/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c4-minmax-rndnu-neon-mlal-ld2r.c | 99 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r() local 182 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r() local 233 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r() local
|
D | 1x16c4-minmax-rndnu-neon-mlal-ld1r.c | 101 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r() local 185 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r() local 236 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r() local
|
D | 1x16c4-minmax-rndnu-neon-mlal-dup.c | 99 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup() local 182 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup() local 233 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld1r.c | 89 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local 140 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld2r.c | 88 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local 139 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 88 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local 139 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
D | 1x16c4s2-minmax-rndnu-neon-mlal.c | 95 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local 174 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld1r.c | 136 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 268 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 343 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-dup.c | 132 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 262 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 337 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld2r.c | 132 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 262 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 337 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
|
D | 2x16c4-minmax-rndnu-neon-mull-dup.c | 114 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local 189 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld1r.c | 116 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local 191 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld2r.c | 114 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local 189 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld2r() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c4-minmax-rndnu-neon-mlal-dup.c | 110 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup() local 193 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup() local 244 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup() local
|
D | 1x16c4-minmax-rndnu-neon-mlal-ld2r.c | 110 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r() local 193 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r() local 244 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld2r() local
|
D | 1x16c4-minmax-rndnu-neon-mlal-ld1r.c | 112 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r() local 196 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r() local 247 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_ld1r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld2r.c | 99 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local 150 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld2r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-ld1r.c | 100 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local 151 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_ld1r() local
|
D | 1x16c4-minmax-rndnu-neon-mull-dup.c | 99 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local 150 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup() local
|
D | 1x16c4s2-minmax-rndnu-neon-mlal.c | 106 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local 185 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-dup.c | 145 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 275 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local 350 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld2r.c | 145 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 275 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local 350 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r() local
|
D | 2x16c4-minmax-rndnu-neon-mlal-ld1r.c | 149 int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 281 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local 356 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r() local
|
D | 2x16c4-minmax-rndnu-neon-mull-dup.c | 127 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local 202 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup() local
|
D | 2x16c4-minmax-rndnu-neon-mull-ld1r.c | 129 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local 204 const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_ld1r() local
|