Allow fp16 denorm inputs to be flushed to zero.
Added a new version of Compare16BitFloat for an fp16 input and a float
result. Previously, the code was using the version for a float input
with an fp result and switching the original and returned values.
But this function is not symmetric: e.g. a denorm input and zero result
should be allowed, but a zero input and denorm result should not be
allowed.
Also removed the rounding modes parameter, since conversion from fp16
to float is exact.
Affects:
dEQP-VK.spirv_assembly.instruction.compute.16bit_storage.uniform_16struct_to_32struct.*
dEQP-VK.spirv_assembly.instruction.compute.16bit_storage.uniform_32struct_to_16struct.*
dEQP-VK.spirv_assembly.instruction.graphics.16bit_storage.uniform_16struct_to_32struct.*
dEQP-VK.spirv_assembly.instruction.graphics.16bit_storage.uniform_32struct_to_16struct.*
Components: Vulkan
VK-GL-CTS issue: 1219
Change-Id: I96d01bb428b550de1e86a601b929d08b3131870d