Lines Matching refs:SIMDINLINE
35 static SIMDINLINE Float SIMDCALL op(Float const &a) \
45 static SIMDINLINE Float SIMDCALL op(Float const &a, Float const &b) \
56 static SIMDINLINE Float SIMDCALL op(Float const &a, Float const &b) \
67 static SIMDINLINE Float SIMDCALL op(Float const &a, Float const &b) \
77 static SIMDINLINE Float SIMDCALL op(Float const &a, Float const &b, Float const &c) \
87 static SIMDINLINE Integer SIMDCALL op(Integer const &a) \
97 static SIMDINLINE Integer SIMDCALL op(Integer const &a, Integer const &b) \
108 static SIMDINLINE Integer SIMDCALL op(Integer const &a, Integer const &b) \
119 static SIMDINLINE Integer SIMDCALL op(Integer const &a, Integer const &b) \
130 static SIMDINLINE Integer SIMDCALL op(Integer const &a, Integer const &b) \
140 static SIMDINLINE Integer SIMDCALL op(Integer const &a, Integer const &b, Integer const &c) \
164 static SIMDINLINE Float SIMDCALL round_ps(Float const &a)
173 static SIMDINLINE Float SIMDCALL ceil_ps(Float const &a) { return round_ps<RoundMode::CEIL_NOEXC>(a); }
174 static SIMDINLINE Float SIMDCALL floor_ps(Float const &a) { return round_ps<RoundMode::FLOOR_NOEXC>(a); }
215 static SIMDINLINE Integer SIMDCALL slli_epi32(Integer const &a) // return a << ImmT
227 static SIMDINLINE Integer SIMDCALL srai_epi32(Integer const &a) // return a >> ImmT (int32)
237 static SIMDINLINE Integer SIMDCALL srli_epi32(Integer const &a) // return a >> ImmT (uint32)
247 static SIMDINLINE Integer SIMDCALL srli_si(Integer const &a) // return a >> (ImmT*8) (uint)
256 static SIMDINLINE Float SIMDCALL srlisi_ps(Float const &a) // same as srli_si, but with Float cast to int
270 static SIMDINLINE Float SIMDCALL castpd_ps(Double const &a) // return *(Float*)(&a)
279 static SIMDINLINE Integer SIMDCALL castps_si(Float const &a) // return *(Integer*)(&a)
288 static SIMDINLINE Double SIMDCALL castsi_pd(Integer const &a) // return *(Double*)(&a)
297 static SIMDINLINE Double SIMDCALL castps_pd(Float const &a) // return *(Double*)(&a)
306 static SIMDINLINE Float SIMDCALL castsi_ps(Integer const &a) // return *(Float*)(&a)
315 static SIMDINLINE Float SIMDCALL cvtepi32_ps(Integer const &a) // return (float)a (int32 --> float)
324 static SIMDINLINE Integer SIMDCALL cvtepu8_epi16(SIMD256Impl::Integer const &a) // return (int16)a (uint8 --> int16)
333 static SIMDINLINE Integer SIMDCALL cvtepu8_epi32(SIMD256Impl::Integer const &a) // return (int32)a (uint8 --> int32)
342 static SIMDINLINE Integer SIMDCALL cvtepu16_epi32(SIMD256Impl::Integer const &a) // return (int32)a (uint16 --> int32)
351 static SIMDINLINE Integer SIMDCALL cvtepu16_epi64(SIMD256Impl::Integer const &a) // return (int64)a (uint16 --> int64)
360 static SIMDINLINE Integer SIMDCALL cvtepu32_epi64(SIMD256Impl::Integer const &a) // return (int64)a (uint32 --> int64)
369 static SIMDINLINE Integer SIMDCALL cvtps_epi32(Float const &a) // return (int32)a (float --> int32)
378 static SIMDINLINE Integer SIMDCALL cvttps_epi32(Float const &a) // return (int32)a (rnd_to_zero(float) --> int32)
391 static SIMDINLINE Float SIMDCALL cmp_ps(Float const &a, Float const &b) // return a (CmpTypeT) b
399 static SIMDINLINE Float SIMDCALL cmplt_ps(Float const &a, Float const &b) { return cmp_ps<CompareType::LT_OQ>(a, b); }
400 static SIMDINLINE Float SIMDCALL cmpgt_ps(Float const &a, Float const &b) { return cmp_ps<CompareType::GT_OQ>(a, b); }
401 static SIMDINLINE Float SIMDCALL cmpneq_ps(Float const &a, Float const &b) { return cmp_ps<CompareType::NEQ_OQ>(a, b); }
402 static SIMDINLINE Float SIMDCALL cmpeq_ps(Float const &a, Float const &b) { return cmp_ps<CompareType::EQ_OQ>(a, b); }
403 static SIMDINLINE Float SIMDCALL cmpge_ps(Float const &a, Float const &b) { return cmp_ps<CompareType::GE_OQ>(a, b); }
404 static SIMDINLINE Float SIMDCALL cmple_ps(Float const &a, Float const &b) { return cmp_ps<CompareType::LE_OQ>(a, b); }
407 static SIMDINLINE Mask SIMDCALL cmp_ps_mask(Float const &a, Float const &b)
423 static SIMDINLINE bool SIMDCALL testz_ps(Float const &a, Float const &b) // return all_lanes_zero(a & b) ? 1 : 0 (float)
429 static SIMDINLINE int SIMDCALL testz_si(Integer const &a, Integer const &b) // return all_lanes_zero(a & b) ? 1 : 0 (int)
441 static SIMDINLINE Integer SIMDCALL blendv_epi32(Integer const &a, Integer const &b, Float const &mask) // return mask ? b : a (int)
450 static SIMDINLINE Integer SIMDCALL blendv_epi32(Integer const &a, Integer const &b, Integer const &mask) // return mask ? b : a (int)
459 static SIMDINLINE Float SIMDCALL broadcast_ss(float const *p) // return *p (all elements in vector get same value)
470 static SIMDINLINE SIMD256Impl::Float SIMDCALL extract_ps(Float const &a)
477 static SIMDINLINE SIMD256Impl::Double SIMDCALL extract_pd(Double const &a)
484 static SIMDINLINE SIMD256Impl::Integer SIMDCALL extract_si(Integer const &a)
491 static SIMDINLINE Float SIMDCALL insert_ps(Float const &a, SIMD256Impl::Float const &b)
500 static SIMDINLINE Double SIMDCALL insert_pd(Double const &a, SIMD256Impl::Double const &b)
509 static SIMDINLINE Integer SIMDCALL insert_si(Integer const &a, SIMD256Impl::Integer const &b)
522 static SIMDINLINE Integer SIMDCALL permute_epi32(Integer const &a, Integer const &swiz) // return a[swiz[i]] for each 32-bit lane i (int32)
527 static SIMDINLINE Float SIMDCALL permute_ps(Float const &a, Integer const &swiz) // return a[swiz[i]] for each 32-bit lane i (float)
567 static SIMDINLINE Float SIMDCALL permute2f128_ps(Float const &a, Float const &b)
577 static SIMDINLINE Double SIMDCALL permute2f128_pd(Double const &a, Double const &b)
587 static SIMDINLINE Integer SIMDCALL permute2f128_si(Integer const &a, Integer const &b)
618 static SIMDINLINE Float SIMDCALL i32gather_ps(float const* p, Integer const &idx) // return *(float*)(((int8*)p) + (idx * ScaleT))
627 static SIMDINLINE Float SIMDCALL load1_ps(float const *p) // return *p (broadcast 1 value to all elements)
632 static SIMDINLINE Float SIMDCALL load_ps(float const *p) // return *p (loads SIMD width elements from memory)
641 static SIMDINLINE Integer SIMDCALL load_si(Integer const *p) // return *p
650 static SIMDINLINE Float SIMDCALL loadu_ps(float const *p) // return *p (same as load_ps but allows for unaligned mem)
659 static SIMDINLINE Integer SIMDCALL loadu_si(Integer const *p) // return *p (same as load_si but allows for unaligned mem)
670 static SIMDINLINE Float SIMDCALL mask_i32gather_ps(Float const &old, float const* p, Integer const &idx, Float const &mask)
679 static SIMDINLINE void SIMDCALL maskstore_ps(float *p, Integer const &mask, Float const &src)
685 static SIMDINLINE uint64_t SIMDCALL movemask_epi8(Integer const &a)
693 static SIMDINLINE uint32_t SIMDCALL movemask_pd(Double const &a)
700 static SIMDINLINE uint32_t SIMDCALL movemask_ps(Float const &a)
708 static SIMDINLINE Integer SIMDCALL set1_epi32(int i) // return i (all elements are same value)
717 static SIMDINLINE Integer SIMDCALL set1_epi8(char i) // return i (all elements are same value)
726 static SIMDINLINE Float SIMDCALL set1_ps(float f) // return f (all elements are same value)
735 static SIMDINLINE Float SIMDCALL setzero_ps() // return 0 (float)
744 static SIMDINLINE Integer SIMDCALL setzero_si() // return 0 (integer)
753 static SIMDINLINE void SIMDCALL store_ps(float *p, Float const &a) // *p = a (stores all elements contiguously in memory)
759 static SIMDINLINE void SIMDCALL store_si(Integer *p, Integer const &a) // *p = a
765 static SIMDINLINE void SIMDCALL stream_ps(float *p, Float const &a) // *p = a (same as store_ps, but doesn't keep memory in cache)
771 static SIMDINLINE Integer SIMDCALL set_epi32(
784 static SIMDINLINE Integer SIMDCALL set_epi32(
792 static SIMDINLINE Float SIMDCALL set_ps(
805 static SIMDINLINE Float SIMDCALL set_ps(
813 static SIMDINLINE Float SIMDCALL vmask_ps(int32_t mask)
