/external/XNNPACK/src/qs8-gemm/gen/ |
D | 3x8c2-minmax-rndnu-neon-mlal-dup.c | 99 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup() local 218 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup() local 295 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup() local
|
D | 3x8c2-minmax-rndnu-neon-mlal-ld4r.c | 99 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r() local 218 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r() local 295 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r() local
|
D | 3x8c2s4-minmax-rndnu-neon-mlal.c | 93 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal() local 204 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal() local
|
D | 3x8c2-minmax-rndnu-neon-mull-ld4r.c | 91 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r() local 168 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r() local
|
D | 3x8c2-minmax-rndnu-neon-mull-dup.c | 91 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup() local 168 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup() local
|
D | 3x8c2-minmax-rndnu-neon-mull-ld2r.c | 94 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld2r() local 171 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld2r() local
|
D | 3x8c2-minmax-rndnu-neon-mlal-ld2r.c | 105 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r() local 227 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r() local 304 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r() local
|
D | 3x8c2-minmax-rndnu-neon-mull-ld1r.c | 100 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld1r() local 177 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld1r() local
|
D | 3x8c2-minmax-rndnu-neon-mlal-ld1r.c | 117 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r() local 245 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r() local 322 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r() local
|
D | 4x8c2-minmax-rndnu-neon-mlal-ld4r.c | 114 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local 264 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local 359 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
|
D | 4x8c2-minmax-rndnu-neon-mlal-dup.c | 114 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local 264 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local 359 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
|
D | 4x8c2s4-minmax-rndnu-neon-mlal.c | 106 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local 247 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
|
D | 4x8c2-minmax-rndnu-neon-mull-ld1r.c | 115 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local 210 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 3x8c2-minmax-rndnu-neon-mlal-ld4r.c | 115 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r() local 234 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r() local 311 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r() local
|
D | 3x8c2-minmax-rndnu-neon-mlal-dup.c | 115 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup() local 234 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup() local 311 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup() local
|
D | 3x8c2s4-minmax-rndnu-neon-mlal.c | 108 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal() local 219 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2s4__neon_mlal() local
|
D | 3x8c2-minmax-rndnu-neon-mull-dup.c | 106 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup() local 183 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_dup() local
|
D | 3x8c2-minmax-rndnu-neon-mull-ld4r.c | 106 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r() local 183 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r() local
|
D | 3x8c2-minmax-rndnu-neon-mlal-ld2r.c | 121 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r() local 243 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r() local 320 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r() local
|
D | 3x8c2-minmax-rndnu-neon-mull-ld2r.c | 109 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld2r() local 186 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld2r() local
|
D | 3x8c2-minmax-rndnu-neon-mull-ld1r.c | 115 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld1r() local 192 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld1r() local
|
D | 3x8c2-minmax-rndnu-neon-mlal-ld1r.c | 133 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r() local 261 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r() local 338 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld1r() local
|
D | 4x8c2-minmax-rndnu-neon-mlal-dup.c | 132 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local 282 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local 377 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup() local
|
D | 4x8c2-minmax-rndnu-neon-mlal-ld4r.c | 132 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local 282 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local 377 const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r() local
|
D | 4x8c2s4-minmax-rndnu-neon-mlal.c | 123 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local 264 int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2s4__neon_mlal() local
|