/external/XNNPACK/src/qs8-gemm/gen/ |
D | 4x8c2-minmax-rndnu-neon-mlal-ld4r.c | 115 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local 265 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local 364 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
|
D | 4x8c2-minmax-rndnu-neon-mlal-dup.c | 115 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local 265 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local 364 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
|
D | 4x8c2s4-minmax-rndnu-neon-mlal.c | 107 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local 248 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
|
D | 4x8c2-minmax-rndnu-neon-mull-ld1r.c | 116 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local 215 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local
|
D | 4x8c2-minmax-rndnu-neon-mull-ld4r.c | 104 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local 203 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local
|
D | 4x8c2-minmax-rndnu-neon-mull-dup.c | 104 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup() local 203 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup() local
|
D | 4x8c2-minmax-rndnu-neon-mull-ld2r.c | 108 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r() local 207 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r() local
|
D | 4x8c2-minmax-rndnu-neon-mlal-ld2r.c | 123 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local 277 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local 376 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
|
D | 4x8c2-minmax-rndnu-neon-mlal-ld1r.c | 139 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local 301 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local 400 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
|
D | 4x8c2s4-minmax-rndnu-neon-mull.c | 98 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull() local
|
D | 4x16c2-minmax-rndnu-neon-mull-dup.c | 120 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local 297 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld4r.c | 120 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local 297 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-dup.c | 131 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local 393 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local 570 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 4x8c2-minmax-rndnu-neon-mlal-dup.c | 133 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local 283 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local 382 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
|
D | 4x8c2-minmax-rndnu-neon-mlal-ld4r.c | 133 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local 283 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local 382 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
|
D | 4x8c2s4-minmax-rndnu-neon-mlal.c | 124 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local 265 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
|
D | 4x8c2-minmax-rndnu-neon-mull-ld4r.c | 121 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local 220 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r() local
|
D | 4x8c2-minmax-rndnu-neon-mull-ld2r.c | 125 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r() local 224 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r() local
|
D | 4x8c2-minmax-rndnu-neon-mull-ld1r.c | 133 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local 232 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local
|
D | 4x8c2-minmax-rndnu-neon-mull-dup.c | 121 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup() local 220 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup() local
|
D | 4x8c2-minmax-rndnu-neon-mlal-ld2r.c | 141 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local 295 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local 394 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld2r() local
|
D | 4x8c2-minmax-rndnu-neon-mlal-ld1r.c | 157 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local 319 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local 418 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld1r() local
|
D | 4x8c2s4-minmax-rndnu-neon-mull.c | 115 int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0x0, va3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull() local
|
D | 4x16c2-minmax-rndnu-neon-mull-dup.c | 137 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local 314 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup() local
|
D | 4x16c2-minmax-rndnu-neon-mull-ld4r.c | 137 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local 314 const int16x8_t vprod3x4567c0 = vmull_s8(vb4567c0, va3c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mull_ld4r() local
|