cuda : vsubss4 for older versions of ROCm/clang

[Engininja2]

The Clang that comes with ROCm 5.2 doesn't have the builtin __builtin_elementwise_sub_sat that's used to polyfill __vsubss4 for HIP. This adds fallback code to compile in that case.

[JohannesGaessler]

This PR looks good to me but please note that I think the implementation could still be made faster. Just tell me when you want to get it merged.

[JohannesGaessler]

If I understand this code correctly it clips the tmp value to be within the legal limits of int8_t. In that case I think something like the following would be faster:

Suggested change

	if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
	if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
	const int smaller_min = tmp < std::numeric_limits<int8_t>::min();
	const int bigger_max = tmp > std::numeric_limits<int8_t>::max();
	tmp = smaller_min * std::numeric_limits<int8_t>::min() + larger_max * std::numeric_limits<int8_t>::max()
	+ !(smaller_min | larger_max) * tmp;

The reason I think it would be faster is that conditional statements are very slow on GPUs. Please note that I did not test this implementation and that you may need to test int vs. int16_t and | vs || for optimal performance.

[Engininja2]

The idea behind writing it that way is for the compiler to recognize that it's been asked to do saturating subtraction and not emit any conditional instructions at all.
The rocm clangs available on compiler explorer emit @llvm.ssub.sat.v4i8: https://godbolt.org/z/azxe38hMs

[Engininja2]

I checked the assembly for that function compiled for gfx803 on rocm-3.5.1 which is probably the oldest anybody might want to use and it's not using conditional instructions either so I think this is good to merge.