Adds validation for float 8

Author: xadupre

_unittests/ut_validation/test_f8.py

@@ -237,7 +237,10 @@ def test_search_float32_into_fe5m2(self):
                         else:
                             add = v - value
                             if len(w) > 0:
-                                raise AssertionError(f"A warning was thrown for v={v}, value={value}, w={w[0]}.")
+                                raise AssertionError(
+                                    f"A warning was thrown for v={v}, "
+                                    f"value={value}, w={w[0]}."
+                                )
                 else:
                     v = value + add
                 b = search_float32_into_fe5m2(v)
@@ -306,9 +309,9 @@ def test_inf_nan(self):
                 0.203125,
                 0.75,
                 numpy.nan,
-                numpy.nan,
-                -numpy.nan,
-                -numpy.nan,
+                max(CastFloat8.values_e4m3fn)[0],
+                max(CastFloat8.values_e4m3fn)[0],
+                min(CastFloat8.values_e4m3fn)[0],
             ],
             dtype=numpy.float32,
         )
@@ -380,26 +383,27 @@ def test_search_e5m2_pow(self):
                 )
 
     def test_float32_to_fe4m3fn_inf(self):
-        mx = 
-        v0 = numpy.float32(448)
+        mx = max(CastFloat8.values_e4m3fn)[0]
+        v0 = numpy.float32(mx)
         v1 = numpy.float32(numpy.inf)
         a = search_float32_into_fe4m3(v0)
         b = search_float32_into_fe4m3(v1)
         self.assertEqual(a, b)
 
-        v0 = numpy.float32(448)
+        v0 = numpy.float32(mx)
         v1 = numpy.float32(numpy.inf)
         a = float32_to_fe4m3(v0)
         b = float32_to_fe4m3(v1)
         self.assertEqual(a, b)
 
-        v0 = numpy.float32(-448)
+        mi = min(CastFloat8.values_e4m3fn)[0]
+        v0 = numpy.float32(mi)
         v1 = numpy.float32(-numpy.inf)
         a = search_float32_into_fe4m3(v0)
         b = search_float32_into_fe4m3(v1)
         self.assertEqual(a, b)
 
-        v0 = numpy.float32(-448)
+        v0 = numpy.float32(mi)
         v1 = numpy.float32(-numpy.inf)
         a = float32_to_fe4m3(v0)
         b = float32_to_fe4m3(v1)
@@ -666,18 +670,32 @@ def test_float32_to_fe4m3fnuz_inf(self):
         self.assertNotEqual(a, b)
 
     def test_float32_to_fe5m2fnuz_inf(self):
-        v0 = numpy.float32(65536)
+        mx = max(CastFloat8.values_e5m2fnuz)[0]
+        v0 = numpy.float32(mx)
         v1 = numpy.float32(numpy.inf)
         a = search_float32_into_fe5m2(v0, fn=True, uz=True)
         b = search_float32_into_fe5m2(v1, fn=True, uz=True)
         self.assertEqual(a, b)
 
-        v0 = numpy.float32(65536)
+        v0 = numpy.float32(mx)
         v1 = numpy.float32(numpy.inf)
         a = float32_to_fe5m2(v0, fn=True, uz=True)
         b = float32_to_fe5m2(v1, fn=True, uz=True)
         self.assertEqual(a, b)
 
+        mi = min(CastFloat8.values_e5m2fnuz)[0]
+        v0 = numpy.float32(mi)
+        v1 = numpy.float32(-numpy.inf)
+        a = search_float32_into_fe5m2(v0, fn=True, uz=True)
+        b = search_float32_into_fe5m2(v1, fn=True, uz=True)
+        self.assertEqual(a, b)
+
+        v0 = numpy.float32(mi)
+        v1 = numpy.float32(-numpy.inf)
+        a = float32_to_fe5m2(v0, fn=True, uz=True)
+        b = float32_to_fe5m2(v1, fn=True, uz=True)
+        self.assertEqual(a, b)
+
         v0 = numpy.float32(numpy.nan)
         v1 = numpy.float32(-numpy.nan)
         a = search_float32_into_fe5m2(v0, fn=True, uz=True)
@@ -688,7 +706,7 @@ def test_float32_to_fe5m2fnuz_inf(self):
         v1 = numpy.float32(-numpy.inf)
         a = search_float32_into_fe5m2(v0, fn=True, uz=True)
         b = search_float32_into_fe5m2(v1, fn=True, uz=True)
-        self.assertEqual(a, b)
+        self.assertNotEqual(a, b)
 
         v0 = numpy.float32(numpy.nan)
         v1 = numpy.float32(-numpy.nan)
@@ -700,7 +718,7 @@ def test_float32_to_fe5m2fnuz_inf(self):
         v1 = numpy.float32(-numpy.inf)
         a = float32_to_fe5m2(v0, fn=True, uz=True)
         b = float32_to_fe5m2(v1, fn=True, uz=True)
-        self.assertEqual(a, b)
+        self.assertNotEqual(a, b)
 
     def test_simple_fe4m3(self):
         values = [448]

onnx_array_api/validation/f8.py

@@ -378,12 +378,16 @@ def search_float32_into_fe4m3(value: float, fn: bool = True, uz: bool = False) -
     b = int.from_bytes(struct.pack("<f", numpy.float32(value)), "little")
     ret = (b & 0x80000000) >> 24  # sign
     if uz:
-        if numpy.isnan(value) or numpy.isinf(value):
+        if numpy.isnan(value):
             return 0x80
+        if numpy.isinf(value):
+            return ret | 0x7F
         set_values = CastFloat8.values_e4m3fnuz
     else:
-        if numpy.isnan(value) or numpy.isinf(value):
+        if numpy.isnan(value):
             return 0x7F | ret
+        if numpy.isinf(value):
+            return 0x7E | ret
         set_values = CastFloat8.values_e4m3fn
     f = numpy.float32(value)
     i = CastFloat8.find_closest_value(f, set_values)
@@ -403,8 +407,10 @@ def search_float32_into_fe5m2(value: float, fn: bool = False, uz: bool = False)
     ret = (b & 0x80000000) >> 24  # sign
 
     if fn and uz:
-        if numpy.isnan(value) or numpy.isinf(value):
+        if numpy.isnan(value):
             return 0x80
+        if numpy.isinf(value):
+            return ret | 0x7F
         set_values = CastFloat8.values_e5m2fnuz
     elif not fn and not uz:
         if numpy.isnan(value):
@@ -435,7 +441,7 @@ def float32_to_fe4m3(x, fn: bool = True, uz: bool = False):
         if (b & 0x7FC00000) == 0x7FC00000:
             return 0x80
         if numpy.isinf(x):
-            return 0x80
+            return ret | 0x7F  # saturation
         e = (b & 0x7F800000) >> 23  # exponent
         m = b & 0x007FFFFF  # mantissa
 
@@ -472,7 +478,7 @@ def float32_to_fe4m3(x, fn: bool = True, uz: bool = False):
         if (b & 0x7FC00000) == 0x7FC00000:
             return 0x7F | ret
         if numpy.isinf(x):
-            return 0x7F | ret
+            return 0x7E | ret  # saturation
         e = (b & 0x7F800000) >> 23  # exponent
         m = b & 0x007FFFFF  # mantissa
 
@@ -524,6 +530,8 @@ def float32_to_fe5m2(x, fn: bool = False, uz: bool = False):
     if fn and uz:
         if (b & 0x7FC00000) == 0x7FC00000:
             return 0x80
+        if (b & 0x7FFFFFFF) == 0x7F800000:
+            return ret | 0x7F
         e = (b & 0x7F800000) >> 23  # exponent
         m = b & 0x007FFFFF  # mantissa