• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c8-neon-mull-padal.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/igemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal(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_3x16c8__neon_mlal_padal(
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 <= 3);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (3 * 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, 8);
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 
54   do {
55     int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56     int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57     int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58     int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59     int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60     int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61     int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62     int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63     int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64     int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65     int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66     int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
67     int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
68     int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
69     int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
70     int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
71     int32x4_t vacc1x0 = vacc0x0;
72     int32x4_t vacc1x1 = vacc0x1;
73     int32x4_t vacc1x2 = vacc0x2;
74     int32x4_t vacc1x3 = vacc0x3;
75     int32x4_t vacc1x4 = vacc0x4;
76     int32x4_t vacc1x5 = vacc0x5;
77     int32x4_t vacc1x6 = vacc0x6;
78     int32x4_t vacc1x7 = vacc0x7;
79     int32x4_t vacc1x8 = vacc0x8;
80     int32x4_t vacc1x9 = vacc0x9;
81     int32x4_t vacc1x10 = vacc0x10;
82     int32x4_t vacc1x11 = vacc0x11;
83     int32x4_t vacc1x12 = vacc0x12;
84     int32x4_t vacc1x13 = vacc0x13;
85     int32x4_t vacc1x14 = vacc0x14;
86     int32x4_t vacc1x15 = vacc0x15;
87     int32x4_t vacc2x0 = vacc0x0;
88     int32x4_t vacc2x1 = vacc0x1;
89     int32x4_t vacc2x2 = vacc0x2;
90     int32x4_t vacc2x3 = vacc0x3;
91     int32x4_t vacc2x4 = vacc0x4;
92     int32x4_t vacc2x5 = vacc0x5;
93     int32x4_t vacc2x6 = vacc0x6;
94     int32x4_t vacc2x7 = vacc0x7;
95     int32x4_t vacc2x8 = vacc0x8;
96     int32x4_t vacc2x9 = vacc0x9;
97     int32x4_t vacc2x10 = vacc0x10;
98     int32x4_t vacc2x11 = vacc0x11;
99     int32x4_t vacc2x12 = vacc0x12;
100     int32x4_t vacc2x13 = vacc0x13;
101     int32x4_t vacc2x14 = vacc0x14;
102     int32x4_t vacc2x15 = vacc0x15;
103 
104     size_t p = ks;
105     do {
106       const int8_t* restrict a0 = a[0];
107       if XNN_UNPREDICTABLE(a0 != zero) {
108         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
109       }
110       const int8_t* restrict a1 = a[1];
111       if XNN_UNPREDICTABLE(a1 != zero) {
112         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
113       }
114       const int8_t* restrict a2 = a[2];
115       if XNN_UNPREDICTABLE(a2 != zero) {
116         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
117       }
118       a += 3;
119 
120       size_t k = kc;
121       // 2x partial unrolled loop to load 16 bytes at a time using MLA.
122       while (k >= 16 * sizeof(int8_t)) {
123         const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
124         const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
125         const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
126         const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
127         const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
128         const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
129 
130         const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
131         const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
132         const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
133         const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
134         const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
135         const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
136         const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
137         const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
138         const int8x8_t vb8x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
139         const int8x8_t vb9x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
140         const int8x8_t vb10x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
141         const int8x8_t vb11x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
142         const int8x8_t vb12x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
143         const int8x8_t vb13x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
144         const int8x8_t vb14x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
145         const int8x8_t vb15x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
146 
147         const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
148         int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
149         int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
150         int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0);
151         vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
152         vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
153         vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1);
154         vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
155         vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
156         vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
157         const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
158         int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
159         int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
160         int16x8_t vprod2x1 = vmull_s8(vb1x0, va2x0);
161         vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
162         vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
163         vprod2x1 = vmlal_s8(vprod2x1, vb1x1, va2x1);
164         vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
165         vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
166         vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
167         const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
168         int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
169         int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
170         int16x8_t vprod2x2 = vmull_s8(vb2x0, va2x0);
171         vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
172         vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
173         vprod2x2 = vmlal_s8(vprod2x2, vb2x1, va2x1);
174         vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
175         vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
176         vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
177         const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
178         int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
179         int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
180         int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0);
181         vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
182         vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
183         vprod2x3 = vmlal_s8(vprod2x3, vb3x1, va2x1);
184         vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
185         vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
186         vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
187         const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
188         int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
189         int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
190         int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0);
191         vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
192         vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
193         vprod2x4 = vmlal_s8(vprod2x4, vb4x1, va2x1);
194         vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
195         vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
196         vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
197         const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
198         int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
199         int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
200         int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0);
201         vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
202         vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
203         vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1);
204         vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
205         vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
206         vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
207         const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
208         int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
209         int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
210         int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0);
211         vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
212         vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
213         vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1);
214         vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
215         vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
216         vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
217         const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
218         int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
219         int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
220         int16x8_t vprod2x7 = vmull_s8(vb7x0, va2x0);
221         vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
222         vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
223         vprod2x7 = vmlal_s8(vprod2x7, vb7x1, va2x1);
224         vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
225         vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
226         vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
227         const int8x8_t vb8x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
228         int16x8_t vprod0x8 = vmull_s8(vb8x0, va0x0);
229         int16x8_t vprod1x8 = vmull_s8(vb8x0, va1x0);
230         int16x8_t vprod2x8 = vmull_s8(vb8x0, va2x0);
231         vprod0x8 = vmlal_s8(vprod0x8, vb8x1, va0x1);
232         vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1);
233         vprod2x8 = vmlal_s8(vprod2x8, vb8x1, va2x1);
234         vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
235         vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
236         vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
237         const int8x8_t vb9x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
238         int16x8_t vprod0x9 = vmull_s8(vb9x0, va0x0);
239         int16x8_t vprod1x9 = vmull_s8(vb9x0, va1x0);
240         int16x8_t vprod2x9 = vmull_s8(vb9x0, va2x0);
241         vprod0x9 = vmlal_s8(vprod0x9, vb9x1, va0x1);
242         vprod1x9 = vmlal_s8(vprod1x9, vb9x1, va1x1);
243         vprod2x9 = vmlal_s8(vprod2x9, vb9x1, va2x1);
244         vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
245         vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
246         vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
247         const int8x8_t vb10x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
248         int16x8_t vprod0x10 = vmull_s8(vb10x0, va0x0);
249         int16x8_t vprod1x10 = vmull_s8(vb10x0, va1x0);
250         int16x8_t vprod2x10 = vmull_s8(vb10x0, va2x0);
251         vprod0x10 = vmlal_s8(vprod0x10, vb10x1, va0x1);
252         vprod1x10 = vmlal_s8(vprod1x10, vb10x1, va1x1);
253         vprod2x10 = vmlal_s8(vprod2x10, vb10x1, va2x1);
254         vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
255         vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
256         vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
257         const int8x8_t vb11x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
258         int16x8_t vprod0x11 = vmull_s8(vb11x0, va0x0);
259         int16x8_t vprod1x11 = vmull_s8(vb11x0, va1x0);
260         int16x8_t vprod2x11 = vmull_s8(vb11x0, va2x0);
261         vprod0x11 = vmlal_s8(vprod0x11, vb11x1, va0x1);
262         vprod1x11 = vmlal_s8(vprod1x11, vb11x1, va1x1);
263         vprod2x11 = vmlal_s8(vprod2x11, vb11x1, va2x1);
264         vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
265         vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
266         vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
267         const int8x8_t vb12x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
268         int16x8_t vprod0x12 = vmull_s8(vb12x0, va0x0);
269         int16x8_t vprod1x12 = vmull_s8(vb12x0, va1x0);
270         int16x8_t vprod2x12 = vmull_s8(vb12x0, va2x0);
271         vprod0x12 = vmlal_s8(vprod0x12, vb12x1, va0x1);
272         vprod1x12 = vmlal_s8(vprod1x12, vb12x1, va1x1);
273         vprod2x12 = vmlal_s8(vprod2x12, vb12x1, va2x1);
274         vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
275         vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
276         vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
277         const int8x8_t vb13x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
278         int16x8_t vprod0x13 = vmull_s8(vb13x0, va0x0);
279         int16x8_t vprod1x13 = vmull_s8(vb13x0, va1x0);
280         int16x8_t vprod2x13 = vmull_s8(vb13x0, va2x0);
281         vprod0x13 = vmlal_s8(vprod0x13, vb13x1, va0x1);
282         vprod1x13 = vmlal_s8(vprod1x13, vb13x1, va1x1);
283         vprod2x13 = vmlal_s8(vprod2x13, vb13x1, va2x1);
284         vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
285         vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
286         vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
287         const int8x8_t vb14x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
288         int16x8_t vprod0x14 = vmull_s8(vb14x0, va0x0);
289         int16x8_t vprod1x14 = vmull_s8(vb14x0, va1x0);
290         int16x8_t vprod2x14 = vmull_s8(vb14x0, va2x0);
291         vprod0x14 = vmlal_s8(vprod0x14, vb14x1, va0x1);
292         vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1);
293         vprod2x14 = vmlal_s8(vprod2x14, vb14x1, va2x1);
294         vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
295         vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
296         vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
297         const int8x8_t vb15x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
298         int16x8_t vprod0x15 = vmull_s8(vb15x0, va0x0);
299         int16x8_t vprod1x15 = vmull_s8(vb15x0, va1x0);
300         int16x8_t vprod2x15 = vmull_s8(vb15x0, va2x0);
301         vprod0x15 = vmlal_s8(vprod0x15, vb15x1, va0x1);
302         vprod1x15 = vmlal_s8(vprod1x15, vb15x1, va1x1);
303         vprod2x15 = vmlal_s8(vprod2x15, vb15x1, va2x1);
304         vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
305         vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
306         vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
307 
308         k -= 16 * sizeof(int8_t);
309       }
310 
311       // Handle 8 bytes at a time using MUL.
312       if (k > 0) {
313         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
314         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
315         const int8x8_t va2 = vld1_s8(a2); a2 += 8;
316 
317         const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
318         const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
319         const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
320         const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
321         vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
322         vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
323         vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
324         const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
325         const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
326         const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
327         const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
328         vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
329         vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
330         vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
331         const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
332         const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
333         const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
334         const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
335         vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
336         vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
337         vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
338         const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
339         const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
340         const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
341         const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
342         vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
343         vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
344         vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
345         const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
346         const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
347         const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
348         const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
349         vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
350         vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
351         vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
352         const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
353         const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
354         const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
355         const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
356         vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
357         vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
358         vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
359         const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
360         const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
361         const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
362         const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
363         vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
364         vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
365         vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
366         const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
367         const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
368         const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
369         const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
370         vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
371         vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
372         vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
373         const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
374         const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
375         const int16x8_t vprod1x8 = vmull_s8(vb8, va1);
376         const int16x8_t vprod2x8 = vmull_s8(vb8, va2);
377         vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
378         vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
379         vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
380         const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
381         const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
382         const int16x8_t vprod1x9 = vmull_s8(vb9, va1);
383         const int16x8_t vprod2x9 = vmull_s8(vb9, va2);
384         vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
385         vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
386         vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
387         const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
388         const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
389         const int16x8_t vprod1x10 = vmull_s8(vb10, va1);
390         const int16x8_t vprod2x10 = vmull_s8(vb10, va2);
391         vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
392         vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
393         vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
394         const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
395         const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
396         const int16x8_t vprod1x11 = vmull_s8(vb11, va1);
397         const int16x8_t vprod2x11 = vmull_s8(vb11, va2);
398         vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
399         vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
400         vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
401         const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
402         const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
403         const int16x8_t vprod1x12 = vmull_s8(vb12, va1);
404         const int16x8_t vprod2x12 = vmull_s8(vb12, va2);
405         vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
406         vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
407         vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
408         const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
409         const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
410         const int16x8_t vprod1x13 = vmull_s8(vb13, va1);
411         const int16x8_t vprod2x13 = vmull_s8(vb13, va2);
412         vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
413         vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
414         vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
415         const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
416         const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
417         const int16x8_t vprod1x14 = vmull_s8(vb14, va1);
418         const int16x8_t vprod2x14 = vmull_s8(vb14, va2);
419         vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
420         vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
421         vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
422         const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
423         const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
424         const int16x8_t vprod1x15 = vmull_s8(vb15, va1);
425         const int16x8_t vprod2x15 = vmull_s8(vb15, va2);
426         vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
427         vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
428         vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
429 
430         k -= 8 * sizeof(int8_t);
431       }
432 
433       p -= 3 * sizeof(void*);
434     } while (p != 0);
435 
436 #if XNN_ARCH_ARM64
437     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
438     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
439     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
440     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
441     const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
442     const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
443     const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
444     const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
445     const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
446     const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
447     const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
448     const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
449     const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9);
450     const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11);
451     const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13);
452     const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, vacc1x15);
453     const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
454     const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
455     const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
456     const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
457     const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9);
458     const int32x4_t vsum2xAB = vpaddq_s32(vacc2x10, vacc2x11);
459     const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13);
460     const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15);
461     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
462     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
463     int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
464     int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
465     int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
466     int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
467     int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB);
468     int32x4_t vacc1xCDEF = vpaddq_s32(vsum1xCD, vsum1xEF);
469     int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
470     int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
471     int32x4_t vacc2x89AB = vpaddq_s32(vsum2x89, vsum2xAB);
472     int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF);
473 #else
474     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
475     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
476     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
477     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
478     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
479     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
480     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
481     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
482     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
483     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
484     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
485     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
486     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
487     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
488     const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
489     const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
490     const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
491     const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
492     const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
493     const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
494     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
495     const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
496     const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
497     const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
498     const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
499     const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
500     const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
501     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
502     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
503     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
504     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
505     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
506     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
507     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
508     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
509     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
510     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
511     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
512     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
513     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
514     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
515     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
516     const int32x2_t vpsum1x8 = vadd_s32(vget_low_s32(vacc1x8), vget_high_s32(vacc1x8));
517     const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9));
518     const int32x2_t vpsum1xA = vadd_s32(vget_low_s32(vacc1x10), vget_high_s32(vacc1x10));
519     const int32x2_t vpsum1xB = vadd_s32(vget_low_s32(vacc1x11), vget_high_s32(vacc1x11));
520     const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9);
521     const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB);
522     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB );
523     const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12));
524     const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13));
525     const int32x2_t vpsum1xE = vadd_s32(vget_low_s32(vacc1x14), vget_high_s32(vacc1x14));
526     const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15));
527     const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD);
528     const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF);
529     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF );
530     const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
531     const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
532     const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
533     const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
534     const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
535     const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
536     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
537     const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
538     const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
539     const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
540     const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
541     const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
542     const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
543     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
544     const int32x2_t vpsum2x8 = vadd_s32(vget_low_s32(vacc2x8), vget_high_s32(vacc2x8));
545     const int32x2_t vpsum2x9 = vadd_s32(vget_low_s32(vacc2x9), vget_high_s32(vacc2x9));
546     const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10));
547     const int32x2_t vpsum2xB = vadd_s32(vget_low_s32(vacc2x11), vget_high_s32(vacc2x11));
548     const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9);
549     const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB);
550     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB );
551     const int32x2_t vpsum2xC = vadd_s32(vget_low_s32(vacc2x12), vget_high_s32(vacc2x12));
552     const int32x2_t vpsum2xD = vadd_s32(vget_low_s32(vacc2x13), vget_high_s32(vacc2x13));
553     const int32x2_t vpsum2xE = vadd_s32(vget_low_s32(vacc2x14), vget_high_s32(vacc2x14));
554     const int32x2_t vpsum2xF = vadd_s32(vget_low_s32(vacc2x15), vget_high_s32(vacc2x15));
555     const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD);
556     const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF);
557     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF );
558 #endif
559 
560     const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
561     vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
562     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
563     vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);
564     vacc0xCDEF = vqrdmulhq_s32(vacc0xCDEF, vmultiplier);
565     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
566     vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
567     vacc1x89AB = vqrdmulhq_s32(vacc1x89AB, vmultiplier);
568     vacc1xCDEF = vqrdmulhq_s32(vacc1xCDEF, vmultiplier);
569     vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier);
570     vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier);
571     vacc2x89AB = vqrdmulhq_s32(vacc2x89AB, vmultiplier);
572     vacc2xCDEF = vqrdmulhq_s32(vacc2xCDEF, vmultiplier);
573 
574     const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
575     const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
576     vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
577     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
578     vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);
579     vacc0xCDEF = vsraq_n_s32(vacc0xCDEF, vbicq_s32(vacc0xCDEF, vzero_shift_mask), 31);
580     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
581     vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
582     vacc1x89AB = vsraq_n_s32(vacc1x89AB, vbicq_s32(vacc1x89AB, vzero_shift_mask), 31);
583     vacc1xCDEF = vsraq_n_s32(vacc1xCDEF, vbicq_s32(vacc1xCDEF, vzero_shift_mask), 31);
584     vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
585     vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
586     vacc2x89AB = vsraq_n_s32(vacc2x89AB, vbicq_s32(vacc2x89AB, vzero_shift_mask), 31);
587     vacc2xCDEF = vsraq_n_s32(vacc2xCDEF, vbicq_s32(vacc2xCDEF, vzero_shift_mask), 31);
588 
589     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
590     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
591     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);
592     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_shift);
593     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
594     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
595     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_shift);
596     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_shift);
597     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
598     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
599     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_shift);
600     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_shift);
601 
602     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
603 #if XNN_ARCH_ARM64
604     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
605     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
606     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
607     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
608     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
609     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
610     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
611     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
612     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
613 #else
614     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
615     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
616     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
617     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
618     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
619     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
620 
621     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
622     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
623     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
624 #endif
625     const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
626     const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
627 
628     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
629     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
630     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
631 
632     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
633     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
634     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
635 
636     if (nc >= 16) {
637       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
638       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
639       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
640 
641       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
642       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
643       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
644 
645       a = (const int8_t**restrict) ((uintptr_t) a - ks);
646 
647       nc -= 16;
648     } else {
649       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
650       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
651       if (nc & 8) {
652         vst1_s8(c2, vout2x01234567); c2 += 8;
653         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
654         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
655         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
656         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
657       }
658       if (nc & 4) {
659         vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
660         vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
661         vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
662         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
663         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
664       }
665       if (nc & 2) {
666         vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
667         vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
668         vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
669         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
670         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
671       }
672       if (nc & 1) {
673         vst1_lane_s8(c2, vout2x01234567, 0);
674         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
675         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
676       }
677 
678       nc = 0;
679     }
680   } while (nc != 0);
681 }
682