Created by: wozna

PR types

New features

PR changes

OPs

Describe

This PR presents:

• fix pybind conversion from uint16 to bfloat16. Numpy doesn't support bfloat16 data type, so in python bfloat16 is represented by uint16. In UT all operations can be done on float32 and then using convert_float32_to_uint16 from OpTest converted to uint16.
• conv2d and dequantize bfloat16 support with unit tests. In both unit tests atol in check output is changed, because we compare native convolution that was calculated on float32 with conv2d bf16 kernel that was calculated on bfloat16.

When the output from the convolution is:

• bfloat16, in Python we get uint16 data type, so atol = 2
• float32, so atol = 1e-2