[Bug c++/108902] New: Conversions std::float16_t<->float with FP16C are not vectorized

[Bug c++/108902] New: Conversions std::float16_t<->float with FP16C are not vectorized

[g.peterhoff@t-online.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108902

            Bug ID: 108902
           Summary: Conversions std::float16_t<->float with FP16C are not
                    vectorized
           Product: gcc
           Version: 13.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: g.peterhoff@t-online.de
  Target Milestone: ---

Please see
https://godbolt.org/z/dGn4qhPef

thx
Gero

[Bug c++/108902] Conversions std::float16_t<->float with FP16C are not vectorized

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108902

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
#include <array>
#include <cstdint>
#include <stdfloat>
#include <algorithm>
#include <execution>
#include <x86intrin.h>


using array_t = std::array<std::float16_t, 128>;


void inc_loop(array_t& arr)    noexcept
{
    for (auto& val : arr)
        ++val;
}


void inc_transform(array_t& arr)    noexcept
{
    std::transform(std::execution::unseq, arr.begin(), arr.end(), arr.begin(),
[](const auto val) constexpr noexcept{return val+1;});
}


void inc_intrinsic(array_t& arr)    noexcept
{
    auto load_cvt = [](const std::float16_t*const ptr) noexcept
    {
        return _mm256_cvtph_ps(*((const __m128i*const)ptr));
    };

    auto save_cvt = [](std::float16_t* ptr, const __m256 arg)    noexcept
    {
        *((__m128i*)ptr) = _mm256_cvtps_ph(arg, _MM_FROUND_CUR_DIRECTION);
    };

    for (std::size_t i=0; i<arr.size(); i+=8)
    {
        __m256
            tmp = load_cvt(&arr[i]);

        ++tmp;
       save_cvt(&arr[i], tmp);
    }
}

[Bug target/108902] Conversions std::float16_t<->float with FP16C are not vectorized

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108902

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
-std=c++23 -march=x86-64-v3 -O3 -mno-vzeroupper

[Bug target/108902] Conversions std::float16_t<->float with FP16C are not vectorized

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108902

Hongtao.liu <crazylht at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |crazylht at gmail dot com

--- Comment #3 from Hongtao.liu <crazylht at gmail dot com> ---
Yes,  in sse.md the corresponding expanders are only defined under
TARGET_AVX512FP16.

7275(define_expand "extend<ssePHmodelower><mode>2"
 7276  [(set (match_operand:VF48H_AVX512VL 0 "register_operand")
 7277        (float_extend:VF48H_AVX512VL
 7278          (match_operand:<ssePHmode> 1 "nonimmediate_operand")))]
 7279  "TARGET_AVX512FP16")


 7369(define_expand "trunc<mode><ssePHmodelower>2"
 7370  [(set (match_operand:<ssePHmode> 0 "register_operand")
 7371        (float_truncate:<ssePHmode>
 7372          (match_operand:VF48H_AVX512VL 1 "nonimmediate_operand")))]
 7373  "TARGET_AVX512FP16")

[Bug target/108902] Conversions std::float16_t<->float with FP16C are not vectorized

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108902

--- Comment #4 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Hongtao.liu from comment #3)
> Yes,  in sse.md the corresponding expanders are only defined under
> TARGET_AVX512FP16.

Even w/ -mavx512fp16, it's still not vectorized since it relied on
vec_unpacks_hi/lo_v*hf for loop vectorizer. BB vectorizer seems ok.

[Bug target/108902] Conversions std::float16_t<->float with FP16C are not vectorized

[g.peterhoff@t-online.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108902

--- Comment #5 from g.peterhoff@t-online.de ---
add test case (https://godbolt.org/z/q65cWKhWx)

void inc_builtin(array_t& arr)    noexcept
{
    auto load_cvt = [](const std::float16_t*const ptr) noexcept
    {
        return __builtin_convertvector(*((const __m128h*const)ptr), __m256);
    };

    auto save_cvt = [](std::float16_t* ptr, const __m256 arg)    noexcept
    {
        *((__m128h*)ptr) = __builtin_convertvector(arg, __m128h);
    };

    for (std::size_t i=0; i<arr.size(); i+=8)
    {
        __m256
            tmp = load_cvt(&arr[i]);

        ++tmp;
        save_cvt(&arr[i], tmp);
    }
}