[mlir][tosa] Add tosa.matmul and tosa.fully_connected lowering

[NatashaKnk]

Summary

Adds lowerings for matmul and fully_connected. Only supports 2D tensors for inputs and weights, and 1D tensors for bias.

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

public:
   using OpRewritePattern<SrcOp>::OpRewritePattern;
 
   LogicalResult matchAndRewrite(SrcOp op,
                                 PatternRewriter &rewriter) const final {
     return elementwiseMatchAndRewriteHelper(op, rewriter);
   }
 };
 
+class MatMulConverter : public OpConversionPattern<tosa::MatMulOp> {
+public:
+  using OpConversionPattern<tosa::MatMulOp>::OpConversionPattern;
+  LogicalResult
+  matchAndRewrite(tosa::MatMulOp op, ArrayRef<Value> args,
+                  ConversionPatternRewriter &rewriter) const final {
+    tosa::MatMulOp::Adaptor adaptor(args);
+
+    Location loc = op.getLoc();
+
+    auto outputTy = op.getType().cast<ShapedType>();
+    auto outputElementTy = outputTy.getElementType();
+    auto zero_attr = rewriter.getZeroAttr(outputElementTy);
+    Value zero = rewriter.create<ConstantOp>(loc, zero_attr);
+    auto initTensor = rewriter.create<linalg::InitTensorOp>(
+        loc, outputTy.getShape(), outputTy.getElementType());
+    Value zeroTensor =
+        rewriter.create<linalg::FillOp>(loc, initTensor, zero).getResult(0);
+    rewriter.replaceOpWithNewOp<linalg::MatmulOp>(
+        op, TypeRange{op.getType()}, ValueRange{adaptor.a(), adaptor.b()},
+        ValueRange{zeroTensor});
+    return success();
+  }
+};
+
+class FullyConnectedConverter
+    : public OpConversionPattern<tosa::FullyConnectedOp> {
+public:
+  using OpConversionPattern<tosa::FullyConnectedOp>::OpConversionPattern;
+  LogicalResult
+  matchAndRewrite(tosa::FullyConnectedOp op, ArrayRef<Value> args,
+                  ConversionPatternRewriter &rewriter) const final {
+    tosa::FullyConnectedOp::Adaptor adaptor(args);
+
+    Location loc = op.getLoc();
+    auto outputTy = op.getType().cast<ShapedType>();
+    auto biasTy = op->getOperand(2).getType().cast<ShapedType>();
+
+    // Reshaping the bias from n to [1, n] for broadcasting
+    SmallVector<int64_t> biasShapeReshaped;
+    biasShapeReshaped.push_back(1);
+    biasShapeReshaped.push_back(biasTy.getShape()[0]);
+
+    RankedTensorType reshapedBias =
+        RankedTensorType::get(biasShapeReshaped, outputTy.getElementType());
+    auto reshapeResult =
+        rewriter.create<tosa::ReshapeOp>(loc, reshapedBias, args[2])
+            ->getResult(0);
+
+    // Creating maps for the output of MatMul and the bias
+    SmallVector<AffineMap, 4> indexingMaps;
+    indexingMaps.push_back(createAffineMapForType(reshapedBias, rewriter));
+    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(outputTy.getRank()));
+
+    auto initTensor =
+        rewriter
+            .create<linalg::InitTensorOp>(loc, outputTy.getShape(),
+                                          outputTy.getElementType())
+            ->getResults();
+
+    auto linalgOp =
+        rewriter
+            .create<linalg::GenericOp>(
+                loc, outputTy, reshapeResult, initTensor, indexingMaps,
+                getNParallelLoopsAttrs(outputTy.getRank()),
+                [&](OpBuilder &nested_builder, Location nested_loc,
+                    ValueRange args) {
+                  nested_builder.create<linalg::YieldOp>(loc, *args.begin());
+                })
+            ->getResults();
+
+    rewriter.replaceOpWithNewOp<linalg::MatmulOp>(
+        op, TypeRange{op.getType()},
+        ValueRange{adaptor.input(), adaptor.weight()}, linalgOp);
+    return success();
+  }
+};
+
 class ReshapeConverter : public OpConversionPattern<tosa::ReshapeOp> {
 public:
   using OpConversionPattern<tosa::ReshapeOp>::OpConversionPattern;
 
   LogicalResult
   matchAndRewrite(tosa::ReshapeOp reshape, ArrayRef<Value> args,
                   ConversionPatternRewriter &rewriter) const final {
     typename tosa::ReshapeOp::Adaptor operands(args);
patterns->add<
       PointwiseConverter<tosa::GreaterEqualOp>,
       PointwiseConverter<tosa::MaximumOp>, PointwiseConverter<tosa::MinimumOp>,
       PointwiseConverter<tosa::CeilOp>, PointwiseConverter<tosa::FloorOp>,
       PointwiseConverter<tosa::ClampOp>, PointwiseConverter<tosa::ReluNOp>,
       IdentityNConverter<tosa::IdentityOp>,
       IdentityNConverter<tosa::IdentityNOp>, ReduceConverter<tosa::ReduceMinOp>,
       ReduceConverter<tosa::ReduceMaxOp>, ReduceConverter<tosa::ReduceSumOp>,
       ReduceConverter<tosa::ReduceProdOp>, ConcatConverter, ReshapeConverter,
-      RescaleConverter, ReverseConverter, TransposeConverter>(
-      patterns->getContext());
+      RescaleConverter, ReverseConverter, TransposeConverter, MatMulConverter,
+      FullyConnectedConverter>(patterns->getContext());
 }

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir

func @reverse(%arg0: tensor<5x4xi32>) -> () {
   // CHECK: [[INIT:%.+]] = linalg.init_tensor [5, 4]
   // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP2]], #[[$MAP1]]], iterator_types = ["parallel", "parallel"]} ins(%arg0 : tensor<5x4xi32>) outs([[INIT]] : tensor<5x4xi32>) {
   // CHECK: ^bb0(%arg1: i32, %arg2: i32):
   // CHECK:   linalg.yield %arg1 : i32
   %1 = "tosa.reverse"(%arg0) {axis = 1 : i64} : (tensor<5x4xi32>) -> tensor<5x4xi32>
 
   return
 }
+
+// -----
+
+
+// CHECK-LABEL: @matmul
+func @matmul(%arg0: tensor<5x3xf32>, %arg1: tensor<3x6xf32>, %arg2: tensor<6xf32>) -> (tensor<5x6xf32>) {
+  // CHECK: [[C0:%.+]] = constant 0
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [5, 6]
+  // CHECK: [[FILLED:%.+]] = linalg.fill([[INIT]], [[C0]]) : tensor<5x6xf32>, f32 -> tensor<5x6xf32>
+  // CHECK: linalg.matmul ins(%arg0, %arg1 : tensor<5x3xf32>, tensor<3x6xf32>) outs([[FILLED]] : tensor<5x6xf32>) -> tensor<5x6xf32>
+  %0 = "tosa.matmul"(%arg0, %arg1) : (tensor<5x3xf32>, tensor<3x6xf32>)  -> (tensor<5x6xf32>)
+  return %0 : tensor<5x6xf32>
+}
+
+// -----
+
+// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0, d1)>
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (0, d1)>
+
+// CHECK-LABEL: @fully_connected
+func @fully_connected(%arg0: tensor<5x3xf32>, %arg1: tensor<3x6xf32>, %arg2: tensor<6xf32>) -> (tensor<5x6xf32>) {
+  // CHECK: [[RS:%.+]] = linalg.tensor_reshape %arg2 [#[[$MAP0]]]
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [5, 6]
+  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP1]], #[[$MAP0]]], iterator_types = ["parallel", "parallel"]} ins([[RS]] : tensor<1x6xf32>) outs([[INIT]] : tensor<5x6xf32>) {
+  // CHECK: ^bb0([[IN:%.+]]: f32, [[MULTIPLIER:%.+]]: f32):
+  // CHECK: linalg.matmul ins(%arg0, %arg1 : tensor<5x3xf32>, tensor<3x6xf32>) outs([[GENERIC]] : tensor<5x6xf32>) -> tensor<5x6xf32>
+  %0 = "tosa.fully_connected"(%arg0, %arg1, %arg2) : (tensor<5x3xf32>, tensor<3x6xf32>, tensor<6xf32>)  -> (tensor<5x6xf32>)
+  return %0 : tensor<5x6xf32>
+}