Home
last modified time | relevance | path

Searched refs:vprod2x5 (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D3x8c8-minmax-neon-mlal-padal.c153 int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
156 vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
159 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
228 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
231 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D3x8c16-minmax-neon-mlal-padal.c144 int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
147 vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
150 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c184 int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
188 vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
192 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
279 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
283 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c174 int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
178 vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
182 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c185 int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
188 vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
191 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
340 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
343 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c176 int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
179 vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
182 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c224 int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
228 vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
232 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
423 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
427 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c127 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
130 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
D4x8c8-minmax-neon-mull-padal.c152 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
156 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c214 int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
218 vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
222 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c151 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
154 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D4x16c8-minmax-neon-mull-padal.c184 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
188 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D3x8c8-minmax-neon-mlal-padal.c168 int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
171 vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
174 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
243 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
246 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D3x8c16-minmax-neon-mlal-padal.c159 int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
162 vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
165 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mlal-padal.c201 int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
205 vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
209 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
296 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
300 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c191 int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
195 vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
199 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c200 int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
203 vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
206 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
355 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
358 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c191 int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
194 vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
197 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c241 int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
245 vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
249 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
440 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
444 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c142 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
145 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
D4x8c8-minmax-neon-mull-padal.c169 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
173 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c231 int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
235 vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
239 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c166 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
169 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D4x16c8-minmax-neon-mull-padal.c201 const int16x8_t vprod2x5 = vmull_s8(vb5, va2); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
205 vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()