/external/XNNPACK/src/qs8-igemm/gen/ |
D | 4x16c2-minmax-rndnu-neon-mull-dup.c | 190 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local 361 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld4r.c | 190 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local 361 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld2r.c | 244 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local 476 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local 647 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld4r.c | 236 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local 464 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local 635 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-dup.c | 236 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local 464 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local 635 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
|
D | 4x16c2s4-minmax-rndnu-neon-mlal.c | 226 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local 445 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld1r.c | 202 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r() local 373 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld2r.c | 194 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r() local 365 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld1r.c | 260 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local 500 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local 671 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
|
D | 4x16c2s4-minmax-rndnu-neon-mull.c | 183 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mull() local
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 4x16c2-minmax-rndnu-neon-mull-dup.c | 173 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local 344 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld4r.c | 173 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local 344 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-dup.c | 218 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local 446 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local 617 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld2r.c | 226 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local 458 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local 629 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld4r.c | 218 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local 446 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local 617 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld2r.c | 177 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r() local 348 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld1r.c | 185 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r() local 356 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r() local
|
D | 4x16c2s4-minmax-rndnu-neon-mlal.c | 209 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local 428 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld1r.c | 242 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local 482 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local 653 const int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
|
D | 4x16c2s4-minmax-rndnu-neon-mull.c | 166 int16x8_t vprod3xCDEFc1 = vmull_s8(vbCDEFc1x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mull() local
|