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_channelwise_layer(pdt, idt, onnx_op_name, scalar_param, exec_mode): ifm_ch = 16 ifm_dim = 5 ishape = (1, ifm_ch, ifm_dim, ifm_dim) if scalar_param: pshape = (1, ) else: pshape = (1, ifm_ch, 1, 1) np.random.seed(0) model = make_single_maxpool_modelwrapper(onnx_op_name, ishape, idt, pdt, pshape) # 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)) else: x = gen_finn_dt_tensor(idt, (1, ifm_ch, ifm_dim, ifm_dim)) input_dict = prepare_inputs(x) y_expected = oxe.execute_onnx(model, input_dict)["outp"] new_model = model.transform(to_hls.InferChannelwiseLinearLayer()) new_model = new_model.transform(GiveUniqueNodeNames()) if exec_mode == "cppsim": new_model = new_model.transform(PrepareCppSim()) new_model = new_model.transform(CompileCppSim()) new_model = new_model.transform(SetExecMode("cppsim")) 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") ctx_produced = oxe.execute_onnx(new_model, input_dict, return_full_exec_context=True) y_produced = ctx_produced["outp"] assert (y_produced == y_expected).all() assert new_model.graph.node[1].op_type == "ChannelwiseOp_Batch"
def step_convert_final_layers(model: ModelWrapper, cfg: DataflowBuildConfig): model = model.transform(to_hls.InferChannelwiseLinearLayer()) model = model.transform(to_hls.InferLabelSelectLayer()) model = model.transform(GiveUniqueNodeNames()) return model
def test_convert_to_hls_layers_synthetic(ch, ifmdim, idt): model = make_model(ch, ifmdim) model.save(export_onnx_path) model = ModelWrapper(export_onnx_path) model = model.transform(InferShapes()) model = model.transform(FoldConstants()) model = model.transform(GiveUniqueNodeNames()) model = model.transform(GiveReadableTensorNames()) model = model.transform(InferDataLayouts()) # model.save("golden.onnx") # generate test vectors of correct shape if ifmdim == -1: input_tensor_shape = (1, ch) else: input_tensor_shape = (1, ch, ifmdim, ifmdim) x = gen_finn_dt_tensor(idt, input_tensor_shape) # generate expected value from streamlined net input_dict = {model.graph.input[0].name: x} output_dict = oxe.execute_onnx(model, input_dict, True) produced_sum = output_dict[model.graph.output[0].name] chw_mul = model.get_initializer(model.graph.node[-1].input[1]) chw_mul = 1 expected_sum = chw_mul * np.sum(2 * (2 * x + 15.0), axis=(2, 3)) / (ifmdim * ifmdim) assert (produced_sum.flatten() == expected_sum.flatten()).all() model = model.transform(InferDataLayouts()) # convert to hls model.set_tensor_datatype(model.graph.input[0].name, idt) # extra streamlining model = model.transform(MoveScalarLinearPastInvariants()) model = model.transform(MoveAddPastMul()) model = model.transform(CollapseRepeatedMul()) model = model.transform(CollapseRepeatedAdd()) # insert top-k node, which should absorb linear ops before it model = model.transform(InferShapes()) model = model.transform(InferDataLayouts()) model = model.transform(InferDataTypes()) model = model.transform(to_hls.InferChannelwiseLinearLayer()) model = model.transform(to_hls.InferAddStreamsLayer()) model = model.transform(to_hls.InferGlobalAccPoolLayer()) model = model.transform(MoveScalarLinearPastInvariants()) model = model.transform(InsertTopK()) model = model.transform(AbsorbScalarMulAddIntoTopK()) model = model.transform(InferDataTypes()) model = model.transform(to_hls.InferLabelSelectLayer()) model = model.transform(AbsorbConsecutiveTransposes()) model = model.transform(InferDataTypes()) model = model.transform(to_hls.InferLabelSelectLayer()) model = model.transform(to_hls.InferDuplicateStreamsLayer()) model = model.transform(SortGraph()) # model.save("golden_hls.onnx") # check topology status finn_nodes = model.get_finn_nodes() assert len(finn_nodes) == 9 add_nodes = model.get_nodes_by_op_type("AddStreams_Batch") assert len(add_nodes) == 1 pool_nodes = model.get_nodes_by_op_type("GlobalAccPool_Batch") assert len(pool_nodes) == 1 label_nodes = model.get_nodes_by_op_type("LabelSelect_Batch") assert len(label_nodes) == 1 channelwise_nodes = model.get_nodes_by_op_type("ChannelwiseOp_Batch") assert len(channelwise_nodes) == 5 dup_nodes = model.get_nodes_by_op_type("DuplicateStreams_Batch") assert len(dup_nodes) == 1 model = model.transform(PrepareCppSim()) model = model.transform(CompileCppSim()) model = model.transform(SetExecMode("cppsim")) output_dict = oxe.execute_onnx(model, input_dict, True) produced_topk_hls = output_dict[model.graph.output[0].name] topk_input = output_dict[model.graph.node[-1].input[0]] assert soft_verify_topk(topk_input, produced_topk_hls, 5) os.remove(export_onnx_path)