/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c2-minmax-rndnu-neon-mlal-ld2r.c | 106 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 191 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 248 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-dup.c | 104 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 188 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 245 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld1r.c | 110 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 197 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 254 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld4r.c | 104 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 188 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 245 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld2r.c | 90 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local 147 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-dup.c | 89 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local 146 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld4r.c | 89 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local 146 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld1r.c | 92 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r() local 149 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r() local
|
D | 1x16c2s4-minmax-rndnu-neon-mlal.c | 101 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local 179 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-dup.c | 141 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 273 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 360 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld4r.c | 141 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 273 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 360 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld2r.c | 145 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 279 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 366 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld1r.c | 153 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local 291 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local 378 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c2-minmax-rndnu-neon-mlal-ld2r.c | 118 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 203 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 260 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld1r.c | 122 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 209 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 266 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-dup.c | 116 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 200 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 257 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld4r.c | 116 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 200 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 257 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld4r.c | 100 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local 157 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld2r.c | 101 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local 158 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld1r.c | 103 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r() local 160 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-dup.c | 100 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local 157 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local
|
D | 1x16c2s4-minmax-rndnu-neon-mlal.c | 112 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local 190 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld4r.c | 155 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 287 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 374 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-dup.c | 155 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 287 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 374 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld2r.c | 159 int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 293 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 380 const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
|