Merge pull request #393 from Xilinx/feature/faster_shape_inf

Faster&smaller shape inference

docker/Dockerfile.finn

@@ -86,7 +86,7 @@ RUN pip install -e git+https://github.com/fbcotter/dataset_loading.git@0.0.4#egg
 
 # git-based Python repo dependencies
 # these are installed in editable mode for easier co-development
-ARG FINN_BASE_COMMIT="1fdf06c068f77ed5a312cd3a6edad098f64b09ed"
+ARG FINN_BASE_COMMIT="7c2603a95e90e4de2575020e575c24eab6a15889"
 ARG FINN_EXP_COMMIT="f82c0d9868bb88ea045dfadb28508d327d287221"
 ARG BREVITAS_COMMIT="462f86cdc60f9915baf13afd1676fb21da44c2ee"
 ARG PYVERILATOR_COMMIT="0c3eb9343500fc1352a02c020a736c8c2db47e8e"

src/finn/custom_op/fpgadataflow/addstreams_batch.py

@@ -29,7 +29,6 @@
 import numpy as np
 import os
 import warnings
-from onnx import TensorProto, helper
 
 from finn.core.datatype import DataType
 from finn.custom_op.fpgadataflow.hlscustomop import HLSCustomOp
@@ -84,19 +83,7 @@ def make_shape_compatible_op(self, model):
         assert ishape == exp_ishape, "Unexpected input1 shape."
         ishape = tuple(model.get_tensor_shape(self.onnx_node.input[1]))
         assert ishape == exp_ishape, "Unexpected input2 shape."
-        # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.float32)
-        return helper.make_node(
-            "Constant",
-            inputs=[],
-            outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.FLOAT,
-                dims=values.shape,
-                vals=values.flatten().astype(float),
-            ),
-        )
+        return super().make_const_shape_op(oshape)
 
     def infer_node_datatype(self, model):
         node = self.onnx_node

src/finn/custom_op/fpgadataflow/channelwise_op_batch.py

@@ -30,7 +30,6 @@
 import os
 import warnings
 from math import ceil
-from onnx import TensorProto, helper
 
 from finn.core.datatype import DataType
 from finn.custom_op.fpgadataflow.hlscustomop import HLSCustomOp
@@ -125,18 +124,7 @@ def calc_tmem(self):
     def make_shape_compatible_op(self, model):
         oshape = self.get_normal_output_shape()
         # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.float32)
-        return helper.make_node(
-            "Constant",
-            inputs=[],
-            outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.FLOAT,
-                dims=values.shape,
-                vals=values.flatten().astype(float),
-            ),
-        )
+        return super().make_const_shape_op(oshape)
 
     def infer_node_datatype(self, model):
         node = self.onnx_node

src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py

@@ -29,7 +29,6 @@
 import math
 import numpy as np
 import os
-from onnx import TensorProto, helper
 
 from finn.core.datatype import DataType
 from finn.custom_op.fpgadataflow.hlscustomop import HLSCustomOp
@@ -148,18 +147,7 @@ def make_shape_compatible_op(self, model):
         ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
         assert ishape == exp_ishape, "Unexpect input shape for ConvInpGen."
         # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.float32)
-        return helper.make_node(
-            "Constant",
-            inputs=[],
-            outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.FLOAT,
-                dims=values.shape,
-                vals=values.flatten().astype(float),
-            ),
-        )
+        return super().make_const_shape_op(oshape)
 
     def infer_node_datatype(self, model):
         node = self.onnx_node

src/finn/custom_op/fpgadataflow/convolutioninputgenerator1d.py

@@ -29,7 +29,6 @@
 import math
 import numpy as np
 import os
-from onnx import TensorProto, helper
 
 from finn.core.datatype import DataType
 from finn.custom_op.fpgadataflow.hlscustomop import HLSCustomOp
@@ -137,19 +136,7 @@ def make_shape_compatible_op(self, model):
         oshape = self.get_normal_output_shape()
         ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
         assert ishape == exp_ishape, "Unexpect input shape for ConvInpGen."
-        # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.float32)
-        return helper.make_node(
-            "Constant",
-            inputs=[],
-            outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.FLOAT,
-                dims=values.shape,
-                vals=values.flatten().astype(float),
-            ),
-        )
+        return super().make_const_shape_op(oshape)
 
     def infer_node_datatype(self, model):
         node = self.onnx_node

src/finn/custom_op/fpgadataflow/downsampler.py

@@ -1,7 +1,6 @@
 import numpy as np
 import os
 import warnings
-from onnx import TensorProto, helper
 
 from finn.core.datatype import DataType
 from finn.custom_op.fpgadataflow.hlscustomop import HLSCustomOp
@@ -83,19 +82,7 @@ def make_shape_compatible_op(self, model):
         oshape = self.get_normal_output_shape()
         ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
         assert ishape == exp_ishape, "Unexpect input shape for DownSampler."
-        # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.float32)
-        return helper.make_node(
-            "Constant",
-            inputs=[],
-            outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.FLOAT,
-                dims=values.shape,
-                vals=values.flatten().astype(float),
-            ),
-        )
+        return super().make_const_shape_op(oshape)
 
     def infer_node_datatype(self, model):
         node = self.onnx_node

src/finn/custom_op/fpgadataflow/fmpadding_batch.py

@@ -1,7 +1,6 @@
 import numpy as np
 import os
 import warnings
-from onnx import TensorProto, helper
 
 from finn.core.datatype import DataType
 from finn.custom_op.fpgadataflow.hlscustomop import HLSCustomOp
@@ -99,19 +98,7 @@ def make_shape_compatible_op(self, model):
         oshape = self.get_normal_output_shape()
         ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
         assert ishape == exp_ishape, "Unexpect input shape for SameResize."
-        # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.float32)
-        return helper.make_node(
-            "Constant",
-            inputs=[],
-            outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.FLOAT,
-                dims=values.shape,
-                vals=values.flatten().astype(float),
-            ),
-        )
+        return super().make_const_shape_op(oshape)
 
     def infer_node_datatype(self, model):
         node = self.onnx_node

src/finn/custom_op/fpgadataflow/globalaccpool_batch.py

@@ -29,7 +29,6 @@
 import numpy as np
 import os
 import warnings
-from onnx import TensorProto, helper
 
 from finn.core.datatype import DataType
 from finn.custom_op.fpgadataflow.hlscustomop import HLSCustomOp
@@ -95,19 +94,7 @@ def make_shape_compatible_op(self, model):
         oshape = self.get_normal_output_shape()
         ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
         assert ishape == exp_ishape, "Unexpected input shape."
-        # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.float32)
-        return helper.make_node(
-            "Constant",
-            inputs=[],
-            outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.FLOAT,
-                dims=values.shape,
-                vals=values.flatten(),
-            ),
-        )
+        return super().make_const_shape_op(oshape)
 
     def infer_node_datatype(self, model):
         node = self.onnx_node

src/finn/custom_op/fpgadataflow/iodma.py

@@ -29,7 +29,6 @@
 import math
 import numpy as np
 import warnings
-from onnx import TensorProto, helper
 
 from finn.core.datatype import DataType
 from finn.custom_op.fpgadataflow.hlscustomop import HLSCustomOp
@@ -146,19 +145,7 @@ def make_shape_compatible_op(self, model):
         oshape = self.get_normal_output_shape()
         ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
         assert ishape == exp_ishape, "Unexpected input shape."
-        # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.float32)
-        return helper.make_node(
-            "Constant",
-            inputs=[],
-            outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.FLOAT,
-                dims=values.shape,
-                vals=values.flatten().astype(float),
-            ),
-        )
+        return super().make_const_shape_op(oshape)
 
     def infer_node_datatype(self, model):
         node = self.onnx_node

src/finn/custom_op/fpgadataflow/labelselect_batch.py

@@ -102,18 +102,14 @@ def make_shape_compatible_op(self, model):
         oshape = self.get_normal_output_shape()
         ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
         assert ishape == exp_ishape, "Unexpected input shape."
-        # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.int64)
         return helper.make_node(
-            "Constant",
+            "RandomNormal",
             inputs=[],
             outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.INT64,
-                dims=values.shape,
-                vals=values.flatten(),
-            ),
+            mean=0.0,
+            scale=1.0,
+            dtype=TensorProto.INT64,
+            shape=list(oshape),
         )
 
     def infer_node_datatype(self, model):

src/finn/custom_op/fpgadataflow/pool_batch.py

@@ -28,7 +28,6 @@
 
 import numpy as np
 import os
-from onnx import TensorProto, helper
 
 from finn.core.datatype import DataType
 from finn.custom_op.fpgadataflow.hlscustomop import HLSCustomOp
@@ -163,19 +162,7 @@ def make_shape_compatible_op(self, model):
         oshape = self.get_normal_output_shape()
         ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
         assert ishape == exp_ishape, "Unexpected input shape for Pool_Batch."
-        # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.float32)
-        return helper.make_node(
-            "Constant",
-            inputs=[],
-            outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.FLOAT,
-                dims=values.shape,
-                vals=values.flatten().astype(float),
-            ),
-        )
+        return super().make_const_shape_op(oshape)
 
     def infer_node_datatype(self, model):
         node = self.onnx_node

src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py

@@ -30,7 +30,6 @@
 import numpy as np
 import os
 import warnings
-from onnx import TensorProto, helper
 
 from finn.core.datatype import DataType
 from finn.custom_op.fpgadataflow.hlscustomop import HLSCustomOp
@@ -165,19 +164,7 @@ def make_shape_compatible_op(self, model):
         oshape = self.get_normal_output_shape()
         ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
         assert ishape == tuple(exp_ishape), "Unexpect input shape for StreamingDWC."
-        # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.float32)
-        return helper.make_node(
-            "Constant",
-            inputs=[],
-            outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.FLOAT,
-                dims=values.shape,
-                vals=values.flatten().astype(float),
-            ),
-        )
+        return super().make_const_shape_op(oshape)
 
     def infer_node_datatype(self, model):
         node = self.onnx_node

src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py

@@ -31,7 +31,6 @@
 import os
 import textwrap
 import warnings
-from onnx import TensorProto, helper
 
 from finn.core.datatype import DataType
 from finn.custom_op.fpgadataflow.hlscustomop import HLSCustomOp
@@ -151,19 +150,7 @@ def calc_tmem(self):
 
     def make_shape_compatible_op(self, model):
         oshape = self.get_normal_output_shape()
-        # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.float32)
-        return helper.make_node(
-            "Constant",
-            inputs=[],
-            outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.FLOAT,
-                dims=values.shape,
-                vals=values.flatten().astype(float),
-            ),
-        )
+        return super().make_const_shape_op(oshape)
 
     def infer_node_datatype(self, model):
         node = self.onnx_node

src/finn/custom_op/fpgadataflow/streamingfifo.py

@@ -30,7 +30,6 @@
 import os
 import subprocess
 import warnings
-from onnx import TensorProto, helper
 from shutil import copy
 
 from finn.core.datatype import DataType
@@ -78,19 +77,7 @@ def make_shape_compatible_op(self, model):
         oshape = self.get_normal_output_shape()
         ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
         assert ishape == tuple(exp_ishape), "Unexpect input shape for StreamingFIFO."
-        # implement tensor with correct shape
-        values = np.random.randn(*oshape).astype(np.float32)
-        return helper.make_node(
-            "Constant",
-            inputs=[],
-            outputs=[self.onnx_node.output[0]],
-            value=helper.make_tensor(
-                name="const_tensor",
-                data_type=TensorProto.FLOAT,
-                dims=values.shape,
-                vals=values.flatten().astype(float),
-            ),
-        )
+        return super().make_const_shape_op(oshape)
 
     def infer_node_datatype(self, model):
         node = self.onnx_node