Home
last modified time | relevance | path

Searched refs:vprod1x7 (Results 1 – 25 of 36) sorted by relevance

12

/external/XNNPACK/src/qs8-gemm/gen/
D2x8c8-minmax-neon-mlal-padal.c135 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
137 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
139 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
186 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
188 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D3x8c8-minmax-neon-mlal-padal.c172 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
175 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
178 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
241 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
244 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D2x8c16-minmax-neon-mlal-padal.c125 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
127 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
129 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D3x8c16-minmax-neon-mlal-padal.c161 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
164 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
167 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c209 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
213 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
217 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
296 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
300 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D2x16c8-minmax-neon-mlal-padal.c159 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
161 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
163 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
266 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
268 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c149 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
151 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
153 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c197 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
201 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
205 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D2x8c8-minmax-neon-mull-padal.c111 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal() local
113 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_2x8c8__neon_mull_padal()
D3x16c8-minmax-neon-mlal-padal.c204 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
207 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
210 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
353 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
356 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c193 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
196 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
199 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c249 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
253 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
257 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
440 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
444 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D2x8c8-minmax-neon-mlal-padal.c148 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
150 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
152 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
199 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal() local
201 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mlal_padal()
D2x8c16-minmax-neon-mlal-padal.c138 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal() local
140 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
142 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_2x8c16__neon_mlal_padal()
D3x8c8-minmax-neon-mlal-padal.c187 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
190 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
193 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
256 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
259 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D3x8c16-minmax-neon-mlal-padal.c176 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
179 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
182 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D2x16c8-minmax-neon-mlal-padal.c172 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
174 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
176 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
279 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
281 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c226 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
230 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
234 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
313 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
317 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c214 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
218 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
222 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c162 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
164 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
166 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D2x8c8-minmax-neon-mull-padal.c124 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal() local
126 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_2x8c8__neon_mull_padal()
D3x16c8-minmax-neon-mlal-padal.c219 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
222 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
225 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
368 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
371 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c208 int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
211 vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
214 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c266 int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
270 vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
274 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
457 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
461 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c155 const int16x8_t vprod1x7 = vmull_s8(vb7, va1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
158 vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()

12