From fa1d6f7b7cfb2bfcf6c7b1c6b7c754ece67066c3 Mon Sep 17 00:00:00 2001
From: auphelia <jakobapk@web.de>
Date: Wed, 22 Apr 2020 13:41:39 +0100
Subject: [PATCH] [Test] Add new test for streaming fclayer decoupled mode with
mh=mw=128
---
.../fpgadataflow/test_fpgadataflow_fclayer.py | 87 +++++++++++++++++++
1 file changed, 87 insertions(+)
diff --git a/tests/fpgadataflow/test_fpgadataflow_fclayer.py b/tests/fpgadataflow/test_fpgadataflow_fclayer.py
index 80c9e84ba..330d9a1ae 100644
--- a/tests/fpgadataflow/test_fpgadataflow_fclayer.py
+++ b/tests/fpgadataflow/test_fpgadataflow_fclayer.py
@@ -300,3 +300,90 @@ def test_fpgadataflow_fclayer_rtlsim(mem_mode, idt, wdt, act, nf, sf, mw, mh):
hls_synt_res_est = model.analysis(hls_synth_res_estimation)
assert "StreamingFCLayer_Batch_0" in hls_synt_res_est
+
+# mem_mode: const or decoupled
+@pytest.mark.parametrize("mem_mode", ["decoupled"])
+# activation: None or DataType
+@pytest.mark.parametrize("act", [DataType.INT4])
+# weight datatype
+@pytest.mark.parametrize("wdt", [DataType.INT4])
+# input datatype
+@pytest.mark.parametrize("idt", [DataType.INT4])
+# neuron folding, -1 is maximum possible
+@pytest.mark.parametrize("nf", [-1])
+# synapse folding, -1 is maximum possible
+@pytest.mark.parametrize("sf", [-1])
+# HLS matrix width (input features)
+@pytest.mark.parametrize("mw", [128])
+# HLS matrix height (output features)
+@pytest.mark.parametrize("mh", [128])
+def test_fpgadataflow_fclayer_large_depth_decoupled_mode(mem_mode, idt, wdt, act, nf, sf, mw, mh):
+ if nf == -1:
+ nf = mh
+ if sf == -1:
+ sf = mw
+ pe = mh // nf
+ simd = mw // sf
+ assert mh % pe == 0
+ assert mw % sf == 0
+ # generate weights
+ W = gen_finn_dt_tensor(wdt, (mw, mh))
+ # generate input data
+ x = gen_finn_dt_tensor(idt, (1, mw))
+ if act is None:
+ # no activation, produce accumulators
+ T = None
+ tdt = None
+ if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
+ odt = DataType.UINT32
+ else:
+ odt = DataType.INT32
+ else:
+ odt = act
+ (min, max) = calculate_signed_dot_prod_range(idt, wdt, mw)
+ n_steps = act.get_num_possible_values() - 1
+ T = np.random.randint(min, max - 1, (mh, n_steps)).astype(np.float32)
+ # provide non-decreasing thresholds
+ T = np.sort(T, axis=1)
+ # generate thresholds for activation
+ if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
+ tdt = DataType.UINT32
+ # bias thresholds to be positive
+ T = np.ceil((T + mw) / 2)
+ assert (T >= 0).all()
+ else:
+ tdt = DataType.INT32
+ model = make_single_fclayer_modelwrapper(W, pe, simd, wdt, idt, odt, T, tdt)
+ for node in model.graph.node:
+ # lookup op_type in registry of CustomOps
+ inst = getCustomOp(node)
+ inst.set_nodeattr("mem_mode", mem_mode)
+
+ # prepare input data
+ input_dict = prepare_inputs(x, idt, wdt)
+ if wdt == DataType.BIPOLAR and idt == DataType.BIPOLAR:
+ # convert inputs to binary and use xnorpopcountmatmul
+ y = xp.xnorpopcountmatmul((x + 1) / 2, (W + 1) / 2)
+ else:
+ y = np.matmul(x, W)
+ if T is not None:
+ y = multithreshold(y, T)
+ if act == DataType.BIPOLAR:
+ # binary to bipolar
+ y = 2 * y - 1
+ else:
+ # signed offset
+ y += act.min()
+ oshape = model.get_tensor_shape("outp")
+ y_expected = y.reshape(oshape)
+ # TODO split up into several dependent tests -- need to check how this
+ # works for parametrized tests...
+ model = model.transform(SetExecMode("rtlsim"))
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(CodeGen_ipgen("xc7z020clg400-1", 5))
+ model = model.transform(HLSSynth_IPGen())
+ y_produced = oxe.execute_onnx(model, input_dict)["outp"]
+ assert (y_produced.reshape(y_expected.shape) == y_expected).all(), "rtlsim failed"
+
+ hls_synt_res_est = model.analysis(hls_synth_res_estimation)
+ assert "StreamingFCLayer_Batch_0" in hls_synt_res_est
--
GitLab