Clarify the denorm support for FP16 (may apply to FP32 too)

[wanghqc]

Moved this topic from Gitlab to Github:

Original post on this topic at Gitlab:

Create this issue based on the discussion on denorm of vstore_half in our weekly meeting. The current OpenCL spec allows FP16/FP32 denorm support to be specified via the clGetDeviceInfo call using CL_DEVICE_HALF_FP_CONFIG/CL_DEVICE_SINGLE_FP_CONFIG. A bit field for CL_FP_DENORM is used to tell whether denorms are supported for FP32/FP16. We need to specify what "support" really means: denorm can be from arithmetic or conversion. One discrepancy is the function like vstore_half: its conversion from float to half requires denorm to be preserved, regardless of the device's denorm handling capability.
We can add words in spec to say conversion is supported regardless the device's denorm capability.
We can also differentiate conversion vs arithmetic denorm capability, which will require new tests and sounds too much

Copy the reply from Ben:

We should be a little careful here, since I think there are multiple types of "conversions":

1. There are the vstore_half and vload_half functions, which involve float <-> half conversions, and are currently specified to NOT allow flushing denorms to zero.

• Reference: https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#the-half-data-type

2. There are the explicit conversion functions, e.g. convert_float. I think these may flush denorms to zero? Maybe this is an area we should improve.

• Reference: https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#explicit-conversions

• Related? https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#edge-case-behavior-in-flush-to-zero-mode

3. There are implicit conversions and explicit casts. Similar to (2), but without an explicit function call.

• https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#implicit-conversions
• https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#explicit-casts

4. There are conversion rules for images. These are allowed to flush denorms to zero, at least for float -> half and float <-> float, though I'm not sure about half -> float.

• https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#conversion-rules-for-half-precision-floating-point-channel-data-type
• https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_C.html#conversion-rules-for-floating-point-channel-data-type