/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c2-minmax-rndnu-neon-mlal-ld2r.c | 125 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 201 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 264 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-dup.c | 123 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 198 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 261 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld1r.c | 129 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 207 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 270 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld4r.c | 123 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 198 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 261 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld2r.c | 100 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local 163 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-dup.c | 99 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local 162 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld4r.c | 99 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local 162 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld1r.c | 102 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r() local 165 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r() local
|
D | 1x16c2s4-minmax-rndnu-neon-mlal.c | 119 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local 188 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-dup.c | 174 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 292 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 385 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld4r.c | 174 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 292 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 385 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld2r.c | 178 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 298 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 391 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld1r.c | 186 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local 310 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld1r() local 403 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); 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 | 137 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 213 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 276 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld1r.c | 141 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 219 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 282 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-dup.c | 135 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 210 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 273 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld4r.c | 135 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 210 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 273 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld4r.c | 110 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local 173 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld2r.c | 111 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local 174 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld1r.c | 113 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r() local 176 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-dup.c | 110 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local 173 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local
|
D | 1x16c2s4-minmax-rndnu-neon-mlal.c | 130 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local 199 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld4r.c | 188 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 306 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 399 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-dup.c | 188 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 306 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 399 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld2r.c | 192 int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 312 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local 405 const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld2r() local
|