The option is controlled through xbar_main.hjson
By default all xbar componentns are marked with
'pipeline' = true
'pipeline_byp' = true
This means all elements (socket1n/socketm1) have both incoming and outgoing
FIFOs that are bypassable if the FIFO is empty
Setting `pipeline_byp` to false makes it a non-bypassable fifo and incurs
extra latency while giving better timing
Setting `pipeline` to false passes through the FIFO entirely and connects
input to output
The current scheme does the following:
remove pipelines for coreI/coreD to minimize process latency
remove pipeliens for ROM / RAM / FLASH to minimize instruction latency
keep pipelines to all other peripherals and memories
diff --git a/util/tlgen/elaborate.py b/util/tlgen/elaborate.py
index 2927fe7..71e95e1 100644
--- a/util/tlgen/elaborate.py
+++ b/util/tlgen/elaborate.py
@@ -82,7 +82,9 @@
new_node = Node(name="sm1_" + str(len(xbar.nodes)),
node_type=NodeType.SOCKET_M1,
clock=xbar.clock)
+ new_node.hdepth = 2
new_node.hpass = 2**len(node.us) - 1
+ new_node.ddepth = 2
new_node.dpass = 1
xbar.insert_node(new_node, node)
process_node(new_node, xbar)
@@ -93,7 +95,9 @@
new_node = Node(name="s1n_" + str(len(xbar.nodes)),
node_type=NodeType.SOCKET_1N,
clock=xbar.clock)
+ new_node.hdepth = 2
new_node.hpass = 1
+ new_node.ddepth = 2
new_node.dpass = 2**len(node.ds) - 1
xbar.insert_node(new_node, node)
@@ -105,42 +109,62 @@
def process_pipeline(xbar):
- """Check if HOST, DEVICE has pipeline key and is True, then propagate it to end
+ """Check if HOST, DEVICE has settings different from default, then propagate it to end
"""
for host in xbar.hosts:
- # go downstream and set the HReqPass at the first instance.
+ # go downstream and change the HReqPass/Depth at the first instance.
# If it is async, skip.
- # If Socket 1N, set hpass to 1 and skip
- # If Socket M1, find position of the host and set 1 of the bit in hpass skip
+ # If Socket 1N,
+ # if pipeline True and bypass false, set hpass to 0
+ # if pipeline is False, set depth to 0
+ # If Socket M1, find position of the host and follow procedure above
# If it is device, it means host and device are directly connected. Ignore now.
# After process node is done, always only one downstream exists in any host node
- if host.pipeline == False:
- # No need to process, default is Pass the req/rsp
+ if host.pipeline == True and host.pipeline_byp == True:
+ # No need to process, same as default
continue
+ no_bypass = (host.pipeline == True and host.pipeline_byp == False)
dnode = host.ds[0].ds
if dnode.node_type == NodeType.SOCKET_1N:
- dnode.hpass = 0
+ dnode.hpass = 0 if no_bypass else dnode.hpass
+
elif dnode.node_type == NodeType.SOCKET_M1:
idx = dnode.us.index(host.ds)
- dnode.hpass = dnode.hpass ^ (1 << idx)
+ dnode.hpass = dnode.hpass ^ (
+ 1 << idx) if no_bypass else dnode.hpass
+
+ # keep variables separate in case we ever need to differentiate
+ dnode.dpass = 0 if no_bypass else dnode.dpass
+ dnode.hdepth = 0 if host.pipeline == False else dnode.hdepth
+ dnode.ddepth = dnode.hdepth
for device in xbar.devices:
# go upstream and set DReq/RspPass at the first instance.
# If it is async, skip
- # If Socket 1N, set dpass to the bit position and skip
- # If Socket M1, set dpass to 1 and skip
+ # If Socket M1
+ # If pipeline True and bypass False, set dpass to 0
+ # If pipeline False, set depth to 0
+ # If Socket 1N, find position of the device and follow procedure above
# If it is host, ignore
- if device.pipeline == False:
+ if device.pipeline == True and device.pipeline_byp == True:
continue
+ no_bypass = (device.pipeline == True and device.pipeline_byp == False)
unode = device.us[0].us
if unode.node_type == NodeType.SOCKET_1N:
idx = unode.ds.index(device.us)
- unode.dpass = unode.dpass ^ (1 << idx)
+ unode.dpass = unode.dpass ^ (
+ 1 << idx) if no_bypass else unode.dpass
+
elif unode.node_type == NodeType.SOCKET_M1:
- unode.dpass = 0
+ unode.dpass = 0 if no_bypass else unode.dpass
+
+ # keep variables separate in case we ever need to differentiate
+ unode.hpass = 0 if no_bypass else unode.hpass
+ unode.ddepth = 0 if device.pipeline == False else unode.ddepth
+ unode.hdepth = unode.ddepth
return xbar
