Merge pull request #409 from ehsan-toosi/enadjara.no_error_for_scalar_with_different_ranks

feat: Generate no error when scalar values with different ranks are provided for scale and zero point

Author: ehsan-toosi

src/Dialect/ONNX/ONNXOps/Quantize/DequantizeLinear.cpp

@@ -92,7 +92,7 @@ LogicalResult ONNXDequantizeLinearOpShapeHelper::computeShape() {
 
 LogicalResult ONNXDequantizeLinearOp::verify() {
   // Is tensor known to be a scalar (rank 0 or rank 1 with 1 element)?
-  auto isScalar = [](RankedTensorType t) -> bool {
+  auto isScalar = [](ShapedType t) -> bool {
     return t.getRank() == 0 || (t.getRank() == 1 && t.getDimSize(0) == 1);
   };
 
@@ -108,7 +108,8 @@ LogicalResult ONNXDequantizeLinearOp::verify() {
   Value zero = getXZeroPoint();
   if (!isNoneValue(zero)) {
     const auto zeroTy = mlir::cast<ShapedType>(zero.getType());
-    if (zeroTy.hasRank() && scaleTy.hasRank() &&
+    if (zeroTy.hasRank() && scaleTy.hasRank() && !isScalar(zeroTy) &&
+        !isScalar(scaleTy) &&
         (zeroTy.getRank() != scaleTy.getRank() ||
             zeroTy.getShape() != scaleTy.getShape())) {
       return emitOpError("x_zero_point must have the same shape as x_scale");

test/mlir/onnx/invalid.mlir

@@ -115,6 +115,16 @@ func.func @test_dequantize_linear_verifier_2(%arg0 : tensor<5x5x1xi32>, %arg1 :
 
 // -----
 
+// We should not get any error for scalar values with different ranks.
+func.func @test_scalar_with_different_rank(%arg0 : tensor<5x5x1xi32>) -> tensor<5x5x1xf32> {
+  %0 = "onnx.Constant"(){ value = dense<1> : tensor<i32>} : () -> tensor<i32>
+  %1 = "onnx.Constant"(){ value = dense<2.0> : tensor<1xf32> } : () -> tensor<1xf32>
+  %2 = "onnx.DequantizeLinear"(%arg0, %1, %0) {} : (tensor<5x5x1xi32>, tensor<1xf32>, tensor<i32>)  -> tensor<5x5x1xf32>
+  "onnx.Return"(%2) : (tensor<5x5x1xf32>) -> ()
+}
+
+// -----
+
 func.func @test_dequantize_linear_verifier_3(%arg0 : tensor<5x5x1xi32>, %arg1 : tensor<3xf32>, %arg2 : tensor<3xi32>) -> tensor<*xf32> {
   // expected-error @+1 {{'onnx.DequantizeLinear' op x_scale and x_zero_point 1-D tensor length must match the input axis dim size}}
   %1 = "onnx.DequantizeLinear"(%arg0, %arg1, %arg2) {} : (tensor<5x5x1xi32>, tensor<3xf32>, tensor<3xi32>)  -> tensor<*xf32>