From: Sebastien Bourdeauducq <sebastien@milkymist.org>
Date: Mon, 13 Feb 2012 15:29:38 +0000 (+0100)
Subject: bus: Wishbone to ASMI caching bridge (untested)
X-Git-Tag: 24jan2021_ls180~2099^2~1028
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=cad9d3b9600022b1c4431e6d7fe50bc89274a32f;p=litex.git

bus: Wishbone to ASMI caching bridge (untested)
---

diff --git a/migen/bus/wishbone2asmi.py b/migen/bus/wishbone2asmi.py
new file mode 100644
index 00000000..5a9875b3
--- /dev/null
+++ b/migen/bus/wishbone2asmi.py
@@ -0,0 +1,147 @@
+from migen.bus import wishbone
+from migen.fhdl.structure import *
+from migen.corelogic.fsm import FSM
+from migen.corelogic.misc import split, displacer, chooser
+from migen.corelogic.record import Record
+
+def _log2_int(n):
+	l = 1
+	r = 0
+	while l < n:
+		l *= 2
+		r += 1
+	if l == n:
+		return r
+	else:
+		raise ValueError("Not a power of 2")
+
+# cachesize (in 32-bit words) is the size of the data store, must be a power of 2
+class WB2ASMI:
+	def __init__(self, cachesize, asmiport):
+		self.wishbone = wishbone.Slave()
+		self.cachesize = cachesize
+		self.asmiport = asmiport
+		if len(self.asmiport.slots) != 1:
+			raise ValueError("ASMI port must have 1 slot")
+		if self.asmiport.hub.dw <= 32:
+			raise ValueError("ASMI data width must be strictly larger than 32")
+		if (self.asmiport.hub.dw % 32) != 0:
+			raise ValueError("ASMI data width must be a multiple of 32")
+
+	def get_fragment(self):
+		comb = []
+		sync = []
+		
+		aaw = self.asmiport.hub.aw
+		adw = self.asmiport.hub.dw
+		
+		# Split address:
+		# TAG | LINE NUMBER | LINE OFFSET
+		offsetbits = _log2_int(adw//32)
+		addressbits = aaw + offsetbits
+		linebits = _log2_int(self.cachesize) - offsetbits
+		tagbits = aaw - linebits
+		adr_offset, adr_line, adr_tag = split(self.wishbone.adr_i, offsetbits, linebits, tagbits)
+		
+		# Data memory
+		data_adr = Signal(BV(linebits))
+		data_do = Signal(BV(adw))
+		data_di = Signal(BV(adw))
+		data_we = Signal(BV(adw//8))
+		data_port = MemoryPort(data_adr, data_do, data_we, data_di, we_granularity=8)
+		data_mem = Memory(adw, 2**linebits, data_port)
+		
+		write_from_asmi = Signal()
+		adr_offset_r = Signal(BV(offsetbits))
+		comb += [
+			data_adr.eq(adr_line),
+			If(write_from_asmi,
+				data_di.eq(self.asmiport.dat_r),
+				data_we.eq(Replicate(1, adw//8))
+			).Else(
+				data_di.eq(Replicate(self.wishbone.dat_i, adw//32),
+				If(self.wishbone.cyc_i & self.wishbone.stb_i & self.wishbone.ack_o,
+					displacer(self.wishbone.we_i, adr_offset, data_we)
+				)
+			),
+			self.asmiport.dat_w.eq(data_do),
+			chooser(data_do, adr_offset_r, self.wishbone.dat_o)
+		]
+		sync += [
+			adr_offset_r.eq(adr_offset)
+		]
+		
+		# Tag memory
+		tag_layout = [("tag", BV(linebits)), ("dirty", BV(1))]
+		tag_do = Record(tag_layout)
+		tag_do_raw = tag_do.to_signal(comb, False)
+		tag_di = Record(tag_layout)
+		tag_di_raw = tag_di.to_signal(comb, True)
+		
+		tag_adr = Signal(BV(linebits))
+		tag_we = Signal()
+		tag_port = MemoryPort(tag_adr, tag_do_raw, tag_we, tag_di_raw)
+		tag_mem = Memory(tagbits+1, 2**linebits, tag_port)
+		
+		comb += [
+			tag_adr.eq(adr_line),
+			tag_di.tag.eq(adr_tag),
+			self.asmiport.adr.eq(Cat(adr_line, tag_do.tag))
+		]
+		
+		# Control FSM
+		fsm = FSM("IDLE", "TEST_HIT",
+			"EVICT_ISSUE", "EVICT_WAIT",
+			"REFILL_WRTAG", "REFILL_ISSUE", "REFILL_WAIT", "REFILL_COMPLETE")
+		
+		fsm.act(fsm.IDLE,
+			If(self.wishbone.cyc_i & self.wishbone.stb_i, fsm.next_state(fsm.TEST_HIT))
+		)
+		fsm.act(fsm.TEST_HIT,
+			If(tag_do.tag == adr_tag,
+				self.wishbone.ack_o.eq(1),
+				If(self.wishbone.we_i,
+					tag_di.dirty.eq(1),
+					tag_we.eq(1)
+				),
+				fsm.next_state(fsm.IDLE)
+			).Else(
+				If(tag_do.dirty,
+					fsm.next_state(fsm.EVICT_ISSUE)
+				).Else(
+					fsm.next_state(fsm.REFILL_WRTAG)
+				)
+			)
+		)
+		
+		fsm.act(fsm.EVICT_ISSUE,
+			self.asmiport.stb.eq(1),
+			self.asmiport.we.eq(1),
+			If(self.asmiport.ack, fsm.next_state(fsm.EVICT_WAIT))
+		)
+		fsm.act(fsm.EVICT_WAIT,
+			# Data is actually sampled by the memory controller in the next state.
+			# But since the data memory has one cycle latency, it gets the data
+			# at the address given during this cycle.
+			If(self.asmiport.get_call_expression(), fsm.next_state(fsm.REFILL_WRTAG))
+		)
+		
+		fsm.act(fsm.REFILL_WRTAG,
+			# Write the tag first to set the ASMI address
+			tag_we.eq(1),
+			fsm.next_state(fsm.REFILL_ISSUE)
+		)
+		fsm.act(fsm.REFILL_ISSUE,
+			self.asmiport.stb.eq(1),
+			If(self.asmiport.ack, fsm.next_state(fsm.REFILL_WAIT))
+		)
+		fsm.act(fsm.REFILL_WAIT,
+			If(self.asmiport.get_call_expression(), fsm.next_state(fsm.REFILL_COMPLETE))
+		)
+		fsm.act(fsm.REFILL_COMPLETE,
+			write_from_asmi.eq(1),
+			fsm.next_state(fsm.TEST_HIT)
+		)
+		
+		return Fragment(comb, sync, memories=[data_mem, tag_mem]) \
+			+ fsm.get_fragment()