Home
last modified time | relevance | path

Searched refs:vxk4x01234567 (Results 1 – 9 of 9) sorted by relevance

/external/XNNPACK/src/qs8-dwconv/gen/
Dup8x9-minmax-sse41-mul16.c142 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
146 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
147 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16()
306 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse41_mul16() local
309 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); 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()
Dup8x9-minmax-sse2-mul16.c144 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
146 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
147 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
328 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
330 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); 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()
Dup8x9-minmax-ssse3-mul16.c144 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
146 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
147 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16()
328 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up8x9__ssse3_mul16() local
330 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); 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()
Dup16x9-minmax-sse41-mul16.c176 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
184 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
185 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16()
407 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse41_mul16() local
411 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); 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()
Dup16x9-minmax-ssse3-mul16.c180 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
184 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
185 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16()
446 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__ssse3_mul16() local
448 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); 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()
Dup16x9-minmax-sse2-mul16.c180 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
184 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
185 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16()
446 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up16x9__sse2_mul16() local
448 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); 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()
Dup24x9-minmax-sse41-mul16.c210 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
222 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
223 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16()
505 const __m128i vxk4x01234567 = _mm_cvtepi8_epi16(vk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse41_mul16() local
509 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); 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()
Dup24x9-minmax-ssse3-mul16.c216 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
222 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
223 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16()
560 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__ssse3_mul16() local
562 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); 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()
Dup24x9-minmax-sse2-mul16.c216 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
222 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
223 const __m128i vp4x01234567hi = _mm_mulhi_epi16(vxi4x01234567, vxk4x01234567); in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16()
560 …const __m128i vxk4x01234567 = _mm_unpacklo_epi8(vk4x01234567, _mm_cmpgt_epi8(_mm_setzero_si128(), … in xnn_qs8_dwconv_minmax_ukernel_up24x9__sse2_mul16() local
562 const __m128i vp4x01234567lo = _mm_mullo_epi16(vxi4x01234567, vxk4x01234567); 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()