/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 150 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local 152 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 166 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 168 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 239 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 241 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 2x16c8-minmax-neon-mull-padal.c | 217 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local 219 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 249 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 251 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 284 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local 286 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 356 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 358 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 332 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 334 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 351 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local 353 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 473 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 475 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 415 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 417 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 590 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 592 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c8-minmax-neon-mull-padal.c | 164 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local 166 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
|
D | 1x16c16-minmax-neon-mlal-padal.c | 180 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local 182 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
|
D | 1x16c8-minmax-neon-mlal-padal.c | 253 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local 255 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
|
D | 2x16c8-minmax-neon-mull-padal.c | 233 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local 235 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
D | 2x16c16-minmax-neon-mlal-padal.c | 265 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local 267 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mull-padal.c | 302 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local 304 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
D | 2x16c8-minmax-neon-mlal-padal.c | 372 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local 374 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
D | 3x16c16-minmax-neon-mlal-padal.c | 350 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local 352 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mull-padal.c | 371 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local 373 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
D | 4x16c16-minmax-neon-mlal-padal.c | 435 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local 437 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
D | 3x16c8-minmax-neon-mlal-padal.c | 491 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local 493 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
D | 4x16c8-minmax-neon-mlal-padal.c | 610 const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local 612 const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|