Home
last modified time | relevance | path

Searched refs:vacc2x23 (Results 1 – 25 of 130) sorted by relevance

123456

/external/XNNPACK/src/qs8-gemm/gen/
D3x8c4s2-minmax-rndnu-neon-mlal.c65 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local
106 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
152 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
201 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
228 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
249 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
265 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
D3x8c4-minmax-rndnu-neon-mull-ld1r.c65 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local
104 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
132 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
181 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
193 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
209 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
D3x8c4-minmax-rndnu-neon-mull-dup.c65 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local
101 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
129 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
178 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
190 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
206 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
D3x8c4-minmax-rndnu-neon-mull-ld2r.c65 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local
101 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
129 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
178 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
190 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
206 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
D3x8c4-minmax-rndnu-neon-mlal-ld2r.c65 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
114 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
161 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
215 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
243 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
292 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
304 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
320 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
D3x8c4-minmax-rndnu-neon-mlal-ld1r.c65 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
120 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
167 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
224 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
252 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
301 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
313 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
329 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
D3x8c4-minmax-rndnu-neon-mlal-dup.c65 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
114 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
161 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
215 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
243 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
292 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
304 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
320 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
D3x8c4s2-minmax-rndnu-neon-mull.c65 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull() local
95 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull()
122 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull()
144 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull()
160 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull()
D4x8c4-minmax-rndnu-neon-mull-ld2r.c71 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r() local
116 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r()
153 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r()
208 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r()
229 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r()
247 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r()
D4x8c4-minmax-rndnu-neon-mull-dup.c71 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup() local
116 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup()
153 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup()
208 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup()
229 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup()
247 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup()
D4x8c4-minmax-rndnu-neon-mull-ld1r.c71 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r() local
120 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r()
157 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r()
212 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r()
233 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r()
251 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r()
D4x8c4s2-minmax-rndnu-neon-mlal.c71 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
123 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
183 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
243 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
279 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
305 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
323 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
D4x8c4s2-minmax-rndnu-neon-mull.c71 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull() local
109 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull()
145 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull()
172 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull()
190 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull()
D4x8c4-minmax-rndnu-neon-mlal-ld2r.c71 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r() local
133 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
194 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
260 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
297 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
352 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
373 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
391 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld2r()
/external/XNNPACK/src/qs8-igemm/gen/
D3x8c4-minmax-rndnu-neon-mull-ld1r.c64 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r() local
119 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
147 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
196 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
210 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
226 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld1r()
D3x8c4-minmax-rndnu-neon-mull-dup.c64 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup() local
116 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
144 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
193 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
207 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
223 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_dup()
D3x8c4-minmax-rndnu-neon-mull-ld2r.c64 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r() local
116 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
144 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
193 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
207 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
223 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r()
D3x8c4-minmax-rndnu-neon-mlal-dup.c64 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup() local
129 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
176 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
230 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
258 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
307 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
321 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
337 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup()
D3x8c4-minmax-rndnu-neon-mlal-ld1r.c64 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r() local
135 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
182 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
239 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
267 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
316 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
330 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
346 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld1r()
D3x8c4-minmax-rndnu-neon-mlal-ld2r.c64 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r() local
129 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
176 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
230 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
258 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
307 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
321 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
337 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_ld2r()
D3x8c4s2-minmax-rndnu-neon-mlal.c64 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal() local
121 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
167 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
216 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
243 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
267 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
283 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal()
D3x8c4s2-minmax-rndnu-neon-mull.c64 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull() local
110 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull()
137 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull()
162 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull()
178 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4s2__neon_mull()
D4x8c4s2-minmax-rndnu-neon-mlal.c68 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal() local
140 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
200 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
260 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
296 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
325 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
343 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mlal()
D4x8c4s2-minmax-rndnu-neon-mull.c68 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull() local
126 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull()
162 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull()
192 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull()
210 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4s2__neon_mull()
D4x8c4-minmax-rndnu-neon-mlal-ld1r.c68 int32x4_t vacc2x23 = vacc0x23; in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r() local
158 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
219 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
289 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
326 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
381 vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
404 int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()
422 const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23)); in xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mlal_ld1r()

123456