xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/8x16c4-minmax-rndnu-neondot.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c4-neondot.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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_8x16c4__neondot(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_8x16c4__neondot(
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 <= 8);
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   const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
61   int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
62   if XNN_UNPREDICTABLE(mr <= 4) {
63     a4 = a3;
64     c4 = c3;
65   }
66   const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
67   int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
68   if XNN_UNPREDICTABLE(mr < 6) {
69     a5 = a4;
70     c5 = c4;
71   }
72   const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
73   int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
74   if XNN_UNPREDICTABLE(mr <= 6) {
75     a6 = a5;
76     c6 = c5;
77   }
78   const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
79   int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
80   if XNN_UNPREDICTABLE(mr != 8) {
81     a7 = a6;
82     c7 = c6;
83   }
84 
85   // Loop over groups of 16 columns.
86   do {
87     // Initialize accumulators with bias. 16 bias values are loaded from the
88     // weight matrix, at the start of the group of 16 columns.
89     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
90     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
91     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
92     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
93     int32x4_t vacc1x0123 = vacc0x0123;
94     int32x4_t vacc1x4567 = vacc0x4567;
95     int32x4_t vacc1x89AB = vacc0x89AB;
96     int32x4_t vacc1xCDEF = vacc0xCDEF;
97     int32x4_t vacc2x0123 = vacc0x0123;
98     int32x4_t vacc2x4567 = vacc0x4567;
99     int32x4_t vacc2x89AB = vacc0x89AB;
100     int32x4_t vacc2xCDEF = vacc0xCDEF;
101     int32x4_t vacc3x0123 = vacc0x0123;
102     int32x4_t vacc3x4567 = vacc0x4567;
103     int32x4_t vacc3x89AB = vacc0x89AB;
104     int32x4_t vacc3xCDEF = vacc0xCDEF;
105     int32x4_t vacc4x0123 = vacc0x0123;
106     int32x4_t vacc4x4567 = vacc0x4567;
107     int32x4_t vacc4x89AB = vacc0x89AB;
108     int32x4_t vacc4xCDEF = vacc0xCDEF;
109     int32x4_t vacc5x0123 = vacc0x0123;
110     int32x4_t vacc5x4567 = vacc0x4567;
111     int32x4_t vacc5x89AB = vacc0x89AB;
112     int32x4_t vacc5xCDEF = vacc0xCDEF;
113     int32x4_t vacc6x0123 = vacc0x0123;
114     int32x4_t vacc6x4567 = vacc0x4567;
115     int32x4_t vacc6x89AB = vacc0x89AB;
116     int32x4_t vacc6xCDEF = vacc0xCDEF;
117     int32x4_t vacc7x0123 = vacc0x0123;
118     int32x4_t vacc7x4567 = vacc0x4567;
119     int32x4_t vacc7x89AB = vacc0x89AB;
120     int32x4_t vacc7xCDEF = vacc0xCDEF;
121 
122     // Inner accumulation loop along the 16 columns.
123     size_t k = kc;
124     // 2x partial unrolled loop to load 8 bytes at a time.
125     while (k >= 8 * sizeof(int8_t)) {
126       // Load a 8x8 block of activations.
127       const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
128       const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
129       const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
130       const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
131       const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
132       const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
133       const int8x8_t va6x01234567 = vld1_s8(a6); a6 += 8;
134       const int8x8_t va7x01234567 = vld1_s8(a7); a7 += 8;
135 
136       // Load a 8x16 block of weights.
137       const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
138       const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
139       const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
140       const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
141       const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
142       const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
143       const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
144       const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
145 
146       // Multiply-accumulate: 8x8 * 8x16 --> 8x16.
147       vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
148       vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
149       vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
150       vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
151       vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
152       vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
153       vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
154       vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
155       vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
156       vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
157       vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
158       vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
159       vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
160       vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
161       vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
162       vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
163       vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
164       vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
165       vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
166       vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
167       vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
168       vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
169       vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
170       vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
171       vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0);
172       vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0);
173       vacc6x89AB = vdotq_lane_s32(vacc6x89AB, vb0123x89AB, va6x01234567, 0);
174       vacc6xCDEF = vdotq_lane_s32(vacc6xCDEF, vb0123xCDEF, va6x01234567, 0);
175       vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0);
176       vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0);
177       vacc7x89AB = vdotq_lane_s32(vacc7x89AB, vb0123x89AB, va7x01234567, 0);
178       vacc7xCDEF = vdotq_lane_s32(vacc7xCDEF, vb0123xCDEF, va7x01234567, 0);
179       vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
180       vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
181       vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
182       vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
183       vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
184       vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
185       vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
186       vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
187       vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
188       vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
189       vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
190       vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
191       vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
192       vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
193       vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
194       vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
195       vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
196       vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
197       vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1);
198       vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
199       vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
200       vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
201       vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb4567x89AB, va5x01234567, 1);
202       vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb4567xCDEF, va5x01234567, 1);
203       vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1);
204       vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb4567x4567, va6x01234567, 1);
205       vacc6x89AB = vdotq_lane_s32(vacc6x89AB, vb4567x89AB, va6x01234567, 1);
206       vacc6xCDEF = vdotq_lane_s32(vacc6xCDEF, vb4567xCDEF, va6x01234567, 1);
207       vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1);
208       vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb4567x4567, va7x01234567, 1);
209       vacc7x89AB = vdotq_lane_s32(vacc7x89AB, vb4567x89AB, va7x01234567, 1);
210       vacc7xCDEF = vdotq_lane_s32(vacc7xCDEF, vb4567xCDEF, va7x01234567, 1);
211 
212       k -= 8 * sizeof(int8_t);
213     }
214     // Handle up to 4 final positions of `k`
215     if XNN_UNLIKELY(k != 0) {
216       // Load a 8x4 block of activations.
217       const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
218       const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
219       const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
220       const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
221       const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 4;
222       const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 4;
223       const int8x8_t va6x01234567 = vld1_s8(a6); a6 += 4;
224       const int8x8_t va7x01234567 = vld1_s8(a7); a7 += 4;
225 
226       // Load a 4x16 block of weights.
227       const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
228       const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
229       const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
230       const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
231 
232       // Multiply-accumulate: 8x4 * 4x16 --> 8x16.
233       vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
234       vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
235       vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
236       vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
237       vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
238       vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
239       vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
240       vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
241       vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
242       vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
243       vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
244       vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
245       vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
246       vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
247       vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
248       vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
249       vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
250       vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
251       vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
252       vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
253       vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
254       vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
255       vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
256       vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
257       vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0);
258       vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0);
259       vacc6x89AB = vdotq_lane_s32(vacc6x89AB, vb0123x89AB, va6x01234567, 0);
260       vacc6xCDEF = vdotq_lane_s32(vacc6xCDEF, vb0123xCDEF, va6x01234567, 0);
261       vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0);
262       vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0);
263       vacc7x89AB = vdotq_lane_s32(vacc7x89AB, vb0123x89AB, va7x01234567, 0);
264       vacc7xCDEF = vdotq_lane_s32(vacc7xCDEF, vb0123xCDEF, va7x01234567, 0);
265     }
266 
267     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
268     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
269     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
270 
271     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
272     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
273     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
274     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
275     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
276     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
277     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
278     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
279     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
280     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
281     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
282     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
283     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
284     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
285     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
286     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
287     vacc4x0123 = vqshlq_s32(vacc4x0123, vright_pre_shift);
288     vacc4x4567 = vqshlq_s32(vacc4x4567, vright_pre_shift);
289     vacc4x89AB = vqshlq_s32(vacc4x89AB, vright_pre_shift);
290     vacc4xCDEF = vqshlq_s32(vacc4xCDEF, vright_pre_shift);
291     vacc5x0123 = vqshlq_s32(vacc5x0123, vright_pre_shift);
292     vacc5x4567 = vqshlq_s32(vacc5x4567, vright_pre_shift);
293     vacc5x89AB = vqshlq_s32(vacc5x89AB, vright_pre_shift);
294     vacc5xCDEF = vqshlq_s32(vacc5xCDEF, vright_pre_shift);
295     vacc6x0123 = vqshlq_s32(vacc6x0123, vright_pre_shift);
296     vacc6x4567 = vqshlq_s32(vacc6x4567, vright_pre_shift);
297     vacc6x89AB = vqshlq_s32(vacc6x89AB, vright_pre_shift);
298     vacc6xCDEF = vqshlq_s32(vacc6xCDEF, vright_pre_shift);
299     vacc7x0123 = vqshlq_s32(vacc7x0123, vright_pre_shift);
300     vacc7x4567 = vqshlq_s32(vacc7x4567, vright_pre_shift);
301     vacc7x89AB = vqshlq_s32(vacc7x89AB, vright_pre_shift);
302     vacc7xCDEF = vqshlq_s32(vacc7xCDEF, vright_pre_shift);
303 
304     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
305     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
306     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
307     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
308     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
309     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
310     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
311     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
312     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
313     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
314     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
315     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
316     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
317     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
318     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
319     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
320     vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
321     vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
322     vacc4x89AB = vqdmulhq_s32(vacc4x89AB, vmultiplier);
323     vacc4xCDEF = vqdmulhq_s32(vacc4xCDEF, vmultiplier);
324     vacc5x0123 = vqdmulhq_s32(vacc5x0123, vmultiplier);
325     vacc5x4567 = vqdmulhq_s32(vacc5x4567, vmultiplier);
326     vacc5x89AB = vqdmulhq_s32(vacc5x89AB, vmultiplier);
327     vacc5xCDEF = vqdmulhq_s32(vacc5xCDEF, vmultiplier);
328     vacc6x0123 = vqdmulhq_s32(vacc6x0123, vmultiplier);
329     vacc6x4567 = vqdmulhq_s32(vacc6x4567, vmultiplier);
330     vacc6x89AB = vqdmulhq_s32(vacc6x89AB, vmultiplier);
331     vacc6xCDEF = vqdmulhq_s32(vacc6xCDEF, vmultiplier);
332     vacc7x0123 = vqdmulhq_s32(vacc7x0123, vmultiplier);
333     vacc7x4567 = vqdmulhq_s32(vacc7x4567, vmultiplier);
334     vacc7x89AB = vqdmulhq_s32(vacc7x89AB, vmultiplier);
335     vacc7xCDEF = vqdmulhq_s32(vacc7xCDEF, vmultiplier);
336 
337     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
338     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
339     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
340     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
341     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
342     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
343     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
344     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
345     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
346     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
347     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
348     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
349     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
350     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
351     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
352     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
353     vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
354     vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
355     vacc4x89AB = vrshlq_s32(vacc4x89AB, vright_post_shift);
356     vacc4xCDEF = vrshlq_s32(vacc4xCDEF, vright_post_shift);
357     vacc5x0123 = vrshlq_s32(vacc5x0123, vright_post_shift);
358     vacc5x4567 = vrshlq_s32(vacc5x4567, vright_post_shift);
359     vacc5x89AB = vrshlq_s32(vacc5x89AB, vright_post_shift);
360     vacc5xCDEF = vrshlq_s32(vacc5xCDEF, vright_post_shift);
361     vacc6x0123 = vrshlq_s32(vacc6x0123, vright_post_shift);
362     vacc6x4567 = vrshlq_s32(vacc6x4567, vright_post_shift);
363     vacc6x89AB = vrshlq_s32(vacc6x89AB, vright_post_shift);
364     vacc6xCDEF = vrshlq_s32(vacc6xCDEF, vright_post_shift);
365     vacc7x0123 = vrshlq_s32(vacc7x0123, vright_post_shift);
366     vacc7x4567 = vrshlq_s32(vacc7x4567, vright_post_shift);
367     vacc7x89AB = vrshlq_s32(vacc7x89AB, vright_post_shift);
368     vacc7xCDEF = vrshlq_s32(vacc7xCDEF, vright_post_shift);
369 
370     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
371 #if XNN_ARCH_ARM64
372     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
373     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
374     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
375     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
376     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
377     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
378     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
379     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
380     const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
381     const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), voutput_zero_point);
382     const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
383     const int16x8_t vacc5x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x89AB), vacc5xCDEF), voutput_zero_point);
384     const int16x8_t vacc6x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x0123), vacc6x4567), voutput_zero_point);
385     const int16x8_t vacc6x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x89AB), vacc6xCDEF), voutput_zero_point);
386     const int16x8_t vacc7x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x0123), vacc7x4567), voutput_zero_point);
387     const int16x8_t vacc7x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x89AB), vacc7xCDEF), voutput_zero_point);
388 
389     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
390     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
391     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
392     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
393     int8x16_t vout4x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc4x89ABCDEF);
394     int8x16_t vout5x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc5x01234567), vacc5x89ABCDEF);
395     int8x16_t vout6x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc6x89ABCDEF);
396     int8x16_t vout7x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc7x01234567), vacc7x89ABCDEF);
397 #else
398     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
399     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
400     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
401     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
402     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
403     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
404     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
405     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
406     const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
407     const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF)), voutput_zero_point);
408     const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
409     const int16x8_t vacc5x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x89AB), vqmovn_s32(vacc5xCDEF)), voutput_zero_point);
410     const int16x8_t vacc6x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x0123), vqmovn_s32(vacc6x4567)), voutput_zero_point);
411     const int16x8_t vacc6x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x89AB), vqmovn_s32(vacc6xCDEF)), voutput_zero_point);
412     const int16x8_t vacc7x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc7x0123), vqmovn_s32(vacc7x4567)), voutput_zero_point);
413     const int16x8_t vacc7x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc7x89AB), vqmovn_s32(vacc7xCDEF)), voutput_zero_point);
414 
415     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
416     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
417     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
418     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
419     int8x16_t vout4x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc4x89ABCDEF));
420     int8x16_t vout5x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc5x01234567), vqmovn_s16(vacc5x89ABCDEF));
421     int8x16_t vout6x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc6x89ABCDEF));
422     int8x16_t vout7x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc7x01234567), vqmovn_s16(vacc7x89ABCDEF));
423 #endif
424     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
425     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
426 
427     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
428     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
429     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
430     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
431     vout4x0123456789ABCDEF = vmaxq_s8(vout4x0123456789ABCDEF, voutput_min);
432     vout5x0123456789ABCDEF = vmaxq_s8(vout5x0123456789ABCDEF, voutput_min);
433     vout6x0123456789ABCDEF = vmaxq_s8(vout6x0123456789ABCDEF, voutput_min);
434     vout7x0123456789ABCDEF = vmaxq_s8(vout7x0123456789ABCDEF, voutput_min);
435 
436     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
437     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
438     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
439     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
440     vout4x0123456789ABCDEF = vminq_s8(vout4x0123456789ABCDEF, voutput_max);
441     vout5x0123456789ABCDEF = vminq_s8(vout5x0123456789ABCDEF, voutput_max);
442     vout6x0123456789ABCDEF = vminq_s8(vout6x0123456789ABCDEF, voutput_max);
443     vout7x0123456789ABCDEF = vminq_s8(vout7x0123456789ABCDEF, voutput_max);
444 
445     if (nc >= 16) {
446       // Main case where there the 16 columns fit in the destination.
447       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
448       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
449       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
450       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
451       vst1q_s8(c4 + 0, vout4x0123456789ABCDEF);
452       vst1q_s8(c5 + 0, vout5x0123456789ABCDEF);
453       vst1q_s8(c6 + 0, vout6x0123456789ABCDEF);
454       vst1q_s8(c7 + 0, vout7x0123456789ABCDEF);
455 
456       // Advance to the next 16 columns.
457       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
458       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
459       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
460       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
461       c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
462       c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
463       c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
464       c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
465 
466       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
467       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
468       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
469       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
470       a4 = (const int8_t*) ((uintptr_t) a4 - kc);
471       a5 = (const int8_t*) ((uintptr_t) a5 - kc);
472       a6 = (const int8_t*) ((uintptr_t) a6 - kc);
473       a7 = (const int8_t*) ((uintptr_t) a7 - kc);
474 
475       nc -= 16;
476     } else {
477       // Final case where not all of the 16 columns fit in the destination.
478       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
479       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
480       int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vget_low_s8(vout4x0123456789ABCDEF), vget_low_s8(vout5x0123456789ABCDEF));
481       int8x16_t vout6x01234567_7x01234567 = vcombine_s8(vget_low_s8(vout6x0123456789ABCDEF), vget_low_s8(vout7x0123456789ABCDEF));
482       if (nc & 8) {
483         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
484         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
485         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
486         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
487         vst1_s8(c4, vget_low_s8(vout4x01234567_5x01234567)); c4 += 8;
488         vst1_s8(c5, vget_high_s8(vout4x01234567_5x01234567)); c5 += 8;
489         vst1_s8(c6, vget_low_s8(vout6x01234567_7x01234567)); c6 += 8;
490         vst1_s8(c7, vget_high_s8(vout6x01234567_7x01234567)); c7 += 8;
491         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
492         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
493         vout4x01234567_5x01234567 = vcombine_s8(vget_high_s8(vout4x0123456789ABCDEF), vget_high_s8(vout5x0123456789ABCDEF));
494         vout6x01234567_7x01234567 = vcombine_s8(vget_high_s8(vout6x0123456789ABCDEF), vget_high_s8(vout7x0123456789ABCDEF));
495       }
496       if (nc & 4) {
497         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
498         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
499         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
500         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
501         vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
502         vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
503         vst1q_lane_u32((void*) c6, vreinterpretq_u32_s8(vout6x01234567_7x01234567), 0); c6 += 4;
504         vst1q_lane_u32((void*) c7, vreinterpretq_u32_s8(vout6x01234567_7x01234567), 2); c7 += 4;
505         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
506         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
507         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
508         vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 4);
509       }
510       if (nc & 2) {
511         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
512         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
513         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
514         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
515         vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
516         vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
517         vst1q_lane_u16((void*) c6, vreinterpretq_u16_s8(vout6x01234567_7x01234567), 0); c6 += 2;
518         vst1q_lane_u16((void*) c7, vreinterpretq_u16_s8(vout6x01234567_7x01234567), 4); c7 += 2;
519         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
520         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
521         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
522         vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 2);
523       }
524       if (nc & 1) {
525         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
526         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
527         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
528         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
529         vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
530         vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
531         vst1q_lane_s8(c6, vout6x01234567_7x01234567, 0);
532         vst1q_lane_s8(c7, vout6x01234567_7x01234567, 8);
533       }
534 
535       nc = 0;
536     }
537   } while (nc != 0);
538 }
539