Merge branch 'dev' into feature/pyxsi_integration

docker/Dockerfile.finn

@@ -124,6 +124,7 @@ RUN pip install pytest-metadata==1.7.0
 RUN pip install pytest-html==3.0.0
 RUN pip install pytest-html-merger==0.0.8
 RUN pip install pytest-cov==4.1.0
+RUN pip install pyyaml==6.0.1
 
 # extra dependencies from other FINN deps
 # installed in Docker image to make entrypoint script go faster

run-docker.sh

@@ -88,7 +88,7 @@ SCRIPTPATH=$(dirname "$SCRIPT")
 : ${PLATFORM_REPO_PATHS="/opt/xilinx/platforms"}
 : ${XRT_DEB_VERSION="xrt_202220.2.14.354_22.04-amd64-xrt"}
 : ${FINN_HOST_BUILD_DIR="/tmp/$DOCKER_INST_NAME"}
-: ${FINN_DOCKER_TAG="xilinx/finn:$(git describe --always --tags --dirty).$XRT_DEB_VERSION"}
+: ${FINN_DOCKER_TAG="xilinx/finn:$(OLD_PWD=$(pwd); cd $SCRIPTPATH; git describe --always --tags --dirty; cd $OLD_PWD).$XRT_DEB_VERSION"}
 : ${FINN_DOCKER_PREBUILT="0"}
 : ${FINN_DOCKER_RUN_AS_ROOT="0"}
 : ${FINN_DOCKER_GPU="$(docker info | grep nvidia | wc -m)"}

src/finn/custom_op/fpgadataflow/hls/lookup_hls.py

@@ -28,6 +28,7 @@
 
 import numpy as np
 import os
+import warnings
 from math import ceil, log2
 from qonnx.core.datatype import DataType
 
@@ -87,31 +88,6 @@ def defines(self, var):
             my_defines.append("#define EmbeddingType %s" % emb_hls_type)
         self.code_gen_dict["$DEFINES$"] = my_defines
 
-    def read_npy_data(self):
-        code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
-        dtype = self.get_input_datatype()
-        if dtype == DataType["BIPOLAR"]:
-            # use binary for bipolar storage
-            dtype = DataType["BINARY"]
-        elem_bits = dtype.bitwidth()
-        packed_bits = self.get_instream_width()
-        packed_hls_type = "ap_uint<%d>" % packed_bits
-        elem_hls_type = dtype.get_hls_datatype_str()
-        npy_type = "int64_t"
-        npy_in = "%s/input_0.npy" % code_gen_dir
-        self.code_gen_dict["$READNPYDATA$"] = []
-        self.code_gen_dict["$READNPYDATA$"].append(
-            'npy2apintstream<%s, %s, %d, %s>("%s", in0_%s);'
-            % (
-                packed_hls_type,
-                elem_hls_type,
-                elem_bits,
-                npy_type,
-                npy_in,
-                self.hls_sname(),
-            )
-        )
-
     def dataoutstrm(self):
         code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
         dtype = self.get_output_datatype()
@@ -273,7 +249,18 @@ def execute_node(self, context, graph):
             )
 
         inp = context[node.input[0]]
-        assert inp.dtype == np.int64, "Inputs must be contained in int64 ndarray"
+
+        # Make sure the input has the right container datatype
+        if inp.dtype is not np.float32:
+            # Issue a warning to make the user aware of this type-cast
+            warnings.warn(
+                f"{node.name}: Changing input container datatype from "
+                f"{inp.dtype} to {np.float32}"
+            )
+            # Convert the input to floating point representation as the
+            # container datatype
+            inp = inp.astype(np.float32)
+
         assert inp.shape == exp_ishape, """Input shape doesn't match expected shape."""
         export_idt = self.get_input_datatype()
         odt = self.get_output_datatype()

src/finn/custom_op/fpgadataflow/rtl/streamingfifo_rtl.py

@@ -127,10 +127,18 @@ def execute_node(self, context, graph):
         elif mode == "rtlsim":
             code_gen_dir = self.get_nodeattr("code_gen_dir_ipgen")
             # create a npy file for the input of the node
-            assert (
-                str(inp.dtype) == "float32"
-            ), """Input datatype is
-                not float32 as expected."""
+
+            # Make sure the input has the right container datatype
+            if inp.dtype is not np.float32:
+                # Issue a warning to make the user aware of this type-cast
+                warnings.warn(
+                    f"{node.name}: Changing input container datatype from "
+                    f"{inp.dtype} to {np.float32}"
+                )
+                # Convert the input to floating point representation as the
+                # container datatype
+                inp = inp.astype(np.float32)
+
             expected_inp_shape = self.get_folded_input_shape()
             reshaped_input = inp.reshape(expected_inp_shape)
             if DataType[self.get_nodeattr("dataType")] == DataType["BIPOLAR"]:

tests/brevitas/test_brevitas_avg_pool_export.py

@@ -41,8 +41,7 @@
 
 import finn.core.onnx_exec as oxe
 from finn.transformation.qonnx.convert_qonnx_to_finn import ConvertQONNXtoFINN
-
-base_export_onnx_path = "test_brevitas_avg_pool_export.onnx"
+from finn.util.basic import make_build_dir
 
 
 @pytest.mark.brevitas_export
@@ -62,7 +61,8 @@ def test_brevitas_avg_pool_export(
     channels,
     idim,
 ):
-    export_onnx_path = base_export_onnx_path.replace(".onnx", "test_QONNX.onnx")
+    build_dir = make_build_dir(prefix="test_brevitas_avg_pool_export")
+    export_onnx_path = os.path.join(build_dir, "test.onnx")
     if signed:
         quant_node = QuantIdentity(
             bit_width=input_bit_width,
@@ -112,5 +112,3 @@ def test_brevitas_avg_pool_export(
     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)

tests/brevitas/test_brevitas_cnv.py

@@ -41,17 +41,18 @@
 
 import finn.core.onnx_exec as oxe
 from finn.transformation.qonnx.convert_qonnx_to_finn import ConvertQONNXtoFINN
+from finn.util.basic import make_build_dir
 from finn.util.test import get_test_model_trained
 
-export_onnx_path = "test_brevitas_cnv.onnx"
-
 
 @pytest.mark.brevitas_export
 @pytest.mark.parametrize("abits", [1, 2])
 @pytest.mark.parametrize("wbits", [1, 2])
 def test_brevitas_cnv_export_exec(wbits, abits):
     if wbits > abits:
         pytest.skip("No wbits > abits cases at the moment")
+    build_dir = make_build_dir("test_brevitas_cnv_export_exec")
+    export_onnx_path = os.path.join(build_dir, "test_brevitas_cnv.onnx")
     cnv = get_test_model_trained("CNV", wbits, abits)
     ishape = (1, 3, 32, 32)
     export_qonnx(cnv, torch.randn(ishape), export_onnx_path)
@@ -78,4 +79,3 @@ def test_brevitas_cnv_export_exec(wbits, abits):
     expected = cnv.forward(input_tensor).detach().numpy()
     assert np.isclose(produced, expected, atol=1e-3).all()
     assert np.argmax(produced) == 3
-    os.remove(export_onnx_path)

tests/brevitas/test_brevitas_debug.py

@@ -42,13 +42,15 @@
 
 import finn.core.onnx_exec as oxe
 from finn.transformation.qonnx.convert_qonnx_to_finn import ConvertQONNXtoFINN
+from finn.util.basic import make_build_dir
 from finn.util.test import get_test_model_trained
 
 
 @pytest.mark.brevitas_export
 @pytest.mark.parametrize("QONNX_FINN_conversion", [False, True])
 def test_brevitas_debug(QONNX_FINN_conversion):
-    finn_onnx = "test_brevitas_debug.onnx"
+    build_dir = make_build_dir("test_brevitas_debug")
+    finn_onnx = os.path.join(build_dir, "test_brevitas_debug.onnx")
     fc = get_test_model_trained("TFC", 2, 2)
     ishape = (1, 1, 28, 28)
     dbg_hook = bo.enable_debug(fc, proxy_level=True)
@@ -94,4 +96,3 @@ def test_brevitas_debug(QONNX_FINN_conversion):
         tensor_pytorch = _unpack_quant_tensor(dbg_hook.values[dbg_name]).detach().numpy()
         tensor_finn = output_dict[dbg_name]
         assert np.isclose(tensor_finn, tensor_pytorch, atol=1e-5).all()
-    os.remove(finn_onnx)

tests/brevitas/test_brevitas_deconv.py

@@ -38,8 +38,7 @@
 
 import finn.core.onnx_exec as oxe
 from finn.transformation.qonnx.convert_qonnx_to_finn import ConvertQONNXtoFINN
-
-export_path = "test_brevitas_deconv.onnx"
+from finn.util.basic import make_build_dir
 
 
 @pytest.mark.brevitas_export
@@ -66,6 +65,8 @@ def test_brevitas_QTransposeConv(ifm_ch, ofm_ch, mh, mw, padding, stride, kw, bi
         padding=padding,
         bias=bias,
     )
+    build_dir = make_build_dir("test_brevitas_QTransposeConv")
+    export_path = os.path.join(build_dir, "test_brevitas_deconv.onnx")
     # outp = el(inp) # expects NCHW data format
     export_qonnx(b_deconv, input_t=inp, export_path=export_path, opset_version=11)
     qonnx_cleanup(export_path, out_file=export_path)
@@ -79,4 +80,3 @@ def test_brevitas_QTransposeConv(ifm_ch, ofm_ch, mh, mw, padding, stride, kw, bi
     inp_tensor = torch.from_numpy(inp_tensor).float()
     expected = b_deconv.forward(inp_tensor).detach().numpy()
     assert np.isclose(produced, expected, atol=1e-3).all()
-    os.remove(export_path)

tests/brevitas/test_brevitas_non_scaled_quanthardtanh_export.py

@@ -43,8 +43,7 @@
 
 import finn.core.onnx_exec as oxe
 from finn.transformation.qonnx.convert_qonnx_to_finn import ConvertQONNXtoFINN
-
-export_onnx_path = "test_brevitas_non_scaled_QuantHardTanh_export.onnx"
+from finn.util.basic import make_build_dir
 
 
 @pytest.mark.brevitas_export
@@ -72,7 +71,8 @@ def get_quant_type(bit_width):
         scaling_impl_type=ScalingImplType.CONST,
         narrow_range=narrow_range,
     )
-    m_path = export_onnx_path
+    build_dir = make_build_dir(prefix="test_brevitas_act_export_qhardtanh_nonscaled")
+    m_path = os.path.join(build_dir, "test_brevitas_non_scaled_QuantHardTanh_export.onnx")
     export_qonnx(b_act, torch.randn(ishape), m_path)
     qonnx_cleanup(m_path, out_file=m_path)
     model = ModelWrapper(m_path)
@@ -85,4 +85,3 @@ def get_quant_type(bit_width):
     inp_tensor = torch.from_numpy(inp_tensor).float()
     expected = b_act.forward(inp_tensor).detach().numpy()
     assert np.isclose(produced, expected, atol=1e-3).all()
-    os.remove(export_onnx_path)

tests/brevitas/test_brevitas_qconv2d.py

@@ -45,8 +45,7 @@
 
 import finn.core.onnx_exec as oxe
 from finn.transformation.qonnx.convert_qonnx_to_finn import ConvertQONNXtoFINN
-
-export_onnx_path = "test_brevitas_conv.onnx"
+from finn.util.basic import make_build_dir
 
 
 @pytest.mark.brevitas_export
@@ -92,7 +91,8 @@ def test_brevitas_QConv2d(dw, bias, in_channels):
     weight_tensor = gen_finn_dt_tensor(DataType["INT4"], w_shape)
     b_conv.weight = torch.nn.Parameter(torch.from_numpy(weight_tensor).float())
     b_conv.eval()
-    m_path = export_onnx_path
+    build_dir = make_build_dir(prefix="test_brevitas_QConv2d")
+    m_path = os.path.join(build_dir, "test_brevitas_conv.onnx")
     export_qonnx(b_conv, torch.randn(ishape), m_path)
     qonnx_cleanup(m_path, out_file=m_path)
     model = ModelWrapper(m_path)
@@ -106,4 +106,3 @@ def test_brevitas_QConv2d(dw, bias, in_channels):
     expected = b_conv.forward(inp_tensor).detach().numpy()
 
     assert np.isclose(produced, expected, atol=1e-3).all()
-    os.remove(export_onnx_path)

tests/brevitas/test_brevitas_qlinear.py

@@ -42,8 +42,7 @@
 
 import finn.core.onnx_exec as oxe
 from finn.transformation.qonnx.convert_qonnx_to_finn import ConvertQONNXtoFINN
-
-export_onnx_path = "test_brevitas_qlinear.onnx"
+from finn.util.basic import make_build_dir
 
 
 @pytest.mark.brevitas_export
@@ -67,7 +66,8 @@ def test_brevitas_qlinear(bias, out_features, in_features, w_bits, i_dtype):
     weight_tensor_fp = np.random.uniform(low=-1.0, high=1.0, size=w_shape).astype(np.float32)
     b_linear.weight.data = torch.from_numpy(weight_tensor_fp)
     b_linear.eval()
-    m_path = export_onnx_path
+    build_dir = make_build_dir(prefix="test_brevitas_qlinear")
+    m_path = os.path.join(build_dir, "test_brevitas_qlinear.onnx")
     export_qonnx(b_linear, torch.randn(i_shape), m_path)
     qonnx_cleanup(m_path, out_file=m_path)
     model = ModelWrapper(m_path)
@@ -81,4 +81,3 @@ def test_brevitas_qlinear(bias, out_features, in_features, w_bits, i_dtype):
     expected = b_linear.forward(inp_tensor).detach().numpy()
 
     assert np.isclose(produced, expected, atol=1e-3).all()
-    os.remove(export_onnx_path)

tests/brevitas/test_brevitas_relu_act_export.py

@@ -41,8 +41,7 @@
 
 import finn.core.onnx_exec as oxe
 from finn.transformation.qonnx.convert_qonnx_to_finn import ConvertQONNXtoFINN
-
-export_onnx_path = "test_brevitas_relu_act_export.onnx"
+from finn.util.basic import make_build_dir
 
 
 @pytest.mark.brevitas_export
@@ -55,7 +54,8 @@ def test_brevitas_act_export_relu(
     b_act = QuantReLU(
         bit_width=abits,
     )
-    m_path = export_onnx_path
+    build_dir = make_build_dir(prefix="test_brevitas_act_export_relu")
+    m_path = os.path.join(build_dir, "test_brevitas_relu_act_export.onnx")
     export_qonnx(b_act, torch.randn(ishape), m_path)
     qonnx_cleanup(m_path, out_file=m_path)
     model = ModelWrapper(m_path)
@@ -70,7 +70,6 @@ def test_brevitas_act_export_relu(
     expected = b_act.forward(inp_tensor).detach().numpy()
 
     assert np.isclose(produced, expected, atol=1e-3).all()
-    os.remove(export_onnx_path)
 
 
 @pytest.mark.brevitas_export
@@ -88,7 +87,8 @@ def test_brevitas_act_export_relu_channel(
         scaling_per_output_channel=True,
         per_channel_broadcastable_shape=(1, ch, 1, 1),
     )
-    m_path = export_onnx_path
+    build_dir = make_build_dir(prefix="test_brevitas_act_export_relu_channel")
+    m_path = os.path.join(build_dir, "test_brevitas_relu_act_export.onnx")
     export_qonnx(b_act, torch.randn(ishape), m_path)
     qonnx_cleanup(m_path, out_file=m_path)
     model = ModelWrapper(m_path)
@@ -103,4 +103,3 @@ def test_brevitas_act_export_relu_channel(
     expected = b_act.forward(inp_tensor).detach().numpy()
 
     assert np.isclose(produced, expected, atol=1e-3).all()
-    os.remove(export_onnx_path)

tests/brevitas/test_brevitas_scaled_qhardtanh_export.py

@@ -43,8 +43,7 @@
 
 import finn.core.onnx_exec as oxe
 from finn.transformation.qonnx.convert_qonnx_to_finn import ConvertQONNXtoFINN
-
-export_onnx_path = "test_brevitas_scaled_QHardTanh_export.onnx"
+from finn.util.basic import make_build_dir
 
 
 @pytest.mark.brevitas_export
@@ -85,7 +84,8 @@ def get_quant_type(bit_width):
             )
         }
         b_act.load_state_dict(checkpoint)
-    m_path = export_onnx_path
+    build_dir = make_build_dir(prefix="test_brevitas_act_export_qhardtanh_scaled")
+    m_path = os.path.join(build_dir, "test_brevitas_scaled_QHardTanh_export.onnx")
     export_qonnx(b_act, torch.randn(ishape), m_path)
     qonnx_cleanup(m_path, out_file=m_path)
     model = ModelWrapper(m_path)
@@ -121,4 +121,3 @@ def get_quant_type(bit_width):
         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)

tests/brevitas/test_brevitas_selu_act_export.py

@@ -40,14 +40,16 @@
 
 import finn.core.onnx_exec as oxe
 from finn.transformation.qonnx.convert_qonnx_to_finn import ConvertQONNXtoFINN
+from finn.util.basic import make_build_dir
 
 
 @pytest.mark.brevitas_export
 @pytest.mark.parametrize("abits", [2, 4, 8])
 @pytest.mark.parametrize("ishape", [(1, 15), (1, 32, 1, 1)])
 @pytest.mark.parametrize("narrow", [True, False])
 def test_brevitas_act_export_selu(abits, ishape, narrow):
-    export_path = "test_brevitas_selu_act_export_%s.onnx" % str(abits)
+    build_dir = make_build_dir(prefix="test_brevitas_act_export_selu")
+    export_path = os.path.join(build_dir, "test_brevitas_selu_act_export_%s.onnx" % str(abits))
     b_act = torch.nn.Sequential(torch.nn.SELU(), QuantIdentity(bit_width=abits, narrow=narrow))
 
     export_qonnx(
@@ -69,4 +71,3 @@ def test_brevitas_act_export_selu(abits, ishape, narrow):
     expected = b_act.forward(inp_tensor).detach().numpy()
 
     assert np.isclose(produced, expected, atol=1e-3).all()
-    os.remove(export_path)