Disable the rewriting of Sqrt into Exp and Log (#2088)

Signed-off-by: Tung D. Le <tung@jp.ibm.com> Co-authored-by: Charles Volzka <42243335+cjvolzka@users.noreply.github.com>
onnx · Mar 24, 2023 · 4e1c970 · 4e1c970
1 parent df58b4d
commit 4e1c970
Show file tree

Hide file tree

Showing 3 changed files with 0 additions and 47 deletions.
diff --git a/src/Accelerators/NNPA/Dialect/ZHigh/ZHighOps/Stick/Stick.cpp b/src/Accelerators/NNPA/Dialect/ZHigh/ZHighOps/Stick/Stick.cpp
@@ -136,7 +136,6 @@ void ZHighStickOp::getCanonicalizationPatterns(
   results.insert<StickUnstickSameLayoutRemovalPattern>(context);
   results.insert<StickUnstickDiffLayoutRemovalPattern>(context);
   results.insert<ReplaceONNXLeakyReluPattern>(context);
-  results.insert<ReplaceONNXSqrtPattern>(context);
   results.insert<ReplaceONNXReciprocalSqrtPattern>(context);
   results.insert<ReshapeTransposeReshape2DTo3DSPattern>(context);
   results.insert<ReshapeTransposeReshape3DSTo2DPattern>(context);

diff --git a/src/Accelerators/NNPA/Dialect/ZHigh/ZHighOps/Stick/ZHighStick.td b/src/Accelerators/NNPA/Dialect/ZHigh/ZHighOps/Stick/ZHighStick.td
@@ -80,17 +80,6 @@ def ReplaceONNXLeakyReluPattern: Pat<
    (SameLayout $X, $stickout)]
 >;
 
-// Since zDNN does not support Sqrt(X), we calculate it by using zDNN-supported
-// operations, i.e. Exp and Log.
-// Formulas: `sqrt(X) = exp(log(x)/2) = exp(0.5 * log(x))`
-def ReplaceONNXSqrtPattern: Pat<
-  (ZHighStickOp:$stick (ONNXSqrtOp (ZHighUnstickOp $X)), $layout),
-  (ZHighExpOp (ZHighMulOp (ZHighLogOp $X, (returnType $X)),
-                          (ZHighStickOp (GetConstantOfType<"0.5"> $X), $layout),
-                          (returnType $X))),
-  [(IsStaticShapeTensor $X), (SameLayout $X, $stick)]
->;
-
 // Calulation of `1/sqrt(X)` or reciprocal square root is often found in
 // deep learning models, but zDNN does not support it. Thus, we rewrite it into
 // zDNN-supported operations.

diff --git a/test/mlir/accelerators/nnpa/transform/zhigh-combine.mlir b/test/mlir/accelerators/nnpa/transform/zhigh-combine.mlir
@@ -130,41 +130,6 @@ func.func @donot_replace_leakyrelu(%arg0 : tensor<1x104x104x128xf32, #zhigh.layo
 
 // -----
 
-func.func @replace_sqrt(%arg0 : tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) -> (tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) {
-  %0 = "zhigh.Unstick"(%arg0) : (tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) -> tensor<4x256x1xf32>
-  %1 = "onnx.Sqrt"(%0) : (tensor<4x256x1xf32>) -> tensor<4x256x1xf32>
-  %2 = "zhigh.Stick"(%1) {layout = "3D"} : (tensor<4x256x1xf32>) -> tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>
-  return %2 : tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>
-
-// CHECK-LABEL:  func.func @replace_sqrt
-// CHECK-SAME:   ([[PARAM_0_:%.+]]: tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) -> tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>> {
-// CHECK-DAG:       [[VAR_0_:%.+]] = "zhigh.Log"([[PARAM_0_]]) : (tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) -> tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>
-// CHECK-DAG:       [[VAR_1_:%.+]] = onnx.Constant dense<5.000000e-01> : tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>
-// CHECK:           [[VAR_2_:%.+]] = "zhigh.Stick"([[VAR_1_]]) {layout = "3D"} : (tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) -> tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>
-// CHECK:           [[VAR_3_:%.+]] = "zhigh.Mul"([[VAR_0_]], [[VAR_2_]]) : (tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>, tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) -> tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>
-// CHECK:           [[VAR_4_:%.+]] = "zhigh.Exp"([[VAR_3_]]) : (tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) -> tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>
-// CHECK:           return [[VAR_4_]] : tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>
-// CHECK:         }
-}
-
-// -----
-
-// Do not replace square root because of unknown dimension.
-// In this case, there is no static shape to create a constant of 2.
-func.func @donot_replace_sqrt(%arg0 : tensor<?x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) -> (tensor<?x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) {
-  %0 = "zhigh.Unstick"(%arg0) : (tensor<?x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) -> tensor<?x256x1xf32>
-  %1 = "onnx.Sqrt"(%0) : (tensor<?x256x1xf32>) -> tensor<?x256x1xf32>
-  %2 = "zhigh.Stick"(%1) {layout = "3D"} : (tensor<?x256x1xf32>) -> tensor<?x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>
-  return %2 : tensor<?x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>
-
-// CHECK-LABEL:  func.func @donot_replace_sqrt
-// CHECK: zhigh.Unstick
-// CHECK: onnx.Sqrt
-// CHECK: zhigh.Stick
-}
-
-// -----
-
 func.func @replace_reciprocal_sqrt(%arg0 : tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) -> (tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) {
   %0 = "zhigh.Unstick"(%arg0) : (tensor<4x256x1xf32, #zhigh.layout<{dataLayout = "3D"}>>) -> tensor<4x256x1xf32>
   %1 = "onnx.Sqrt"(%0) : (tensor<4x256x1xf32>) -> tensor<4x256x1xf32>