Home
last modified time | relevance | path

Searched refs:vprod2x6 (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D3x8c8-minmax-neon-mlal-padal.c163 int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
166 vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
169 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
235 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
238 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c197 int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
201 vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
205 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
288 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
292 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D3x8c16-minmax-neon-mlal-padal.c153 int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
156 vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
159 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c186 int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
190 vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
194 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c195 int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
198 vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
201 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
347 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
350 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c185 int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
188 vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
191 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c237 int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
241 vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
245 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
432 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
436 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c134 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
137 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c226 int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
230 vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
234 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c161 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
165 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
D3x16c8-minmax-neon-mull-padal.c158 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
161 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D4x16c8-minmax-neon-mull-padal.c193 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
197 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D3x8c8-minmax-neon-mlal-padal.c178 int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
181 vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
184 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
250 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
253 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c214 int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
218 vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
222 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
305 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
309 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D3x8c16-minmax-neon-mlal-padal.c168 int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
171 vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
174 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c203 int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
207 vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
211 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c210 int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
213 vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
216 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
362 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
365 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c200 int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
203 vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
206 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c254 int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
258 vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
262 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
449 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
453 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c149 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
152 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c243 int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
247 vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
251 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c178 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
182 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
D3x16c8-minmax-neon-mull-padal.c173 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
176 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D4x16c8-minmax-neon-mull-padal.c210 const int16x8_t vprod2x6 = vmull_s8(vb6, va2); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
214 vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()