/external/XNNPACK/src/qs8-igemm/gen/ |
D | 4x16c2-minmax-rndnu-neon-mull-dup.c | 182 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local 359 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld4r.c | 182 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local 359 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld2r.c | 231 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local 468 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local 645 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld4r.c | 223 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local 456 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local 633 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-dup.c | 223 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local 456 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local 633 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
|
D | 4x16c2s4-minmax-rndnu-neon-mlal.c | 213 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local 437 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld1r.c | 194 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r() local 371 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld2r.c | 186 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r() local 363 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld1r.c | 247 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local 492 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local 669 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
|
D | 4x16c2s4-minmax-rndnu-neon-mull.c | 175 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, 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 | 165 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local 342 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld4r.c | 165 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local 342 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-dup.c | 205 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local 438 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local 615 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld2r.c | 213 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local 450 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local 627 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld4r.c | 205 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local 438 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local 615 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld2r.c | 169 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r() local 346 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld2r() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld1r.c | 177 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r() local 354 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld1r() local
|
D | 4x16c2s4-minmax-rndnu-neon-mlal.c | 196 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local 420 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld1r.c | 229 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local 474 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local 651 const int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1, va3c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r() local
|
D | 4x16c2s4-minmax-rndnu-neon-mull.c | 158 int16x8_t vprod3x89ABc1 = vmull_s8(vb89ABc1x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mull() local
|