/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x8c8-minmax-neon-mull-padal.c | 105 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local 107 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal()
|
D | 1x8c16-minmax-neon-mlal-padal.c | 113 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local 115 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
|
D | 1x8c8-minmax-neon-mlal-padal.c | 154 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local 156 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
|
D | 2x8c8-minmax-neon-mull-padal.c | 142 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local 144 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal()
|
D | 1x16c8-minmax-neon-mull-padal.c | 143 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 145 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 2x8c16-minmax-neon-mlal-padal.c | 158 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local 160 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 159 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 161 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 2x8c8-minmax-neon-mlal-padal.c | 217 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 219 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
|
D | 3x8c8-minmax-neon-mull-padal.c | 179 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local 181 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
|
D | 3x8c16-minmax-neon-mlal-padal.c | 203 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 205 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 232 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 234 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 4x8c8-minmax-neon-mull-padal.c | 216 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local 218 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x8c8-minmax-neon-mull-padal.c | 119 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local 121 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal()
|
D | 1x8c16-minmax-neon-mlal-padal.c | 127 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local 129 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
|
D | 1x8c8-minmax-neon-mlal-padal.c | 168 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local 170 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
|
D | 2x8c8-minmax-neon-mull-padal.c | 158 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local 160 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal()
|
D | 1x16c8-minmax-neon-mull-padal.c | 157 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 159 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 2x8c16-minmax-neon-mlal-padal.c | 174 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local 176 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 173 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 175 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 2x8c8-minmax-neon-mlal-padal.c | 233 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 235 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
|
D | 3x8c8-minmax-neon-mull-padal.c | 197 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local 199 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 246 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 248 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 3x8c16-minmax-neon-mlal-padal.c | 221 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 223 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
|
D | 4x8c8-minmax-neon-mull-padal.c | 236 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local 238 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mull-padal.c | 226 const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 228 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
|