[BUG] cuNumeric and Numpy get different output on BinaryOp nextafter when the input data type is float16

[XiaLuNV]

Software versions

Python : 3.10.9 | packaged by conda-forge
Platform : Linux 5.15.0-83-generic #92-Ubuntu
Cunumeric : 24.01
Numpy : 1.23.5
CTK package : cuda-version-12.4
GPU driver : 550.00
GPU devices :
GPU 0: NVIDIA H100
GPU 1: NVIDIA H100

Jupyter notebook / Jupyter Lab version

No response

Expected behavior

For tests
https://github.com/nv-legate/cunumeric/blob/branch-24.01/tests/integration/test_binary_ufunc.py
for op = 'nextafter'
When the type of input data is float16, the results between Numpy and cuNumeric are different.
It failed on allclose(out_np, out_num).

Observed behavior

First 5 differences for allclose:
Numpy cuNumeric
index (0, 0): 0.07916259765625 0.0791015625
index (0, 1): 0.373291015625 0.373046875
index (0, 2): 0.93408203125 0.93359375
index (0, 3): 0.4189453125 0.418701171875
index (0, 4): 0.234375 0.2342529296875

With diff_limit=5

cunumeric.nextafter failed the test
Inputs:
[[0.0791 0.373 0.9336 0.4187 0.2343 ]
[0.5723 0.5723 0.417 0.626 0.2203 ]
[0.622 0.4778 0.974 0.773 0.02715]
[0.2211 0.1203 0.1753 0.4294 0.6577 ]]
dtype: float16
2
dtype: uint64
NumPy output:
[[0.07916 0.3733 0.934 0.419 0.2344 ]
[0.5728 0.5728 0.4172 0.6265 0.2205 ]
[0.6226 0.478 0.9746 0.7734 0.02716]
[0.2212 0.12036 0.1754 0.4297 0.658 ]]
dtype: float16
cuNumeric output:
[[0.0791 0.373 0.9336 0.4187 0.2343 ]
[0.5723 0.5723 0.417 0.626 0.2203 ]
[0.622 0.4778 0.974 0.773 0.02715]
[0.2211 0.1203 0.1753 0.4294 0.6577 ]]
dtype: float16

Example code or instructions

Adds 
np.random.uniform(size=(4, 5)).astype("e")
in https://github.com/nv-legate/cunumeric/blob/branch-24.01/tests/integration/test_binary_ufunc.py#L131

Stack traceback or browser console output

No response

[manopapad]

This is an issue with limited accuracy of operations on half-width floats. Can you please try doing it similarly to what the einsum test is doing? https://github.com/nv-legate/cunumeric/blob/branch-24.01/tests/integration/test_einsum.py#L234

[XiaLuNV]

It works. Thanks.

[manopapad]

Excellent! Please make a PR with the fixes when you can.

[XiaLuNV]

The PR about this was merged. Say #1100.