/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x16c4-minmax-rndnu-neon-mlal-ld1r.c | 178 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 362 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 497 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld2r.c | 172 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 353 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 488 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-dup.c | 172 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 353 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 488 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
|
D | 3x16c4-minmax-rndnu-neon-mull-dup.c | 151 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local 286 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local
|
D | 3x16c4-minmax-rndnu-neon-mull-ld2r.c | 151 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local 286 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mull-ld1r.c | 154 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local 289 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local
|
D | 4x16c4-minmax-rndnu-neon-mlal-ld1r.c | 212 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local 445 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local 606 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
|
D | 4x16c4-minmax-rndnu-neon-mlal-dup.c | 204 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local 433 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local 594 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
|
D | 4x16c4-minmax-rndnu-neon-mlal-ld2r.c | 204 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local 433 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local 594 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
|
D | 3x16c4s2-minmax-rndnu-neon-mlal.c | 164 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local 339 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
|
D | 4x16c4-minmax-rndnu-neon-mull-ld2r.c | 177 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local 338 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
|
D | 4x16c4-minmax-rndnu-neon-mull-ld1r.c | 181 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local 342 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
|
/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x16c4-minmax-rndnu-neon-mlal-ld2r.c | 157 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 338 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 473 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-dup.c | 157 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 338 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 473 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld1r.c | 163 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 347 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 482 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
|
D | 3x16c4-minmax-rndnu-neon-mull-ld1r.c | 139 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local 274 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local
|
D | 3x16c4-minmax-rndnu-neon-mull-ld2r.c | 136 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local 271 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mull-dup.c | 136 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local 271 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local
|
D | 4x16c4-minmax-rndnu-neon-mlal-dup.c | 187 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local 416 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local 577 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
|
D | 4x16c4-minmax-rndnu-neon-mlal-ld1r.c | 195 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local 428 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local 589 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
|
D | 4x16c4-minmax-rndnu-neon-mlal-ld2r.c | 187 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local 416 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local 577 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
|
D | 3x16c4s2-minmax-rndnu-neon-mlal.c | 149 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local 324 int16x8_t vprod2x89c0 = vmull_s8(vb89c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
|
D | 4x16c4-minmax-rndnu-neon-mull-ld2r.c | 160 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local 321 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
|
D | 4x16c4-minmax-rndnu-neon-mull-dup.c | 160 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local 321 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
|
D | 4x16c4-minmax-rndnu-neon-mull-ld1r.c | 164 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local 325 const int16x8_t vprod2x89c0 = vmull_s8(vb89c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
|