| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 1x16c4-minmax-rndnu-neon-mlal-ld1r.c | 105 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r() local 187 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r() local 238 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r() local
|
| D | 1x16c4-minmax-rndnu-neon-mlal-ld2r.c | 103 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r() local 184 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r() local 235 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r() local
|
| D | 1x16c4-minmax-rndnu-neon-mlal-dup.c | 103 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup() local 184 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup() local 235 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup() local
|
| D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 90 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local 141 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
| D | 1x16c4-minmax-rndnu-neon-mull-ld2r.c | 90 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local 141 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
|
| D | 1x16c4-minmax-rndnu-neon-mull-ld1r.c | 91 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local 142 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
|
| D | 1x16c4s2-minmax-rndnu-neon-mlal.c | 99 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local 176 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local
|
| D | 2x16c4-minmax-rndnu-neon-mlal-dup.c | 139 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 266 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 339 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
|
| D | 2x16c4-minmax-rndnu-neon-mlal-ld2r.c | 139 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 266 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 339 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
|
| D | 2x16c4-minmax-rndnu-neon-mlal-ld1r.c | 143 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 272 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 345 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
|
| D | 2x16c4-minmax-rndnu-neon-mull-dup.c | 118 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local 191 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local
|
| D | 2x16c4-minmax-rndnu-neon-mull-ld2r.c | 118 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local 191 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local
|
| D | 2x16c4-minmax-rndnu-neon-mull-ld1r.c | 120 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local 193 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local
|
| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 1x16c4-minmax-rndnu-neon-mlal-ld1r.c | 116 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r() local 198 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r() local 249 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r() local
|
| D | 1x16c4-minmax-rndnu-neon-mlal-ld2r.c | 114 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r() local 195 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r() local 246 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r() local
|
| D | 1x16c4-minmax-rndnu-neon-mlal-dup.c | 114 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup() local 195 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup() local 246 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup() local
|
| D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 101 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local 152 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
| D | 1x16c4-minmax-rndnu-neon-mull-ld2r.c | 101 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local 152 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
|
| D | 1x16c4-minmax-rndnu-neon-mull-ld1r.c | 102 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local 153 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
|
| D | 1x16c4s2-minmax-rndnu-neon-mlal.c | 110 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local 187 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local
|
| D | 2x16c4-minmax-rndnu-neon-mlal-dup.c | 152 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 279 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 352 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
|
| D | 2x16c4-minmax-rndnu-neon-mlal-ld2r.c | 152 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 279 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 352 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
|
| D | 2x16c4-minmax-rndnu-neon-mlal-ld1r.c | 156 int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 285 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 358 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
|
| D | 2x16c4-minmax-rndnu-neon-mull-dup.c | 131 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local 204 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local
|
| D | 2x16c4-minmax-rndnu-neon-mull-ld2r.c | 131 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local 204 const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local
|