Reduce crossing and queue depths to save space and ease timing

src/main/scala/devices/xilinx/xilinxml507mig/XilinxML507MIG.scala

@@ -0,0 +1,199 @@
+// See LICENSE.SiFive for license details.
+
+package sifive.fpgashells.devices.xilinx.xilinxml507mig
+
+import Chisel._
+import chisel3.core.{Input, Output}
+import freechips.rocketchip.config.Parameters
+import freechips.rocketchip.diplomacy._
+import freechips.rocketchip.subsystem.{AsynchronousCrossing, HasCrossing}
+import freechips.rocketchip.tilelink._
+import freechips.rocketchip.util._
+
+case class XilinxML507MIGParams(
+    address: Seq[AddressSet]
+)
+
+class MemoryController extends BlackBox {
+    val io = IO(new Bundle {
+        val sys = new MemorySysIO
+        val ddr2 = new MemoryDDR2IO
+
+        val request_addr = Input(UInt(28.W))
+        val request_type = Input(Bool())
+        val request_data = Input(UInt(256.W))
+        val request_mask = Input(UInt(32.W))
+        val request_valid = Input(Bool())
+        val request_ready = Output(Bool())
+
+        val response_data = Output(UInt(256.W))
+        val response_valid = Output(Bool())
+        // no ready, as the mig does not wait
+    })
+
+    override def desiredName: String = "memory_controller"
+}
+
+class ResponseQueueIO extends Bundle {
+    val read = Bool()
+    val source = UInt()
+    val size = UInt()
+}
+
+class XilinxML507MIGToTL(c: XilinxML507MIGParams)(implicit p: Parameters) extends LazyModule with HasCrossing {
+    // Corresponds to MIG interface with 64 bit width and a burst length of 4
+    val width = 256
+    val beatBytes = width/8 // 32 byte (half a cache-line, fragmented)
+
+    val address_range = AddressRange.fromSets(c.address).head
+    require(log2Ceil(address_range.size) == 28, "Max 256MiB DIMMs supported")
+    val crossing = AsynchronousCrossing(1)
+
+    val device = new MemoryDevice
+    val node = TLManagerNode(
+        Seq(TLManagerPortParameters(
+            Seq(TLManagerParameters(
+                address         = c.address,
+                resources       = device.reg,
+                regionType      = RegionType.UNCACHED,
+                executable      = true,
+                supportsGet     = TransferSizes(1, beatBytes),
+                supportsPutFull = TransferSizes(1, beatBytes),
+                fifoId          = Some(0) // in-order
+            )),
+            beatBytes = beatBytes
+        ))
+    )
+    // We could possibly also support supportsPutPartial, as we need support
+    // for masks anyway because of the possibility of transfers smaller that
+    // the data width (size signal, see below).
+    // Seems we can: TL$7.3
+
+    lazy val module = new LazyModuleImp(this) {
+        val io = IO(new Bundle {
+            val port_sys = new MemorySysIO
+            val port_ddr2 = new MemoryDDR2IO
+        })
+
+        val controller = Module(new MemoryController)
+        io.port_sys <> controller.io.sys
+        io.port_ddr2 <> controller.io.ddr2
+
+        // in: TLBundle, edge: TLEdgeIn
+        val (in, edge) = node.in(0)
+
+        // Due to the TLFragmenter defined below, all messages are 32 bytes or
+        // smaller. The data signal of the TL channels is also 32 bytes, so
+        // all messages will be transfered in a single beat.
+        // Also, TL guarantees (see TL$4.6) that the payload of a data message
+        // is always aligned to the width of the beat, e.g. in case of a 32
+        // byte data signal, data[7:0] will always have address 0x***00000 and
+        // data[255:247] address 0x***11111. It is also guaranteed that the
+        // mask bits always correctly reflect the active bytes inside the beat
+        // with respect to the size and address. So we can directly forward
+        // the mask, (relative) address and data to the MIG interface.
+
+        // An AddressSet is always aligned, so we don't need to subtract the
+        // base address, we can just take the lower bits. The lowest 5 bits
+        // are used for indexing the 32 byte word of the MIG.
+        val address = in.a.bits.address(27, 0) & "hFFFFFE0".U
+
+        // Save the source, size and type of the requests in a queue so we
+        // can synthesize the right responses in fifo order. The length also
+        // determines the maximum number of in-flight requests.
+        val ack_queue = Module(new Queue(new ResponseQueueIO, 2))
+
+        // Pass data directly to the controller
+        controller.io.request_addr := address
+        controller.io.request_type := !edge.hasData(in.a.bits)
+        controller.io.request_data := in.a.bits.data
+        // TL uses high to indicate valid data while mig uses low
+        controller.io.request_mask := ~ in.a.bits.mask
+
+        ack_queue.io.enq.bits.read   := !edge.hasData(in.a.bits)
+        ack_queue.io.enq.bits.source := in.a.bits.source
+        ack_queue.io.enq.bits.size   := in.a.bits.size
+
+        // We are ready when the controller and the queue input are ready
+        in.a.ready := controller.io.request_ready && ack_queue.io.enq.ready
+        // Both queues only latch data if the other is ready, so that data
+        // is latched into both queues or not at all
+        controller.io.request_valid := in.a.valid && ack_queue.io.enq.ready
+        ack_queue.io.enq.valid      := in.a.valid && controller.io.request_ready
+
+
+        // We have to buffer the responses from the MIG as it has no internal
+        // buffer and will output its read responses only for one cycle. To
+        // avoid losing any responses, this queue *must* be at least as wide
+        // as the ack queue, so that we can catch all responses, even if the
+        // ack queue is completely filled with read requests.
+        val response_queue = Module(new Queue(controller.io.response_data, 2))
+
+        response_queue.io.enq.bits  := controller.io.response_data
+        response_queue.io.enq.valid := controller.io.response_valid
+        // MIG does not support delaying a response, so we ignore enq.ready.
+        // This will result in lost reads and returning wrong data in further
+        // AccessAckData messages, so this must be avoided (see above).
+
+        // Acks may or may not contain data depending on the request, but we
+        // can always pass the data, even if it is invalid in the write case,
+        // because it is ignored for AccessAck responses
+        val response_read = ack_queue.io.deq.bits.read
+        in.d.bits.opcode := Mux(response_read, TLMessages.AccessAckData, TLMessages.AccessAck)
+        in.d.bits.param  := UInt(0) // reserved, must be 0
+        in.d.bits.size   := ack_queue.io.deq.bits.size
+        in.d.bits.source := ack_queue.io.deq.bits.source
+        in.d.bits.sink   := UInt(0) // ignored
+        in.d.bits.data   := response_queue.io.deq.bits
+        in.d.bits.error  := Bool(false)
+
+        // The data is valid when the ack queue data is valid (write case) or
+        // when the ack *and* response queues are valid (read case)
+        in.d.valid       := ack_queue.io.deq.valid && (!response_read ||
+                            response_queue.io.deq.valid)
+        // Let the ack queue dequeue when the master is ready (write case) or
+        // when the master is ready *and* there is a valid response (read case)
+        ack_queue.io.deq.ready      := in.d.ready && (!response_read ||
+                                       response_queue.io.deq.valid)
+        // Let the response queue dequeue when the master is ready and there
+        // is a valid read ack waiting
+        response_queue.io.deq.ready := in.d.ready && response_read &&
+                                       ack_queue.io.deq.valid
+
+
+        // Tie off unused channels
+        in.b.valid := Bool(false)
+        in.c.ready := Bool(true)
+        in.e.ready := Bool(true)
+    }
+}
+
+class XilinxML507MIG(c : XilinxML507MIGParams)(implicit p: Parameters) extends LazyModule {
+    // The Fragmenter will not fragment messages <= 32 bytes, so all
+    // slaves have to support this size. 64 byte specifies the maximum
+    // supported transfer size that the slave side of the fragmenter supports
+    // against the master (here the main memory bus). Specifying alwaysMin as
+    // true results in all messages being fragmented to the minimal size
+    // (32 byte). In TL1 terms, slaves correspond roughly to managers and
+    // masters to clients (confusingly…).
+    val fragmenter = LazyModule(new TLFragmenter(32, 64, alwaysMin=true))
+    val island = LazyModule(new XilinxML507MIGToTL(c))
+
+    val node: TLInwardNode =
+        island.node := island.crossTLIn := fragmenter.node
+
+    lazy val module = new LazyModuleImp(this) {
+        val io = IO(new Bundle {
+            val port_sys = new MemorySysIO
+            val port_ddr2 = new MemoryDDR2IO
+        })
+
+        io.port_sys <> island.module.io.port_sys
+        io.port_ddr2 <> island.module.io.port_ddr2
+
+        // The MIGToTL module lives in a separate clock domain together with
+        // the MIG, which is why it is called "island".
+        island.module.clock := io.port_sys.clk0
+        island.module.reset := io.port_sys.reset
+    }
+}

src/main/scala/devices/xilinx/xilinxml507mig/XilinxML507MIGPeriphery.scala

@@ -0,0 +1,58 @@
+// See LICENSE.SiFive for license details.
+
+package sifive.fpgashells.devices.xilinx.xilinxml507mig
+
+import Chisel._
+import chisel3.core.{Input, Output}
+import chisel3.experimental.Analog
+import freechips.rocketchip.config.Field
+import freechips.rocketchip.diplomacy._
+import freechips.rocketchip.subsystem.BaseSubsystem
+
+case object MemoryML507Key extends Field[XilinxML507MIGParams]
+
+trait HasMemoryML507 { this: BaseSubsystem =>
+    val memory = LazyModule(new XilinxML507MIG(p(MemoryML507Key)))
+
+    memory.node := memBuses.head.toDRAMController(Some("xilinxml507mig"))()
+}
+
+class MemorySysIO extends Bundle {
+    val clk0 = Input(Clock())
+    val clk90 = Input(Clock())
+    val clkdiv0 = Input(Clock())
+    val clk_locked = Input(Bool())
+    val clk_idelay = Input(Clock())
+    val reset = Input(Bool())
+}
+
+class MemoryDDR2IO extends Bundle {
+    val dq = Analog(64.W)
+    val a = Output(Bits(13.W))
+    val ba = Output(Bits(2.W))
+    val ras_n = Output(Bits(1.W))
+    val cas_n = Output(Bits(1.W))
+    val we_n = Output(Bits(1.W))
+    val cs_n = Output(Bits(1.W))
+    val odt = Output(Bits(1.W))
+    val cke = Output(Bits(1.W))
+    val dm = Output(Bits(8.W))
+    val dqs = Analog(8.W)
+    val dqs_n = Analog(8.W)
+    val ck = Output(Bits(2.W))
+    val ck_n = Output(Bits(2.W))
+}
+
+trait HasMemoryML507Bundle {
+    val ddr_sys: MemorySysIO
+    val ddr2: MemoryDDR2IO
+}
+
+trait HasMemoryML507ModuleImp extends LazyModuleImp with HasMemoryML507Bundle {
+    val outer: HasMemoryML507
+    val ddr_sys = IO(new MemorySysIO)
+    val ddr2 = IO(new MemoryDDR2IO)
+
+    ddr_sys <> outer.memory.module.io.port_sys
+    ddr2 <> outer.memory.module.io.port_ddr2
+}

src/main/scala/ip/xilinx/Xilinx.scala

@@ -114,6 +114,17 @@ class ml507_dvi_clock extends BlackBox {
   }
 }
 
+class ml507_ddr2_clock extends BlackBox {
+  val io = new Bundle {
+    val CLKIN_P_IN = Clock(INPUT)
+    val CLKIN_N_IN = Clock(INPUT)
+    val CLK0_OUT   = Clock(OUTPUT)
+    val CLK90_OUT  = Clock(OUTPUT)
+    val CLKDV_OUT  = Clock(OUTPUT)
+    val LOCKED_OUT = Bool(OUTPUT)
+  }
+}
+
 //-------------------------------------------------------------------------
 // vc707_sys_clock_mmcm
 //-------------------------------------------------------------------------

src/main/scala/shell/xilinx/ML507Shell.scala

@@ -15,7 +15,8 @@ import sifive.blocks.devices.uart._
 import sifive.blocks.devices.chiplink._
 import sifive.blocks.devices.terminal._
 
-import sifive.fpgashells.ip.xilinx.{PowerOnResetFPGAOnly, sdio_spi_bridge, ml507_dvi_clock, ml507_sys_clock, vc707reset}
+import sifive.fpgashells.devices.xilinx.xilinxml507mig._
+import sifive.fpgashells.ip.xilinx.{PowerOnResetFPGAOnly, sdio_spi_bridge, ml507_ddr2_clock, ml507_dvi_clock, ml507_sys_clock, vc707reset}
 
 //-------------------------------------------------------------------------
 // ML507Shell
@@ -56,11 +57,18 @@ abstract class ML507Shell(implicit val p: Parameters) extends RawModule {
   // 100Mhz sysclk
   val sys_clock            = IO(Input(Clock()))
 
+  // 200MHz ddrclk
+  val ddr_clock_p          = IO(Input(Clock()))
+  val ddr_clock_n          = IO(Input(Clock()))
+
   // active high async reset
   val reset                = IO(Input(Bool()))
 
   // LED
-  val led                  = IO(Vec(8, Output(Bool())))
+  val led                  = IO(Output(Vec(8, Bool())))
+
+  // DIP switches
+  val dip                  = IO(Input(Vec(8, Bool())))
 
   // UART
   val uart_tx              = IO(Output(Bool()))
@@ -71,27 +79,12 @@ abstract class ML507Shell(implicit val p: Parameters) extends RawModule {
   val sdio_cmd             = IO(Analog(1.W))
   val sdio_dat             = IO(Analog(4.W))
 
-  //Buttons
-  val btn_0                = IO(Input(Bool()))
-  val btn_1                = IO(Input(Bool()))
-  val btn_2                = IO(Input(Bool()))
-  val btn_3                = IO(Input(Bool()))
-
-  //Sliding switches
-  val sw_0                 = IO(Input(Bool()))
-  val sw_1                 = IO(Input(Bool()))
-  val sw_2                 = IO(Input(Bool()))
-  val sw_3                 = IO(Input(Bool()))
-  val sw_4                 = IO(Input(Bool()))
-  val sw_5                 = IO(Input(Bool()))
-  val sw_6                 = IO(Input(Bool()))
-  val sw_7                 = IO(Input(Bool()))
-
   // Feedback
   val clock_led            = IO(Output(Clock()))
   val reset_led            = IO(Output(Bool()))
 
   val dvi                  = IO(new TerminalDVIIO)
+  val ddr2                 = IO(new MemoryDDR2IO)
 
   //-----------------------------------------------------------------------
   // Wire declrations
@@ -108,6 +101,12 @@ abstract class ML507Shell(implicit val p: Parameters) extends RawModule {
   val dvi_clock       = Wire(Clock())
   val dvi_reset       = Wire(Bool())
 
+  val ddr_clk0        = Wire(Clock())
+  val ddr_clk90       = Wire(Clock())
+  val ddr_clkdiv0     = Wire(Clock())
+  val ddr_clk_locked  = Wire(Bool())
+  val ddr_reset       = Wire(Bool())
+
   val sd_spi_sck      = Wire(Bool())
   val sd_spi_cs       = Wire(Bool())
   val sd_spi_dq_i     = Wire(Vec(4, Bool()))
@@ -137,9 +136,19 @@ abstract class ML507Shell(implicit val p: Parameters) extends RawModule {
   ml507_dvi_clock.io.CLKIN_IN := sys_clock
   dvi_clock := ml507_dvi_clock.io.CLKFX_OUT
 
+  // 200 MHz (DDR2 and IDELAY clock)
+  val ml507_ddr2_clock = Module(new ml507_ddr2_clock)
+  ml507_ddr2_clock.io.CLKIN_P_IN := ddr_clock_p
+  ml507_ddr2_clock.io.CLKIN_N_IN := ddr_clock_n
+  ddr_clk0 := ml507_ddr2_clock.io.CLK0_OUT
+  ddr_clk90 := ml507_ddr2_clock.io.CLK90_OUT
+  ddr_clkdiv0 := ml507_ddr2_clock.io.CLKDV_OUT
+  ddr_clk_locked := ml507_ddr2_clock.io.LOCKED_OUT
+
   // Clocks locked?
   clk_locked := ml507_sys_clock.io.LOCKED_OUT &
-                ml507_dvi_clock.io.LOCKED_OUT
+                ml507_dvi_clock.io.LOCKED_OUT &
+                ddr_clk_locked
 
   //-----------------------------------------------------------------------
   // System reset
@@ -151,7 +160,8 @@ abstract class ML507Shell(implicit val p: Parameters) extends RawModule {
   val safe_reset = Module(new vc707reset)
 
   safe_reset.io.areset := do_reset
-  safe_reset.io.clock1 := dut_clock
+  safe_reset.io.clock1 := ddr_clk0
+  ddr_reset            := safe_reset.io.reset1
   safe_reset.io.clock2 := dut_clock
   safe_reset.io.clock3 := dvi_clock
   dvi_reset            := safe_reset.io.reset3
@@ -172,6 +182,21 @@ abstract class ML507Shell(implicit val p: Parameters) extends RawModule {
     dut.terminal.reset := dvi_reset
   }
 
+  //-----------------------------------------------------------------------
+  // Memory controller
+  //-----------------------------------------------------------------------
+
+  def connectDDRMemory(dut: HasMemoryML507ModuleImp): Unit = {
+    ddr2 <> dut.ddr2
+
+    dut.ddr_sys.clk0 := ddr_clk0
+    dut.ddr_sys.clk90 := ddr_clk90
+    dut.ddr_sys.clkdiv0 := ddr_clkdiv0
+    dut.ddr_sys.clk_locked := ddr_clk_locked
+    dut.ddr_sys.clk_idelay := ddr_clk0
+    dut.ddr_sys.reset := ddr_reset
+  }
+
   //-----------------------------------------------------------------------
   // UART
   //-----------------------------------------------------------------------