Skip to content

perf(avx512f): vectorize 16x16 float / 8x8 double transpose #1371

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 1 commit into from
Jul 7, 2026
+77 −0
Merged

perf(avx512f): vectorize 16x16 float / 8x8 double transpose #1371

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
77 changes: 77 additions & 0 deletions include/xsimd/arch/xsimd_avx512f.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2787,6 +2787,83 @@ namespace xsimd
return transpose(reinterpret_cast<batch<uint8_t, A>*>(matrix_begin), reinterpret_cast<batch<uint8_t, A>*>(matrix_end), A {});
}

template <class A>
XSIMD_INLINE void transpose(batch<float, A>* matrix_begin, batch<float, A>* matrix_end, requires_arch<avx512f>) noexcept
{
assert((matrix_end - matrix_begin == batch<float, A>::size) && "correctly sized matrix");
(void)matrix_end;
__m512 t[16], u[16];
for (int k = 0; k < 8; ++k)
{
t[2 * k + 0] = _mm512_unpacklo_ps(matrix_begin[2 * k], matrix_begin[2 * k + 1]);
t[2 * k + 1] = _mm512_unpackhi_ps(matrix_begin[2 * k], matrix_begin[2 * k + 1]);
}
for (int k = 0; k < 4; ++k)
{
u[4 * k + 0] = _mm512_shuffle_ps(t[4 * k + 0], t[4 * k + 2], 0x44);
u[4 * k + 1] = _mm512_shuffle_ps(t[4 * k + 0], t[4 * k + 2], 0xEE);
u[4 * k + 2] = _mm512_shuffle_ps(t[4 * k + 1], t[4 * k + 3], 0x44);
u[4 * k + 3] = _mm512_shuffle_ps(t[4 * k + 1], t[4 * k + 3], 0xEE);
}
for (int k = 0; k < 4; ++k)
{
t[k + 0] = _mm512_shuffle_f32x4(u[k + 0], u[k + 4], 0x88);
t[k + 4] = _mm512_shuffle_f32x4(u[k + 0], u[k + 4], 0xdd);
t[k + 8] = _mm512_shuffle_f32x4(u[k + 8], u[k + 12], 0x88);
t[k + 12] = _mm512_shuffle_f32x4(u[k + 8], u[k + 12], 0xdd);
}
for (int k = 0; k < 8; ++k)
{
matrix_begin[k + 0] = _mm512_shuffle_f32x4(t[k], t[k + 8], 0x88);
matrix_begin[k + 8] = _mm512_shuffle_f32x4(t[k], t[k + 8], 0xdd);
}
}
template <class A>
XSIMD_INLINE void transpose(batch<uint32_t, A>* matrix_begin, batch<uint32_t, A>* matrix_end, requires_arch<avx512f>) noexcept
{
return transpose(reinterpret_cast<batch<float, A>*>(matrix_begin), reinterpret_cast<batch<float, A>*>(matrix_end), A {});
}
template <class A>
XSIMD_INLINE void transpose(batch<int32_t, A>* matrix_begin, batch<int32_t, A>* matrix_end, requires_arch<avx512f>) noexcept
{
return transpose(reinterpret_cast<batch<float, A>*>(matrix_begin), reinterpret_cast<batch<float, A>*>(matrix_end), A {});
}

template <class A>
XSIMD_INLINE void transpose(batch<double, A>* matrix_begin, batch<double, A>* matrix_end, requires_arch<avx512f>) noexcept
{
assert((matrix_end - matrix_begin == batch<double, A>::size) && "correctly sized matrix");
(void)matrix_end;
__m512d t[8], u[8];
for (int k = 0; k < 4; ++k)
{
t[2 * k + 0] = _mm512_unpacklo_pd(matrix_begin[2 * k], matrix_begin[2 * k + 1]);
t[2 * k + 1] = _mm512_unpackhi_pd(matrix_begin[2 * k], matrix_begin[2 * k + 1]);
}
for (int k = 0; k < 2; ++k)
{
u[4 * k + 0] = _mm512_shuffle_f64x2(t[4 * k + 0], t[4 * k + 2], 0x88);
u[4 * k + 1] = _mm512_shuffle_f64x2(t[4 * k + 1], t[4 * k + 3], 0x88);
u[4 * k + 2] = _mm512_shuffle_f64x2(t[4 * k + 0], t[4 * k + 2], 0xdd);
u[4 * k + 3] = _mm512_shuffle_f64x2(t[4 * k + 1], t[4 * k + 3], 0xdd);
}
for (int k = 0; k < 4; ++k)
{
matrix_begin[k + 0] = _mm512_shuffle_f64x2(u[k], u[k + 4], 0x88);
matrix_begin[k + 4] = _mm512_shuffle_f64x2(u[k], u[k + 4], 0xdd);
}
}
template <class A>
XSIMD_INLINE void transpose(batch<uint64_t, A>* matrix_begin, batch<uint64_t, A>* matrix_end, requires_arch<avx512f>) noexcept
{
return transpose(reinterpret_cast<batch<double, A>*>(matrix_begin), reinterpret_cast<batch<double, A>*>(matrix_end), A {});
}
template <class A>
XSIMD_INLINE void transpose(batch<int64_t, A>* matrix_begin, batch<int64_t, A>* matrix_end, requires_arch<avx512f>) noexcept
{
return transpose(reinterpret_cast<batch<double, A>*>(matrix_begin), reinterpret_cast<batch<double, A>*>(matrix_end), A {});
}

// trunc
template <class A>
XSIMD_INLINE batch<float, A>
Expand Down
Loading