Viewing file:  simd_x86_conversions.h (80.53 KB) -rw-r--r--
Select action/file-type:
 (+) |  (+) |  (+) | Code (+) | Session (+) |  (+) | SDB (+) |  (+) |  (+) |  (+) |  (+) |  (+) |
// x86 specific conversion optimizations -*- C++ -*-
// Copyright (C) 2020-2021 Free Software Foundation, Inc.
//
// This file is part of the GNU ISO C++ Library. This library is free
// software; you can redistribute it and/or modify it under the
// terms of the GNU General Public License as published by the
// Free Software Foundation; either version 3, or (at your option)
// any later version.
// This library is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
// Under Section 7 of GPL version 3, you are granted additional
// permissions described in the GCC Runtime Library Exception, version
// 3.1, as published by the Free Software Foundation.
// You should have received a copy of the GNU General Public License and
// a copy of the GCC Runtime Library Exception along with this program;
// see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
// <http://www.gnu.org/licenses/>.
#ifndef _GLIBCXX_EXPERIMENTAL_SIMD_X86_CONVERSIONS_H
#define _GLIBCXX_EXPERIMENTAL_SIMD_X86_CONVERSIONS_H
#if __cplusplus >= 201703L
// work around PR85827
// 1-arg __convert_x86 {{{1
template <typename _To, typename _V, typename _Traits>
_GLIBCXX_SIMD_INTRINSIC _To
__convert_x86(_V __v)
{
static_assert(__is_vector_type_v<_V>);
using _Tp = typename _Traits::value_type;
constexpr size_t _Np = _Traits::_S_full_size;
[[maybe_unused]] const auto __intrin = __to_intrin(__v);
using _Up = typename _VectorTraits<_To>::value_type;
constexpr size_t _M = _VectorTraits<_To>::_S_full_size;
// [xyz]_to_[xyz] {{{2
[[maybe_unused]] constexpr bool __x_to_x
= sizeof(__v) <= 16 && sizeof(_To) <= 16;
[[maybe_unused]] constexpr bool __x_to_y
= sizeof(__v) <= 16 && sizeof(_To) == 32;
[[maybe_unused]] constexpr bool __x_to_z
= sizeof(__v) <= 16 && sizeof(_To) == 64;
[[maybe_unused]] constexpr bool __y_to_x
= sizeof(__v) == 32 && sizeof(_To) <= 16;
[[maybe_unused]] constexpr bool __y_to_y
= sizeof(__v) == 32 && sizeof(_To) == 32;
[[maybe_unused]] constexpr bool __y_to_z
= sizeof(__v) == 32 && sizeof(_To) == 64;
[[maybe_unused]] constexpr bool __z_to_x
= sizeof(__v) == 64 && sizeof(_To) <= 16;
[[maybe_unused]] constexpr bool __z_to_y
= sizeof(__v) == 64 && sizeof(_To) == 32;
[[maybe_unused]] constexpr bool __z_to_z
= sizeof(__v) == 64 && sizeof(_To) == 64;
// iX_to_iX {{{2
[[maybe_unused]] constexpr bool __i_to_i
= is_integral_v<_Up> && is_integral_v<_Tp>;
[[maybe_unused]] constexpr bool __i8_to_i16
= __i_to_i && sizeof(_Tp) == 1 && sizeof(_Up) == 2;
[[maybe_unused]] constexpr bool __i8_to_i32
= __i_to_i && sizeof(_Tp) == 1 && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __i8_to_i64
= __i_to_i && sizeof(_Tp) == 1 && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __i16_to_i8
= __i_to_i && sizeof(_Tp) == 2 && sizeof(_Up) == 1;
[[maybe_unused]] constexpr bool __i16_to_i32
= __i_to_i && sizeof(_Tp) == 2 && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __i16_to_i64
= __i_to_i && sizeof(_Tp) == 2 && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __i32_to_i8
= __i_to_i && sizeof(_Tp) == 4 && sizeof(_Up) == 1;
[[maybe_unused]] constexpr bool __i32_to_i16
= __i_to_i && sizeof(_Tp) == 4 && sizeof(_Up) == 2;
[[maybe_unused]] constexpr bool __i32_to_i64
= __i_to_i && sizeof(_Tp) == 4 && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __i64_to_i8
= __i_to_i && sizeof(_Tp) == 8 && sizeof(_Up) == 1;
[[maybe_unused]] constexpr bool __i64_to_i16
= __i_to_i && sizeof(_Tp) == 8 && sizeof(_Up) == 2;
[[maybe_unused]] constexpr bool __i64_to_i32
= __i_to_i && sizeof(_Tp) == 8 && sizeof(_Up) == 4;
// [fsu]X_to_[fsu]X {{{2
// ibw = integral && byte or word, i.e. char and short with any signedness
[[maybe_unused]] constexpr bool __s64_to_f32
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __s32_to_f32
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __s16_to_f32
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 2
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __s8_to_f32
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 1
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __u64_to_f32
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __u32_to_f32
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __u16_to_f32
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 2
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __u8_to_f32
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 1
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __s64_to_f64
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __s32_to_f64
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __u64_to_f64
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __u32_to_f64
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __f32_to_s64
= is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) == 8
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f32_to_s32
= is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) == 4
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f32_to_u64
= is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) == 8
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f32_to_u32
= is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) == 4
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f64_to_s64
= is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) == 8
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f64_to_s32
= is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) == 4
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f64_to_u64
= is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) == 8
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f64_to_u32
= is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) == 4
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __ibw_to_f32
= is_integral_v<_Tp> && sizeof(_Tp) <= 2
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __ibw_to_f64
= is_integral_v<_Tp> && sizeof(_Tp) <= 2
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __f32_to_ibw
= is_integral_v<_Up> && sizeof(_Up) <= 2
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f64_to_ibw
= is_integral_v<_Up> && sizeof(_Up) <= 2
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f32_to_f64
= is_floating_point_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __f64_to_f32
= is_floating_point_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
if constexpr (__i_to_i && __y_to_x && !__have_avx2) //{{{2
return __convert_x86<_To>(__lo128(__v), __hi128(__v));
else if constexpr (__i_to_i && __x_to_y && !__have_avx2) //{{{2
return __concat(__convert_x86<__vector_type_t<_Up, _M / 2>>(__v),
__convert_x86<__vector_type_t<_Up, _M / 2>>(
__extract_part<1, _Np / _M * 2>(__v)));
else if constexpr (__i_to_i) //{{{2
{
static_assert(__x_to_x || __have_avx2,
"integral conversions with ymm registers require AVX2");
static_assert(__have_avx512bw
|| ((sizeof(_Tp) >= 4 || sizeof(__v) < 64)
&& (sizeof(_Up) >= 4 || sizeof(_To) < 64)),
"8/16-bit integers in zmm registers require AVX512BW");
static_assert((sizeof(__v) < 64 && sizeof(_To) < 64) || __have_avx512f,
"integral conversions with ymm registers require AVX2");
}
if constexpr (is_floating_point_v<_Tp> == is_floating_point_v<_Up> && //{{{2
sizeof(_Tp) == sizeof(_Up))
{
// conversion uses simple bit reinterpretation (or no conversion at all)
if constexpr (_Np >= _M)
return __intrin_bitcast<_To>(__v);
else
return __zero_extend(__vector_bitcast<_Up>(__v));
}
else if constexpr (_Np < _M && sizeof(_To) > 16) //{{{2
// zero extend (eg. xmm -> ymm)
return __zero_extend(
__convert_x86<__vector_type_t<
_Up, (16 / sizeof(_Up) > _Np) ? 16 / sizeof(_Up) : _Np>>(__v));
else if constexpr (_Np > _M && sizeof(__v) > 16) //{{{2
// partial input (eg. ymm -> xmm)
return __convert_x86<_To>(__extract_part<0, _Np / _M>(__v));
else if constexpr (__i64_to_i32) //{{{2
{
if constexpr (__x_to_x && __have_avx512vl)
return __intrin_bitcast<_To>(_mm_cvtepi64_epi32(__intrin));
else if constexpr (__x_to_x)
return __auto_bitcast(
_mm_shuffle_ps(__vector_bitcast<float>(__v), __m128(), 8));
else if constexpr (__y_to_x && __have_avx512vl)
return __intrin_bitcast<_To>(_mm256_cvtepi64_epi32(__intrin));
else if constexpr (__y_to_x && __have_avx512f)
return __intrin_bitcast<_To>(
__lo128(_mm512_cvtepi64_epi32(__auto_bitcast(__v))));
else if constexpr (__y_to_x)
return __intrin_bitcast<_To>(
__lo128(_mm256_permute4x64_epi64(_mm256_shuffle_epi32(__intrin, 8),
0 + 4 * 2)));
else if constexpr (__z_to_y)
return __intrin_bitcast<_To>(_mm512_cvtepi64_epi32(__intrin));
}
else if constexpr (__i64_to_i16) //{{{2
{
if constexpr (__x_to_x && __have_avx512vl)
return __intrin_bitcast<_To>(_mm_cvtepi64_epi16(__intrin));
else if constexpr (__x_to_x && __have_avx512f)
return __intrin_bitcast<_To>(
__lo128(_mm512_cvtepi64_epi16(__auto_bitcast(__v))));
else if constexpr (__x_to_x && __have_ssse3)
{
return __intrin_bitcast<_To>(
_mm_shuffle_epi8(__intrin,
_mm_setr_epi8(0, 1, 8, 9, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80)));
// fallback without SSSE3
}
else if constexpr (__y_to_x && __have_avx512vl)
return __intrin_bitcast<_To>(_mm256_cvtepi64_epi16(__intrin));
else if constexpr (__y_to_x && __have_avx512f)
return __intrin_bitcast<_To>(
__lo128(_mm512_cvtepi64_epi16(__auto_bitcast(__v))));
else if constexpr (__y_to_x)
{
const auto __a = _mm256_shuffle_epi8(
__intrin,
_mm256_setr_epi8(0, 1, 8, 9, -0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, 0, 1, 8, 9, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
-0x80));
return __intrin_bitcast<_To>(__lo128(__a) | __hi128(__a));
}
else if constexpr (__z_to_x)
return __intrin_bitcast<_To>(_mm512_cvtepi64_epi16(__intrin));
}
else if constexpr (__i64_to_i8) //{{{2
{
if constexpr (__x_to_x && __have_avx512vl)
return __intrin_bitcast<_To>(_mm_cvtepi64_epi8(__intrin));
else if constexpr (__x_to_x && __have_avx512f)
return __intrin_bitcast<_To>(
__lo128(_mm512_cvtepi64_epi8(__zero_extend(__intrin))));
else if constexpr (__y_to_x && __have_avx512vl)
return __intrin_bitcast<_To>(_mm256_cvtepi64_epi8(__intrin));
else if constexpr (__y_to_x && __have_avx512f)
return __intrin_bitcast<_To>(
_mm512_cvtepi64_epi8(__zero_extend(__intrin)));
else if constexpr (__z_to_x)
return __intrin_bitcast<_To>(_mm512_cvtepi64_epi8(__intrin));
}
else if constexpr (__i32_to_i64) //{{{2
{
if constexpr (__have_sse4_1 && __x_to_x)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm_cvtepi32_epi64(__intrin)
: _mm_cvtepu32_epi64(__intrin));
else if constexpr (__x_to_x)
{
return __intrin_bitcast<_To>(
_mm_unpacklo_epi32(__intrin, is_signed_v<_Tp>
? _mm_srai_epi32(__intrin, 31)
: __m128i()));
}
else if constexpr (__x_to_y)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm256_cvtepi32_epi64(__intrin)
: _mm256_cvtepu32_epi64(__intrin));
else if constexpr (__y_to_z)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm512_cvtepi32_epi64(__intrin)
: _mm512_cvtepu32_epi64(__intrin));
}
else if constexpr (__i32_to_i16) //{{{2
{
if constexpr (__x_to_x && __have_avx512vl)
return __intrin_bitcast<_To>(_mm_cvtepi32_epi16(__intrin));
else if constexpr (__x_to_x && __have_avx512f)
return __intrin_bitcast<_To>(
__lo128(_mm512_cvtepi32_epi16(__auto_bitcast(__v))));
else if constexpr (__x_to_x && __have_ssse3)
return __intrin_bitcast<_To>(_mm_shuffle_epi8(
__intrin, _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, -0x80)));
else if constexpr (__x_to_x)
{
auto __a = _mm_unpacklo_epi16(__intrin, __m128i()); // 0o.o 1o.o
auto __b = _mm_unpackhi_epi16(__intrin, __m128i()); // 2o.o 3o.o
auto __c = _mm_unpacklo_epi16(__a, __b); // 02oo ..oo
auto __d = _mm_unpackhi_epi16(__a, __b); // 13oo ..oo
return __intrin_bitcast<_To>(
_mm_unpacklo_epi16(__c, __d)); // 0123 oooo
}
else if constexpr (__y_to_x && __have_avx512vl)
return __intrin_bitcast<_To>(_mm256_cvtepi32_epi16(__intrin));
else if constexpr (__y_to_x && __have_avx512f)
return __intrin_bitcast<_To>(
__lo128(_mm512_cvtepi32_epi16(__auto_bitcast(__v))));
else if constexpr (__y_to_x)
{
auto __a = _mm256_shuffle_epi8(
__intrin,
_mm256_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, 0, 1, 4, 5, 8,
9, 12, 13, -0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80));
return __intrin_bitcast<_To>(__lo128(
_mm256_permute4x64_epi64(__a,
0xf8))); // __a[0] __a[2] | __a[3] __a[3]
}
else if constexpr (__z_to_y)
return __intrin_bitcast<_To>(_mm512_cvtepi32_epi16(__intrin));
}
else if constexpr (__i32_to_i8) //{{{2
{
if constexpr (__x_to_x && __have_avx512vl)
return __intrin_bitcast<_To>(_mm_cvtepi32_epi8(__intrin));
else if constexpr (__x_to_x && __have_avx512f)
return __intrin_bitcast<_To>(
__lo128(_mm512_cvtepi32_epi8(__zero_extend(__intrin))));
else if constexpr (__x_to_x && __have_ssse3)
{
return __intrin_bitcast<_To>(
_mm_shuffle_epi8(__intrin,
_mm_setr_epi8(0, 4, 8, 12, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80)));
}
else if constexpr (__x_to_x)
{
const auto __a
= _mm_unpacklo_epi8(__intrin, __intrin); // 0... .... 1... ....
const auto __b
= _mm_unpackhi_epi8(__intrin, __intrin); // 2... .... 3... ....
const auto __c = _mm_unpacklo_epi8(__a, __b); // 02.. .... .... ....
const auto __d = _mm_unpackhi_epi8(__a, __b); // 13.. .... .... ....
const auto __e = _mm_unpacklo_epi8(__c, __d); // 0123 .... .... ....
return __intrin_bitcast<_To>(__e & _mm_cvtsi32_si128(-1));
}
else if constexpr (__y_to_x && __have_avx512vl)
return __intrin_bitcast<_To>(_mm256_cvtepi32_epi8(__intrin));
else if constexpr (__y_to_x && __have_avx512f)
return __intrin_bitcast<_To>(
_mm512_cvtepi32_epi8(__zero_extend(__intrin)));
else if constexpr (__z_to_x)
return __intrin_bitcast<_To>(_mm512_cvtepi32_epi8(__intrin));
}
else if constexpr (__i16_to_i64) //{{{2
{
if constexpr (__x_to_x && __have_sse4_1)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm_cvtepi16_epi64(__intrin)
: _mm_cvtepu16_epi64(__intrin));
else if constexpr (__x_to_x && is_signed_v<_Tp>)
{
auto __x = _mm_srai_epi16(__intrin, 15);
auto __y = _mm_unpacklo_epi16(__intrin, __x);
__x = _mm_unpacklo_epi16(__x, __x);
return __intrin_bitcast<_To>(_mm_unpacklo_epi32(__y, __x));
}
else if constexpr (__x_to_x)
return __intrin_bitcast<_To>(
_mm_unpacklo_epi32(_mm_unpacklo_epi16(__intrin, __m128i()),
__m128i()));
else if constexpr (__x_to_y)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm256_cvtepi16_epi64(__intrin)
: _mm256_cvtepu16_epi64(__intrin));
else if constexpr (__x_to_z)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm512_cvtepi16_epi64(__intrin)
: _mm512_cvtepu16_epi64(__intrin));
}
else if constexpr (__i16_to_i32) //{{{2
{
if constexpr (__x_to_x && __have_sse4_1)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm_cvtepi16_epi32(__intrin)
: _mm_cvtepu16_epi32(__intrin));
else if constexpr (__x_to_x && is_signed_v<_Tp>)
return __intrin_bitcast<_To>(
_mm_srai_epi32(_mm_unpacklo_epi16(__intrin, __intrin), 16));
else if constexpr (__x_to_x && is_unsigned_v<_Tp>)
return __intrin_bitcast<_To>(_mm_unpacklo_epi16(__intrin, __m128i()));
else if constexpr (__x_to_y)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm256_cvtepi16_epi32(__intrin)
: _mm256_cvtepu16_epi32(__intrin));
else if constexpr (__y_to_z)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm512_cvtepi16_epi32(__intrin)
: _mm512_cvtepu16_epi32(__intrin));
}
else if constexpr (__i16_to_i8) //{{{2
{
if constexpr (__x_to_x && __have_avx512bw_vl)
return __intrin_bitcast<_To>(_mm_cvtepi16_epi8(__intrin));
else if constexpr (__x_to_x && __have_avx512bw)
return __intrin_bitcast<_To>(
__lo128(_mm512_cvtepi16_epi8(__zero_extend(__intrin))));
else if constexpr (__x_to_x && __have_ssse3)
return __intrin_bitcast<_To>(_mm_shuffle_epi8(
__intrin, _mm_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, -0x80)));
else if constexpr (__x_to_x)
{
auto __a
= _mm_unpacklo_epi8(__intrin, __intrin); // 00.. 11.. 22.. 33..
auto __b
= _mm_unpackhi_epi8(__intrin, __intrin); // 44.. 55.. 66.. 77..
auto __c = _mm_unpacklo_epi8(__a, __b); // 0404 .... 1515 ....
auto __d = _mm_unpackhi_epi8(__a, __b); // 2626 .... 3737 ....
auto __e = _mm_unpacklo_epi8(__c, __d); // 0246 0246 .... ....
auto __f = _mm_unpackhi_epi8(__c, __d); // 1357 1357 .... ....
return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__e, __f));
}
else if constexpr (__y_to_x && __have_avx512bw_vl)
return __intrin_bitcast<_To>(_mm256_cvtepi16_epi8(__intrin));
else if constexpr (__y_to_x && __have_avx512bw)
return __intrin_bitcast<_To>(
__lo256(_mm512_cvtepi16_epi8(__zero_extend(__intrin))));
else if constexpr (__y_to_x)
{
auto __a = _mm256_shuffle_epi8(
__intrin,
_mm256_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, -0x80, 0, 2,
4, 6, 8, 10, 12, 14));
return __intrin_bitcast<_To>(__lo128(__a) | __hi128(__a));
}
else if constexpr (__z_to_y && __have_avx512bw)
return __intrin_bitcast<_To>(_mm512_cvtepi16_epi8(__intrin));
else if constexpr (__z_to_y)
__assert_unreachable<_Tp>();
}
else if constexpr (__i8_to_i64) //{{{2
{
if constexpr (__x_to_x && __have_sse4_1)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm_cvtepi8_epi64(__intrin)
: _mm_cvtepu8_epi64(__intrin));
else if constexpr (__x_to_x && is_signed_v<_Tp>)
{
if constexpr (__have_ssse3)
{
auto __dup = _mm_unpacklo_epi8(__intrin, __intrin);
auto __epi16 = _mm_srai_epi16(__dup, 8);
_mm_shuffle_epi8(__epi16,
_mm_setr_epi8(0, 1, 1, 1, 1, 1, 1, 1, 2, 3, 3,
3, 3, 3, 3, 3));
}
else
{
auto __x = _mm_unpacklo_epi8(__intrin, __intrin);
__x = _mm_unpacklo_epi16(__x, __x);
return __intrin_bitcast<_To>(
_mm_unpacklo_epi32(_mm_srai_epi32(__x, 24),
_mm_srai_epi32(__x, 31)));
}
}
else if constexpr (__x_to_x)
{
return __intrin_bitcast<_To>(_mm_unpacklo_epi32(
_mm_unpacklo_epi16(_mm_unpacklo_epi8(__intrin, __m128i()),
__m128i()),
__m128i()));
}
else if constexpr (__x_to_y)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm256_cvtepi8_epi64(__intrin)
: _mm256_cvtepu8_epi64(__intrin));
else if constexpr (__x_to_z)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm512_cvtepi8_epi64(__intrin)
: _mm512_cvtepu8_epi64(__intrin));
}
else if constexpr (__i8_to_i32) //{{{2
{
if constexpr (__x_to_x && __have_sse4_1)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm_cvtepi8_epi32(__intrin)
: _mm_cvtepu8_epi32(__intrin));
else if constexpr (__x_to_x && is_signed_v<_Tp>)
{
const auto __x = _mm_unpacklo_epi8(__intrin, __intrin);
return __intrin_bitcast<_To>(
_mm_srai_epi32(_mm_unpacklo_epi16(__x, __x), 24));
}
else if constexpr (__x_to_x && is_unsigned_v<_Tp>)
return __intrin_bitcast<_To>(
_mm_unpacklo_epi16(_mm_unpacklo_epi8(__intrin, __m128i()),
__m128i()));
else if constexpr (__x_to_y)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm256_cvtepi8_epi32(__intrin)
: _mm256_cvtepu8_epi32(__intrin));
else if constexpr (__x_to_z)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm512_cvtepi8_epi32(__intrin)
: _mm512_cvtepu8_epi32(__intrin));
}
else if constexpr (__i8_to_i16) //{{{2
{
if constexpr (__x_to_x && __have_sse4_1)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm_cvtepi8_epi16(__intrin)
: _mm_cvtepu8_epi16(__intrin));
else if constexpr (__x_to_x && is_signed_v<_Tp>)
return __intrin_bitcast<_To>(
_mm_srai_epi16(_mm_unpacklo_epi8(__intrin, __intrin), 8));
else if constexpr (__x_to_x && is_unsigned_v<_Tp>)
return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__intrin, __m128i()));
else if constexpr (__x_to_y)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm256_cvtepi8_epi16(__intrin)
: _mm256_cvtepu8_epi16(__intrin));
else if constexpr (__y_to_z && __have_avx512bw)
return __intrin_bitcast<_To>(is_signed_v<_Tp>
? _mm512_cvtepi8_epi16(__intrin)
: _mm512_cvtepu8_epi16(__intrin));
else if constexpr (__y_to_z)
__assert_unreachable<_Tp>();
}
else if constexpr (__f32_to_s64) //{{{2
{
if constexpr (__have_avx512dq_vl && __x_to_x)
return __intrin_bitcast<_To>(_mm_cvttps_epi64(__intrin));
else if constexpr (__have_avx512dq_vl && __x_to_y)
return __intrin_bitcast<_To>(_mm256_cvttps_epi64(__intrin));
else if constexpr (__have_avx512dq && __y_to_z)
return __intrin_bitcast<_To>(_mm512_cvttps_epi64(__intrin));
// else use scalar fallback
}
else if constexpr (__f32_to_u64) //{{{2
{
if constexpr (__have_avx512dq_vl && __x_to_x)
return __intrin_bitcast<_To>(_mm_cvttps_epu64(__intrin));
else if constexpr (__have_avx512dq_vl && __x_to_y)
return __intrin_bitcast<_To>(_mm256_cvttps_epu64(__intrin));
else if constexpr (__have_avx512dq && __y_to_z)
return __intrin_bitcast<_To>(_mm512_cvttps_epu64(__intrin));
// else use scalar fallback
}
else if constexpr (__f32_to_s32) //{{{2
{
if constexpr (__x_to_x || __y_to_y || __z_to_z)
{
// go to fallback, it does the right thing
}
else
__assert_unreachable<_Tp>();
}
else if constexpr (__f32_to_u32) //{{{2
{
if constexpr (__have_avx512vl && __x_to_x)
return __auto_bitcast(_mm_cvttps_epu32(__intrin));
else if constexpr (__have_avx512f && __x_to_x)
return __auto_bitcast(
__lo128(_mm512_cvttps_epu32(__auto_bitcast(__v))));
else if constexpr (__have_avx512vl && __y_to_y)
return __vector_bitcast<_Up>(_mm256_cvttps_epu32(__intrin));
else if constexpr (__have_avx512f && __y_to_y)
return __vector_bitcast<_Up>(
__lo256(_mm512_cvttps_epu32(__auto_bitcast(__v))));
else if constexpr (__x_to_x || __y_to_y || __z_to_z)
{
// go to fallback, it does the right thing. We can't use the
// _mm_floor_ps - 0x8000'0000 trick for f32->u32 because it would
// discard small input values (only 24 mantissa bits)
}
else
__assert_unreachable<_Tp>();
}
else if constexpr (__f32_to_ibw) //{{{2
return __convert_x86<_To>(__convert_x86<__vector_type_t<int, _Np>>(__v));
else if constexpr (__f64_to_s64) //{{{2
{
if constexpr (__have_avx512dq_vl && __x_to_x)
return __intrin_bitcast<_To>(_mm_cvttpd_epi64(__intrin));
else if constexpr (__have_avx512dq_vl && __y_to_y)
return __intrin_bitcast<_To>(_mm256_cvttpd_epi64(__intrin));
else if constexpr (__have_avx512dq && __z_to_z)
return __intrin_bitcast<_To>(_mm512_cvttpd_epi64(__intrin));
// else use scalar fallback
}
else if constexpr (__f64_to_u64) //{{{2
{
if constexpr (__have_avx512dq_vl && __x_to_x)
return __intrin_bitcast<_To>(_mm_cvttpd_epu64(__intrin));
else if constexpr (__have_avx512dq_vl && __y_to_y)
return __intrin_bitcast<_To>(_mm256_cvttpd_epu64(__intrin));
else if constexpr (__have_avx512dq && __z_to_z)
return __intrin_bitcast<_To>(_mm512_cvttpd_epu64(__intrin));
// else use scalar fallback
}
else if constexpr (__f64_to_s32) //{{{2
{
if constexpr (__x_to_x)
return __intrin_bitcast<_To>(_mm_cvttpd_epi32(__intrin));
else if constexpr (__y_to_x)
return __intrin_bitcast<_To>(_mm256_cvttpd_epi32(__intrin));
else if constexpr (__z_to_y)
return __intrin_bitcast<_To>(_mm512_cvttpd_epi32(__intrin));
}
else if constexpr (__f64_to_u32) //{{{2
{
if constexpr (__have_avx512vl && __x_to_x)
return __intrin_bitcast<_To>(_mm_cvttpd_epu32(__intrin));
else if constexpr (__have_sse4_1 && __x_to_x)
return __vector_bitcast<_Up, _M>(
_mm_cvttpd_epi32(_mm_floor_pd(__intrin) - 0x8000'0000u))
^ 0x8000'0000u;
else if constexpr (__x_to_x)
{
// use scalar fallback: it's only 2 values to convert, can't get
// much better than scalar decomposition
}
else if constexpr (__have_avx512vl && __y_to_x)
return __intrin_bitcast<_To>(_mm256_cvttpd_epu32(__intrin));
else if constexpr (__y_to_x)
{
return __intrin_bitcast<_To>(
__vector_bitcast<_Up>(
_mm256_cvttpd_epi32(_mm256_floor_pd(__intrin) - 0x8000'0000u))
^ 0x8000'0000u);
}
else if constexpr (__z_to_y)
return __intrin_bitcast<_To>(_mm512_cvttpd_epu32(__intrin));
}
else if constexpr (__f64_to_ibw) //{{{2
{
return __convert_x86<_To>(
__convert_x86<__vector_type_t<int, (_Np < 4 ? 4 : _Np)>>(__v));
}
else if constexpr (__s64_to_f32) //{{{2
{
if constexpr (__x_to_x && __have_avx512dq_vl)
return __intrin_bitcast<_To>(_mm_cvtepi64_ps(__intrin));
else if constexpr (__y_to_x && __have_avx512dq_vl)
return __intrin_bitcast<_To>(_mm256_cvtepi64_ps(__intrin));
else if constexpr (__z_to_y && __have_avx512dq)
return __intrin_bitcast<_To>(_mm512_cvtepi64_ps(__intrin));
else if constexpr (__z_to_y)
return __intrin_bitcast<_To>(
_mm512_cvtpd_ps(__convert_x86<__vector_type_t<double, 8>>(__v)));
}
else if constexpr (__u64_to_f32) //{{{2
{
if constexpr (__x_to_x && __have_avx512dq_vl)
return __intrin_bitcast<_To>(_mm_cvtepu64_ps(__intrin));
else if constexpr (__y_to_x && __have_avx512dq_vl)
return __intrin_bitcast<_To>(_mm256_cvtepu64_ps(__intrin));
else if constexpr (__z_to_y && __have_avx512dq)
return __intrin_bitcast<_To>(_mm512_cvtepu64_ps(__intrin));
else if constexpr (__z_to_y)
{
return __intrin_bitcast<_To>(
__lo256(_mm512_cvtepu32_ps(__auto_bitcast(
_mm512_cvtepi64_epi32(_mm512_srai_epi64(__intrin, 32)))))
* 0x100000000LL
+ __lo256(_mm512_cvtepu32_ps(
__auto_bitcast(_mm512_cvtepi64_epi32(__intrin)))));
}
}
else if constexpr (__s32_to_f32) //{{{2
{
// use fallback (builtin conversion)
}
else if constexpr (__u32_to_f32) //{{{2
{
if constexpr (__x_to_x && __have_avx512vl)
{
// use fallback
}
else if constexpr (__x_to_x && __have_avx512f)
return __intrin_bitcast<_To>(
__lo128(_mm512_cvtepu32_ps(__auto_bitcast(__v))));
else if constexpr (__x_to_x && (__have_fma || __have_fma4))
// work around PR85819
return __auto_bitcast(0x10000
* _mm_cvtepi32_ps(__to_intrin(__v >> 16))
+ _mm_cvtepi32_ps(__to_intrin(__v & 0xffff)));
else if constexpr (__y_to_y && __have_avx512vl)
{
// use fallback
}
else if constexpr (__y_to_y && __have_avx512f)
return __intrin_bitcast<_To>(
__lo256(_mm512_cvtepu32_ps(__auto_bitcast(__v))));
else if constexpr (__y_to_y)
// work around PR85819
return 0x10000 * _mm256_cvtepi32_ps(__to_intrin(__v >> 16))
+ _mm256_cvtepi32_ps(__to_intrin(__v & 0xffff));
// else use fallback (builtin conversion)
}
else if constexpr (__ibw_to_f32) //{{{2
{
if constexpr (_M <= 4 || __have_avx2)
return __convert_x86<_To>(
__convert_x86<__vector_type_t<int, _M>>(__v));
else
{
static_assert(__x_to_y);
__m128i __a, __b;
if constexpr (__have_sse4_1)
{
__a = sizeof(_Tp) == 2
? (is_signed_v<_Tp> ? _mm_cvtepi16_epi32(__intrin)
: _mm_cvtepu16_epi32(__intrin))
: (is_signed_v<_Tp> ? _mm_cvtepi8_epi32(__intrin)
: _mm_cvtepu8_epi32(__intrin));
const auto __w
= _mm_shuffle_epi32(__intrin, sizeof(_Tp) == 2 ? 0xee : 0xe9);
__b = sizeof(_Tp) == 2
? (is_signed_v<_Tp> ? _mm_cvtepi16_epi32(__w)
: _mm_cvtepu16_epi32(__w))
: (is_signed_v<_Tp> ? _mm_cvtepi8_epi32(__w)
: _mm_cvtepu8_epi32(__w));
}
else
{
__m128i __tmp;
if constexpr (sizeof(_Tp) == 1)
{
__tmp = is_signed_v<_Tp>
? _mm_srai_epi16(_mm_unpacklo_epi8(__intrin,
__intrin),
8)
: _mm_unpacklo_epi8(__intrin, __m128i());
}
else
{
static_assert(sizeof(_Tp) == 2);
__tmp = __intrin;
}
__a = is_signed_v<_Tp>
? _mm_srai_epi32(_mm_unpacklo_epi16(__tmp, __tmp), 16)
: _mm_unpacklo_epi16(__tmp, __m128i());
__b = is_signed_v<_Tp>
? _mm_srai_epi32(_mm_unpackhi_epi16(__tmp, __tmp), 16)
: _mm_unpackhi_epi16(__tmp, __m128i());
}
return __convert_x86<_To>(__vector_bitcast<int>(__a),
__vector_bitcast<int>(__b));
}
}
else if constexpr (__s64_to_f64) //{{{2
{
if constexpr (__x_to_x && __have_avx512dq_vl)
return __intrin_bitcast<_To>(_mm_cvtepi64_pd(__intrin));
else if constexpr (__y_to_y && __have_avx512dq_vl)
return __intrin_bitcast<_To>(_mm256_cvtepi64_pd(__intrin));
else if constexpr (__z_to_z && __have_avx512dq)
return __intrin_bitcast<_To>(_mm512_cvtepi64_pd(__intrin));
else if constexpr (__z_to_z)
{
return __intrin_bitcast<_To>(
_mm512_cvtepi32_pd(_mm512_cvtepi64_epi32(__to_intrin(__v >> 32)))
* 0x100000000LL
+ _mm512_cvtepu32_pd(_mm512_cvtepi64_epi32(__intrin)));
}
}
else if constexpr (__u64_to_f64) //{{{2
{
if constexpr (__x_to_x && __have_avx512dq_vl)
return __intrin_bitcast<_To>(_mm_cvtepu64_pd(__intrin));
else if constexpr (__y_to_y && __have_avx512dq_vl)
return __intrin_bitcast<_To>(_mm256_cvtepu64_pd(__intrin));
else if constexpr (__z_to_z && __have_avx512dq)
return __intrin_bitcast<_To>(_mm512_cvtepu64_pd(__intrin));
else if constexpr (__z_to_z)
{
return __intrin_bitcast<_To>(
_mm512_cvtepu32_pd(_mm512_cvtepi64_epi32(__to_intrin(__v >> 32)))
* 0x100000000LL
+ _mm512_cvtepu32_pd(_mm512_cvtepi64_epi32(__intrin)));
}
}
else if constexpr (__s32_to_f64) //{{{2
{
if constexpr (__x_to_x)
return __intrin_bitcast<_To>(_mm_cvtepi32_pd(__intrin));
else if constexpr (__x_to_y)
return __intrin_bitcast<_To>(_mm256_cvtepi32_pd(__intrin));
else if constexpr (__y_to_z)
return __intrin_bitcast<_To>(_mm512_cvtepi32_pd(__intrin));
}
else if constexpr (__u32_to_f64) //{{{2
{
if constexpr (__x_to_x && __have_avx512vl)
return __intrin_bitcast<_To>(_mm_cvtepu32_pd(__intrin));
else if constexpr (__x_to_x && __have_avx512f)
return __intrin_bitcast<_To>(
__lo128(_mm512_cvtepu32_pd(__auto_bitcast(__v))));
else if constexpr (__x_to_x)
return __intrin_bitcast<_To>(
_mm_cvtepi32_pd(__to_intrin(__v ^ 0x8000'0000u)) + 0x8000'0000u);
else if constexpr (__x_to_y && __have_avx512vl)
return __intrin_bitcast<_To>(_mm256_cvtepu32_pd(__intrin));
else if constexpr (__x_to_y && __have_avx512f)
return __intrin_bitcast<_To>(
__lo256(_mm512_cvtepu32_pd(__auto_bitcast(__v))));
else if constexpr (__x_to_y)
return __intrin_bitcast<_To>(
_mm256_cvtepi32_pd(__to_intrin(__v ^ 0x8000'0000u)) + 0x8000'0000u);
else if constexpr (__y_to_z)
return __intrin_bitcast<_To>(_mm512_cvtepu32_pd(__intrin));
}
else if constexpr (__ibw_to_f64) //{{{2
{
return __convert_x86<_To>(
__convert_x86<__vector_type_t<int, std::max(size_t(4), _M)>>(__v));
}
else if constexpr (__f32_to_f64) //{{{2
{
if constexpr (__x_to_x)
return __intrin_bitcast<_To>(_mm_cvtps_pd(__intrin));
else if constexpr (__x_to_y)
return __intrin_bitcast<_To>(_mm256_cvtps_pd(__intrin));
else if constexpr (__y_to_z)
return __intrin_bitcast<_To>(_mm512_cvtps_pd(__intrin));
}
else if constexpr (__f64_to_f32) //{{{2
{
if constexpr (__x_to_x)
return __intrin_bitcast<_To>(_mm_cvtpd_ps(__intrin));
else if constexpr (__y_to_x)
return __intrin_bitcast<_To>(_mm256_cvtpd_ps(__intrin));
else if constexpr (__z_to_y)
return __intrin_bitcast<_To>(_mm512_cvtpd_ps(__intrin));
}
else //{{{2
__assert_unreachable<_Tp>();
// fallback:{{{2
return __vector_convert<_To>(__v, make_index_sequence<std::min(_M, _Np)>());
//}}}
}
// }}}
// 2-arg __convert_x86 {{{1
template <typename _To, typename _V, typename _Traits>
_GLIBCXX_SIMD_INTRINSIC _To
__convert_x86(_V __v0, _V __v1)
{
static_assert(__is_vector_type_v<_V>);
using _Tp = typename _Traits::value_type;
constexpr size_t _Np = _Traits::_S_full_size;
[[maybe_unused]] const auto __i0 = __to_intrin(__v0);
[[maybe_unused]] const auto __i1 = __to_intrin(__v1);
using _Up = typename _VectorTraits<_To>::value_type;
constexpr size_t _M = _VectorTraits<_To>::_S_full_size;
static_assert(2 * _Np <= _M,
"__v1 would be discarded; use the one-argument "
"__convert_x86 overload instead");
// [xyz]_to_[xyz] {{{2
[[maybe_unused]] constexpr bool __x_to_x
= sizeof(__v0) <= 16 && sizeof(_To) <= 16;
[[maybe_unused]] constexpr bool __x_to_y
= sizeof(__v0) <= 16 && sizeof(_To) == 32;
[[maybe_unused]] constexpr bool __x_to_z
= sizeof(__v0) <= 16 && sizeof(_To) == 64;
[[maybe_unused]] constexpr bool __y_to_x
= sizeof(__v0) == 32 && sizeof(_To) <= 16;
[[maybe_unused]] constexpr bool __y_to_y
= sizeof(__v0) == 32 && sizeof(_To) == 32;
[[maybe_unused]] constexpr bool __y_to_z
= sizeof(__v0) == 32 && sizeof(_To) == 64;
[[maybe_unused]] constexpr bool __z_to_x
= sizeof(__v0) == 64 && sizeof(_To) <= 16;
[[maybe_unused]] constexpr bool __z_to_y
= sizeof(__v0) == 64 && sizeof(_To) == 32;
[[maybe_unused]] constexpr bool __z_to_z
= sizeof(__v0) == 64 && sizeof(_To) == 64;
// iX_to_iX {{{2
[[maybe_unused]] constexpr bool __i_to_i
= is_integral_v<_Up> && is_integral_v<_Tp>;
[[maybe_unused]] constexpr bool __i8_to_i16
= __i_to_i && sizeof(_Tp) == 1 && sizeof(_Up) == 2;
[[maybe_unused]] constexpr bool __i8_to_i32
= __i_to_i && sizeof(_Tp) == 1 && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __i8_to_i64
= __i_to_i && sizeof(_Tp) == 1 && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __i16_to_i8
= __i_to_i && sizeof(_Tp) == 2 && sizeof(_Up) == 1;
[[maybe_unused]] constexpr bool __i16_to_i32
= __i_to_i && sizeof(_Tp) == 2 && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __i16_to_i64
= __i_to_i && sizeof(_Tp) == 2 && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __i32_to_i8
= __i_to_i && sizeof(_Tp) == 4 && sizeof(_Up) == 1;
[[maybe_unused]] constexpr bool __i32_to_i16
= __i_to_i && sizeof(_Tp) == 4 && sizeof(_Up) == 2;
[[maybe_unused]] constexpr bool __i32_to_i64
= __i_to_i && sizeof(_Tp) == 4 && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __i64_to_i8
= __i_to_i && sizeof(_Tp) == 8 && sizeof(_Up) == 1;
[[maybe_unused]] constexpr bool __i64_to_i16
= __i_to_i && sizeof(_Tp) == 8 && sizeof(_Up) == 2;
[[maybe_unused]] constexpr bool __i64_to_i32
= __i_to_i && sizeof(_Tp) == 8 && sizeof(_Up) == 4;
// [fsu]X_to_[fsu]X {{{2
// ibw = integral && byte or word, i.e. char and short with any signedness
[[maybe_unused]] constexpr bool __i64_to_f32
= is_integral_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __s32_to_f32
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __s16_to_f32
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 2
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __s8_to_f32
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 1
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __u32_to_f32
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __u16_to_f32
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 2
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __u8_to_f32
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 1
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __s64_to_f64
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __s32_to_f64
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __s16_to_f64
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 2
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __s8_to_f64
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 1
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __u64_to_f64
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __u32_to_f64
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __u16_to_f64
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 2
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __u8_to_f64
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 1
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __f32_to_s64
= is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) == 8
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f32_to_s32
= is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) == 4
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f32_to_u64
= is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) == 8
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f32_to_u32
= is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) == 4
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f64_to_s64
= is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) == 8
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f64_to_s32
= is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) == 4
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f64_to_u64
= is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) == 8
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f64_to_u32
= is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) == 4
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f32_to_ibw
= is_integral_v<_Up> && sizeof(_Up) <= 2
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f64_to_ibw
= is_integral_v<_Up> && sizeof(_Up) <= 2
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f32_to_f64
= is_floating_point_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __f64_to_f32
= is_floating_point_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
if constexpr (__i_to_i && __y_to_x && !__have_avx2) //{{{2
// <double, 4>, <double, 4> => <short, 8>
return __convert_x86<_To>(__lo128(__v0), __hi128(__v0), __lo128(__v1),
__hi128(__v1));
else if constexpr (__i_to_i) // assert ISA {{{2
{
static_assert(__x_to_x || __have_avx2,
"integral conversions with ymm registers require AVX2");
static_assert(__have_avx512bw
|| ((sizeof(_Tp) >= 4 || sizeof(__v0) < 64)
&& (sizeof(_Up) >= 4 || sizeof(_To) < 64)),
"8/16-bit integers in zmm registers require AVX512BW");
static_assert((sizeof(__v0) < 64 && sizeof(_To) < 64) || __have_avx512f,
"integral conversions with ymm registers require AVX2");
}
// concat => use 1-arg __convert_x86 {{{2
if constexpr (sizeof(__v0) < 16 || (sizeof(__v0) == 16 && __have_avx2)
|| (sizeof(__v0) == 16 && __have_avx
&& is_floating_point_v<_Tp>)
|| (sizeof(__v0) == 32 && __have_avx512f
&& (sizeof(_Tp) >= 4 || __have_avx512bw)))
{
// The ISA can handle wider input registers, so concat and use one-arg
// implementation. This reduces code duplication considerably.
return __convert_x86<_To>(__concat(__v0, __v1));
}
else //{{{2
{
// conversion using bit reinterpretation (or no conversion at all)
// should all go through the concat branch above:
static_assert(
!(is_floating_point_v<
_Tp> == is_floating_point_v<_Up> && sizeof(_Tp) == sizeof(_Up)));
// handle all zero extension{{{2
if constexpr (2 * _Np < _M && sizeof(_To) > 16)
{
constexpr size_t Min = 16 / sizeof(_Up);
return __zero_extend(
__convert_x86<
__vector_type_t<_Up, (Min > 2 * _Np) ? Min : 2 * _Np>>(__v0,
__v1));
}
else if constexpr (__i64_to_i32) //{{{2
{
if constexpr (__x_to_x)
return __auto_bitcast(_mm_shuffle_ps(__auto_bitcast(__v0),
__auto_bitcast(__v1), 0x88));
else if constexpr (__y_to_y)
{
// AVX512F is not available (would concat otherwise)
return __auto_bitcast(
__xzyw(_mm256_shuffle_ps(__auto_bitcast(__v0),
__auto_bitcast(__v1), 0x88)));
// alternative:
// const auto v0_abxxcdxx = _mm256_shuffle_epi32(__v0, 8);
// const auto v1_efxxghxx = _mm256_shuffle_epi32(__v1, 8);
// const auto v_abefcdgh = _mm256_unpacklo_epi64(v0_abxxcdxx,
// v1_efxxghxx); return _mm256_permute4x64_epi64(v_abefcdgh,
// 0x01 * 0 + 0x04 * 2 + 0x10 * 1 + 0x40 * 3); // abcdefgh
}
else if constexpr (__z_to_z)
return __intrin_bitcast<_To>(
__concat(_mm512_cvtepi64_epi32(__i0),
_mm512_cvtepi64_epi32(__i1)));
}
else if constexpr (__i64_to_i16) //{{{2
{
if constexpr (__x_to_x)
{
// AVX2 is not available (would concat otherwise)
if constexpr (__have_sse4_1)
{
return __intrin_bitcast<_To>(_mm_shuffle_epi8(
_mm_blend_epi16(__i0, _mm_slli_si128(__i1, 4), 0x44),
_mm_setr_epi8(0, 1, 8, 9, 4, 5, 12, 13, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, -0x80)));
}
else
{
return __vector_type_t<_Up, _M>{_Up(__v0[0]), _Up(__v0[1]),
_Up(__v1[0]), _Up(__v1[1])};
}
}
else if constexpr (__y_to_x)
{
auto __a
= _mm256_unpacklo_epi16(__i0, __i1); // 04.. .... 26.. ....
auto __b
= _mm256_unpackhi_epi16(__i0, __i1); // 15.. .... 37.. ....
auto __c
= _mm256_unpacklo_epi16(__a, __b); // 0145 .... 2367 ....
return __intrin_bitcast<_To>(
_mm_unpacklo_epi32(__lo128(__c), __hi128(__c))); // 0123 4567
}
else if constexpr (__z_to_y)
return __intrin_bitcast<_To>(
__concat(_mm512_cvtepi64_epi16(__i0),
_mm512_cvtepi64_epi16(__i1)));
}
else if constexpr (__i64_to_i8) //{{{2
{
if constexpr (__x_to_x && __have_sse4_1)
{
return __intrin_bitcast<_To>(_mm_shuffle_epi8(
_mm_blend_epi16(__i0, _mm_slli_si128(__i1, 4), 0x44),
_mm_setr_epi8(0, 8, 4, 12, -0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
-0x80)));
}
else if constexpr (__x_to_x && __have_ssse3)
{
return __intrin_bitcast<_To>(_mm_unpacklo_epi16(
_mm_shuffle_epi8(
__i0, _mm_setr_epi8(0, 8, -0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80)),
_mm_shuffle_epi8(
__i1, _mm_setr_epi8(0, 8, -0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80))));
}
else if constexpr (__x_to_x)
{
return __vector_type_t<_Up, _M>{_Up(__v0[0]), _Up(__v0[1]),
_Up(__v1[0]), _Up(__v1[1])};
}
else if constexpr (__y_to_x)
{
const auto __a = _mm256_shuffle_epi8(
_mm256_blend_epi32(__i0, _mm256_slli_epi64(__i1, 32), 0xAA),
_mm256_setr_epi8(0, 8, -0x80, -0x80, 4, 12, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, 0, 8, -0x80,
-0x80, 4, 12, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80));
return __intrin_bitcast<_To>(__lo128(__a) | __hi128(__a));
} // __z_to_x uses concat fallback
}
else if constexpr (__i32_to_i16) //{{{2
{
if constexpr (__x_to_x)
{
// AVX2 is not available (would concat otherwise)
if constexpr (__have_sse4_1)
{
return __intrin_bitcast<_To>(_mm_shuffle_epi8(
_mm_blend_epi16(__i0, _mm_slli_si128(__i1, 2), 0xaa),
_mm_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10,
11, 14, 15)));
}
else if constexpr (__have_ssse3)
{
return __intrin_bitcast<_To>(
_mm_hadd_epi16(__to_intrin(__v0 << 16),
__to_intrin(__v1 << 16)));
/*
return _mm_unpacklo_epi64(
_mm_shuffle_epi8(__i0, _mm_setr_epi8(0, 1, 4, 5, 8, 9,
12, 13, 8, 9, 12, 13, 12, 13, 14, 15)),
_mm_shuffle_epi8(__i1, _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12,
13, 8, 9, 12, 13, 12, 13, 14, 15)));
*/
}
else
{
auto __a = _mm_unpacklo_epi16(__i0, __i1); // 04.. 15..
auto __b = _mm_unpackhi_epi16(__i0, __i1); // 26.. 37..
auto __c = _mm_unpacklo_epi16(__a, __b); // 0246 ....
auto __d = _mm_unpackhi_epi16(__a, __b); // 1357 ....
return __intrin_bitcast<_To>(
_mm_unpacklo_epi16(__c, __d)); // 0123 4567
}
}
else if constexpr (__y_to_y)
{
const auto __shuf
= _mm256_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
0, 1, 4, 5, 8, 9, 12, 13, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, -0x80);
auto __a = _mm256_shuffle_epi8(__i0, __shuf);
auto __b = _mm256_shuffle_epi8(__i1, __shuf);
return __intrin_bitcast<_To>(
__xzyw(_mm256_unpacklo_epi64(__a, __b)));
} // __z_to_z uses concat fallback
}
else if constexpr (__i32_to_i8) //{{{2
{
if constexpr (__x_to_x && __have_ssse3)
{
const auto shufmask
= _mm_setr_epi8(0, 4, 8, 12, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80);
return __intrin_bitcast<_To>(
_mm_unpacklo_epi32(_mm_shuffle_epi8(__i0, shufmask),
_mm_shuffle_epi8(__i1, shufmask)));
}
else if constexpr (__x_to_x)
{
auto __a = _mm_unpacklo_epi8(__i0, __i1); // 04.. .... 15.. ....
auto __b = _mm_unpackhi_epi8(__i0, __i1); // 26.. .... 37.. ....
auto __c = _mm_unpacklo_epi8(__a, __b); // 0246 .... .... ....
auto __d = _mm_unpackhi_epi8(__a, __b); // 1357 .... .... ....
auto __e = _mm_unpacklo_epi8(__c, __d); // 0123 4567 .... ....
return __intrin_bitcast<_To>(__e & __m128i{-1, 0});
}
else if constexpr (__y_to_x)
{
const auto __a = _mm256_shuffle_epi8(
_mm256_blend_epi16(__i0, _mm256_slli_epi32(__i1, 16), 0xAA),
_mm256_setr_epi8(0, 4, 8, 12, -0x80, -0x80, -0x80, -0x80, 2,
6, 10, 14, -0x80, -0x80, -0x80, -0x80, -0x80,
-0x80, -0x80, -0x80, 0, 4, 8, 12, -0x80,
-0x80, -0x80, -0x80, 2, 6, 10, 14));
return __intrin_bitcast<_To>(__lo128(__a) | __hi128(__a));
} // __z_to_y uses concat fallback
}
else if constexpr (__i16_to_i8) //{{{2
{
if constexpr (__x_to_x && __have_ssse3)
{
const auto __shuf = reinterpret_cast<__m128i>(
__vector_type_t<_UChar, 16>{0, 2, 4, 6, 8, 10, 12, 14, 0x80,
0x80, 0x80, 0x80, 0x80, 0x80,
0x80, 0x80});
return __intrin_bitcast<_To>(
_mm_unpacklo_epi64(_mm_shuffle_epi8(__i0, __shuf),
_mm_shuffle_epi8(__i1, __shuf)));
}
else if constexpr (__x_to_x)
{
auto __a = _mm_unpacklo_epi8(__i0, __i1); // 08.. 19.. 2A.. 3B..
auto __b = _mm_unpackhi_epi8(__i0, __i1); // 4C.. 5D.. 6E.. 7F..
auto __c = _mm_unpacklo_epi8(__a, __b); // 048C .... 159D ....
auto __d = _mm_unpackhi_epi8(__a, __b); // 26AE .... 37BF ....
auto __e = _mm_unpacklo_epi8(__c, __d); // 0246 8ACE .... ....
auto __f = _mm_unpackhi_epi8(__c, __d); // 1357 9BDF .... ....
return __intrin_bitcast<_To>(_mm_unpacklo_epi8(__e, __f));
}
else if constexpr (__y_to_y)
{
return __intrin_bitcast<_To>(__xzyw(_mm256_shuffle_epi8(
(__to_intrin(__v0) & _mm256_set1_epi32(0x00ff00ff))
| _mm256_slli_epi16(__i1, 8),
_mm256_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11,
13, 15, 0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5,
7, 9, 11, 13, 15))));
} // __z_to_z uses concat fallback
}
else if constexpr (__i64_to_f32) //{{{2
{
if constexpr (__x_to_x)
return __make_wrapper<float>(__v0[0], __v0[1], __v1[0], __v1[1]);
else if constexpr (__y_to_y)
{
static_assert(__y_to_y && __have_avx2);
const auto __a = _mm256_unpacklo_epi32(__i0, __i1); // aeAE cgCG
const auto __b = _mm256_unpackhi_epi32(__i0, __i1); // bfBF dhDH
const auto __lo32
= _mm256_unpacklo_epi32(__a, __b); // abef cdgh
const auto __hi32 = __vector_bitcast<
conditional_t<is_signed_v<_Tp>, int, _UInt>>(
_mm256_unpackhi_epi32(__a, __b)); // ABEF CDGH
const auto __hi
= 0x100000000LL
* __convert_x86<__vector_type_t<float, 8>>(__hi32);
const auto __mid
= 0x10000 * _mm256_cvtepi32_ps(_mm256_srli_epi32(__lo32, 16));
const auto __lo
= _mm256_cvtepi32_ps(_mm256_set1_epi32(0x0000ffffu) & __lo32);
return __xzyw((__hi + __mid) + __lo);
}
else if constexpr (__z_to_z && __have_avx512dq)
{
return is_signed_v<_Tp> ? __concat(_mm512_cvtepi64_ps(__i0),
_mm512_cvtepi64_ps(__i1))
: __concat(_mm512_cvtepu64_ps(__i0),
_mm512_cvtepu64_ps(__i1));
}
else if constexpr (__z_to_z && is_signed_v<_Tp>)
{
const __m512 __hi32 = _mm512_cvtepi32_ps(
__concat(_mm512_cvtepi64_epi32(__to_intrin(__v0 >> 32)),
_mm512_cvtepi64_epi32(__to_intrin(__v1 >> 32))));
const __m512i __lo32 = __concat(_mm512_cvtepi64_epi32(__i0),
_mm512_cvtepi64_epi32(__i1));
// split low 32-bits, because if __hi32 is a small negative
// number, the 24-bit mantissa may lose important information if
// any of the high 8 bits of __lo32 is set, leading to
// catastrophic cancelation in the FMA
const __m512 __hi16
= _mm512_cvtepu32_ps(_mm512_set1_epi32(0xffff0000u) & __lo32);
const __m512 __lo16
= _mm512_cvtepi32_ps(_mm512_set1_epi32(0x0000ffffu) & __lo32);
return (__hi32 * 0x100000000LL + __hi16) + __lo16;
}
else if constexpr (__z_to_z && is_unsigned_v<_Tp>)
{
return __intrin_bitcast<_To>(
_mm512_cvtepu32_ps(__concat(
_mm512_cvtepi64_epi32(_mm512_srai_epi64(__i0, 32)),
_mm512_cvtepi64_epi32(_mm512_srai_epi64(__i1, 32))))
* 0x100000000LL
+ _mm512_cvtepu32_ps(__concat(_mm512_cvtepi64_epi32(__i0),
_mm512_cvtepi64_epi32(__i1))));
}
}
else if constexpr (__f64_to_s32) //{{{2
{
// use concat fallback
}
else if constexpr (__f64_to_u32) //{{{2
{
if constexpr (__x_to_x && __have_sse4_1)
{
return __vector_bitcast<_Up, _M>(_mm_unpacklo_epi64(
_mm_cvttpd_epi32(_mm_floor_pd(__i0) - 0x8000'0000u),
_mm_cvttpd_epi32(_mm_floor_pd(__i1) - 0x8000'0000u)))
^ 0x8000'0000u;
// without SSE4.1 just use the scalar fallback, it's only four
// values
}
else if constexpr (__y_to_y)
{
return __vector_bitcast<_Up>(
__concat(_mm256_cvttpd_epi32(_mm256_floor_pd(__i0)
- 0x8000'0000u),
_mm256_cvttpd_epi32(_mm256_floor_pd(__i1)
- 0x8000'0000u)))
^ 0x8000'0000u;
} // __z_to_z uses fallback
}
else if constexpr (__f64_to_ibw) //{{{2
{
// one-arg __f64_to_ibw goes via _SimdWrapper<int, ?>. The fallback
// would go via two independet conversions to _SimdWrapper<_To> and
// subsequent interleaving. This is better, because f64->__i32
// allows to combine __v0 and __v1 into one register: if constexpr
// (__z_to_x || __y_to_x) {
return __convert_x86<_To>(
__convert_x86<__vector_type_t<int, _Np * 2>>(__v0, __v1));
//}
}
else if constexpr (__f32_to_ibw) //{{{2
{
return __convert_x86<_To>(
__convert_x86<__vector_type_t<int, _Np>>(__v0),
__convert_x86<__vector_type_t<int, _Np>>(__v1));
} //}}}
// fallback: {{{2
if constexpr (sizeof(_To) >= 32)
// if _To is ymm or zmm, then _SimdWrapper<_Up, _M / 2> is xmm or ymm
return __concat(__convert_x86<__vector_type_t<_Up, _M / 2>>(__v0),
__convert_x86<__vector_type_t<_Up, _M / 2>>(__v1));
else if constexpr (sizeof(_To) == 16)
{
const auto __lo = __to_intrin(__convert_x86<_To>(__v0));
const auto __hi = __to_intrin(__convert_x86<_To>(__v1));
if constexpr (sizeof(_Up) * _Np == 8)
{
if constexpr (is_floating_point_v<_Up>)
return __auto_bitcast(
_mm_unpacklo_pd(__vector_bitcast<double>(__lo),
__vector_bitcast<double>(__hi)));
else
return __intrin_bitcast<_To>(_mm_unpacklo_epi64(__lo, __hi));
}
else if constexpr (sizeof(_Up) * _Np == 4)
{
if constexpr (is_floating_point_v<_Up>)
return __auto_bitcast(
_mm_unpacklo_ps(__vector_bitcast<float>(__lo),
__vector_bitcast<float>(__hi)));
else
return __intrin_bitcast<_To>(_mm_unpacklo_epi32(__lo, __hi));
}
else if constexpr (sizeof(_Up) * _Np == 2)
return __intrin_bitcast<_To>(_mm_unpacklo_epi16(__lo, __hi));
else
__assert_unreachable<_Tp>();
}
else
return __vector_convert<_To>(__v0, __v1, make_index_sequence<_Np>());
//}}}
}
}
//}}}1
// 4-arg __convert_x86 {{{1
template <typename _To, typename _V, typename _Traits>
_GLIBCXX_SIMD_INTRINSIC _To
__convert_x86(_V __v0, _V __v1, _V __v2, _V __v3)
{
static_assert(__is_vector_type_v<_V>);
using _Tp = typename _Traits::value_type;
constexpr size_t _Np = _Traits::_S_full_size;
[[maybe_unused]] const auto __i0 = __to_intrin(__v0);
[[maybe_unused]] const auto __i1 = __to_intrin(__v1);
[[maybe_unused]] const auto __i2 = __to_intrin(__v2);
[[maybe_unused]] const auto __i3 = __to_intrin(__v3);
using _Up = typename _VectorTraits<_To>::value_type;
constexpr size_t _M = _VectorTraits<_To>::_S_full_size;
static_assert(4 * _Np <= _M,
"__v2/__v3 would be discarded; use the two/one-argument "
"__convert_x86 overload instead");
// [xyz]_to_[xyz] {{{2
[[maybe_unused]] constexpr bool __x_to_x
= sizeof(__v0) <= 16 && sizeof(_To) <= 16;
[[maybe_unused]] constexpr bool __x_to_y
= sizeof(__v0) <= 16 && sizeof(_To) == 32;
[[maybe_unused]] constexpr bool __x_to_z
= sizeof(__v0) <= 16 && sizeof(_To) == 64;
[[maybe_unused]] constexpr bool __y_to_x
= sizeof(__v0) == 32 && sizeof(_To) <= 16;
[[maybe_unused]] constexpr bool __y_to_y
= sizeof(__v0) == 32 && sizeof(_To) == 32;
[[maybe_unused]] constexpr bool __y_to_z
= sizeof(__v0) == 32 && sizeof(_To) == 64;
[[maybe_unused]] constexpr bool __z_to_x
= sizeof(__v0) == 64 && sizeof(_To) <= 16;
[[maybe_unused]] constexpr bool __z_to_y
= sizeof(__v0) == 64 && sizeof(_To) == 32;
[[maybe_unused]] constexpr bool __z_to_z
= sizeof(__v0) == 64 && sizeof(_To) == 64;
// iX_to_iX {{{2
[[maybe_unused]] constexpr bool __i_to_i
= is_integral_v<_Up> && is_integral_v<_Tp>;
[[maybe_unused]] constexpr bool __i8_to_i16
= __i_to_i && sizeof(_Tp) == 1 && sizeof(_Up) == 2;
[[maybe_unused]] constexpr bool __i8_to_i32
= __i_to_i && sizeof(_Tp) == 1 && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __i8_to_i64
= __i_to_i && sizeof(_Tp) == 1 && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __i16_to_i8
= __i_to_i && sizeof(_Tp) == 2 && sizeof(_Up) == 1;
[[maybe_unused]] constexpr bool __i16_to_i32
= __i_to_i && sizeof(_Tp) == 2 && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __i16_to_i64
= __i_to_i && sizeof(_Tp) == 2 && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __i32_to_i8
= __i_to_i && sizeof(_Tp) == 4 && sizeof(_Up) == 1;
[[maybe_unused]] constexpr bool __i32_to_i16
= __i_to_i && sizeof(_Tp) == 4 && sizeof(_Up) == 2;
[[maybe_unused]] constexpr bool __i32_to_i64
= __i_to_i && sizeof(_Tp) == 4 && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __i64_to_i8
= __i_to_i && sizeof(_Tp) == 8 && sizeof(_Up) == 1;
[[maybe_unused]] constexpr bool __i64_to_i16
= __i_to_i && sizeof(_Tp) == 8 && sizeof(_Up) == 2;
[[maybe_unused]] constexpr bool __i64_to_i32
= __i_to_i && sizeof(_Tp) == 8 && sizeof(_Up) == 4;
// [fsu]X_to_[fsu]X {{{2
// ibw = integral && byte or word, i.e. char and short with any signedness
[[maybe_unused]] constexpr bool __i64_to_f32
= is_integral_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __s32_to_f32
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __s16_to_f32
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 2
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __s8_to_f32
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 1
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __u32_to_f32
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __u16_to_f32
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 2
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __u8_to_f32
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 1
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
[[maybe_unused]] constexpr bool __s64_to_f64
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __s32_to_f64
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __s16_to_f64
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 2
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __s8_to_f64
= is_integral_v<_Tp> && is_signed_v<_Tp> && sizeof(_Tp) == 1
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __u64_to_f64
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __u32_to_f64
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __u16_to_f64
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 2
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __u8_to_f64
= is_integral_v<_Tp> && is_unsigned_v<_Tp> && sizeof(_Tp) == 1
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __f32_to_s64
= is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) == 8
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f32_to_s32
= is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) == 4
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f32_to_u64
= is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) == 8
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f32_to_u32
= is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) == 4
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f64_to_s64
= is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) == 8
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f64_to_s32
= is_integral_v<_Up> && is_signed_v<_Up> && sizeof(_Up) == 4
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f64_to_u64
= is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) == 8
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f64_to_u32
= is_integral_v<_Up> && is_unsigned_v<_Up> && sizeof(_Up) == 4
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f32_to_ibw
= is_integral_v<_Up> && sizeof(_Up) <= 2
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 4;
[[maybe_unused]] constexpr bool __f64_to_ibw
= is_integral_v<_Up> && sizeof(_Up) <= 2
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
[[maybe_unused]] constexpr bool __f32_to_f64
= is_floating_point_v<_Tp> && sizeof(_Tp) == 4
&& is_floating_point_v<_Up> && sizeof(_Up) == 8;
[[maybe_unused]] constexpr bool __f64_to_f32
= is_floating_point_v<_Tp> && sizeof(_Tp) == 8
&& is_floating_point_v<_Up> && sizeof(_Up) == 4;
if constexpr (__i_to_i && __y_to_x && !__have_avx2) //{{{2
{
// <double, 4>, <double, 4>, <double, 4>, <double, 4> => <char, 16>
return __convert_x86<_To>(__lo128(__v0), __hi128(__v0), __lo128(__v1),
__hi128(__v1), __lo128(__v2), __hi128(__v2),
__lo128(__v3), __hi128(__v3));
}
else if constexpr (__i_to_i) // assert ISA {{{2
{
static_assert(__x_to_x || __have_avx2,
"integral conversions with ymm registers require AVX2");
static_assert(__have_avx512bw
|| ((sizeof(_Tp) >= 4 || sizeof(__v0) < 64)
&& (sizeof(_Up) >= 4 || sizeof(_To) < 64)),
"8/16-bit integers in zmm registers require AVX512BW");
static_assert((sizeof(__v0) < 64 && sizeof(_To) < 64) || __have_avx512f,
"integral conversions with ymm registers require AVX2");
}
// concat => use 2-arg __convert_x86 {{{2
if constexpr (sizeof(__v0) < 16 || (sizeof(__v0) == 16 && __have_avx2)
|| (sizeof(__v0) == 16 && __have_avx
&& is_floating_point_v<_Tp>)
|| (sizeof(__v0) == 32 && __have_avx512f))
{
// The ISA can handle wider input registers, so concat and use two-arg
// implementation. This reduces code duplication considerably.
return __convert_x86<_To>(__concat(__v0, __v1), __concat(__v2, __v3));
}
else //{{{2
{
// conversion using bit reinterpretation (or no conversion at all)
// should all go through the concat branch above:
static_assert(
!(is_floating_point_v<
_Tp> == is_floating_point_v<_Up> && sizeof(_Tp) == sizeof(_Up)));
// handle all zero extension{{{2
if constexpr (4 * _Np < _M && sizeof(_To) > 16)
{
constexpr size_t Min = 16 / sizeof(_Up);
return __zero_extend(
__convert_x86<
__vector_type_t<_Up, (Min > 4 * _Np) ? Min : 4 * _Np>>(
__v0, __v1, __v2, __v3));
}
else if constexpr (__i64_to_i16) //{{{2
{
if constexpr (__x_to_x && __have_sse4_1)
{
return __intrin_bitcast<_To>(_mm_shuffle_epi8(
_mm_blend_epi16(
_mm_blend_epi16(__i0, _mm_slli_si128(__i1, 2), 0x22),
_mm_blend_epi16(_mm_slli_si128(__i2, 4),
_mm_slli_si128(__i3, 6), 0x88),
0xcc),
_mm_setr_epi8(0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7,
14, 15)));
}
else if constexpr (__y_to_y && __have_avx2)
{
return __intrin_bitcast<_To>(_mm256_shuffle_epi8(
__xzyw(_mm256_blend_epi16(
__auto_bitcast(
_mm256_shuffle_ps(__vector_bitcast<float>(__v0),
__vector_bitcast<float>(__v2),
0x88)), // 0.1. 8.9. 2.3. A.B.
__to_intrin(__vector_bitcast<int>(_mm256_shuffle_ps(
__vector_bitcast<float>(__v1),
__vector_bitcast<float>(__v3), 0x88))
<< 16), // .4.5 .C.D .6.7 .E.F
0xaa) // 0415 8C9D 2637 AEBF
), // 0415 2637 8C9D AEBF
_mm256_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11,
14, 15, 0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7,
10, 11, 14, 15)));
/*
auto __a = _mm256_unpacklo_epi16(__v0, __v1); // 04.. .... 26..
.... auto __b = _mm256_unpackhi_epi16(__v0, __v1); // 15..
.... 37.. .... auto __c = _mm256_unpacklo_epi16(__v2, __v3); //
8C.. .... AE.. .... auto __d = _mm256_unpackhi_epi16(__v2,
__v3);
// 9D.. .... BF.. .... auto __e = _mm256_unpacklo_epi16(__a,
__b);
// 0145 .... 2367 .... auto __f = _mm256_unpacklo_epi16(__c,
__d);
// 89CD .... ABEF .... auto __g = _mm256_unpacklo_epi64(__e,
__f);
// 0145 89CD 2367 ABEF return __concat(
_mm_unpacklo_epi32(__lo128(__g), __hi128(__g)),
_mm_unpackhi_epi32(__lo128(__g), __hi128(__g))); // 0123
4567 89AB CDEF
*/
} // else use fallback
}
else if constexpr (__i64_to_i8) //{{{2
{
if constexpr (__x_to_x)
{
// TODO: use fallback for now
}
else if constexpr (__y_to_x)
{
auto __a
= _mm256_srli_epi32(_mm256_slli_epi32(__i0, 24), 24)
| _mm256_srli_epi32(_mm256_slli_epi32(__i1, 24), 16)
| _mm256_srli_epi32(_mm256_slli_epi32(__i2, 24), 8)
| _mm256_slli_epi32(
__i3, 24); // 048C .... 159D .... 26AE .... 37BF ....
/*return _mm_shuffle_epi8(
_mm_blend_epi32(__lo128(__a) << 32, __hi128(__a), 0x5),
_mm_setr_epi8(4, 12, 0, 8, 5, 13, 1, 9, 6, 14, 2, 10, 7, 15,
3, 11));*/
auto __b = _mm256_unpackhi_epi64(
__a, __a); // 159D .... 159D .... 37BF .... 37BF ....
auto __c = _mm256_unpacklo_epi8(
__a, __b); // 0145 89CD .... .... 2367 ABEF .... ....
return __intrin_bitcast<_To>(
_mm_unpacklo_epi16(__lo128(__c),
__hi128(__c))); // 0123 4567 89AB CDEF
}
}
else if constexpr (__i32_to_i8) //{{{2
{
if constexpr (__x_to_x)
{
if constexpr (__have_ssse3)
{
const auto __x0 = __vector_bitcast<_UInt>(__v0) & 0xff;
const auto __x1 = (__vector_bitcast<_UInt>(__v1) & 0xff)
<< 8;
const auto __x2 = (__vector_bitcast<_UInt>(__v2) & 0xff)
<< 16;
const auto __x3 = __vector_bitcast<_UInt>(__v3) << 24;
return __intrin_bitcast<_To>(
_mm_shuffle_epi8(__to_intrin(__x0 | __x1 | __x2 | __x3),
_mm_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13,
2, 6, 10, 14, 3, 7, 11,
15)));
}
else
{
auto __a
= _mm_unpacklo_epi8(__i0, __i2); // 08.. .... 19.. ....
auto __b
= _mm_unpackhi_epi8(__i0, __i2); // 2A.. .... 3B.. ....
auto __c
= _mm_unpacklo_epi8(__i1, __i3); // 4C.. .... 5D.. ....
auto __d
= _mm_unpackhi_epi8(__i1, __i3); // 6E.. .... 7F.. ....
auto __e
= _mm_unpacklo_epi8(__a, __c); // 048C .... .... ....
auto __f
= _mm_unpackhi_epi8(__a, __c); // 159D .... .... ....
auto __g
= _mm_unpacklo_epi8(__b, __d); // 26AE .... .... ....
auto __h
= _mm_unpackhi_epi8(__b, __d); // 37BF .... .... ....
return __intrin_bitcast<_To>(_mm_unpacklo_epi8(
_mm_unpacklo_epi8(__e, __g), // 0246 8ACE .... ....
_mm_unpacklo_epi8(__f, __h) // 1357 9BDF .... ....
)); // 0123 4567 89AB CDEF
}
}
else if constexpr (__y_to_y)
{
const auto __a = _mm256_shuffle_epi8(
__to_intrin((__vector_bitcast<_UShort>(_mm256_blend_epi16(
__i0, _mm256_slli_epi32(__i1, 16), 0xAA))
& 0xff)
| (__vector_bitcast<_UShort>(_mm256_blend_epi16(
__i2, _mm256_slli_epi32(__i3, 16), 0xAA))
<< 8)),
_mm256_setr_epi8(0, 4, 8, 12, 2, 6, 10, 14, 1, 5, 9, 13, 3, 7,
11, 15, 0, 4, 8, 12, 2, 6, 10, 14, 1, 5, 9,
13, 3, 7, 11, 15));
return __intrin_bitcast<_To>(_mm256_permutevar8x32_epi32(
__a, _mm256_setr_epi32(0, 4, 1, 5, 2, 6, 3, 7)));
}
}
else if constexpr (__i64_to_f32) //{{{2
{
// this branch is only relevant with AVX and w/o AVX2 (i.e. no ymm
// integers)
if constexpr (__x_to_y)
{
return __make_wrapper<float>(__v0[0], __v0[1], __v1[0], __v1[1],
__v2[0], __v2[1], __v3[0],
__v3[1]);
const auto __a = _mm_unpacklo_epi32(__i0, __i1); // acAC
const auto __b = _mm_unpackhi_epi32(__i0, __i1); // bdBD
const auto __c = _mm_unpacklo_epi32(__i2, __i3); // egEG
const auto __d = _mm_unpackhi_epi32(__i2, __i3); // fhFH
const auto __lo32a = _mm_unpacklo_epi32(__a, __b); // abcd
const auto __lo32b = _mm_unpacklo_epi32(__c, __d); // efgh
const auto __hi32 = __vector_bitcast<
conditional_t<is_signed_v<_Tp>, int, _UInt>>(
__concat(_mm_unpackhi_epi32(__a, __b),
_mm_unpackhi_epi32(__c, __d))); // ABCD EFGH
const auto __hi
= 0x100000000LL
* __convert_x86<__vector_type_t<float, 8>>(__hi32);
const auto __mid
= 0x10000
* _mm256_cvtepi32_ps(__concat(_mm_srli_epi32(__lo32a, 16),
_mm_srli_epi32(__lo32b, 16)));
const auto __lo = _mm256_cvtepi32_ps(
__concat(_mm_set1_epi32(0x0000ffffu) & __lo32a,
_mm_set1_epi32(0x0000ffffu) & __lo32b));
return (__hi + __mid) + __lo;
}
}
else if constexpr (__f64_to_ibw) //{{{2
{
return __convert_x86<_To>(
__convert_x86<__vector_type_t<int, _Np * 2>>(__v0, __v1),
__convert_x86<__vector_type_t<int, _Np * 2>>(__v2, __v3));
}
else if constexpr (__f32_to_ibw) //{{{2
{
return __convert_x86<_To>(
__convert_x86<__vector_type_t<int, _Np>>(__v0),
__convert_x86<__vector_type_t<int, _Np>>(__v1),
__convert_x86<__vector_type_t<int, _Np>>(__v2),
__convert_x86<__vector_type_t<int, _Np>>(__v3));
} //}}}
// fallback: {{{2
if constexpr (sizeof(_To) >= 32)
// if _To is ymm or zmm, then _SimdWrapper<_Up, _M / 2> is xmm or ymm
return __concat(__convert_x86<__vector_type_t<_Up, _M / 2>>(__v0,
__v1),
__convert_x86<__vector_type_t<_Up, _M / 2>>(__v2,
__v3));
else if constexpr (sizeof(_To) == 16)
{
const auto __lo = __to_intrin(__convert_x86<_To>(__v0, __v1));
const auto __hi = __to_intrin(__convert_x86<_To>(__v2, __v3));
if constexpr (sizeof(_Up) * _Np * 2 == 8)
{
if constexpr (is_floating_point_v<_Up>)
return __auto_bitcast(_mm_unpacklo_pd(__lo, __hi));
else
return __intrin_bitcast<_To>(_mm_unpacklo_epi64(__lo, __hi));
}
else if constexpr (sizeof(_Up) * _Np * 2 == 4)
{
if constexpr (is_floating_point_v<_Up>)
return __auto_bitcast(_mm_unpacklo_ps(__lo, __hi));
else
return __intrin_bitcast<_To>(_mm_unpacklo_epi32(__lo, __hi));
}
else
__assert_unreachable<_Tp>();
}
else
return __vector_convert<_To>(__v0, __v1, __v2, __v3,
make_index_sequence<_Np>());
//}}}2
}
}
//}}}
// 8-arg __convert_x86 {{{1
template <typename _To, typename _V, typename _Traits>
_GLIBCXX_SIMD_INTRINSIC _To
__convert_x86(_V __v0, _V __v1, _V __v2, _V __v3, _V __v4, _V __v5, _V __v6,
_V __v7)
{
static_assert(__is_vector_type_v<_V>);
using _Tp = typename _Traits::value_type;
constexpr size_t _Np = _Traits::_S_full_size;
[[maybe_unused]] const auto __i0 = __to_intrin(__v0);
[[maybe_unused]] const auto __i1 = __to_intrin(__v1);
[[maybe_unused]] const auto __i2 = __to_intrin(__v2);
[[maybe_unused]] const auto __i3 = __to_intrin(__v3);
[[maybe_unused]] const auto __i4 = __to_intrin(__v4);
[[maybe_unused]] const auto __i5 = __to_intrin(__v5);
[[maybe_unused]] const auto __i6 = __to_intrin(__v6);
[[maybe_unused]] const auto __i7 = __to_intrin(__v7);
using _Up = typename _VectorTraits<_To>::value_type;
constexpr size_t _M = _VectorTraits<_To>::_S_full_size;
static_assert(8 * _Np <= _M,
"__v4-__v7 would be discarded; use the four/two/one-argument "
"__convert_x86 overload instead");
// [xyz]_to_[xyz] {{{2
[[maybe_unused]] constexpr bool __x_to_x
= sizeof(__v0) <= 16 && sizeof(_To) <= 16;
[[maybe_unused]] constexpr bool __x_to_y
= sizeof(__v0) <= 16 && sizeof(_To) == 32;
[[maybe_unused]] constexpr bool __x_to_z
= sizeof(__v0) <= 16 && sizeof(_To) == 64;
[[maybe_unused]] constexpr bool __y_to_x
= sizeof(__v0) == 32 && sizeof(_To) <= 16;
[[maybe_unused]] constexpr bool __y_to_y
= sizeof(__v0) == 32 && sizeof(_To) == 32;
[[maybe_unused]] constexpr bool __y_to_z
= sizeof(__v0) == 32 && sizeof(_To) == 64;
[[maybe_unused]] constexpr bool __z_to_x
= sizeof(__v0) == 64 && sizeof(_To) <= 16;
[[maybe_unused]] constexpr bool __z_to_y
= sizeof(__v0) == 64 && sizeof(_To) == 32;
[[maybe_unused]] constexpr bool __z_to_z
= sizeof(__v0) == 64 && sizeof(_To) == 64;
// [if]X_to_i8 {{{2
[[maybe_unused]] constexpr bool __i_to_i
= is_integral_v<_Up> && is_integral_v<_Tp>;
[[maybe_unused]] constexpr bool __i64_to_i8
= __i_to_i && sizeof(_Tp) == 8 && sizeof(_Up) == 1;
[[maybe_unused]] constexpr bool __f64_to_i8
= is_integral_v<_Up> && sizeof(_Up) == 1
&& is_floating_point_v<_Tp> && sizeof(_Tp) == 8;
if constexpr (__i_to_i) // assert ISA {{{2
{
static_assert(__x_to_x || __have_avx2,
"integral conversions with ymm registers require AVX2");
static_assert(__have_avx512bw
|| ((sizeof(_Tp) >= 4 || sizeof(__v0) < 64)
&& (sizeof(_Up) >= 4 || sizeof(_To) < 64)),
"8/16-bit integers in zmm registers require AVX512BW");
static_assert((sizeof(__v0) < 64 && sizeof(_To) < 64) || __have_avx512f,
"integral conversions with ymm registers require AVX2");
}
// concat => use 4-arg __convert_x86 {{{2
if constexpr (sizeof(__v0) < 16 || (sizeof(__v0) == 16 && __have_avx2)
|| (sizeof(__v0) == 16 && __have_avx
&& is_floating_point_v<_Tp>)
|| (sizeof(__v0) == 32 && __have_avx512f))
{
// The ISA can handle wider input registers, so concat and use two-arg
// implementation. This reduces code duplication considerably.
return __convert_x86<_To>(__concat(__v0, __v1), __concat(__v2, __v3),
__concat(__v4, __v5), __concat(__v6, __v7));
}
else //{{{2
{
// conversion using bit reinterpretation (or no conversion at all)
// should all go through the concat branch above:
static_assert(
!(is_floating_point_v<
_Tp> == is_floating_point_v<_Up> && sizeof(_Tp) == sizeof(_Up)));
static_assert(!(8 * _Np < _M && sizeof(_To) > 16),
"zero extension should be impossible");
if constexpr (__i64_to_i8) //{{{2
{
if constexpr (__x_to_x && __have_ssse3)
{
// unsure whether this is better than the variant below
return __intrin_bitcast<_To>(_mm_shuffle_epi8(
__to_intrin(
(((__v0 & 0xff) | ((__v1 & 0xff) << 8))
| (((__v2 & 0xff) << 16) | ((__v3 & 0xff) << 24)))
| ((((__v4 & 0xff) << 32) | ((__v5 & 0xff) << 40))
| (((__v6 & 0xff) << 48) | (__v7 << 56)))),
_mm_setr_epi8(0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14,
7, 15)));
}
else if constexpr (__x_to_x)
{
const auto __a = _mm_unpacklo_epi8(__i0, __i1); // ac
const auto __b = _mm_unpackhi_epi8(__i0, __i1); // bd
const auto __c = _mm_unpacklo_epi8(__i2, __i3); // eg
const auto __d = _mm_unpackhi_epi8(__i2, __i3); // fh
const auto __e = _mm_unpacklo_epi8(__i4, __i5); // ik
const auto __f = _mm_unpackhi_epi8(__i4, __i5); // jl
const auto __g = _mm_unpacklo_epi8(__i6, __i7); // mo
const auto __h = _mm_unpackhi_epi8(__i6, __i7); // np
return __intrin_bitcast<_To>(_mm_unpacklo_epi64(
_mm_unpacklo_epi32(_mm_unpacklo_epi8(__a, __b), // abcd
_mm_unpacklo_epi8(__c, __d)), // efgh
_mm_unpacklo_epi32(_mm_unpacklo_epi8(__e, __f), // ijkl
_mm_unpacklo_epi8(__g, __h)) // mnop
));
}
else if constexpr (__y_to_y)
{
auto __a = // 048C GKOS 159D HLPT 26AE IMQU 37BF JNRV
__to_intrin(
(((__v0 & 0xff) | ((__v1 & 0xff) << 8))
| (((__v2 & 0xff) << 16) | ((__v3 & 0xff) << 24)))
| ((((__v4 & 0xff) << 32) | ((__v5 & 0xff) << 40))
| (((__v6 & 0xff) << 48) | ((__v7 << 56)))));
/*
auto __b = _mm256_unpackhi_epi64(__a, __a); // 159D HLPT 159D
HLPT 37BF JNRV 37BF JNRV auto __c = _mm256_unpacklo_epi8(__a,
__b); // 0145 89CD GHKL OPST 2367 ABEF IJMN QRUV auto __d =
__xzyw(__c); // 0145 89CD 2367 ABEF GHKL OPST IJMN QRUV return
_mm256_shuffle_epi8(
__d, _mm256_setr_epi8(0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12,
13, 6, 7, 14, 15, 0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7,
14, 15));
*/
auto __b = _mm256_shuffle_epi8( // 0145 89CD GHKL OPST 2367 ABEF
// IJMN QRUV
__a, _mm256_setr_epi8(0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13,
6, 14, 7, 15, 0, 8, 1, 9, 2, 10, 3, 11,
4, 12, 5, 13, 6, 14, 7, 15));
auto __c
= __xzyw(__b); // 0145 89CD 2367 ABEF GHKL OPST IJMN QRUV
return __intrin_bitcast<_To>(_mm256_shuffle_epi8(
__c, _mm256_setr_epi8(0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13,
6, 7, 14, 15, 0, 1, 8, 9, 2, 3, 10, 11,
4, 5, 12, 13, 6, 7, 14, 15)));
}
else if constexpr (__z_to_z)
{
return __concat(
__convert_x86<__vector_type_t<_Up, _M / 2>>(__v0, __v1, __v2,
__v3),
__convert_x86<__vector_type_t<_Up, _M / 2>>(__v4, __v5, __v6,
__v7));
}
}
else if constexpr (__f64_to_i8) //{{{2
{
return __convert_x86<_To>(
__convert_x86<__vector_type_t<int, _Np * 2>>(__v0, __v1),
__convert_x86<__vector_type_t<int, _Np * 2>>(__v2, __v3),
__convert_x86<__vector_type_t<int, _Np * 2>>(__v4, __v5),
__convert_x86<__vector_type_t<int, _Np * 2>>(__v6, __v7));
}
else // unreachable {{{2
__assert_unreachable<_Tp>();
//}}}
// fallback: {{{2
if constexpr (sizeof(_To) >= 32)
// if _To is ymm or zmm, then _SimdWrapper<_Up, _M / 2> is xmm or ymm
return __concat(
__convert_x86<__vector_type_t<_Up, _M / 2>>(__v0, __v1, __v2, __v3),
__convert_x86<__vector_type_t<_Up, _M / 2>>(__v4, __v5, __v6,
__v7));
else if constexpr (sizeof(_To) == 16)
{
const auto __lo
= __to_intrin(__convert_x86<_To>(__v0, __v1, __v2, __v3));
const auto __hi
= __to_intrin(__convert_x86<_To>(__v4, __v5, __v6, __v7));
static_assert(sizeof(_Up) == 1 && _Np == 2);
return __intrin_bitcast<_To>(_mm_unpacklo_epi64(__lo, __hi));
}
else
{
__assert_unreachable<_Tp>();
// return __vector_convert<_To>(__v0, __v1, __v2, __v3, __v4, __v5,
// __v6, __v7,
// make_index_sequence<_Np>());
} //}}}2
}
}
//}}}
// 16-arg __convert_x86 {{{1
template <typename _To, typename _V, typename _Traits>
_GLIBCXX_SIMD_INTRINSIC _To
__convert_x86(_V __v0, _V __v1, _V __v2, _V __v3, _V __v4, _V __v5, _V __v6,
_V __v7, _V __v8, _V __v9, _V __v10, _V __v11, _V __v12,
_V __v13, _V __v14, _V __v15)
{
// concat => use 8-arg __convert_x86
return __convert_x86<_To>(__concat(__v0, __v1), __concat(__v2, __v3),
__concat(__v4, __v5), __concat(__v6, __v7),
__concat(__v8, __v9), __concat(__v10, __v11),
__concat(__v12, __v13), __concat(__v14, __v15));
}
//}}}
#endif // __cplusplus >= 201703L
#endif // _GLIBCXX_EXPERIMENTAL_SIMD_X86_CONVERSIONS_H
// vim: foldmethod=marker
|