Skip to content

avx512bw,vbmi: Optimize shift_left and shift_right. #1147

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
serge-sans-paille merged 2 commits into from
Jul 22, 2025
Merged

avx512bw,vbmi: Optimize shift_left and shift_right. #1147

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6af69c2
avx512bw: Optimize shift_left and shift_right.
degasus Jul 21, 2025
1d957d8
avx512bw,vbmi: Use make_batch_constant.
degasus Jul 21, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
82 changes: 40 additions & 42 deletions include/xsimd/arch/xsimd_avx512bw.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -486,33 +486,32 @@ namespace xsimd
// slide_left
namespace detail
{
template <size_t... Is>
constexpr std::array<uint64_t, sizeof...(Is)> make_slide_perm_hi(::xsimd::detail::index_sequence<Is...>)
struct make_slide_perm_hi
{
return { (Is == 0 ? 8 : Is - 1)... };
}
static constexpr uint64_t get(size_t i, size_t)
{
return i == 0 ? 8 : i - 1;
}
};

template <size_t N, size_t... Is>
constexpr std::array<uint16_t, sizeof...(Is)> make_slide_left_pattern(::xsimd::detail::index_sequence<Is...>)
{
return { (Is >= N ? Is - N : 0)... };
}
template <size_t N, size_t... Is>
constexpr std::array<uint16_t, sizeof...(Is)> make_slide_left_mask(::xsimd::detail::index_sequence<Is...>)
template <size_t N>
struct make_slide_left_pattern
{
return { (Is >= N ? 0xFFFF : 0x0000)... };
}
static constexpr uint16_t get(size_t i, size_t)
{
return i >= N ? i - N : 0;
}
};
}

template <size_t N, class A, class T>
XSIMD_INLINE batch<T, A> slide_left(batch<T, A> const& x, requires_arch<avx512bw>) noexcept
{
constexpr unsigned BitCount = N * 8;
if (BitCount == 0)
if (N == 0)
{
return x;
}
if (BitCount >= 512)
if (N >= 64)
{
return batch<T, A>(T(0));
}
Expand All @@ -526,10 +525,10 @@ namespace xsimd
buffer[0] = buffer[0] << 8;
xx = _mm512_load_epi64(&buffer[0]);

alignas(A::alignment()) auto slide_perm = detail::make_slide_perm_hi(::xsimd::detail::make_index_sequence<512 / 64>());
auto slide_perm = xsimd::make_batch_constant<uint64_t, detail::make_slide_perm_hi, A>();
__m512i xl = _mm512_slli_epi64(x, 8);
__m512i xr = _mm512_srli_epi64(x, 56);
xr = _mm512_permutex2var_epi64(xr, _mm512_load_epi64(slide_perm.data()), _mm512_setzero_si512());
xr = _mm512_permutex2var_epi64(xr, slide_perm.as_batch(), _mm512_setzero_si512());
xx = _mm512_or_si512(xr, xl);
if (N == 1)
return xx;
Expand All @@ -538,50 +537,49 @@ namespace xsimd
{
xx = x;
}
alignas(A::alignment()) auto slide_pattern = detail::make_slide_left_pattern<N / 2>(::xsimd::detail::make_index_sequence<512 / 16>());
alignas(A::alignment()) auto slide_mask = detail::make_slide_left_mask<N / 2>(::xsimd::detail::make_index_sequence<512 / 16>());
return _mm512_and_si512(_mm512_permutexvar_epi16(_mm512_load_epi32(slide_pattern.data()), xx), _mm512_load_epi32(slide_mask.data()));
__mmask32 mask = 0xFFFFFFFFu << ((N / 2) & 31);
auto slide_pattern = xsimd::make_batch_constant<uint16_t, detail::make_slide_left_pattern<N / 2>, A>();
return _mm512_maskz_permutexvar_epi16(mask, slide_pattern.as_batch(), xx);
}

// slide_right
namespace detail
{
template <size_t... Is>
constexpr std::array<uint64_t, sizeof...(Is)> make_slide_perm_low(::xsimd::detail::index_sequence<Is...>)
struct make_slide_perm_low
{
return { (Is + 1)... };
}
static constexpr uint64_t get(size_t i, size_t)
{
return i + 1;
}
};

template <size_t N, size_t... Is>
constexpr std::array<uint16_t, sizeof...(Is)> make_slide_right_pattern(::xsimd::detail::index_sequence<Is...>)
{
return { (Is < (32 - N) ? Is + N : 0)... };
}
template <size_t N, size_t... Is>
constexpr std::array<uint16_t, sizeof...(Is)> make_slide_right_mask(::xsimd::detail::index_sequence<Is...>)
template <size_t N>
struct make_slide_right_pattern
{
return { (Is < 32 - N ? 0xFFFF : 0x0000)... };
}
static constexpr uint16_t get(size_t i, size_t n)
{
return i < (n - N) ? i + N : 0;
}
};
}
template <size_t N, class A, class T>
XSIMD_INLINE batch<T, A> slide_right(batch<T, A> const& x, requires_arch<avx512bw>) noexcept
{
constexpr unsigned BitCount = N * 8;
if (BitCount == 0)
if (N == 0)
{
return x;
}
if (BitCount >= 512)
if (N >= 64)
{
return batch<T, A>(T(0));
}
batch<T, A> xx;
if (N & 1)
{
alignas(A::alignment()) auto slide_perm = detail::make_slide_perm_low(::xsimd::detail::make_index_sequence<512 / 64>());
auto slide_perm = xsimd::make_batch_constant<uint64_t, detail::make_slide_perm_low, A>();
__m512i xr = _mm512_srli_epi64(x, 8);
__m512i xl = _mm512_slli_epi64(x, 56);
xl = _mm512_permutex2var_epi64(xl, _mm512_load_epi64(slide_perm.data()), _mm512_setzero_si512());
xl = _mm512_permutex2var_epi64(xl, slide_perm.as_batch(), _mm512_setzero_si512());
xx = _mm512_or_si512(xr, xl);
if (N == 1)
return xx;
Expand All @@ -590,9 +588,9 @@ namespace xsimd
{
xx = x;
}
alignas(A::alignment()) auto slide_pattern = detail::make_slide_right_pattern<N / 2>(::xsimd::detail::make_index_sequence<512 / 16>());
alignas(A::alignment()) auto slide_mask = detail::make_slide_right_mask<N / 2>(::xsimd::detail::make_index_sequence<512 / 16>());
return _mm512_and_si512(_mm512_permutexvar_epi16(_mm512_load_epi32(slide_pattern.data()), xx), _mm512_load_epi32(slide_mask.data()));
__mmask32 mask = 0xFFFFFFFFu >> ((N / 2) & 31);
auto slide_pattern = xsimd::make_batch_constant<uint16_t, detail::make_slide_right_pattern<N / 2>, A>();
return _mm512_maskz_permutexvar_epi16(mask, slide_pattern.as_batch(), xx);
}

// ssub
Expand Down
30 changes: 18 additions & 12 deletions include/xsimd/arch/xsimd_avx512vbmi.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,17 +26,23 @@ namespace xsimd

namespace detail
{
template <size_t N, size_t... Is>
constexpr std::array<uint8_t, sizeof...(Is)> make_slide_left_bytes_pattern(::xsimd::detail::index_sequence<Is...>)
template <size_t N>
struct make_slide_left_bytes_pattern
{
return { (Is >= N ? Is - N : 0)... };
}
static constexpr uint8_t get(size_t i, size_t)
{
return i >= N ? i - N : 0;
}
};

template <size_t N, size_t... Is>
constexpr std::array<uint8_t, sizeof...(Is)> make_slide_right_bytes_pattern(::xsimd::detail::index_sequence<Is...>)
template <size_t N>
struct make_slide_right_bytes_pattern
{
return { (Is < (64 - N) ? Is + N : 0)... };
}
static constexpr uint8_t get(size_t i, size_t n)
{
return i < (n - N) ? i + N : 0;
}
};
}

// slide_left
Expand All @@ -53,8 +59,8 @@ namespace xsimd
}

__mmask64 mask = 0xFFFFFFFFFFFFFFFFull << (N & 63);
alignas(A::alignment()) auto slide_pattern = detail::make_slide_left_bytes_pattern<N>(::xsimd::detail::make_index_sequence<512 / 8>());
return _mm512_maskz_permutexvar_epi8(mask, _mm512_load_epi32(slide_pattern.data()), x);
auto slide_pattern = xsimd::make_batch_constant<uint8_t, detail::make_slide_left_bytes_pattern<N>, A>();
return _mm512_maskz_permutexvar_epi8(mask, slide_pattern.as_batch(), x);
}

// slide_right
Expand All @@ -70,8 +76,8 @@ namespace xsimd
return batch<T, A>(T(0));
}
__mmask64 mask = 0xFFFFFFFFFFFFFFFFull >> (N & 63);
alignas(A::alignment()) auto slide_pattern = detail::make_slide_right_bytes_pattern<N>(::xsimd::detail::make_index_sequence<512 / 8>());
return _mm512_maskz_permutexvar_epi8(mask, _mm512_load_epi32(slide_pattern.data()), x);
auto slide_pattern = xsimd::make_batch_constant<uint8_t, detail::make_slide_right_bytes_pattern<N>, A>();
return _mm512_maskz_permutexvar_epi8(mask, slide_pattern.as_batch(), x);
}

// swizzle (dynamic version)
Expand Down