/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x16c2s4-minmax-rndnu-neon-mlal.c | 255 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local 393 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local
|
D | 4x16c2s4-minmax-rndnu-neon-mlal.c | 315 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local 491 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local
|
D | 3x16c2s4-minmax-rndnu-neon-mull.c | 187 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull() local
|
D | 3x16c2-minmax-rndnu-neon-mlal-dup.c | 264 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local 410 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
|
D | 3x16c2-minmax-rndnu-neon-mlal-ld4r.c | 264 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local 410 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
|
D | 3x16c2-minmax-rndnu-neon-mlal-ld2r.c | 270 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local 419 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
|
D | 3x16c2-minmax-rndnu-neon-mlal-ld1r.c | 282 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local 437 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
|
D | 4x16c2s4-minmax-rndnu-neon-mull.c | 229 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2s4__neon_mull() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-dup.c | 326 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local 511 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld2r.c | 334 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local 523 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld4r.c | 326 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local 511 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
|
D | 3x16c2-minmax-rndnu-neon-mull-dup.c | 195 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local
|
D | 3x16c2-minmax-rndnu-neon-mull-ld2r.c | 198 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x16c2s4-minmax-rndnu-neon-mlal.c | 270 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local 408 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local
|
D | 4x16c2s4-minmax-rndnu-neon-mlal.c | 332 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local 508 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal() local
|
D | 3x16c2s4-minmax-rndnu-neon-mull.c | 202 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull() local
|
D | 3x16c2-minmax-rndnu-neon-mlal-ld4r.c | 280 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local 426 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld4r() local
|
D | 3x16c2-minmax-rndnu-neon-mlal-dup.c | 280 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local 426 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_dup() local
|
D | 3x16c2-minmax-rndnu-neon-mlal-ld1r.c | 298 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local 453 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld1r() local
|
D | 3x16c2-minmax-rndnu-neon-mlal-ld2r.c | 286 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local 435 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mlal_ld2r() local
|
D | 4x16c2s4-minmax-rndnu-neon-mull.c | 246 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mull() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld2r.c | 352 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local 541 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld2r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-ld4r.c | 344 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local 529 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld4r() local
|
D | 4x16c2-minmax-rndnu-neon-mlal-dup.c | 344 int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3x0, va2c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local 529 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_dup() local
|
D | 3x16c2-minmax-rndnu-neon-mull-dup.c | 210 const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup() local
|