Fix computing intermediate zero in DynamicQuantizeLinear (#2535)

Signed-off-by: Tung D. Le <[email protected]>
onnx · Sep 27, 2023 · 6ad3bbe · 6ad3bbe
1 parent a3930ad
commit 6ad3bbe
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 4 deletions.
diff --git a/src/Conversion/ONNXToKrnl/Quantization/DynamicQuantizeLinear.cpp b/src/Conversion/ONNXToKrnl/Quantization/DynamicQuantizeLinear.cpp
@@ -105,7 +105,7 @@ struct ONNXDynamicQuantizeLinearOpLowering
     create.krnl.store(scale, YScale);
 
     // Compute y_zero_point.
-    Value interZeroPoint = create.math.div(create.math.sub(qMin, xMin), scale);
+    Value interZeroPoint = create.math.sub(qMin, create.math.div(xMin, scale));
     // Saturate zero point.
     Value saturateZeroPoint =
         create.onnx.clip(interZeroPoint, qMin, qMax, /*scalarType=*/true);

diff --git a/...icDequantizeLinear_with_canonicalize.mlir → ...amicQuantizeLinear_with_canonicalize.mlir b/...icDequantizeLinear_with_canonicalize.mlir → ...amicQuantizeLinear_with_canonicalize.mlir
@@ -7,7 +7,6 @@ func.func @test_dynamic_quantize_linear(%arg0: tensor<?x2xf32>) -> (tensor<?x2xu
   %y, %y_scale, %y_zero_point = "onnx.DynamicQuantizeLinear"(%arg0) : (tensor<?x2xf32>) -> (tensor<?x2xui8>, tensor<f32>, tensor<ui8>)
   return %y, %y_scale, %y_zero_point:  tensor<?x2xui8>, tensor<f32>, tensor<ui8>
 
-// mlir2FileCheck.py
 // CHECK-DAG:   [[MAP_0_:#.+]] = affine_map<(d0) -> (d0)>
 // CHECK-LABEL:  func.func @test_dynamic_quantize_linear
 // CHECK-SAME:   ([[PARAM_0_:%.+]]: memref<?x2xf32>) -> (memref<?x2xui8>, memref<f32>, memref<ui8>) {
@@ -68,8 +67,8 @@ func.func @test_dynamic_quantize_linear(%arg0: tensor<?x2xf32>) -> (tensor<?x2xu
 // CHECK:           [[VAR_8_:%.+]] = arith.subf [[VAR_5_]], [[VAR_7_]] : f32
 // CHECK:           [[VAR_9_:%.+]] = arith.divf [[VAR_8_]], [[CST_2_dot_550000_]] : f32
 // CHECK:           krnl.store [[VAR_9_]], [[RES_1_]][] : memref<f32>
-// CHECK:           [[VAR_10_:%.+]] = arith.subf [[CST_0_dot_000000_]], [[VAR_7_]] : f32
-// CHECK:           [[VAR_11_:%.+]] = arith.divf [[VAR_10_]], [[VAR_9_]] : f32
+// CHECK:           [[VAR_10_:%.+]] = arith.divf [[VAR_7_]], [[VAR_9_]] : f32
+// CHECK:           [[VAR_11_:%.+]] = arith.subf [[CST_0_dot_000000_]], [[VAR_10_]] : f32
 // CHECK:           [[VAR_12_:%.+]] = arith.cmpf olt, [[VAR_11_]], [[CST_0_dot_000000_]] : f32
 // CHECK:           [[VAR_13_:%.+]] = arith.select [[VAR_12_]], [[CST_0_dot_000000_]], [[VAR_11_]] : f32
 // CHECK:           [[VAR_14_:%.+]] = arith.cmpf olt, [[VAR_13_]], [[CST_2_dot_550000_]] : f32