/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 148 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 151 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 164 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 167 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 237 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 240 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 2x16c8-minmax-neon-mull-padal.c | 215 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local 218 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 247 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 250 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 282 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 285 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 354 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 357 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 330 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 333 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 349 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 352 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 471 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 474 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 413 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 416 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 588 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 591 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 162 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 165 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 178 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 181 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 251 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 254 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 2x16c8-minmax-neon-mull-padal.c | 231 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 234 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 263 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 266 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 300 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 303 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 370 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 373 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 348 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 351 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 369 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 372 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 433 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 436 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 489 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 492 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 608 const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 611 const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|