diff --git a/src/finn/custom_op/fpgadataflow/__init__.py b/src/finn/custom_op/fpgadataflow/__init__.py
index 65c898a8c453420ed96ca22715ef2595c5840288..7de6cce936ee54d58d9a526e926ff79dcd35b90d 100644
--- a/src/finn/custom_op/fpgadataflow/__init__.py
+++ b/src/finn/custom_op/fpgadataflow/__init__.py
@@ -308,6 +308,12 @@ class HLSCustomOp(CustomOp):
f.close()
self.code_gen_dict.clear()
+ def code_generation_ipi(self):
+ """Constructs and returns the TCL for node instantiation in Vivado IPI."""
+ vlnv = self.get_nodeattr("ip_vlnv")
+ cmd = ["create_bd_cell -type ip -vlnv %s %s" % (vlnv, self.onnx_node.name)]
+ return cmd
+
def compile_singlenode_code(self):
"""Builds the bash script for compilation using the CppBuilder from
finn.util.basic and executes the script to produce the executable."""
diff --git a/src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py b/src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py
index f666becdbcceca6ca202907610595f8c0069c5a0..5f1697f819d229d6d7c3b8907abcb541061ecbb3 100644
--- a/src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingdatawidthconverter_batch.py
@@ -51,6 +51,10 @@ class StreamingDataWidthConverter_Batch(HLSCustomOp):
"outWidth": ("i", True, 0),
# FINN DataTypes for inputs/outputs
"dataType": ("s", True, ""),
+ # Toggle between hls or IPI implementation
+ # hls - use the hls generated IP during stitching
+ # vivado - use the AXI Infrastructure DWC
+ "impl_style": ("s", False, "hls"),
}
my_attrs.update(super().get_nodeattr_types())
return my_attrs
@@ -381,3 +385,65 @@ class StreamingDataWidthConverter_Batch(HLSCustomOp):
exp_shape
), """Output
shape doesn't match expected shape, should be same as input shape"""
+
+ def code_generation_ipi(self):
+ impl_style = self.get_nodeattr("impl_style")
+ if impl_style == "hls":
+ return super().code_generation_ipi()
+ elif impl_style == "vivado":
+ cmd = []
+ node_name = self.onnx_node.name
+ # create a hierarchy for this layer, with the same port names
+ clk_name = self.get_verilog_top_module_intf_names()["clk"][0]
+ rst_name = self.get_verilog_top_module_intf_names()["rst"][0]
+ dout_name = self.get_verilog_top_module_intf_names()["m_axis"][0]
+ din_name = self.get_verilog_top_module_intf_names()["s_axis"][0]
+ cmd.append("create_bd_cell -type hier %s" % node_name)
+ cmd.append("create_bd_pin -dir I -type clk /%s/%s" % (node_name, clk_name))
+ cmd.append("create_bd_pin -dir I -type rst /%s/%s" % (node_name, rst_name))
+ cmd.append(
+ "create_bd_intf_pin -mode Master "
+ "-vlnv xilinx.com:interface:axis_rtl:1.0 /%s/%s"
+ % (node_name, dout_name)
+ )
+ cmd.append(
+ "create_bd_intf_pin -mode Slave "
+ "-vlnv xilinx.com:interface:axis_rtl:1.0 /%s/%s" % (node_name, din_name)
+ )
+ # instantiate and configure DWC
+ cmd.append(
+ "create_bd_cell -type ip "
+ "-vlnv xilinx.com:ip:axis_dwidth_converter:1.1 /%s/dwc" % node_name
+ )
+ cmd.append(
+ "set_property -dict "
+ "[list CONFIG.S_TDATA_NUM_BYTES.VALUE_SRC PROPAGATED] "
+ "[get_bd_cells /%s/dwc]" % node_name
+ )
+ cmd.append(
+ "set_property -dict "
+ "[list CONFIG.M_TDATA_NUM_BYTES {%d}] [get_bd_cells /%s/dwc]"
+ % (np.ceil(self.get_outstream_width() / 8), node_name)
+ )
+ cmd.append(
+ "connect_bd_intf_net [get_bd_intf_pins %s/dwc/M_AXIS] "
+ "[get_bd_intf_pins %s/%s]" % (node_name, node_name, dout_name)
+ )
+ cmd.append(
+ "connect_bd_intf_net [get_bd_intf_pins %s/dwc/S_AXIS] "
+ "[get_bd_intf_pins %s/%s]" % (node_name, node_name, din_name)
+ )
+ cmd.append(
+ "connect_bd_net [get_bd_pins %s/%s] [get_bd_pins %s/dwc/aresetn]"
+ % (node_name, rst_name, node_name)
+ )
+ cmd.append(
+ "connect_bd_net [get_bd_pins %s/%s] [get_bd_pins %s/dwc/aclk]"
+ % (node_name, clk_name, node_name)
+ )
+ return cmd
+ else:
+ raise Exception(
+ "DWC implementation style %s not supported, please use hls or vivado"
+ % impl_style
+ )
diff --git a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
index de640f22ae87c89721f8b7ed2b3b270b54000ebb..87f52eeea591ba42bf5374df3c93bcc3e4f8e944 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
@@ -28,9 +28,6 @@
import math
import os
-import subprocess
-from shutil import copy
-
import numpy as np
from onnx import TensorProto, helper
@@ -100,23 +97,6 @@ class StreamingFCLayer_Batch(HLSCustomOp):
my_attrs.update(super().get_nodeattr_types())
return my_attrs
- def get_verilog_top_module_name(self):
- "Return the Verilog top module name for this node."
-
- node = self.onnx_node
- # set top name depending on mem_mode
- mem_mode = self.get_nodeattr("mem_mode")
- if mem_mode == "const" or mem_mode == "external":
- prefixed_top_name = "%s_%s" % (node.name, node.name)
- elif mem_mode == "decoupled":
- prefixed_top_name = "%s_memstream" % (node.name)
- else:
- raise Exception(
- """Please set mem_mode to "const", "decoupled", or "external",
- currently no other parameter value is supported!"""
- )
- return prefixed_top_name
-
def calc_wmem(self):
"""Calculates and returns WMEM."""
mw = self.get_nodeattr("MW")
@@ -705,7 +685,7 @@ class StreamingFCLayer_Batch(HLSCustomOp):
)
super().reset_rtlsim(sim)
super().toggle_clk(sim)
- if mem_mode == "external":
+ if mem_mode == "external" or mem_mode == "decoupled":
wnbits = self.get_weightstream_width()
export_wdt = self.get_weight_datatype()
# we have converted bipolar weights to binary for export,
@@ -1008,112 +988,96 @@ class StreamingFCLayer_Batch(HLSCustomOp):
)
)
- def code_generation_ipgen(self, model, fpgapart, clk):
- # generate code for all mem_mode of MVAU/FCLayer unit
- super().code_generation_ipgen(model, fpgapart, clk)
-
- # if mem_mode = "decoupled" generate code for verilog wrapper
+ def code_generation_ipi(self):
+ cmd = []
+ # add streamer if needed
mem_mode = self.get_nodeattr("mem_mode")
if mem_mode == "decoupled":
- # empty code gen dictionary for new entries
- self.code_gen_dict.clear()
- self.code_gen_dict["$TOPNAME$"] = [
- "{}_memstream".format(self.onnx_node.name)
- ]
- self.code_gen_dict["$LAYER_NAME$"] = [
- "{}_{}".format(self.onnx_node.name, self.onnx_node.name)
- ]
- # make instream width a multiple of 8 for AXI stream interface
- in_width = self.get_instream_width_padded()
- self.code_gen_dict["$IN_RANGE$"] = ["[{}:0]".format(in_width - 1)]
- self.code_gen_dict["$OUT_RANGE$"] = [
- "[{}:0]".format(self.get_outstream_width_padded() - 1)
- ]
- # make weight stream width a multiple of 8 for AXI stream interface
- weight_width = self.get_weightstream_width_padded()
- self.code_gen_dict["$WEIGHT_RANGE$"] = ["[{}:0]".format(weight_width - 1)]
- self.code_gen_dict["$WEIGHT_WIDTH$"] = [str(weight_width)]
- self.code_gen_dict["$WSTREAM_DEPTH$"] = [str(self.calc_wmem())]
- self.code_gen_dict["$MEM_DEPTH$"] = [str(self.calc_wmem())]
- self.code_gen_dict["$RAM_STYLE$"] = [self.get_nodeattr("ram_style")]
-
- template = self.decoupled_wrapper
-
- for key in self.code_gen_dict:
- # transform list into long string separated by '\n'
- code_gen_line = "\n".join(self.code_gen_dict[key])
- template = template.replace(key, code_gen_line)
- code_gen_dir = self.get_nodeattr("code_gen_dir_ipgen")
- f = open(
- os.path.join(
- code_gen_dir, "{}_memstream.v".format(self.onnx_node.name)
- ),
- "w",
+ node_name = self.onnx_node.name
+ # create a hierarchy for this layer, with the same port names
+ clk_name = self.get_verilog_top_module_intf_names()["clk"][0]
+ rst_name = self.get_verilog_top_module_intf_names()["rst"][0]
+ dout_name = self.get_verilog_top_module_intf_names()["m_axis"][0]
+ din_name = self.get_verilog_top_module_intf_names()["s_axis"][0]
+ cmd.append("create_bd_cell -type hier %s" % node_name)
+ cmd.append("create_bd_pin -dir I -type clk /%s/%s" % (node_name, clk_name))
+ cmd.append("create_bd_pin -dir I -type rst /%s/%s" % (node_name, rst_name))
+ cmd.append(
+ "create_bd_intf_pin -mode Master "
+ "-vlnv xilinx.com:interface:axis_rtl:1.0 /%s/%s"
+ % (node_name, dout_name)
)
- f.write(template)
- f.close()
- self.code_gen_dict.clear()
-
- def ipgen_singlenode_code(self):
- # generate ip block of MVAU/FCLayer unit for all mem modes
- super().ipgen_singlenode_code()
-
- mem_mode = self.get_nodeattr("mem_mode")
- if mem_mode == "decoupled":
- # copy necessary verilog and .dat files
- # into verilog folder in code generation folder
- code_gen_dir = self.get_nodeattr("code_gen_dir_ipgen")
- verilog_folder = "{}/project_{}/sol1/impl/verilog/".format(
- code_gen_dir, self.onnx_node.name
+ cmd.append(
+ "create_bd_intf_pin -mode Slave "
+ "-vlnv xilinx.com:interface:axis_rtl:1.0 /%s/%s" % (node_name, din_name)
)
- # copy memstream components from finn-rtllib
- memstream_dir = "/workspace/finn/finn-rtllib/memstream/hdl/"
- for file in os.listdir(memstream_dir):
- if file.endswith(".v"):
- verilog_file = os.path.join(memstream_dir, file)
- copy(verilog_file, verilog_folder)
- # copy .dat files of weights
- for file in os.listdir(code_gen_dir):
- if file.endswith(".dat"):
- dat_file = os.path.join(code_gen_dir, file)
- copy(dat_file, verilog_folder)
- # copy verilog wrapper
- verilog_wrapper = "{}/{}_memstream.v".format(
- code_gen_dir, self.onnx_node.name
+ # instantiate the hls ip
+ cmd.append(
+ "create_bd_cell -type ip -vlnv %s /%s/%s"
+ % (self.get_nodeattr("ip_vlnv"), node_name, node_name)
)
- copy(verilog_wrapper, verilog_folder)
- # prepare the IP packaging tcl template
- template = templates.ip_package_tcl
- self.code_gen_dict["$TOPNAME$"] = [
- "{}_memstream".format(self.onnx_node.name)
- ]
- self.code_gen_dict["$VERILOG_DIR$"] = [verilog_folder]
- for key in self.code_gen_dict:
- # transform list into long string separated by '\n'
- code_gen_line = "\n".join(self.code_gen_dict[key])
- template = template.replace(key, code_gen_line)
- code_gen_dir = self.get_nodeattr("code_gen_dir_ipgen")
- f = open(os.path.join(verilog_folder, "package_ip.tcl"), "w")
- f.write(template)
- f.close()
- # create a shell script and call Vivado to invoke the IP pkg script
- make_project_sh = verilog_folder + "/make_ip.sh"
- working_dir = os.environ["PWD"]
- with open(make_project_sh, "w") as f:
- f.write("#!/bin/bash \n")
- f.write("cd {}\n".format(verilog_folder))
- f.write("vivado -mode batch -source package_ip.tcl\n")
- f.write("cd {}\n".format(working_dir))
- bash_command = ["bash", make_project_sh]
- process_compile = subprocess.Popen(bash_command, stdout=subprocess.PIPE)
- process_compile.communicate()
- # re-set ip_path to point to the new packaged IP
- self.set_nodeattr("ip_path", verilog_folder)
- vlnv = "xilinx.com:hls:%s:1.0" % (
- "{}_memstream".format(self.onnx_node.name)
+ # instantiate a streamer and connect it to the HLS IP
+ strm_vlnv = "xilinx.com:user:memstream:1.0"
+ strm_inst = node_name + "_wstrm"
+ cmd.append(
+ "create_bd_cell -type ip -vlnv %s /%s/%s"
+ % (strm_vlnv, node_name, strm_inst)
+ )
+ cmd.append(
+ "set_property -dict [list "
+ "CONFIG.NSTREAMS {1} "
+ "CONFIG.MEM_DEPTH {%d} "
+ "CONFIG.MEM_WIDTH {%d} "
+ "CONFIG.MEM_INIT {%s} "
+ "CONFIG.RAM_STYLE {%s} "
+ "CONFIG.STRM0_DEPTH {%d} "
+ "CONFIG.STRM0_WIDTH {%d} "
+ "CONFIG.STRM0_OFFSET {0} "
+ "] [get_bd_cells /%s/%s]"
+ % (
+ self.calc_wmem(),
+ self.get_weightstream_width_padded(),
+ self.get_nodeattr("code_gen_dir_ipgen") + "/",
+ self.get_nodeattr("ram_style"),
+ self.calc_wmem(),
+ self.get_weightstream_width_padded(),
+ node_name,
+ strm_inst,
+ )
+ )
+ cmd.append(
+ "connect_bd_intf_net [get_bd_intf_pins %s/%s/m_axis_0] "
+ "[get_bd_intf_pins %s/%s/weights_V_V]"
+ % (node_name, strm_inst, node_name, node_name)
+ )
+ cmd.append(
+ "connect_bd_net [get_bd_pins %s/%s] [get_bd_pins %s/%s/aresetn]"
+ % (node_name, rst_name, node_name, strm_inst)
+ )
+ cmd.append(
+ "connect_bd_net [get_bd_pins %s/%s] [get_bd_pins %s/%s/aclk]"
+ % (node_name, clk_name, node_name, strm_inst)
+ )
+ cmd.append(
+ "connect_bd_net [get_bd_pins %s/%s] [get_bd_pins %s/%s/%s]"
+ % (node_name, rst_name, node_name, node_name, rst_name)
+ )
+ cmd.append(
+ "connect_bd_net [get_bd_pins %s/%s] [get_bd_pins %s/%s/%s]"
+ % (node_name, clk_name, node_name, node_name, clk_name)
+ )
+ cmd.append(
+ "connect_bd_intf_net [get_bd_intf_pins %s/%s] "
+ "[get_bd_intf_pins %s/%s/%s]"
+ % (node_name, din_name, node_name, node_name, din_name)
+ )
+ cmd.append(
+ "connect_bd_intf_net [get_bd_intf_pins %s/%s] "
+ "[get_bd_intf_pins %s/%s/%s]"
+ % (node_name, dout_name, node_name, node_name, dout_name)
)
- self.set_nodeattr("ip_vlnv", vlnv)
- self.code_gen_dict.clear()
+ cmd.append("save_bd_design")
+ return cmd
def get_verilog_top_module_intf_names(self):
intf_names = super().get_verilog_top_module_intf_names()
diff --git a/src/finn/custom_op/fpgadataflow/streamingfifo.py b/src/finn/custom_op/fpgadataflow/streamingfifo.py
index 1f734b548f923341687843c538d1887fcc069bee..c9011b50d06a55c34bdd49c8ea374bdf81ea5f4f 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfifo.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfifo.py
@@ -51,6 +51,16 @@ class StreamingFIFO(HLSCustomOp):
"folded_shape": ("ints", True, []),
# FINN DataTypes for inputs/outputs
"dataType": ("s", True, ""),
+ # Toggle between hls or IPI implementation
+ # rtl - use the hls generated IP during stitching
+ # vivado - use the AXI Infrastructure FIFO
+ "impl_style": ("s", False, "rtl"),
+ # FPGA resource type for FIFOs when impl_style is vivado
+ # auto -- let Vivado decide
+ # block -- use BRAM
+ # distributed -- use LUTRAM
+ # ultra -- use URAM (on UltraScale+)
+ "ram_style": ("s", False, "auto"),
}
my_attrs.update(super().get_nodeattr_types())
@@ -306,3 +316,71 @@ class StreamingFIFO(HLSCustomOp):
def pragmas(self):
pass
+
+ def code_generation_ipi(self):
+ impl_style = self.get_nodeattr("impl_style")
+ if impl_style == "rtl":
+ return super().code_generation_ipi()
+ elif impl_style == "vivado":
+ cmd = []
+ node_name = self.onnx_node.name
+ depth = self.get_nodeattr("depth")
+ ram_style = self.get_nodeattr("ram_style")
+ # create a hierarchy for this layer, with the same port names
+ clk_name = self.get_verilog_top_module_intf_names()["clk"][0]
+ rst_name = self.get_verilog_top_module_intf_names()["rst"][0]
+ dout_name = self.get_verilog_top_module_intf_names()["m_axis"][0]
+ din_name = self.get_verilog_top_module_intf_names()["s_axis"][0]
+ cmd.append("create_bd_cell -type hier %s" % node_name)
+ cmd.append("create_bd_pin -dir I -type clk /%s/%s" % (node_name, clk_name))
+ cmd.append("create_bd_pin -dir I -type rst /%s/%s" % (node_name, rst_name))
+ cmd.append(
+ "create_bd_intf_pin -mode Master "
+ "-vlnv xilinx.com:interface:axis_rtl:1.0 /%s/%s"
+ % (node_name, dout_name)
+ )
+ cmd.append(
+ "create_bd_intf_pin -mode Slave "
+ "-vlnv xilinx.com:interface:axis_rtl:1.0 /%s/%s" % (node_name, din_name)
+ )
+ # instantiate and configure DWC
+ cmd.append(
+ "create_bd_cell -type ip "
+ "-vlnv xilinx.com:ip:axis_data_fifo:2.0 /%s/fifo" % node_name
+ )
+ cmd.append(
+ "set_property -dict [list CONFIG.FIFO_DEPTH {%d}] "
+ "[get_bd_cells /%s/fifo]" % (depth, node_name)
+ )
+ cmd.append(
+ "set_property -dict [list CONFIG.FIFO_MEMORY_TYPE {%s}] "
+ "[get_bd_cells /%s/fifo]" % (ram_style, node_name)
+ )
+ cmd.append(
+ "set_property -dict [list CONFIG.TDATA_NUM_BYTES {%d}] "
+ "[get_bd_cells /%s/fifo]"
+ % (np.ceil(self.get_outstream_width() / 8), node_name)
+ )
+ cmd.append(
+ "connect_bd_intf_net [get_bd_intf_pins %s/fifo/M_AXIS] "
+ "[get_bd_intf_pins %s/%s]" % (node_name, node_name, dout_name)
+ )
+ cmd.append(
+ "connect_bd_intf_net [get_bd_intf_pins %s/fifo/S_AXIS] "
+ "[get_bd_intf_pins %s/%s]" % (node_name, node_name, din_name)
+ )
+ cmd.append(
+ "connect_bd_net [get_bd_pins %s/%s] "
+ "[get_bd_pins %s/fifo/s_axis_aresetn]"
+ % (node_name, rst_name, node_name)
+ )
+ cmd.append(
+ "connect_bd_net [get_bd_pins %s/%s] "
+ "[get_bd_pins %s/fifo/s_axis_aclk]" % (node_name, clk_name, node_name)
+ )
+ return cmd
+ else:
+ raise Exception(
+ "FIFO implementation style %s not supported, please use rtl or vivado"
+ % impl_style
+ )
diff --git a/src/finn/transformation/fpgadataflow/create_stitched_ip.py b/src/finn/transformation/fpgadataflow/create_stitched_ip.py
index 90b4b6c47e6e353c1b606d6918eb271e9c0619c5..6e9ce35634760c06c7a409fb5befdb94a08e9c7d 100644
--- a/src/finn/transformation/fpgadataflow/create_stitched_ip.py
+++ b/src/finn/transformation/fpgadataflow/create_stitched_ip.py
@@ -182,6 +182,8 @@ class CreateStitchedIP(Transformation):
def apply(self, model):
ip_dirs = ["list"]
+ # add RTL streamer IP
+ ip_dirs.append("/workspace/finn/finn-rtllib/memstream")
# ensure that all nodes are fpgadataflow, and that IPs are generated
for node in model.graph.node:
assert node.domain == "finn", 'Node domain is not set to "finn"'
@@ -196,10 +198,7 @@ class CreateStitchedIP(Transformation):
ip_dir_value = node_inst.get_nodeattr("ip_path")
assert os.path.isdir(ip_dir_value), "IP generation directory doesn't exist."
ip_dirs += [ip_dir_value]
- vlnv = node_inst.get_nodeattr("ip_vlnv")
- inst_name = node.name
- create_cmd = "create_bd_cell -type ip -vlnv %s %s" % (vlnv, inst_name)
- self.create_cmds += [create_cmd]
+ self.create_cmds += node_inst.code_generation_ipi()
my_producer = model.find_producer(node.input[0])
self.connect_clk_rst(node)
self.connect_axi(node)
@@ -223,6 +222,7 @@ class CreateStitchedIP(Transformation):
# find index of producer output connected to our target input
# get names of hdl interfaces for input and producer output
# issue a TCL directive to connect input to output
+ # if FC layer with mode "decoupled", add a streamer on input 1
for i in range(len(node.input)):
producer = model.find_producer(node.input[i])
if producer is None: