/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x8c8-minmax-neon-mull-padal.c | 91 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 1x8c16-minmax-neon-mlal-padal.c | 99 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 1x8c8-minmax-neon-mlal-padal.c | 140 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 2x8c8-minmax-neon-mull-padal.c | 122 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 2x8c16-minmax-neon-mlal-padal.c | 138 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 1x16c16-minmax-neon-mlal-padal.c | 139 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 1x16c8-minmax-neon-mull-padal.c | 123 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 2x8c8-minmax-neon-mlal-padal.c | 197 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 177 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 1x16c8-minmax-neon-mlal-padal.c | 212 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 3x8c8-minmax-neon-mull-padal.c | 153 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 2x16c16-minmax-neon-mlal-padal.c | 210 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 244 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mlal-padal.c | 254 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 282 const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x8c8-minmax-neon-mull-padal.c | 105 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 1x8c16-minmax-neon-mlal-padal.c | 113 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 1x8c8-minmax-neon-mlal-padal.c | 154 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 1x16c8-minmax-neon-mull-padal.c | 137 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 1x16c16-minmax-neon-mlal-padal.c | 153 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 2x8c16-minmax-neon-mlal-padal.c | 154 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 2x8c8-minmax-neon-mull-padal.c | 138 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 2x8c8-minmax-neon-mlal-padal.c | 213 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 3x8c8-minmax-neon-mull-padal.c | 171 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 195 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 1x16c8-minmax-neon-mlal-padal.c | 226 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|
D | 4x8c8-minmax-neon-mull-padal.c | 204 const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, 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() local
|