1
0

tilelink: WidthWidget rewrite beat merging

- errors are properly OR reduced
- registers latched only as needed (was previously a shift register)
- combines beats without inspecting address (removes addr_lo dependency)
This commit is contained in:
Wesley W. Terpstra 2017-07-25 16:54:30 -07:00
parent f0ffb7e31e
commit 0f5065fbf3

View File

@ -26,58 +26,55 @@ class TLWidthWidget(innerBeatBytes: Int)(implicit p: Parameters) extends LazyMod
val inBytes = edgeIn.manager.beatBytes
val outBytes = edgeOut.manager.beatBytes
val ratio = outBytes / inBytes
val keepBits = log2Ceil(outBytes)
val dropBits = log2Ceil(inBytes)
val countBits = log2Ceil(ratio)
val rdata = Reg(UInt(width = (ratio-1)*inBytes*8))
val rmask = Reg(UInt(width = (ratio-1)*inBytes))
val data = Cat(edgeIn.data(in.bits), rdata)
val mask = Cat(edgeIn.mask(in.bits), rmask)
val address = edgeIn.address(in.bits)
val size = edgeIn.size(in.bits)
val hasData = edgeIn.hasData(in.bits)
val limit = UIntToOH1(size, keepBits) >> dropBits
val count = RegInit(UInt(0, width = log2Ceil(ratio)))
val count = RegInit(UInt(0, width = countBits))
val first = count === UInt(0)
val limit = UIntToOH1(size, log2Ceil(outBytes)) >> log2Ceil(inBytes)
val last = count === limit || !hasData
val enable = Seq.tabulate(ratio) { i => !((count ^ UInt(i)) & limit).orR }
when (in.fire()) {
rdata := data >> inBytes*8
rmask := mask >> inBytes
count := count + UInt(1)
when (last) { count := UInt(0) }
}
val cases = Seq.tabulate(log2Ceil(ratio)+1) { i =>
val high = outBytes
val take = (1 << i)*inBytes
(Fill(1 << (log2Ceil(ratio)-i), data(high*8-1, (high-take)*8)),
Fill(1 << (log2Ceil(ratio)-i), mask(high -1, (high-take))))
def helper(idata: UInt): UInt = {
val odata = Seq.fill(ratio) { idata }
val rdata = Reg(Vec(ratio-1, idata))
val pdata = rdata :+ idata
val mdata = (enable zip (odata zip pdata)) map { case (e, (o, p)) => Mux(e, o, p) }
when (in.fire() && !last) {
(rdata zip mdata) foreach { case (r, m) => r := m }
}
val dataMux = Vec.tabulate(log2Ceil(edgeIn.maxTransfer)+1) { lgSize =>
cases(min(max(lgSize - log2Ceil(inBytes), 0), log2Ceil(ratio)))._1
}
val maskMux = Vec.tabulate(log2Ceil(edgeIn.maxTransfer)+1) { lgSize =>
cases(min(max(lgSize - log2Ceil(inBytes), 0), log2Ceil(ratio)))._2
Cat(mdata.reverse)
}
val dataOut = if (edgeIn.staticHasData(in.bits) == Some(false)) UInt(0) else dataMux(size)
lazy val maskFull = edgeOut.mask(address, size)
lazy val maskOut = Mux(hasData, maskMux(size) & maskFull, maskFull)
def reduce(i: Bool): Bool = {
val state = Reg(Bool())
val next = i || (!first && state)
when (in.fire()) { state := next }
next
}
in.ready := out.ready || !last
out.valid := in.valid && last
out.bits := in.bits
edgeOut.data(out.bits) := dataOut
out.bits match {
case a: TLBundleA => a.mask := maskOut
case b: TLBundleB => b.mask := maskOut
case c: TLBundleC => ()
case d: TLBundleD => ()
// addr_lo gets padded with 0s on D channel, the only lossy transform in this core
// this should be safe, because we only care about addr_lo on D to determine which
// piece of data to extract when the D data bus is narrowed. Since we duplicated the
// data to all locations, addr_lo still points at a valid copy.
// Don't put down hardware if we never carry data
edgeOut.data(out.bits) := (if (edgeIn.staticHasData(in.bits) == Some(false)) UInt(0) else helper(edgeIn.data(in.bits)))
(out.bits, in.bits) match {
case (o: TLBundleA, i: TLBundleA) => o.mask := edgeOut.mask(o.address, o.size) & Mux(hasData, helper(i.mask), ~UInt(0, width=outBytes))
case (o: TLBundleB, i: TLBundleB) => o.mask := edgeOut.mask(o.address, o.size) & Mux(hasData, helper(i.mask), ~UInt(0, width=outBytes))
case (o: TLBundleC, i: TLBundleC) => o.error := reduce(i.error)
case (o: TLBundleD, i: TLBundleD) => o.error := reduce(i.error)
case _ => require(false, "Impossible bundle combination in WidthWidget")
}
}