| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 1x8c2-minmax-fp32-neonv8-mlal-ld4r.c | 79 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r() local
81 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
82 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
131 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r() local
132 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
168 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r() local
169 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
|
| D | 1x8c2-minmax-fp32-neonv8-mlal-ld1r.c | 85 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
87 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
88 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
140 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
141 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
177 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
178 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
|
| D | 1x8c2-minmax-fp32-neon-mlal-ld4r.c | 78 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
80 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
81 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
130 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
131 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
167 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
168 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
|
| D | 1x8c2-minmax-rndnu-neon-mlal-dup.c | 78 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup() local
80 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup()
81 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup()
130 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup() local
131 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup()
167 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup() local
168 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup()
|
| D | 1x8c2-minmax-rndnu-neon-mlal-ld2r.c | 80 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r() local
82 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r()
83 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r()
133 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r() local
134 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r()
170 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r() local
171 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r()
|
| D | 1x8c2-minmax-fp32-neon-mlal-dup.c | 78 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup() local
80 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup()
81 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup()
130 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup() local
131 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup()
167 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup() local
168 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup()
|
| D | 1x8c2-minmax-fp32-neonv8-mlal-dup.c | 79 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup() local
81 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
82 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
131 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup() local
132 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
168 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup() local
169 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
|
| /external/XNNPACK/src/qc8-igemm/gen/ |
| D | 1x8c2-minmax-fp32-neonv8-mlal-ld4r.c | 91 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r() local
93 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
94 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
143 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r() local
144 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
180 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r() local
181 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
|
| D | 1x8c2-minmax-fp32-neonv8-mlal-ld1r.c | 97 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
99 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
100 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
152 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
153 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
189 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
190 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
|
| D | 1x8c2-minmax-fp32-neonv8-mlal-dup.c | 91 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup() local
93 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
94 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
143 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup() local
144 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
180 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup() local
181 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
|
| D | 1x8c2-minmax-fp32-neon-mlal-ld1r.c | 96 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r() local
98 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r()
99 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r()
151 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r() local
152 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r()
188 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r() local
189 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r()
|
| D | 1x8c2-minmax-fp32-neonv8-mlal-ld2r.c | 93 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r() local
95 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r()
96 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r()
146 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r() local
147 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r()
183 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r() local
184 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r()
|
| D | 1x8c2-minmax-fp32-neon-mlal-ld4r.c | 90 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
92 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
93 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
142 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
143 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
179 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
180 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
|
| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 1x8c2-minmax-fp32-neon-mlal-ld2r.c | 92 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r() local
94 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r()
95 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r()
145 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r() local
146 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r()
182 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r() local
183 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r()
|
| D | 1x8c2-minmax-fp32-neonv8-mlal-ld1r.c | 97 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
99 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
100 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
152 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
153 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
189 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
190 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
|
| D | 1x8c2-minmax-fp32-neonv8-mlal-dup.c | 91 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup() local
93 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
94 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
143 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup() local
144 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
180 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup() local
181 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_dup()
|
| D | 1x8c2-minmax-fp32-neonv8-mlal-ld4r.c | 91 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r() local
93 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
94 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
143 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r() local
144 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
180 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r() local
181 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld4r()
|
| D | 1x8c2-minmax-fp32-neon-mlal-dup.c | 90 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup() local
92 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup()
93 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup()
142 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup() local
143 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup()
179 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup() local
180 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_dup()
|
| D | 1x8c2-minmax-fp32-neonv8-mlal-ld2r.c | 93 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r() local
95 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r()
96 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r()
146 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r() local
147 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r()
183 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r() local
184 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r()
|
| D | 1x8c2-minmax-fp32-neon-mlal-ld4r.c | 90 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
92 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
93 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
142 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
143 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
179 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
180 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
|
| D | 1x8c2-minmax-rndnu-neon-mlal-ld2r.c | 92 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r() local
94 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r()
95 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r()
145 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r() local
146 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r()
182 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r() local
183 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r()
|
| /external/XNNPACK/src/qc8-gemm/gen/ |
| D | 1x8c2-minmax-fp32-neon-mlal-ld4r.c | 78 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
80 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
81 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
130 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
131 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
167 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r() local
168 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld4r()
|
| D | 1x8c2-minmax-fp32-neonv8-mlal-ld1r.c | 85 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
87 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
88 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
140 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
141 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
177 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r() local
178 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld1r()
|
| D | 1x8c2-minmax-fp32-neon-mlal-ld1r.c | 84 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r() local
86 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r()
87 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r()
139 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r() local
140 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r()
176 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r() local
177 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld1r()
|
| D | 1x8c2-minmax-fp32-neon-mlal-ld2r.c | 80 int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r() local
82 vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r()
83 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r()
133 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r() local
134 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r()
170 const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r() local
171 vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1); in xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r()
|