/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x16c4s2-minmax-rndnu-neon-mlal.c | 194 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local 290 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld1r.c | 205 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 308 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-dup.c | 201 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 302 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld2r.c | 201 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 302 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
|
D | 3x16c4s2-minmax-rndnu-neon-mlal.c | 254 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local 386 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
|
D | 2x16c4s2-minmax-rndnu-neon-mull.c | 148 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld2r.c | 263 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 401 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-dup.c | 263 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 401 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
|
D | 4x16c4s2-minmax-rndnu-neon-mlal.c | 314 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local 482 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld1r.c | 269 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 410 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
|
D | 2x16c4-minmax-rndnu-neon-mull-dup.c | 154 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld1r.c | 156 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld2r.c | 154 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, 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 | 207 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local 303 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mlal() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-dup.c | 214 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 315 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld2r.c | 214 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 315 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld1r.c | 218 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 321 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
|
D | 3x16c4s2-minmax-rndnu-neon-mlal.c | 269 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local 401 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
|
D | 2x16c4s2-minmax-rndnu-neon-mull.c | 161 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4s2__neon_mull() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld1r.c | 284 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 425 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld2r.c | 278 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 416 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
|
D | 4x16c4s2-minmax-rndnu-neon-mlal.c | 331 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local 499 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-dup.c | 278 int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 416 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mull-dup.c | 167 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld1r.c | 169 const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local
|