| 1 | //===---- CIRGenBuiltinNVPTX.cpp - Emit CIR for NVPTX builtins ------------===// |
| 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 contains code to emit NVPTX Builtin calls. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "CIRGenFunction.h" |
| 14 | |
| 15 | #include "mlir/IR/Value.h" |
| 16 | #include "clang/Basic/TargetBuiltins.h" |
| 17 | #include "clang/CIR/Dialect/IR/CIRDialect.h" |
| 18 | #include "llvm/IR/Intrinsics.h" |
| 19 | #include "llvm/IR/IntrinsicsNVPTX.h" |
| 20 | #include "llvm/Support/NVPTXAddrSpace.h" |
| 21 | |
| 22 | #include <algorithm> |
| 23 | |
| 24 | using namespace clang; |
| 25 | using namespace clang::CIRGen; |
| 26 | |
| 27 | static mlir::Value makeLdu(CIRGenFunction &cgf, const CallExpr *expr, |
| 28 | llvm::StringRef intrinsicName) { |
| 29 | auto &builder = cgf.getBuilder(); |
| 30 | Address ptr = cgf.emitPointerWithAlignment(expr: expr->getArg(Arg: 0)); |
| 31 | QualType argType = expr->getArg(Arg: 0)->getType(); |
| 32 | mlir::Type elemTy = cgf.convertTypeForMem(t: argType->getPointeeType()); |
| 33 | clang::CharUnits align = ptr.getAlignment(); |
| 34 | mlir::Location loc = cgf.getLoc(srcLoc: expr->getExprLoc()); |
| 35 | mlir::Value alignVal = |
| 36 | builder.getConstantInt(loc, ty: builder.getSInt32Ty(), value: align.getQuantity()); |
| 37 | return cir::LLVMIntrinsicCallOp::create( |
| 38 | builder, location: loc, intrinsic_name: builder.getStringAttr(bytes: intrinsicName), resType: elemTy, |
| 39 | operands: {ptr.emitRawPointer(), alignVal}) |
| 40 | .getResult(); |
| 41 | } |
| 42 | |
| 43 | static mlir::Value makeLdg(CIRGenFunction &cgf, const CallExpr *expr) { |
| 44 | auto &builder = cgf.getBuilder(); |
| 45 | Address ptr = cgf.emitPointerWithAlignment(expr: expr->getArg(Arg: 0)); |
| 46 | QualType argType = expr->getArg(Arg: 0)->getType(); |
| 47 | mlir::Type elemTy = cgf.convertTypeForMem(t: argType->getPointeeType()); |
| 48 | mlir::Location loc = cgf.getLoc(srcLoc: expr->getExprLoc()); |
| 49 | |
| 50 | // Use addrspace(1) for NVPTX ADDRESS_SPACE_GLOBAL. |
| 51 | mlir::Type globalPtrTy = cir::PointerType::get( |
| 52 | pointee: elemTy, addrSpace: cir::TargetAddressSpaceAttr::get( |
| 53 | context: builder.getContext(), value: llvm::NVPTXAS::ADDRESS_SPACE_GLOBAL)); |
| 54 | mlir::Value asc = |
| 55 | builder.createAddrSpaceCast(loc, src: ptr.getPointer(), newTy: globalPtrTy); |
| 56 | cir::LoadOp load = |
| 57 | builder.createAlignedLoad(loc, ty: elemTy, ptr: asc, align: ptr.getAlignment()); |
| 58 | load.setInvariant(true); |
| 59 | return load.getResult(); |
| 60 | } |
| 61 | |
| 62 | /// Emit a CIR LLVMIntrinsicCallOp for a unary NVVM intrinsic. |
| 63 | /// The result type is inferred from the single argument. |
| 64 | static mlir::Value emitUnaryNVVMIntrinsic(CIRGenFunction &cgf, |
| 65 | const CallExpr *expr, |
| 66 | llvm::StringRef intrinsicName) { |
| 67 | auto &builder = cgf.getBuilder(); |
| 68 | mlir::Value arg = cgf.emitScalarExpr(e: expr->getArg(Arg: 0)); |
| 69 | return cir::LLVMIntrinsicCallOp::create( |
| 70 | builder, location: cgf.getLoc(srcLoc: expr->getExprLoc()), |
| 71 | intrinsic_name: builder.getStringAttr(bytes: intrinsicName), resType: arg.getType(), operands: {arg}) |
| 72 | .getResult(); |
| 73 | } |
| 74 | |
| 75 | /// Emit a CIR LLVMIntrinsicCallOp for an NVVM fadd/fmul intrinsic, which takes |
| 76 | /// the rounding mode as a trailing operand. |
| 77 | static mlir::Value emitNVVMFPArith(CIRGenFunction &cgf, const CallExpr *expr, |
| 78 | llvm::StringRef intrinsicName, |
| 79 | llvm::APFloat::roundingMode rm) { |
| 80 | auto &builder = cgf.getBuilder(); |
| 81 | mlir::Location loc = cgf.getLoc(srcLoc: expr->getExprLoc()); |
| 82 | mlir::Value lhs = cgf.emitScalarExpr(e: expr->getArg(Arg: 0)); |
| 83 | mlir::Value rhs = cgf.emitScalarExpr(e: expr->getArg(Arg: 1)); |
| 84 | mlir::Value rnd = |
| 85 | builder.getConstInt(loc, t: builder.getSInt32Ty(), c: static_cast<int>(rm)); |
| 86 | return cir::LLVMIntrinsicCallOp::create(builder, location: loc, |
| 87 | intrinsic_name: builder.getStringAttr(bytes: intrinsicName), |
| 88 | resType: lhs.getType(), operands: {lhs, rhs, rnd}) |
| 89 | .getResult(); |
| 90 | } |
| 91 | |
| 92 | /// Emit a CIR LLVMIntrinsicCallOp for an NVVM fma intrinsic, which takes the |
| 93 | /// rounding mode as a trailing operand. |
| 94 | static mlir::Value emitNVVMFMA(CIRGenFunction &cgf, const CallExpr *expr, |
| 95 | llvm::StringRef intrinsicName, |
| 96 | llvm::APFloat::roundingMode rm) { |
| 97 | auto &builder = cgf.getBuilder(); |
| 98 | mlir::Location loc = cgf.getLoc(srcLoc: expr->getExprLoc()); |
| 99 | mlir::Value a = cgf.emitScalarExpr(e: expr->getArg(Arg: 0)); |
| 100 | mlir::Value b = cgf.emitScalarExpr(e: expr->getArg(Arg: 1)); |
| 101 | mlir::Value c = cgf.emitScalarExpr(e: expr->getArg(Arg: 2)); |
| 102 | mlir::Value rnd = |
| 103 | builder.getConstInt(loc, t: builder.getSInt32Ty(), c: static_cast<int>(rm)); |
| 104 | return cir::LLVMIntrinsicCallOp::create(builder, location: loc, |
| 105 | intrinsic_name: builder.getStringAttr(bytes: intrinsicName), |
| 106 | resType: a.getType(), operands: {a, b, c, rnd}) |
| 107 | .getResult(); |
| 108 | } |
| 109 | |
| 110 | static mlir::Value emitBar0Reduction(CIRGenFunction &cgf, const CallExpr *expr, |
| 111 | llvm::StringRef intrinsicName, |
| 112 | bool returnsPred) { |
| 113 | CIRGenBuilderTy &builder = cgf.getBuilder(); |
| 114 | mlir::Location loc = cgf.getLoc(srcLoc: expr->getExprLoc()); |
| 115 | mlir::Type si32Ty = builder.getSInt32Ty(); |
| 116 | mlir::Value zero = builder.getNullValue(ty: si32Ty, loc); |
| 117 | mlir::Value pred = builder.createCompare( |
| 118 | loc, kind: cir::CmpOpKind::ne, lhs: cgf.emitScalarExpr(e: expr->getArg(Arg: 0)), rhs: zero); |
| 119 | mlir::Type resultTy = returnsPred ? mlir::Type(builder.getBoolTy()) : si32Ty; |
| 120 | mlir::Value result = builder.emitIntrinsicCallOp( |
| 121 | loc, str: intrinsicName, resTy: resultTy, op: mlir::ValueRange{zero, pred}); |
| 122 | if (returnsPred) |
| 123 | result = builder.createBoolToInt(src: result, newTy: si32Ty); |
| 124 | return result; |
| 125 | } |
| 126 | |
| 127 | static mlir::Value makeScopedAtomicRMW(CIRGenFunction &cgf, |
| 128 | const CallExpr *expr, |
| 129 | cir::AtomicFetchKind kind, |
| 130 | cir::SyncScopeKind scope) { |
| 131 | auto &builder = cgf.getBuilder(); |
| 132 | Address destAddr = cgf.emitPointerWithAlignment(expr: expr->getArg(Arg: 0)); |
| 133 | mlir::Value destValue = destAddr.emitRawPointer(); |
| 134 | mlir::Value val = cgf.emitScalarExpr(e: expr->getArg(Arg: 1)); |
| 135 | auto rmwi = cir::AtomicFetchOp::create( |
| 136 | builder, location: cgf.getLoc(srcLoc: expr->getSourceRange()), ptr: destValue, val, binop: kind, |
| 137 | mem_order: cir::MemOrder::Relaxed, sync_scope: scope, /*is_volatile=*/false, |
| 138 | /*fetch_first=*/true); |
| 139 | return rmwi->getResult(idx: 0); |
| 140 | } |
| 141 | |
| 142 | static mlir::Value makeScopedAtomicXchg(CIRGenFunction &cgf, |
| 143 | const CallExpr *expr, |
| 144 | cir::SyncScopeKind scope) { |
| 145 | auto &builder = cgf.getBuilder(); |
| 146 | Address destAddr = cgf.emitPointerWithAlignment(expr: expr->getArg(Arg: 0)); |
| 147 | mlir::Value destValue = destAddr.emitRawPointer(); |
| 148 | mlir::Value val = cgf.emitScalarExpr(e: expr->getArg(Arg: 1)); |
| 149 | auto xchg = cir::AtomicXchgOp::create( |
| 150 | builder, location: cgf.getLoc(srcLoc: expr->getSourceRange()), ptr: destValue, val, |
| 151 | mem_order: cir::MemOrder::Relaxed, sync_scope: scope, /*is_volatile=*/false); |
| 152 | return xchg.getResult(); |
| 153 | } |
| 154 | |
| 155 | // Mirrors the WMMA load mappings in classic CodeGen. |
| 156 | struct NVPTXMmaLdstInfo { |
| 157 | unsigned NumResults; // Number of elements to load/store |
| 158 | // Intrinsic IDs for row/col variants. 0 if particular layout is unsupported. |
| 159 | unsigned IID_col; |
| 160 | unsigned IID_row; |
| 161 | }; |
| 162 | |
| 163 | #define MMA_INTR(geom_op_type, layout) \ |
| 164 | llvm::Intrinsic::nvvm_wmma_##geom_op_type##_##layout##_stride |
| 165 | #define MMA_LDST(n, geom_op_type) \ |
| 166 | {n, MMA_INTR(geom_op_type, col), MMA_INTR(geom_op_type, row)} |
| 167 | |
| 168 | static NVPTXMmaLdstInfo getNVPTXMmaLdstInfo(unsigned BuiltinID) { |
| 169 | switch (BuiltinID) { |
| 170 | // FP MMA loads |
| 171 | case NVPTX::BI__hmma_m16n16k16_ld_a: |
| 172 | return MMA_LDST(8, m16n16k16_load_a_f16); |
| 173 | case NVPTX::BI__hmma_m16n16k16_ld_b: |
| 174 | return MMA_LDST(8, m16n16k16_load_b_f16); |
| 175 | case NVPTX::BI__hmma_m16n16k16_ld_c_f16: |
| 176 | return MMA_LDST(4, m16n16k16_load_c_f16); |
| 177 | case NVPTX::BI__hmma_m16n16k16_ld_c_f32: |
| 178 | return MMA_LDST(8, m16n16k16_load_c_f32); |
| 179 | case NVPTX::BI__hmma_m32n8k16_ld_a: |
| 180 | return MMA_LDST(8, m32n8k16_load_a_f16); |
| 181 | case NVPTX::BI__hmma_m32n8k16_ld_b: |
| 182 | return MMA_LDST(8, m32n8k16_load_b_f16); |
| 183 | case NVPTX::BI__hmma_m32n8k16_ld_c_f16: |
| 184 | return MMA_LDST(4, m32n8k16_load_c_f16); |
| 185 | case NVPTX::BI__hmma_m32n8k16_ld_c_f32: |
| 186 | return MMA_LDST(8, m32n8k16_load_c_f32); |
| 187 | case NVPTX::BI__hmma_m8n32k16_ld_a: |
| 188 | return MMA_LDST(8, m8n32k16_load_a_f16); |
| 189 | case NVPTX::BI__hmma_m8n32k16_ld_b: |
| 190 | return MMA_LDST(8, m8n32k16_load_b_f16); |
| 191 | case NVPTX::BI__hmma_m8n32k16_ld_c_f16: |
| 192 | return MMA_LDST(4, m8n32k16_load_c_f16); |
| 193 | case NVPTX::BI__hmma_m8n32k16_ld_c_f32: |
| 194 | return MMA_LDST(8, m8n32k16_load_c_f32); |
| 195 | |
| 196 | // Integer MMA loads |
| 197 | case NVPTX::BI__imma_m16n16k16_ld_a_s8: |
| 198 | return MMA_LDST(2, m16n16k16_load_a_s8); |
| 199 | case NVPTX::BI__imma_m16n16k16_ld_a_u8: |
| 200 | return MMA_LDST(2, m16n16k16_load_a_u8); |
| 201 | case NVPTX::BI__imma_m16n16k16_ld_b_s8: |
| 202 | return MMA_LDST(2, m16n16k16_load_b_s8); |
| 203 | case NVPTX::BI__imma_m16n16k16_ld_b_u8: |
| 204 | return MMA_LDST(2, m16n16k16_load_b_u8); |
| 205 | case NVPTX::BI__imma_m16n16k16_ld_c: |
| 206 | return MMA_LDST(8, m16n16k16_load_c_s32); |
| 207 | case NVPTX::BI__imma_m32n8k16_ld_a_s8: |
| 208 | return MMA_LDST(4, m32n8k16_load_a_s8); |
| 209 | case NVPTX::BI__imma_m32n8k16_ld_a_u8: |
| 210 | return MMA_LDST(4, m32n8k16_load_a_u8); |
| 211 | case NVPTX::BI__imma_m32n8k16_ld_b_s8: |
| 212 | return MMA_LDST(1, m32n8k16_load_b_s8); |
| 213 | case NVPTX::BI__imma_m32n8k16_ld_b_u8: |
| 214 | return MMA_LDST(1, m32n8k16_load_b_u8); |
| 215 | case NVPTX::BI__imma_m32n8k16_ld_c: |
| 216 | return MMA_LDST(8, m32n8k16_load_c_s32); |
| 217 | case NVPTX::BI__imma_m8n32k16_ld_a_s8: |
| 218 | return MMA_LDST(1, m8n32k16_load_a_s8); |
| 219 | case NVPTX::BI__imma_m8n32k16_ld_a_u8: |
| 220 | return MMA_LDST(1, m8n32k16_load_a_u8); |
| 221 | case NVPTX::BI__imma_m8n32k16_ld_b_s8: |
| 222 | return MMA_LDST(4, m8n32k16_load_b_s8); |
| 223 | case NVPTX::BI__imma_m8n32k16_ld_b_u8: |
| 224 | return MMA_LDST(4, m8n32k16_load_b_u8); |
| 225 | case NVPTX::BI__imma_m8n32k16_ld_c: |
| 226 | return MMA_LDST(8, m8n32k16_load_c_s32); |
| 227 | |
| 228 | // Sub-integer MMA loads. |
| 229 | // Only row/col layout is supported by A/B fragments. |
| 230 | case NVPTX::BI__imma_m8n8k32_ld_a_s4: |
| 231 | return {.NumResults: 1, .IID_col: 0, MMA_INTR(m8n8k32_load_a_s4, row)}; |
| 232 | case NVPTX::BI__imma_m8n8k32_ld_a_u4: |
| 233 | return {.NumResults: 1, .IID_col: 0, MMA_INTR(m8n8k32_load_a_u4, row)}; |
| 234 | case NVPTX::BI__imma_m8n8k32_ld_b_s4: |
| 235 | return {.NumResults: 1, MMA_INTR(m8n8k32_load_b_s4, col), .IID_row: 0}; |
| 236 | case NVPTX::BI__imma_m8n8k32_ld_b_u4: |
| 237 | return {.NumResults: 1, MMA_INTR(m8n8k32_load_b_u4, col), .IID_row: 0}; |
| 238 | case NVPTX::BI__imma_m8n8k32_ld_c: |
| 239 | return MMA_LDST(2, m8n8k32_load_c_s32); |
| 240 | case NVPTX::BI__bmma_m8n8k128_ld_a_b1: |
| 241 | return {.NumResults: 1, .IID_col: 0, MMA_INTR(m8n8k128_load_a_b1, row)}; |
| 242 | case NVPTX::BI__bmma_m8n8k128_ld_b_b1: |
| 243 | return {.NumResults: 1, MMA_INTR(m8n8k128_load_b_b1, col), .IID_row: 0}; |
| 244 | case NVPTX::BI__bmma_m8n8k128_ld_c: |
| 245 | return MMA_LDST(2, m8n8k128_load_c_s32); |
| 246 | |
| 247 | // Double MMA loads |
| 248 | case NVPTX::BI__dmma_m8n8k4_ld_a: |
| 249 | return MMA_LDST(1, m8n8k4_load_a_f64); |
| 250 | case NVPTX::BI__dmma_m8n8k4_ld_b: |
| 251 | return MMA_LDST(1, m8n8k4_load_b_f64); |
| 252 | case NVPTX::BI__dmma_m8n8k4_ld_c: |
| 253 | return MMA_LDST(2, m8n8k4_load_c_f64); |
| 254 | |
| 255 | // Alternate float MMA loads |
| 256 | case NVPTX::BI__mma_bf16_m16n16k16_ld_a: |
| 257 | return MMA_LDST(4, m16n16k16_load_a_bf16); |
| 258 | case NVPTX::BI__mma_bf16_m16n16k16_ld_b: |
| 259 | return MMA_LDST(4, m16n16k16_load_b_bf16); |
| 260 | case NVPTX::BI__mma_bf16_m8n32k16_ld_a: |
| 261 | return MMA_LDST(2, m8n32k16_load_a_bf16); |
| 262 | case NVPTX::BI__mma_bf16_m8n32k16_ld_b: |
| 263 | return MMA_LDST(8, m8n32k16_load_b_bf16); |
| 264 | case NVPTX::BI__mma_bf16_m32n8k16_ld_a: |
| 265 | return MMA_LDST(8, m32n8k16_load_a_bf16); |
| 266 | case NVPTX::BI__mma_bf16_m32n8k16_ld_b: |
| 267 | return MMA_LDST(2, m32n8k16_load_b_bf16); |
| 268 | case NVPTX::BI__mma_tf32_m16n16k8_ld_a: |
| 269 | return MMA_LDST(4, m16n16k8_load_a_tf32); |
| 270 | case NVPTX::BI__mma_tf32_m16n16k8_ld_b: |
| 271 | return MMA_LDST(4, m16n16k8_load_b_tf32); |
| 272 | case NVPTX::BI__mma_tf32_m16n16k8_ld_c: |
| 273 | return MMA_LDST(8, m16n16k8_load_c_f32); |
| 274 | |
| 275 | default: |
| 276 | llvm_unreachable("Unknown WMMA load builtin" ); |
| 277 | } |
| 278 | } |
| 279 | #undef MMA_LDST |
| 280 | #undef MMA_INTR |
| 281 | |
| 282 | static void storeWMMAFragment(CIRGenBuilderTy &builder, mlir::Location loc, |
| 283 | mlir::Value result, Address dst, |
| 284 | unsigned numRegisters) { |
| 285 | CharUnits alignment = |
| 286 | std::min(a: dst.getAlignment(), b: CharUnits::fromQuantity(Quantity: 4)); |
| 287 | if (numRegisters == 1) { |
| 288 | builder.createStore(loc, val: result, dst: dst.withAlignment(newAlignment: alignment)); |
| 289 | return; |
| 290 | } |
| 291 | |
| 292 | // FP16 intrinsic registers are <2 x half>, stored in i32 fragment slots. |
| 293 | for (unsigned i = 0; i < numRegisters; ++i) { |
| 294 | mlir::Value value = cir::ExtractMemberOp::create(builder, location: loc, record: result, index: i); |
| 295 | value = builder.createBitcast(loc, src: value, newTy: dst.getElementType()); |
| 296 | mlir::Value ptr = builder.createPtrStride(loc, base: dst.emitRawPointer(), |
| 297 | stride: builder.getSInt32(c: i, loc)); |
| 298 | builder.createStore(loc, val: value, dst: Address(ptr, alignment)); |
| 299 | } |
| 300 | } |
| 301 | |
| 302 | static std::optional<mlir::Value> |
| 303 | emitWMMALoad(CIRGenFunction &cgf, const CallExpr *expr, unsigned builtinId) { |
| 304 | auto &builder = cgf.getBuilder(); |
| 305 | mlir::Location loc = cgf.getLoc(srcLoc: expr->getExprLoc()); |
| 306 | Address dst = cgf.emitPointerWithAlignment(expr: expr->getArg(Arg: 0)); |
| 307 | mlir::Value src = cgf.emitScalarExpr(e: expr->getArg(Arg: 1)); |
| 308 | mlir::Value ldm = cgf.emitScalarExpr(e: expr->getArg(Arg: 2)); |
| 309 | std::optional<llvm::APSInt> isColMajorArg = |
| 310 | expr->getArg(Arg: 3)->getIntegerConstantExpr(Ctx: cgf.getContext()); |
| 311 | if (!isColMajorArg) |
| 312 | return std::nullopt; |
| 313 | bool isColMajor = isColMajorArg->getSExtValue(); |
| 314 | |
| 315 | NVPTXMmaLdstInfo info = getNVPTXMmaLdstInfo(BuiltinID: builtinId); |
| 316 | unsigned intrinsicId = isColMajor ? info.IID_col : info.IID_row; |
| 317 | if (!intrinsicId) |
| 318 | return std::nullopt; |
| 319 | |
| 320 | cir::FuncType intrinsicType = |
| 321 | cgf.getIntrinsicType(id: intrinsicId, overloadTys: {src.getType()}); |
| 322 | llvm::StringRef name = llvm::Intrinsic::getBaseName(id: intrinsicId); |
| 323 | assert(name.starts_with("llvm." ) && "Not an LLVM intrinsic!" ); |
| 324 | name = name.drop_front(N: 5); |
| 325 | mlir::Value result = builder.emitIntrinsicCallOp( |
| 326 | loc, str: name, resTy: intrinsicType.getReturnType(), op: mlir::ValueRange{src, ldm}); |
| 327 | storeWMMAFragment(builder, loc, result, dst, numRegisters: info.NumResults); |
| 328 | return mlir::Value{}; |
| 329 | } |
| 330 | |
| 331 | std::optional<mlir::Value> |
| 332 | CIRGenFunction::emitNVPTXBuiltinExpr(unsigned builtinId, const CallExpr *expr) { |
| 333 | switch (builtinId) { |
| 334 | case NVPTX::BI__nvvm_atom_add_gen_i: |
| 335 | case NVPTX::BI__nvvm_atom_add_gen_l: |
| 336 | case NVPTX::BI__nvvm_atom_add_gen_ll: |
| 337 | return makeBinaryAtomicValue(kind: cir::AtomicFetchKind::Add, expr, |
| 338 | /*originalArgType=*/nullptr, |
| 339 | /*emittedArgValue=*/nullptr, |
| 340 | ordering: cir::MemOrder::Relaxed); |
| 341 | case NVPTX::BI__nvvm_atom_sub_gen_i: |
| 342 | case NVPTX::BI__nvvm_atom_sub_gen_l: |
| 343 | case NVPTX::BI__nvvm_atom_sub_gen_ll: |
| 344 | return makeBinaryAtomicValue(kind: cir::AtomicFetchKind::Sub, expr, |
| 345 | /*originalArgType=*/nullptr, |
| 346 | /*emittedArgValue=*/nullptr, |
| 347 | ordering: cir::MemOrder::Relaxed); |
| 348 | case NVPTX::BI__nvvm_atom_and_gen_i: |
| 349 | case NVPTX::BI__nvvm_atom_and_gen_l: |
| 350 | case NVPTX::BI__nvvm_atom_and_gen_ll: |
| 351 | return makeBinaryAtomicValue(kind: cir::AtomicFetchKind::And, expr, |
| 352 | /*originalArgType=*/nullptr, |
| 353 | /*emittedArgValue=*/nullptr, |
| 354 | ordering: cir::MemOrder::Relaxed); |
| 355 | case NVPTX::BI__nvvm_atom_or_gen_i: |
| 356 | case NVPTX::BI__nvvm_atom_or_gen_l: |
| 357 | case NVPTX::BI__nvvm_atom_or_gen_ll: |
| 358 | return makeBinaryAtomicValue(kind: cir::AtomicFetchKind::Or, expr, |
| 359 | /*originalArgType=*/nullptr, |
| 360 | /*emittedArgValue=*/nullptr, |
| 361 | ordering: cir::MemOrder::Relaxed); |
| 362 | case NVPTX::BI__nvvm_atom_xor_gen_i: |
| 363 | case NVPTX::BI__nvvm_atom_xor_gen_l: |
| 364 | case NVPTX::BI__nvvm_atom_xor_gen_ll: |
| 365 | return makeBinaryAtomicValue(kind: cir::AtomicFetchKind::Xor, expr, |
| 366 | /*originalArgType=*/nullptr, |
| 367 | /*emittedArgValue=*/nullptr, |
| 368 | ordering: cir::MemOrder::Relaxed); |
| 369 | case NVPTX::BI__nvvm_atom_xchg_gen_i: |
| 370 | case NVPTX::BI__nvvm_atom_xchg_gen_l: |
| 371 | case NVPTX::BI__nvvm_atom_xchg_gen_ll: |
| 372 | return makeScopedAtomicXchg(cgf&: *this, expr, scope: cir::SyncScopeKind::System); |
| 373 | case NVPTX::BI__nvvm_atom_max_gen_i: |
| 374 | case NVPTX::BI__nvvm_atom_max_gen_l: |
| 375 | case NVPTX::BI__nvvm_atom_max_gen_ll: |
| 376 | return makeBinaryAtomicValue(kind: cir::AtomicFetchKind::Max, expr, |
| 377 | /*originalArgType=*/nullptr, |
| 378 | /*emittedArgValue=*/nullptr, |
| 379 | ordering: cir::MemOrder::Relaxed); |
| 380 | case NVPTX::BI__nvvm_atom_max_gen_ui: |
| 381 | case NVPTX::BI__nvvm_atom_max_gen_ul: |
| 382 | case NVPTX::BI__nvvm_atom_max_gen_ull: |
| 383 | return makeBinaryAtomicValue(kind: cir::AtomicFetchKind::Max, expr, |
| 384 | /*originalArgType=*/nullptr, |
| 385 | /*emittedArgValue=*/nullptr, |
| 386 | ordering: cir::MemOrder::Relaxed); |
| 387 | case NVPTX::BI__nvvm_atom_min_gen_i: |
| 388 | case NVPTX::BI__nvvm_atom_min_gen_l: |
| 389 | case NVPTX::BI__nvvm_atom_min_gen_ll: |
| 390 | return makeBinaryAtomicValue(kind: cir::AtomicFetchKind::Min, expr, |
| 391 | /*originalArgType=*/nullptr, |
| 392 | /*emittedArgValue=*/nullptr, |
| 393 | ordering: cir::MemOrder::Relaxed); |
| 394 | case NVPTX::BI__nvvm_atom_min_gen_ui: |
| 395 | case NVPTX::BI__nvvm_atom_min_gen_ul: |
| 396 | case NVPTX::BI__nvvm_atom_min_gen_ull: |
| 397 | return makeBinaryAtomicValue(kind: cir::AtomicFetchKind::Min, expr, |
| 398 | /*originalArgType=*/nullptr, |
| 399 | /*emittedArgValue=*/nullptr, |
| 400 | ordering: cir::MemOrder::Relaxed); |
| 401 | case NVPTX::BI__nvvm_atom_cas_gen_us: |
| 402 | case NVPTX::BI__nvvm_atom_cas_gen_i: |
| 403 | case NVPTX::BI__nvvm_atom_cas_gen_l: |
| 404 | case NVPTX::BI__nvvm_atom_cas_gen_ll: |
| 405 | return emitAtomicCmpXchg(expr, /*returnBool=*/false, successOrder: cir::MemOrder::Relaxed, |
| 406 | failureOrder: cir::MemOrder::Relaxed, |
| 407 | scope: cir::SyncScopeKind::System); |
| 408 | case NVPTX::BI__nvvm_atom_add_gen_f: |
| 409 | case NVPTX::BI__nvvm_atom_add_gen_d: |
| 410 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Add, |
| 411 | scope: cir::SyncScopeKind::System); |
| 412 | case NVPTX::BI__nvvm_atom_inc_gen_ui: |
| 413 | return makeBinaryAtomicValue(kind: cir::AtomicFetchKind::UIncWrap, expr, |
| 414 | /*originalArgType=*/nullptr, |
| 415 | /*emittedArgValue=*/nullptr, |
| 416 | ordering: cir::MemOrder::Relaxed); |
| 417 | case NVPTX::BI__nvvm_atom_dec_gen_ui: |
| 418 | return makeBinaryAtomicValue(kind: cir::AtomicFetchKind::UDecWrap, expr, |
| 419 | /*originalArgType=*/nullptr, |
| 420 | /*emittedArgValue=*/nullptr, |
| 421 | ordering: cir::MemOrder::Relaxed); |
| 422 | case NVPTX::BI__nvvm_ldg_c: |
| 423 | case NVPTX::BI__nvvm_ldg_sc: |
| 424 | case NVPTX::BI__nvvm_ldg_c2: |
| 425 | case NVPTX::BI__nvvm_ldg_sc2: |
| 426 | case NVPTX::BI__nvvm_ldg_c4: |
| 427 | case NVPTX::BI__nvvm_ldg_sc4: |
| 428 | case NVPTX::BI__nvvm_ldg_s: |
| 429 | case NVPTX::BI__nvvm_ldg_s2: |
| 430 | case NVPTX::BI__nvvm_ldg_s4: |
| 431 | case NVPTX::BI__nvvm_ldg_i: |
| 432 | case NVPTX::BI__nvvm_ldg_i2: |
| 433 | case NVPTX::BI__nvvm_ldg_i4: |
| 434 | case NVPTX::BI__nvvm_ldg_l: |
| 435 | case NVPTX::BI__nvvm_ldg_l2: |
| 436 | case NVPTX::BI__nvvm_ldg_ll: |
| 437 | case NVPTX::BI__nvvm_ldg_ll2: |
| 438 | case NVPTX::BI__nvvm_ldg_uc: |
| 439 | case NVPTX::BI__nvvm_ldg_uc2: |
| 440 | case NVPTX::BI__nvvm_ldg_uc4: |
| 441 | case NVPTX::BI__nvvm_ldg_us: |
| 442 | case NVPTX::BI__nvvm_ldg_us2: |
| 443 | case NVPTX::BI__nvvm_ldg_us4: |
| 444 | case NVPTX::BI__nvvm_ldg_ui: |
| 445 | case NVPTX::BI__nvvm_ldg_ui2: |
| 446 | case NVPTX::BI__nvvm_ldg_ui4: |
| 447 | case NVPTX::BI__nvvm_ldg_ul: |
| 448 | case NVPTX::BI__nvvm_ldg_ul2: |
| 449 | case NVPTX::BI__nvvm_ldg_ull: |
| 450 | case NVPTX::BI__nvvm_ldg_ull2: |
| 451 | case NVPTX::BI__nvvm_ldg_f: |
| 452 | case NVPTX::BI__nvvm_ldg_f2: |
| 453 | case NVPTX::BI__nvvm_ldg_f4: |
| 454 | case NVPTX::BI__nvvm_ldg_d: |
| 455 | case NVPTX::BI__nvvm_ldg_d2: |
| 456 | return makeLdg(cgf&: *this, expr); |
| 457 | case NVPTX::BI__nvvm_ldu_c: |
| 458 | case NVPTX::BI__nvvm_ldu_sc: |
| 459 | case NVPTX::BI__nvvm_ldu_c2: |
| 460 | case NVPTX::BI__nvvm_ldu_sc2: |
| 461 | case NVPTX::BI__nvvm_ldu_c4: |
| 462 | case NVPTX::BI__nvvm_ldu_sc4: |
| 463 | case NVPTX::BI__nvvm_ldu_s: |
| 464 | case NVPTX::BI__nvvm_ldu_s2: |
| 465 | case NVPTX::BI__nvvm_ldu_s4: |
| 466 | case NVPTX::BI__nvvm_ldu_i: |
| 467 | case NVPTX::BI__nvvm_ldu_i2: |
| 468 | case NVPTX::BI__nvvm_ldu_i4: |
| 469 | case NVPTX::BI__nvvm_ldu_l: |
| 470 | case NVPTX::BI__nvvm_ldu_l2: |
| 471 | case NVPTX::BI__nvvm_ldu_ll: |
| 472 | case NVPTX::BI__nvvm_ldu_ll2: |
| 473 | case NVPTX::BI__nvvm_ldu_uc: |
| 474 | case NVPTX::BI__nvvm_ldu_uc2: |
| 475 | case NVPTX::BI__nvvm_ldu_uc4: |
| 476 | case NVPTX::BI__nvvm_ldu_us: |
| 477 | case NVPTX::BI__nvvm_ldu_us2: |
| 478 | case NVPTX::BI__nvvm_ldu_us4: |
| 479 | case NVPTX::BI__nvvm_ldu_ui: |
| 480 | case NVPTX::BI__nvvm_ldu_ui2: |
| 481 | case NVPTX::BI__nvvm_ldu_ui4: |
| 482 | case NVPTX::BI__nvvm_ldu_ul: |
| 483 | case NVPTX::BI__nvvm_ldu_ul2: |
| 484 | case NVPTX::BI__nvvm_ldu_ull: |
| 485 | case NVPTX::BI__nvvm_ldu_ull2: |
| 486 | return makeLdu(cgf&: *this, expr, intrinsicName: "nvvm.ldu.global.i" ); |
| 487 | case NVPTX::BI__nvvm_ldu_f: |
| 488 | case NVPTX::BI__nvvm_ldu_f2: |
| 489 | case NVPTX::BI__nvvm_ldu_f4: |
| 490 | case NVPTX::BI__nvvm_ldu_d: |
| 491 | case NVPTX::BI__nvvm_ldu_d2: |
| 492 | return makeLdu(cgf&: *this, expr, intrinsicName: "nvvm.ldu.global.f" ); |
| 493 | case NVPTX::BI__nvvm_atom_cta_add_gen_i: |
| 494 | case NVPTX::BI__nvvm_atom_cta_add_gen_l: |
| 495 | case NVPTX::BI__nvvm_atom_cta_add_gen_ll: |
| 496 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Add, |
| 497 | scope: cir::SyncScopeKind::Workgroup); |
| 498 | case NVPTX::BI__nvvm_atom_sys_add_gen_i: |
| 499 | case NVPTX::BI__nvvm_atom_sys_add_gen_l: |
| 500 | case NVPTX::BI__nvvm_atom_sys_add_gen_ll: |
| 501 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Add, |
| 502 | scope: cir::SyncScopeKind::System); |
| 503 | case NVPTX::BI__nvvm_atom_cta_add_gen_f: |
| 504 | case NVPTX::BI__nvvm_atom_cta_add_gen_d: |
| 505 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Add, |
| 506 | scope: cir::SyncScopeKind::Workgroup); |
| 507 | case NVPTX::BI__nvvm_atom_sys_add_gen_f: |
| 508 | case NVPTX::BI__nvvm_atom_sys_add_gen_d: |
| 509 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Add, |
| 510 | scope: cir::SyncScopeKind::System); |
| 511 | case NVPTX::BI__nvvm_atom_cta_xchg_gen_i: |
| 512 | case NVPTX::BI__nvvm_atom_cta_xchg_gen_l: |
| 513 | case NVPTX::BI__nvvm_atom_cta_xchg_gen_ll: |
| 514 | return makeScopedAtomicXchg(cgf&: *this, expr, scope: cir::SyncScopeKind::Workgroup); |
| 515 | case NVPTX::BI__nvvm_atom_sys_xchg_gen_i: |
| 516 | case NVPTX::BI__nvvm_atom_sys_xchg_gen_l: |
| 517 | case NVPTX::BI__nvvm_atom_sys_xchg_gen_ll: |
| 518 | return makeScopedAtomicXchg(cgf&: *this, expr, scope: cir::SyncScopeKind::System); |
| 519 | case NVPTX::BI__nvvm_atom_cta_max_gen_i: |
| 520 | case NVPTX::BI__nvvm_atom_cta_max_gen_ui: |
| 521 | case NVPTX::BI__nvvm_atom_cta_max_gen_l: |
| 522 | case NVPTX::BI__nvvm_atom_cta_max_gen_ul: |
| 523 | case NVPTX::BI__nvvm_atom_cta_max_gen_ll: |
| 524 | case NVPTX::BI__nvvm_atom_cta_max_gen_ull: |
| 525 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Max, |
| 526 | scope: cir::SyncScopeKind::Workgroup); |
| 527 | case NVPTX::BI__nvvm_atom_sys_max_gen_i: |
| 528 | case NVPTX::BI__nvvm_atom_sys_max_gen_ui: |
| 529 | case NVPTX::BI__nvvm_atom_sys_max_gen_l: |
| 530 | case NVPTX::BI__nvvm_atom_sys_max_gen_ul: |
| 531 | case NVPTX::BI__nvvm_atom_sys_max_gen_ll: |
| 532 | case NVPTX::BI__nvvm_atom_sys_max_gen_ull: |
| 533 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Max, |
| 534 | scope: cir::SyncScopeKind::System); |
| 535 | case NVPTX::BI__nvvm_atom_cta_min_gen_i: |
| 536 | case NVPTX::BI__nvvm_atom_cta_min_gen_ui: |
| 537 | case NVPTX::BI__nvvm_atom_cta_min_gen_l: |
| 538 | case NVPTX::BI__nvvm_atom_cta_min_gen_ul: |
| 539 | case NVPTX::BI__nvvm_atom_cta_min_gen_ll: |
| 540 | case NVPTX::BI__nvvm_atom_cta_min_gen_ull: |
| 541 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Min, |
| 542 | scope: cir::SyncScopeKind::Workgroup); |
| 543 | case NVPTX::BI__nvvm_atom_sys_min_gen_i: |
| 544 | case NVPTX::BI__nvvm_atom_sys_min_gen_ui: |
| 545 | case NVPTX::BI__nvvm_atom_sys_min_gen_l: |
| 546 | case NVPTX::BI__nvvm_atom_sys_min_gen_ul: |
| 547 | case NVPTX::BI__nvvm_atom_sys_min_gen_ll: |
| 548 | case NVPTX::BI__nvvm_atom_sys_min_gen_ull: |
| 549 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Min, |
| 550 | scope: cir::SyncScopeKind::System); |
| 551 | case NVPTX::BI__nvvm_atom_cta_inc_gen_ui: |
| 552 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::UIncWrap, |
| 553 | scope: cir::SyncScopeKind::Workgroup); |
| 554 | case NVPTX::BI__nvvm_atom_cta_dec_gen_ui: |
| 555 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::UDecWrap, |
| 556 | scope: cir::SyncScopeKind::Workgroup); |
| 557 | case NVPTX::BI__nvvm_atom_sys_inc_gen_ui: |
| 558 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::UIncWrap, |
| 559 | scope: cir::SyncScopeKind::System); |
| 560 | case NVPTX::BI__nvvm_atom_sys_dec_gen_ui: |
| 561 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::UDecWrap, |
| 562 | scope: cir::SyncScopeKind::System); |
| 563 | case NVPTX::BI__nvvm_atom_cta_and_gen_i: |
| 564 | case NVPTX::BI__nvvm_atom_cta_and_gen_l: |
| 565 | case NVPTX::BI__nvvm_atom_cta_and_gen_ll: |
| 566 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::And, |
| 567 | scope: cir::SyncScopeKind::Workgroup); |
| 568 | case NVPTX::BI__nvvm_atom_sys_and_gen_i: |
| 569 | case NVPTX::BI__nvvm_atom_sys_and_gen_l: |
| 570 | case NVPTX::BI__nvvm_atom_sys_and_gen_ll: |
| 571 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::And, |
| 572 | scope: cir::SyncScopeKind::System); |
| 573 | case NVPTX::BI__nvvm_atom_cta_or_gen_i: |
| 574 | case NVPTX::BI__nvvm_atom_cta_or_gen_l: |
| 575 | case NVPTX::BI__nvvm_atom_cta_or_gen_ll: |
| 576 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Or, |
| 577 | scope: cir::SyncScopeKind::Workgroup); |
| 578 | case NVPTX::BI__nvvm_atom_sys_or_gen_i: |
| 579 | case NVPTX::BI__nvvm_atom_sys_or_gen_l: |
| 580 | case NVPTX::BI__nvvm_atom_sys_or_gen_ll: |
| 581 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Or, |
| 582 | scope: cir::SyncScopeKind::System); |
| 583 | case NVPTX::BI__nvvm_atom_cta_xor_gen_i: |
| 584 | case NVPTX::BI__nvvm_atom_cta_xor_gen_l: |
| 585 | case NVPTX::BI__nvvm_atom_cta_xor_gen_ll: |
| 586 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Xor, |
| 587 | scope: cir::SyncScopeKind::Workgroup); |
| 588 | case NVPTX::BI__nvvm_atom_sys_xor_gen_i: |
| 589 | case NVPTX::BI__nvvm_atom_sys_xor_gen_l: |
| 590 | case NVPTX::BI__nvvm_atom_sys_xor_gen_ll: |
| 591 | return makeScopedAtomicRMW(cgf&: *this, expr, kind: cir::AtomicFetchKind::Xor, |
| 592 | scope: cir::SyncScopeKind::System); |
| 593 | case NVPTX::BI__nvvm_atom_cta_cas_gen_us: |
| 594 | case NVPTX::BI__nvvm_atom_cta_cas_gen_i: |
| 595 | case NVPTX::BI__nvvm_atom_cta_cas_gen_l: |
| 596 | case NVPTX::BI__nvvm_atom_cta_cas_gen_ll: |
| 597 | return emitAtomicCmpXchg(expr, /*returnBool=*/false, successOrder: cir::MemOrder::Relaxed, |
| 598 | failureOrder: cir::MemOrder::Relaxed, |
| 599 | scope: cir::SyncScopeKind::Workgroup); |
| 600 | case NVPTX::BI__nvvm_atom_sys_cas_gen_us: |
| 601 | case NVPTX::BI__nvvm_atom_sys_cas_gen_i: |
| 602 | case NVPTX::BI__nvvm_atom_sys_cas_gen_l: |
| 603 | case NVPTX::BI__nvvm_atom_sys_cas_gen_ll: |
| 604 | return emitAtomicCmpXchg(expr, /*returnBool=*/false, successOrder: cir::MemOrder::Relaxed, |
| 605 | failureOrder: cir::MemOrder::Relaxed, |
| 606 | scope: cir::SyncScopeKind::System); |
| 607 | case NVPTX::BI__nvvm_match_all_sync_i32p: |
| 608 | case NVPTX::BI__nvvm_match_all_sync_i64p: |
| 609 | cgm.errorNYI(expr->getSourceRange(), |
| 610 | std::string("unimplemented NVPTX builtin call: " ) + |
| 611 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 612 | return mlir::Value{}; |
| 613 | // WMMA fragment loads. |
| 614 | case NVPTX::BI__hmma_m16n16k16_ld_a: |
| 615 | case NVPTX::BI__hmma_m16n16k16_ld_b: |
| 616 | case NVPTX::BI__hmma_m16n16k16_ld_c_f16: |
| 617 | case NVPTX::BI__hmma_m16n16k16_ld_c_f32: |
| 618 | case NVPTX::BI__hmma_m32n8k16_ld_a: |
| 619 | case NVPTX::BI__hmma_m32n8k16_ld_b: |
| 620 | case NVPTX::BI__hmma_m32n8k16_ld_c_f16: |
| 621 | case NVPTX::BI__hmma_m32n8k16_ld_c_f32: |
| 622 | case NVPTX::BI__hmma_m8n32k16_ld_a: |
| 623 | case NVPTX::BI__hmma_m8n32k16_ld_b: |
| 624 | case NVPTX::BI__hmma_m8n32k16_ld_c_f16: |
| 625 | case NVPTX::BI__hmma_m8n32k16_ld_c_f32: |
| 626 | case NVPTX::BI__imma_m16n16k16_ld_a_s8: |
| 627 | case NVPTX::BI__imma_m16n16k16_ld_a_u8: |
| 628 | case NVPTX::BI__imma_m16n16k16_ld_b_s8: |
| 629 | case NVPTX::BI__imma_m16n16k16_ld_b_u8: |
| 630 | case NVPTX::BI__imma_m16n16k16_ld_c: |
| 631 | case NVPTX::BI__imma_m32n8k16_ld_a_s8: |
| 632 | case NVPTX::BI__imma_m32n8k16_ld_a_u8: |
| 633 | case NVPTX::BI__imma_m32n8k16_ld_b_s8: |
| 634 | case NVPTX::BI__imma_m32n8k16_ld_b_u8: |
| 635 | case NVPTX::BI__imma_m32n8k16_ld_c: |
| 636 | case NVPTX::BI__imma_m8n32k16_ld_a_s8: |
| 637 | case NVPTX::BI__imma_m8n32k16_ld_a_u8: |
| 638 | case NVPTX::BI__imma_m8n32k16_ld_b_s8: |
| 639 | case NVPTX::BI__imma_m8n32k16_ld_b_u8: |
| 640 | case NVPTX::BI__imma_m8n32k16_ld_c: |
| 641 | case NVPTX::BI__imma_m8n8k32_ld_a_s4: |
| 642 | case NVPTX::BI__imma_m8n8k32_ld_a_u4: |
| 643 | case NVPTX::BI__imma_m8n8k32_ld_b_s4: |
| 644 | case NVPTX::BI__imma_m8n8k32_ld_b_u4: |
| 645 | case NVPTX::BI__imma_m8n8k32_ld_c: |
| 646 | case NVPTX::BI__bmma_m8n8k128_ld_a_b1: |
| 647 | case NVPTX::BI__bmma_m8n8k128_ld_b_b1: |
| 648 | case NVPTX::BI__bmma_m8n8k128_ld_c: |
| 649 | case NVPTX::BI__dmma_m8n8k4_ld_a: |
| 650 | case NVPTX::BI__dmma_m8n8k4_ld_b: |
| 651 | case NVPTX::BI__dmma_m8n8k4_ld_c: |
| 652 | case NVPTX::BI__mma_bf16_m16n16k16_ld_a: |
| 653 | case NVPTX::BI__mma_bf16_m16n16k16_ld_b: |
| 654 | case NVPTX::BI__mma_bf16_m8n32k16_ld_a: |
| 655 | case NVPTX::BI__mma_bf16_m8n32k16_ld_b: |
| 656 | case NVPTX::BI__mma_bf16_m32n8k16_ld_a: |
| 657 | case NVPTX::BI__mma_bf16_m32n8k16_ld_b: |
| 658 | case NVPTX::BI__mma_tf32_m16n16k8_ld_a: |
| 659 | case NVPTX::BI__mma_tf32_m16n16k8_ld_b: |
| 660 | case NVPTX::BI__mma_tf32_m16n16k8_ld_c: |
| 661 | return emitWMMALoad(cgf&: *this, expr, builtinId); |
| 662 | case NVPTX::BI__hmma_m16n16k16_st_c_f16: |
| 663 | case NVPTX::BI__hmma_m16n16k16_st_c_f32: |
| 664 | case NVPTX::BI__hmma_m32n8k16_st_c_f16: |
| 665 | case NVPTX::BI__hmma_m32n8k16_st_c_f32: |
| 666 | case NVPTX::BI__hmma_m8n32k16_st_c_f16: |
| 667 | case NVPTX::BI__hmma_m8n32k16_st_c_f32: |
| 668 | case NVPTX::BI__imma_m16n16k16_st_c_i32: |
| 669 | case NVPTX::BI__imma_m32n8k16_st_c_i32: |
| 670 | case NVPTX::BI__imma_m8n32k16_st_c_i32: |
| 671 | case NVPTX::BI__imma_m8n8k32_st_c_i32: |
| 672 | case NVPTX::BI__bmma_m8n8k128_st_c_i32: |
| 673 | case NVPTX::BI__dmma_m8n8k4_st_c_f64: |
| 674 | case NVPTX::BI__mma_m16n16k8_st_c_f32: |
| 675 | cgm.errorNYI(expr->getSourceRange(), |
| 676 | std::string("unimplemented NVPTX builtin call: " ) + |
| 677 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 678 | return mlir::Value{}; |
| 679 | // BI__hmma_m16n16k16_mma_<Dtype><CType>(d, a, b, c, layout, satf) --> |
| 680 | // Intrinsic::nvvm_wmma_m16n16k16_mma_sync<layout A,B><DType><CType><Satf> |
| 681 | case NVPTX::BI__hmma_m16n16k16_mma_f16f16: |
| 682 | case NVPTX::BI__hmma_m16n16k16_mma_f32f16: |
| 683 | case NVPTX::BI__hmma_m16n16k16_mma_f32f32: |
| 684 | case NVPTX::BI__hmma_m16n16k16_mma_f16f32: |
| 685 | case NVPTX::BI__hmma_m32n8k16_mma_f16f16: |
| 686 | case NVPTX::BI__hmma_m32n8k16_mma_f32f16: |
| 687 | case NVPTX::BI__hmma_m32n8k16_mma_f32f32: |
| 688 | case NVPTX::BI__hmma_m32n8k16_mma_f16f32: |
| 689 | case NVPTX::BI__hmma_m8n32k16_mma_f16f16: |
| 690 | case NVPTX::BI__hmma_m8n32k16_mma_f32f16: |
| 691 | case NVPTX::BI__hmma_m8n32k16_mma_f32f32: |
| 692 | case NVPTX::BI__hmma_m8n32k16_mma_f16f32: |
| 693 | case NVPTX::BI__imma_m16n16k16_mma_s8: |
| 694 | case NVPTX::BI__imma_m16n16k16_mma_u8: |
| 695 | case NVPTX::BI__imma_m32n8k16_mma_s8: |
| 696 | case NVPTX::BI__imma_m32n8k16_mma_u8: |
| 697 | case NVPTX::BI__imma_m8n32k16_mma_s8: |
| 698 | case NVPTX::BI__imma_m8n32k16_mma_u8: |
| 699 | case NVPTX::BI__imma_m8n8k32_mma_s4: |
| 700 | case NVPTX::BI__imma_m8n8k32_mma_u4: |
| 701 | case NVPTX::BI__bmma_m8n8k128_mma_xor_popc_b1: |
| 702 | case NVPTX::BI__bmma_m8n8k128_mma_and_popc_b1: |
| 703 | case NVPTX::BI__dmma_m8n8k4_mma_f64: |
| 704 | case NVPTX::BI__mma_bf16_m16n16k16_mma_f32: |
| 705 | case NVPTX::BI__mma_bf16_m8n32k16_mma_f32: |
| 706 | case NVPTX::BI__mma_bf16_m32n8k16_mma_f32: |
| 707 | case NVPTX::BI__mma_tf32_m16n16k8_mma_f32: |
| 708 | cgm.errorNYI(expr->getSourceRange(), |
| 709 | std::string("unimplemented NVPTX builtin call: " ) + |
| 710 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 711 | return mlir::Value{}; |
| 712 | // The following builtins require half type support |
| 713 | case NVPTX::BI__nvvm_ff2f16x2_rn: |
| 714 | cgm.errorNYI(expr->getSourceRange(), |
| 715 | std::string("unimplemented NVPTX builtin call: " ) + |
| 716 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 717 | return mlir::Value{}; |
| 718 | case NVPTX::BI__nvvm_ff2f16x2_rn_relu: |
| 719 | cgm.errorNYI(expr->getSourceRange(), |
| 720 | std::string("unimplemented NVPTX builtin call: " ) + |
| 721 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 722 | return mlir::Value{}; |
| 723 | case NVPTX::BI__nvvm_ff2f16x2_rz: |
| 724 | cgm.errorNYI(expr->getSourceRange(), |
| 725 | std::string("unimplemented NVPTX builtin call: " ) + |
| 726 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 727 | return mlir::Value{}; |
| 728 | case NVPTX::BI__nvvm_ff2f16x2_rz_relu: |
| 729 | cgm.errorNYI(expr->getSourceRange(), |
| 730 | std::string("unimplemented NVPTX builtin call: " ) + |
| 731 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 732 | return mlir::Value{}; |
| 733 | case NVPTX::BI__nvvm_fma_rn_f16: |
| 734 | case NVPTX::BI__nvvm_fma_rn_f16x2: |
| 735 | case NVPTX::BI__nvvm_fma_rn_bf16: |
| 736 | case NVPTX::BI__nvvm_fma_rn_bf16x2: |
| 737 | case NVPTX::BI__nvvm_fma_rn_f: |
| 738 | case NVPTX::BI__nvvm_fma_rn_d: |
| 739 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma" , |
| 740 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 741 | case NVPTX::BI__nvvm_fma_rz_f: |
| 742 | case NVPTX::BI__nvvm_fma_rz_d: |
| 743 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma" , rm: llvm::APFloat::rmTowardZero); |
| 744 | case NVPTX::BI__nvvm_fma_rm_f: |
| 745 | case NVPTX::BI__nvvm_fma_rm_d: |
| 746 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma" , |
| 747 | rm: llvm::APFloat::rmTowardNegative); |
| 748 | case NVPTX::BI__nvvm_fma_rp_f: |
| 749 | case NVPTX::BI__nvvm_fma_rp_d: |
| 750 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma" , |
| 751 | rm: llvm::APFloat::rmTowardPositive); |
| 752 | case NVPTX::BI__nvvm_fma_rn_ftz_f16: |
| 753 | case NVPTX::BI__nvvm_fma_rn_ftz_f16x2: |
| 754 | case NVPTX::BI__nvvm_fma_rn_ftz_f: |
| 755 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.ftz" , |
| 756 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 757 | case NVPTX::BI__nvvm_fma_rz_ftz_f: |
| 758 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.ftz" , |
| 759 | rm: llvm::APFloat::rmTowardZero); |
| 760 | case NVPTX::BI__nvvm_fma_rm_ftz_f: |
| 761 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.ftz" , |
| 762 | rm: llvm::APFloat::rmTowardNegative); |
| 763 | case NVPTX::BI__nvvm_fma_rp_ftz_f: |
| 764 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.ftz" , |
| 765 | rm: llvm::APFloat::rmTowardPositive); |
| 766 | case NVPTX::BI__nvvm_fma_rn_sat_f16: |
| 767 | case NVPTX::BI__nvvm_fma_rn_sat_f16x2: |
| 768 | case NVPTX::BI__nvvm_fma_rn_sat_f: |
| 769 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.sat" , |
| 770 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 771 | case NVPTX::BI__nvvm_fma_rz_sat_f: |
| 772 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.sat" , |
| 773 | rm: llvm::APFloat::rmTowardZero); |
| 774 | case NVPTX::BI__nvvm_fma_rm_sat_f: |
| 775 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.sat" , |
| 776 | rm: llvm::APFloat::rmTowardNegative); |
| 777 | case NVPTX::BI__nvvm_fma_rp_sat_f: |
| 778 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.sat" , |
| 779 | rm: llvm::APFloat::rmTowardPositive); |
| 780 | case NVPTX::BI__nvvm_fma_rn_ftz_sat_f16: |
| 781 | case NVPTX::BI__nvvm_fma_rn_ftz_sat_f16x2: |
| 782 | case NVPTX::BI__nvvm_fma_rn_ftz_sat_f: |
| 783 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.ftz.sat" , |
| 784 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 785 | case NVPTX::BI__nvvm_fma_rz_ftz_sat_f: |
| 786 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.ftz.sat" , |
| 787 | rm: llvm::APFloat::rmTowardZero); |
| 788 | case NVPTX::BI__nvvm_fma_rm_ftz_sat_f: |
| 789 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.ftz.sat" , |
| 790 | rm: llvm::APFloat::rmTowardNegative); |
| 791 | case NVPTX::BI__nvvm_fma_rp_ftz_sat_f: |
| 792 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.ftz.sat" , |
| 793 | rm: llvm::APFloat::rmTowardPositive); |
| 794 | case NVPTX::BI__nvvm_fma_rn_relu_f16: |
| 795 | case NVPTX::BI__nvvm_fma_rn_relu_f16x2: |
| 796 | case NVPTX::BI__nvvm_fma_rn_relu_bf16: |
| 797 | case NVPTX::BI__nvvm_fma_rn_relu_bf16x2: |
| 798 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.relu" , |
| 799 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 800 | case NVPTX::BI__nvvm_fma_rn_ftz_relu_f16: |
| 801 | case NVPTX::BI__nvvm_fma_rn_ftz_relu_f16x2: |
| 802 | return emitNVVMFMA(cgf&: *this, expr, intrinsicName: "nvvm.fma.ftz.relu" , |
| 803 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 804 | case NVPTX::BI__nvvm_fma_rn_oob_f16: |
| 805 | case NVPTX::BI__nvvm_fma_rn_oob_f16x2: |
| 806 | case NVPTX::BI__nvvm_fma_rn_oob_bf16: |
| 807 | case NVPTX::BI__nvvm_fma_rn_oob_bf16x2: |
| 808 | return emitBuiltinWithOneOverloadedType<3>(e: expr, intrinName: "nvvm.fma.rn.oob" ) |
| 809 | .getValue(); |
| 810 | case NVPTX::BI__nvvm_fma_rn_oob_relu_f16: |
| 811 | case NVPTX::BI__nvvm_fma_rn_oob_relu_f16x2: |
| 812 | case NVPTX::BI__nvvm_fma_rn_oob_relu_bf16: |
| 813 | case NVPTX::BI__nvvm_fma_rn_oob_relu_bf16x2: |
| 814 | return emitBuiltinWithOneOverloadedType<3>(e: expr, intrinName: "nvvm.fma.rn.oob.relu" ) |
| 815 | .getValue(); |
| 816 | case NVPTX::BI__nvvm_fmax_f16: |
| 817 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmax.f16" ) |
| 818 | .getValue(); |
| 819 | case NVPTX::BI__nvvm_fmax_f16x2: |
| 820 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmax.f16x2" ) |
| 821 | .getValue(); |
| 822 | case NVPTX::BI__nvvm_fmax_ftz_f16: |
| 823 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmax.ftz.f16" ) |
| 824 | .getValue(); |
| 825 | case NVPTX::BI__nvvm_fmax_ftz_f16x2: |
| 826 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmax.ftz.f16x2" ) |
| 827 | .getValue(); |
| 828 | case NVPTX::BI__nvvm_fmax_ftz_nan_f16: |
| 829 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmax.ftz.nan.f16" ) |
| 830 | .getValue(); |
| 831 | case NVPTX::BI__nvvm_fmax_ftz_nan_f16x2: |
| 832 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmax.ftz.nan.f16x2" ) |
| 833 | .getValue(); |
| 834 | case NVPTX::BI__nvvm_fmax_ftz_nan_xorsign_abs_f16: |
| 835 | return emitBuiltinWithOneOverloadedType<2>( |
| 836 | e: expr, intrinName: "nvvm.fmax.ftz.nan.xorsign.abs.f16" ) |
| 837 | .getValue(); |
| 838 | case NVPTX::BI__nvvm_fmax_ftz_nan_xorsign_abs_f16x2: |
| 839 | return emitBuiltinWithOneOverloadedType<2>( |
| 840 | e: expr, intrinName: "nvvm.fmax.ftz.nan.xorsign.abs.f16x2" ) |
| 841 | .getValue(); |
| 842 | case NVPTX::BI__nvvm_fmax_ftz_xorsign_abs_f16: |
| 843 | return emitBuiltinWithOneOverloadedType<2>(e: expr, |
| 844 | intrinName: "nvvm.fmax.ftz.xorsign.abs.f16" ) |
| 845 | .getValue(); |
| 846 | case NVPTX::BI__nvvm_fmax_ftz_xorsign_abs_f16x2: |
| 847 | return emitBuiltinWithOneOverloadedType<2>( |
| 848 | e: expr, intrinName: "nvvm.fmax.ftz.xorsign.abs.f16x2" ) |
| 849 | .getValue(); |
| 850 | case NVPTX::BI__nvvm_fmax_nan_f16: |
| 851 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmax.nan.f16" ) |
| 852 | .getValue(); |
| 853 | case NVPTX::BI__nvvm_fmax_nan_f16x2: |
| 854 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmax.nan.f16x2" ) |
| 855 | .getValue(); |
| 856 | case NVPTX::BI__nvvm_fmax_nan_xorsign_abs_f16: |
| 857 | return emitBuiltinWithOneOverloadedType<2>(e: expr, |
| 858 | intrinName: "nvvm.fmax.nan.xorsign.abs.f16" ) |
| 859 | .getValue(); |
| 860 | case NVPTX::BI__nvvm_fmax_nan_xorsign_abs_f16x2: |
| 861 | return emitBuiltinWithOneOverloadedType<2>( |
| 862 | e: expr, intrinName: "nvvm.fmax.nan.xorsign.abs.f16x2" ) |
| 863 | .getValue(); |
| 864 | case NVPTX::BI__nvvm_fmax_xorsign_abs_f16: |
| 865 | return emitBuiltinWithOneOverloadedType<2>(e: expr, |
| 866 | intrinName: "nvvm.fmax.xorsign.abs.f16" ) |
| 867 | .getValue(); |
| 868 | case NVPTX::BI__nvvm_fmax_xorsign_abs_f16x2: |
| 869 | return emitBuiltinWithOneOverloadedType<2>(e: expr, |
| 870 | intrinName: "nvvm.fmax.xorsign.abs.f16x2" ) |
| 871 | .getValue(); |
| 872 | case NVPTX::BI__nvvm_fmin_f16: |
| 873 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmin.f16" ) |
| 874 | .getValue(); |
| 875 | case NVPTX::BI__nvvm_fmin_f16x2: |
| 876 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmin.f16x2" ) |
| 877 | .getValue(); |
| 878 | case NVPTX::BI__nvvm_fmin_ftz_f16: |
| 879 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmin.ftz.f16" ) |
| 880 | .getValue(); |
| 881 | case NVPTX::BI__nvvm_fmin_ftz_f16x2: |
| 882 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmin.ftz.f16x2" ) |
| 883 | .getValue(); |
| 884 | case NVPTX::BI__nvvm_fmin_ftz_nan_f16: |
| 885 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmin.ftz.nan.f16" ) |
| 886 | .getValue(); |
| 887 | case NVPTX::BI__nvvm_fmin_ftz_nan_f16x2: |
| 888 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmin.ftz.nan.f16x2" ) |
| 889 | .getValue(); |
| 890 | case NVPTX::BI__nvvm_fmin_ftz_nan_xorsign_abs_f16: |
| 891 | return emitBuiltinWithOneOverloadedType<2>( |
| 892 | e: expr, intrinName: "nvvm.fmin.ftz.nan.xorsign.abs.f16" ) |
| 893 | .getValue(); |
| 894 | case NVPTX::BI__nvvm_fmin_ftz_nan_xorsign_abs_f16x2: |
| 895 | return emitBuiltinWithOneOverloadedType<2>( |
| 896 | e: expr, intrinName: "nvvm.fmin.ftz.nan.xorsign.abs.f16x2" ) |
| 897 | .getValue(); |
| 898 | case NVPTX::BI__nvvm_fmin_ftz_xorsign_abs_f16: |
| 899 | return emitBuiltinWithOneOverloadedType<2>(e: expr, |
| 900 | intrinName: "nvvm.fmin.ftz.xorsign.abs.f16" ) |
| 901 | .getValue(); |
| 902 | case NVPTX::BI__nvvm_fmin_ftz_xorsign_abs_f16x2: |
| 903 | return emitBuiltinWithOneOverloadedType<2>( |
| 904 | e: expr, intrinName: "nvvm.fmin.ftz.xorsign.abs.f16x2" ) |
| 905 | .getValue(); |
| 906 | case NVPTX::BI__nvvm_fmin_nan_f16: |
| 907 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmin.nan.f16" ) |
| 908 | .getValue(); |
| 909 | case NVPTX::BI__nvvm_fmin_nan_f16x2: |
| 910 | return emitBuiltinWithOneOverloadedType<2>(e: expr, intrinName: "nvvm.fmin.nan.f16x2" ) |
| 911 | .getValue(); |
| 912 | case NVPTX::BI__nvvm_fmin_nan_xorsign_abs_f16: |
| 913 | return emitBuiltinWithOneOverloadedType<2>(e: expr, |
| 914 | intrinName: "nvvm.fmin.nan.xorsign.abs.f16" ) |
| 915 | .getValue(); |
| 916 | case NVPTX::BI__nvvm_fmin_nan_xorsign_abs_f16x2: |
| 917 | return emitBuiltinWithOneOverloadedType<2>( |
| 918 | e: expr, intrinName: "nvvm.fmin.nan.xorsign.abs.f16x2" ) |
| 919 | .getValue(); |
| 920 | case NVPTX::BI__nvvm_fmin_xorsign_abs_f16: |
| 921 | return emitBuiltinWithOneOverloadedType<2>(e: expr, |
| 922 | intrinName: "nvvm.fmin.xorsign.abs.f16" ) |
| 923 | .getValue(); |
| 924 | case NVPTX::BI__nvvm_fmin_xorsign_abs_f16x2: |
| 925 | return emitBuiltinWithOneOverloadedType<2>(e: expr, |
| 926 | intrinName: "nvvm.fmin.xorsign.abs.f16x2" ) |
| 927 | .getValue(); |
| 928 | case NVPTX::BI__nvvm_fabs_f: |
| 929 | case NVPTX::BI__nvvm_abs_bf16: |
| 930 | case NVPTX::BI__nvvm_abs_bf16x2: |
| 931 | case NVPTX::BI__nvvm_fabs_f16: |
| 932 | case NVPTX::BI__nvvm_fabs_f16x2: |
| 933 | return emitUnaryNVVMIntrinsic(cgf&: *this, expr, intrinsicName: "nvvm.fabs" ); |
| 934 | case NVPTX::BI__nvvm_fabs_ftz_f: |
| 935 | case NVPTX::BI__nvvm_fabs_ftz_f16: |
| 936 | case NVPTX::BI__nvvm_fabs_ftz_f16x2: |
| 937 | return emitUnaryNVVMIntrinsic(cgf&: *this, expr, intrinsicName: "nvvm.fabs.ftz" ); |
| 938 | case NVPTX::BI__nvvm_fabs_d: |
| 939 | return emitUnaryNVVMIntrinsic(cgf&: *this, expr, intrinsicName: "fabs" ); |
| 940 | case NVPTX::BI__nvvm_ex2_approx_d: |
| 941 | case NVPTX::BI__nvvm_ex2_approx_f: |
| 942 | case NVPTX::BI__nvvm_ex2_approx_f16: |
| 943 | case NVPTX::BI__nvvm_ex2_approx_f16x2: |
| 944 | return emitUnaryNVVMIntrinsic(cgf&: *this, expr, intrinsicName: "nvvm.ex2.approx" ); |
| 945 | case NVPTX::BI__nvvm_ex2_approx_ftz_f: |
| 946 | return emitUnaryNVVMIntrinsic(cgf&: *this, expr, intrinsicName: "nvvm.ex2.approx.ftz" ); |
| 947 | case NVPTX::BI__nvvm_add_rn_f: |
| 948 | case NVPTX::BI__nvvm_add_rn_d: |
| 949 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd" , |
| 950 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 951 | case NVPTX::BI__nvvm_add_rz_f: |
| 952 | case NVPTX::BI__nvvm_add_rz_d: |
| 953 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd" , |
| 954 | rm: llvm::APFloat::rmTowardZero); |
| 955 | case NVPTX::BI__nvvm_add_rm_f: |
| 956 | case NVPTX::BI__nvvm_add_rm_d: |
| 957 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd" , |
| 958 | rm: llvm::APFloat::rmTowardNegative); |
| 959 | case NVPTX::BI__nvvm_add_rp_f: |
| 960 | case NVPTX::BI__nvvm_add_rp_d: |
| 961 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd" , |
| 962 | rm: llvm::APFloat::rmTowardPositive); |
| 963 | case NVPTX::BI__nvvm_add_rn_ftz_f: |
| 964 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd.ftz" , |
| 965 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 966 | case NVPTX::BI__nvvm_add_rz_ftz_f: |
| 967 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd.ftz" , |
| 968 | rm: llvm::APFloat::rmTowardZero); |
| 969 | case NVPTX::BI__nvvm_add_rm_ftz_f: |
| 970 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd.ftz" , |
| 971 | rm: llvm::APFloat::rmTowardNegative); |
| 972 | case NVPTX::BI__nvvm_add_rp_ftz_f: |
| 973 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd.ftz" , |
| 974 | rm: llvm::APFloat::rmTowardPositive); |
| 975 | case NVPTX::BI__nvvm_add_rn_sat_f: |
| 976 | case NVPTX::BI__nvvm_add_rn_sat_f16: |
| 977 | case NVPTX::BI__nvvm_add_rn_sat_v2f16: |
| 978 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd.sat" , |
| 979 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 980 | case NVPTX::BI__nvvm_add_rz_sat_f: |
| 981 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd.sat" , |
| 982 | rm: llvm::APFloat::rmTowardZero); |
| 983 | case NVPTX::BI__nvvm_add_rm_sat_f: |
| 984 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd.sat" , |
| 985 | rm: llvm::APFloat::rmTowardNegative); |
| 986 | case NVPTX::BI__nvvm_add_rp_sat_f: |
| 987 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd.sat" , |
| 988 | rm: llvm::APFloat::rmTowardPositive); |
| 989 | case NVPTX::BI__nvvm_add_rn_ftz_sat_f: |
| 990 | case NVPTX::BI__nvvm_add_rn_ftz_sat_f16: |
| 991 | case NVPTX::BI__nvvm_add_rn_ftz_sat_v2f16: |
| 992 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd.ftz.sat" , |
| 993 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 994 | case NVPTX::BI__nvvm_add_rz_ftz_sat_f: |
| 995 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd.ftz.sat" , |
| 996 | rm: llvm::APFloat::rmTowardZero); |
| 997 | case NVPTX::BI__nvvm_add_rm_ftz_sat_f: |
| 998 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd.ftz.sat" , |
| 999 | rm: llvm::APFloat::rmTowardNegative); |
| 1000 | case NVPTX::BI__nvvm_add_rp_ftz_sat_f: |
| 1001 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fadd.ftz.sat" , |
| 1002 | rm: llvm::APFloat::rmTowardPositive); |
| 1003 | case NVPTX::BI__nvvm_mul_rn_f: |
| 1004 | case NVPTX::BI__nvvm_mul_rn_d: |
| 1005 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fmul" , |
| 1006 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 1007 | case NVPTX::BI__nvvm_mul_rz_f: |
| 1008 | case NVPTX::BI__nvvm_mul_rz_d: |
| 1009 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fmul" , |
| 1010 | rm: llvm::APFloat::rmTowardZero); |
| 1011 | case NVPTX::BI__nvvm_mul_rm_f: |
| 1012 | case NVPTX::BI__nvvm_mul_rm_d: |
| 1013 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fmul" , |
| 1014 | rm: llvm::APFloat::rmTowardNegative); |
| 1015 | case NVPTX::BI__nvvm_mul_rp_f: |
| 1016 | case NVPTX::BI__nvvm_mul_rp_d: |
| 1017 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fmul" , |
| 1018 | rm: llvm::APFloat::rmTowardPositive); |
| 1019 | case NVPTX::BI__nvvm_mul_rn_ftz_f: |
| 1020 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fmul.ftz" , |
| 1021 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 1022 | case NVPTX::BI__nvvm_mul_rz_ftz_f: |
| 1023 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fmul.ftz" , |
| 1024 | rm: llvm::APFloat::rmTowardZero); |
| 1025 | case NVPTX::BI__nvvm_mul_rm_ftz_f: |
| 1026 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fmul.ftz" , |
| 1027 | rm: llvm::APFloat::rmTowardNegative); |
| 1028 | case NVPTX::BI__nvvm_mul_rp_ftz_f: |
| 1029 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fmul.ftz" , |
| 1030 | rm: llvm::APFloat::rmTowardPositive); |
| 1031 | case NVPTX::BI__nvvm_mul_rn_sat_f16: |
| 1032 | case NVPTX::BI__nvvm_mul_rn_sat_v2f16: |
| 1033 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fmul.sat" , |
| 1034 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 1035 | case NVPTX::BI__nvvm_mul_rn_ftz_sat_f16: |
| 1036 | case NVPTX::BI__nvvm_mul_rn_ftz_sat_v2f16: |
| 1037 | return emitNVVMFPArith(cgf&: *this, expr, intrinsicName: "nvvm.fmul.ftz.sat" , |
| 1038 | rm: llvm::APFloat::rmNearestTiesToEven); |
| 1039 | case NVPTX::BI__nvvm_ldg_h: |
| 1040 | case NVPTX::BI__nvvm_ldg_h2: |
| 1041 | cgm.errorNYI(expr->getSourceRange(), |
| 1042 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1043 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1044 | return mlir::Value{}; |
| 1045 | case NVPTX::BI__nvvm_ldu_h: |
| 1046 | case NVPTX::BI__nvvm_ldu_h2: |
| 1047 | return makeLdu(cgf&: *this, expr, intrinsicName: "nvvm.ldu.global.f" ); |
| 1048 | case NVPTX::BI__nvvm_cp_async_ca_shared_global_4: |
| 1049 | cgm.errorNYI(expr->getSourceRange(), |
| 1050 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1051 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1052 | return mlir::Value{}; |
| 1053 | case NVPTX::BI__nvvm_cp_async_ca_shared_global_8: |
| 1054 | cgm.errorNYI(expr->getSourceRange(), |
| 1055 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1056 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1057 | return mlir::Value{}; |
| 1058 | case NVPTX::BI__nvvm_cp_async_ca_shared_global_16: |
| 1059 | cgm.errorNYI(expr->getSourceRange(), |
| 1060 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1061 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1062 | return mlir::Value{}; |
| 1063 | case NVPTX::BI__nvvm_cp_async_cg_shared_global_16: |
| 1064 | cgm.errorNYI(expr->getSourceRange(), |
| 1065 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1066 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1067 | return mlir::Value{}; |
| 1068 | case NVPTX::BI__nvvm_read_ptx_sreg_clusterid_x: |
| 1069 | cgm.errorNYI(expr->getSourceRange(), |
| 1070 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1071 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1072 | return mlir::Value{}; |
| 1073 | case NVPTX::BI__nvvm_read_ptx_sreg_clusterid_y: |
| 1074 | cgm.errorNYI(expr->getSourceRange(), |
| 1075 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1076 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1077 | return mlir::Value{}; |
| 1078 | case NVPTX::BI__nvvm_read_ptx_sreg_clusterid_z: |
| 1079 | cgm.errorNYI(expr->getSourceRange(), |
| 1080 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1081 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1082 | return mlir::Value{}; |
| 1083 | case NVPTX::BI__nvvm_read_ptx_sreg_clusterid_w: |
| 1084 | cgm.errorNYI(expr->getSourceRange(), |
| 1085 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1086 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1087 | return mlir::Value{}; |
| 1088 | case NVPTX::BI__nvvm_read_ptx_sreg_nclusterid_x: |
| 1089 | cgm.errorNYI(expr->getSourceRange(), |
| 1090 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1091 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1092 | return mlir::Value{}; |
| 1093 | case NVPTX::BI__nvvm_read_ptx_sreg_nclusterid_y: |
| 1094 | cgm.errorNYI(expr->getSourceRange(), |
| 1095 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1096 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1097 | return mlir::Value{}; |
| 1098 | case NVPTX::BI__nvvm_read_ptx_sreg_nclusterid_z: |
| 1099 | cgm.errorNYI(expr->getSourceRange(), |
| 1100 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1101 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1102 | return mlir::Value{}; |
| 1103 | case NVPTX::BI__nvvm_read_ptx_sreg_nclusterid_w: |
| 1104 | cgm.errorNYI(expr->getSourceRange(), |
| 1105 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1106 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1107 | return mlir::Value{}; |
| 1108 | case NVPTX::BI__nvvm_read_ptx_sreg_cluster_ctaid_x: |
| 1109 | cgm.errorNYI(expr->getSourceRange(), |
| 1110 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1111 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1112 | return mlir::Value{}; |
| 1113 | case NVPTX::BI__nvvm_read_ptx_sreg_cluster_ctaid_y: |
| 1114 | cgm.errorNYI(expr->getSourceRange(), |
| 1115 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1116 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1117 | return mlir::Value{}; |
| 1118 | case NVPTX::BI__nvvm_read_ptx_sreg_cluster_ctaid_z: |
| 1119 | cgm.errorNYI(expr->getSourceRange(), |
| 1120 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1121 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1122 | return mlir::Value{}; |
| 1123 | case NVPTX::BI__nvvm_read_ptx_sreg_cluster_ctaid_w: |
| 1124 | cgm.errorNYI(expr->getSourceRange(), |
| 1125 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1126 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1127 | return mlir::Value{}; |
| 1128 | case NVPTX::BI__nvvm_read_ptx_sreg_cluster_nctaid_x: |
| 1129 | cgm.errorNYI(expr->getSourceRange(), |
| 1130 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1131 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1132 | return mlir::Value{}; |
| 1133 | case NVPTX::BI__nvvm_read_ptx_sreg_cluster_nctaid_y: |
| 1134 | cgm.errorNYI(expr->getSourceRange(), |
| 1135 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1136 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1137 | return mlir::Value{}; |
| 1138 | case NVPTX::BI__nvvm_read_ptx_sreg_cluster_nctaid_z: |
| 1139 | cgm.errorNYI(expr->getSourceRange(), |
| 1140 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1141 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1142 | return mlir::Value{}; |
| 1143 | case NVPTX::BI__nvvm_read_ptx_sreg_cluster_nctaid_w: |
| 1144 | cgm.errorNYI(expr->getSourceRange(), |
| 1145 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1146 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1147 | return mlir::Value{}; |
| 1148 | case NVPTX::BI__nvvm_read_ptx_sreg_cluster_ctarank: |
| 1149 | cgm.errorNYI(expr->getSourceRange(), |
| 1150 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1151 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1152 | return mlir::Value{}; |
| 1153 | case NVPTX::BI__nvvm_read_ptx_sreg_cluster_nctarank: |
| 1154 | cgm.errorNYI(expr->getSourceRange(), |
| 1155 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1156 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1157 | return mlir::Value{}; |
| 1158 | case NVPTX::BI__nvvm_is_explicit_cluster: |
| 1159 | cgm.errorNYI(expr->getSourceRange(), |
| 1160 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1161 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1162 | return mlir::Value{}; |
| 1163 | case NVPTX::BI__nvvm_isspacep_shared_cluster: |
| 1164 | cgm.errorNYI(expr->getSourceRange(), |
| 1165 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1166 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1167 | return mlir::Value{}; |
| 1168 | case NVPTX::BI__nvvm_mapa: |
| 1169 | cgm.errorNYI(expr->getSourceRange(), |
| 1170 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1171 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1172 | return mlir::Value{}; |
| 1173 | case NVPTX::BI__nvvm_mapa_shared_cluster: |
| 1174 | cgm.errorNYI(expr->getSourceRange(), |
| 1175 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1176 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1177 | return mlir::Value{}; |
| 1178 | case NVPTX::BI__nvvm_getctarank: |
| 1179 | cgm.errorNYI(expr->getSourceRange(), |
| 1180 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1181 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1182 | return mlir::Value{}; |
| 1183 | case NVPTX::BI__nvvm_getctarank_shared_cluster: |
| 1184 | cgm.errorNYI(expr->getSourceRange(), |
| 1185 | std::string("unimplemented NVPTX builtin call: " ) + |
| 1186 | getContext().BuiltinInfo.getName(ID: builtinId)); |
| 1187 | return mlir::Value{}; |
| 1188 | case NVPTX::BI__nvvm_barrier_cluster_arrive: |
| 1189 | return builder.emitIntrinsicCallOp(loc: getLoc(srcLoc: expr->getExprLoc()), |
| 1190 | str: "nvvm.barrier.cluster.arrive" , |
| 1191 | resTy: builder.getVoidTy()); |
| 1192 | case NVPTX::BI__nvvm_barrier_cluster_arrive_relaxed: |
| 1193 | return builder.emitIntrinsicCallOp(loc: getLoc(srcLoc: expr->getExprLoc()), |
| 1194 | str: "nvvm.barrier.cluster.arrive.relaxed" , |
| 1195 | resTy: builder.getVoidTy()); |
| 1196 | case NVPTX::BI__nvvm_barrier_cluster_wait: |
| 1197 | return builder.emitIntrinsicCallOp(loc: getLoc(srcLoc: expr->getExprLoc()), |
| 1198 | str: "nvvm.barrier.cluster.wait" , |
| 1199 | resTy: builder.getVoidTy()); |
| 1200 | case NVPTX::BI__nvvm_fence_sc_cluster: |
| 1201 | return builder.emitIntrinsicCallOp(loc: getLoc(srcLoc: expr->getExprLoc()), |
| 1202 | str: "nvvm.fence.sc.cluster" , |
| 1203 | resTy: builder.getVoidTy()); |
| 1204 | case NVPTX::BI__nvvm_bar_sync: |
| 1205 | return builder.emitIntrinsicCallOp( |
| 1206 | loc: getLoc(srcLoc: expr->getExprLoc()), str: "nvvm.barrier.cta.sync.aligned.all" , |
| 1207 | resTy: builder.getVoidTy(), op: mlir::ValueRange{emitScalarExpr(e: expr->getArg(Arg: 0))}); |
| 1208 | case NVPTX::BI__syncthreads: |
| 1209 | return builder.emitIntrinsicCallOp( |
| 1210 | loc: getLoc(srcLoc: expr->getExprLoc()), str: "nvvm.barrier.cta.sync.aligned.all" , |
| 1211 | resTy: builder.getVoidTy(), |
| 1212 | op: mlir::ValueRange{builder.getConstInt(loc: getLoc(srcLoc: expr->getExprLoc()), |
| 1213 | t: builder.getSInt32Ty(), c: 0)}); |
| 1214 | case NVPTX::BI__nvvm_barrier_sync: |
| 1215 | return builder.emitIntrinsicCallOp( |
| 1216 | loc: getLoc(srcLoc: expr->getExprLoc()), str: "nvvm.barrier.cta.sync.all" , |
| 1217 | resTy: builder.getVoidTy(), op: mlir::ValueRange{emitScalarExpr(e: expr->getArg(Arg: 0))}); |
| 1218 | case NVPTX::BI__nvvm_barrier_sync_cnt: |
| 1219 | return builder.emitIntrinsicCallOp( |
| 1220 | loc: getLoc(srcLoc: expr->getExprLoc()), str: "nvvm.barrier.cta.sync.count" , |
| 1221 | resTy: builder.getVoidTy(), |
| 1222 | op: mlir::ValueRange{emitScalarExpr(e: expr->getArg(Arg: 0)), |
| 1223 | emitScalarExpr(e: expr->getArg(Arg: 1))}); |
| 1224 | case NVPTX::BI__nvvm_bar0_and: |
| 1225 | return emitBar0Reduction(cgf&: *this, expr, |
| 1226 | intrinsicName: "nvvm.barrier.cta.red.and.aligned.all" , |
| 1227 | /*returnsPred=*/true); |
| 1228 | case NVPTX::BI__nvvm_bar0_or: |
| 1229 | return emitBar0Reduction(cgf&: *this, expr, intrinsicName: "nvvm.barrier.cta.red.or.aligned.all" , |
| 1230 | /*returnsPred=*/true); |
| 1231 | case NVPTX::BI__nvvm_bar0_popc: |
| 1232 | return emitBar0Reduction(cgf&: *this, expr, |
| 1233 | intrinsicName: "nvvm.barrier.cta.red.popc.aligned.all" , |
| 1234 | /*returnsPred=*/false); |
| 1235 | |
| 1236 | default: |
| 1237 | return std::nullopt; |
| 1238 | } |
| 1239 | } |
| 1240 | |