1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-padal-dup.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_gemm_minmax_ukernel_3x8c2__neon_mlal_padal_dup(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_3x8c2__neon_mlal_padal_dup(
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, 2);
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 vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
58     int32x4_t vacc1x0123 = vacc0x0123;
59     int32x4_t vacc1x4567 = vacc0x4567;
60     int32x4_t vacc2x0123 = vacc0x0123;
61     int32x4_t vacc2x4567 = vacc0x4567;
62 
63     size_t k = kc;
64 
65     while (k >= 16 * sizeof(int8_t)) {
66       const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
67       const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
68       const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
69       const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
70       const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
71       const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
72 
73       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81 
82       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0)));
83       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0)));
84       int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0)));
85       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));
87       vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));
88       vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0)));
89       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
90       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
91       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
92       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0)));
93       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0)));
94       int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0)));
95       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0)));
97       vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0)));
98       vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0)));
99       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
100       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
101       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
102       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1)));
103       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1)));
104       int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1)));
105       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));
107       vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));
108       vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1)));
109       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
110       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
111       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
112       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1)));
113       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1)));
114       int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1)));
115       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1)));
117       vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1)));
118       vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1)));
119       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
120       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
121       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
122       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2)));
123       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2)));
124       int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2)));
125       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));
127       vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));
128       vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2)));
129       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
130       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
131       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
132       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2)));
133       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2)));
134       int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2)));
135       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2)));
137       vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2)));
138       vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2)));
139       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
140       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
141       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
142       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3)));
143       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3)));
144       int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3)));
145       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));
147       vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));
148       vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3)));
149       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
150       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
151       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
152       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3)));
153       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3)));
154       int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3)));
155       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3)));
157       vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3)));
158       vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3)));
159       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
160       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
161       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
162 
163       k -= 16 * sizeof(int8_t);
164     }
165 
166     if (k >= 8 * sizeof(int8_t)) {
167       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
168       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
169       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
170 
171       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
175       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
179 
180       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
181       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
182       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
183       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
184       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
185       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
186       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
187       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
188       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
189       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
190       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
191       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3)));
192       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
193       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
194       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
195       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
196       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
197       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
198       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
199       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
200       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
201       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
202       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
203       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
204       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
205       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
206       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
207       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3)));
208       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
209       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
210       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
211       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
212       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
213       const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
214       const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
215       const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3)));
216       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
217       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
218       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
219       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
220       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
221       const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
222       const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
223       const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3)));
224       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
225       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
226       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
227       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
228 
229       k -= 8 * sizeof(int8_t);
230     }
231 
232     if XNN_UNLIKELY(k != 0) {
233       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
234       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
235       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
236 
237       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
238       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
239 
240       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
241       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
242       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0)));
243       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
244       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
245       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
246       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0)));
247       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
248       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
249       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
250       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0)));
251       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
252 
253       if (k > 2 * sizeof(int8_t)) {
254         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
255         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
256 
257         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
258         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
259         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1)));
260         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
261         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
262         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
263         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1)));
264         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
265         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
266         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
267         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1)));
268         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
269 
270         if (k > 4 * sizeof(int8_t)) {
271           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
272           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
273 
274           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
275           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
276           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2)));
277           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
278           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
279           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
280           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2)));
281           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
282           const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
283           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
284           const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2)));
285           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
286         }
287       }
288     }
289     const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
290     vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
291     vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
292     vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
293     vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
294     vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier);
295     vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier);
296 
297     const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
298     const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
299     vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
300     vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
301     vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
302     vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
303     vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
304     vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
305 
306     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
307     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
308     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
309     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
310     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
311     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
312 
313     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
314 #if XNN_ARCH_ARM64
315     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
316     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
317     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
318 
319     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
320     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
321 #else
322     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
323     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
324     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
325 
326     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
327     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
328 #endif
329     const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.output_min);
330     const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.output_max);
331 
332     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
333     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
334 
335     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
336     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
337 
338     if (nc >= 8) {
339       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
340       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
341       vst1_s8(c2 + 0, vout2x01234567);
342 
343       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
344       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
345       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
346 
347       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
348       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
349       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
350 
351       nc -= 8;
352     } else {
353       if (nc & 4) {
354         vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
355         vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
356         vst1_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
357         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
358         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
359       }
360       if (nc & 2) {
361         vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
362         vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
363         vst1_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
364         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
365         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
366       }
367       if (nc & 1) {
368         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
369         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
370         vst1_lane_s8(c2, vout2x01234567, 0);
371       }
372 
373       nc = 0;
374     }
375   } while (nc != 0);
376 }
377