Add fragmenter in front of TLMemoryML507 and implementation notes

src/main/scala/unleashed/u500ml507devkit/fpga/Memory.scala

@@ -18,13 +18,23 @@ case object MemoryML507Key extends Field[MemoryML507Params]
 trait HasMemoryML507 { this: BaseSubsystem =>
     val memory = LazyModule(new TLMemoryML507(p(MemoryML507Key)))
 
+    // 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 = TLFragmenter(32, 64, alwaysMin=true)
+
     // TODO: right TL/memory node chain?
-    memory.node := memBuses.head.toDRAMController(Some("ml507mig"))()
+    memory.node := fragmenter := memBuses.head.toDRAMController(Some("ml507mig"))()
 }
 
 class TLMemoryML507(c: MemoryML507Params)(implicit p: Parameters) extends LazyModule {
-    val width = 512
+    // Corresponds to MIG interface with 64 bit width and a burst length of 4
-    val beatBytes = width/8 // TODO: To wide? TLFragmenter? fixedSize?
+    val width = 256
+    val beatBytes = width/8 // 32 byte (half a cache-line, fragmented)
 
     val device = new MemoryDevice
     val node = TLManagerNode(
@@ -41,15 +51,54 @@ class TLMemoryML507(c: MemoryML507Params)(implicit p: Parameters) extends LazyMo
             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).
 
     lazy val module = new LazyModuleImp(this) {
-        val (in, edge)= node.in(0)
+        // in: TLBundle, edge: TLEdgeIn
+        val (in, edge) = node.in(0)
+
+        // Due to the Fragmenter defined above, 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 possibly
+        // data to the MIG interface.
+        // Put requests can be acknowledged as soon as they are latched into
+        // the write fifo of the MIG (possibly combinatorily).
+        // For read requests, we have to store the source id and size in a
+        // queue for later acknowledgment.
+        // We are ready if both the MIG and the response data queue are not
+        // full.
+
+        // Widths of the A channel:
+        // addressBits: 32
+        // dataBits:    256
+        // sourceBits:  6
+        // sinkBits:    1
+        // sizeBits:    3
+
+        // source (from): in.a.bits.source
+        // adresse (to): edgeIn.address(in.a.bits)
+        // size: edgeIn.size(in.a.bits)
+        // isPut: edgeIn.hasData(in.a.bits)
+
+        // bits kommt von Decoupled: ready, valid + bits
+
+        println("a parameters: " + in.a.bits.params)
+
+        in.a.ready := Bool(false)
+        in.d.valid := Bool(false)
 
         // Tie off unused channels
-        in.a.ready := Bool(false)
         in.b.valid := Bool(false)
-        in.c.ready := Bool(false)
+        in.c.ready := Bool(true)
-        in.d.valid := Bool(false)
+        in.e.ready := Bool(true)
-        in.e.ready := Bool(false)
     }
 }