• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/MRxNRc4-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/igemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_ukernel_6x16c4__neondot(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const int8_t** restrict a,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const int8_t* zero,
30     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
31 {
32   assert(mr != 0);
33   assert(mr <= 6);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (6 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(int8_t) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   kc = round_up_po2(kc, 4);
44   int8_t* c0 = c;
45   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr < 2) {
47     c1 = c0;
48   }
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     c2 = c1;
52   }
53   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
54   if XNN_UNPREDICTABLE(mr < 4) {
55     c3 = c2;
56   }
57   int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
58   if XNN_UNPREDICTABLE(mr <= 4) {
59     c4 = c3;
60   }
61   int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
62   if XNN_UNPREDICTABLE(mr != 6) {
63     c5 = c4;
64   }
65 
66   do {
67     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
68     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
69     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
70     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
71     int32x4_t vacc1x0123 = vacc0x0123;
72     int32x4_t vacc1x4567 = vacc0x4567;
73     int32x4_t vacc1x89AB = vacc0x89AB;
74     int32x4_t vacc1xCDEF = vacc0xCDEF;
75     int32x4_t vacc2x0123 = vacc0x0123;
76     int32x4_t vacc2x4567 = vacc0x4567;
77     int32x4_t vacc2x89AB = vacc0x89AB;
78     int32x4_t vacc2xCDEF = vacc0xCDEF;
79     int32x4_t vacc3x0123 = vacc0x0123;
80     int32x4_t vacc3x4567 = vacc0x4567;
81     int32x4_t vacc3x89AB = vacc0x89AB;
82     int32x4_t vacc3xCDEF = vacc0xCDEF;
83     int32x4_t vacc4x0123 = vacc0x0123;
84     int32x4_t vacc4x4567 = vacc0x4567;
85     int32x4_t vacc4x89AB = vacc0x89AB;
86     int32x4_t vacc4xCDEF = vacc0xCDEF;
87     int32x4_t vacc5x0123 = vacc0x0123;
88     int32x4_t vacc5x4567 = vacc0x4567;
89     int32x4_t vacc5x89AB = vacc0x89AB;
90     int32x4_t vacc5xCDEF = vacc0xCDEF;
91 
92     size_t p = ks;
93     do {
94       const int8_t* restrict a0 = a[0];
95       if XNN_UNPREDICTABLE(a0 != zero) {
96         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
97       }
98       const int8_t* restrict a1 = a[1];
99       if XNN_UNPREDICTABLE(a1 != zero) {
100         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
101       }
102       const int8_t* restrict a2 = a[2];
103       if XNN_UNPREDICTABLE(a2 != zero) {
104         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
105       }
106       const int8_t* restrict a3 = a[3];
107       if XNN_UNPREDICTABLE(a3 != zero) {
108         a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
109       }
110       const int8_t* restrict a4 = a[4];
111       if XNN_UNPREDICTABLE(a4 != zero) {
112         a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
113       }
114       const int8_t* restrict a5 = a[5];
115       if XNN_UNPREDICTABLE(a5 != zero) {
116         a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
117       }
118       a += 6;
119 
120       // Inner accumulation loop along the 16 columns.
121       size_t k = kc;
122       // 2x partial unrolled loop to load 8 bytes at a time.
123       while (k >= 8 * sizeof(int8_t)) {
124         // Load a 6x8 block of activations.
125         const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
126         const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
127         const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
128         const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
129         const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
130         const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
131 
132         // Load a 8x16 block of weights.
133         const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
134         const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
135         const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
136         const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
137         const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
138         const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
139         const int8x16_t vb4567x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
140         const int8x16_t vb4567xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
141 
142         // Multiply-accumulate: 6x8 * 8x16 --> 6x16.
143         vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
144         vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
145         vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
146         vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
147         vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
148         vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
149         vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
150         vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
151         vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
152         vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
153         vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
154         vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
155         vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
156         vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
157         vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
158         vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
159         vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
160         vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
161         vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
162         vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
163         vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
164         vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
165         vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
166         vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
167         vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
168         vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
169         vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
170         vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
171         vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
172         vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
173         vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
174         vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
175         vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
176         vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
177         vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
178         vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
179         vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
180         vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
181         vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb4567x89AB, va3x01234567, 1);
182         vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
183         vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
184         vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
185         vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb4567x89AB, va4x01234567, 1);
186         vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
187         vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
188         vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
189         vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb4567x89AB, va5x01234567, 1);
190         vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb4567xCDEF, va5x01234567, 1);
191 
192         k -= 8 * sizeof(int8_t);
193       }
194       // Handle up to 4 final positions of `k`
195       if XNN_UNLIKELY(k != 0) {
196         // Load a 6x4 block of activations.
197         const int8x8_t va0x01234567 = vld1_s8(a0);
198         const int8x8_t va1x01234567 = vld1_s8(a1);
199         const int8x8_t va2x01234567 = vld1_s8(a2);
200         const int8x8_t va3x01234567 = vld1_s8(a3);
201         const int8x8_t va4x01234567 = vld1_s8(a4);
202         const int8x8_t va5x01234567 = vld1_s8(a5);
203 
204         // Load a 4x16 block of weights.
205         const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
206         const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
207         const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
208         const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
209 
210         // Multiply-accumulate: 6x4 * 4x16 --> 6x16.
211         vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
212         vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
213         vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
214         vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
215         vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
216         vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
217         vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
218         vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
219         vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
220         vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
221         vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
222         vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
223         vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
224         vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
225         vacc3x89AB = vdotq_lane_s32(vacc3x89AB, vb0123x89AB, va3x01234567, 0);
226         vacc3xCDEF = vdotq_lane_s32(vacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
227         vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
228         vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
229         vacc4x89AB = vdotq_lane_s32(vacc4x89AB, vb0123x89AB, va4x01234567, 0);
230         vacc4xCDEF = vdotq_lane_s32(vacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
231         vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
232         vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
233         vacc5x89AB = vdotq_lane_s32(vacc5x89AB, vb0123x89AB, va5x01234567, 0);
234         vacc5xCDEF = vdotq_lane_s32(vacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
235       }
236       p -= 6 * sizeof(void*);
237     } while (p != 0);
238 
239     const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
240     vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
241     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
242     vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);
243     vacc0xCDEF = vqrdmulhq_s32(vacc0xCDEF, vmultiplier);
244     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
245     vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
246     vacc1x89AB = vqrdmulhq_s32(vacc1x89AB, vmultiplier);
247     vacc1xCDEF = vqrdmulhq_s32(vacc1xCDEF, vmultiplier);
248     vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier);
249     vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier);
250     vacc2x89AB = vqrdmulhq_s32(vacc2x89AB, vmultiplier);
251     vacc2xCDEF = vqrdmulhq_s32(vacc2xCDEF, vmultiplier);
252     vacc3x0123 = vqrdmulhq_s32(vacc3x0123, vmultiplier);
253     vacc3x4567 = vqrdmulhq_s32(vacc3x4567, vmultiplier);
254     vacc3x89AB = vqrdmulhq_s32(vacc3x89AB, vmultiplier);
255     vacc3xCDEF = vqrdmulhq_s32(vacc3xCDEF, vmultiplier);
256     vacc4x0123 = vqrdmulhq_s32(vacc4x0123, vmultiplier);
257     vacc4x4567 = vqrdmulhq_s32(vacc4x4567, vmultiplier);
258     vacc4x89AB = vqrdmulhq_s32(vacc4x89AB, vmultiplier);
259     vacc4xCDEF = vqrdmulhq_s32(vacc4xCDEF, vmultiplier);
260     vacc5x0123 = vqrdmulhq_s32(vacc5x0123, vmultiplier);
261     vacc5x4567 = vqrdmulhq_s32(vacc5x4567, vmultiplier);
262     vacc5x89AB = vqrdmulhq_s32(vacc5x89AB, vmultiplier);
263     vacc5xCDEF = vqrdmulhq_s32(vacc5xCDEF, vmultiplier);
264 
265     const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
266     const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
267     vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
268     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
269     vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);
270     vacc0xCDEF = vsraq_n_s32(vacc0xCDEF, vbicq_s32(vacc0xCDEF, vzero_shift_mask), 31);
271     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
272     vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
273     vacc1x89AB = vsraq_n_s32(vacc1x89AB, vbicq_s32(vacc1x89AB, vzero_shift_mask), 31);
274     vacc1xCDEF = vsraq_n_s32(vacc1xCDEF, vbicq_s32(vacc1xCDEF, vzero_shift_mask), 31);
275     vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
276     vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
277     vacc2x89AB = vsraq_n_s32(vacc2x89AB, vbicq_s32(vacc2x89AB, vzero_shift_mask), 31);
278     vacc2xCDEF = vsraq_n_s32(vacc2xCDEF, vbicq_s32(vacc2xCDEF, vzero_shift_mask), 31);
279     vacc3x0123 = vsraq_n_s32(vacc3x0123, vbicq_s32(vacc3x0123, vzero_shift_mask), 31);
280     vacc3x4567 = vsraq_n_s32(vacc3x4567, vbicq_s32(vacc3x4567, vzero_shift_mask), 31);
281     vacc3x89AB = vsraq_n_s32(vacc3x89AB, vbicq_s32(vacc3x89AB, vzero_shift_mask), 31);
282     vacc3xCDEF = vsraq_n_s32(vacc3xCDEF, vbicq_s32(vacc3xCDEF, vzero_shift_mask), 31);
283     vacc4x0123 = vsraq_n_s32(vacc4x0123, vbicq_s32(vacc4x0123, vzero_shift_mask), 31);
284     vacc4x4567 = vsraq_n_s32(vacc4x4567, vbicq_s32(vacc4x4567, vzero_shift_mask), 31);
285     vacc4x89AB = vsraq_n_s32(vacc4x89AB, vbicq_s32(vacc4x89AB, vzero_shift_mask), 31);
286     vacc4xCDEF = vsraq_n_s32(vacc4xCDEF, vbicq_s32(vacc4xCDEF, vzero_shift_mask), 31);
287     vacc5x0123 = vsraq_n_s32(vacc5x0123, vbicq_s32(vacc5x0123, vzero_shift_mask), 31);
288     vacc5x4567 = vsraq_n_s32(vacc5x4567, vbicq_s32(vacc5x4567, vzero_shift_mask), 31);
289     vacc5x89AB = vsraq_n_s32(vacc5x89AB, vbicq_s32(vacc5x89AB, vzero_shift_mask), 31);
290     vacc5xCDEF = vsraq_n_s32(vacc5xCDEF, vbicq_s32(vacc5xCDEF, vzero_shift_mask), 31);
291 
292     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
293     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
294     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);
295     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_shift);
296     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
297     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
298     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_shift);
299     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_shift);
300     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
301     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
302     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_shift);
303     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_shift);
304     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_shift);
305     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_shift);
306     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_shift);
307     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_shift);
308     vacc4x0123 = vrshlq_s32(vacc4x0123, vright_shift);
309     vacc4x4567 = vrshlq_s32(vacc4x4567, vright_shift);
310     vacc4x89AB = vrshlq_s32(vacc4x89AB, vright_shift);
311     vacc4xCDEF = vrshlq_s32(vacc4xCDEF, vright_shift);
312     vacc5x0123 = vrshlq_s32(vacc5x0123, vright_shift);
313     vacc5x4567 = vrshlq_s32(vacc5x4567, vright_shift);
314     vacc5x89AB = vrshlq_s32(vacc5x89AB, vright_shift);
315     vacc5xCDEF = vrshlq_s32(vacc5xCDEF, vright_shift);
316 
317     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
318 #if XNN_ARCH_ARM64
319     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
320     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
321     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
322     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
323     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
324     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
325     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
326     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
327     const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
328     const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), voutput_zero_point);
329     const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
330     const int16x8_t vacc5x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x89AB), vacc5xCDEF), voutput_zero_point);
331 
332     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
333     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
334     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
335     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
336     int8x16_t vout4x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc4x89ABCDEF);
337     int8x16_t vout5x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc5x01234567), vacc5x89ABCDEF);
338 #else
339     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
340     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
341     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
342     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
343     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
344     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
345     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
346     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
347     const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
348     const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF)), voutput_zero_point);
349     const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
350     const int16x8_t vacc5x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x89AB), vqmovn_s32(vacc5xCDEF)), voutput_zero_point);
351 
352     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
353     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
354     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
355     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
356     int8x16_t vout4x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc4x89ABCDEF));
357     int8x16_t vout5x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc5x01234567), vqmovn_s16(vacc5x89ABCDEF));
358 #endif
359     const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
360     const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
361 
362     vout5x0123456789ABCDEF = vmaxq_s8(vout5x0123456789ABCDEF, voutput_min);
363     vout4x0123456789ABCDEF = vmaxq_s8(vout4x0123456789ABCDEF, voutput_min);
364     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
365     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
366     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
367     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
368 
369     vout5x0123456789ABCDEF = vminq_s8(vout5x0123456789ABCDEF, voutput_max);
370     vout4x0123456789ABCDEF = vminq_s8(vout4x0123456789ABCDEF, voutput_max);
371     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
372     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
373     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
374     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
375 
376     if (nc >= 16) {
377       vst1q_s8(c5 + 0, vout5x0123456789ABCDEF);
378       vst1q_s8(c4 + 0, vout4x0123456789ABCDEF);
379       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
380       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
381       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
382       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
383 
384       c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
385       c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
386       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
387       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
388       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
389       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
390 
391       a = (const int8_t**restrict) ((uintptr_t) a - ks);
392 
393       nc -= 16;
394     } else {
395       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
396       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
397       int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vget_low_s8(vout4x0123456789ABCDEF), vget_low_s8(vout5x0123456789ABCDEF));
398       if (nc & 8) {
399         vst1_s8(c5, vget_high_s8(vout4x01234567_5x01234567)); c5 += 8;
400         vst1_s8(c4, vget_low_s8(vout4x01234567_5x01234567)); c4 += 8;
401         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
402         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
403         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
404         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
405         vout4x01234567_5x01234567 = vcombine_s8(vget_high_s8(vout4x0123456789ABCDEF), vget_high_s8(vout5x0123456789ABCDEF));
406         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
407         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
408       }
409       if (nc & 4) {
410         vst1q_lane_u32(__builtin_assume_aligned(c5, 1), vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
411         vst1q_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
412         vst1q_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
413         vst1q_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
414         vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
415         vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
416         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
417         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
418         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
419       }
420       if (nc & 2) {
421         vst1q_lane_u16(__builtin_assume_aligned(c5, 1), vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
422         vst1q_lane_u16(__builtin_assume_aligned(c4, 1), vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
423         vst1q_lane_u16(__builtin_assume_aligned(c3, 1), vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
424         vst1q_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
425         vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
426         vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
427         vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
428         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
429         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
430       }
431       if (nc & 1) {
432         vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
433         vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
434         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
435         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
436         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
437         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
438       }
439 
440       nc = 0;
441     }
442   } while (nc != 0);
443 }
444