Home
last modified time | relevance | path

Searched refs:vacc2x7 (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D3x8c8-minmax-neon-mlal-padal.c79 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
179 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
245 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
262 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
308 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D3x8c8-minmax-neon-mull-padal.c79 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal() local
144 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
161 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
207 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c16-minmax-neon-mlal-padal.c79 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
168 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
185 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
231 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c85 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal() local
174 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
192 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
244 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mull_padal()
D4x8c8-minmax-neon-mlal-padal.c85 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
218 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
301 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
319 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
371 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c85 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
206 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
224 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
276 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c95 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
168 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
249 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
333 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D3x16c8-minmax-neon-mlal-padal.c95 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
211 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
357 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
438 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
522 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c95 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
200 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
297 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
381 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c101 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
206 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
304 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
400 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
D4x16c8-minmax-neon-mlal-padal.c101 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
258 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
445 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
543 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
639 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
D4x16c16-minmax-neon-mlal-padal.c101 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
246 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
368 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
464 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D3x8c8-minmax-neon-mull-padal.c78 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal() local
159 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
179 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
225 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mull_padal()
D3x8c8-minmax-neon-mlal-padal.c78 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal() local
194 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
260 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
280 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
326 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_3x8c8__neon_mlal_padal()
D3x8c16-minmax-neon-mlal-padal.c78 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal() local
183 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
203 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
249 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_3x8c16__neon_mlal_padal()
D4x8c8-minmax-neon-mull-padal.c82 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal() local
191 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
212 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
264 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mull_padal()
D4x8c8-minmax-neon-mlal-padal.c82 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal() local
235 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
318 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
339 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
391 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_4x8c8__neon_mlal_padal()
D4x8c16-minmax-neon-mlal-padal.c82 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal() local
223 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
244 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
296 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_4x8c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c94 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
183 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
267 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
351 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D3x16c8-minmax-neon-mlal-padal.c94 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
226 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
372 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
456 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
540 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c94 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
215 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
315 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
399 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c98 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
223 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
324 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
420 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
D4x16c8-minmax-neon-mlal-padal.c98 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
275 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
462 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
563 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
659 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
D4x16c16-minmax-neon-mlal-padal.c98 int32x4_t vacc2x7 = vacc0x7; in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
263 vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
388 const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
484 const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()