From: Luke Kenneth Casson Leighton <lkcl@lkcl.net>
Date: Tue, 5 Oct 2021 17:09:24 +0000 (+0100)
Subject: add PartitionedRepl first version, no unit test just demo
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=4842290ca7444952b0a70984e8b5afa2d4a51fa0;p=ieee754fpu.git

add PartitionedRepl first version, no unit test just demo
---

diff --git a/src/ieee754/part_repl/repl.py b/src/ieee754/part_repl/repl.py
new file mode 100644
index 00000000..364b7721
--- /dev/null
+++ b/src/ieee754/part_repl/repl.py
@@ -0,0 +1,188 @@
+# SPDX-License-Identifier: LGPL-2.1-or-later
+# See Notices.txt for copyright information
+
+"""
+Copyright (C) 2021 Luke Kenneth Casson Leighton <lkcl@lkcl.net>
+
+dynamically-partitionable "repl" class, directly equivalent
+to nmigen Repl
+
+See:
+
+* http://libre-riscv.org/3d_gpu/architecture/dynamic_simd/repl
+* http://bugs.libre-riscv.org/show_bug.cgi?id=709
+
+"""
+
+from nmigen import Signal, Module, Elaboratable, Cat, Repl
+from nmigen.back.pysim import Simulator, Settle
+from nmigen.cli import rtlil
+
+from ieee754.part_mul_add.partpoints import PartitionPoints
+from ieee754.part.partsig import PartitionedSignal
+
+
+def get_runlengths(pbit, size):
+    res = []
+    count = 1
+    # identify where the 1s are, which indicates "start of a new partition"
+    # we want a list of the lengths of all partitions
+    for i in range(size):
+        if pbit & (1<<i): # it's a 1: ends old partition, starts new
+            res.append(count) # add partition
+            count = 1 # start again
+        else:
+            count += 1
+    # end reached, add whatever is left. could have done this by creating
+    # "fake" extra bit on the partitions, but hey
+    res.append(count)
+
+    print ("get_runlengths", bin(pbit), size, res)
+
+    return res
+
+
+class PartitionedRepl(Elaboratable):
+    def __init__(self, repl, qty, mask):
+        """Create a ``PartitionedRepl`` operator
+        """
+        # work out the length (total of all PartitionedSignals)
+        self.repl = repl
+        self.qty = qty
+        width, signed = repl.shape()
+        if isinstance(mask, dict):
+            mask = list(mask.values())
+        self.mask = mask
+        self.shape = (width * qty), signed
+        self.output = PartitionedSignal(mask, self.shape, reset_less=True)
+        self.partition_points = self.output.partpoints
+        self.mwidth = len(self.partition_points)+1
+
+    def get_chunk(self, y, numparts):
+        x = self.repl
+        if not isinstance(x, PartitionedSignal):
+            # assume Scalar. totally different rules
+            end = numparts * (len(x) // self.mwidth)
+            return x[:end]
+        # PartitionedSignal: start at partition point
+        keys = [0] + list(x.partpoints.keys()) + [len(x)]
+        # get current index and increment it (for next Repl chunk)
+        upto = y[0]
+        y[0] += numparts
+        print ("getting", upto, numparts, keys, len(x))
+        # get the partition point as far as we are up to
+        start = keys[upto]
+        end = keys[upto+numparts]
+        print ("start end", start, end, len(x))
+        return x[start:end]
+
+    def elaborate(self, platform):
+        m = Module()
+        comb = m.d.comb
+
+        keys = list(self.partition_points.keys())
+        print ("keys", keys, "values", self.partition_points.values())
+        print ("mask", self.mask)
+        outpartsize = len(self.output) // self.mwidth
+        width, signed = self.output.shape()
+        print ("width, signed", width, signed)
+
+        with m.Switch(Cat(self.mask)):
+            # for each partition possibility, create a Repl sequence
+            for pbit in range(1<<len(keys)):
+                # set up some indices pointing to where things have got
+                # then when called below in the inner nested loop they give
+                # the relevant sequential chunk
+                output = []
+                y = [0]
+                # get a list of the length of each partition run
+                runlengths = get_runlengths(pbit, len(keys))
+                print ("pbit", bin(pbit), "runs", runlengths)
+                for i in runlengths:                     # for each partition
+                    thing = self.get_chunk(y, i)         # get sequential chunk
+                    output.append(Repl(thing, self.qty)) # and replicate it
+                with m.Case(pbit):
+                    # direct access to the underlying Signal
+                    comb += self.output.sig.eq(Cat(*output)) # cat all chunks
+
+        return m
+
+    def ports(self):
+        if isinstance(self.repl, PartitionedSignal):
+            return [self.repl.lower(), self.output.lower()]
+        return [self.repl, self.output.lower()]
+
+
+if __name__ == "__main__":
+    from ieee754.part.test.test_partsig import create_simulator
+    m = Module()
+    mask = Signal(3)
+    a = PartitionedSignal(mask, 32)
+    m.submodules.repl = repl = PartitionedRepl(a, 2, mask)
+    omask = (1<<len(repl.output))-1
+
+    traces = repl.ports()
+    vl = rtlil.convert(repl, ports=traces)
+    with open("part_repl.il", "w") as f:
+        f.write(vl)
+
+    sim = create_simulator(m, traces, "partrepl")
+
+    def process():
+        yield mask.eq(0b000)
+        yield a.sig.eq(0xa12345c7)
+        yield Settle()
+        out = yield repl.output.sig
+        print("out 000", bin(out&omask), hex(out&omask))
+        yield mask.eq(0b010)
+        yield Settle()
+        out = yield repl.output.sig
+        print("out 010", bin(out&omask), hex(out&omask))
+        yield mask.eq(0b110)
+        yield Settle()
+        out = yield repl.output.sig
+        print("out 110", bin(out&omask), hex(out&omask))
+        yield mask.eq(0b111)
+        yield Settle()
+        out = yield repl.output.sig
+        print("out 111", bin(out&omask), hex(out&omask))
+
+    sim.add_process(process)
+    with sim.write_vcd("partition_repl.vcd", "partition_repl.gtkw",
+                        traces=traces):
+        sim.run()
+
+    # Scalar
+    m = Module()
+    mask = Signal(3)
+    a = Signal(32)
+    m.submodules.ass = ass = PartitionedRepl(a, 2, mask)
+    omask = (1<<len(ass.output))-1
+
+    traces = ass.ports()
+    sim = create_simulator(m, traces, "partass")
+
+    def process():
+        yield mask.eq(0b000)
+        yield a.eq(0xa12345c7)
+        yield Settle()
+        out = yield ass.output.sig
+        print("out 000", bin(out&omask), hex(out&omask))
+        yield mask.eq(0b010)
+        yield Settle()
+        out = yield ass.output.sig
+        print("out 010", bin(out&omask), hex(out&omask))
+        yield mask.eq(0b110)
+        yield Settle()
+        out = yield ass.output.sig
+        print("out 110", bin(out&omask), hex(out&omask))
+        yield mask.eq(0b111)
+        yield Settle()
+        out = yield ass.output.sig
+        print("out 111", bin(out&omask), hex(out&omask))
+
+    sim.add_process(process)
+    with sim.write_vcd("partition_repl_scalar.vcd",
+                       "partition_repl_scalar.gtkw",
+                        traces=traces):
+        sim.run()