fir: automatically use transposed topology

artiq/gateware/dsp/fir.py

@@ -1,5 +1,6 @@
 from operator import add
 from functools import reduce
+from collections import namedtuple
 import numpy as np
 from migen import *
 
@@ -38,7 +39,10 @@ def halfgen4(width, n):
 class FIR(Module):
     """Full-rate finite impulse response filter.
 
-    :param coefficients: integer taps.
+    Tries to use transposed form (adder chain instead of adder tree)
+    as much as possible.
+
+    :param coefficients: integer taps, increasing delay.
     :param width: bit width of input and output.
     :param shift: scale factor (as power of two).
     """
@@ -47,37 +51,46 @@ class FIR(Module):
         self.i = Signal((width, True))
         self.o = Signal((width, True))
         n = len(coefficients)
-        self.latency = (n + 1)//2 + 2
+        self.latency = n//2 + 3
 
         ###
 
+        if shift is None:
+            shift = bits_for(sum(abs(c) for c in coefficients)) - 1
+
         # Delay line: increasing delay
         x = [Signal((width, True)) for _ in range(n)]
         self.sync += [xi.eq(xj) for xi, xj in zip(x, [self.i] + x)]
 
-        if shift is None:
-            shift = width - 1
-
+        o = Signal((width + shift + 1, True))
+        self.comb += self.o.eq(o >> shift)
+        delay = -1
         # Make products
-        o = []
         for i, c in enumerate(coefficients):
             # simplify for halfband and symmetric filters
-            if c == 0 or c in coefficients[i + 1:]:
+            if not c or c in coefficients[:i]:
                 continue
-            m = Signal((width + shift, True))
-            self.sync += m.eq(c*reduce(add, [
-                xj for xj, cj in zip(x[::-1], coefficients) if cj == c
-            ]))
-            o.append(m)
-
-        # Make sum
-        self.sync += self.o.eq(reduce(add, o) >> shift)
+            js = [j for j, cj in enumerate(coefficients) if cj == c]
+            m = Signal.like(o)
+            o0, o = o, Signal.like(o)
+            if delay < js[0]:
+                self.sync += o0.eq(o + m)
+                delay += 1
+            else:
+                self.comb += o0.eq(o + m)
+            assert js[0] - delay >= 0
+            self.sync += m.eq(c*reduce(add, [x[j - delay] for j in js]))
+        # symmetric rounding
+        if shift:
+            self.comb += o.eq((1 << shift - 1) - 1)
 
 
 class ParallelFIR(Module):
     """Full-rate parallelized finite impulse response filter.
 
-    :param coefficients: integer taps.
+    Tries to use transposed form as much as possible.
+
+    :param coefficients: integer taps, increasing delay.
     :param parallelism: number of samples per cycle.
     :param width: bit width of input and output.
     :param shift: scale factor (as power of two).
@@ -86,34 +99,43 @@ class ParallelFIR(Module):
         self.width = width
         self.parallelism = p = parallelism
         n = len(coefficients)
-        # input and output: old to young, decreasing delay
+        # input and output: old to new, decreasing delay
         self.i = [Signal((width, True)) for i in range(p)]
         self.o = [Signal((width, True)) for i in range(p)]
-        self.latency = (n + 1)//2//parallelism + 3  # minus one sample
+        self.latency = (n + 1)//2//p + 2
+        # ... plus one sample
 
         ###
 
-        # Delay line: young to old, increasing delay
+        if shift is None:
+            shift = bits_for(sum(abs(c) for c in coefficients)) - 1
+
+        # Delay line: increasing delay
         x = [Signal((width, True)) for _ in range(n + p - 1)]
         self.sync += [xi.eq(xj) for xi, xj in zip(x, self.i[::-1] + x)]
 
-        if shift is None:
-            shift = width - 1
-
-        for j in range(p):
+        for delay in range(p):
+            o = Signal((width + shift + 1, True))
+            self.comb += self.o[delay].eq(o >> shift)
             # Make products
-            o = []
             for i, c in enumerate(coefficients):
                 # simplify for halfband and symmetric filters
-                if c == 0 or c in coefficients[i + 1:]:
+                if not c or c in coefficients[:i]:
                     continue
-                m = Signal((width + shift, True))
-                self.sync += m.eq(c*reduce(add, [
-                    xj for xj, cj in zip(x[-1 - j::-1], coefficients) if cj == c
-                ]))
-                o.append(m)
-            # Make sum
-            self.sync += self.o[j].eq(reduce(add, o) >> shift)
+                js = [j + p - 1 for j, cj in enumerate(coefficients)
+                      if cj == c]
+                m = Signal.like(o)
+                o0, o = o, Signal.like(o)
+                if delay + p <= js[0]:
+                    self.sync += o0.eq(o + m)
+                    delay += p
+                else:
+                    self.comb += o0.eq(o + m)
+                assert js[0] - delay >= 0
+                self.sync += m.eq(c*reduce(add, [x[j - delay] for j in js]))
+            # symmetric rounding
+            if shift:
+                self.comb += o.eq((1 << shift - 1) - 1)
 
 
 def halfgen4_cascade(rate, width, order=None):