Home
last modified time | relevance | path

Searched refs:vprod0x5 (Results 1 – 25 of 48) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D1x8c8-minmax-neon-mlal-padal.c89 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
90 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
91 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
124 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
125 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal()
D1x8c16-minmax-neon-mlal-padal.c82 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
83 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
84 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c120 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
122 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
124 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
175 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
177 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D1x16c8-minmax-neon-mlal-padal.c105 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
106 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
107 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
172 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
173 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D3x8c8-minmax-neon-mlal-padal.c151 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
154 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
157 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
226 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
229 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D2x8c16-minmax-neon-mlal-padal.c112 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
114 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
116 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D1x16c16-minmax-neon-mlal-padal.c98 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
99 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
100 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D1x8c8-minmax-neon-mull-padal.c75 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal() local
76 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c142 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
145 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
148 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c182 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
186 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
190 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
277 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
281 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D2x16c8-minmax-neon-mlal-padal.c144 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
146 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
148 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
255 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
257 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c136 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
138 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
140 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c172 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
176 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
180 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x8c8-minmax-neon-mlal-padal.c100 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
101 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
102 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
135 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal() local
136 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mlal_padal()
D1x8c16-minmax-neon-mlal-padal.c93 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal() local
94 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
95 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mlal-padal.c133 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
135 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
137 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
188 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
190 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D1x16c8-minmax-neon-mlal-padal.c116 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
117 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
118 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
183 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
184 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D2x8c16-minmax-neon-mlal-padal.c125 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
127 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
129 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D1x16c16-minmax-neon-mlal-padal.c109 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
110 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
111 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D3x8c8-minmax-neon-mlal-padal.c166 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
169 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
172 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
241 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
244 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D1x8c8-minmax-neon-mull-padal.c86 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal() local
87 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_1x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c157 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
160 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
163 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D2x16c8-minmax-neon-mlal-padal.c157 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
159 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
161 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
268 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
270 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c199 int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
203 vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
207 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
294 const int16x8_t vprod0x5 = vmull_s8(vb5, va0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
298 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c189 int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
193 vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
197 vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()

12