Home
last modified time | relevance | path

Searched refs:vprod1x01c0 (Results 1 – 25 of 128) sorted by relevance

123456

/external/XNNPACK/src/qc8-gemm/gen/
D2x8c4-minmax-fp32-neonv8-mlal-dup.c83 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
86 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
88 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
164 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
166 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
221 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
222 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
D2x8c4-minmax-fp32-neon-mlal-ld2r.c82 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
85 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
87 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
163 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
165 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
220 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
221 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
D2x8c4-minmax-fp32-neon-mlal-ld1r.c86 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local
89 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r()
91 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r()
169 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local
171 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r()
226 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local
227 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r()
D2x8c4-minmax-fp32-neonv8-mlal-ld2r.c83 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
86 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r()
88 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r()
164 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
166 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r()
221 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
222 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r()
D2x8c4-minmax-fp32-neonv8-mlal-ld1r.c87 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
90 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r()
92 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r()
170 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
172 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r()
227 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
228 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r()
D2x8c4-minmax-fp32-neon-mlal-dup.c82 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
85 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
87 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
163 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
165 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
220 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
221 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
/external/XNNPACK/src/qs8-igemm/gen/
D2x8c4-minmax-rndnu-neon-mlal-ld1r.c99 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local
102 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r()
104 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r()
182 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local
184 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r()
239 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r() local
240 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld1r()
D2x8c4-minmax-rndnu-neon-mlal-ld2r.c95 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local
98 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r()
100 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r()
176 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local
178 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r()
233 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r() local
234 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r()
D2x8c4-minmax-rndnu-neon-mlal-dup.c95 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local
98 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup()
100 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup()
176 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local
178 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup()
233 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local
234 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup()
D2x8c4-minmax-fp32-neon-mlal-ld2r.c95 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
98 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
100 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
176 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
178 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
233 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
234 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
D2x8c4-minmax-fp32-neonv8-mlal-dup.c96 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
99 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
101 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
177 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
179 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
234 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
235 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
D2x8c4-minmax-fp32-neon-mlal-ld1r.c99 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local
102 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r()
104 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r()
182 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local
184 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r()
239 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local
240 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r()
D2x8c4-minmax-fp32-neon-mlal-dup.c95 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
98 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
100 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
176 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
178 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
233 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
234 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
/external/XNNPACK/src/qc8-igemm/gen/
D2x8c4-minmax-fp32-neon-mlal-ld1r.c99 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local
102 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r()
104 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r()
182 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local
184 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r()
239 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r() local
240 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld1r()
D2x8c4-minmax-fp32-neonv8-mlal-dup.c96 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
99 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
101 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
177 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
179 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
234 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
235 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
D2x8c4-minmax-fp32-neonv8-mlal-ld2r.c96 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
99 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r()
101 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r()
177 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
179 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r()
234 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
235 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r()
D2x8c4-minmax-fp32-neonv8-mlal-ld1r.c100 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
103 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r()
105 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r()
183 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
185 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r()
240 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
241 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r()
D2x8c4-minmax-fp32-neon-mlal-dup.c95 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
98 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
100 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
176 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
178 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
233 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
234 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
D2x8c4-minmax-fp32-neon-mlal-ld2r.c95 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
98 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
100 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
176 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
178 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
233 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
234 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qc8_igemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
/external/XNNPACK/src/qs8-gemm/gen/
D2x8c4-minmax-fp32-neonv8-mlal-dup.c83 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
86 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
88 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
164 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
166 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
221 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup() local
222 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_dup()
D2x8c4-minmax-rndnu-neon-mlal-dup.c82 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local
85 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup()
87 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup()
163 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local
165 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup()
220 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup() local
221 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_dup()
D2x8c4-minmax-fp32-neon-mlal-dup.c82 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
85 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
87 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
163 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
165 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
220 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup() local
221 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_dup()
D2x8c4-minmax-fp32-neonv8-mlal-ld2r.c83 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
86 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r()
88 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r()
164 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
166 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r()
221 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r() local
222 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld2r()
D2x8c4-minmax-fp32-neon-mlal-ld2r.c82 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
85 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
87 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
163 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
165 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
220 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r() local
221 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neon_mlal_ld2r()
D2x8c4-minmax-fp32-neonv8-mlal-ld1r.c87 int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
90 vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r()
92 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r()
170 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
172 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r()
227 const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r() local
228 vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0); in xnn_qs8_gemm_minmax_fp32_ukernel_2x8c4__neonv8_mlal_ld1r()

123456