fix: preserve original dtype for all-zero input type inference

Author: lee2716

Deeploy/DeeployTypes.py

@@ -3105,16 +3105,9 @@ def _exportGraph(self, folderPath, fileName):
             raise OSError(f"Error exporting the context to: {absoluteOnnxPath}")
 
         # VJUNG: ONNX-Graphsurgeon needs tensors to be in their export types
-<<<<<<< HEAD
-        constTensors = [tensor for tensor in self.graph.tensors().values() if isinstance(tensor, gs.Constant)]
-        for tensor in constTensors:
-            if tensor.dtype != tensor.export_dtype:
-=======
-        # Added hasattr check for compatibility with different onnx-graphsurgeon versions
         constTensors = [tensor for tensor in self.graph.tensors().values() if isinstance(tensor, gs.Constant)]
         for tensor in constTensors:
             if hasattr(tensor, 'export_dtype') and tensor.dtype != tensor.export_dtype:
->>>>>>> 937e3cb3 (refactor: restore Snitch framework code to origin/devel)
                 tensor.values = tensor.values.astype(tensor.export_dtype)
 
         model = gs.export_onnx(self.graph)

DeeployTest/generateNetwork.py

@@ -73,7 +73,8 @@ def generateNetwork(args):
         test_inputs, test_outputs, graph = generateDebugConfig(inputs, outputs, activations, graph)
 
     else:
-        # Load as float64 and infer types later
+        # Load as float64 for uniform handling, but preserve original dtypes for type inference
+        test_input_original_dtypes = [inputs[x].dtype for x in inputs.files]
         test_inputs = [inputs[x].reshape(-1).astype(np.float64) for x in inputs.files]
         test_outputs = [outputs[x].reshape(-1).astype(np.float64) for x in outputs.files]
 
@@ -122,7 +123,8 @@ def generateNetwork(args):
 
             _type = PointerClass(_type)
         else:
-            _type, offset = inferTypeAndOffset(values, signProp)
+            original_dtype = test_input_original_dtypes[index] if index < len(test_input_original_dtypes) else None
+            _type, offset = inferTypeAndOffset(values, signProp, original_dtype = original_dtype)
 
         inputTypes[f"input_{index}"] = _type
         inputOffsets[f"input_{index}"] = offset

DeeployTest/testMVP.py

@@ -68,7 +68,8 @@ def setupDeployer(graph: gs.Graph, memoryHierarchy: MemoryHierarchy, defaultTarg
     inputs = np.load(f'{args.dir}/inputs.npz')
     tensors = graph.tensors()
 
-    # Load as int64 and infer types later
+    # Load as float64 for uniform handling, but preserve original dtypes for type inference
+    test_input_original_dtypes = [inputs[x].dtype for x in inputs.files]
     test_inputs = [inputs[x].reshape(-1).astype(np.float64) for x in inputs.files]
 
     platform, signProp = mapPlatform(args.platform)
@@ -83,7 +84,8 @@ def setupDeployer(graph: gs.Graph, memoryHierarchy: MemoryHierarchy, defaultTarg
         cluster.n_cores = args.cores
 
     for index, num in enumerate(test_inputs):
-        _type, offset = inferTypeAndOffset(num, signProp)
+        original_dtype = test_input_original_dtypes[index] if index < len(test_input_original_dtypes) else None
+        _type, offset = inferTypeAndOffset(num, signProp, original_dtype = original_dtype)
         inputTypes[f"input_{index}"] = _type
         inputOffsets[f"input_{index}"] = offset
 
@@ -241,7 +243,8 @@ def setupDeployer(graph: gs.Graph, memoryHierarchy: MemoryHierarchy, defaultTarg
     if args.debug:
         test_inputs, test_outputs, graph = generateDebugConfig(inputs, outputs, activations, graph)
     else:
-        # Load as int64 and infer types later
+        # Load as float64 for uniform handling, but preserve original dtypes for type inference
+        test_input_original_dtypes = [inputs[x].dtype for x in inputs.files]
         test_inputs = [inputs[x].reshape(-1).astype(np.float64) for x in inputs.files]
         test_outputs = [outputs[x].reshape(-1).astype(np.float64) for x in outputs.files]
 
@@ -280,7 +283,8 @@ def setupDeployer(graph: gs.Graph, memoryHierarchy: MemoryHierarchy, defaultTarg
     log.debug(f"Deployer: {deployer}")
 
     for index, num in enumerate(test_inputs):
-        _type, offset = inferTypeAndOffset(num, signProp)
+        original_dtype = test_input_original_dtypes[index] if index < len(test_input_original_dtypes) else None
+        _type, offset = inferTypeAndOffset(num, signProp, original_dtype = original_dtype)
         inputTypes[f"input_{index}"] = _type
         inputOffsets[f"input_{index}"] = offset
 

DeeployTest/testUtils/typeMapping.py

@@ -42,12 +42,20 @@ def isInteger(x: npt.NDArray) -> bool:
     return np.abs((x.astype(int) - x)).max() <= 0.001
 
 
-def inferMinimalType(values: np.ndarray, default: Type[BaseType] = int8_t) -> Type[BaseType]:
+def inferMinimalType(values: np.ndarray,
+                     default: Type[BaseType] = int8_t,
+                     original_dtype: np.dtype = None) -> Type[BaseType]:
     # WIESEP: We cannot do type inference for empty arrays.
     if np.prod(values.shape) == 0:
         print(f"Warning: Empty input array for type inference for {values}!")
         return default
 
+    # For all-zero arrays, use original dtype to distinguish int vs float
+    if np.all(values == 0) and original_dtype is not None:
+        if np.issubdtype(original_dtype, np.floating):
+            return minimalFloatType(values)
+        return minimalIntegerType(values)
+
     if isInteger(values):
         return minimalIntegerType(values)
     else:
@@ -67,7 +75,9 @@ def signPropTypeAndOffset(_type: Type[IntegerImmediate]) -> Tuple[Type[IntegerIm
     return signedType, 2**(signedType.typeWidth - 1)
 
 
-def inferTypeAndOffset(values: np.ndarray, signProp: bool = False) -> Tuple[Type[Pointer], int]:
+def inferTypeAndOffset(values: np.ndarray,
+                       signProp: bool = False,
+                       original_dtype: np.dtype = None) -> Tuple[Type[Pointer], int]:
     """Infers the data type of the provided input array.
 
     Parameters
@@ -77,13 +87,17 @@ def inferTypeAndOffset(values: np.ndarray, signProp: bool = False) -> Tuple[Type
 
     signProp : bool
         Whether to consider signedness when inferring the data type.
+
+    original_dtype : np.dtype, optional
+        Original numpy dtype before float64 cast, used to resolve all-zero ambiguity.
+
     Returns
     -------
     Tuple[Type[BaseType], int]
         The inferred type and offset
     """
 
-    _type = inferMinimalType(values)
+    _type = inferMinimalType(values, original_dtype = original_dtype)
 
     if signProp and issubclass(_type, IntegerImmediate):
         _type, offset = signPropTypeAndOffset(_type)