Remove the compile option -nnpa-clip-to-dlfloat-range (#3075)

Signed-off-by: Tung D. Le <[email protected]>

src/Accelerators/NNPA/Compiler/NNPACompilerOptions.cpp

@@ -30,13 +30,6 @@ 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> nnpaClipToDLFloatRange("nnpa-clip-to-dlfloat-range",
-    llvm::cl::desc("Clip CPU tensors to dlfloat range before stickification to "
-                   "avoid out-of-range. Only clip Softmax inputs at this "
-                   "moment. Default is true. This option will be removed and "
-                   "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::desc(
         "Enabling this will convert a pattern `stick -> element-wise op -> "

src/Accelerators/NNPA/Compiler/NNPACompilerOptions.hpp

@@ -67,7 +67,6 @@ 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> nnpaClipToDLFloatRange;
 extern llvm::cl::opt<bool> nnpaEnableZHighToOnnx;
 extern llvm::cl::opt<bool> nnpaEnableZHighDecomposeStickUnstick;
 extern llvm::cl::opt<bool> nnpaEnableCompilerStickUnstick;

src/Accelerators/NNPA/Compiler/NNPACompilerUtils.cpp

@@ -145,16 +145,6 @@ void addONNXToZHighPasses(mlir::PassManager &pm) {
   pm.addNestedPass<func::FuncOp>(onnx_mlir::createShapeInferencePass());
   pm.addPass(mlir::createCanonicalizerPass());
 
-  // Clip zhigh.Stick inputs if required. This is to avoid out-of-range of
-  // dlfloat. Do constant propagation after clipping to remove ONNX ops used for
-  // clipping such as ONNXMax if applicable.
-  // This pass will be removed and replaced by nnpa-saturation in the future.
-  if (!nnpaEnableSaturation && nnpaClipToDLFloatRange) {
-    pm.addNestedPass<func::FuncOp>(
-        onnx_mlir::zhigh::createZHighClipToDLFloatPass());
-    pm.addNestedPass<func::FuncOp>(onnx_mlir::createConstPropONNXToONNXPass());
-  }
-
   // One more call to ONNX shape inference/canonicalization/... to update shape
   // if possible.
   if (enableONNXHybridPass) {

src/Accelerators/NNPA/NNPAAccelerator.cpp

@@ -125,10 +125,6 @@ void NNPAAccelerator::registerPasses(int optLevel) const {
     return onnx_mlir::zhigh::createZHighLayoutPropagationPass();
   });
 
-  mlir::registerPass([]() -> std::unique_ptr<mlir::Pass> {
-    return onnx_mlir::zhigh::createZHighClipToDLFloatPass();
-  });
-
   mlir::registerPass([]() -> std::unique_ptr<mlir::Pass> {
     return onnx_mlir::zhigh::createZHighDecomposeStickUnstickPass();
   });

src/Accelerators/NNPA/Pass/NNPAPasses.hpp

@@ -55,9 +55,6 @@ std::unique_ptr<mlir::Pass> createZHighConstPropagationPass();
 std::unique_ptr<mlir::Pass> createZHighScrubDisposablePass(
     bool closeAfter = true);
 
-/// Pass for clipping values to dlfloat before stickification at ZHighIR.
-std::unique_ptr<mlir::Pass> createZHighClipToDLFloatPass();
-
 /// Pass for decomposing stick/unstick at ZHighIR.
 std::unique_ptr<mlir::Pass> createZHighDecomposeStickUnstickPass();
 

src/Accelerators/NNPA/Transform/ZHigh/CMakeLists.txt

@@ -38,17 +38,6 @@ add_onnx_mlir_library(OMZHighLayoutPropagation
   ${NNPA_INCLUDE_PATH}
   )
 
-add_onnx_mlir_rewriter(ZHighClipToDLFloat)
-add_onnx_mlir_library(OMZHighClipToDLFloat
-  ZHighClipToDLFloat.cpp
-
-  LINK_LIBS PUBLIC
-  MLIRRewrite
-  MLIRTransformUtils
-  OMZHighOps
-  OMONNXOps
-  )
-
 add_onnx_mlir_rewriter(ZHighDecomposeStickUnstick)
 add_onnx_mlir_library(OMZHighDecomposeStickUnstick
   ZHighDecomposeStickUnstick.cpp