Home
last modified time | relevance | path

Searched refs:vprod1x14 (Results 1 – 18 of 18) sorted by relevance

/external/XNNPACK/src/qs8-igemm/gen/
D2x16c8-minmax-neon-mlal-padal.c221 int16x8_t vprod1x14 = vmull_s8(vb14x0, va1x0); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
223 vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
225 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
314 const int16x8_t vprod1x14 = vmull_s8(vb14, va1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
316 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c204 int16x8_t vprod1x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
206 vprod1x14 = vmlal_s8(vprod1x14, vget_high_s8(vb14), vget_high_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
208 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c289 int16x8_t vprod1x14 = vmull_s8(vb14x0, va1x0); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
292 vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
295 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
417 const int16x8_t vprod1x14 = vmull_s8(vb14, va1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
420 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c271 int16x8_t vprod1x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
274 vprod1x14 = vmlal_s8(vprod1x14, vget_high_s8(vb14), vget_high_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
277 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c357 int16x8_t vprod1x14 = vmull_s8(vb14x0, va1x0); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
361 vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
365 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
520 const int16x8_t vprod1x14 = vmull_s8(vb14, va1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
524 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
D2x16c8-minmax-neon-mull-padal.c175 const int16x8_t vprod1x14 = vmull_s8(vb14, va1); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
177 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c338 int16x8_t vprod1x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
342 vprod1x14 = vmlal_s8(vprod1x14, vget_high_s8(vb14), vget_high_s8(va1)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
346 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c228 const int16x8_t vprod1x14 = vmull_s8(vb14, va1); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
231 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D4x16c8-minmax-neon-mull-padal.c281 const int16x8_t vprod1x14 = vmull_s8(vb14, va1); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
285 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
/external/XNNPACK/src/qs8-gemm/gen/
D2x16c8-minmax-neon-mlal-padal.c208 int16x8_t vprod1x14 = vmull_s8(vb14x0, va1x0); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
210 vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
212 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
301 const int16x8_t vprod1x14 = vmull_s8(vb14, va1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
303 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D2x16c16-minmax-neon-mlal-padal.c191 int16x8_t vprod1x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
193 vprod1x14 = vmlal_s8(vprod1x14, vget_high_s8(vb14), vget_high_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
195 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c274 int16x8_t vprod1x14 = vmull_s8(vb14x0, va1x0); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
277 vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
280 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
402 const int16x8_t vprod1x14 = vmull_s8(vb14, va1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
405 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c256 int16x8_t vprod1x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
259 vprod1x14 = vmlal_s8(vprod1x14, vget_high_s8(vb14), vget_high_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
262 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c340 int16x8_t vprod1x14 = vmull_s8(vb14x0, va1x0); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
344 vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
348 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
503 const int16x8_t vprod1x14 = vmull_s8(vb14, va1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
507 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
D2x16c8-minmax-neon-mull-padal.c162 const int16x8_t vprod1x14 = vmull_s8(vb14, va1); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
164 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c321 int16x8_t vprod1x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
325 vprod1x14 = vmlal_s8(vprod1x14, vget_high_s8(vb14), vget_high_s8(va1)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
329 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c213 const int16x8_t vprod1x14 = vmull_s8(vb14, va1); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
216 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D4x16c8-minmax-neon-mull-padal.c264 const int16x8_t vprod1x14 = vmull_s8(vb14, va1); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
268 vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()