. .indexTypeUint8 = true,
  386. };
  387. SetNext(next, index_type_uint8);
  388. } else {
  389. LOG_INFO(Render_Vulkan, "Device doesn't support uint8 indexes");
  390. }
  391. VkPhysicalDeviceTransformFeedbackFeaturesEXT transform_feedback;
  392. if (ext_transform_feedback) {
  393. transform_feedback = {
  394. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT,
  395. .pNext = nullptr,
  396. .transformFeedback = true,
  397. .geometryStreams = true,
  398. };
  399. SetNext(next, transform_feedback);
  400. } else {
  401. LOG_INFO(Render_Vulkan, "Device doesn't support transform feedbacks");
  402. }
  403. VkPhysicalDeviceCustomBorderColorFeaturesEXT custom_border;
  404. if (ext_custom_border_color) {
  405. custom_border = {
  406. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT,
  407. .pNext = nullptr,
  408. .customBorderColors = VK_TRUE,
  409. .customBorderColorWithoutFormat = VK_TRUE,
  410. };
  411. SetNext(next, custom_border);
  412. } else {
  413. LOG_INFO(Render_Vulkan, "Device doesn't support custom border colors");
  414. }
  415. VkPhysicalDeviceExtendedDynamicStateFeaturesEXT dynamic_state;
  416. if (ext_extended_dynamic_state) {
  417. dynamic_state = {
  418. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT,
  419. .pNext = nullptr,
  420. .extendedDynamicState = VK_TRUE,
  421. };
  422. SetNext(next, dynamic_state);
  423. } else {
  424. LOG_INFO(Render_Vulkan, "Device doesn't support extended dynamic state");
  425. }
  426. VkPhysicalDeviceLineRasterizationFeaturesEXT line_raster;
  427. if (ext_line_rasterization) {
  428. line_raster = {
  429. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT,
  430. .pNext = nullptr,
  431. .rectangularLines = VK_TRUE,
  432. .bresenhamLines = VK_FALSE,
  433. .smoothLines = VK_TRUE,
  434. .stippledRectangularLines = VK_FALSE,
  435. .stippledBresenhamLines = VK_FALSE,
  436. .stippledSmoothLines = VK_FALSE,
  437. };
  438. SetNext(next, line_raster);
  439. } else {
  440. LOG_INFO(Render_Vulkan, "Device doesn't support smooth lines");
  441. }
  442. if (!ext_conservative_rasterization) {
  443. LOG_INFO(Render_Vulkan, "Device doesn't support conservative rasterization");
  444. }
  445. VkPhysicalDeviceProvokingVertexFeaturesEXT provoking_vertex;
  446. if (ext_provoking_vertex) {
  447. provoking_vertex = {
  448. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_FEATURES_EXT,
  449. .pNext = nullptr,
  450. .provokingVertexLast = VK_TRUE,
  451. .transformFeedbackPreservesProvokingVertex = VK_TRUE,
  452. };
  453. SetNext(next, provoking_vertex);
  454. } else {
  455. LOG_INFO(Render_Vulkan, "Device doesn't support provoking vertex last");
  456. }
  457. VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT vertex_input_dynamic;
  458. if (ext_vertex_input_dynamic_state) {
  459. vertex_input_dynamic = {
  460. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_INPUT_DYNAMIC_STATE_FEATURES_EXT,
  461. .pNext = nullptr,
  462. .vertexInputDynamicState = VK_TRUE,
  463. };
  464. SetNext(next, vertex_input_dynamic);
  465. } else {
  466. LOG_INFO(Render_Vulkan, "Device doesn't support vertex input dynamic state");
  467. }
  468. VkPhysicalDeviceShaderAtomicInt64FeaturesKHR atomic_int64;
  469. if (ext_shader_atomic_int64) {
  470. atomic_int64 = {
  471. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR,
  472. .pNext = nullptr,
  473. .shaderBufferInt64Atomics = VK_TRUE,
  474. .shaderSharedInt64Atomics = VK_TRUE,
  475. };
  476. SetNext(next, atomic_int64);
  477. }
  478. VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR workgroup_layout;
  479. if (khr_workgroup_memory_explicit_layout) {
  480. workgroup_layout = {
  481. .sType =
  482. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR,
  483. .pNext = nullptr,
  484. .workgroupMemoryExplicitLayout = VK_TRUE,
  485. .workgroupMemoryExplicitLayoutScalarBlockLayout = VK_TRUE,
  486. .workgroupMemoryExplicitLayout8BitAccess = VK_TRUE,
  487. .workgroupMemoryExplicitLayout16BitAccess = VK_TRUE,
  488. };
  489. SetNext(next, workgroup_layout);
  490. }
  491. VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR executable_properties;
  492. if (khr_pipeline_executable_properties) {
  493. LOG_INFO(Render_Vulkan, "Enabling shader feedback, expect slower shader build times");
  494. executable_properties = {
  495. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR,
  496. .pNext = nullptr,
  497. .pipelineExecutableInfo = VK_TRUE,
  498. };
  499. SetNext(next, executable_properties);
  500. }
  501. if (!ext_depth_range_unrestricted) {
  502. LOG_INFO(Render_Vulkan, "Device doesn't support depth range unrestricted");
  503. }
  504. VkDeviceDiagnosticsConfigCreateInfoNV diagnostics_nv;
  505. if (Settings::values.enable_nsight_aftermath && nv_device_diagnostics_config) {
  506. nsight_aftermath_tracker = std::make_unique<NsightAftermathTracker>();
  507. diagnostics_nv = {
  508. .sType = VK_STRUCTURE_TYPE_DEVICE_DIAGNOSTICS_CONFIG_CREATE_INFO_NV,
  509. .pNext = &features2,
  510. .flags = VK_DEVICE_DIAGNOSTICS_CONFIG_ENABLE_SHADER_DEBUG_INFO_BIT_NV |
  511. VK_DEVICE_DIAGNOSTICS_CONFIG_ENABLE_RESOURCE_TRACKING_BIT_NV |
  512. VK_DEVICE_DIAGNOSTICS_CONFIG_ENABLE_AUTOMATIC_CHECKPOINTS_BIT_NV,
  513. };
  514. first_next = &diagnostics_nv;
  515. }
  516. logical = vk::Device::Create(physical, queue_cis, extensions, first_next, dld);
  517. CollectPhysicalMemoryInfo();
  518. CollectTelemetryParameters();
  519. CollectToolingInfo();
  520. if (driver_id == VK_DRIVER_ID_NVIDIA_PROPRIETARY_KHR) {
  521. const auto arch = GetNvidiaArchitecture(physical, supported_extensions);
  522. switch (arch) {
  523. case NvidiaArchitecture::AmpereOrNewer:
  524. LOG_WARNING(Render_Vulkan, "Blacklisting Ampere devices from float16 math");
  525. is_float16_supported = false;
  526. break;
  527. case NvidiaArchitecture::Turing:
  528. break;
  529. case NvidiaArchitecture::VoltaOrOlder:
  530. LOG_WARNING(Render_Vulkan, "Blacklisting Volta and older from VK_KHR_push_descriptor");
  531. khr_push_descriptor = false;
  532. break;
  533. }
  534. }
  535. if (ext_extended_dynamic_state && driver_id == VK_DRIVER_ID_MESA_RADV) {
  536. // Mask driver version variant
  537. const u32 version = (properties.driverVersion << 3) >> 3;
  538. if (version < VK_MAKE_API_VERSION(0, 21, 2, 0)) {
  539. LOG_WARNING(Render_Vulkan,
  540. "RADV versions older than 21.2 have broken VK_EXT_extended_dynamic_state");
  541. ext_extended_dynamic_state = false;
  542. }
  543. }
  544. if (ext_vertex_input_dynamic_state && driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
  545. LOG_WARNING(Render_Vulkan, "Blacklisting Intel for VK_EXT_vertex_input_dynamic_state");
  546. ext_vertex_input_dynamic_state = false;
  547. }
  548. if (is_float16_supported && driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
  549. // Intel's compiler crashes when using fp16 on Astral Chain, disable it for the time being.
  550. LOG_WARNING(Render_Vulkan, "Blacklisting Intel proprietary from float16 math");
  551. is_float16_supported = false;
  552. }
  553. graphics_queue = logical.GetQueue(graphics_family);
  554. present_queue = logical.GetQueue(present_family);
  555. }
  556. Device::~Device() = default;
  557. VkFormat Device::GetSupportedFormat(VkFormat wanted_format, VkFormatFeatureFlags wanted_usage,
  558. FormatType format_type) const {
  559. if (IsFormatSupported(wanted_format, wanted_usage, format_type)) {
  560. return wanted_format;
  561. }
  562. // The wanted format is not supported by hardware, search for alternatives
  563. const VkFormat* alternatives = GetFormatAlternatives(wanted_format);
  564. if (alternatives == nullptr) {
  565. UNREACHABLE_MSG("Format={} with usage={} and type={} has no defined alternatives and host "
  566. "hardware does not support it",
  567. wanted_format, wanted_usage, format_type);
  568. return wanted_format;
  569. }
  570. std::size_t i = 0;
  571. for (VkFormat alternative = *alternatives; alternative; alternative = alternatives[++i]) {
  572. if (!IsFormatSupported(alternative, wanted_usage, format_type)) {
  573. continue;
  574. }
  575. LOG_WARNING(Render_Vulkan,
  576. "Emulating format={} with alternative format={} with usage={} and type={}",
  577. wanted_format, alternative, wanted_usage, format_type);
  578. return alternative;
  579. }
  580. // No alternatives found, panic
  581. UNREACHABLE_MSG("Format={} with usage={} and type={} is not supported by the host hardware and "
  582. "doesn't support any of the alternatives",
  583. wanted_format, wanted_usage, format_type);
  584. return wanted_format;
  585. }
  586. void Device::ReportLoss() const {
  587. LOG_CRITICAL(Render_Vulkan, "Device loss occured!");
  588. // Wait for the log to flush and for Nsight Aftermath to dump the results
  589. std::this_thread::sleep_for(std::chrono::seconds{15});
  590. }
  591. void Device::SaveShader(std::span<const u32> spirv) const {
  592. if (nsight_aftermath_tracker) {
  593. nsight_aftermath_tracker->SaveShader(spirv);
  594. }
  595. }
  596. bool Device::IsOptimalAstcSupported(const VkPhysicalDeviceFeatures& features) const {
  597. // Disable for now to avoid converting ASTC twice.
  598. static constexpr std::array astc_formats = {
  599. VK_FORMAT_ASTC_4x4_UNORM_BLOCK, VK_FORMAT_ASTC_4x4_SRGB_BLOCK,
  600. VK_FORMAT_ASTC_5x4_UNORM_BLOCK, VK_FORMAT_ASTC_5x4_SRGB_BLOCK,
  601. VK_FORMAT_ASTC_5x5_UNORM_BLOCK, VK_FORMAT_ASTC_5x5_SRGB_BLOCK,
  602. VK_FORMAT_ASTC_6x5_UNORM_BLOCK, VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
  603. VK_FORMAT_ASTC_6x6_UNORM_BLOCK, VK_FORMAT_ASTC_6x6_SRGB_BLOCK,
  604. VK_FORMAT_ASTC_8x5_UNORM_BLOCK, VK_FORMAT_ASTC_8x5_SRGB_BLOCK,
  605. VK_FORMAT_ASTC_8x6_UNORM_BLOCK, VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
  606. VK_FORMAT_ASTC_8x8_UNORM_BLOCK, VK_FORMAT_ASTC_8x8_SRGB_BLOCK,
  607. VK_FORMAT_ASTC_10x5_UNORM_BLOCK, VK_FORMAT_ASTC_10x5_SRGB_BLOCK,
  608. VK_FORMAT_ASTC_10x6_UNORM_BLOCK, VK_FORMAT_ASTC_10x6_SRGB_BLOCK,
  609. VK_FORMAT_ASTC_10x8_UNORM_BLOCK, VK_FORMAT_ASTC_10x8_SRGB_BLOCK,
  610. VK_FORMAT_ASTC_10x10_UNORM_BLOCK, VK_FORMAT_ASTC_10x10_SRGB_BLOCK,
  611. VK_FORMAT_ASTC_12x10_UNORM_BLOCK, VK_FORMAT_ASTC_12x10_SRGB_BLOCK,
  612. VK_FORMAT_ASTC_12x12_UNORM_BLOCK, VK_FORMAT_ASTC_12x12_SRGB_BLOCK,
  613. };
  614. if (!features.textureCompressionASTC_LDR) {
  615. return false;
  616. }
  617. const auto format_feature_usage{
  618. VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT | VK_FORMAT_FEATURE_BLIT_SRC_BIT |
  619. VK_FORMAT_FEATURE_BLIT_DST_BIT | VK_FORMAT_FEATURE_TRANSFER_SRC_BIT |
  620. VK_FORMAT_FEATURE_TRANSFER_DST_BIT};
  621. for (const auto format : astc_formats) {
  622. const auto physical_format_properties{physical.GetFormatProperties(format)};
  623. if ((physical_format_properties.optimalTilingFeatures & format_feature_usage) == 0) {
  624. return false;
  625. }
  626. }
  627. return true;
  628. }
  629. bool Device::TestDepthStencilBlits() const {
  630. static constexpr VkFormatFeatureFlags required_features =
  631. VK_FORMAT_FEATURE_BLIT_SRC_BIT | VK_FORMAT_FEATURE_BLIT_DST_BIT;
  632. const auto test_features = [](VkFormatProperties props) {
  633. return (props.optimalTilingFeatures & required_features) == required_features;
  634. };
  635. return test_features(format_properties.at(VK_FORMAT_D32_SFLOAT_S8_UINT)) &&
  636. test_features(format_properties.at(VK_FORMAT_D24_UNORM_S8_UINT));
  637. }
  638. bool Device::IsFormatSupported(VkFormat wanted_format, VkFormatFeatureFlags wanted_usage,
  639. FormatType format_type) const {
  640. const auto it = format_properties.find(wanted_format);
  641. if (it == format_properties.end()) {
  642. UNIMPLEMENTED_MSG("Unimplemented format query={}", wanted_format);
  643. return true;
  644. }
  645. const auto supported_usage = GetFormatFeatures(it->second, format_type);
  646. return (supported_usage & wanted_usage) == wanted_usage;
  647. }
  648. std::string Device::GetDriverName() const {
  649. switch (driver_id) {
  650. case VK_DRIVER_ID_AMD_PROPRIETARY:
  651. return "AMD";
  652. case VK_DRIVER_ID_AMD_OPEN_SOURCE:
  653. return "AMDVLK";
  654. case VK_DRIVER_ID_MESA_RADV:
  655. return "RADV";
  656. case VK_DRIVER_ID_NVIDIA_PROPRIETARY:
  657. return "NVIDIA";
  658. case VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS:
  659. return "INTEL";
  660. case VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA:
  661. return "ANV";
  662. case VK_DRIVER_ID_MESA_LLVMPIPE:
  663. return "LAVAPIPE";
  664. default:
  665. return vendor_name;
  666. }
  667. }
  668. void Device::CheckSuitability(bool requires_swapchain) const {
  669. std::bitset<REQUIRED_EXTENSIONS.size()> available_extensions;
  670. bool has_swapchain = false;
  671. for (const VkExtensionProperties& property : physical.EnumerateDeviceExtensionProperties()) {
  672. const std::string_view name{property.extensionName};
  673. for (size_t i = 0; i < REQUIRED_EXTENSIONS.size(); ++i) {
  674. if (available_extensions[i]) {
  675. continue;
  676. }
  677. available_extensions[i] = name == REQUIRED_EXTENSIONS[i];
  678. }
  679. has_swapchain = has_swapchain || name == VK_KHR_SWAPCHAIN_EXTENSION_NAME;
  680. }
  681. for (size_t i = 0; i < REQUIRED_EXTENSIONS.size(); ++i) {
  682. if (available_extensions[i]) {
  683. continue;
  684. }
  685. LOG_ERROR(Render_Vulkan, "Missing required extension: {}", REQUIRED_EXTENSIONS[i]);
  686. throw vk::Exception(VK_ERROR_EXTENSION_NOT_PRESENT);
  687. }
  688. if (requires_swapchain && !has_swapchain) {
  689. LOG_ERROR(Render_Vulkan, "Missing required extension: VK_KHR_swapchain");
  690. throw vk::Exception(VK_ERROR_EXTENSION_NOT_PRESENT);
  691. }
  692. struct LimitTuple {
  693. u32 minimum;
  694. u32 value;
  695. const char* name;
  696. };
  697. const VkPhysicalDeviceLimits& limits{properties.limits};
  698. const std::array limits_report{
  699. LimitTuple{65536, limits.maxUniformBufferRange, "maxUniformBufferRange"},
  700. LimitTuple{16, limits.maxViewports, "maxViewports"},
  701. LimitTuple{8, limits.maxColorAttachments, "maxColorAttachments"},
  702. LimitTuple{8, limits.maxClipDistances, "maxClipDistances"},
  703. };
  704. for (const auto& tuple : limits_report) {
  705. if (tuple.value < tuple.minimum) {
  706. LOG_ERROR(Render_Vulkan, "{} has to be {} or greater but it is {}", tuple.name,
  707. tuple.minimum, tuple.value);
  708. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  709. }
  710. }
  711. VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT demote{};
  712. demote.sType =
  713. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT;
  714. demote.pNext = nullptr;
  715. VkPhysicalDeviceVariablePointerFeaturesKHR variable_pointers{};
  716. variable_pointers.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES_KHR;
  717. variable_pointers.pNext = &demote;
  718. VkPhysicalDeviceRobustness2FeaturesEXT robustness2{};
  719. robustness2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT;
  720. robustness2.pNext = &variable_pointers;
  721. VkPhysicalDeviceFeatures2KHR features2{};
  722. features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2;
  723. features2.pNext = &robustness2;
  724. physical.GetFeatures2KHR(features2);
  725. const VkPhysicalDeviceFeatures& features{features2.features};
  726. const std::array feature_report{
  727. std::make_pair(features.robustBufferAccess, "robustBufferAccess"),
  728. std::make_pair(features.vertexPipelineStoresAndAtomics, "vertexPipelineStoresAndAtomics"),
  729. std::make_pair(features.imageCubeArray, "imageCubeArray"),
  730. std::make_pair(features.independentBlend, "independentBlend"),
  731. std::make_pair(features.depthClamp, "depthClamp"),
  732. std::make_pair(features.samplerAnisotropy, "samplerAnisotropy"),
  733. std::make_pair(features.largePoints, "largePoints"),
  734. std::make_pair(features.multiViewport, "multiViewport"),
  735. std::make_pair(features.depthBiasClamp, "depthBiasClamp"),
  736. std::make_pair(features.fillModeNonSolid, "fillModeNonSolid"),
  737. std::make_pair(features.wideLines, "wideLines"),
  738. std::make_pair(features.geometryShader, "geometryShader"),
  739. std::make_pair(features.tessellationShader, "tessellationShader"),
  740. std::make_pair(features.sampleRateShading, "sampleRateShading"),
  741. std::make_pair(features.dualSrcBlend, "dualSrcBlend"),
  742. std::make_pair(features.occlusionQueryPrecise, "occlusionQueryPrecise"),
  743. std::make_pair(features.fragmentStoresAndAtomics, "fragmentStoresAndAtomics"),
  744. std::make_pair(features.shaderImageGatherExtended, "shaderImageGatherExtended"),
  745. std::make_pair(features.shaderStorageImageWriteWithoutFormat,
  746. "shaderStorageImageWriteWithoutFormat"),
  747. std::make_pair(features.shaderClipDistance, "shaderClipDistance"),
  748. std::make_pair(features.shaderCullDistance, "shaderCullDistance"),
  749. std::make_pair(demote.shaderDemoteToHelperInvocation, "shaderDemoteToHelperInvocation"),
  750. std::make_pair(variable_pointers.variablePointers, "variablePointers"),
  751. std::make_pair(variable_pointers.variablePointersStorageBuffer,
  752. "variablePointersStorageBuffer"),
  753. std::make_pair(robustness2.robustBufferAccess2, "robustBufferAccess2"),
  754. std::make_pair(robustness2.robustImageAccess2, "robustImageAccess2"),
  755. std::make_pair(robustness2.nullDescriptor, "nullDescriptor"),
  756. };
  757. for (const auto& [is_supported, name] : feature_report) {
  758. if (is_supported) {
  759. continue;
  760. }
  761. LOG_ERROR(Render_Vulkan, "Missing required feature: {}", name);
  762. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  763. }
  764. }
  765. std::vector<const char*> Device::LoadExtensions(bool requires_surface) {
  766. std::vector<const char*> extensions;
  767. extensions.reserve(8 + REQUIRED_EXTENSIONS.size());
  768. extensions.insert(extensions.begin(), REQUIRED_EXTENSIONS.begin(), REQUIRED_EXTENSIONS.end());
  769. if (requires_surface) {
  770. extensions.push_back(VK_KHR_SWAPCHAIN_EXTENSION_NAME);
  771. }
  772. bool has_khr_shader_float16_int8{};
  773. bool has_khr_workgroup_memory_explicit_layout{};
  774. bool has_khr_pipeline_executable_properties{};
  775. bool has_ext_subgroup_size_control{};
  776. bool has_ext_transform_feedback{};
  777. bool has_ext_custom_border_color{};
  778. bool has_ext_extended_dynamic_state{};
  779. bool has_ext_shader_atomic_int64{};
  780. bool has_ext_provoking_vertex{};
  781. bool has_ext_vertex_input_dynamic_state{};
  782. bool has_ext_line_rasterization{};
  783. for (const std::string& extension : supported_extensions) {
  784. const auto test = [&](std::optional<std::reference_wrapper<bool>> status, const char* name,
  785. bool push) {
  786. if (extension != name) {
  787. return;
  788. }
  789. if (push) {
  790. extensions.push_back(name);
  791. }
  792. if (status) {
  793. status->get() = true;
  794. }
  795. };
  796. test(nv_viewport_swizzle, VK_NV_VIEWPORT_SWIZZLE_EXTENSION_NAME, true);
  797. test(nv_viewport_array2, VK_NV_VIEWPORT_ARRAY2_EXTENSION_NAME, true);
  798. test(nv_geometry_shader_passthrough, VK_NV_GEOMETRY_SHADER_PASSTHROUGH_EXTENSION_NAME,
  799. true);
  800. test(khr_uniform_buffer_standard_layout,
  801. VK_KHR_UNIFORM_BUFFER_STANDARD_LAYOUT_EXTENSION_NAME, true);
  802. test(khr_spirv_1_4, VK_KHR_SPIRV_1_4_EXTENSION_NAME, true);
  803. test(khr_push_descriptor, VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME, true);
  804. test(has_khr_shader_float16_int8, VK_KHR_SHADER_FLOAT16_INT8_EXTENSION_NAME, false);
  805. test(ext_depth_range_unrestricted, VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME, true);
  806. test(ext_index_type_uint8, VK_EXT_INDEX_TYPE_UINT8_EXTENSION_NAME, true);
  807. test(ext_sampler_filter_minmax, VK_EXT_SAMPLER_FILTER_MINMAX_EXTENSION_NAME, true);
  808. test(ext_shader_viewport_index_layer, VK_EXT_SHADER_VIEWPORT_INDEX_LAYER_EXTENSION_NAME,
  809. true);
  810. test(ext_tooling_info, VK_EXT_TOOLING_INFO_EXTENSION_NAME, true);
  811. test(ext_shader_stencil_export, VK_EXT_SHADER_STENCIL_EXPORT_EXTENSION_NAME, true);
  812. test(ext_conservative_rasterization, VK_EXT_CONSERVATIVE_RASTERIZATION_EXTENSION_NAME,
  813. true);
  814. test(has_ext_transform_feedback, VK_EXT_TRANSFORM_FEEDBACK_EXTENSION_NAME, false);
  815. test(has_ext_custom_border_color, VK_EXT_CUSTOM_BORDER_COLOR_EXTENSION_NAME, false);
  816. test(has_ext_extended_dynamic_state, VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME, false);
  817. test(has_ext_subgroup_size_control, VK_EXT_SUBGROUP_SIZE_CONTROL_EXTENSION_NAME, false);
  818. test(has_ext_provoking_vertex, VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME, false);
  819. test(has_ext_vertex_input_dynamic_state, VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME,
  820. false);
  821. test(has_ext_shader_atomic_int64, VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME, false);
  822. test(has_khr_workgroup_memory_explicit_layout,
  823. VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME, false);
  824. test(has_ext_line_rasterization, VK_EXT_LINE_RASTERIZATION_EXTENSION_NAME, false);
  825. if (Settings::values.enable_nsight_aftermath) {
  826. test(nv_device_diagnostics_config, VK_NV_DEVICE_DIAGNOSTICS_CONFIG_EXTENSION_NAME,
  827. true);
  828. }
  829. if (Settings::values.renderer_shader_feedback) {
  830. test(has_khr_pipeline_executable_properties,
  831. VK_KHR_PIPELINE_EXECUTABLE_PROPERTIES_EXTENSION_NAME, false);
  832. }
  833. }
  834. VkPhysicalDeviceFeatures2KHR features{};
  835. features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2_KHR;
  836. VkPhysicalDeviceProperties2KHR physical_properties;
  837. physical_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR;
  838. if (has_khr_shader_float16_int8) {
  839. VkPhysicalDeviceFloat16Int8FeaturesKHR float16_int8_features;
  840. float16_int8_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR;
  841. float16_int8_features.pNext = nullptr;
  842. features.pNext = &float16_int8_features;
  843. physical.GetFeatures2KHR(features);
  844. is_float16_supported = float16_int8_features.shaderFloat16;
  845. extensions.push_back(VK_KHR_SHADER_FLOAT16_INT8_EXTENSION_NAME);
  846. }
  847. if (has_ext_subgroup_size_control) {
  848. VkPhysicalDeviceSubgroupSizeControlFeaturesEXT subgroup_features;
  849. subgroup_features.sType =
  850. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT;
  851. subgroup_features.pNext = nullptr;
  852. features.pNext = &subgroup_features;
  853. physical.GetFeatures2KHR(features);
  854. VkPhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_properties;
  855. subgroup_properties.sType =
  856. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT;
  857. subgroup_properties.pNext = nullptr;
  858. physical_properties.pNext = &subgroup_properties;
  859. physical.GetProperties2KHR(physical_properties);
  860. is_warp_potentially_bigger = subgroup_properties.maxSubgroupSize > GuestWarpSize;
  861. if (subgroup_features.subgroupSizeControl &&
  862. subgroup_properties.minSubgroupSize <= GuestWarpSize &&
  863. subgroup_properties.maxSubgroupSize >= GuestWarpSize) {
  864. extensions.push_back(VK_EXT_SUBGROUP_SIZE_CONTROL_EXTENSION_NAME);
  865. guest_warp_stages = subgroup_properties.requiredSubgroupSizeStages;
  866. ext_subgroup_size_control = true;
  867. }
  868. } else {
  869. is_warp_potentially_bigger = true;
  870. }
  871. if (has_ext_provoking_vertex) {
  872. VkPhysicalDeviceProvokingVertexFeaturesEXT provoking_vertex;
  873. provoking_vertex.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_FEATURES_EXT;
  874. provoking_vertex.pNext = nullptr;
  875. features.pNext = &provoking_vertex;
  876. physical.GetFeatures2KHR(features);
  877. if (provoking_vertex.provokingVertexLast &&
  878. provoking_vertex.transformFeedbackPreservesProvokingVertex) {
  879. extensions.push_back(VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME);
  880. ext_provoking_vertex = true;
  881. }
  882. }
  883. if (has_ext_vertex_input_dynamic_state) {
  884. VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT vertex_input;
  885. vertex_input.sType =
  886. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_INPUT_DYNAMIC_STATE_FEATURES_EXT;
  887. vertex_input.pNext = nullptr;
  888. features.pNext = &vertex_input;
  889. physical.GetFeatures2KHR(features);
  890. if (vertex_input.vertexInputDynamicState) {
  891. extensions.push_back(VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME);
  892. ext_vertex_input_dynamic_state = true;
  893. }
  894. }
  895. if (has_ext_shader_atomic_int64) {
  896. VkPhysicalDeviceShaderAtomicInt64Features atomic_int64;
  897. atomic_int64.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT;
  898. atomic_int64.pNext = nullptr;
  899. features.pNext = &atomic_int64;
  900. physical.GetFeatures2KHR(features);
  901. if (atomic_int64.shaderBufferInt64Atomics && atomic_int64.shaderSharedInt64Atomics) {
  902. extensions.push_back(VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME);
  903. ext_shader_atomic_int64 = true;
  904. }
  905. }
  906. if (has_ext_transform_feedback) {
  907. VkPhysicalDeviceTransformFeedbackFeaturesEXT tfb_features;
  908. tfb_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT;
  909. tfb_features.pNext = nullptr;
  910. features.pNext = &tfb_features;
  911. physical.GetFeatures2KHR(features);
  912. VkPhysicalDeviceTransformFeedbackPropertiesEXT tfb_properties;
  913. tfb_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT;
  914. tfb_properties.pNext = nullptr;
  915. physical_properties.pNext = &tfb_properties;
  916. physical.GetProperties2KHR(physical_properties);
  917. if (tfb_features.transformFeedback && tfb_features.geometryStreams &&
  918. tfb_properties.maxTransformFeedbackStreams >= 4 &&
  919. tfb_properties.maxTransformFeedbackBuffers && tfb_properties.transformFeedbackQueries &&
  920. tfb_properties.transformFeedbackDraw) {
  921. extensions.push_back(VK_EXT_TRANSFORM_FEEDBACK_EXTENSION_NAME);
  922. ext_transform_feedback = true;
  923. }
  924. }
  925. if (has_ext_custom_border_color) {
  926. VkPhysicalDeviceCustomBorderColorFeaturesEXT border_features;
  927. border_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT;
  928. border_features.pNext = nullptr;
  929. features.pNext = &border_features;
  930. physical.GetFeatures2KHR(features);
  931. if (border_features.customBorderColors && border_features.customBorderColorWithoutFormat) {
  932. extensions.push_back(VK_EXT_CUSTOM_BORDER_COLOR_EXTENSION_NAME);
  933. ext_custom_border_color = true;
  934. }
  935. }
  936. if (has_ext_extended_dynamic_state) {
  937. VkPhysicalDeviceExtendedDynamicStateFeaturesEXT extended_dynamic_state;
  938. extended_dynamic_state.sType =
  939. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT;
  940. extended_dynamic_state.pNext = nullptr;
  941. features.pNext = &extended_dynamic_state;
  942. physical.GetFeatures2KHR(features);
  943. if (extended_dynamic_state.extendedDynamicState) {
  944. extensions.push_back(VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME);
  945. ext_extended_dynamic_state = true;
  946. }
  947. }
  948. if (has_ext_line_rasterization) {
  949. VkPhysicalDeviceLineRasterizationFeaturesEXT line_raster;
  950. line_raster.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT;
  951. line_raster.pNext = nullptr;
  952. features.pNext = &line_raster;
  953. physical.GetFeatures2KHR(features);
  954. if (line_raster.rectangularLines && line_raster.smoothLines) {
  955. extensions.push_back(VK_EXT_LINE_RASTERIZATION_EXTENSION_NAME);
  956. ext_line_rasterization = true;
  957. }
  958. }
  959. if (has_khr_workgroup_memory_explicit_layout) {
  960. VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR layout;
  961. layout.sType =
  962. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR;
  963. layout.pNext = nullptr;
  964. features.pNext = &layout;
  965. physical.GetFeatures2KHR(features);
  966. if (layout.workgroupMemoryExplicitLayout &&
  967. layout.workgroupMemoryExplicitLayout8BitAccess &&
  968. layout.workgroupMemoryExplicitLayout16BitAccess &&
  969. layout.workgroupMemoryExplicitLayoutScalarBlockLayout) {
  970. extensions.push_back(VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME);
  971. khr_workgroup_memory_explicit_layout = true;
  972. }
  973. }
  974. if (has_khr_pipeline_executable_properties) {
  975. VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR executable_properties;
  976. executable_properties.sType =
  977. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR;
  978. executable_properties.pNext = nullptr;
  979. features.pNext = &executable_properties;
  980. physical.GetFeatures2KHR(features);
  981. if (executable_properties.pipelineExecutableInfo) {
  982. extensions.push_back(VK_KHR_PIPELINE_EXECUTABLE_PROPERTIES_EXTENSION_NAME);
  983. khr_pipeline_executable_properties = true;
  984. }
  985. }
  986. if (khr_push_descriptor) {
  987. VkPhysicalDevicePushDescriptorPropertiesKHR push_descriptor;
  988. push_descriptor.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR;
  989. push_descriptor.pNext = nullptr;
  990. physical_properties.pNext = &push_descriptor;
  991. physical.GetProperties2KHR(physical_properties);
  992. max_push_descriptors = push_descriptor.maxPushDescriptors;
  993. }
  994. return extensions;
  995. }
  996. void Device::SetupFamilies(VkSurfaceKHR surface) {
  997. const std::vector queue_family_properties = physical.GetQueueFamilyProperties();
  998. std::optional<u32> graphics;
  999. std::optional<u32> present;
  1000. for (u32 index = 0; index < static_cast<u32>(queue_family_properties.size()); ++index) {
  1001. if (graphics && (present || !surface)) {
  1002. break;
  1003. }
  1004. const VkQueueFamilyProperties& queue_family = queue_family_properties[index];
  1005. if (queue_family.queueCount == 0) {
  1006. continue;
  1007. }
  1008. if (queue_family.queueFlags & VK_QUEUE_GRAPHICS_BIT) {
  1009. graphics = index;
  1010. }
  1011. if (surface && physical.GetSurfaceSupportKHR(index, surface)) {
  1012. present = index;
  1013. }
  1014. }
  1015. if (!graphics) {
  1016. LOG_ERROR(Render_Vulkan, "Device lacks a graphics queue");
  1017. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  1018. }
  1019. if (surface && !present) {
  1020. LOG_ERROR(Render_Vulkan, "Device lacks a present queue");
  1021. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  1022. }
  1023. graphics_family = *graphics;
  1024. present_family = *present;
  1025. }
  1026. void Device::SetupFeatures() {
  1027. const VkPhysicalDeviceFeatures features{physical.GetFeatures()};
  1028. is_depth_bounds_supported = features.depthBounds;
  1029. is_formatless_image_load_supported = features.shaderStorageImageReadWithoutFormat;
  1030. is_shader_float64_supported = features.shaderFloat64;
  1031. is_shader_int64_supported = features.shaderInt64;
  1032. is_shader_int16_supported = features.shaderInt16;
  1033. is_shader_storage_image_multisample = features.shaderStorageImageMultisample;
  1034. is_blit_depth_stencil_supported = TestDepthStencilBlits();
  1035. is_optimal_astc_supported = IsOptimalAstcSupported(features);
  1036. }
  1037. void Device::SetupProperties() {
  1038. float_controls.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR;
  1039. VkPhysicalDeviceProperties2KHR properties2{};
  1040. properties2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR;
  1041. properties2.pNext = &float_controls;
  1042. physical.GetProperties2KHR(properties2);
  1043. }
  1044. void Device::CollectTelemetryParameters() {
  1045. VkPhysicalDeviceDriverPropertiesKHR driver{
  1046. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR,
  1047. .pNext = nullptr,
  1048. .driverID = {},
  1049. .driverName = {},
  1050. .driverInfo = {},
  1051. .conformanceVersion = {},
  1052. };
  1053. VkPhysicalDeviceProperties2KHR device_properties{
  1054. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR,
  1055. .pNext = &driver,
  1056. .properties = {},
  1057. };
  1058. physical.GetProperties2KHR(device_properties);
  1059. driver_id = driver.driverID;
  1060. vendor_name = driver.driverName;
  1061. }
  1062. void Device::CollectPhysicalMemoryInfo() {
  1063. const auto mem_properties = physical.GetMemoryProperties();
  1064. const size_t num_properties = mem_properties.memoryHeapCount;
  1065. device_access_memory = 0;
  1066. for (size_t element = 0; element < num_properties; ++element) {
  1067. if ((mem_properties.memoryHeaps[element].flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) != 0) {
  1068. device_access_memory += mem_properties.memoryHeaps[element].size;
  1069. }
  1070. }
  1071. }
  1072. void Device::CollectToolingInfo() {
  1073. if (!ext_tooling_info) {
  1074. return;
  1075. }
  1076. const auto vkGetPhysicalDeviceToolPropertiesEXT =
  1077. reinterpret_cast<PFN_vkGetPhysicalDeviceToolPropertiesEXT>(
  1078. dld.vkGetInstanceProcAddr(instance, "vkGetPhysicalDeviceToolPropertiesEXT"));
  1079. if (!vkGetPhysicalDeviceToolPropertiesEXT) {
  1080. return;
  1081. }
  1082. u32 tool_count = 0;
  1083. if (vkGetPhysicalDeviceToolPropertiesEXT(physical, &tool_count, nullptr) != VK_SUCCESS) {
  1084. return;
  1085. }
  1086. std::vector<VkPhysicalDeviceToolPropertiesEXT> tools(tool_count);
  1087. if (vkGetPhysicalDeviceToolPropertiesEXT(physical, &tool_count, tools.data()) != VK_SUCCESS) {
  1088. return;
  1089. }
  1090. for (const VkPhysicalDeviceToolPropertiesEXT& tool : tools) {
  1091. const std::string_view name = tool.name;
  1092. LOG_INFO(Render_Vulkan, "{}", name);
  1093. has_renderdoc = has_renderdoc || name == "RenderDoc";
  1094. has_nsight_graphics = has_nsight_graphics || name == "NVIDIA Nsight Graphics";
  1095. }
  1096. }
  1097. std::vector<VkDeviceQueueCreateInfo> Device::GetDeviceQueueCreateInfos() const {
  1098. static constexpr float QUEUE_PRIORITY = 1.0f;
  1099. std::unordered_set<u32> unique_queue_families{graphics_family, present_family};
  1100. std::vector<VkDeviceQueueCreateInfo> queue_cis;
  1101. queue_cis.reserve(unique_queue_families.size());
  1102. for (const u32 queue_family : unique_queue_families) {
  1103. auto& ci = queue_cis.emplace_back(VkDeviceQueueCreateInfo{
  1104. .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
  1105. .pNext = nullptr,
  1106. .flags = 0,
  1107. .queueFamilyIndex = queue_family,
  1108. .queueCount = 1,
  1109. .pQueuePriorities = nullptr,
  1110. });
  1111. ci.pQueuePriorities = &QUEUE_PRIORITY;
  1112. }
  1113. return queue_cis;
  1114. }
  1115. } // namespace Vulkan