Home
last modified time | relevance | path

Searched refs:vp1x01234567hi (Results 1 – 9 of 9) sorted by relevance

/external/XNNPACK/src/qs8-dwconv/gen/
Dup8x9-minmax-sse41-mul16.c108 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
110 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
111 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
274 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
276 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
277 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
Dup8x9-minmax-sse2-mul16.c108 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
110 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
111 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
295 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
297 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
298 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
Dup8x9-minmax-ssse3-mul16.c108 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
110 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
111 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
295 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
297 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
298 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
Dup16x9-minmax-sse41-mul16.c122 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
126 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
127 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
373 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
375 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
376 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
Dup16x9-minmax-ssse3-mul16.c122 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
126 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
127 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
410 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
412 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
413 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
Dup16x9-minmax-sse2-mul16.c122 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
126 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
127 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
410 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
412 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
413 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
Dup24x9-minmax-sse41-mul16.c136 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
142 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
143 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
471 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
473 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
474 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
Dup24x9-minmax-ssse3-mul16.c136 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
142 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
143 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
524 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
526 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
527 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
Dup24x9-minmax-sse2-mul16.c136 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
142 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
143 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
524 const __m128i vp1x01234567hi = _mm_mulhi_epi16(vxi1x01234567, vxk1x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
526 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
527 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp1x01234567lo, vp1x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()