add first cut (untested) of PartitionedXOR
authorLuke Kenneth Casson Leighton <lkcl@lkcl.net>
Sat, 16 Jan 2021 18:51:34 +0000 (18:51 +0000)
committerLuke Kenneth Casson Leighton <lkcl@lkcl.net>
Sat, 16 Jan 2021 18:51:34 +0000 (18:51 +0000)
src/ieee754/part_bits/xor.py [new file with mode: 0644]

diff --git a/src/ieee754/part_bits/xor.py b/src/ieee754/part_bits/xor.py
new file mode 100644 (file)
index 0000000..c5471f5
--- /dev/null
@@ -0,0 +1,57 @@
+# SPDX-License-Identifier: LGPL-2.1-or-later
+# See Notices.txt for copyright information
+
+"""
+Copyright (C) 2020 Luke Kenneth Casson Leighton <lkcl@lkcl.net>
+
+dynamically-partitionable "xor" class, directly equivalent
+to Signal.xor() except SIMD-partitionable
+
+See:
+
+* http://libre-riscv.org/3d_gpu/architecture/dynamic_simd/logicops
+* http://bugs.libre-riscv.org/show_bug.cgi?id=176
+"""
+
+from nmigen import Signal, Module, Elaboratable, Cat, C, Mux, Repl
+from nmigen.cli import main
+
+from ieee754.part_mul_add.partpoints import PartitionPoints
+from ieee754.part_cmp.experiments.eq_combiner import XORCombiner
+
+
+class PartitionedXOR(Elaboratable):
+
+    def __init__(self, width, partition_points):
+        """Create a ``PartitionedXOR`` operator
+        """
+        self.width = width
+        self.a = Signal(width, reset_less=True)
+        self.partition_points = PartitionPoints(partition_points)
+        self.mwidth = len(self.partition_points)+1
+        self.output = Signal(self.mwidth, reset_less=True)
+        if not self.partition_points.fits_in_width(width):
+            raise ValueError("partition_points doesn't fit in width")
+
+    def elaborate(self, platform):
+        m = Module()
+        comb = m.d.comb
+        m.submodules.xorc = xorc = XORCombiner(self.mwidth)
+
+        # make a series of "xor", splitting a and b into partition chunks
+        xors = Signal(self.mwidth, reset_less=True)
+        xorl = []
+        keys = list(self.partition_points.keys()) + [self.width]
+        start = 0
+        for i in range(len(keys)):
+            end = keys[i]
+            xorl.append(self.a[start:end].xor())
+            start = end # for next time round loop
+        comb += xors.eq(Cat(*xorl))
+
+        # put the partial results through the combiner
+        comb += xorc.gates.eq(self.partition_points.as_sig())
+        comb += xorc.neqs.eq(xors)
+        comb += self.output.eq(xorc.outputs)
+
+        return m