/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 141 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 173 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
|
D | 1x16c16-minmax-neon-mlal-padal.c | 157 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 189 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
|
D | 1x16c8-minmax-neon-mlal-padal.c | 230 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 262 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
|
D | 2x16c8-minmax-neon-mull-padal.c | 198 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 242 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
|
D | 2x16c16-minmax-neon-mlal-padal.c | 230 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 274 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
|
D | 2x16c8-minmax-neon-mlal-padal.c | 337 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 381 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 255 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 311 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 303 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 359 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 312 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 380 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 376 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 444 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 444 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 500 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 551 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 619 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c16-minmax-neon-mlal-padal.c | 143 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 175 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
|
D | 1x16c8-minmax-neon-mull-padal.c | 127 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 159 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
|
D | 1x16c8-minmax-neon-mlal-padal.c | 216 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 248 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
|
D | 2x16c16-minmax-neon-mlal-padal.c | 214 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 258 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
|
D | 2x16c8-minmax-neon-mull-padal.c | 182 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local 226 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
|
D | 3x16c8-minmax-neon-mull-padal.c | 237 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 293 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
|
D | 2x16c8-minmax-neon-mlal-padal.c | 321 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 365 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
|
D | 3x16c16-minmax-neon-mlal-padal.c | 285 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 341 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
|
D | 4x16c16-minmax-neon-mlal-padal.c | 356 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 424 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mull-padal.c | 292 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 360 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
|
D | 3x16c8-minmax-neon-mlal-padal.c | 426 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 482 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
|
D | 4x16c8-minmax-neon-mlal-padal.c | 531 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 599 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
|