Lines Matching refs:__m512i
59 …const __m512i vmultiplier = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.mu… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
60 …const __m512i vrounding = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.roun… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
61 …const __m512i vremainder_mask = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
62 …const __m512i vremainder_threshold = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
64 …const __m512i voutput_zero_point = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
65 …const __m512i voutput_min = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
66 …const __m512i voutput_max = _mm512_broadcast_i32x4(_mm_load_si128((const __m128i*) params->sse2.ou… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
68 __m512i vacc0x0123 = _mm512_maskz_expandloadu_epi32(vbias_mask, w); in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
69 …__m512i vacc0x4567 = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 4 *… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
70 …__m512i vacc0x89AB = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 8 *… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
71 …__m512i vacc0xCDEF = _mm512_maskz_expandloadu_epi32(vbias_mask, (const void*) ((uintptr_t) w + 12 … in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
72 __m512i vacc1x0123 = vacc0x0123; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
73 __m512i vacc1x4567 = vacc0x4567; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
74 __m512i vacc1x89AB = vacc0x89AB; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
75 __m512i vacc1xCDEF = vacc0xCDEF; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
76 __m512i vacc2x0123 = vacc0x0123; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
77 __m512i vacc2x4567 = vacc0x4567; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
78 __m512i vacc2x89AB = vacc0x89AB; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
79 __m512i vacc2xCDEF = vacc0xCDEF; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
80 __m512i vacc3x0123 = vacc0x0123; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
81 __m512i vacc3x4567 = vacc0x4567; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
82 __m512i vacc3x89AB = vacc0x89AB; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
83 __m512i vacc3xCDEF = vacc0xCDEF; in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
108 …const __m512i va0 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a0))… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
110 …const __m512i va1 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a1))… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
112 …const __m512i va2 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a2))… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
114 …const __m512i va3 = _mm512_broadcast_i32x4(_mm_cvtepi8_epi16(_mm_loadl_epi64((const __m128i*) a3))… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
117 const __m512i vb0123 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) w)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
123 …const __m512i vb4567 = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + 32… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
129 …const __m512i vb89AB = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + 64… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
135 …const __m512i vbCDEF = _mm512_cvtepi8_epi16(_mm256_load_si256((const __m256i*) ((uintptr_t) w + 96… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
148 …const __m512i vacc0x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x0123, vacc0x4567), _mm… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
149 …const __m512i vacc0x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x89AB, vacc0xCDEF), _mm… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
150 …const __m512i vacc1x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x0123, vacc1x4567), _mm… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
151 …const __m512i vacc1x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x89AB, vacc1xCDEF), _mm… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
152 …const __m512i vacc2x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x0123, vacc2x4567), _mm… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
153 …const __m512i vacc2x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x89AB, vacc2xCDEF), _mm… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
154 …const __m512i vacc3x04152637 = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x0123, vacc3x4567), _mm… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
155 …const __m512i vacc3x8C9DAEBF = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x89AB, vacc3xCDEF), _mm… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
157 …__m512i vacc0x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc0x04152637, vacc0x8C9D… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
158 …__m512i vacc1x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc1x04152637, vacc1x8C9D… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
159 …__m512i vacc2x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc2x04152637, vacc2x8C9D… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
160 …__m512i vacc3x084C195D2A6E3B7F = _mm512_add_epi32(_mm512_unpacklo_epi32(vacc3x04152637, vacc3x8C9D… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
162 …const __m512i vacc0x88CC99DDAAEEBBFF = _mm512_shuffle_epi32(vacc0x084C195D2A6E3B7F, _MM_SHUFFLE(3,… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
163 …const __m512i vacc1x88CC99DDAAEEBBFF = _mm512_shuffle_epi32(vacc1x084C195D2A6E3B7F, _MM_SHUFFLE(3,… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
164 …const __m512i vacc2x88CC99DDAAEEBBFF = _mm512_shuffle_epi32(vacc2x084C195D2A6E3B7F, _MM_SHUFFLE(3,… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
165 …const __m512i vacc3x88CC99DDAAEEBBFF = _mm512_shuffle_epi32(vacc3x084C195D2A6E3B7F, _MM_SHUFFLE(3,… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
167 …const __m512i vprod0x04152637 = _mm512_add_epi64(_mm512_mul_epi32(vacc0x084C195D2A6E3B7F, vmultipl… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
168 …const __m512i vprod1x04152637 = _mm512_add_epi64(_mm512_mul_epi32(vacc1x084C195D2A6E3B7F, vmultipl… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
169 …const __m512i vprod2x04152637 = _mm512_add_epi64(_mm512_mul_epi32(vacc2x084C195D2A6E3B7F, vmultipl… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
170 …const __m512i vprod3x04152637 = _mm512_add_epi64(_mm512_mul_epi32(vacc3x084C195D2A6E3B7F, vmultipl… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
172 …const __m512i vprod0x8C9DAEBF = _mm512_add_epi64(_mm512_mul_epi32(vacc0x88CC99DDAAEEBBFF, vmultipl… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
173 …const __m512i vprod1x8C9DAEBF = _mm512_add_epi64(_mm512_mul_epi32(vacc1x88CC99DDAAEEBBFF, vmultipl… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
174 …const __m512i vprod2x8C9DAEBF = _mm512_add_epi64(_mm512_mul_epi32(vacc2x88CC99DDAAEEBBFF, vmultipl… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
175 …const __m512i vprod3x8C9DAEBF = _mm512_add_epi64(_mm512_mul_epi32(vacc3x88CC99DDAAEEBBFF, vmultipl… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
177 const __m512i vq31prod0x04152637 = _mm512_srli_epi64(vprod0x04152637, 31); in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
178 const __m512i vq31prod0x8C9DAEBF = _mm512_add_epi64(vprod0x8C9DAEBF, vprod0x8C9DAEBF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
179 const __m512i vq31prod1x04152637 = _mm512_srli_epi64(vprod1x04152637, 31); in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
180 const __m512i vq31prod1x8C9DAEBF = _mm512_add_epi64(vprod1x8C9DAEBF, vprod1x8C9DAEBF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
181 const __m512i vq31prod2x04152637 = _mm512_srli_epi64(vprod2x04152637, 31); in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
182 const __m512i vq31prod2x8C9DAEBF = _mm512_add_epi64(vprod2x8C9DAEBF, vprod2x8C9DAEBF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
183 const __m512i vq31prod3x04152637 = _mm512_srli_epi64(vprod3x04152637, 31); in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
184 const __m512i vq31prod3x8C9DAEBF = _mm512_add_epi64(vprod3x8C9DAEBF, vprod3x8C9DAEBF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
186 …const __m512i vq31prod0x084C195D2A6E3B7F = _mm512_mask_blend_epi32(vblend_mask, vq31prod0x04152637… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
187 …const __m512i vq31prod1x084C195D2A6E3B7F = _mm512_mask_blend_epi32(vblend_mask, vq31prod1x04152637… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
188 …const __m512i vq31prod2x084C195D2A6E3B7F = _mm512_mask_blend_epi32(vblend_mask, vq31prod2x04152637… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
189 …const __m512i vq31prod3x084C195D2A6E3B7F = _mm512_mask_blend_epi32(vblend_mask, vq31prod3x04152637… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
191 const __m512i vrem0x084C195D2A6E3B7F = in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
193 const __m512i vrem1x084C195D2A6E3B7F = in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
195 const __m512i vrem2x084C195D2A6E3B7F = in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
197 const __m512i vrem3x084C195D2A6E3B7F = in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
205 const __m512i vminus_one = _mm512_set1_epi32(-1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
215 …__m512i vacc01x084Cx195Dx2A6Ex3B7F = _mm512_adds_epi16(_mm512_packs_epi32(vacc0x084C195D2A6E3B7F, … in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
216 …__m512i vacc23x084Cx195Dx2A6Ex3B7F = _mm512_adds_epi16(_mm512_packs_epi32(vacc2x084C195D2A6E3B7F, … in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
221 …__m512i vout0123x084Cx195Dx2A6Ex3B7F = _mm512_packs_epi16(vacc01x084Cx195Dx2A6Ex3B7F, vacc23x084Cx… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()
223 …const __m512i vout0123x0123456789ABCDEF = _mm512_shuffle_epi8(vout0123x084Cx195Dx2A6Ex3B7F, _mm512… in xnn_qs8_igemm_minmax_ukernel_4x16c8__avx512skx()