/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x16c4-minmax-rndnu-neon-mlal-ld2r.c | 177 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 350 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 477 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-dup.c | 177 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 350 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 477 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld1r.c | 183 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 359 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 486 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
|
D | 3x16c4-minmax-rndnu-neon-mull-dup.c | 148 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local 275 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local
|
D | 3x16c4-minmax-rndnu-neon-mull-ld2r.c | 148 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local 275 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mull-ld1r.c | 151 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local 278 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local
|
D | 4x16c4-minmax-rndnu-neon-mlal-ld1r.c | 221 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local 444 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local 593 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
|
D | 4x16c4-minmax-rndnu-neon-mlal-dup.c | 213 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local 432 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local 581 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
|
D | 4x16c4-minmax-rndnu-neon-mlal-ld2r.c | 213 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local 432 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local 581 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
|
D | 3x16c4s2-minmax-rndnu-neon-mlal.c | 169 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local 336 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
|
D | 4x16c4-minmax-rndnu-neon-mull-ld2r.c | 176 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local 325 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
|
D | 4x16c4-minmax-rndnu-neon-mull-dup.c | 176 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local 325 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x16c4-minmax-rndnu-neon-mlal-ld1r.c | 198 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 374 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local 501 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld1r() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-dup.c | 192 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 365 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local 492 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_dup() local
|
D | 3x16c4-minmax-rndnu-neon-mlal-ld2r.c | 192 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 365 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local 492 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mlal_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mull-dup.c | 163 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local 290 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_dup() local
|
D | 3x16c4-minmax-rndnu-neon-mull-ld2r.c | 163 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local 290 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld2r() local
|
D | 3x16c4-minmax-rndnu-neon-mull-ld1r.c | 166 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local 293 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r() local
|
D | 4x16c4-minmax-rndnu-neon-mlal-dup.c | 230 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local 449 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local 598 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_dup() local
|
D | 4x16c4-minmax-rndnu-neon-mlal-ld2r.c | 230 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local 449 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local 598 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld2r() local
|
D | 4x16c4-minmax-rndnu-neon-mlal-ld1r.c | 238 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local 461 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local 610 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mlal_ld1r() local
|
D | 3x16c4s2-minmax-rndnu-neon-mlal.c | 184 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local 351 int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal() local
|
D | 4x16c4-minmax-rndnu-neon-mull-dup.c | 193 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local 342 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_dup() local
|
D | 4x16c4-minmax-rndnu-neon-mull-ld2r.c | 193 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local 342 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld2r() local
|
D | 4x16c4-minmax-rndnu-neon-mull-ld1r.c | 197 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local 346 const int16x8_t vprod2xCDc0 = vmull_s8(vbCDc0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neon_mull_ld1r() local
|