Lines Matching refs:vout

62 …__m256i vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]} = _mm256_adds_epi16(_mm2…
64 …__m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[N:N+8]…
66 …__m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[N:N+8]…
70 …vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]} = _mm256_min_epi16(_mm256_max_ep…
72 …vout${ABC[N:N+8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[N:N+8]}, _mm256_castsi256_si128(voutput_…
74 …     vout${ABC[N:N+8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[N:N+8]}, voutput_min), voutput_max);
78 …vout${ABC[N:N+16]} = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout${ABC[N:N+4]}${A…
80 …    __m128i vout${ABC[N:N+8]}${ABC[N:N+8]} = _mm_packs_epi16(vout${ABC[N:N+8]}, vout${ABC[N:N+8]});
83       _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]});
85       _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
88         _mm_storeu_si128((__m128i*) (output + ${N}), vout${ABC[N:N+16]});
90         _mm_storel_epi64((__m128i*) (output + ${N}), vout${ABC[N:N+8]}${ABC[N:N+8]});
110 …__m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), …
111 …vout${ABC[0:8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[0:8]}, _mm256_castsi256_si128(voutput_min)…
113 …__m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), …
114         vout${ABC[0:8]} = _mm_min_epi16(_mm_max_epi16(vout${ABC[0:8]}, voutput_min), voutput_max);
115       __m128i vout${ABC[0:8]}${ABC[0:8]} = _mm_packs_epi16(vout${ABC[0:8]}, vout${ABC[0:8]});
119           _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]});
124             *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
125             vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
129             *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0);
130             vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
134             *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);
140           *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout${ABC[0:8]}${ABC[0:8]});
141           vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32);
145           *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout${ABC[0:8]}${ABC[0:8]}, 0);
146           vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16);
150           *output = (int8_t) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0);