diff --git a/docs/finn/getting_started.rst b/docs/finn/getting_started.rst
index 95594bb67a2be3a4c3fbba488c75a704f623c136..feba21ea20d6445b88f96cf8ed2ca57598e614f4 100644
--- a/docs/finn/getting_started.rst
+++ b/docs/finn/getting_started.rst
@@ -18,6 +18,7 @@ Requirements
* A working Vivado 2019.1 installation
* A `VIVADO_PATH` environment variable pointing to the Vivado installation directory (e.g. the directory where settings64.sh is located)
* (optional) A PYNQ board with a network connection
+ * the ``bitstring`` package must be installed on the PYNQ: ``sudo pip install bitstring``
Running FINN in Docker
======================
diff --git a/src/finn/custom_op/fpgadataflow/tlastmarker.py b/src/finn/custom_op/fpgadataflow/tlastmarker.py
index 25ea05e3607a52731ae1b64de421837bf137ee2b..17ba44b959577faf573d77ae222f7b2a3be6669d 100644
--- a/src/finn/custom_op/fpgadataflow/tlastmarker.py
+++ b/src/finn/custom_op/fpgadataflow/tlastmarker.py
@@ -30,20 +30,30 @@ from finn.custom_op.fpgadataflow import HLSCustomOp
class TLastMarker(HLSCustomOp):
- """Class that corresponds to the TLastMarker node that needs to be
- inserted at the end of the model for rtlsim with stitched IP.
- It marks the end of the current image/input sample."""
+ """Node that adds/removes AXI stream TLAST signals where needed. Its behavior
+ is transparent in node-by-node execution, only visible in IP-stitched rtlsim or
+ actual hardware.
+ This node may be needed at the end of the network to signal a DMA write (needed by the
+ FINN PYNQ shell) or at the beginning to remove the end-of-burst from DMA read."""
def __init__(self, onnx_node):
super().__init__(onnx_node)
def get_nodeattr_types(self):
my_attrs = {
+ # number of (static) iterations until TLAST=1 is generated for Direction=out
"NumIters": ("i", True, 0),
+ # whether static or dynamic (from AXI lite) number of iterations are used
+ "DynIters": ("i", False, 1),
+ # direction: whether to insert or remove TLAST
+ "Direction": ("s", False, "out"),
# width of input-output data streams, in bits
"StreamWidth": ("i", True, 0),
# width of individual element in stream, in bits
"ElemWidth": ("i", True, 0),
+ # Protocol: external or internal
+ # Vitis docs recommend using qdma_axis for external, ap_axiu for internal
+ "Protocol": ("s", False, "external"),
}
my_attrs.update(super().get_nodeattr_types())
return my_attrs
@@ -76,12 +86,33 @@ class TLastMarker(HLSCustomOp):
def defines(self, var):
stream_width = self.get_nodeattr("StreamWidth")
+ direction = self.get_nodeattr("Direction")
+ protocol = self.get_nodeattr("Protocol")
# output stream must have TLAST, so we use this stream data type:
# qdma_axis<stream_data_width,0,0,0 >
- out_stream_dtype = "qdma_axis<%d,0,0,0>" % stream_width
+ if direction == "out":
+ if protocol == "external":
+ out_stream_dtype = "qdma_axis<%d,0,0,0>" % stream_width
+ elif protocol == "internal":
+ out_stream_dtype = "ap_axiu<%d,0,0,0>" % stream_width
+ else:
+ raise Exception("Unrecognized Protocol in TLastMarker")
+ in_stream_dtype = "ap_uint<%d>" % stream_width
+ elif direction == "in":
+ out_stream_dtype = "ap_uint<%d>" % stream_width
+ if protocol == "external":
+ in_stream_dtype = "qdma_axis<%d,0,0,0>" % stream_width
+ elif protocol == "internal":
+ in_stream_dtype = "ap_axiu<%d,0,0,0>" % stream_width
+ else:
+ raise Exception("Unrecognized Protocol in TLastMarker")
+ else:
+ raise Exception("Unrecognized Direction in TLastMarker")
+
self.code_gen_dict["$DEFINES$"] = [
"#define StreamWidth %d" % stream_width,
"#define OutDType %s" % out_stream_dtype,
+ "#define InDType %s" % in_stream_dtype,
"#define NumItersPerImg %d" % self.get_nodeattr("NumIters"),
]
@@ -89,27 +120,60 @@ class TLastMarker(HLSCustomOp):
self.code_gen_dict["$READNPYDATA$"] = []
def docompute(self):
- self.code_gen_dict["$DOCOMPUTE$"] = [
- "unsigned int n = 1;",
- "OutDType t;",
- "t.set_keep(-1);",
- "io_section: { // start of cycle accurate region",
- "#pragma HLS protocol fixed",
- "// do a first read from stream before we decide on numIters",
- "// giving software a chance to set up the numIters prior to startup",
- "t.set_data(in0.read());",
- "n = (numIters == 0 ? NumItersPerImg : numIters);",
- "t.set_last(n==1);",
- "out.write(t);",
- "} // end of cycle accurate region",
- "// do one less iteration than spec since we already did one",
- "for(unsigned int i=1; i<n; i++) {",
- "#pragma HLS PIPELINE II=1",
- "t.set_data(in0.read());",
- "t.set_last(i==(n-1));",
- "out.write(t);",
- "}",
- ]
+ dyn_iters = self.get_nodeattr("DynIters")
+ direction = self.get_nodeattr("Direction")
+ use_qdma_axis = self.get_nodeattr("Protocol") == "external"
+ if direction == "in":
+ # read from input and just pass data along; ignore tlast
+ # no dyn iters on input, it doesnt make sense
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ "for(unsigned int i=0; i<NumItersPerImg; i++) {",
+ "#pragma HLS PIPELINE II=1",
+ "out.write(in0.read().get_data());"
+ if use_qdma_axis
+ else "out.write(in0.read().data);",
+ "}",
+ ]
+
+ elif dyn_iters == 1:
+ # output, with dynamic iteration counts
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ "unsigned int n = 1;",
+ "OutDType t;",
+ "t.set_keep(-1);" if use_qdma_axis else "t.keep = -1;",
+ "io_section: { // start of cycle accurate region",
+ "#pragma HLS protocol fixed",
+ "// do a first read from stream before we decide on numIters",
+ "// giving software a chance to set up the numIters prior to startup",
+ "t.set_data(in0.read());" if use_qdma_axis else "t.data = in0.read();",
+ "n = (numIters == 0 ? NumItersPerImg : numIters);",
+ "t.set_last(n==1);" if use_qdma_axis else "t.last = (n==1);",
+ "out.write(t);",
+ "} // end of cycle accurate region",
+ "// do one less iteration than spec since we already did one",
+ "for(unsigned int i=1; i<n; i++) {",
+ "#pragma HLS PIPELINE II=1",
+ "t.set_data(in0.read());" if use_qdma_axis else "t.data = in0.read();",
+ "t.set_last(i==(n-1));" if use_qdma_axis else "t.last = (i==(n-1));",
+ "out.write(t);",
+ "}",
+ ]
+
+ else:
+ # output, with static iteration counts
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ "unsigned int n = 1;",
+ "OutDType t;",
+ "t.set_keep(-1);" if use_qdma_axis else "t.keep = -1;",
+ "for(unsigned int i=0; i<NumItersPerImg; i++) {",
+ "#pragma HLS PIPELINE II=1",
+ "t.set_data(in0.read());" if use_qdma_axis else "t.data = in0.read();",
+ "t.set_last(i==(NumItersPerImg-1));"
+ if use_qdma_axis
+ else "t.last = (i==(NumItersPerImg-1));",
+ "out.write(t);",
+ "}",
+ ]
def dataoutstrm(self):
self.code_gen_dict["$DATAOUTSTREAM$"] = []
@@ -118,18 +182,30 @@ class TLastMarker(HLSCustomOp):
self.code_gen_dict["$SAVEASCNPY$"] = []
def blackboxfunction(self):
- self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
- """void %s(hls::stream<ap_uint<StreamWidth> > &in0,
- hls::stream<OutDType> &out, unsigned int numIters)"""
- % self.onnx_node.name
- ]
+ dyn_iters = self.get_nodeattr("DynIters")
+
+ if dyn_iters == 1:
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ """void %s(hls::stream<InDType> &in0,
+ hls::stream<OutDType> &out, unsigned int numIters)"""
+ % self.onnx_node.name
+ ]
+ else:
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ """void %s(hls::stream<InDType> &in0, hls::stream<OutDType> &out)"""
+ % self.onnx_node.name
+ ]
def pragmas(self):
self.code_gen_dict["$PRAGMAS$"] = ["#pragma HLS INTERFACE axis port=in0"]
self.code_gen_dict["$PRAGMAS$"].append("#pragma HLS INTERFACE axis port=out")
- self.code_gen_dict["$PRAGMAS$"].append(
- "#pragma HLS INTERFACE s_axilite port=numIters bundle=control"
- )
+
+ dyn_iters = self.get_nodeattr("DynIters")
+ if dyn_iters == 1:
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE s_axilite port=numIters bundle=control"
+ )
+
self.code_gen_dict["$PRAGMAS$"].append(
"#pragma HLS INTERFACE ap_ctrl_none port=return"
)
@@ -158,7 +234,7 @@ class TLastMarker(HLSCustomOp):
def strm_decl(self):
self.code_gen_dict["$STREAMDECLARATIONS$"] = []
self.code_gen_dict["$STREAMDECLARATIONS$"].append(
- 'hls::stream<ap_uint<{}>> in0 ("in0");'.format(self.get_instream_width())
+ 'hls::stream<InDType> in0 ("in0");'
)
self.code_gen_dict["$STREAMDECLARATIONS$"].append(
'hls::stream<OutDType> out ("out");'
diff --git a/src/finn/transformation/fpgadataflow/insert_tlastmarker.py b/src/finn/transformation/fpgadataflow/insert_tlastmarker.py
index 32f32ece585a93465ba32fede45d5eb606a2b0a3..04dd437af27b9fbe18b2255c20a8e4acda03b3d0 100644
--- a/src/finn/transformation/fpgadataflow/insert_tlastmarker.py
+++ b/src/finn/transformation/fpgadataflow/insert_tlastmarker.py
@@ -31,23 +31,34 @@ from onnx import helper as oh
from finn.custom_op.registry import getCustomOp
from finn.transformation import Transformation
+from finn.util.basic import get_by_name
+
+import numpy as np
class InsertTLastMarker(Transformation):
- """Ensure that the graph is terminated with a TLastMarker node, inserting
- one if necessary."""
+ """Ensure that the graph is started/terminated with a TLastMarker node, inserting
+ one if necessary. Use constructor args to determine type of TLastMarker to be inserted.
+ More information available on the TLastMarker documentation.
+ """
- def __init__(self):
+ def __init__(self, both=False, external=True, dynamic=True):
super().__init__()
+ self.dyniters = dynamic
+ self.external = external
+ self.both = both
def apply(self, model):
# TODO only makes sense for a pure fpgadataflow graph -- check!
graph_out_name = model.graph.output[0].name
final_node = model.find_producer(graph_out_name)
- if final_node.op_type == "TLastMarker":
- # TODO maybe check the correctness of properties
- return (model, False)
- else:
+ graph_modified = False
+ if final_node.op_type != "TLastMarker" and not (
+ final_node.op_type == "IODMA"
+ and get_by_name(final_node.attribute, "direction").s.decode("UTF-8")
+ == "out"
+ ):
+
custom_op = getCustomOp(final_node)
num_iters = int(custom_op.get_number_output_values())
stream_width = int(custom_op.get_outstream_width())
@@ -69,8 +80,51 @@ class InsertTLastMarker(Transformation):
NumIters=num_iters,
StreamWidth=stream_width,
ElemWidth=elem_width,
+ DynIters=(1 if self.dyniters else 0),
+ Direction="out",
+ Protocol=("external" if self.external else "internal"),
domain="finn",
backend="fpgadataflow",
)
model.graph.node.append(tlast_node)
- return (model, True)
+ graph_modified = True
+ # if both is True, also insert marker on input
+ if self.both:
+ graph_in_name = model.graph.input[0].name
+ first_node = model.find_consumer(graph_in_name)
+ if first_node.op_type != "TLastMarker" and not (
+ first_node.op_type == "IODMA"
+ and get_by_name(first_node.attribute, "direction").s.decode("UTF-8")
+ == "in"
+ ):
+
+ custom_op = getCustomOp(first_node)
+ num_iters = np.prod(custom_op.get_folded_input_shape()[1:-1])
+ stream_width = int(custom_op.get_instream_width())
+ in_shape = model.get_tensor_shape(graph_in_name)
+ in_dtype = model.get_tensor_datatype(graph_in_name)
+ elem_width = in_dtype.bitwidth()
+ # make new buffer
+ first_node_in = oh.make_tensor_value_info(
+ model.make_new_valueinfo_name(), TensorProto.FLOAT, in_shape
+ )
+ model.graph.value_info.append(first_node_in)
+ model.set_tensor_datatype(first_node_in.name, in_dtype)
+ # reroute final node output to first_node_in_name
+ first_node.input[0] = first_node_in.name
+ tlast_node = oh.make_node(
+ "TLastMarker",
+ [graph_in_name],
+ [first_node_in.name],
+ NumIters=num_iters,
+ StreamWidth=stream_width,
+ ElemWidth=elem_width,
+ DynIters=(1 if self.dyniters else 0),
+ Direction="in",
+ Protocol=("external" if self.external else "internal"),
+ domain="finn",
+ backend="fpgadataflow",
+ )
+ model.graph.node.insert(0, tlast_node)
+ graph_modified = True
+ return (model, graph_modified)