/external/XNNPACK/src/qs8-gemm/gen/ |
D | 1x16c2s4-minmax-rndnu-neon-mlal.c | 137 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local 197 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld2r.c | 144 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 211 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-dup.c | 142 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 208 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld1r.c | 148 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 217 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld4r.c | 142 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 208 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local
|
D | 2x16c2s4-minmax-rndnu-neon-mlal.c | 200 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local 297 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local
|
D | 1x16c2s4-minmax-rndnu-neon-mull.c | 103 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull() local
|
D | 3x16c2s4-minmax-rndnu-neon-mlal.c | 263 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local 397 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local
|
D | 2x16c2s4-minmax-rndnu-neon-mull.c | 147 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-dup.c | 207 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local 311 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_dup() local
|
D | 2x16c2-minmax-rndnu-neon-mlal-ld4r.c | 207 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local 311 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c2__neon_mlal_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld2r.c | 110 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-dup.c | 109 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_dup() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld4r.c | 109 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 1x16c2s4-minmax-rndnu-neon-mlal.c | 148 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local 208 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mlal() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld2r.c | 156 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local 223 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld2r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld1r.c | 160 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local 229 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld1r() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-dup.c | 154 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local 220 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup() local
|
D | 1x16c2-minmax-rndnu-neon-mlal-ld4r.c | 154 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local 220 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_ld4r() local
|
D | 2x16c2s4-minmax-rndnu-neon-mlal.c | 213 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local 310 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mlal() local
|
D | 1x16c2s4-minmax-rndnu-neon-mull.c | 114 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull() local
|
D | 3x16c2s4-minmax-rndnu-neon-mlal.c | 278 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local 412 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2s4__neon_mlal() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld4r.c | 120 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld4r() local
|
D | 1x16c2-minmax-rndnu-neon-mull-ld2r.c | 121 const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld2r() local
|
D | 2x16c2s4-minmax-rndnu-neon-mull.c | 160 int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2s4__neon_mull() local
|