| 1 | //===- Utils.cpp - Utilities to support the Linalg dialect ----------------===// |
| 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 utilities for the Linalg dialect. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "mlir/Dialect/Arith/Utils/Utils.h" |
| 14 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 15 | #include "mlir/Dialect/Complex/IR/Complex.h" |
| 16 | #include "mlir/Dialect/Utils/StaticValueUtils.h" |
| 17 | #include "mlir/IR/ImplicitLocOpBuilder.h" |
| 18 | #include "llvm/ADT/SmallBitVector.h" |
| 19 | #include <numeric> |
| 20 | |
| 21 | using namespace mlir; |
| 22 | |
| 23 | std::optional<SmallVector<OpFoldResult>> |
| 24 | mlir::inferExpandShapeOutputShape(OpBuilder &b, Location loc, |
| 25 | ShapedType expandedType, |
| 26 | ArrayRef<ReassociationIndices> reassociation, |
| 27 | ArrayRef<OpFoldResult> inputShape) { |
| 28 | |
| 29 | SmallVector<Value> outputShapeValues; |
| 30 | SmallVector<int64_t> outputShapeInts; |
| 31 | // For zero-rank inputs, all dims in result shape are unit extent. |
| 32 | if (inputShape.empty()) { |
| 33 | outputShapeInts.resize(expandedType.getRank(), 1); |
| 34 | return getMixedValues(staticValues: outputShapeInts, dynamicValues: outputShapeValues, b); |
| 35 | } |
| 36 | |
| 37 | // Check for all static shapes. |
| 38 | if (expandedType.hasStaticShape()) { |
| 39 | ArrayRef<int64_t> staticShape = expandedType.getShape(); |
| 40 | outputShapeInts.assign(in_start: staticShape.begin(), in_end: staticShape.end()); |
| 41 | return getMixedValues(staticValues: outputShapeInts, dynamicValues: outputShapeValues, b); |
| 42 | } |
| 43 | |
| 44 | outputShapeInts.resize(expandedType.getRank(), ShapedType::kDynamic); |
| 45 | for (const auto &it : llvm::enumerate(First&: reassociation)) { |
| 46 | ReassociationIndices indexGroup = it.value(); |
| 47 | |
| 48 | int64_t indexGroupStaticSizesProductInt = 1; |
| 49 | bool foundDynamicShape = false; |
| 50 | for (int64_t index : indexGroup) { |
| 51 | int64_t outputDimSize = expandedType.getDimSize(index); |
| 52 | // Cannot infer expanded shape with multiple dynamic dims in the |
| 53 | // same reassociation group! |
| 54 | if (ShapedType::isDynamic(outputDimSize)) { |
| 55 | if (foundDynamicShape) |
| 56 | return std::nullopt; |
| 57 | foundDynamicShape = true; |
| 58 | } else { |
| 59 | outputShapeInts[index] = outputDimSize; |
| 60 | indexGroupStaticSizesProductInt *= outputDimSize; |
| 61 | } |
| 62 | } |
| 63 | if (!foundDynamicShape) |
| 64 | continue; |
| 65 | |
| 66 | int64_t inputIndex = it.index(); |
| 67 | // Call get<Value>() under the assumption that we're not casting |
| 68 | // dynamism. |
| 69 | Value indexGroupSize = cast<Value>(Val: inputShape[inputIndex]); |
| 70 | Value indexGroupStaticSizesProduct = |
| 71 | b.create<arith::ConstantIndexOp>(location: loc, args&: indexGroupStaticSizesProductInt); |
| 72 | Value dynamicDimSize = b.createOrFold<arith::DivSIOp>( |
| 73 | loc, indexGroupSize, indexGroupStaticSizesProduct); |
| 74 | outputShapeValues.push_back(Elt: dynamicDimSize); |
| 75 | } |
| 76 | |
| 77 | if ((int64_t)outputShapeValues.size() != |
| 78 | llvm::count(outputShapeInts, ShapedType::kDynamic)) |
| 79 | return std::nullopt; |
| 80 | |
| 81 | return getMixedValues(staticValues: outputShapeInts, dynamicValues: outputShapeValues, b); |
| 82 | } |
| 83 | |
| 84 | /// Matches a ConstantIndexOp. |
| 85 | /// TODO: This should probably just be a general matcher that uses matchConstant |
| 86 | /// and checks the operation for an index type. |
| 87 | detail::op_matcher<arith::ConstantIndexOp> mlir::matchConstantIndex() { |
| 88 | return detail::op_matcher<arith::ConstantIndexOp>(); |
| 89 | } |
| 90 | |
| 91 | llvm::SmallBitVector mlir::getPositionsOfShapeOne(unsigned rank, |
| 92 | ArrayRef<int64_t> shape) { |
| 93 | llvm::SmallBitVector dimsToProject(shape.size()); |
| 94 | for (unsigned pos = 0, e = shape.size(); pos < e && rank > 0; ++pos) { |
| 95 | if (shape[pos] == 1) { |
| 96 | dimsToProject.set(pos); |
| 97 | --rank; |
| 98 | } |
| 99 | } |
| 100 | return dimsToProject; |
| 101 | } |
| 102 | |
| 103 | Value mlir::getValueOrCreateConstantIntOp(OpBuilder &b, Location loc, |
| 104 | OpFoldResult ofr) { |
| 105 | if (auto value = dyn_cast_if_present<Value>(Val&: ofr)) |
| 106 | return value; |
| 107 | auto attr = cast<IntegerAttr>(cast<Attribute>(Val&: ofr)); |
| 108 | return b.create<arith::ConstantOp>( |
| 109 | loc, b.getIntegerAttr(attr.getType(), attr.getValue().getSExtValue())); |
| 110 | } |
| 111 | |
| 112 | Value mlir::getValueOrCreateConstantIndexOp(OpBuilder &b, Location loc, |
| 113 | OpFoldResult ofr) { |
| 114 | if (auto value = dyn_cast_if_present<Value>(Val&: ofr)) |
| 115 | return value; |
| 116 | auto attr = cast<IntegerAttr>(cast<Attribute>(Val&: ofr)); |
| 117 | return b.create<arith::ConstantIndexOp>(loc, attr.getValue().getSExtValue()); |
| 118 | } |
| 119 | |
| 120 | Value mlir::getValueOrCreateCastToIndexLike(OpBuilder &b, Location loc, |
| 121 | Type targetType, Value value) { |
| 122 | if (targetType == value.getType()) |
| 123 | return value; |
| 124 | |
| 125 | bool targetIsIndex = targetType.isIndex(); |
| 126 | bool valueIsIndex = value.getType().isIndex(); |
| 127 | if (targetIsIndex ^ valueIsIndex) |
| 128 | return b.create<arith::IndexCastOp>(loc, targetType, value); |
| 129 | |
| 130 | auto targetIntegerType = dyn_cast<IntegerType>(targetType); |
| 131 | auto valueIntegerType = dyn_cast<IntegerType>(value.getType()); |
| 132 | assert(targetIntegerType && valueIntegerType && |
| 133 | "unexpected cast between types other than integers and index" ); |
| 134 | assert(targetIntegerType.getSignedness() == valueIntegerType.getSignedness()); |
| 135 | |
| 136 | if (targetIntegerType.getWidth() > valueIntegerType.getWidth()) |
| 137 | return b.create<arith::ExtSIOp>(loc, targetIntegerType, value); |
| 138 | return b.create<arith::TruncIOp>(loc, targetIntegerType, value); |
| 139 | } |
| 140 | |
| 141 | static Value convertScalarToIntDtype(ImplicitLocOpBuilder &b, Value operand, |
| 142 | IntegerType toType, bool isUnsigned) { |
| 143 | // If operand is floating point, cast directly to the int type. |
| 144 | if (isa<FloatType>(Val: operand.getType())) { |
| 145 | if (isUnsigned) |
| 146 | return b.create<arith::FPToUIOp>(toType, operand); |
| 147 | return b.create<arith::FPToSIOp>(toType, operand); |
| 148 | } |
| 149 | // Cast index operands directly to the int type. |
| 150 | if (operand.getType().isIndex()) |
| 151 | return b.create<arith::IndexCastOp>(toType, operand); |
| 152 | if (auto fromIntType = dyn_cast<IntegerType>(operand.getType())) { |
| 153 | // Either extend or truncate. |
| 154 | if (toType.getWidth() > fromIntType.getWidth()) { |
| 155 | if (isUnsigned) |
| 156 | return b.create<arith::ExtUIOp>(toType, operand); |
| 157 | return b.create<arith::ExtSIOp>(toType, operand); |
| 158 | } |
| 159 | if (toType.getWidth() < fromIntType.getWidth()) |
| 160 | return b.create<arith::TruncIOp>(toType, operand); |
| 161 | return operand; |
| 162 | } |
| 163 | |
| 164 | return {}; |
| 165 | } |
| 166 | |
| 167 | static Value convertScalarToFpDtype(ImplicitLocOpBuilder &b, Value operand, |
| 168 | FloatType toType, bool isUnsigned) { |
| 169 | // If operand is integer, cast directly to the float type. |
| 170 | // Note that it is unclear how to cast from BF16<->FP16. |
| 171 | if (isa<IntegerType>(Val: operand.getType())) { |
| 172 | if (isUnsigned) |
| 173 | return b.create<arith::UIToFPOp>(toType, operand); |
| 174 | return b.create<arith::SIToFPOp>(toType, operand); |
| 175 | } |
| 176 | if (auto fromFpTy = dyn_cast<FloatType>(operand.getType())) { |
| 177 | if (toType.getWidth() > fromFpTy.getWidth()) |
| 178 | return b.create<arith::ExtFOp>(toType, operand); |
| 179 | if (toType.getWidth() < fromFpTy.getWidth()) |
| 180 | return b.create<arith::TruncFOp>(toType, operand); |
| 181 | return operand; |
| 182 | } |
| 183 | |
| 184 | return {}; |
| 185 | } |
| 186 | |
| 187 | static Value convertScalarToComplexDtype(ImplicitLocOpBuilder &b, Value operand, |
| 188 | ComplexType targetType, |
| 189 | bool isUnsigned) { |
| 190 | if (auto fromComplexType = dyn_cast<ComplexType>(operand.getType())) { |
| 191 | if (isa<FloatType>(targetType.getElementType()) && |
| 192 | isa<FloatType>(fromComplexType.getElementType())) { |
| 193 | Value real = b.create<complex::ReOp>(operand); |
| 194 | Value imag = b.create<complex::ImOp>(operand); |
| 195 | Type targetETy = targetType.getElementType(); |
| 196 | if (targetType.getElementType().getIntOrFloatBitWidth() < |
| 197 | fromComplexType.getElementType().getIntOrFloatBitWidth()) { |
| 198 | real = b.create<arith::TruncFOp>(targetETy, real); |
| 199 | imag = b.create<arith::TruncFOp>(targetETy, imag); |
| 200 | } else { |
| 201 | real = b.create<arith::ExtFOp>(targetETy, real); |
| 202 | imag = b.create<arith::ExtFOp>(targetETy, imag); |
| 203 | } |
| 204 | return b.create<complex::CreateOp>(targetType, real, imag); |
| 205 | } |
| 206 | } |
| 207 | |
| 208 | if (isa<FloatType>(Val: operand.getType())) { |
| 209 | FloatType toFpTy = cast<FloatType>(targetType.getElementType()); |
| 210 | auto toBitwidth = toFpTy.getIntOrFloatBitWidth(); |
| 211 | Value from = operand; |
| 212 | if (from.getType().getIntOrFloatBitWidth() < toBitwidth) { |
| 213 | from = b.create<arith::ExtFOp>(toFpTy, from); |
| 214 | } |
| 215 | if (from.getType().getIntOrFloatBitWidth() > toBitwidth) { |
| 216 | from = b.create<arith::TruncFOp>(toFpTy, from); |
| 217 | } |
| 218 | Value zero = b.create<mlir::arith::ConstantFloatOp>( |
| 219 | mlir::APFloat(toFpTy.getFloatSemantics(), 0), toFpTy); |
| 220 | return b.create<complex::CreateOp>(targetType, from, zero); |
| 221 | } |
| 222 | |
| 223 | if (isa<IntegerType>(Val: operand.getType())) { |
| 224 | FloatType toFpTy = cast<FloatType>(targetType.getElementType()); |
| 225 | Value from = operand; |
| 226 | if (isUnsigned) { |
| 227 | from = b.create<arith::UIToFPOp>(toFpTy, from); |
| 228 | } else { |
| 229 | from = b.create<arith::SIToFPOp>(toFpTy, from); |
| 230 | } |
| 231 | Value zero = b.create<mlir::arith::ConstantFloatOp>( |
| 232 | mlir::APFloat(toFpTy.getFloatSemantics(), 0), toFpTy); |
| 233 | return b.create<complex::CreateOp>(targetType, from, zero); |
| 234 | } |
| 235 | |
| 236 | return {}; |
| 237 | } |
| 238 | |
| 239 | Value mlir::convertScalarToDtype(OpBuilder &b, Location loc, Value operand, |
| 240 | Type toType, bool isUnsignedCast) { |
| 241 | if (operand.getType() == toType) |
| 242 | return operand; |
| 243 | ImplicitLocOpBuilder ib(loc, b); |
| 244 | Value result; |
| 245 | if (auto intTy = dyn_cast<IntegerType>(toType)) { |
| 246 | result = convertScalarToIntDtype(ib, operand, intTy, isUnsignedCast); |
| 247 | } else if (auto floatTy = dyn_cast<FloatType>(toType)) { |
| 248 | result = convertScalarToFpDtype(ib, operand, floatTy, isUnsignedCast); |
| 249 | } else if (auto complexTy = dyn_cast<ComplexType>(toType)) { |
| 250 | result = |
| 251 | convertScalarToComplexDtype(ib, operand, complexTy, isUnsignedCast); |
| 252 | } |
| 253 | |
| 254 | if (result) |
| 255 | return result; |
| 256 | |
| 257 | emitWarning(loc) << "could not cast operand of type " << operand.getType() |
| 258 | << " to " << toType; |
| 259 | return operand; |
| 260 | } |
| 261 | |
| 262 | SmallVector<Value> |
| 263 | mlir::getValueOrCreateConstantIndexOp(OpBuilder &b, Location loc, |
| 264 | ArrayRef<OpFoldResult> valueOrAttrVec) { |
| 265 | return llvm::to_vector<4>( |
| 266 | Range: llvm::map_range(C&: valueOrAttrVec, F: [&](OpFoldResult value) -> Value { |
| 267 | return getValueOrCreateConstantIndexOp(b, loc, ofr: value); |
| 268 | })); |
| 269 | } |
| 270 | |
| 271 | Value mlir::createScalarOrSplatConstant(OpBuilder &builder, Location loc, |
| 272 | Type type, const APInt &value) { |
| 273 | TypedAttr attr; |
| 274 | if (isa<IntegerType>(Val: type)) { |
| 275 | attr = builder.getIntegerAttr(type, value); |
| 276 | } else { |
| 277 | auto vecTy = cast<ShapedType>(type); |
| 278 | attr = SplatElementsAttr::get(vecTy, value); |
| 279 | } |
| 280 | |
| 281 | return builder.create<arith::ConstantOp>(loc, attr); |
| 282 | } |
| 283 | |
| 284 | Value mlir::createScalarOrSplatConstant(OpBuilder &builder, Location loc, |
| 285 | Type type, int64_t value) { |
| 286 | unsigned elementBitWidth = 0; |
| 287 | if (auto intTy = dyn_cast<IntegerType>(type)) |
| 288 | elementBitWidth = intTy.getWidth(); |
| 289 | else |
| 290 | elementBitWidth = cast<ShapedType>(type).getElementTypeBitWidth(); |
| 291 | |
| 292 | return createScalarOrSplatConstant(builder, loc, type, |
| 293 | value: APInt(elementBitWidth, value)); |
| 294 | } |
| 295 | |
| 296 | Value mlir::createScalarOrSplatConstant(OpBuilder &builder, Location loc, |
| 297 | Type type, const APFloat &value) { |
| 298 | if (isa<FloatType>(type)) |
| 299 | return builder.createOrFold<arith::ConstantOp>( |
| 300 | loc, type, builder.getFloatAttr(type, value)); |
| 301 | TypedAttr splat = SplatElementsAttr::get(cast<ShapedType>(type), value); |
| 302 | return builder.createOrFold<arith::ConstantOp>(loc, type, splat); |
| 303 | } |
| 304 | |
| 305 | Type mlir::getType(OpFoldResult ofr) { |
| 306 | if (auto value = dyn_cast_if_present<Value>(Val&: ofr)) |
| 307 | return value.getType(); |
| 308 | auto attr = cast<IntegerAttr>(cast<Attribute>(Val&: ofr)); |
| 309 | return attr.getType(); |
| 310 | } |
| 311 | |
| 312 | Value ArithBuilder::_and(Value lhs, Value rhs) { |
| 313 | return b.create<arith::AndIOp>(loc, lhs, rhs); |
| 314 | } |
| 315 | Value ArithBuilder::add(Value lhs, Value rhs) { |
| 316 | if (isa<FloatType>(lhs.getType())) |
| 317 | return b.create<arith::AddFOp>(loc, lhs, rhs); |
| 318 | return b.create<arith::AddIOp>(loc, lhs, rhs, ovf); |
| 319 | } |
| 320 | Value ArithBuilder::sub(Value lhs, Value rhs) { |
| 321 | if (isa<FloatType>(lhs.getType())) |
| 322 | return b.create<arith::SubFOp>(loc, lhs, rhs); |
| 323 | return b.create<arith::SubIOp>(loc, lhs, rhs, ovf); |
| 324 | } |
| 325 | Value ArithBuilder::mul(Value lhs, Value rhs) { |
| 326 | if (isa<FloatType>(lhs.getType())) |
| 327 | return b.create<arith::MulFOp>(loc, lhs, rhs); |
| 328 | return b.create<arith::MulIOp>(loc, lhs, rhs, ovf); |
| 329 | } |
| 330 | Value ArithBuilder::sgt(Value lhs, Value rhs) { |
| 331 | if (isa<FloatType>(lhs.getType())) |
| 332 | return b.create<arith::CmpFOp>(loc, arith::CmpFPredicate::OGT, lhs, rhs); |
| 333 | return b.create<arith::CmpIOp>(loc, arith::CmpIPredicate::sgt, lhs, rhs); |
| 334 | } |
| 335 | Value ArithBuilder::slt(Value lhs, Value rhs) { |
| 336 | if (isa<FloatType>(lhs.getType())) |
| 337 | return b.create<arith::CmpFOp>(loc, arith::CmpFPredicate::OLT, lhs, rhs); |
| 338 | return b.create<arith::CmpIOp>(loc, arith::CmpIPredicate::slt, lhs, rhs); |
| 339 | } |
| 340 | Value ArithBuilder::select(Value cmp, Value lhs, Value rhs) { |
| 341 | return b.create<arith::SelectOp>(loc, cmp, lhs, rhs); |
| 342 | } |
| 343 | |
| 344 | namespace mlir::arith { |
| 345 | |
| 346 | Value createProduct(OpBuilder &builder, Location loc, ArrayRef<Value> values) { |
| 347 | return createProduct(builder, loc, values, resultType: values.front().getType()); |
| 348 | } |
| 349 | |
| 350 | Value createProduct(OpBuilder &builder, Location loc, ArrayRef<Value> values, |
| 351 | Type resultType) { |
| 352 | Value one = builder.create<ConstantOp>(loc, resultType, |
| 353 | builder.getOneAttr(resultType)); |
| 354 | ArithBuilder arithBuilder(builder, loc); |
| 355 | return std::accumulate( |
| 356 | first: values.begin(), last: values.end(), init: one, |
| 357 | binary_op: [&arithBuilder](Value acc, Value v) { return arithBuilder.mul(acc, v); }); |
| 358 | } |
| 359 | |
| 360 | /// Map strings to float types. |
| 361 | std::optional<FloatType> parseFloatType(MLIRContext *ctx, StringRef name) { |
| 362 | Builder b(ctx); |
| 363 | return llvm::StringSwitch<std::optional<FloatType>>(name) |
| 364 | .Case("f4E2M1FN" , b.getType<Float4E2M1FNType>()) |
| 365 | .Case("f6E2M3FN" , b.getType<Float6E2M3FNType>()) |
| 366 | .Case("f6E3M2FN" , b.getType<Float6E3M2FNType>()) |
| 367 | .Case("f8E5M2" , b.getType<Float8E5M2Type>()) |
| 368 | .Case("f8E4M3" , b.getType<Float8E4M3Type>()) |
| 369 | .Case("f8E4M3FN" , b.getType<Float8E4M3FNType>()) |
| 370 | .Case("f8E5M2FNUZ" , b.getType<Float8E5M2FNUZType>()) |
| 371 | .Case("f8E4M3FNUZ" , b.getType<Float8E4M3FNUZType>()) |
| 372 | .Case("f8E3M4" , b.getType<Float8E3M4Type>()) |
| 373 | .Case("f8E8M0FNU" , b.getType<Float8E8M0FNUType>()) |
| 374 | .Case("bf16" , b.getType<BFloat16Type>()) |
| 375 | .Case("f16" , b.getType<Float16Type>()) |
| 376 | .Case("f32" , b.getType<Float32Type>()) |
| 377 | .Case("f64" , b.getType<Float64Type>()) |
| 378 | .Case("f80" , b.getType<Float80Type>()) |
| 379 | .Case("f128" , b.getType<Float128Type>()) |
| 380 | .Default(std::nullopt); |
| 381 | } |
| 382 | |
| 383 | } // namespace mlir::arith |
| 384 | |