/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x16c4-minmax-rndnu-neon-mlal-dup.c | 126 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 259 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 352 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld1r.c | 130 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 265 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 358 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld2r.c | 126 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 259 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 352 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
|
D | 2x16c4-minmax-rndnu-neon-mull-dup.c | 111 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local 204 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld1r.c | 113 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local 206 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld2r.c | 111 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local 204 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld2r.c | 156 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 337 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 456 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-dup.c | 156 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 337 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 456 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld1r.c | 162 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 346 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 465 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
|
D | 2x16c4s2-minmax-rndnu-neon-mlal.c | 120 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local 248 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
|
D | 3x16c4-minmax-rndnu-neon-mull-dup.c | 135 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local 254 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local
|
D | 3x16c4-minmax-rndnu-neon-mull-ld2r.c | 135 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local 254 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mull-ld1r.c | 138 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local 257 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 2x16c4-minmax-rndnu-neon-mlal-dup.c | 139 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 272 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 365 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld2r.c | 139 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 272 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 365 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld1r.c | 143 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 278 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 371 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld2r.c | 124 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local 217 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local
|
D | 2x16c4-minmax-rndnu-neon-mull-dup.c | 124 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local 217 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld1r.c | 126 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local 219 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld1r.c | 177 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 361 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 480 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-dup.c | 171 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 352 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 471 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld2r.c | 171 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 352 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 471 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
|
D | 2x16c4s2-minmax-rndnu-neon-mlal.c | 133 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local 261 int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
|
D | 3x16c4-minmax-rndnu-neon-mull-dup.c | 150 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local 269 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local
|
D | 3x16c4-minmax-rndnu-neon-mull-ld2r.c | 150 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local 269 const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local
|