Lines Matching refs:SIMDINLINE
53 static SIMDINLINE Float SIMDCALL op(Float a) \
62 static SIMDINLINE Float SIMDCALL op(Float a, Float b) \
69 static SIMDINLINE Float SIMDCALL op(Float a, Float b) \
76 static SIMDINLINE Double SIMDCALL op(Double a, Double b) \
83 static SIMDINLINE Float SIMDCALL op(Float a, Float b) \
91 static SIMDINLINE Double SIMDCALL op(Double a, Double b) \
98 static SIMDINLINE Float SIMDCALL op(Float a, Float b, Float c) \
104 static SIMDINLINE Integer SIMDCALL op(Integer a) \
109 static SIMDINLINE Integer SIMDCALL op(SIMD256Impl::Integer a) \
115 static SIMDINLINE Integer SIMDCALL op(SIMD128Impl::Integer a) \
122 static SIMDINLINE Integer SIMDCALL op(Integer a) \
129 static SIMDINLINE Integer SIMDCALL op(Integer a, Integer b) \
136 static SIMDINLINE Integer SIMDCALL op(Integer a, Integer b) \
142 static SIMDINLINE Integer SIMDCALL op(Integer a, Integer b) \
149 static SIMDINLINE Integer SIMDCALL op(Integer a, Integer b) \
156 static SIMDINLINE Integer vmask(__mmask16 m)
161 static SIMDINLINE Integer vmask(__mmask8 m)
182 static SIMDINLINE Float SIMDCALL round_ps(Float a)
187 static SIMDINLINE Float SIMDCALL ceil_ps(Float a) { return round_ps<RoundMode::CEIL_NOEXC>(a); }
188 static SIMDINLINE Float SIMDCALL floor_ps(Float a) { return round_ps<RoundMode::FLOOR_NOEXC>(a); }
238 static SIMDINLINE Float SIMDCALL srlisi_ps(Float a)
249 static SIMDINLINE Float SIMDCALL castpd_ps(Double a) // return *(Float*)(&a)
254 static SIMDINLINE Integer SIMDCALL castps_si(Float a) // return *(Integer*)(&a)
259 static SIMDINLINE Double SIMDCALL castsi_pd(Integer a) // return *(Double*)(&a)
264 static SIMDINLINE Double SIMDCALL castps_pd(Float a) // return *(Double*)(&a)
269 static SIMDINLINE Integer SIMDCALL castpd_si(Double a) // return *(Integer*)(&a)
274 static SIMDINLINE Float SIMDCALL castsi_ps(Integer a) // return *(Float*)(&a)
279 static SIMDINLINE Float SIMDCALL cvtepi32_ps(Integer a) // return (float)a (int32 --> float)
290 static SIMDINLINE Integer SIMDCALL cvtps_epi32(Float a) // return (int32)a (float --> int32)
295 static SIMDINLINE Integer SIMDCALL cvttps_epi32(Float a) // return (int32)a (rnd_to_zero(float) --> int32)
304 static SIMDINLINE Mask SIMDCALL cmp_ps_mask(Float a, Float b)
310 static SIMDINLINE Float SIMDCALL cmp_ps(Float a, Float b) // return a (CmpTypeT) b
317 static SIMDINLINE Float SIMDCALL cmplt_ps(Float a, Float b) { return cmp_ps<CompareType::LT_OQ>(a, b); }
318 static SIMDINLINE Float SIMDCALL cmpgt_ps(Float a, Float b) { return cmp_ps<CompareType::GT_OQ>(a, b); }
319 static SIMDINLINE Float SIMDCALL cmpneq_ps(Float a, Float b) { return cmp_ps<CompareType::NEQ_OQ>(a, b); }
320 static SIMDINLINE Float SIMDCALL cmpeq_ps(Float a, Float b) { return cmp_ps<CompareType::EQ_OQ>(a, b); }
321 static SIMDINLINE Float SIMDCALL cmpge_ps(Float a, Float b) { return cmp_ps<CompareType::GE_OQ>(a, b); }
322 static SIMDINLINE Float SIMDCALL cmple_ps(Float a, Float b) { return cmp_ps<CompareType::LE_OQ>(a, b); }
325 static SIMDINLINE Integer SIMDCALL cmp_epi32(Integer a, Integer b)
332 static SIMDINLINE Integer SIMDCALL cmp_epi64(Integer a, Integer b)
349 static SIMDINLINE bool SIMDCALL testz_ps(Float a, Float b) // return all_lanes_zero(a & b) ? 1 : 0 (float)
354 static SIMDINLINE bool SIMDCALL testz_si(Integer a, Integer b) // return all_lanes_zero(a & b) ? 1 : 0 (int)
363 static SIMDINLINE Float blend_ps(Float a, Float b) // return ImmT ? b : a (float)
369 static SIMDINLINE Integer blend_epi32(Integer a, Integer b) // return ImmT ? b : a (int32)
374 static SIMDINLINE Float blendv_ps(Float a, Float b, Float mask) // return mask ? b : a (float)
380 static SIMDINLINE Integer SIMDCALL blendv_epi32(Integer a, Integer b, Float mask) // return mask ? b : a (int)
385 static SIMDINLINE Integer SIMDCALL blendv_epi32(Integer a, Integer b, Integer mask) // return mask ? b : a (int)
390 static SIMDINLINE Float SIMDCALL broadcast_ss(float const *p) // return *p (all elements in vector get same value)
396 static SIMDINLINE SIMD256Impl::Float SIMDCALL extract_ps(Float a)
402 static SIMDINLINE SIMD256Impl::Double SIMDCALL extract_pd(Double a)
408 static SIMDINLINE SIMD256Impl::Integer SIMDCALL extract_si(Integer a)
414 static SIMDINLINE Float SIMDCALL insert_ps(Float a, SIMD256Impl::Float b)
420 static SIMDINLINE Double SIMDCALL insert_pd(Double a, SIMD256Impl::Double b)
426 static SIMDINLINE Integer SIMDCALL insert_si(Integer a, SIMD256Impl::Integer b)
436 static SIMDINLINE Integer SIMDCALL permute_epi32(Integer a, Integer swiz) // return a[swiz[i]] for each 32-bit lane i (float)
441 static SIMDINLINE Float SIMDCALL permute_ps(Float a, Integer swiz) // return a[swiz[i]] for each 32-bit lane i (float)
457 static SIMDINLINE Integer SIMDCALL shuffle_epi64(Integer a, Integer b)
465 static SIMDINLINE Integer SIMDCALL unpackhi_epi32(Integer a, Integer b)
485 static SIMDINLINE Float SIMDCALL i32gather_ps(float const* p, Integer idx) // return *(float*)(((int8*)p) + (idx * ScaleT))
490 static SIMDINLINE Float SIMDCALL load1_ps(float const *p) // return *p (broadcast 1 value to all elements)
495 static SIMDINLINE Float SIMDCALL load_ps(float const *p) // return *p (loads SIMD width elements from memory)
500 static SIMDINLINE Integer SIMDCALL load_si(Integer const *p) // return *p
505 static SIMDINLINE Float SIMDCALL loadu_ps(float const *p) // return *p (same as load_ps but allows for unaligned mem)
510 static SIMDINLINE Integer SIMDCALL loadu_si(Integer const *p) // return *p (same as load_si but allows for unaligned mem)
517 static SIMDINLINE Float SIMDCALL mask_i32gather_ps(Float old, float const* p, Integer idx, Float mask)
524 static SIMDINLINE void SIMDCALL maskstore_ps(float *p, Integer mask, Float src)
530 //static SIMDINLINE uint64_t SIMDCALL movemask_epi8(Integer a)
536 static SIMDINLINE uint32_t SIMDCALL movemask_pd(Double a)
541 static SIMDINLINE uint32_t SIMDCALL movemask_ps(Float a)
547 static SIMDINLINE Integer SIMDCALL set1_epi64(long long i) // return i (all elements are same value)
552 static SIMDINLINE Integer SIMDCALL set1_epi32(int i) // return i (all elements are same value)
557 static SIMDINLINE Integer SIMDCALL set1_epi8(char i) // return i (all elements are same value)
562 static SIMDINLINE Float SIMDCALL set1_ps(float f) // return f (all elements are same value)
567 static SIMDINLINE Double SIMDCALL setzero_pd() // return 0 (double)
572 static SIMDINLINE Float SIMDCALL setzero_ps() // return 0 (float)
577 static SIMDINLINE Integer SIMDCALL setzero_si() // return 0 (integer)
582 static SIMDINLINE void SIMDCALL store_ps(float *p, Float a) // *p = a (stores all elements contiguously in memory)
587 static SIMDINLINE void SIMDCALL store_si(Integer *p, Integer a) // *p = a
592 static SIMDINLINE void SIMDCALL storeu_si(Integer *p, Integer a) // *p = a (same as store_si but allows for unaligned mem)
597 static SIMDINLINE void SIMDCALL stream_ps(float *p, Float a) // *p = a (same as store_ps, but doesn't keep memory in cache)
602 static SIMDINLINE Integer SIMDCALL set_epi32(
611 static SIMDINLINE Integer SIMDCALL set_epi32(
619 static SIMDINLINE Float SIMDCALL set_ps(
628 static SIMDINLINE Float SIMDCALL set_ps(
636 static SIMDINLINE Float SIMDCALL vmask_ps(int32_t mask)
