/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 154 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 158 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 170 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 174 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 243 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 247 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 2x16c8-minmax-neon-mull-padal.c | 221 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local 225 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 253 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 257 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 288 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 292 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 360 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 364 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 336 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 340 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 355 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 359 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 477 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 481 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 419 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 423 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 594 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 598 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 168 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 172 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 184 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 188 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 257 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 261 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 2x16c8-minmax-neon-mull-padal.c | 237 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 241 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 269 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 273 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 306 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 310 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 376 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 380 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 354 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 358 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 375 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 379 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 439 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 443 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 495 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 499 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 614 const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 618 const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|