| 1 | //===----------------------------------------------------------------------===// |
| 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 functions associated with NVVM Intrinsics. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "llvm/IR/NVVMIntrinsicUtils.h" |
| 14 | #include "llvm/ADT/StringRef.h" |
| 15 | |
| 16 | using namespace llvm; |
| 17 | using namespace nvvm; |
| 18 | |
| 19 | void nvvm::printEvictPolicyType(raw_ostream &OS, const Constant *ImmArgVal) { |
| 20 | const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal); |
| 21 | if (!CI || |
| 22 | CI->getZExtValue() > static_cast<uint64_t>(EvictPolicyType::EVICT_LAST)) { |
| 23 | OS << "Unsupported evict policy" ; |
| 24 | return; |
| 25 | } |
| 26 | OS << getEvictPolicyName(Policy: static_cast<EvictPolicyType>(CI->getZExtValue())); |
| 27 | } |
| 28 | |
| 29 | void nvvm::printTMAReductionOp(raw_ostream &OS, const Constant *ImmArgVal) { |
| 30 | const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal); |
| 31 | if (!CI || CI->getZExtValue() > static_cast<uint64_t>(TMAReductionOp::XOR)) |
| 32 | llvm_unreachable( |
| 33 | "printTMAReductionOp called with invalid value for immediate argument" ); |
| 34 | |
| 35 | OS << getTMATensorReductionOpName( |
| 36 | Op: static_cast<TMAReductionOp>(CI->getZExtValue())); |
| 37 | } |
| 38 | |
| 39 | void nvvm::printTcgen05MMAKind(raw_ostream &OS, const Constant *ImmArgVal) { |
| 40 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 41 | uint64_t Val = CI->getZExtValue(); |
| 42 | switch (static_cast<Tcgen05MMAKind>(Val)) { |
| 43 | case Tcgen05MMAKind::F16: |
| 44 | OS << "f16" ; |
| 45 | return; |
| 46 | case Tcgen05MMAKind::TF32: |
| 47 | OS << "tf32" ; |
| 48 | return; |
| 49 | case Tcgen05MMAKind::F8F6F4: |
| 50 | OS << "f8f6f4" ; |
| 51 | return; |
| 52 | case Tcgen05MMAKind::I8: |
| 53 | OS << "i8" ; |
| 54 | return; |
| 55 | case Tcgen05MMAKind::TI16: |
| 56 | OS << "ti16" ; |
| 57 | return; |
| 58 | } |
| 59 | } |
| 60 | } |
| 61 | |
| 62 | void nvvm::printTcgen05CollectorUsageOp(raw_ostream &OS, |
| 63 | const Constant *ImmArgVal) { |
| 64 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 65 | uint64_t Val = CI->getZExtValue(); |
| 66 | switch (static_cast<Tcgen05CollectorUsageOp>(Val)) { |
| 67 | case Tcgen05CollectorUsageOp::DISCARD: |
| 68 | OS << "discard" ; |
| 69 | return; |
| 70 | case Tcgen05CollectorUsageOp::LASTUSE: |
| 71 | OS << "lastuse" ; |
| 72 | return; |
| 73 | case Tcgen05CollectorUsageOp::FILL: |
| 74 | OS << "fill" ; |
| 75 | return; |
| 76 | case Tcgen05CollectorUsageOp::USE: |
| 77 | OS << "use" ; |
| 78 | return; |
| 79 | } |
| 80 | } |
| 81 | } |
| 82 | |
| 83 | void nvvm::printTcgen05MMACollectorBBuffer(raw_ostream &OS, |
| 84 | const Constant *ImmArgVal) { |
| 85 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 86 | uint64_t Val = CI->getZExtValue(); |
| 87 | switch (static_cast<Tcgen05MMACollectorBBuffer>(Val)) { |
| 88 | case Tcgen05MMACollectorBBuffer::B0: |
| 89 | OS << "b0" ; |
| 90 | return; |
| 91 | case Tcgen05MMACollectorBBuffer::B1: |
| 92 | OS << "b1" ; |
| 93 | return; |
| 94 | case Tcgen05MMACollectorBBuffer::B2: |
| 95 | OS << "b2" ; |
| 96 | return; |
| 97 | case Tcgen05MMACollectorBBuffer::B3: |
| 98 | OS << "b3" ; |
| 99 | return; |
| 100 | } |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | void nvvm::printTensormapElemType(raw_ostream &OS, const Constant *ImmArgVal) { |
| 105 | static constexpr StringRef TensormapElemTypes[] = { |
| 106 | "u8" , "u16" , "u32" , "s32" , "u64" , "s64" , |
| 107 | "f16" , "f32" , "f32.ftz" , "f64" , "bf16" , "tf32" , |
| 108 | "tf32.ftz" , "b4x16" , "b4x16_p64" , "b6x16_p32" }; |
| 109 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 110 | uint64_t Val = CI->getZExtValue(); |
| 111 | if (Val <= static_cast<uint64_t>(nvvm::TensormapElemType::B6x16_p32)) { |
| 112 | OS << TensormapElemTypes[Val]; |
| 113 | return; |
| 114 | } |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | void nvvm::printTensormapInterleaveLayout(raw_ostream &OS, |
| 119 | const Constant *ImmArgVal) { |
| 120 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 121 | uint64_t Val = CI->getZExtValue(); |
| 122 | switch (static_cast<TensormapInterleaveLayout>(Val)) { |
| 123 | case TensormapInterleaveLayout::NO_INTERLEAVE: |
| 124 | OS << "No interleave" ; |
| 125 | return; |
| 126 | case TensormapInterleaveLayout::INTERLEAVE_16B: |
| 127 | OS << "16B interleave" ; |
| 128 | return; |
| 129 | case TensormapInterleaveLayout::INTERLEAVE_32B: |
| 130 | OS << "32B interleave" ; |
| 131 | return; |
| 132 | } |
| 133 | } |
| 134 | } |
| 135 | |
| 136 | void nvvm::printTensormapSwizzleMode(raw_ostream &OS, |
| 137 | const Constant *ImmArgVal) { |
| 138 | static constexpr StringRef TensormapSwizzleModes[] = { |
| 139 | "No swizzling" , "32B swizzling" , "64B swizzling" , "128B swizzling" , |
| 140 | "96B swizzling" }; |
| 141 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 142 | uint64_t Val = CI->getZExtValue(); |
| 143 | if (Val <= static_cast<uint64_t>(nvvm::TensormapSwizzleMode::SWIZZLE_96B)) { |
| 144 | OS << TensormapSwizzleModes[Val]; |
| 145 | return; |
| 146 | } |
| 147 | } |
| 148 | } |
| 149 | |
| 150 | void nvvm::printTensormapSwizzleAtomicity(raw_ostream &OS, |
| 151 | const Constant *ImmArgVal) { |
| 152 | static constexpr StringRef TensormapSwizzleAtomicities[] = { |
| 153 | "16B" , "32B" , "32B + 8B flip" , "64B" }; |
| 154 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 155 | uint64_t Val = CI->getZExtValue(); |
| 156 | if (Val <= static_cast<uint64_t>( |
| 157 | nvvm::TensormapSwizzleAtomicity::SWIZZLE_ATOMICITY_64B)) { |
| 158 | OS << TensormapSwizzleAtomicities[Val]; |
| 159 | return; |
| 160 | } |
| 161 | } |
| 162 | } |
| 163 | |
| 164 | void nvvm::printTensormapFillMode(raw_ostream &OS, const Constant *ImmArgVal) { |
| 165 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 166 | uint64_t Val = CI->getZExtValue(); |
| 167 | OS << (Val == static_cast<uint64_t>(TensormapFillMode::ZERO_FILL) |
| 168 | ? "Zero fill" |
| 169 | : "OOB-NaN fill" ); |
| 170 | return; |
| 171 | } |
| 172 | } |
| 173 | |