Use python ints throughout, adding float64 to test

(Fixes #25, newly ailing test for float64 depending on later PR)

Author: awf

src/gfloat/decode.py

@@ -21,6 +21,8 @@ def decode_float(fi: FormatInfo, i: int) -> FloatValue:
       ValueError:
         If :paramref:`i` is outside the range of valid code points in :paramref:`fi`.
     """
+    assert isinstance(i, int)
+
     k = fi.k
     p = fi.precision
     t = p - 1  # Trailing significand field width
@@ -56,13 +58,10 @@ def decode_float(fi: FormatInfo, i: int) -> FloatValue:
         expval = exp - expBias
         fsignificand = 1.0 + significand * 2**-t
 
-    # val: the raw value excluding specials
-    val = sign * fsignificand * 2.0**expval
-
-    # Now overwrite the raw value with specials: Infs, NaN, -0, NaN_0
+    # Handle specials: Infs, NaN, -0, NaN_0
     signed_infinity = -np.inf if signbit else np.inf
 
-    fval = val
+    fval = None
     # All-bits-special exponent (ABSE)
     if w > 0 and exp == 2**w - 1:
         min_i_with_nan = 2 ** (p - 1) - fi.num_high_nans
@@ -78,6 +77,10 @@ def decode_float(fi: FormatInfo, i: int) -> FloatValue:
         else:
             fval = np.nan
 
+    # In range - compute value
+    if fval is None:
+        fval = sign * fsignificand * 2.0**expval
+
     # Compute FloatClass
     fclass = None
     if fval == 0:

src/gfloat/formats.py

@@ -3,6 +3,20 @@
 from .block import BlockFormatInfo
 from .types import FormatInfo
 
+#: FormatInfo for IEEE-754 Binary64 format
+format_info_binary64 = FormatInfo(
+    name="binary64",
+    k=64,
+    precision=53,
+    emax=1023,
+    has_nz=True,
+    has_infs=True,
+    num_high_nans=2**52 - 1,
+    has_subnormals=True,
+    is_signed=True,
+    is_twos_complement=False,
+)
+
 #: FormatInfo for IEEE-754 Binary32 format
 format_info_binary32 = FormatInfo(
     name="binary32",
@@ -204,6 +218,7 @@ def format_info_p3109(precision: int) -> FormatInfo:
     *_fp8_formats,
     *_fp16_formats,
     format_info_binary32,
+    format_info_binary64,
 ]
 
 # ------

src/gfloat/round.py

@@ -191,8 +191,9 @@ def encode_float(fi: FormatInfo, v: float) -> int:
         isig = 0
         biased_exp = 0
     else:
-        assert fi.bits < 64  # TODO: check implementation if fi is binary64
         sig, exp = np.frexp(vpos)
+        exp = int(exp)  # All calculations in Python ints
+
         # sig in range [0.5, 1)
         sig *= 2
         exp -= 1
@@ -226,6 +227,6 @@ def encode_float(fi: FormatInfo, v: float) -> int:
         isig = (1 << t) - isig
 
     # Pack values into a single integer
-    code = (sign << (k - 1)) | (biased_exp << t) | (isig << 0)
+    code = (int(sign) << (k - 1)) | (biased_exp << t) | (isig << 0)
 
     return code

test/test_encode.py

@@ -18,6 +18,8 @@ def test_encode(fi: FormatInfo) -> None:
         step = 13
     elif fi.bits <= 32:
         step = 73013
+    elif fi.bits <= 64:
+        step = (73013 << 32) + 39
 
     for i in range(0, 2**fi.bits, step):
         fv = decode_float(fi, i)

test/test_round.py

@@ -446,6 +446,8 @@ def test_round_roundtrip(fi: FormatInfo) -> None:
         step = 13
     elif fi.bits <= 32:
         step = 73013
+    elif fi.bits <= 64:
+        step = (73013 << 32) + 39
 
     for i in range(0, 2**fi.bits, step):
         fv = decode_float(fi, i)