diff --git a/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py b/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
index 073d6620ac3c2a4f62ac544e74ecf21b6e36d58f..ad91013a2e796a60df30bd42595d41e46c1c2ee6 100644
--- a/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
+++ b/src/finn/custom_op/fpgadataflow/channelwise_op_batch.py
@@ -59,7 +59,7 @@ def get_smallest_possible(vals):
for k in DataType.__members__:
dt = DataType[k]
- if dt in [DataType.BIPOLAR, DataType.TERNARY, DataType.FLOAT32]:
+ if dt in [DataType["BIPOLAR"], DataType["TERNARY"], DataType["FLOAT32"]]:
# not currently supported
continue
@@ -75,9 +75,9 @@ def get_smallest_possible(vals):
)
if (0 <= vals).all():
- return DataType.UINT64
+ return DataType["UINT64"]
else:
- return DataType.INT64
+ return DataType["INT64"]
class ChannelwiseOp_Batch(HLSCustomOp):
@@ -347,8 +347,8 @@ class ChannelwiseOp_Batch(HLSCustomOp):
)
# get input data type
export_idt = self.get_input_datatype()
- if self.get_input_datatype() == DataType.BIPOLAR:
- export_idt = DataType.BINARY
+ if self.get_input_datatype() == DataType["BIPOLAR"]:
+ export_idt = DataType["BINARY"]
idt_hls = export_idt.get_hls_datatype_str()
# write parameters into params.h
@@ -356,8 +356,8 @@ class ChannelwiseOp_Batch(HLSCustomOp):
pdt_hls = pdt.get_hls_datatype_str()
# use binary to export bipolar activations
export_odt = self.get_output_datatype()
- if self.get_output_datatype() == DataType.BIPOLAR:
- export_odt = DataType.BINARY
+ if self.get_output_datatype() == DataType["BIPOLAR"]:
+ export_odt = DataType["BINARY"]
odt_hls = export_odt.get_hls_datatype_str()
# get desired function
func = self.get_nodeattr("Func")
@@ -438,7 +438,7 @@ class ChannelwiseOp_Batch(HLSCustomOp):
# load output npy file
super().npy_to_dynamic_output(context)
# reinterpret binary output as bipolar where needed
- if self.get_output_datatype() == DataType.BIPOLAR:
+ if self.get_output_datatype() == DataType["BIPOLAR"]:
out = context[node.output[0]]
out = 2 * out - 1
context[node.output[0]] = out
@@ -546,9 +546,9 @@ class ChannelwiseOp_Batch(HLSCustomOp):
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_output_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_outstream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
diff --git a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
index 9ec7bc662d95b1c94ca17bc3c9a1a7b6199cc18a..19732e44398665cfd9b97f9a1abcec56372e2523 100644
--- a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
+++ b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator.py
@@ -319,10 +319,10 @@ class ConvolutionInputGenerator(HLSCustomOp):
inp.shape == exp_ishape
), """Input shape doesn't
match expected shape (1, ifm_dim_h, ifm_dim_w, ifm_ch)."""
- if self.get_input_datatype() == DataType.BIPOLAR:
+ if self.get_input_datatype() == DataType["BIPOLAR"]:
# store bipolar activations as binary
inp = (inp + 1) / 2
- export_idt = DataType.BINARY
+ export_idt = DataType["BINARY"]
else:
export_idt = self.get_input_datatype()
# reshape input into folded form
@@ -370,7 +370,7 @@ class ConvolutionInputGenerator(HLSCustomOp):
)
)
# binary -> bipolar if needed
- if self.get_output_datatype() == DataType.BIPOLAR:
+ if self.get_output_datatype() == DataType["BIPOLAR"]:
out = context[node.output[0]]
out = 2 * out - 1
context[node.output[0]] = out
@@ -404,9 +404,9 @@ class ConvolutionInputGenerator(HLSCustomOp):
def read_npy_data(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_input_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_instream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
@@ -465,9 +465,9 @@ class ConvolutionInputGenerator(HLSCustomOp):
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_output_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_outstream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
diff --git a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator1d.py b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator1d.py
index b428210acfd70186f68e7f1b35cfcd945d0a77d9..1f9fcade0f974df2b2f21171d42e63f4af5e7eac 100644
--- a/src/finn/custom_op/fpgadataflow/convolutioninputgenerator1d.py
+++ b/src/finn/custom_op/fpgadataflow/convolutioninputgenerator1d.py
@@ -345,10 +345,10 @@ class ConvolutionInputGenerator1D(HLSCustomOp):
inp.shape == exp_ishape
), """Input shape doesn't
match expected shape (1, ifm_dim, ifm_dim, ifm_ch)."""
- if self.get_input_datatype() == DataType.BIPOLAR:
+ if self.get_input_datatype() == DataType["BIPOLAR"]:
# store bipolar activations as binary
inp = (inp + 1) / 2
- export_idt = DataType.BINARY
+ export_idt = DataType["BINARY"]
else:
export_idt = self.get_input_datatype()
# reshape input into folded form
@@ -396,7 +396,7 @@ class ConvolutionInputGenerator1D(HLSCustomOp):
)
)
# binary -> bipolar if needed
- if self.get_output_datatype() == DataType.BIPOLAR:
+ if self.get_output_datatype() == DataType["BIPOLAR"]:
out = context[node.output[0]]
out = 2 * out - 1
context[node.output[0]] = out
@@ -502,9 +502,9 @@ class ConvolutionInputGenerator1D(HLSCustomOp):
def read_npy_data(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_input_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_instream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
@@ -572,9 +572,9 @@ class ConvolutionInputGenerator1D(HLSCustomOp):
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_output_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_outstream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
diff --git a/src/finn/custom_op/fpgadataflow/downsampler.py b/src/finn/custom_op/fpgadataflow/downsampler.py
index 2313ab28b41668b93a55298aa2b589dac999070e..e8948c322a2543ee8ecbf682ee4a7989270ee3ad 100644
--- a/src/finn/custom_op/fpgadataflow/downsampler.py
+++ b/src/finn/custom_op/fpgadataflow/downsampler.py
@@ -164,9 +164,9 @@ class DownSampler(HLSCustomOp):
def read_npy_data(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_input_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_instream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
@@ -197,9 +197,9 @@ class DownSampler(HLSCustomOp):
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_output_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_outstream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
diff --git a/src/finn/custom_op/fpgadataflow/fmpadding_batch.py b/src/finn/custom_op/fpgadataflow/fmpadding_batch.py
index ca0b2f12ab6e84bab0b87e5a34917619c2ba289d..03d3436346aa716e9ea9e49027fdbb17bee74311 100644
--- a/src/finn/custom_op/fpgadataflow/fmpadding_batch.py
+++ b/src/finn/custom_op/fpgadataflow/fmpadding_batch.py
@@ -210,9 +210,9 @@ class FMPadding_Batch(HLSCustomOp):
def read_npy_data(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_input_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_instream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
@@ -261,9 +261,9 @@ class FMPadding_Batch(HLSCustomOp):
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_output_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_outstream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
diff --git a/src/finn/custom_op/fpgadataflow/pool_batch.py b/src/finn/custom_op/fpgadataflow/pool_batch.py
index cef964acd5192ad254e1086dacead590b51e7ec1..f4638e6de3616c568da295f091a1ad39262e6dd8 100644
--- a/src/finn/custom_op/fpgadataflow/pool_batch.py
+++ b/src/finn/custom_op/fpgadataflow/pool_batch.py
@@ -235,9 +235,9 @@ class Pool_Batch(HLSCustomOp):
def read_npy_data(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_input_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_instream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
@@ -296,9 +296,9 @@ class Pool_Batch(HLSCustomOp):
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_output_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_outstream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
diff --git a/src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py b/src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py
index 67e3cd36549d34cab55a931cc040fee5d14ca06f..11809b9bc267d00c8cd630163cc969187efc7417 100644
--- a/src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py
@@ -228,9 +228,9 @@ class StreamingDataWidthConverter_Batch(HLSCustomOp):
def read_npy_data(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_input_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_instream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
@@ -262,9 +262,9 @@ class StreamingDataWidthConverter_Batch(HLSCustomOp):
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_output_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_outstream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
@@ -331,10 +331,10 @@ class StreamingDataWidthConverter_Batch(HLSCustomOp):
exp_shape
), "Input shape does not match expected shape."
- if self.get_input_datatype() == DataType.BIPOLAR:
+ if self.get_input_datatype() == DataType["BIPOLAR"]:
# store bipolar activations as binary
inp = (inp + 1) / 2
- export_idt = DataType.BINARY
+ export_idt = DataType["BINARY"]
else:
export_idt = self.get_input_datatype()
# reshape input into folded shape
@@ -377,7 +377,7 @@ class StreamingDataWidthConverter_Batch(HLSCustomOp):
)
)
# binary -> bipolar if needed
- if self.get_output_datatype() == DataType.BIPOLAR:
+ if self.get_output_datatype() == DataType["BIPOLAR"]:
out = context[node.output[0]]
out = 2 * out - 1
context[node.output[0]] = out
diff --git a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
index 96594d441345332bbe5873570156e07cacbb385d..968c9a6bad0364813c4e70829628da3d07152fbd 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
@@ -509,15 +509,15 @@ class StreamingFCLayer_Batch(HLSCustomOp):
ret = dict()
inp_hls_str = self.get_input_datatype().get_hls_datatype_str()
out_hls_str = self.get_output_datatype().get_hls_datatype_str()
- inp_is_binary = self.get_input_datatype() == DataType.BINARY
- # out_is_binary = self.get_output_datatype() == DataType.BINARY
- wt_is_binary = self.get_weight_datatype() == DataType.BINARY
+ inp_is_binary = self.get_input_datatype() == DataType["BINARY"]
+ # out_is_binary = self.get_output_datatype() == DataType["BINARY"]
+ wt_is_binary = self.get_weight_datatype() == DataType["BINARY"]
bin_xnor_mode = self.get_nodeattr("binaryXnorMode") == 1
if (inp_is_binary or wt_is_binary) and (not bin_xnor_mode):
raise Exception("True binary (non-bipolar) inputs not yet supported")
- inp_is_bipolar = self.get_input_datatype() == DataType.BIPOLAR
- # out_is_bipolar = self.get_output_datatype() == DataType.BIPOLAR
- wt_is_bipolar = self.get_weight_datatype() == DataType.BIPOLAR
+ inp_is_bipolar = self.get_input_datatype() == DataType["BIPOLAR"]
+ # out_is_bipolar = self.get_output_datatype() == DataType["BIPOLAR"]
+ wt_is_bipolar = self.get_weight_datatype() == DataType["BIPOLAR"]
# reinterpret inp/wt as bipolar if bin_xnor_mode is iset
inp_is_bipolar = inp_is_bipolar or (inp_is_binary and bin_xnor_mode)
wt_is_bipolar = wt_is_bipolar or (wt_is_binary and bin_xnor_mode)
@@ -567,7 +567,7 @@ class StreamingFCLayer_Batch(HLSCustomOp):
# ONNX uses (in_features, out_features) and matmul(x, W)
# finn-hlslib uses (out_features, in_features) and matmul(W, x)
ret = orig_weight_matrix.T
- if self.get_weight_datatype() == DataType.BIPOLAR:
+ if self.get_weight_datatype() == DataType["BIPOLAR"]:
# convert bipolar to binary
ret = (ret + 1) / 2
# interleave rows between PEs and reshape
@@ -658,11 +658,11 @@ class StreamingFCLayer_Batch(HLSCustomOp):
), """Threshold matrix dimension is
not as expected (2)."""
n_thres_steps = orig_thres_matrix.shape[1]
- inp_is_bipolar = self.get_input_datatype() == DataType.BIPOLAR
- wt_is_bipolar = self.get_weight_datatype() == DataType.BIPOLAR
+ inp_is_bipolar = self.get_input_datatype() == DataType["BIPOLAR"]
+ wt_is_bipolar = self.get_weight_datatype() == DataType["BIPOLAR"]
# reinterpret inp/wt as bipolar if bin_xnor_mode is iset
- inp_is_binary = self.get_input_datatype() == DataType.BINARY
- wt_is_binary = self.get_weight_datatype() == DataType.BINARY
+ inp_is_binary = self.get_input_datatype() == DataType["BINARY"]
+ wt_is_binary = self.get_weight_datatype() == DataType["BINARY"]
bin_xnor_mode = self.get_nodeattr("binaryXnorMode") == 1
inp_is_bipolar = inp_is_bipolar or (inp_is_binary and bin_xnor_mode)
wt_is_bipolar = wt_is_bipolar or (wt_is_binary and bin_xnor_mode)
@@ -717,8 +717,8 @@ class StreamingFCLayer_Batch(HLSCustomOp):
export_wdt = self.get_weight_datatype()
# we have converted bipolar weights to binary for export,
# so use it as such for weight generation
- if self.get_weight_datatype() == DataType.BIPOLAR:
- export_wdt = DataType.BINARY
+ if self.get_weight_datatype() == DataType["BIPOLAR"]:
+ export_wdt = DataType["BINARY"]
if weight_file_mode == "hls_header":
weight_hls_code = numpy_to_hls_code(
weight_tensor, export_wdt, "weights", True, True
@@ -847,11 +847,11 @@ class StreamingFCLayer_Batch(HLSCustomOp):
if thresholds is not None:
threshold_tensor = self.get_hls_compatible_threshold_tensor(thresholds)
# use UINT32 threshold export for bipolar times bipolar
- inp_is_bipolar = self.get_input_datatype() == DataType.BIPOLAR
- wt_is_bipolar = self.get_weight_datatype() == DataType.BIPOLAR
+ inp_is_bipolar = self.get_input_datatype() == DataType["BIPOLAR"]
+ wt_is_bipolar = self.get_weight_datatype() == DataType["BIPOLAR"]
# reinterpret inp/wt as bipolar if bin_xnor_mode is iset
- inp_is_binary = self.get_input_datatype() == DataType.BINARY
- wt_is_binary = self.get_weight_datatype() == DataType.BINARY
+ inp_is_binary = self.get_input_datatype() == DataType["BINARY"]
+ wt_is_binary = self.get_weight_datatype() == DataType["BINARY"]
bin_xnor_mode = self.get_nodeattr("binaryXnorMode") == 1
inp_is_bipolar = inp_is_bipolar or (inp_is_binary and bin_xnor_mode)
wt_is_bipolar = wt_is_bipolar or (wt_is_binary and bin_xnor_mode)
@@ -872,8 +872,8 @@ class StreamingFCLayer_Batch(HLSCustomOp):
tdt_hls = tdt.get_hls_datatype_str()
# use binary to export bipolar activations
export_odt = self.get_output_datatype()
- if self.get_output_datatype() == DataType.BIPOLAR:
- export_odt = DataType.BINARY
+ if self.get_output_datatype() == DataType["BIPOLAR"]:
+ export_odt = DataType["BINARY"]
odt_hls = export_odt.get_hls_datatype_str()
f_thresh.write(
"static ThresholdsActivation<{},{},{},{},{},{},{}> threshs \
@@ -921,10 +921,10 @@ class StreamingFCLayer_Batch(HLSCustomOp):
not float32 as expected."""
expected_inp_shape = self.get_folded_input_shape()
reshaped_input = context[inputs].reshape(expected_inp_shape)
- if self.get_input_datatype() == DataType.BIPOLAR:
+ if self.get_input_datatype() == DataType["BIPOLAR"]:
# store bipolar activations as binary
reshaped_input = (reshaped_input + 1) / 2
- export_idt = DataType.BINARY
+ export_idt = DataType["BINARY"]
else:
export_idt = self.get_input_datatype()
# make copy before saving the array
@@ -943,7 +943,7 @@ class StreamingFCLayer_Batch(HLSCustomOp):
# load output npy file
super().npy_to_dynamic_output(context)
# reinterpret binary output as bipolar where needed
- if self.get_output_datatype() == DataType.BIPOLAR:
+ if self.get_output_datatype() == DataType["BIPOLAR"]:
out = context[node.output[0]]
out = 2 * out - 1
context[node.output[0]] = out
@@ -966,8 +966,8 @@ class StreamingFCLayer_Batch(HLSCustomOp):
export_wdt = self.get_weight_datatype()
# we have converted bipolar weights to binary for export,
# so use it as such for weight generation
- if self.get_weight_datatype() == DataType.BIPOLAR:
- export_wdt = DataType.BINARY
+ if self.get_weight_datatype() == DataType["BIPOLAR"]:
+ export_wdt = DataType["BINARY"]
wei = npy_to_rtlsim_input(
"{}/weights.npy".format(code_gen_dir), export_wdt, wnbits
)
@@ -1058,9 +1058,9 @@ class StreamingFCLayer_Batch(HLSCustomOp):
def read_npy_data(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_input_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_instream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
@@ -1134,8 +1134,8 @@ class StreamingFCLayer_Batch(HLSCustomOp):
]
elif mem_mode == "decoupled" or mem_mode == "external":
wdt = self.get_weight_datatype()
- if wdt == DataType.BIPOLAR:
- export_wdt = DataType.BINARY
+ if wdt == DataType["BIPOLAR"]:
+ export_wdt = DataType["BINARY"]
else:
export_wdt = wdt
wdtype_hls_str = export_wdt.get_hls_datatype_str()
@@ -1160,9 +1160,9 @@ class StreamingFCLayer_Batch(HLSCustomOp):
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_output_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_outstream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
diff --git a/src/finn/custom_op/fpgadataflow/streamingfifo.py b/src/finn/custom_op/fpgadataflow/streamingfifo.py
index 9653d698f54e2b97c66bff62c0b3c11057b36aad..c8ae83cc90e1199831da5286d5051fdb969e825a 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfifo.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfifo.py
@@ -261,10 +261,10 @@ class StreamingFIFO(HLSCustomOp):
not float32 as expected."""
expected_inp_shape = self.get_folded_input_shape()
reshaped_input = inp.reshape(expected_inp_shape)
- if DataType[self.get_nodeattr("dataType")] == DataType.BIPOLAR:
+ if DataType[self.get_nodeattr("dataType")] == DataType["BIPOLAR"]:
# store bipolar activations as binary
reshaped_input = (reshaped_input + 1) / 2
- export_idt = DataType.BINARY
+ export_idt = DataType["BINARY"]
else:
export_idt = DataType[self.get_nodeattr("dataType")]
# make copy before saving the array
diff --git a/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py b/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
index 19a42fe2d6b53879d401ec8bd462ddd59623dc1e..87ecde8f9c7c50b7a22213db2b856d7439050421 100644
--- a/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
@@ -205,9 +205,9 @@ class StreamingMaxPool_Batch(HLSCustomOp):
def read_npy_data(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_input_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_instream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
@@ -255,9 +255,9 @@ class StreamingMaxPool_Batch(HLSCustomOp):
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_output_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_outstream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
@@ -323,10 +323,10 @@ class StreamingMaxPool_Batch(HLSCustomOp):
inp.shape == exp_ishape
), """Input shape doesn't
match expected shape (1, ifm_dim, ifm_dim, ifm_ch)."""
- if self.get_input_datatype() == DataType.BIPOLAR:
+ if self.get_input_datatype() == DataType["BIPOLAR"]:
# store bipolar activations as binary
inp = (inp + 1) / 2
- export_idt = DataType.BINARY
+ export_idt = DataType["BINARY"]
else:
export_idt = self.get_input_datatype()
# no reshaping for input since assuming no folding on input
@@ -373,7 +373,7 @@ class StreamingMaxPool_Batch(HLSCustomOp):
)
)
# binary -> bipolar if needed
- if self.get_output_datatype() == DataType.BIPOLAR:
+ if self.get_output_datatype() == DataType["BIPOLAR"]:
out = context[node.output[0]]
out = 2 * out - 1
context[node.output[0]] = out
diff --git a/src/finn/custom_op/fpgadataflow/thresholding_batch.py b/src/finn/custom_op/fpgadataflow/thresholding_batch.py
index 7fb7634dc22e1d00569e7bb755bf120d6de4f808..81ce7fe8c5730c774f80d2b1640d5241e02b0b85 100644
--- a/src/finn/custom_op/fpgadataflow/thresholding_batch.py
+++ b/src/finn/custom_op/fpgadataflow/thresholding_batch.py
@@ -390,8 +390,8 @@ class Thresholding_Batch(HLSCustomOp):
tdt_hls = tdt.get_hls_datatype_str()
# use binary to export bipolar activations
export_odt = self.get_output_datatype()
- if self.get_output_datatype() == DataType.BIPOLAR:
- export_odt = DataType.BINARY
+ if self.get_output_datatype() == DataType["BIPOLAR"]:
+ export_odt = DataType["BINARY"]
odt_hls = export_odt.get_hls_datatype_str()
f_thresh.write(
"static ThresholdsActivation<{},{},{},{},{},{},{}> threshs \
@@ -515,10 +515,10 @@ class Thresholding_Batch(HLSCustomOp):
not float32 as expected."""
expected_inp_shape = self.get_folded_input_shape()
reshaped_input = context[inputs].reshape(expected_inp_shape)
- if self.get_input_datatype() == DataType.BIPOLAR:
+ if self.get_input_datatype() == DataType["BIPOLAR"]:
# store bipolar activations as binary
reshaped_input = (reshaped_input + 1) / 2
- export_idt = DataType.BINARY
+ export_idt = DataType["BINARY"]
else:
export_idt = self.get_input_datatype()
# make copy before saving the array
@@ -537,7 +537,7 @@ class Thresholding_Batch(HLSCustomOp):
# load output npy file
super().npy_to_dynamic_output(context)
# reinterpret binary output as bipolar where needed
- if self.get_output_datatype() == DataType.BIPOLAR:
+ if self.get_output_datatype() == DataType["BIPOLAR"]:
out = context[node.output[0]]
out = 2 * out - 1
context[node.output[0]] = out
@@ -711,9 +711,9 @@ class Thresholding_Batch(HLSCustomOp):
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_output_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_outstream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
diff --git a/src/finn/custom_op/fpgadataflow/upsampler.py b/src/finn/custom_op/fpgadataflow/upsampler.py
index e8aa09b1c0754b68e37d01551afe90811f22e7cd..8331610dc11bb440e6bb56924bbf8efeed2653c2 100644
--- a/src/finn/custom_op/fpgadataflow/upsampler.py
+++ b/src/finn/custom_op/fpgadataflow/upsampler.py
@@ -147,9 +147,9 @@ class UpsampleNearestNeighbour_Batch(HLSCustomOp):
def read_npy_data(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_input_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_instream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
@@ -180,9 +180,9 @@ class UpsampleNearestNeighbour_Batch(HLSCustomOp):
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
dtype = self.get_output_datatype()
- if dtype == DataType.BIPOLAR:
+ if dtype == DataType["BIPOLAR"]:
# use binary for bipolar storage
- dtype = DataType.BINARY
+ dtype = DataType["BINARY"]
elem_bits = dtype.bitwidth()
packed_bits = self.get_outstream_width()
packed_hls_type = "ap_uint<%d>" % packed_bits
diff --git a/src/finn/custom_op/fpgadataflow/vector_vector_activate_batch.py b/src/finn/custom_op/fpgadataflow/vector_vector_activate_batch.py
index 9fc133b9bc09ca556df5779970d68d4b62131659..fa990f28087c0ec47a071e85ed1af6ff9328f70f 100644
--- a/src/finn/custom_op/fpgadataflow/vector_vector_activate_batch.py
+++ b/src/finn/custom_op/fpgadataflow/vector_vector_activate_batch.py
@@ -236,8 +236,8 @@ class Vector_Vector_Activate_Batch(HLSCustomOp):
ret = dict()
inp_hls_str = self.get_input_datatype().get_hls_datatype_str()
out_hls_str = self.get_output_datatype().get_hls_datatype_str()
- inp_is_bipolar = self.get_input_datatype() == DataType.BIPOLAR
- wt_is_bipolar = self.get_weight_datatype() == DataType.BIPOLAR
+ inp_is_bipolar = self.get_input_datatype() == DataType["BIPOLAR"]
+ wt_is_bipolar = self.get_weight_datatype() == DataType["BIPOLAR"]
# fill in TSrcI and TWeightI
# TODO handle bipolar inputs
if inp_is_bipolar or wt_is_bipolar:
diff --git a/src/finn/qnn-data/templates/driver/driver_base.py b/src/finn/qnn-data/templates/driver/driver_base.py
index df1ab15c7892d66a668d82040b0da93366942cb7..b6dd8350809a33ab5dad3e21b0f52f41cbe872ec 100644
--- a/src/finn/qnn-data/templates/driver/driver_base.py
+++ b/src/finn/qnn-data/templates/driver/driver_base.py
@@ -456,9 +456,9 @@ class FINNExampleOverlay(Overlay):
# also benchmark driver-related overheads
input_npy = gen_finn_dt_tensor(self.idt(), self.ishape_normal())
# provide as int8/uint8 to support fast packing path where possible
- if self.idt() == DataType.UINT8:
+ if self.idt() == DataType["UINT8"]:
input_npy = input_npy.astype(np.uint8)
- elif self.idt() == DataType.INT8:
+ elif self.idt() == DataType["INT8"]:
input_npy = input_npy.astype(np.int8)
start = time.time()
ibuf_folded = self.fold_input(input_npy)
diff --git a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
index 3cb193055f3a455d95f7735ab38b2601809dbabd..4ab9c8fe27fce2a71cd9e21e6a94c1eb706cc47a 100644
--- a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
+++ b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
@@ -589,18 +589,18 @@ class InferBinaryStreamingFCLayer(Transformation):
mm_output = n.output[0]
mm_in_shape = model.get_tensor_shape(mm_input)
mm_out_shape = model.get_tensor_shape(mm_output)
- assert model.get_tensor_datatype(mm_input) == DataType.BINARY, (
+ assert model.get_tensor_datatype(mm_input) == DataType["BINARY"], (
n.name
+ """: First
input for xnorpopcount is not set to FINN DataType BINARY."""
)
- assert model.get_tensor_datatype(mm_weight) == DataType.BINARY, (
+ assert model.get_tensor_datatype(mm_weight) == DataType["BINARY"], (
n.name
+ """: Second
input (weights) for xnorpopcount is not set to FINN DataType BINARY."""
)
- idt = DataType.BINARY
- wdt = DataType.BINARY
+ idt = DataType["BINARY"]
+ wdt = DataType["BINARY"]
mm_output = n.output[0]
W = model.get_initializer(mm_weight)
# extract weight shape, note that ONNX and finn-hlslib
@@ -766,7 +766,7 @@ class InferQuantizedStreamingFCLayer(Transformation):
+ ": out_bias must be integer for HLS conversion."
)
actval = int(actval)
- odt_is_bipolar = odt == DataType.BIPOLAR
+ odt_is_bipolar = odt == DataType["BIPOLAR"]
bipolar_ok = (
odt_is_bipolar and (scale == 2.0) and (actval == -1)
)
@@ -1254,7 +1254,7 @@ class InferChannelwiseLinearLayer(Transformation):
for k in DataType.__members__:
dt = DataType[k]
- if dt in [DataType.BIPOLAR, DataType.TERNARY, DataType.FLOAT32]:
+ if dt in [DataType["BIPOLAR"], DataType["TERNARY"], DataType["FLOAT32"]]:
# not currently supported
continue
@@ -1270,9 +1270,9 @@ class InferChannelwiseLinearLayer(Transformation):
)
if (0 <= vals).all():
- return DataType.UINT64
+ return DataType["UINT64"]
else:
- return DataType.INT64
+ return DataType["INT64"]
def apply(self, model):
graph = model.graph
diff --git a/src/finn/transformation/streamline/absorb.py b/src/finn/transformation/streamline/absorb.py
index cf712b38054c78c6e414ad914ab67378daec5d12..cba9648187ffadcff048b045f0c85c81c770e44d 100644
--- a/src/finn/transformation/streamline/absorb.py
+++ b/src/finn/transformation/streamline/absorb.py
@@ -205,7 +205,7 @@ class FactorOutMulSignMagnitude(Transformation):
actual_ndims = len(tuple(filter(lambda x: x > 1, A.shape)))
is_1d = actual_ndims == 1
is_not_bipolar = (
- model.get_tensor_datatype(mul_weight_name) != DataType.BIPOLAR
+ model.get_tensor_datatype(mul_weight_name) != DataType["BIPOLAR"]
)
is_signed = (A < 0).any()
if is_signed and (is_scalar or is_1d) and is_not_bipolar:
@@ -217,7 +217,7 @@ class FactorOutMulSignMagnitude(Transformation):
# create new mul node with sign(A) as the operand
sgn = np.sign(A)
model.set_initializer(sign_mul_param_name, sgn)
- model.set_tensor_datatype(sign_mul_param_name, DataType.BIPOLAR)
+ model.set_tensor_datatype(sign_mul_param_name, DataType["BIPOLAR"])
# replace original mul weight by magnitudes
model.set_initializer(mul_weight_name, np.abs(A))
new_mul = oh.make_node(
@@ -457,7 +457,7 @@ class AbsorbScalarMulAddIntoTopK(Transformation):
graph.node.remove(prod)
n.input[0] = prod_input
# to avoid error the dataype is set to float32
- model.set_tensor_datatype(n.input[0], DataType.FLOAT32)
+ model.set_tensor_datatype(n.input[0], DataType["FLOAT32"])
graph_modified = True
if graph_modified:
model = model.transform(InferShapes())
diff --git a/src/finn/transformation/streamline/collapse_repeated.py b/src/finn/transformation/streamline/collapse_repeated.py
index 50265046d94db1e7233a45b934fd68f08431a95d..92c48c84ffa1a161f623ef6b22caaeb92f4a8199 100644
--- a/src/finn/transformation/streamline/collapse_repeated.py
+++ b/src/finn/transformation/streamline/collapse_repeated.py
@@ -85,8 +85,8 @@ class CollapseRepeatedOp(Transformation):
# replace parameter value
model.set_initializer(new_node_param_name, new_param)
# be conservative with param/output DataTypes
- model.set_tensor_datatype(new_node_param_name, DataType.FLOAT32)
- model.set_tensor_datatype(end_name, DataType.FLOAT32)
+ model.set_tensor_datatype(new_node_param_name, DataType["FLOAT32"])
+ model.set_tensor_datatype(end_name, DataType["FLOAT32"])
# remove old nodes
graph.node.remove(n)
graph.node.remove(consumer)
diff --git a/src/finn/transformation/streamline/reorder.py b/src/finn/transformation/streamline/reorder.py
index cd44e115eed42d1c0529b86a49e8855ff7c492ce..416755ae21002716e1535094d58aa8593ce3221a 100644
--- a/src/finn/transformation/streamline/reorder.py
+++ b/src/finn/transformation/streamline/reorder.py
@@ -408,16 +408,16 @@ class MoveMulPastDWConv(Transformation):
# rewire mul input to be conv input
conv_node.input[0] = start_name
model.set_tensor_shape(start_name, conv_in_shape)
- model.set_tensor_datatype(start_name, DataType.FLOAT32)
+ model.set_tensor_datatype(start_name, DataType["FLOAT32"])
# use old conv input tensor as conv output
conv_node.output[0] = conv_in_name
model.set_tensor_shape(conv_in_name, conv_out_shape)
- model.set_tensor_datatype(conv_in_name, DataType.FLOAT32)
+ model.set_tensor_datatype(conv_in_name, DataType["FLOAT32"])
# use new conv output as new mul node input
mul_node.input[0] = conv_in_name
# use old conv output as new mul node output
mul_node.output[0] = conv_out_name
- model.set_tensor_datatype(conv_out_name, DataType.FLOAT32)
+ model.set_tensor_datatype(conv_out_name, DataType["FLOAT32"])
# move mul node past conv node
graph.node.remove(mul_node)
graph.node.insert(node_ind, mul_node)
@@ -482,16 +482,16 @@ class MoveMulPastMaxPool(Transformation):
# rewire mul input to be maxpool input
maxpool_node.input[0] = start_name
model.set_tensor_shape(start_name, maxpool_in_shape)
- model.set_tensor_datatype(start_name, DataType.FLOAT32)
+ model.set_tensor_datatype(start_name, DataType["FLOAT32"])
# use old maxpool input tensor as maxpool output
maxpool_node.output[0] = maxpool_in_name
model.set_tensor_shape(maxpool_in_name, maxpool_out_shape)
- model.set_tensor_datatype(maxpool_in_name, DataType.FLOAT32)
+ model.set_tensor_datatype(maxpool_in_name, DataType["FLOAT32"])
# use new maxpool output as new mul node input
mul_node.input[0] = maxpool_in_name
# use old maxpool output as new mul node output
mul_node.output[0] = maxpool_out_name
- model.set_tensor_datatype(maxpool_out_name, DataType.FLOAT32)
+ model.set_tensor_datatype(maxpool_out_name, DataType["FLOAT32"])
# move mul node past maxpool node
graph.node.remove(mul_node)
graph.node.insert(node_ind, mul_node)
@@ -638,7 +638,7 @@ class MoveScalarLinearPastInvariants(Transformation):
model.set_tensor_shape(n.output[0], out_shape)
model.set_tensor_shape(prod0.output[0], out_shape)
model.set_tensor_datatype(prod0.output[0], scalar_op_odt)
- model.set_tensor_datatype(n.output[0], DataType.FLOAT32)
+ model.set_tensor_datatype(n.output[0], DataType["FLOAT32"])
graph.node.remove(prod0)
graph.node.insert(node_ind - 1, prod0)
graph_modified = True
diff --git a/src/finn/transformation/streamline/sign_to_thres.py b/src/finn/transformation/streamline/sign_to_thres.py
index 13f2e8524af7ce2d3457d0637f1c6d02733f504b..61d7eb35430262b1ee90dfa478076fb6f7556612 100644
--- a/src/finn/transformation/streamline/sign_to_thres.py
+++ b/src/finn/transformation/streamline/sign_to_thres.py
@@ -69,6 +69,6 @@ class ConvertSignToThres(Transformation):
graph.node.insert(node_ind, mt_node)
graph.node.remove(n)
# add quantization annotations
- model.set_tensor_datatype(sign_out_name, DataType.BIPOLAR)
+ model.set_tensor_datatype(sign_out_name, DataType["BIPOLAR"])
graph_modified = True
return (model, graph_modified)
diff --git a/src/finn/util/create.py b/src/finn/util/create.py
index d9c5d7b1b59916edfc8730992535f3ddb57c4d60..62229a69b68c26dd191b3e1d4a44f1bb8b19ed07 100644
--- a/src/finn/util/create.py
+++ b/src/finn/util/create.py
@@ -49,10 +49,10 @@ def hls_random_mlp_maker(layer_spec):
# no activation, produce accumulators
T = None
tdt = None
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
- odt = DataType.UINT32
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
+ odt = DataType["UINT32"]
else:
- odt = DataType.INT32
+ odt = DataType["INT32"]
else:
odt = act
(min, max) = calculate_signed_dot_prod_range(idt, wdt, mw)
@@ -61,13 +61,13 @@ def hls_random_mlp_maker(layer_spec):
# provide non-decreasing thresholds
T = np.sort(T, axis=1)
# generate thresholds for activation
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
- tdt = DataType.UINT32
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
+ tdt = DataType["UINT32"]
# bias thresholds to be positive
T = np.ceil((T + mw) / 2)
assert (T >= 0).all()
else:
- tdt = DataType.INT32
+ tdt = DataType["INT32"]
lyr["T"] = T
lyr["tdt"] = tdt
lyr["odt"] = odt
@@ -120,11 +120,11 @@ def hls_mlp_maker(layer_spec):
# StreamingFC:
# - specify their datatypes as such
# - specify their datatypes as BINARY as use binaryXnorMode
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
# we'll internally convert weights/inputs to binary and specify the
# datatypes as such, and also set the binaryXnorMode attribute to 1
- export_wdt = DataType.BINARY
- export_idt = DataType.BINARY
+ export_wdt = DataType["BINARY"]
+ export_idt = DataType["BINARY"]
binary_xnor_mode = 1
else:
export_wdt = wdt
@@ -134,7 +134,7 @@ def hls_mlp_maker(layer_spec):
if T is not None:
no_act = 0
node_inp_list = [current_in_name, current_W_name, current_T_name]
- if odt == DataType.BIPOLAR:
+ if odt == DataType["BIPOLAR"]:
actval = 0
else:
actval = odt.min()
diff --git a/tests/brevitas/test_brevitas_QConv2d.py b/tests/brevitas/test_brevitas_QConv2d.py
index c1f790946bfa5f53194b96b1fea9c1722797a4a0..9de2efbcee627384cb76d1e05d0495cf6b40b169 100644
--- a/tests/brevitas/test_brevitas_QConv2d.py
+++ b/tests/brevitas/test_brevitas_QConv2d.py
@@ -86,7 +86,7 @@ def test_brevitas_QConv2d(dw, bias, in_channels):
weight_narrow_range=True,
weight_scaling_min_val=2e-16,
)
- weight_tensor = gen_finn_dt_tensor(DataType.INT4, w_shape)
+ weight_tensor = gen_finn_dt_tensor(DataType["INT4"], w_shape)
b_conv.weight = torch.nn.Parameter(torch.from_numpy(weight_tensor).float())
b_conv.eval()
bo.export_finn_onnx(b_conv, ishape, export_onnx_path)
diff --git a/tests/brevitas/test_brevitas_mobilenet.py b/tests/brevitas/test_brevitas_mobilenet.py
index eb642adada9bd9abb8a328518770899d3da96ada..108c97c2e83b7f3ca9dd6ead746b3ef8b4d10af5 100644
--- a/tests/brevitas/test_brevitas_mobilenet.py
+++ b/tests/brevitas/test_brevitas_mobilenet.py
@@ -78,7 +78,9 @@ def test_brevitas_mobilenet():
bo.export_finn_onnx(preproc, (1, 3, 224, 224), preproc_onnx)
preproc_model = ModelWrapper(preproc_onnx)
# set input finn datatype to UINT8
- preproc_model.set_tensor_datatype(preproc_model.graph.input[0].name, DataType.UINT8)
+ preproc_model.set_tensor_datatype(
+ preproc_model.graph.input[0].name, DataType["UINT8"]
+ )
preproc_model = preproc_model.transform(InferShapes())
preproc_model = preproc_model.transform(GiveUniqueNodeNames())
preproc_model = preproc_model.transform(GiveUniqueParameterTensors())
diff --git a/tests/brevitas/test_brevitas_qlinear.py b/tests/brevitas/test_brevitas_qlinear.py
index 873866b37727730b7cedd035f5edd93f7c1afe32..67e4d04d066a2a1a6baf78429d91e724a9d80e7f 100644
--- a/tests/brevitas/test_brevitas_qlinear.py
+++ b/tests/brevitas/test_brevitas_qlinear.py
@@ -48,7 +48,7 @@ export_onnx_path = "test_brevitas_qlinear.onnx"
@pytest.mark.parametrize("out_features", [4])
@pytest.mark.parametrize("in_features", [3])
@pytest.mark.parametrize("w_bits", [4])
-@pytest.mark.parametrize("i_dtype", [DataType.UINT4])
+@pytest.mark.parametrize("i_dtype", [DataType["UINT4"]])
def test_brevitas_qlinear(bias, out_features, in_features, w_bits, i_dtype):
i_shape = (1, in_features)
w_shape = (out_features, in_features)
diff --git a/tests/end2end/test_end2end_bnn_pynq.py b/tests/end2end/test_end2end_bnn_pynq.py
index 6790485ceab373cd539727aefc59fa8999b3b192..14e10da86ec5cec14dfabdd0b4f6c2b92dd42519 100644
--- a/tests/end2end/test_end2end_bnn_pynq.py
+++ b/tests/end2end/test_end2end_bnn_pynq.py
@@ -352,7 +352,7 @@ class TestEnd2End:
model = model.transform(MergeONNXModels(pre_model))
# add input quantization annotation: UINT8 for all BNN-PYNQ models
global_inp_name = model.graph.input[0].name
- model.set_tensor_datatype(global_inp_name, DataType.UINT8)
+ model.set_tensor_datatype(global_inp_name, DataType["UINT8"])
# postprocessing: insert Top-1 node at the end
model = model.transform(InsertTopK(k=1))
chkpt_name = get_checkpoint_name(topology, wbits, abits, "pre_post")
diff --git a/tests/end2end/test_end2end_cybsec_mlp.py b/tests/end2end/test_end2end_cybsec_mlp.py
index 7b4cebb52b3e4758746d4054827c6f96e8a4d681..23a5d23f1a91798b834797b3a8ccc35d07b2e61a 100644
--- a/tests/end2end/test_end2end_cybsec_mlp.py
+++ b/tests/end2end/test_end2end_cybsec_mlp.py
@@ -141,9 +141,11 @@ def test_end2end_cybsec_mlp_export():
assert finn_model.graph.node[3].op_type == "MatMul"
assert finn_model.graph.node[-1].op_type == "MultiThreshold"
# verify datatypes on some tensors
- assert finn_model.get_tensor_datatype(finnonnx_in_tensor_name) == DataType.BIPOLAR
+ assert (
+ finn_model.get_tensor_datatype(finnonnx_in_tensor_name) == DataType["BIPOLAR"]
+ )
first_matmul_w_name = finn_model.graph.node[3].input[1]
- assert finn_model.get_tensor_datatype(first_matmul_w_name) == DataType.INT2
+ assert finn_model.get_tensor_datatype(first_matmul_w_name) == DataType["INT2"]
@pytest.mark.slow
diff --git a/tests/end2end/test_end2end_mobilenet_v1.py b/tests/end2end/test_end2end_mobilenet_v1.py
index 760c77ea406386ce4886b21bcb042808984dcb7f..0d639ae084af257db61c85ff0b5c0d5101539b71 100644
--- a/tests/end2end/test_end2end_mobilenet_v1.py
+++ b/tests/end2end/test_end2end_mobilenet_v1.py
@@ -97,7 +97,9 @@ def test_end2end_mobilenet_export():
bo.export_finn_onnx(preproc, (1, 3, 224, 224), preproc_onnx)
preproc_model = ModelWrapper(preproc_onnx)
# set input finn datatype to UINT8
- preproc_model.set_tensor_datatype(preproc_model.graph.input[0].name, DataType.UINT8)
+ preproc_model.set_tensor_datatype(
+ preproc_model.graph.input[0].name, DataType["UINT8"]
+ )
preproc_model = preproc_model.transform(InferShapes())
preproc_model = preproc_model.transform(FoldConstants())
preproc_model = preproc_model.transform(GiveUniqueNodeNames())
diff --git a/tests/fpgadataflow/test_code_gen_trafo.py b/tests/fpgadataflow/test_code_gen_trafo.py
index 89fab37d6d5225383ccb13a748c83573d6ee4516..5ddff3d36f03d17833e17bc98649a64dabf31577 100644
--- a/tests/fpgadataflow/test_code_gen_trafo.py
+++ b/tests/fpgadataflow/test_code_gen_trafo.py
@@ -39,7 +39,7 @@ from finn.transformation.fpgadataflow.prepare_cppsim import PrepareCppSim
@pytest.mark.vivado
def test_code_gen_trafo():
- idt = wdt = odt = DataType.BIPOLAR
+ idt = wdt = odt = DataType["BIPOLAR"]
mw = 8
mh = 8
pe = 4
diff --git a/tests/fpgadataflow/test_compilation_trafo.py b/tests/fpgadataflow/test_compilation_trafo.py
index 6284748b9ccdc422b42bd9e301eb395d8dd1ad45..81e2ff9a7c5829982cdb6121378e9e9e3af81632 100644
--- a/tests/fpgadataflow/test_compilation_trafo.py
+++ b/tests/fpgadataflow/test_compilation_trafo.py
@@ -40,7 +40,7 @@ from finn.transformation.fpgadataflow.prepare_cppsim import PrepareCppSim
@pytest.mark.vivado
def test_compilation_trafo():
- idt = wdt = odt = DataType.BIPOLAR
+ idt = wdt = odt = DataType["BIPOLAR"]
mw = 8
mh = 8
pe = 4
diff --git a/tests/fpgadataflow/test_convert_to_hls_1d_conv_layer.py b/tests/fpgadataflow/test_convert_to_hls_1d_conv_layer.py
index 4e7030449c87b81d7a492b0e76dd05a047be3858..5cc5f8fa6c1ccd3e5a9e154b6fb2773caf4668a9 100644
--- a/tests/fpgadataflow/test_convert_to_hls_1d_conv_layer.py
+++ b/tests/fpgadataflow/test_convert_to_hls_1d_conv_layer.py
@@ -72,7 +72,7 @@ from finn.util.basic import gen_finn_dt_tensor
def test_convert_to_hls_1d_conv_layer(conv_config, depthwise, exec_mode):
pad, kernel_size, stride, dilation = conv_config
np.random.seed(0)
- idt = DataType.UINT4
+ idt = DataType["UINT4"]
in_feature_dim_h, in_feature_dim_w = [10, 1]
in_chn = 16
@@ -101,7 +101,7 @@ def test_convert_to_hls_1d_conv_layer(conv_config, depthwise, exec_mode):
input_shape = [1, in_chn, in_feature_dim_h, in_feature_dim_w]
output_shape = [1, out_chn, out_feature_dim_h, out_feature_dim_w]
- conv_weight_dt = DataType.UINT4
+ conv_weight_dt = DataType["UINT4"]
conv_config = {}
conv_config["dilations"] = [dilation_h, dilation_w]
diff --git a/tests/fpgadataflow/test_convert_to_hls_channelwise_layer.py b/tests/fpgadataflow/test_convert_to_hls_channelwise_layer.py
index 8dd927fa7628d1500fe644b030278fbaa3f18810..bf690d1d68bc0f580663735c3596c1dfc0a651e8 100644
--- a/tests/fpgadataflow/test_convert_to_hls_channelwise_layer.py
+++ b/tests/fpgadataflow/test_convert_to_hls_channelwise_layer.py
@@ -76,9 +76,13 @@ def make_single_maxpool_modelwrapper(onnx_op_name, ishape, idt, pdt, pshape):
# parameter datatype
-@pytest.mark.parametrize("pdt", [DataType.BIPOLAR, DataType.UINT4, DataType.INT2])
+@pytest.mark.parametrize(
+ "pdt", [DataType["BIPOLAR"], DataType["UINT4"], DataType["INT2"]]
+)
# input datatype
-@pytest.mark.parametrize("idt", [DataType.INT32, DataType.UINT4, DataType.INT4])
+@pytest.mark.parametrize(
+ "idt", [DataType["INT32"], DataType["UINT4"], DataType["INT4"]]
+)
# function
@pytest.mark.parametrize("onnx_op_name", ["Add", "Mul"])
# vector parameter or scalar parameter (broadcast)
@@ -103,10 +107,10 @@ def test_convert_to_hls_channelwise_layer(
# Since the aren't Data types with a bit width of a non power of 2,
# there are cases where the input won't use it full range.
- if idt == DataType.INT32:
- x = gen_finn_dt_tensor(DataType.INT16, (1, ifm_ch, ifm_dim, ifm_dim))
- elif idt == DataType.UINT32:
- x = gen_finn_dt_tensor(DataType.UINT16, (1, ifm_ch, ifm_dim, ifm_dim))
+ if idt == DataType["INT32"]:
+ x = gen_finn_dt_tensor(DataType["INT16"], (1, ifm_ch, ifm_dim, ifm_dim))
+ elif idt == DataType["UINT32"]:
+ x = gen_finn_dt_tensor(DataType["UINT16"], (1, ifm_ch, ifm_dim, ifm_dim))
else:
x = gen_finn_dt_tensor(idt, (1, ifm_ch, ifm_dim, ifm_dim))
diff --git a/tests/fpgadataflow/test_convert_to_hls_conv_fc_transition.py b/tests/fpgadataflow/test_convert_to_hls_conv_fc_transition.py
index cf2903a5789d7d3892ac549338b274268c1661b3..9b0f3d68aed655f0b36857d50a085093ea94aecb 100755
--- a/tests/fpgadataflow/test_convert_to_hls_conv_fc_transition.py
+++ b/tests/fpgadataflow/test_convert_to_hls_conv_fc_transition.py
@@ -79,10 +79,10 @@ def get_multithreshold_rand_params(channels, num_of_thres, seed=None):
@pytest.mark.slow
def test_convert_to_hls_conv_fc_transition(conv_config, depthwise, use_reshape):
np.random.seed(0)
- idt = DataType.UINT4
- odt = DataType.UINT4
- conv_weight_dt = DataType.INT4
- fc_weight_dt = DataType.INT4
+ idt = DataType["UINT4"]
+ odt = DataType["UINT4"]
+ conv_weight_dt = DataType["INT4"]
+ fc_weight_dt = DataType["INT4"]
input_shape, kernel_shape, stride, pad = conv_config
kernel_size_h, kernel_size_w = kernel_shape
@@ -186,8 +186,8 @@ def test_convert_to_hls_conv_fc_transition(conv_config, depthwise, use_reshape):
model.set_tensor_datatype("global_out", odt)
model.set_tensor_datatype("conv_param", conv_weight_dt)
model.set_tensor_datatype("matmul_param", fc_weight_dt)
- model.set_tensor_datatype("thres1_param", DataType.INT32)
- model.set_tensor_datatype("thres2_param", DataType.INT32)
+ model.set_tensor_datatype("thres1_param", DataType["INT32"])
+ model.set_tensor_datatype("thres2_param", DataType["INT32"])
model.set_initializer(
"conv_param", gen_finn_dt_tensor(conv_weight_dt, conv_param_shape)
diff --git a/tests/fpgadataflow/test_convert_to_hls_conv_layer.py b/tests/fpgadataflow/test_convert_to_hls_conv_layer.py
index deca7c96127fdf03d9feb7504d5a6daebb41a5d5..d96bc987567cdcfcd18a404986c954c7527c7354 100644
--- a/tests/fpgadataflow/test_convert_to_hls_conv_layer.py
+++ b/tests/fpgadataflow/test_convert_to_hls_conv_layer.py
@@ -63,7 +63,7 @@ from finn.util.basic import gen_finn_dt_tensor
def test_convert_to_hls_conv_layer(conv_config, depthwise, exec_mode):
kernel_size, stride, pad = conv_config
np.random.seed(0)
- idt = DataType.UINT4
+ idt = DataType["UINT4"]
in_feature_dim = 7
in_chn = 16
@@ -84,7 +84,7 @@ def test_convert_to_hls_conv_layer(conv_config, depthwise, exec_mode):
input_shape = [1, in_chn, in_feature_dim, in_feature_dim]
output_shape = [1, out_chn, out_feature_dim, out_feature_dim]
- conv_weight_dt = DataType.UINT4
+ conv_weight_dt = DataType["UINT4"]
conv_config = {}
conv_config["dilations"] = [1, 1]
diff --git a/tests/fpgadataflow/test_convert_to_hls_layers_synthetic.py b/tests/fpgadataflow/test_convert_to_hls_layers_synthetic.py
index b0780c073114351ba136fefe6973114bd1a8505b..6089901566cb412e63cd8acc7a8260081248ba52 100644
--- a/tests/fpgadataflow/test_convert_to_hls_layers_synthetic.py
+++ b/tests/fpgadataflow/test_convert_to_hls_layers_synthetic.py
@@ -138,7 +138,7 @@ def make_model(ch, ifmdim):
# data types
-@pytest.mark.parametrize("idt", [DataType.UINT2])
+@pytest.mark.parametrize("idt", [DataType["UINT2"]])
# channels
@pytest.mark.parametrize("ch", [16])
# ifmdim
diff --git a/tests/fpgadataflow/test_convert_to_hls_pool_batch.py b/tests/fpgadataflow/test_convert_to_hls_pool_batch.py
index 70716e88a4de827be37416b63a925b30d01c342a..3efafc040df07a7d56638bf5ce0b1ce01887343c 100644
--- a/tests/fpgadataflow/test_convert_to_hls_pool_batch.py
+++ b/tests/fpgadataflow/test_convert_to_hls_pool_batch.py
@@ -118,9 +118,9 @@ def prepare_inputs(input_tensor):
# input datatype
-@pytest.mark.parametrize("idt", [DataType.UINT4, DataType.INT4, DataType.INT8])
+@pytest.mark.parametrize("idt", [DataType["UINT4"], DataType["INT4"], DataType["INT8"]])
# output datatype
-@pytest.mark.parametrize("odt", [DataType.UINT4, DataType.INT4])
+@pytest.mark.parametrize("odt", [DataType["UINT4"], DataType["INT4"]])
# pool configuration: ( k,stride, pad, ifm_dim )
@pytest.mark.parametrize("pool_config", [(7, 7, 0, 7), (3, 2, 1, 5)])
# input channels
diff --git a/tests/fpgadataflow/test_depthwise_convolution.py b/tests/fpgadataflow/test_depthwise_convolution.py
index 75ce055c0e9a093a5ddeab6b13af8d36d6152fb8..633db668d3bc5de815a313743c06cd74a7166c9c 100644
--- a/tests/fpgadataflow/test_depthwise_convolution.py
+++ b/tests/fpgadataflow/test_depthwise_convolution.py
@@ -60,14 +60,14 @@ def set_up_reference_model(act, idt, wdt, k, ifm_dim, ifm_ch, stride, padding):
ofm_dim = compute_conv_output_dim(ifm_dim, k, stride, total_pad=total_pad)
if act is None:
- odt = DataType.INT32
+ odt = DataType["INT32"]
else:
odt = act
out_act = oh.make_tensor_value_info(
"out_act", TensorProto.FLOAT, [1, ofm_dim, ofm_dim, ofm_ch]
)
T = oh.make_tensor_value_info("T", TensorProto.FLOAT, [ofm_ch, 15])
- tdt = DataType.INT32
+ tdt = DataType["INT32"]
thresh_node = oh.make_node(
"MultiThreshold",
domain="finn.custom_op.general",
@@ -161,7 +161,7 @@ def set_up_reference_model(act, idt, wdt, k, ifm_dim, ifm_ch, stride, padding):
# PE
@pytest.mark.parametrize("pe", [1, 2, 4])
# Output activation
-@pytest.mark.parametrize("act", [None, DataType.UINT4])
+@pytest.mark.parametrize("act", [None, DataType["UINT4"]])
# kernel size
@pytest.mark.parametrize("k", [2, 4])
# stride
@@ -171,7 +171,7 @@ def set_up_reference_model(act, idt, wdt, k, ifm_dim, ifm_ch, stride, padding):
@pytest.mark.slow
@pytest.mark.vivado
def test_depthwise_conv_hls_cppsim(act, pe, k, stride, padding):
- idt = wdt = DataType.INT4
+ idt = wdt = DataType["INT4"]
ifm_dim = 6
ifm_ch = 4
@@ -203,7 +203,7 @@ def test_depthwise_conv_hls_cppsim(act, pe, k, stride, padding):
# PE
@pytest.mark.parametrize("pe", [1, 2, 4])
# Output activation
-@pytest.mark.parametrize("act", [None, DataType.UINT4])
+@pytest.mark.parametrize("act", [None, DataType["UINT4"]])
# kernel size
@pytest.mark.parametrize("k", [2, 4])
# stride
@@ -213,7 +213,7 @@ def test_depthwise_conv_hls_cppsim(act, pe, k, stride, padding):
@pytest.mark.slow
@pytest.mark.vivado
def test_depthwise_conv_hls_rtlsim(act, pe, k, stride, padding):
- idt = wdt = DataType.INT4
+ idt = wdt = DataType["INT4"]
ifm_dim = 6
ifm_ch = 4
diff --git a/tests/fpgadataflow/test_fpgadataflow_addstreams.py b/tests/fpgadataflow/test_fpgadataflow_addstreams.py
index 021d58b4a382f2fe3d1a2c3c2a4ce8d7f3c87ae5..8cbf54ec188b12c67e02a33e3540718e9b08f382 100644
--- a/tests/fpgadataflow/test_fpgadataflow_addstreams.py
+++ b/tests/fpgadataflow/test_fpgadataflow_addstreams.py
@@ -82,7 +82,7 @@ def prepare_inputs(input1, input2):
# data types
-@pytest.mark.parametrize("idt", [DataType.UINT4, DataType.UINT8])
+@pytest.mark.parametrize("idt", [DataType["UINT4"], DataType["UINT8"]])
# channels
@pytest.mark.parametrize("ch", [1, 64])
# folding
diff --git a/tests/fpgadataflow/test_fpgadataflow_channelwise_ops.py b/tests/fpgadataflow/test_fpgadataflow_channelwise_ops.py
index 15bcd5fa8a937aa313f2c73f253f934f6bbd332b..949046d4ae313b852471e7d8a93e44fea48f7b0f 100644
--- a/tests/fpgadataflow/test_fpgadataflow_channelwise_ops.py
+++ b/tests/fpgadataflow/test_fpgadataflow_channelwise_ops.py
@@ -85,11 +85,11 @@ def make_modelwrapper(C, pe, idt, odt, pdt, func, vecs):
# activation: None or DataType
-@pytest.mark.parametrize("act", [DataType.INT8])
+@pytest.mark.parametrize("act", [DataType["INT8"]])
# input datatype
-@pytest.mark.parametrize("idt", [DataType.INT4])
+@pytest.mark.parametrize("idt", [DataType["INT4"]])
# param datatype
-@pytest.mark.parametrize("pdt", [DataType.INT4])
+@pytest.mark.parametrize("pdt", [DataType["INT4"]])
# folding, -1 is maximum possible
@pytest.mark.parametrize("nf", [-1, 2])
# number of input features
diff --git a/tests/fpgadataflow/test_fpgadataflow_convinputgenerator.py b/tests/fpgadataflow/test_fpgadataflow_convinputgenerator.py
index 86622cf6d44dbda3af417283f5ceea1d1ebc3bf0..47cd7e7ba1df76cc793cd0946581239a6883874e 100644
--- a/tests/fpgadataflow/test_fpgadataflow_convinputgenerator.py
+++ b/tests/fpgadataflow/test_fpgadataflow_convinputgenerator.py
@@ -131,7 +131,7 @@ def prepare_inputs(input_tensor):
# input datatype
-@pytest.mark.parametrize("idt", [DataType.BIPOLAR, DataType.INT2])
+@pytest.mark.parametrize("idt", [DataType["BIPOLAR"], DataType["INT2"]])
# kernel size
@pytest.mark.parametrize("k", [2, 3])
# input dimension
diff --git a/tests/fpgadataflow/test_fpgadataflow_convinputgenerator1d.py b/tests/fpgadataflow/test_fpgadataflow_convinputgenerator1d.py
index b3d695469b7a4fa1f4235feee29e7fc3dece0df5..8440ac1fe46a0d1ea4db3d76489dfc4ce68ff642 100644
--- a/tests/fpgadataflow/test_fpgadataflow_convinputgenerator1d.py
+++ b/tests/fpgadataflow/test_fpgadataflow_convinputgenerator1d.py
@@ -144,8 +144,8 @@ def prepare_inputs(input_tensor):
# input datatype
-# @pytest.mark.parametrize("idt", [DataType.BIPOLAR, DataType.INT8])
-@pytest.mark.parametrize("idt", [DataType.INT8])
+# @pytest.mark.parametrize("idt", [DataType["BIPOLAR"], DataType["INT8"]])
+@pytest.mark.parametrize("idt", [DataType["INT8"]])
# kernel size
@pytest.mark.parametrize("k", [[4, 1]])
# input dimension
diff --git a/tests/fpgadataflow/test_fpgadataflow_duplicatestreams.py b/tests/fpgadataflow/test_fpgadataflow_duplicatestreams.py
index 6b776e8827d8e76102bd069ae8567051ed0580ba..73bf1165afa9418be0c89f77797de538275fd220 100644
--- a/tests/fpgadataflow/test_fpgadataflow_duplicatestreams.py
+++ b/tests/fpgadataflow/test_fpgadataflow_duplicatestreams.py
@@ -85,7 +85,7 @@ def prepare_inputs(input_tensor, idt):
# data type
-@pytest.mark.parametrize("idt", [DataType.INT4, DataType.UINT16])
+@pytest.mark.parametrize("idt", [DataType["INT4"], DataType["UINT16"]])
# channels
@pytest.mark.parametrize("ch", [64])
# folding
diff --git a/tests/fpgadataflow/test_fpgadataflow_dwc.py b/tests/fpgadataflow/test_fpgadataflow_dwc.py
index b0af4382383d8935c69e362b1a43db536979c784..248b591eb48d7cfd6f121738a9bca525c38a45f8 100644
--- a/tests/fpgadataflow/test_fpgadataflow_dwc.py
+++ b/tests/fpgadataflow/test_fpgadataflow_dwc.py
@@ -82,7 +82,7 @@ def prepare_inputs(input_tensor, dt):
# outWidth
@pytest.mark.parametrize("OUTWidth", [2, 4])
# finn_dtype
-@pytest.mark.parametrize("finn_dtype", [DataType.BIPOLAR, DataType.INT2])
+@pytest.mark.parametrize("finn_dtype", [DataType["BIPOLAR"], DataType["INT2"]])
@pytest.mark.slow
@pytest.mark.vivado
def test_fpgadataflow_dwc_rtlsim(Shape, INWidth, OUTWidth, finn_dtype):
diff --git a/tests/fpgadataflow/test_fpgadataflow_fclayer.py b/tests/fpgadataflow/test_fpgadataflow_fclayer.py
index 49c326d2a34e7262826505ae32f2509b42ae0a35..02c3a3dc9506152fe999873df0612e76a5c9cefd 100644
--- a/tests/fpgadataflow/test_fpgadataflow_fclayer.py
+++ b/tests/fpgadataflow/test_fpgadataflow_fclayer.py
@@ -59,11 +59,11 @@ def make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T=None, tdt=Non
# StreamingFC:
# - specify their datatypes as such
# - specify their datatypes as BINARY as use binaryXnorMode
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
# we'll internally convert weights/inputs to binary and specify the
# datatypes as such, and also set the binaryXnorMode attribute to 1
- export_wdt = DataType.BINARY
- export_idt = DataType.BINARY
+ export_wdt = DataType["BINARY"]
+ export_idt = DataType["BINARY"]
binary_xnor_mode = 1
else:
export_wdt = wdt
@@ -75,7 +75,7 @@ def make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T=None, tdt=Non
if T is not None:
no_act = 0
node_inp_list = ["inp", "weights", "thresh"]
- if odt == DataType.BIPOLAR:
+ if odt == DataType["BIPOLAR"]:
actval = 0
else:
actval = odt.min()
@@ -123,7 +123,7 @@ def make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T=None, tdt=Non
def prepare_inputs(input_tensor, idt, wdt):
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
# convert bipolar to binary
return {"inp": (input_tensor + 1) / 2}
else:
@@ -133,11 +133,11 @@ def prepare_inputs(input_tensor, idt, wdt):
# mem_mode: const or decoupled
@pytest.mark.parametrize("mem_mode", ["const", "decoupled", "external"])
# activation: None or DataType
-@pytest.mark.parametrize("act", [None, DataType.BIPOLAR, DataType.INT4])
+@pytest.mark.parametrize("act", [None, DataType["BIPOLAR"], DataType["INT4"]])
# weight datatype
-@pytest.mark.parametrize("wdt", [DataType.BIPOLAR, DataType.INT4])
+@pytest.mark.parametrize("wdt", [DataType["BIPOLAR"], DataType["INT4"]])
# input datatype
-@pytest.mark.parametrize("idt", [DataType.BIPOLAR, DataType.INT4])
+@pytest.mark.parametrize("idt", [DataType["BIPOLAR"], DataType["INT4"]])
# neuron folding, -1 is maximum possible
@pytest.mark.parametrize("nf", [-1, 2, 1])
# synapse folding, -1 is maximum possible
@@ -165,10 +165,10 @@ def test_fpgadataflow_fclayer_cppsim(mem_mode, idt, wdt, act, nf, sf, mw, mh):
# no activation, produce accumulators
T = None
tdt = None
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
- odt = DataType.UINT32
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
+ odt = DataType["UINT32"]
else:
- odt = DataType.INT32
+ odt = DataType["INT32"]
else:
odt = act
(min, max) = calculate_signed_dot_prod_range(idt, wdt, mw)
@@ -177,13 +177,13 @@ def test_fpgadataflow_fclayer_cppsim(mem_mode, idt, wdt, act, nf, sf, mw, mh):
# provide non-decreasing thresholds
T = np.sort(T, axis=1)
# generate thresholds for activation
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
- tdt = DataType.UINT32
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
+ tdt = DataType["UINT32"]
# bias thresholds to be positive
T = np.ceil((T + mw) / 2)
assert (T >= 0).all()
else:
- tdt = DataType.INT32
+ tdt = DataType["INT32"]
model = make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T, tdt)
for node in model.graph.node:
# lookup op_type in registry of CustomOps
@@ -194,14 +194,14 @@ def test_fpgadataflow_fclayer_cppsim(mem_mode, idt, wdt, act, nf, sf, mw, mh):
model = model.transform(CompileCppSim())
# prepare input data
input_dict = prepare_inputs(x, idt, wdt)
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
# convert inputs to binary and use xnorpopcountmatmul
y = xp.xnorpopcountmatmul((x + 1) / 2, (W + 1) / 2)
else:
y = np.matmul(x, W)
if T is not None:
y = multithreshold(y, T)
- if act == DataType.BIPOLAR:
+ if act == DataType["BIPOLAR"]:
# binary to bipolar
y = 2 * y - 1
else:
@@ -220,11 +220,11 @@ def test_fpgadataflow_fclayer_cppsim(mem_mode, idt, wdt, act, nf, sf, mw, mh):
# mem_mode: const or decoupled
@pytest.mark.parametrize("mem_mode", ["const", "decoupled", "external"])
# activation: None or DataType
-@pytest.mark.parametrize("act", [None, DataType.BIPOLAR, DataType.INT4])
+@pytest.mark.parametrize("act", [None, DataType["BIPOLAR"], DataType["INT4"]])
# weight datatype
-@pytest.mark.parametrize("wdt", [DataType.BIPOLAR, DataType.INT4])
+@pytest.mark.parametrize("wdt", [DataType["BIPOLAR"], DataType["INT4"]])
# input datatype
-@pytest.mark.parametrize("idt", [DataType.BIPOLAR, DataType.INT4])
+@pytest.mark.parametrize("idt", [DataType["BIPOLAR"], DataType["INT4"]])
# neuron folding, -1 is maximum possible
@pytest.mark.parametrize("nf", [-1, 2, 1])
# synapse folding, -1 is maximum possible
@@ -252,10 +252,10 @@ def test_fpgadataflow_fclayer_rtlsim(mem_mode, idt, wdt, act, nf, sf, mw, mh):
# no activation, produce accumulators
T = None
tdt = None
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
- odt = DataType.UINT32
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
+ odt = DataType["UINT32"]
else:
- odt = DataType.INT32
+ odt = DataType["INT32"]
else:
odt = act
(min, max) = calculate_signed_dot_prod_range(idt, wdt, mw)
@@ -264,13 +264,13 @@ def test_fpgadataflow_fclayer_rtlsim(mem_mode, idt, wdt, act, nf, sf, mw, mh):
# provide non-decreasing thresholds
T = np.sort(T, axis=1)
# generate thresholds for activation
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
- tdt = DataType.UINT32
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
+ tdt = DataType["UINT32"]
# bias thresholds to be positive
T = np.ceil((T + mw) / 2)
assert (T >= 0).all()
else:
- tdt = DataType.INT32
+ tdt = DataType["INT32"]
model = make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T, tdt)
for node in model.graph.node:
# lookup op_type in registry of CustomOps
@@ -279,14 +279,14 @@ def test_fpgadataflow_fclayer_rtlsim(mem_mode, idt, wdt, act, nf, sf, mw, mh):
# prepare input data
input_dict = prepare_inputs(x, idt, wdt)
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
# convert inputs to binary and use xnorpopcountmatmul
y = xp.xnorpopcountmatmul((x + 1) / 2, (W + 1) / 2)
else:
y = np.matmul(x, W)
if T is not None:
y = multithreshold(y, T)
- if act == DataType.BIPOLAR:
+ if act == DataType["BIPOLAR"]:
# binary to bipolar
y = 2 * y - 1
else:
@@ -319,11 +319,11 @@ def test_fpgadataflow_fclayer_rtlsim(mem_mode, idt, wdt, act, nf, sf, mw, mh):
# mem_mode: const or decoupled
@pytest.mark.parametrize("mem_mode", ["decoupled"])
# activation: None or DataType
-@pytest.mark.parametrize("act", [DataType.INT4])
+@pytest.mark.parametrize("act", [DataType["INT4"]])
# weight datatype
-@pytest.mark.parametrize("wdt", [DataType.INT4])
+@pytest.mark.parametrize("wdt", [DataType["INT4"]])
# input datatype
-@pytest.mark.parametrize("idt", [DataType.INT4])
+@pytest.mark.parametrize("idt", [DataType["INT4"]])
# neuron folding, -1 is maximum possible
@pytest.mark.parametrize("nf", [-1])
# synapse folding, -1 is maximum possible
@@ -352,10 +352,10 @@ def test_fpgadataflow_fclayer_large_depth_decoupled_mode_rtlsim(
# no activation, produce accumulators
T = None
tdt = None
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
- odt = DataType.UINT32
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
+ odt = DataType["UINT32"]
else:
- odt = DataType.INT32
+ odt = DataType["INT32"]
else:
odt = act
(min, max) = calculate_signed_dot_prod_range(idt, wdt, mw)
@@ -364,13 +364,13 @@ def test_fpgadataflow_fclayer_large_depth_decoupled_mode_rtlsim(
# provide non-decreasing thresholds
T = np.sort(T, axis=1)
# generate thresholds for activation
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
- tdt = DataType.UINT32
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
+ tdt = DataType["UINT32"]
# bias thresholds to be positive
T = np.ceil((T + mw) / 2)
assert (T >= 0).all()
else:
- tdt = DataType.INT32
+ tdt = DataType["INT32"]
model = make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T, tdt)
for node in model.graph.node:
# lookup op_type in registry of CustomOps
@@ -379,14 +379,14 @@ def test_fpgadataflow_fclayer_large_depth_decoupled_mode_rtlsim(
# prepare input data
input_dict = prepare_inputs(x, idt, wdt)
- if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
+ if wdt == DataType["BIPOLAR"] and idt == DataType["BIPOLAR"]:
# convert inputs to binary and use xnorpopcountmatmul
y = xp.xnorpopcountmatmul((x + 1) / 2, (W + 1) / 2)
else:
y = np.matmul(x, W)
if T is not None:
y = multithreshold(y, T)
- if act == DataType.BIPOLAR:
+ if act == DataType["BIPOLAR"]:
# binary to bipolar
y = 2 * y - 1
else:
diff --git a/tests/fpgadataflow/test_fpgadataflow_fifo.py b/tests/fpgadataflow/test_fpgadataflow_fifo.py
index 81f66c42ca76d42fe8ee50576d72007f6ca6c12f..4d3074fe14617df4386f060b6a476734931fb4ca 100644
--- a/tests/fpgadataflow/test_fpgadataflow_fifo.py
+++ b/tests/fpgadataflow/test_fpgadataflow_fifo.py
@@ -86,7 +86,7 @@ def prepare_inputs(input_tensor, dt):
# outWidth
@pytest.mark.parametrize("depth", [16])
# finn_dtype
-@pytest.mark.parametrize("finn_dtype", [DataType.BIPOLAR]) # , DataType.INT2])
+@pytest.mark.parametrize("finn_dtype", [DataType["BIPOLAR"]]) # , DataType["INT2"]])
@pytest.mark.slow
@pytest.mark.vivado
def test_fpgadataflow_fifo_rtlsim(Shape, folded_shape, depth, finn_dtype):
diff --git a/tests/fpgadataflow/test_fpgadataflow_fmpadding.py b/tests/fpgadataflow/test_fpgadataflow_fmpadding.py
index 5db12ee22828e43e276ed85f04f985653fe0a2dd..b564273c0927938859dc438dce619e7067a7ad74 100644
--- a/tests/fpgadataflow/test_fpgadataflow_fmpadding.py
+++ b/tests/fpgadataflow/test_fpgadataflow_fmpadding.py
@@ -108,7 +108,7 @@ def make_single_fmpadding_modelwrapper(idim, padding, num_ch, simd, idt, pad_sty
# PaddingStyle: selects behavior when (odim-idim)%2 != 0
@pytest.mark.parametrize("pad_style", [2])
# FINN input datatype
-@pytest.mark.parametrize("idt", [DataType.INT2, DataType.INT4])
+@pytest.mark.parametrize("idt", [DataType["INT2"], DataType["INT4"]])
# execution mode
@pytest.mark.parametrize("mode", ["cppsim", "rtlsim"])
@pytest.mark.slow
diff --git a/tests/fpgadataflow/test_fpgadataflow_globalaccpool.py b/tests/fpgadataflow/test_fpgadataflow_globalaccpool.py
index f1373123a69f4c3d02b191c0f0560b59d2c9a7b2..2299cc6e8f397df718d2fd65be8a562c2457e42d 100644
--- a/tests/fpgadataflow/test_fpgadataflow_globalaccpool.py
+++ b/tests/fpgadataflow/test_fpgadataflow_globalaccpool.py
@@ -78,7 +78,7 @@ def prepare_inputs(input_tensor, idt):
# data type
-@pytest.mark.parametrize("idt", [DataType.UINT4, DataType.UINT16])
+@pytest.mark.parametrize("idt", [DataType["UINT4"], DataType["UINT16"]])
# channels
@pytest.mark.parametrize("ch", [64])
# folding
@@ -127,7 +127,7 @@ def test_fpgadataflow_globalaccpool(idt, ch, fold, imdim, exec_mode):
exp_cycles_dict = model.analysis(exp_cycles_per_layer)
exp_cycles = exp_cycles_dict[node.name]
# commented out, needs performance debug:
- # test_fpgadataflow_globalaccpool[rtlsim-7-1-64-DataType.UINT4]
+ # test_fpgadataflow_globalaccpool[rtlsim-7-1-64-DataType["UINT4"]]
# assert False where False =
# <function isclose at 0x7eff26d5ca60>(50, 103, atol=(0.1 * 103))
# assert np.isclose(exp_cycles, cycles_rtlsim, atol=0.1 * cycles_rtlsim)
diff --git a/tests/fpgadataflow/test_fpgadataflow_ipstitch.py b/tests/fpgadataflow/test_fpgadataflow_ipstitch.py
index 9a6050a55dd86ca5064b293f87304cbb1365edea..933da667ecb54d471d9e6d48ac4462421addad7e 100644
--- a/tests/fpgadataflow/test_fpgadataflow_ipstitch.py
+++ b/tests/fpgadataflow/test_fpgadataflow_ipstitch.py
@@ -69,9 +69,9 @@ ip_stitch_model_dir = os.environ["FINN_BUILD_DIR"]
def create_one_fc_model(mem_mode="const"):
# create a model with a StreamingFCLayer instance with no activation
# the wider range of the full accumulator makes debugging a bit easier
- wdt = DataType.INT2
- idt = DataType.INT32
- odt = DataType.INT32
+ wdt = DataType["INT2"]
+ idt = DataType["INT32"]
+ odt = DataType["INT32"]
m = 4
no_act = 1
binary_xnor_mode = 0
@@ -122,9 +122,9 @@ def create_one_fc_model(mem_mode="const"):
def create_two_fc_model(mem_mode="decoupled"):
# create a model with two StreamingFCLayer instances
- wdt = DataType.INT2
- idt = DataType.INT32
- odt = DataType.INT32
+ wdt = DataType["INT2"]
+ idt = DataType["INT32"]
+ odt = DataType["INT32"]
m = 4
actval = 0
no_act = 1
diff --git a/tests/fpgadataflow/test_fpgadataflow_labelselect.py b/tests/fpgadataflow/test_fpgadataflow_labelselect.py
index 8997208a648fa79439a882de23865496ba527858..8ed06c8bdf1c0dbfab2f8141bf724132f4a24705 100644
--- a/tests/fpgadataflow/test_fpgadataflow_labelselect.py
+++ b/tests/fpgadataflow/test_fpgadataflow_labelselect.py
@@ -81,7 +81,9 @@ def prepare_inputs(input_tensor, idt):
return {"inp": input_tensor}
-@pytest.mark.parametrize("idt", [DataType.UINT8, DataType.UINT16, DataType.INT16])
+@pytest.mark.parametrize(
+ "idt", [DataType["UINT8"], DataType["UINT16"], DataType["INT16"]]
+)
# labels
@pytest.mark.parametrize("labels", [10, 100])
# folding
diff --git a/tests/fpgadataflow/test_fpgadataflow_res_estimate.py b/tests/fpgadataflow/test_fpgadataflow_res_estimate.py
index 9def746c1c872a8b99b5bab48e8d0bd20798cedd..fe52a73fc07df8551442e975c5eb378c132a56d7 100644
--- a/tests/fpgadataflow/test_fpgadataflow_res_estimate.py
+++ b/tests/fpgadataflow/test_fpgadataflow_res_estimate.py
@@ -54,9 +54,9 @@ def test_res_estimate():
mw = mh = 4
simd = 1
pe = 1
- idt = DataType.INT2
- wdt = DataType.INT2
- odt = DataType.INT2
+ idt = DataType["INT2"]
+ wdt = DataType["INT2"]
+ odt = DataType["INT2"]
actval = odt.min()
inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, mw])
diff --git a/tests/fpgadataflow/test_fpgadataflow_thresholding.py b/tests/fpgadataflow/test_fpgadataflow_thresholding.py
index 012bc3e2e140c3fa63729584629613e3046f8838..341bd3f37041c9b5a1526e99b2c4bad4d3dd3029 100644
--- a/tests/fpgadataflow/test_fpgadataflow_thresholding.py
+++ b/tests/fpgadataflow/test_fpgadataflow_thresholding.py
@@ -98,9 +98,9 @@ def make_single_thresholding_modelwrapper(T, pe, idt, odt, actval, mem_mode):
# activation: None or DataType
-@pytest.mark.parametrize("act", [DataType.INT4, DataType.BIPOLAR])
+@pytest.mark.parametrize("act", [DataType["INT4"], DataType["BIPOLAR"]])
# input datatype
-@pytest.mark.parametrize("idt", [DataType.INT16, DataType.UINT16])
+@pytest.mark.parametrize("idt", [DataType["INT16"], DataType["UINT16"]])
# folding, -1 is maximum possible
@pytest.mark.parametrize("nf", [-1, 2, 1])
# number of input features
@@ -125,12 +125,12 @@ def test_fpgadataflow_thresholding(idt, act, nf, ich, exec_mode, mem_mode):
T = np.random.randint(idt.min(), idt.max() + 1, (ich, n_steps)).astype(np.float32)
# make the vivado_hls threshold bug appear (incorrect rtlsim result when first
# threshold of first channel is zero, while using BIPOLAR output)
- if act == DataType.BIPOLAR:
+ if act == DataType["BIPOLAR"]:
T[0][0] = 0
# provide non-decreasing thresholds
T = np.sort(T, axis=1)
- if odt == DataType.BIPOLAR:
+ if odt == DataType["BIPOLAR"]:
actval = 0
else:
actval = odt.min()
@@ -154,7 +154,7 @@ def test_fpgadataflow_thresholding(idt, act, nf, ich, exec_mode, mem_mode):
input_dict = {"inp": x}
y = multithreshold(x, T)
- if act == DataType.BIPOLAR:
+ if act == DataType["BIPOLAR"]:
# binary to bipolar
y = 2 * y - 1
else:
@@ -186,8 +186,8 @@ def test_fpgadataflow_thresholding(idt, act, nf, ich, exec_mode, mem_mode):
@pytest.mark.vivado
def test_runtime_thresholds_single_layer():
mem_mode = "decoupled"
- act = DataType.INT4
- idt = DataType.INT16
+ act = DataType["INT4"]
+ idt = DataType["INT16"]
nf = 8
ich = 16
pe = ich // nf
@@ -202,7 +202,7 @@ def test_runtime_thresholds_single_layer():
# provide non-decreasing thresholds
T = np.sort(T, axis=1)
- if odt == DataType.BIPOLAR:
+ if odt == DataType["BIPOLAR"]:
actval = 0
else:
actval = odt.min()
@@ -245,7 +245,7 @@ def test_runtime_thresholds_single_layer():
# old weights (see above)
y = exec_ctx["outp"][1]
expected = multithreshold(in_tensor, T)[1]
- if act == DataType.BIPOLAR:
+ if act == DataType["BIPOLAR"]:
# binary to bipolar
expected = 2 * expected - 1
else:
@@ -274,7 +274,7 @@ def test_runtime_thresholds_single_layer():
rtlsim_exec(model, exec_ctx, pre_hook=write_weights)
y = exec_ctx["outp"][1]
expected = multithreshold(in_tensor, new_weights)[1]
- if act == DataType.BIPOLAR:
+ if act == DataType["BIPOLAR"]:
# binary to bipolar
expected = 2 * expected - 1
else:
diff --git a/tests/fpgadataflow/test_fpgadataflow_upsampler.py b/tests/fpgadataflow/test_fpgadataflow_upsampler.py
index f7e06adb816cfa664187f39a9567d4f742e4043b..1709cfe32904a5ed369f8399150a8a1d05f4b781 100644
--- a/tests/fpgadataflow/test_fpgadataflow_upsampler.py
+++ b/tests/fpgadataflow/test_fpgadataflow_upsampler.py
@@ -60,7 +60,7 @@ class ForceDataTypeForTensors(Transformation):
Forces a certain datatype for all tensors in a model.
"""
- def __init__(self, dType=DataType.INT8):
+ def __init__(self, dType=DataType["INT8"]):
super().__init__()
self._dType = dType
@@ -116,7 +116,7 @@ class PyTorchTestModel(nn.Module):
# param datatype
-@pytest.mark.parametrize("dt", [DataType.INT8])
+@pytest.mark.parametrize("dt", [DataType["INT8"]])
# Width/height of square input feature map
@pytest.mark.parametrize("IFMDim", [3, 5])
# upscaling factor
diff --git a/tests/fpgadataflow/test_fpgadataflow_vvau.py b/tests/fpgadataflow/test_fpgadataflow_vvau.py
index 36b844deab4e28ff35290a170f713a64be839e8a..6f39994bf27594a063a1e66c5bba7867eaabef6e 100644
--- a/tests/fpgadataflow/test_fpgadataflow_vvau.py
+++ b/tests/fpgadataflow/test_fpgadataflow_vvau.py
@@ -141,11 +141,11 @@ def prepare_inputs(input_tensor):
# mem_mode: const or decoupled
-@pytest.mark.parametrize("idt", [DataType.UINT4, DataType.UINT8])
+@pytest.mark.parametrize("idt", [DataType["UINT4"], DataType["UINT8"]])
# weight datatype
-@pytest.mark.parametrize("wdt", [DataType.INT4])
+@pytest.mark.parametrize("wdt", [DataType["INT4"]])
# activation: None or DataType
-@pytest.mark.parametrize("act", [DataType.UINT4, None])
+@pytest.mark.parametrize("act", [DataType["UINT4"], None])
# PE
@pytest.mark.parametrize("pe", [1, "channels"])
# Input image shape
@@ -187,14 +187,14 @@ def test_fpgadataflow_vvau(
if act is None:
T = None
tdt = None
- odt = DataType.INT32
+ odt = DataType["INT32"]
else:
odt = act
(min_v, max_v) = _calculate_dot_prod_range(idt, wdt, k_h * k_w * channels)
n_steps = act.get_num_possible_values() - 1
T = np.random.randint(min_v, max_v - 1, (channels, n_steps)).astype(np.float32)
T = np.sort(T, axis=1)
- tdt = DataType.INT32
+ tdt = DataType["INT32"]
model = _make_single_vvau_modelwrapper(
W, pe, k_h, k_w, channels, dim_h, dim_w, wdt, idt, odt, T, tdt
diff --git a/tests/fpgadataflow/test_layer_streaming_maxpool_batch.py b/tests/fpgadataflow/test_layer_streaming_maxpool_batch.py
index 556e15f13607caa556daff079026f0b2bacb1b2b..236eb2a0342a2782f106761f4cd356888a2f8630 100644
--- a/tests/fpgadataflow/test_layer_streaming_maxpool_batch.py
+++ b/tests/fpgadataflow/test_layer_streaming_maxpool_batch.py
@@ -121,7 +121,7 @@ def prepare_inputs(input_tensor):
# input datatype
-@pytest.mark.parametrize("idt", [DataType.BIPOLAR, DataType.INT4])
+@pytest.mark.parametrize("idt", [DataType["BIPOLAR"], DataType["INT4"]])
# 1d maxpool
@pytest.mark.parametrize("dim_1d", [False, True])
# kernel size
@@ -151,7 +151,7 @@ def test_fpgadataflow_streamingmaxpool(idt, dim_1d, k, ifm_dim, ifm_ch, exec_mod
ofm_dim_h = int(((ifm_dim_h - k_h) / stride_h) + 1)
ofm_dim_w = int(((ifm_dim_w - k_w) / stride_w) + 1)
ofm_dim = (ofm_dim_h, ofm_dim_w)
- if idt == DataType.BIPOLAR and dim_1d:
+ if idt == DataType["BIPOLAR"] and dim_1d:
pytest.skip("Skipping binary StreamingMaxPool_1d (not implemented)")
if ifm_dim_h % k_h != 0 or ifm_dim_w % k_w != 0:
pytest.skip("Skipping StreamingMaxPool test w/ ImgDim % PoolDim != 0")
diff --git a/tests/fpgadataflow/test_runtime_weights.py b/tests/fpgadataflow/test_runtime_weights.py
index 706d11114f2f08df700efd40afb8dea218efbf42..0196a78d5c4254d7cb116641f946bcccb9e1ebc9 100644
--- a/tests/fpgadataflow/test_runtime_weights.py
+++ b/tests/fpgadataflow/test_runtime_weights.py
@@ -49,8 +49,8 @@ target_clk_ns = 5
@pytest.mark.vivado
def test_runtime_weights_single_layer():
- idt = DataType.UINT32
- wdt = DataType.UINT4
+ idt = DataType["UINT32"]
+ wdt = DataType["UINT4"]
act = None
mw = 64
mh = 32
diff --git a/tests/fpgadataflow/test_set_folding.py b/tests/fpgadataflow/test_set_folding.py
index 8f4d57d3f84dd5c5a167b0e35b775def8ed27c5d..66fd5b43a1b8b8c8986bf9c9b9d0e9efd7a744a6 100644
--- a/tests/fpgadataflow/test_set_folding.py
+++ b/tests/fpgadataflow/test_set_folding.py
@@ -115,7 +115,7 @@ def make_multi_fclayer_model(ch, wdt, adt, tdt, nnodes):
def test_set_folding(target_fps, platform):
model = make_multi_fclayer_model(
- 128, DataType.INT4, DataType.INT2, DataType.INT16, 5
+ 128, DataType["INT4"], DataType["INT2"], DataType["INT16"], 5
)
model = model.transform(GiveUniqueNodeNames())
diff --git a/tests/transformation/streamline/test_move_flatten_past_affine.py b/tests/transformation/streamline/test_move_flatten_past_affine.py
index 1971ecfaa181d6ee799a9191b63d2482629b1e1c..ef01436dc9435676b562e2b635a8cf12e901046b 100644
--- a/tests/transformation/streamline/test_move_flatten_past_affine.py
+++ b/tests/transformation/streamline/test_move_flatten_past_affine.py
@@ -77,14 +77,14 @@ def test_move_flatten_past_affine(data_layout, batch_size):
model = ModelWrapper(model)
# initialize values
- a0_values = gen_finn_dt_tensor(DataType.TERNARY, [1024, 1000])
+ a0_values = gen_finn_dt_tensor(DataType["TERNARY"], [1024, 1000])
model.set_initializer("a0", a0_values)
a1_values = np.random.uniform(low=0.1, high=0.99, size=(1)).astype(np.float32)
model.set_initializer("a1", a1_values)
a2_values = np.random.uniform(low=-1, high=1, size=(1000)).astype(np.float32)
model.set_initializer("a2", a2_values)
- model.set_tensor_datatype("inp", DataType.INT2)
+ model.set_tensor_datatype("inp", DataType["INT2"])
model.set_tensor_layout("inp", data_layout)
model = model.transform(InferShapes())
model = model.transform(InferDataTypes())
@@ -93,7 +93,7 @@ def test_move_flatten_past_affine(data_layout, batch_size):
model = model.transform(GiveReadableTensorNames())
# compare execution before and after transformation
- inp_values = gen_finn_dt_tensor(DataType.INT2, ishape)
+ inp_values = gen_finn_dt_tensor(DataType["INT2"], ishape)
idict = {model.graph.input[0].name: inp_values}
model_transformed = model.transform(MoveFlattenPastAffine())
assert oxe.compare_execution(model, model_transformed, idict)
diff --git a/tests/transformation/streamline/test_move_flatten_past_topk.py b/tests/transformation/streamline/test_move_flatten_past_topk.py
index 5e0211ad8857653ce75af2f5a7de0c6439770108..6086f7804eda4447de8f5948f521f0b003f65020 100644
--- a/tests/transformation/streamline/test_move_flatten_past_topk.py
+++ b/tests/transformation/streamline/test_move_flatten_past_topk.py
@@ -69,7 +69,7 @@ def test_move_flatten_past_affine(data_layout, batch_size):
model = helper.make_model(graph, producer_name="move_flatten_model")
model = ModelWrapper(model)
- model.set_tensor_datatype("inp", DataType.INT2)
+ model.set_tensor_datatype("inp", DataType["INT2"])
model.set_tensor_layout("inp", data_layout)
model = model.transform(InsertTopK())
model = model.transform(InferShapes())
@@ -79,7 +79,7 @@ def test_move_flatten_past_affine(data_layout, batch_size):
model = model.transform(GiveReadableTensorNames())
# compare execution before and after transformation
- inp_values = gen_finn_dt_tensor(DataType.INT2, ishape)
+ inp_values = gen_finn_dt_tensor(DataType["INT2"], ishape)
idict = {model.graph.input[0].name: inp_values}
model_transformed = model.transform(MoveFlattenPastTopK())
assert oxe.compare_execution(model, model_transformed, idict)
diff --git a/tests/transformation/streamline/test_move_mul_past_dw_conv.py b/tests/transformation/streamline/test_move_mul_past_dw_conv.py
index cb9beed713eb448b49015a7de601a4d15edc035b..e9e956d845ef8e56d2078bcd738ad3bb0ff72bfa 100644
--- a/tests/transformation/streamline/test_move_mul_past_dw_conv.py
+++ b/tests/transformation/streamline/test_move_mul_past_dw_conv.py
@@ -68,9 +68,9 @@ def test_move_mul_past_dw_conv(ifm_dim, ifm_ch, k, stride, pad_amt, dw):
model = helper.make_model(graph, producer_name="mulpastconv-model")
model = ModelWrapper(model)
- inp_values = gen_finn_dt_tensor(DataType.INT2, [1, ifm_ch, ifm_dim, ifm_dim])
- mul_values = gen_finn_dt_tensor(DataType.INT2, [1, ifm_ch, 1, 1])
- W_values = gen_finn_dt_tensor(DataType.INT2, W_shape)
+ inp_values = gen_finn_dt_tensor(DataType["INT2"], [1, ifm_ch, ifm_dim, ifm_dim])
+ mul_values = gen_finn_dt_tensor(DataType["INT2"], [1, ifm_ch, 1, 1])
+ W_values = gen_finn_dt_tensor(DataType["INT2"], W_shape)
model.set_initializer("W", W_values)
model.set_initializer("mul", mul_values)
model = model.transform(InferShapes())
diff --git a/tests/transformation/streamline/test_move_mul_past_maxpool.py b/tests/transformation/streamline/test_move_mul_past_maxpool.py
index 81f18842ed8ba2b5230f3a853076244d0a0ab8d9..2c51aaf36a79591fd0fd0cea368d5e23da0d07c3 100755
--- a/tests/transformation/streamline/test_move_mul_past_maxpool.py
+++ b/tests/transformation/streamline/test_move_mul_past_maxpool.py
@@ -66,7 +66,7 @@ def test_move_mul_past_maxpool(ifm_dim, ifm_ch, k, stride, pad, cw, negative):
model = helper.make_model(graph, producer_name="mulpastmaxpool-model")
model = ModelWrapper(model)
- inp_values = gen_finn_dt_tensor(DataType.INT2, [1, ifm_ch, ifm_dim, ifm_dim])
+ inp_values = gen_finn_dt_tensor(DataType["INT2"], [1, ifm_ch, ifm_dim, ifm_dim])
mul_values = np.random.random_sample(mul_shape).astype(np.float32)
if negative == 1:
mul_values = mul_values * (-1)
diff --git a/tests/transformation/streamline/test_remove_identity_ops.py b/tests/transformation/streamline/test_remove_identity_ops.py
index ad7c20fb51902f22c20896bdfb3321dc74d0572d..ee4e42fc8417017184594b6e754f3e8270a46ee1 100644
--- a/tests/transformation/streamline/test_remove_identity_ops.py
+++ b/tests/transformation/streamline/test_remove_identity_ops.py
@@ -69,11 +69,11 @@ def test_remove_identity_ops(op, as_first_node, approx):
model = helper.make_model(graph, producer_name="mulpastconv-model")
model = ModelWrapper(model)
- inp_values = gen_finn_dt_tensor(DataType.INT2, [1, 4, 1, 1])
+ inp_values = gen_finn_dt_tensor(DataType["INT2"], [1, 4, 1, 1])
mul_values = np.random.uniform(low=0.1, high=0.99, size=(1)).astype(np.float32)
shape_values = np.asarray([1, -1], dtype=np.int64)
div_values = np.random.uniform(low=0.1, high=0.99, size=(1)).astype(np.float32)
- matmul_values = gen_finn_dt_tensor(DataType.INT2, [4, 2])
+ matmul_values = gen_finn_dt_tensor(DataType["INT2"], [4, 2])
model.set_initializer("mul", mul_values)
model.set_initializer("shape", shape_values)
model.set_initializer("div", div_values)
diff --git a/tests/transformation/streamline/test_round_thresholds.py b/tests/transformation/streamline/test_round_thresholds.py
index f9259908a2b4e4d716e3fb9ae7ec28cd9ec85d03..2e57f1c85f6ac197ca7a4cf15e595c34cc0fb564 100644
--- a/tests/transformation/streamline/test_round_thresholds.py
+++ b/tests/transformation/streamline/test_round_thresholds.py
@@ -47,17 +47,17 @@ def test_round_thresholds():
model = ModelWrapper(model_def)
threshold_val = np.asarray([[-1.1], [0.7], [2.3], [5.1]], dtype=np.float32)
model.set_initializer("thresholds", threshold_val)
- model.set_tensor_datatype("v", DataType.INT8)
+ model.set_tensor_datatype("v", DataType["INT8"])
inp_dict_f = {"v": np.floor(threshold_val).T}
inp_dict_n = {"v": np.round(threshold_val).T}
inp_dict_c = {"v": np.ceil(threshold_val).T}
orig_f = oxe.execute_onnx(model, inp_dict_f)["out"]
orig_n = oxe.execute_onnx(model, inp_dict_n)["out"]
orig_c = oxe.execute_onnx(model, inp_dict_c)["out"]
- assert model.get_tensor_datatype("thresholds") == DataType.FLOAT32
+ assert model.get_tensor_datatype("thresholds") == DataType["FLOAT32"]
new_model = model.transform(RoundAndClipThresholds())
# rounded up thresholds should have same dtype as input
- assert new_model.get_tensor_datatype("thresholds") == DataType.INT8
+ assert new_model.get_tensor_datatype("thresholds") == DataType["INT8"]
new_f = oxe.execute_onnx(new_model, inp_dict_f)["out"]
new_n = oxe.execute_onnx(new_model, inp_dict_n)["out"]
new_c = oxe.execute_onnx(new_model, inp_dict_c)["out"]
diff --git a/tests/transformation/test_infer_datatypes_lfc.py b/tests/transformation/test_infer_datatypes_lfc.py
index 00715e3e3ca3626e1b76bf3b23bae4dc1d65b053..8883dac7a54eafaaa768c8ae991b2030e385b318 100644
--- a/tests/transformation/test_infer_datatypes_lfc.py
+++ b/tests/transformation/test_infer_datatypes_lfc.py
@@ -49,12 +49,12 @@ def test_infer_datatypes_lfc():
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
- assert model.get_tensor_datatype("MatMul_0_out0") == DataType.INT32
- assert model.get_tensor_datatype("MatMul_1_out0") == DataType.INT32
- assert model.get_tensor_datatype("MatMul_2_out0") == DataType.INT32
- assert model.get_tensor_datatype("MatMul_3_out0") == DataType.INT32
- assert model.get_tensor_datatype("MultiThreshold_0_out0") == DataType.BIPOLAR
- assert model.get_tensor_datatype("MultiThreshold_1_out0") == DataType.BIPOLAR
- assert model.get_tensor_datatype("MultiThreshold_2_out0") == DataType.BIPOLAR
- assert model.get_tensor_datatype("MultiThreshold_3_out0") == DataType.BIPOLAR
+ assert model.get_tensor_datatype("MatMul_0_out0") == DataType["INT32"]
+ assert model.get_tensor_datatype("MatMul_1_out0") == DataType["INT32"]
+ assert model.get_tensor_datatype("MatMul_2_out0") == DataType["INT32"]
+ assert model.get_tensor_datatype("MatMul_3_out0") == DataType["INT32"]
+ assert model.get_tensor_datatype("MultiThreshold_0_out0") == DataType["BIPOLAR"]
+ assert model.get_tensor_datatype("MultiThreshold_1_out0") == DataType["BIPOLAR"]
+ assert model.get_tensor_datatype("MultiThreshold_2_out0") == DataType["BIPOLAR"]
+ assert model.get_tensor_datatype("MultiThreshold_3_out0") == DataType["BIPOLAR"]
os.remove(export_onnx_path)
diff --git a/tests/util/test_create.py b/tests/util/test_create.py
index 42a288b74ecda9746296519b1b86563c75b2752e..c11e60175ea3ac94b6686ec5f8401a7c134fe53e 100644
--- a/tests/util/test_create.py
+++ b/tests/util/test_create.py
@@ -32,7 +32,9 @@ import finn.util.create as create
from finn.core.datatype import DataType
-@pytest.mark.parametrize("bitwidth", [DataType.BIPOLAR, DataType.INT2, DataType.INT4])
+@pytest.mark.parametrize(
+ "bitwidth", [DataType["BIPOLAR"], DataType["INT2"], DataType["INT4"]]
+)
def test_hls_random_mlp_maker(bitwidth):
w = bitwidth
a = bitwidth
@@ -42,7 +44,7 @@ def test_hls_random_mlp_maker(bitwidth):
"mh": 100,
"simd": 185,
"pe": 100,
- "idt": DataType.BIPOLAR,
+ "idt": DataType["BIPOLAR"],
"wdt": w,
"act": a,
},
@@ -56,7 +58,7 @@ def test_hls_random_mlp_maker(bitwidth):
"pe": 1,
"idt": a,
"wdt": w,
- "act": DataType.BIPOLAR,
+ "act": DataType["BIPOLAR"],
},
]
diff --git a/tests/util/test_data_packing_hls.py b/tests/util/test_data_packing_hls.py
index 3221eda34c85ed9d65b258b6489699cda8400517..9c47bb293e7640a30d6741fbc9ae4b9801f8bd7a 100644
--- a/tests/util/test_data_packing_hls.py
+++ b/tests/util/test_data_packing_hls.py
@@ -38,7 +38,9 @@ from finn.core.datatype import DataType
from finn.util.data_packing import numpy_to_hls_code
-@pytest.mark.parametrize("dtype", [DataType.BINARY, DataType.INT2, DataType.INT32])
+@pytest.mark.parametrize(
+ "dtype", [DataType["BINARY"], DataType["INT2"], DataType["INT32"]]
+)
@pytest.mark.parametrize("test_shape", [(1, 2, 4), (1, 1, 64), (2, 64)])
@pytest.mark.vivado
def test_npy2apintstream(test_shape, dtype):
@@ -119,17 +121,17 @@ def test_numpy_to_hls_code():
return "".join(s.split())
A = [[1, 1, 1, 0], [0, 1, 1, 0]]
- ret = numpy_to_hls_code(A, DataType.BINARY, "test", True)
+ ret = numpy_to_hls_code(A, DataType["BINARY"], "test", True)
eA = """ap_uint<4> test[2] =
{ap_uint<4>("0xe", 16), ap_uint<4>("0x6", 16)};"""
assert remove_all_whitespace(ret) == remove_all_whitespace(eA)
B = [[[3, 3], [3, 3]], [[1, 3], [3, 1]]]
- ret = numpy_to_hls_code(B, DataType.UINT2, "test", True)
+ ret = numpy_to_hls_code(B, DataType["UINT2"], "test", True)
eB = """ap_uint<4> test[2][2] =
{{ap_uint<4>("0xf", 16), ap_uint<4>("0xf", 16)},
{ap_uint<4>("0x7", 16), ap_uint<4>("0xd", 16)}};"""
assert remove_all_whitespace(ret) == remove_all_whitespace(eB)
- ret = numpy_to_hls_code(B, DataType.UINT2, "test", True, True)
+ ret = numpy_to_hls_code(B, DataType["UINT2"], "test", True, True)
eB = """{{ap_uint<4>("0xf", 16), ap_uint<4>("0xf", 16)},
{ap_uint<4>("0x7", 16), ap_uint<4>("0xd", 16)}};"""
assert remove_all_whitespace(ret) == remove_all_whitespace(eB)