1//===- ConversionUtils.cpp ------------------------------------------------===//
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// Utility functions for TOSA lowering
10//
11//===----------------------------------------------------------------------===//
12
13#include "mlir/Dialect/Tosa/Utils/ConversionUtils.h"
14#include "mlir/Dialect/Tosa/IR/TosaOps.h"
15
16using namespace mlir;
17using namespace mlir::tosa;
18
19SmallVector<utils::IteratorType>
20mlir::tosa::getNParallelLoopsAttrs(unsigned nParallelLoops) {
21 return SmallVector<utils::IteratorType>(nParallelLoops,
22 utils::IteratorType::parallel);
23}
24
25SmallVector<Value>
26mlir::tosa::condenseValues(const SmallVector<Value> &values) {
27 SmallVector<Value> condensedValues;
28 for (auto value : values)
29 if (value)
30 condensedValues.push_back(Elt: value);
31 return condensedValues;
32}
33
34Value mlir::tosa::clampFloatHelper(Location loc, Value arg, Value min,
35 Value max, OpBuilder &rewriter) {
36 Value minValue = rewriter.create<arith::MinimumFOp>(location: loc, args&: arg, args&: max);
37 return rewriter.create<arith::MaximumFOp>(location: loc, args&: minValue, args&: min);
38}
39
40Value mlir::tosa::clampIntHelper(Location loc, Value arg, Value min, Value max,
41 OpBuilder &rewriter, bool isUnsigned) {
42 if (isUnsigned) {
43 auto minOrArg = rewriter.create<arith::MaxUIOp>(location: loc, args&: min, args&: arg);
44 return rewriter.create<arith::MinUIOp>(location: loc, args&: max, args&: minOrArg);
45 }
46 auto minOrArg = rewriter.create<arith::MaxSIOp>(location: loc, args&: min, args&: arg);
47 return rewriter.create<arith::MinSIOp>(location: loc, args&: max, args&: minOrArg);
48}
49
50bool mlir::tosa::validIntegerRange(IntegerType ty, int64_t value) {
51 uint64_t bitwidth = ty.getIntOrFloatBitWidth();
52 if (ty.getSignedness() == IntegerType::Unsigned) {
53 uint64_t uvalue = value;
54 APInt intMin = APInt::getMinValue(numBits: bitwidth);
55 APInt intMax = APInt::getMaxValue(numBits: bitwidth);
56 return uvalue >= intMin.getZExtValue() && uvalue <= intMax.getZExtValue();
57 }
58
59 APInt intMin = APInt::getSignedMinValue(numBits: bitwidth);
60 APInt intMax = APInt::getSignedMaxValue(numBits: bitwidth);
61 return value >= intMin.getSExtValue() && value <= intMax.getSExtValue();
62}
63
64namespace {
65// Given two tensors of high and low ranks, derive the output shape
66// to reshape the lower rank to.
67// Examples:
68// If lower=[c], higher=[a, b, c], [c] reshaped into [1, 1, c].
69// If lower=[b, c], higher=[a, b, c], [b, c] reshaped into [1, b, c].
70// If lower=[a], higher=[a, a], [a] reshaped into [1, a].
71// If lower=[a], target=[a, b, a], [a] reshaped into [1, 1, a].
72// If lower=[], target=[a, b, c], [] reshaped into [1, 1, 1].
73LogicalResult
74computeReshapeOutput(ArrayRef<int64_t> higherRankShape,
75 ArrayRef<int64_t> lowerRankShape,
76 SmallVectorImpl<int64_t> &reshapeOutputShape) {
77 // Initialize new shapes with [1] * higherRank.
78 int64_t higherRank = higherRankShape.size();
79 int64_t lowerRank = lowerRankShape.size();
80 reshapeOutputShape.assign(NumElts: higherRank, Elt: 1);
81
82 int64_t higherRankDim;
83 int64_t lowerRankDim;
84 const int64_t rankDiff = higherRank - lowerRank;
85
86 for (int64_t i = lowerRank - 1; i >= 0; i--) {
87 higherRankDim = higherRankShape[i + rankDiff];
88 lowerRankDim = lowerRankShape[i];
89
90 if (lowerRankDim != 1 && higherRankDim != 1 &&
91 lowerRankDim != higherRankDim)
92 return failure();
93
94 reshapeOutputShape[i + rankDiff] = lowerRankDim == 1 ? 1 : lowerRankDim;
95 }
96 return success();
97}
98} // namespace
99
100LogicalResult mlir::tosa::EqualizeRanks(PatternRewriter &rewriter, Location loc,
101 Value &input1, Value &input2) {
102 ImplicitLocOpBuilder builder(loc, rewriter);
103 return EqualizeRanks(builder, input1, input2);
104}
105
106LogicalResult mlir::tosa::EqualizeRanks(ImplicitLocOpBuilder &builder,
107 Value &input1, Value &input2) {
108 auto input1Ty = llvm::dyn_cast<RankedTensorType>(Val: input1.getType());
109 auto input2Ty = llvm::dyn_cast<RankedTensorType>(Val: input2.getType());
110
111 if (!input1Ty || !input2Ty) {
112 return failure();
113 }
114
115 int64_t input1Rank = input1Ty.getRank();
116 int64_t input2Rank = input2Ty.getRank();
117
118 if (input1Rank == input2Rank)
119 return success();
120
121 Value higherTensorValue, lowerTensorValue;
122 if (input1Rank > input2Rank) {
123 higherTensorValue = input1;
124 lowerTensorValue = input2;
125 } else {
126 higherTensorValue = input2;
127 lowerTensorValue = input1;
128 }
129
130 ArrayRef<int64_t> higherRankShape =
131 llvm::cast<RankedTensorType>(Val: higherTensorValue.getType()).getShape();
132 ArrayRef<int64_t> lowerRankShape =
133 llvm::cast<RankedTensorType>(Val: lowerTensorValue.getType()).getShape();
134
135 SmallVector<int64_t, 4> reshapeOutputShape;
136
137 if (computeReshapeOutput(higherRankShape, lowerRankShape, reshapeOutputShape)
138 .failed())
139 return failure();
140
141 auto reshapeInputType =
142 llvm::cast<RankedTensorType>(Val: lowerTensorValue.getType());
143 auto reshapeOutputType = RankedTensorType::get(
144 shape: ArrayRef<int64_t>(reshapeOutputShape), elementType: reshapeInputType.getElementType());
145 auto reshapeOutputShapeValue = getTosaConstShape(builder, shape: reshapeOutputShape);
146
147 auto reshapeLower = builder.create<tosa::ReshapeOp>(
148 args&: reshapeOutputType, args&: lowerTensorValue, args&: reshapeOutputShapeValue);
149
150 if (input1Rank > input2Rank) {
151 input1 = higherTensorValue;
152 input2 = reshapeLower.getResult();
153 } else {
154 input1 = reshapeLower.getResult();
155 input2 = higherTensorValue;
156 }
157
158 return success();
159}
160
161Value mlir::tosa::getTosaConstShape(ImplicitLocOpBuilder &builder,
162 llvm::ArrayRef<int64_t> shape) {
163 auto attr = builder.getIndexTensorAttr(values: convertFromMlirShape(shape));
164 auto type = mlir::tosa::shapeType::get(context: builder.getContext(), rank: shape.size());
165 mlir::Operation *mlir_op = builder.create<tosa::ConstShapeOp>(args&: type, args&: attr);
166 return mlir_op->getResult(idx: 0);
167}
168
169Value mlir::tosa::getTosaConstShape(PatternRewriter &rewriter, Location loc,
170 llvm::ArrayRef<int64_t> shape) {
171 ImplicitLocOpBuilder builder(loc, rewriter);
172 return getTosaConstShape(builder, shape);
173}
174
175SmallVector<int64_t> mlir::tosa::convertFromMlirShape(ArrayRef<int64_t> shape) {
176 return to_vector(Range: llvm::map_range(C&: shape, F: [](int64_t dim) {
177 return ShapedType::isDynamic(dValue: dim) ? -1 : dim;
178 }));
179}
180
181bool mlir::tosa::getConstShapeValues(Operation *op,
182 llvm::SmallVector<int64_t> &result_shape) {
183 if (!op) {
184 return false;
185 }
186 if (auto constOp = mlir::dyn_cast<tosa::ConstShapeOp>(Val: op)) {
187 Attribute constOpAttr = constOp->getAttr(name: "values");
188 DenseElementsAttr elementsAttr = cast<DenseElementsAttr>(Val&: constOpAttr);
189 for (int i = 0; i < elementsAttr.size(); i++) {
190 int64_t val = elementsAttr.getValues<int64_t>()[i];
191 result_shape.push_back(Elt: val);
192 }
193 return true;
194 }
195 // for undefined op, return false.
196 return false;
197}
198
199// returns a small vector of int64_t values that attr contains
200SmallVector<int64_t>
201mlir::tosa::convertFromIntAttr(const DenseElementsAttr &attr, const int rank) {
202 if (attr.isSplat()) {
203 int64_t v = attr.getSplatValue<APInt>().getSExtValue();
204 return SmallVector<int64_t>(rank, v);
205 }
206
207 if (auto int_array_attr = llvm::dyn_cast<DenseIntElementsAttr>(Val: attr)) {
208 SmallVector<int64_t> vec;
209 for (APInt val : int_array_attr.getValues<APInt>()) {
210 vec.push_back(Elt: val.getSExtValue());
211 }
212 return vec;
213 }
214 return {};
215}
216
217bool mlir::tosa::hasUniqueConstantScatterIndices(
218 ShapedType indicesType, DenseIntElementsAttr indicesAttr) {
219 llvm::ArrayRef<int64_t> const indicesShape = indicesType.getShape();
220 const unsigned int indicesRank = indicesShape.size();
221 const unsigned int lastDimSize = indicesShape[indicesRank - 1];
222
223 // check each batch of indices from the flat indicesAttr values
224 // for duplicates
225 auto const indicesValues = indicesAttr.getValues<int32_t>();
226 assert(
227 (indicesValues.size() % lastDimSize == 0) &&
228 "Constant indices data length should be a multiple of indicesShape[-1]");
229
230 std::vector<uint64_t> indices(lastDimSize);
231 for (auto beg = indicesValues.begin(); beg < indicesValues.end();
232 beg += lastDimSize) {
233 std::copy(first: beg, last: beg + lastDimSize, result: indices.begin());
234 std::sort(first: indices.begin(), last: indices.end());
235 if (std::adjacent_find(first: indices.begin(), last: indices.end()) != indices.end()) {
236 // found duplicate values in indices in batch
237 return false;
238 }
239 }
240
241 return true;
242}
243

source code of mlir/lib/Dialect/Tosa/Utils/ConversionUtils.cpp