/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x16c2-minmax-rndnu-neon-mlal-dup.c | 135 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 270 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 367 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld4r.c | 135 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 270 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 367 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld2r.c | 139 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 276 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 373 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld1r.c | 147 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local 288 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local 385 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
|
D | 2x16c2-minmax-rndnu-neon-mull-dup.c | 113 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local 210 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
|
D | 2x16c2-minmax-rndnu-neon-mull-ld1r.c | 119 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local 216 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
|
D | 2x16c2-minmax-rndnu-neon-mull-ld4r.c | 113 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local 210 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
|
D | 2x16c2-minmax-rndnu-neon-mull-ld2r.c | 115 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r() local 212 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r() local
|
D | 2x16c2s4-minmax-rndnu-neon-mlal.c | 130 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local 258 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local
|
D | 3x16c2-minmax-rndnu-neon-mlal-dup.c | 169 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local 353 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local 482 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
|
D | 3x16c2-minmax-rndnu-neon-mlal-ld4r.c | 169 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local 353 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local 482 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
|
D | 3x16c2-minmax-rndnu-neon-mull-dup.c | 138 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local 267 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local
|
D | 3x16c2-minmax-rndnu-neon-mull-ld2r.c | 141 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r() local 270 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 2x16c2-minmax-rndnu-neon-mlal-ld4r.c | 149 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 284 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 381 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-dup.c | 149 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 284 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 381 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld2r.c | 153 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 290 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 387 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
|
D | 2x16c2-minmax-rndnu-neon-mull-dup.c | 126 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local 223 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup() local
|
D | 2x16c2-minmax-rndnu-neon-mull-ld4r.c | 126 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local 223 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld4r() local
|
D | 2x16c2-minmax-rndnu-neon-mull-ld2r.c | 128 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r() local 225 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld2r() local
|
D | 2x16c2-minmax-rndnu-neon-mull-ld1r.c | 132 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local 229 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_ld1r() local
|
D | 2x16c2s4-minmax-rndnu-neon-mlal.c | 143 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local 271 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld1r.c | 161 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local 302 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local 399 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
|
D | 3x16c2-minmax-rndnu-neon-mlal-ld4r.c | 185 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local 369 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local 498 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
|
D | 3x16c2-minmax-rndnu-neon-mlal-dup.c | 185 int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local 369 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local 498 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
|
D | 3x16c2-minmax-rndnu-neon-mull-dup.c | 153 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local 282 const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local
|