1 //===-- TargetParser - Parser for target features ---------------*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements a target parser to recognise hardware features such as 10 // FPU/CPU/ARCH names as well as specific support such as HDIV, etc. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/TargetParser/TargetParser.h" 15 #include "llvm/ADT/ArrayRef.h" 16 #include "llvm/TargetParser/Triple.h" 17 18 using namespace llvm; 19 using namespace AMDGPU; 20 21 namespace { 22 23 struct GPUInfo { 24 StringLiteral Name; 25 StringLiteral CanonicalName; 26 AMDGPU::GPUKind Kind; 27 unsigned Features; 28 }; 29 30 constexpr GPUInfo R600GPUs[] = { 31 // Name Canonical Kind Features 32 // Name 33 {{"r600"}, {"r600"}, GK_R600, FEATURE_NONE }, 34 {{"rv630"}, {"r600"}, GK_R600, FEATURE_NONE }, 35 {{"rv635"}, {"r600"}, GK_R600, FEATURE_NONE }, 36 {{"r630"}, {"r630"}, GK_R630, FEATURE_NONE }, 37 {{"rs780"}, {"rs880"}, GK_RS880, FEATURE_NONE }, 38 {{"rs880"}, {"rs880"}, GK_RS880, FEATURE_NONE }, 39 {{"rv610"}, {"rs880"}, GK_RS880, FEATURE_NONE }, 40 {{"rv620"}, {"rs880"}, GK_RS880, FEATURE_NONE }, 41 {{"rv670"}, {"rv670"}, GK_RV670, FEATURE_NONE }, 42 {{"rv710"}, {"rv710"}, GK_RV710, FEATURE_NONE }, 43 {{"rv730"}, {"rv730"}, GK_RV730, FEATURE_NONE }, 44 {{"rv740"}, {"rv770"}, GK_RV770, FEATURE_NONE }, 45 {{"rv770"}, {"rv770"}, GK_RV770, FEATURE_NONE }, 46 {{"cedar"}, {"cedar"}, GK_CEDAR, FEATURE_NONE }, 47 {{"palm"}, {"cedar"}, GK_CEDAR, FEATURE_NONE }, 48 {{"cypress"}, {"cypress"}, GK_CYPRESS, FEATURE_FMA }, 49 {{"hemlock"}, {"cypress"}, GK_CYPRESS, FEATURE_FMA }, 50 {{"juniper"}, {"juniper"}, GK_JUNIPER, FEATURE_NONE }, 51 {{"redwood"}, {"redwood"}, GK_REDWOOD, FEATURE_NONE }, 52 {{"sumo"}, {"sumo"}, GK_SUMO, FEATURE_NONE }, 53 {{"sumo2"}, {"sumo"}, GK_SUMO, FEATURE_NONE }, 54 {{"barts"}, {"barts"}, GK_BARTS, FEATURE_NONE }, 55 {{"caicos"}, {"caicos"}, GK_CAICOS, FEATURE_NONE }, 56 {{"aruba"}, {"cayman"}, GK_CAYMAN, FEATURE_FMA }, 57 {{"cayman"}, {"cayman"}, GK_CAYMAN, FEATURE_FMA }, 58 {{"turks"}, {"turks"}, GK_TURKS, FEATURE_NONE } 59 }; 60 61 // This table should be sorted by the value of GPUKind 62 // Don't bother listing the implicitly true features 63 constexpr GPUInfo AMDGCNGPUs[] = { 64 // clang-format off 65 // Name Canonical Kind Features 66 // Name 67 {{"gfx600"}, {"gfx600"}, GK_GFX600, FEATURE_FAST_FMA_F32}, 68 {{"tahiti"}, {"gfx600"}, GK_GFX600, FEATURE_FAST_FMA_F32}, 69 {{"gfx601"}, {"gfx601"}, GK_GFX601, FEATURE_NONE}, 70 {{"pitcairn"}, {"gfx601"}, GK_GFX601, FEATURE_NONE}, 71 {{"verde"}, {"gfx601"}, GK_GFX601, FEATURE_NONE}, 72 {{"gfx602"}, {"gfx602"}, GK_GFX602, FEATURE_NONE}, 73 {{"hainan"}, {"gfx602"}, GK_GFX602, FEATURE_NONE}, 74 {{"oland"}, {"gfx602"}, GK_GFX602, FEATURE_NONE}, 75 {{"gfx700"}, {"gfx700"}, GK_GFX700, FEATURE_NONE}, 76 {{"kaveri"}, {"gfx700"}, GK_GFX700, FEATURE_NONE}, 77 {{"gfx701"}, {"gfx701"}, GK_GFX701, FEATURE_FAST_FMA_F32}, 78 {{"hawaii"}, {"gfx701"}, GK_GFX701, FEATURE_FAST_FMA_F32}, 79 {{"gfx702"}, {"gfx702"}, GK_GFX702, FEATURE_FAST_FMA_F32}, 80 {{"gfx703"}, {"gfx703"}, GK_GFX703, FEATURE_NONE}, 81 {{"kabini"}, {"gfx703"}, GK_GFX703, FEATURE_NONE}, 82 {{"mullins"}, {"gfx703"}, GK_GFX703, FEATURE_NONE}, 83 {{"gfx704"}, {"gfx704"}, GK_GFX704, FEATURE_NONE}, 84 {{"bonaire"}, {"gfx704"}, GK_GFX704, FEATURE_NONE}, 85 {{"gfx705"}, {"gfx705"}, GK_GFX705, FEATURE_NONE}, 86 {{"gfx801"}, {"gfx801"}, GK_GFX801, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK}, 87 {{"carrizo"}, {"gfx801"}, GK_GFX801, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK}, 88 {{"gfx802"}, {"gfx802"}, GK_GFX802, FEATURE_FAST_DENORMAL_F32}, 89 {{"iceland"}, {"gfx802"}, GK_GFX802, FEATURE_FAST_DENORMAL_F32}, 90 {{"tonga"}, {"gfx802"}, GK_GFX802, FEATURE_FAST_DENORMAL_F32}, 91 {{"gfx803"}, {"gfx803"}, GK_GFX803, FEATURE_FAST_DENORMAL_F32}, 92 {{"fiji"}, {"gfx803"}, GK_GFX803, FEATURE_FAST_DENORMAL_F32}, 93 {{"polaris10"}, {"gfx803"}, GK_GFX803, FEATURE_FAST_DENORMAL_F32}, 94 {{"polaris11"}, {"gfx803"}, GK_GFX803, FEATURE_FAST_DENORMAL_F32}, 95 {{"gfx805"}, {"gfx805"}, GK_GFX805, FEATURE_FAST_DENORMAL_F32}, 96 {{"tongapro"}, {"gfx805"}, GK_GFX805, FEATURE_FAST_DENORMAL_F32}, 97 {{"gfx810"}, {"gfx810"}, GK_GFX810, FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK}, 98 {{"stoney"}, {"gfx810"}, GK_GFX810, FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK}, 99 {{"gfx900"}, {"gfx900"}, GK_GFX900, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK}, 100 {{"gfx902"}, {"gfx902"}, GK_GFX902, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK}, 101 {{"gfx904"}, {"gfx904"}, GK_GFX904, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK}, 102 {{"gfx906"}, {"gfx906"}, GK_GFX906, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK|FEATURE_SRAMECC}, 103 {{"gfx908"}, {"gfx908"}, GK_GFX908, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK|FEATURE_SRAMECC}, 104 {{"gfx909"}, {"gfx909"}, GK_GFX909, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK}, 105 {{"gfx90a"}, {"gfx90a"}, GK_GFX90A, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK|FEATURE_SRAMECC}, 106 {{"gfx90c"}, {"gfx90c"}, GK_GFX90C, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK}, 107 {{"gfx940"}, {"gfx940"}, GK_GFX940, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK|FEATURE_SRAMECC}, 108 {{"gfx941"}, {"gfx941"}, GK_GFX941, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK|FEATURE_SRAMECC}, 109 {{"gfx942"}, {"gfx942"}, GK_GFX942, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_XNACK|FEATURE_SRAMECC}, 110 {{"gfx1010"}, {"gfx1010"}, GK_GFX1010, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_XNACK|FEATURE_WGP}, 111 {{"gfx1011"}, {"gfx1011"}, GK_GFX1011, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_XNACK|FEATURE_WGP}, 112 {{"gfx1012"}, {"gfx1012"}, GK_GFX1012, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_XNACK|FEATURE_WGP}, 113 {{"gfx1013"}, {"gfx1013"}, GK_GFX1013, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_XNACK|FEATURE_WGP}, 114 {{"gfx1030"}, {"gfx1030"}, GK_GFX1030, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 115 {{"gfx1031"}, {"gfx1031"}, GK_GFX1031, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 116 {{"gfx1032"}, {"gfx1032"}, GK_GFX1032, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 117 {{"gfx1033"}, {"gfx1033"}, GK_GFX1033, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 118 {{"gfx1034"}, {"gfx1034"}, GK_GFX1034, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 119 {{"gfx1035"}, {"gfx1035"}, GK_GFX1035, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 120 {{"gfx1036"}, {"gfx1036"}, GK_GFX1036, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 121 {{"gfx1100"}, {"gfx1100"}, GK_GFX1100, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 122 {{"gfx1101"}, {"gfx1101"}, GK_GFX1101, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 123 {{"gfx1102"}, {"gfx1102"}, GK_GFX1102, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 124 {{"gfx1103"}, {"gfx1103"}, GK_GFX1103, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 125 {{"gfx1150"}, {"gfx1150"}, GK_GFX1150, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 126 {{"gfx1151"}, {"gfx1151"}, GK_GFX1151, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 127 {{"gfx1200"}, {"gfx1200"}, GK_GFX1200, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 128 {{"gfx1201"}, {"gfx1201"}, GK_GFX1201, FEATURE_FAST_FMA_F32|FEATURE_FAST_DENORMAL_F32|FEATURE_WAVE32|FEATURE_WGP}, 129 // clang-format on 130 }; 131 132 const GPUInfo *getArchEntry(AMDGPU::GPUKind AK, ArrayRef<GPUInfo> Table) { 133 GPUInfo Search = { {""}, {""}, AK, AMDGPU::FEATURE_NONE }; 134 135 auto I = 136 llvm::lower_bound(Table, Search, [](const GPUInfo &A, const GPUInfo &B) { 137 return A.Kind < B.Kind; 138 }); 139 140 if (I == Table.end() || I->Kind != Search.Kind) 141 return nullptr; 142 return I; 143 } 144 145 } // namespace 146 147 StringRef llvm::AMDGPU::getArchNameAMDGCN(GPUKind AK) { 148 if (const auto *Entry = getArchEntry(AK, AMDGCNGPUs)) 149 return Entry->CanonicalName; 150 return ""; 151 } 152 153 StringRef llvm::AMDGPU::getArchNameR600(GPUKind AK) { 154 if (const auto *Entry = getArchEntry(AK, R600GPUs)) 155 return Entry->CanonicalName; 156 return ""; 157 } 158 159 AMDGPU::GPUKind llvm::AMDGPU::parseArchAMDGCN(StringRef CPU) { 160 for (const auto &C : AMDGCNGPUs) { 161 if (CPU == C.Name) 162 return C.Kind; 163 } 164 165 return AMDGPU::GPUKind::GK_NONE; 166 } 167 168 AMDGPU::GPUKind llvm::AMDGPU::parseArchR600(StringRef CPU) { 169 for (const auto &C : R600GPUs) { 170 if (CPU == C.Name) 171 return C.Kind; 172 } 173 174 return AMDGPU::GPUKind::GK_NONE; 175 } 176 177 unsigned AMDGPU::getArchAttrAMDGCN(GPUKind AK) { 178 if (const auto *Entry = getArchEntry(AK, AMDGCNGPUs)) 179 return Entry->Features; 180 return FEATURE_NONE; 181 } 182 183 unsigned AMDGPU::getArchAttrR600(GPUKind AK) { 184 if (const auto *Entry = getArchEntry(AK, R600GPUs)) 185 return Entry->Features; 186 return FEATURE_NONE; 187 } 188 189 void AMDGPU::fillValidArchListAMDGCN(SmallVectorImpl<StringRef> &Values) { 190 // XXX: Should this only report unique canonical names? 191 for (const auto &C : AMDGCNGPUs) 192 Values.push_back(C.Name); 193 } 194 195 void AMDGPU::fillValidArchListR600(SmallVectorImpl<StringRef> &Values) { 196 for (const auto &C : R600GPUs) 197 Values.push_back(C.Name); 198 } 199 200 AMDGPU::IsaVersion AMDGPU::getIsaVersion(StringRef GPU) { 201 AMDGPU::GPUKind AK = parseArchAMDGCN(GPU); 202 if (AK == AMDGPU::GPUKind::GK_NONE) { 203 if (GPU == "generic-hsa") 204 return {7, 0, 0}; 205 if (GPU == "generic") 206 return {6, 0, 0}; 207 return {0, 0, 0}; 208 } 209 210 // clang-format off 211 switch (AK) { 212 case GK_GFX600: return {6, 0, 0}; 213 case GK_GFX601: return {6, 0, 1}; 214 case GK_GFX602: return {6, 0, 2}; 215 case GK_GFX700: return {7, 0, 0}; 216 case GK_GFX701: return {7, 0, 1}; 217 case GK_GFX702: return {7, 0, 2}; 218 case GK_GFX703: return {7, 0, 3}; 219 case GK_GFX704: return {7, 0, 4}; 220 case GK_GFX705: return {7, 0, 5}; 221 case GK_GFX801: return {8, 0, 1}; 222 case GK_GFX802: return {8, 0, 2}; 223 case GK_GFX803: return {8, 0, 3}; 224 case GK_GFX805: return {8, 0, 5}; 225 case GK_GFX810: return {8, 1, 0}; 226 case GK_GFX900: return {9, 0, 0}; 227 case GK_GFX902: return {9, 0, 2}; 228 case GK_GFX904: return {9, 0, 4}; 229 case GK_GFX906: return {9, 0, 6}; 230 case GK_GFX908: return {9, 0, 8}; 231 case GK_GFX909: return {9, 0, 9}; 232 case GK_GFX90A: return {9, 0, 10}; 233 case GK_GFX90C: return {9, 0, 12}; 234 case GK_GFX940: return {9, 4, 0}; 235 case GK_GFX941: return {9, 4, 1}; 236 case GK_GFX942: return {9, 4, 2}; 237 case GK_GFX1010: return {10, 1, 0}; 238 case GK_GFX1011: return {10, 1, 1}; 239 case GK_GFX1012: return {10, 1, 2}; 240 case GK_GFX1013: return {10, 1, 3}; 241 case GK_GFX1030: return {10, 3, 0}; 242 case GK_GFX1031: return {10, 3, 1}; 243 case GK_GFX1032: return {10, 3, 2}; 244 case GK_GFX1033: return {10, 3, 3}; 245 case GK_GFX1034: return {10, 3, 4}; 246 case GK_GFX1035: return {10, 3, 5}; 247 case GK_GFX1036: return {10, 3, 6}; 248 case GK_GFX1100: return {11, 0, 0}; 249 case GK_GFX1101: return {11, 0, 1}; 250 case GK_GFX1102: return {11, 0, 2}; 251 case GK_GFX1103: return {11, 0, 3}; 252 case GK_GFX1150: return {11, 5, 0}; 253 case GK_GFX1151: return {11, 5, 1}; 254 case GK_GFX1200: return {12, 0, 0}; 255 case GK_GFX1201: return {12, 0, 1}; 256 default: return {0, 0, 0}; 257 } 258 // clang-format on 259 } 260 261 StringRef AMDGPU::getCanonicalArchName(const Triple &T, StringRef Arch) { 262 assert(T.isAMDGPU()); 263 auto ProcKind = T.isAMDGCN() ? parseArchAMDGCN(Arch) : parseArchR600(Arch); 264 if (ProcKind == GK_NONE) 265 return StringRef(); 266 267 return T.isAMDGCN() ? getArchNameAMDGCN(ProcKind) : getArchNameR600(ProcKind); 268 } 269 270 void AMDGPU::fillAMDGPUFeatureMap(StringRef GPU, const Triple &T, 271 StringMap<bool> &Features) { 272 // XXX - What does the member GPU mean if device name string passed here? 273 if (T.isAMDGCN()) { 274 switch (parseArchAMDGCN(GPU)) { 275 case GK_GFX1201: 276 case GK_GFX1200: 277 Features["ci-insts"] = true; 278 Features["dot7-insts"] = true; 279 Features["dot8-insts"] = true; 280 Features["dot9-insts"] = true; 281 Features["dot10-insts"] = true; 282 Features["dl-insts"] = true; 283 Features["atomic-ds-pk-add-16-insts"] = true; 284 Features["atomic-flat-pk-add-16-insts"] = true; 285 Features["atomic-buffer-global-pk-add-f16-insts"] = true; 286 Features["atomic-global-pk-add-bf16-inst"] = true; 287 Features["16-bit-insts"] = true; 288 Features["dpp"] = true; 289 Features["gfx8-insts"] = true; 290 Features["gfx9-insts"] = true; 291 Features["gfx10-insts"] = true; 292 Features["gfx10-3-insts"] = true; 293 Features["gfx11-insts"] = true; 294 Features["gfx12-insts"] = true; 295 Features["atomic-fadd-rtn-insts"] = true; 296 Features["image-insts"] = true; 297 Features["fp8-conversion-insts"] = true; 298 break; 299 case GK_GFX1151: 300 case GK_GFX1150: 301 case GK_GFX1103: 302 case GK_GFX1102: 303 case GK_GFX1101: 304 case GK_GFX1100: 305 Features["ci-insts"] = true; 306 Features["dot5-insts"] = true; 307 Features["dot7-insts"] = true; 308 Features["dot8-insts"] = true; 309 Features["dot9-insts"] = true; 310 Features["dot10-insts"] = true; 311 Features["dl-insts"] = true; 312 Features["16-bit-insts"] = true; 313 Features["dpp"] = true; 314 Features["gfx8-insts"] = true; 315 Features["gfx9-insts"] = true; 316 Features["gfx10-insts"] = true; 317 Features["gfx10-3-insts"] = true; 318 Features["gfx11-insts"] = true; 319 Features["atomic-fadd-rtn-insts"] = true; 320 Features["image-insts"] = true; 321 Features["gws"] = true; 322 break; 323 case GK_GFX1036: 324 case GK_GFX1035: 325 case GK_GFX1034: 326 case GK_GFX1033: 327 case GK_GFX1032: 328 case GK_GFX1031: 329 case GK_GFX1030: 330 Features["ci-insts"] = true; 331 Features["dot1-insts"] = true; 332 Features["dot2-insts"] = true; 333 Features["dot5-insts"] = true; 334 Features["dot6-insts"] = true; 335 Features["dot7-insts"] = true; 336 Features["dot10-insts"] = true; 337 Features["dl-insts"] = true; 338 Features["16-bit-insts"] = true; 339 Features["dpp"] = true; 340 Features["gfx8-insts"] = true; 341 Features["gfx9-insts"] = true; 342 Features["gfx10-insts"] = true; 343 Features["gfx10-3-insts"] = true; 344 Features["image-insts"] = true; 345 Features["s-memrealtime"] = true; 346 Features["s-memtime-inst"] = true; 347 Features["gws"] = true; 348 break; 349 case GK_GFX1012: 350 case GK_GFX1011: 351 Features["dot1-insts"] = true; 352 Features["dot2-insts"] = true; 353 Features["dot5-insts"] = true; 354 Features["dot6-insts"] = true; 355 Features["dot7-insts"] = true; 356 Features["dot10-insts"] = true; 357 [[fallthrough]]; 358 case GK_GFX1013: 359 case GK_GFX1010: 360 Features["dl-insts"] = true; 361 Features["ci-insts"] = true; 362 Features["16-bit-insts"] = true; 363 Features["dpp"] = true; 364 Features["gfx8-insts"] = true; 365 Features["gfx9-insts"] = true; 366 Features["gfx10-insts"] = true; 367 Features["image-insts"] = true; 368 Features["s-memrealtime"] = true; 369 Features["s-memtime-inst"] = true; 370 Features["gws"] = true; 371 break; 372 case GK_GFX942: 373 case GK_GFX941: 374 case GK_GFX940: 375 Features["gfx940-insts"] = true; 376 Features["fp8-insts"] = true; 377 Features["fp8-conversion-insts"] = true; 378 Features["atomic-ds-pk-add-16-insts"] = true; 379 Features["atomic-flat-pk-add-16-insts"] = true; 380 Features["atomic-global-pk-add-bf16-inst"] = true; 381 Features["gfx90a-insts"] = true; 382 Features["atomic-buffer-global-pk-add-f16-insts"] = true; 383 Features["atomic-fadd-rtn-insts"] = true; 384 Features["dot3-insts"] = true; 385 Features["dot4-insts"] = true; 386 Features["dot5-insts"] = true; 387 Features["dot6-insts"] = true; 388 Features["mai-insts"] = true; 389 Features["dl-insts"] = true; 390 Features["dot1-insts"] = true; 391 Features["dot2-insts"] = true; 392 Features["dot7-insts"] = true; 393 Features["dot10-insts"] = true; 394 Features["gfx9-insts"] = true; 395 Features["gfx8-insts"] = true; 396 Features["16-bit-insts"] = true; 397 Features["dpp"] = true; 398 Features["s-memrealtime"] = true; 399 Features["ci-insts"] = true; 400 Features["s-memtime-inst"] = true; 401 Features["gws"] = true; 402 break; 403 case GK_GFX90A: 404 Features["gfx90a-insts"] = true; 405 Features["atomic-buffer-global-pk-add-f16-insts"] = true; 406 Features["atomic-fadd-rtn-insts"] = true; 407 [[fallthrough]]; 408 case GK_GFX908: 409 Features["dot3-insts"] = true; 410 Features["dot4-insts"] = true; 411 Features["dot5-insts"] = true; 412 Features["dot6-insts"] = true; 413 Features["mai-insts"] = true; 414 [[fallthrough]]; 415 case GK_GFX906: 416 Features["dl-insts"] = true; 417 Features["dot1-insts"] = true; 418 Features["dot2-insts"] = true; 419 Features["dot7-insts"] = true; 420 Features["dot10-insts"] = true; 421 [[fallthrough]]; 422 case GK_GFX90C: 423 case GK_GFX909: 424 case GK_GFX904: 425 case GK_GFX902: 426 case GK_GFX900: 427 Features["gfx9-insts"] = true; 428 [[fallthrough]]; 429 case GK_GFX810: 430 case GK_GFX805: 431 case GK_GFX803: 432 case GK_GFX802: 433 case GK_GFX801: 434 Features["gfx8-insts"] = true; 435 Features["16-bit-insts"] = true; 436 Features["dpp"] = true; 437 Features["s-memrealtime"] = true; 438 [[fallthrough]]; 439 case GK_GFX705: 440 case GK_GFX704: 441 case GK_GFX703: 442 case GK_GFX702: 443 case GK_GFX701: 444 case GK_GFX700: 445 Features["ci-insts"] = true; 446 [[fallthrough]]; 447 case GK_GFX602: 448 case GK_GFX601: 449 case GK_GFX600: 450 Features["image-insts"] = true; 451 Features["s-memtime-inst"] = true; 452 Features["gws"] = true; 453 break; 454 case GK_NONE: 455 break; 456 default: 457 llvm_unreachable("Unhandled GPU!"); 458 } 459 } else { 460 if (GPU.empty()) 461 GPU = "r600"; 462 463 switch (llvm::AMDGPU::parseArchR600(GPU)) { 464 case GK_CAYMAN: 465 case GK_CYPRESS: 466 case GK_RV770: 467 case GK_RV670: 468 // TODO: Add fp64 when implemented. 469 break; 470 case GK_TURKS: 471 case GK_CAICOS: 472 case GK_BARTS: 473 case GK_SUMO: 474 case GK_REDWOOD: 475 case GK_JUNIPER: 476 case GK_CEDAR: 477 case GK_RV730: 478 case GK_RV710: 479 case GK_RS880: 480 case GK_R630: 481 case GK_R600: 482 break; 483 default: 484 llvm_unreachable("Unhandled GPU!"); 485 } 486 } 487 } 488 489 static bool isWave32Capable(StringRef GPU, const Triple &T) { 490 bool IsWave32Capable = false; 491 // XXX - What does the member GPU mean if device name string passed here? 492 if (T.isAMDGCN()) { 493 switch (parseArchAMDGCN(GPU)) { 494 case GK_GFX1201: 495 case GK_GFX1200: 496 case GK_GFX1151: 497 case GK_GFX1150: 498 case GK_GFX1103: 499 case GK_GFX1102: 500 case GK_GFX1101: 501 case GK_GFX1100: 502 case GK_GFX1036: 503 case GK_GFX1035: 504 case GK_GFX1034: 505 case GK_GFX1033: 506 case GK_GFX1032: 507 case GK_GFX1031: 508 case GK_GFX1030: 509 case GK_GFX1012: 510 case GK_GFX1011: 511 case GK_GFX1013: 512 case GK_GFX1010: 513 IsWave32Capable = true; 514 break; 515 default: 516 break; 517 } 518 } 519 return IsWave32Capable; 520 } 521 522 bool AMDGPU::insertWaveSizeFeature(StringRef GPU, const Triple &T, 523 StringMap<bool> &Features, 524 std::string &ErrorMsg) { 525 bool IsWave32Capable = isWave32Capable(GPU, T); 526 const bool IsNullGPU = GPU.empty(); 527 // FIXME: Not diagnosing wavefrontsize32 on wave64 only targets. 528 const bool HaveWave32 = 529 (IsWave32Capable || IsNullGPU) && Features.count("wavefrontsize32"); 530 const bool HaveWave64 = Features.count("wavefrontsize64"); 531 if (HaveWave32 && HaveWave64) { 532 ErrorMsg = "'wavefrontsize32' and 'wavefrontsize64' are mutually exclusive"; 533 return false; 534 } 535 // Don't assume any wavesize with an unknown subtarget. 536 if (!IsNullGPU) { 537 // Default to wave32 if available, or wave64 if not 538 if (!HaveWave32 && !HaveWave64) { 539 StringRef DefaultWaveSizeFeature = 540 IsWave32Capable ? "wavefrontsize32" : "wavefrontsize64"; 541 Features.insert(std::make_pair(DefaultWaveSizeFeature, true)); 542 } 543 } 544 return true; 545 } 546