• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c4-neon-mull-dup.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld2r(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 4);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 4 * sizeof(int8_t));
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr < 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     a2 = a1;
52     c2 = c1;
53   }
54   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
55   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56   if XNN_UNPREDICTABLE(mr != 4) {
57     a3 = a2;
58     c3 = c2;
59   }
60 
61   do {
62     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
63     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
64     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
65     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
66     int32x4_t vacc1x01 = vacc0x01;
67     int32x4_t vacc1x23 = vacc0x23;
68     int32x4_t vacc1x45 = vacc0x45;
69     int32x4_t vacc1x67 = vacc0x67;
70     int32x4_t vacc2x01 = vacc0x01;
71     int32x4_t vacc2x23 = vacc0x23;
72     int32x4_t vacc2x45 = vacc0x45;
73     int32x4_t vacc2x67 = vacc0x67;
74     int32x4_t vacc3x01 = vacc0x01;
75     int32x4_t vacc3x23 = vacc0x23;
76     int32x4_t vacc3x45 = vacc0x45;
77     int32x4_t vacc3x67 = vacc0x67;
78 
79     size_t k = kc;
80 
81 
82     while (k >= 8 * sizeof(int8_t)) {
83       const int32x2x2_t va0 = vld2_dup_s32((const void*)a0); a0 += 8;
84       const int32x2x2_t va1 = vld2_dup_s32((const void*)a1); a1 += 8;
85       const int32x2x2_t va2 = vld2_dup_s32((const void*)a2); a2 += 8;
86       const int32x2x2_t va3 = vld2_dup_s32((const void*)a3); a3 += 8;
87 
88       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92       const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93       const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94       const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95       const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96 
97       const int8x8_t va0c0 = vreinterpret_s8_s32(va0.val[0]);
98       const int8x8_t va1c0 = vreinterpret_s8_s32(va1.val[0]);
99       const int8x8_t va2c0 = vreinterpret_s8_s32(va2.val[0]);
100       const int8x8_t va3c0 = vreinterpret_s8_s32(va3.val[0]);
101 
102       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
103       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
104       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
105       const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
106       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
107       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
108       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
109       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
110       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
111       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
112       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
113       const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
114       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
115       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
116       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
117       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
118       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
119       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
120       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
121       const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
122       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
123       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
124       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
125       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
126       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
127       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
128       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
129       const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
130       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
131       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
132       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
133       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
134       const int8x8_t va0c1 = vreinterpret_s8_s32(va0.val[1]);
135       const int8x8_t va1c1 = vreinterpret_s8_s32(va1.val[1]);
136       const int8x8_t va2c1 = vreinterpret_s8_s32(va2.val[1]);
137       const int8x8_t va3c1 = vreinterpret_s8_s32(va3.val[1]);
138 
139       const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
140       const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
141       const int16x8_t vprod2x01c1 = vmull_s8(vb01c1, va2c1);
142       const int16x8_t vprod3x01c1 = vmull_s8(vb01c1, va3c1);
143       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
144       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
145       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c1);
146       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c1);
147       const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
148       const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
149       const int16x8_t vprod2x23c1 = vmull_s8(vb23c1, va2c1);
150       const int16x8_t vprod3x23c1 = vmull_s8(vb23c1, va3c1);
151       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
152       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
153       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c1);
154       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c1);
155       const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
156       const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
157       const int16x8_t vprod2x45c1 = vmull_s8(vb45c1, va2c1);
158       const int16x8_t vprod3x45c1 = vmull_s8(vb45c1, va3c1);
159       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
160       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
161       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c1);
162       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c1);
163       const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
164       const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
165       const int16x8_t vprod2x67c1 = vmull_s8(vb67c1, va2c1);
166       const int16x8_t vprod3x67c1 = vmull_s8(vb67c1, va3c1);
167       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
168       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
169       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c1);
170       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c1);
171 
172       k -= 8 * sizeof(int8_t);
173     }
174 
175     if XNN_UNLIKELY(k != 0) {
176       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
177       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
178       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
179       const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
180 
181       const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182       const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
183       const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
184       const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
185 
186       const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
187       const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
188       vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
189       const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
190       vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
191       const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
192       vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
193       const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
194       vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
195       const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
196       const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
197       vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
198       const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
199       vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
200       const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
201       vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
202       const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
203       vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
204       const int8x8_t va2c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va2), 0));
205       const int16x8_t vprod2x01c0 = vmull_s8(vb01c0, va2c0);
206       vacc2x01 = vpadalq_s16(vacc2x01, vprod2x01c0);
207       const int16x8_t vprod2x23c0 = vmull_s8(vb23c0, va2c0);
208       vacc2x23 = vpadalq_s16(vacc2x23, vprod2x23c0);
209       const int16x8_t vprod2x45c0 = vmull_s8(vb45c0, va2c0);
210       vacc2x45 = vpadalq_s16(vacc2x45, vprod2x45c0);
211       const int16x8_t vprod2x67c0 = vmull_s8(vb67c0, va2c0);
212       vacc2x67 = vpadalq_s16(vacc2x67, vprod2x67c0);
213       const int8x8_t va3c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va3), 0));
214       const int16x8_t vprod3x01c0 = vmull_s8(vb01c0, va3c0);
215       vacc3x01 = vpadalq_s16(vacc3x01, vprod3x01c0);
216       const int16x8_t vprod3x23c0 = vmull_s8(vb23c0, va3c0);
217       vacc3x23 = vpadalq_s16(vacc3x23, vprod3x23c0);
218       const int16x8_t vprod3x45c0 = vmull_s8(vb45c0, va3c0);
219       vacc3x45 = vpadalq_s16(vacc3x45, vprod3x45c0);
220       const int16x8_t vprod3x67c0 = vmull_s8(vb67c0, va3c0);
221       vacc3x67 = vpadalq_s16(vacc3x67, vprod3x67c0);
222     }
223 
224 #if XNN_ARCH_ARM64
225     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
226     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
227     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
228     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
229     int32x4_t vacc2x0123 = vpaddq_s32(vacc2x01, vacc2x23);
230     int32x4_t vacc2x4567 = vpaddq_s32(vacc2x45, vacc2x67);
231     int32x4_t vacc3x0123 = vpaddq_s32(vacc3x01, vacc3x23);
232     int32x4_t vacc3x4567 = vpaddq_s32(vacc3x45, vacc3x67);
233 #else
234     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
235     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
236     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
237     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
238     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
239     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
240     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
241     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
242     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
243     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
244     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
245     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
246     const int32x2_t vsum2x01 = vpadd_s32(vget_low_s32(vacc2x01), vget_high_s32(vacc2x01));
247     const int32x2_t vsum2x23 = vpadd_s32(vget_low_s32(vacc2x23), vget_high_s32(vacc2x23));
248     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23);
249     const int32x2_t vsum2x45 = vpadd_s32(vget_low_s32(vacc2x45), vget_high_s32(vacc2x45));
250     const int32x2_t vsum2x67 = vpadd_s32(vget_low_s32(vacc2x67), vget_high_s32(vacc2x67));
251     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67);
252     const int32x2_t vsum3x01 = vpadd_s32(vget_low_s32(vacc3x01), vget_high_s32(vacc3x01));
253     const int32x2_t vsum3x23 = vpadd_s32(vget_low_s32(vacc3x23), vget_high_s32(vacc3x23));
254     int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23);
255     const int32x2_t vsum3x45 = vpadd_s32(vget_low_s32(vacc3x45), vget_high_s32(vacc3x45));
256     const int32x2_t vsum3x67 = vpadd_s32(vget_low_s32(vacc3x67), vget_high_s32(vacc3x67));
257     int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67);
258 #endif
259 
260     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
261     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
262     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
263 
264     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
265     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
266     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
267     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
268     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
269     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
270     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
271     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
272 
273     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
274     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
275     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
276     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
277     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
278     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
279     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
280     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
281 
282     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
283     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
284     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
285     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
286     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
287     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
288     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
289     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
290 
291     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
292 #if XNN_ARCH_ARM64
293     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
294     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
295     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
296     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
297 
298     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
299     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
300     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
301     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
302 
303     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
304     int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
305 #else
306     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
307     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
308     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
309     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
310 
311     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
312     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
313     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
314     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
315 
316     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
317     int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
318 #endif
319 
320     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
321     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
322     vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
323 
324     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
325     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
326     vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
327 
328     if (nc >= 8) {
329       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
330       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
331       vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
332       vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
333 
334       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
335       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
336       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
337       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
338 
339       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
340       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
341       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
342       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
343 
344       nc -= 8;
345     } else {
346       // Final case where not all of the 8 columns fit in the destination.
347       if (nc & 4) {
348         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
349         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
350         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
351         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
352         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
353         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
354       }
355       if (nc & 2) {
356         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
357         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
358         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
359         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
360         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
361         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
362       }
363       if (nc & 1) {
364         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
365         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
366         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
367         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
368       }
369 
370       nc = 0;
371     }
372   } while (nc != 0);
373 }
374