From: Luke Kenneth Casson Leighton <lkcl@lkcl.net>
Date: Mon, 3 Feb 2020 11:51:51 +0000 (+0000)
Subject: add first untested ge.py for partitioned signnal
X-Git-Tag: ls180-24jan2020~291
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=6a42f101b9735924d3b936b58b62dcb2b2e27152;p=ieee754fpu.git

add first untested ge.py for partitioned signnal
---

diff --git a/src/ieee754/part_cmp/ge.py b/src/ieee754/part_cmp/ge.py
new file mode 100644
index 00000000..8c3efb31
--- /dev/null
+++ b/src/ieee754/part_cmp/ge.py
@@ -0,0 +1,117 @@
+# 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 "comparison" class, directly equivalent
+to Signal.__ge__ except SIMD-partitionable
+
+See:
+
+* http://libre-riscv.org/3d_gpu/architecture/dynamic_simd/ge
+* http://bugs.libre-riscv.org/show_bug.cgi?id=132
+"""
+
+from nmigen import Signal, Module, Elaboratable, Cat, C, Mux, Repl
+from nmigen.cli import main
+
+from ieee754.part_mul_add.partpoints import PartitionPoints
+
+def create_ge(nes, les, start, count):
+    """create a greater-than-or-equal from partitioned eqs and greaterthans
+
+        this works by doing:  lt3 |
+                             (lt2 & eq3) |
+                             (lt1 & eq3 & eq2) |
+                             (lt0 & eq3 & eq2 & eq1)
+    """
+    res = []
+    for i in range(count-1):
+        ands = [les[start+count-1]] # always one, and it's the end one
+        res.append(les[start+i])
+
+
+class PartitionedGe(Elaboratable):
+
+    def __init__(self, width, partition_points):
+        """Create a ``PartitionedGe`` operator
+        """
+        self.width = width
+        self.a = Signal(width, reset_less=True)
+        self.b = 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
+
+        # see equal.py notes
+
+        # prepare the output bits (name them for convenience)
+        gtsigs = []
+        for i in range(self.mwidth):
+            gtsig = Signal(name="gtsig%d"%i, reset_less=True)
+            gtsigs.append(gtsig)
+
+        # make series of !eqs and !gts, splitting a and b into partition chunks
+        nes = Signal(self.mwidth, reset_less=True)
+        les = Signal(self.mwidth, reset_less=True)
+        nel = []
+        lel = []
+        keys = list(self.partition_points.keys()) + [self.width]
+        start = 0
+        for i in range(len(keys)):
+            end = keys[i]
+            nel.append(self.a[start:end] != self.b[start:end]) # see bool below
+            lel.append(self.a[start:end] <= self.b[start:end]) # see bool below
+            start = end # for next time round loop
+        comb += nes.eq(Cat(*nel))
+        comb += les.eq(Cat(*nel))
+
+        # now, based on the partition points, create the (multi-)boolean result
+        # this is a terrible way to do it, it's very laborious.  however it
+        # will actually "work".  optimisations come later
+
+        # we want just the partition points, as a number
+        ppoints = Signal(self.mwidth-1)
+        comb += ppoints.eq(self.partition_points.as_sig())
+
+        with m.Switch(ppoints):
+            for pval in range(1<<(self.mwidth-1)): # for each partition point
+                # identify (find-first) transition points, and how
+                # long each partition is
+                start = 0
+                count = 1
+                idx = [0] * self.mwidth
+                for ipdx in range((self.mwidth-1)):
+                    if (pval & (1<<ipdx)):
+                        idx[start] = count
+                        start = ipdx + 1
+                        count = 1
+                    else:
+                        count += 1
+                idx[start] = count # update last point (or create it)
+
+                # now for each partition combination,
+                with m.Case(pval):
+                    #print (pval, bin(pval), idx)
+                    for i in range(self.mwidth):
+                        n = "andsig_%d_%d" % (pval, i)
+                        if not idx[start]:
+                            continue
+                        ands = create_ge(nes, les, i, idx[start])
+                        andsig = Signal(len(ands), name=n, reset_less=True)
+                        ands = ands.bool() # create an AND cascade
+                        #print ("ands", pval, i, ands)
+                        comb += andsig.eq(ands)
+                        comb += gtsigs[i].eq(~andsig) # here's the inversion
+
+        # assign cascade-SIMD-compares to output
+        comb += self.output.eq(Cat(*gtsigs))
+
+        return m