[BUG] Incorrect Results Using __mmask16 Mask in if_else Function with AVX-512

[Panhaolin2001]

Here is a example: https://godbolt.org/z/nd9Kc3cqq

[Panhaolin2001]

The output should not be:
4,6,8,10,12,14,16,18,20,23,29,31,37,41,43,47,53,59,61,67,71,73,79,83,89,97,101,103,107,109,113,117 ?

[jfalcou]

Two things happens :

• raw mask types are in fact integers so the value of the mask is turned into true inside the if_else
• If I fix this by filling a logical manually, you ends up with the 1 not in the correct place as internally EVE is using big endian.

If your mask is always half true/half false, you can use the relative conditional mask liek keep_last :

https://godbolt.org/z/YMvja9sYo

[DenisYaroshevskiy]

Do you think you want to implicitly cast __mask16 to logical? I think that'd make sense but I also know that it doesn't work.
I appreciate that we like the implicit conversions and in other places you can pass a register directly as a parameter.

What are you thinking, @jfalcou ?

[jfalcou]

Maybe we need to make the wrapper type over the mmask type public

[DenisYaroshevskiy]

I think we'd need to put up a pr and then see what can we do. I don't think I understand your comment like this.

@Panhaolin2001 - you probably know this, but here is how to write what you meant today: https://godbolt.org/z/off5xv6nG

UPD: I would also recommend looking into our formula generation constructor, since this seems like what you want (from your example): https://godbolt.org/z/PhjhPxohP

[Panhaolin2001]

I'm sorry I may not have expressed myself too well. I currently made a type Mask<T> wrapped around __mmask8 __mmask16 __mmask32 and __mmask64 on avx512, and I want to make it compatible with the type eve::logical<eve::wide<T>>. One question I have is that when I do the auto mask = a > b; operation, the mask type is of type eve::logical<eve::wide<T>> and the internal storage_type is of type __mmask family, so why is it that when constructing the mask by using the auto mask = a > b if_else function returns the correct value, but not when __mmask16{0b0000000001111111} is specified manually?

[jfalcou]

We already have such a type. We may just need to make it public api and amend logical functions to handle it.

As for why __mask dont work, it is because they are just plain intégral so the constructor used to convert them to logical think you are doing logical(some.number) which is eqv to using logical(true).

[Panhaolin2001]

Thank you for your response! Looking forward to you exposing it as a public API and improving the handling of logic functions.

[jfalcou]

@DenisYaroshevskiy In fact, maybe we can just have an API that let us build logical from a bit set, using maybe a tagged constructor, and optimize it on AVX512 to not do anything and, for the other arch, to do the bit to mask conversion ?

logical<wide<float>> l( as_mask{0b01101101} );

?

[DenisYaroshevskiy]

I don't like this bits constructor. I think the formula/initializer list constructor do this already.

This is very length specific

[DenisYaroshevskiy]

@Panhaolin2001

I don't know how much I understood.

If you want to write a convrersion from your type to logical, provide a conversion operator

struct my {
  __mmask16 m;

  using eve_logical_t = eve::logical<eve::wide<float, eve::fixed<16>>>;
  
  operator eve_logical_t () {
    return eve_logical_t{m};
  }
};

[jfalcou]

It has to be something like

struct my {
  __mmask16 m;

  using eve_logical_t = eve::logical<eve::wide<float, eve::fixed<16>>>;
  
  operator eve_logical_t () {
    eve_logical_t that;
    that.storage().value = m;
    return that;
  }
};

or else the mask is seen as an integer then turned to true/false.

[DenisYaroshevskiy]

You are right, I thought what I wrote works. This is horrible. It works on any other platform. We need to at least stop it from compiling

[jfalcou]

It is complicated cause for most compiler the __mask* stuff are typedefs to integral :/

[DenisYaroshevskiy]

are typedefs to integral

we should only convert from bool and not from integrals

[jfalcou]

Ok thus

eve/include/eve/arch/cpu/logical_wide.hpp

Line 126 in c29cd38

template<scalar_value S>

Should go and the bool one after it should.use std::same_as to prevent conversion.
Will make a PR

[Panhaolin2001]

Is there any plan to convert from clang builtin vector extension to logical type?

typedef float vec_type __attribute__((vector_size(32)));
vec_type a{1,2,3,4,5,6,7,8}, b{2,3,1,4,6,7,9,4};
auto mask = a < b; 
auto eve_mask = eve::logical<eve::wide<T, eve::fixed<8>>>{mask};

[DenisYaroshevskiy]

Is there any plan to convert from clang builtin vector extension to logical type?

No plans.

I looked into it.
Clang always uses ints for mask type, even on avx512.

Here is how I'd do it:

  typedef float vec_type __attribute__((vector_size(32)));
  using eve_vec_type = eve::wide<float, eve::fixed<8>>;
  using eve_int_type = eve::wide<std::uint32_t, eve::fixed<8>>;

  vec_type a{1, 2, 3, 4, 5, 6, 7, 8}, b{2, 3, 1, 4, 6, 7, 9, 4};

  auto mask = a < b;
  auto clang_mask = eve::bit_cast(mask, eve::as<eve_int_type>{});
  auto eve_mask = eve::bit_cast(clang_mask != 0, eve::as<eve::logical<eve_vec_type>>{});

https://godbolt.org/z/jx4839jEh

[Panhaolin2001]

Excuse me, I have an off-topic question. May I ask, on avx512 platform wide<float, fixed<128>> represents a vector type with 128 elements, but avx512 can only support 16 elements of float type, how does wide<float, fixed<128>> utilize avx512 registers? Splicing?

[DenisYaroshevskiy]

I'm not 100% what you mean by splicing but they are represented as an array of registers.

So 128 / 16 = 8 registers.

We call this "aggregation".

We usually don't test that many registers at once, it should work but do test your code just in case.

If you can tell me more about the problem you are trying to solve, maybe I can give you a better answer.