Home
last modified time | relevance | path

Searched refs:vq31prod46 (Results 1 – 9 of 9) sorted by relevance

/external/XNNPACK/src/qs8-dwconv/gen/
Dup8x9-minmax-sse41-mul16.c218 const __m128i vq31prod46 = _mm_srli_epi64(vprod46, 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
222 const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
378 const __m128i vq31prod46 = _mm_srli_epi64(vprod46, 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
382 const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
Dup8x9-minmax-sse2-mul16.c234 const __m128i vq31prod46 = _mm_srli_epi64(_mm_add_epi64(vprod46, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
240 _mm_castsi128_ps(vq31prod46), _mm_castsi128_ps(vq31prod57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
414 const __m128i vq31prod46 = _mm_srli_epi64(_mm_add_epi64(vprod46, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
420 _mm_castsi128_ps(vq31prod46), _mm_castsi128_ps(vq31prod57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
Dup8x9-minmax-ssse3-mul16.c234 const __m128i vq31prod46 = _mm_srli_epi64(_mm_add_epi64(vprod46, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
240 _mm_castsi128_ps(vq31prod46), _mm_castsi128_ps(vq31prod57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
414 const __m128i vq31prod46 = _mm_srli_epi64(_mm_add_epi64(vprod46, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
420 _mm_castsi128_ps(vq31prod46), _mm_castsi128_ps(vq31prod57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
Dup16x9-minmax-sse41-mul16.c298 const __m128i vq31prod46 = _mm_srli_epi64(vprod46, 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
306 const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
486 const __m128i vq31prod46 = _mm_srli_epi64(vprod46, 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
490 const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
Dup16x9-minmax-ssse3-mul16.c326 const __m128i vq31prod46 = _mm_srli_epi64(_mm_add_epi64(vprod46, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
336 _mm_castsi128_ps(vq31prod46), _mm_castsi128_ps(vq31prod57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
538 const __m128i vq31prod46 = _mm_srli_epi64(_mm_add_epi64(vprod46, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
544 _mm_castsi128_ps(vq31prod46), _mm_castsi128_ps(vq31prod57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
Dup16x9-minmax-sse2-mul16.c326 const __m128i vq31prod46 = _mm_srli_epi64(_mm_add_epi64(vprod46, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
336 _mm_castsi128_ps(vq31prod46), _mm_castsi128_ps(vq31prod57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
538 const __m128i vq31prod46 = _mm_srli_epi64(_mm_add_epi64(vprod46, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
544 _mm_castsi128_ps(vq31prod46), _mm_castsi128_ps(vq31prod57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
Dup24x9-minmax-sse41-mul16.c378 const __m128i vq31prod46 = _mm_srli_epi64(vprod46, 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
390 const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
584 const __m128i vq31prod46 = _mm_srli_epi64(vprod46, 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
588 const __m128i vq31prod4567 = _mm_blend_epi16(vq31prod46, vq31prod57, 0xCC); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
Dup24x9-minmax-ssse3-mul16.c418 const __m128i vq31prod46 = _mm_srli_epi64(_mm_add_epi64(vprod46, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
432 _mm_castsi128_ps(vq31prod46), _mm_castsi128_ps(vq31prod57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
652 const __m128i vq31prod46 = _mm_srli_epi64(_mm_add_epi64(vprod46, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
658 _mm_castsi128_ps(vq31prod46), _mm_castsi128_ps(vq31prod57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
Dup24x9-minmax-sse2-mul16.c418 const __m128i vq31prod46 = _mm_srli_epi64(_mm_add_epi64(vprod46, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
432 _mm_castsi128_ps(vq31prod46), _mm_castsi128_ps(vq31prod57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
652 const __m128i vq31prod46 = _mm_srli_epi64(_mm_add_epi64(vprod46, vrounding), 31); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
658 _mm_castsi128_ps(vq31prod46), _mm_castsi128_ps(vq31prod57), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()