Home
last modified time | relevance | path

Searched refs:vp3x01234567hi (Results 1 – 9 of 9) sorted by relevance

/external/XNNPACK/src/qs8-dwconv/gen/
Dup8x9-minmax-sse41-mul16.c134 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
136 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
137 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
298 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
300 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
301 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
Dup8x9-minmax-sse2-mul16.c134 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
136 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
137 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
319 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
321 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
322 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
Dup8x9-minmax-ssse3-mul16.c134 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
136 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
137 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
319 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
321 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
322 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
Dup16x9-minmax-sse41-mul16.c164 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
168 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
169 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
399 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
401 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
402 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
Dup16x9-minmax-ssse3-mul16.c164 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
168 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
169 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
436 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
438 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
439 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
Dup16x9-minmax-sse2-mul16.c164 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
168 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
169 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
436 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
438 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
439 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
Dup24x9-minmax-sse41-mul16.c194 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
200 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
201 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
497 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
499 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
500 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
Dup24x9-minmax-ssse3-mul16.c194 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
200 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
201 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
550 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
552 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
553 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
Dup24x9-minmax-sse2-mul16.c194 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
200 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
201 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
550 const __m128i vp3x01234567hi = _mm_mulhi_epi16(vxi3x01234567, vxk3x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
552 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
553 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp3x01234567lo, vp3x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()