Values in (2^15, 2^23) are not rounded by rounding ops

[jdh8]

Currently, rounding ops ignore values outside range of i16, i.e. [-2¹⁵, 2¹⁵).

tt-metal/tt_metal/hw/ckernels/wormhole_b0/metal/llk_api/llk_sfpu/ckernel_sfpu_floor.h

Lines 31 to 33 in 3d33e8d

	v_if (v <= SHRT_MIN || v >= SHRT_MAX){
	result = v;
	}

Since there are 24 significand bits (23 explicit) in f32, the maximum finite non-integer in f32 is 2²³ − 0.5. As a result, non-integers with absolute value in (2¹⁵, 2²³) are returned as is and not rounded.

Either of the following would solve this issue:

• Work on i32 instead of i16
• A second iteration that works on magnitudes > 2¹⁵

[mouliraj-mcw]

Hi @jdh8
As mentioned in the C kernel, since support for float-to-int32 conversion is currently unavailable, we will wait until this support is added.

Hi @rdjogoTT, @rtawfik01
Is support for float-to-int32 conversion possible?

cc: @eyonland

[rdjogoTT]

@mouliraj-mcw I do not believe support for float-to-int32 is going to become available as a single call similar to float_to_int16, because the underlying instruction does not support int32 as output to the conversion. We do have a manually implemented conversion that I made for the Typecast OP found here: https://github.com/tenstorrent/tt-llk-wh-b0/blob/eadfdb466dcc738803969f9eda76aacbb33b5ad6/common/inc/sfpu/ckernel_sfpu_typecast.h#L62-L97
I wouldn't recommend just dropping in this kernel in place of float_to_int, because that would be very slow. It would be better to combine the two so that there isn't as much of a perf hit.