/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x8c4-minmax-rndnu-neon-mlal-ld2r.c | 89 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local 167 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local 222 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local
|
D | 2x8c4-minmax-rndnu-neon-mlal-dup.c | 89 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local 167 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local 222 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neon-mlal-ld2r.c | 89 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 167 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 222 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
|
D | 2x8c4-minmax-fp32-neon-mlal-dup.c | 89 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 167 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 222 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-dup.c | 90 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 168 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 223 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-ld2r.c | 90 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 168 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 223 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-ld1r.c | 94 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local 174 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local 229 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
|
D | 2x8c4-minmax-rndnu-neon-mlal-ld1r.c | 93 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local 173 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local 228 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local
|
D | 2x8c4-minmax-fp32-neon-mlal-ld1r.c | 93 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local 173 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local 228 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local
|
/external/XNNPACK/src/qc8-gemm/gen/ |
D | 2x8c4-minmax-fp32-neon-mlal-ld2r.c | 89 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 167 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 222 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-ld2r.c | 90 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 168 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 223 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-dup.c | 90 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 168 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 223 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-ld1r.c | 94 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local 174 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local 229 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
|
D | 2x8c4-minmax-fp32-neon-mlal-dup.c | 89 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 167 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 222 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 2x8c4-minmax-fp32-neon-mlal-ld2r.c | 102 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 180 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 235 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
|
D | 2x8c4-minmax-rndnu-neon-mlal-ld1r.c | 106 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local 186 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local 241 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-dup.c | 103 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 181 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 236 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
|
D | 2x8c4-minmax-rndnu-neon-mlal-dup.c | 102 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local 180 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local 235 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-ld2r.c | 103 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 181 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 236 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
|
D | 2x8c4-minmax-fp32-neon-mlal-dup.c | 102 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 180 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 235 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
|
D | 2x8c4-minmax-rndnu-neon-mlal-ld2r.c | 102 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local 180 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local 235 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local
|
/external/XNNPACK/src/qc8-igemm/gen/ |
D | 2x8c4-minmax-fp32-neon-mlal-dup.c | 102 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 180 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 235 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-dup.c | 103 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 181 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 236 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-ld2r.c | 103 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 181 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 236 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
|
D | 2x8c4-minmax-fp32-neon-mlal-ld2r.c | 102 int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 180 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 235 const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
|