/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x8c8-minmax-neon-mull-padal.c | 225 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local 227 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
|
D | 3x8c16-minmax-neon-mlal-padal.c | 249 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 251 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
|
D | 4x8c8-minmax-neon-mull-padal.c | 264 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local 266 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
|
D | 3x8c8-minmax-neon-mlal-padal.c | 326 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 328 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
|
D | 4x8c16-minmax-neon-mlal-padal.c | 296 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local 298 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 351 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 353 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 4x8c8-minmax-neon-mlal-padal.c | 391 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 393 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 399 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 401 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 420 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 422 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 484 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 486 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 540 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 542 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 659 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 661 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x8c8-minmax-neon-mull-padal.c | 207 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local 209 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
|
D | 3x8c16-minmax-neon-mlal-padal.c | 231 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local 233 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
|
D | 4x8c8-minmax-neon-mull-padal.c | 244 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local 246 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
|
D | 3x8c8-minmax-neon-mlal-padal.c | 308 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local 310 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
|
D | 4x8c16-minmax-neon-mlal-padal.c | 276 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local 278 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
|
D | 4x8c8-minmax-neon-mlal-padal.c | 371 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local 373 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 333 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 335 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 381 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 383 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 400 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 402 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 522 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 524 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 464 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 466 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 639 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 641 const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|