diff --git a/src/finn/builder/build_dataflow_config.py b/src/finn/builder/build_dataflow_config.py
index 13946c9d1e3780f5e45ee2ca863190d36849ea4f..51e751610108289c04fa0ae72fe593cbb0c16a86 100644
--- a/src/finn/builder/build_dataflow_config.py
+++ b/src/finn/builder/build_dataflow_config.py
@@ -257,7 +257,7 @@ class DataflowBuildConfig:
#: setting the FIFO sizes.
auto_fifo_strategy: Optional[
AutoFIFOSizingMethod
- ] = AutoFIFOSizingMethod.CHARACTERIZE
+ ] = AutoFIFOSizingMethod.LARGEFIFO_RTLSIM
#: Memory resource type for large FIFOs
#: Only relevant when `auto_fifo_depths = True`
diff --git a/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py b/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
index 3ed76db2982e411b711be5bd78e39dd866332714..9d08a2432834f0786695d6300c8bcd68c18452d5 100644
--- a/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
+++ b/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
@@ -103,8 +103,8 @@ class ChannelwiseOp_Batch(HLSCustomOp):
"paramDataType": ("s", True, ""),
"outputDataType": ("s", True, ""),
# input and output FIFO depths
- "inFIFODepth": ("i", False, 0),
- "outFIFODepth": ("i", False, 0),
+ "inFIFODepths": ("ints", False, [0]),
+ "outFIFODepths": ("ints", False, [0]),
# number of input vectors, examples:
# [1] is a single vector (like a FC layer with batch=1)
# [4] is four vectors (like a FC layer with batch=4)
diff --git a/src/finn/custom_op/fpgadataflow/hlscustomop.py b/src/finn/custom_op/fpgadataflow/hlscustomop.py
index 79ae6957564ce07c18b76552089f64107fe51356..7d322dc372ba91d1613fd5cf9601ff2fc0d023ea 100644
--- a/src/finn/custom_op/fpgadataflow/hlscustomop.py
+++ b/src/finn/custom_op/fpgadataflow/hlscustomop.py
@@ -108,12 +108,9 @@ class HLSCustomOp(CustomOp):
# ID of FPGA device to which this Op is allocated, in
# a multi-FPGA setting
"device_id": ("i", False, 0),
- # input and output FIFO depths
- "inFIFODepth": ("i", False, 2),
- "outFIFODepth": ("i", False, 2),
# input and output FIFO depths for multi-I/O nodes
- "inFIFODepths": ("ints", False, []),
- "outFIFODepths": ("ints", False, []),
+ "inFIFODepths": ("ints", False, [2]),
+ "outFIFODepths": ("ints", False, [2]),
"output_hook": ("s", False, ""),
# accumulated characteristic function over two periods
"io_chrc_in": ("t", False, np.asarray([], dtype=np.int32)),
diff --git a/src/finn/custom_op/fpgadataflow/matrixvectoractivation.py b/src/finn/custom_op/fpgadataflow/matrixvectoractivation.py
index d5dbc86c4ec2469f8588b82e76d0295bafb1328e..e78a918e81385a5b6e837a384c6c1a238073a80d 100644
--- a/src/finn/custom_op/fpgadataflow/matrixvectoractivation.py
+++ b/src/finn/custom_op/fpgadataflow/matrixvectoractivation.py
@@ -1227,8 +1227,11 @@ class MatrixVectorActivation(HLSCustomOp):
self.code_gen_dict["$PRAGMAS$"].append(
"#pragma HLS INTERFACE axis port=out name=out_" + self.hls_sname()
)
- in_fifo_depth = self.get_nodeattr("inFIFODepth")
- out_fifo_depth = self.get_nodeattr("outFIFODepth")
+ # TODO can we deprecate this entirely? this looks like legacy code
+ # that does not really serve a purpose - FIFO sizes are not typically
+ # allocated at this point; at best they are set to 2 as the default
+ in_fifo_depth = 2
+ out_fifo_depth = 2
# insert depth pragmas only if specified
if in_fifo_depth != 0:
self.code_gen_dict["$PRAGMAS$"].append(
diff --git a/src/finn/custom_op/fpgadataflow/thresholding_batch.py b/src/finn/custom_op/fpgadataflow/thresholding_batch.py
index 828ddd97375a8466079a44f297d9c4a0e2a59ea1..110e456cbd3885e1bf434743ee422a425c2d7b99 100644
--- a/src/finn/custom_op/fpgadataflow/thresholding_batch.py
+++ b/src/finn/custom_op/fpgadataflow/thresholding_batch.py
@@ -76,8 +76,8 @@ class Thresholding_Batch(HLSCustomOp):
"weightDataType": ("s", True, ""),
"outputDataType": ("s", True, ""),
# input and output FIFO depths
- "inFIFODepth": ("i", False, 0),
- "outFIFODepth": ("i", False, 0),
+ "inFIFODepths": ("ints", False, [0]),
+ "outFIFODepths": ("ints", False, [0]),
# number of input vectors, examples:
# [1] is a single vector (like a FC layer with batch=1)
# [4] is four vectors (like a FC layer with batch=4)
diff --git a/src/finn/custom_op/fpgadataflow/vectorvectoractivation.py b/src/finn/custom_op/fpgadataflow/vectorvectoractivation.py
index abcb1c756d6f2caad3a68e32c0409e5930929c6a..6391f27bbb80aa0d8cc8c30e752d8fe4ff142fae 100644
--- a/src/finn/custom_op/fpgadataflow/vectorvectoractivation.py
+++ b/src/finn/custom_op/fpgadataflow/vectorvectoractivation.py
@@ -578,8 +578,11 @@ class VectorVectorActivation(HLSCustomOp):
self.code_gen_dict["$PRAGMAS$"].append(
"#pragma HLS INTERFACE axis port=out name=out_" + self.hls_sname()
)
- in_fifo_depth = self.get_nodeattr("inFIFODepth")
- out_fifo_depth = self.get_nodeattr("outFIFODepth")
+ # TODO can we deprecate this entirely? this looks like legacy code
+ # that does not really serve a purpose - FIFO sizes are not typically
+ # allocated at this point; at best they are set to 2 as the default
+ in_fifo_depth = 2
+ out_fifo_depth = 2
# insert depth pragmas only if specified
if in_fifo_depth != 0:
self.code_gen_dict["$PRAGMAS$"].append(
diff --git a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
index 429bc34ffc59b5d98bb559f36ac557de4dbba92f..753dbb0f87a877f69a24bb520d09a173925cafd4 100644
--- a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
+++ b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
@@ -1234,6 +1234,7 @@ class InferDuplicateStreamsLayer(Transformation):
inputDataType=dt.name,
numInputVectors=vecs,
NumOutputStreams=n_outputs,
+ outFIFODepths=[2] * n_outputs,
name="DuplicateStreams_Batch_" + node.name,
)
diff --git a/src/finn/transformation/fpgadataflow/insert_fifo.py b/src/finn/transformation/fpgadataflow/insert_fifo.py
index b9222cf3eef1a8566e698b4f58a85d100f824243..e75d1880cb1b7c5c07d3c7a4e155216e4adc740d 100644
--- a/src/finn/transformation/fpgadataflow/insert_fifo.py
+++ b/src/finn/transformation/fpgadataflow/insert_fifo.py
@@ -138,14 +138,8 @@ class InsertFIFO(Transformation):
# check if outFIFOdepth attribute of first node
# and inFIFOdepth attribute of consumer node is equal
- if idx_out == 0:
- n0_depth = n0.get_nodeattr("outFIFODepth")
- else:
- n0_depth = n0.get_nodeattr("outFIFODepths")[idx_out]
- if idx_inp == 0:
- n1_depth = n1.get_nodeattr("inFIFODepth")
- else:
- n1_depth = n1.get_nodeattr("inFIFODepths")[idx_inp]
+ n0_depth = n0.get_nodeattr("outFIFODepths")[idx_out]
+ n1_depth = n1.get_nodeattr("inFIFODepths")[idx_inp]
if n0_depth == n1_depth:
fifo_depth = n0_depth
@@ -187,18 +181,12 @@ class InsertFIFO(Transformation):
if inp == output_name:
consumer.input[idx] = fifo_output_tensor.name
# ensure created FIFO depth is reflected on both sides
- if idx_out == 0:
- n0.set_nodeattr("outFIFODepth", fifo_depth)
- else:
- odepths = n0.get_nodeattr("outFIFODepths")
- odepths[idx_out] = fifo_depth
- n0.set_nodeattr("outFIFODepths", odepths)
- if idx_inp == 0:
- n1.set_nodeattr("inFIFODepth", fifo_depth)
- else:
- idepths = n1.get_nodeattr("inFIFODepths")
- idepths[idx_inp] = fifo_depth
- n1.set_nodeattr("inFIFODepths", idepths)
+ odepths = n0.get_nodeattr("outFIFODepths")
+ odepths[idx_out] = fifo_depth
+ n0.set_nodeattr("outFIFODepths", odepths)
+ idepths = n1.get_nodeattr("inFIFODepths")
+ idepths[idx_inp] = fifo_depth
+ n1.set_nodeattr("inFIFODepths", idepths)
graph_modified = True
@@ -221,7 +209,7 @@ class InsertFIFO(Transformation):
else:
fld_shape = n0.get_folded_input_shape(inp_ind)
dtype = n0.get_input_datatype(inp_ind)
- fifo_depth = n0.get_nodeattr("inFIFODepth")
+ fifo_depth = n0.get_nodeattr("inFIFODepths")[inp_ind]
if fifo_depth <= 2:
warnings.warn("Overriding input FIFO depth to 32")
diff --git a/src/finn/transformation/fpgadataflow/set_fifo_depths.py b/src/finn/transformation/fpgadataflow/set_fifo_depths.py
index 90ea853b6072b145df64a8a73ee93c65989fe447..f715aaeffb6d4d00f2e14c5fb25ec931443d5d97 100644
--- a/src/finn/transformation/fpgadataflow/set_fifo_depths.py
+++ b/src/finn/transformation/fpgadataflow/set_fifo_depths.py
@@ -250,14 +250,21 @@ class InsertAndSetFIFODepths(Transformation):
)
assert node.op_type != "StreamingFIFO", "Found existing StreamingFIFO node"
node = getCustomOp(node)
+ ifd = node.get_nodeattr("inFIFODepths")
+ ofd = node.get_nodeattr("outFIFODepths")
if self.max_depth is not None:
- node.set_nodeattr("inFIFODepth", self.max_depth)
- node.set_nodeattr("outFIFODepth", self.max_depth)
+ ifd = [self.max_depth] * len(ifd)
+ ofd = [self.max_depth] * len(ofd)
else:
- i_depth = np.prod(node.get_folded_input_shape()[:-1])
- o_depth = np.prod(node.get_folded_output_shape()[:-1])
- node.set_nodeattr("inFIFODepth", i_depth)
- node.set_nodeattr("outFIFODepth", o_depth)
+ # set each FIFO to its tensor size
+ # (except stream width hence the :-1)
+ for i in range(len(ifd)):
+ ifd[i] = np.prod(node.get_folded_input_shape(i)[:-1])
+ for o in range(len(ofd)):
+ ofd[o] = np.prod(node.get_folded_output_shape(o)[:-1])
+ node.set_nodeattr("inFIFODepths", ifd)
+ node.set_nodeattr("outFIFODepths", ofd)
+
if node.onnx_node.op_type in extw_optypes:
mmode = node.get_nodeattr("mem_mode")
if mmode == "external":
@@ -380,8 +387,11 @@ class InsertAndSetFIFODepths(Transformation):
reset_implementation(node_inst)
del fifos[node.name]
else:
- getCustomOp(node).set_nodeattr("inFIFODepth", 0)
- getCustomOp(node).set_nodeattr("outFIFODepth", 0)
+ inst = getCustomOp(node)
+ ifd = inst.get_nodeattr("inFIFODepths")
+ ofd = inst.get_nodeattr("outFIFODepths")
+ inst.set_nodeattr("inFIFODepths", [0] * len(ifd))
+ inst.set_nodeattr("outFIFODepths", [0] * len(ofd))
# for every extw node we changed from external to decoupled,
# change back and reset implementation
if node.op_type in extw_optypes:
diff --git a/tests/end2end/test_end2end_bnn_pynq.py b/tests/end2end/test_end2end_bnn_pynq.py
index 103f18b514c23c4e1ad35a85d020dc0481aa9c47..5f787d1f889645d04884aed9b89a0b1c91d1f418 100644
--- a/tests/end2end/test_end2end_bnn_pynq.py
+++ b/tests/end2end/test_end2end_bnn_pynq.py
@@ -569,8 +569,8 @@ class TestEnd2End:
for node in hls_layers:
if node.op_type != "StreamingFIFO":
op_inst = getCustomOp(node)
- assert op_inst.get_nodeattr("inFIFODepth") == 0
- assert op_inst.get_nodeattr("outFIFODepth") == 0
+ assert op_inst.get_nodeattr("inFIFODepths") == [0]
+ assert op_inst.get_nodeattr("outFIFODepths") == [0]
model.save(
get_checkpoint_name(
topology, wbits, abits, QONNX_export, "fifodepth_" + kind