vulkan_device.cpp 50 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221
  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. if (is_int8_supported || is_float16_supported) {
  349. VkPhysicalDeviceFloat16Int8FeaturesKHR float16_int8{
  350. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR,
  351. .pNext = nullptr,
  352. .shaderFloat16 = is_float16_supported,
  353. .shaderInt8 = is_int8_supported,
  354. };
  355. SetNext(next, float16_int8);
  356. }
  357. if (!is_float16_supported) {
  358. LOG_INFO(Render_Vulkan, "Device doesn't support float16 natively");
  359. }
  360. if (!is_int8_supported) {
  361. LOG_INFO(Render_Vulkan, "Device doesn't support int8 natively");
  362. }
  363. if (!nv_viewport_swizzle) {
  364. LOG_INFO(Render_Vulkan, "Device doesn't support viewport swizzles");
  365. }
  366. if (!nv_viewport_array2) {
  367. LOG_INFO(Render_Vulkan, "Device doesn't support viewport masks");
  368. }
  369. if (!nv_geometry_shader_passthrough) {
  370. LOG_INFO(Render_Vulkan, "Device doesn't support passthrough geometry shaders");
  371. }
  372. VkPhysicalDeviceUniformBufferStandardLayoutFeaturesKHR std430_layout;
  373. if (khr_uniform_buffer_standard_layout) {
  374. std430_layout = {
  375. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_UNIFORM_BUFFER_STANDARD_LAYOUT_FEATURES_KHR,
  376. .pNext = nullptr,
  377. .uniformBufferStandardLayout = true,
  378. };
  379. SetNext(next, std430_layout);
  380. } else {
  381. LOG_INFO(Render_Vulkan, "Device doesn't support packed UBOs");
  382. }
  383. VkPhysicalDeviceIndexTypeUint8FeaturesEXT index_type_uint8;
  384. if (ext_index_type_uint8) {
  385. index_type_uint8 = {
  386. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT,
  387. .pNext = nullptr,
  388. .indexTypeUint8 = true,
  389. };
  390. SetNext(next, index_type_uint8);
  391. } else {
  392. LOG_INFO(Render_Vulkan, "Device doesn't support uint8 indexes");
  393. }
  394. VkPhysicalDeviceTransformFeedbackFeaturesEXT transform_feedback;
  395. if (ext_transform_feedback) {
  396. transform_feedback = {
  397. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT,
  398. .pNext = nullptr,
  399. .transformFeedback = true,
  400. .geometryStreams = true,
  401. };
  402. SetNext(next, transform_feedback);
  403. } else {
  404. LOG_INFO(Render_Vulkan, "Device doesn't support transform feedbacks");
  405. }
  406. VkPhysicalDeviceCustomBorderColorFeaturesEXT custom_border;
  407. if (ext_custom_border_color) {
  408. custom_border = {
  409. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT,
  410. .pNext = nullptr,
  411. .customBorderColors = VK_TRUE,
  412. .customBorderColorWithoutFormat = VK_TRUE,
  413. };
  414. SetNext(next, custom_border);
  415. } else {
  416. LOG_INFO(Render_Vulkan, "Device doesn't support custom border colors");
  417. }
  418. VkPhysicalDeviceExtendedDynamicStateFeaturesEXT dynamic_state;
  419. if (ext_extended_dynamic_state) {
  420. dynamic_state = {
  421. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT,
  422. .pNext = nullptr,
  423. .extendedDynamicState = VK_TRUE,
  424. };
  425. SetNext(next, dynamic_state);
  426. } else {
  427. LOG_INFO(Render_Vulkan, "Device doesn't support extended dynamic state");
  428. }
  429. VkPhysicalDeviceLineRasterizationFeaturesEXT line_raster;
  430. if (ext_line_rasterization) {
  431. line_raster = {
  432. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT,
  433. .pNext = nullptr,
  434. .rectangularLines = VK_TRUE,
  435. .bresenhamLines = VK_FALSE,
  436. .smoothLines = VK_TRUE,
  437. .stippledRectangularLines = VK_FALSE,
  438. .stippledBresenhamLines = VK_FALSE,
  439. .stippledSmoothLines = VK_FALSE,
  440. };
  441. SetNext(next, line_raster);
  442. } else {
  443. LOG_INFO(Render_Vulkan, "Device doesn't support smooth lines");
  444. }
  445. if (!ext_conservative_rasterization) {
  446. LOG_INFO(Render_Vulkan, "Device doesn't support conservative rasterization");
  447. }
  448. VkPhysicalDeviceProvokingVertexFeaturesEXT provoking_vertex;
  449. if (ext_provoking_vertex) {
  450. provoking_vertex = {
  451. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_FEATURES_EXT,
  452. .pNext = nullptr,
  453. .provokingVertexLast = VK_TRUE,
  454. .transformFeedbackPreservesProvokingVertex = VK_TRUE,
  455. };
  456. SetNext(next, provoking_vertex);
  457. } else {
  458. LOG_INFO(Render_Vulkan, "Device doesn't support provoking vertex last");
  459. }
  460. VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT vertex_input_dynamic;
  461. if (ext_vertex_input_dynamic_state) {
  462. vertex_input_dynamic = {
  463. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_INPUT_DYNAMIC_STATE_FEATURES_EXT,
  464. .pNext = nullptr,
  465. .vertexInputDynamicState = VK_TRUE,
  466. };
  467. SetNext(next, vertex_input_dynamic);
  468. } else {
  469. LOG_INFO(Render_Vulkan, "Device doesn't support vertex input dynamic state");
  470. }
  471. VkPhysicalDeviceShaderAtomicInt64FeaturesKHR atomic_int64;
  472. if (ext_shader_atomic_int64) {
  473. atomic_int64 = {
  474. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_ATOMIC_INT64_FEATURES_KHR,
  475. .pNext = nullptr,
  476. .shaderBufferInt64Atomics = VK_TRUE,
  477. .shaderSharedInt64Atomics = VK_TRUE,
  478. };
  479. SetNext(next, atomic_int64);
  480. }
  481. VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR workgroup_layout;
  482. if (khr_workgroup_memory_explicit_layout) {
  483. workgroup_layout = {
  484. .sType =
  485. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR,
  486. .pNext = nullptr,
  487. .workgroupMemoryExplicitLayout = VK_TRUE,
  488. .workgroupMemoryExplicitLayoutScalarBlockLayout = VK_TRUE,
  489. .workgroupMemoryExplicitLayout8BitAccess = VK_TRUE,
  490. .workgroupMemoryExplicitLayout16BitAccess = VK_TRUE,
  491. };
  492. SetNext(next, workgroup_layout);
  493. }
  494. VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR executable_properties;
  495. if (khr_pipeline_executable_properties) {
  496. LOG_INFO(Render_Vulkan, "Enabling shader feedback, expect slower shader build times");
  497. executable_properties = {
  498. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR,
  499. .pNext = nullptr,
  500. .pipelineExecutableInfo = VK_TRUE,
  501. };
  502. SetNext(next, executable_properties);
  503. }
  504. if (!ext_depth_range_unrestricted) {
  505. LOG_INFO(Render_Vulkan, "Device doesn't support depth range unrestricted");
  506. }
  507. VkDeviceDiagnosticsConfigCreateInfoNV diagnostics_nv;
  508. if (Settings::values.enable_nsight_aftermath && nv_device_diagnostics_config) {
  509. nsight_aftermath_tracker = std::make_unique<NsightAftermathTracker>();
  510. diagnostics_nv = {
  511. .sType = VK_STRUCTURE_TYPE_DEVICE_DIAGNOSTICS_CONFIG_CREATE_INFO_NV,
  512. .pNext = &features2,
  513. .flags = VK_DEVICE_DIAGNOSTICS_CONFIG_ENABLE_SHADER_DEBUG_INFO_BIT_NV |
  514. VK_DEVICE_DIAGNOSTICS_CONFIG_ENABLE_RESOURCE_TRACKING_BIT_NV |
  515. VK_DEVICE_DIAGNOSTICS_CONFIG_ENABLE_AUTOMATIC_CHECKPOINTS_BIT_NV,
  516. };
  517. first_next = &diagnostics_nv;
  518. }
  519. logical = vk::Device::Create(physical, queue_cis, extensions, first_next, dld);
  520. CollectPhysicalMemoryInfo();
  521. CollectTelemetryParameters();
  522. CollectToolingInfo();
  523. if (driver_id == VK_DRIVER_ID_NVIDIA_PROPRIETARY_KHR) {
  524. const auto arch = GetNvidiaArchitecture(physical, supported_extensions);
  525. switch (arch) {
  526. case NvidiaArchitecture::AmpereOrNewer:
  527. LOG_WARNING(Render_Vulkan, "Blacklisting Ampere devices from float16 math");
  528. is_float16_supported = false;
  529. break;
  530. case NvidiaArchitecture::Turing:
  531. break;
  532. case NvidiaArchitecture::VoltaOrOlder:
  533. LOG_WARNING(Render_Vulkan, "Blacklisting Volta and older from VK_KHR_push_descriptor");
  534. khr_push_descriptor = false;
  535. break;
  536. }
  537. }
  538. if (ext_extended_dynamic_state && driver_id == VK_DRIVER_ID_MESA_RADV) {
  539. // Mask driver version variant
  540. const u32 version = (properties.driverVersion << 3) >> 3;
  541. if (version < VK_MAKE_API_VERSION(0, 21, 2, 0)) {
  542. LOG_WARNING(Render_Vulkan,
  543. "RADV versions older than 21.2 have broken VK_EXT_extended_dynamic_state");
  544. ext_extended_dynamic_state = false;
  545. }
  546. }
  547. if (ext_vertex_input_dynamic_state && driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
  548. LOG_WARNING(Render_Vulkan, "Blacklisting Intel for VK_EXT_vertex_input_dynamic_state");
  549. ext_vertex_input_dynamic_state = false;
  550. }
  551. if (is_float16_supported && driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
  552. // Intel's compiler crashes when using fp16 on Astral Chain, disable it for the time being.
  553. LOG_WARNING(Render_Vulkan, "Blacklisting Intel proprietary from float16 math");
  554. is_float16_supported = false;
  555. }
  556. graphics_queue = logical.GetQueue(graphics_family);
  557. present_queue = logical.GetQueue(present_family);
  558. sets_per_pool = 64;
  559. if (driver_id == VK_DRIVER_ID_AMD_PROPRIETARY || driver_id == VK_DRIVER_ID_AMD_OPEN_SOURCE) {
  560. // AMD drivers need a higher amount of Sets per Pool in certain circunstances like in XC2.
  561. sets_per_pool = 96;
  562. }
  563. }
  564. Device::~Device() = default;
  565. VkFormat Device::GetSupportedFormat(VkFormat wanted_format, VkFormatFeatureFlags wanted_usage,
  566. FormatType format_type) const {
  567. if (IsFormatSupported(wanted_format, wanted_usage, format_type)) {
  568. return wanted_format;
  569. }
  570. // The wanted format is not supported by hardware, search for alternatives
  571. const VkFormat* alternatives = GetFormatAlternatives(wanted_format);
  572. if (alternatives == nullptr) {
  573. UNREACHABLE_MSG("Format={} with usage={} and type={} has no defined alternatives and host "
  574. "hardware does not support it",
  575. wanted_format, wanted_usage, format_type);
  576. return wanted_format;
  577. }
  578. std::size_t i = 0;
  579. for (VkFormat alternative = *alternatives; alternative; alternative = alternatives[++i]) {
  580. if (!IsFormatSupported(alternative, wanted_usage, format_type)) {
  581. continue;
  582. }
  583. LOG_WARNING(Render_Vulkan,
  584. "Emulating format={} with alternative format={} with usage={} and type={}",
  585. wanted_format, alternative, wanted_usage, format_type);
  586. return alternative;
  587. }
  588. // No alternatives found, panic
  589. UNREACHABLE_MSG("Format={} with usage={} and type={} is not supported by the host hardware and "
  590. "doesn't support any of the alternatives",
  591. wanted_format, wanted_usage, format_type);
  592. return wanted_format;
  593. }
  594. void Device::ReportLoss() const {
  595. LOG_CRITICAL(Render_Vulkan, "Device loss occured!");
  596. // Wait for the log to flush and for Nsight Aftermath to dump the results
  597. std::this_thread::sleep_for(std::chrono::seconds{15});
  598. }
  599. void Device::SaveShader(std::span<const u32> spirv) const {
  600. if (nsight_aftermath_tracker) {
  601. nsight_aftermath_tracker->SaveShader(spirv);
  602. }
  603. }
  604. bool Device::IsOptimalAstcSupported(const VkPhysicalDeviceFeatures& features) const {
  605. // Disable for now to avoid converting ASTC twice.
  606. static constexpr std::array astc_formats = {
  607. VK_FORMAT_ASTC_4x4_UNORM_BLOCK, VK_FORMAT_ASTC_4x4_SRGB_BLOCK,
  608. VK_FORMAT_ASTC_5x4_UNORM_BLOCK, VK_FORMAT_ASTC_5x4_SRGB_BLOCK,
  609. VK_FORMAT_ASTC_5x5_UNORM_BLOCK, VK_FORMAT_ASTC_5x5_SRGB_BLOCK,
  610. VK_FORMAT_ASTC_6x5_UNORM_BLOCK, VK_FORMAT_ASTC_6x5_SRGB_BLOCK,
  611. VK_FORMAT_ASTC_6x6_UNORM_BLOCK, VK_FORMAT_ASTC_6x6_SRGB_BLOCK,
  612. VK_FORMAT_ASTC_8x5_UNORM_BLOCK, VK_FORMAT_ASTC_8x5_SRGB_BLOCK,
  613. VK_FORMAT_ASTC_8x6_UNORM_BLOCK, VK_FORMAT_ASTC_8x6_SRGB_BLOCK,
  614. VK_FORMAT_ASTC_8x8_UNORM_BLOCK, VK_FORMAT_ASTC_8x8_SRGB_BLOCK,
  615. VK_FORMAT_ASTC_10x5_UNORM_BLOCK, VK_FORMAT_ASTC_10x5_SRGB_BLOCK,
  616. VK_FORMAT_ASTC_10x6_UNORM_BLOCK, VK_FORMAT_ASTC_10x6_SRGB_BLOCK,
  617. VK_FORMAT_ASTC_10x8_UNORM_BLOCK, VK_FORMAT_ASTC_10x8_SRGB_BLOCK,
  618. VK_FORMAT_ASTC_10x10_UNORM_BLOCK, VK_FORMAT_ASTC_10x10_SRGB_BLOCK,
  619. VK_FORMAT_ASTC_12x10_UNORM_BLOCK, VK_FORMAT_ASTC_12x10_SRGB_BLOCK,
  620. VK_FORMAT_ASTC_12x12_UNORM_BLOCK, VK_FORMAT_ASTC_12x12_SRGB_BLOCK,
  621. };
  622. if (!features.textureCompressionASTC_LDR) {
  623. return false;
  624. }
  625. const auto format_feature_usage{
  626. VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT | VK_FORMAT_FEATURE_BLIT_SRC_BIT |
  627. VK_FORMAT_FEATURE_BLIT_DST_BIT | VK_FORMAT_FEATURE_TRANSFER_SRC_BIT |
  628. VK_FORMAT_FEATURE_TRANSFER_DST_BIT};
  629. for (const auto format : astc_formats) {
  630. const auto physical_format_properties{physical.GetFormatProperties(format)};
  631. if ((physical_format_properties.optimalTilingFeatures & format_feature_usage) == 0) {
  632. return false;
  633. }
  634. }
  635. return true;
  636. }
  637. bool Device::TestDepthStencilBlits() const {
  638. static constexpr VkFormatFeatureFlags required_features =
  639. VK_FORMAT_FEATURE_BLIT_SRC_BIT | VK_FORMAT_FEATURE_BLIT_DST_BIT;
  640. const auto test_features = [](VkFormatProperties props) {
  641. return (props.optimalTilingFeatures & required_features) == required_features;
  642. };
  643. return test_features(format_properties.at(VK_FORMAT_D32_SFLOAT_S8_UINT)) &&
  644. test_features(format_properties.at(VK_FORMAT_D24_UNORM_S8_UINT));
  645. }
  646. bool Device::IsFormatSupported(VkFormat wanted_format, VkFormatFeatureFlags wanted_usage,
  647. FormatType format_type) const {
  648. const auto it = format_properties.find(wanted_format);
  649. if (it == format_properties.end()) {
  650. UNIMPLEMENTED_MSG("Unimplemented format query={}", wanted_format);
  651. return true;
  652. }
  653. const auto supported_usage = GetFormatFeatures(it->second, format_type);
  654. return (supported_usage & wanted_usage) == wanted_usage;
  655. }
  656. std::string Device::GetDriverName() const {
  657. switch (driver_id) {
  658. case VK_DRIVER_ID_AMD_PROPRIETARY:
  659. return "AMD";
  660. case VK_DRIVER_ID_AMD_OPEN_SOURCE:
  661. return "AMDVLK";
  662. case VK_DRIVER_ID_MESA_RADV:
  663. return "RADV";
  664. case VK_DRIVER_ID_NVIDIA_PROPRIETARY:
  665. return "NVIDIA";
  666. case VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS:
  667. return "INTEL";
  668. case VK_DRIVER_ID_INTEL_OPEN_SOURCE_MESA:
  669. return "ANV";
  670. case VK_DRIVER_ID_MESA_LLVMPIPE:
  671. return "LAVAPIPE";
  672. default:
  673. return vendor_name;
  674. }
  675. }
  676. void Device::CheckSuitability(bool requires_swapchain) const {
  677. std::bitset<REQUIRED_EXTENSIONS.size()> available_extensions;
  678. bool has_swapchain = false;
  679. for (const VkExtensionProperties& property : physical.EnumerateDeviceExtensionProperties()) {
  680. const std::string_view name{property.extensionName};
  681. for (size_t i = 0; i < REQUIRED_EXTENSIONS.size(); ++i) {
  682. if (available_extensions[i]) {
  683. continue;
  684. }
  685. available_extensions[i] = name == REQUIRED_EXTENSIONS[i];
  686. }
  687. has_swapchain = has_swapchain || name == VK_KHR_SWAPCHAIN_EXTENSION_NAME;
  688. }
  689. for (size_t i = 0; i < REQUIRED_EXTENSIONS.size(); ++i) {
  690. if (available_extensions[i]) {
  691. continue;
  692. }
  693. LOG_ERROR(Render_Vulkan, "Missing required extension: {}", REQUIRED_EXTENSIONS[i]);
  694. throw vk::Exception(VK_ERROR_EXTENSION_NOT_PRESENT);
  695. }
  696. if (requires_swapchain && !has_swapchain) {
  697. LOG_ERROR(Render_Vulkan, "Missing required extension: VK_KHR_swapchain");
  698. throw vk::Exception(VK_ERROR_EXTENSION_NOT_PRESENT);
  699. }
  700. struct LimitTuple {
  701. u32 minimum;
  702. u32 value;
  703. const char* name;
  704. };
  705. const VkPhysicalDeviceLimits& limits{properties.limits};
  706. const std::array limits_report{
  707. LimitTuple{65536, limits.maxUniformBufferRange, "maxUniformBufferRange"},
  708. LimitTuple{16, limits.maxViewports, "maxViewports"},
  709. LimitTuple{8, limits.maxColorAttachments, "maxColorAttachments"},
  710. LimitTuple{8, limits.maxClipDistances, "maxClipDistances"},
  711. };
  712. for (const auto& tuple : limits_report) {
  713. if (tuple.value < tuple.minimum) {
  714. LOG_ERROR(Render_Vulkan, "{} has to be {} or greater but it is {}", tuple.name,
  715. tuple.minimum, tuple.value);
  716. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  717. }
  718. }
  719. VkPhysicalDeviceShaderDemoteToHelperInvocationFeaturesEXT demote{};
  720. demote.sType =
  721. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DEMOTE_TO_HELPER_INVOCATION_FEATURES_EXT;
  722. demote.pNext = nullptr;
  723. VkPhysicalDeviceVariablePointerFeaturesKHR variable_pointers{};
  724. variable_pointers.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES_KHR;
  725. variable_pointers.pNext = &demote;
  726. VkPhysicalDeviceRobustness2FeaturesEXT robustness2{};
  727. robustness2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT;
  728. robustness2.pNext = &variable_pointers;
  729. VkPhysicalDeviceFeatures2KHR features2{};
  730. features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2;
  731. features2.pNext = &robustness2;
  732. physical.GetFeatures2KHR(features2);
  733. const VkPhysicalDeviceFeatures& features{features2.features};
  734. const std::array feature_report{
  735. std::make_pair(features.robustBufferAccess, "robustBufferAccess"),
  736. std::make_pair(features.vertexPipelineStoresAndAtomics, "vertexPipelineStoresAndAtomics"),
  737. std::make_pair(features.imageCubeArray, "imageCubeArray"),
  738. std::make_pair(features.independentBlend, "independentBlend"),
  739. std::make_pair(features.depthClamp, "depthClamp"),
  740. std::make_pair(features.samplerAnisotropy, "samplerAnisotropy"),
  741. std::make_pair(features.largePoints, "largePoints"),
  742. std::make_pair(features.multiViewport, "multiViewport"),
  743. std::make_pair(features.depthBiasClamp, "depthBiasClamp"),
  744. std::make_pair(features.fillModeNonSolid, "fillModeNonSolid"),
  745. std::make_pair(features.wideLines, "wideLines"),
  746. std::make_pair(features.geometryShader, "geometryShader"),
  747. std::make_pair(features.tessellationShader, "tessellationShader"),
  748. std::make_pair(features.sampleRateShading, "sampleRateShading"),
  749. std::make_pair(features.dualSrcBlend, "dualSrcBlend"),
  750. std::make_pair(features.occlusionQueryPrecise, "occlusionQueryPrecise"),
  751. std::make_pair(features.fragmentStoresAndAtomics, "fragmentStoresAndAtomics"),
  752. std::make_pair(features.shaderImageGatherExtended, "shaderImageGatherExtended"),
  753. std::make_pair(features.shaderStorageImageWriteWithoutFormat,
  754. "shaderStorageImageWriteWithoutFormat"),
  755. std::make_pair(features.shaderClipDistance, "shaderClipDistance"),
  756. std::make_pair(features.shaderCullDistance, "shaderCullDistance"),
  757. std::make_pair(demote.shaderDemoteToHelperInvocation, "shaderDemoteToHelperInvocation"),
  758. std::make_pair(variable_pointers.variablePointers, "variablePointers"),
  759. std::make_pair(variable_pointers.variablePointersStorageBuffer,
  760. "variablePointersStorageBuffer"),
  761. std::make_pair(robustness2.robustBufferAccess2, "robustBufferAccess2"),
  762. std::make_pair(robustness2.robustImageAccess2, "robustImageAccess2"),
  763. std::make_pair(robustness2.nullDescriptor, "nullDescriptor"),
  764. };
  765. for (const auto& [is_supported, name] : feature_report) {
  766. if (is_supported) {
  767. continue;
  768. }
  769. LOG_ERROR(Render_Vulkan, "Missing required feature: {}", name);
  770. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  771. }
  772. }
  773. std::vector<const char*> Device::LoadExtensions(bool requires_surface) {
  774. std::vector<const char*> extensions;
  775. extensions.reserve(8 + REQUIRED_EXTENSIONS.size());
  776. extensions.insert(extensions.begin(), REQUIRED_EXTENSIONS.begin(), REQUIRED_EXTENSIONS.end());
  777. if (requires_surface) {
  778. extensions.push_back(VK_KHR_SWAPCHAIN_EXTENSION_NAME);
  779. }
  780. bool has_khr_shader_float16_int8{};
  781. bool has_khr_workgroup_memory_explicit_layout{};
  782. bool has_khr_pipeline_executable_properties{};
  783. bool has_ext_subgroup_size_control{};
  784. bool has_ext_transform_feedback{};
  785. bool has_ext_custom_border_color{};
  786. bool has_ext_extended_dynamic_state{};
  787. bool has_ext_shader_atomic_int64{};
  788. bool has_ext_provoking_vertex{};
  789. bool has_ext_vertex_input_dynamic_state{};
  790. bool has_ext_line_rasterization{};
  791. for (const std::string& extension : supported_extensions) {
  792. const auto test = [&](std::optional<std::reference_wrapper<bool>> status, const char* name,
  793. bool push) {
  794. if (extension != name) {
  795. return;
  796. }
  797. if (push) {
  798. extensions.push_back(name);
  799. }
  800. if (status) {
  801. status->get() = true;
  802. }
  803. };
  804. test(nv_viewport_swizzle, VK_NV_VIEWPORT_SWIZZLE_EXTENSION_NAME, true);
  805. test(nv_viewport_array2, VK_NV_VIEWPORT_ARRAY2_EXTENSION_NAME, true);
  806. test(nv_geometry_shader_passthrough, VK_NV_GEOMETRY_SHADER_PASSTHROUGH_EXTENSION_NAME,
  807. true);
  808. test(khr_uniform_buffer_standard_layout,
  809. VK_KHR_UNIFORM_BUFFER_STANDARD_LAYOUT_EXTENSION_NAME, true);
  810. test(khr_spirv_1_4, VK_KHR_SPIRV_1_4_EXTENSION_NAME, true);
  811. test(khr_push_descriptor, VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME, true);
  812. test(has_khr_shader_float16_int8, VK_KHR_SHADER_FLOAT16_INT8_EXTENSION_NAME, false);
  813. test(ext_depth_range_unrestricted, VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME, true);
  814. test(ext_index_type_uint8, VK_EXT_INDEX_TYPE_UINT8_EXTENSION_NAME, true);
  815. test(ext_sampler_filter_minmax, VK_EXT_SAMPLER_FILTER_MINMAX_EXTENSION_NAME, true);
  816. test(ext_shader_viewport_index_layer, VK_EXT_SHADER_VIEWPORT_INDEX_LAYER_EXTENSION_NAME,
  817. true);
  818. test(ext_tooling_info, VK_EXT_TOOLING_INFO_EXTENSION_NAME, true);
  819. test(ext_shader_stencil_export, VK_EXT_SHADER_STENCIL_EXPORT_EXTENSION_NAME, true);
  820. test(ext_conservative_rasterization, VK_EXT_CONSERVATIVE_RASTERIZATION_EXTENSION_NAME,
  821. true);
  822. test(has_ext_transform_feedback, VK_EXT_TRANSFORM_FEEDBACK_EXTENSION_NAME, false);
  823. test(has_ext_custom_border_color, VK_EXT_CUSTOM_BORDER_COLOR_EXTENSION_NAME, false);
  824. test(has_ext_extended_dynamic_state, VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME, false);
  825. test(has_ext_subgroup_size_control, VK_EXT_SUBGROUP_SIZE_CONTROL_EXTENSION_NAME, false);
  826. test(has_ext_provoking_vertex, VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME, false);
  827. test(has_ext_vertex_input_dynamic_state, VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME,
  828. false);
  829. test(has_ext_shader_atomic_int64, VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME, false);
  830. test(has_khr_workgroup_memory_explicit_layout,
  831. VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME, false);
  832. test(has_ext_line_rasterization, VK_EXT_LINE_RASTERIZATION_EXTENSION_NAME, false);
  833. if (Settings::values.enable_nsight_aftermath) {
  834. test(nv_device_diagnostics_config, VK_NV_DEVICE_DIAGNOSTICS_CONFIG_EXTENSION_NAME,
  835. true);
  836. }
  837. if (Settings::values.renderer_shader_feedback) {
  838. test(has_khr_pipeline_executable_properties,
  839. VK_KHR_PIPELINE_EXECUTABLE_PROPERTIES_EXTENSION_NAME, false);
  840. }
  841. }
  842. VkPhysicalDeviceFeatures2KHR features{};
  843. features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2_KHR;
  844. VkPhysicalDeviceProperties2KHR physical_properties;
  845. physical_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR;
  846. if (has_khr_shader_float16_int8) {
  847. VkPhysicalDeviceFloat16Int8FeaturesKHR float16_int8_features;
  848. float16_int8_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT16_INT8_FEATURES_KHR;
  849. float16_int8_features.pNext = nullptr;
  850. features.pNext = &float16_int8_features;
  851. physical.GetFeatures2KHR(features);
  852. is_float16_supported = float16_int8_features.shaderFloat16;
  853. is_int8_supported = float16_int8_features.shaderInt8;
  854. extensions.push_back(VK_KHR_SHADER_FLOAT16_INT8_EXTENSION_NAME);
  855. }
  856. if (has_ext_subgroup_size_control) {
  857. VkPhysicalDeviceSubgroupSizeControlFeaturesEXT subgroup_features;
  858. subgroup_features.sType =
  859. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT;
  860. subgroup_features.pNext = nullptr;
  861. features.pNext = &subgroup_features;
  862. physical.GetFeatures2KHR(features);
  863. VkPhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_properties;
  864. subgroup_properties.sType =
  865. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_PROPERTIES_EXT;
  866. subgroup_properties.pNext = nullptr;
  867. physical_properties.pNext = &subgroup_properties;
  868. physical.GetProperties2KHR(physical_properties);
  869. is_warp_potentially_bigger = subgroup_properties.maxSubgroupSize > GuestWarpSize;
  870. if (subgroup_features.subgroupSizeControl &&
  871. subgroup_properties.minSubgroupSize <= GuestWarpSize &&
  872. subgroup_properties.maxSubgroupSize >= GuestWarpSize) {
  873. extensions.push_back(VK_EXT_SUBGROUP_SIZE_CONTROL_EXTENSION_NAME);
  874. guest_warp_stages = subgroup_properties.requiredSubgroupSizeStages;
  875. ext_subgroup_size_control = true;
  876. }
  877. } else {
  878. is_warp_potentially_bigger = true;
  879. }
  880. if (has_ext_provoking_vertex) {
  881. VkPhysicalDeviceProvokingVertexFeaturesEXT provoking_vertex;
  882. provoking_vertex.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_FEATURES_EXT;
  883. provoking_vertex.pNext = nullptr;
  884. features.pNext = &provoking_vertex;
  885. physical.GetFeatures2KHR(features);
  886. if (provoking_vertex.provokingVertexLast &&
  887. provoking_vertex.transformFeedbackPreservesProvokingVertex) {
  888. extensions.push_back(VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME);
  889. ext_provoking_vertex = true;
  890. }
  891. }
  892. if (has_ext_vertex_input_dynamic_state) {
  893. VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT vertex_input;
  894. vertex_input.sType =
  895. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_INPUT_DYNAMIC_STATE_FEATURES_EXT;
  896. vertex_input.pNext = nullptr;
  897. features.pNext = &vertex_input;
  898. physical.GetFeatures2KHR(features);
  899. if (vertex_input.vertexInputDynamicState) {
  900. extensions.push_back(VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME);
  901. ext_vertex_input_dynamic_state = true;
  902. }
  903. }
  904. if (has_ext_shader_atomic_int64) {
  905. VkPhysicalDeviceShaderAtomicInt64Features atomic_int64;
  906. atomic_int64.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT;
  907. atomic_int64.pNext = nullptr;
  908. features.pNext = &atomic_int64;
  909. physical.GetFeatures2KHR(features);
  910. if (atomic_int64.shaderBufferInt64Atomics && atomic_int64.shaderSharedInt64Atomics) {
  911. extensions.push_back(VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME);
  912. ext_shader_atomic_int64 = true;
  913. }
  914. }
  915. if (has_ext_transform_feedback) {
  916. VkPhysicalDeviceTransformFeedbackFeaturesEXT tfb_features;
  917. tfb_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT;
  918. tfb_features.pNext = nullptr;
  919. features.pNext = &tfb_features;
  920. physical.GetFeatures2KHR(features);
  921. VkPhysicalDeviceTransformFeedbackPropertiesEXT tfb_properties;
  922. tfb_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT;
  923. tfb_properties.pNext = nullptr;
  924. physical_properties.pNext = &tfb_properties;
  925. physical.GetProperties2KHR(physical_properties);
  926. if (tfb_features.transformFeedback && tfb_features.geometryStreams &&
  927. tfb_properties.maxTransformFeedbackStreams >= 4 &&
  928. tfb_properties.maxTransformFeedbackBuffers && tfb_properties.transformFeedbackQueries &&
  929. tfb_properties.transformFeedbackDraw) {
  930. extensions.push_back(VK_EXT_TRANSFORM_FEEDBACK_EXTENSION_NAME);
  931. ext_transform_feedback = true;
  932. }
  933. }
  934. if (has_ext_custom_border_color) {
  935. VkPhysicalDeviceCustomBorderColorFeaturesEXT border_features;
  936. border_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT;
  937. border_features.pNext = nullptr;
  938. features.pNext = &border_features;
  939. physical.GetFeatures2KHR(features);
  940. if (border_features.customBorderColors && border_features.customBorderColorWithoutFormat) {
  941. extensions.push_back(VK_EXT_CUSTOM_BORDER_COLOR_EXTENSION_NAME);
  942. ext_custom_border_color = true;
  943. }
  944. }
  945. if (has_ext_extended_dynamic_state) {
  946. VkPhysicalDeviceExtendedDynamicStateFeaturesEXT extended_dynamic_state;
  947. extended_dynamic_state.sType =
  948. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT;
  949. extended_dynamic_state.pNext = nullptr;
  950. features.pNext = &extended_dynamic_state;
  951. physical.GetFeatures2KHR(features);
  952. if (extended_dynamic_state.extendedDynamicState) {
  953. extensions.push_back(VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME);
  954. ext_extended_dynamic_state = true;
  955. }
  956. }
  957. if (has_ext_line_rasterization) {
  958. VkPhysicalDeviceLineRasterizationFeaturesEXT line_raster;
  959. line_raster.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT;
  960. line_raster.pNext = nullptr;
  961. features.pNext = &line_raster;
  962. physical.GetFeatures2KHR(features);
  963. if (line_raster.rectangularLines && line_raster.smoothLines) {
  964. extensions.push_back(VK_EXT_LINE_RASTERIZATION_EXTENSION_NAME);
  965. ext_line_rasterization = true;
  966. }
  967. }
  968. if (has_khr_workgroup_memory_explicit_layout) {
  969. VkPhysicalDeviceWorkgroupMemoryExplicitLayoutFeaturesKHR layout;
  970. layout.sType =
  971. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_FEATURES_KHR;
  972. layout.pNext = nullptr;
  973. features.pNext = &layout;
  974. physical.GetFeatures2KHR(features);
  975. if (layout.workgroupMemoryExplicitLayout &&
  976. layout.workgroupMemoryExplicitLayout8BitAccess &&
  977. layout.workgroupMemoryExplicitLayout16BitAccess &&
  978. layout.workgroupMemoryExplicitLayoutScalarBlockLayout) {
  979. extensions.push_back(VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME);
  980. khr_workgroup_memory_explicit_layout = true;
  981. }
  982. }
  983. if (has_khr_pipeline_executable_properties) {
  984. VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR executable_properties;
  985. executable_properties.sType =
  986. VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR;
  987. executable_properties.pNext = nullptr;
  988. features.pNext = &executable_properties;
  989. physical.GetFeatures2KHR(features);
  990. if (executable_properties.pipelineExecutableInfo) {
  991. extensions.push_back(VK_KHR_PIPELINE_EXECUTABLE_PROPERTIES_EXTENSION_NAME);
  992. khr_pipeline_executable_properties = true;
  993. }
  994. }
  995. if (khr_push_descriptor) {
  996. VkPhysicalDevicePushDescriptorPropertiesKHR push_descriptor;
  997. push_descriptor.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR;
  998. push_descriptor.pNext = nullptr;
  999. physical_properties.pNext = &push_descriptor;
  1000. physical.GetProperties2KHR(physical_properties);
  1001. max_push_descriptors = push_descriptor.maxPushDescriptors;
  1002. }
  1003. return extensions;
  1004. }
  1005. void Device::SetupFamilies(VkSurfaceKHR surface) {
  1006. const std::vector queue_family_properties = physical.GetQueueFamilyProperties();
  1007. std::optional<u32> graphics;
  1008. std::optional<u32> present;
  1009. for (u32 index = 0; index < static_cast<u32>(queue_family_properties.size()); ++index) {
  1010. if (graphics && (present || !surface)) {
  1011. break;
  1012. }
  1013. const VkQueueFamilyProperties& queue_family = queue_family_properties[index];
  1014. if (queue_family.queueCount == 0) {
  1015. continue;
  1016. }
  1017. if (queue_family.queueFlags & VK_QUEUE_GRAPHICS_BIT) {
  1018. graphics = index;
  1019. }
  1020. if (surface && physical.GetSurfaceSupportKHR(index, surface)) {
  1021. present = index;
  1022. }
  1023. }
  1024. if (!graphics) {
  1025. LOG_ERROR(Render_Vulkan, "Device lacks a graphics queue");
  1026. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  1027. }
  1028. if (surface && !present) {
  1029. LOG_ERROR(Render_Vulkan, "Device lacks a present queue");
  1030. throw vk::Exception(VK_ERROR_FEATURE_NOT_PRESENT);
  1031. }
  1032. graphics_family = *graphics;
  1033. present_family = *present;
  1034. }
  1035. void Device::SetupFeatures() {
  1036. const VkPhysicalDeviceFeatures features{physical.GetFeatures()};
  1037. is_depth_bounds_supported = features.depthBounds;
  1038. is_formatless_image_load_supported = features.shaderStorageImageReadWithoutFormat;
  1039. is_shader_float64_supported = features.shaderFloat64;
  1040. is_shader_int64_supported = features.shaderInt64;
  1041. is_shader_int16_supported = features.shaderInt16;
  1042. is_shader_storage_image_multisample = features.shaderStorageImageMultisample;
  1043. is_blit_depth_stencil_supported = TestDepthStencilBlits();
  1044. is_optimal_astc_supported = IsOptimalAstcSupported(features);
  1045. }
  1046. void Device::SetupProperties() {
  1047. float_controls.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FLOAT_CONTROLS_PROPERTIES_KHR;
  1048. VkPhysicalDeviceProperties2KHR properties2{};
  1049. properties2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR;
  1050. properties2.pNext = &float_controls;
  1051. physical.GetProperties2KHR(properties2);
  1052. }
  1053. void Device::CollectTelemetryParameters() {
  1054. VkPhysicalDeviceDriverPropertiesKHR driver{
  1055. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRIVER_PROPERTIES_KHR,
  1056. .pNext = nullptr,
  1057. .driverID = {},
  1058. .driverName = {},
  1059. .driverInfo = {},
  1060. .conformanceVersion = {},
  1061. };
  1062. VkPhysicalDeviceProperties2KHR device_properties{
  1063. .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR,
  1064. .pNext = &driver,
  1065. .properties = {},
  1066. };
  1067. physical.GetProperties2KHR(device_properties);
  1068. driver_id = driver.driverID;
  1069. vendor_name = driver.driverName;
  1070. }
  1071. void Device::CollectPhysicalMemoryInfo() {
  1072. const auto mem_properties = physical.GetMemoryProperties();
  1073. const size_t num_properties = mem_properties.memoryHeapCount;
  1074. device_access_memory = 0;
  1075. for (size_t element = 0; element < num_properties; ++element) {
  1076. if ((mem_properties.memoryHeaps[element].flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) != 0) {
  1077. device_access_memory += mem_properties.memoryHeaps[element].size;
  1078. }
  1079. }
  1080. }
  1081. void Device::CollectToolingInfo() {
  1082. if (!ext_tooling_info) {
  1083. return;
  1084. }
  1085. const auto vkGetPhysicalDeviceToolPropertiesEXT =
  1086. reinterpret_cast<PFN_vkGetPhysicalDeviceToolPropertiesEXT>(
  1087. dld.vkGetInstanceProcAddr(instance, "vkGetPhysicalDeviceToolPropertiesEXT"));
  1088. if (!vkGetPhysicalDeviceToolPropertiesEXT) {
  1089. return;
  1090. }
  1091. u32 tool_count = 0;
  1092. if (vkGetPhysicalDeviceToolPropertiesEXT(physical, &tool_count, nullptr) != VK_SUCCESS) {
  1093. return;
  1094. }
  1095. std::vector<VkPhysicalDeviceToolPropertiesEXT> tools(tool_count);
  1096. if (vkGetPhysicalDeviceToolPropertiesEXT(physical, &tool_count, tools.data()) != VK_SUCCESS) {
  1097. return;
  1098. }
  1099. for (const VkPhysicalDeviceToolPropertiesEXT& tool : tools) {
  1100. const std::string_view name = tool.name;
  1101. LOG_INFO(Render_Vulkan, "{}", name);
  1102. has_renderdoc = has_renderdoc || name == "RenderDoc";
  1103. has_nsight_graphics = has_nsight_graphics || name == "NVIDIA Nsight Graphics";
  1104. }
  1105. }
  1106. std::vector<VkDeviceQueueCreateInfo> Device::GetDeviceQueueCreateInfos() const {
  1107. static constexpr float QUEUE_PRIORITY = 1.0f;
  1108. std::unordered_set<u32> unique_queue_families{graphics_family, present_family};
  1109. std::vector<VkDeviceQueueCreateInfo> queue_cis;
  1110. queue_cis.reserve(unique_queue_families.size());
  1111. for (const u32 queue_family : unique_queue_families) {
  1112. auto& ci = queue_cis.emplace_back(VkDeviceQueueCreateInfo{
  1113. .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO,
  1114. .pNext = nullptr,
  1115. .flags = 0,
  1116. .queueFamilyIndex = queue_family,
  1117. .queueCount = 1,
  1118. .pQueuePriorities = nullptr,
  1119. });
  1120. ci.pQueuePriorities = &QUEUE_PRIORITY;
  1121. }
  1122. return queue_cis;
  1123. }
  1124. } // namespace Vulkan