| 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::printTcgen05MMAKind(raw_ostream &OS, const Constant *ImmArgVal) { |
| 20 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 21 | uint64_t Val = CI->getZExtValue(); |
| 22 | switch (static_cast<Tcgen05MMAKind>(Val)) { |
| 23 | case Tcgen05MMAKind::F16: |
| 24 | OS << "f16" ; |
| 25 | return; |
| 26 | case Tcgen05MMAKind::TF32: |
| 27 | OS << "tf32" ; |
| 28 | return; |
| 29 | case Tcgen05MMAKind::F8F6F4: |
| 30 | OS << "f8f6f4" ; |
| 31 | return; |
| 32 | case Tcgen05MMAKind::I8: |
| 33 | OS << "i8" ; |
| 34 | return; |
| 35 | } |
| 36 | } |
| 37 | llvm_unreachable( |
| 38 | "printTcgen05MMAKind called with invalid value for immediate argument" ); |
| 39 | } |
| 40 | |
| 41 | void nvvm::printTcgen05CollectorUsageOp(raw_ostream &OS, |
| 42 | const Constant *ImmArgVal) { |
| 43 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 44 | uint64_t Val = CI->getZExtValue(); |
| 45 | switch (static_cast<Tcgen05CollectorUsageOp>(Val)) { |
| 46 | case Tcgen05CollectorUsageOp::DISCARD: |
| 47 | OS << "discard" ; |
| 48 | return; |
| 49 | case Tcgen05CollectorUsageOp::LASTUSE: |
| 50 | OS << "lastuse" ; |
| 51 | return; |
| 52 | case Tcgen05CollectorUsageOp::FILL: |
| 53 | OS << "fill" ; |
| 54 | return; |
| 55 | case Tcgen05CollectorUsageOp::USE: |
| 56 | OS << "use" ; |
| 57 | return; |
| 58 | } |
| 59 | } |
| 60 | llvm_unreachable("printTcgen05CollectorUsageOp called with invalid value for " |
| 61 | "immediate argument" ); |
| 62 | } |
| 63 | |
| 64 | void nvvm::printTensormapElemType(raw_ostream &OS, const Constant *ImmArgVal) { |
| 65 | static constexpr StringRef TensormapElemTypes[] = { |
| 66 | "u8" , "u16" , "u32" , "s32" , "u64" , "s64" , |
| 67 | "f16" , "f32" , "f32.ftz" , "f64" , "bf16" , "tf32" , |
| 68 | "tf32.ftz" , "b4x16" , "b4x16_p64" , "b6x16_p32" }; |
| 69 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 70 | uint64_t Val = CI->getZExtValue(); |
| 71 | if (Val <= static_cast<uint64_t>(nvvm::TensormapElemType::B6x16_p32)) { |
| 72 | OS << TensormapElemTypes[Val]; |
| 73 | return; |
| 74 | } |
| 75 | } |
| 76 | } |
| 77 | |
| 78 | void nvvm::printTensormapInterleaveLayout(raw_ostream &OS, |
| 79 | const Constant *ImmArgVal) { |
| 80 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 81 | uint64_t Val = CI->getZExtValue(); |
| 82 | switch (static_cast<TensormapInterleaveLayout>(Val)) { |
| 83 | case TensormapInterleaveLayout::NO_INTERLEAVE: |
| 84 | OS << "No interleave" ; |
| 85 | return; |
| 86 | case TensormapInterleaveLayout::INTERLEAVE_16B: |
| 87 | OS << "16B interleave" ; |
| 88 | return; |
| 89 | case TensormapInterleaveLayout::INTERLEAVE_32B: |
| 90 | OS << "32B interleave" ; |
| 91 | return; |
| 92 | } |
| 93 | } |
| 94 | } |
| 95 | |
| 96 | void nvvm::printTensormapSwizzleMode(raw_ostream &OS, |
| 97 | const Constant *ImmArgVal) { |
| 98 | static constexpr StringRef TensormapSwizzleModes[] = { |
| 99 | "No swizzling" , "32B swizzling" , "64B swizzling" , "128B swizzling" , |
| 100 | "96B swizzling" }; |
| 101 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 102 | uint64_t Val = CI->getZExtValue(); |
| 103 | if (Val <= static_cast<uint64_t>(nvvm::TensormapSwizzleMode::SWIZZLE_96B)) { |
| 104 | OS << TensormapSwizzleModes[Val]; |
| 105 | return; |
| 106 | } |
| 107 | } |
| 108 | } |
| 109 | |
| 110 | void nvvm::printTensormapSwizzleAtomicity(raw_ostream &OS, |
| 111 | const Constant *ImmArgVal) { |
| 112 | static constexpr StringRef TensormapSwizzleAtomicities[] = { |
| 113 | "16B" , "32B" , "32B + 8B flip" , "64B" }; |
| 114 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 115 | uint64_t Val = CI->getZExtValue(); |
| 116 | if (Val <= static_cast<uint64_t>( |
| 117 | nvvm::TensormapSwizzleAtomicity::SWIZZLE_ATOMICITY_64B)) { |
| 118 | OS << TensormapSwizzleAtomicities[Val]; |
| 119 | return; |
| 120 | } |
| 121 | } |
| 122 | } |
| 123 | |
| 124 | void nvvm::printTensormapFillMode(raw_ostream &OS, const Constant *ImmArgVal) { |
| 125 | if (const auto *CI = dyn_cast<ConstantInt>(Val: ImmArgVal)) { |
| 126 | uint64_t Val = CI->getZExtValue(); |
| 127 | OS << (Val == static_cast<uint64_t>(TensormapFillMode::ZERO_FILL) |
| 128 | ? "Zero fill" |
| 129 | : "OOB-NaN fill" ); |
| 130 | return; |
| 131 | } |
| 132 | } |
| 133 | |