xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x8c2-minmax-rndnu-neon-mlal-dup.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-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 
xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     const int8_t* restrict a,
22     size_t a_stride,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(mr != 0);
30   assert(mr <= 3);
31   assert(nc != 0);
32   assert(kc != 0);
33   assert(kc % sizeof(int8_t) == 0);
34   assert(a != NULL);
35   assert(w != NULL);
36   assert(c != NULL);
37 
38   kc = round_up_po2(kc, 2 * sizeof(int8_t));
39   const int8_t* a0 = a;
40   int8_t* c0 = c;
41   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr < 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     a2 = a1;
51     c2 = c1;
52   }
53 
54   do {
55     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
56     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57     int32x4_t vacc1x0123 = vacc0x0123;
58     int32x4_t vacc1x4567 = vacc0x4567;
59     int32x4_t vacc2x0123 = vacc0x0123;
60     int32x4_t vacc2x4567 = vacc0x4567;
61 
62     size_t k = kc;
63 
64     while (k >= 16 * sizeof(int8_t)) {
65       const int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
66       const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
67       const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
68       const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
69       const int8x8_t va2x0 = vld1_s8(a2); a2 += 8;
70       const int8x8_t va2x1 = vld1_s8(a2); a2 += 8;
71       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79 
80       const int8x8_t va0c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 0));
81       const int8x8_t va0c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 0));
82       const int8x8_t va1c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 0));
83       const int8x8_t va1c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 0));
84       const int8x8_t va2c0x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 0));
85       const int8x8_t va2c0x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 0));
86 
87       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
88       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
89       int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0);
90       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
92       vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
93       vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2c0x1);
94       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
95       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
96       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
97       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
98       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
99       int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0);
100       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
102       vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
103       vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2c0x1);
104       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
105       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
106       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
107       const int8x8_t va0c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 1));
108       const int8x8_t va0c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 1));
109       const int8x8_t va1c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 1));
110       const int8x8_t va1c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 1));
111       const int8x8_t va2c1x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 1));
112       const int8x8_t va2c1x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 1));
113 
114       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
115       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
116       int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0);
117       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
119       vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
120       vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1);
121       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
122       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
123       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
124       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
125       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
126       int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2c1x0);
127       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
129       vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
130       vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2c1x1);
131       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
132       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
133       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
134       const int8x8_t va0c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 2));
135       const int8x8_t va0c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 2));
136       const int8x8_t va1c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 2));
137       const int8x8_t va1c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 2));
138       const int8x8_t va2c2x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 2));
139       const int8x8_t va2c2x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 2));
140 
141       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
142       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
143       int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2c2x0);
144       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
145       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
146       vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
147       vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2c2x1);
148       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
149       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
150       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
151       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
152       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
153       int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2c2x0);
154       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
156       vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
157       vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2c2x1);
158       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
159       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
160       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
161       const int8x8_t va0c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x0), 3));
162       const int8x8_t va0c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0x1), 3));
163       const int8x8_t va1c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x0), 3));
164       const int8x8_t va1c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1x1), 3));
165       const int8x8_t va2c3x0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x0), 3));
166       const int8x8_t va2c3x1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2x1), 3));
167 
168       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
169       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
170       int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2c3x0);
171       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
172       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
173       vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
174       vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2c3x1);
175       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
176       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
177       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
178       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
179       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
180       int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2c3x0);
181       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
182       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
183       vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
184       vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2c3x1);
185       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
186       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
187       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
188 
189       k -= 16 * sizeof(int8_t);
190     }
191 
192     if (k >= 8 * sizeof(int8_t)) {
193       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
194       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
195       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
196 
197       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
198       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
199       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
200       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
201       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
202       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
203       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
204       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
205 
206       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
207       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
208       const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
209 
210       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
211       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
212       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
213       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
214       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
215       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
216       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
217       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
218       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
219       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
220       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
221       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
222       const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
223       const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
224       const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
225 
226       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
227       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
228       const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
229       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
230       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
231       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
232       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
233       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
234       const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
235       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
236       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
237       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
238       const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
239       const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
240       const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
241 
242       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
243       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
244       const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
245       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
246       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
247       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
248       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
249       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
250       const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
251       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
252       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
253       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
254       const int8x8_t va0c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 3));
255       const int8x8_t va1c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 3));
256       const int8x8_t va2c3 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 3));
257 
258       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
259       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
260       const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
261       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
262       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
263       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
264       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
265       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
266       const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
267       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
268       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
269       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
270 
271       k -= 8 * sizeof(int8_t);
272     }
273 
274     if XNN_UNLIKELY(k != 0) {
275       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
276       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
277       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
278 
279       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
280       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
281 
282       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
283       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
284       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
285       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
286       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
287       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
288       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
289       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
290       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
291       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
292       const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
293       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
294       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
295       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
296       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
297 
298       if (k > 2 * sizeof(int8_t)) {
299         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
300         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
301 
302         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
303         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
304         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
305         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
306         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
307         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
308         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
309         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
310         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
311         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
312         const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
313         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
314         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
315         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
316         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
317 
318         if (k > 4 * sizeof(int8_t)) {
319           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
320           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
321 
322           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
323           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
324           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
325           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
326           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
327           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
328           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
329           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
330           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
331           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
332           const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
333           const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
334           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
335           const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
336           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
337         }
338       }
339     }
340 
341     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
342     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
343     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
344 
345     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
346     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
347     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
348     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
349     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
350     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
351 
352     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
353     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
354     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
355     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
356     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
357     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
358 
359     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
360     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
361     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
362     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
363     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
364     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
365 
366     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
367 #if XNN_ARCH_ARM64
368     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
369     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
370     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
371 
372     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
373     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
374     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
375 
376     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
377     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
378 #else
379     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
380     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
381     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
382 
383     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
384     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
385     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
386 
387     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
388     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
389 #endif
390 
391     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
392     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
393     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
394 
395     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
396     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
397     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
398 
399     if (nc >= 8) {
400       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
401       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
402       vst1_s8(c2 + 0, vout2x01234567);
403 
404       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
405       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
406       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
407 
408       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
409       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
410       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
411 
412       nc -= 8;
413     } else {
414       // Final case where not all of the 8 columns fit in the destination.
415       if (nc & 4) {
416         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
417         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
418         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
419         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
420         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
421       }
422       if (nc & 2) {
423         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
424         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
425         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
426         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
427         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
428       }
429       if (nc & 1) {
430         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
431         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
432         vst1_lane_s8(c2, vout2x01234567, 0);
433       }
434 
435       nc = 0;
436     }
437   } while (nc != 0);
438 }
439