| /external/XNNPACK/src/qs8-igemm/gen/ |
| D | 2x16c8-minmax-neon-mull-padal.c | 76 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
152 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
203 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
259 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
| D | 2x16c8-minmax-neon-mlal-padal.c | 76 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
190 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
291 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
342 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
398 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
| D | 2x16c16-minmax-neon-mlal-padal.c | 76 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
178 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
235 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
291 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
| D | 3x16c8-minmax-neon-mull-padal.c | 80 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
196 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
260 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
328 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
| D | 3x16c8-minmax-neon-mlal-padal.c | 80 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
245 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
385 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
449 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
517 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
| D | 3x16c16-minmax-neon-mlal-padal.c | 80 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
232 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
308 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
376 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
| D | 4x16c8-minmax-neon-mull-padal.c | 84 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
240 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
317 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
397 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
| D | 4x16c8-minmax-neon-mlal-padal.c | 84 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
300 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
479 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
556 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
636 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
| D | 4x16c16-minmax-neon-mlal-padal.c | 84 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
286 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
381 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
461 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|
| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 2x16c8-minmax-neon-mull-padal.c | 75 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
139 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
187 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
243 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
|
| D | 2x16c8-minmax-neon-mlal-padal.c | 75 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
177 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
278 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
326 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
382 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
|
| D | 2x16c16-minmax-neon-mlal-padal.c | 75 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
165 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
219 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
275 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
|
| D | 3x16c8-minmax-neon-mull-padal.c | 81 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
181 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
242 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
310 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
|
| D | 3x16c8-minmax-neon-mlal-padal.c | 81 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
230 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
370 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
431 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
499 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
|
| D | 3x16c16-minmax-neon-mlal-padal.c | 81 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
217 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
290 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
358 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
|
| D | 4x16c8-minmax-neon-mull-padal.c | 87 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
223 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
297 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
377 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
|
| D | 4x16c8-minmax-neon-mlal-padal.c | 87 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
283 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
462 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
536 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
616 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
|
| D | 4x16c16-minmax-neon-mlal-padal.c | 87 int32x4_t vacc1x9 = vacc0x9; in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
269 vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
361 const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
441 const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9)); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
|