Home
last modified time | relevance | path

Searched refs:vp4x01234567hi (Results 1 – 9 of 9) sorted by relevance

/external/XNNPACK/src/qs8-dwconv/gen/
Dup8x9-minmax-sse41-mul16.c147 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
149 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
150 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
310 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
312 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
313 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
Dup8x9-minmax-sse2-mul16.c147 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
149 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
150 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
331 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
333 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
334 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
Dup8x9-minmax-ssse3-mul16.c147 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
149 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
150 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
331 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
333 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
334 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
Dup16x9-minmax-sse41-mul16.c185 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
189 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
190 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
412 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
414 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
415 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
Dup16x9-minmax-ssse3-mul16.c185 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
189 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
190 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
449 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
451 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
452 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
Dup16x9-minmax-sse2-mul16.c185 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
189 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
190 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
449 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
451 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
452 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
Dup24x9-minmax-sse41-mul16.c223 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
229 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
230 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
510 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
512 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
513 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
Dup24x9-minmax-ssse3-mul16.c223 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
229 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
230 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
563 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
565 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
566 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
Dup24x9-minmax-sse2-mul16.c223 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
229 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
230 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
563 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
565 vacc0123 = _mm_add_epi32(vacc0123, _mm_unpacklo_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
566 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()