def test_end2end_cnv_w1a1_verify_dataflow_part():
model = ModelWrapper(build_dir + "/end2end_cnv_w1a1_ipstitch.onnx")
x = np.zeros((1, 32, 32, 3), dtype=np.float32)
inp_name = model.graph.input[0].name
out_name = model.graph.output[0].name
inp_dict = {inp_name: x}
# cppsim
model = model.transform(PrepareCppSim())
model = model.transform(CompileCppSim())
model = model.transform(SetExecMode("cppsim"))
model.save(build_dir + "/end2end_cnv_w1a1_ipgen_cppsim.onnx")
ret_cppsim = execute_onnx(model, inp_dict, True)
res_cppsim = ret_cppsim[out_name]
# node-by-node rtlsim
model = model.transform(SetExecMode("rtlsim"))
model = model.transform(PrepareRTLSim())
model.save(build_dir + "/end2end_cnv_w1a1_ipgen_nodebynode_rtlsim.onnx")
ret_rtlsim_nodebynode = execute_onnx(model, inp_dict, True)
res_rtlsim_nodebynode = ret_rtlsim_nodebynode[out_name]
# whole-network (ip-stitched) rtlsim
model.set_metadata_prop("exec_mode", "rtlsim")
model.save(build_dir + "/end2end_cnv_w1a1_ipstitch_whole_rtlsim.onnx")
# this is a particularly long-running test, set liveness thr. to unlimited
os.environ["LIVENESS_THRESHOLD"] = "-1"
ret_rtlsim_whole = execute_onnx(model, inp_dict, True)
res_rtlsim_whole = ret_rtlsim_whole[out_name]
assert np.isclose(res_cppsim, res_rtlsim_nodebynode).all()
assert np.isclose(res_cppsim, res_rtlsim_whole).all()
def test_end2end_cnv_w1a1_fold_and_tlastmarker():
model = ModelWrapper(build_dir + "/end2end_cnv_w1a1_dataflow_model.onnx")
fc_layers = model.get_nodes_by_op_type("StreamingFCLayer_Batch")
# each tuple is (PE, SIMD, in_fifo_depth) for a layer
folding = [
(16, 3, 128),
(32, 32, 128),
(16, 32, 128),
(16, 32, 128),
(4, 32, 81),
(1, 32, 2),
(1, 4, 2),
(1, 8, 128),
(5, 1, 3),
]
for fcl, (pe, simd, ififodepth) in zip(fc_layers, folding):
fcl_inst = getCustomOp(fcl)
fcl_inst.set_nodeattr("PE", pe)
fcl_inst.set_nodeattr("SIMD", simd)
fcl_inst.set_nodeattr("inFIFODepth", ififodepth)
swg_layers = model.get_nodes_by_op_type("ConvolutionInputGenerator")
for i in range(len(swg_layers)):
swg_inst = getCustomOp(swg_layers[i])
simd = folding[i][1]
swg_inst.set_nodeattr("SIMD", simd)
model = model.transform(InsertDWC())
model = model.transform(InsertFIFO())
model = model.transform(InsertTLastMarker())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(AnnotateResources("estimate"))
model.save(build_dir + "/end2end_cnv_w1a1_folded.onnx")
def test_renaming():
# load the onnx model
raw_m = get_data("finn.data", "onnx/mnist-conv/model.onnx")
model = ModelWrapper(raw_m)
model = model.transform(InferShapes())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
# do some basic checks
assert model.graph.input[0].name == "global_in"
assert model.graph.output[0].name == "global_out"
assert model.graph.node[1].op_type == "Conv"
assert model.graph.node[1].name == "Conv_0"
assert model.graph.node[1].input[1] == "Conv_0_param0"
assert model.graph.node[6].op_type == "Add"
assert model.graph.node[6].name == "Add_1"
assert model.graph.node[6].input[1] == "Add_1_param0"
# ensure running renaming twice still yields the same names
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
assert model.graph.node[1].op_type == "Conv"
assert model.graph.node[1].name == "Conv_0"
assert model.graph.node[1].input[1] == "Conv_0_param0"
assert model.graph.node[6].op_type == "Add"
assert model.graph.node[6].name == "Add_1"
assert model.graph.node[6].input[1] == "Add_1_param0"
# run renamed model to make sure we did not mess up the topology
raw_i = get_data("finn.data", "onnx/mnist-conv/test_data_set_0/input_0.pb")
raw_o = get_data("finn.data", "onnx/mnist-conv/test_data_set_0/output_0.pb")
input_tensor = onnx.load_tensor_from_string(raw_i)
output_tensor = onnx.load_tensor_from_string(raw_o)
input_dict = {"global_in": np_helper.to_array(input_tensor)}
output_dict = oxe.execute_onnx(model, input_dict)
assert np.isclose(
np_helper.to_array(output_tensor), output_dict["global_out"], atol=1e-3
).all()
def test_xnorpopcountmatmul():
M = 1
K = 3
N = 3
x = helper.make_tensor_value_info("x", TensorProto.FLOAT, [M, K])
W = helper.make_tensor_value_info("W", TensorProto.FLOAT, [K, N])
out = helper.make_tensor_value_info("out", TensorProto.FLOAT, ["x", "y"])
node_def = helper.make_node("XnorPopcountMatMul", ["x", "W"], ["out"],
domain="finn.custom_op.general")
modelproto = helper.make_model(
helper.make_graph([node_def], "test_model", [x], [out],
value_info=[W]))
model = ModelWrapper(modelproto)
model.set_tensor_datatype("x", DataType.BINARY)
model.set_tensor_datatype("W", DataType.BINARY)
W_data = np.asarray([[1, 0, 0], [0, 1, 0], [0, 0, 1]], dtype=np.float32)
model.set_initializer("W", W_data)
# test shape inference
model = model.transform(InferShapes())
assert model.get_tensor_shape("out") == [M, N]
# test datatype inference
assert model.get_tensor_datatype("out") is DataType.FLOAT32
model = model.transform(InferDataTypes())
assert model.get_tensor_datatype("out") is DataType.UINT32
# test execution
x_data = np.asarray([[1, 0, 0]], dtype=np.float32)
inp_dict = {"x": x_data}
out_dict = oxe.execute_onnx(model, inp_dict)
Wb = 2 * W_data - 1
xb = 2 * x_data - 1
rb = np.matmul(xb, Wb)
assert (2 * out_dict["out"] - K == rb).all()
def step_resnet50_streamline_linear(model: ModelWrapper,
cfg: DataflowBuildConfig):
streamline_transformations = [
AbsorbScalarMulAddIntoTopK(
), # before MoveAddPastMul to avoid int->float
ConvertSubToAdd(),
ConvertDivToMul(),
RemoveIdentityOps(),
CollapseRepeatedMul(),
BatchNormToAffine(),
ConvertSignToThres(),
MoveAddPastMul(),
MoveScalarAddPastMatMul(),
MoveAddPastConv(),
MoveScalarMulPastMatMul(),
MoveScalarMulPastConv(),
MoveScalarLinearPastInvariants(),
MoveAddPastMul(),
CollapseRepeatedAdd(),
CollapseRepeatedMul(),
AbsorbAddIntoMultiThreshold(),
FactorOutMulSignMagnitude(),
MoveMaxPoolPastMultiThreshold(),
AbsorbMulIntoMultiThreshold(),
Absorb1BitMulIntoMatMul(),
Absorb1BitMulIntoConv(),
RoundAndClipThresholds(),
]
for trn in streamline_transformations:
model = model.transform(trn)
model = model.transform(GiveUniqueNodeNames())
return model
def step_streamline(model: ModelWrapper, cfg: DataflowBuildConfig):
"""Run streamlining on given model. Streamlining involves moving floating point
scale/shift parameters around, collapsing adjacent ones into a single parameter,
then absorbing the scale/shift into the following `MultiThreshold` node.
Streamlining requires careful topology design and cannot be applied to all
topologies.
"""
model = model.transform(MoveScalarLinearPastInvariants())
model = model.transform(Streamline())
need_lowering = len(model.get_nodes_by_op_type("Conv")) > 0
if need_lowering:
model = model.transform(LowerConvsToMatMul())
model = model.transform(MakeMaxPoolNHWC())
model = model.transform(absorb.AbsorbTransposeIntoMultiThreshold())
model = model.transform(ConvertBipolarMatMulToXnorPopcount())
model = model.transform(Streamline())
# absorb final add-mul nodes into TopK
model = model.transform(absorb.AbsorbScalarMulAddIntoTopK())
model = model.transform(InferDataLayouts())
model = model.transform(RemoveUnusedTensors())
if VerificationStepType.STREAMLINED_PYTHON in cfg._resolve_verification_steps(
):
verify_step(model, cfg, "streamlined_python", need_parent=False)
return model
def test_brevitas_debug():
finn_onnx = "test_brevitas_debug.onnx"
fc = get_test_model_trained("TFC", 2, 2)
dbg_hook = bo.enable_debug(fc)
bo.export_finn_onnx(fc, (1, 1, 28, 28), finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
assert len(model.graph.input) == 1
assert len(model.graph.output) == 1
# load one of the test vectors
raw_i = get_data("finn", "data/onnx/mnist-conv/test_data_set_0/input_0.pb")
input_tensor = onnx.load_tensor_from_string(raw_i)
# run using FINN-based execution
input_dict = {"0": nph.to_array(input_tensor)}
output_dict = oxe.execute_onnx(model, input_dict, return_full_exec_context=True)
produced = output_dict[model.graph.output[0].name]
# run using PyTorch/Brevitas
input_tensor = torch.from_numpy(nph.to_array(input_tensor)).float()
assert input_tensor.shape == (1, 1, 28, 28)
# do forward pass in PyTorch/Brevitas
expected = fc.forward(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=1e-3).all()
# check all tensors at debug markers
names_brevitas = set(dbg_hook.values.keys())
names_finn = set(output_dict.keys())
names_common = names_brevitas.intersection(names_finn)
assert len(names_common) == 16
for dbg_name in names_common:
tensor_pytorch = dbg_hook.values[dbg_name].detach().numpy()
tensor_finn = output_dict[dbg_name]
assert np.isclose(tensor_finn, tensor_pytorch, atol=1e-5).all()
os.remove(finn_onnx)
def test_streamline_fc(size, wbits, abits):
if size == "LFC" and wbits == 2 and abits == 2:
pytest.skip("No LFC-w2a2 present at the moment")
if wbits > abits:
pytest.skip("No wbits > abits cases at the moment")
nname = "%s_%dW%dA" % (size, wbits, abits)
finn_onnx = export_onnx_path + "/%s.onnx" % nname
fc = get_test_model_trained(size, wbits, abits)
bo.export_finn_onnx(fc, (1, 1, 28, 28), finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
# load one of the test vectors
raw_i = get_data("finn", "data/onnx/mnist-conv/test_data_set_0/input_0.pb")
input_tensor = onnx.load_tensor_from_string(raw_i)
# run using FINN-based execution
input_dict = {"global_in": nph.to_array(input_tensor)}
expected_ctx = oxe.execute_onnx(model, input_dict, True)
expected = expected_ctx[model.graph.output[0].name]
model = model.transform(Streamline())
produced_ctx = oxe.execute_onnx(model, input_dict, True)
produced = produced_ctx[model.graph.output[0].name]
assert np.isclose(expected, produced, atol=1e-3).all()
def test_brevitas_fc_onnx_export_and_exec(size, wbits, abits):
if size == "LFC" and wbits == 2 and abits == 2:
pytest.skip("No LFC-w2a2 present at the moment")
if wbits > abits:
pytest.skip("No wbits > abits cases at the moment")
nname = "%s_%dW%dA" % (size, wbits, abits)
finn_onnx = export_onnx_path + "/%s.onnx" % nname
fc = get_test_model_trained(size, wbits, abits)
bo.export_finn_onnx(fc, (1, 1, 28, 28), finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
# load one of the test vectors
raw_i = get_data("finn", "data/onnx/mnist-conv/test_data_set_0/input_0.pb")
input_tensor = onnx.load_tensor_from_string(raw_i)
# run using FINN-based execution
input_dict = {"0": nph.to_array(input_tensor)}
output_dict = oxe.execute_onnx(model, input_dict)
produced = output_dict[list(output_dict.keys())[0]]
# run using PyTorch/Brevitas
input_tensor = torch.from_numpy(nph.to_array(input_tensor)).float()
assert input_tensor.shape == (1, 1, 28, 28)
# do forward pass in PyTorch/Brevitas
expected = fc.forward(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=1e-3).all()
def test_end2end_tfc_w1a2_verify_dataflow_part():
model = ModelWrapper(build_dir + "/end2end_tfc_w1a2_ipstitch.onnx")
x = np.zeros((1, 784), dtype=np.float32)
inp_name = model.graph.input[0].name
out_name = model.graph.output[0].name
inp_dict = {inp_name: x}
# npysim
model = model.transform(CodeGen_npysim())
model = model.transform(Compile())
model = model.transform(SetExecMode("npysim"))
model.save(build_dir + "/end2end_tfc_w1a2_ipstitch_npysim.onnx")
ret_npysim = execute_onnx(model, inp_dict, True)
res_npysim = ret_npysim[out_name]
# node-by-node rtlsim
model = model.transform(SetExecMode("rtlsim"))
getCustomOp(model.graph.node[0]).set_nodeattr("rtlsim_trace", "default")
getCustomOp(model.graph.node[1]).set_nodeattr("rtlsim_trace", "default")
getCustomOp(model.graph.node[2]).set_nodeattr("rtlsim_trace", "default")
getCustomOp(model.graph.node[3]).set_nodeattr("rtlsim_trace", "default")
model.save(build_dir + "/end2end_tfc_w1a2_ipstitch_nodebynode_rtlsim.onnx")
ret_rtlsim_nodebynode = execute_onnx(model, inp_dict, True)
res_rtlsim_nodebynode = ret_rtlsim_nodebynode[out_name]
# whole-network (ip-stitched) rtlsim
model.set_metadata_prop("exec_mode", "rtlsim")
model.set_metadata_prop("rtlsim_trace", "whole_trace.vcd")
model.save(build_dir + "/end2end_tfc_w1a2_ipstitch_whole_rtlsim.onnx")
ret_rtlsim_whole = execute_onnx(model, inp_dict, True)
res_rtlsim_whole = ret_rtlsim_whole[out_name]
assert np.isclose(res_npysim, res_rtlsim_nodebynode).all()
assert np.isclose(res_npysim, res_rtlsim_whole).all()
def test_move_chw_add_past_conv(idim, k, s, ich, och):
odim = compute_conv_output_dim(idim, k, s)
ishape = [1, ich, idim, idim]
oshape = [1, och, odim, odim]
add_param_shape = [1, ich, 1, 1]
conv_param_shape = [och, ich, k, k]
inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, ishape)
outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, oshape)
a0 = helper.make_tensor_value_info("a0", TensorProto.FLOAT,
add_param_shape)
a1 = helper.make_tensor_value_info("a1", TensorProto.FLOAT,
conv_param_shape)
conv_config = {}
conv_config["dilations"] = [1, 1]
conv_config["group"] = 1
conv_config["kernel_shape"] = [k, k]
conv_config["pads"] = [0, 0, 0, 0]
conv_config["strides"] = [s, s]
add_node = helper.make_node("Add", ["inp", "a0"], ["add_out"])
conv_node = helper.make_node("Conv", ["add_out", "a1"], ["outp"],
**conv_config)
model = helper.make_model(
helper.make_graph(
nodes=[add_node, conv_node],
name="move-add-graph",
inputs=[inp],
outputs=[outp],
value_info=[a0, a1],
))
model = ModelWrapper(model)
# initialize model
a0_values = np.random.uniform(
low=0, high=1, size=tuple(add_param_shape)).astype(np.float32)
model.set_initializer("a0", a0_values)
a1_values = np.random.uniform(
low=0, high=1, size=tuple(conv_param_shape)).astype(np.float32)
model.set_initializer("a1", a1_values)
model = model.transform(InferShapes())
# execution before transformation
inp_values = np.random.uniform(low=0, high=1,
size=tuple(ishape)).astype(np.float32)
idict = {model.graph.input[0].name: inp_values}
odict = oxe.execute_onnx(model, idict)
y_before = odict[model.graph.output[0].name]
model = model.transform(MoveAddPastConv())
odict = oxe.execute_onnx(model, idict)
y_after = odict[model.graph.output[0].name]
assert np.isclose(y_before, y_after).all()
assert model.graph.node[0].op_type == "Conv"
assert model.graph.node[1].op_type == "Add"
def test_add_pre_and_postproc(self, topology, wbits, abits):
prev_chkpt_name = get_checkpoint_name(topology, wbits, abits,
"import_and_tidy")
model = load_test_checkpoint_or_skip(prev_chkpt_name)
global_inp_name = model.graph.input[0].name
ishape = model.get_tensor_shape(global_inp_name)
# preprocessing: torchvision's ToTensor divides uint8 inputs by 255
totensor_pyt = ToTensor()
chkpt_preproc_name = get_checkpoint_name(topology, wbits, abits,
"preproc")
bo.export_finn_onnx(totensor_pyt, ishape, chkpt_preproc_name)
assert os.path.isfile(chkpt_preproc_name)
# join preprocessing and core model
pre_model = ModelWrapper(chkpt_preproc_name)
pre_model = pre_model.transform(InferShapes())
pre_model = pre_model.transform(FoldConstants())
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)
# postprocessing: insert Top-1 node at the end
model = model.transform(InsertTopK(k=1))
chkpt_name = get_checkpoint_name(topology, wbits, abits, "pre_post")
# tidy-up again
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
model = model.transform(RemoveStaticGraphInputs())
model.save(chkpt_name)
assert os.path.isfile(chkpt_name)
def test_collapse_repeated_op():
top_in = oh.make_tensor_value_info("top_in", TensorProto.FLOAT, [2])
add_param_0 = oh.make_tensor_value_info("add_param_0", TensorProto.FLOAT, [2])
mul_param_0 = oh.make_tensor_value_info("mul_param_0", TensorProto.FLOAT, [2])
add_param_1 = oh.make_tensor_value_info("add_param_1", TensorProto.FLOAT, [2])
mul_param_1 = oh.make_tensor_value_info("mul_param_1", TensorProto.FLOAT, [2])
top_out = oh.make_tensor_value_info("top_out", TensorProto.FLOAT, [2])
modelproto = oh.make_model(
oh.make_graph(
name="test",
inputs=[top_in],
outputs=[top_out],
value_info=[add_param_0, mul_param_0, add_param_1, mul_param_1],
nodes=[
oh.make_node("Add", ["top_in", "add_param_0"], ["middle_0"]),
oh.make_node("Add", ["middle_0", "add_param_1"], ["middle_1"]),
oh.make_node("Mul", ["middle_1", "mul_param_0"], ["middle_2"]),
oh.make_node("Mul", ["middle_2", "mul_param_1"], ["top_out"]),
],
)
)
model = ModelWrapper(modelproto)
model = model.transform(InferShapes())
model.set_initializer("add_param_0", np.asarray([1, 3], dtype=np.float32))
model.set_initializer("add_param_1", np.asarray([-1, 3], dtype=np.float32))
model.set_initializer("mul_param_0", np.asarray([2, 4], dtype=np.float32))
model.set_initializer("mul_param_1", np.asarray([2, -4], dtype=np.float32))
new_model = model.transform(CollapseRepeatedAdd())
new_model = new_model.transform(CollapseRepeatedMul())
inp_dict = {"top_in": np.asarray([-1.0, 1.0], dtype=np.float32)}
assert ox.compare_execution(model, new_model, inp_dict)
def test_brevitas_fc_onnx_export_and_exec(size, wbits, abits, pretrained):
if size == "LFC" and wbits == 2 and abits == 2:
pytest.skip(f"No LFC_{MAX_WBITS}W{MAX_ABITS}A present.")
if wbits > abits:
pytest.skip("No wbits > abits cases.")
nname = f"{size}_{wbits}W{abits}A"
finn_onnx = nname + ".onnx"
fc, _ = model_with_cfg(nname.lower(), pretrained=pretrained)
FINNManager.export_onnx(fc, FC_INPUT_SIZE, finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(GiveUniqueNodeNames())
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
# load a random test vector
input_tensor = np.random.uniform(MIN_INP_VAL,
MAX_INP_VAL,
size=FC_INPUT_SIZE).astype(np.float32)
# run using FINN-based execution
input_dict = {"0": input_tensor}
output_dict = oxe.execute_onnx(model, input_dict)
produced = output_dict[list(output_dict.keys())[0]]
# run using PyTorch/Brevitas
input_tensor = torch.from_numpy(input_tensor).float()
# do forward pass in PyTorch/Brevitas
expected = fc.forward(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=ATOL).all()
def test_brevitas_cnv_export_exec(wbits, abits):
if wbits > abits:
pytest.skip("No wbits > abits cases at the moment")
cnv = get_test_model_trained("CNV", wbits, abits)
bo.export_finn_onnx(cnv, (1, 3, 32, 32), export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(GiveUniqueNodeNames())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
assert len(model.graph.input) == 1
assert len(model.graph.output) == 1
fn = pk.resource_filename("finn", "data/cifar10/cifar10-test-data-class3.npz")
input_tensor = np.load(fn)["arr_0"].astype(np.float32)
input_tensor = input_tensor / 255
assert input_tensor.shape == (1, 3, 32, 32)
# run using FINN-based execution
input_dict = {model.graph.input[0].name: input_tensor}
output_dict = oxe.execute_onnx(model, input_dict, True)
produced = output_dict[model.graph.output[0].name]
# do forward pass in PyTorch/Brevitas
input_tensor = torch.from_numpy(input_tensor).float()
expected = cnv.forward(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=1e-3).all()
assert np.argmax(produced) == 3
os.remove(export_onnx_path)
def test_onnx_exec_internal_rounding():
inp0 = onnx.helper.make_tensor_value_info("inp0", onnx.TensorProto.FLOAT,
[2, 2])
inp1 = onnx.helper.make_tensor_value_info("inp1", onnx.TensorProto.FLOAT,
[1])
outp = onnx.helper.make_tensor_value_info("outp", onnx.TensorProto.FLOAT,
[2, 2])
mul_node = onnx.helper.make_node("Mul",
inputs=["inp0", "inp1"],
outputs=["outp"])
graph = onnx.helper.make_graph(nodes=[mul_node],
name="mul_graph",
inputs=[inp0, inp1],
outputs=[outp])
model = onnx.helper.make_model(graph, producer_name="mul-model")
model = ModelWrapper(model)
idt = DataType.INT2
model.set_tensor_datatype("inp0", idt)
model.set_tensor_datatype("inp1", idt)
model.transform(InferShapes())
mul_value = np.asarray([-1], dtype=np.float32)
inp_int = gen_finn_dt_tensor(idt, [2, 2])
scale = np.random.uniform(low=0, high=1, size=(2, 2)).astype(np.float32)
inp_rounded = (inp_int * scale) / (scale + 1e-7)
input_dict = {"inp0": inp_rounded, "inp1": mul_value}
output_dict = oxe.execute_onnx(model, input_dict)
produced = output_dict["outp"]
expected = np.multiply(inp_int, mul_value)
assert (produced == expected).all()
def make_dupstreams_modelwrapper(ch, pe, idim, idt):
shape = [1, idim, idim, ch]
inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, shape)
outp0 = helper.make_tensor_value_info("outp0", TensorProto.FLOAT, shape)
outp1 = helper.make_tensor_value_info("outp1", TensorProto.FLOAT, shape)
dupstrm_node = helper.make_node(
"DuplicateStreams_Batch",
["inp"],
["outp0", "outp1"],
domain="finn.custom_op.fpgadataflow",
backend="fpgadataflow",
NumChannels=ch,
PE=pe,
inputDataType=idt.name,
numInputVectors=[1, idim, idim],
)
graph = helper.make_graph(nodes=[dupstrm_node],
name="graph",
inputs=[inp],
outputs=[outp0, outp1])
model = helper.make_model(graph, producer_name="addstreams-model")
model = ModelWrapper(model)
model.set_tensor_datatype("inp", idt)
model = model.transform(InferShapes())
model = model.transform(InferDataTypes())
return model
def test_move_scalar_add_past_matmul():
top_in = oh.make_tensor_value_info("top_in", TensorProto.FLOAT, [1, 2])
add_param = oh.make_tensor_value_info("add_param", TensorProto.FLOAT, [1, 1])
matmul_param = oh.make_tensor_value_info("matmul_param", TensorProto.FLOAT, [2, 2])
top_out = oh.make_tensor_value_info("top_out", TensorProto.FLOAT, [1, 2])
modelproto = oh.make_model(
oh.make_graph(
name="test",
inputs=[top_in],
outputs=[top_out],
value_info=[add_param, matmul_param],
nodes=[
oh.make_node("Add", ["top_in", "add_param"], ["middle"]),
oh.make_node("MatMul", ["middle", "matmul_param"], ["top_out"]),
],
)
)
model = ModelWrapper(modelproto)
model = model.transform(InferShapes())
model.set_initializer("add_param", np.asarray([[3]], dtype=np.float32))
model.set_initializer(
"matmul_param", np.asarray([[2, 4], [-1, 1]], dtype=np.float32)
)
new_model = model.transform(MoveScalarAddPastMatMul())
inp_dict = {"top_in": np.asarray([[-1.0, 1.0]], dtype=np.float32)}
assert ox.compare_execution(model, new_model, inp_dict)
assert new_model.graph.node[0].op_type == "MatMul"
assert new_model.graph.node[1].op_type == "Add"
assert new_model.graph.node[0].output[0] == new_model.graph.node[1].input[0]
def make_lookup_model(embeddings, ishape, idt, edt):
num_embeddings, embedding_dim = embeddings.shape
class LookupModel(nn.Module):
def __init__(self, num_embeddings, embedding_dim):
super().__init__()
self.lookup = nn.Embedding(
num_embeddings=num_embeddings, embedding_dim=embedding_dim
)
def forward(self, x):
x = self.lookup(x)
return x
torch_model = LookupModel(num_embeddings, embedding_dim)
input_t = torch.zeros(ishape, dtype=torch.int64)
ret = FINNManager.export(torch_model, input_t=input_t, opset_version=11)
model = ModelWrapper(ret)
iname = model.graph.input[0].name
ename = model.graph.node[0].input[0]
model.set_tensor_datatype(iname, idt)
eshape = model.get_tensor_shape(ename)
assert tuple(eshape) == embeddings.shape
model.set_initializer(ename, embeddings)
model.set_tensor_datatype(ename, edt)
model = model.transform(InferShapes())
model = model.transform(InferDataTypes())
return model
def test_quartznet_asr_4b(pretrained):
finn_onnx = "quant_quartznet_perchannelscaling_4b.onnx"
quartznet = quant_quartznet_perchannelscaling_4b(pretrained,
export_mode=True)
quartznet.eval()
FINNManager.export(quartznet, QUARTZNET_POSTPROCESSED_INPUT_SIZE,
finn_onnx)
model = ModelWrapper(finn_onnx)
model = model.transform(GiveUniqueNodeNames())
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
#load a random test vector
input_tensor = np.random.uniform(
MIN_INP_VAL, MAX_INP_VAL,
size=QUARTZNET_POSTPROCESSED_INPUT_SIZE).astype(np.float32)
# run using FINN-based execution
input_dict = {"0": input_tensor}
output_dict = oxe.execute_onnx(model, input_dict)
produced = output_dict[list(output_dict.keys())[0]]
# run using PyTorch/Brevitas
input_tensor = torch.from_numpy(input_tensor).float()
# do forward pass in PyTorch/Brevitas
expected = quartznet(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=ATOL).all()
def test_end2end_tfc_w1a2_verify_dataflow_part():
model = ModelWrapper(build_dir + "/end2end_tfc_w1a2_ipstitch.onnx")
x = np.zeros((1, 784), dtype=np.float32)
inp_name = model.graph.input[0].name
out_name = model.graph.output[0].name
inp_dict = {inp_name: x}
# cppsim
model = model.transform(PrepareCppSim())
model = model.transform(CompileCppSim())
model = model.transform(SetExecMode("cppsim"))
model.save(build_dir + "/end2end_tfc_w1a2_ipstitch_cppsim.onnx")
ret_cppsim = execute_onnx(model, inp_dict, True)
res_cppsim = ret_cppsim[out_name]
# node-by-node rtlsim
model = model.transform(SetExecMode("rtlsim"))
model = model.transform(PrepareRTLSim())
model.save(build_dir + "/end2end_tfc_w1a2_ipstitch_nodebynode_rtlsim.onnx")
ret_rtlsim_nodebynode = execute_onnx(model, inp_dict, True)
res_rtlsim_nodebynode = ret_rtlsim_nodebynode[out_name]
# whole-network (ip-stitched) rtlsim
model.set_metadata_prop("exec_mode", "rtlsim")
model.save(build_dir + "/end2end_tfc_w1a2_ipstitch_whole_rtlsim.onnx")
ret_rtlsim_whole = execute_onnx(model, inp_dict, True)
res_rtlsim_whole = ret_rtlsim_whole[out_name]
assert np.isclose(res_cppsim, res_rtlsim_nodebynode).all()
assert np.isclose(res_cppsim, res_rtlsim_whole).all()
def test_brevitas_cnv_onnx_export_and_exec(wbits, abits, pretrained):
if wbits > abits:
pytest.skip("No wbits > abits cases.")
nname = f"CNV_{wbits}W{abits}A"
finn_onnx = nname + ".onnx"
cnv, _ = model_with_cfg(nname.lower(), pretrained=pretrained)
cnv.eval()
# load a random int test vector
input_a = np.random.randint(MIN_INP_VAL, MAX_INP_VAL, size=CNV_INPUT_SIZE).astype(np.float32)
scale = 1. / 255
input_t = torch.from_numpy(input_a * scale)
input_qt = QuantTensor(
input_t, scale=torch.tensor(scale), bit_width=torch.tensor(8.0), signed=False)
FINNManager.export(cnv, export_path=finn_onnx, input_t=input_qt)
model = ModelWrapper(finn_onnx)
model = model.transform(GiveUniqueNodeNames())
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(RemoveStaticGraphInputs())
# run using FINN-based execution
input_dict = {"0": input_a}
output_dict = oxe.execute_onnx(model, input_dict)
produced = output_dict[list(output_dict.keys())[0]]
# do forward pass in PyTorch/Brevitas
expected = cnv(input_t).detach().numpy()
assert np.isclose(produced, expected, atol=ATOL).all()
def test_brevitas_QConv2d(dw, bias, in_channels, QONNX_export):
ishape = (1, 32, 111, 111)
if dw is True:
groups = in_channels
out_channels = in_channels
kernel_size = 3
padding = 1
stride = 1
w_shape = (32, 1, 3, 3)
else:
groups = 1
out_channels = 64
kernel_size = 1
padding = 0
stride = 1
w_shape = (64, 32, 1, 1)
b_conv = QuantConv2d(
in_channels=in_channels,
out_channels=out_channels,
groups=groups,
kernel_size=kernel_size,
padding=padding,
stride=stride,
bias=bias,
bias_quant_type=QuantType.FP,
weight_bit_width=4,
weight_quant_type=QuantType.INT,
weight_scaling_impl_type=ScalingImplType.STATS,
weight_scaling_stats_op=StatsOp.MAX,
weight_scaling_per_output_channel=True,
weight_restrict_scaling_type=RestrictValueType.LOG_FP,
weight_narrow_range=True,
weight_scaling_min_val=2e-16,
)
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()
if QONNX_export:
m_path = export_onnx_path
BrevitasONNXManager.export(b_conv, ishape, m_path)
qonnx_cleanup(m_path, out_file=m_path)
model = ModelWrapper(m_path)
model = model.transform(ConvertQONNXtoFINN())
model.save(m_path)
else:
bo.export_finn_onnx(b_conv, ishape, export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(InferShapes())
inp_tensor = np.random.uniform(low=-1.0, high=1.0, size=ishape).astype(np.float32)
idict = {model.graph.input[0].name: inp_tensor}
odict = oxe.execute_onnx(model, idict, True)
produced = odict[model.graph.output[0].name]
inp_tensor = torch.from_numpy(inp_tensor).float()
expected = b_conv.forward(inp_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=1e-3).all()
os.remove(export_onnx_path)
def test_end2end_cnv_w1a1_convert_to_hls_layers():
model = ModelWrapper(build_dir + "/end2end_cnv_w1a1_streamlined.onnx")
model = model.transform(to_hls.InferBinaryStreamingFCLayer(mem_mode))
model = model.transform(to_hls.InferQuantizedStreamingFCLayer(mem_mode))
model = model.transform(to_hls.InferConvInpGen())
model = model.transform(to_hls.InferStreamingMaxPool())
model = model.transform(MoveReshape())
model.save(build_dir + "/end2end_cnv_w1a1_hls_layers.onnx")
def test_end2end_tfc_w1a2_import_and_tidy():
model = ModelWrapper(build_dir + "/end2end_tfc_w1a2_export.onnx")
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
model.save(build_dir + "/end2end_tfc_w1a2_tidy.onnx")
def test_end2end_tfc_w1a1_convert_to_hls_layers():
model = ModelWrapper(build_dir + "/end2end_tfc_w1a1_streamlined.onnx")
model = model.transform(ConvertBipolarMatMulToXnorPopcount())
model = model.transform(absorb.AbsorbAddIntoMultiThreshold())
model = model.transform(absorb.AbsorbMulIntoMultiThreshold())
model = model.transform(RoundAndClipThresholds())
model = model.transform(to_hls.InferBinaryStreamingFCLayer())
model.save(build_dir + "/end2end_tfc_w1a1_hls_layers.onnx")
def test_end2end_cnv_w1a1_import_and_tidy():
model = ModelWrapper(build_dir + "/end2end_cnv_w1a1_export.onnx")
model = model.transform(DoubleToSingleFloat())
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model.save(build_dir + "/end2end_cnv_w1a1_tidy.onnx")
def step_resnet50_set_fifo_depths(model: ModelWrapper,
cfg: DataflowBuildConfig):
"""
Depending on the auto_fifo_depths setting, do one of the following:
* if auto_fifo_depths=True: Run the `InsertAndSetFIFODepths` transformation
to attempt to determine the FIFO sizes that provide full throughput. Involves
running stitched-IP rtlsim and may take a long time.
* if auto_fifo_depths=False: Assume the folding config file contains FIFO
sizes as well. Runs the `InsertFIFO` transformation, then
`ApplyConfig(cfg.folding_config_file)`, and finally `RemoveShallowFIFOs`.
Coherency with config file node naming is ensured by calling
`GiveUniqueNodeNames`.
"""
if cfg.auto_fifo_depths:
model = model.transform(
InsertAndSetFIFODepths(
cfg._resolve_fpga_part(),
cfg._resolve_hls_clk_period(),
vivado_ram_style=cfg.large_fifo_mem_style.value,
))
else:
# assume folding cfg json contains FIFO sizes too
# insert DWCs, FIFOs and run ApplyConfig once more
model = model.transform(InsertDWC())
# need to make sure all FIFOs are created so that their depth can be
# set by ApplyConfig, so create_shallow_fifos=True
model = model.transform(InsertFIFO(create_shallow_fifos=True))
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
if cfg.folding_config_file is not None:
model = model.transform(ApplyConfig(cfg.folding_config_file))
# remove any shallow FIFOs
model = model.transform(RemoveShallowFIFOs())
# extract the final configuration and save it as json
hw_attrs = [
"PE",
"SIMD",
"ram_style",
"depth",
"impl_style",
"resType",
"mem_mode",
"runtime_writeable_weights",
]
extract_model_config_to_json(model,
cfg.output_dir + "/final_hw_config.json",
hw_attrs)
# after FIFOs are ready to go, call PrepareIP and HLSSynthIP again
# this will only run for the new nodes (e.g. FIFOs and DWCs)
model = model.transform(
PrepareIP(cfg._resolve_fpga_part(), cfg._resolve_hls_clk_period()))
model = model.transform(HLSSynthIP())
model = model.transform(ReplaceVerilogRelPaths())
return model
def test_conv_lowering_conv_1x1():
np.random.seed(0)
in_feature_dim_h = 7
in_feature_dim_w = 7
in_chn = 3
kernel_size = 1
out_feature_dim_h = in_feature_dim_h
out_feature_dim_w = in_feature_dim_w
input_shape = [1, in_chn, in_feature_dim_h, in_feature_dim_w]
output_shape = [1, in_chn, out_feature_dim_h, out_feature_dim_w]
conv_param_shape = [in_chn, in_chn, kernel_size, kernel_size]
conv_config = {}
conv_config["dilations"] = [1, 1]
conv_config["group"] = 1
conv_config["kernel_shape"] = [kernel_size, kernel_size]
conv_config["pads"] = [0, 0, 0, 0]
conv_config["strides"] = [1, 1]
top_in = oh.make_tensor_value_info("top_in", TensorProto.FLOAT,
input_shape)
top_out = oh.make_tensor_value_info("top_out", TensorProto.FLOAT,
output_shape)
value_info = [
oh.make_tensor_value_info("p1", TensorProto.FLOAT, conv_param_shape)
]
modelproto = oh.make_model(
oh.make_graph(
name="test",
inputs=[top_in],
outputs=[top_out],
value_info=value_info,
nodes=[
oh.make_node("Conv", ["top_in", "p1"], ["top_out"],
**conv_config)
],
))
model = ModelWrapper(modelproto)
model = model.transform(InferShapes())
model.set_initializer("p1",
np.random.rand(*conv_param_shape).astype(np.float32))
new_model = model.transform(LowerConvsToMatMul())
inp_dict = {"top_in": np.random.rand(*input_shape).astype(np.float32)}
assert oxe.compare_execution(model, new_model, inp_dict)
assert new_model.graph.node[0].op_type == "Transpose"
assert new_model.graph.node[1].op_type == "MatMul"
assert new_model.graph.node[2].op_type == "Transpose"
assert len(new_model.graph.node) == 3
def test_const_folding_shapes():
lfc = get_test_model_untrained("LFC", 1, 1)
bo.export_finn_onnx(lfc, (1, 1, 28, 28), export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(InferShapes())
model = model.transform(FoldConstants())
assert model.graph.node[0].op_type == "Reshape"
assert list(model.get_tensor_shape("0")) == [1, 1, 28, 28]
assert list(model.get_tensor_shape("27")) == [1, 784]
os.remove(export_onnx_path)
