Merge pull request #1030 from iksnagreb/fix/round-and-clip-thresholds

Fix RoundAndClipThresholds Transformation

src/finn/builder/build_dataflow_steps.py

@@ -121,6 +121,7 @@
 )
 from finn.transformation.streamline import Streamline
 from finn.transformation.streamline.reorder import MakeMaxPoolNHWC
+from finn.transformation.streamline.round_thresholds import RoundAndClipThresholds
 from finn.util.basic import (
     get_rtlsim_trace_depth,
     pyverilate_get_liveness_threshold_cycles,
@@ -503,6 +504,7 @@ def step_minimize_bit_width(model: ModelWrapper, cfg: DataflowBuildConfig):
     if cfg.minimize_bit_width:
         model = model.transform(MinimizeWeightBitWidth())
         model = model.transform(MinimizeAccumulatorWidth())
+        model = model.transform(RoundAndClipThresholds())
         # make sure the changed datatypes are propagated through the network
         model = model.transform(InferDataTypes())
     return model

src/finn/transformation/streamline/round_thresholds.py

@@ -1,4 +1,5 @@
-# Copyright (c) 2020, Xilinx
+# Copyright (c) 2020-2022, Xilinx
+# Copyright (C) 2022-2024, Advanced Micro Devices, Inc.
 # All rights reserved.
 #
 # Redistribution and use in source and binary forms, with or without
@@ -27,42 +28,67 @@
 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 import numpy as np
+from qonnx.core.datatype import DataType
+from qonnx.core.modelwrapper import ModelWrapper
+from qonnx.custom_op.registry import getCustomOp
 from qonnx.transformation.base import Transformation
+from qonnx.transformation.infer_datatypes import InferDataTypes
 
 
 class RoundAndClipThresholds(Transformation):
     """For MultiThreshold nodes operating on integer inputs, round up
     thresholds values to the nearest integer. Additionally, if the input
-    is unsigned, sets negative thresholds to zero."""
+    is unsigned, sets negative thresholds to zero. Type-casts thresholds (back)
+    to the float32 container type (this is separate from the quantization
+    annotation). Runs InferDataTypes() afterward to propagate any changes to the
+    quantization data types."""
 
-    def apply(self, model):
+    def apply(self, model: ModelWrapper):  # noqa
         graph = model.graph
         graph_modified = False
-        for n in graph.node:
-            if n.op_type == "MultiThreshold":
-                idtype = model.get_tensor_datatype(n.input[0])
-                T = model.get_initializer(n.input[1])
-                Tnew = np.ceil(T)
-                if idtype.is_integer() and (T != Tnew).any():
-                    # round up the thresholds to nearest integer
-                    model.set_initializer(n.input[1], Tnew)
-                    # use same datatype as inputs for thresholds
-                    model.set_tensor_datatype(n.input[1], idtype)
+        for index, node in enumerate(graph.node):
+            op_type = node.op_type
+            if op_type == "MultiThreshold" or op_type.startswith("Thresholding"):
+                thresholds = model.get_initializer(node.input[1])
+                if thresholds is None:
+                    continue
+                dtype = model.get_tensor_datatype(node.input[0])
+                # This transformation only applies to thresholding operations
+                # operating on integer inputs
+                if not dtype.is_integer():
+                    continue
+                # Round thresholds up to nearest integer and clip thresholds
+                # outside the input range
+                #   Note: This might promote the thresholds to float64 and
+                #   introduce extra inaccuracies due to large integers not being
+                #   exactly representable in floating-point representation.
+                #   See for example: np.ceil(np.float32(16777217)) == 16777216
+                new_thresholds = np.clip(np.ceil(thresholds), dtype.min(), dtype.max() + 1)
+                # Convert back to the preferred float32 container type
+                new_thresholds = new_thresholds.astype(np.float32)
+                # Insert the rounded and clipped thresholds back into the model
+                model.set_initializer(node.input[1], new_thresholds)
+                # The rounded and clipped thresholds now fit into a data type
+                # that is one bit bigger than the input datatype
+                # Determine new max_value
+                max_val = dtype.max() + 1
+                if not dtype.signed():
+                    tdt = DataType.get_smallest_possible(max_val)
+                else:
+                    tdt = DataType.get_smallest_possible(-(max_val) - 1)
+                model.set_tensor_datatype(node.input[1], tdt)
+                # If hw op we need to set the weight data type attribute as well
+                if op_type.startswith("Thresholding"):
+                    inst = getCustomOp(node)
+                    inst.set_nodeattr("weightDataType", tdt.name)
+                # ones
+                if np.any(new_thresholds != thresholds):
+                    # Track the graph has been modified to inform the transform
+                    # container to exhaustively repeat this transformation until
+                    # no changes are possible
                     graph_modified = True
-                if idtype.is_integer() and not idtype.signed() and (Tnew < 0).any():
-                    # clip any negative thresholds if input is unsigned
-                    Tnew = np.clip(Tnew, 0, None)
-                    model.set_initializer(n.input[1], Tnew)
-                    # use same datatype as inputs for thresholds
-                    model.set_tensor_datatype(n.input[1], idtype)
-                    graph_modified = True
-                if idtype.is_integer() and (
-                    (Tnew < (idtype.min() - 1)).any() or (Tnew > (idtype.max() + 1)).any()
-                ):
-                    # clip any large thresholds to input range + 1
-                    Tnew = np.clip(Tnew, idtype.min() - 1, idtype.max() + 1)
-                    model.set_initializer(n.input[1], Tnew)
-                    # use same datatype as inputs for thresholds
-                    model.set_tensor_datatype(n.input[1], idtype)
-                    graph_modified = True
-        return (model, graph_modified)
+                    # Immediately exit here to propagate the data type changes
+                    # before considering the next node
+                    break
+        model = model.transform(InferDataTypes())
+        return model, graph_modified

tests/end2end/test_end2end_bnn_pynq.py

@@ -94,6 +94,7 @@
     MakeMaxPoolNHWC,
     MoveScalarLinearPastInvariants,
 )
+from finn.transformation.streamline.round_thresholds import RoundAndClipThresholds
 from finn.util.basic import get_finn_root, make_build_dir, test_board_map
 from finn.util.pytorch import ToTensor
 from finn.util.test import (
@@ -672,6 +673,7 @@ def test_minimize_bit_width(self, topology, wbits, abits, board):
         model = load_test_checkpoint_or_skip(prev_chkpt_name)
         model = model.transform(MinimizeAccumulatorWidth())
         model = model.transform(MinimizeWeightBitWidth())
+        model = model.transform(RoundAndClipThresholds())
         curr_chkpt_name = get_checkpoint_name(topology, wbits, abits, "minimize_bit_width")
         model.save(curr_chkpt_name)
 

tests/end2end/test_end2end_mobilenet_v1.py

@@ -353,6 +353,7 @@ def test_end2end_mobilenet_minimize_bit_width():
     model = load_test_checkpoint_or_skip(build_dir + "/end2end_mobilenet_folded.onnx")
     model = model.transform(MinimizeAccumulatorWidth())
     model = model.transform(MinimizeWeightBitWidth())
+    model = model.transform(RoundAndClipThresholds())
     model.save(build_dir + "/end2end_mobilenet_minimize_bitwidth.onnx")
 
 

tests/fpgadataflow/test_fpgadataflow_thresholding.py

@@ -49,6 +49,7 @@
 from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
 from finn.transformation.fpgadataflow.set_exec_mode import SetExecMode
 from finn.transformation.fpgadataflow.specialize_layers import SpecializeLayers
+from finn.transformation.streamline.round_thresholds import RoundAndClipThresholds
 
 test_fpga_part = "xczu3eg-sbva484-1-e"
 target_clk_ns = 5
@@ -133,10 +134,8 @@ def make_single_multithresholding_modelwrapper(
 @pytest.mark.parametrize(
     "idt_tdt_cfg",
     [
-        (DataType["INT8"], DataType["INT8"]),
-        (DataType["INT8"], DataType["INT9"]),
-        (DataType["UINT5"], DataType["UINT5"]),
-        (DataType["UINT5"], DataType["UINT6"]),
+        (DataType["INT8"], DataType["INT25"]),
+        (DataType["UINT5"], DataType["UINT8"]),
     ],
 )
 @pytest.mark.parametrize("fold", [-1, 1, 2])
@@ -145,6 +144,7 @@ def make_single_multithresholding_modelwrapper(
 @pytest.mark.parametrize("impl_style", ["hls", "rtl"])
 @pytest.mark.parametrize("exec_mode", ["cppsim", "rtlsim"])
 @pytest.mark.parametrize("mem_mode", ["internal_embedded", "internal_decoupled"])
+@pytest.mark.parametrize("round_thresh", [True, False])
 @pytest.mark.fpgadataflow
 @pytest.mark.vivado
 @pytest.mark.slow
@@ -159,6 +159,7 @@ def test_fpgadataflow_thresholding(
     impl_style,
     exec_mode,
     mem_mode,
+    round_thresh,
 ):
     # the mem_mode parameter can only be used for the hls thresholding
     # so the test will only be executed once for impl_style=rtl and once skipped
@@ -234,6 +235,8 @@ def test_fpgadataflow_thresholding(
     node = model.get_nodes_by_op_type(model.graph.node[0].op_type)[0]
     inst = getCustomOp(node)
     inst.set_nodeattr("PE", pe)
+    if round_thresh is True:
+        model = model.transform(RoundAndClipThresholds())
     model = model.transform(GiveUniqueNodeNames())
 
     if impl_style == "hls":