Home
last modified time | relevance | path

Searched refs:vq31prod02 (Results 1 – 9 of 9) sorted by relevance

/external/XNNPACK/src/qs8-dwconv/gen/
Dup8x9-minmax-sse41-mul16.c216 const __m128i vq31prod02 = _mm_srli_epi64(vprod02, 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
221 const __m128i vq31prod0123 = _mm_blend_epi16(vq31prod02, vq31prod13, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
376 const __m128i vq31prod02 = _mm_srli_epi64(vprod02, 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
381 const __m128i vq31prod0123 = _mm_blend_epi16(vq31prod02, vq31prod13, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
Dup8x9-minmax-sse2-mul16.c232 const __m128i vq31prod02 = _mm_srli_epi64(_mm_add_epi64(vprod02, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
238 _mm_castsi128_ps(vq31prod02), _mm_castsi128_ps(vq31prod13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
412 const __m128i vq31prod02 = _mm_srli_epi64(_mm_add_epi64(vprod02, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
418 _mm_castsi128_ps(vq31prod02), _mm_castsi128_ps(vq31prod13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
Dup8x9-minmax-ssse3-mul16.c232 const __m128i vq31prod02 = _mm_srli_epi64(_mm_add_epi64(vprod02, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
238 _mm_castsi128_ps(vq31prod02), _mm_castsi128_ps(vq31prod13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
412 const __m128i vq31prod02 = _mm_srli_epi64(_mm_add_epi64(vprod02, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
418 _mm_castsi128_ps(vq31prod02), _mm_castsi128_ps(vq31prod13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
Dup16x9-minmax-sse41-mul16.c296 const __m128i vq31prod02 = _mm_srli_epi64(vprod02, 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
305 const __m128i vq31prod0123 = _mm_blend_epi16(vq31prod02, vq31prod13, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
484 const __m128i vq31prod02 = _mm_srli_epi64(vprod02, 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
489 const __m128i vq31prod0123 = _mm_blend_epi16(vq31prod02, vq31prod13, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
Dup16x9-minmax-ssse3-mul16.c324 const __m128i vq31prod02 = _mm_srli_epi64(_mm_add_epi64(vprod02, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
334 _mm_castsi128_ps(vq31prod02), _mm_castsi128_ps(vq31prod13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
536 const __m128i vq31prod02 = _mm_srli_epi64(_mm_add_epi64(vprod02, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
542 _mm_castsi128_ps(vq31prod02), _mm_castsi128_ps(vq31prod13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
Dup16x9-minmax-sse2-mul16.c324 const __m128i vq31prod02 = _mm_srli_epi64(_mm_add_epi64(vprod02, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
334 _mm_castsi128_ps(vq31prod02), _mm_castsi128_ps(vq31prod13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
536 const __m128i vq31prod02 = _mm_srli_epi64(_mm_add_epi64(vprod02, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
542 _mm_castsi128_ps(vq31prod02), _mm_castsi128_ps(vq31prod13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
Dup24x9-minmax-sse41-mul16.c376 const __m128i vq31prod02 = _mm_srli_epi64(vprod02, 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
389 const __m128i vq31prod0123 = _mm_blend_epi16(vq31prod02, vq31prod13, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
582 const __m128i vq31prod02 = _mm_srli_epi64(vprod02, 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
587 const __m128i vq31prod0123 = _mm_blend_epi16(vq31prod02, vq31prod13, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
Dup24x9-minmax-ssse3-mul16.c416 const __m128i vq31prod02 = _mm_srli_epi64(_mm_add_epi64(vprod02, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
430 _mm_castsi128_ps(vq31prod02), _mm_castsi128_ps(vq31prod13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
650 const __m128i vq31prod02 = _mm_srli_epi64(_mm_add_epi64(vprod02, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
656 _mm_castsi128_ps(vq31prod02), _mm_castsi128_ps(vq31prod13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
Dup24x9-minmax-sse2-mul16.c416 const __m128i vq31prod02 = _mm_srli_epi64(_mm_add_epi64(vprod02, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
430 _mm_castsi128_ps(vq31prod02), _mm_castsi128_ps(vq31prod13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
650 const __m128i vq31prod02 = _mm_srli_epi64(_mm_add_epi64(vprod02, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
656 _mm_castsi128_ps(vq31prod02), _mm_castsi128_ps(vq31prod13), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()