/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x8c8-minmax-neon-mull-padal.c | 89 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local 100 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local
|
D | 1x8c16-minmax-neon-mlal-padal.c | 97 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local 108 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
|
D | 1x8c8-minmax-neon-mlal-padal.c | 138 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local 149 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
|
D | 2x8c8-minmax-neon-mull-padal.c | 120 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local 137 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
|
D | 2x8c16-minmax-neon-mlal-padal.c | 136 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local 153 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
|
D | 1x16c16-minmax-neon-mlal-padal.c | 137 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 154 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
|
D | 1x16c8-minmax-neon-mull-padal.c | 121 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 138 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
|
D | 2x8c8-minmax-neon-mlal-padal.c | 195 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 212 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 175 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 198 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
D | 1x16c8-minmax-neon-mlal-padal.c | 210 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 227 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mull-padal.c | 151 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local 174 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
|
D | 2x16c16-minmax-neon-mlal-padal.c | 208 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 237 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mlal-padal.c | 252 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 275 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x8c8-minmax-neon-mull-padal.c | 103 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local 114 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local
|
D | 1x8c16-minmax-neon-mlal-padal.c | 111 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local 122 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
|
D | 1x8c8-minmax-neon-mlal-padal.c | 152 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local 163 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
|
D | 1x16c8-minmax-neon-mull-padal.c | 135 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 152 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
|
D | 1x16c16-minmax-neon-mlal-padal.c | 151 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 168 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
|
D | 2x8c16-minmax-neon-mlal-padal.c | 152 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local 169 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
|
D | 2x8c8-minmax-neon-mull-padal.c | 136 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local 153 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
|
D | 2x8c8-minmax-neon-mlal-padal.c | 211 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 228 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
|
D | 3x8c8-minmax-neon-mull-padal.c | 169 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local 192 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
|
D | 3x8c16-minmax-neon-mlal-padal.c | 193 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 216 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
|
D | 1x16c8-minmax-neon-mlal-padal.c | 224 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 241 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
|
D | 4x8c8-minmax-neon-mull-padal.c | 202 const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local 231 const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
|