onnx · tungld · Feb 13, 2026 · Feb 13, 2026 · Feb 13, 2026 · Feb 13, 2026
diff --git a/src/Accelerators/NNPA/Conversion/ONNXToZHigh/ONNXToZHigh.cpp b/src/Accelerators/NNPA/Conversion/ONNXToZHigh/ONNXToZHigh.cpp
@@ -1686,7 +1686,6 @@ void ONNXToZHighLoweringPass::runOnOperation() {
   // prohibit the combined ops lowering happened.
   RewritePatternSet combinedPatterns(&getContext());
   onnx_mlir::getONNXToZHighMultipleOpPatterns(combinedPatterns);
-
   // It's ok to fail.
   (void)applyPatternsGreedily(module, std::move(combinedPatterns));
 

diff --git a/src/Accelerators/NNPA/Conversion/ONNXToZHigh/ONNXToZHigh.td b/src/Accelerators/NNPA/Conversion/ONNXToZHigh/ONNXToZHigh.td
@@ -344,21 +344,22 @@ def replaceONNXMaxPattern : Pat<
 //                                 (ZHighInvSqrtOp (ZHighStickOp %X))
 //===----------------------------------------------------------------------===//
 def replaceDiv1SqrtPattern : Pat<
-  (ONNXDivOp $a, (ONNXSqrtOp $x)),
+  (ONNXDivOp:$divOp $a, (ONNXSqrtOp:$sqrtOp $x)),
   (ZHighUnstickOp (ZHighInvSqrtOp (ZHighStickOp:$s_x $x, (NoneLayoutAttr), (GetDefaultSaturation)),
                                (returnType $s_x))),
-  [(IsCompatibleWithNNPALevelArch15),(IsConstOfOnes:$a),(IsFloatType:$a)]
+  [(IsCompatibleWithNNPALevelArch15),(IsConstOfOnes:$a),(IsFloatType:$a),
+   (OpIsBigEnoughForNNPA:$sqrtOp),(OpIsBigEnoughForNNPA:$divOp)]
 >;
 
 //===----------------------------------------------------------------------===//
 // ONNXReciprocalOp(ONNXSqrtOp %X) = ZHighUnstickOp
 //                                   (ZHighInvSqrtOp (ZHighStickOp %X))
 //===----------------------------------------------------------------------===//
 def replaceReciprocalSqrtPattern : Pat<
-  (ONNXReciprocalOp (ONNXSqrtOp $x)),
+  (ONNXReciprocalOp (ONNXSqrtOp:$sqrtOp $x)),
   (ZHighUnstickOp (ZHighInvSqrtOp (ZHighStickOp:$s_x $x, (NoneLayoutAttr), (GetDefaultSaturation)),
                                    (returnType $s_x))),
-  [(IsCompatibleWithNNPALevelArch15)]
+  [(IsCompatibleWithNNPALevelArch15), (OpIsBigEnoughForNNPA:$sqrtOp)]
 >;
 
 //===----------------------------------------------------------------------===//

diff --git a/src/Accelerators/NNPA/Conversion/ONNXToZHigh/ONNXToZHighCommon.cpp b/src/Accelerators/NNPA/Conversion/ONNXToZHigh/ONNXToZHighCommon.cpp
@@ -129,4 +129,13 @@ void configureONNXToZHighLoweringPass(bool optReportNNPAUnsupportedOps,
   }
 }
 
+bool isTooSmallOpForNNPA(mlir::Operation *op) {
+  if (!op)
+    return false;
+  // Operation whose all inputs are scalar.
+  bool scalarOp = llvm::all_of(
+      op->getOperands(), [](mlir::Value v) { return isScalarTensor(v); });
+  return scalarOp;
+}
+
 } // namespace onnx_mlir
diff --git a/src/Accelerators/NNPA/Conversion/ONNXToZHigh/ONNXToZHighCommon.hpp b/src/Accelerators/NNPA/Conversion/ONNXToZHigh/ONNXToZHighCommon.hpp
@@ -43,6 +43,36 @@ struct ONNXToZHighLoweringConfiguration {
   };
 };
 
+/// Get transposed tensor by using a permutation array.
+mlir::Value emitONNXTranspose(mlir::Location loc,
+    mlir::PatternRewriter &rewriter, mlir::Value x,
+    mlir::ArrayRef<int64_t> perms);
+
+/// Get transposed tensor by using a permutation array and a result type.
+mlir::Value emitONNXTransposeWithType(mlir::Location loc,
+    mlir::PatternRewriter &rewriter, mlir::Type transposedType, mlir::Value x,
+    mlir::ArrayRef<int64_t> perms);
+
+/// Split a tensor along an axis in which each chunk has a size of
+/// NNPAGetMaxForDim and the last chuck can be smaller.
+mlir::ValueRange splitAlongAxis(
+    onnx_mlir::MultiDialectBuilder<onnx_mlir::OnnxBuilder> &create,
+    mlir::Value X, int64_t axis);
+
+// Check if a value is a constant tensor of a single f32 value or not.
+bool isF32ScalarConstantTensor(mlir::Value v);
+
+// Get FloatAttr from a constant tensor of a single f32 value.
+mlir::FloatAttr getScalarF32AttrFromConstant(mlir::Value v);
+
+// Emit ONNX Concat to store the shape of the input x.
+mlir::Value getDynShape(
+    mlir::Location loc, mlir::PatternRewriter &rewriter, mlir::Value x);
+
+// Check if an operation is too small for NNPA, e.g. operation witl all scalar
+// tensors.
+bool isTooSmallOpForNNPA(mlir::Operation *op);
+
 template <typename OP_TYPE>
 void addDynamicallyLegalOpFor(mlir::ConversionTarget *target,
     const onnx_mlir::DimAnalysis *dimAnalysis,
@@ -66,6 +96,11 @@ void addDynamicallyLegalOpFor(mlir::ConversionTarget *target,
 
     // If not CPU, check if the op is legal for NNPA.
     bool isLegalForNNPA = false;
+
+    // Operations whose inputs are small, do not use NNPA.
+    if (isTooSmallOpForNNPA(genericOp))
+      return true;
+
     if (checkLegalityFn)
       isLegalForNNPA = !checkLegalityFn(op, dimAnalysis);
     else {
@@ -97,31 +132,5 @@ void addDynamicallyLegalOpFor(mlir::ConversionTarget *target,
   });
 }
 
-/// Get transposed tensor by using a permutation array.
-mlir::Value emitONNXTranspose(mlir::Location loc,
-    mlir::PatternRewriter &rewriter, mlir::Value x,
-    mlir::ArrayRef<int64_t> perms);
-
-/// Get transposed tensor by using a permutation array and a result type.
-mlir::Value emitONNXTransposeWithType(mlir::Location loc,
-    mlir::PatternRewriter &rewriter, mlir::Type transposedType, mlir::Value x,
-    mlir::ArrayRef<int64_t> perms);
-
-/// Split a tensor along an axis in which each chunk has a size of
-/// NNPAGetMaxForDim and the last chuck can be smaller.
-mlir::ValueRange splitAlongAxis(
-    onnx_mlir::MultiDialectBuilder<onnx_mlir::OnnxBuilder> &create,
-    mlir::Value X, int64_t axis);
-
-// Check if a value is a constant tensor of a single f32 value or not.
-bool isF32ScalarConstantTensor(mlir::Value v);
-
-// Get FloatAttr from a constant tensor of a single f32 value.
-mlir::FloatAttr getScalarF32AttrFromConstant(mlir::Value v);
-
-// Emit ONNX Concat to store the shape of the input x.
-mlir::Value getDynShape(
-    mlir::Location loc, mlir::PatternRewriter &rewriter, mlir::Value x);
-
 } // namespace onnx_mlir
 #endif
diff --git a/src/Accelerators/NNPA/Conversion/ONNXToZHigh/ONNXToZHighCommon.td b/src/Accelerators/NNPA/Conversion/ONNXToZHigh/ONNXToZHighCommon.td
@@ -58,4 +58,6 @@ def ACT_RELUAttr: NativeCodeCall<"$_builder.getStringAttr(\"ACT_RELU\")">;
 
 def GetTypeOf : NativeCodeCall<"$0.getType()" >;
 
+def OpIsBigEnoughForNNPA: Constraint<CPred<"!isTooSmallOpForNNPA($_self.getDefiningOp())">>;
+
 #endif // ONNX_TO_ZHIGH_COMMON
diff --git a/test/accelerators/NNPA/numerical/CMakeLists.txt b/test/accelerators/NNPA/numerical/CMakeLists.txt
@@ -110,7 +110,7 @@ add_numerical_test(TestGRUNNPA
 
 
 # LeakyRelu
-set(TestLeakyReluNNPA_instruction zdnn_mul)
+set(TestLeakyReluNNPA_instruction zdnnx_leaky_relu)
 # Automatically set following config when using --maccel=NNPA
 add_numerical_test(TestLeakyReluNNPA
   ${ONNX_NUMERICALTEST_SRC_DIR}/TestLeakyRelu.cpp

diff --git a/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/add.mlir b/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/add.mlir
@@ -64,3 +64,30 @@ func.func @test_exceed_limit_add(%arg0 : tensor<32769x10xf32>, %arg1 : tensor<32
 // CHECK-LABEL:  func @test_exceed_limit_add
 // CHECK:        "onnx.Add"
 }
+
+// -----
+
+func.func @test_scalar_add_1(%arg0 : tensor<f32>, %arg1 : tensor<f32>) -> tensor<f32> {
+  %x = "onnx.Add"(%arg0, %arg1) : (tensor<f32>, tensor<f32>) -> tensor<f32>
+  "func.return"(%x) : (tensor<f32>) -> ()
+
+// CHECK-LABEL:  func.func @test_scalar_add_1
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<f32>, [[PARAM_1_:%.+]]: tensor<f32>) -> tensor<f32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.Add"([[PARAM_0_]], [[PARAM_1_]]) : (tensor<f32>, tensor<f32>) -> tensor<f32>
+// CHECK:           return [[VAR_0_]] : tensor<f32>
+// CHECK:         }
+}
+
+// -----
+
+func.func @test_scalar_add_2(%arg0 : tensor<1xf32>, %arg1 : tensor<1xf32>) -> tensor<1xf32> {
+  %x = "onnx.Add"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
+  "func.return"(%x) : (tensor<1xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_scalar_add_2
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>, [[PARAM_1_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.Add"([[PARAM_0_]], [[PARAM_1_]]) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           return [[VAR_0_]] : tensor<1xf32>
+// CHECK:         }
+}
+
diff --git a/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/div.mlir b/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/div.mlir
@@ -64,3 +64,17 @@ func.func @test_exceed_limit_div(%arg0 : tensor<32769x10xf32>, %arg1 : tensor<32
 // CHECK-LABEL:  func @test_exceed_limit_div
 // CHECK:        "onnx.Div"
 }
+
+// -----
+
+func.func @test_scalar_div(%arg0 : tensor<1xf32>, %arg1 : tensor<1xf32>) -> tensor<1xf32> {
+  %x = "onnx.Div"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
+  "func.return"(%x) : (tensor<1xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_scalar_div
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>, [[PARAM_1_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.Div"([[PARAM_0_]], [[PARAM_1_]]) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           return [[VAR_0_]] : tensor<1xf32>
+// CHECK:         }
+}
+
diff --git a/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/exp.mlir b/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/exp.mlir
@@ -35,3 +35,17 @@ func.func @test_exceed_limit_exp(%arg0 : tensor<32769x10xf32>) -> tensor<*xf32>
 // CHECK-LABEL:  func @test_exceed_limit_exp
 // CHECK:        "onnx.Exp"
 }
+
+// -----
+
+func.func @test_scalar_exp(%arg0 : tensor<1xf32>) -> tensor<1xf32> {
+  %x = "onnx.Exp"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>
+  "func.return"(%x) : (tensor<1xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_scalar_exp
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.Exp"([[PARAM_0_]]) : (tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           return [[VAR_0_]] : tensor<1xf32>
+// CHECK:         }
+}
+
diff --git a/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/gelu.mlir b/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/gelu.mlir
@@ -28,3 +28,29 @@ func.func @test_gelu_tanh_arch15(%arg0 : tensor<1x2xf32>) -> tensor<1x2xf32> {
 // CHECK:           return [[VAR_2_]] : tensor<1x2xf32>
 // CHECK:         }
 }
+
+// -----
+
+func.func @test_gelu_erf_arch15_scalar(%arg0 : tensor<1xf32>) -> tensor<1xf32>{
+  %0 ="onnx.Gelu"(%arg0) {approximate = "none"} : (tensor<1xf32>) -> tensor<1xf32>
+ "func.return"(%0) : (tensor<1xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_gelu_erf_arch15_scalar
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.Gelu"([[PARAM_0_]]) <{approximate = "none"}> : (tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           return [[VAR_0_]] : tensor<1xf32>
+// CHECK:         }
+}
+
+// -----
+
+func.func @test_gelu_tanh_arch15_scalar(%arg0 : tensor<1xf32>) -> tensor<1xf32>{
+  %0 ="onnx.Gelu"(%arg0) {approximate = "tanh"} : (tensor<1xf32>) -> tensor<1xf32>
+ "func.return"(%0) : (tensor<1xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_gelu_tanh_arch15_scalar
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.Gelu"([[PARAM_0_]]) <{approximate = "tanh"}> : (tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           return [[VAR_0_]] : tensor<1xf32>
+// CHECK:         }
+}
diff --git a/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/invsqrt.mlir b/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/invsqrt.mlir
@@ -43,3 +43,21 @@ func.func @test_invsqrt_div2(%arg0 : tensor<1x2xf32>) -> tensor<*xf32> {
 // CHECK:           return [[VAR_2_]] : tensor<1x2xf32>
 // CHECK:         }
 }
+
+// -----
+
+func.func @test_scalar_invsqrt(%arg0 : tensor<1xf32>) -> tensor<1xf32> {
+  %a = onnx.Constant dense<[1.0]> : tensor<1xf32>
+  %x = "onnx.Sqrt"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>
+  %y = "onnx.Div"(%a, %x) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
+  "func.return"(%y) : (tensor<1xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_scalar_invsqrt
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK-DAG:       [[VAR_0_:%.+]] = onnx.Constant dense<1.000000e+00> : tensor<1xf32>
+// CHECK-DAG:       [[VAR_1_:%.+]] = "onnx.Sqrt"([[PARAM_0_]]) : (tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           [[VAR_2_:%.+]] = "onnx.Div"([[VAR_0_]], [[VAR_1_]]) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           return [[VAR_2_]] : tensor<1xf32>
+// CHECK:         }
+}
+
diff --git a/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/leakyrelu.mlir b/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/leakyrelu.mlir
@@ -40,3 +40,15 @@ func.func @test_leakyrelu_default(%arg0 : tensor<10x10xf32>) -> tensor<*xf32> {
 // CHECK:         }
 }
 
+// -----
+
+func.func @test_leakyrelu_scalar(%arg0 : tensor<1xf32>) -> tensor<*xf32> {
+  %0 = "onnx.LeakyRelu"(%arg0) : (tensor<1xf32>) -> tensor<*xf32>
+  "func.return"(%0) : (tensor<*xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_leakyrelu_scalar
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.LeakyRelu"([[PARAM_0_]]) <{alpha = 0.00999999977 : f32}> : (tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           return [[VAR_0_]] : tensor<1xf32>
+// CHECK:         }
+}
diff --git a/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/log.mlir b/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/log.mlir
@@ -36,3 +36,17 @@ func.func @test_exceed_limit_log(%arg0 : tensor<32769x10xf32>) -> tensor<*xf32>
 // CHECK-LABEL:  func @test_exceed_limit_log
 // CHECK:        "onnx.Log"
 }
+
+// -----
+
+func.func @test_scalar_log(%arg0 : tensor<1xf32>) -> tensor<1xf32> {
+  %x = "onnx.Log"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>
+  "func.return"(%x) : (tensor<1xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_scalar_log
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.Log"([[PARAM_0_]]) : (tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           return [[VAR_0_]] : tensor<1xf32>
+// CHECK:         }
+}
+
diff --git a/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/max.mlir b/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/max.mlir
@@ -64,3 +64,16 @@ func.func @test_exceed_limit_max(%arg0 : tensor<32769x10xf32>, %arg1 : tensor<32
 // CHECK-LABEL:  func @test_exceed_limit_max
 // CHECK:        "onnx.Max"
 }
+
+// -----
+
+func.func @test_max_scalar(%arg0 : tensor<1xf32>, %arg1 : tensor<1xf32>) -> tensor<*xf32> {
+  %0 = "onnx.Max"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<*xf32>
+  "func.return"(%0) : (tensor<*xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_max_scalar
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>, [[PARAM_1_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.Max"([[PARAM_0_]], [[PARAM_1_]]) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           return [[VAR_0_]] : tensor<1xf32>
+// CHECK:         }
+}
diff --git a/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/min.mlir b/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/min.mlir
@@ -64,3 +64,16 @@ func.func @test_exceed_limit_min(%arg0 : tensor<32769x10xf32>, %arg1 : tensor<32
 // CHECK-LABEL:  func @test_exceed_limit_min
 // CHECK:        "onnx.Min"
 }
+
+// -----
+
+func.func @test_min_scalar(%arg0 : tensor<1xf32>, %arg1 : tensor<1xf32>) -> tensor<*xf32> {
+  %0 = "onnx.Min"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<*xf32>
+  "func.return"(%0) : (tensor<*xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_min_scalar
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>, [[PARAM_1_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.Min"([[PARAM_0_]], [[PARAM_1_]]) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           return [[VAR_0_]] : tensor<1xf32>
+// CHECK:         }
+}
diff --git a/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/mul.mlir b/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/mul.mlir
@@ -64,3 +64,17 @@ func.func @test_exceed_limit_mul(%arg0 : tensor<32769x10xf32>, %arg1 : tensor<32
 // CHECK-LABEL:  func @test_exceed_limit_mul
 // CHECK:        "onnx.Mul"
 }
+
+// -----
+
+func.func @test_scalar_mul(%arg0 : tensor<1xf32>, %arg1 : tensor<1xf32>) -> tensor<1xf32> {
+  %x = "onnx.Mul"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
+  "func.return"(%x) : (tensor<1xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_scalar_mul
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>, [[PARAM_1_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.Mul"([[PARAM_0_]], [[PARAM_1_]]) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           return [[VAR_0_]] : tensor<1xf32>
+// CHECK:         }
+}
+
diff --git a/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/reducemax.mlir b/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/reducemax.mlir
@@ -73,3 +73,18 @@ func.func @test_reduce_max_axes_not_lowered_not_multiple_axes(%arg0 : tensor<3x2
 // CHECK:           return [[VAR_1_]] : tensor<1x2x1xf32>
 // CHECK:         }
 }
+
+// -----
+
+func.func @test_reduce_max_scalar(%arg0 : tensor<1xf32>) -> tensor<*xf32> {
+   %cst = "onnx.Constant"() {value = dense<[0]> : tensor<1xi64> } : () -> tensor<1xi64>
+   %0 ="onnx.ReduceMax"(%arg0, %cst) {keepdims = 1 : si64, noop_with_empty_axes = 0 : si64} : (tensor<1xf32>, tensor<1xi64>)-> tensor<*xf32>
+   "func.return"(%0) : (tensor<*xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_reduce_max_scalar
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK:           [[VAR_0_:%.+]] = onnx.Constant dense<0> : tensor<1xi64>
+// CHECK:           [[VAR_1_:%.+]] = "onnx.ReduceMax"([[PARAM_0_]], [[VAR_0_]]) <{keepdims = 1 : si64, noop_with_empty_axes = 0 : si64}> : (tensor<1xf32>, tensor<1xi64>) -> tensor<1xf32>
+// CHECK:           return [[VAR_1_]] : tensor<1xf32>
+// CHECK:         }
+}
diff --git a/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/reducemean.mlir b/test/mlir/accelerators/nnpa/conversion/onnx-to-zhigh/reducemean.mlir
@@ -70,3 +70,16 @@ func.func @test_exceed_limit_reducemean_v13(%arg0 : tensor<32769x3x5x7xf32>) ->
 // CHECK-LABEL:  func @test_exceed_limit_reducemean_v13
 // CHECK:        "onnx.ReduceMeanV13"
 }
+
+// -----
+
+func.func @test_reduce_mean_scalar(%arg0 : tensor<1xf32>) -> tensor<*xf32> {
+   %0 ="onnx.ReduceMeanV13"(%arg0) { axes = [0] } : (tensor<1xf32>)-> tensor<*xf32>
+   "func.return"(%0) : (tensor<*xf32>) -> ()
+
+// CHECK-LABEL:  func.func @test_reduce_mean_scalar
+// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<1xf32>) -> tensor<1xf32> {
+// CHECK:           [[VAR_0_:%.+]] = "onnx.ReduceMeanV13"([[PARAM_0_]]) <{axes = [0], keepdims = 1 : si64}> : (tensor<1xf32>) -> tensor<1xf32>
+// CHECK:           return [[VAR_0_]] : tensor<1xf32>
+// CHECK:         }
+}
Original file line number	Diff line number	Diff line change
Expand Up		@@ -58,4 +58,6 @@ def ACT_RELUAttr: NativeCodeCall<"$_builder.getStringAttr(\"ACT_RELU\")">;

		def GetTypeOf : NativeCodeCall<"$0.getType()" >;

		def OpIsBigEnoughForNNPA: Constraint<CPred<"!isTooSmallOpForNNPA($_self.getDefiningOp())">>;

		#endif // ONNX_TO_ZHIGH_COMMON