/external/XNNPACK/src/qs8-gemm/gen/ |
D | 2x8c4-minmax-rndnu-neon-mlal-ld2r.c | 103 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local 175 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local 226 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local
|
D | 2x8c4-minmax-rndnu-neon-mlal-dup.c | 103 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local 175 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local 226 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neon-mlal-ld2r.c | 103 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 175 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 226 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
|
D | 2x8c4-minmax-fp32-neon-mlal-dup.c | 103 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 175 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 226 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-dup.c | 104 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 176 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 227 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-ld2r.c | 104 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 176 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 227 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-ld1r.c | 108 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local 182 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local 233 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
|
D | 2x8c4-minmax-rndnu-neon-mlal-ld1r.c | 107 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local 181 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local 232 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local
|
D | 2x8c4-minmax-fp32-neon-mlal-ld1r.c | 107 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local 181 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local 232 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, 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 | 103 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 175 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 226 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-ld2r.c | 104 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 176 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 227 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-dup.c | 104 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 176 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 227 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-ld1r.c | 108 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local 182 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local 233 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
|
D | 2x8c4-minmax-fp32-neon-mlal-dup.c | 103 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 175 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 226 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, 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 | 116 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 188 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 239 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
|
D | 2x8c4-minmax-rndnu-neon-mlal-ld1r.c | 120 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local 194 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local 245 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-dup.c | 117 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 189 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 240 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
|
D | 2x8c4-minmax-rndnu-neon-mlal-dup.c | 116 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local 188 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local 239 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-ld2r.c | 117 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 189 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 240 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
|
D | 2x8c4-minmax-fp32-neon-mlal-dup.c | 116 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 188 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 239 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
|
D | 2x8c4-minmax-rndnu-neon-mlal-ld2r.c | 116 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local 188 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local 239 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, 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 | 116 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 188 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local 239 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-dup.c | 117 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 189 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local 240 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
|
D | 2x8c4-minmax-fp32-neonv8-mlal-ld2r.c | 117 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 189 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local 240 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
|
D | 2x8c4-minmax-fp32-neon-mlal-ld2r.c | 116 int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 188 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local 239 const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
|