diff --git a/tests/fpgadataflow/test_fpgadataflow_fclayer.py b/tests/fpgadataflow/test_fpgadataflow_fclayer.py
index 545a196e11f5cda99b0c3e749acd6aed437e7ad1..ce37455e2bd395ba3fc7cd6fda1e55a8c411e9dd 100644
--- a/tests/fpgadataflow/test_fpgadataflow_fclayer.py
+++ b/tests/fpgadataflow/test_fpgadataflow_fclayer.py
@@ -1,3 +1,5 @@
+import pytest
+
import numpy as np
from onnx import TensorProto, helper
@@ -5,11 +7,7 @@ import finn.core.onnx_exec as oxe
import finn.custom_op.xnorpopcount as xp
from finn.core.datatype import DataType
from finn.core.modelwrapper import ModelWrapper
-from finn.core.utils import (
- gen_finn_dt_tensor,
- interleave_matrix_outer_dim_from_partitions
-)
-from finn.custom_op.multithreshold import multithreshold
+from finn.core.utils import gen_finn_dt_tensor
def make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T=None, tdt=None):
@@ -23,10 +21,8 @@ def make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T=None, tdt=Non
sf = mw // simd
if T is not None:
tmem = nf
- n_thres_steps = T.shape[1]
else:
tmem = 0
- n_thres_steps = 0
inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, sf, simd])
outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, nf, pe])
@@ -53,7 +49,7 @@ def make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T=None, tdt=Non
outputDataType=odt.name,
)
graph = helper.make_graph(
- nodes=[FCLayer_node], name="fclayer_graph", inputs=[inp], outputs=[outp],
+ nodes=[FCLayer_node], name="fclayer_graph", inputs=[inp], outputs=[outp]
)
model = helper.make_model(graph, producer_name="fclayer-model")
@@ -74,10 +70,12 @@ def prepare_inputs(model, input_tensor, idt):
input_tensor = (np.asarray(input_tensor, dtype=np.float32)).reshape(*ishape)
return {"inp": input_tensor}
-def create_noactivation_testcases(idt, wdt, odt):
- mh = 8
- mw = 8
+def create_noactivation_testcases(idt, wdt, mh=8, mw=8):
+ if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
+ odt = DataType.UINT32
+ else:
+ odt = DataType.INT32
# generate weights
W = gen_finn_dt_tensor(wdt, (mw, mh))
# generate input data
@@ -88,13 +86,13 @@ def create_noactivation_testcases(idt, wdt, odt):
simd_values = [1, int(mw / 2), mw]
for pe in pe_values:
for simd in simd_values:
- model = make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T, tdt)
+ model = make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt)
# prepare input data
input_dict = prepare_inputs(model, x, idt)
if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
# convert inputs to binary and use xnorpopcountmatmul
- y = xp.xnorpopcountmatmul((x+1)/2, (W+1)/2)
+ y = xp.xnorpopcountmatmul((x + 1) / 2, (W + 1) / 2)
else:
y = np.matmul(x, W)
oshape = model.get_tensor_shape("outp")
@@ -104,191 +102,7 @@ def create_noactivation_testcases(idt, wdt, odt):
assert (y_produced.reshape(y_expected.shape) == y_expected).all()
-# no activation cases
-
-# no act -all bipolar
-def test_fpgadataflow_fclayer_ibp_wbp_noact():
- wdt = idt = DataType.BIPOLAR
- odt = DataType.UINT32
-
- create_noativation_testcases(idt, wdt, odt)
-
-
-# no act - all signed
-def test_fpgadataflow_fclayer_iint2_wint2_noact():
- wdt = idt = DataType.INT2
- odt = DataType.INT32
- create_noativation_testcases(idt, wdt, odt)
-
-
-# no act - all ternary
-
-
-def test_fpgadataflow_fclayer_it_wt_noact():
- wdt = idt = DataType.TERNARY
- odt = DataType.INT32
- create_noativation_testcases(idt, wdt, odt)
-
-
-# no act - idt: bipolar wdt: signed
-
-
-def test_fpgadataflow_fclayer_ibp_wint2_noact():
- wdt = DataType.INT2
- idt = DataType.BIPOLAR
- odt = DataType.INT32
- create_noativation_testcases(idt, wdt, odt)
-
-
-# no act - idt: bipolar wdt: ternary
-
-
-def test_fpgadataflow_fclayer_ibp_wt_noact():
- wdt = DataType.INT2
- idt = DataType.TERNARY
- odt = DataType.INT32
- create_noativation_testcases(idt, wdt, odt)
-
-
-# no act - idt: signed wdt: bipolar
-
-
-def test_fpgadataflow_fclayer_it_wbp_noact():
- wdt = DataType.BIPOLAR
- idt = DataType.INT2
- odt = DataType.INT32
- create_noativation_testcases(idt, wdt, odt)
-
-
-# no act - idt: signed wdt: ternary
-
-
-def test_fpgadataflow_fclayer_it_wbp_noact():
- wdt = DataType.TERNARY
- idt = DataType.INT2
- odt = DataType.INT32
- create_noativation_testcases(idt, wdt, odt)
-
-
-# no act - idt: ternary wdt: bipolar
-
-
-def test_fpgadataflow_fclayer_it_wbp_noact():
- wdt = DataType.BIPOLAR
- idt = DataType.TERNARY
- odt = DataType.INT32
- create_noativation_testcases(idt, wdt, odt)
-
-
-# no act - idt: ternary wdt: signed
-
-
-def test_fpgadataflow_fclayer_it_wbp_noact():
- wdt = DataType.INT2
- idt = DataType.TERNARY
- odt = DataType.INT32
- create_noativation_testcases(idt, wdt, odt)
-
-
-def test_fpgadataflow_fclayer_all_bipolar():
- mh = 8
- mw = 8
- wdt = idt = odt = DataType.BIPOLAR
- tdt = DataType.UINT32
- # generate weights
- W = gen_finn_dt_tensor(wdt, (mw, mh))
- # single global threshold at zero
- T = np.zeros((1, 1))
-
- # generate input data
- x = gen_finn_dt_tensor(idt, (1, mw))
-
- # set up layers with different pe and simd
- pe_values = [1, int(mh / 2), mh]
- simd_values = [1, int(mw / 2), mw]
- for pe in pe_values:
- for simd in simd_values:
- model = make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T, tdt)
-
- # prepare input data
- input_dict = prepare_inputs(model, x, idt)
-
- # execute model
- produced = oxe.execute_onnx(model, input_dict)["outp"]
- # TODO fix this to resemble the noact tester above
- # expected output
- # correction of bipolar values to enable xnorpopcountmutmal
- Wb = (W + 1) * 0.5
- xb = (x + 1) * 0.5
- y = xp.xnorpopcountmatmul(Wb, xb.reshape(-1, 1))
- expected = multithreshold(y.reshape(1, mh), T)
-
- assert (produced.reshape(expected.shape) == expected).all()
-
- wdt = idt = odt = DataType.BIPOLAR
- create_testcases(idt, wdt, odt)
-
-def test_fpgadataflow_fclayer_all_signed():
- mh = 8
- mw = 8
- wdt = idt = odt = DataType.INT2
- tdt = DataType.INT32
- # generate weights
- W = gen_finn_dt_tensor(wdt, [mh, mw])
- # single global threshold at zero
- T = np.zeros((1, 1))
-
- # generate input data
- x = gen_finn_dt_tensor(idt, mw)
-
- # set up layers with different pe and simd
- pe_values = [1, int(mh / 2), mh]
- simd_values = [1, int(mw / 2), mw]
- for pe in pe_values:
- for simd in simd_values:
- model = make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T, tdt)
- # prepare input data
- input_dict = prepare_inputs(model, x, idt)
-
- # execute model
- produced = oxe.execute_onnx(model, input_dict)["outp"]
-
- # expected output
- oshape = model.get_tensor_shape("outp")
- y = np.dot(W, x).reshape(oshape.shape)
- expected = multithreshold(y.reshape(1, mh), T)
-
- assert (produced.reshape(expected.shape) == expected).all()
-
-
-def test_fpgadataflow_fclayer_all_ternary():
- mh = 8
- mw = 8
- wdt = idt = odt = DataType.TERNARY
- tdt = DataType.INT32
- # generate weights
- W = gen_finn_dt_tensor(wdt, [mh, mw])
- # single global threshold at zero
- T = np.zeros((1, 1))
-
- # generate input data
- x = gen_finn_dt_tensor(idt, mw)
-
- # set up layers with different pe and simd
- pe_values = [1, int(mh / 2), mh]
- simd_values = [1, int(mw / 2), mw]
- for pe in pe_values:
- for simd in simd_values:
- model = make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T, tdt)
- # prepare input data
- input_dict = prepare_inputs(model, x, idt)
-
- # execute model
- produced = oxe.execute_onnx(model, input_dict)["outp"]
-
- # expected output
- oshape = model.get_tensor_shape("outp")
- y = np.dot(W, x).reshape(oshape.shape)
- expected = multithreshold(y.reshape(1, mh), T)
-
- assert (produced.reshape(expected.shape) == expected).all()
+@pytest.mark.parametrize("wdt", [DataType.BIPOLAR, DataType.INT2])
+@pytest.mark.parametrize("idt", [DataType.BIPOLAR, DataType.INT2])
+def test_fpgadataflow_fclayer_noact(wdt, idt):
+ create_noactivation_testcases(idt, wdt)