Home
last modified time | relevance | path

Searched refs:vsum0x01 (Results 1 – 25 of 48) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D1x8c8-minmax-neon-mull-padal.c88 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local
92 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal()
99 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local
101 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal()
D1x8c16-minmax-neon-mlal-padal.c96 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
100 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
107 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
109 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
D1x8c8-minmax-neon-mlal-padal.c137 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
141 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
148 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
150 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
D2x8c8-minmax-neon-mull-padal.c119 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
127 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal()
136 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
138 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal()
D1x16c8-minmax-neon-mull-padal.c120 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
128 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
137 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
139 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
D2x8c16-minmax-neon-mlal-padal.c135 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
143 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
152 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
154 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D1x16c16-minmax-neon-mlal-padal.c136 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
144 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
153 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
155 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c194 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
202 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
211 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
213 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c150 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
162 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
173 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
175 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c174 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
186 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
197 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
199 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D1x16c8-minmax-neon-mlal-padal.c209 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
217 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
226 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
228 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c181 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
197 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
210 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
212 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x8c8-minmax-neon-mull-padal.c102 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local
106 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal()
113 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local
115 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal()
D1x8c16-minmax-neon-mlal-padal.c110 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
114 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
121 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
123 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
D1x8c8-minmax-neon-mlal-padal.c151 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
155 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
162 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
164 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
D2x8c8-minmax-neon-mull-padal.c135 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
143 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal()
152 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
154 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal()
D1x16c8-minmax-neon-mull-padal.c134 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
142 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
151 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
153 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
D2x8c16-minmax-neon-mlal-padal.c151 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
159 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
168 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
170 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D1x16c16-minmax-neon-mlal-padal.c150 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
158 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
167 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
169 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c210 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
218 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
227 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
229 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c168 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
180 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
191 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
193 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
D1x16c8-minmax-neon-mlal-padal.c223 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
231 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
240 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
242 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D3x8c16-minmax-neon-mlal-padal.c192 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
204 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
215 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
217 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c201 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
217 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
230 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
232 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
D2x16c8-minmax-neon-mull-padal.c191 const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
207 int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
220 const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
222 int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()

12