• Home
  • Raw
  • Download

Lines Matching refs:vacc4567

84       __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 4 * sizeof(int32_t)));  in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()  local
98 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp0x01234567lo, vp0x01234567hi)); 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()
124 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); 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()
150 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
163 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp5x01234567lo, vp5x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
176 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp6x01234567lo, vp6x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
189 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp7x01234567lo, vp7x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
202 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp8x01234567lo, vp8x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
210 const __m128i vnmask4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
213 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vnmask4567), vnmask4567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
255 vacc4567 = in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
259 … __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
273 __m128i vacc4567 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 4 * sizeof(int32_t))); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16() local
286 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp0x01234567lo, vp0x01234567hi)); 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()
310 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp2x01234567lo, vp2x01234567hi)); 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()
334 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp4x01234567lo, vp4x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
346 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp5x01234567lo, vp5x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
358 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp6x01234567lo, vp6x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
370 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp7x01234567lo, vp7x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
382 vacc4567 = _mm_add_epi32(vacc4567, _mm_unpackhi_epi16(vp8x01234567lo, vp8x01234567hi)); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
389 const __m128i vnmask4567 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc4567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
392 const __m128i vabsacc4567 = _mm_sub_epi32(_mm_xor_si128(vacc4567, vnmask4567), vnmask4567); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
435 vacc4567 = in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()
439 … __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); in xnn_qs8_dwconv_minmax_ukernel_up8x9__sse2_mul16()