From 2bede1aa5c5593e3bda95243ab8cf8547c984b48 Mon Sep 17 00:00:00 2001
From: Alexandre Eichenberger <alexe@us.ibm.com>
Date: Thu, 13 Feb 2025 16:05:39 -0500
Subject: [PATCH] ZHigh to ONNX optimization is default on. Switch flag from
 enable to disable (#3074)

Signed-off-by: Alexandre Eichenberger <alexe@us.ibm.com>
---
 src/Accelerators/NNPA/Compiler/NNPACompilerOptions.cpp | 8 ++++----
 src/Accelerators/NNPA/Compiler/NNPACompilerOptions.hpp | 2 +-
 src/Accelerators/NNPA/Compiler/NNPACompilerUtils.cpp   | 2 +-
 3 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/src/Accelerators/NNPA/Compiler/NNPACompilerOptions.cpp b/src/Accelerators/NNPA/Compiler/NNPACompilerOptions.cpp
index 29b48073ac..2461e39ea1 100644
--- a/src/Accelerators/NNPA/Compiler/NNPACompilerOptions.cpp
+++ b/src/Accelerators/NNPA/Compiler/NNPACompilerOptions.cpp
@@ -30,13 +30,13 @@ llvm::cl::opt<NNPAEmissionTargetType> nnpaEmissionTarget(
         clEnumVal(EmitZNONE, "Do not emit NNPA-related target (default)")),
     llvm::cl::init(EmitZNONE), llvm::cl::cat(OnnxMlirOptions));
 
-llvm::cl::opt<bool> nnpaEnableZHighToOnnx("enable-zhigh-to-onnx",
+llvm::cl::opt<bool> nnpaDisableZHighToOnnx("disable-zhigh-to-onnx",
     llvm::cl::desc(
-        "Enabling this will convert a pattern `stick -> element-wise op -> "
+        "By default we convert a pattern `stick -> element-wise op -> "
         "unstick` back to an ONNX element-wise op. This conversion is called "
         "after applying all optimizations to remove stick/unstick at ZHigh "
-        "level. Default is true."),
-    llvm::cl::init(true), llvm::cl::cat(OnnxMlirOptions));
+        "level. Use this option to disable this optimization."),
+    llvm::cl::init(false), llvm::cl::cat(OnnxMlirOptions));
 
 llvm::cl::opt<bool> nnpaEnableZHighDecomposeStickUnstick(
     "enable-zhigh-decompose-stick-unstick",
diff --git a/src/Accelerators/NNPA/Compiler/NNPACompilerOptions.hpp b/src/Accelerators/NNPA/Compiler/NNPACompilerOptions.hpp
index 1e07a92d7e..3958cc2bd5 100644
--- a/src/Accelerators/NNPA/Compiler/NNPACompilerOptions.hpp
+++ b/src/Accelerators/NNPA/Compiler/NNPACompilerOptions.hpp
@@ -67,7 +67,7 @@ typedef enum {
 extern llvm::cl::OptionCategory OnnxMlirOptions;
 extern llvm::cl::OptionCategory OnnxMlirCommonOptions;
 extern llvm::cl::opt<onnx_mlir::NNPAEmissionTargetType> nnpaEmissionTarget;
-extern llvm::cl::opt<bool> nnpaEnableZHighToOnnx;
+extern llvm::cl::opt<bool> nnpaDisableZHighToOnnx;
 extern llvm::cl::opt<bool> nnpaEnableZHighDecomposeStickUnstick;
 extern llvm::cl::opt<bool> nnpaEnableCompilerStickUnstick;
 extern llvm::cl::opt<bool> nnpaEnableScalarBcastBinary;
diff --git a/src/Accelerators/NNPA/Compiler/NNPACompilerUtils.cpp b/src/Accelerators/NNPA/Compiler/NNPACompilerUtils.cpp
index 141e733937..1d52d60700 100644
--- a/src/Accelerators/NNPA/Compiler/NNPACompilerUtils.cpp
+++ b/src/Accelerators/NNPA/Compiler/NNPACompilerUtils.cpp
@@ -173,7 +173,7 @@ void addONNXToZHighPasses(mlir::PassManager &pm) {
   // sub, ...) that are of `stick -> light-weight op -> unstick`, it's better to
   // use CPU instead of NNPA to avoid stick/unstick. CPU is efficient to handle
   // these ops, e.g vectorize the computation.
-  if (nnpaEnableZHighToOnnx)
+  if (!nnpaDisableZHighToOnnx)
     pm.addNestedPass<func::FuncOp>(onnx_mlir::createZHighToONNXPass());
 
   // Constant propagation at ZHighIR: constant stickify.