vram/VRAM: finish the top-level VRAM module

Well, for now at least. It can build 112KiB and 128KiB memories that
seem to synthesize to something reasonable.

experiments/vram/Top.bsv

@@ -4,35 +4,10 @@ import VRAM::*;
 import ECP5_RAM::*;
 import TriState::*;
 
-(* always_enabled *)
-interface Top;
-   method Action phi2(bit v);
-   method Action we(bit we);
-   method Action addr(UInt#(24) addr);
-   interface InOut#(Bit#(8)) data();
-endinterface
-
 (* synthesize *)
-module mkTop(Top);
+module mkTop(VRAM);
-   Reg#(PortReq) reqA <- mkRegU();
+   let _ret <- mkVRAM(112);
-   Reg#(VRAMData) respA <- mkRegU();
+   return _ret;
-
-   let _ret <- mkByteRAMArray(8);
-
-   rule putA;
-      _ret.portA.put(reqA.chip_select, reqA.write, reqA.addr, reqA.datain);
-   endrule
-
-   rule getA;
-      respA <= _ret.portA.read();
-   endrule
-
-   method portA_read = respA._read;
-   method Action portA_put(cs, w, a, d);
-      reqA <= PortReq{chip_select: cs, write: w, addr: a, datain: d};
-   endmethod
-   method portB_read = _ret.portB.read;
-   method portB_put = _ret.portB.put;
 endmodule
 
 endpackage

tasks.py

@@ -50,7 +50,7 @@ def find_verilog_modules(c, target_dir, modules):
         module_path = None
         verilog_path = Path(module).with_suffix(".v")
         # Try preferred libpaths first.
-        for p in libpaths:
+        for p in preferred_libpaths:
             f = p / verilog_path
             if f.is_file():
                 module_path = f
@@ -214,6 +214,7 @@ def synth(c, target):
 
 @task
 def test(c, target):
+    to_run = []
     for target in expand_test_target(target):
         out_info, out_sim, out_bsc = ensure_build_dirs(target, "info", "sim", "bsc")
         libdirs = bluespec_libdirs(target, "sim")
@@ -226,8 +227,11 @@ def test(c, target):
         testdata_tgt = testdata_tgt.relative_to(out_sim, walk_up=True)
         testdata.unlink(missing_ok=True)
         testdata.symlink_to(testdata_tgt, target_is_directory=True)
-        with c.cd(out_sim):
+        to_run.append((out_sim, f"./TB -V {target.stem}.vcd"))
-            c.run(f"./TB -V {target.stem}.vcd")
+    for d, cmd in to_run:
+        print("")
+        with c.cd(d):
+            c.run(cmd)
 
 @task
 def clean(c):

vram/VRAM.bsv

@@ -7,11 +7,19 @@ import BRAM::*;
 import Vector::*;
 import FIFOF::*;
 import SpecialFIFOs::*;
+import Real::*;
+import Printf::*;
 
 import DelayLine::*;
 import ECP5_RAM::*;
 
-typedef UInt#(17) VRAMAddr;
+export VRAMAddr;
+export VRAMData;
+export mkVRAM;
+export VRAMRequest;
+export VRAMResponse;
+export VRAMServer;
+export VRAM;
 
 typedef Bit#(8) VRAMData;
 
@@ -19,9 +27,7 @@ typedef Bit#(8) VRAMData;
 // address bits.
 typedef UInt#(12) ByteAddr;
 
-// The difference between ByteRAM_Addr and VRAMAddr is the chip
+typedef UInt#(3) ChipAddr;
-// select ID.
-typedef UInt#(5) ChipAddr;
 
 // ByteRAM is two EBRs glued together to make a whole-byte memory.
 typedef EBR#(ByteAddr, VRAMData, ByteAddr, VRAMData) ByteRAM;
@@ -30,14 +36,14 @@ typedef EBR#(ByteAddr, VRAMData, ByteAddr, VRAMData) ByteRAM;
 // block. Like the underlying ECP5 EBRs, callers must bring their own
 // flow control to read out responses one cycle after putting a read
 // request.
-module mkByteRAM(UInt#(3) chip_addr, ByteRAM ifc);
+module mkByteRAM(ChipAddr chip_addr, ByteRAM ifc);
    EBRPortConfig cfg = defaultValue;
    cfg.chip_select_addr = chip_addr;
    EBR#(ByteAddr, Bit#(4), ByteAddr, Bit#(4)) upper <- mkEBRCore(cfg, cfg);
    EBR#(ByteAddr, Bit#(4), ByteAddr, Bit#(4)) lower <- mkEBRCore(cfg, cfg);
 
    interface EBRPort portA;
-      method Action put(UInt#(3) chip_select, Bool write, ByteAddr addr, VRAMData data_in);
+      method Action put(ChipAddr chip_select, Bool write, ByteAddr addr, VRAMData data_in);
          upper.portA.put(chip_select, write, addr, truncate(data_in>>4));
          lower.portA.put(chip_select, write, addr, truncate(data_in));
       endmethod
@@ -48,7 +54,7 @@ module mkByteRAM(UInt#(3) chip_addr, ByteRAM ifc);
    endinterface
 
    interface EBRPort portB;
-      method Action put(UInt#(3) chip_select, Bool write, ByteAddr addr, VRAMData data_in);
+      method Action put(ChipAddr chip_select, Bool write, ByteAddr addr, VRAMData data_in);
          upper.portB.put(chip_select, write, addr, truncate(data_in>>4));
          lower.portB.put(chip_select, write, addr, truncate(data_in));
       endmethod
@@ -59,6 +65,10 @@ module mkByteRAM(UInt#(3) chip_addr, ByteRAM ifc);
    endinterface
 endmodule : mkByteRAM
 
+// mkByteRAMArray arrays up to 8 mkByteRAMs together, using the
+// hardwired chip select lines to route inputs appropriately and a mux
+// tree to collect outputs. With num_chips=8, the resulting ByteRAM is
+// 32768x8b.
 module mkByteRAMArray(Integer num_chips, ByteRAM ifc);
    if (num_chips > 8)
       error("mkByteRAMArray can only array 8 raw ByteRAMs");
@@ -67,11 +77,11 @@ module mkByteRAMArray(Integer num_chips, ByteRAM ifc);
    for (Integer i=0; i<num_chips; i=i+1)
       blocks[i] <- mkByteRAM(fromInteger(i));
 
-   DelayLine#(UInt#(3)) read_chip_A <- mkDelayLine(1);
+   DelayLine#(ChipAddr) read_chip_A <- mkDelayLine(1);
-   DelayLine#(UInt#(3)) read_chip_B <- mkDelayLine(1);
+   DelayLine#(ChipAddr) read_chip_B <- mkDelayLine(1);
 
    interface EBRPort portA;
-      method Action put(UInt#(3) chip_select, Bool write, ByteAddr addr, VRAMData data_in);
+      method Action put(ChipAddr chip_select, Bool write, ByteAddr addr, VRAMData data_in);
          for (Integer i=0; i<num_chips; i=i+1)
             blocks[i].portA.put(chip_select, write, addr, data_in);
          if (write)
@@ -89,7 +99,7 @@ module mkByteRAMArray(Integer num_chips, ByteRAM ifc);
    endinterface
 
    interface EBRPort portB;
-      method Action put(UInt#(3) chip_select, Bool write, ByteAddr addr, VRAMData data_in);
+      method Action put(ChipAddr chip_select, Bool write, ByteAddr addr, VRAMData data_in);
          for (Integer i=0; i<num_chips; i=i+1)
             blocks[i].portB.put(chip_select, write, addr, data_in);
          if (write)
@@ -107,6 +117,10 @@ module mkByteRAMArray(Integer num_chips, ByteRAM ifc);
    endinterface
 endmodule
 
+typedef UInt#(2) ArrayAddr;
+
+typedef UInt#(17) VRAMAddr;
+
 typedef struct {
    VRAMAddr addr;
    Maybe#(VRAMData) data;
@@ -124,38 +138,58 @@ interface VRAM;
    interface VRAMServer portB;
 endinterface
 
-module mkVRAM(Integer num_4kB_blocks, VRAM ifc);
+// mkVRAM creates a dual port VRAM of the specified size, using ECP5
-   if (num_4kB_blocks > 32)
+// EBR memory primitives. The memory size must be a multiple of 4KiB,
-      error("maximum number of blocks is 32 (128KiB)");
+// with a maximum of 128KiB.
-   UInt#(TAdd#(SizeOf#(VRAMAddr), 1)) max_request_addr = fromInteger((4096 * num_4kB_blocks));
+//
+// The returned VRAM servers implement flow control. As long as
+// responses are processed as soon as they're available, each port can
+// process one memory operation per cycle.
+//
+// The VRAM does not prevent write-write or write-read conflicts
+// between the ports. The outcome of a simultaneous write to the same
+// address is unspecified, as is the read output in a simultaneous
+// read and write of the same address. The caller must use external
+// arbitration to avoid such accesses.
+module mkVRAM(Integer num_kilobytes, VRAM ifc);
+   if (num_kilobytes > 128)
+      error("maximum VRAM size is 128KiB");
+   let num_bytes = num_kilobytes*1024;
+   if (num_bytes % 4096 != 0)
+      error("VRAM must be a multiple of 4096b");
+   let num_byterams = num_bytes/4096;
+   let num_arrays = ceil(fromInteger(num_byterams) / 8);
 
-   function Tuple2#(ChipAddr, ByteAddr) split_addr(VRAMAddr a);
+   function Tuple3#(ArrayAddr, ChipAddr, ByteAddr) split_addr(VRAMAddr a);
-      UInt#(TAdd#(SizeOf#(VRAMAddr), 1)) expanded = extend(a);
+      if (num_bytes < 128*1024)
-      VRAMAddr wrapped = truncate(expanded % max_request_addr);
+         a = a % fromInteger(num_bytes);
-      match {.chip, .off} = split(pack(wrapped));
+      match {.top, .byteaddr} = split(pack(a));
-      return tuple2(unpack(chip), unpack(off));
+      Tuple2#(Bit#(SizeOf#(ArrayAddr)), Bit#(SizeOf#(ChipAddr))) route = split(top);
+      return tuple3(unpack(tpl_1(route)), unpack(tpl_2(route)), unpack(byteaddr));
    endfunction
 
-   ByteRAM blocks[num_4kB_blocks];
+   ByteRAM arrays[num_arrays];
-   for (Integer i=0; i<num_4kB_blocks; i=i+1)
+   for (Integer i=0; i<num_arrays; i=i+1) begin
-      blocks[i] <- mkByteRAM(0);
+      let array_size = min(num_byterams - (i*8), 8);
+      arrays[i] <- mkByteRAMArray(array_size);
+   end
 
-   Reg#(Maybe#(ChipAddr)) inflight_A[2] <- mkCReg(2, tagged Invalid);
+   Reg#(Maybe#(ArrayAddr)) inflight_A[2] <- mkCReg(2, tagged Invalid);
-   Reg#(Maybe#(ChipAddr)) inflight_B[2] <- mkCReg(2, tagged Invalid);
+   Reg#(Maybe#(ArrayAddr)) inflight_B[2] <- mkCReg(2, tagged Invalid);
 
    interface VRAMServer portA;
       interface Put request;
          method Action put(VRAMRequest req) if (inflight_A[1] matches tagged Invalid);
-            match {.chip, .off} = split_addr(req.addr);
+            match {.array, .chip, .byteaddr} = split_addr(req.addr);
-            blocks[chip].portA.put(0, isValid(req.data), off, fromMaybe(0, req.data));
+            arrays[array].portA.put(chip, isValid(req.data), byteaddr, fromMaybe(0, req.data));
             if (!isValid(req.data))
-               inflight_A[1] <= tagged Valid chip;
+               inflight_A[1] <= tagged Valid array;
          endmethod
       endinterface
       interface Get response;
-         method ActionValue#(VRAMResponse) get() if (inflight_A[0] matches tagged Valid .chip);
+         method ActionValue#(VRAMResponse) get() if (inflight_A[0] matches tagged Valid .array);
             inflight_A[0] <= tagged Invalid;
-            return VRAMResponse{data: blocks[chip].portA.read()};
+            return VRAMResponse{data: arrays[array].portA.read()};
          endmethod
       endinterface
    endinterface
@@ -163,16 +197,16 @@ module mkVRAM(Integer num_4kB_blocks, VRAM ifc);
    interface VRAMServer portB;
       interface Put request;
          method Action put(VRAMRequest req) if (inflight_B[1] matches tagged Invalid);
-            match {.chip, .off} = split_addr(req.addr);
+            match {.array, .chip, .byteaddr} = split_addr(req.addr);
-            blocks[chip].portB.put(0, isValid(req.data), off, fromMaybe(0, req.data));
+            arrays[array].portB.put(0, isValid(req.data), byteaddr, fromMaybe(0, req.data));
             if (!isValid(req.data))
-               inflight_B[1] <= tagged Valid chip;
+               inflight_B[1] <= tagged Valid array;
          endmethod
       endinterface
       interface Get response;
-         method ActionValue#(VRAMResponse) get() if (inflight_B[0] matches tagged Valid .chip);
+         method ActionValue#(VRAMResponse) get() if (inflight_B[0] matches tagged Valid .array);
             inflight_B[0] <= tagged Invalid;
-            return VRAMResponse{data: blocks[chip].portB.read()};
+            return VRAMResponse{data: arrays[array].portB.read()};
          endmethod
       endinterface
    endinterface