def test_end2end_mobilenet_convert_to_hls_layers():
model = load_test_checkpoint_or_skip(build_dir + "/end2end_mobilenet_lowered.onnx")
model = model.transform(to_hls.InferPool_Batch())
model = model.transform(to_hls.InferConvInpGen())
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())
model.save(build_dir + "/end2end_mobilenet_hls_layers.onnx")
def step_mobilenet_convert_to_hls_layers(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.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_convert_to_hls_pool_batch(idt, odt, pool_config, ifm_ch, pe, op_type,
exec_mode):
k, stride, pad, ifm_dim = pool_config
if ifm_ch % pe != 0:
pytest.skip("ifm_ch%pe != 0. Skipping")
if pad != 0 and idt.signed():
pytest.skip("No support for pal_val != 0. Skipping")
np.random.seed(0)
ofm_dim = int(((ifm_dim + 2 * pad - k) / stride) + 1)
x = gen_finn_dt_tensor(idt, (1, ifm_ch, ifm_dim, ifm_dim))
# prepare input data
input_dict = prepare_inputs(x)
if op_type == "MaxPool":
# if idt.signed():
# pytest.skip("""No support for signed input (see accu initialization
# in Pool_batch HLSLIB function). Skipping""")
if idt != odt:
pytest.skip("Skipping Maxpool with idt != odt")
model = make_single_maxpool_modelwrapper(k, stride, pad, ifm_ch,
ifm_dim, ofm_dim, idt)
elif op_type == "QuantAvgPool2d":
if pad != 0:
pytest.skip("No padding support for QuantAvgPool2d. Skipping")
if idt.signed() != odt.signed():
pytest.skip(
"Skipping QuantAvgPool2d with idt.signed() != odt.signed()")
model = make_single_quantavpool_modelwrapper(k, stride, ifm_ch,
ifm_dim, ofm_dim, idt,
odt)
else:
assert False, "{} is not a supported op_type".format(op_type)
y_expected = oxe.execute_onnx(model, input_dict)["outp"]
new_model = model.transform(to_hls.InferPool_Batch())
new_model = new_model.transform(GiveUniqueNodeNames())
if ifm_ch != pe:
new_model = new_model.transform(to_hls.InferConvInpGen())
# Folding
for n in new_model.graph.node:
if n.op_type == "ConvolutionInputGenerator":
inst = getCustomOp(n)
inst.set_nodeattr("SIMD", pe)
elif n.op_type == "Pool_Batch":
inst = getCustomOp(n)
inst.set_nodeattr("PE", pe)
if exec_mode == "cppsim":
new_model = new_model.transform(SetExecMode("cppsim"))
new_model = new_model.transform(PrepareCppSim())
new_model = new_model.transform(CompileCppSim())
elif exec_mode == "rtlsim":
new_model = new_model.transform(SetExecMode("rtlsim"))
new_model = new_model.transform(GiveUniqueNodeNames())
new_model = new_model.transform(PrepareIP("xc7z020clg400-1", 5))
new_model = new_model.transform(HLSSynthIP())
new_model = new_model.transform(PrepareRTLSim())
else:
raise Exception("Unknown exec_mode")
# execute new_model
y_produced = oxe.execute_onnx(new_model, input_dict)["outp"]
assert (y_produced == y_expected).all()
if stride <= k:
if pad == 0 or ifm_ch == pe:
assert len(new_model.graph.node) == 4
else:
assert len(new_model.graph.node) == 5
else:
assert len(new_model.graph.node) == 1
if exec_mode == "rtlsim":
node = new_model.get_nodes_by_op_type("Pool_Batch")[0]
inst = getCustomOp(node)
cycles_rtlsim = inst.get_nodeattr("cycles_rtlsim")
exp_cycles_dict = new_model.analysis(exp_cycles_per_layer)
exp_cycles = exp_cycles_dict[node.name]
assert np.isclose(exp_cycles, cycles_rtlsim, atol=10)