diff --git a/src/Conversion/ONNXToKrnl/Quantization/QuantizeLinear.cpp b/src/Conversion/ONNXToKrnl/Quantization/QuantizeLinear.cpp
index d7af519ea9..01293c81e9 100644
--- a/src/Conversion/ONNXToKrnl/Quantization/QuantizeLinear.cpp
+++ b/src/Conversion/ONNXToKrnl/Quantization/QuantizeLinear.cpp
@@ -71,15 +71,8 @@ void emitQuantizationLinearScalarParameters(ConversionPatternRewriter &rewriter,
   DimsExpr outputAF;
   outputAF.emplace_back(zero);
 
-  // faster than original loop on z16, takes 124us for 64k vals
-  // Allocate output buffers.
-  MemRefType flatBufferType = llvm::cast<MemRefType>(flatInput.getType());
-  Value flatBuffer = create.mem.alignedAlloc(flatBufferType, flatInputDims);
-  DimsExpr bufferAF;
-  bufferAF.emplace_back(zero);
-
   create.krnl.simdIterateIE(simdLb, simdUb, totVL, simdOnly, enableParallel,
-      {flatInput}, {inputAF}, {flatBuffer}, {bufferAF},
+      {flatInput}, {inputAF}, {flatAlloc}, {outputAF},
       {[&](const KrnlBuilder &kb, ArrayRef<Value> inputVals, int64_t VL) {
         MultiDialectBuilder<MathBuilder> create(kb);
         Value x = inputVals[0];
@@ -95,29 +88,10 @@ void emitQuantizationLinearScalarParameters(ConversionPatternRewriter &rewriter,
           adjustX = roundX;
         // Saturate: use max into a min.
         Value saturateX = create.math.clip(adjustX, qMin, qMax);
-        // Old approach.
-        // return create.math.cast(quantizedElementType, saturateX);
-        return saturateX;
+        // Convert into quantized type.
+        return create.math.cast(quantizedElementType, saturateX);
       }});
 
-  // A second loop that performs scalar float to int performs better than the
-  // compiler's attempt to generate SIMD conversion code. This might not hold
-  // with all data types, but is definitely noticeable with uint8.
-  //
-  // Investigate further: we might save the vector to a buffer on the fly
-  // (avoiding a second loop as below), and then reload each value as scalar and
-  // then saved them as scalar (thus avoiding the insert/extract SIMD operations
-  // that also do not perform well). We can have a SIMD buffer in memory for the
-  // non-quantized and quantized simd values, but then we also need to privatize
-  // it, which is also not easy in this scheme. So ignore this for now.
-  create.krnl.forLoopIE(simdLb, simdUb, 1, enableParallel,
-      [&](const KrnlBuilder &kb, ValueRange loopInd) {
-        MultiDialectBuilder<KrnlBuilder, MemRefBuilder, MathBuilder> create(kb);
-        Value buffVal = create.krnl.loadIE(flatBuffer, {zero}, {loopInd[0]});
-        Value res = create.math.cast(quantizedElementType, buffVal);
-        create.krnl.storeIE(res, flatAlloc, {zero}, {loopInd[0]});
-      });
-
   if (totVL > 1)
     onnxToKrnlSimdReport(op, /*successful*/ true, totVL,
         simdLoopStaticTripCount, "quantizationLinear whole tensor");
@@ -202,9 +176,10 @@ struct ONNXQuantizeLinearOpLowering
       hasZeroPoint = true;
     }
     if (disableQuantZeroPoint) {
-      // TODO: should we expect to disable hasZeroPoint forcefully, or generate
-      // an error if we had a zero point? Right now, just forcefully assert we
-      // have no zero point, i.e. ignore one even if we had a zero point.
+      // TODO: should we expect to disable hasZeroPoint forcefully, or
+      // generate an error if we had a zero point? Right now, just forcefully
+      // assert we have no zero point, i.e. ignore one even if we had a zero
+      // point.
       hasZeroPoint = false;
     }
     emitQuantizationLinearScalarParameters(rewriter, loc, op, xMemRefType,
diff --git a/src/Dialect/Mlir/DialectBuilder.cpp b/src/Dialect/Mlir/DialectBuilder.cpp
index 1075a25203..ff425513a5 100644
--- a/src/Dialect/Mlir/DialectBuilder.cpp
+++ b/src/Dialect/Mlir/DialectBuilder.cpp
@@ -886,6 +886,46 @@ Value MathBuilder::cast(Type destType, Value src) const {
   LLVM_DEBUG(llvm::dbgs() << "srcType: " << srcType << "\n";
              llvm::dbgs() << "destType: " << destType << "\n";);
 
+  // Before we process with the actual cast, there is a special case that we
+  // want to handle here. Cast from float to int that have different width, llvm
+  // generate better patterns if we first cast from float to int of the same
+  // width, and then from int to a different size int.
+  // Skip that optimization if the result is a 1 bit (boolean).
+  if (mlir::isa<FloatType>(srcElemType) &&
+      mlir::isa<IntegerType>(destElemType) && bitTrunc && destElemWidth > 1) {
+    // Quantization: float to smaller int. First determine the intermediary
+    // type, same integer type as destination type, with the same type width as
+    // the source float type.
+    Type step1ElementType;
+    IntegerType destIntType = mlir::cast<IntegerType>(destElemType);
+    bool destIssSigned = destIntType.isSignless() || destIntType.isSigned();
+    if (destIssSigned)
+      step1ElementType = b().getIntegerType(srcElemWidth);
+    else
+      step1ElementType = b().getIntegerType(srcElemWidth, false);
+    // Perform (recursively) the 2 step conversion. Exceptionally ok here to use
+    // element type here as cast will promote it to a vector if src is a vector.
+    Value step1Val = cast(step1ElementType, src);
+    return cast(destType, step1Val);
+  }
+  if (mlir::isa<IntegerType>(srcElemType) &&
+      mlir::isa<FloatType>(destElemType) && bitExtend) {
+    // Dequantization: small int to a float. First determine the intermediary
+    // type,  same integer type as source type, with the same type width as
+    // the destination float type.
+    Type step1ElementType;
+    IntegerType srcIntType = mlir::cast<IntegerType>(srcElemType);
+    bool srcIssSigned = srcIntType.isSignless() || srcIntType.isSigned();
+    if (srcIssSigned)
+      step1ElementType = b().getIntegerType(destElemWidth);
+    else
+      step1ElementType = b().getIntegerType(destElemWidth, false);
+    // Perform (recursively) the 2 step conversion. Exceptionally ok here to use
+    // element type here as cast will promote it to a vector if src is a vector.
+    Value step1Val = cast(step1ElementType, src);
+    return cast(destType, step1Val);
+  }
+
   // Handle boolean first because they need special handling.
   // Boolean to int/float conversions. Boolean are unsigned.
   if (srcElemType.isInteger(1)) {
diff --git a/test/mlir/conversion/onnx_to_krnl/Quantization/DequantizeLinear_with_canonicalize.mlir b/test/mlir/conversion/onnx_to_krnl/Quantization/DequantizeLinear_with_canonicalize.mlir
index 93d38fc77a..4873f5a1a4 100644
--- a/test/mlir/conversion/onnx_to_krnl/Quantization/DequantizeLinear_with_canonicalize.mlir
+++ b/test/mlir/conversion/onnx_to_krnl/Quantization/DequantizeLinear_with_canonicalize.mlir
@@ -5,10 +5,12 @@
 
 // -----
 
+
 func.func @test_dequantizelinear_i8(%arg0: tensor<4xi8>, %arg1: tensor<f32>, %arg2: tensor<i8>) -> tensor<4xf32> {
   %0 = "onnx.DequantizeLinear"(%arg0, %arg1, %arg2) {axis = 1 : si64} : (tensor<4xi8>, tensor<f32>, tensor<i8>) -> tensor<4xf32>
   return %0 : tensor<4xf32>
 
+// mlir2FileCheck.py
 // CHECK-LABEL:  func.func @test_dequantizelinear_i8
 // CHECK-SAME:   ([[PARAM_0_:%.+]]: memref<4xi8>, [[PARAM_1_:%.+]]: memref<f32>, [[PARAM_2_:%.+]]: memref<i8>) -> memref<4xf32> {
 // CHECK-DAG:       [[RES_:%.+]] = memref.alloc() {{.*}}: memref<4xf32>
@@ -18,12 +20,13 @@ func.func @test_dequantizelinear_i8(%arg0: tensor<4xi8>, %arg1: tensor<f32>, %ar
 // CHECK-DAG:         [[LOAD_PARAM_0_MEM_:%.+]] = krnl.load [[PARAM_0_]]{{.}}[[VAR_1_]]{{.}} : memref<4xi8>
 // CHECK-DAG:         [[LOAD_PARAM_1_MEM_:%.+]] = krnl.load [[PARAM_1_]][] : memref<f32>
 // CHECK-DAG:         [[LOAD_PARAM_2_MEM_:%.+]] = krnl.load [[PARAM_2_]][] : memref<i8>
-// CHECK-NOT: separator of consecutive DAGs
-// CHECK-DAG:         [[VAR_5_:%.+]] = arith.sitofp [[LOAD_PARAM_2_MEM_]] : i8 to f32
-// CHECK-DAG:         [[VAR_6_:%.+]] = arith.sitofp [[LOAD_PARAM_0_MEM_]] : i8 to f32
-// CHECK:             [[VAR_7_:%.+]] = arith.subf [[VAR_6_]], [[VAR_5_]] : f32
-// CHECK:             [[VAR_8_:%.+]] = arith.mulf [[VAR_7_]], [[LOAD_PARAM_1_MEM_]] : f32
-// CHECK:             krnl.store [[VAR_8_]], [[RES_]]{{.}}[[VAR_1_]]{{.}} : memref<4xf32>
+// CHECK:             [[VAR_5_:%.+]] = arith.extsi [[LOAD_PARAM_0_MEM_]] : i8 to i32
+// CHECK-DAG:         [[VAR_6_:%.+]] = arith.sitofp [[VAR_5_]] : i32 to f32
+// CHECK-DAG:         [[VAR_7_:%.+]] = arith.extsi [[LOAD_PARAM_2_MEM_]] : i8 to i32
+// CHECK:             [[VAR_8_:%.+]] = arith.sitofp [[VAR_7_]] : i32 to f32
+// CHECK:             [[VAR_9_:%.+]] = arith.subf [[VAR_6_]], [[VAR_8_]] : f32
+// CHECK:             [[VAR_10_:%.+]] = arith.mulf [[VAR_9_]], [[LOAD_PARAM_1_MEM_]] : f32
+// CHECK:             krnl.store [[VAR_10_]], [[RES_]]{{.}}[[VAR_1_]]{{.}} : memref<4xf32>
 // CHECK:           }
 // CHECK:           return [[RES_]] : memref<4xf32>
 // CHECK:         }
@@ -47,12 +50,14 @@ func.func @test_dequantizelinear_ui8(%arg0: tensor<4xui8>, %arg1: tensor<f32>, %
 // CHECK-DAG:         [[LOAD_PARAM_1_MEM_:%.+]] = krnl.load [[PARAM_1_]][] : memref<f32>
 // CHECK-DAG:         [[LOAD_PARAM_2_MEM_:%.+]] = krnl.load [[PARAM_2_]][] : memref<ui8>
 // CHECK:             [[VAR_5_:%.+]] = builtin.unrealized_conversion_cast [[LOAD_PARAM_0_MEM_]] : ui8 to i8
-// CHECK-DAG:         [[VAR_6_:%.+]] = arith.uitofp [[VAR_5_]] : i8 to f32
-// CHECK-DAG:         [[VAR_7_:%.+]] = builtin.unrealized_conversion_cast [[LOAD_PARAM_2_MEM_]] : ui8 to i8
-// CHECK:             [[VAR_8_:%.+]] = arith.uitofp [[VAR_7_]] : i8 to f32
-// CHECK:             [[VAR_9_:%.+]] = arith.subf [[VAR_6_]], [[VAR_8_]] : f32
-// CHECK:             [[VAR_10_:%.+]] = arith.mulf [[VAR_9_]], [[LOAD_PARAM_1_MEM_]] : f32
-// CHECK:             krnl.store [[VAR_10_]], [[RES_]]{{.}}[[VAR_1_]]{{.}} : memref<4xf32>
+// CHECK:             [[VAR_6_:%.+]] = arith.extui [[VAR_5_]] : i8 to i32
+// CHECK-DAG:         [[VAR_7_:%.+]] = arith.uitofp [[VAR_6_]] : i32 to f32
+// CHECK-DAG:         [[VAR_8_:%.+]] = builtin.unrealized_conversion_cast [[LOAD_PARAM_2_MEM_]] : ui8 to i8
+// CHECK:             [[VAR_9_:%.+]] = arith.extui [[VAR_8_]] : i8 to i32
+// CHECK:             [[VAR_10_:%.+]] = arith.uitofp [[VAR_9_]] : i32 to f32
+// CHECK:             [[VAR_11_:%.+]] = arith.subf [[VAR_7_]], [[VAR_10_]] : f32
+// CHECK:             [[VAR_12_:%.+]] = arith.mulf [[VAR_11_]], [[LOAD_PARAM_1_MEM_]] : f32
+// CHECK:             krnl.store [[VAR_12_]], [[RES_]]{{.}}[[VAR_1_]]{{.}} : memref<4xf32>
 // CHECK:           }
 // CHECK:           return [[RES_]] : memref<4xf32>
 // CHECK:         }
diff --git a/test/mlir/conversion/onnx_to_krnl/Quantization/DynamicQuantizeLinear_with_canonicalize.mlir b/test/mlir/conversion/onnx_to_krnl/Quantization/DynamicQuantizeLinear_with_canonicalize.mlir
index 934fba2240..d8dd788672 100644
--- a/test/mlir/conversion/onnx_to_krnl/Quantization/DynamicQuantizeLinear_with_canonicalize.mlir
+++ b/test/mlir/conversion/onnx_to_krnl/Quantization/DynamicQuantizeLinear_with_canonicalize.mlir
@@ -75,19 +75,19 @@ func.func @test_dynamic_quantize_linear(%arg0: tensor<?x2xf32>) -> (tensor<?x2xu
 // CHECK-DAG:       [[VAR_23_:%.+]] = arith.select [[VAR_21_]], [[VAR_22_]], [[VAR_12_]] : f32
 // CHECK-DAG:       [[VAR_24_:%.+]] = arith.cmpf oeq, [[VAR_13_]], [[CST_5_dot_000000_]] : f32
 // CHECK:           [[VAR_25_:%.+]] = arith.select [[VAR_24_]], [[VAR_23_]], [[VAR_16_]] : f32
-// CHECK:           [[VAR_26_:%.+]] = arith.fptoui [[VAR_25_]] : f32 to i8
-// CHECK:           [[VAR_27_:%.+]] = builtin.unrealized_conversion_cast [[VAR_26_]] : i8 to ui8
+// CHECK:           [[VAR_26_:%.+]] = arith.fptoui [[VAR_25_]] : f32 to i32
+// CHECK:           [[VAR_27_:%.+]] = arith.trunci [[VAR_26_]] : i32 to i8
+// CHECK:           [[VAR_28_:%.+]] = builtin.unrealized_conversion_cast [[VAR_27_]] : i8 to ui8
 // CHECK:           krnl.store [[VAR_7_]], [[RES_1_]][] : memref<f32>
-// CHECK:           krnl.store [[VAR_27_]], [[RES_2_]][] : memref<ui8>
-// CHECK-DAG:       [[VAR_28_:%.+]] = affine.apply [[MAP_0_]](){{.}}[[VAR_dim_]]{{.}}
+// CHECK:           krnl.store [[VAR_28_]], [[RES_2_]][] : memref<ui8>
+// CHECK-DAG:       [[VAR_29_:%.+]] = affine.apply [[MAP_0_]](){{.}}[[VAR_dim_]]{{.}}
 // CHECK-DAG:       [[RES_5_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
-// CHECK:           affine.store [[VAR_28_]], [[RES_5_]][0] : memref<1xindex>
+// CHECK:           affine.store [[VAR_29_]], [[RES_5_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_:%.+]] = memref.reshape [[PARAM_0_]]([[RES_5_]]) : (memref<?x2xf32>, memref<1xindex>) -> memref<?xf32>
-// CHECK-DAG:       [[VAR_29_:%.+]] = affine.apply [[MAP_0_]](){{.}}[[VAR_dim_]]{{.}}
+// CHECK-DAG:       [[VAR_30_:%.+]] = affine.apply [[MAP_0_]](){{.}}[[VAR_dim_]]{{.}}
 // CHECK-DAG:       [[RES_6_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
-// CHECK:           affine.store [[VAR_29_]], [[RES_6_]][0] : memref<1xindex>
+// CHECK:           affine.store [[VAR_30_]], [[RES_6_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_14_:%.+]] = memref.reshape [[RES_]]([[RES_]]_13) : (memref<?x2xui8>, memref<1xindex>) -> memref<?xui8>
-// CHECK-DAG:       [[RES_7_:%.+]] = memref.alloc([[VAR_28_]]) {{.*}}: memref<?xf32>
 // CHECK-DAG:       [[LOOP_2_:%.+]] = krnl.define_loops 1
 // CHECK:           krnl.iterate([[LOOP_2_]]) with ([[LOOP_2_]] -> [[I_4_:%.+]] = 0 to [[MAP_1_]]([[VAR_dim_]])){
 // CHECK:             [[VAR_32_2_:%.+]] = krnl.get_induction_var_value([[LOOP_2_]]) : (!krnl.loop) -> index
@@ -112,15 +112,10 @@ func.func @test_dynamic_quantize_linear(%arg0: tensor<?x2xf32>) -> (tensor<?x2xu
 // CHECK:             [[VAR_49_:%.+]] = arith.addf [[VAR_48_]], [[VAR_25_]] : f32
 // CHECK:             [[VAR_50_:%.+]] = arith.maxnumf [[VAR_49_]], [[CST_0_dot_000000_]] : f32
 // CHECK:             [[VAR_51_:%.+]] = arith.minnumf [[VAR_50_]], [[CST_2_dot_550000_]] : f32
-// CHECK:             krnl.store [[VAR_51_]], [[RES_7_]]{{.}}[[VAR_32_2_]]{{.}} : memref<?xf32>
-// CHECK:           }
-// CHECK:           [[LOOP_3_:%.+]] = krnl.define_loops 1
-// CHECK:           krnl.iterate([[LOOP_3_]]) with ([[LOOP_3_]] -> [[I_5_:%.+]] = 0 to [[MAP_1_]]([[VAR_dim_]])){
-// CHECK:             [[VAR_32_3_:%.+]] = krnl.get_induction_var_value([[LOOP_3_]]) : (!krnl.loop) -> index
-// CHECK:             [[LOAD_PARAM_0_MEM_1_1_:%.+]] = krnl.load [[RES_7_]]{{.}}[[VAR_32_3_]]{{.}} : memref<?xf32>
-// CHECK:             [[LOAD_RES_3_MEM_1_1_:%.+]] = arith.fptoui [[LOAD_PARAM_0_MEM_1_1_]] : f32 to i8
-// CHECK:             [[VAR_35_3_:%.+]] = builtin.unrealized_conversion_cast [[LOAD_RES_3_MEM_1_1_]] : i8 to ui8
-// CHECK:             krnl.store [[VAR_35_3_]], [[VAR_reshape_14_]]{{.}}[[VAR_32_3_]]{{.}} : memref<?xui8>
+// CHECK:             [[VAR_52_:%.+]] = arith.fptoui [[VAR_51_]] : f32 to i32
+// CHECK:             [[VAR_53_:%.+]] = arith.trunci [[VAR_52_]] : i32 to i8
+// CHECK:             [[VAR_54_:%.+]] = builtin.unrealized_conversion_cast [[VAR_53_]] : i8 to ui8
+// CHECK:             krnl.store [[VAR_54_]], [[VAR_reshape_14_]]{{.}}[[VAR_32_2_]]{{.}} : memref<?xui8>
 // CHECK:           }
 // CHECK:           return [[RES_]], [[RES_]]_6, [[RES_]]_7 : memref<?x2xui8>, memref<f32>, memref<ui8>
 // CHECK:         }
diff --git a/test/mlir/conversion/onnx_to_krnl/Quantization/DynamicQuantizeLinear_with_simd_canonicalize.mlir b/test/mlir/conversion/onnx_to_krnl/Quantization/DynamicQuantizeLinear_with_simd_canonicalize.mlir
index 637cd5fdaf..b0bea0a414 100644
--- a/test/mlir/conversion/onnx_to_krnl/Quantization/DynamicQuantizeLinear_with_simd_canonicalize.mlir
+++ b/test/mlir/conversion/onnx_to_krnl/Quantization/DynamicQuantizeLinear_with_simd_canonicalize.mlir
@@ -87,17 +87,17 @@ func.func @test_dynamic_quantize_linear_simd_only(%arg0: tensor<256x16xf32>) ->
 // CHECK-DAG:       [[VAR_26_:%.+]] = arith.select [[VAR_24_]], [[VAR_25_]], [[VAR_15_]] : f32
 // CHECK-DAG:       [[VAR_27_:%.+]] = arith.cmpf oeq, [[VAR_16_]], [[CST_5_dot_000000_]] : f32
 // CHECK:           [[VAR_28_:%.+]] = arith.select [[VAR_27_]], [[VAR_26_]], [[VAR_19_]] : f32
-// CHECK:           [[VAR_29_:%.+]] = arith.fptoui [[VAR_28_]] : f32 to i8
-// CHECK:           [[VAR_30_:%.+]] = builtin.unrealized_conversion_cast [[VAR_29_]] : i8 to ui8
+// CHECK:           [[VAR_29_:%.+]] = arith.fptoui [[VAR_28_]] : f32 to i32
+// CHECK:           [[VAR_30_:%.+]] = arith.trunci [[VAR_29_]] : i32 to i8
+// CHECK:           [[VAR_31_:%.+]] = builtin.unrealized_conversion_cast [[VAR_30_]] : i8 to ui8
 // CHECK:           krnl.store [[VAR_10_]], [[RES_1_]][] : memref<f32>
-// CHECK:           krnl.store [[VAR_30_]], [[RES_2_]][] : memref<ui8>
+// CHECK:           krnl.store [[VAR_31_]], [[RES_2_]][] : memref<ui8>
 // CHECK:           [[RES_8_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[CST_4096_]], [[RES_8_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_19_:%.+]] = memref.reshape [[PARAM_0_]]([[RES_8_]]) : (memref<256x16xf32>, memref<1xindex>) -> memref<4096xf32>
 // CHECK-DAG:       [[RES_9_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[CST_4096_]], [[RES_9_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_21_:%.+]] = memref.reshape [[RES_]]([[RES_]]_20) : (memref<256x16xui8>, memref<1xindex>) -> memref<4096xui8>
-// CHECK-DAG:       [[RES_10_:%.+]] = memref.alloc() {{.*}}: memref<4096xf32>
 // CHECK-DAG:       [[LOOP_1_:%.+]] = krnl.define_loops 1
 // CHECK:           [[BLOCK_TILE__1_:%.+]], [[BLOCK_IN__1_:%.+]] = krnl.block [[LOOP_1_]] 16 : (!krnl.loop) -> (!krnl.loop, !krnl.loop)
 // CHECK:           krnl.iterate([[BLOCK_TILE__1_]]) with ([[LOOP_1_]] -> [[I_1_:%.+]] = 0 to 4096){
@@ -126,15 +126,10 @@ func.func @test_dynamic_quantize_linear_simd_only(%arg0: tensor<256x16xf32>) ->
 // CHECK:             [[VAR_52_:%.+]] = arith.addf [[VAR_50_]], [[VAR_51_]] : vector<16xf32>
 // CHECK:             [[VAR_53_:%.+]] = arith.maxnumf [[VAR_52_]], [[VAR_cst_0_]] : vector<16xf32>
 // CHECK:             [[VAR_54_:%.+]] = arith.minnumf [[VAR_53_]], [[VAR_cst_]] : vector<16xf32>
-// CHECK:             vector.store [[VAR_54_]], [[RES_10_]]{{.}}[[VAR_33_1_]]{{.}} : memref<4096xf32>, vector<16xf32>
-// CHECK:           }
-// CHECK:           [[LOOP_2_:%.+]] = krnl.define_loops 1
-// CHECK:           krnl.iterate([[LOOP_2_]]) with ([[LOOP_2_]] -> [[I_2_:%.+]] = 0 to 4096){
-// CHECK:             [[VAR_33_2_:%.+]] = krnl.get_induction_var_value([[LOOP_2_]]) : (!krnl.loop) -> index
-// CHECK:             [[LOAD_VAR_reshape_MEM_2_:%.+]] = krnl.load [[RES_10_]]{{.}}[[VAR_33_2_]]{{.}} : memref<4096xf32>
-// CHECK:             [[LOAD_VAR_reshape_MEM_1_1_:%.+]] = arith.fptoui [[LOAD_VAR_reshape_MEM_2_]] : f32 to i8
-// CHECK:             [[LOAD_RES_4_MEM_2_:%.+]] = builtin.unrealized_conversion_cast [[LOAD_VAR_reshape_MEM_1_1_]] : i8 to ui8
-// CHECK:             krnl.store [[LOAD_RES_4_MEM_2_]], [[VAR_reshape_21_]]{{.}}[[VAR_33_2_]]{{.}} : memref<4096xui8>
+// CHECK:             [[VAR_55_:%.+]] = arith.fptoui [[VAR_54_]] : vector<16xf32> to vector<16xi32>
+// CHECK:             [[VAR_56_:%.+]] = arith.trunci [[VAR_55_]] : vector<16xi32> to vector<16xi8>
+// CHECK:             [[VAR_57_:%.+]] = builtin.unrealized_conversion_cast [[VAR_56_]] : vector<16xi8> to vector<16xui8>
+// CHECK:             vector.store [[VAR_57_]], [[VAR_reshape_21_]]{{.}}[[VAR_33_1_]]{{.}} : memref<4096xui8>, vector<16xui8>
 // CHECK:           }
 // CHECK:           return [[RES_]], [[RES_]]_11, [[RES_]]_12 : memref<256x16xui8>, memref<f32>, memref<ui8>
 // CHECK:         }
@@ -237,17 +232,17 @@ func.func @test_dynamic_quantize_linear_simd_and_scalar(%arg0: tensor<255x17xf32
 // CHECK-DAG:       [[VAR_27_:%.+]] = arith.select [[VAR_25_]], [[VAR_26_]], [[VAR_16_]] : f32
 // CHECK-DAG:       [[VAR_28_:%.+]] = arith.cmpf oeq, [[VAR_17_]], [[CST_5_dot_000000_]] : f32
 // CHECK:           [[VAR_29_:%.+]] = arith.select [[VAR_28_]], [[VAR_27_]], [[VAR_20_]] : f32
-// CHECK:           [[VAR_30_:%.+]] = arith.fptoui [[VAR_29_]] : f32 to i8
-// CHECK:           [[VAR_31_:%.+]] = builtin.unrealized_conversion_cast [[VAR_30_]] : i8 to ui8
+// CHECK:           [[VAR_30_:%.+]] = arith.fptoui [[VAR_29_]] : f32 to i32
+// CHECK:           [[VAR_31_:%.+]] = arith.trunci [[VAR_30_]] : i32 to i8
+// CHECK:           [[VAR_32_:%.+]] = builtin.unrealized_conversion_cast [[VAR_31_]] : i8 to ui8
 // CHECK:           krnl.store [[VAR_11_]], [[RES_1_]][] : memref<f32>
-// CHECK:           krnl.store [[VAR_31_]], [[RES_2_]][] : memref<ui8>
+// CHECK:           krnl.store [[VAR_32_]], [[RES_2_]][] : memref<ui8>
 // CHECK:           [[RES_8_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[CST_4335_]], [[RES_8_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_19_:%.+]] = memref.reshape [[PARAM_0_]]([[RES_8_]]) : (memref<255x17xf32>, memref<1xindex>) -> memref<4335xf32>
 // CHECK-DAG:       [[RES_9_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[CST_4335_]], [[RES_9_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_21_:%.+]] = memref.reshape [[RES_]]([[RES_]]_20) : (memref<255x17xui8>, memref<1xindex>) -> memref<4335xui8>
-// CHECK-DAG:       [[RES_10_:%.+]] = memref.alloc() {{.*}}: memref<4335xf32>
 // CHECK-DAG:       [[LOOP_2_:%.+]] = krnl.define_loops 1
 // CHECK:           [[BLOCK_TILE__1_:%.+]], [[BLOCK_IN__1_:%.+]] = krnl.block [[LOOP_2_]] 16 : (!krnl.loop) -> (!krnl.loop, !krnl.loop)
 // CHECK:           krnl.iterate([[BLOCK_TILE__1_]]) with ([[LOOP_2_]] -> [[I_2_:%.+]] = 0 to 4320){
@@ -276,7 +271,10 @@ func.func @test_dynamic_quantize_linear_simd_and_scalar(%arg0: tensor<255x17xf32
 // CHECK:             [[VAR_54_:%.+]] = arith.addf [[VAR_52_]], [[VAR_53_]] : vector<16xf32>
 // CHECK:             [[VAR_55_:%.+]] = arith.maxnumf [[VAR_54_]], [[VAR_cst_0_]] : vector<16xf32>
 // CHECK:             [[VAR_56_:%.+]] = arith.minnumf [[VAR_55_]], [[VAR_cst_]] : vector<16xf32>
-// CHECK:             vector.store [[VAR_56_]], [[RES_10_]]{{.}}[[VAR_35_2_]]{{.}} : memref<4335xf32>, vector<16xf32>
+// CHECK:             [[VAR_57_:%.+]] = arith.fptoui [[VAR_56_]] : vector<16xf32> to vector<16xi32>
+// CHECK:             [[VAR_58_:%.+]] = arith.trunci [[VAR_57_]] : vector<16xi32> to vector<16xi8>
+// CHECK:             [[VAR_59_:%.+]] = builtin.unrealized_conversion_cast [[VAR_58_]] : vector<16xi8> to vector<16xui8>
+// CHECK:             vector.store [[VAR_59_]], [[VAR_reshape_21_]]{{.}}[[VAR_35_2_]]{{.}} : memref<4335xui8>, vector<16xui8>
 // CHECK:           }
 // CHECK:           [[LOOP_3_:%.+]] = krnl.define_loops 1
 // CHECK:           krnl.iterate([[LOOP_3_]]) with ([[LOOP_3_]] -> [[I_3_:%.+]] = 4320 to 4335){
@@ -302,15 +300,10 @@ func.func @test_dynamic_quantize_linear_simd_and_scalar(%arg0: tensor<255x17xf32
 // CHECK:             [[VAR_52_1_:%.+]] = arith.addf [[VAR_51_1_]], [[VAR_29_]] : f32
 // CHECK:             [[VAR_53_1_:%.+]] = arith.maxnumf [[VAR_52_1_]], [[CST_0_dot_000000_]] : f32
 // CHECK:             [[VAR_54_1_:%.+]] = arith.minnumf [[VAR_53_1_]], [[CST_2_dot_550000_]] : f32
-// CHECK:             krnl.store [[VAR_54_1_]], [[RES_10_]]{{.}}[[VAR_35_3_]]{{.}} : memref<4335xf32>
-// CHECK:           }
-// CHECK:           [[LOOP_4_:%.+]] = krnl.define_loops 1
-// CHECK:           krnl.iterate([[LOOP_4_]]) with ([[LOOP_4_]] -> [[I_4_:%.+]] = 0 to 4335){
-// CHECK:             [[VAR_35_4_:%.+]] = krnl.get_induction_var_value([[LOOP_4_]]) : (!krnl.loop) -> index
-// CHECK:             [[LOAD_VAR_reshape_MEM_2_1_:%.+]] = krnl.load [[RES_10_]]{{.}}[[VAR_35_4_]]{{.}} : memref<4335xf32>
-// CHECK:             [[LOAD_VAR_reshape_MEM_3_1_:%.+]] = arith.fptoui [[LOAD_VAR_reshape_MEM_2_1_]] : f32 to i8
-// CHECK:             [[LOAD_RES_4_MEM_1_1_:%.+]] = builtin.unrealized_conversion_cast [[LOAD_VAR_reshape_MEM_3_1_]] : i8 to ui8
-// CHECK:             krnl.store [[LOAD_RES_4_MEM_1_1_]], [[VAR_reshape_21_]]{{.}}[[VAR_35_4_]]{{.}} : memref<4335xui8>
+// CHECK:             [[VAR_55_1_:%.+]] = arith.fptoui [[VAR_54_1_]] : f32 to i32
+// CHECK:             [[VAR_56_1_:%.+]] = arith.trunci [[VAR_55_1_]] : i32 to i8
+// CHECK:             [[VAR_57_1_:%.+]] = builtin.unrealized_conversion_cast [[VAR_56_1_]] : i8 to ui8
+// CHECK:             krnl.store [[VAR_57_1_]], [[VAR_reshape_21_]]{{.}}[[VAR_35_3_]]{{.}} : memref<4335xui8>
 // CHECK:           }
 // CHECK:           return [[RES_]], [[RES_]]_11, [[RES_]]_12 : memref<255x17xui8>, memref<f32>, memref<ui8>
 // CHECK:         }
@@ -400,17 +393,17 @@ func.func @test_dynamic_quantize_linear_reduced_simd_only(%arg0: tensor<1x8xf32>
 // CHECK-DAG:       [[VAR_26_:%.+]] = arith.select [[VAR_24_]], [[VAR_25_]], [[VAR_15_]] : f32
 // CHECK-DAG:       [[VAR_27_:%.+]] = arith.cmpf oeq, [[VAR_16_]], [[CST_5_dot_000000_]] : f32
 // CHECK:           [[VAR_28_:%.+]] = arith.select [[VAR_27_]], [[VAR_26_]], [[VAR_19_]] : f32
-// CHECK:           [[VAR_29_:%.+]] = arith.fptoui [[VAR_28_]] : f32 to i8
-// CHECK:           [[VAR_30_:%.+]] = builtin.unrealized_conversion_cast [[VAR_29_]] : i8 to ui8
+// CHECK:           [[VAR_29_:%.+]] = arith.fptoui [[VAR_28_]] : f32 to i32
+// CHECK:           [[VAR_30_:%.+]] = arith.trunci [[VAR_29_]] : i32 to i8
+// CHECK:           [[VAR_31_:%.+]] = builtin.unrealized_conversion_cast [[VAR_30_]] : i8 to ui8
 // CHECK:           krnl.store [[VAR_10_]], [[RES_1_]][] : memref<f32>
-// CHECK:           krnl.store [[VAR_30_]], [[RES_2_]][] : memref<ui8>
+// CHECK:           krnl.store [[VAR_31_]], [[RES_2_]][] : memref<ui8>
 // CHECK:           [[RES_8_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[CST_8_]], [[RES_8_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_19_:%.+]] = memref.reshape [[PARAM_0_]]([[RES_8_]]) : (memref<1x8xf32>, memref<1xindex>) -> memref<8xf32>
 // CHECK-DAG:       [[RES_9_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[CST_8_]], [[RES_9_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_21_:%.+]] = memref.reshape [[RES_]]([[RES_]]_20) : (memref<1x8xui8>, memref<1xindex>) -> memref<8xui8>
-// CHECK-DAG:       [[RES_10_:%.+]] = memref.alloc() {{.*}}: memref<8xf32>
 // CHECK-DAG:       [[LOOP_1_:%.+]] = krnl.define_loops 1
 // CHECK:           [[BLOCK_TILE__1_:%.+]], [[BLOCK_IN__1_:%.+]] = krnl.block [[LOOP_1_]] 8 : (!krnl.loop) -> (!krnl.loop, !krnl.loop)
 // CHECK:           krnl.iterate([[BLOCK_TILE__1_]]) with ([[LOOP_1_]] -> [[I_1_:%.+]] = 0 to 8){
@@ -439,15 +432,10 @@ func.func @test_dynamic_quantize_linear_reduced_simd_only(%arg0: tensor<1x8xf32>
 // CHECK:             [[VAR_52_:%.+]] = arith.addf [[VAR_50_]], [[VAR_51_]] : vector<8xf32>
 // CHECK:             [[VAR_53_:%.+]] = arith.maxnumf [[VAR_52_]], [[VAR_cst_0_]] : vector<8xf32>
 // CHECK:             [[VAR_54_:%.+]] = arith.minnumf [[VAR_53_]], [[VAR_cst_]] : vector<8xf32>
-// CHECK:             vector.store [[VAR_54_]], [[RES_10_]]{{.}}[[VAR_33_1_]]{{.}} : memref<8xf32>, vector<8xf32>
-// CHECK:           }
-// CHECK:           [[LOOP_2_:%.+]] = krnl.define_loops 1
-// CHECK:           krnl.iterate([[LOOP_2_]]) with ([[LOOP_2_]] -> [[I_2_:%.+]] = 0 to 8){
-// CHECK:             [[VAR_33_2_:%.+]] = krnl.get_induction_var_value([[LOOP_2_]]) : (!krnl.loop) -> index
-// CHECK:             [[LOAD_VAR_reshape_MEM_2_:%.+]] = krnl.load [[RES_10_]]{{.}}[[VAR_33_2_]]{{.}} : memref<8xf32>
-// CHECK:             [[LOAD_VAR_reshape_MEM_1_1_:%.+]] = arith.fptoui [[LOAD_VAR_reshape_MEM_2_]] : f32 to i8
-// CHECK:             [[LOAD_RES_4_MEM_2_:%.+]] = builtin.unrealized_conversion_cast [[LOAD_VAR_reshape_MEM_1_1_]] : i8 to ui8
-// CHECK:             krnl.store [[LOAD_RES_4_MEM_2_]], [[VAR_reshape_21_]]{{.}}[[VAR_33_2_]]{{.}} : memref<8xui8>
+// CHECK:             [[VAR_55_:%.+]] = arith.fptoui [[VAR_54_]] : vector<8xf32> to vector<8xi32>
+// CHECK:             [[VAR_56_:%.+]] = arith.trunci [[VAR_55_]] : vector<8xi32> to vector<8xi8>
+// CHECK:             [[VAR_57_:%.+]] = builtin.unrealized_conversion_cast [[VAR_56_]] : vector<8xi8> to vector<8xui8>
+// CHECK:             vector.store [[VAR_57_]], [[VAR_reshape_21_]]{{.}}[[VAR_33_1_]]{{.}} : memref<8xui8>, vector<8xui8>
 // CHECK:           }
 // CHECK:           return [[RES_]], [[RES_]]_11, [[RES_]]_12 : memref<1x8xui8>, memref<f32>, memref<ui8>
 // CHECK:         }
diff --git a/test/mlir/conversion/onnx_to_krnl/Quantization/DynamicQuantizeLinear_with_simd_parallel_canonicalize.mlir b/test/mlir/conversion/onnx_to_krnl/Quantization/DynamicQuantizeLinear_with_simd_parallel_canonicalize.mlir
index dd048e3ab7..14d809207c 100644
--- a/test/mlir/conversion/onnx_to_krnl/Quantization/DynamicQuantizeLinear_with_simd_parallel_canonicalize.mlir
+++ b/test/mlir/conversion/onnx_to_krnl/Quantization/DynamicQuantizeLinear_with_simd_parallel_canonicalize.mlir
@@ -141,17 +141,17 @@ func.func @test_dynamic_quantize_linear_simd_only(%arg0: tensor<256x16xf32>) ->
 // CHECK-DAG:       [[VAR_27_:%.+]] = arith.select [[VAR_25_]], [[VAR_26_]], [[VAR_16_]] : f32
 // CHECK-DAG:       [[VAR_28_:%.+]] = arith.cmpf oeq, [[VAR_17_]], [[CST_5_dot_000000_]] : f32
 // CHECK:           [[VAR_29_:%.+]] = arith.select [[VAR_28_]], [[VAR_27_]], [[VAR_20_]] : f32
-// CHECK:           [[VAR_30_:%.+]] = arith.fptoui [[VAR_29_]] : f32 to i8
-// CHECK:           [[VAR_31_:%.+]] = builtin.unrealized_conversion_cast [[VAR_30_]] : i8 to ui8
+// CHECK:           [[VAR_30_:%.+]] = arith.fptoui [[VAR_29_]] : f32 to i32
+// CHECK:           [[VAR_31_:%.+]] = arith.trunci [[VAR_30_]] : i32 to i8
+// CHECK:           [[VAR_32_:%.+]] = builtin.unrealized_conversion_cast [[VAR_31_]] : i8 to ui8
 // CHECK:           krnl.store [[VAR_11_]], [[RES_1_]][] : memref<f32>
-// CHECK:           krnl.store [[VAR_31_]], [[RES_2_]][] : memref<ui8>
+// CHECK:           krnl.store [[VAR_32_]], [[RES_2_]][] : memref<ui8>
 // CHECK:           [[RES_10_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[CST_4096_]], [[RES_10_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_23_:%.+]] = memref.reshape [[PARAM_0_]]([[RES_10_]]) : (memref<256x16xf32>, memref<1xindex>) -> memref<4096xf32>
 // CHECK-DAG:       [[RES_11_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[CST_4096_]], [[RES_11_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_25_:%.+]] = memref.reshape [[RES_]]([[RES_]]_24) : (memref<256x16xui8>, memref<1xindex>) -> memref<4096xui8>
-// CHECK-DAG:       [[RES_12_:%.+]] = memref.alloc() {{.*}}: memref<4096xf32>
 // CHECK-DAG:       [[LOOP_2_:%.+]] = krnl.define_loops 1
 // CHECK:           [[BLOCK_TILE__0_:%.+]], [[BLOCK_IN__0_:%.+]] = krnl.block [[LOOP_2_]] 16 : (!krnl.loop) -> (!krnl.loop, !krnl.loop)
 // CHECK:           krnl.parallel([[BLOCK_TILE__0_]]) : !krnl.loop
@@ -181,16 +181,10 @@ func.func @test_dynamic_quantize_linear_simd_only(%arg0: tensor<256x16xf32>) ->
 // CHECK:             [[VAR_53_:%.+]] = arith.addf [[VAR_51_2_]], [[VAR_52_2_]] : vector<16xf32>
 // CHECK:             [[VAR_54_:%.+]] = arith.maxnumf [[VAR_53_]], [[VAR_cst_0_]] : vector<16xf32>
 // CHECK:             [[VAR_55_:%.+]] = arith.minnumf [[VAR_54_]], [[VAR_cst_]] : vector<16xf32>
-// CHECK:             vector.store [[VAR_55_]], [[RES_12_]]{{.}}[[VAR_34_2_]]{{.}} : memref<4096xf32>, vector<16xf32>
-// CHECK:           }
-// CHECK:           [[LOOP_3_:%.+]] = krnl.define_loops 1
-// CHECK:           krnl.parallel([[LOOP_3_]]) : !krnl.loop
-// CHECK:           krnl.iterate([[LOOP_3_]]) with ([[LOOP_3_]] -> [[I_5_:%.+]] = 0 to 4096){
-// CHECK:             [[VAR_34_3_:%.+]] = krnl.get_induction_var_value([[LOOP_3_]]) : (!krnl.loop) -> index
-// CHECK:             [[VAR_35_1_1_:%.+]] = krnl.load [[RES_12_]]{{.}}[[VAR_34_3_]]{{.}} : memref<4096xf32>
-// CHECK:             [[VAR_36_3_:%.+]] = arith.fptoui [[VAR_35_1_1_]] : f32 to i8
-// CHECK:             [[VAR_37_2_:%.+]] = builtin.unrealized_conversion_cast [[VAR_36_3_]] : i8 to ui8
-// CHECK:             krnl.store [[VAR_37_2_]], [[VAR_reshape_25_]]{{.}}[[VAR_34_3_]]{{.}} : memref<4096xui8>
+// CHECK:             [[VAR_56_:%.+]] = arith.fptoui [[VAR_55_]] : vector<16xf32> to vector<16xi32>
+// CHECK:             [[VAR_57_:%.+]] = arith.trunci [[VAR_56_]] : vector<16xi32> to vector<16xi8>
+// CHECK:             [[VAR_58_:%.+]] = builtin.unrealized_conversion_cast [[VAR_57_]] : vector<16xi8> to vector<16xui8>
+// CHECK:             vector.store [[VAR_58_]], [[VAR_reshape_25_]]{{.}}[[VAR_34_2_]]{{.}} : memref<4096xui8>, vector<16xui8>
 // CHECK:           }
 // CHECK:           return [[RES_]], [[RES_]]_13, [[RES_]]_14 : memref<256x16xui8>, memref<f32>, memref<ui8>
 // CHECK:         }
@@ -334,17 +328,17 @@ func.func @test_dynamic_quantize_linear_simd_and_scalar(%arg0: tensor<255x17xf32
 // CHECK-DAG:       [[VAR_27_:%.+]] = arith.select [[VAR_25_]], [[VAR_26_]], [[VAR_16_]] : f32
 // CHECK-DAG:       [[VAR_28_:%.+]] = arith.cmpf oeq, [[VAR_17_]], [[CST_5_dot_000000_]] : f32
 // CHECK:           [[VAR_29_:%.+]] = arith.select [[VAR_28_]], [[VAR_27_]], [[VAR_20_]] : f32
-// CHECK:           [[VAR_30_:%.+]] = arith.fptoui [[VAR_29_]] : f32 to i8
-// CHECK:           [[VAR_31_:%.+]] = builtin.unrealized_conversion_cast [[VAR_30_]] : i8 to ui8
+// CHECK:           [[VAR_30_:%.+]] = arith.fptoui [[VAR_29_]] : f32 to i32
+// CHECK:           [[VAR_31_:%.+]] = arith.trunci [[VAR_30_]] : i32 to i8
+// CHECK:           [[VAR_32_:%.+]] = builtin.unrealized_conversion_cast [[VAR_31_]] : i8 to ui8
 // CHECK:           krnl.store [[VAR_11_]], [[RES_1_]][] : memref<f32>
-// CHECK:           krnl.store [[VAR_31_]], [[RES_2_]][] : memref<ui8>
+// CHECK:           krnl.store [[VAR_32_]], [[RES_2_]][] : memref<ui8>
 // CHECK:           [[RES_10_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[CST_4335_]], [[RES_10_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_23_:%.+]] = memref.reshape [[PARAM_0_]]([[RES_10_]]) : (memref<255x17xf32>, memref<1xindex>) -> memref<4335xf32>
 // CHECK-DAG:       [[RES_11_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[CST_4335_]], [[RES_11_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_25_:%.+]] = memref.reshape [[RES_]]([[RES_]]_24) : (memref<255x17xui8>, memref<1xindex>) -> memref<4335xui8>
-// CHECK-DAG:       [[RES_12_:%.+]] = memref.alloc() {{.*}}: memref<4335xf32>
 // CHECK-DAG:       [[LOOP_2_:%.+]] = krnl.define_loops 1
 // CHECK:           [[BLOCK_TILE__0_:%.+]], [[BLOCK_IN__0_:%.+]] = krnl.block [[LOOP_2_]] 16 : (!krnl.loop) -> (!krnl.loop, !krnl.loop)
 // CHECK:           krnl.parallel([[BLOCK_TILE__0_]]) : !krnl.loop
@@ -374,7 +368,10 @@ func.func @test_dynamic_quantize_linear_simd_and_scalar(%arg0: tensor<255x17xf32
 // CHECK:             [[VAR_54_:%.+]] = arith.addf [[VAR_52_2_]], [[VAR_53_2_]] : vector<16xf32>
 // CHECK:             [[VAR_55_:%.+]] = arith.maxnumf [[VAR_54_]], [[VAR_cst_0_]] : vector<16xf32>
 // CHECK:             [[VAR_56_:%.+]] = arith.minnumf [[VAR_55_]], [[VAR_cst_]] : vector<16xf32>
-// CHECK:             vector.store [[VAR_56_]], [[RES_12_]]{{.}}[[VAR_35_2_]]{{.}} : memref<4335xf32>, vector<16xf32>
+// CHECK:             [[VAR_57_:%.+]] = arith.fptoui [[VAR_56_]] : vector<16xf32> to vector<16xi32>
+// CHECK:             [[VAR_58_:%.+]] = arith.trunci [[VAR_57_]] : vector<16xi32> to vector<16xi8>
+// CHECK:             [[VAR_59_:%.+]] = builtin.unrealized_conversion_cast [[VAR_58_]] : vector<16xi8> to vector<16xui8>
+// CHECK:             vector.store [[VAR_59_]], [[VAR_reshape_25_]]{{.}}[[VAR_35_2_]]{{.}} : memref<4335xui8>, vector<16xui8>
 // CHECK:           }
 // CHECK:           [[LOOP_3_:%.+]] = krnl.define_loops 1
 // CHECK:           krnl.iterate([[LOOP_3_]]) with ([[LOOP_3_]] -> [[I_5_:%.+]] = 4320 to 4335){
@@ -400,16 +397,10 @@ func.func @test_dynamic_quantize_linear_simd_and_scalar(%arg0: tensor<255x17xf32
 // CHECK:             [[VAR_52_3_:%.+]] = arith.addf [[LOAD_RES_6_MEM_1_1_]], [[VAR_29_]] : f32
 // CHECK:             [[VAR_53_3_:%.+]] = arith.maxnumf [[VAR_52_3_]], [[CST_0_dot_000000_]] : f32
 // CHECK:             [[VAR_54_1_:%.+]] = arith.minnumf [[VAR_53_3_]], [[CST_2_dot_550000_]] : f32
-// CHECK:             krnl.store [[VAR_54_1_]], [[RES_12_]]{{.}}[[VAR_35_3_]]{{.}} : memref<4335xf32>
-// CHECK:           }
-// CHECK:           [[LOOP_4_:%.+]] = krnl.define_loops 1
-// CHECK:           krnl.parallel([[LOOP_4_]]) : !krnl.loop
-// CHECK:           krnl.iterate([[LOOP_4_]]) with ([[LOOP_4_]] -> [[I_6_:%.+]] = 0 to 4335){
-// CHECK:             [[VAR_35_4_:%.+]] = krnl.get_induction_var_value([[LOOP_4_]]) : (!krnl.loop) -> index
-// CHECK:             [[VAR_36_1_1_:%.+]] = krnl.load [[RES_12_]]{{.}}[[VAR_35_4_]]{{.}} : memref<4335xf32>
-// CHECK:             [[VAR_37_4_:%.+]] = arith.fptoui [[VAR_36_1_1_]] : f32 to i8
-// CHECK:             [[VAR_38_3_:%.+]] = builtin.unrealized_conversion_cast [[VAR_37_4_]] : i8 to ui8
-// CHECK:             krnl.store [[VAR_38_3_]], [[VAR_reshape_25_]]{{.}}[[VAR_35_4_]]{{.}} : memref<4335xui8>
+// CHECK:             [[VAR_55_1_:%.+]] = arith.fptoui [[VAR_54_1_]] : f32 to i32
+// CHECK:             [[VAR_56_1_:%.+]] = arith.trunci [[VAR_55_1_]] : i32 to i8
+// CHECK:             [[VAR_57_1_:%.+]] = builtin.unrealized_conversion_cast [[VAR_56_1_]] : i8 to ui8
+// CHECK:             krnl.store [[VAR_57_1_]], [[VAR_reshape_25_]]{{.}}[[VAR_35_3_]]{{.}} : memref<4335xui8>
 // CHECK:           }
 // CHECK:           return [[RES_]], [[RES_]]_13, [[RES_]]_14 : memref<255x17xui8>, memref<f32>, memref<ui8>
 // CHECK:         }
@@ -499,17 +490,17 @@ func.func @test_dynamic_quantize_linear_reduced_simd_only(%arg0: tensor<1x8xf32>
 // CHECK-DAG:       [[VAR_26_:%.+]] = arith.select [[VAR_24_]], [[VAR_25_]], [[VAR_15_]] : f32
 // CHECK-DAG:       [[VAR_27_:%.+]] = arith.cmpf oeq, [[VAR_16_]], [[CST_5_dot_000000_]] : f32
 // CHECK:           [[VAR_28_:%.+]] = arith.select [[VAR_27_]], [[VAR_26_]], [[VAR_19_]] : f32
-// CHECK:           [[VAR_29_:%.+]] = arith.fptoui [[VAR_28_]] : f32 to i8
-// CHECK:           [[VAR_30_:%.+]] = builtin.unrealized_conversion_cast [[VAR_29_]] : i8 to ui8
+// CHECK:           [[VAR_29_:%.+]] = arith.fptoui [[VAR_28_]] : f32 to i32
+// CHECK:           [[VAR_30_:%.+]] = arith.trunci [[VAR_29_]] : i32 to i8
+// CHECK:           [[VAR_31_:%.+]] = builtin.unrealized_conversion_cast [[VAR_30_]] : i8 to ui8
 // CHECK:           krnl.store [[VAR_10_]], [[RES_1_]][] : memref<f32>
-// CHECK:           krnl.store [[VAR_30_]], [[RES_2_]][] : memref<ui8>
+// CHECK:           krnl.store [[VAR_31_]], [[RES_2_]][] : memref<ui8>
 // CHECK:           [[RES_8_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[CST_8_]], [[RES_8_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_19_:%.+]] = memref.reshape [[PARAM_0_]]([[RES_8_]]) : (memref<1x8xf32>, memref<1xindex>) -> memref<8xf32>
 // CHECK-DAG:       [[RES_9_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[CST_8_]], [[RES_9_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_21_:%.+]] = memref.reshape [[RES_]]([[RES_]]_20) : (memref<1x8xui8>, memref<1xindex>) -> memref<8xui8>
-// CHECK-DAG:       [[RES_10_:%.+]] = memref.alloc() {{.*}}: memref<8xf32>
 // CHECK-DAG:       [[LOOP_1_:%.+]] = krnl.define_loops 1
 // CHECK:           [[BLOCK_TILE__1_:%.+]], [[BLOCK_IN__1_:%.+]] = krnl.block [[LOOP_1_]] 8 : (!krnl.loop) -> (!krnl.loop, !krnl.loop)
 // CHECK:           krnl.parallel([[BLOCK_TILE__1_]]) : !krnl.loop
@@ -539,16 +530,10 @@ func.func @test_dynamic_quantize_linear_reduced_simd_only(%arg0: tensor<1x8xf32>
 // CHECK:             [[VAR_52_:%.+]] = arith.addf [[VAR_50_]], [[VAR_51_]] : vector<8xf32>
 // CHECK:             [[VAR_53_:%.+]] = arith.maxnumf [[VAR_52_]], [[VAR_cst_0_]] : vector<8xf32>
 // CHECK:             [[VAR_54_:%.+]] = arith.minnumf [[VAR_53_]], [[VAR_cst_]] : vector<8xf32>
-// CHECK:             vector.store [[VAR_54_]], [[RES_10_]]{{.}}[[VAR_33_1_]]{{.}} : memref<8xf32>, vector<8xf32>
-// CHECK:           }
-// CHECK:           [[LOOP_2_:%.+]] = krnl.define_loops 1
-// CHECK:           krnl.parallel([[LOOP_2_]]) : !krnl.loop
-// CHECK:           krnl.iterate([[LOOP_2_]]) with ([[LOOP_2_]] -> [[I_2_:%.+]] = 0 to 8){
-// CHECK:             [[VAR_33_2_:%.+]] = krnl.get_induction_var_value([[LOOP_2_]]) : (!krnl.loop) -> index
-// CHECK:             [[LOAD_VAR_reshape_MEM_2_:%.+]] = krnl.load [[RES_10_]]{{.}}[[VAR_33_2_]]{{.}} : memref<8xf32>
-// CHECK:             [[LOAD_VAR_reshape_MEM_1_1_:%.+]] = arith.fptoui [[LOAD_VAR_reshape_MEM_2_]] : f32 to i8
-// CHECK:             [[LOAD_RES_4_MEM_2_:%.+]] = builtin.unrealized_conversion_cast [[LOAD_VAR_reshape_MEM_1_1_]] : i8 to ui8
-// CHECK:             krnl.store [[LOAD_RES_4_MEM_2_]], [[VAR_reshape_21_]]{{.}}[[VAR_33_2_]]{{.}} : memref<8xui8>
+// CHECK:             [[VAR_55_:%.+]] = arith.fptoui [[VAR_54_]] : vector<8xf32> to vector<8xi32>
+// CHECK:             [[VAR_56_:%.+]] = arith.trunci [[VAR_55_]] : vector<8xi32> to vector<8xi8>
+// CHECK:             [[VAR_57_:%.+]] = builtin.unrealized_conversion_cast [[VAR_56_]] : vector<8xi8> to vector<8xui8>
+// CHECK:             vector.store [[VAR_57_]], [[VAR_reshape_21_]]{{.}}[[VAR_33_1_]]{{.}} : memref<8xui8>, vector<8xui8>
 // CHECK:           }
 // CHECK:           return [[RES_]], [[RES_]]_11, [[RES_]]_12 : memref<1x8xui8>, memref<f32>, memref<ui8>
 // CHECK:         }
diff --git a/test/mlir/conversion/onnx_to_krnl/Quantization/QuantizationWithoutZeroPoint.mlir b/test/mlir/conversion/onnx_to_krnl/Quantization/QuantizationWithoutZeroPoint.mlir
index ff079672bd..15fedeab1b 100644
--- a/test/mlir/conversion/onnx_to_krnl/Quantization/QuantizationWithoutZeroPoint.mlir
+++ b/test/mlir/conversion/onnx_to_krnl/Quantization/QuantizationWithoutZeroPoint.mlir
@@ -22,9 +22,10 @@ func.func @test_dequantizelinear_ui8(%arg0: tensor<4xui8>, %arg1: tensor<f32>, %
 // CHECK-DAG:         [[LOAD_PARAM_0_MEM_:%.+]] = krnl.load [[PARAM_0_]]{{.}}[[VAR_1_]]{{.}} : memref<4xui8>
 // CHECK-DAG:         [[LOAD_PARAM_1_MEM_:%.+]] = krnl.load [[PARAM_1_]][] : memref<f32>
 // CHECK:             [[VAR_4_:%.+]] = builtin.unrealized_conversion_cast [[LOAD_PARAM_0_MEM_]] : ui8 to i8
-// CHECK:             [[VAR_5_:%.+]] = arith.uitofp [[VAR_4_]] : i8 to f32
-// CHECK:             [[VAR_6_:%.+]] = arith.mulf [[VAR_5_]], [[LOAD_PARAM_1_MEM_]] : f32
-// CHECK:             krnl.store [[VAR_6_]], [[RES_]]{{.}}[[VAR_1_]]{{.}} : memref<4xf32>
+// CHECK:             [[VAR_5_:%.+]] = arith.extui [[VAR_4_]] : i8 to i32
+// CHECK:             [[VAR_6_:%.+]] = arith.uitofp [[VAR_5_]] : i32 to f32
+// CHECK:             [[VAR_7_:%.+]] = arith.mulf [[VAR_6_]], [[LOAD_PARAM_1_MEM_]] : f32
+// CHECK:             krnl.store [[VAR_7_]], [[RES_]]{{.}}[[VAR_1_]]{{.}} : memref<4xf32>
 // CHECK:           }
 // CHECK:           return [[RES_]] : memref<4xf32>
 // CHECK:         }
@@ -60,22 +61,22 @@ func.func @test_dynamic_quantize_linear(%arg0: tensor<?x2xf32>) -> (tensor<?x2xu
 // CHECK-DAG:       [[LOOP_0_:%.+]]:2 = krnl.define_loops 2
 // CHECK-DAG:       [[VAR_dim_9_:%.+]] = memref.dim [[PARAM_0_]], [[CST_0_3_]] : memref<?x2xf32>
 // CHECK:           krnl.iterate([[LOOP_0_]]#0, [[LOOP_0_]]#1) with ([[LOOP_0_]]#0 -> [[I_0_:%.+]] = 0 to [[VAR_dim_9_]], [[LOOP_0_]]#1 -> [[I_1_:%.+]] = 0 to 2){
-// CHECK:             [[VAR_13_:%.+]]:2 = krnl.get_induction_var_value([[LOOP_0_]]#0, [[LOOP_0_]]#1) : (!krnl.loop, !krnl.loop) -> (index, index)
-// CHECK-DAG:         [[LOAD_PARAM_0_MEM_:%.+]] = krnl.load [[PARAM_0_]]{{.}}[[VAR_13_]]#0, [[VAR_13_]]#1] : memref<?x2xf32>
+// CHECK:             [[VAR_12_:%.+]]:2 = krnl.get_induction_var_value([[LOOP_0_]]#0, [[LOOP_0_]]#1) : (!krnl.loop, !krnl.loop) -> (index, index)
+// CHECK-DAG:         [[LOAD_PARAM_0_MEM_:%.+]] = krnl.load [[PARAM_0_]]{{.}}[[VAR_12_]]#0, [[VAR_12_]]#1] : memref<?x2xf32>
 // CHECK-DAG:         [[LOAD_RES_3_MEM_:%.+]] = krnl.load [[RES_3_]][] : memref<f32>
-// CHECK:             [[VAR_16_:%.+]] = arith.minnumf [[LOAD_RES_3_MEM_]], [[LOAD_PARAM_0_MEM_]] : f32
-// CHECK:             krnl.store [[VAR_16_]], [[RES_3_]][] : memref<f32>
+// CHECK:             [[VAR_15_:%.+]] = arith.minnumf [[LOAD_RES_3_MEM_]], [[LOAD_PARAM_0_MEM_]] : f32
+// CHECK:             krnl.store [[VAR_15_]], [[RES_3_]][] : memref<f32>
 // CHECK:           }
 // CHECK:           [[RES_4_:%.+]] = memref.alloc() : memref<f32>
 // CHECK:           krnl.memset [[RES_4_]], [[CST_0_1_]] : memref<f32>
 // CHECK-DAG:       [[LOOP_1_:%.+]]:2 = krnl.define_loops 2
 // CHECK-DAG:       [[VAR_dim_11_:%.+]] = memref.dim [[PARAM_0_]], [[CST_0_3_]] : memref<?x2xf32>
 // CHECK:           krnl.iterate([[LOOP_1_]]#0, [[LOOP_1_]]#1) with ([[LOOP_1_]]#0 -> [[I_2_:%.+]] = 0 to [[VAR_dim_11_]], [[LOOP_1_]]#1 -> [[I_3_:%.+]] = 0 to 2){
-// CHECK:             [[VAR_13_1_:%.+]]:2 = krnl.get_induction_var_value([[LOOP_1_]]#0, [[LOOP_1_]]#1) : (!krnl.loop, !krnl.loop) -> (index, index)
-// CHECK-DAG:         [[LOAD_PARAM_0_MEM_1_:%.+]] = krnl.load [[PARAM_0_]]{{.}}[[VAR_13_1_]]#0, [[VAR_13_1_]]#1] : memref<?x2xf32>
+// CHECK:             [[VAR_12_1_:%.+]]:2 = krnl.get_induction_var_value([[LOOP_1_]]#0, [[LOOP_1_]]#1) : (!krnl.loop, !krnl.loop) -> (index, index)
+// CHECK-DAG:         [[LOAD_PARAM_0_MEM_1_:%.+]] = krnl.load [[PARAM_0_]]{{.}}[[VAR_12_1_]]#0, [[VAR_12_1_]]#1] : memref<?x2xf32>
 // CHECK-DAG:         [[LOAD_RES_3_MEM_1_:%.+]] = krnl.load [[RES_4_]][] : memref<f32>
-// CHECK:             [[VAR_16_1_:%.+]] = arith.maxnumf [[LOAD_RES_3_MEM_1_]], [[LOAD_PARAM_0_MEM_1_]] : f32
-// CHECK:             krnl.store [[VAR_16_1_]], [[RES_4_]][] : memref<f32>
+// CHECK:             [[VAR_15_1_:%.+]] = arith.maxnumf [[LOAD_RES_3_MEM_1_]], [[LOAD_PARAM_0_MEM_1_]] : f32
+// CHECK:             krnl.store [[VAR_15_1_]], [[RES_4_]][] : memref<f32>
 // CHECK:           }
 // CHECK-DAG:       [[LOAD_RES_3_MEM_2_:%.+]] = krnl.load [[RES_3_]][] : memref<f32>
 // CHECK-DAG:       [[LOAD_RES_4_MEM_:%.+]] = krnl.load [[RES_4_]][] : memref<f32>
@@ -95,39 +96,33 @@ func.func @test_dynamic_quantize_linear(%arg0: tensor<?x2xf32>) -> (tensor<?x2xu
 // CHECK-DAG:       [[RES_6_:%.+]] = memref.alloc() {{.*}}: memref<1xindex>
 // CHECK:           affine.store [[VAR_10_]], [[RES_6_]][0] : memref<1xindex>
 // CHECK-DAG:       [[VAR_reshape_14_:%.+]] = memref.reshape [[RES_]]([[RES_]]_13) : (memref<?x2xui8>, memref<1xindex>) -> memref<?xui8>
-// CHECK-DAG:       [[RES_7_:%.+]] = memref.alloc([[VAR_9_]]) {{.*}}: memref<?xf32>
 // CHECK-DAG:       [[LOOP_2_:%.+]] = krnl.define_loops 1
 // CHECK:           krnl.iterate([[LOOP_2_]]) with ([[LOOP_2_]] -> [[I_4_:%.+]] = 0 to [[MAP_1_]]([[VAR_dim_]])){
-// CHECK:             [[VAR_13_2_:%.+]] = krnl.get_induction_var_value([[LOOP_2_]]) : (!krnl.loop) -> index
-// CHECK:             [[LOAD_PARAM_0_MEM_1_:%.+]] = krnl.load [[VAR_reshape_]]{{.}}[[VAR_13_2_]]{{.}} : memref<?xf32>
+// CHECK:             [[VAR_12_2_:%.+]] = krnl.get_induction_var_value([[LOOP_2_]]) : (!krnl.loop) -> index
+// CHECK:             [[LOAD_PARAM_0_MEM_1_:%.+]] = krnl.load [[VAR_reshape_]]{{.}}[[VAR_12_2_]]{{.}} : memref<?xf32>
 // CHECK:             [[LOAD_RES_3_MEM_1_:%.+]] = arith.divf [[LOAD_PARAM_0_MEM_1_]], [[VAR_7_]] : f32
-// CHECK:             [[VAR_16_2_:%.+]] = math.floor [[LOAD_RES_3_MEM_1_]] : f32
-// CHECK:             [[VAR_17_:%.+]] = arith.subf [[LOAD_RES_3_MEM_1_]], [[VAR_16_2_]] : f32
-// CHECK-DAG:         [[VAR_18_:%.+]] = arith.cmpf ogt, [[VAR_17_]], [[CST_5_dot_000000_]] : f32
-// CHECK-DAG:         [[VAR_19_:%.+]] = arith.addf [[VAR_16_2_]], [[CST_1_dot_000000_]] : f32
+// CHECK:             [[VAR_15_2_:%.+]] = math.floor [[LOAD_RES_3_MEM_1_]] : f32
+// CHECK:             [[VAR_16_:%.+]] = arith.subf [[LOAD_RES_3_MEM_1_]], [[VAR_15_2_]] : f32
+// CHECK-DAG:         [[VAR_17_:%.+]] = arith.cmpf ogt, [[VAR_16_]], [[CST_5_dot_000000_]] : f32
+// CHECK-DAG:         [[VAR_18_:%.+]] = arith.addf [[VAR_15_2_]], [[CST_1_dot_000000_]] : f32
 // CHECK-NOT: separator of consecutive DAGs
-// CHECK-DAG:         [[VAR_20_:%.+]] = arith.select [[VAR_18_]], [[VAR_19_]], [[VAR_16_2_]] : f32
-// CHECK-DAG:         [[VAR_21_:%.+]] = arith.mulf [[VAR_16_2_]], [[CST_5_dot_000000_]] : f32
-// CHECK:             [[VAR_22_:%.+]] = math.floor [[VAR_21_]] : f32
-// CHECK:             [[VAR_23_:%.+]] = arith.mulf [[VAR_22_]], [[CST_2_dot_000000_]] : f32
-// CHECK:             [[VAR_24_:%.+]] = arith.subf [[VAR_16_2_]], [[VAR_23_]] : f32
-// CHECK-DAG:         [[VAR_25_:%.+]] = arith.cmpf oeq, [[VAR_24_]], [[CST_1_dot_000000_]] : f32
-// CHECK-DAG:         [[VAR_26_:%.+]] = arith.addf [[VAR_16_2_]], [[CST_1_dot_000000_]] : f32
+// CHECK-DAG:         [[VAR_19_:%.+]] = arith.select [[VAR_17_]], [[VAR_18_]], [[VAR_15_2_]] : f32
+// CHECK-DAG:         [[VAR_20_:%.+]] = arith.mulf [[VAR_15_2_]], [[CST_5_dot_000000_]] : f32
+// CHECK:             [[VAR_21_:%.+]] = math.floor [[VAR_20_]] : f32
+// CHECK:             [[VAR_22_:%.+]] = arith.mulf [[VAR_21_]], [[CST_2_dot_000000_]] : f32
+// CHECK:             [[VAR_23_:%.+]] = arith.subf [[VAR_15_2_]], [[VAR_22_]] : f32
+// CHECK-DAG:         [[VAR_24_:%.+]] = arith.cmpf oeq, [[VAR_23_]], [[CST_1_dot_000000_]] : f32
+// CHECK-DAG:         [[VAR_25_:%.+]] = arith.addf [[VAR_15_2_]], [[CST_1_dot_000000_]] : f32
 // CHECK-NOT: separator of consecutive DAGs
-// CHECK-DAG:         [[VAR_27_:%.+]] = arith.select [[VAR_25_]], [[VAR_26_]], [[VAR_16_2_]] : f32
-// CHECK-DAG:         [[VAR_28_:%.+]] = arith.cmpf oeq, [[VAR_17_]], [[CST_5_dot_000000_]] : f32
-// CHECK:             [[VAR_29_:%.+]] = arith.select [[VAR_28_]], [[VAR_27_]], [[VAR_20_]] : f32
-// CHECK:             [[VAR_30_:%.+]] = arith.maxnumf [[VAR_29_]], [[CST_0_dot_000000_]] : f32
-// CHECK:             [[VAR_31_:%.+]] = arith.minnumf [[VAR_30_]], [[CST_2_dot_550000_]] : f32
-// CHECK:             krnl.store [[VAR_31_]], [[RES_7_]]{{.}}[[VAR_13_2_]]{{.}} : memref<?xf32>
-// CHECK:           }
-// CHECK:           [[LOOP_3_:%.+]] = krnl.define_loops 1
-// CHECK:           krnl.iterate([[LOOP_3_]]) with ([[LOOP_3_]] -> [[I_5_:%.+]] = 0 to [[MAP_1_]]([[VAR_dim_]])){
-// CHECK:             [[VAR_13_3_:%.+]] = krnl.get_induction_var_value([[LOOP_3_]]) : (!krnl.loop) -> index
-// CHECK:             [[LOAD_PARAM_0_MEM_1_1_:%.+]] = krnl.load [[RES_7_]]{{.}}[[VAR_13_3_]]{{.}} : memref<?xf32>
-// CHECK:             [[LOAD_RES_3_MEM_1_1_:%.+]] = arith.fptoui [[LOAD_PARAM_0_MEM_1_1_]] : f32 to i8
-// CHECK:             [[VAR_16_3_:%.+]] = builtin.unrealized_conversion_cast [[LOAD_RES_3_MEM_1_1_]] : i8 to ui8
-// CHECK:             krnl.store [[VAR_16_3_]], [[VAR_reshape_14_]]{{.}}[[VAR_13_3_]]{{.}} : memref<?xui8>
+// CHECK-DAG:         [[VAR_26_:%.+]] = arith.select [[VAR_24_]], [[VAR_25_]], [[VAR_15_2_]] : f32
+// CHECK-DAG:         [[VAR_27_:%.+]] = arith.cmpf oeq, [[VAR_16_]], [[CST_5_dot_000000_]] : f32
+// CHECK:             [[VAR_28_:%.+]] = arith.select [[VAR_27_]], [[VAR_26_]], [[VAR_19_]] : f32
+// CHECK:             [[VAR_29_:%.+]] = arith.maxnumf [[VAR_28_]], [[CST_0_dot_000000_]] : f32
+// CHECK:             [[VAR_30_:%.+]] = arith.minnumf [[VAR_29_]], [[CST_2_dot_550000_]] : f32
+// CHECK:             [[VAR_31_:%.+]] = arith.fptoui [[VAR_30_]] : f32 to i32
+// CHECK:             [[VAR_32_:%.+]] = arith.trunci [[VAR_31_]] : i32 to i8
+// CHECK:             [[VAR_33_:%.+]] = builtin.unrealized_conversion_cast [[VAR_32_]] : i8 to ui8
+// CHECK:             krnl.store [[VAR_33_]], [[VAR_reshape_14_]]{{.}}[[VAR_12_2_]]{{.}} : memref<?xui8>
 // CHECK:           }
 // CHECK:           return [[RES_]], [[RES_]]_6, [[RES_]]_7 : memref<?x2xui8>, memref<f32>, memref<ui8>
 // CHECK:         }
@@ -150,39 +145,33 @@ func.func @test_quantize_linear_ui8(%arg0: tensor<6xf32>, %arg1: tensor<f32>, %a
 // CHECK-DAG:       [[CST_2_dot_550000_:%.+]] = arith.constant 2.550000e+02 : f32
 // CHECK-DAG:       [[RES_:%.+]] = memref.alloc() {{.*}}: memref<6xui8>
 // CHECK-DAG:       [[LOAD_PARAM_1_MEM_:%.+]] = krnl.load [[PARAM_1_]][] : memref<f32>
-// CHECK-DAG:       [[RES_1_:%.+]] = memref.alloc() {{.*}}: memref<6xf32>
 // CHECK-DAG:       [[LOOP_0_:%.+]] = krnl.define_loops 1
 // CHECK:           krnl.iterate([[LOOP_0_]]) with ([[LOOP_0_]] -> [[I_0_:%.+]] = 0 to 6){
-// CHECK:             [[VAR_3_:%.+]] = krnl.get_induction_var_value([[LOOP_0_]]) : (!krnl.loop) -> index
-// CHECK:             [[LOAD_PARAM_0_MEM_:%.+]] = krnl.load [[PARAM_0_]]{{.}}[[VAR_3_]]{{.}} : memref<6xf32>
-// CHECK:             [[VAR_5_:%.+]] = arith.divf [[LOAD_PARAM_0_MEM_]], [[LOAD_PARAM_1_MEM_]] : f32
-// CHECK:             [[VAR_6_:%.+]] = math.floor [[VAR_5_]] : f32
-// CHECK:             [[VAR_7_:%.+]] = arith.subf [[VAR_5_]], [[VAR_6_]] : f32
-// CHECK-DAG:         [[VAR_8_:%.+]] = arith.cmpf ogt, [[VAR_7_]], [[CST_5_dot_000000_]] : f32
-// CHECK-DAG:         [[VAR_9_:%.+]] = arith.addf [[VAR_6_]], [[CST_1_dot_000000_]] : f32
+// CHECK:             [[VAR_2_:%.+]] = krnl.get_induction_var_value([[LOOP_0_]]) : (!krnl.loop) -> index
+// CHECK:             [[LOAD_PARAM_0_MEM_:%.+]] = krnl.load [[PARAM_0_]]{{.}}[[VAR_2_]]{{.}} : memref<6xf32>
+// CHECK:             [[VAR_4_:%.+]] = arith.divf [[LOAD_PARAM_0_MEM_]], [[LOAD_PARAM_1_MEM_]] : f32
+// CHECK:             [[VAR_5_:%.+]] = math.floor [[VAR_4_]] : f32
+// CHECK:             [[VAR_6_:%.+]] = arith.subf [[VAR_4_]], [[VAR_5_]] : f32
+// CHECK-DAG:         [[VAR_7_:%.+]] = arith.cmpf ogt, [[VAR_6_]], [[CST_5_dot_000000_]] : f32
+// CHECK-DAG:         [[VAR_8_:%.+]] = arith.addf [[VAR_5_]], [[CST_1_dot_000000_]] : f32
 // CHECK-NOT: separator of consecutive DAGs
-// CHECK-DAG:         [[VAR_10_:%.+]] = arith.select [[VAR_8_]], [[VAR_9_]], [[VAR_6_]] : f32
-// CHECK-DAG:         [[VAR_11_:%.+]] = arith.mulf [[VAR_6_]], [[CST_5_dot_000000_]] : f32
-// CHECK:             [[VAR_12_:%.+]] = math.floor [[VAR_11_]] : f32
-// CHECK:             [[VAR_13_:%.+]] = arith.mulf [[VAR_12_]], [[CST_2_dot_000000_]] : f32
-// CHECK:             [[VAR_14_:%.+]] = arith.subf [[VAR_6_]], [[VAR_13_]] : f32
-// CHECK-DAG:         [[VAR_15_:%.+]] = arith.cmpf oeq, [[VAR_14_]], [[CST_1_dot_000000_]] : f32
-// CHECK-DAG:         [[VAR_16_:%.+]] = arith.addf [[VAR_6_]], [[CST_1_dot_000000_]] : f32
+// CHECK-DAG:         [[VAR_9_:%.+]] = arith.select [[VAR_7_]], [[VAR_8_]], [[VAR_5_]] : f32
+// CHECK-DAG:         [[VAR_10_:%.+]] = arith.mulf [[VAR_5_]], [[CST_5_dot_000000_]] : f32
+// CHECK:             [[VAR_11_:%.+]] = math.floor [[VAR_10_]] : f32
+// CHECK:             [[VAR_12_:%.+]] = arith.mulf [[VAR_11_]], [[CST_2_dot_000000_]] : f32
+// CHECK:             [[VAR_13_:%.+]] = arith.subf [[VAR_5_]], [[VAR_12_]] : f32
+// CHECK-DAG:         [[VAR_14_:%.+]] = arith.cmpf oeq, [[VAR_13_]], [[CST_1_dot_000000_]] : f32
+// CHECK-DAG:         [[VAR_15_:%.+]] = arith.addf [[VAR_5_]], [[CST_1_dot_000000_]] : f32
 // CHECK-NOT: separator of consecutive DAGs
-// CHECK-DAG:         [[VAR_17_:%.+]] = arith.select [[VAR_15_]], [[VAR_16_]], [[VAR_6_]] : f32
-// CHECK-DAG:         [[VAR_18_:%.+]] = arith.cmpf oeq, [[VAR_7_]], [[CST_5_dot_000000_]] : f32
-// CHECK:             [[VAR_19_:%.+]] = arith.select [[VAR_18_]], [[VAR_17_]], [[VAR_10_]] : f32
-// CHECK:             [[VAR_20_:%.+]] = arith.maxnumf [[VAR_19_]], [[CST_0_dot_000000_]] : f32
-// CHECK:             [[VAR_21_:%.+]] = arith.minnumf [[VAR_20_]], [[CST_2_dot_550000_]] : f32
-// CHECK:             krnl.store [[VAR_21_]], [[RES_1_]]{{.}}[[VAR_3_]]{{.}} : memref<6xf32>
-// CHECK:           }
-// CHECK:           [[LOOP_1_:%.+]] = krnl.define_loops 1
-// CHECK:           krnl.iterate([[LOOP_1_]]) with ([[LOOP_1_]] -> [[I_1_:%.+]] = 0 to 6){
-// CHECK:             [[VAR_3_1_:%.+]] = krnl.get_induction_var_value([[LOOP_1_]]) : (!krnl.loop) -> index
-// CHECK:             [[LOAD_PARAM_0_MEM_1_:%.+]] = krnl.load [[RES_1_]]{{.}}[[VAR_3_1_]]{{.}} : memref<6xf32>
-// CHECK:             [[VAR_5_1_:%.+]] = arith.fptoui [[LOAD_PARAM_0_MEM_1_]] : f32 to i8
-// CHECK:             [[VAR_6_1_:%.+]] = builtin.unrealized_conversion_cast [[VAR_5_1_]] : i8 to ui8
-// CHECK:             krnl.store [[VAR_6_1_]], [[RES_]]{{.}}[[VAR_3_1_]]{{.}} : memref<6xui8>
+// CHECK-DAG:         [[VAR_16_:%.+]] = arith.select [[VAR_14_]], [[VAR_15_]], [[VAR_5_]] : f32
+// CHECK-DAG:         [[VAR_17_:%.+]] = arith.cmpf oeq, [[VAR_6_]], [[CST_5_dot_000000_]] : f32
+// CHECK:             [[VAR_18_:%.+]] = arith.select [[VAR_17_]], [[VAR_16_]], [[VAR_9_]] : f32
+// CHECK:             [[VAR_19_:%.+]] = arith.maxnumf [[VAR_18_]], [[CST_0_dot_000000_]] : f32
+// CHECK:             [[VAR_20_:%.+]] = arith.minnumf [[VAR_19_]], [[CST_2_dot_550000_]] : f32
+// CHECK:             [[VAR_21_:%.+]] = arith.fptoui [[VAR_20_]] : f32 to i32
+// CHECK:             [[VAR_22_:%.+]] = arith.trunci [[VAR_21_]] : i32 to i8
+// CHECK:             [[VAR_23_:%.+]] = builtin.unrealized_conversion_cast [[VAR_22_]] : i8 to ui8
+// CHECK:             krnl.store [[VAR_23_]], [[RES_]]{{.}}[[VAR_2_]]{{.}} : memref<6xui8>
 // CHECK:           }
 // CHECK:           return [[RES_]] : memref<6xui8>
 // CHECK:         }
diff --git a/test/mlir/conversion/onnx_to_krnl/Quantization/QuantizeLinear_with_canonicalize.mlir b/test/mlir/conversion/onnx_to_krnl/Quantization/QuantizeLinear_with_canonicalize.mlir
index 6ec8f9d8cb..42d7cad73e 100644
--- a/test/mlir/conversion/onnx_to_krnl/Quantization/QuantizeLinear_with_canonicalize.mlir
+++ b/test/mlir/conversion/onnx_to_krnl/Quantization/QuantizeLinear_with_canonicalize.mlir
@@ -22,8 +22,8 @@ func.func @test_quantize_linear_ui8(%arg0: tensor<6xf32>, %arg1: tensor<f32>, %a
 // CHECK-DAG:       [[LOAD_PARAM_1_MEM_:%.+]] = krnl.load [[PARAM_1_]][] : memref<f32>
 // CHECK-DAG:       [[LOAD_PARAM_2_MEM_:%.+]] = krnl.load [[PARAM_2_]][] : memref<ui8>
 // CHECK:           [[VAR_2_:%.+]] = builtin.unrealized_conversion_cast [[LOAD_PARAM_2_MEM_]] : ui8 to i8
-// CHECK-DAG:       [[VAR_3_:%.+]] = arith.uitofp [[VAR_2_]] : i8 to f32
-// CHECK-DAG:       [[RES_1_:%.+]] = memref.alloc() {{.*}}: memref<6xf32>
+// CHECK:           [[VAR_3_:%.+]] = arith.extui [[VAR_2_]] : i8 to i32
+// CHECK-DAG:       [[VAR_4_:%.+]] = arith.uitofp [[VAR_3_]] : i32 to f32
 // CHECK-DAG:       [[LOOP_0_:%.+]] = krnl.define_loops 1
 // CHECK:           krnl.iterate([[LOOP_0_]]) with ([[LOOP_0_]] -> [[I_0_:%.+]] = 0 to 6){
 // CHECK:             [[VAR_6_:%.+]] = krnl.get_induction_var_value([[LOOP_0_]]) : (!krnl.loop) -> index
@@ -45,18 +45,13 @@ func.func @test_quantize_linear_ui8(%arg0: tensor<6xf32>, %arg1: tensor<f32>, %a
 // CHECK-DAG:         [[VAR_20_:%.+]] = arith.select [[VAR_18_]], [[VAR_19_]], [[VAR_9_]] : f32
 // CHECK-DAG:         [[VAR_21_:%.+]] = arith.cmpf oeq, [[VAR_10_]], [[CST_5_dot_000000_]] : f32
 // CHECK:             [[VAR_22_:%.+]] = arith.select [[VAR_21_]], [[VAR_20_]], [[VAR_13_]] : f32
-// CHECK:             [[VAR_23_:%.+]] = arith.addf [[VAR_22_]], [[VAR_3_]] : f32
+// CHECK:             [[VAR_23_:%.+]] = arith.addf [[VAR_22_]], [[VAR_4_]] : f32
 // CHECK:             [[VAR_24_:%.+]] = arith.maxnumf [[VAR_23_]], [[CST_0_dot_000000_]] : f32
 // CHECK:             [[VAR_25_:%.+]] = arith.minnumf [[VAR_24_]], [[CST_2_dot_550000_]] : f32
-// CHECK:             krnl.store [[VAR_25_]], [[RES_1_]]{{.}}[[VAR_6_]]{{.}} : memref<6xf32>
-// CHECK:           }
-// CHECK:           [[LOOP_1_:%.+]] = krnl.define_loops 1
-// CHECK:           krnl.iterate([[LOOP_1_]]) with ([[LOOP_1_]] -> [[I_1_:%.+]] = 0 to 6){
-// CHECK:             [[VAR_6_1_:%.+]] = krnl.get_induction_var_value([[LOOP_1_]]) : (!krnl.loop) -> index
-// CHECK:             [[LOAD_PARAM_0_MEM_1_:%.+]] = krnl.load [[RES_1_]]{{.}}[[VAR_6_1_]]{{.}} : memref<6xf32>
-// CHECK:             [[VAR_8_1_:%.+]] = arith.fptoui [[LOAD_PARAM_0_MEM_1_]] : f32 to i8
-// CHECK:             [[VAR_9_1_:%.+]] = builtin.unrealized_conversion_cast [[VAR_8_1_]] : i8 to ui8
-// CHECK:             krnl.store [[VAR_9_1_]], [[RES_]]{{.}}[[VAR_6_1_]]{{.}} : memref<6xui8>
+// CHECK:             [[VAR_26_:%.+]] = arith.fptoui [[VAR_25_]] : f32 to i32
+// CHECK:             [[VAR_27_:%.+]] = arith.trunci [[VAR_26_]] : i32 to i8
+// CHECK:             [[VAR_28_:%.+]] = builtin.unrealized_conversion_cast [[VAR_27_]] : i8 to ui8
+// CHECK:             krnl.store [[VAR_28_]], [[RES_]]{{.}}[[VAR_6_]]{{.}} : memref<6xui8>
 // CHECK:           }
 // CHECK:           return [[RES_]] : memref<6xui8>
 // CHECK:         }
@@ -80,9 +75,8 @@ func.func @test_quantize_linear_i8(%arg0: tensor<6xf32>, %arg1: tensor<f32>, %ar
 // CHECK-DAG:       [[RES_:%.+]] = memref.alloc() {{.*}}: memref<6xi8>
 // CHECK-DAG:       [[LOAD_PARAM_1_MEM_:%.+]] = krnl.load [[PARAM_1_]][] : memref<f32>
 // CHECK-DAG:       [[LOAD_PARAM_2_MEM_:%.+]] = krnl.load [[PARAM_2_]][] : memref<i8>
-// CHECK-NOT: separator of consecutive DAGs
-// CHECK-DAG:       [[VAR_2_:%.+]] = arith.sitofp [[LOAD_PARAM_2_MEM_]] : i8 to f32
-// CHECK-DAG:       [[RES_1_:%.+]] = memref.alloc() {{.*}}: memref<6xf32>
+// CHECK:           [[VAR_2_:%.+]] = arith.extsi [[LOAD_PARAM_2_MEM_]] : i8 to i32
+// CHECK-DAG:       [[VAR_3_:%.+]] = arith.sitofp [[VAR_2_]] : i32 to f32
 // CHECK-DAG:       [[LOOP_0_:%.+]] = krnl.define_loops 1
 // CHECK:           krnl.iterate([[LOOP_0_]]) with ([[LOOP_0_]] -> [[I_0_:%.+]] = 0 to 6){
 // CHECK:             [[VAR_5_:%.+]] = krnl.get_induction_var_value([[LOOP_0_]]) : (!krnl.loop) -> index
@@ -104,17 +98,12 @@ func.func @test_quantize_linear_i8(%arg0: tensor<6xf32>, %arg1: tensor<f32>, %ar
 // CHECK-DAG:         [[VAR_19_:%.+]] = arith.select [[VAR_17_]], [[VAR_18_]], [[VAR_8_]] : f32
 // CHECK-DAG:         [[VAR_20_:%.+]] = arith.cmpf oeq, [[VAR_9_]], [[CST_5_dot_000000_]] : f32
 // CHECK:             [[VAR_21_:%.+]] = arith.select [[VAR_20_]], [[VAR_19_]], [[VAR_12_]] : f32
-// CHECK:             [[VAR_22_:%.+]] = arith.addf [[VAR_21_]], [[VAR_2_]] : f32
+// CHECK:             [[VAR_22_:%.+]] = arith.addf [[VAR_21_]], [[VAR_3_]] : f32
 // CHECK:             [[VAR_23_:%.+]] = arith.maxnumf [[VAR_22_]], [[CST_minus_1_dot_280000_]] : f32
 // CHECK:             [[VAR_24_:%.+]] = arith.minnumf [[VAR_23_]], [[CST_1_dot_270000_]] : f32
-// CHECK:             krnl.store [[VAR_24_]], [[RES_1_]]{{.}}[[VAR_5_]]{{.}} : memref<6xf32>
-// CHECK:           }
-// CHECK:           [[LOOP_1_:%.+]] = krnl.define_loops 1
-// CHECK:           krnl.iterate([[LOOP_1_]]) with ([[LOOP_1_]] -> [[I_1_:%.+]] = 0 to 6){
-// CHECK:             [[VAR_5_1_:%.+]] = krnl.get_induction_var_value([[LOOP_1_]]) : (!krnl.loop) -> index
-// CHECK:             [[LOAD_PARAM_0_MEM_1_:%.+]] = krnl.load [[RES_1_]]{{.}}[[VAR_5_1_]]{{.}} : memref<6xf32>
-// CHECK:             [[VAR_7_1_:%.+]] = arith.fptosi [[LOAD_PARAM_0_MEM_1_]] : f32 to i8
-// CHECK:             krnl.store [[VAR_7_1_]], [[RES_]]{{.}}[[VAR_5_1_]]{{.}} : memref<6xi8>
+// CHECK:             [[VAR_25_:%.+]] = arith.fptosi [[VAR_24_]] : f32 to i32
+// CHECK:             [[VAR_26_:%.+]] = arith.trunci [[VAR_25_]] : i32 to i8
+// CHECK:             krnl.store [[VAR_26_]], [[RES_]]{{.}}[[VAR_5_]]{{.}} : memref<6xi8>
 // CHECK:           }
 // CHECK:           return [[RES_]] : memref<6xi8>
 // CHECK:         }
diff --git a/test/mlir/conversion/onnx_to_krnl/onnx_lowering_fuse.mlir b/test/mlir/conversion/onnx_to_krnl/onnx_lowering_fuse.mlir
index 5309b274fd..f319459c68 100644
--- a/test/mlir/conversion/onnx_to_krnl/onnx_lowering_fuse.mlir
+++ b/test/mlir/conversion/onnx_to_krnl/onnx_lowering_fuse.mlir
@@ -124,8 +124,9 @@ func.func @test_fuse_element8(%arg0: tensor<?xf32>, %arg1: tensor<1xf32>) -> ten
 // CHECK-DAG:         [[LOAD_PARAM_0_MEM_:%.+]] = krnl.load [[PARAM_0_]]{{.}}[[VAR_1_]]{{.}} : memref<?xf32>
 // CHECK-DAG:         [[LOAD_PARAM_1_MEM_:%.+]] = krnl.load [[PARAM_1_]]{{.}}[[CST_0_]]{{.}} : memref<1xf32>
 // CHECK:             [[VAR_4_:%.+]] = math.powf [[LOAD_PARAM_0_MEM_]], [[LOAD_PARAM_1_MEM_]] : f32
-// CHECK:             [[VAR_5_:%.+]] = arith.fptosi [[VAR_4_]] : f32 to i8
-// CHECK:             krnl.store [[VAR_5_]], [[RES_]]{{.}}[[VAR_1_]]{{.}} : memref<?xi8>
+// CHECK:             [[VAR_5_:%.+]] = arith.fptosi [[VAR_4_]] : f32 to i32
+// CHECK:             [[VAR_6_:%.+]] = arith.trunci [[VAR_5_]] : i32 to i8
+// CHECK:             krnl.store [[VAR_6_]], [[RES_]]{{.}}[[VAR_1_]]{{.}} : memref<?xi8>
 // CHECK:           }
 // CHECK:           return [[RES_]] : memref<?xi8>
 // CHECK:         }
@@ -322,13 +323,14 @@ func.func @fuse_element_20(%533: tensor<?x?x768xf32>, %537 : tensor<?x?x1xf32>,
 
 // -----
 
+
 func.func @test_fuse_element21(%arg0: tensor<?xf32>, %arg1: tensor<1xf32>, %arg2 : tensor<1xi8>) -> tensor<?xi8> {
-  %0 = "onnx.Pow"(%arg0, %arg1) : (tensor<?xf32>, tensor<1xf32>) -> tensor<?xf32> 
+  %0 = "onnx.Pow"(%arg0, %arg1) : (tensor<?xf32>, tensor<1xf32>) -> tensor<?xf32>
   %1 = "onnx.Cast"(%0) {to = i8} : (tensor<?xf32>) -> tensor<?xi8>
   %2 = "onnx.Add"(%1, %arg2) : (tensor<?xi8>, tensor<1xi8>) -> tensor<?xi8>
   return %2 : tensor<?xi8>
-}
 
+// mlir2FileCheck.py
 // CHECK-DAG:   [[MAP_0_:#.+]] = affine_map<(d0) -> (d0)>
 // CHECK-LABEL:  func.func @test_fuse_element21
 // CHECK-SAME:   ([[PARAM_0_:%.+]]: memref<?xf32>, [[PARAM_1_:%.+]]: memref<1xf32>, [[PARAM_2_:%.+]]: memref<1xi8>) -> memref<?xi8> {
@@ -341,12 +343,13 @@ func.func @test_fuse_element21(%arg0: tensor<?xf32>, %arg1: tensor<1xf32>, %arg2
 // CHECK-DAG:         [[LOAD_PARAM_0_MEM_:%.+]] = krnl.load [[PARAM_0_]]{{.}}[[VAR_1_]]{{.}} : memref<?xf32>
 // CHECK-DAG:         [[LOAD_PARAM_1_MEM_:%.+]] = krnl.load [[PARAM_1_]]{{.}}[[CST_0_]]{{.}} : memref<1xf32>
 // CHECK:             [[VAR_4_:%.+]] = math.powf [[LOAD_PARAM_0_MEM_]], [[LOAD_PARAM_1_MEM_]] : f32
-// CHECK-DAG:         [[VAR_5_:%.+]] = arith.fptosi [[VAR_4_]] : f32 to i8
+// CHECK:             [[VAR_5_:%.+]] = arith.fptosi [[VAR_4_]] : f32 to i32
+// CHECK-DAG:         [[VAR_6_:%.+]] = arith.trunci [[VAR_5_]] : i32 to i8
 // CHECK-DAG:         [[LOAD_PARAM_2_MEM_:%.+]] = krnl.load [[PARAM_2_]]{{.}}[[CST_0_]]{{.}} : memref<1xi8>
-// CHECK:             [[VAR_7_:%.+]] = arith.addi [[VAR_5_]], [[LOAD_PARAM_2_MEM_]] : i8
-// CHECK:             krnl.store [[VAR_7_]], [[RES_]]{{.}}[[VAR_1_]]{{.}} : memref<?xi8>
+// CHECK:             [[VAR_8_:%.+]] = arith.addi [[VAR_6_]], [[LOAD_PARAM_2_MEM_]] : i8
+// CHECK:             krnl.store [[VAR_8_]], [[RES_]]{{.}}[[VAR_1_]]{{.}} : memref<?xi8>
 // CHECK:           }
 // CHECK:           return [[RES_]] : memref<?xi8>
 // CHECK:         }
-
+}
 
diff --git a/utils/fixLitTest.py b/utils/fixLitTest.py
index 717f2c3567..b9e5ab3b46 100755
--- a/utils/fixLitTest.py
+++ b/utils/fixLitTest.py
@@ -425,6 +425,7 @@ def main(argv):
             dprint("\n>> Tested with " + str(test_error_num) + " errors:")
             for f in test_error_functions:
                 dprint(">>   " + f)
+    dprint(">> Completed processing of " + lit_test_filename + "\n")
 
 
 if __name__ == "__main__":