From e54f2c022828dd2e6a3f2bef011e0555cc5159e4 Mon Sep 17 00:00:00 2001
From: Luke Kenneth Casson Leighton <lkcl@lkcl.net>
Date: Wed, 27 Feb 2019 12:15:34 +0000
Subject: [PATCH] split out first stage normalisation to module and use it

---
 src/add/fpbase.py                |  6 ++++
 src/add/nmigen_add_experiment.py | 57 ++++++++++++++++++++++++++++++--
 src/add/test_add.py              |  2 +-
 3 files changed, 62 insertions(+), 3 deletions(-)

diff --git a/src/add/fpbase.py b/src/add/fpbase.py
index 9ee5092b..242f41d8 100644
--- a/src/add/fpbase.py
+++ b/src/add/fpbase.py
@@ -376,6 +376,12 @@ class Overflow:
 
         self.roundz = Signal(reset_less=True)
 
+    def copy(self, inp):
+        return [self.guard.eq(inp.guard),
+                self.round_bit.eq(inp.round_bit),
+                self.sticky.eq(inp.sticky),
+                self.m0.eq(inp.m0)]
+
     def elaborate(self, platform):
         m = Module()
         m.d.comb += self.roundz.eq(self.guard & \
diff --git a/src/add/nmigen_add_experiment.py b/src/add/nmigen_add_experiment.py
index 7394f2ba..360e2939 100644
--- a/src/add/nmigen_add_experiment.py
+++ b/src/add/nmigen_add_experiment.py
@@ -274,10 +274,61 @@ class FPAddStage1(FPState):
         ]
 
 
+class FPNorm1Mod:
+
+    def __init__(self, width):
+        self.out_norm = Signal(reset_less=True)
+        self.in_z = FPNumBase(width, False)
+        self.out_z = FPNumBase(width, False)
+        self.in_of = Overflow()
+        self.out_of = Overflow()
+
+    def setup(self, m, in_z, out_z, in_of, out_of, out_norm):
+        """ links module to inputs and outputs
+        """
+        m.d.comb += self.in_z.copy(in_z)
+        m.d.comb += out_z.copy(self.out_z)
+        m.d.comb += self.in_of.copy(in_of)
+        m.d.comb += out_of.copy(self.out_of)
+        m.d.comb += out_norm.eq(self.out_norm)
+
+    def elaborate(self, platform):
+        m = Module()
+        m.submodules.norm1_in_overflow = self.in_of
+        m.submodules.norm1_out_overflow = self.out_of
+        m.submodules.norm1_in_z = self.in_z
+        m.submodules.norm1_out_z = self.out_z
+        m.d.comb += self.out_z.copy(self.in_z)
+        m.d.comb += self.out_of.copy(self.in_of)
+        m.d.comb += self.out_norm.eq((self.in_z.m[-1] == 0) & \
+                                     (self.in_z.e > self.in_z.N126))
+        with m.If(self.out_norm):
+            m.d.comb += [
+                self.out_z.e.eq(self.in_z.e - 1),  # DECREASE exponent
+                self.out_z.m.eq(self.in_z.m << 1), # shift mantissa UP
+                self.out_z.m[0].eq(self.in_of.guard), # steal guard (was tot[2])
+                self.out_of.guard.eq(self.in_of.round_bit), # round (was tot[1])
+                self.out_of.round_bit.eq(0),        # reset round bit
+                self.out_of.m0.eq(self.in_of.guard),
+            ]
+
+        return m
+
+
 class FPNorm1(FPState):
 
+    def __init__(self, width):
+        FPState.__init__(self, "normalise_1")
+        self.mod = FPNorm1Mod(width)
+        self.out_norm = Signal(reset_less=True)
+        self.out_z = FPNumBase(width)
+        self.out_of = Overflow()
+
     def action(self, m):
-        self.normalise_1(m, self.z, self.of, "normalise_2")
+        m.d.sync += self.of.copy(self.out_of)
+        m.d.sync += self.z.copy(self.out_z)
+        with m.If(~self.out_norm):
+            m.next = "normalise_2"
 
 
 class FPNorm2(FPState):
@@ -473,9 +524,11 @@ class FPADD:
         add1.set_inputs({"tot": tot, "z": z}) # Z input passes through
         add1.set_outputs({"z": z, "of": of})  # XXX Z as output
 
-        n1 = self.add_state(FPNorm1("normalise_1"))
+        n1 = self.add_state(FPNorm1(self.width))
         n1.set_inputs({"z": z, "of": of})  # XXX Z as output
         n1.set_outputs({"z": z})  # XXX Z as output
+        n1.mod.setup(m, z, n1.out_z, of, n1.out_of, n1.out_norm)
+        m.submodules.normalise_1 = n1.mod
 
         n2 = self.add_state(FPNorm2("normalise_2"))
         n2.set_inputs({"z": z, "of": of})  # XXX Z as output
diff --git a/src/add/test_add.py b/src/add/test_add.py
index 39c779db..bd430c61 100644
--- a/src/add/test_add.py
+++ b/src/add/test_add.py
@@ -67,6 +67,6 @@ def testbench(dut):
     yield from run_edge_cases(dut, count, add)
 
 if __name__ == '__main__':
-    dut = FPADD(width=32, single_cycle=True)
+    dut = FPADD(width=32, single_cycle=False)
     run_simulation(dut, testbench(dut), vcd_name="test_add.vcd")
 
-- 
2.30.2