Merge pull request #649 from Xilinx/feature/decoupled-vvau

Decoupled mode support for VVAU

fetch-repos.sh

@@ -32,7 +32,7 @@ FINN_EXP_COMMIT="9cbd2787b5160e2b44e0e8164a0df1457dbd5366"
 BREVITAS_COMMIT="a5b71d6de1389d3e7db898fef72e014842670f03"
 PYVERILATOR_COMMIT="64b8294ff1afebb47be76fcad6ae87027e0402c2"
 CNPY_COMMIT="4e8810b1a8637695171ed346ce68f6984e585ef4"
-HLSLIB_COMMIT="36e6c8cb1019ba0307e1886011692a58e02f3bfa"
+HLSLIB_COMMIT="bb43a97f799b63f536885919f03ecdfcfb04f405"
 OMX_COMMIT="d1065a788219ca0eb54d5e57600b1f9d7f67d4cc"
 AVNET_BDF_COMMIT="2d49cfc25766f07792c0b314489f21fe916b639b"
 XIL_BDF_COMMIT="8cf4bb674a919ac34e3d99d8d71a9e60af93d14e"

src/finn/transformation/fpgadataflow/convert_to_hls_layers.py

@@ -910,6 +910,10 @@ class InferVectorVectorActivation(Transformation):
     a depthwise convolution. Any immediately following MultiThreshold
     layers will also be absorbed into the VVAU."""
 
+    def __init__(self, mem_mode="const"):
+        super().__init__()
+        self.mem_mode = mem_mode
+
     def apply(self, model):
         graph = model.graph
         node_ind = 0
@@ -1010,6 +1014,7 @@ class InferVectorVectorActivation(Transformation):
                             ActVal=actval,
                             noActivation=0,
                             name="VectorVectorActivation_" + n.name,
+                            mem_mode=self.mem_mode,
                         )
                         graph.node.insert(node_ind, new_node)
                         # remove old nodes

tests/fpgadataflow/test_fpgadataflow_vvau.py

@@ -75,7 +75,19 @@ def _calculate_dot_prod_range(dt_a, dt_b, len):
 
 
 def _make_single_vvau_modelwrapper(
-    W, pe, k_h, k_w, channels, dim_h, dim_w, wdt, idt, odt, T=None, tdt=None
+    W,
+    pe,
+    k_h,
+    k_w,
+    channels,
+    dim_h,
+    dim_w,
+    wdt,
+    idt,
+    odt,
+    T=None,
+    tdt=None,
+    mem_mode="const",
 ):
     in_shape = [1, dim_h, dim_w, k_h * k_w * channels]  # [N, H, W, K*K*CH]
     out_shape = [
@@ -113,6 +125,7 @@ def _make_single_vvau_modelwrapper(
         weightDataType=wdt.name,
         outputDataType=odt.name,
         noActivation=no_act,
+        mem_mode=mem_mode,
     )
 
     graph = helper.make_graph(
@@ -140,7 +153,7 @@ def prepare_inputs(input_tensor):
     return {"inp": input_tensor}
 
 
-# mem_mode: const or decoupled
+# input datatype
 @pytest.mark.parametrize("idt", [DataType["UINT4"], DataType["UINT8"]])
 # weight datatype
 @pytest.mark.parametrize("wdt", [DataType["INT4"]])
@@ -156,13 +169,15 @@ def prepare_inputs(input_tensor):
 @pytest.mark.parametrize("k_w", [3, 1])
 # Number of input and output channels
 @pytest.mark.parametrize("channels", [3, 4])
+# memory mode
+@pytest.mark.parametrize("mem_mode", ["const", "decoupled"])
 # execution mode
 @pytest.mark.parametrize("exec_mode", ["cppsim", "rtlsim"])
 @pytest.mark.fpgadataflow
 @pytest.mark.slow
 @pytest.mark.vivado
 def test_fpgadataflow_vvau(
-    idt, wdt, act, pe, dim_h, dim_w, k_h, k_w, channels, exec_mode
+    idt, wdt, act, pe, dim_h, dim_w, k_h, k_w, channels, mem_mode, exec_mode
 ):
     if pe == "channels":
         pe = channels
@@ -198,7 +213,7 @@ def test_fpgadataflow_vvau(
         tdt = DataType["INT32"]
 
     model = _make_single_vvau_modelwrapper(
-        W, pe, k_h, k_w, channels, dim_h, dim_w, wdt, idt, odt, T, tdt
+        W, pe, k_h, k_w, channels, dim_h, dim_w, wdt, idt, odt, T, tdt, mem_mode
     )
 
     if exec_mode == "cppsim":