25 #ifndef _GLIBCXX_EXPERIMENTAL_SIMD_NEON_H_
26 #define _GLIBCXX_EXPERIMENTAL_SIMD_NEON_H_
28 #if __cplusplus >= 201703L
30 #if !_GLIBCXX_SIMD_HAVE_NEON
31 #error "simd_neon.h may only be included when NEON on ARM is available"
34 _GLIBCXX_SIMD_BEGIN_NAMESPACE
37 struct _CommonImplNeon : _CommonImplBuiltin
40 using _CommonImplBuiltin::_S_store;
47 template <
typename _Abi,
typename>
48 struct _SimdImplNeon : _SimdImplBuiltin<_Abi>
50 using _Base = _SimdImplBuiltin<_Abi>;
52 template <
typename _Tp>
53 using _MaskMember =
typename _Base::template _MaskMember<_Tp>;
55 template <
typename _Tp>
56 static constexpr
size_t _S_max_store_size = 16;
59 template <
typename _Tp,
size_t _Np,
typename _Up>
60 static inline _SimdWrapper<_Tp, _Np>
61 _S_masked_load(_SimdWrapper<_Tp, _Np> __merge, _MaskMember<_Tp> __k,
62 const _Up* __mem) noexcept
64 __execute_n_times<_Np>([&](
auto __i) {
66 __merge._M_set(__i,
static_cast<_Tp
>(__mem[__i]));
73 template <
typename _Tp,
size_t _Np>
74 _GLIBCXX_SIMD_INTRINSIC
static void
75 _S_masked_store_nocvt(_SimdWrapper<_Tp, _Np> __v, _Tp* __mem,
78 __execute_n_times<_Np>([&](
auto __i) {
80 __mem[__i] = __v[__i];
86 template <
typename _Tp,
typename _BinaryOperation>
87 _GLIBCXX_SIMD_INTRINSIC
static _Tp
88 _S_reduce(simd<_Tp, _Abi> __x, _BinaryOperation&& __binary_op)
90 constexpr
size_t _Np = __x.size();
91 if constexpr (
sizeof(__x) == 16 && _Np >= 4
92 && !_Abi::template _S_is_partial<_Tp>)
94 const auto __halves = split<simd<_Tp, simd_abi::_Neon<8>>>(__x);
95 const auto __y = __binary_op(__halves[0], __halves[1]);
96 return _SimdImplNeon<simd_abi::_Neon<8>>::_S_reduce(
97 __y,
static_cast<_BinaryOperation&&
>(__binary_op));
99 else if constexpr (_Np == 8)
101 __x = __binary_op(__x, _Base::template _M_make_simd<_Tp, _Np>(
102 __vector_permute<1, 0, 3, 2, 5, 4, 7, 6>(
104 __x = __binary_op(__x, _Base::template _M_make_simd<_Tp, _Np>(
105 __vector_permute<3, 2, 1, 0, 7, 6, 5, 4>(
107 __x = __binary_op(__x, _Base::template _M_make_simd<_Tp, _Np>(
108 __vector_permute<7, 6, 5, 4, 3, 2, 1, 0>(
112 else if constexpr (_Np == 4)
115 = __binary_op(__x, _Base::template _M_make_simd<_Tp, _Np>(
116 __vector_permute<1, 0, 3, 2>(__x._M_data)));
118 = __binary_op(__x, _Base::template _M_make_simd<_Tp, _Np>(
119 __vector_permute<3, 2, 1, 0>(__x._M_data)));
122 else if constexpr (_Np == 2)
124 __x = __binary_op(__x, _Base::template _M_make_simd<_Tp, _Np>(
125 __vector_permute<1, 0>(__x._M_data)));
129 return _Base::_S_reduce(__x,
130 static_cast<_BinaryOperation&&
>(__binary_op));
136 template <
typename _Tp,
typename _TVT = _VectorTraits<_Tp>>
137 _GLIBCXX_SIMD_INTRINSIC
static _Tp _S_sqrt(_Tp __x)
139 if constexpr (__have_neon_a64)
141 const auto __intrin = __to_intrin(__x);
142 if constexpr (_TVT::template _S_is<float, 2>)
143 return vsqrt_f32(__intrin);
144 else if constexpr (_TVT::template _S_is<float, 4>)
145 return vsqrtq_f32(__intrin);
146 else if constexpr (_TVT::template _S_is<double, 1>)
147 return vsqrt_f64(__intrin);
148 else if constexpr (_TVT::template _S_is<double, 2>)
149 return vsqrtq_f64(__intrin);
151 __assert_unreachable<_Tp>();
154 return _Base::_S_sqrt(__x);
159 template <
typename _TW,
typename _TVT = _VectorTraits<_TW>>
160 _GLIBCXX_SIMD_INTRINSIC
static _TW _S_trunc(_TW __x)
162 using _Tp =
typename _TVT::value_type;
163 if constexpr (__have_neon_a32)
165 const auto __intrin = __to_intrin(__x);
166 if constexpr (_TVT::template _S_is<float, 2>)
167 return vrnd_f32(__intrin);
168 else if constexpr (_TVT::template _S_is<float, 4>)
169 return vrndq_f32(__intrin);
170 else if constexpr (_TVT::template _S_is<double, 1>)
171 return vrnd_f64(__intrin);
172 else if constexpr (_TVT::template _S_is<double, 2>)
173 return vrndq_f64(__intrin);
175 __assert_unreachable<_Tp>();
177 else if constexpr (is_same_v<_Tp, float>)
179 auto __intrin = __to_intrin(__x);
180 if constexpr (
sizeof(__x) == 16)
181 __intrin = vcvtq_f32_s32(vcvtq_s32_f32(__intrin));
183 __intrin = vcvt_f32_s32(vcvt_s32_f32(__intrin));
184 return _Base::_S_abs(__x)._M_data < 0x1p23f
185 ? __vector_bitcast<
float>(__intrin)
189 return _Base::_S_trunc(__x);
194 template <typename _Tp,
size_t _Np>
195 _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
196 _S_round(_SimdWrapper<_Tp, _Np> __x)
198 if constexpr (__have_neon_a32)
200 const auto __intrin = __to_intrin(__x);
201 if constexpr (
sizeof(_Tp) == 4 &&
sizeof(__x) == 8)
202 return vrnda_f32(__intrin);
203 else if constexpr (sizeof(_Tp) == 4 && sizeof(__x) == 16)
204 return vrndaq_f32(__intrin);
205 else if constexpr (sizeof(_Tp) == 8 && sizeof(__x) == 8)
206 return vrnda_f64(__intrin);
207 else if constexpr (sizeof(_Tp) == 8 && sizeof(__x) == 16)
208 return vrndaq_f64(__intrin);
210 __assert_unreachable<_Tp>();
213 return _Base::_S_round(__x);
218 template <typename _Tp, typename _TVT = _VectorTraits<_Tp>>
219 _GLIBCXX_SIMD_INTRINSIC static _Tp _S_floor(_Tp __x)
221 if constexpr (__have_neon_a32)
223 const auto __intrin = __to_intrin(__x);
224 if constexpr (_TVT::template _S_is<float, 2>)
225 return vrndm_f32(__intrin);
226 else if constexpr (_TVT::template _S_is<float, 4>)
227 return vrndmq_f32(__intrin);
228 else if constexpr (_TVT::template _S_is<double, 1>)
229 return vrndm_f64(__intrin);
230 else if constexpr (_TVT::template _S_is<double, 2>)
231 return vrndmq_f64(__intrin);
233 __assert_unreachable<_Tp>();
236 return _Base::_S_floor(__x);
241 template <
typename _Tp,
typename _TVT = _VectorTraits<_Tp>>
242 _GLIBCXX_SIMD_INTRINSIC
static _Tp _S_ceil(_Tp __x)
244 if constexpr (__have_neon_a32)
246 const auto __intrin = __to_intrin(__x);
247 if constexpr (_TVT::template _S_is<float, 2>)
248 return vrndp_f32(__intrin);
249 else if constexpr (_TVT::template _S_is<float, 4>)
250 return vrndpq_f32(__intrin);
251 else if constexpr (_TVT::template _S_is<double, 1>)
252 return vrndp_f64(__intrin);
253 else if constexpr (_TVT::template _S_is<double, 2>)
254 return vrndpq_f64(__intrin);
256 __assert_unreachable<_Tp>();
259 return _Base::_S_ceil(__x);
265 struct _MaskImplNeonMixin
267 using _Base = _MaskImplBuiltinMixin;
269 template <
typename _Tp,
size_t _Np>
270 _GLIBCXX_SIMD_INTRINSIC
static constexpr _SanitizedBitMask<_Np>
271 _S_to_bits(_SimdWrapper<_Tp, _Np> __x)
273 if (__builtin_is_constant_evaluated())
274 return _Base::_S_to_bits(__x);
276 using _I = __int_for_sizeof_t<_Tp>;
277 if constexpr (
sizeof(__x) == 16)
279 auto __asint = __vector_bitcast<_I>(__x);
281 [[maybe_unused]] constexpr
auto __zero = decltype(__asint)();
283 [[maybe_unused]] constexpr
auto __zero = decltype(__lo64(__asint))();
285 if constexpr (
sizeof(_Tp) == 1)
287 constexpr
auto __bitsel
288 = __generate_from_n_evaluations<16, __vector_type_t<_I, 16>>(
290 return static_cast<_I
>(
291 __i < _Np ? (__i < 8 ? 1 << __i : 1 << (__i - 8)) : 0);
295 return __vector_bitcast<_UShort>(
296 vpaddq_s8(vpaddq_s8(vpaddq_s8(__asint, __zero), __zero),
299 return __vector_bitcast<_UShort>(
300 vpadd_s8(vpadd_s8(vpadd_s8(__lo64(__asint), __hi64(__asint)),
305 else if constexpr (
sizeof(_Tp) == 2)
307 constexpr
auto __bitsel
308 = __generate_from_n_evaluations<8, __vector_type_t<_I, 8>>(
310 return static_cast<_I
>(__i < _Np ? 1 << __i : 0);
314 return vaddvq_s16(__asint);
317 vpadd_s16(vpadd_s16(__lo64(__asint), __hi64(__asint)), __zero),
321 else if constexpr (
sizeof(_Tp) == 4)
323 constexpr
auto __bitsel
324 = __generate_from_n_evaluations<4, __vector_type_t<_I, 4>>(
326 return static_cast<_I
>(__i < _Np ? 1 << __i : 0);
330 return vaddvq_s32(__asint);
332 return vpadd_s32(vpadd_s32(__lo64(__asint), __hi64(__asint)),
336 else if constexpr (
sizeof(_Tp) == 8)
337 return (__asint[0] & 1) | (__asint[1] & 2);
339 __assert_unreachable<_Tp>();
341 else if constexpr (sizeof(__x) == 8)
343 auto __asint = __vector_bitcast<_I>(__x);
344 [[maybe_unused]] constexpr
auto __zero = decltype(__asint)();
345 if constexpr (
sizeof(_Tp) == 1)
347 constexpr
auto __bitsel
348 = __generate_from_n_evaluations<8, __vector_type_t<_I, 8>>(
350 return static_cast<_I
>(__i < _Np ? 1 << __i : 0);
354 return vaddv_s8(__asint);
356 return vpadd_s8(vpadd_s8(vpadd_s8(__asint, __zero), __zero),
360 else if constexpr (
sizeof(_Tp) == 2)
362 constexpr
auto __bitsel
363 = __generate_from_n_evaluations<4, __vector_type_t<_I, 4>>(
365 return static_cast<_I
>(__i < _Np ? 1 << __i : 0);
369 return vaddv_s16(__asint);
371 return vpadd_s16(vpadd_s16(__asint, __zero), __zero)[0];
374 else if constexpr (
sizeof(_Tp) == 4)
376 __asint &= __make_vector<_I>(0x1, 0x2);
378 return vaddv_s32(__asint);
380 return vpadd_s32(__asint, __zero)[0];
384 __assert_unreachable<_Tp>();
387 return _Base::_S_to_bits(__x);
393 template <
typename _Abi,
typename>
394 struct _MaskImplNeon : _MaskImplNeonMixin, _MaskImplBuiltin<_Abi>
396 using _MaskImplBuiltinMixin::_S_to_maskvector;
397 using _MaskImplNeonMixin::_S_to_bits;
398 using _Base = _MaskImplBuiltin<_Abi>;
399 using _Base::_S_convert;
402 template <
typename _Tp>
403 _GLIBCXX_SIMD_INTRINSIC
static bool _S_all_of(simd_mask<_Tp, _Abi> __k)
406 = __vector_bitcast<char>(__k._M_data)
407 | ~__vector_bitcast<char>(_Abi::template _S_implicit_mask<_Tp>());
408 if constexpr (
sizeof(__k) == 16)
410 const auto __x = __vector_bitcast<long long>(__kk);
411 return __x[0] + __x[1] == -2;
413 else if constexpr (
sizeof(__k) <= 8)
414 return __bit_cast<__int_for_sizeof_t<decltype(__kk)>>(__kk) == -1;
416 __assert_unreachable<_Tp>();
421 template <typename _Tp>
422 _GLIBCXX_SIMD_INTRINSIC static
bool _S_any_of(simd_mask<_Tp, _Abi> __k)
425 = __vector_bitcast<char>(__k._M_data)
426 | ~__vector_bitcast<char>(_Abi::template _S_implicit_mask<_Tp>());
427 if constexpr (
sizeof(__k) == 16)
429 const auto __x = __vector_bitcast<long long>(__kk);
430 return (__x[0] | __x[1]) != 0;
432 else if constexpr (
sizeof(__k) <= 8)
433 return __bit_cast<__int_for_sizeof_t<decltype(__kk)>>(__kk) != 0;
435 __assert_unreachable<_Tp>();
440 template <typename _Tp>
441 _GLIBCXX_SIMD_INTRINSIC static
bool _S_none_of(simd_mask<_Tp, _Abi> __k)
443 const auto __kk = _Abi::_S_masked(__k._M_data);
444 if constexpr (
sizeof(__k) == 16)
446 const auto __x = __vector_bitcast<long long>(__kk);
447 return (__x[0] | __x[1]) == 0;
449 else if constexpr (
sizeof(__k) <= 8)
450 return __bit_cast<__int_for_sizeof_t<decltype(__kk)>>(__kk) == 0;
452 __assert_unreachable<_Tp>();
457 template <typename _Tp>
458 _GLIBCXX_SIMD_INTRINSIC static
bool _S_some_of(simd_mask<_Tp, _Abi> __k)
460 if constexpr (
sizeof(__k) <= 8)
462 const auto __kk = __vector_bitcast<char>(__k._M_data)
463 | ~__vector_bitcast<char>(
464 _Abi::template _S_implicit_mask<_Tp>());
466 return __bit_cast<_Up>(__kk) + 1 > 1;
469 return _Base::_S_some_of(__k);
474 template <
typename _Tp>
475 _GLIBCXX_SIMD_INTRINSIC
static int _S_popcount(simd_mask<_Tp, _Abi> __k)
477 if constexpr (
sizeof(_Tp) == 1)
479 const auto __s8 = __vector_bitcast<_SChar>(__k._M_data);
480 int8x8_t __tmp = __lo64(__s8) + __hi64z(__s8);
481 return -vpadd_s8(vpadd_s8(vpadd_s8(__tmp, int8x8_t()), int8x8_t()),
484 else if constexpr (
sizeof(_Tp) == 2)
486 const auto __s16 = __vector_bitcast<short>(__k._M_data);
487 int16x4_t __tmp = __lo64(__s16) + __hi64z(__s16);
488 return -vpadd_s16(vpadd_s16(__tmp, int16x4_t()), int16x4_t())[0];
490 else if constexpr (
sizeof(_Tp) == 4)
492 const auto __s32 = __vector_bitcast<int>(__k._M_data);
493 int32x2_t __tmp = __lo64(__s32) + __hi64z(__s32);
494 return -vpadd_s32(__tmp, int32x2_t())[0];
496 else if constexpr (
sizeof(_Tp) == 8)
498 static_assert(
sizeof(__k) == 16);
499 const auto __s64 = __vector_bitcast<long>(__k._M_data);
500 return -(__s64[0] + __s64[1]);
506 template <
typename _Tp>
507 _GLIBCXX_SIMD_INTRINSIC
static int
508 _S_find_first_set(simd_mask<_Tp, _Abi> __k)
511 return _Base::_S_find_first_set(__k);
516 template <
typename _Tp>
517 _GLIBCXX_SIMD_INTRINSIC
static int
518 _S_find_last_set(simd_mask<_Tp, _Abi> __k)
521 return _Base::_S_find_last_set(__k);
527 _GLIBCXX_SIMD_END_NAMESPACE
typename make_unsigned< _Tp >::type make_unsigned_t
Alias template for make_unsigned.