vulkan_device.cpp 52 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266
  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 STENCIL8_UINT{
  22. VK_FORMAT_D16_UNORM_S8_UINT,
  23. VK_FORMAT_D24_UNORM_S8_UINT,
  24. VK_FORMAT_D32_SFLOAT_S8_UINT,
  25. VK_FORMAT_UNDEFINED,
  26. };
  27. constexpr std::array DEPTH24_UNORM_STENCIL8_UINT{
  28. VK_FORMAT_D32_SFLOAT_S8_UINT,
  29. VK_FORMAT_D16_UNORM_S8_UINT,
  30. VK_FORMAT_UNDEFINED,
  31. };
  32. constexpr std::array DEPTH16_UNORM_STENCIL8_UINT{
  33. VK_FORMAT_D24_UNORM_S8_UINT,
  34. VK_FORMAT_D32_SFLOAT_S8_UINT,
  35. VK_FORMAT_UNDEFINED,
  36. };
  37. } // namespace Alternatives
  38. enum class NvidiaArchitecture {
  39. AmpereOrNewer,
  40. Turing,
  41. VoltaOrOlder,
  42. };
  43. constexpr std::array REQUIRED_EXTENSIONS{
  44. VK_KHR_MAINTENANCE1_EXTENSION_NAME,
  45. VK_KHR_STORAGE_BUFFER_STORAGE_CLASS_EXTENSION_NAME,
  46. VK_KHR_SHADER_DRAW_PARAMETERS_EXTENSION_NAME,
  47. VK_KHR_16BIT_STORAGE_EXTENSION_NAME,
  48. VK_KHR_8BIT_STORAGE_EXTENSION_NAME,
  49. VK_KHR_DRIVER_PROPERTIES_EXTENSION_NAME,
  50. VK_KHR_DESCRIPTOR_UPDATE_TEMPLATE_EXTENSION_NAME,
  51. VK_KHR_TIMELINE_SEMAPHORE_EXTENSION_NAME,
  52. VK_KHR_SAMPLER_MIRROR_CLAMP_TO_EDGE_EXTENSION_NAME,
  53. VK_KHR_SHADER_FLOAT_CONTROLS_EXTENSION_NAME,
  54. VK_KHR_VARIABLE_POINTERS_EXTENSION_NAME,
  55. VK_EXT_VERTEX_ATTRIBUTE_DIVISOR_EXTENSION_NAME,
  56. VK_EXT_SHADER_SUBGROUP_BALLOT_EXTENSION_NAME,
  57. VK_EXT_SHADER_SUBGROUP_VOTE_EXTENSION_NAME,
  58. VK_EXT_ROBUSTNESS_2_EXTENSION_NAME,
  59. VK_EXT_HOST_QUERY_RESET_EXTENSION_NAME,
  60. VK_EXT_SHADER_DEMOTE_TO_HELPER_INVOCATION_EXTENSION_NAME,
  61. #ifdef _WIN32
  62. VK_KHR_EXTERNAL_MEMORY_WIN32_EXTENSION_NAME,
  63. #endif
  64. #ifdef __unix__
  65. VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME,
  66. #endif
  67. };
  68. template <typename T>
  69. void SetNext(void**& next, T& data) {
  70. *next = &data;
  71. next = &data.pNext;
  72. }
  73. constexpr const VkFormat* GetFormatAlternatives(VkFormat format) {
  74. switch (format) {
  75. case VK_FORMAT_S8_UINT:
  76. return Alternatives::STENCIL8_UINT.data();
  77. case VK_FORMAT_D24_UNORM_S8_UINT:
  78. return Alternatives::DEPTH24_UNORM_STENCIL8_UINT.data();
  79. case VK_FORMAT_D16_UNORM_S8_UINT:
  80. return Alternatives::DEPTH16_UNORM_STENCIL8_UINT.data();
  81. default:
  82. return nullptr;
  83. }
  84. }
  85. VkFormatFeatureFlags GetFormatFeatures(VkFormatProperties properties, FormatType format_type) {
  86. switch (format_type) {
  87. case FormatType::Linear:
  88. return properties.linearTilingFeatures;
  89. case FormatType::Optimal:
  90. return properties.optimalTilingFeatures;
  91. case FormatType::Buffer:
  92. return properties.bufferFeatures;
  93. default:
  94. return {};
  95. }
  96. }
  97. std::unordered_map<VkFormat, VkFormatProperties> GetFormatProperties(vk::PhysicalDevice physical) {
  98. static constexpr std::array formats{
  99. VK_FORMAT_A8B8G8R8_UNORM_PACK32,
  100. VK_FORMAT_A8B8G8R8_UINT_PACK32,
  101. VK_FORMAT_A8B8G8R8_SNORM_PACK32,
  102. VK_FORMAT_A8B8G8R8_SINT_PACK32,
  103. VK_FORMAT_A8B8G8R8_SRGB_PACK32,
  104. VK_FORMAT_R5G6B5_UNORM_PACK16,
  105. VK_FORMAT_B5G6R5_UNORM_PACK16,
  106. VK_FORMAT_A2B10G10R10_UNORM_PACK32,
  107. VK_FORMAT_A2B10G10R10_UINT_PACK32,
  108. VK_FORMAT_A1R5G5B5_UNORM_PACK16,
  109. VK_FORMAT_R32G32B32A32_SFLOAT,
  110. VK_FORMAT_R32G32B32A32_SINT,
  111. VK_FORMAT_R32G32B32A32_UINT,
  112. VK_FORMAT_R32G32_SFLOAT,
  113. VK_FORMAT_R32G32_SINT,
  114. VK_FORMAT_R32G32_UINT,
  115. VK_FORMAT_R16G16B16A16_SINT,
  116. VK_FORMAT_R16G16B16A16_UINT,
  117. VK_FORMAT_R16G16B16A16_SNORM,
  118. VK_FORMAT_R16G16B16A16_UNORM,
  119. VK_FORMAT_R16G16_UNORM,
  120. VK_FORMAT_R16G16_SNORM,
  121. VK_FORMAT_R16G16_SFLOAT,
  122. VK_FORMAT_R16G16_SINT,
  123. VK_FORMAT_R16_UNORM,
  124. VK_FORMAT_R16_SNORM,
  125. VK_FORMAT_R16_UINT,
  126. VK_FORMAT_R8G8B8A8_SRGB,
  127. VK_FORMAT_R8G8_UNORM,
  128. VK_FORMAT_R8G8_SNORM,
  129. VK_FORMAT_R8G8_SINT,
  130. VK_FORMAT_R8G8_UINT,
  131. VK_FORMAT_R8_UNORM,
  132. VK_FORMAT_R8_SNORM,
  133. VK_FORMAT_R8_SINT,
  134. VK_FORMAT_R8_UINT,
  135. VK_FORMAT_B10G11R11_UFLOAT_PACK32,
  136. VK_FORMAT_R32_SFLOAT,
  137. VK_FORMAT_R32_UINT,
  138. VK_FORMAT_R32_SINT,
  139. VK_FORMAT_R16_SFLOAT,
  140. VK_FORMAT_R16G16B16A16_SFLOAT,
  141. VK_FORMAT_B8G8R8A8_UNORM,
  142. VK_FORMAT_B8G8R8A8_SRGB,
  143. VK_FORMAT_R4G4B4A4_UNORM_PACK16,
  144. VK_FORMAT_D32_SFLOAT,
  145. VK_FORMAT_D16_UNORM,
  146. VK_FORMAT_S8_UINT,
  147. VK_FORMAT_D16_UNORM_S8_UINT,
  148. VK_FORMAT_D24_UNORM_S8_UINT,
  149. VK_FORMAT_D32_SFLOAT_S8_UINT,
  150. VK_FORMAT_BC1_RGBA_UNORM_BLOCK,
  151. VK_FORMAT_BC2_UNORM_BLOCK,
  152. VK_FORMAT_BC3_UNORM_BLOCK,
  153. VK_FORMAT_BC4_UNORM_BLOCK,
  154. VK_FORMAT_BC4_SNORM_BLOCK,
  155. VK_FORMAT_BC5_UNORM_BLOCK,
  156. VK_FORMAT_BC5_SNORM_BLOCK,
  157. VK_FORMAT_BC7_UNORM_BLOCK,
  158. VK_FORMAT_BC6H_UFLOAT_BLOCK,
  159. VK_FORMAT_BC6H_SFLOAT_BLOCK,
  160. VK_FORMAT_BC1_RGBA_SRGB_BLOCK,
  161. VK_FORMAT_BC2_SRGB_BLOCK,
  162. VK_FORMAT_BC3_SRGB_BLOCK,
  163. VK_FORMAT_BC7_SRGB_BLOCK,
  164. VK_FORMAT_ASTC_4x4_UNORM_BLOCK,
  165. VK_FORMAT_ASTC_4x4_SRGB_BLOCK,
  166. VK_FORMAT_ASTC_5x4_UNORM_BLOCK,
  167. VK_FORMAT_ASTC_5x4_SRGB_BLOCK,
  168. VK_FORMAT_ASTC_5x5_UNORM_BLOCK,
  169. VK_FORMAT_ASTC_5x5_SRGB_BLOCK,
  170. VK_FORMAT_ASTC_6x5_UNORM_BLOCK,
  171. VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
  172. VK_FORMAT_ASTC_6x6_UNORM_BLOCK,
  173. VK_FORMAT_ASTC_6x6_SRGB_BLOCK,
  174. VK_FORMAT_ASTC_8x5_UNORM_BLOCK,
  175. VK_FORMAT_ASTC_8x5_SRGB_BLOCK,
  176. VK_FORMAT_ASTC_8x6_UNORM_BLOCK,
  177. VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
  178. VK_FORMAT_ASTC_8x8_UNORM_BLOCK,
  179. VK_FORMAT_ASTC_8x8_SRGB_BLOCK,
  180. VK_FORMAT_ASTC_10x5_UNORM_BLOCK,
  181. VK_FORMAT_ASTC_10x5_SRGB_BLOCK,
  182. VK_FORMAT_ASTC_10x6_UNORM_BLOCK,
  183. VK_FORMAT_ASTC_10x6_SRGB_BLOCK,
  184. VK_FORMAT_ASTC_10x8_UNORM_BLOCK,
  185. VK_FORMAT_ASTC_10x8_SRGB_BLOCK,
  186. VK_FORMAT_ASTC_10x10_UNORM_BLOCK,
  187. VK_FORMAT_ASTC_10x10_SRGB_BLOCK,
  188. VK_FORMAT_ASTC_12x10_UNORM_BLOCK,
  189. VK_FORMAT_ASTC_12x10_SRGB_BLOCK,
  190. VK_FORMAT_ASTC_12x12_UNORM_BLOCK,
  191. VK_FORMAT_ASTC_12x12_SRGB_BLOCK,
  192. VK_FORMAT_ASTC_8x6_UNORM_BLOCK,
  193. VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
  194. VK_FORMAT_ASTC_6x5_UNORM_BLOCK,
  195. VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
  196. VK_FORMAT_E5B9G9R9_UFLOAT_PACK32,
  197. };
  198. std::unordered_map<VkFormat, VkFormatProperties> format_properties;
  199. for (const auto format : formats) {
  200. format_properties.emplace(format, physical.GetFormatProperties(format));
  201. }
  202. return format_properties;
  203. }
  204. std::vector<std::string> GetSupportedExtensions(vk::PhysicalDevice physical) {
  205. const std::vector extensions = physical.EnumerateDeviceExtensionProperties();
  206. std::vector<std::string> supported_extensions;
  207. supported_extensions.reserve(extensions.size());
  208. for (const auto& extension : extensions) {
  209. supported_extensions.emplace_back(extension.extensionName);
  210. }
  211. return supported_extensions;
  212. }
  213. NvidiaArchitecture GetNvidiaArchitecture(vk::PhysicalDevice physical,
  214. std::span<const std::string> exts) {
  215. if (std::ranges::find(exts, VK_KHR_FRAGMENT_SHADING_RATE_EXTENSION_NAME) != exts.end()) {
  216. VkPhysicalDeviceFragmentShadingRatePropertiesKHR shading_rate_props{};
  217. shading_rate_props.sType =
  218. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FRAGMENT_SHADING_RATE_PROPERTIES_KHR;
  219. VkPhysicalDeviceProperties2KHR physical_properties{};
  220. physical_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR;
  221. physical_properties.pNext = &shading_rate_props;
  222. physical.GetProperties2KHR(physical_properties);
  223. if (shading_rate_props.primitiveFragmentShadingRateWithMultipleViewports) {
  224. // Only Ampere and newer support this feature
  225. return NvidiaArchitecture::AmpereOrNewer;
  226. }
  227. }
  228. if (std::ranges::find(exts, VK_NV_SHADING_RATE_IMAGE_EXTENSION_NAME) != exts.end()) {
  229. return NvidiaArchitecture::Turing;
  230. }
  231. return NvidiaArchitecture::VoltaOrOlder;
  232. }
  233. } // Anonymous namespace
  234. Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR surface,
  235. const vk::InstanceDispatch& dld_)
  236. : instance{instance_}, dld{dld_}, physical{physical_}, properties{physical.GetProperties()},
  237. supported_extensions{GetSupportedExtensions(physical)},
  238. format_properties(GetFormatProperties(physical)) {
  239. CheckSuitability(surface != nullptr);
  240. SetupFamilies(surface);
  241. SetupFeatures();
  242. SetupProperties();
  243. const auto queue_cis = GetDeviceQueueCreateInfos();
  244. const std::vector extensions = LoadExtensions(surface != nullptr);
  245. VkPhysicalDeviceFeatures2 features2{
  246. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2,
  247. .pNext = nullptr,
  248. .features{
  249. .robustBufferAccess = true,
  250. .fullDrawIndexUint32 = false,
  251. .imageCubeArray = true,
  252. .independentBlend = true,
  253. .geometryShader = true,
  254. .tessellationShader = true,
  255. .sampleRateShading = true,
  256. .dualSrcBlend = true,
  257. .logicOp = false,
  258. .multiDrawIndirect = false,
  259. .drawIndirectFirstInstance = false,
  260. .depthClamp = true,
  261. .depthBiasClamp = true,
  262. .fillModeNonSolid = true,
  263. .depthBounds = is_depth_bounds_supported,
  264. .wideLines = true,
  265. .largePoints = true,
  266. .alphaToOne = false,
  267. .multiViewport = true,
  268. .samplerAnisotropy = true,
  269. .textureCompressionETC2 = false,
  270. .textureCompressionASTC_LDR = is_optimal_astc_supported,
  271. .textureCompressionBC = false,
  272. .occlusionQueryPrecise = true,
  273. .pipelineStatisticsQuery = false,
  274. .vertexPipelineStoresAndAtomics = true,
  275. .fragmentStoresAndAtomics = true,
  276. .shaderTessellationAndGeometryPointSize = false,
  277. .shaderImageGatherExtended = true,
  278. .shaderStorageImageExtendedFormats = false,
  279. .shaderStorageImageMultisample = is_shader_storage_image_multisample,
  280. .shaderStorageImageReadWithoutFormat = is_formatless_image_load_supported,
  281. .shaderStorageImageWriteWithoutFormat = true,
  282. .shaderUniformBufferArrayDynamicIndexing = false,
  283. .shaderSampledImageArrayDynamicIndexing = false,
  284. .shaderStorageBufferArrayDynamicIndexing = false,
  285. .shaderStorageImageArrayDynamicIndexing = false,
  286. .shaderClipDistance = true,
  287. .shaderCullDistance = true,
  288. .shaderFloat64 = is_shader_float64_supported,
  289. .shaderInt64 = is_shader_int64_supported,
  290. .shaderInt16 = is_shader_int16_supported,
  291. .shaderResourceResidency = false,
  292. .shaderResourceMinLod = false,
  293. .sparseBinding = false,
  294. .sparseResidencyBuffer = false,
  295. .sparseResidencyImage2D = false,
  296. .sparseResidencyImage3D = false,
  297. .sparseResidency2Samples = false,
  298. .sparseResidency4Samples = false,
  299. .sparseResidency8Samples = false,
  300. .sparseResidency16Samples = false,
  301. .sparseResidencyAliased = false,
  302. .variableMultisampleRate = false,
  303. .inheritedQueries = false,
  304. },
  305. };
  306. const void* first_next = &features2;
  307. void** next = &features2.pNext;
  308. VkPhysicalDeviceTimelineSemaphoreFeaturesKHR timeline_semaphore{
  309. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES_KHR,
  310. .pNext = nullptr,
  311. .timelineSemaphore = true,
  312. };
  313. SetNext(next, timeline_semaphore);
  314. VkPhysicalDevice16BitStorageFeaturesKHR bit16_storage{
  315. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES_KHR,
  316. .pNext = nullptr,
  317. .storageBuffer16BitAccess = true,
  318. .uniformAndStorageBuffer16BitAccess = true,
  319. .storagePushConstant16 = false,
  320. .storageInputOutput16 = false,
  321. };
  322. SetNext(next, bit16_storage);
  323. VkPhysicalDevice8BitStorageFeaturesKHR bit8_storage{
  324. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_8BIT_STORAGE_FEATURES_KHR,
  325. .pNext = nullptr,
  326. .storageBuffer8BitAccess = false,
  327. .uniformAndStorageBuffer8BitAccess = true,
  328. .storagePushConstant8 = false,
  329. };
  330. SetNext(next, bit8_storage);
  331. VkPhysicalDeviceRobustness2FeaturesEXT robustness2{
  332. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT,
  333. .pNext = nullptr,
  334. .robustBufferAccess2 = true,
  335. .robustImageAccess2 = true,
  336. .nullDescriptor = true,
  337. };
  338. SetNext(next, robustness2);
  339. VkPhysicalDeviceHostQueryResetFeaturesEXT host_query_reset{
  340. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_HOST_QUERY_RESET_FEATURES_EXT,
  341. .pNext = nullptr,
  342. .hostQueryReset = true,
  343. };
  344. SetNext(next, host_query_reset);
  345. VkPhysicalDeviceVariablePointerFeaturesKHR variable_pointers{
  346. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES_KHR,
  347. .pNext = nullptr,
  348. .variablePointersStorageBuffer = VK_TRUE,
  349. .variablePointers = VK_TRUE,
  350. };
  351. SetNext(next, variable_pointers);
  352. VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT demote{
  353. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT,
  354. .pNext = nullptr,
  355. .shaderDemoteToHelperInvocation = true,
  356. };
  357. SetNext(next, demote);
  358. VkPhysicalDeviceFloat16Int8FeaturesKHR float16_int8;
  359. if (is_int8_supported || is_float16_supported) {
  360. float16_int8 = {
  361. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR,
  362. .pNext = nullptr,
  363. .shaderFloat16 = is_float16_supported,
  364. .shaderInt8 = is_int8_supported,
  365. };
  366. SetNext(next, float16_int8);
  367. }
  368. if (!is_float16_supported) {
  369. LOG_INFO(Render_Vulkan, "Device doesn't support float16 natively");
  370. }
  371. if (!is_int8_supported) {
  372. LOG_INFO(Render_Vulkan, "Device doesn't support int8 natively");
  373. }
  374. if (!nv_viewport_swizzle) {
  375. LOG_INFO(Render_Vulkan, "Device doesn't support viewport swizzles");
  376. }
  377. if (!nv_viewport_array2) {
  378. LOG_INFO(Render_Vulkan, "Device doesn't support viewport masks");
  379. }
  380. if (!nv_geometry_shader_passthrough) {
  381. LOG_INFO(Render_Vulkan, "Device doesn't support passthrough geometry shaders");
  382. }
  383. VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR std430_layout;
  384. if (khr_uniform_buffer_standard_layout) {
  385. std430_layout = {
  386. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR,
  387. .pNext = nullptr,
  388. .uniformBufferStandardLayout = true,
  389. };
  390. SetNext(next, std430_layout);
  391. } else {
  392. LOG_INFO(Render_Vulkan, "Device doesn't support packed UBOs");
  393. }
  394. VkPhysicalDeviceIndexTypeUint8FeaturesEXT index_type_uint8;
  395. if (ext_index_type_uint8) {
  396. index_type_uint8 = {
  397. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT,
  398. .pNext = nullptr,
  399. .indexTypeUint8 = true,
  400. };
  401. SetNext(next, index_type_uint8);
  402. } else {
  403. LOG_INFO(Render_Vulkan, "Device doesn't support uint8 indexes");
  404. }
  405. VkPhysicalDeviceTransformFeedbackFeaturesEXT transform_feedback;
  406. if (ext_transform_feedback) {
  407. transform_feedback = {
  408. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT,
  409. .pNext = nullptr,
  410. .transformFeedback = true,
  411. .geometryStreams = true,
  412. };
  413. SetNext(next, transform_feedback);
  414. } else {
  415. LOG_INFO(Render_Vulkan, "Device doesn't support transform feedbacks");
  416. }
  417. VkPhysicalDeviceCustomBorderColorFeaturesEXT custom_border;
  418. if (ext_custom_border_color) {
  419. custom_border = {
  420. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT,
  421. .pNext = nullptr,
  422. .customBorderColors = VK_TRUE,
  423. .customBorderColorWithoutFormat = VK_TRUE,
  424. };
  425. SetNext(next, custom_border);
  426. } else {
  427. LOG_INFO(Render_Vulkan, "Device doesn't support custom border colors");
  428. }
  429. VkPhysicalDeviceExtendedDynamicStateFeaturesEXT dynamic_state;
  430. if (ext_extended_dynamic_state) {
  431. dynamic_state = {
  432. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT,
  433. .pNext = nullptr,
  434. .extendedDynamicState = VK_TRUE,
  435. };
  436. SetNext(next, dynamic_state);
  437. } else {
  438. LOG_INFO(Render_Vulkan, "Device doesn't support extended dynamic state");
  439. }
  440. VkPhysicalDeviceLineRasterizationFeaturesEXT line_raster;
  441. if (ext_line_rasterization) {
  442. line_raster = {
  443. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT,
  444. .pNext = nullptr,
  445. .rectangularLines = VK_TRUE,
  446. .bresenhamLines = VK_FALSE,
  447. .smoothLines = VK_TRUE,
  448. .stippledRectangularLines = VK_FALSE,
  449. .stippledBresenhamLines = VK_FALSE,
  450. .stippledSmoothLines = VK_FALSE,
  451. };
  452. SetNext(next, line_raster);
  453. } else {
  454. LOG_INFO(Render_Vulkan, "Device doesn't support smooth lines");
  455. }
  456. if (!ext_conservative_rasterization) {
  457. LOG_INFO(Render_Vulkan, "Device doesn't support conservative rasterization");
  458. }
  459. VkPhysicalDeviceProvokingVertexFeaturesEXT provoking_vertex;
  460. if (ext_provoking_vertex) {
  461. provoking_vertex = {
  462. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_FEATURES_EXT,
  463. .pNext = nullptr,
  464. .provokingVertexLast = VK_TRUE,
  465. .transformFeedbackPreservesProvokingVertex = VK_TRUE,
  466. };
  467. SetNext(next, provoking_vertex);
  468. } else {
  469. LOG_INFO(Render_Vulkan, "Device doesn't support provoking vertex last");
  470. }
  471. VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT vertex_input_dynamic;
  472. if (ext_vertex_input_dynamic_state) {
  473. vertex_input_dynamic = {
  474. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_INPUT_DYNAMIC_STATE_FEATURES_EXT,
  475. .pNext = nullptr,
  476. .vertexInputDynamicState = VK_TRUE,
  477. };
  478. SetNext(next, vertex_input_dynamic);
  479. } else {
  480. LOG_INFO(Render_Vulkan, "Device doesn't support vertex input dynamic state");
  481. }
  482. VkPhysicalDeviceShaderAtomicInt64FeaturesKHR atomic_int64;
  483. if (ext_shader_atomic_int64) {
  484. atomic_int64 = {
  485. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR,
  486. .pNext = nullptr,
  487. .shaderBufferInt64Atomics = VK_TRUE,
  488. .shaderSharedInt64Atomics = VK_TRUE,
  489. };
  490. SetNext(next, atomic_int64);
  491. }
  492. VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR workgroup_layout;
  493. if (khr_workgroup_memory_explicit_layout) {
  494. workgroup_layout = {
  495. .sType =
  496. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR,
  497. .pNext = nullptr,
  498. .workgroupMemoryExplicitLayout = VK_TRUE,
  499. .workgroupMemoryExplicitLayoutScalarBlockLayout = VK_TRUE,
  500. .workgroupMemoryExplicitLayout8BitAccess = VK_TRUE,
  501. .workgroupMemoryExplicitLayout16BitAccess = VK_TRUE,
  502. };
  503. SetNext(next, workgroup_layout);
  504. }
  505. VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR executable_properties;
  506. if (khr_pipeline_executable_properties) {
  507. LOG_INFO(Render_Vulkan, "Enabling shader feedback, expect slower shader build times");
  508. executable_properties = {
  509. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR,
  510. .pNext = nullptr,
  511. .pipelineExecutableInfo = VK_TRUE,
  512. };
  513. SetNext(next, executable_properties);
  514. }
  515. if (!ext_depth_range_unrestricted) {
  516. LOG_INFO(Render_Vulkan, "Device doesn't support depth range unrestricted");
  517. }
  518. VkDeviceDiagnosticsConfigCreateInfoNV diagnostics_nv;
  519. if (Settings::values.enable_nsight_aftermath && nv_device_diagnostics_config) {
  520. nsight_aftermath_tracker = std::make_unique<NsightAftermathTracker>();
  521. diagnostics_nv = {
  522. .sType = VK_STRUCTURE_TYPE_DEVICE_DIAGNOSTICS_CONFIG_CREATE_INFO_NV,
  523. .pNext = &features2,
  524. .flags = VK_DEVICE_DIAGNOSTICS_CONFIG_ENABLE_SHADER_DEBUG_INFO_BIT_NV |
  525. VK_DEVICE_DIAGNOSTICS_CONFIG_ENABLE_RESOURCE_TRACKING_BIT_NV |
  526. VK_DEVICE_DIAGNOSTICS_CONFIG_ENABLE_AUTOMATIC_CHECKPOINTS_BIT_NV,
  527. };
  528. first_next = &diagnostics_nv;
  529. }
  530. logical = vk::Device::Create(physical, queue_cis, extensions, first_next, dld);
  531. CollectPhysicalMemoryInfo();
  532. CollectTelemetryParameters();
  533. CollectToolingInfo();
  534. if (driver_id == VK_DRIVER_ID_NVIDIA_PROPRIETARY_KHR) {
  535. const auto arch = GetNvidiaArchitecture(physical, supported_extensions);
  536. switch (arch) {
  537. case NvidiaArchitecture::AmpereOrNewer:
  538. LOG_WARNING(Render_Vulkan, "Blacklisting Ampere devices from float16 math");
  539. is_float16_supported = false;
  540. break;
  541. case NvidiaArchitecture::Turing:
  542. break;
  543. case NvidiaArchitecture::VoltaOrOlder:
  544. LOG_WARNING(Render_Vulkan, "Blacklisting Volta and older from VK_KHR_push_descriptor");
  545. khr_push_descriptor = false;
  546. break;
  547. }
  548. }
  549. if (ext_extended_dynamic_state && driver_id == VK_DRIVER_ID_MESA_RADV) {
  550. // Mask driver version variant
  551. const u32 version = (properties.driverVersion << 3) >> 3;
  552. if (version < VK_MAKE_API_VERSION(0, 21, 2, 0)) {
  553. LOG_WARNING(Render_Vulkan,
  554. "RADV versions older than 21.2 have broken VK_EXT_extended_dynamic_state");
  555. ext_extended_dynamic_state = false;
  556. }
  557. }
  558. sets_per_pool = 64;
  559. const bool is_amd =
  560. driver_id == VK_DRIVER_ID_AMD_PROPRIETARY || driver_id == VK_DRIVER_ID_AMD_OPEN_SOURCE;
  561. if (is_amd) {
  562. // AMD drivers need a higher amount of Sets per Pool in certain circunstances like in XC2.
  563. sets_per_pool = 96;
  564. // Disable VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT on AMD GCN4 and lower as it is broken.
  565. if (!is_float16_supported) {
  566. LOG_WARNING(
  567. Render_Vulkan,
  568. "AMD GCN4 and earlier do not properly support VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT");
  569. has_broken_cube_compatibility = true;
  570. }
  571. }
  572. const bool is_amd_or_radv = is_amd || driver_id == VK_DRIVER_ID_MESA_RADV;
  573. if (ext_sampler_filter_minmax && is_amd_or_radv) {
  574. // Disable ext_sampler_filter_minmax on AMD GCN4 and lower as it is broken.
  575. if (!is_float16_supported) {
  576. LOG_WARNING(Render_Vulkan,
  577. "Blacklisting AMD GCN4 and earlier for VK_EXT_sampler_filter_minmax");
  578. ext_sampler_filter_minmax = false;
  579. }
  580. }
  581. if (ext_vertex_input_dynamic_state && driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
  582. LOG_WARNING(Render_Vulkan, "Blacklisting Intel for VK_EXT_vertex_input_dynamic_state");
  583. ext_vertex_input_dynamic_state = false;
  584. }
  585. if (is_float16_supported && driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
  586. // Intel's compiler crashes when using fp16 on Astral Chain, disable it for the time being.
  587. LOG_WARNING(Render_Vulkan, "Blacklisting Intel proprietary from float16 math");
  588. is_float16_supported = false;
  589. }
  590. supports_d24_depth =
  591. IsFormatSupported(VK_FORMAT_D24_UNORM_S8_UINT,
  592. VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT, FormatType::Optimal);
  593. graphics_queue = logical.GetQueue(graphics_family);
  594. present_queue = logical.GetQueue(present_family);
  595. }
  596. Device::~Device() = default;
  597. VkFormat Device::GetSupportedFormat(VkFormat wanted_format, VkFormatFeatureFlags wanted_usage,
  598. FormatType format_type) const {
  599. if (IsFormatSupported(wanted_format, wanted_usage, format_type)) {
  600. return wanted_format;
  601. }
  602. // The wanted format is not supported by hardware, search for alternatives
  603. const VkFormat* alternatives = GetFormatAlternatives(wanted_format);
  604. if (alternatives == nullptr) {
  605. UNREACHABLE_MSG("Format={} with usage={} and type={} has no defined alternatives and host "
  606. "hardware does not support it",
  607. wanted_format, wanted_usage, format_type);
  608. return wanted_format;
  609. }
  610. std::size_t i = 0;
  611. for (VkFormat alternative = *alternatives; alternative; alternative = alternatives[++i]) {
  612. if (!IsFormatSupported(alternative, wanted_usage, format_type)) {
  613. continue;
  614. }
  615. LOG_WARNING(Render_Vulkan,
  616. "Emulating format={} with alternative format={} with usage={} and type={}",
  617. wanted_format, alternative, wanted_usage, format_type);
  618. return alternative;
  619. }
  620. // No alternatives found, panic
  621. UNREACHABLE_MSG("Format={} with usage={} and type={} is not supported by the host hardware and "
  622. "doesn't support any of the alternatives",
  623. wanted_format, wanted_usage, format_type);
  624. return wanted_format;
  625. }
  626. void Device::ReportLoss() const {
  627. LOG_CRITICAL(Render_Vulkan, "Device loss occured!");
  628. // Wait for the log to flush and for Nsight Aftermath to dump the results
  629. std::this_thread::sleep_for(std::chrono::seconds{15});
  630. }
  631. void Device::SaveShader(std::span<const u32> spirv) const {
  632. if (nsight_aftermath_tracker) {
  633. nsight_aftermath_tracker->SaveShader(spirv);
  634. }
  635. }
  636. bool Device::IsOptimalAstcSupported(const VkPhysicalDeviceFeatures& features) const {
  637. // Disable for now to avoid converting ASTC twice.
  638. static constexpr std::array astc_formats = {
  639. VK_FORMAT_ASTC_4x4_UNORM_BLOCK, VK_FORMAT_ASTC_4x4_SRGB_BLOCK,
  640. VK_FORMAT_ASTC_5x4_UNORM_BLOCK, VK_FORMAT_ASTC_5x4_SRGB_BLOCK,
  641. VK_FORMAT_ASTC_5x5_UNORM_BLOCK, VK_FORMAT_ASTC_5x5_SRGB_BLOCK,
  642. VK_FORMAT_ASTC_6x5_UNORM_BLOCK, VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
  643. VK_FORMAT_ASTC_6x6_UNORM_BLOCK, VK_FORMAT_ASTC_6x6_SRGB_BLOCK,
  644. VK_FORMAT_ASTC_8x5_UNORM_BLOCK, VK_FORMAT_ASTC_8x5_SRGB_BLOCK,
  645. VK_FORMAT_ASTC_8x6_UNORM_BLOCK, VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
  646. VK_FORMAT_ASTC_8x8_UNORM_BLOCK, VK_FORMAT_ASTC_8x8_SRGB_BLOCK,
  647. VK_FORMAT_ASTC_10x5_UNORM_BLOCK, VK_FORMAT_ASTC_10x5_SRGB_BLOCK,
  648. VK_FORMAT_ASTC_10x6_UNORM_BLOCK, VK_FORMAT_ASTC_10x6_SRGB_BLOCK,
  649. VK_FORMAT_ASTC_10x8_UNORM_BLOCK, VK_FORMAT_ASTC_10x8_SRGB_BLOCK,
  650. VK_FORMAT_ASTC_10x10_UNORM_BLOCK, VK_FORMAT_ASTC_10x10_SRGB_BLOCK,
  651. VK_FORMAT_ASTC_12x10_UNORM_BLOCK, VK_FORMAT_ASTC_12x10_SRGB_BLOCK,
  652. VK_FORMAT_ASTC_12x12_UNORM_BLOCK, VK_FORMAT_ASTC_12x12_SRGB_BLOCK,
  653. };
  654. if (!features.textureCompressionASTC_LDR) {
  655. return false;
  656. }
  657. const auto format_feature_usage{
  658. VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT | VK_FORMAT_FEATURE_BLIT_SRC_BIT |
  659. VK_FORMAT_FEATURE_BLIT_DST_BIT | VK_FORMAT_FEATURE_TRANSFER_SRC_BIT |
  660. VK_FORMAT_FEATURE_TRANSFER_DST_BIT};
  661. for (const auto format : astc_formats) {
  662. const auto physical_format_properties{physical.GetFormatProperties(format)};
  663. if ((physical_format_properties.optimalTilingFeatures & format_feature_usage) == 0) {
  664. return false;
  665. }
  666. }
  667. return true;
  668. }
  669. bool Device::TestDepthStencilBlits() const {
  670. static constexpr VkFormatFeatureFlags required_features =
  671. VK_FORMAT_FEATURE_BLIT_SRC_BIT | VK_FORMAT_FEATURE_BLIT_DST_BIT;
  672. const auto test_features = [](VkFormatProperties props) {
  673. return (props.optimalTilingFeatures & required_features) == required_features;
  674. };
  675. return test_features(format_properties.at(VK_FORMAT_D32_SFLOAT_S8_UINT)) &&
  676. test_features(format_properties.at(VK_FORMAT_D24_UNORM_S8_UINT));
  677. }
  678. bool Device::IsFormatSupported(VkFormat wanted_format, VkFormatFeatureFlags wanted_usage,
  679. FormatType format_type) const {
  680. const auto it = format_properties.find(wanted_format);
  681. if (it == format_properties.end()) {
  682. UNIMPLEMENTED_MSG("Unimplemented format query={}", wanted_format);
  683. return true;
  684. }
  685. const auto supported_usage = GetFormatFeatures(it->second, format_type);
  686. return (supported_usage & wanted_usage) == wanted_usage;
  687. }
  688. std::string Device::GetDriverName() const {
  689. switch (driver_id) {
  690. case VK_DRIVER_ID_AMD_PROPRIETARY:
  691. return "AMD";
  692. case VK_DRIVER_ID_AMD_OPEN_SOURCE:
  693. return "AMDVLK";
  694. case VK_DRIVER_ID_MESA_RADV:
  695. return "RADV";
  696. case VK_DRIVER_ID_NVIDIA_PROPRIETARY:
  697. return "NVIDIA";
  698. case VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS:
  699. return "INTEL";
  700. case VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA:
  701. return "ANV";
  702. case VK_DRIVER_ID_MESA_LLVMPIPE:
  703. return "LAVAPIPE";
  704. default:
  705. return vendor_name;
  706. }
  707. }
  708. void Device::CheckSuitability(bool requires_swapchain) const {
  709. std::bitset<REQUIRED_EXTENSIONS.size()> available_extensions;
  710. bool has_swapchain = false;
  711. for (const VkExtensionProperties& property : physical.EnumerateDeviceExtensionProperties()) {
  712. const std::string_view name{property.extensionName};
  713. for (size_t i = 0; i < REQUIRED_EXTENSIONS.size(); ++i) {
  714. if (available_extensions[i]) {
  715. continue;
  716. }
  717. available_extensions[i] = name == REQUIRED_EXTENSIONS[i];
  718. }
  719. has_swapchain = has_swapchain || name == VK_KHR_SWAPCHAIN_EXTENSION_NAME;
  720. }
  721. for (size_t i = 0; i < REQUIRED_EXTENSIONS.size(); ++i) {
  722. if (available_extensions[i]) {
  723. continue;
  724. }
  725. LOG_ERROR(Render_Vulkan, "Missing required extension: {}", REQUIRED_EXTENSIONS[i]);
  726. throw vk::Exception(VK_ERROR_EXTENSION_NOT_PRESENT);
  727. }
  728. if (requires_swapchain && !has_swapchain) {
  729. LOG_ERROR(Render_Vulkan, "Missing required extension: VK_KHR_swapchain");
  730. throw vk::Exception(VK_ERROR_EXTENSION_NOT_PRESENT);
  731. }
  732. struct LimitTuple {
  733. u32 minimum;
  734. u32 value;
  735. const char* name;
  736. };
  737. const VkPhysicalDeviceLimits& limits{properties.limits};
  738. const std::array limits_report{
  739. LimitTuple{65536, limits.maxUniformBufferRange, "maxUniformBufferRange"},
  740. LimitTuple{16, limits.maxViewports, "maxViewports"},
  741. LimitTuple{8, limits.maxColorAttachments, "maxColorAttachments"},
  742. LimitTuple{8, limits.maxClipDistances, "maxClipDistances"},
  743. };
  744. for (const auto& tuple : limits_report) {
  745. if (tuple.value < tuple.minimum) {
  746. LOG_ERROR(Render_Vulkan, "{} has to be {} or greater but it is {}", tuple.name,
  747. tuple.minimum, tuple.value);
  748. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  749. }
  750. }
  751. VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT demote{};
  752. demote.sType =
  753. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT;
  754. demote.pNext = nullptr;
  755. VkPhysicalDeviceVariablePointerFeaturesKHR variable_pointers{};
  756. variable_pointers.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES_KHR;
  757. variable_pointers.pNext = &demote;
  758. VkPhysicalDeviceRobustness2FeaturesEXT robustness2{};
  759. robustness2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT;
  760. robustness2.pNext = &variable_pointers;
  761. VkPhysicalDeviceFeatures2KHR features2{};
  762. features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2;
  763. features2.pNext = &robustness2;
  764. physical.GetFeatures2KHR(features2);
  765. const VkPhysicalDeviceFeatures& features{features2.features};
  766. const std::array feature_report{
  767. std::make_pair(features.robustBufferAccess, "robustBufferAccess"),
  768. std::make_pair(features.vertexPipelineStoresAndAtomics, "vertexPipelineStoresAndAtomics"),
  769. std::make_pair(features.imageCubeArray, "imageCubeArray"),
  770. std::make_pair(features.independentBlend, "independentBlend"),
  771. std::make_pair(features.depthClamp, "depthClamp"),
  772. std::make_pair(features.samplerAnisotropy, "samplerAnisotropy"),
  773. std::make_pair(features.largePoints, "largePoints"),
  774. std::make_pair(features.multiViewport, "multiViewport"),
  775. std::make_pair(features.depthBiasClamp, "depthBiasClamp"),
  776. std::make_pair(features.fillModeNonSolid, "fillModeNonSolid"),
  777. std::make_pair(features.wideLines, "wideLines"),
  778. std::make_pair(features.geometryShader, "geometryShader"),
  779. std::make_pair(features.tessellationShader, "tessellationShader"),
  780. std::make_pair(features.sampleRateShading, "sampleRateShading"),
  781. std::make_pair(features.dualSrcBlend, "dualSrcBlend"),
  782. std::make_pair(features.occlusionQueryPrecise, "occlusionQueryPrecise"),
  783. std::make_pair(features.fragmentStoresAndAtomics, "fragmentStoresAndAtomics"),
  784. std::make_pair(features.shaderImageGatherExtended, "shaderImageGatherExtended"),
  785. std::make_pair(features.shaderStorageImageWriteWithoutFormat,
  786. "shaderStorageImageWriteWithoutFormat"),
  787. std::make_pair(features.shaderClipDistance, "shaderClipDistance"),
  788. std::make_pair(features.shaderCullDistance, "shaderCullDistance"),
  789. std::make_pair(demote.shaderDemoteToHelperInvocation, "shaderDemoteToHelperInvocation"),
  790. std::make_pair(variable_pointers.variablePointers, "variablePointers"),
  791. std::make_pair(variable_pointers.variablePointersStorageBuffer,
  792. "variablePointersStorageBuffer"),
  793. std::make_pair(robustness2.robustBufferAccess2, "robustBufferAccess2"),
  794. std::make_pair(robustness2.robustImageAccess2, "robustImageAccess2"),
  795. std::make_pair(robustness2.nullDescriptor, "nullDescriptor"),
  796. };
  797. for (const auto& [is_supported, name] : feature_report) {
  798. if (is_supported) {
  799. continue;
  800. }
  801. LOG_ERROR(Render_Vulkan, "Missing required feature: {}", name);
  802. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  803. }
  804. }
  805. std::vector<const char*> Device::LoadExtensions(bool requires_surface) {
  806. std::vector<const char*> extensions;
  807. extensions.reserve(8 + REQUIRED_EXTENSIONS.size());
  808. extensions.insert(extensions.begin(), REQUIRED_EXTENSIONS.begin(), REQUIRED_EXTENSIONS.end());
  809. if (requires_surface) {
  810. extensions.push_back(VK_KHR_SWAPCHAIN_EXTENSION_NAME);
  811. }
  812. bool has_khr_shader_float16_int8{};
  813. bool has_khr_workgroup_memory_explicit_layout{};
  814. bool has_khr_pipeline_executable_properties{};
  815. bool has_khr_image_format_list{};
  816. bool has_khr_swapchain_mutable_format{};
  817. bool has_ext_subgroup_size_control{};
  818. bool has_ext_transform_feedback{};
  819. bool has_ext_custom_border_color{};
  820. bool has_ext_extended_dynamic_state{};
  821. bool has_ext_shader_atomic_int64{};
  822. bool has_ext_provoking_vertex{};
  823. bool has_ext_vertex_input_dynamic_state{};
  824. bool has_ext_line_rasterization{};
  825. for (const std::string& extension : supported_extensions) {
  826. const auto test = [&](std::optional<std::reference_wrapper<bool>> status, const char* name,
  827. bool push) {
  828. if (extension != name) {
  829. return;
  830. }
  831. if (push) {
  832. extensions.push_back(name);
  833. }
  834. if (status) {
  835. status->get() = true;
  836. }
  837. };
  838. test(nv_viewport_swizzle, VK_NV_VIEWPORT_SWIZZLE_EXTENSION_NAME, true);
  839. test(nv_viewport_array2, VK_NV_VIEWPORT_ARRAY2_EXTENSION_NAME, true);
  840. test(nv_geometry_shader_passthrough, VK_NV_GEOMETRY_SHADER_PASSTHROUGH_EXTENSION_NAME,
  841. true);
  842. test(khr_uniform_buffer_standard_layout,
  843. VK_KHR_UNIFORM_BUFFER_STANDARD_LAYOUT_EXTENSION_NAME, true);
  844. test(khr_spirv_1_4, VK_KHR_SPIRV_1_4_EXTENSION_NAME, true);
  845. test(khr_push_descriptor, VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME, true);
  846. test(has_khr_shader_float16_int8, VK_KHR_SHADER_FLOAT16_INT8_EXTENSION_NAME, false);
  847. test(ext_depth_range_unrestricted, VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME, true);
  848. test(ext_index_type_uint8, VK_EXT_INDEX_TYPE_UINT8_EXTENSION_NAME, true);
  849. test(ext_sampler_filter_minmax, VK_EXT_SAMPLER_FILTER_MINMAX_EXTENSION_NAME, true);
  850. test(ext_shader_viewport_index_layer, VK_EXT_SHADER_VIEWPORT_INDEX_LAYER_EXTENSION_NAME,
  851. true);
  852. test(ext_tooling_info, VK_EXT_TOOLING_INFO_EXTENSION_NAME, true);
  853. test(ext_shader_stencil_export, VK_EXT_SHADER_STENCIL_EXPORT_EXTENSION_NAME, true);
  854. test(ext_conservative_rasterization, VK_EXT_CONSERVATIVE_RASTERIZATION_EXTENSION_NAME,
  855. true);
  856. test(has_ext_transform_feedback, VK_EXT_TRANSFORM_FEEDBACK_EXTENSION_NAME, false);
  857. test(has_ext_custom_border_color, VK_EXT_CUSTOM_BORDER_COLOR_EXTENSION_NAME, false);
  858. test(has_ext_extended_dynamic_state, VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME, false);
  859. test(has_ext_subgroup_size_control, VK_EXT_SUBGROUP_SIZE_CONTROL_EXTENSION_NAME, false);
  860. test(has_ext_provoking_vertex, VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME, false);
  861. test(has_ext_vertex_input_dynamic_state, VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME,
  862. false);
  863. test(has_ext_shader_atomic_int64, VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME, false);
  864. test(has_khr_workgroup_memory_explicit_layout,
  865. VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME, false);
  866. test(has_khr_image_format_list, VK_KHR_IMAGE_FORMAT_LIST_EXTENSION_NAME, false);
  867. test(has_khr_swapchain_mutable_format, VK_KHR_SWAPCHAIN_MUTABLE_FORMAT_EXTENSION_NAME,
  868. false);
  869. test(has_ext_line_rasterization, VK_EXT_LINE_RASTERIZATION_EXTENSION_NAME, false);
  870. if (Settings::values.enable_nsight_aftermath) {
  871. test(nv_device_diagnostics_config, VK_NV_DEVICE_DIAGNOSTICS_CONFIG_EXTENSION_NAME,
  872. true);
  873. }
  874. if (Settings::values.renderer_shader_feedback) {
  875. test(has_khr_pipeline_executable_properties,
  876. VK_KHR_PIPELINE_EXECUTABLE_PROPERTIES_EXTENSION_NAME, false);
  877. }
  878. }
  879. VkPhysicalDeviceFeatures2KHR features{};
  880. features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2_KHR;
  881. VkPhysicalDeviceProperties2KHR physical_properties;
  882. physical_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR;
  883. if (has_khr_shader_float16_int8) {
  884. VkPhysicalDeviceFloat16Int8FeaturesKHR float16_int8_features;
  885. float16_int8_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR;
  886. float16_int8_features.pNext = nullptr;
  887. features.pNext = &float16_int8_features;
  888. physical.GetFeatures2KHR(features);
  889. is_float16_supported = float16_int8_features.shaderFloat16;
  890. is_int8_supported = float16_int8_features.shaderInt8;
  891. extensions.push_back(VK_KHR_SHADER_FLOAT16_INT8_EXTENSION_NAME);
  892. }
  893. if (has_ext_subgroup_size_control) {
  894. VkPhysicalDeviceSubgroupSizeControlFeaturesEXT subgroup_features;
  895. subgroup_features.sType =
  896. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT;
  897. subgroup_features.pNext = nullptr;
  898. features.pNext = &subgroup_features;
  899. physical.GetFeatures2KHR(features);
  900. VkPhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_properties;
  901. subgroup_properties.sType =
  902. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT;
  903. subgroup_properties.pNext = nullptr;
  904. physical_properties.pNext = &subgroup_properties;
  905. physical.GetProperties2KHR(physical_properties);
  906. is_warp_potentially_bigger = subgroup_properties.maxSubgroupSize > GuestWarpSize;
  907. if (subgroup_features.subgroupSizeControl &&
  908. subgroup_properties.minSubgroupSize <= GuestWarpSize &&
  909. subgroup_properties.maxSubgroupSize >= GuestWarpSize) {
  910. extensions.push_back(VK_EXT_SUBGROUP_SIZE_CONTROL_EXTENSION_NAME);
  911. guest_warp_stages = subgroup_properties.requiredSubgroupSizeStages;
  912. ext_subgroup_size_control = true;
  913. }
  914. } else {
  915. is_warp_potentially_bigger = true;
  916. }
  917. if (has_ext_provoking_vertex) {
  918. VkPhysicalDeviceProvokingVertexFeaturesEXT provoking_vertex;
  919. provoking_vertex.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_FEATURES_EXT;
  920. provoking_vertex.pNext = nullptr;
  921. features.pNext = &provoking_vertex;
  922. physical.GetFeatures2KHR(features);
  923. if (provoking_vertex.provokingVertexLast &&
  924. provoking_vertex.transformFeedbackPreservesProvokingVertex) {
  925. extensions.push_back(VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME);
  926. ext_provoking_vertex = true;
  927. }
  928. }
  929. if (has_ext_vertex_input_dynamic_state) {
  930. VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT vertex_input;
  931. vertex_input.sType =
  932. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_INPUT_DYNAMIC_STATE_FEATURES_EXT;
  933. vertex_input.pNext = nullptr;
  934. features.pNext = &vertex_input;
  935. physical.GetFeatures2KHR(features);
  936. if (vertex_input.vertexInputDynamicState) {
  937. extensions.push_back(VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME);
  938. ext_vertex_input_dynamic_state = true;
  939. }
  940. }
  941. if (has_ext_shader_atomic_int64) {
  942. VkPhysicalDeviceShaderAtomicInt64Features atomic_int64;
  943. atomic_int64.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT;
  944. atomic_int64.pNext = nullptr;
  945. features.pNext = &atomic_int64;
  946. physical.GetFeatures2KHR(features);
  947. if (atomic_int64.shaderBufferInt64Atomics && atomic_int64.shaderSharedInt64Atomics) {
  948. extensions.push_back(VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME);
  949. ext_shader_atomic_int64 = true;
  950. }
  951. }
  952. if (has_ext_transform_feedback) {
  953. VkPhysicalDeviceTransformFeedbackFeaturesEXT tfb_features;
  954. tfb_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT;
  955. tfb_features.pNext = nullptr;
  956. features.pNext = &tfb_features;
  957. physical.GetFeatures2KHR(features);
  958. VkPhysicalDeviceTransformFeedbackPropertiesEXT tfb_properties;
  959. tfb_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT;
  960. tfb_properties.pNext = nullptr;
  961. physical_properties.pNext = &tfb_properties;
  962. physical.GetProperties2KHR(physical_properties);
  963. if (tfb_features.transformFeedback && tfb_features.geometryStreams &&
  964. tfb_properties.maxTransformFeedbackStreams >= 4 &&
  965. tfb_properties.maxTransformFeedbackBuffers && tfb_properties.transformFeedbackQueries &&
  966. tfb_properties.transformFeedbackDraw) {
  967. extensions.push_back(VK_EXT_TRANSFORM_FEEDBACK_EXTENSION_NAME);
  968. ext_transform_feedback = true;
  969. }
  970. }
  971. if (has_ext_custom_border_color) {
  972. VkPhysicalDeviceCustomBorderColorFeaturesEXT border_features;
  973. border_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT;
  974. border_features.pNext = nullptr;
  975. features.pNext = &border_features;
  976. physical.GetFeatures2KHR(features);
  977. if (border_features.customBorderColors && border_features.customBorderColorWithoutFormat) {
  978. extensions.push_back(VK_EXT_CUSTOM_BORDER_COLOR_EXTENSION_NAME);
  979. ext_custom_border_color = true;
  980. }
  981. }
  982. if (has_ext_extended_dynamic_state) {
  983. VkPhysicalDeviceExtendedDynamicStateFeaturesEXT extended_dynamic_state;
  984. extended_dynamic_state.sType =
  985. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT;
  986. extended_dynamic_state.pNext = nullptr;
  987. features.pNext = &extended_dynamic_state;
  988. physical.GetFeatures2KHR(features);
  989. if (extended_dynamic_state.extendedDynamicState) {
  990. extensions.push_back(VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME);
  991. ext_extended_dynamic_state = true;
  992. }
  993. }
  994. if (has_ext_line_rasterization) {
  995. VkPhysicalDeviceLineRasterizationFeaturesEXT line_raster;
  996. line_raster.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT;
  997. line_raster.pNext = nullptr;
  998. features.pNext = &line_raster;
  999. physical.GetFeatures2KHR(features);
  1000. if (line_raster.rectangularLines && line_raster.smoothLines) {
  1001. extensions.push_back(VK_EXT_LINE_RASTERIZATION_EXTENSION_NAME);
  1002. ext_line_rasterization = true;
  1003. }
  1004. }
  1005. if (has_khr_workgroup_memory_explicit_layout) {
  1006. VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR layout;
  1007. layout.sType =
  1008. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR;
  1009. layout.pNext = nullptr;
  1010. features.pNext = &layout;
  1011. physical.GetFeatures2KHR(features);
  1012. if (layout.workgroupMemoryExplicitLayout &&
  1013. layout.workgroupMemoryExplicitLayout8BitAccess &&
  1014. layout.workgroupMemoryExplicitLayout16BitAccess &&
  1015. layout.workgroupMemoryExplicitLayoutScalarBlockLayout) {
  1016. extensions.push_back(VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME);
  1017. khr_workgroup_memory_explicit_layout = true;
  1018. }
  1019. }
  1020. if (has_khr_pipeline_executable_properties) {
  1021. VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR executable_properties;
  1022. executable_properties.sType =
  1023. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR;
  1024. executable_properties.pNext = nullptr;
  1025. features.pNext = &executable_properties;
  1026. physical.GetFeatures2KHR(features);
  1027. if (executable_properties.pipelineExecutableInfo) {
  1028. extensions.push_back(VK_KHR_PIPELINE_EXECUTABLE_PROPERTIES_EXTENSION_NAME);
  1029. khr_pipeline_executable_properties = true;
  1030. }
  1031. }
  1032. if (has_khr_image_format_list && has_khr_swapchain_mutable_format) {
  1033. extensions.push_back(VK_KHR_IMAGE_FORMAT_LIST_EXTENSION_NAME);
  1034. extensions.push_back(VK_KHR_SWAPCHAIN_MUTABLE_FORMAT_EXTENSION_NAME);
  1035. khr_swapchain_mutable_format = true;
  1036. }
  1037. if (khr_push_descriptor) {
  1038. VkPhysicalDevicePushDescriptorPropertiesKHR push_descriptor;
  1039. push_descriptor.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR;
  1040. push_descriptor.pNext = nullptr;
  1041. physical_properties.pNext = &push_descriptor;
  1042. physical.GetProperties2KHR(physical_properties);
  1043. max_push_descriptors = push_descriptor.maxPushDescriptors;
  1044. }
  1045. return extensions;
  1046. }
  1047. void Device::SetupFamilies(VkSurfaceKHR surface) {
  1048. const std::vector queue_family_properties = physical.GetQueueFamilyProperties();
  1049. std::optional<u32> graphics;
  1050. std::optional<u32> present;
  1051. for (u32 index = 0; index < static_cast<u32>(queue_family_properties.size()); ++index) {
  1052. if (graphics && (present || !surface)) {
  1053. break;
  1054. }
  1055. const VkQueueFamilyProperties& queue_family = queue_family_properties[index];
  1056. if (queue_family.queueCount == 0) {
  1057. continue;
  1058. }
  1059. if (queue_family.queueFlags & VK_QUEUE_GRAPHICS_BIT) {
  1060. graphics = index;
  1061. }
  1062. if (surface && physical.GetSurfaceSupportKHR(index, surface)) {
  1063. present = index;
  1064. }
  1065. }
  1066. if (!graphics) {
  1067. LOG_ERROR(Render_Vulkan, "Device lacks a graphics queue");
  1068. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  1069. }
  1070. if (surface && !present) {
  1071. LOG_ERROR(Render_Vulkan, "Device lacks a present queue");
  1072. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  1073. }
  1074. graphics_family = *graphics;
  1075. present_family = *present;
  1076. }
  1077. void Device::SetupFeatures() {
  1078. const VkPhysicalDeviceFeatures features{physical.GetFeatures()};
  1079. is_depth_bounds_supported = features.depthBounds;
  1080. is_formatless_image_load_supported = features.shaderStorageImageReadWithoutFormat;
  1081. is_shader_float64_supported = features.shaderFloat64;
  1082. is_shader_int64_supported = features.shaderInt64;
  1083. is_shader_int16_supported = features.shaderInt16;
  1084. is_shader_storage_image_multisample = features.shaderStorageImageMultisample;
  1085. is_blit_depth_stencil_supported = TestDepthStencilBlits();
  1086. is_optimal_astc_supported = IsOptimalAstcSupported(features);
  1087. }
  1088. void Device::SetupProperties() {
  1089. float_controls.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR;
  1090. VkPhysicalDeviceProperties2KHR properties2{};
  1091. properties2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR;
  1092. properties2.pNext = &float_controls;
  1093. physical.GetProperties2KHR(properties2);
  1094. }
  1095. void Device::CollectTelemetryParameters() {
  1096. VkPhysicalDeviceDriverPropertiesKHR driver{
  1097. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR,
  1098. .pNext = nullptr,
  1099. .driverID = {},
  1100. .driverName = {},
  1101. .driverInfo = {},
  1102. .conformanceVersion = {},
  1103. };
  1104. VkPhysicalDeviceProperties2KHR device_properties{
  1105. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR,
  1106. .pNext = &driver,
  1107. .properties = {},
  1108. };
  1109. physical.GetProperties2KHR(device_properties);
  1110. driver_id = driver.driverID;
  1111. vendor_name = driver.driverName;
  1112. }
  1113. void Device::CollectPhysicalMemoryInfo() {
  1114. const auto mem_properties = physical.GetMemoryProperties();
  1115. const size_t num_properties = mem_properties.memoryHeapCount;
  1116. device_access_memory = 0;
  1117. for (size_t element = 0; element < num_properties; ++element) {
  1118. if ((mem_properties.memoryHeaps[element].flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) != 0) {
  1119. device_access_memory += mem_properties.memoryHeaps[element].size;
  1120. }
  1121. }
  1122. }
  1123. void Device::CollectToolingInfo() {
  1124. if (!ext_tooling_info) {
  1125. return;
  1126. }
  1127. const auto vkGetPhysicalDeviceToolPropertiesEXT =
  1128. reinterpret_cast<PFN_vkGetPhysicalDeviceToolPropertiesEXT>(
  1129. dld.vkGetInstanceProcAddr(instance, "vkGetPhysicalDeviceToolPropertiesEXT"));
  1130. if (!vkGetPhysicalDeviceToolPropertiesEXT) {
  1131. return;
  1132. }
  1133. u32 tool_count = 0;
  1134. if (vkGetPhysicalDeviceToolPropertiesEXT(physical, &tool_count, nullptr) != VK_SUCCESS) {
  1135. return;
  1136. }
  1137. std::vector<VkPhysicalDeviceToolPropertiesEXT> tools(tool_count);
  1138. if (vkGetPhysicalDeviceToolPropertiesEXT(physical, &tool_count, tools.data()) != VK_SUCCESS) {
  1139. return;
  1140. }
  1141. for (const VkPhysicalDeviceToolPropertiesEXT& tool : tools) {
  1142. const std::string_view name = tool.name;
  1143. LOG_INFO(Render_Vulkan, "{}", name);
  1144. has_renderdoc = has_renderdoc || name == "RenderDoc";
  1145. has_nsight_graphics = has_nsight_graphics || name == "NVIDIA Nsight Graphics";
  1146. }
  1147. }
  1148. std::vector<VkDeviceQueueCreateInfo> Device::GetDeviceQueueCreateInfos() const {
  1149. static constexpr float QUEUE_PRIORITY = 1.0f;
  1150. std::unordered_set<u32> unique_queue_families{graphics_family, present_family};
  1151. std::vector<VkDeviceQueueCreateInfo> queue_cis;
  1152. queue_cis.reserve(unique_queue_families.size());
  1153. for (const u32 queue_family : unique_queue_families) {
  1154. auto& ci = queue_cis.emplace_back(VkDeviceQueueCreateInfo{
  1155. .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
  1156. .pNext = nullptr,
  1157. .flags = 0,
  1158. .queueFamilyIndex = queue_family,
  1159. .queueCount = 1,
  1160. .pQueuePriorities = nullptr,
  1161. });
  1162. ci.pQueuePriorities = &QUEUE_PRIORITY;
  1163. }
  1164. return queue_cis;
  1165. }
  1166. } // namespace Vulkan