collect_shader_info_pass.cpp 35 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936
  1. // Copyright 2021 yuzu Emulator Project
  2. // Licensed under GPLv2 or any later version
  3. // Refer to the license.txt file included.
  4. #include "common/alignment.h"
  5. #include "shader_recompiler/environment.h"
  6. #include "shader_recompiler/frontend/ir/modifiers.h"
  7. #include "shader_recompiler/frontend/ir/program.h"
  8. #include "shader_recompiler/frontend/ir/value.h"
  9. #include "shader_recompiler/ir_opt/passes.h"
  10. #include "shader_recompiler/shader_info.h"
  11. namespace Shader::Optimization {
  12. namespace {
  13. void AddConstantBufferDescriptor(Info& info, u32 index, u32 count) {
  14. if (count != 1) {
  15. throw NotImplementedException("Constant buffer descriptor indexing");
  16. }
  17. if ((info.constant_buffer_mask & (1U << index)) != 0) {
  18. return;
  19. }
  20. info.constant_buffer_mask |= 1U << index;
  21. auto& cbufs{info.constant_buffer_descriptors};
  22. cbufs.insert(std::ranges::lower_bound(cbufs, index, {}, &ConstantBufferDescriptor::index),
  23. ConstantBufferDescriptor{
  24. .index = index,
  25. .count = 1,
  26. });
  27. }
  28. void GetPatch(Info& info, IR::Patch patch) {
  29. if (!IR::IsGeneric(patch)) {
  30. throw NotImplementedException("Reading non-generic patch {}", patch);
  31. }
  32. info.uses_patches.at(IR::GenericPatchIndex(patch)) = true;
  33. }
  34. void SetPatch(Info& info, IR::Patch patch) {
  35. if (IR::IsGeneric(patch)) {
  36. info.uses_patches.at(IR::GenericPatchIndex(patch)) = true;
  37. return;
  38. }
  39. switch (patch) {
  40. case IR::Patch::TessellationLodLeft:
  41. case IR::Patch::TessellationLodTop:
  42. case IR::Patch::TessellationLodRight:
  43. case IR::Patch::TessellationLodBottom:
  44. info.stores_tess_level_outer = true;
  45. break;
  46. case IR::Patch::TessellationLodInteriorU:
  47. case IR::Patch::TessellationLodInteriorV:
  48. info.stores_tess_level_inner = true;
  49. break;
  50. default:
  51. throw NotImplementedException("Set patch {}", patch);
  52. }
  53. }
  54. void CheckCBufNVN(Info& info, IR::Inst& inst) {
  55. const IR::Value cbuf_index{inst.Arg(0)};
  56. if (!cbuf_index.IsImmediate()) {
  57. info.nvn_buffer_used.set();
  58. return;
  59. }
  60. const u32 index{cbuf_index.U32()};
  61. if (index != 0) {
  62. return;
  63. }
  64. const IR::Value cbuf_offset{inst.Arg(1)};
  65. if (!cbuf_offset.IsImmediate()) {
  66. info.nvn_buffer_used.set();
  67. return;
  68. }
  69. const u32 offset{cbuf_offset.U32()};
  70. const u32 descriptor_size{0x10};
  71. const u32 upper_limit{info.nvn_buffer_base + descriptor_size * 16};
  72. if (offset >= info.nvn_buffer_base && offset < upper_limit) {
  73. const std::size_t nvn_index{(offset - info.nvn_buffer_base) / descriptor_size};
  74. info.nvn_buffer_used.set(nvn_index, true);
  75. }
  76. }
  77. void VisitUsages(Info& info, IR::Inst& inst) {
  78. switch (inst.GetOpcode()) {
  79. case IR::Opcode::CompositeConstructF16x2:
  80. case IR::Opcode::CompositeConstructF16x3:
  81. case IR::Opcode::CompositeConstructF16x4:
  82. case IR::Opcode::CompositeExtractF16x2:
  83. case IR::Opcode::CompositeExtractF16x3:
  84. case IR::Opcode::CompositeExtractF16x4:
  85. case IR::Opcode::CompositeInsertF16x2:
  86. case IR::Opcode::CompositeInsertF16x3:
  87. case IR::Opcode::CompositeInsertF16x4:
  88. case IR::Opcode::SelectF16:
  89. case IR::Opcode::BitCastU16F16:
  90. case IR::Opcode::BitCastF16U16:
  91. case IR::Opcode::PackFloat2x16:
  92. case IR::Opcode::UnpackFloat2x16:
  93. case IR::Opcode::ConvertS16F16:
  94. case IR::Opcode::ConvertS32F16:
  95. case IR::Opcode::ConvertS64F16:
  96. case IR::Opcode::ConvertU16F16:
  97. case IR::Opcode::ConvertU32F16:
  98. case IR::Opcode::ConvertU64F16:
  99. case IR::Opcode::ConvertF16S8:
  100. case IR::Opcode::ConvertF16S16:
  101. case IR::Opcode::ConvertF16S32:
  102. case IR::Opcode::ConvertF16S64:
  103. case IR::Opcode::ConvertF16U8:
  104. case IR::Opcode::ConvertF16U16:
  105. case IR::Opcode::ConvertF16U32:
  106. case IR::Opcode::ConvertF16U64:
  107. case IR::Opcode::ConvertF16F32:
  108. case IR::Opcode::ConvertF32F16:
  109. case IR::Opcode::FPAbs16:
  110. case IR::Opcode::FPAdd16:
  111. case IR::Opcode::FPCeil16:
  112. case IR::Opcode::FPFloor16:
  113. case IR::Opcode::FPFma16:
  114. case IR::Opcode::FPMul16:
  115. case IR::Opcode::FPNeg16:
  116. case IR::Opcode::FPRoundEven16:
  117. case IR::Opcode::FPSaturate16:
  118. case IR::Opcode::FPClamp16:
  119. case IR::Opcode::FPTrunc16:
  120. case IR::Opcode::FPOrdEqual16:
  121. case IR::Opcode::FPUnordEqual16:
  122. case IR::Opcode::FPOrdNotEqual16:
  123. case IR::Opcode::FPUnordNotEqual16:
  124. case IR::Opcode::FPOrdLessThan16:
  125. case IR::Opcode::FPUnordLessThan16:
  126. case IR::Opcode::FPOrdGreaterThan16:
  127. case IR::Opcode::FPUnordGreaterThan16:
  128. case IR::Opcode::FPOrdLessThanEqual16:
  129. case IR::Opcode::FPUnordLessThanEqual16:
  130. case IR::Opcode::FPOrdGreaterThanEqual16:
  131. case IR::Opcode::FPUnordGreaterThanEqual16:
  132. case IR::Opcode::FPIsNan16:
  133. case IR::Opcode::GlobalAtomicAddF16x2:
  134. case IR::Opcode::GlobalAtomicMinF16x2:
  135. case IR::Opcode::GlobalAtomicMaxF16x2:
  136. case IR::Opcode::StorageAtomicAddF16x2:
  137. case IR::Opcode::StorageAtomicMinF16x2:
  138. case IR::Opcode::StorageAtomicMaxF16x2:
  139. info.uses_fp16 = true;
  140. break;
  141. case IR::Opcode::CompositeConstructF64x2:
  142. case IR::Opcode::CompositeConstructF64x3:
  143. case IR::Opcode::CompositeConstructF64x4:
  144. case IR::Opcode::CompositeExtractF64x2:
  145. case IR::Opcode::CompositeExtractF64x3:
  146. case IR::Opcode::CompositeExtractF64x4:
  147. case IR::Opcode::CompositeInsertF64x2:
  148. case IR::Opcode::CompositeInsertF64x3:
  149. case IR::Opcode::CompositeInsertF64x4:
  150. case IR::Opcode::SelectF64:
  151. case IR::Opcode::BitCastU64F64:
  152. case IR::Opcode::BitCastF64U64:
  153. case IR::Opcode::PackDouble2x32:
  154. case IR::Opcode::UnpackDouble2x32:
  155. case IR::Opcode::FPAbs64:
  156. case IR::Opcode::FPAdd64:
  157. case IR::Opcode::FPCeil64:
  158. case IR::Opcode::FPFloor64:
  159. case IR::Opcode::FPFma64:
  160. case IR::Opcode::FPMax64:
  161. case IR::Opcode::FPMin64:
  162. case IR::Opcode::FPMul64:
  163. case IR::Opcode::FPNeg64:
  164. case IR::Opcode::FPRecip64:
  165. case IR::Opcode::FPRecipSqrt64:
  166. case IR::Opcode::FPRoundEven64:
  167. case IR::Opcode::FPSaturate64:
  168. case IR::Opcode::FPClamp64:
  169. case IR::Opcode::FPTrunc64:
  170. case IR::Opcode::FPOrdEqual64:
  171. case IR::Opcode::FPUnordEqual64:
  172. case IR::Opcode::FPOrdNotEqual64:
  173. case IR::Opcode::FPUnordNotEqual64:
  174. case IR::Opcode::FPOrdLessThan64:
  175. case IR::Opcode::FPUnordLessThan64:
  176. case IR::Opcode::FPOrdGreaterThan64:
  177. case IR::Opcode::FPUnordGreaterThan64:
  178. case IR::Opcode::FPOrdLessThanEqual64:
  179. case IR::Opcode::FPUnordLessThanEqual64:
  180. case IR::Opcode::FPOrdGreaterThanEqual64:
  181. case IR::Opcode::FPUnordGreaterThanEqual64:
  182. case IR::Opcode::FPIsNan64:
  183. case IR::Opcode::ConvertS16F64:
  184. case IR::Opcode::ConvertS32F64:
  185. case IR::Opcode::ConvertS64F64:
  186. case IR::Opcode::ConvertU16F64:
  187. case IR::Opcode::ConvertU32F64:
  188. case IR::Opcode::ConvertU64F64:
  189. case IR::Opcode::ConvertF32F64:
  190. case IR::Opcode::ConvertF64F32:
  191. case IR::Opcode::ConvertF64S8:
  192. case IR::Opcode::ConvertF64S16:
  193. case IR::Opcode::ConvertF64S32:
  194. case IR::Opcode::ConvertF64S64:
  195. case IR::Opcode::ConvertF64U8:
  196. case IR::Opcode::ConvertF64U16:
  197. case IR::Opcode::ConvertF64U32:
  198. case IR::Opcode::ConvertF64U64:
  199. info.uses_fp64 = true;
  200. break;
  201. default:
  202. break;
  203. }
  204. switch (inst.GetOpcode()) {
  205. case IR::Opcode::GetCbufU8:
  206. case IR::Opcode::GetCbufS8:
  207. case IR::Opcode::UndefU8:
  208. case IR::Opcode::LoadGlobalU8:
  209. case IR::Opcode::LoadGlobalS8:
  210. case IR::Opcode::WriteGlobalU8:
  211. case IR::Opcode::WriteGlobalS8:
  212. case IR::Opcode::LoadStorageU8:
  213. case IR::Opcode::LoadStorageS8:
  214. case IR::Opcode::WriteStorageU8:
  215. case IR::Opcode::WriteStorageS8:
  216. case IR::Opcode::LoadSharedU8:
  217. case IR::Opcode::LoadSharedS8:
  218. case IR::Opcode::WriteSharedU8:
  219. case IR::Opcode::SelectU8:
  220. case IR::Opcode::ConvertF16S8:
  221. case IR::Opcode::ConvertF16U8:
  222. case IR::Opcode::ConvertF32S8:
  223. case IR::Opcode::ConvertF32U8:
  224. case IR::Opcode::ConvertF64S8:
  225. case IR::Opcode::ConvertF64U8:
  226. info.uses_int8 = true;
  227. break;
  228. default:
  229. break;
  230. }
  231. switch (inst.GetOpcode()) {
  232. case IR::Opcode::GetCbufU16:
  233. case IR::Opcode::GetCbufS16:
  234. case IR::Opcode::UndefU16:
  235. case IR::Opcode::LoadGlobalU16:
  236. case IR::Opcode::LoadGlobalS16:
  237. case IR::Opcode::WriteGlobalU16:
  238. case IR::Opcode::WriteGlobalS16:
  239. case IR::Opcode::LoadStorageU16:
  240. case IR::Opcode::LoadStorageS16:
  241. case IR::Opcode::WriteStorageU16:
  242. case IR::Opcode::WriteStorageS16:
  243. case IR::Opcode::LoadSharedU16:
  244. case IR::Opcode::LoadSharedS16:
  245. case IR::Opcode::WriteSharedU16:
  246. case IR::Opcode::SelectU16:
  247. case IR::Opcode::BitCastU16F16:
  248. case IR::Opcode::BitCastF16U16:
  249. case IR::Opcode::ConvertS16F16:
  250. case IR::Opcode::ConvertS16F32:
  251. case IR::Opcode::ConvertS16F64:
  252. case IR::Opcode::ConvertU16F16:
  253. case IR::Opcode::ConvertU16F32:
  254. case IR::Opcode::ConvertU16F64:
  255. case IR::Opcode::ConvertF16S16:
  256. case IR::Opcode::ConvertF16U16:
  257. case IR::Opcode::ConvertF32S16:
  258. case IR::Opcode::ConvertF32U16:
  259. case IR::Opcode::ConvertF64S16:
  260. case IR::Opcode::ConvertF64U16:
  261. info.uses_int16 = true;
  262. break;
  263. default:
  264. break;
  265. }
  266. switch (inst.GetOpcode()) {
  267. case IR::Opcode::UndefU64:
  268. case IR::Opcode::LoadGlobalU8:
  269. case IR::Opcode::LoadGlobalS8:
  270. case IR::Opcode::LoadGlobalU16:
  271. case IR::Opcode::LoadGlobalS16:
  272. case IR::Opcode::LoadGlobal32:
  273. case IR::Opcode::LoadGlobal64:
  274. case IR::Opcode::LoadGlobal128:
  275. case IR::Opcode::WriteGlobalU8:
  276. case IR::Opcode::WriteGlobalS8:
  277. case IR::Opcode::WriteGlobalU16:
  278. case IR::Opcode::WriteGlobalS16:
  279. case IR::Opcode::WriteGlobal32:
  280. case IR::Opcode::WriteGlobal64:
  281. case IR::Opcode::WriteGlobal128:
  282. case IR::Opcode::SelectU64:
  283. case IR::Opcode::BitCastU64F64:
  284. case IR::Opcode::BitCastF64U64:
  285. case IR::Opcode::PackUint2x32:
  286. case IR::Opcode::UnpackUint2x32:
  287. case IR::Opcode::IAdd64:
  288. case IR::Opcode::ISub64:
  289. case IR::Opcode::INeg64:
  290. case IR::Opcode::ShiftLeftLogical64:
  291. case IR::Opcode::ShiftRightLogical64:
  292. case IR::Opcode::ShiftRightArithmetic64:
  293. case IR::Opcode::ConvertS64F16:
  294. case IR::Opcode::ConvertS64F32:
  295. case IR::Opcode::ConvertS64F64:
  296. case IR::Opcode::ConvertU64F16:
  297. case IR::Opcode::ConvertU64F32:
  298. case IR::Opcode::ConvertU64F64:
  299. case IR::Opcode::ConvertU64U32:
  300. case IR::Opcode::ConvertU32U64:
  301. case IR::Opcode::ConvertF16U64:
  302. case IR::Opcode::ConvertF32U64:
  303. case IR::Opcode::ConvertF64U64:
  304. case IR::Opcode::SharedAtomicExchange64:
  305. case IR::Opcode::GlobalAtomicIAdd64:
  306. case IR::Opcode::GlobalAtomicSMin64:
  307. case IR::Opcode::GlobalAtomicUMin64:
  308. case IR::Opcode::GlobalAtomicSMax64:
  309. case IR::Opcode::GlobalAtomicUMax64:
  310. case IR::Opcode::GlobalAtomicAnd64:
  311. case IR::Opcode::GlobalAtomicOr64:
  312. case IR::Opcode::GlobalAtomicXor64:
  313. case IR::Opcode::GlobalAtomicExchange64:
  314. case IR::Opcode::StorageAtomicIAdd64:
  315. case IR::Opcode::StorageAtomicSMin64:
  316. case IR::Opcode::StorageAtomicUMin64:
  317. case IR::Opcode::StorageAtomicSMax64:
  318. case IR::Opcode::StorageAtomicUMax64:
  319. case IR::Opcode::StorageAtomicAnd64:
  320. case IR::Opcode::StorageAtomicOr64:
  321. case IR::Opcode::StorageAtomicXor64:
  322. case IR::Opcode::StorageAtomicExchange64:
  323. info.uses_int64 = true;
  324. break;
  325. default:
  326. break;
  327. }
  328. switch (inst.GetOpcode()) {
  329. case IR::Opcode::WriteGlobalU8:
  330. case IR::Opcode::WriteGlobalS8:
  331. case IR::Opcode::WriteGlobalU16:
  332. case IR::Opcode::WriteGlobalS16:
  333. case IR::Opcode::WriteGlobal32:
  334. case IR::Opcode::WriteGlobal64:
  335. case IR::Opcode::WriteGlobal128:
  336. case IR::Opcode::GlobalAtomicIAdd32:
  337. case IR::Opcode::GlobalAtomicSMin32:
  338. case IR::Opcode::GlobalAtomicUMin32:
  339. case IR::Opcode::GlobalAtomicSMax32:
  340. case IR::Opcode::GlobalAtomicUMax32:
  341. case IR::Opcode::GlobalAtomicInc32:
  342. case IR::Opcode::GlobalAtomicDec32:
  343. case IR::Opcode::GlobalAtomicAnd32:
  344. case IR::Opcode::GlobalAtomicOr32:
  345. case IR::Opcode::GlobalAtomicXor32:
  346. case IR::Opcode::GlobalAtomicExchange32:
  347. case IR::Opcode::GlobalAtomicIAdd64:
  348. case IR::Opcode::GlobalAtomicSMin64:
  349. case IR::Opcode::GlobalAtomicUMin64:
  350. case IR::Opcode::GlobalAtomicSMax64:
  351. case IR::Opcode::GlobalAtomicUMax64:
  352. case IR::Opcode::GlobalAtomicAnd64:
  353. case IR::Opcode::GlobalAtomicOr64:
  354. case IR::Opcode::GlobalAtomicXor64:
  355. case IR::Opcode::GlobalAtomicExchange64:
  356. case IR::Opcode::GlobalAtomicAddF32:
  357. case IR::Opcode::GlobalAtomicAddF16x2:
  358. case IR::Opcode::GlobalAtomicAddF32x2:
  359. case IR::Opcode::GlobalAtomicMinF16x2:
  360. case IR::Opcode::GlobalAtomicMinF32x2:
  361. case IR::Opcode::GlobalAtomicMaxF16x2:
  362. case IR::Opcode::GlobalAtomicMaxF32x2:
  363. info.stores_global_memory = true;
  364. [[fallthrough]];
  365. case IR::Opcode::LoadGlobalU8:
  366. case IR::Opcode::LoadGlobalS8:
  367. case IR::Opcode::LoadGlobalU16:
  368. case IR::Opcode::LoadGlobalS16:
  369. case IR::Opcode::LoadGlobal32:
  370. case IR::Opcode::LoadGlobal64:
  371. case IR::Opcode::LoadGlobal128:
  372. info.uses_int64 = true;
  373. info.uses_global_memory = true;
  374. info.used_constant_buffer_types |= IR::Type::U32 | IR::Type::U32x2;
  375. info.used_storage_buffer_types |= IR::Type::U32 | IR::Type::U32x2 | IR::Type::U32x4;
  376. break;
  377. default:
  378. break;
  379. }
  380. switch (inst.GetOpcode()) {
  381. case IR::Opcode::DemoteToHelperInvocation:
  382. info.uses_demote_to_helper_invocation = true;
  383. break;
  384. case IR::Opcode::GetAttribute:
  385. case IR::Opcode::GetAttributeU32:
  386. info.loads.mask[static_cast<size_t>(inst.Arg(0).Attribute())] = true;
  387. break;
  388. case IR::Opcode::SetAttribute:
  389. info.stores.mask[static_cast<size_t>(inst.Arg(0).Attribute())] = true;
  390. break;
  391. case IR::Opcode::GetPatch:
  392. GetPatch(info, inst.Arg(0).Patch());
  393. break;
  394. case IR::Opcode::SetPatch:
  395. SetPatch(info, inst.Arg(0).Patch());
  396. break;
  397. case IR::Opcode::GetAttributeIndexed:
  398. info.loads_indexed_attributes = true;
  399. break;
  400. case IR::Opcode::SetAttributeIndexed:
  401. info.stores_indexed_attributes = true;
  402. break;
  403. case IR::Opcode::SetFragColor:
  404. info.stores_frag_color[inst.Arg(0).U32()] = true;
  405. break;
  406. case IR::Opcode::SetSampleMask:
  407. info.stores_sample_mask = true;
  408. break;
  409. case IR::Opcode::SetFragDepth:
  410. info.stores_frag_depth = true;
  411. break;
  412. case IR::Opcode::WorkgroupId:
  413. info.uses_workgroup_id = true;
  414. break;
  415. case IR::Opcode::LocalInvocationId:
  416. info.uses_local_invocation_id = true;
  417. break;
  418. case IR::Opcode::InvocationId:
  419. info.uses_invocation_id = true;
  420. break;
  421. case IR::Opcode::SampleId:
  422. info.uses_sample_id = true;
  423. break;
  424. case IR::Opcode::IsHelperInvocation:
  425. info.uses_is_helper_invocation = true;
  426. break;
  427. case IR::Opcode::ResolutionDownFactor:
  428. case IR::Opcode::IsTextureScaled:
  429. case IR::Opcode::IsImageScaled:
  430. info.uses_rescaling_uniform = true;
  431. break;
  432. case IR::Opcode::LaneId:
  433. info.uses_subgroup_invocation_id = true;
  434. break;
  435. case IR::Opcode::ShuffleIndex:
  436. case IR::Opcode::ShuffleUp:
  437. case IR::Opcode::ShuffleDown:
  438. case IR::Opcode::ShuffleButterfly:
  439. info.uses_subgroup_shuffles = true;
  440. break;
  441. case IR::Opcode::GetCbufU8:
  442. case IR::Opcode::GetCbufS8:
  443. case IR::Opcode::GetCbufU16:
  444. case IR::Opcode::GetCbufS16:
  445. case IR::Opcode::GetCbufU32:
  446. case IR::Opcode::GetCbufF32:
  447. case IR::Opcode::GetCbufU32x2: {
  448. const IR::Value index{inst.Arg(0)};
  449. const IR::Value offset{inst.Arg(1)};
  450. if (!index.IsImmediate()) {
  451. throw NotImplementedException("Constant buffer with non-immediate index");
  452. }
  453. AddConstantBufferDescriptor(info, index.U32(), 1);
  454. u32 element_size{};
  455. switch (inst.GetOpcode()) {
  456. case IR::Opcode::GetCbufU8:
  457. case IR::Opcode::GetCbufS8:
  458. info.used_constant_buffer_types |= IR::Type::U8;
  459. element_size = 1;
  460. break;
  461. case IR::Opcode::GetCbufU16:
  462. case IR::Opcode::GetCbufS16:
  463. info.used_constant_buffer_types |= IR::Type::U16;
  464. element_size = 2;
  465. break;
  466. case IR::Opcode::GetCbufU32:
  467. info.used_constant_buffer_types |= IR::Type::U32;
  468. element_size = 4;
  469. break;
  470. case IR::Opcode::GetCbufF32:
  471. info.used_constant_buffer_types |= IR::Type::F32;
  472. element_size = 4;
  473. break;
  474. case IR::Opcode::GetCbufU32x2:
  475. info.used_constant_buffer_types |= IR::Type::U32x2;
  476. element_size = 8;
  477. break;
  478. default:
  479. break;
  480. }
  481. u32& size{info.constant_buffer_used_sizes[index.U32()]};
  482. if (offset.IsImmediate()) {
  483. size = Common::AlignUp(std::max(size, offset.U32() + element_size), 16u);
  484. } else {
  485. size = 0x10'000;
  486. }
  487. break;
  488. }
  489. case IR::Opcode::BindlessImageSampleImplicitLod:
  490. case IR::Opcode::BindlessImageSampleExplicitLod:
  491. case IR::Opcode::BindlessImageSampleDrefImplicitLod:
  492. case IR::Opcode::BindlessImageSampleDrefExplicitLod:
  493. case IR::Opcode::BindlessImageGather:
  494. case IR::Opcode::BindlessImageGatherDref:
  495. case IR::Opcode::BindlessImageFetch:
  496. case IR::Opcode::BindlessImageQueryDimensions:
  497. case IR::Opcode::BindlessImageQueryLod:
  498. case IR::Opcode::BindlessImageGradient:
  499. case IR::Opcode::BoundImageSampleImplicitLod:
  500. case IR::Opcode::BoundImageSampleExplicitLod:
  501. case IR::Opcode::BoundImageSampleDrefImplicitLod:
  502. case IR::Opcode::BoundImageSampleDrefExplicitLod:
  503. case IR::Opcode::BoundImageGather:
  504. case IR::Opcode::BoundImageGatherDref:
  505. case IR::Opcode::BoundImageFetch:
  506. case IR::Opcode::BoundImageQueryDimensions:
  507. case IR::Opcode::BoundImageQueryLod:
  508. case IR::Opcode::BoundImageGradient:
  509. case IR::Opcode::ImageGather:
  510. case IR::Opcode::ImageGatherDref:
  511. case IR::Opcode::ImageFetch:
  512. case IR::Opcode::ImageQueryDimensions:
  513. case IR::Opcode::ImageGradient: {
  514. const TextureType type{inst.Flags<IR::TextureInstInfo>().type};
  515. info.uses_sampled_1d |= type == TextureType::Color1D || type == TextureType::ColorArray1D;
  516. info.uses_sparse_residency |=
  517. inst.GetAssociatedPseudoOperation(IR::Opcode::GetSparseFromOp) != nullptr;
  518. break;
  519. }
  520. case IR::Opcode::ImageSampleImplicitLod:
  521. case IR::Opcode::ImageSampleExplicitLod:
  522. case IR::Opcode::ImageSampleDrefImplicitLod:
  523. case IR::Opcode::ImageSampleDrefExplicitLod:
  524. case IR::Opcode::ImageQueryLod: {
  525. const auto flags{inst.Flags<IR::TextureInstInfo>()};
  526. const TextureType type{flags.type};
  527. info.uses_sampled_1d |= type == TextureType::Color1D || type == TextureType::ColorArray1D;
  528. info.uses_shadow_lod |= flags.is_depth != 0;
  529. info.uses_sparse_residency |=
  530. inst.GetAssociatedPseudoOperation(IR::Opcode::GetSparseFromOp) != nullptr;
  531. break;
  532. }
  533. case IR::Opcode::ImageRead: {
  534. const auto flags{inst.Flags<IR::TextureInstInfo>()};
  535. info.uses_typeless_image_reads |= flags.image_format == ImageFormat::Typeless;
  536. info.uses_sparse_residency |=
  537. inst.GetAssociatedPseudoOperation(IR::Opcode::GetSparseFromOp) != nullptr;
  538. break;
  539. }
  540. case IR::Opcode::ImageWrite: {
  541. const auto flags{inst.Flags<IR::TextureInstInfo>()};
  542. info.uses_typeless_image_writes |= flags.image_format == ImageFormat::Typeless;
  543. info.uses_image_buffers |= flags.type == TextureType::Buffer;
  544. break;
  545. }
  546. case IR::Opcode::SubgroupEqMask:
  547. case IR::Opcode::SubgroupLtMask:
  548. case IR::Opcode::SubgroupLeMask:
  549. case IR::Opcode::SubgroupGtMask:
  550. case IR::Opcode::SubgroupGeMask:
  551. info.uses_subgroup_mask = true;
  552. break;
  553. case IR::Opcode::VoteAll:
  554. case IR::Opcode::VoteAny:
  555. case IR::Opcode::VoteEqual:
  556. case IR::Opcode::SubgroupBallot:
  557. info.uses_subgroup_vote = true;
  558. break;
  559. case IR::Opcode::FSwizzleAdd:
  560. info.uses_fswzadd = true;
  561. break;
  562. case IR::Opcode::DPdxFine:
  563. case IR::Opcode::DPdyFine:
  564. case IR::Opcode::DPdxCoarse:
  565. case IR::Opcode::DPdyCoarse:
  566. info.uses_derivatives = true;
  567. break;
  568. case IR::Opcode::LoadStorageU8:
  569. case IR::Opcode::LoadStorageS8:
  570. case IR::Opcode::WriteStorageU8:
  571. case IR::Opcode::WriteStorageS8:
  572. info.used_storage_buffer_types |= IR::Type::U8;
  573. break;
  574. case IR::Opcode::LoadStorageU16:
  575. case IR::Opcode::LoadStorageS16:
  576. case IR::Opcode::WriteStorageU16:
  577. case IR::Opcode::WriteStorageS16:
  578. info.used_storage_buffer_types |= IR::Type::U16;
  579. break;
  580. case IR::Opcode::LoadStorage32:
  581. case IR::Opcode::WriteStorage32:
  582. case IR::Opcode::StorageAtomicIAdd32:
  583. case IR::Opcode::StorageAtomicUMin32:
  584. case IR::Opcode::StorageAtomicUMax32:
  585. case IR::Opcode::StorageAtomicAnd32:
  586. case IR::Opcode::StorageAtomicOr32:
  587. case IR::Opcode::StorageAtomicXor32:
  588. case IR::Opcode::StorageAtomicExchange32:
  589. info.used_storage_buffer_types |= IR::Type::U32;
  590. break;
  591. case IR::Opcode::LoadStorage64:
  592. case IR::Opcode::WriteStorage64:
  593. info.used_storage_buffer_types |= IR::Type::U32x2;
  594. break;
  595. case IR::Opcode::LoadStorage128:
  596. case IR::Opcode::WriteStorage128:
  597. info.used_storage_buffer_types |= IR::Type::U32x4;
  598. break;
  599. case IR::Opcode::SharedAtomicSMin32:
  600. info.uses_atomic_s32_min = true;
  601. break;
  602. case IR::Opcode::SharedAtomicSMax32:
  603. info.uses_atomic_s32_max = true;
  604. break;
  605. case IR::Opcode::SharedAtomicInc32:
  606. info.uses_shared_increment = true;
  607. break;
  608. case IR::Opcode::SharedAtomicDec32:
  609. info.uses_shared_decrement = true;
  610. break;
  611. case IR::Opcode::SharedAtomicExchange64:
  612. info.uses_int64_bit_atomics = true;
  613. break;
  614. case IR::Opcode::GlobalAtomicInc32:
  615. case IR::Opcode::StorageAtomicInc32:
  616. info.used_storage_buffer_types |= IR::Type::U32;
  617. info.uses_global_increment = true;
  618. break;
  619. case IR::Opcode::GlobalAtomicDec32:
  620. case IR::Opcode::StorageAtomicDec32:
  621. info.used_storage_buffer_types |= IR::Type::U32;
  622. info.uses_global_decrement = true;
  623. break;
  624. case IR::Opcode::GlobalAtomicAddF32:
  625. case IR::Opcode::StorageAtomicAddF32:
  626. info.used_storage_buffer_types |= IR::Type::U32;
  627. info.uses_atomic_f32_add = true;
  628. break;
  629. case IR::Opcode::GlobalAtomicAddF16x2:
  630. case IR::Opcode::StorageAtomicAddF16x2:
  631. info.used_storage_buffer_types |= IR::Type::U32;
  632. info.uses_atomic_f16x2_add = true;
  633. break;
  634. case IR::Opcode::GlobalAtomicAddF32x2:
  635. case IR::Opcode::StorageAtomicAddF32x2:
  636. info.used_storage_buffer_types |= IR::Type::U32;
  637. info.uses_atomic_f32x2_add = true;
  638. break;
  639. case IR::Opcode::GlobalAtomicMinF16x2:
  640. case IR::Opcode::StorageAtomicMinF16x2:
  641. info.used_storage_buffer_types |= IR::Type::U32;
  642. info.uses_atomic_f16x2_min = true;
  643. break;
  644. case IR::Opcode::GlobalAtomicMinF32x2:
  645. case IR::Opcode::StorageAtomicMinF32x2:
  646. info.used_storage_buffer_types |= IR::Type::U32;
  647. info.uses_atomic_f32x2_min = true;
  648. break;
  649. case IR::Opcode::GlobalAtomicMaxF16x2:
  650. case IR::Opcode::StorageAtomicMaxF16x2:
  651. info.used_storage_buffer_types |= IR::Type::U32;
  652. info.uses_atomic_f16x2_max = true;
  653. break;
  654. case IR::Opcode::GlobalAtomicMaxF32x2:
  655. case IR::Opcode::StorageAtomicMaxF32x2:
  656. info.used_storage_buffer_types |= IR::Type::U32;
  657. info.uses_atomic_f32x2_max = true;
  658. break;
  659. case IR::Opcode::StorageAtomicSMin32:
  660. info.used_storage_buffer_types |= IR::Type::U32;
  661. info.uses_atomic_s32_min = true;
  662. break;
  663. case IR::Opcode::StorageAtomicSMax32:
  664. info.used_storage_buffer_types |= IR::Type::U32;
  665. info.uses_atomic_s32_max = true;
  666. break;
  667. case IR::Opcode::GlobalAtomicIAdd64:
  668. case IR::Opcode::GlobalAtomicSMin64:
  669. case IR::Opcode::GlobalAtomicUMin64:
  670. case IR::Opcode::GlobalAtomicSMax64:
  671. case IR::Opcode::GlobalAtomicUMax64:
  672. case IR::Opcode::GlobalAtomicAnd64:
  673. case IR::Opcode::GlobalAtomicOr64:
  674. case IR::Opcode::GlobalAtomicXor64:
  675. case IR::Opcode::GlobalAtomicExchange64:
  676. case IR::Opcode::StorageAtomicIAdd64:
  677. case IR::Opcode::StorageAtomicSMin64:
  678. case IR::Opcode::StorageAtomicUMin64:
  679. case IR::Opcode::StorageAtomicSMax64:
  680. case IR::Opcode::StorageAtomicUMax64:
  681. case IR::Opcode::StorageAtomicAnd64:
  682. case IR::Opcode::StorageAtomicOr64:
  683. case IR::Opcode::StorageAtomicXor64:
  684. info.used_storage_buffer_types |= IR::Type::U64;
  685. info.uses_int64_bit_atomics = true;
  686. break;
  687. case IR::Opcode::BindlessImageAtomicIAdd32:
  688. case IR::Opcode::BindlessImageAtomicSMin32:
  689. case IR::Opcode::BindlessImageAtomicUMin32:
  690. case IR::Opcode::BindlessImageAtomicSMax32:
  691. case IR::Opcode::BindlessImageAtomicUMax32:
  692. case IR::Opcode::BindlessImageAtomicInc32:
  693. case IR::Opcode::BindlessImageAtomicDec32:
  694. case IR::Opcode::BindlessImageAtomicAnd32:
  695. case IR::Opcode::BindlessImageAtomicOr32:
  696. case IR::Opcode::BindlessImageAtomicXor32:
  697. case IR::Opcode::BindlessImageAtomicExchange32:
  698. case IR::Opcode::BoundImageAtomicIAdd32:
  699. case IR::Opcode::BoundImageAtomicSMin32:
  700. case IR::Opcode::BoundImageAtomicUMin32:
  701. case IR::Opcode::BoundImageAtomicSMax32:
  702. case IR::Opcode::BoundImageAtomicUMax32:
  703. case IR::Opcode::BoundImageAtomicInc32:
  704. case IR::Opcode::BoundImageAtomicDec32:
  705. case IR::Opcode::BoundImageAtomicAnd32:
  706. case IR::Opcode::BoundImageAtomicOr32:
  707. case IR::Opcode::BoundImageAtomicXor32:
  708. case IR::Opcode::BoundImageAtomicExchange32:
  709. case IR::Opcode::ImageAtomicIAdd32:
  710. case IR::Opcode::ImageAtomicSMin32:
  711. case IR::Opcode::ImageAtomicUMin32:
  712. case IR::Opcode::ImageAtomicSMax32:
  713. case IR::Opcode::ImageAtomicUMax32:
  714. case IR::Opcode::ImageAtomicInc32:
  715. case IR::Opcode::ImageAtomicDec32:
  716. case IR::Opcode::ImageAtomicAnd32:
  717. case IR::Opcode::ImageAtomicOr32:
  718. case IR::Opcode::ImageAtomicXor32:
  719. case IR::Opcode::ImageAtomicExchange32:
  720. info.uses_atomic_image_u32 = true;
  721. break;
  722. default:
  723. break;
  724. }
  725. }
  726. void VisitFpModifiers(Info& info, IR::Inst& inst) {
  727. switch (inst.GetOpcode()) {
  728. case IR::Opcode::FPAdd16:
  729. case IR::Opcode::FPFma16:
  730. case IR::Opcode::FPMul16:
  731. case IR::Opcode::FPRoundEven16:
  732. case IR::Opcode::FPFloor16:
  733. case IR::Opcode::FPCeil16:
  734. case IR::Opcode::FPTrunc16: {
  735. const auto control{inst.Flags<IR::FpControl>()};
  736. switch (control.fmz_mode) {
  737. case IR::FmzMode::DontCare:
  738. break;
  739. case IR::FmzMode::FTZ:
  740. case IR::FmzMode::FMZ:
  741. info.uses_fp16_denorms_flush = true;
  742. break;
  743. case IR::FmzMode::None:
  744. info.uses_fp16_denorms_preserve = true;
  745. break;
  746. }
  747. break;
  748. }
  749. case IR::Opcode::FPAdd32:
  750. case IR::Opcode::FPFma32:
  751. case IR::Opcode::FPMul32:
  752. case IR::Opcode::FPRoundEven32:
  753. case IR::Opcode::FPFloor32:
  754. case IR::Opcode::FPCeil32:
  755. case IR::Opcode::FPTrunc32:
  756. case IR::Opcode::FPOrdEqual32:
  757. case IR::Opcode::FPUnordEqual32:
  758. case IR::Opcode::FPOrdNotEqual32:
  759. case IR::Opcode::FPUnordNotEqual32:
  760. case IR::Opcode::FPOrdLessThan32:
  761. case IR::Opcode::FPUnordLessThan32:
  762. case IR::Opcode::FPOrdGreaterThan32:
  763. case IR::Opcode::FPUnordGreaterThan32:
  764. case IR::Opcode::FPOrdLessThanEqual32:
  765. case IR::Opcode::FPUnordLessThanEqual32:
  766. case IR::Opcode::FPOrdGreaterThanEqual32:
  767. case IR::Opcode::FPUnordGreaterThanEqual32:
  768. case IR::Opcode::ConvertF16F32:
  769. case IR::Opcode::ConvertF64F32: {
  770. const auto control{inst.Flags<IR::FpControl>()};
  771. switch (control.fmz_mode) {
  772. case IR::FmzMode::DontCare:
  773. break;
  774. case IR::FmzMode::FTZ:
  775. case IR::FmzMode::FMZ:
  776. info.uses_fp32_denorms_flush = true;
  777. break;
  778. case IR::FmzMode::None:
  779. info.uses_fp32_denorms_preserve = true;
  780. break;
  781. }
  782. break;
  783. }
  784. default:
  785. break;
  786. }
  787. }
  788. void VisitCbufs(Info& info, IR::Inst& inst) {
  789. switch (inst.GetOpcode()) {
  790. case IR::Opcode::GetCbufU8:
  791. case IR::Opcode::GetCbufS8:
  792. case IR::Opcode::GetCbufU16:
  793. case IR::Opcode::GetCbufS16:
  794. case IR::Opcode::GetCbufU32:
  795. case IR::Opcode::GetCbufF32:
  796. case IR::Opcode::GetCbufU32x2: {
  797. CheckCBufNVN(info, inst);
  798. break;
  799. }
  800. default:
  801. break;
  802. }
  803. }
  804. void Visit(Info& info, IR::Inst& inst) {
  805. VisitUsages(info, inst);
  806. VisitFpModifiers(info, inst);
  807. VisitCbufs(info, inst);
  808. }
  809. void GatherInfoFromHeader(Environment& env, Info& info) {
  810. Stage stage{env.ShaderStage()};
  811. if (stage == Stage::Compute) {
  812. return;
  813. }
  814. const auto& header{env.SPH()};
  815. if (stage == Stage::Fragment) {
  816. if (!info.loads_indexed_attributes) {
  817. return;
  818. }
  819. for (size_t index = 0; index < IR::NUM_GENERICS; ++index) {
  820. const size_t offset{static_cast<size_t>(IR::Attribute::Generic0X) + index * 4};
  821. const auto vector{header.ps.imap_generic_vector[index]};
  822. info.loads.mask[offset + 0] = vector.x != PixelImap::Unused;
  823. info.loads.mask[offset + 1] = vector.y != PixelImap::Unused;
  824. info.loads.mask[offset + 2] = vector.z != PixelImap::Unused;
  825. info.loads.mask[offset + 3] = vector.w != PixelImap::Unused;
  826. }
  827. return;
  828. }
  829. if (info.loads_indexed_attributes) {
  830. for (size_t index = 0; index < IR::NUM_GENERICS; ++index) {
  831. const IR::Attribute attribute{IR::Attribute::Generic0X + index * 4};
  832. const auto mask = header.vtg.InputGeneric(index);
  833. for (size_t i = 0; i < 4; ++i) {
  834. info.loads.Set(attribute + i, mask[i]);
  835. }
  836. }
  837. for (size_t index = 0; index < 8; ++index) {
  838. const u16 mask{header.vtg.clip_distances};
  839. info.loads.Set(IR::Attribute::ClipDistance0 + index, ((mask >> index) & 1) != 0);
  840. }
  841. info.loads.Set(IR::Attribute::PrimitiveId, header.vtg.imap_systemb.primitive_array_id != 0);
  842. info.loads.Set(IR::Attribute::Layer, header.vtg.imap_systemb.rt_array_index != 0);
  843. info.loads.Set(IR::Attribute::ViewportIndex, header.vtg.imap_systemb.viewport_index != 0);
  844. info.loads.Set(IR::Attribute::PointSize, header.vtg.imap_systemb.point_size != 0);
  845. info.loads.Set(IR::Attribute::PositionX, header.vtg.imap_systemb.position_x != 0);
  846. info.loads.Set(IR::Attribute::PositionY, header.vtg.imap_systemb.position_y != 0);
  847. info.loads.Set(IR::Attribute::PositionZ, header.vtg.imap_systemb.position_z != 0);
  848. info.loads.Set(IR::Attribute::PositionW, header.vtg.imap_systemb.position_w != 0);
  849. info.loads.Set(IR::Attribute::PointSpriteS, header.vtg.point_sprite_s != 0);
  850. info.loads.Set(IR::Attribute::PointSpriteT, header.vtg.point_sprite_t != 0);
  851. info.loads.Set(IR::Attribute::FogCoordinate, header.vtg.fog_coordinate != 0);
  852. info.loads.Set(IR::Attribute::TessellationEvaluationPointU,
  853. header.vtg.tessellation_eval_point_u != 0);
  854. info.loads.Set(IR::Attribute::TessellationEvaluationPointV,
  855. header.vtg.tessellation_eval_point_v != 0);
  856. info.loads.Set(IR::Attribute::InstanceId, header.vtg.instance_id != 0);
  857. info.loads.Set(IR::Attribute::VertexId, header.vtg.vertex_id != 0);
  858. // TODO: Legacy varyings
  859. }
  860. if (info.stores_indexed_attributes) {
  861. for (size_t index = 0; index < IR::NUM_GENERICS; ++index) {
  862. const IR::Attribute attribute{IR::Attribute::Generic0X + index * 4};
  863. const auto mask{header.vtg.OutputGeneric(index)};
  864. for (size_t i = 0; i < 4; ++i) {
  865. info.stores.Set(attribute + i, mask[i]);
  866. }
  867. }
  868. for (size_t index = 0; index < 8; ++index) {
  869. const u16 mask{header.vtg.omap_systemc.clip_distances};
  870. info.stores.Set(IR::Attribute::ClipDistance0 + index, ((mask >> index) & 1) != 0);
  871. }
  872. info.stores.Set(IR::Attribute::PrimitiveId,
  873. header.vtg.omap_systemb.primitive_array_id != 0);
  874. info.stores.Set(IR::Attribute::Layer, header.vtg.omap_systemb.rt_array_index != 0);
  875. info.stores.Set(IR::Attribute::ViewportIndex, header.vtg.omap_systemb.viewport_index != 0);
  876. info.stores.Set(IR::Attribute::PointSize, header.vtg.omap_systemb.point_size != 0);
  877. info.stores.Set(IR::Attribute::PositionX, header.vtg.omap_systemb.position_x != 0);
  878. info.stores.Set(IR::Attribute::PositionY, header.vtg.omap_systemb.position_y != 0);
  879. info.stores.Set(IR::Attribute::PositionZ, header.vtg.omap_systemb.position_z != 0);
  880. info.stores.Set(IR::Attribute::PositionW, header.vtg.omap_systemb.position_w != 0);
  881. info.stores.Set(IR::Attribute::PointSpriteS, header.vtg.omap_systemc.point_sprite_s != 0);
  882. info.stores.Set(IR::Attribute::PointSpriteT, header.vtg.omap_systemc.point_sprite_t != 0);
  883. info.stores.Set(IR::Attribute::FogCoordinate, header.vtg.omap_systemc.fog_coordinate != 0);
  884. info.stores.Set(IR::Attribute::TessellationEvaluationPointU,
  885. header.vtg.omap_systemc.tessellation_eval_point_u != 0);
  886. info.stores.Set(IR::Attribute::TessellationEvaluationPointV,
  887. header.vtg.omap_systemc.tessellation_eval_point_v != 0);
  888. info.stores.Set(IR::Attribute::InstanceId, header.vtg.omap_systemc.instance_id != 0);
  889. info.stores.Set(IR::Attribute::VertexId, header.vtg.omap_systemc.vertex_id != 0);
  890. // TODO: Legacy varyings
  891. }
  892. }
  893. } // Anonymous namespace
  894. void CollectShaderInfoPass(Environment& env, IR::Program& program) {
  895. Info& info{program.info};
  896. const u32 base{[&] {
  897. switch (program.stage) {
  898. case Stage::VertexA:
  899. case Stage::VertexB:
  900. return 0x110u;
  901. case Stage::TessellationControl:
  902. return 0x210u;
  903. case Stage::TessellationEval:
  904. return 0x310u;
  905. case Stage::Geometry:
  906. return 0x410u;
  907. case Stage::Fragment:
  908. return 0x510u;
  909. case Stage::Compute:
  910. return 0x310u;
  911. }
  912. throw InvalidArgument("Invalid stage {}", program.stage);
  913. }()};
  914. info.nvn_buffer_base = base;
  915. for (IR::Block* const block : program.post_order_blocks) {
  916. for (IR::Inst& inst : block->Instructions()) {
  917. Visit(info, inst);
  918. }
  919. }
  920. GatherInfoFromHeader(env, info);
  921. }
  922. } // namespace Shader::Optimization