diff --git a/src/finn/custom_op/im2col.py b/src/finn/custom_op/im2col.py
index 82a6b140f7af1be4e5c0f429d077b99c7865383e..8ed0041704d421dab587f08bcbcd9e739e8434e9 100644
--- a/src/finn/custom_op/im2col.py
+++ b/src/finn/custom_op/im2col.py
@@ -80,6 +80,8 @@ class Im2Col(CustomOp):
"input_shape": ("s", True, ""),
"pad_amount": ("i", False, 0),
"pad_value": ("i", False, 0),
+ # depthwise: if != 0, infer ConvolutionInputGenerator with depthwise == 1
+ "depthwise": ("i", False, 0),
}
def make_shape_compatible_op(self, model):
diff --git a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
index d421a5f3ef8ca980b399087de1482b2ae913da1b..b70b126680d650547cf376dd601c048c73a1cfd4 100644
--- a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
+++ b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
@@ -34,6 +34,7 @@ from finn.custom_op.registry import getCustomOp
from finn.transformation.infer_shapes import InferShapes
from finn.transformation.infer_datatypes import InferDataTypes
import finn.core.data_layout as DataLayout
+from finn.util.basic import get_by_name
class InferConvInpGen(Transformation):
@@ -56,6 +57,7 @@ class InferConvInpGen(Transformation):
k = i2c_inst.get_nodeattr("kernel_size")
pad = i2c_inst.get_nodeattr("pad_amount")
pad_val = i2c_inst.get_nodeattr("pad_value")
+ depthwise = i2c_inst.get_nodeattr("depthwise")
ifm_ch = i2c_in_shape[-1]
ifm_dim = i2c_in_shape[1]
ofm_dim = i2c_out_shape[1]
@@ -67,7 +69,11 @@ class InferConvInpGen(Transformation):
if pad > 0:
# if padding enabled, ensure pad_val supported by DataType
- assert dt.allowed(pad_val), "Im2Col DataType must support pad_val"
+ # assert dt.allowed(pad_val),"""FMPadding_Batch DataType
+ # must support pad_val"""
+ assert (
+ pad_val == 0
+ ), "FMPadding_Batch doesn't currently support pad_val!= 0"
odim_padding = ifm_dim + 2 * pad
@@ -112,6 +118,7 @@ class InferConvInpGen(Transformation):
Stride=stride,
inputDataType=dt.name,
outputDataType=dt.name,
+ depthwise=depthwise,
)
graph.node.insert(ConvInpGen_node_idx, ConvInpGen_node)
# remove old nodes
@@ -169,6 +176,137 @@ class InferStreamingMaxPool(Transformation):
return (model, graph_modified)
+class InferPool_Batch(Transformation):
+ """If kernel_shape > strides, replace Pool layer with with of Im2col
+ + pool(with kernel_shape == strides), plus Transpose layers to keep the original
+ data layout."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ node_ind += 1
+ if n.op_type in ["MaxPool"]:
+ # extract pool parameters
+ k = get_by_name(n.attribute, "kernel_shape").ints[-1]
+ stride = get_by_name(n.attribute, "strides").ints[-1]
+
+ if k <= stride:
+ continue
+
+ try:
+ pad = get_by_name(n.attribute, "pads").ints[-1]
+ except AttributeError:
+ pad = 0
+
+ node_input = n.input[0]
+ node_output = n.output[0]
+ idt = model.get_tensor_datatype(node_input)
+ if not idt.is_integer():
+ continue
+
+ # odt = model.get_tensor_datatype(node_output)
+
+ ifm_ch = model.get_tensor_shape(n.input[0])[1] # assume NCHW
+ ofm_ch = ifm_ch
+ ifm_dim = model.get_tensor_shape(n.input[0])[-1] # assume NCHW
+ ofm_dim = model.get_tensor_shape(n.output[0])[-1] # assume NCHW
+ # create new intermediate values
+ inp_trans_out = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ TensorProto.FLOAT,
+ (1, ifm_dim, ifm_dim, ifm_ch), # NHWC
+ )
+ graph.value_info.append(inp_trans_out)
+ inp_trans_out = inp_trans_out.name
+ model.set_tensor_datatype(inp_trans_out, idt)
+
+ im2col_out = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ TensorProto.FLOAT,
+ (1, ofm_dim, ofm_dim, ifm_ch * k * k),
+ )
+ graph.value_info.append(im2col_out)
+ im2col_out = im2col_out.name
+ model.set_tensor_datatype(im2col_out, idt)
+
+ pool_output = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ TensorProto.FLOAT,
+ (1, ofm_dim, ofm_dim, ofm_ch),
+ )
+ graph.value_info.append(pool_output)
+ pool_output = pool_output.name
+ # model.set_tensor_datatype(pool_output, odt)
+
+ # create new nodes
+ # NCHW -> NHWC
+ inp_trans_node = helper.make_node(
+ "Transpose", [node_input], [inp_trans_out], perm=[0, 2, 3, 1]
+ )
+
+ if n.op_type == "MaxPool":
+ pool_fxn = "MaxPool"
+ pad_value = idt.min()
+ else:
+ raise Exception(
+ "pad_value and pool_fxn not configured for {}".format(n.op_type)
+ )
+
+ # format input tensor
+ im2col_node = helper.make_node(
+ "Im2Col",
+ [inp_trans_out],
+ [im2col_out],
+ domain="finn",
+ stride=stride,
+ kernel_size=k,
+ pad_amount=pad,
+ pad_value=pad_value,
+ depthwise=1,
+ input_shape="(1,{},{},{})".format(ifm_dim, ifm_dim, ifm_ch),
+ )
+
+ # Warning PE has to be equal to ifm_ch until Im2Col is replaced by
+ # ConvolutionInputGenerator with depthwise=1.
+ # For other settings the output will be incorrect due to incorrect input
+ # data layout
+ pool_node = helper.make_node(
+ "Pool_Batch",
+ [im2col_out],
+ [pool_output],
+ domain="finn",
+ backend="fpgadataflow",
+ dataType=idt.name,
+ Channels=ifm_ch,
+ PE=ifm_ch,
+ KernelSize=k,
+ Function=pool_fxn,
+ OutImgDim=ofm_dim,
+ BatchSize=1,
+ )
+
+ # NHWC -> NCHW
+ out_trans_node = helper.make_node(
+ "Transpose", [pool_output], [node_output], perm=[0, 3, 1, 2]
+ )
+
+ # insert nodes where the conv is to preserve topological ordering
+ graph.node.insert(node_ind, inp_trans_node)
+ graph.node.insert(node_ind + 1, im2col_node)
+ graph.node.insert(node_ind + 2, pool_node)
+ graph.node.insert(node_ind + 3, out_trans_node)
+ # remove old node
+ graph.node.remove(n)
+ graph_modified = True
+
+ if graph_modified:
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ return (model, graph_modified)
+
+
class InferBinaryStreamingFCLayer(Transformation):
"""Convert XnorPopcountMatMul layers to
StreamingFCLayer_Batch layers. Any immediately following MultiThreshold