1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/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/gemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal(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_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal(
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_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
29 {
30   assert(mr != 0);
31   assert(mr <= 3);
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, 8);
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 
55   do {
56     int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57     int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58     int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59     int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60     int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61     int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62     int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63     int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64     int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65     int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66     int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
67     int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
68     int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
69     int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
70     int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
71     int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
72     int32x4_t vacc1x0 = vacc0x0;
73     int32x4_t vacc1x1 = vacc0x1;
74     int32x4_t vacc1x2 = vacc0x2;
75     int32x4_t vacc1x3 = vacc0x3;
76     int32x4_t vacc1x4 = vacc0x4;
77     int32x4_t vacc1x5 = vacc0x5;
78     int32x4_t vacc1x6 = vacc0x6;
79     int32x4_t vacc1x7 = vacc0x7;
80     int32x4_t vacc1x8 = vacc0x8;
81     int32x4_t vacc1x9 = vacc0x9;
82     int32x4_t vacc1x10 = vacc0x10;
83     int32x4_t vacc1x11 = vacc0x11;
84     int32x4_t vacc1x12 = vacc0x12;
85     int32x4_t vacc1x13 = vacc0x13;
86     int32x4_t vacc1x14 = vacc0x14;
87     int32x4_t vacc1x15 = vacc0x15;
88     int32x4_t vacc2x0 = vacc0x0;
89     int32x4_t vacc2x1 = vacc0x1;
90     int32x4_t vacc2x2 = vacc0x2;
91     int32x4_t vacc2x3 = vacc0x3;
92     int32x4_t vacc2x4 = vacc0x4;
93     int32x4_t vacc2x5 = vacc0x5;
94     int32x4_t vacc2x6 = vacc0x6;
95     int32x4_t vacc2x7 = vacc0x7;
96     int32x4_t vacc2x8 = vacc0x8;
97     int32x4_t vacc2x9 = vacc0x9;
98     int32x4_t vacc2x10 = vacc0x10;
99     int32x4_t vacc2x11 = vacc0x11;
100     int32x4_t vacc2x12 = vacc0x12;
101     int32x4_t vacc2x13 = vacc0x13;
102     int32x4_t vacc2x14 = vacc0x14;
103     int32x4_t vacc2x15 = vacc0x15;
104 
105     size_t k = kc;
106     // 2x partial unrolled loop to load 16 bytes at a time using MLA.
107     while (k >= 16 * sizeof(int8_t)) {
108       const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
109       const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
110       const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
111       const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
112       const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
113       const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
114 
115       const int8x8_t vb0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
116       const int8x8_t vb1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
117       const int8x8_t vb2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
118       const int8x8_t vb3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
119       const int8x8_t vb4x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
120       const int8x8_t vb5x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
121       const int8x8_t vb6x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
122       const int8x8_t vb7x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
123       const int8x8_t vb8x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
124       const int8x8_t vb9x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
125       const int8x8_t vb10x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
126       const int8x8_t vb11x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
127       const int8x8_t vb12x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
128       const int8x8_t vb13x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
129       const int8x8_t vb14x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
130       const int8x8_t vb15x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
131 
132       const int8x8_t vb0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
133       int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0);
134       int16x8_t vprod1x0 = vmull_s8(vb0x0, va1x0);
135       int16x8_t vprod2x0 = vmull_s8(vb0x0, va2x0);
136       vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1);
137       vprod1x0 = vmlal_s8(vprod1x0, vb0x1, va1x1);
138       vprod2x0 = vmlal_s8(vprod2x0, vb0x1, va2x1);
139       vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
140       vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
141       vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
142       const int8x8_t vb1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
143       int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0);
144       int16x8_t vprod1x1 = vmull_s8(vb1x0, va1x0);
145       int16x8_t vprod2x1 = vmull_s8(vb1x0, va2x0);
146       vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1);
147       vprod1x1 = vmlal_s8(vprod1x1, vb1x1, va1x1);
148       vprod2x1 = vmlal_s8(vprod2x1, vb1x1, va2x1);
149       vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
150       vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
151       vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
152       const int8x8_t vb2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
153       int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0);
154       int16x8_t vprod1x2 = vmull_s8(vb2x0, va1x0);
155       int16x8_t vprod2x2 = vmull_s8(vb2x0, va2x0);
156       vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1);
157       vprod1x2 = vmlal_s8(vprod1x2, vb2x1, va1x1);
158       vprod2x2 = vmlal_s8(vprod2x2, vb2x1, va2x1);
159       vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
160       vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
161       vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
162       const int8x8_t vb3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
163       int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0);
164       int16x8_t vprod1x3 = vmull_s8(vb3x0, va1x0);
165       int16x8_t vprod2x3 = vmull_s8(vb3x0, va2x0);
166       vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1);
167       vprod1x3 = vmlal_s8(vprod1x3, vb3x1, va1x1);
168       vprod2x3 = vmlal_s8(vprod2x3, vb3x1, va2x1);
169       vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
170       vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
171       vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
172       const int8x8_t vb4x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
173       int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0);
174       int16x8_t vprod1x4 = vmull_s8(vb4x0, va1x0);
175       int16x8_t vprod2x4 = vmull_s8(vb4x0, va2x0);
176       vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1);
177       vprod1x4 = vmlal_s8(vprod1x4, vb4x1, va1x1);
178       vprod2x4 = vmlal_s8(vprod2x4, vb4x1, va2x1);
179       vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
180       vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
181       vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
182       const int8x8_t vb5x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
183       int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0);
184       int16x8_t vprod1x5 = vmull_s8(vb5x0, va1x0);
185       int16x8_t vprod2x5 = vmull_s8(vb5x0, va2x0);
186       vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1);
187       vprod1x5 = vmlal_s8(vprod1x5, vb5x1, va1x1);
188       vprod2x5 = vmlal_s8(vprod2x5, vb5x1, va2x1);
189       vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
190       vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
191       vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
192       const int8x8_t vb6x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
193       int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0);
194       int16x8_t vprod1x6 = vmull_s8(vb6x0, va1x0);
195       int16x8_t vprod2x6 = vmull_s8(vb6x0, va2x0);
196       vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1);
197       vprod1x6 = vmlal_s8(vprod1x6, vb6x1, va1x1);
198       vprod2x6 = vmlal_s8(vprod2x6, vb6x1, va2x1);
199       vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
200       vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
201       vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
202       const int8x8_t vb7x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
203       int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0);
204       int16x8_t vprod1x7 = vmull_s8(vb7x0, va1x0);
205       int16x8_t vprod2x7 = vmull_s8(vb7x0, va2x0);
206       vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1);
207       vprod1x7 = vmlal_s8(vprod1x7, vb7x1, va1x1);
208       vprod2x7 = vmlal_s8(vprod2x7, vb7x1, va2x1);
209       vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
210       vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
211       vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
212       const int8x8_t vb8x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
213       int16x8_t vprod0x8 = vmull_s8(vb8x0, va0x0);
214       int16x8_t vprod1x8 = vmull_s8(vb8x0, va1x0);
215       int16x8_t vprod2x8 = vmull_s8(vb8x0, va2x0);
216       vprod0x8 = vmlal_s8(vprod0x8, vb8x1, va0x1);
217       vprod1x8 = vmlal_s8(vprod1x8, vb8x1, va1x1);
218       vprod2x8 = vmlal_s8(vprod2x8, vb8x1, va2x1);
219       vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
220       vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
221       vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
222       const int8x8_t vb9x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
223       int16x8_t vprod0x9 = vmull_s8(vb9x0, va0x0);
224       int16x8_t vprod1x9 = vmull_s8(vb9x0, va1x0);
225       int16x8_t vprod2x9 = vmull_s8(vb9x0, va2x0);
226       vprod0x9 = vmlal_s8(vprod0x9, vb9x1, va0x1);
227       vprod1x9 = vmlal_s8(vprod1x9, vb9x1, va1x1);
228       vprod2x9 = vmlal_s8(vprod2x9, vb9x1, va2x1);
229       vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
230       vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
231       vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
232       const int8x8_t vb10x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
233       int16x8_t vprod0x10 = vmull_s8(vb10x0, va0x0);
234       int16x8_t vprod1x10 = vmull_s8(vb10x0, va1x0);
235       int16x8_t vprod2x10 = vmull_s8(vb10x0, va2x0);
236       vprod0x10 = vmlal_s8(vprod0x10, vb10x1, va0x1);
237       vprod1x10 = vmlal_s8(vprod1x10, vb10x1, va1x1);
238       vprod2x10 = vmlal_s8(vprod2x10, vb10x1, va2x1);
239       vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
240       vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
241       vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
242       const int8x8_t vb11x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
243       int16x8_t vprod0x11 = vmull_s8(vb11x0, va0x0);
244       int16x8_t vprod1x11 = vmull_s8(vb11x0, va1x0);
245       int16x8_t vprod2x11 = vmull_s8(vb11x0, va2x0);
246       vprod0x11 = vmlal_s8(vprod0x11, vb11x1, va0x1);
247       vprod1x11 = vmlal_s8(vprod1x11, vb11x1, va1x1);
248       vprod2x11 = vmlal_s8(vprod2x11, vb11x1, va2x1);
249       vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
250       vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
251       vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
252       const int8x8_t vb12x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
253       int16x8_t vprod0x12 = vmull_s8(vb12x0, va0x0);
254       int16x8_t vprod1x12 = vmull_s8(vb12x0, va1x0);
255       int16x8_t vprod2x12 = vmull_s8(vb12x0, va2x0);
256       vprod0x12 = vmlal_s8(vprod0x12, vb12x1, va0x1);
257       vprod1x12 = vmlal_s8(vprod1x12, vb12x1, va1x1);
258       vprod2x12 = vmlal_s8(vprod2x12, vb12x1, va2x1);
259       vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
260       vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
261       vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
262       const int8x8_t vb13x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
263       int16x8_t vprod0x13 = vmull_s8(vb13x0, va0x0);
264       int16x8_t vprod1x13 = vmull_s8(vb13x0, va1x0);
265       int16x8_t vprod2x13 = vmull_s8(vb13x0, va2x0);
266       vprod0x13 = vmlal_s8(vprod0x13, vb13x1, va0x1);
267       vprod1x13 = vmlal_s8(vprod1x13, vb13x1, va1x1);
268       vprod2x13 = vmlal_s8(vprod2x13, vb13x1, va2x1);
269       vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
270       vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
271       vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
272       const int8x8_t vb14x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
273       int16x8_t vprod0x14 = vmull_s8(vb14x0, va0x0);
274       int16x8_t vprod1x14 = vmull_s8(vb14x0, va1x0);
275       int16x8_t vprod2x14 = vmull_s8(vb14x0, va2x0);
276       vprod0x14 = vmlal_s8(vprod0x14, vb14x1, va0x1);
277       vprod1x14 = vmlal_s8(vprod1x14, vb14x1, va1x1);
278       vprod2x14 = vmlal_s8(vprod2x14, vb14x1, va2x1);
279       vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
280       vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
281       vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
282       const int8x8_t vb15x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(  int8_t));
283       int16x8_t vprod0x15 = vmull_s8(vb15x0, va0x0);
284       int16x8_t vprod1x15 = vmull_s8(vb15x0, va1x0);
285       int16x8_t vprod2x15 = vmull_s8(vb15x0, va2x0);
286       vprod0x15 = vmlal_s8(vprod0x15, vb15x1, va0x1);
287       vprod1x15 = vmlal_s8(vprod1x15, vb15x1, va1x1);
288       vprod2x15 = vmlal_s8(vprod2x15, vb15x1, va2x1);
289       vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
290       vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
291       vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
292 
293       k -= 16 * sizeof(int8_t);
294     }
295 
296     // Handle 8 bytes at a time using MUL.
297     if (k > 0) {
298       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
299       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
300       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
301 
302       const int8x8_t vb0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
303       const int16x8_t vprod0x0 = vmull_s8(vb0, va0);
304       const int16x8_t vprod1x0 = vmull_s8(vb0, va1);
305       const int16x8_t vprod2x0 = vmull_s8(vb0, va2);
306       vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
307       vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
308       vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
309       const int8x8_t vb1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
310       const int16x8_t vprod0x1 = vmull_s8(vb1, va0);
311       const int16x8_t vprod1x1 = vmull_s8(vb1, va1);
312       const int16x8_t vprod2x1 = vmull_s8(vb1, va2);
313       vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
314       vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
315       vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
316       const int8x8_t vb2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
317       const int16x8_t vprod0x2 = vmull_s8(vb2, va0);
318       const int16x8_t vprod1x2 = vmull_s8(vb2, va1);
319       const int16x8_t vprod2x2 = vmull_s8(vb2, va2);
320       vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
321       vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
322       vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
323       const int8x8_t vb3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
324       const int16x8_t vprod0x3 = vmull_s8(vb3, va0);
325       const int16x8_t vprod1x3 = vmull_s8(vb3, va1);
326       const int16x8_t vprod2x3 = vmull_s8(vb3, va2);
327       vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
328       vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
329       vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
330       const int8x8_t vb4 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
331       const int16x8_t vprod0x4 = vmull_s8(vb4, va0);
332       const int16x8_t vprod1x4 = vmull_s8(vb4, va1);
333       const int16x8_t vprod2x4 = vmull_s8(vb4, va2);
334       vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
335       vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
336       vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
337       const int8x8_t vb5 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
338       const int16x8_t vprod0x5 = vmull_s8(vb5, va0);
339       const int16x8_t vprod1x5 = vmull_s8(vb5, va1);
340       const int16x8_t vprod2x5 = vmull_s8(vb5, va2);
341       vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
342       vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
343       vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
344       const int8x8_t vb6 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
345       const int16x8_t vprod0x6 = vmull_s8(vb6, va0);
346       const int16x8_t vprod1x6 = vmull_s8(vb6, va1);
347       const int16x8_t vprod2x6 = vmull_s8(vb6, va2);
348       vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
349       vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
350       vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
351       const int8x8_t vb7 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
352       const int16x8_t vprod0x7 = vmull_s8(vb7, va0);
353       const int16x8_t vprod1x7 = vmull_s8(vb7, va1);
354       const int16x8_t vprod2x7 = vmull_s8(vb7, va2);
355       vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
356       vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
357       vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
358       const int8x8_t vb8 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
359       const int16x8_t vprod0x8 = vmull_s8(vb8, va0);
360       const int16x8_t vprod1x8 = vmull_s8(vb8, va1);
361       const int16x8_t vprod2x8 = vmull_s8(vb8, va2);
362       vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
363       vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
364       vacc2x8 = vpadalq_s16(vacc2x8, vprod2x8);
365       const int8x8_t vb9 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
366       const int16x8_t vprod0x9 = vmull_s8(vb9, va0);
367       const int16x8_t vprod1x9 = vmull_s8(vb9, va1);
368       const int16x8_t vprod2x9 = vmull_s8(vb9, va2);
369       vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
370       vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
371       vacc2x9 = vpadalq_s16(vacc2x9, vprod2x9);
372       const int8x8_t vb10 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
373       const int16x8_t vprod0x10 = vmull_s8(vb10, va0);
374       const int16x8_t vprod1x10 = vmull_s8(vb10, va1);
375       const int16x8_t vprod2x10 = vmull_s8(vb10, va2);
376       vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
377       vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
378       vacc2x10 = vpadalq_s16(vacc2x10, vprod2x10);
379       const int8x8_t vb11 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
380       const int16x8_t vprod0x11 = vmull_s8(vb11, va0);
381       const int16x8_t vprod1x11 = vmull_s8(vb11, va1);
382       const int16x8_t vprod2x11 = vmull_s8(vb11, va2);
383       vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
384       vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
385       vacc2x11 = vpadalq_s16(vacc2x11, vprod2x11);
386       const int8x8_t vb12 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
387       const int16x8_t vprod0x12 = vmull_s8(vb12, va0);
388       const int16x8_t vprod1x12 = vmull_s8(vb12, va1);
389       const int16x8_t vprod2x12 = vmull_s8(vb12, va2);
390       vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
391       vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
392       vacc2x12 = vpadalq_s16(vacc2x12, vprod2x12);
393       const int8x8_t vb13 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
394       const int16x8_t vprod0x13 = vmull_s8(vb13, va0);
395       const int16x8_t vprod1x13 = vmull_s8(vb13, va1);
396       const int16x8_t vprod2x13 = vmull_s8(vb13, va2);
397       vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
398       vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
399       vacc2x13 = vpadalq_s16(vacc2x13, vprod2x13);
400       const int8x8_t vb14 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
401       const int16x8_t vprod0x14 = vmull_s8(vb14, va0);
402       const int16x8_t vprod1x14 = vmull_s8(vb14, va1);
403       const int16x8_t vprod2x14 = vmull_s8(vb14, va2);
404       vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
405       vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
406       vacc2x14 = vpadalq_s16(vacc2x14, vprod2x14);
407       const int8x8_t vb15 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
408       const int16x8_t vprod0x15 = vmull_s8(vb15, va0);
409       const int16x8_t vprod1x15 = vmull_s8(vb15, va1);
410       const int16x8_t vprod2x15 = vmull_s8(vb15, va2);
411       vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
412       vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
413       vacc2x15 = vpadalq_s16(vacc2x15, vprod2x15);
414 
415       k -= 8 * sizeof(int8_t);
416     }
417 
418 #if XNN_ARCH_ARM64
419     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
420     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
421     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
422     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
423     const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
424     const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
425     const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
426     const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
427     const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
428     const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
429     const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
430     const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
431     const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9);
432     const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11);
433     const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13);
434     const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, vacc1x15);
435     const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
436     const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
437     const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
438     const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
439     const int32x4_t vsum2x89 = vpaddq_s32(vacc2x8, vacc2x9);
440     const int32x4_t vsum2xAB = vpaddq_s32(vacc2x10, vacc2x11);
441     const int32x4_t vsum2xCD = vpaddq_s32(vacc2x12, vacc2x13);
442     const int32x4_t vsum2xEF = vpaddq_s32(vacc2x14, vacc2x15);
443     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
444     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
445     int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
446     int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
447     int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
448     int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
449     int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB);
450     int32x4_t vacc1xCDEF = vpaddq_s32(vsum1xCD, vsum1xEF);
451     int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
452     int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
453     int32x4_t vacc2x89AB = vpaddq_s32(vsum2x89, vsum2xAB);
454     int32x4_t vacc2xCDEF = vpaddq_s32(vsum2xCD, vsum2xEF);
455 #else
456     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
457     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
458     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
459     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
460     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
461     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
462     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
463     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
464     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
465     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
466     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
467     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
468     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
469     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
470     const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
471     const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
472     const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
473     const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
474     const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
475     const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
476     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
477     const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
478     const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
479     const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
480     const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
481     const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
482     const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
483     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
484     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
485     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
486     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
487     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
488     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
489     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
490     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
491     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
492     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
493     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
494     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
495     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
496     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
497     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
498     const int32x2_t vpsum1x8 = vadd_s32(vget_low_s32(vacc1x8), vget_high_s32(vacc1x8));
499     const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9));
500     const int32x2_t vpsum1xA = vadd_s32(vget_low_s32(vacc1x10), vget_high_s32(vacc1x10));
501     const int32x2_t vpsum1xB = vadd_s32(vget_low_s32(vacc1x11), vget_high_s32(vacc1x11));
502     const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9);
503     const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB);
504     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB );
505     const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12));
506     const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13));
507     const int32x2_t vpsum1xE = vadd_s32(vget_low_s32(vacc1x14), vget_high_s32(vacc1x14));
508     const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15));
509     const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD);
510     const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF);
511     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF );
512     const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
513     const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
514     const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
515     const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
516     const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
517     const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
518     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
519     const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
520     const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
521     const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
522     const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
523     const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
524     const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
525     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
526     const int32x2_t vpsum2x8 = vadd_s32(vget_low_s32(vacc2x8), vget_high_s32(vacc2x8));
527     const int32x2_t vpsum2x9 = vadd_s32(vget_low_s32(vacc2x9), vget_high_s32(vacc2x9));
528     const int32x2_t vpsum2xA = vadd_s32(vget_low_s32(vacc2x10), vget_high_s32(vacc2x10));
529     const int32x2_t vpsum2xB = vadd_s32(vget_low_s32(vacc2x11), vget_high_s32(vacc2x11));
530     const int32x2_t vsum2x89 = vpadd_s32(vpsum2x8, vpsum2x9);
531     const int32x2_t vsum2xAB = vpadd_s32(vpsum2xA, vpsum2xB);
532     int32x4_t vacc2x89AB = vcombine_s32(vsum2x89, vsum2xAB );
533     const int32x2_t vpsum2xC = vadd_s32(vget_low_s32(vacc2x12), vget_high_s32(vacc2x12));
534     const int32x2_t vpsum2xD = vadd_s32(vget_low_s32(vacc2x13), vget_high_s32(vacc2x13));
535     const int32x2_t vpsum2xE = vadd_s32(vget_low_s32(vacc2x14), vget_high_s32(vacc2x14));
536     const int32x2_t vpsum2xF = vadd_s32(vget_low_s32(vacc2x15), vget_high_s32(vacc2x15));
537     const int32x2_t vsum2xCD = vpadd_s32(vpsum2xC, vpsum2xD);
538     const int32x2_t vsum2xEF = vpadd_s32(vpsum2xE, vpsum2xF);
539     int32x4_t vacc2xCDEF = vcombine_s32(vsum2xCD, vsum2xEF );
540 #endif
541 
542     const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
543     vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
544     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
545     vacc0x89AB = vqrdmulhq_s32(vacc0x89AB, vmultiplier);
546     vacc0xCDEF = vqrdmulhq_s32(vacc0xCDEF, vmultiplier);
547     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
548     vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
549     vacc1x89AB = vqrdmulhq_s32(vacc1x89AB, vmultiplier);
550     vacc1xCDEF = vqrdmulhq_s32(vacc1xCDEF, vmultiplier);
551     vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier);
552     vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier);
553     vacc2x89AB = vqrdmulhq_s32(vacc2x89AB, vmultiplier);
554     vacc2xCDEF = vqrdmulhq_s32(vacc2xCDEF, vmultiplier);
555 
556     const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
557     const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
558     vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
559     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
560     vacc0x89AB = vsraq_n_s32(vacc0x89AB, vbicq_s32(vacc0x89AB, vzero_shift_mask), 31);
561     vacc0xCDEF = vsraq_n_s32(vacc0xCDEF, vbicq_s32(vacc0xCDEF, vzero_shift_mask), 31);
562     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
563     vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
564     vacc1x89AB = vsraq_n_s32(vacc1x89AB, vbicq_s32(vacc1x89AB, vzero_shift_mask), 31);
565     vacc1xCDEF = vsraq_n_s32(vacc1xCDEF, vbicq_s32(vacc1xCDEF, vzero_shift_mask), 31);
566     vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
567     vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
568     vacc2x89AB = vsraq_n_s32(vacc2x89AB, vbicq_s32(vacc2x89AB, vzero_shift_mask), 31);
569     vacc2xCDEF = vsraq_n_s32(vacc2xCDEF, vbicq_s32(vacc2xCDEF, vzero_shift_mask), 31);
570 
571     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
572     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
573     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_shift);
574     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_shift);
575     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
576     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
577     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_shift);
578     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_shift);
579     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
580     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
581     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_shift);
582     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_shift);
583 
584     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
585 #if XNN_ARCH_ARM64
586     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
587     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
588     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
589     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
590     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
591     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
592     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
593     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
594     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
595 #else
596     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
597     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
598     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
599     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
600     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
601     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
602 
603     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
604     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
605     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
606 #endif
607     const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
608     const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
609 
610     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
611     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
612     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
613 
614     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
615     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
616     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
617 
618     if (nc >= 16) {
619       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
620       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
621       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
622 
623       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
624       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
625       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
626 
627       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
628       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
629       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
630 
631       nc -= 16;
632     } else {
633       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
634       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
635       if (nc & 8) {
636         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
637         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
638         vst1_s8(c2, vout2x01234567); c2 += 8;
639         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
640         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
641       }
642       if (nc & 4) {
643         vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
644         vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
645         vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
646         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
647         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
648       }
649       if (nc & 2) {
650         vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
651         vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
652         vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
653         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
654         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
655       }
656       if (nc & 1) {
657         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
658         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
659         vst1_lane_s8(c2, vout2x01234567, 0);
660       }
661 
662       nc = 0;
663     }
664   } while (nc != 0);
665 }
666