diff --git a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
index 29773f297364a1fa299fcf4e618c1c9875793ecd..0ac85daba2424df9d8f87dcaa96dfc1a46a1acb0 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
@@ -14,8 +14,6 @@ class StreamingFCLayer_Batch(HLSCustomOp):
def get_nodeattr_types(self):
my_attrs = {
- "WMEM": ("i", True, 0),
- "TMEM": ("i", True, 0),
"PE": ("i", True, 0),
"SIMD": ("i", True, 0),
"MW": ("i", True, 0),
@@ -29,10 +27,30 @@ class StreamingFCLayer_Batch(HLSCustomOp):
# use xnor-popcount for binary weights/inputs, thus treating them
# as bipolar
"binaryXnorMode": ("i", False, 0),
+ # no-activation mode (produce accumulators)
+ "noActivation": ("i", False, 0),
}
my_attrs.update(super().get_nodeattr_types())
return my_attrs
+ def calc_wmem(self):
+ mw = self.get_nodeattr("MW")
+ mh = self.get_nodeattr("MH")
+ pe = self.get_nodeattr("PE")
+ simd = self.get_nodeattr("SIMD")
+ assert mh % pe == 0
+ assert mw % simd == 0
+ wmem = mw * mh // (pe * simd)
+ return wmem
+
+ def calc_tmem(self):
+ if self.get_nodeattr("noActivation") == 1:
+ return 0
+ else:
+ mh = self.get_nodeattr("MH")
+ pe = self.get_nodeattr("PE")
+ return mh // pe
+
def make_shape_compatible_op(self):
pass
@@ -106,7 +124,7 @@ class StreamingFCLayer_Batch(HLSCustomOp):
mh = self.get_nodeattr("MH")
pe = self.get_nodeattr("PE")
simd = self.get_nodeattr("SIMD")
- wmem = mw * mh // (pe * simd)
+ wmem = self.calc_wmem()
assert orig_weight_matrix.shape == (mw, mh)
assert mw % simd == 0
assert mh % pe == 0
@@ -184,15 +202,13 @@ class StreamingFCLayer_Batch(HLSCustomOp):
self.get_nodeattr("SIMD"),
export_wdt.get_hls_datatype_str(),
self.get_nodeattr("PE"),
- self.get_nodeattr("WMEM"),
+ self.calc_wmem(),
)
)
else:
f_weights.write(
"static BinaryWeights<{},{},{}> weights = ".format(
- self.get_nodeattr("SIMD"),
- self.get_nodeattr("PE"),
- self.get_nodeattr("WMEM"),
+ self.get_nodeattr("SIMD"), self.get_nodeattr("PE"), self.calc_wmem()
)
)
f_weights.write(weight_hls_code)
@@ -229,7 +245,7 @@ class StreamingFCLayer_Batch(HLSCustomOp):
f_thresh.write(
"static ThresholdsActivation<{},{},{},{},{},{},{}> threshs \
= ".format(
- self.get_nodeattr("TMEM"),
+ self.calc_tmem(),
self.get_nodeattr("PE"),
threshold_tensor.shape[-1],
tdt_hls,
@@ -291,10 +307,8 @@ class StreamingFCLayer_Batch(HLSCustomOp):
def global_includes(self):
self.code_gen_dict["$GLOBALS$"] = ['#include "weights.hpp"']
self.code_gen_dict["$GLOBALS$"] += ['#include "activations.hpp"']
- if self.get_nodeattr("WMEM") != 0:
- # TODO find a better way of checking for no pregenerated weights
- self.code_gen_dict["$GLOBALS$"] += ['#include "params.h"']
- if self.get_nodeattr("TMEM") != 0:
+ self.code_gen_dict["$GLOBALS$"] += ['#include "params.h"']
+ if self.calc_tmem() != 0:
# TODO find a better way of checking for no pregenerated thresholds
self.code_gen_dict["$GLOBALS$"] += ['#include "thresh.h"']
@@ -308,8 +322,8 @@ class StreamingFCLayer_Batch(HLSCustomOp):
self.get_nodeattr("MH"),
self.get_nodeattr("SIMD"),
self.get_nodeattr("PE"),
- self.get_nodeattr("WMEM"),
- self.get_nodeattr("TMEM"),
+ self.calc_wmem(),
+ self.calc_tmem(),
numReps,
)
]
@@ -344,7 +358,7 @@ class StreamingFCLayer_Batch(HLSCustomOp):
def docompute(self):
node = self.onnx_node
tmpl_args = self.get_template_param_values()
- if self.get_nodeattr("TMEM") == 0:
+ if self.calc_tmem() == 0:
odtype_hls_str = self.get_output_datatype().get_hls_datatype_str()
threshs = "PassThroughActivation<%s>()" % odtype_hls_str
else:
diff --git a/tests/fpgadataflow/test_code_gen_trafo.py b/tests/fpgadataflow/test_code_gen_trafo.py
index 53f74ffcbd72cb1338bc39cbe012880b892fc737..533710605e2bc514ba0fb0c8784c378d07451951 100644
--- a/tests/fpgadataflow/test_code_gen_trafo.py
+++ b/tests/fpgadataflow/test_code_gen_trafo.py
@@ -15,12 +15,9 @@ def test_code_gen_trafo():
mh = 8
pe = 4
simd = 4
- wmem = mw * mh // (pe * simd)
- nf = mh // pe
- sf = mw // simd
- inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, sf, simd])
- outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, nf, pe])
+ inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, mw])
+ outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, mh])
node_inp_list = ["inp", "weights", "thresh"]
FCLayer_node = helper.make_node(
"StreamingFCLayer_Batch",
@@ -35,11 +32,10 @@ def test_code_gen_trafo():
MH=mh,
SIMD=simd,
PE=pe,
- WMEM=wmem,
- TMEM=0,
inputDataType=idt.name,
weightDataType=wdt.name,
outputDataType=odt.name,
+ noActivation=1,
)
graph = helper.make_graph(
nodes=[FCLayer_node], name="fclayer_graph", inputs=[inp], outputs=[outp]
diff --git a/tests/fpgadataflow/test_compilation_trafo.py b/tests/fpgadataflow/test_compilation_trafo.py
index fd4174f27b2b9db1fae4f95589a5c9832679cc32..f84ce34b54b3496f7e277e55ac574124e09c25d3 100644
--- a/tests/fpgadataflow/test_compilation_trafo.py
+++ b/tests/fpgadataflow/test_compilation_trafo.py
@@ -16,12 +16,9 @@ def test_compilation_trafo():
mh = 8
pe = 4
simd = 4
- wmem = mw * mh // (pe * simd)
- nf = mh // pe
- sf = mw // simd
- inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, sf, simd])
- outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, nf, pe])
+ inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, mw])
+ outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, mh])
node_inp_list = ["inp", "weights", "thresh"]
FCLayer_node = helper.make_node(
"StreamingFCLayer_Batch",
@@ -36,11 +33,10 @@ def test_compilation_trafo():
MH=mh,
SIMD=simd,
PE=pe,
- WMEM=wmem,
- TMEM=0,
inputDataType=idt.name,
weightDataType=wdt.name,
outputDataType=odt.name,
+ noActivation=1,
)
graph = helper.make_graph(
nodes=[FCLayer_node], name="fclayer_graph", inputs=[inp], outputs=[outp]
diff --git a/tests/fpgadataflow/test_fpgadataflow_fclayer.py b/tests/fpgadataflow/test_fpgadataflow_fclayer.py
index 5ce7d9ff23c3877791a3d825ebe91cab9af56e92..0df66c4af2cfbadead8e95322c433cf69d4d2715 100644
--- a/tests/fpgadataflow/test_fpgadataflow_fclayer.py
+++ b/tests/fpgadataflow/test_fpgadataflow_fclayer.py
@@ -19,13 +19,6 @@ def make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T=None, tdt=Non
mh = W.shape[1]
assert mh % pe == 0
assert mw % simd == 0
- wmem = mw * mh // (pe * simd)
- assert mw * mh == wmem * pe * simd
- nf = mh // pe
- if T is not None:
- tmem = nf
- else:
- tmem = 0
# there are two ways to implement bipolar weights and inputs for
# StreamingFC:
@@ -45,6 +38,7 @@ def make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T=None, tdt=Non
inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, mw])
outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, mh])
if T is not None:
+ no_act = 0
node_inp_list = ["inp", "weights", "thresh"]
if odt == DataType.BIPOLAR:
actval = 0
@@ -54,6 +48,7 @@ def make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T=None, tdt=Non
# no thresholds
node_inp_list = ["inp", "weights"]
actval = 0
+ no_act = 1
FCLayer_node = helper.make_node(
"StreamingFCLayer_Batch",
node_inp_list,
@@ -65,13 +60,12 @@ def make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T=None, tdt=Non
MH=mh,
SIMD=simd,
PE=pe,
- WMEM=wmem,
- TMEM=tmem,
inputDataType=export_idt.name,
weightDataType=export_wdt.name,
outputDataType=odt.name,
ActVal=actval,
binaryXnorMode=binary_xnor_mode,
+ noActivation=no_act,
)
graph = helper.make_graph(
nodes=[FCLayer_node], name="fclayer_graph", inputs=[inp], outputs=[outp]
diff --git a/tests/test_set_attribute.py b/tests/test_set_attribute.py
index 43688f212c612959291938e44d34f6badedb6396..7e56fdbb938441115bf458532560af8ac803b4f9 100644
--- a/tests/test_set_attribute.py
+++ b/tests/test_set_attribute.py
@@ -10,13 +10,10 @@ def test_set_attribute():
mh = 8
pe = 4
simd = 4
- wmem = mw * mh // (pe * simd)
- nf = mh // pe
- sf = mw // simd
idt = odt = wdt = DataType.BIPOLAR
- inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, sf, simd])
- outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, nf, pe])
+ inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, mw])
+ outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, mh])
node_inp_list = ["inp", "weights"]
FCLayer_node = helper.make_node(
@@ -32,7 +29,6 @@ def test_set_attribute():
MH=mh,
SIMD=simd,
PE=pe,
- WMEM=wmem,
inputDataType=idt.name,
weightDataType=wdt.name,
outputDataType=odt.name,