xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/2x16c16-minmax-rndnu-neon-mlal.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c16-neon-mlal.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_rndnu_ukernel_2x16c16__neon_mlal(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c16__neon_mlal(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 2);
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, 16 * sizeof(int8_t));
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr != 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48 
49   do {
50     int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
51     int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
52     int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
53     int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
54     int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
55     int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
56     int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
57     int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
58     int32x4_t vacc0x8 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
59     int32x4_t vacc0x9 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
60     int32x4_t vacc0x10 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
61     int32x4_t vacc0x11 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
62     int32x4_t vacc0x12 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63     int32x4_t vacc0x13 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64     int32x4_t vacc0x14 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65     int32x4_t vacc0x15 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66     int32x4_t vacc1x0 = vacc0x0;
67     int32x4_t vacc1x1 = vacc0x1;
68     int32x4_t vacc1x2 = vacc0x2;
69     int32x4_t vacc1x3 = vacc0x3;
70     int32x4_t vacc1x4 = vacc0x4;
71     int32x4_t vacc1x5 = vacc0x5;
72     int32x4_t vacc1x6 = vacc0x6;
73     int32x4_t vacc1x7 = vacc0x7;
74     int32x4_t vacc1x8 = vacc0x8;
75     int32x4_t vacc1x9 = vacc0x9;
76     int32x4_t vacc1x10 = vacc0x10;
77     int32x4_t vacc1x11 = vacc0x11;
78     int32x4_t vacc1x12 = vacc0x12;
79     int32x4_t vacc1x13 = vacc0x13;
80     int32x4_t vacc1x14 = vacc0x14;
81     int32x4_t vacc1x15 = vacc0x15;
82 
83     // KC loop of 16
84     size_t k = kc;
85     while (k != 0) {
86       const int8x16_t va0 = vld1q_s8(a0); a0 += 16;
87       const int8x16_t va1 = vld1q_s8(a1); a1 += 16;
88 
89       const int8x16_t vb0 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
90       const int8x16_t vb1 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
91       const int8x16_t vb2 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
92       const int8x16_t vb3 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
93       const int8x16_t vb4 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
94       const int8x16_t vb5 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
95       const int8x16_t vb6 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
96       const int8x16_t vb7 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
97       const int8x16_t vb8 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
98       const int8x16_t vb9 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
99       const int8x16_t vb10 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
100       const int8x16_t vb11 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
101       const int8x16_t vb12 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
102       const int8x16_t vb13 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
103       const int8x16_t vb14 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
104       const int8x16_t vb15 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
105 
106       int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0));
107       int16x8_t vprod1x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va1));
108       vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0));
109       vprod1x0 = vmlal_s8(vprod1x0, vget_high_s8(vb0), vget_high_s8(va1));
110       vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
111       vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
112       int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0));
113       int16x8_t vprod1x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va1));
114       vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0));
115       vprod1x1 = vmlal_s8(vprod1x1, vget_high_s8(vb1), vget_high_s8(va1));
116       vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
117       vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
118       int16x8_t vprod0x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va0));
119       int16x8_t vprod1x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va1));
120       vprod0x2 = vmlal_s8(vprod0x2, vget_high_s8(vb2), vget_high_s8(va0));
121       vprod1x2 = vmlal_s8(vprod1x2, vget_high_s8(vb2), vget_high_s8(va1));
122       vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
123       vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
124       int16x8_t vprod0x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va0));
125       int16x8_t vprod1x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va1));
126       vprod0x3 = vmlal_s8(vprod0x3, vget_high_s8(vb3), vget_high_s8(va0));
127       vprod1x3 = vmlal_s8(vprod1x3, vget_high_s8(vb3), vget_high_s8(va1));
128       vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
129       vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
130       int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0));
131       int16x8_t vprod1x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va1));
132       vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0));
133       vprod1x4 = vmlal_s8(vprod1x4, vget_high_s8(vb4), vget_high_s8(va1));
134       vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
135       vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
136       int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0));
137       int16x8_t vprod1x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va1));
138       vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0));
139       vprod1x5 = vmlal_s8(vprod1x5, vget_high_s8(vb5), vget_high_s8(va1));
140       vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
141       vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
142       int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0));
143       int16x8_t vprod1x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va1));
144       vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0));
145       vprod1x6 = vmlal_s8(vprod1x6, vget_high_s8(vb6), vget_high_s8(va1));
146       vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
147       vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
148       int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0));
149       int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1));
150       vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0));
151       vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1));
152       vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
153       vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
154       int16x8_t vprod0x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va0));
155       int16x8_t vprod1x8 = vmull_s8(vget_low_s8(vb8), vget_low_s8(va1));
156       vprod0x8 = vmlal_s8(vprod0x8, vget_high_s8(vb8), vget_high_s8(va0));
157       vprod1x8 = vmlal_s8(vprod1x8, vget_high_s8(vb8), vget_high_s8(va1));
158       vacc0x8 = vpadalq_s16(vacc0x8, vprod0x8);
159       vacc1x8 = vpadalq_s16(vacc1x8, vprod1x8);
160       int16x8_t vprod0x9 = vmull_s8(vget_low_s8(vb9), vget_low_s8(va0));
161       int16x8_t vprod1x9 = vmull_s8(vget_low_s8(vb9), vget_low_s8(va1));
162       vprod0x9 = vmlal_s8(vprod0x9, vget_high_s8(vb9), vget_high_s8(va0));
163       vprod1x9 = vmlal_s8(vprod1x9, vget_high_s8(vb9), vget_high_s8(va1));
164       vacc0x9 = vpadalq_s16(vacc0x9, vprod0x9);
165       vacc1x9 = vpadalq_s16(vacc1x9, vprod1x9);
166       int16x8_t vprod0x10 = vmull_s8(vget_low_s8(vb10), vget_low_s8(va0));
167       int16x8_t vprod1x10 = vmull_s8(vget_low_s8(vb10), vget_low_s8(va1));
168       vprod0x10 = vmlal_s8(vprod0x10, vget_high_s8(vb10), vget_high_s8(va0));
169       vprod1x10 = vmlal_s8(vprod1x10, vget_high_s8(vb10), vget_high_s8(va1));
170       vacc0x10 = vpadalq_s16(vacc0x10, vprod0x10);
171       vacc1x10 = vpadalq_s16(vacc1x10, vprod1x10);
172       int16x8_t vprod0x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va0));
173       int16x8_t vprod1x11 = vmull_s8(vget_low_s8(vb11), vget_low_s8(va1));
174       vprod0x11 = vmlal_s8(vprod0x11, vget_high_s8(vb11), vget_high_s8(va0));
175       vprod1x11 = vmlal_s8(vprod1x11, vget_high_s8(vb11), vget_high_s8(va1));
176       vacc0x11 = vpadalq_s16(vacc0x11, vprod0x11);
177       vacc1x11 = vpadalq_s16(vacc1x11, vprod1x11);
178       int16x8_t vprod0x12 = vmull_s8(vget_low_s8(vb12), vget_low_s8(va0));
179       int16x8_t vprod1x12 = vmull_s8(vget_low_s8(vb12), vget_low_s8(va1));
180       vprod0x12 = vmlal_s8(vprod0x12, vget_high_s8(vb12), vget_high_s8(va0));
181       vprod1x12 = vmlal_s8(vprod1x12, vget_high_s8(vb12), vget_high_s8(va1));
182       vacc0x12 = vpadalq_s16(vacc0x12, vprod0x12);
183       vacc1x12 = vpadalq_s16(vacc1x12, vprod1x12);
184       int16x8_t vprod0x13 = vmull_s8(vget_low_s8(vb13), vget_low_s8(va0));
185       int16x8_t vprod1x13 = vmull_s8(vget_low_s8(vb13), vget_low_s8(va1));
186       vprod0x13 = vmlal_s8(vprod0x13, vget_high_s8(vb13), vget_high_s8(va0));
187       vprod1x13 = vmlal_s8(vprod1x13, vget_high_s8(vb13), vget_high_s8(va1));
188       vacc0x13 = vpadalq_s16(vacc0x13, vprod0x13);
189       vacc1x13 = vpadalq_s16(vacc1x13, vprod1x13);
190       int16x8_t vprod0x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va0));
191       int16x8_t vprod1x14 = vmull_s8(vget_low_s8(vb14), vget_low_s8(va1));
192       vprod0x14 = vmlal_s8(vprod0x14, vget_high_s8(vb14), vget_high_s8(va0));
193       vprod1x14 = vmlal_s8(vprod1x14, vget_high_s8(vb14), vget_high_s8(va1));
194       vacc0x14 = vpadalq_s16(vacc0x14, vprod0x14);
195       vacc1x14 = vpadalq_s16(vacc1x14, vprod1x14);
196       int16x8_t vprod0x15 = vmull_s8(vget_low_s8(vb15), vget_low_s8(va0));
197       int16x8_t vprod1x15 = vmull_s8(vget_low_s8(vb15), vget_low_s8(va1));
198       vprod0x15 = vmlal_s8(vprod0x15, vget_high_s8(vb15), vget_high_s8(va0));
199       vprod1x15 = vmlal_s8(vprod1x15, vget_high_s8(vb15), vget_high_s8(va1));
200       vacc0x15 = vpadalq_s16(vacc0x15, vprod0x15);
201       vacc1x15 = vpadalq_s16(vacc1x15, vprod1x15);
202 
203       k -= 16 * sizeof(int8_t);
204     }
205 
206 #if XNN_ARCH_ARM64
207     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
208     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
209     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
210     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
211     const int32x4_t vsum0x89 = vpaddq_s32(vacc0x8, vacc0x9);
212     const int32x4_t vsum0xAB = vpaddq_s32(vacc0x10, vacc0x11);
213     const int32x4_t vsum0xCD = vpaddq_s32(vacc0x12, vacc0x13);
214     const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15);
215     const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
216     const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
217     const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
218     const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
219     const int32x4_t vsum1x89 = vpaddq_s32(vacc1x8, vacc1x9);
220     const int32x4_t vsum1xAB = vpaddq_s32(vacc1x10, vacc1x11);
221     const int32x4_t vsum1xCD = vpaddq_s32(vacc1x12, vacc1x13);
222     const int32x4_t vsum1xEF = vpaddq_s32(vacc1x14, vacc1x15);
223     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
224     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
225     int32x4_t vacc0x89AB = vpaddq_s32(vsum0x89, vsum0xAB);
226     int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF);
227     int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
228     int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
229     int32x4_t vacc1x89AB = vpaddq_s32(vsum1x89, vsum1xAB);
230     int32x4_t vacc1xCDEF = vpaddq_s32(vsum1xCD, vsum1xEF);
231 #else
232     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
233     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
234     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
235     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
236     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
237     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
238     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
239     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
240     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
241     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
242     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
243     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
244     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
245     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
246     const int32x2_t vpsum0x8 = vadd_s32(vget_low_s32(vacc0x8), vget_high_s32(vacc0x8));
247     const int32x2_t vpsum0x9 = vadd_s32(vget_low_s32(vacc0x9), vget_high_s32(vacc0x9));
248     const int32x2_t vpsum0xA = vadd_s32(vget_low_s32(vacc0x10), vget_high_s32(vacc0x10));
249     const int32x2_t vpsum0xB = vadd_s32(vget_low_s32(vacc0x11), vget_high_s32(vacc0x11));
250     const int32x2_t vsum0x89 = vpadd_s32(vpsum0x8, vpsum0x9);
251     const int32x2_t vsum0xAB = vpadd_s32(vpsum0xA, vpsum0xB);
252     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB );
253     const int32x2_t vpsum0xC = vadd_s32(vget_low_s32(vacc0x12), vget_high_s32(vacc0x12));
254     const int32x2_t vpsum0xD = vadd_s32(vget_low_s32(vacc0x13), vget_high_s32(vacc0x13));
255     const int32x2_t vpsum0xE = vadd_s32(vget_low_s32(vacc0x14), vget_high_s32(vacc0x14));
256     const int32x2_t vpsum0xF = vadd_s32(vget_low_s32(vacc0x15), vget_high_s32(vacc0x15));
257     const int32x2_t vsum0xCD = vpadd_s32(vpsum0xC, vpsum0xD);
258     const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF);
259     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF );
260     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
261     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
262     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
263     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
264     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
265     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
266     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
267     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
268     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
269     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
270     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
271     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
272     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
273     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
274     const int32x2_t vpsum1x8 = vadd_s32(vget_low_s32(vacc1x8), vget_high_s32(vacc1x8));
275     const int32x2_t vpsum1x9 = vadd_s32(vget_low_s32(vacc1x9), vget_high_s32(vacc1x9));
276     const int32x2_t vpsum1xA = vadd_s32(vget_low_s32(vacc1x10), vget_high_s32(vacc1x10));
277     const int32x2_t vpsum1xB = vadd_s32(vget_low_s32(vacc1x11), vget_high_s32(vacc1x11));
278     const int32x2_t vsum1x89 = vpadd_s32(vpsum1x8, vpsum1x9);
279     const int32x2_t vsum1xAB = vpadd_s32(vpsum1xA, vpsum1xB);
280     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB );
281     const int32x2_t vpsum1xC = vadd_s32(vget_low_s32(vacc1x12), vget_high_s32(vacc1x12));
282     const int32x2_t vpsum1xD = vadd_s32(vget_low_s32(vacc1x13), vget_high_s32(vacc1x13));
283     const int32x2_t vpsum1xE = vadd_s32(vget_low_s32(vacc1x14), vget_high_s32(vacc1x14));
284     const int32x2_t vpsum1xF = vadd_s32(vget_low_s32(vacc1x15), vget_high_s32(vacc1x15));
285     const int32x2_t vsum1xCD = vpadd_s32(vpsum1xC, vpsum1xD);
286     const int32x2_t vsum1xEF = vpadd_s32(vpsum1xE, vpsum1xF);
287     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF );
288 #endif
289 
290     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
291     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
292     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
293 
294     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
295     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
296     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
297     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
298     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
299     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
300     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
301     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
302 
303     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
304     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
305     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
306     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
307     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
308     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
309     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
310     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
311 
312     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
313     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
314     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
315     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
316     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
317     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
318     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
319     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
320 
321     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
322 #if XNN_ARCH_ARM64
323     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
324     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
325     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
326     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
327     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
328     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
329 #else
330     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
331     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
332     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
333     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
334 
335     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
336     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
337 #endif
338     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
339     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
340 
341     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
342     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
343 
344     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
345     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
346 
347     if (nc >= 16) {
348       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
349       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
350 
351       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
352       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
353 
354       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
355       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
356 
357       nc -= 16;
358     } else {
359       // Final case where not all of the 16 columns fit in the destination.
360       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
361       if (nc & 8) {
362         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
363         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
364         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
365       }
366       if (nc & 4) {
367         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
368         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
369         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
370       }
371       if (nc & 2) {
372         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
373         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
374         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
375       }
376       if (nc & 1) {
377         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
378         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
379       }
380 
381       nc = 0;
382     }
383   } while (nc != 0);
384 }
385