ZHigh to ONNX optimization is default on. Switch flag from enable to disable

src/Accelerators/NNPA/Compiler/NNPACompilerOptions.cpp

@@ -37,13 +37,13 @@ llvm::cl::opt<bool> nnpaClipToDLFloatRange("nnpa-clip-to-dlfloat-range",
                    "replaced by --nnpa-saturation in the future."),
     llvm::cl::init(true), 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",

src/Accelerators/NNPA/Compiler/NNPACompilerOptions.hpp

@@ -68,7 +68,7 @@ extern llvm::cl::OptionCategory OnnxMlirOptions;
 extern llvm::cl::OptionCategory OnnxMlirCommonOptions;
 extern llvm::cl::opt<onnx_mlir::NNPAEmissionTargetType> nnpaEmissionTarget;
 extern llvm::cl::opt<bool> nnpaClipToDLFloatRange;
-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;

src/Accelerators/NNPA/Compiler/NNPACompilerUtils.cpp

@@ -183,7 +183,7 @@ 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.