/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x8c4-minmax-rndnu-neon-mlal-ld2r.c | 127 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local 224 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local 295 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
|
D | 3x8c4-minmax-rndnu-neon-mlal-dup.c | 127 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local 224 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local 295 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
|
D | 3x8c4-minmax-rndnu-neon-mlal-ld1r.c | 133 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local 233 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local 304 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
|
D | 3x8c4-minmax-rndnu-neon-mull-dup.c | 110 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local 181 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local
|
D | 3x8c4-minmax-rndnu-neon-mull-ld2r.c | 110 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local 181 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local
|
D | 3x8c4-minmax-rndnu-neon-mull-ld1r.c | 113 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local 184 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local
|
D | 4x8c4-minmax-rndnu-neon-mlal-ld2r.c | 150 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local 272 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local 355 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
|
D | 4x8c4-minmax-rndnu-neon-mlal-dup.c | 150 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local 272 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local 355 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
|
D | 3x8c4s2-minmax-rndnu-neon-mlal.c | 119 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local 210 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local
|
D | 4x8c4-minmax-rndnu-neon-mull-ld2r.c | 128 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r() local 211 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r() local
|
D | 4x8c4-minmax-rndnu-neon-mull-dup.c | 128 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup() local 211 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup() local
|
D | 4x8c4-minmax-rndnu-neon-mlal-ld1r.c | 158 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local 284 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local 367 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x8c4-minmax-rndnu-neon-mlal-dup.c | 142 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local 239 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local 310 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
|
D | 3x8c4-minmax-rndnu-neon-mlal-ld2r.c | 142 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local 239 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local 310 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
|
D | 3x8c4-minmax-rndnu-neon-mlal-ld1r.c | 148 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local 248 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local 319 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
|
D | 3x8c4-minmax-rndnu-neon-mull-ld1r.c | 128 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local 199 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local
|
D | 3x8c4-minmax-rndnu-neon-mull-dup.c | 125 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local 196 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local
|
D | 3x8c4-minmax-rndnu-neon-mull-ld2r.c | 125 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local 196 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local
|
D | 4x8c4-minmax-rndnu-neon-mlal-dup.c | 167 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local 289 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local 372 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_dup() local
|
D | 4x8c4-minmax-rndnu-neon-mlal-ld2r.c | 167 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local 289 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local 372 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
|
D | 3x8c4s2-minmax-rndnu-neon-mlal.c | 134 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local 225 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local
|
D | 4x8c4-minmax-rndnu-neon-mlal-ld1r.c | 175 int16x8_t vprod2x67c0 = vmull_s8(vb67c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local 301 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local 384 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
|
D | 4x8c4-minmax-rndnu-neon-mull-ld2r.c | 145 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r() local 228 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r() local
|
D | 4x8c4-minmax-rndnu-neon-mull-dup.c | 145 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup() local 228 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup() local
|
D | 4x8c4-minmax-rndnu-neon-mull-ld1r.c | 149 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r() local 232 const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r() local
|