Home
last modified time | relevance | path

Searched refs:vprod1x8 (Results 1 – 18 of 18) sorted by relevance

/external/XNNPACK/src/qs8-igemm/gen/
D2x16c8-minmax-neon-mlal-padal.c179 int16x8_t vprod1x8 = vmull_s8(vb8x0, va1x0); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
181 vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
183 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
284 const int16x8_t vprod1x8 = vmull_s8(vb8, va1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
286 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c168 int16x8_t vprod1x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
170 vprod1x8 = vmlal_s8(vprod1x8, vget_high_s8(vb8), vget_high_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
172 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c229 int16x8_t vprod1x8 = vmull_s8(vb8x0, va1x0); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
232 vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
235 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
375 const int16x8_t vprod1x8 = vmull_s8(vb8, va1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
378 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c217 int16x8_t vprod1x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
220 vprod1x8 = vmlal_s8(vprod1x8, vget_high_s8(vb8), vget_high_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
223 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c279 int16x8_t vprod1x8 = vmull_s8(vb8x0, va1x0); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
283 vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
287 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
466 const int16x8_t vprod1x8 = vmull_s8(vb8, va1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
470 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
D2x16c8-minmax-neon-mull-padal.c145 const int16x8_t vprod1x8 = vmull_s8(vb8, va1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
147 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c266 int16x8_t vprod1x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
270 vprod1x8 = vmlal_s8(vprod1x8, vget_high_s8(vb8), vget_high_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
274 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c186 const int16x8_t vprod1x8 = vmull_s8(vb8, va1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
189 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D4x16c8-minmax-neon-mull-padal.c227 const int16x8_t vprod1x8 = vmull_s8(vb8, va1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
231 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
/external/XNNPACK/src/qs8-gemm/gen/
D2x16c8-minmax-neon-mlal-padal.c166 int16x8_t vprod1x8 = vmull_s8(vb8x0, va1x0); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
168 vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
170 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
271 const int16x8_t vprod1x8 = vmull_s8(vb8, va1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
273 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c155 int16x8_t vprod1x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
157 vprod1x8 = vmlal_s8(vprod1x8, vget_high_s8(vb8), vget_high_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
159 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c214 int16x8_t vprod1x8 = vmull_s8(vb8x0, va1x0); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
217 vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
220 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
360 const int16x8_t vprod1x8 = vmull_s8(vb8, va1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
363 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c202 int16x8_t vprod1x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
205 vprod1x8 = vmlal_s8(vprod1x8, vget_high_s8(vb8), vget_high_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
208 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c262 int16x8_t vprod1x8 = vmull_s8(vb8x0, va1x0); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
266 vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
270 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
449 const int16x8_t vprod1x8 = vmull_s8(vb8, va1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
453 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
D2x16c8-minmax-neon-mull-padal.c132 const int16x8_t vprod1x8 = vmull_s8(vb8, va1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
134 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c249 int16x8_t vprod1x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
253 vprod1x8 = vmlal_s8(vprod1x8, vget_high_s8(vb8), vget_high_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
257 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c171 const int16x8_t vprod1x8 = vmull_s8(vb8, va1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
174 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D4x16c8-minmax-neon-mull-padal.c210 const int16x8_t vprod1x8 = vmull_s8(vb8, va1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
214 vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()