Lines Matching refs:__A
31 /// \param __A
36 /// conversion of __B, and higher 64 bits come from conversion of __A.
38 _mm_cvtne2ps_pbh(__m128 __A, __m128 __B) {
39 return (__m128bh)__builtin_ia32_cvtne2ps2bf16_128((__v4sf) __A,
49 /// \param __A
57 /// A 1 means conversion of __A or __B. A 0 means element from __W.
59 /// conversion of __B, and higher 64 bits come from conversion of __A.
61 _mm_mask_cvtne2ps_pbh(__m128bh __W, __mmask8 __U, __m128 __A, __m128 __B) {
63 (__v8bf)_mm_cvtne2ps_pbh(__A, __B),
73 /// \param __A
79 /// A 1 means conversion of __A or __B. A 0 means element is zero.
81 /// conversion of __B, and higher 64 bits come from conversion of __A.
83 _mm_maskz_cvtne2ps_pbh(__mmask8 __U, __m128 __A, __m128 __B) {
85 (__v8bf)_mm_cvtne2ps_pbh(__A, __B),
95 /// \param __A
100 /// conversion of __B, and higher 128 bits come from conversion of __A.
102 _mm256_cvtne2ps_pbh(__m256 __A, __m256 __B) {
103 return (__m256bh)__builtin_ia32_cvtne2ps2bf16_256((__v8sf) __A,
113 /// \param __A
121 /// A 1 means conversion of __A or __B. A 0 means element from __W.
123 /// conversion of __B, and higher 128 bits come from conversion of __A.
125 _mm256_mask_cvtne2ps_pbh(__m256bh __W, __mmask16 __U, __m256 __A, __m256 __B) {
127 (__v16bf)_mm256_cvtne2ps_pbh(__A, __B),
137 /// \param __A
143 /// A 1 means conversion of __A or __B. A 0 means element is zero.
145 /// conversion of __B, and higher 128 bits come from conversion of __A.
147 _mm256_maskz_cvtne2ps_pbh(__mmask16 __U, __m256 __A, __m256 __B) {
149 (__v16bf)_mm256_cvtne2ps_pbh(__A, __B),
159 /// \param __A
162 /// conversion of __A, and higher 64 bits are 0.
172 /// \param __A
178 /// A 1 means conversion of __A. A 0 means element from __W.
180 /// conversion of __A, and higher 64 bits are 0.
182 _mm_mask_cvtneps_pbh(__m128bh __W, __mmask8 __U, __m128 __A) {
183 return (__m128bh)__builtin_ia32_cvtneps2bf16_128_mask((__v4sf) __A,
194 /// \param __A
198 /// A 1 means conversion of __A. A 0 means element is zero.
200 /// conversion of __A, and higher 64 bits are 0.
202 _mm_maskz_cvtneps_pbh(__mmask8 __U, __m128 __A) {
203 return (__m128bh)__builtin_ia32_cvtneps2bf16_128_mask((__v4sf) __A,
214 /// \param __A
216 /// \returns A 128-bit vector of [8 x bfloat] comes from conversion of __A.
226 /// \param __A
232 /// A 1 means conversion of __A. A 0 means element from __W.
233 /// \returns A 128-bit vector of [8 x bfloat] comes from conversion of __A.
235 _mm256_mask_cvtneps_pbh(__m128bh __W, __mmask8 __U, __m256 __A) {
236 return (__m128bh)__builtin_ia32_cvtneps2bf16_256_mask((__v8sf)__A,
247 /// \param __A
251 /// A 1 means conversion of __A. A 0 means element is zero.
252 /// \returns A 128-bit vector of [8 x bfloat] comes from conversion of __A.
254 _mm256_maskz_cvtneps_pbh(__mmask8 __U, __m256 __A) {
255 return (__m128bh)__builtin_ia32_cvtneps2bf16_256_mask((__v8sf)__A,
266 /// \param __A
273 /// __A, __B and __D
275 _mm_dpbf16_ps(__m128 __D, __m128bh __A, __m128bh __B) {
277 (__v8bf)__A,
287 /// \param __A
295 /// A 1 means __A and __B's dot product accumulated with __D. A 0 means __D.
297 /// __A, __B and __D
299 _mm_mask_dpbf16_ps(__m128 __D, __mmask8 __U, __m128bh __A, __m128bh __B) {
301 (__v4sf)_mm_dpbf16_ps(__D, __A, __B),
311 /// \param __A
319 /// A 1 means __A and __B's dot product accumulated with __D. A 0 means 0.
321 /// __A, __B and __D
323 _mm_maskz_dpbf16_ps(__mmask8 __U, __m128 __D, __m128bh __A, __m128bh __B) {
325 (__v4sf)_mm_dpbf16_ps(__D, __A, __B),
335 /// \param __A
342 /// __A, __B and __D
344 _mm256_dpbf16_ps(__m256 __D, __m256bh __A, __m256bh __B) {
346 (__v16bf)__A,
356 /// \param __A
364 /// A 1 means __A and __B's dot product accumulated with __D. A 0 means __D.
366 /// __A, __B and __D
368 _mm256_mask_dpbf16_ps(__m256 __D, __mmask8 __U, __m256bh __A, __m256bh __B) {
370 (__v8sf)_mm256_dpbf16_ps(__D, __A, __B),
380 /// \param __A
388 /// A 1 means __A and __B's dot product accumulated with __D. A 0 means 0.
390 /// __A, __B and __D
392 _mm256_maskz_dpbf16_ps(__mmask8 __U, __m256 __D, __m256bh __A, __m256bh __B) {
394 (__v8sf)_mm256_dpbf16_ps(__D, __A, __B),
404 /// \param __A
408 static __inline__ __bf16 __DEFAULT_FN_ATTRS128 _mm_cvtness_sbh(float __A) {
409 __v4sf __V = {__A, 0, 0, 0};
419 /// \param __A
421 /// \returns A 128-bit vector of [4 x float] come from conversion of __A
422 static __inline__ __m128 __DEFAULT_FN_ATTRS128 _mm_cvtpbh_ps(__m128bh __A) {
424 (__m128i)_mm_slli_epi32((__m128i)_mm_cvtepi16_epi32((__m128i)__A), 16));
431 /// \param __A
433 /// \returns A 256-bit vector of [8 x float] come from conversion of __A
434 static __inline__ __m256 __DEFAULT_FN_ATTRS256 _mm256_cvtpbh_ps(__m128bh __A) {
436 (__m256i)_mm256_cvtepi16_epi32((__m128i)__A), 16));
446 /// \param __A
448 /// \returns A 128-bit vector of [4 x float] come from conversion of __A
450 _mm_maskz_cvtpbh_ps(__mmask8 __U, __m128bh __A) {
452 (__m128i)_mm_maskz_cvtepi16_epi32((__mmask8)__U, (__m128i)__A), 16));
462 /// \param __A
464 /// \returns A 256-bit vector of [8 x float] come from conversion of __A
466 _mm256_maskz_cvtpbh_ps(__mmask8 __U, __m128bh __A) {
468 (__m256i)_mm256_maskz_cvtepi16_epi32((__mmask8)__U, (__m128i)__A), 16));
481 /// \param __A
483 /// \returns A 128-bit vector of [4 x float] come from conversion of __A
485 _mm_mask_cvtpbh_ps(__m128 __S, __mmask8 __U, __m128bh __A) {
487 (__m128i)__S, (__mmask8)__U, (__m128i)_mm_cvtepi16_epi32((__m128i)__A),
501 /// \param __A
503 /// \returns A 256-bit vector of [8 x float] come from conversion of __A
505 _mm256_mask_cvtpbh_ps(__m256 __S, __mmask8 __U, __m128bh __A) {
507 (__m256i)__S, (__mmask8)__U, (__m256i)_mm256_cvtepi16_epi32((__m128i)__A),