gateware.accumulator: new module.

Wide counters (24-bit and wider) are problematic on the fairly slow
iCE40 architecture, and this module allows using counters of any width
by pipelining them.

Co-authored-by: Wanda <wanda@phinode.net>

software/glasgow/gateware/accumulator.py

@@ -0,0 +1,64 @@
+import operator
+
+from amaranth import *
+from amaranth.lib import wiring
+from amaranth.lib.wiring import In, Out
+
+
+class Accumulator(wiring.Component):
+    """Pipelined arithmetic accumulator.
+
+    Computes :py:`new_sum = old_sum + addend` using at most :py:`stage_width` wide adders, with
+    a latency of :py:`(width + stage_width - 1) // stage_width` cycles and throughput of one
+    addition per cycle.
+
+    Members
+    -------
+    addend : In(width)
+        Addend.
+    sum : Out(width)
+        Accumulated sum.
+    """
+    def __init__(self, width, *, stage_width=16):
+        self._width = operator.index(width)
+        self._stage_width = operator.index(stage_width)
+        assert self._width >= 1 and self._stage_width >= 1
+        self._stages = 1 + (self._width + self._stage_width - 1) // self._stage_width
+        super().__init__({
+            "addend": In(self._width),
+            "sum": Out(self._width)
+        })
+
+    @property
+    def stages(self):
+        return self._stages
+
+    def elaborate(self, platform):
+        m = Module()
+
+        carry = Const(0)
+        addend = Signal.like(self.addend)
+        result = Cat()
+
+        m.d.sync += addend.eq(self.addend)
+
+        for index, start_at in enumerate(range(0, self._width, self._stage_width)):
+            stage_width = min(self._width - start_at, self._stage_width)
+
+            carry_next = Signal(name=f"carry{index}")
+            addend_next = Signal.like(addend[stage_width:], name=f"addend{index}")
+            result_next = Signal.like(result, name=f"result{index}")
+            stage = Signal(stage_width, name=f"stage{index}")
+
+            m.d.sync += Cat(stage, carry_next).eq(stage + addend[:stage_width] + carry)
+            m.d.sync += addend_next.eq(addend[stage_width:])
+            m.d.sync += result_next.eq(result)
+
+            carry = carry_next
+            addend = addend_next
+            result = Cat(result_next, stage)
+
+        m.d.comb += self.sum.eq(result)
+
+        return m
+

software/tests/gateware/test_accumulator.py

@@ -0,0 +1,25 @@
+import unittest
+from amaranth import *
+from amaranth.sim import Tick
+
+from glasgow.gateware import simulation_test
+from glasgow.gateware.accumulator import Accumulator
+
+
+class AccumulatorTestCase(unittest.TestCase):
+    def setUp(self):
+        self.tb = Accumulator(5, stage_width=2)
+
+    @simulation_test()
+    def test_counter(self, tb):
+        total = 0
+        queue = [0] * (self.tb.stages + 1)
+        for i in range(100):
+            addend = i * 2137 % 32
+            total += addend
+            total %= 32
+            queue.append(total)
+            self.assertEqual(queue[0], (yield self.tb.sum))
+            del queue[0]
+            yield self.tb.addend.eq(addend)
+            yield Tick()