def execute_node(self, context, graph):
node = self.onnx_node
inp_name = node.input[0]
out_name = node.output[0]
inp = context[inp_name]
dummy_out = context[out_name]
# convert i/o NHWC -> NCHW
inp = np.transpose(inp, (0, 3, 1, 2))
dummy_out = np.transpose(dummy_out, (0, 3, 1, 2))
# execute as regular MaxPool
orig_domain = node.domain
node.domain = ""
node.op_type = "MaxPool"
inp_vi = helper.make_tensor_value_info(inp_name, TensorProto.FLOAT,
inp.shape)
out_vi = helper.make_tensor_value_info(out_name, TensorProto.FLOAT,
dummy_out.shape)
tmp_graph = helper.make_graph(nodes=[node],
name="tmp_graph",
inputs=[inp_vi],
outputs=[out_vi])
tmp_model = helper.make_model(tmp_graph, producer_name="finn")
tmp_model = ModelWrapper(tmp_model)
new_ctx = {inp_name: inp}
from finn.core.onnx_exec import execute_onnx
ret = execute_onnx(tmp_model, new_ctx)
# restore original node props
node.domain = orig_domain
node.op_type = "MaxPoolNHWC"
outp = ret[out_name]
# convert output NCHW -> NHWC
outp = np.transpose(outp, (0, 2, 3, 1))
context[out_name] = outp
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)
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_fpgadataflow_ipstitch_gen_model(): # exec_mode):
model = create_one_fc_model()
if model.graph.node[0].op_type == "StreamingDataflowPartition":
sdp_node = getCustomOp(model.graph.node[0])
assert sdp_node.__class__.__name__ == "StreamingDataflowPartition"
assert os.path.isfile(sdp_node.get_nodeattr("model"))
model = ModelWrapper(sdp_node.get_nodeattr("model"))
model.set_metadata_prop("exec_mode", "remote_pynq")
model = model.transform(InsertTLastMarker())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(PrepareIP(test_fpga_part, 5))
model = model.transform(HLSSynthIP())
assert model.graph.node[0].op_type == "StreamingFCLayer_Batch"
assert model.graph.node[-1].op_type == "TLastMarker"
model.save(ip_stitch_model_dir +
"/test_fpgadataflow_ipstitch_gen_model.onnx")
def step_hls_codegen(model: ModelWrapper, cfg: DataflowBuildConfig):
"Generate Vivado HLS code to prepare HLSCustomOp nodes for IP generation."
model = model.transform(
PrepareIP(cfg._resolve_fpga_part(), cfg._resolve_hls_clk_period())
)
return model
def step_apply_folding_config(model: ModelWrapper, cfg: DataflowBuildConfig):
"""Apply the folding configuration file onto the model to set folding (parallelization)
and other attributes, if config file is specified."""
if cfg.folding_config_file is not None:
model = model.transform(GiveUniqueNodeNames())
model = model.transform(ApplyConfig(cfg.folding_config_file))
if VerificationStepType.FOLDED_HLS_CPPSIM in cfg._resolve_verification_steps(
):
# prepare cppsim
model = model.transform(PrepareCppSim())
model = model.transform(CompileCppSim())
model = model.transform(SetExecMode("cppsim"))
verify_step(model, cfg, "folded_hls_cppsim", need_parent=True)
return model
def step_target_fps_parallelization(model: ModelWrapper, cfg: DataflowBuildConfig):
"""If target_fps was specified, use the SetFolding transformation to determine
parallelization attributes. The auto-generated config will be saved under
auto_folding_config.json under the outputs, which can serve as a basis for
customizing the folding factors further."""
target_cycles_per_frame = cfg._resolve_cycles_per_frame()
if target_cycles_per_frame is not None:
model = model.transform(
SetFolding(
target_cycles_per_frame,
mvau_wwidth_max=cfg.mvau_wwidth_max,
two_pass_relaxation=cfg.folding_two_pass_relaxation,
)
)
# extract the suggested configuration and save it as json
hw_attrs = [
"PE",
"SIMD",
"ram_style",
"resType",
"mem_mode",
"runtime_writeable_weights",
]
extract_model_config_to_json(
model, cfg.output_dir + "/auto_folding_config.json", hw_attrs
)
return model
def test_modelwrapper_detect_forks_n_joins():
# create small network with properties to be tested
Neg_node = onnx.helper.make_node("Neg", inputs=["in1"], outputs=["neg1"])
Round_node = onnx.helper.make_node("Round",
inputs=["neg1"],
outputs=["round1"])
Ceil_node = onnx.helper.make_node("Ceil",
inputs=["neg1"],
outputs=["ceil1"])
Add_node = onnx.helper.make_node("Add",
inputs=["round1", "ceil1"],
outputs=["out1"])
in1 = onnx.helper.make_tensor_value_info("in1", onnx.TensorProto.FLOAT,
[4, 4])
out1 = onnx.helper.make_tensor_value_info("out1", onnx.TensorProto.FLOAT,
[4, 4])
graph = onnx.helper.make_graph(
nodes=[Neg_node, Round_node, Ceil_node, Add_node],
name="simple_graph",
inputs=[in1],
outputs=[out1],
value_info=[
onnx.helper.make_tensor_value_info("neg1", onnx.TensorProto.FLOAT,
[4, 4]),
onnx.helper.make_tensor_value_info("round1",
onnx.TensorProto.FLOAT, [4, 4]),
onnx.helper.make_tensor_value_info("ceil1", onnx.TensorProto.FLOAT,
[4, 4]),
],
)
onnx_model = onnx.helper.make_model(graph, producer_name="simple-model")
model = ModelWrapper(onnx_model)
# test
assert model.is_fork_node(Neg_node)
assert not model.is_fork_node(Round_node)
assert not model.is_fork_node(Ceil_node)
assert not model.is_fork_node(Add_node)
assert not model.is_join_node(Neg_node)
assert not model.is_join_node(Round_node)
assert not model.is_join_node(Ceil_node)
assert model.is_join_node(Add_node)
def test_brevitas_act_export_relu(abits, max_val, scaling_impl_type):
min_val = -1.0
ishape = (1, 15)
b_act = QuantReLU(
bit_width=abits,
max_val=max_val,
scaling_impl_type=scaling_impl_type,
restrict_scaling_type=RestrictValueType.LOG_FP,
quant_type=QuantType.INT,
)
if scaling_impl_type == ScalingImplType.PARAMETER:
checkpoint = {
"act_quant_proxy.fused_activation_quant_proxy.tensor_quant.\
scaling_impl.learned_value":
torch.tensor(0.49).type(torch.FloatTensor)
}
b_act.load_state_dict(checkpoint)
bo.export_finn_onnx(b_act, ishape, export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(InferShapes())
inp_tensor = np.random.uniform(low=min_val, high=max_val,
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()
b_act.eval()
expected = b_act.forward(inp_tensor).detach().numpy()
if not np.isclose(produced, expected, atol=1e-3).all():
print(abits, max_val, scaling_impl_type)
print("scale: ",
b_act.quant_act_scale().type(torch.FloatTensor).detach())
if abits < 5:
print(
"thres:",
", ".join(["{:8.4f}".format(x)
for x in b_act.export_thres[0]]),
)
print("input:",
", ".join(["{:8.4f}".format(x) for x in inp_tensor[0]]))
print("prod :", ", ".join(["{:8.4f}".format(x) for x in produced[0]]))
print("expec:", ", ".join(["{:8.4f}".format(x) for x in expected[0]]))
assert np.isclose(produced, expected, atol=1e-3).all()
os.remove(export_onnx_path)
def test_move_scalar_past_matmul_only_if_linear(test_args):
scalar_op = test_args[0]
transf_fxn = test_args[1]
input_shape = [1, 2]
matmul_shape = [2, 2]
top_in = oh.make_tensor_value_info("top_in", TensorProto.FLOAT, input_shape)
top_out = oh.make_tensor_value_info("top_out", TensorProto.FLOAT, input_shape)
p1 = oh.make_tensor_value_info("p1", TensorProto.FLOAT, [1, 1])
p2 = oh.make_tensor_value_info("p2", TensorProto.FLOAT, matmul_shape)
p3 = oh.make_tensor_value_info("p3", TensorProto.FLOAT, matmul_shape)
p4 = oh.make_tensor_value_info("p4", TensorProto.FLOAT, matmul_shape)
modelproto = oh.make_model(
oh.make_graph(
name="test",
inputs=[top_in],
outputs=[top_out],
value_info=[p1, p2, p3, p4],
nodes=[
oh.make_node(scalar_op, ["top_in", "p1"], ["t1"]),
oh.make_node("MatMul", ["t1", "p2"], ["fork"]),
oh.make_node("MatMul", ["fork", "p3"], ["t3"]),
oh.make_node(scalar_op, ["t3", "fork"], ["t4"]),
oh.make_node("MatMul", ["t4", "p4"], ["top_out"]),
],
)
)
model = ModelWrapper(modelproto)
model = model.transform(InferShapes())
np.random.seed(0)
model.set_initializer("p1", np.random.rand(1, 1).astype(np.float32))
model.set_initializer("p2", np.random.rand(*matmul_shape).astype(np.float32))
model.set_initializer("p3", np.random.rand(*matmul_shape).astype(np.float32))
model.set_initializer("p4", np.random.rand(*matmul_shape).astype(np.float32))
# Transform
new_model = model.transform(transf_fxn)
# Test
inp_dict = {"top_in": np.random.rand(*input_shape).astype(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 == scalar_op
assert new_model.graph.node[2].op_type == "MatMul"
assert new_model.graph.node[3].op_type == scalar_op
assert new_model.graph.node[4].op_type == "MatMul"
def load_test_checkpoint_or_skip(filename):
"Try to load given .onnx and return ModelWrapper, else skip current test."
if os.path.isfile(filename):
model = ModelWrapper(filename)
return model
else:
warnings.warn(filename + " not found from previous test step, skipping")
pytest.skip(filename + " not found from previous test step, skipping")
def test_end2end_tfc_w1a2_deploy_on_pynq():
model = ModelWrapper(build_dir + "/end2end_tfc_w1a2_pynq_driver.onnx")
try:
ip = os.environ[
"PYNQ_IP"] # no fault for this one; skip if not defined
if ip == "":
pytest.skip("PYNQ board IP address not specified")
username = os.getenv("PYNQ_USERNAME", "xilinx")
password = os.getenv("PYNQ_PASSWORD", "xilinx")
port = os.getenv("PYNQ_PORT", 22)
target_dir = os.getenv("PYNQ_TARGET_DIR", "/home/xilinx/finn")
model = model.transform(
DeployToPYNQ(ip, port, username, password, target_dir))
# save the model to be able to link it to the parent
model.save(build_dir + "/end2end_tfc_w1a2_pynq_deploy.onnx")
except KeyError:
pytest.skip("PYNQ board IP address not specified")
def load_test_checkpoint_or_skip(filename):
"Try to load given .onnx and return ModelWrapper, else skip current test."
try:
model = ModelWrapper(filename)
return model
except FileNotFoundError:
warnings.warn(filename + " not found from previous test step, skipping")
pytest.skip(filename + " not found from previous test step, skipping")
def apply(self, model):
# hide your riches!
hidden_ops = _hide_finn_ops(model)
# call regular ONNX shape inference
model = ModelWrapper(si.infer_shapes(model.model))
# bring back hidden ops
_restore_finn_ops(model, hidden_ops)
return (model, False)
def test_fpgadataflow_ipstitch_remote_execution():
try:
ip = os.environ["PYNQ_IP"] # NOQA
if ip == "":
pytest.skip("PYNQ board IP address not specified")
model = ModelWrapper(
ip_stitch_model_dir +
"/test_fpgadataflow_ipstitch_pynq_deployment.onnx")
iname = "inp"
idt = model.get_tensor_datatype(iname)
ishape = model.get_tensor_shape(iname)
x = gen_finn_dt_tensor(idt, ishape)
input_dict = {"inp": x}
outp = execute_onnx(model, input_dict)
assert np.isclose(outp["outp"], x).all()
except KeyError:
pytest.skip("PYNQ board IP address not specified")
def test_change_datalayout_quantavgpool(s, k, ibits, obits, signed, c, idim):
n = 1
odim = compute_pool_output_dim(idim, k, s)
# determine input FINN datatype
if signed is True:
prefix = "INT"
else:
prefix = "UINT"
dt_name = prefix + str(ibits)
dtype = DataType[dt_name]
inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [n, c, idim, idim])
outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [n, c, odim, odim])
node = helper.make_node(
"QuantAvgPool2d",
["inp"],
["outp"],
domain="finn.custom_op.general",
stride=s,
kernel=k,
ibits=ibits,
obits=obits,
signed=signed,
data_layout="NCHW",
)
graph = helper.make_graph(
nodes=[node], name="single-quantavgpool", inputs=[inp], outputs=[outp]
)
model = helper.make_model(graph)
model = ModelWrapper(model)
model = model.transform(InferShapes())
model = model.transform(InferDataTypes())
model = model.transform(InferDataLayouts())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model_transformed = model.transform(ChangeDataLayoutQuantAvgPool2d())
model_transformed = model_transformed.transform(InferShapes())
model_transformed = model_transformed.transform(InferDataTypes())
model_transformed = model_transformed.transform(InferDataLayouts())
model_transformed = model_transformed.transform(GiveUniqueNodeNames())
model_transformed = model_transformed.transform(GiveReadableTensorNames())
inp_values = gen_finn_dt_tensor(dtype, [n, c, idim, idim])
idict = {"inp": inp_values}
assert oxe.compare_execution(model, model_transformed, idict)
assert len(model.graph.node) + 2 == len(model_transformed.graph.node)
assert model_transformed.graph.node[-1].op_type == "Transpose"
assert model_transformed.graph.node[0].op_type == "Transpose"
# check if QuantAvgPool2d node has datalayout set correctly
node = model_transformed.graph.node[1]
d_layout = get_by_name(node.attribute, "data_layout").s.decode("UTF-8")
assert d_layout == "NHWC"
assert model_transformed.get_tensor_layout(node.input[0]) == DataLayout.NHWC
assert model_transformed.get_tensor_layout(node.output[0]) == DataLayout.NHWC
def step_mobilenet_lower_convs(model: ModelWrapper, cfg: DataflowBuildConfig):
model = model.transform(LowerConvsToMatMul())
model = model.transform(absorb.AbsorbTransposeIntoMultiThreshold())
model = model.transform(absorb.AbsorbConsecutiveTransposes())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
model = model.transform(RoundAndClipThresholds())
model = model.transform(InferDataLayouts())
return model
def make_single_maxpool_modelwrapper(k, stride, pad, ifm_ch, ifm_dim, ofm_dim,
idt):
odt = idt
inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT,
[1, ifm_ch, ifm_dim, ifm_dim])
outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT,
[1, ifm_ch, ofm_dim, ofm_dim])
mp_node = helper.make_node(
"MaxPool",
["inp"],
["outp"],
kernel_shape=[k, k],
pads=[pad, pad, pad, pad],
strides=[stride, stride],
)
graph = helper.make_graph(nodes=[mp_node],
name="mp_graph",
inputs=[inp],
outputs=[outp])
model = helper.make_model(graph, producer_name="mp-model")
model = ModelWrapper(model)
model.set_tensor_datatype("inp", idt)
model.set_tensor_datatype("outp", odt)
model = model.transform(InferShapes())
return model
def make_single_quantavpool_modelwrapper(k, stride, ifm_ch, ifm_dim, ofm_dim,
idt, odt):
inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT,
[1, ifm_ch, ifm_dim, ifm_dim])
outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT,
[1, ifm_ch, ofm_dim, ofm_dim])
mp_node = helper.make_node(
"QuantAvgPool2d",
["inp"],
["outp"],
domain="finn.custom_op.general",
stride=stride,
kernel=k,
ibits=idt.bitwidth(),
obits=odt.bitwidth(),
signed=1 if idt.signed() else 0,
data_layout="NCHW",
)
graph = helper.make_graph(nodes=[mp_node],
name="mp_graph",
inputs=[inp],
outputs=[outp])
model = helper.make_model(graph, producer_name="mp-model")
model = ModelWrapper(model)
model.set_tensor_datatype("inp", idt)
model.set_tensor_datatype("outp", odt)
model = model.transform(InferShapes())
return model
def step_mobilenet_convert_to_hls_layers_separate_th(model: ModelWrapper,
cfg: DataflowBuildConfig):
mem_mode = cfg.default_mem_mode.value
model = model.transform(to_hls.InferPool_Batch())
model = model.transform(to_hls.InferConvInpGen())
model = model.transform(to_hls.InferThresholdingLayer())
model = model.transform(to_hls.InferVVAU())
model = model.transform(to_hls.InferQuantizedStreamingFCLayer(mem_mode))
model = model.transform(to_hls.InferChannelwiseLinearLayer())
model = model.transform(to_hls.InferLabelSelectLayer())
model = model.transform(InferShapes())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
return model
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_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_avg_pool_export(kernel_size, stride, signed, bit_width,
input_bit_width, channels, idim):
quant_avgpool = QuantAvgPool2d(kernel_size=kernel_size,
stride=stride,
bit_width=bit_width)
quant_avgpool.eval()
# determine input
prefix = 'INT' if signed else 'UINT'
dt_name = prefix + str(input_bit_width)
dtype = DataType[dt_name]
input_shape = (1, channels, idim, idim)
input_array = gen_finn_dt_tensor(dtype, input_shape)
scale_array = np.random.uniform(low=0, high=1, size=(1, channels, 1,
1)).astype(np.float32)
input_tensor = torch.from_numpy(input_array * scale_array).float()
scale_tensor = torch.from_numpy(scale_array).float()
zp = torch.tensor(0.)
input_quant_tensor = QuantTensor(input_tensor,
scale_tensor,
zp,
input_bit_width,
signed,
training=False)
# export
FINNManager.export(quant_avgpool,
export_path=export_onnx_path,
input_t=input_quant_tensor)
model = ModelWrapper(export_onnx_path)
model = model.transform(InferShapes())
model = model.transform(InferDataTypes())
# reference brevitas output
ref_output_array = quant_avgpool(
input_quant_tensor).tensor.detach().numpy()
# finn output
idict = {model.graph.input[0].name: input_array}
odict = oxe.execute_onnx(model, idict, True)
finn_output = odict[model.graph.output[0].name]
# compare outputs
assert np.isclose(ref_output_array, finn_output).all()
# cleanup
os.remove(export_onnx_path)
def step_create_dataflow_partition(model: ModelWrapper,
cfg: DataflowBuildConfig):
"""Separate consecutive groups of HLSCustomOp nodes into StreamingDataflowPartition
nodes, which point to a separate ONNX file. Dataflow accelerator synthesis
can only be performed on those HLSCustomOp sub-graphs."""
parent_model = model.transform(CreateDataflowPartition())
sdp_nodes = parent_model.get_nodes_by_op_type("StreamingDataflowPartition")
assert len(
sdp_nodes) == 1, "Only a single StreamingDataflowPartition supported."
sdp_node = sdp_nodes[0]
sdp_node = getCustomOp(sdp_node)
dataflow_model_filename = sdp_node.get_nodeattr("model")
if cfg.save_intermediate_models:
parent_model.save(cfg.output_dir +
"/intermediate_models/dataflow_parent.onnx")
model = ModelWrapper(dataflow_model_filename)
return model
def step_synthesize_bitfile(model: ModelWrapper, cfg: DataflowBuildConfig):
"""Synthesize a bitfile for the using the specified shell flow, using either
Vivado or Vitis, to target the specified board."""
if DataflowOutputType.BITFILE in cfg.generate_outputs:
bitfile_dir = cfg.output_dir + "/bitfile"
os.makedirs(bitfile_dir, exist_ok=True)
report_dir = cfg.output_dir + "/report"
os.makedirs(report_dir, exist_ok=True)
partition_model_dir = cfg.output_dir + "/intermediate_models/kernel_partitions"
if cfg.shell_flow_type == ShellFlowType.VIVADO_ZYNQ:
model = model.transform(
ZynqBuild(
cfg.board,
cfg.synth_clk_period_ns,
cfg.enable_hw_debug,
partition_model_dir=partition_model_dir,
)
)
copy(model.get_metadata_prop("bitfile"), bitfile_dir + "/finn-accel.bit")
copy(model.get_metadata_prop("hw_handoff"), bitfile_dir + "/finn-accel.hwh")
copy(
model.get_metadata_prop("vivado_synth_rpt"),
report_dir + "/post_synth_resources.xml",
)
vivado_pynq_proj_dir = model.get_metadata_prop("vivado_pynq_proj")
timing_rpt = (
"%s/finn_zynq_link.runs/impl_1/top_wrapper_timing_summary_routed.rpt"
% vivado_pynq_proj_dir
)
copy(timing_rpt, report_dir + "/post_route_timing.rpt")
elif cfg.shell_flow_type == ShellFlowType.VITIS_ALVEO:
model = model.transform(
VitisBuild(
cfg._resolve_fpga_part(),
cfg.synth_clk_period_ns,
cfg.vitis_platform,
strategy=cfg._resolve_vitis_opt_strategy(),
enable_debug=cfg.enable_hw_debug,
floorplan_file=cfg.vitis_floorplan_file,
partition_model_dir=partition_model_dir,
)
)
copy(model.get_metadata_prop("bitfile"), bitfile_dir + "/finn-accel.xclbin")
copy(
model.get_metadata_prop("vivado_synth_rpt"),
report_dir + "/post_synth_resources.xml",
)
else:
raise Exception("Unrecognized shell_flow_type: " + str(cfg.shell_flow_type))
print("Bitfile written into " + bitfile_dir)
return model
def test_batchnorm_to_affine_epsilon(epsilon):
""" Dummy batchnorm node to test out the epsilon attribute. """
batchnorm_node = onnx.helper.make_node(
"BatchNormalization",
inputs=["x", "s", "bias", "mean", "var"],
outputs=["y"],
epsilon=epsilon,
)
x = onnx.helper.make_tensor_value_info("x", onnx.TensorProto.FLOAT, [1, 3, 5, 5])
s = onnx.helper.make_tensor_value_info("s", onnx.TensorProto.FLOAT, [3])
bias = onnx.helper.make_tensor_value_info("bias", onnx.TensorProto.FLOAT, [3])
mean = onnx.helper.make_tensor_value_info("mean", onnx.TensorProto.FLOAT, [3])
var = onnx.helper.make_tensor_value_info("var", onnx.TensorProto.FLOAT, [3])
y = onnx.helper.make_tensor_value_info("y", onnx.TensorProto.FLOAT, [1, 3, 5, 5])
# Graph
graph = onnx.helper.make_graph(
nodes=[batchnorm_node],
name="test_batchnorm_graph",
inputs=[x],
outputs=[y],
value_info=[s, bias, mean, var],
)
onnx_model = onnx.helper.make_model(graph, producer_name="test_batchnorm-model")
model = ModelWrapper(onnx_model)
model.set_initializer("s", np.array([1, 2, 3]).astype(np.float32))
model.set_initializer("bias", np.array([1, 2, 3]).astype(np.float32))
model.set_initializer("mean", np.array([3, 4, 5]).astype(np.float32))
model.set_initializer("var", np.array([0.5, 0.7, 0.3]).astype(np.float32))
i_val = np.arange(0, 3 * 5 * 5, dtype=np.float32)
i_val = np.reshape(i_val, [1, 3, 5, 5])
input_dict = {"x": i_val}
output_node_name = "y"
output_dict = oxe.execute_onnx(model, input_dict, return_full_exec_context=True)
output_original = output_dict[output_node_name]
model_lowered = model.transform(BatchNormToAffine())
output_dict = oxe.execute_onnx(
model_lowered, input_dict, return_full_exec_context=True
)
output_lowered = output_dict[output_node_name]
assert (output_original == output_lowered).all()
def test_infer_shapes():
# load the onnx model
raw_m = get_data("finn.data", "onnx/mnist-conv/model.onnx")
model = ModelWrapper(raw_m)
graph = model.graph
# multi-thresholding node to be inserted between the first Relu and MaxPool node
# get Relu node to use data
Relu_node = graph.node[3]
assert Relu_node.op_type == "Relu", "The wrong model was chosen for the check"
# create thresholds tensor as constant
mt_thresh0 = helper.make_tensor_value_info("mt_thresh0", TensorProto.FLOAT,
[8, 7])
# random numbers for the thresholds
# thresholds for one channel have to be sorted to guarantee the correct behavior
mt_thresh0_values = np.empty([8, 7], dtype=np.float32)
for i in range(len(mt_thresh0_values)):
mt_thresh0_values[i] = np.sort(np.random.random_sample(7) * 10)
model.set_initializer(mt_thresh0.name, mt_thresh0_values)
# add multi-thresholding node and change Relu node
mt_node = helper.make_node(
"MultiThreshold",
["mt_v0", "mt_thresh0"],
[Relu_node.output[0]],
domain="finn.custom_op.general",
)
Relu_node.output[0] = "mt_v0"
# explicitly remove any present shape from ReLU and MultiThreshold outputs
util.remove_by_name(model.graph.value_info, Relu_node.output[0])
util.remove_by_name(model.graph.value_info, mt_node.output[0])
graph.node.insert(4, mt_node)
# first check routine
# check if at least one shape is not specified
assert not (
model.check_all_tensor_shapes_specified()
), "All tensors are already specified before the shape inference execution"
# perform shape inference on mixed model
model = model.transform(InferShapes())
# second check routine
# now all shapes should be specified and mt_node output shape is (1,8,28,28)
assert (model.check_all_tensor_shapes_specified()
), "There are still tensors that are not specified"
assert (model.get_tensor_shape(mt_node.output[0])) == ([
1, 8, 28, 28
]), "output of multi-thresholding node has wrong shape"
def test_end2end_tfc_w1a2_create_dataflow_partition():
model = ModelWrapper(build_dir + "/end2end_tfc_w1a2_hls_layers.onnx")
parent_model = model.transform(CreateDataflowPartition())
parent_model.save(build_dir + "/end2end_tfc_w1a2_dataflow_parent.onnx")
sdp_node = getCustomOp(parent_model.graph.node[2])
dataflow_model_filename = sdp_node.get_nodeattr("model")
dataflow_model = ModelWrapper(dataflow_model_filename)
dataflow_model.save(build_dir + "/end2end_tfc_w1a2_dataflow_model.onnx")
def test_ipstitch_rtlsim(self, topology, wbits, abits, kind):
prev_chkpt_name = get_checkpoint_name(
topology, wbits, abits, "fifodepth_" + kind
)
model = load_test_checkpoint_or_skip(prev_chkpt_name)
test_fpga_part = get_build_env(kind, target_clk_ns)["part"]
model = model.transform(InsertDWC())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(AnnotateCycles())
perf = model.analysis(dataflow_performance)
latency = perf["critical_path_cycles"]
# rtlsim only supports impl_style=rtl for StreamingFIFO, ensure that
for fifo_layer in model.get_nodes_by_op_type("StreamingFIFO"):
getCustomOp(fifo_layer).set_nodeattr("impl_style", "rtl")
model = model.transform(PrepareIP(test_fpga_part, target_clk_ns))
model = model.transform(HLSSynthIP())
model = model.transform(CreateStitchedIP(test_fpga_part, target_clk_ns))
model = model.transform(PrepareRTLSim())
model.set_metadata_prop("exec_mode", "rtlsim")
os.environ["LIVENESS_THRESHOLD"] = str(int(latency * 1.1))
if rtlsim_trace:
model.set_metadata_prop(
"rtlsim_trace", "%s_w%da%d.vcd" % (topology, wbits, abits)
)
os.environ["RTLSIM_TRACE_DEPTH"] = "3"
rtlsim_chkpt = get_checkpoint_name(
topology, wbits, abits, "ipstitch_rtlsim_" + kind
)
model.save(rtlsim_chkpt)
parent_chkpt = get_checkpoint_name(topology, wbits, abits, "dataflow_parent")
(input_tensor_npy, output_tensor_npy) = get_golden_io_pair(
topology, wbits, abits, return_topk=1
)
y = execute_parent(parent_chkpt, rtlsim_chkpt, input_tensor_npy)
model = ModelWrapper(rtlsim_chkpt)
perf["cycles_rtlsim"] = model.get_metadata_prop("cycles_rtlsim")
# warnings.warn("Estimated & rtlsim performance: " + str(perf))
# for (k, v) in perf.items():
# update_dashboard_data(topology, wbits, abits, k, v)
update_dashboard_data(
topology, wbits, abits, "cycles_rtlsim", perf["cycles_rtlsim"]
)
assert np.isclose(y, output_tensor_npy).all()
def make_randomly_sorted_linear_model(num_of_nodes, seed=None):
if seed is not None:
np.random.seed(seed)
ch = 2
ifmdim = 16
input_shape = (1, ch, ifmdim, ifmdim)
top_in = helper.make_tensor_value_info("t0", TensorProto.FLOAT, input_shape)
top_out = helper.make_tensor_value_info(
"t" + str(num_of_nodes), TensorProto.FLOAT, input_shape
)
value_info = []
nodes = []
for i in range(num_of_nodes):
nodes += [
helper.make_node("Add", ["t" + str(i), "p" + str(i)], ["t" + str(i + 1)])
]
value_info += [
helper.make_tensor_value_info("p" + str(i), TensorProto.FLOAT, input_shape)
]
nodes = np.random.permutation(nodes)
modelproto = helper.make_model(
helper.make_graph(
name="test",
inputs=[top_in],
outputs=[top_out],
value_info=value_info,
nodes=nodes,
)
)
model = ModelWrapper(modelproto)
model = model.transform(InferShapes())
for i in range(num_of_nodes):
model.set_initializer(
"p" + str(i), np.random.rand(*input_shape).astype(np.float32)
)
return model
def test_end2end_cnv_w1a1_verify_all():
# use the streamlined model as the "golden" model for right answers
golden = ModelWrapper(build_dir + "/end2end_cnv_w1a1_streamlined.onnx")
iname = golden.graph.input[0].name
oname = golden.graph.output[0].name
# load one of the test vectors
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)
x = input_tensor
# x = np.zeros(ishape, dtype=np.float32)
ret_golden = execute_onnx(golden, {iname: x}, True)
y_golden = ret_golden[oname]
# set up parent+child graph to test
# we'll use models from the previous step as the child model
parent_model = ModelWrapper(build_dir + "/end2end_cnv_w1a1_dataflow_parent.onnx")
iname = parent_model.graph.input[0].name
oname = parent_model.graph.output[0].name
# produce results with cppsim
sdp_node = parent_model.get_nodes_by_op_type("StreamingDataflowPartition")[0]
sdp_node = getCustomOp(sdp_node)
sdp_node.set_nodeattr("model", build_dir + "/end2end_cnv_w1a1_ipgen_cppsim.onnx")
ret_cppsim = execute_onnx(parent_model, {iname: x}, True)
y_cppsim = ret_cppsim[oname]
# produce results with node-by-node rtlsim
sdp_node.set_nodeattr(
"model", build_dir + "/end2end_cnv_w1a1_ipgen_nodebynode_rtlsim.onnx"
)
ret_nodebynode_rtlsim = execute_onnx(parent_model, {iname: x}, True)
y_nodebynode_rtlsim = ret_nodebynode_rtlsim[oname]
# produce results with whole-network (stitched ip) rtlsim
sdp_node.set_nodeattr(
"model", 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_whole_rtlsim = execute_onnx(parent_model, {iname: x}, True)
y_whole_rtlsim = ret_whole_rtlsim[oname]
assert np.isclose(y_golden, y_cppsim).all()
assert np.isclose(y_golden, y_nodebynode_rtlsim).all()
assert np.isclose(y_golden, y_whole_rtlsim).all()
assert np.argmax(y_golden) == 3
