/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x8c8-minmax-neon-mull-padal.c | 102 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local 106 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal()
|
D | 1x8c16-minmax-neon-mlal-padal.c | 110 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local 114 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
|
D | 1x8c8-minmax-neon-mlal-padal.c | 151 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local 155 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
|
D | 2x8c8-minmax-neon-mull-padal.c | 139 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local 143 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal()
|
D | 1x16c8-minmax-neon-mull-padal.c | 140 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 144 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 2x8c16-minmax-neon-mlal-padal.c | 155 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local 159 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 156 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 160 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 2x8c8-minmax-neon-mlal-padal.c | 214 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 218 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
|
D | 3x8c8-minmax-neon-mull-padal.c | 176 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local 180 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
|
D | 3x8c16-minmax-neon-mlal-padal.c | 200 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 204 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 229 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 233 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 4x8c8-minmax-neon-mull-padal.c | 213 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local 217 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x8c8-minmax-neon-mull-padal.c | 116 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local 120 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal()
|
D | 1x8c16-minmax-neon-mlal-padal.c | 124 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local 128 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
|
D | 1x8c8-minmax-neon-mlal-padal.c | 165 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local 169 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
|
D | 2x8c8-minmax-neon-mull-padal.c | 155 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local 159 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal()
|
D | 1x16c8-minmax-neon-mull-padal.c | 154 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 158 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 2x8c16-minmax-neon-mlal-padal.c | 171 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local 175 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 170 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 174 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 2x8c8-minmax-neon-mlal-padal.c | 230 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local 234 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
|
D | 3x8c8-minmax-neon-mull-padal.c | 194 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local 198 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 243 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 247 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 3x8c16-minmax-neon-mlal-padal.c | 218 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 222 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
|
D | 4x8c8-minmax-neon-mull-padal.c | 233 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local 237 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mull-padal.c | 223 const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 227 const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
|