diff --git a/src/finn/analysis/fpgadataflow/hls_synth_res_estimation.py b/src/finn/analysis/fpgadataflow/hls_synth_res_estimation.py
index 71e80d7da18c82d8f69a6c37829c4c6fc58b398d..78fc2ccfc92f9b7ca3ae6beafe7d24bdbfada2bc 100644
--- a/src/finn/analysis/fpgadataflow/hls_synth_res_estimation.py
+++ b/src/finn/analysis/fpgadataflow/hls_synth_res_estimation.py
@@ -74,5 +74,4 @@ def hls_synth_res_estimation(model):
for this node. Please run "CodeGen_ipgen" transformation and
"HLSSynth_IPGen" first to generate the report files"""
)
-
return res_dict
diff --git a/src/finn/custom_op/fpgadataflow/__init__.py b/src/finn/custom_op/fpgadataflow/__init__.py
index 38c2ed6638bef5cd709a7da83218145de91dce0a..9a6f66087fafff3745e239da4cb9f05c4ec73451 100644
--- a/src/finn/custom_op/fpgadataflow/__init__.py
+++ b/src/finn/custom_op/fpgadataflow/__init__.py
@@ -31,7 +31,7 @@ import numpy as np
import os
import subprocess
from finn.custom_op import CustomOp
-from finn.util.basic import CppBuilder, make_build_dir
+from finn.util.basic import CppBuilder, make_build_dir, roundup_to_integer_multiple
from finn.util.fpgadataflow import (
IPGenBuilder,
pyverilate_get_liveness_threshold_cycles,
@@ -493,15 +493,28 @@ compilation transformations?
"""Returns folded output shape (according to neuron folding), if implemented."""
raise Exception("get_folded_output_shape not implemented for this op")
- def get_instream_width(self, axi_strm_padding=False):
+ def get_instream_width(self):
"""Returns input stream width, if implemented."""
raise Exception("get_instream_width not implemented for this op")
- def get_outstream_width(self, axi_strm_padding=False):
+ def get_outstream_width(self):
"""Returns output stream width, if implemented."""
raise Exception("get_outstream_width not implemented for this op")
+ def get_instream_width_padded(self):
+ """Returns input stream width padded to a multiple of 8. This is required
+ by the AXI Stream spec."""
+ in_width = self.get_instream_width()
+ return roundup_to_integer_multiple(in_width, 8)
+
+ def get_outstream_width_padded(self):
+ """Returns output stream width padded to a multiple of 8. This is required
+ by the AXI Stream spec."""
+ out_width = self.get_outstream_width()
+ return roundup_to_integer_multiple(out_width, 8)
+
def get_ap_int_max_w(self):
+ "Return the maximum width of any ap_int used in this module."
instream = self.get_instream_width()
outstream = self.get_outstream_width()
return max([instream, outstream])
diff --git a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
index 0c8fe863fef9f2be97ab3d29e00d72b877d8108a..66daa9f7b408a1d17ee3cca6aea5ab4a843f7e4f 100644
--- a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
+++ b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
@@ -34,7 +34,6 @@ from finn.core.datatype import DataType
from finn.custom_op.fpgadataflow import HLSCustomOp
from finn.custom_op.im2col import compute_conv_output_dim
from onnx import TensorProto, helper
-from finn.util.basic import roundup_to_integer_multiple
from finn.util.data_packing import npy_to_rtlsim_input, rtlsim_output_to_npy
# ONNX i/o tensor shape assumptions for ConvolutionInputGenerator:
@@ -142,7 +141,7 @@ class ConvolutionInputGenerator(HLSCustomOp):
"""Returns FINN DataType of output."""
return DataType[self.get_nodeattr("outputDataType")]
- def get_instream_width(self, axi_strm_padding=False):
+ def get_instream_width(self):
"""Returns stream width, input and output stream width are equal for
the sliding window function"""
ibits = self.get_input_datatype().bitwidth()
@@ -150,15 +149,13 @@ class ConvolutionInputGenerator(HLSCustomOp):
ifm_ch = self.get_nodeattr("IFMChannels")
assert ifm_ch % simd == 0, "SIMD must divide IFMChannels"
in_width = simd * ibits
- if axi_strm_padding is True:
- in_width = roundup_to_integer_multiple(in_width, 8)
return in_width
- def get_outstream_width(self, axi_strm_padding=False):
+ def get_outstream_width(self):
"""Returns stream width, input and output stream width are equal for
the sliding window function, so the function to determine the input
stream width can be reused."""
- return self.get_instream_width(axi_strm_padding)
+ return self.get_instream_width()
def get_number_output_values(self):
folded_oshape = self.get_folded_output_shape()
diff --git a/src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py b/src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py
index ce4f883fa029225a5748c08463858e3bf1bfd35c..f30871909b1c70f3b5df148f1b6eae22fdbadc25 100644
--- a/src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py
@@ -32,7 +32,6 @@ import numpy as np
from finn.custom_op.fpgadataflow import HLSCustomOp
from finn.core.datatype import DataType
from onnx import TensorProto, helper
-from finn.util.basic import roundup_to_integer_multiple
from finn.util.data_packing import npy_to_rtlsim_input, rtlsim_output_to_npy
# does not do anything at the ONNX node-by-node level, and input-output
@@ -151,16 +150,12 @@ class StreamingDataWidthConverter_Batch(HLSCustomOp):
folded_ishape = self.get_folded_input_shape()
return np.prod(folded_ishape[:-1])
- def get_instream_width(self, axi_strm_padding=False):
+ def get_instream_width(self):
in_width = self.get_nodeattr("inWidth")
- if axi_strm_padding is True:
- in_width = roundup_to_integer_multiple(in_width, 8)
return in_width
- def get_outstream_width(self, axi_strm_padding=False):
+ def get_outstream_width(self):
out_width = self.get_nodeattr("outWidth")
- if axi_strm_padding is True:
- out_width = roundup_to_integer_multiple(out_width, 8)
return out_width
def make_shape_compatible_op(self, model):
diff --git a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
index f04ee7ca7830760f4ed2804b8b71f8fe5d29325f..46920711e13057178be9fca5fe3a18ce3e14feda 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
@@ -279,29 +279,27 @@ class StreamingFCLayer_Batch(HLSCustomOp):
"""Returns FINN DataType of output."""
return DataType[self.get_nodeattr("outputDataType")]
- def get_instream_width(self, axi_strm_padding=False):
+ def get_instream_width(self):
i_bits = self.get_input_datatype().bitwidth()
in_width = i_bits * self.get_nodeattr("SIMD")
- if axi_strm_padding is True:
- in_width = roundup_to_integer_multiple(in_width, 8)
return in_width
- def get_outstream_width(self, axi_strm_padding=False):
+ def get_outstream_width(self):
o_bits = self.get_output_datatype().bitwidth()
out_width = o_bits * self.get_nodeattr("PE")
- if axi_strm_padding is True:
- out_width = roundup_to_integer_multiple(out_width, 8)
return out_width
- def get_weightstream_width(self, axi_strm_padding=False):
+ def get_weightstream_width(self):
pe = self.get_nodeattr("PE")
simd = self.get_nodeattr("SIMD")
wp = self.get_weight_datatype().bitwidth()
w_width = pe * simd * wp
- if axi_strm_padding is True:
- w_width = roundup_to_integer_multiple(w_width, 8)
return w_width
+ def get_weightstream_width_padded(self):
+ weight_width = self.get_weightstream_width()
+ return roundup_to_integer_multiple(weight_width, 8)
+
def get_ap_int_max_w(self):
temp_value = super().get_ap_int_max_w()
weightstream = self.get_weightstream_width()
@@ -982,13 +980,13 @@ class StreamingFCLayer_Batch(HLSCustomOp):
"{}_{}".format(self.onnx_node.name, self.onnx_node.name)
]
# make instream width a multiple of 8 for AXI stream interface
- in_width = roundup_to_integer_multiple(self.get_instream_width(), 8)
+ in_width = self.get_instream_width_padded()
self.code_gen_dict["$IN_RANGE$"] = ["[{}:0]".format(in_width - 1)]
self.code_gen_dict["$OUT_RANGE$"] = [
- "[{}:0]".format(self.get_outstream_width(axi_strm_padding=True) - 1)
+ "[{}:0]".format(self.get_outstream_width_padded() - 1)
]
# make weight stream width a multiple of 8 for AXI stream interface
- weight_width = roundup_to_integer_multiple(self.get_weightstream_width(), 8)
+ weight_width = self.get_weightstream_width_padded()
self.code_gen_dict["$WEIGHT_RANGE$"] = ["[{}:0]".format(weight_width - 1)]
self.code_gen_dict["$WEIGHT_WIDTH$"] = [str(weight_width)]
self.code_gen_dict["$WSTREAM_DEPTH$"] = [str(self.calc_wmem())]
diff --git a/src/finn/custom_op/fpgadataflow/streamingfifo.py b/src/finn/custom_op/fpgadataflow/streamingfifo.py
index 6e004c47b1e13d95efa356b6b8984688f54027cc..eb96c6c04eb0b7b83c3f925e10f86b17ec399e42 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfifo.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfifo.py
@@ -33,7 +33,6 @@ import subprocess
from finn.custom_op.fpgadataflow import HLSCustomOp
from finn.core.datatype import DataType
from onnx import TensorProto, helper
-from finn.util.basic import roundup_to_integer_multiple
from finn.util.data_packing import npy_to_rtlsim_input, rtlsim_output_to_npy
from . import templates
@@ -110,7 +109,7 @@ class StreamingFIFO(HLSCustomOp):
"{}_{}".format(self.onnx_node.name, self.onnx_node.name)
]
# make instream width a multiple of 8 for axi interface
- in_width = self.get_instream_width(axi_strm_padding=True)
+ in_width = self.get_instream_width_padded()
self.code_gen_dict["$IN_RANGE$"] = ["[{}:0]".format(in_width - 1)]
self.code_gen_dict["$OUT_RANGE$"] = ["[{}:0]".format(in_width - 1)]
self.code_gen_dict["$WIDTH$"] = [str(in_width)]
@@ -164,6 +163,8 @@ class StreamingFIFO(HLSCustomOp):
def get_normal_input_shape(self):
depth = self.get_nodeattr("depth")
+ # depth has to be between 2 and 256 with the current
+ # StreamingFIFO implementation
assert (
depth >= 2
), """Depth is too low. Please set node attribute "depth" to a value
@@ -172,10 +173,22 @@ class StreamingFIFO(HLSCustomOp):
depth <= 256
), """Depth is too high. Please set node attribute "depth" to a value
between 2 and 256"""
+ # derive normal shape from folded shape
+ # StreamingFIFOs are inserted in between fpgadataflow nodes
+ # the folded shape could be for example (1, nf, pe)
+ # with nf (neuron folding): mh // pe
+ # the normal input shape is in this case (1, mh)
+ # so to achieve this the two inner dimensions are multiplied
+ # and together with all previous dimensions
+ # this gives the normal input shape
+
folded_shape = self.get_nodeattr("folded_shape")
+ # extract inner dimension
inner_dim = folded_shape[-1]
+ # multiply with the next inner dimension
folding_factor = folded_shape[-2] * inner_dim
normal_ishape = []
+ # create the normal_ishape
for i in range(len(folded_shape) - 2):
normal_ishape.append(folded_shape[i])
normal_ishape.append(folding_factor)
@@ -191,20 +204,16 @@ class StreamingFIFO(HLSCustomOp):
def get_folded_output_shape(self):
return self.get_nodeattr("folded_shape")
- def get_instream_width(self, axi_strm_padding=False):
+ def get_instream_width(self):
dtype = DataType[self.get_nodeattr("dataType")]
folded_shape = self.get_nodeattr("folded_shape")
in_width = folded_shape[-1] * dtype.bitwidth()
- if axi_strm_padding is True:
- in_width = roundup_to_integer_multiple(in_width, 8)
return in_width
- def get_outstream_width(self, axi_strm_padding=False):
+ def get_outstream_width(self):
dtype = DataType[self.get_nodeattr("dataType")]
folded_shape = self.get_nodeattr("folded_shape")
in_width = folded_shape[-1] * dtype.bitwidth()
- if axi_strm_padding is True:
- in_width = roundup_to_integer_multiple(in_width, 8)
return in_width
def execute_node(self, context, graph):
diff --git a/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py b/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
index ef1a5ee1bdc0bbe5c773aa375bf4402a8cb16ddb..83bc19030ebba66907e08c5b1e52d7c0ff9207a6 100644
--- a/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
@@ -33,7 +33,6 @@ from finn.custom_op.fpgadataflow import HLSCustomOp
from finn.custom_op.im2col import compute_conv_output_dim
from finn.core.datatype import DataType
from onnx import TensorProto, helper
-from finn.util.basic import roundup_to_integer_multiple
from finn.util.data_packing import npy_to_rtlsim_input, rtlsim_output_to_npy
@@ -88,17 +87,15 @@ class StreamingMaxPool_Batch(HLSCustomOp):
folded_oshape = self.get_folded_output_shape()
return np.prod(folded_oshape[:-1])
- def get_instream_width(self, axi_strm_padding=False):
+ def get_instream_width(self):
dt_bits = self.get_input_datatype().bitwidth()
ifm_ch = self.get_nodeattr("NumChannels")
in_width = int(dt_bits * ifm_ch)
- if axi_strm_padding is True:
- in_width = roundup_to_integer_multiple(in_width, 8)
return in_width
- def get_outstream_width(self, axi_strm_padding=False):
+ def get_outstream_width(self):
"""For streaming maxpool out stream with is the same as in stream width"""
- return self.get_instream_width(axi_strm_padding)
+ return self.get_instream_width()
def make_shape_compatible_op(self, model):
exp_ishape = self.get_normal_input_shape()
diff --git a/src/finn/custom_op/fpgadataflow/tlastmarker.py b/src/finn/custom_op/fpgadataflow/tlastmarker.py
index e5a5fed6c9d5d31fbf0082707879480e0c0a2dc7..25ea05e3607a52731ae1b64de421837bf137ee2b 100644
--- a/src/finn/custom_op/fpgadataflow/tlastmarker.py
+++ b/src/finn/custom_op/fpgadataflow/tlastmarker.py
@@ -27,7 +27,6 @@
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
from finn.custom_op.fpgadataflow import HLSCustomOp
-from finn.util.basic import roundup_to_integer_multiple
class TLastMarker(HLSCustomOp):
@@ -148,16 +147,12 @@ class TLastMarker(HLSCustomOp):
def get_folded_output_shape(self):
return self.get_folded_input_shape()
- def get_instream_width(self, axi_strm_padding=False):
+ def get_instream_width(self):
stream_width = self.get_nodeattr("StreamWidth")
- if axi_strm_padding is True:
- stream_width = roundup_to_integer_multiple(stream_width, 8)
return stream_width
- def get_outstream_width(self, axi_strm_padding=False):
+ def get_outstream_width(self):
stream_width = self.get_nodeattr("StreamWidth")
- if axi_strm_padding is True:
- stream_width = roundup_to_integer_multiple(stream_width, 8)
return stream_width
def strm_decl(self):
diff --git a/src/finn/transformation/fpgadataflow/insert_fifo.py b/src/finn/transformation/fpgadataflow/insert_fifo.py
index b80f2fbc1d9893d304f28c3494c44a69a1db052e..f66d0dc087ecbdd112422484ee1e01cb5ceef1c0 100644
--- a/src/finn/transformation/fpgadataflow/insert_fifo.py
+++ b/src/finn/transformation/fpgadataflow/insert_fifo.py
@@ -27,14 +27,20 @@ def _suitable_node(node):
class InsertFIFO(Transformation):
- """Ensure that the graph is terminated with a TLastMarker node, inserting
- one if necessary."""
+ """Inserting FIFOs in the beginning and end of the graph as well as
+ between fpgadataflow nodes.
+
+ Takes the setting for the depth from the surrounding nodes by extracting
+ node attribute 'outFIFODepth' of the previous and node attribute 'inFIFODepth'
+ of the subsequent node. max() of these two values sets the FIFO depth.
+
+ The other node attributes necessary to create a FIFO node are taking from the
+ node the FIFO node is inserted after: 'folded_shape' and 'dtype'"""
def __init__(self):
super().__init__()
def apply(self, model):
- # default depth for FIFOs
graph = model.graph
node_ind = -1
graph_modified = False
@@ -50,10 +56,19 @@ class InsertFIFO(Transformation):
fld_shape = n0.get_folded_output_shape()
dtype = n0.get_output_datatype()
+ # check if folded_shape of output of first node and
+ # input of the second node is equal
+ n1 = getCustomOp(consumer)
+ assert (
+ fld_shape == n1.get_folded_input_shape()
+ ), """The
+ folded output shape of the first node is not the same as the
+ folded output shape of the second node. A streaming fifo can't
+ be implemented in between these nodes."""
+
# check if outFIFOdepth attribute of first node
# and inFIFOdepth attribute of consumer node is equal
n0_depth = n0.get_nodeattr("outFIFODepth")
- n1 = getCustomOp(consumer)
n1_depth = n1.get_nodeattr("inFIFODepth")
if n0_depth == n1_depth:
fifo_depth = n0_depth
@@ -69,6 +84,7 @@ class InsertFIFO(Transformation):
n0.get_normal_output_shape(),
)
graph.value_info.append(fifo_output_tensor)
+ model.set_tensor_datatype(fifo_output_tensor.name, dtype)
fifo_node = oh.make_node(
"StreamingFIFO",
@@ -104,6 +120,7 @@ class InsertFIFO(Transformation):
n0.get_normal_input_shape(),
)
graph.value_info.append(fifo_output_tensor)
+ model.set_tensor_datatype(fifo_output_tensor.name, dtype)
fifo_node = oh.make_node(
"StreamingFIFO",
@@ -142,6 +159,7 @@ class InsertFIFO(Transformation):
n0.get_normal_output_shape(),
)
graph.value_info.append(fifo_input_tensor)
+ model.set_tensor_datatype(fifo_output_tensor.name, dtype)
fifo_node = oh.make_node(
"StreamingFIFO",
diff --git a/tests/end2end/test_end2end_tfc_w1a1_throughput_test.py b/tests/end2end/test_end2end_tfc_w1a1_throughput_test.py
index 26fd247aaf517158078a8cdbb577a55f1fdae6fa..6d72d9983d7a99d495b4e03e5ff0b5b633ee16ae 100644
--- a/tests/end2end/test_end2end_tfc_w1a1_throughput_test.py
+++ b/tests/end2end/test_end2end_tfc_w1a1_throughput_test.py
@@ -219,7 +219,7 @@ def test_end2end_tfc_w1a1_verify_dataflow_part():
ret_rtlsim_whole = execute_onnx(model, inp_dict, True)
res_rtlsim_whole = ret_rtlsim_whole[out_name]
assert np.isclose(res_npysim, res_rtlsim_nodebynode).all()
- assert np.isclose(res_rtlsim_nodebynode, res_rtlsim_whole).all()
+ assert np.isclose(res_npysim, res_rtlsim_whole).all()
def test_end2end_tfc_w1a1_verify_all():