| 1 | //=====-- NVPTXSubtarget.h - Define Subtarget for the NVPTX ---*- 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 declares the NVPTX specific subclass of TargetSubtarget. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #ifndef LLVM_LIB_TARGET_NVPTX_NVPTXSUBTARGET_H |
| 14 | #define LLVM_LIB_TARGET_NVPTX_NVPTXSUBTARGET_H |
| 15 | |
| 16 | #include "NVPTX.h" |
| 17 | #include "NVPTXFrameLowering.h" |
| 18 | #include "NVPTXISelLowering.h" |
| 19 | #include "NVPTXInstrInfo.h" |
| 20 | #include "NVPTXRegisterInfo.h" |
| 21 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
| 22 | #include "llvm/IR/DataLayout.h" |
| 23 | #include "llvm/IR/NVVMIntrinsicUtils.h" |
| 24 | #include "llvm/Support/NVPTXAddrSpace.h" |
| 25 | #include <string> |
| 26 | |
| 27 | #define |
| 28 | #include "NVPTXGenSubtargetInfo.inc" |
| 29 | |
| 30 | namespace llvm { |
| 31 | |
| 32 | class NVPTXSubtarget : public NVPTXGenSubtargetInfo { |
| 33 | virtual void anchor(); |
| 34 | std::string TargetName; |
| 35 | |
| 36 | // PTX version x.y is represented as 10*x+y, e.g. 3.1 == 31 |
| 37 | unsigned PTXVersion; |
| 38 | |
| 39 | // Full SM version x.y is represented as 100*x+10*y+feature, e.g. 3.1 == 310 |
| 40 | // sm_90a == 901 |
| 41 | unsigned int FullSmVersion; |
| 42 | |
| 43 | // SM version x.y is represented as 10*x+y, e.g. 3.1 == 31. Derived from |
| 44 | // FullSmVersion. |
| 45 | unsigned int SmVersion; |
| 46 | |
| 47 | NVPTXInstrInfo InstrInfo; |
| 48 | NVPTXTargetLowering TLInfo; |
| 49 | std::unique_ptr<const SelectionDAGTargetInfo> TSInfo; |
| 50 | |
| 51 | // NVPTX does not have any call stack frame, but need a NVPTX specific |
| 52 | // FrameLowering class because TargetFrameLowering is abstract. |
| 53 | NVPTXFrameLowering FrameLowering; |
| 54 | |
| 55 | public: |
| 56 | /// This constructor initializes the data members to match that |
| 57 | /// of the specified module. |
| 58 | /// |
| 59 | NVPTXSubtarget(const Triple &TT, const std::string &CPU, |
| 60 | const std::string &FS, const NVPTXTargetMachine &TM); |
| 61 | |
| 62 | ~NVPTXSubtarget() override; |
| 63 | |
| 64 | const TargetFrameLowering *getFrameLowering() const override { |
| 65 | return &FrameLowering; |
| 66 | } |
| 67 | const NVPTXInstrInfo *getInstrInfo() const override { return &InstrInfo; } |
| 68 | const NVPTXRegisterInfo *getRegisterInfo() const override { |
| 69 | return &InstrInfo.getRegisterInfo(); |
| 70 | } |
| 71 | const NVPTXTargetLowering *getTargetLowering() const override { |
| 72 | return &TLInfo; |
| 73 | } |
| 74 | |
| 75 | const SelectionDAGTargetInfo *getSelectionDAGInfo() const override; |
| 76 | |
| 77 | // Checks PTX version and family-specific and architecture-specific SM |
| 78 | // versions. For example, sm_100{f/a} and any future variants in the same |
| 79 | // family will match for any PTX version greater than or equal to |
| 80 | // `PTXVersion`. |
| 81 | bool hasPTXWithFamilySMs(unsigned PTXVersion, |
| 82 | ArrayRef<unsigned> SMVersions) const; |
| 83 | // Checks PTX version and architecture-specific SM versions. |
| 84 | // For example, sm_100{a} will match for any PTX version greater than or equal |
| 85 | // to `PTXVersion`. |
| 86 | bool hasPTXWithAccelSMs(unsigned PTXVersion, |
| 87 | ArrayRef<unsigned> SMVersions) const; |
| 88 | |
| 89 | bool has256BitVectorLoadStore(unsigned AS) const { |
| 90 | return SmVersion >= 100 && PTXVersion >= 88 && |
| 91 | AS == NVPTXAS::ADDRESS_SPACE_GLOBAL; |
| 92 | } |
| 93 | bool hasUsedBytesMaskPragma() const { |
| 94 | return SmVersion >= 50 && PTXVersion >= 83; |
| 95 | } |
| 96 | bool hasAtomAddF64() const { return SmVersion >= 60; } |
| 97 | bool hasAtomScope() const { return SmVersion >= 60; } |
| 98 | bool hasAtomBitwise64() const { return SmVersion >= 32; } |
| 99 | bool hasAtomMinMax64() const { return SmVersion >= 32; } |
| 100 | bool hasAtomCas16() const { return SmVersion >= 70 && PTXVersion >= 63; } |
| 101 | bool hasAtomSwap128() const { return SmVersion >= 90 && PTXVersion >= 83; } |
| 102 | bool hasClusters() const { return SmVersion >= 90 && PTXVersion >= 78; } |
| 103 | bool hasLDG() const { return SmVersion >= 32; } |
| 104 | bool hasHWROT32() const { return SmVersion >= 32; } |
| 105 | bool hasBrx() const { return SmVersion >= 30 && PTXVersion >= 60; } |
| 106 | bool hasFP16Math() const { return SmVersion >= 53; } |
| 107 | bool hasBF16Math() const { return SmVersion >= 80; } |
| 108 | bool allowFP16Math() const; |
| 109 | bool hasMaskOperator() const { return PTXVersion >= 71; } |
| 110 | bool hasNoReturn() const { return SmVersion >= 30 && PTXVersion >= 64; } |
| 111 | // Does SM & PTX support memory orderings (weak and atomic: relaxed, acquire, |
| 112 | // release, acq_rel, sc) ? |
| 113 | bool hasMemoryOrdering() const { return SmVersion >= 70 && PTXVersion >= 60; } |
| 114 | // Does SM & PTX support .acquire and .release qualifiers for fence? |
| 115 | bool hasSplitAcquireAndReleaseFences() const { |
| 116 | return SmVersion >= 90 && PTXVersion >= 86; |
| 117 | } |
| 118 | // Does SM & PTX support atomic relaxed MMIO operations ? |
| 119 | bool hasRelaxedMMIO() const { return SmVersion >= 70 && PTXVersion >= 82; } |
| 120 | bool hasDotInstructions() const { |
| 121 | return SmVersion >= 61 && PTXVersion >= 50; |
| 122 | } |
| 123 | |
| 124 | // Checks following instructions support: |
| 125 | // - tcgen05.ld/st |
| 126 | // - tcgen05.alloc/dealloc/relinquish |
| 127 | // - tcgen05.cp |
| 128 | // - tcgen05.fence/wait |
| 129 | // - tcgen05.commit |
| 130 | // - tcgen05.mma |
| 131 | bool hasTcgen05InstSupport() const { |
| 132 | // sm_101 renamed to sm_110 in PTX 9.0 |
| 133 | return hasPTXWithFamilySMs(PTXVersion: 90, SMVersions: {100, 110}) || |
| 134 | hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {100, 101}) || |
| 135 | hasPTXWithAccelSMs(PTXVersion: 86, SMVersions: {100, 101}); |
| 136 | } |
| 137 | |
| 138 | // Checks tcgen05.shift instruction support. |
| 139 | bool hasTcgen05ShiftSupport() const { |
| 140 | // sm_101 renamed to sm_110 in PTX 9.0 |
| 141 | return hasPTXWithAccelSMs(PTXVersion: 90, SMVersions: {100, 110, 103}) || |
| 142 | hasPTXWithAccelSMs(PTXVersion: 88, SMVersions: {100, 101, 103}) || |
| 143 | hasPTXWithAccelSMs(PTXVersion: 86, SMVersions: {100, 101}); |
| 144 | } |
| 145 | |
| 146 | bool hasTcgen05MMAScaleInputDImm() const { |
| 147 | return hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {100}) || hasPTXWithAccelSMs(PTXVersion: 86, SMVersions: {100}); |
| 148 | } |
| 149 | |
| 150 | bool hasTcgen05MMAI8Kind() const { |
| 151 | return hasPTXWithAccelSMs(PTXVersion: 90, SMVersions: {100, 110}) || |
| 152 | hasPTXWithAccelSMs(PTXVersion: 86, SMVersions: {100, 101}); |
| 153 | } |
| 154 | |
| 155 | bool hasTcgen05MMASparseMxf4nvf4() const { |
| 156 | return hasPTXWithAccelSMs(PTXVersion: 90, SMVersions: {100, 110, 103}) || |
| 157 | hasPTXWithAccelSMs(PTXVersion: 87, SMVersions: {100, 101, 103}); |
| 158 | } |
| 159 | |
| 160 | bool hasTcgen05MMASparseMxf4() const { |
| 161 | return hasPTXWithAccelSMs(PTXVersion: 90, SMVersions: {100, 110, 103}) || |
| 162 | hasPTXWithAccelSMs(PTXVersion: 86, SMVersions: {100, 101, 103}); |
| 163 | } |
| 164 | |
| 165 | bool hasTcgen05LdRedSupport() const { |
| 166 | return hasPTXWithFamilySMs(PTXVersion: 90, SMVersions: {110, 103}) || |
| 167 | hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {101, 103}); |
| 168 | } |
| 169 | |
| 170 | bool hasReduxSyncF32() const { |
| 171 | return hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {100}) || hasPTXWithAccelSMs(PTXVersion: 86, SMVersions: {100}); |
| 172 | } |
| 173 | |
| 174 | bool hasMMABlockScale() const { |
| 175 | return hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {120}) || hasPTXWithAccelSMs(PTXVersion: 87, SMVersions: {120}); |
| 176 | } |
| 177 | |
| 178 | bool hasMMASparseBlockScaleF4() const { |
| 179 | return hasPTXWithAccelSMs(PTXVersion: 87, SMVersions: {120, 121}); |
| 180 | } |
| 181 | |
| 182 | // f32x2 instructions in Blackwell family |
| 183 | bool hasF32x2Instructions() const; |
| 184 | |
| 185 | // Checks support for following in TMA: |
| 186 | // - cta_group::1/2 support |
| 187 | // - im2col_w/w_128 mode support |
| 188 | // - tile_gather4 mode support |
| 189 | // - tile_scatter4 mode support |
| 190 | bool hasTMABlackwellSupport() const { |
| 191 | return hasPTXWithFamilySMs(PTXVersion: 90, SMVersions: {100, 110}) || |
| 192 | hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {100, 101}) || |
| 193 | hasPTXWithAccelSMs(PTXVersion: 86, SMVersions: {100, 101}); |
| 194 | } |
| 195 | |
| 196 | // Checks support for conversions involving e4m3x2 and e5m2x2. |
| 197 | bool hasFP8ConversionSupport() const { |
| 198 | if (PTXVersion >= 81) |
| 199 | return SmVersion >= 89; |
| 200 | |
| 201 | if (PTXVersion >= 78) |
| 202 | return SmVersion >= 90; |
| 203 | |
| 204 | return false; |
| 205 | } |
| 206 | |
| 207 | // Checks support for conversions involving the following types: |
| 208 | // - e2m3x2/e3m2x2 |
| 209 | // - e2m1x2 |
| 210 | // - ue8m0x2 |
| 211 | bool hasNarrowFPConversionSupport() const { |
| 212 | return hasPTXWithFamilySMs(PTXVersion: 90, SMVersions: {100, 110, 120}) || |
| 213 | hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {100, 101, 120}) || |
| 214 | hasPTXWithAccelSMs(PTXVersion: 86, SMVersions: {100, 101, 120}); |
| 215 | } |
| 216 | |
| 217 | bool hasTensormapReplaceSupport() const { |
| 218 | return hasPTXWithFamilySMs(PTXVersion: 90, SMVersions: {90, 100, 110, 120}) || |
| 219 | hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {90, 100, 101, 120}) || |
| 220 | hasPTXWithAccelSMs(PTXVersion: 83, SMVersions: {90, 100, 101, 120}); |
| 221 | } |
| 222 | |
| 223 | bool hasTensormapReplaceElemtypeSupport(unsigned value) const { |
| 224 | if (value >= static_cast<unsigned>(nvvm::TensormapElemType::B4x16)) |
| 225 | return hasPTXWithFamilySMs(PTXVersion: 90, SMVersions: {100, 110, 120}) || |
| 226 | hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {100, 101, 120}) || |
| 227 | hasPTXWithAccelSMs(PTXVersion: 87, SMVersions: {100, 101, 120}); |
| 228 | |
| 229 | return hasTensormapReplaceSupport(); |
| 230 | } |
| 231 | |
| 232 | bool hasTensormapReplaceSwizzleAtomicitySupport() const { |
| 233 | return hasPTXWithFamilySMs(PTXVersion: 90, SMVersions: {100, 110, 120}) || |
| 234 | hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {100, 101, 120}) || |
| 235 | hasPTXWithAccelSMs(PTXVersion: 87, SMVersions: {100, 101, 120}); |
| 236 | } |
| 237 | |
| 238 | bool hasTensormapReplaceSwizzleModeSupport(unsigned value) const { |
| 239 | if (value == static_cast<unsigned>(nvvm::TensormapSwizzleMode::SWIZZLE_96B)) |
| 240 | return hasPTXWithAccelSMs(PTXVersion: 88, SMVersions: {103}); |
| 241 | |
| 242 | return hasTensormapReplaceSupport(); |
| 243 | } |
| 244 | |
| 245 | bool hasClusterLaunchControlTryCancelMulticastSupport() const { |
| 246 | return hasPTXWithFamilySMs(PTXVersion: 90, SMVersions: {100, 110, 120}) || |
| 247 | hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {100, 101, 120}) || |
| 248 | hasPTXWithAccelSMs(PTXVersion: 86, SMVersions: {100, 101, 120}); |
| 249 | } |
| 250 | |
| 251 | bool hasSetMaxNRegSupport() const { |
| 252 | return hasPTXWithFamilySMs(PTXVersion: 90, SMVersions: {100, 110, 120}) || |
| 253 | hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {100, 101, 120}) || |
| 254 | hasPTXWithAccelSMs(PTXVersion: 86, SMVersions: {100, 101, 120}) || |
| 255 | hasPTXWithAccelSMs(PTXVersion: 80, SMVersions: {90}); |
| 256 | } |
| 257 | |
| 258 | bool hasLdStmatrixBlackwellSupport() const { |
| 259 | return hasPTXWithFamilySMs(PTXVersion: 90, SMVersions: {100, 110, 120}) || |
| 260 | hasPTXWithFamilySMs(PTXVersion: 88, SMVersions: {100, 101, 120}) || |
| 261 | hasPTXWithAccelSMs(PTXVersion: 86, SMVersions: {100, 101, 120}); |
| 262 | } |
| 263 | |
| 264 | // Prior to CUDA 12.3 ptxas did not recognize that the trap instruction |
| 265 | // terminates a basic block. Instead, it would assume that control flow |
| 266 | // continued to the next instruction. The next instruction could be in the |
| 267 | // block that's lexically below it. This would lead to a phantom CFG edges |
| 268 | // being created within ptxas. This issue was fixed in CUDA 12.3. Thus, when |
| 269 | // PTX ISA versions 8.3+ we can confidently say that the bug will not be |
| 270 | // present. |
| 271 | bool hasPTXASUnreachableBug() const { return PTXVersion < 83; } |
| 272 | bool hasCvtaParam() const { return SmVersion >= 70 && PTXVersion >= 77; } |
| 273 | bool hasConvertWithStochasticRounding() const { |
| 274 | return hasPTXWithAccelSMs(PTXVersion: 87, SMVersions: {100, 103}); |
| 275 | } |
| 276 | unsigned int getFullSmVersion() const { return FullSmVersion; } |
| 277 | unsigned int getSmVersion() const { return getFullSmVersion() / 10; } |
| 278 | unsigned int getSmFamilyVersion() const { return getFullSmVersion() / 100; } |
| 279 | // GPUs with "a" suffix have architecture-accelerated features that are |
| 280 | // supported on the specified architecture only, hence such targets do not |
| 281 | // follow the onion layer model. hasArchAccelFeatures() allows distinguishing |
| 282 | // such GPU variants from the base GPU architecture. |
| 283 | // - false represents non-accelerated architecture. |
| 284 | // - true represents architecture-accelerated variant. |
| 285 | bool hasArchAccelFeatures() const { |
| 286 | return (getFullSmVersion() & 1) && PTXVersion >= 80; |
| 287 | } |
| 288 | // GPUs with 'f' suffix have architecture-accelerated features which are |
| 289 | // portable across all future architectures under same SM major. For example, |
| 290 | // sm_100f features will work for sm_10X*f*/sm_10X*a* future architectures. |
| 291 | // - false represents non-family-specific architecture. |
| 292 | // - true represents family-specific variant. |
| 293 | bool hasFamilySpecificFeatures() const { |
| 294 | return getFullSmVersion() % 10 == 2 ? PTXVersion >= 88 |
| 295 | : hasArchAccelFeatures(); |
| 296 | } |
| 297 | // If the user did not provide a target we default to the `sm_75` target. |
| 298 | std::string getTargetName() const { |
| 299 | return TargetName.empty() ? "sm_75" : TargetName; |
| 300 | } |
| 301 | bool hasTargetName() const { return !TargetName.empty(); } |
| 302 | |
| 303 | bool hasNativeBF16Support(int Opcode) const; |
| 304 | |
| 305 | // Get maximum value of required alignments among the supported data types. |
| 306 | // From the PTX ISA doc, section 8.2.3: |
| 307 | // The memory consistency model relates operations executed on memory |
| 308 | // locations with scalar data-types, which have a maximum size and alignment |
| 309 | // of 64 bits. Memory operations with a vector data-type are modelled as a |
| 310 | // set of equivalent memory operations with a scalar data-type, executed in |
| 311 | // an unspecified order on the elements in the vector. |
| 312 | unsigned getMaxRequiredAlignment() const { return 8; } |
| 313 | // Get the smallest cmpxchg word size that the hardware supports. |
| 314 | unsigned getMinCmpXchgSizeInBits() const { return 32; } |
| 315 | |
| 316 | unsigned getPTXVersion() const { return PTXVersion; } |
| 317 | |
| 318 | NVPTXSubtarget &initializeSubtargetDependencies(StringRef CPU, StringRef FS); |
| 319 | void ParseSubtargetFeatures(StringRef CPU, StringRef TuneCPU, StringRef FS); |
| 320 | |
| 321 | void failIfClustersUnsupported(std::string const &FailureMessage) const; |
| 322 | }; |
| 323 | |
| 324 | } // End llvm namespace |
| 325 | |
| 326 | #endif |
| 327 | |