[MLIR] Add support for Relu (#392)

* Add support for Relu

* Add comments

src/compiler/dialect/onnx/gen_doc.py

@@ -263,7 +263,7 @@ def  collect_types(schema, input) :
     return allowedTypeStr
 
 def gen_schema(schema) :
-    ShapeInferenceList=['Exp', 'Tanh', 'Sinh', 'Cosh', 'Sigmoid',
+    ShapeInferenceList=['Exp', 'Tanh', 'Sinh', 'Cosh', 'Sigmoid', 'Relu',
                         'Add', 'Mul', 'Div', 'Sub', 'And', 'Or', 'Xor',
                         'MatMul', 'Gemm']
     CanonicalList=['Add', 'Identity']

src/compiler/dialect/onnx/onnx_ops.cpp

@@ -78,6 +78,14 @@ void ONNXSigmoidOp::inferShapes() {
   getResult()->setType(getOperand()->getType());
 }
 
+//===----------------------------------------------------------------------===//
+// Relu
+/// Infer the output shape of the ONNXReluOp. This method is required by the
+/// shape inference interface.
+void ONNXReluOp::inferShapes() {
+  getResult()->setType(getOperand()->getType());
+}
+
 //===----------------------------------------------------------------------===//
 // Add
 /// Infer the output shape of the ONNXAddOp. This method is required by the

src/compiler/dialect/onnx/onnxop.inc

@@ -2185,7 +2185,7 @@ def ONNXReduceSumSquareOp:ONNX_Op<"ReduceSumSquare",
 }
 
 def ONNXReluOp:ONNX_Op<"Relu", 
-    [NoSideEffect]> {
+    [NoSideEffect, DeclareOpInterfaceMethods<ShapeInferenceOpInterface>]> {
   let summary = "ONNX Relu operation";
   let description = [{
     "Relu takes one input data (Tensor<T>) and produces one output data"

src/compiler/pass/lower_frontend_to_krnl.cpp

@@ -251,6 +251,23 @@ Value* mapToLowerScalarOp<ONNXSigmoidOp>(Location loc,
   return result;
 }
 
+//===----------------------------------------------------------------------===//
+// Scalar unary ops for lowering ONNXReluOp
+//===----------------------------------------------------------------------===//
+template <>
+Value* mapToLowerScalarOp<ONNXReluOp>(Location loc, ArrayRef<Type> result_types,
+    ArrayRef<Value*> operands, ConversionPatternRewriter& rewriter) {
+  // ONNXReluOp(%X) = SelectOp(CmpFOp(OLT, %X, ConstantOp 0),
+  //                           ConstantOp 0,
+  //                           %X)
+  Value* operand = operands[0];
+  auto zero = rewriter.create<ConstantOp>(loc, rewriter.getF32FloatAttr(0.0f));
+  auto lessThanZero =
+      rewriter.create<CmpFOp>(loc, CmpFPredicate::OLT, operand, zero);
+  auto result = rewriter.create<SelectOp>(loc, lessThanZero, zero, operand);
+  return result;
+}
+
 //===----------------------------------------------------------------------===//
 // Element-wise n-ary ops lowering to Krnl dialect.
 //===----------------------------------------------------------------------===//
@@ -452,6 +469,7 @@ void FrontendToKrnlLoweringPass::runOnModule() {
       ONNXElementwiseUnaryOpLowering<mlir::ONNXSinhOp>,
       ONNXElementwiseUnaryOpLowering<mlir::ONNXCoshOp>,
       ONNXElementwiseUnaryOpLowering<mlir::ONNXSigmoidOp>,
+      ONNXElementwiseUnaryOpLowering<mlir::ONNXReluOp>,
       ONNXElementwiseBinaryOpLowering<mlir::ONNXAddOp>,
       ONNXElementwiseBinaryOpLowering<mlir::ONNXMulOp>,
       ONNXElementwiseBinaryOpLowering<mlir::ONNXDivOp>,

src/compiler/pass/shape_inference_pass.cpp

@@ -89,20 +89,21 @@ class ShapeInferencePass : public mlir::FunctionPass<ShapeInferencePass> {
     // shaped outputs. All those operation need to implement the inferShape()
     // method.
     if (op->getName().getStringRef() != "onnx.Exp" &&
-	op->getName().getStringRef() != "onnx.Tanh" &&
+        op->getName().getStringRef() != "onnx.Tanh" &&
-	op->getName().getStringRef() != "onnx.Sinh" &&
+        op->getName().getStringRef() != "onnx.Sinh" &&
-	op->getName().getStringRef() != "onnx.Cosh" &&
+        op->getName().getStringRef() != "onnx.Cosh" &&
-	op->getName().getStringRef() != "onnx.Sigmoid" &&
+        op->getName().getStringRef() != "onnx.Sigmoid" &&
-	op->getName().getStringRef() != "onnx.Mul" &&
+        op->getName().getStringRef() != "onnx.Relu" &&
-	op->getName().getStringRef() != "onnx.Add" &&
+        op->getName().getStringRef() != "onnx.Mul" &&
-	op->getName().getStringRef() != "onnx.Div" &&
+        op->getName().getStringRef() != "onnx.Add" &&
-	op->getName().getStringRef() != "onnx.Sub" &&
+        op->getName().getStringRef() != "onnx.Div" &&
-	op->getName().getStringRef() != "onnx.And" &&
+        op->getName().getStringRef() != "onnx.Sub" &&
-	op->getName().getStringRef() != "onnx.Or" &&
+        op->getName().getStringRef() != "onnx.And" &&
-	op->getName().getStringRef() != "onnx.Xor" &&
+        op->getName().getStringRef() != "onnx.Or" &&
-	op->getName().getStringRef() != "onnx.MatMul" &&
+        op->getName().getStringRef() != "onnx.Xor" &&
-	op->getName().getStringRef() != "onnx.Gemm" &&
+        op->getName().getStringRef() != "onnx.MatMul" &&
-	op->getName().getStringRef() != "onnx.FullGemm")
+        op->getName().getStringRef() != "onnx.Gemm" &&
+        op->getName().getStringRef() != "onnx.FullGemm")
       return false;
     return llvm::any_of(op->getResultTypes(),
         [](Type result_type) { return !result_type.isa<RankedTensorType>(); });
@@ -118,4 +119,4 @@ std::unique_ptr<mlir::Pass> mlir::createShapeInferencePass() {
 }
 
 static PassRegistration<ShapeInferencePass> pass(
-     "shape-inference", "Shape inference for frontend dialects.");
+    "shape-inference", "Shape inference for frontend dialects.");

test/mlir/onnx/onnx_lowering.mlir

@@ -256,4 +256,25 @@ func @test_sigmoid(%arg0 : tensor<?x10xf32>) -> tensor<*xf32> {
   // CHECK: [[SIGMOID_RES:%.+]] = divf [[ONE]], [[DIVISOR]] : f32
   // CHECK: store [[SIGMOID_RES]], [[RES]][%arg1, %arg2] : memref<?x10xf32>
   // CHECK: return [[RES]] : memref<?x10xf32>
-}
+}
+
+func @test_relu(%arg0 : tensor<?x10xf32>) -> tensor<*xf32> {
+  %0 = "onnx.Relu"(%arg0) : (tensor<?x10xf32>) -> tensor<*xf32>
+  "std.return"(%0) : (tensor<*xf32>) -> ()
+
+  // CHECK-LABEL: test_relu
+  // CHECK: [[DIM_0:%.+]] = dim %arg0, 0 : memref<?x10xf32>
+  // CHECK: [[RES:%.+]] = alloc([[DIM_0]]) : memref<?x10xf32>
+  // CHECK: [[DEF_LOOPS:%.+]]:2 = krnl.define_loops 2
+  // CHECK: [[OPT_LOOPS:%.+]]:2 = krnl.optimize_loops  {
+  // CHECK:   krnl.return_loops [[DEF_LOOPS]]#0, [[DEF_LOOPS]]#1
+  // CHECK: } : () -> (!krnl.loop, !krnl.loop)
+  // CHECK: [[DIM_2:%.+]] = dim %arg0, 0 : memref<?x10xf32>
+  // CHECK: krnl.iterate([[OPT_LOOPS]]#0, [[OPT_LOOPS]]#1) with ([[DEF_LOOPS]]#0 -> %arg1 = 0 to [[DIM_2]], [[DEF_LOOPS]]#1 -> %arg2 = 0 to 10) {
+  // CHECK: [[LOAD:%.+]] = load %arg0[%arg1, %arg2] : memref<?x10xf32>
+  // CHECK: [[ZERO:%.+]] = constant {{0.+}} : f32
+  // CHECK: [[LTZERO:%.+]] = cmpf "olt", [[LOAD]], [[ZERO]] : f32
+  // CHECK: [[RELU_RES:%.+]] = select [[LTZERO]], [[ZERO]], [[LOAD]] : f32
+  // CHECK: store [[RELU_RES]], [[RES]][%arg1, %arg2] : memref<?x10xf32>
+  // CHECK: return [[RES]] : memref<?x10xf32>
+}

test/mlir/onnx/onnx_lowering_with_dealloc.mlir

@@ -527,4 +527,47 @@ func @test_sigmoid_sigmoid(%arg0 : tensor<?x10xf32>) -> tensor<*xf32> {
   // CHECK-NOT: dealloc [[RET_RES]] : memref<?x10xf32>
 
   // CHECK: return [[RET_RES]] : memref<?x10xf32>
-}
+}
+
+func @test_relu_relu(%arg0 : tensor<?x10xf32>) -> tensor<*xf32> {
+  %0 = "onnx.Relu"(%arg0) : (tensor<?x10xf32>) -> tensor<*xf32>
+  %1 = "onnx.Relu"(%0) : (tensor<*xf32>) -> tensor<*xf32>
+  "std.return"(%1) : (tensor<*xf32>) -> ()
+
+  // CHECK-LABEL: test_relu_relu
+  /// First Relu
+  // CHECK: [[DIM_0:%.+]] = dim %arg0, 0 : memref<?x10xf32>
+  // CHECK: [[RES:%.+]] = alloc([[DIM_0]]) : memref<?x10xf32>
+  // CHECK: [[DEF_LOOPS:%.+]]:2 = krnl.define_loops 2
+  // CHECK: [[OPT_LOOPS:%.+]]:2 = krnl.optimize_loops  {
+  // CHECK:   krnl.return_loops [[DEF_LOOPS]]#0, [[DEF_LOOPS]]#1
+  // CHECK: } : () -> (!krnl.loop, !krnl.loop)
+  // CHECK: [[DIM_2:%.+]] = dim %arg0, 0 : memref<?x10xf32>
+  // CHECK: krnl.iterate([[OPT_LOOPS]]#0, [[OPT_LOOPS]]#1) with ([[DEF_LOOPS]]#0 -> %arg1 = 0 to [[DIM_2]], [[DEF_LOOPS]]#1 -> %arg2 = 0 to 10) {
+  // CHECK: [[LOAD:%.+]] = load %arg0[%arg1, %arg2] : memref<?x10xf32>
+  // CHECK: [[ZERO:%.+]] = constant {{0.+}} : f32
+  // CHECK: [[LTZERO:%.+]] = cmpf "olt", [[LOAD]], [[ZERO]] : f32
+  // CHECK: [[RELU_RES:%.+]] = select [[LTZERO]], [[ZERO]], [[LOAD]] : f32
+  // CHECK: store [[RELU_RES]], [[RES]][%arg1, %arg2] : memref<?x10xf32>
+
+  /// Second Relu
+  // CHECK: [[DIM_0:%.+]] = dim [[RES]], 0 : memref<?x10xf32>
+  // CHECK: [[RET_RES:%.+]] = alloc([[DIM_0]]) : memref<?x10xf32>
+  // CHECK: [[DEF_LOOPS:%.+]]:2 = krnl.define_loops 2
+  // CHECK: [[OPT_LOOPS:%.+]]:2 = krnl.optimize_loops  {
+  // CHECK:   krnl.return_loops [[DEF_LOOPS]]#0, [[DEF_LOOPS]]#1
+  // CHECK: } : () -> (!krnl.loop, !krnl.loop)
+  // CHECK: [[DIM_2:%.+]] = dim [[RES]], 0 : memref<?x10xf32>
+  // CHECK: krnl.iterate([[OPT_LOOPS]]#0, [[OPT_LOOPS]]#1) with ([[DEF_LOOPS]]#0 -> %arg1 = 0 to [[DIM_2]], [[DEF_LOOPS]]#1 -> %arg2 = 0 to 10) {
+  // CHECK: [[LOAD:%.+]] = load [[RES]][%arg1, %arg2] : memref<?x10xf32>
+  // CHECK: [[ZERO:%.+]] = constant {{0.+}} : f32
+  // CHECK: [[LTZERO:%.+]] = cmpf "olt", [[LOAD]], [[ZERO]] : f32
+  // CHECK: [[RELU_RES:%.+]] = select [[LTZERO]], [[ZERO]], [[LOAD]] : f32
+  // CHECK: store [[RELU_RES]], [[RET_RES]][%arg1, %arg2] : memref<?x10xf32>
+
+  /// Dealloc of first result.
+  // CHECK: dealloc [[RES]] : memref<?x10xf32>
+  // CHECK-NOT: dealloc [[RET_RES]] : memref<?x10xf32>
+
+  // CHECK: return [[RET_RES]] : memref<?x10xf32>
+}