/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c2-minmax-rndnu-neon-mlal-ld2r.c | 102 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 189 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 246 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-dup.c | 100 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 186 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 243 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld1r.c | 106 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 195 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 252 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld4r.c | 100 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 186 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 243 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld2r.c | 88 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local 145 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-dup.c | 87 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local 144 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld4r.c | 87 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local 144 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld1r.c | 90 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r() local 147 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r() local
|
D | 1x16c2s4-minmax-rndnu-neon-mlal.c | 97 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local 177 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-dup.c | 134 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 269 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 358 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld4r.c | 134 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 269 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 358 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld2r.c | 138 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 275 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 364 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld1r.c | 146 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local 287 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local 376 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c2-minmax-rndnu-neon-mlal-ld2r.c | 114 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 201 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 258 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld1r.c | 118 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 207 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 264 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-dup.c | 112 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 198 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 255 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld4r.c | 112 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 198 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 255 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld4r.c | 98 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local 155 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld2r.c | 99 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local 156 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld1r.c | 101 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r() local 158 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-dup.c | 98 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local 155 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local
|
D | 1x16c2s4-minmax-rndnu-neon-mlal.c | 108 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local 188 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld4r.c | 148 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 283 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 372 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-dup.c | 148 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 283 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 372 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld2r.c | 152 int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 289 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 378 const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
|