[FEATURE] Do we really need TUMU mode in RISCV masked load ?

[SadiinsoSnowfall]

The current implementation of the masked loads under riscv uses _tumu intrinsics, Tail Undisturbed + Mask Undisturbed, ie we set to zero the inactive lanes w/rt the mask and the vector len.

I'm not sure we really need this as most EVE functions assume that there can be garbage in the inactive lanes anyway.

[DenisYaroshevskiy]

All functions in eve have to assume garbage there.
Garbage might happen even if the initial values were 0s - think cos.

I couldn't find why we did this.

What is the alternative? My rvv knowledge is still almost not existent.

tagging @ita-sc if he maybe has time

[ita-sc]

Hi!

There are several places to think about:
For this code part:

eve/include/eve/detail/function/simd/riscv/load.hpp

Line 76 in ce425a8

return perform_alternative_load(rvv_true<T, N>(), tgt, ptr);

This is just code minimization, actually, here we can use unconditional load intrinsics (some kind of __riscv_vle16_v_i16m2 for example). I've not done this during original review, as this would require a huge if else block, as we do not have overload variants of intrinsics for these code (we do not have an information about required rvv mode here without mask register).

For this part of code:

eve/include/eve/detail/function/simd/riscv/load.hpp

Line 78 in ce425a8

else return perform_alternative_load(cond.mask(tgt), tgt, ptr);

I'm not sure. If it is ok that we could have undefined values in places that were not masked - we can remove filling with zeros, I guess.

For _tu part, this could be removed, I believe. This was mainly done because I allowed to have bit_cast between types with same sizeof, but not the same N::value*sizeof(scalar_type) (initially can happen if we have really small wide, so lmul < 1, now prohibited).

I still have plans to revisit this codebase, however I can't provide a definite timeframe for this activity, given my current schedule. I guess if someone already has patches to address these issues - they are very welcome.

[DenisYaroshevskiy]

@ita-sc thanks for taking the time. No worries, this is open source, ppl are coming and going depending on wether they have time and energy.

We appreciate your impact