[FixedPt] add the fixed-point quantize transformation and unit tests

src/qonnx/transformation/fixedpt_quantize.py

@@ -0,0 +1,118 @@
+# Copyright (c) 2025 Advanced Micro Devices, Inc.
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+#   list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+#   this list of conditions and the following disclaimer in the documentation
+#   and/or other materials provided with the distribution.
+#
+# * Neither the name of qonnx nor the names of its
+#   contributors may be used to endorse or promote products derived from
+#   this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+import numpy as np
+from warnings import warn
+
+from qonnx.core.modelwrapper import ModelWrapper
+from qonnx.transformation.base import Transformation
+from qonnx.custom_op.general.intquant import resolve_rounding_mode
+from qonnx.core.datatype import DataType
+
+
+def default_op_filter(op):
+    return op.op_type in ["Add", "Mul"]
+
+
+class FixedPointQuantizeParamsFromDict(Transformation):
+    """
+    Quantize model parameters to a given fixed-point representation.
+    The self.max_err dictionary stores the maximum error for each quantized input after calling.
+    Parameters:
+        fixedpt_dict: Dictionary containing tensor names and their corresponding target fixed-point data type or its canonical name
+        rounding_mode: Rounding mode used for conversion into fixed point.
+                       Default is "ROUND",
+                       possible values: ["ROUND", "HALF_EVEN", "CEIL", "FLOOR", "UP", "DOWN", "HALF_UP", "HALF_DOWN"]
+    """
+
+    def __init__(self, fixedpt_dict, rounding_mode="ROUND"):
+        super().__init__()
+        self.fixedpt_dict = fixedpt_dict
+        self.max_err = {}
+        self.round_func = resolve_rounding_mode(rounding_mode)
+
+    def apply(self, model: ModelWrapper):
+        for tname, tdtype in self.fixedpt_dict.items():
+            if (in1_t := model.get_initializer(tname)) is not None:
+                if isinstance(tdtype, str):
+                    tdtype = DataType[tdtype]
+                current_dtype = model.get_tensor_datatype(tname)
+                if current_dtype.is_fixed_point():
+                    warn(f"Tensor {tname} is already a {current_dtype.get_canonical_name()} type. Recasting to {tdtype.get_canonical_name()}")
+
+                in1_t_new = self.round_func(in1_t.astype(np.float32) / tdtype.scale_factor()) * tdtype.scale_factor()
+                if (in1_t_new.max() > tdtype.max()) or (in1_t_new.min() < tdtype.min()):
+                    warn(
+                        f"Range of {tname} [{in1_t_new.min():.3f}, {in1_t_new.max():.3f}] greater than"
+                        f"{tdtype.get_canonical_name()} [{tdtype.min():.3f}, {tdtype:.max():.3f}], clipping.")
+                    in1_t_new = np.clip(in1_t_new, tdtype.min(), tdtype.max())
+                model.set_tensor_datatype(tname, tdtype)
+                model.set_initializer(tname, in1_t_new)
+
+                self.max_err[tname] = np.linalg.norm(in1_t.flatten() - in1_t_new.flatten(), ord=np.inf)
+
+        return (model, False)
+
+class FixedPointQuantizeParams(Transformation):
+    """
+    Quantize model parameters to a given fixed-point representation.
+    Identifies specific operations in a model (e.g., "Add", "Mul") using a filter function,
+    and quantizes any non-quantized input initializers to the given fixed-point representation.
+    The self.max_err dictionary stores the maximum error for each quantized input after calling.
+    Parameters:
+        fixedpt_dtype: The fixed-point data type or its canonical name to use for quantization.
+        op_filter: A lambda function to filter operations in the model graph
+                   that should be quantized. By default, it selects operations
+                   of type "Add" and "Mul".
+        rounding_mode: Rounding mode used for conversion into fixed point.
+                       Default is "ROUND",
+                       possible values: ["ROUND", "HALF_EVEN", "CEIL", "FLOOR", "UP", "DOWN", "HALF_UP", "HALF_DOWN"]
+    """
+    def __init__(self, fixedpt_dtype, op_filter=default_op_filter, rounding_mode="ROUND"):
+        super().__init__()
+        if isinstance(fixedpt_dtype, str):
+            self.fixedpt_dtype = DataType[fixedpt_dtype]
+        else:
+            self.fixedpt_dtype = fixedpt_dtype
+        self.op_filter = op_filter
+        self.max_err = {}
+        self.rounding_mode = rounding_mode
+
+    def apply(self, model: ModelWrapper):
+        ops = [op for op in model.graph.node if self.op_filter(op)]
+        fixedpt_dict = {}
+        for op in ops:
+            for inp_name in op.input:
+                if (model.get_initializer(inp_name)) is not None:
+                    fixedpt_dict[inp_name] = self.fixedpt_dtype
+
+        fxpdict_transform = FixedPointQuantizeParamsFromDict(fixedpt_dict, self.rounding_mode)
+        model = model.transform(fxpdict_transform)
+        self.max_err = fxpdict_transform.max_err
+
+        return (model, False)