Home
last modified time | relevance | path

Searched refs:vb4567x0123 (Results 1 – 25 of 74) sorted by relevance

123

/external/XNNPACK/src/qc8-gemm/gen/
D8x8c4-minmax-fp32-neondot.c124 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot() local
144 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot()
146 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot()
148 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot()
150 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot()
152 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot()
154 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot()
156 vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot()
158 vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot()
D6x8c4-minmax-fp32-neondot.c106 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot() local
122 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot()
124 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot()
126 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot()
128 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot()
130 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot()
132 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot()
D4x8c4-minmax-fp32-neondot.c88 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__neondot() local
100 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__neondot()
102 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__neondot()
104 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__neondot()
106 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__neondot()
/external/XNNPACK/src/qc8-igemm/gen/
D8x8c4-minmax-fp32-neondot.c146 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot() local
166 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot()
168 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot()
170 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot()
172 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot()
174 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot()
176 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot()
178 vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot()
180 vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot()
D6x8c4-minmax-fp32-neondot.c124 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot() local
140 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot()
142 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot()
144 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot()
146 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot()
148 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot()
150 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot()
D4x8c4-minmax-fp32-neondot.c102 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qc8_igemm_minmax_fp32_ukernel_4x8c4__neondot() local
114 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x8c4__neondot()
116 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x8c4__neondot()
118 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x8c4__neondot()
120 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qc8_igemm_minmax_fp32_ukernel_4x8c4__neondot()
/external/XNNPACK/src/qs8-igemm/gen/
D8x8c4-minmax-rndnu-neondot.c145 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot() local
165 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
167 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
169 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
171 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
173 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
175 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
177 vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
179 vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
D6x8c4-minmax-rndnu-neondot.c123 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot() local
139 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
141 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
143 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
145 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
147 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
149 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
D4x8c4-minmax-rndnu-neondot.c101 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neondot() local
113 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
115 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
117 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
119 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
/external/XNNPACK/src/qs8-gemm/gen/
D8x8c4-minmax-rndnu-neondot.c123 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot() local
143 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
145 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
147 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
149 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
151 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
153 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
155 vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
157 vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
D6x8c4-minmax-rndnu-neondot.c105 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot() local
121 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
123 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
125 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
127 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
129 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
131 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
D4x8c4-minmax-rndnu-neondot.c87 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neondot() local
99 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
101 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
103 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
105 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
D8x16c4-minmax-rndnu-neondot.c141 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot() local
179 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
183 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
187 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
191 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
195 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
199 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
203 vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
207 vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1); in xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot()
/external/XNNPACK/src/qu8-igemm/gen/
D8x8c4-minmax-rndnu-neondot.c157 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot() local
164 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
169 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
174 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
179 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
184 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
189 vpacc5x0123 = vdotq_lane_u32(vpacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
194 vpacc6x0123 = vdotq_lane_u32(vpacc6x0123, vb4567x0123, va6x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
199 vpacc7x0123 = vdotq_lane_u32(vpacc7x0123, vb4567x0123, va7x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_8x8c4__neondot()
D6x8c4-minmax-rndnu-neondot.c133 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot() local
140 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
145 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
150 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
155 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
160 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
165 vpacc5x0123 = vdotq_lane_u32(vpacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot()
D5x8c4-minmax-rndnu-neondot.c121 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot() local
128 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
133 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
138 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
143 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
148 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot()
D4x8c4-minmax-rndnu-neondot.c109 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot() local
116 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
121 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
126 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
131 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot()
D3x8c4-minmax-rndnu-neondot.c97 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot() local
104 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot()
109 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot()
114 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_3x8c4__neondot()
D2x8c4-minmax-rndnu-neondot.c85 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot() local
92 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot()
97 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qu8_igemm_minmax_rndnu_ukernel_2x8c4__neondot()
/external/XNNPACK/src/qu8-gemm/gen/
D8x8c4-minmax-rndnu-neondot.c133 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot() local
140 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
145 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
150 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
155 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
160 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
165 vpacc5x0123 = vdotq_lane_u32(vpacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
170 vpacc6x0123 = vdotq_lane_u32(vpacc6x0123, vb4567x0123, va6x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
175 vpacc7x0123 = vdotq_lane_u32(vpacc7x0123, vb4567x0123, va7x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_8x8c4__neondot()
D6x8c4-minmax-rndnu-neondot.c113 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot() local
120 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
125 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
130 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
135 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
140 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
145 vpacc5x0123 = vdotq_lane_u32(vpacc5x0123, vb4567x0123, va5x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot()
D5x8c4-minmax-rndnu-neondot.c103 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot() local
110 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
115 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
120 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
125 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
130 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot()
D4x8c4-minmax-rndnu-neondot.c93 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot() local
100 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
105 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
110 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
115 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot()
D3x8c4-minmax-rndnu-neondot.c83 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot() local
90 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot()
95 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot()
100 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot()
D2x8c4-minmax-rndnu-neondot.c73 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot() local
80 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot()
85 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1); in xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot()

123