Home
last modified time | relevance | path

Searched refs:vp2x01234567lo (Results 1 – 9 of 9) sorted by relevance

/external/XNNPACK/src/qs8-dwconv/gen/
Dup8x9-minmax-sse41-mul16.c120 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
123 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
124 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
285 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
288 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
289 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
Dup8x9-minmax-sse2-mul16.c120 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
123 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
124 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
306 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
309 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
310 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
Dup8x9-minmax-ssse3-mul16.c120 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
123 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
124 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
306 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
309 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
310 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
Dup16x9-minmax-sse41-mul16.c142 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
147 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
148 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
385 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
388 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
389 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
Dup16x9-minmax-ssse3-mul16.c142 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
147 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
148 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
422 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
425 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
426 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
Dup16x9-minmax-sse2-mul16.c142 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
147 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
148 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
422 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
425 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
426 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
Dup24x9-minmax-sse41-mul16.c164 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
171 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
172 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
483 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
486 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
487 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
Dup24x9-minmax-ssse3-mul16.c164 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
171 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
172 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
536 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
539 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
540 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
Dup24x9-minmax-sse2-mul16.c164 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
171 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
172 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
536 const __m128i vp2x01234567lo = _mm_mullo_epi16(vxi2x01234567, vxk2x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
539 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
540 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()