xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/4x8c16-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_4x8c16__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_4x8c16__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 <= 4);
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   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   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
55   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56   if XNN_UNPREDICTABLE(mr != 4) {
57     a3 = a2;
58     c3 = c2;
59   }
60 
61   do {
62     int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
63     int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
64     int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
65     int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
66     int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
67     int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
68     int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
69     int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const void*) ((uintptr_t) w + sizeof(int32_t));
70     int32x4_t vacc1x0 = vacc0x0;
71     int32x4_t vacc1x1 = vacc0x1;
72     int32x4_t vacc1x2 = vacc0x2;
73     int32x4_t vacc1x3 = vacc0x3;
74     int32x4_t vacc1x4 = vacc0x4;
75     int32x4_t vacc1x5 = vacc0x5;
76     int32x4_t vacc1x6 = vacc0x6;
77     int32x4_t vacc1x7 = vacc0x7;
78     int32x4_t vacc2x0 = vacc0x0;
79     int32x4_t vacc2x1 = vacc0x1;
80     int32x4_t vacc2x2 = vacc0x2;
81     int32x4_t vacc2x3 = vacc0x3;
82     int32x4_t vacc2x4 = vacc0x4;
83     int32x4_t vacc2x5 = vacc0x5;
84     int32x4_t vacc2x6 = vacc0x6;
85     int32x4_t vacc2x7 = vacc0x7;
86     int32x4_t vacc3x0 = vacc0x0;
87     int32x4_t vacc3x1 = vacc0x1;
88     int32x4_t vacc3x2 = vacc0x2;
89     int32x4_t vacc3x3 = vacc0x3;
90     int32x4_t vacc3x4 = vacc0x4;
91     int32x4_t vacc3x5 = vacc0x5;
92     int32x4_t vacc3x6 = vacc0x6;
93     int32x4_t vacc3x7 = vacc0x7;
94 
95     // KC loop of 16
96     size_t k = kc;
97     while (k != 0) {
98       const int8x16_t va0 = vld1q_s8(a0); a0 += 16;
99       const int8x16_t va1 = vld1q_s8(a1); a1 += 16;
100       const int8x16_t va2 = vld1q_s8(a2); a2 += 16;
101       const int8x16_t va3 = vld1q_s8(a3); a3 += 16;
102 
103       const int8x16_t vb0 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
104       const int8x16_t vb1 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
105       const int8x16_t vb2 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
106       const int8x16_t vb3 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
107       const int8x16_t vb4 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
108       const int8x16_t vb5 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
109       const int8x16_t vb6 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
110       const int8x16_t vb7 = vld1q_s8(w); w = (const void*) ((uintptr_t) w + 16 * sizeof(int8_t));
111 
112       int16x8_t vprod0x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va0));
113       int16x8_t vprod1x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va1));
114       int16x8_t vprod2x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va2));
115       int16x8_t vprod3x0 = vmull_s8(vget_low_s8(vb0), vget_low_s8(va3));
116       vprod0x0 = vmlal_s8(vprod0x0, vget_high_s8(vb0), vget_high_s8(va0));
117       vprod1x0 = vmlal_s8(vprod1x0, vget_high_s8(vb0), vget_high_s8(va1));
118       vprod2x0 = vmlal_s8(vprod2x0, vget_high_s8(vb0), vget_high_s8(va2));
119       vprod3x0 = vmlal_s8(vprod3x0, vget_high_s8(vb0), vget_high_s8(va3));
120       vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0);
121       vacc1x0 = vpadalq_s16(vacc1x0, vprod1x0);
122       vacc2x0 = vpadalq_s16(vacc2x0, vprod2x0);
123       vacc3x0 = vpadalq_s16(vacc3x0, vprod3x0);
124       int16x8_t vprod0x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va0));
125       int16x8_t vprod1x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va1));
126       int16x8_t vprod2x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va2));
127       int16x8_t vprod3x1 = vmull_s8(vget_low_s8(vb1), vget_low_s8(va3));
128       vprod0x1 = vmlal_s8(vprod0x1, vget_high_s8(vb1), vget_high_s8(va0));
129       vprod1x1 = vmlal_s8(vprod1x1, vget_high_s8(vb1), vget_high_s8(va1));
130       vprod2x1 = vmlal_s8(vprod2x1, vget_high_s8(vb1), vget_high_s8(va2));
131       vprod3x1 = vmlal_s8(vprod3x1, vget_high_s8(vb1), vget_high_s8(va3));
132       vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1);
133       vacc1x1 = vpadalq_s16(vacc1x1, vprod1x1);
134       vacc2x1 = vpadalq_s16(vacc2x1, vprod2x1);
135       vacc3x1 = vpadalq_s16(vacc3x1, vprod3x1);
136       int16x8_t vprod0x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va0));
137       int16x8_t vprod1x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va1));
138       int16x8_t vprod2x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va2));
139       int16x8_t vprod3x2 = vmull_s8(vget_low_s8(vb2), vget_low_s8(va3));
140       vprod0x2 = vmlal_s8(vprod0x2, vget_high_s8(vb2), vget_high_s8(va0));
141       vprod1x2 = vmlal_s8(vprod1x2, vget_high_s8(vb2), vget_high_s8(va1));
142       vprod2x2 = vmlal_s8(vprod2x2, vget_high_s8(vb2), vget_high_s8(va2));
143       vprod3x2 = vmlal_s8(vprod3x2, vget_high_s8(vb2), vget_high_s8(va3));
144       vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2);
145       vacc1x2 = vpadalq_s16(vacc1x2, vprod1x2);
146       vacc2x2 = vpadalq_s16(vacc2x2, vprod2x2);
147       vacc3x2 = vpadalq_s16(vacc3x2, vprod3x2);
148       int16x8_t vprod0x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va0));
149       int16x8_t vprod1x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va1));
150       int16x8_t vprod2x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va2));
151       int16x8_t vprod3x3 = vmull_s8(vget_low_s8(vb3), vget_low_s8(va3));
152       vprod0x3 = vmlal_s8(vprod0x3, vget_high_s8(vb3), vget_high_s8(va0));
153       vprod1x3 = vmlal_s8(vprod1x3, vget_high_s8(vb3), vget_high_s8(va1));
154       vprod2x3 = vmlal_s8(vprod2x3, vget_high_s8(vb3), vget_high_s8(va2));
155       vprod3x3 = vmlal_s8(vprod3x3, vget_high_s8(vb3), vget_high_s8(va3));
156       vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3);
157       vacc1x3 = vpadalq_s16(vacc1x3, vprod1x3);
158       vacc2x3 = vpadalq_s16(vacc2x3, vprod2x3);
159       vacc3x3 = vpadalq_s16(vacc3x3, vprod3x3);
160       int16x8_t vprod0x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va0));
161       int16x8_t vprod1x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va1));
162       int16x8_t vprod2x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va2));
163       int16x8_t vprod3x4 = vmull_s8(vget_low_s8(vb4), vget_low_s8(va3));
164       vprod0x4 = vmlal_s8(vprod0x4, vget_high_s8(vb4), vget_high_s8(va0));
165       vprod1x4 = vmlal_s8(vprod1x4, vget_high_s8(vb4), vget_high_s8(va1));
166       vprod2x4 = vmlal_s8(vprod2x4, vget_high_s8(vb4), vget_high_s8(va2));
167       vprod3x4 = vmlal_s8(vprod3x4, vget_high_s8(vb4), vget_high_s8(va3));
168       vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4);
169       vacc1x4 = vpadalq_s16(vacc1x4, vprod1x4);
170       vacc2x4 = vpadalq_s16(vacc2x4, vprod2x4);
171       vacc3x4 = vpadalq_s16(vacc3x4, vprod3x4);
172       int16x8_t vprod0x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va0));
173       int16x8_t vprod1x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va1));
174       int16x8_t vprod2x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va2));
175       int16x8_t vprod3x5 = vmull_s8(vget_low_s8(vb5), vget_low_s8(va3));
176       vprod0x5 = vmlal_s8(vprod0x5, vget_high_s8(vb5), vget_high_s8(va0));
177       vprod1x5 = vmlal_s8(vprod1x5, vget_high_s8(vb5), vget_high_s8(va1));
178       vprod2x5 = vmlal_s8(vprod2x5, vget_high_s8(vb5), vget_high_s8(va2));
179       vprod3x5 = vmlal_s8(vprod3x5, vget_high_s8(vb5), vget_high_s8(va3));
180       vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5);
181       vacc1x5 = vpadalq_s16(vacc1x5, vprod1x5);
182       vacc2x5 = vpadalq_s16(vacc2x5, vprod2x5);
183       vacc3x5 = vpadalq_s16(vacc3x5, vprod3x5);
184       int16x8_t vprod0x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va0));
185       int16x8_t vprod1x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va1));
186       int16x8_t vprod2x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va2));
187       int16x8_t vprod3x6 = vmull_s8(vget_low_s8(vb6), vget_low_s8(va3));
188       vprod0x6 = vmlal_s8(vprod0x6, vget_high_s8(vb6), vget_high_s8(va0));
189       vprod1x6 = vmlal_s8(vprod1x6, vget_high_s8(vb6), vget_high_s8(va1));
190       vprod2x6 = vmlal_s8(vprod2x6, vget_high_s8(vb6), vget_high_s8(va2));
191       vprod3x6 = vmlal_s8(vprod3x6, vget_high_s8(vb6), vget_high_s8(va3));
192       vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6);
193       vacc1x6 = vpadalq_s16(vacc1x6, vprod1x6);
194       vacc2x6 = vpadalq_s16(vacc2x6, vprod2x6);
195       vacc3x6 = vpadalq_s16(vacc3x6, vprod3x6);
196       int16x8_t vprod0x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va0));
197       int16x8_t vprod1x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va1));
198       int16x8_t vprod2x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va2));
199       int16x8_t vprod3x7 = vmull_s8(vget_low_s8(vb7), vget_low_s8(va3));
200       vprod0x7 = vmlal_s8(vprod0x7, vget_high_s8(vb7), vget_high_s8(va0));
201       vprod1x7 = vmlal_s8(vprod1x7, vget_high_s8(vb7), vget_high_s8(va1));
202       vprod2x7 = vmlal_s8(vprod2x7, vget_high_s8(vb7), vget_high_s8(va2));
203       vprod3x7 = vmlal_s8(vprod3x7, vget_high_s8(vb7), vget_high_s8(va3));
204       vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7);
205       vacc1x7 = vpadalq_s16(vacc1x7, vprod1x7);
206       vacc2x7 = vpadalq_s16(vacc2x7, vprod2x7);
207       vacc3x7 = vpadalq_s16(vacc3x7, vprod3x7);
208 
209       k -= 16 * sizeof(int8_t);
210     }
211 
212 #if XNN_ARCH_ARM64
213     const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1);
214     const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3);
215     const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5);
216     const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7);
217     const int32x4_t vsum1x01 = vpaddq_s32(vacc1x0, vacc1x1);
218     const int32x4_t vsum1x23 = vpaddq_s32(vacc1x2, vacc1x3);
219     const int32x4_t vsum1x45 = vpaddq_s32(vacc1x4, vacc1x5);
220     const int32x4_t vsum1x67 = vpaddq_s32(vacc1x6, vacc1x7);
221     const int32x4_t vsum2x01 = vpaddq_s32(vacc2x0, vacc2x1);
222     const int32x4_t vsum2x23 = vpaddq_s32(vacc2x2, vacc2x3);
223     const int32x4_t vsum2x45 = vpaddq_s32(vacc2x4, vacc2x5);
224     const int32x4_t vsum2x67 = vpaddq_s32(vacc2x6, vacc2x7);
225     const int32x4_t vsum3x01 = vpaddq_s32(vacc3x0, vacc3x1);
226     const int32x4_t vsum3x23 = vpaddq_s32(vacc3x2, vacc3x3);
227     const int32x4_t vsum3x45 = vpaddq_s32(vacc3x4, vacc3x5);
228     const int32x4_t vsum3x67 = vpaddq_s32(vacc3x6, vacc3x7);
229     int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23);
230     int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67);
231     int32x4_t vacc1x0123 = vpaddq_s32(vsum1x01, vsum1x23);
232     int32x4_t vacc1x4567 = vpaddq_s32(vsum1x45, vsum1x67);
233     int32x4_t vacc2x0123 = vpaddq_s32(vsum2x01, vsum2x23);
234     int32x4_t vacc2x4567 = vpaddq_s32(vsum2x45, vsum2x67);
235     int32x4_t vacc3x0123 = vpaddq_s32(vsum3x01, vsum3x23);
236     int32x4_t vacc3x4567 = vpaddq_s32(vsum3x45, vsum3x67);
237 #else
238     const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0));
239     const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1));
240     const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2));
241     const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3));
242     const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1);
243     const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3);
244     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 );
245     const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4));
246     const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5));
247     const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6));
248     const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7));
249     const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5);
250     const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7);
251     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 );
252     const int32x2_t vpsum1x0 = vadd_s32(vget_low_s32(vacc1x0), vget_high_s32(vacc1x0));
253     const int32x2_t vpsum1x1 = vadd_s32(vget_low_s32(vacc1x1), vget_high_s32(vacc1x1));
254     const int32x2_t vpsum1x2 = vadd_s32(vget_low_s32(vacc1x2), vget_high_s32(vacc1x2));
255     const int32x2_t vpsum1x3 = vadd_s32(vget_low_s32(vacc1x3), vget_high_s32(vacc1x3));
256     const int32x2_t vsum1x01 = vpadd_s32(vpsum1x0, vpsum1x1);
257     const int32x2_t vsum1x23 = vpadd_s32(vpsum1x2, vpsum1x3);
258     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23 );
259     const int32x2_t vpsum1x4 = vadd_s32(vget_low_s32(vacc1x4), vget_high_s32(vacc1x4));
260     const int32x2_t vpsum1x5 = vadd_s32(vget_low_s32(vacc1x5), vget_high_s32(vacc1x5));
261     const int32x2_t vpsum1x6 = vadd_s32(vget_low_s32(vacc1x6), vget_high_s32(vacc1x6));
262     const int32x2_t vpsum1x7 = vadd_s32(vget_low_s32(vacc1x7), vget_high_s32(vacc1x7));
263     const int32x2_t vsum1x45 = vpadd_s32(vpsum1x4, vpsum1x5);
264     const int32x2_t vsum1x67 = vpadd_s32(vpsum1x6, vpsum1x7);
265     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67 );
266     const int32x2_t vpsum2x0 = vadd_s32(vget_low_s32(vacc2x0), vget_high_s32(vacc2x0));
267     const int32x2_t vpsum2x1 = vadd_s32(vget_low_s32(vacc2x1), vget_high_s32(vacc2x1));
268     const int32x2_t vpsum2x2 = vadd_s32(vget_low_s32(vacc2x2), vget_high_s32(vacc2x2));
269     const int32x2_t vpsum2x3 = vadd_s32(vget_low_s32(vacc2x3), vget_high_s32(vacc2x3));
270     const int32x2_t vsum2x01 = vpadd_s32(vpsum2x0, vpsum2x1);
271     const int32x2_t vsum2x23 = vpadd_s32(vpsum2x2, vpsum2x3);
272     int32x4_t vacc2x0123 = vcombine_s32(vsum2x01, vsum2x23 );
273     const int32x2_t vpsum2x4 = vadd_s32(vget_low_s32(vacc2x4), vget_high_s32(vacc2x4));
274     const int32x2_t vpsum2x5 = vadd_s32(vget_low_s32(vacc2x5), vget_high_s32(vacc2x5));
275     const int32x2_t vpsum2x6 = vadd_s32(vget_low_s32(vacc2x6), vget_high_s32(vacc2x6));
276     const int32x2_t vpsum2x7 = vadd_s32(vget_low_s32(vacc2x7), vget_high_s32(vacc2x7));
277     const int32x2_t vsum2x45 = vpadd_s32(vpsum2x4, vpsum2x5);
278     const int32x2_t vsum2x67 = vpadd_s32(vpsum2x6, vpsum2x7);
279     int32x4_t vacc2x4567 = vcombine_s32(vsum2x45, vsum2x67 );
280     const int32x2_t vpsum3x0 = vadd_s32(vget_low_s32(vacc3x0), vget_high_s32(vacc3x0));
281     const int32x2_t vpsum3x1 = vadd_s32(vget_low_s32(vacc3x1), vget_high_s32(vacc3x1));
282     const int32x2_t vpsum3x2 = vadd_s32(vget_low_s32(vacc3x2), vget_high_s32(vacc3x2));
283     const int32x2_t vpsum3x3 = vadd_s32(vget_low_s32(vacc3x3), vget_high_s32(vacc3x3));
284     const int32x2_t vsum3x01 = vpadd_s32(vpsum3x0, vpsum3x1);
285     const int32x2_t vsum3x23 = vpadd_s32(vpsum3x2, vpsum3x3);
286     int32x4_t vacc3x0123 = vcombine_s32(vsum3x01, vsum3x23 );
287     const int32x2_t vpsum3x4 = vadd_s32(vget_low_s32(vacc3x4), vget_high_s32(vacc3x4));
288     const int32x2_t vpsum3x5 = vadd_s32(vget_low_s32(vacc3x5), vget_high_s32(vacc3x5));
289     const int32x2_t vpsum3x6 = vadd_s32(vget_low_s32(vacc3x6), vget_high_s32(vacc3x6));
290     const int32x2_t vpsum3x7 = vadd_s32(vget_low_s32(vacc3x7), vget_high_s32(vacc3x7));
291     const int32x2_t vsum3x45 = vpadd_s32(vpsum3x4, vpsum3x5);
292     const int32x2_t vsum3x67 = vpadd_s32(vpsum3x6, vpsum3x7);
293     int32x4_t vacc3x4567 = vcombine_s32(vsum3x45, vsum3x67 );
294 #endif
295 
296     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
297     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
298     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
299 
300     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
301     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
302     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
303     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
304     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
305     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
306     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
307     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
308 
309     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
310     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
311     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
312     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
313     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
314     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
315     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
316     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
317 
318     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
319     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
320     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
321     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
322     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
323     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
324     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
325     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
326 
327     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
328 #if XNN_ARCH_ARM64
329     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
330     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
331     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
332     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
333     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
334     int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
335 #else
336     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
337     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
338     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
339     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
340 
341     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
342     int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
343 #endif
344     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
345     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
346 
347     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
348     vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
349 
350     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
351     vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
352 
353     if (nc >= 8) {
354       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
355       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
356       vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
357       vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
358 
359       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
360       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
361       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
362       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
363 
364       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
365       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
366       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
367       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
368 
369       nc -= 8;
370     } else {
371       // Final case where not all of the 8 columns fit in the destination.
372       if (nc & 4) {
373         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
374         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
375         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
376         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
377         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
378         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
379       }
380       if (nc & 2) {
381         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
382         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
383         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
384         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
385         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
386         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
387       }
388       if (nc & 1) {
389         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
390         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
391         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
392         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
393       }
394 
395       nc = 0;
396     }
397   } while (nc != 0);
398 }
399