/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x16c4s2-minmax-rndnu-neon-mlal.c | 180 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local 282 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld1r.c | 191 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 300 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-dup.c | 187 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 294 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld2r.c | 187 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 294 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
|
D | 3x16c4s2-minmax-rndnu-neon-mlal.c | 234 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local 374 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
|
D | 2x16c4s2-minmax-rndnu-neon-mull.c | 140 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld2r.c | 243 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 389 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-dup.c | 243 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 389 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
|
D | 4x16c4s2-minmax-rndnu-neon-mlal.c | 288 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local 466 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld1r.c | 249 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 398 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
|
D | 2x16c4-minmax-rndnu-neon-mull-dup.c | 146 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld1r.c | 148 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld2r.c | 146 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 2x16c4s2-minmax-rndnu-neon-mlal.c | 193 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local 295 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-dup.c | 200 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 307 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld2r.c | 200 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 307 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld1r.c | 204 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 313 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
|
D | 3x16c4s2-minmax-rndnu-neon-mlal.c | 249 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local 389 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
|
D | 2x16c4s2-minmax-rndnu-neon-mull.c | 153 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld1r.c | 264 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 413 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld2r.c | 258 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 404 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
|
D | 4x16c4s2-minmax-rndnu-neon-mlal.c | 305 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local 483 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-dup.c | 258 int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 404 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mull-dup.c | 159 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld1r.c | 161 const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local
|