vectorize cast function; type annotation changes; add to exclusion list

Signed-off-by: Yuan Yao <[email protected]>

docs/Operators.md

@@ -3718,7 +3718,7 @@ for from_type, to_type in test_cases:
                 "-INF",
                 "-4",
                 "0.01",
-                "-1000000",
+                "-0.0",
             ],
             dtype=np.float32,
         )
@@ -3746,7 +3746,7 @@ for from_type, to_type in test_cases:
             raise ValueError(
                 f"Conversion from {from_type} to {to_type} is not tested."
             )
-        expected = evaluate_float4e2m1_from_bits(
+        expected = unpacked_float4e2m1_to_float32(
             subbyte.float32_to_float4e2m1_unpacked(np_fp32)
         )
         output = make_tensor(

docs/TestCoverage.md

@@ -2587,7 +2587,7 @@ for from_type, to_type in test_cases:
                 "-INF",
                 "-4",
                 "0.01",
-                "-1000000",
+                "-0.0",
             ],
             dtype=np.float32,
         )
@@ -2615,7 +2615,7 @@ for from_type, to_type in test_cases:
             raise ValueError(
                 f"Conversion from {from_type} to {to_type} is not tested."
             )
-        expected = evaluate_float4e2m1_from_bits(
+        expected = unpacked_float4e2m1_to_float32(
             subbyte.float32_to_float4e2m1_unpacked(np_fp32)
         )
         output = make_tensor(

onnx/backend/test/case/node/cast.py

@@ -19,9 +19,9 @@
     tensor_dtype_to_field,
 )
 from onnx.numpy_helper import (
-    evaluate_float4e2m1_from_bits,
     float8e4m3_to_float32,
     float8e5m2_to_float32,
+    unpacked_float4e2m1_to_float32,
 )
 
 
@@ -303,7 +303,7 @@ def export() -> None:
                         "-INF",
                         "-4",
                         "0.01",
-                        "-1000000",
+                        "-0.0",
                     ],
                     dtype=np.float32,
                 )
@@ -331,7 +331,7 @@ def export() -> None:
                     raise ValueError(
                         f"Conversion from {from_type} to {to_type} is not tested."
                     )
-                expected = evaluate_float4e2m1_from_bits(
+                expected = unpacked_float4e2m1_to_float32(
                     subbyte.float32_to_float4e2m1_unpacked(np_fp32)
                 )
                 output = make_tensor(

onnx/backend/test/data/node/test_cast_FLOAT16_to_FLOAT4E2M1/test_data_set_0/input_0.pb

@@ -1,2 +1,2 @@
 
-*'�o�h�x������
��
��
��
��
�����B��B
x
+*'�o�h�x������
��
��
��
��
�����B��B
x

onnx/backend/test/data/node/test_cast_FLOAT16_to_FLOAT4E2M1/test_data_set_0/output_0.pb

@@ -1,2 +1,2 @@
 *
-
�
w�
B
y
+
�
w�
B
y

onnx/backend/test/data/node/test_cast_FLOAT4E2M1_to_FLOAT/test_data_set_0/input_0.pb

@@ -1,2 +1,2 @@
 *
-
�
w�
B
x
+
�
w�
B
x

onnx/backend/test/data/node/test_cast_FLOAT4E2M1_to_FLOAT16/test_data_set_0/input_0.pb

@@ -1,2 +1,2 @@
 *
-
�
w�
B
x
+
�
w�
B
x

onnx/backend/test/data/node/test_cast_FLOAT_to_FLOAT4E2M1/test_data_set_0/output_0.pb

@@ -1,2 +1,2 @@
 *
-
�
w�
B
y
+
�
w�
B
y

onnx/numpy_helper.py

@@ -221,22 +221,26 @@ def unpack_int4(
     return res
 
 
-def evaluate_float4e2m1_from_bits(x: np.ndarray[np.uint8]) -> np.ndarray[np.float32]:
-    """Evaluate the numerical value of a single float4e2m1 element represented as uint8
+def unpacked_float4e2m1_to_float32(x: np.ndarray) -> np.ndarray:
+    """Evaluate the numerical value of an array of unpacked float4e2m1 values (as uint8)
     See :ref:`onnx-detail-int4` for technical details.
 
     Args:
-        x: a uint8 element representing a float4e2m1 (using the 4 LSB)
+        x: an array of uint8 elements representing a float4e2m1 (using the 4 LSB)
 
     Returns:
-        A float32 element representing the value of the float4e2m1 input.
+        An array of float32 elements representing the values of the float4e2m1 input.
     """
     # x is stored in 4 LSB of int
-    S = np.where(np.bitwise_and(x, 0x08), -1, 1)
-    M = x & 0x01
-    E = (x & 0x06) >> 1
-
-    val = np.where(E==0, S*(M/2.0), S*(1.0+M/2.0) *2.0 **(E-1))  # denormalized, normalized
+    sign = np.where(np.bitwise_and(x, 0x08), -1, 1)
+    mantissa = x & 0x01
+    exponent = (x & 0x06) >> 1
+
+    val = np.where(
+        exponent == 0,
+        sign * (mantissa / 2.0),
+        sign * (1.0 + mantissa / 2.0) * 2.0 ** (exponent - 1),
+    )  # denormalized, normalized
     return val
 
 
@@ -258,8 +262,8 @@ def unpack_float4e2m1(
     res_high, res_low = subbyte.unpack_single_4bitx2(data.ravel(), False)
     res = np.empty((res_high.size + res_low.size,), dtype=np.float32)
 
-    res[0::2] = evaluate_float4e2m1_from_bits(res_high)
-    res[1::2] = evaluate_float4e2m1_from_bits(res_low)
+    res[0::2] = unpacked_float4e2m1_to_float32(res_high)
+    res[1::2] = unpacked_float4e2m1_to_float32(res_low)
 
     if (
         res.size == np.prod(dims) + 1

onnx/reference/ops/op_cast.py

@@ -24,9 +24,9 @@
 )
 from onnx.numpy_helper import (
     bfloat16_to_float32,
-    evaluate_float4e2m1_from_bits,
     float8e4m3_to_float32,
     float8e5m2_to_float32,
+    unpacked_float4e2m1_to_float32,
 )
 from onnx.onnx_pb import TensorProto
 from onnx.reference.op_run import OpRun
@@ -131,7 +131,7 @@ def cast_to(x, to, saturate):  # noqa: PLR0911
     if x.dtype == float4e2m1 and x.dtype.descr[0][0] == "float4e2m1":
         if to == TensorProto.FLOAT4E2M1:
             return x
-        res = evaluate_float4e2m1_from_bits(x)
+        res = unpacked_float4e2m1_to_float32(x)
         if to == TensorProto.FLOAT:
             return res.astype(np.float32)
         elif to == TensorProto.FLOAT16:

onnx/reference/ops/op_dequantize_linear.py

@@ -17,9 +17,9 @@
 )
 from onnx.helper import np_dtype_to_tensor_dtype
 from onnx.numpy_helper import (
-    evaluate_float4e2m1_from_bits,
     float8e4m3_to_float32,
     float8e5m2_to_float32,
+    unpacked_float4e2m1_to_float32,
 )
 from onnx.reference.op_run import OpRun
 from onnx.reference.ops.op_quantize_linear import reshape_input
@@ -93,7 +93,7 @@ def _run(
             elif x_type == TensorProto.FLOAT8E5M2FNUZ:
                 dx = float8e5m2_to_float32(x, fn=True, uz=True)
             elif x_type == TensorProto.FLOAT4E2M1:
-                dx = evaluate_float4e2m1_from_bits(x)
+                dx = unpacked_float4e2m1_to_float32(x)
             else:
                 dx = x.astype(np.float32)
         y = dx * reshape_input(x_scale, x.shape, axis, block_size)

onnx/subbyte.py

@@ -72,7 +72,7 @@ def unpack_single_4bitx2(
     return (x_low.astype(dtype), x_high.astype(dtype))
 
 
-def float32_to_float4e2m1_unpacked(x: np.ndarray | np.dtype) -> np.ndarray:
+def float32_to_float4e2m1_unpacked_slow(x: np.ndarray | np.dtype) -> np.ndarray:
     """Cast float32 to float4e2m1 (without packing).
 
     Args:
@@ -85,7 +85,7 @@ def float32_to_float4e2m1_unpacked(x: np.ndarray | np.dtype) -> np.ndarray:
     def float32_to_float4e2m1(value):
         if np.isnan(value):
             return 0x7
-        s = 0x0 if value >= 0 else 0x8
+        s = 0x8 if np.signbit(value) else 0x0
         magnitude = np.abs(value)
         if np.isinf(magnitude):
             ret = 0x7
@@ -116,14 +116,38 @@ def float32_to_float4e2m1(value):
     return y.astype(np.uint8)  # type: ignore[no-any-return]
 
 
-def float32x2_to_float4e2m1x2(val_low: np.dtype, val_high: np.dtype) -> np.ndarray:
+def float32_to_float4e2m1_unpacked(values: np.ndarray) -> np.ndarray:
+    """Cast float32 to float4e2m1 (without packing).
+
+    Args:
+        values: element or array to be converted
+
+    Returns:
+        An ndarray with unpacked float4e2m1 elements (as uint8)
+    """
+    sign = np.where(np.signbit(values), 0x8, 0x0).astype(np.uint8)
+    magnitude = np.abs(values)
+    res = np.zeros(values.shape, dtype=np.uint8)
+    res[(magnitude > 0.25) & (magnitude < 0.75)] = 0x1
+    res[(magnitude >= 0.75) & (magnitude <= 1.25)] = 0x2
+    res[(magnitude > 1.25) & (magnitude < 1.75)] = 0x3
+    res[(magnitude >= 1.75) & (magnitude <= 2.5)] = 0x4
+    res[(magnitude > 2.5) & (magnitude < 3.5)] = 0x5
+    res[(magnitude >= 3.5) & (magnitude <= 5.0)] = 0x6
+    res[magnitude > 5.0] = 0x7
+    res |= sign
+    res[np.isnan(values)] = 0x7
+    return res
+
+
+def float32x2_to_float4e2m1x2(val_low: np.ndarray, val_high: np.ndarray) -> np.ndarray:
     """Cast two elements to float4e2m1 and pack to a single byte
     Args:
         val_low: element to be packed in the 4 LSB
         val_high: element to be packed in the 4 MSB
 
     Returns:
-        An ndarray with a single uint8 element, containing both float4e2m1 elements
+        An ndarray with uint8 elements, containing both float4e2m1 elements
     """
     i8_high = float32_to_float4e2m1_unpacked(val_high)
     i8_low = float32_to_float4e2m1_unpacked(val_low)

onnx/test/test_backend_onnxruntime.py

@@ -164,6 +164,10 @@ def run_node(cls, node, inputs, device=None, outputs_info=None, **kwargs):
         "|test_cast_UINT4_to_FLOAT16"  # No corresponding Numpy type for Tensor Type.
         "|test_cast_INT4_to_FLOAT16"  # No corresponding Numpy type for Tensor Type.
         "|test_maxpool_2d_ceil_output_size_reduce_by_one"  # TODO: remove after https://github.com/microsoft/onnxruntime/pull/18377 in Ort release.
+        "|test_quantizeLinear_float4e2m1"  # No corresponding Numpy type for Tensor Type.
+        "|test_dequantizeLinear_float4e2m1"  # No corresponding Numpy type for Tensor Type.
+        "|cast_float4e2m1"  # No corresponding Numpy type for Tensor Type.
+        "|to_float4e2m1"  # No corresponding Numpy type for Tensor Type.
         ")"
     )