/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 157 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 173 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 246 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 2x16c8-minmax-neon-mull-padal.c | 224 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 256 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 3x16c8-minmax-neon-mull-padal.c | 291 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 363 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 339 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 4x16c8-minmax-neon-mull-padal.c | 358 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 480 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 422 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 597 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 171 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 187 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 260 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 2x16c8-minmax-neon-mull-padal.c | 240 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 272 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 3x16c8-minmax-neon-mull-padal.c | 309 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 379 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 357 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 4x16c8-minmax-neon-mull-padal.c | 378 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 442 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 498 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 617 const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(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()
|