xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x8c2-minmax-rndnu-neon-mlal-ld1r.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_ld1r(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_ld1r(
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 int16x4_t va00x0 = vld1_dup_s16((const void*)a0);
66       const int16x4_t va01x0 = vld1_dup_s16((const void*)(a0 + 2));
67       const int16x4_t va02x0 = vld1_dup_s16((const void*)(a0 + 4));
68       const int16x4_t va03x0 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
69       const int16x4_t va00x1 = vld1_dup_s16((const void*)a0);
70       const int16x4_t va01x1 = vld1_dup_s16((const void*)(a0 + 2));
71       const int16x4_t va02x1 = vld1_dup_s16((const void*)(a0 + 4));
72       const int16x4_t va03x1 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
73       const int16x4_t va10x0 = vld1_dup_s16((const void*)a1);
74       const int16x4_t va11x0 = vld1_dup_s16((const void*)(a1 + 2));
75       const int16x4_t va12x0 = vld1_dup_s16((const void*)(a1 + 4));
76       const int16x4_t va13x0 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
77       const int16x4_t va10x1 = vld1_dup_s16((const void*)a1);
78       const int16x4_t va11x1 = vld1_dup_s16((const void*)(a1 + 2));
79       const int16x4_t va12x1 = vld1_dup_s16((const void*)(a1 + 4));
80       const int16x4_t va13x1 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
81       const int16x4_t va20x0 = vld1_dup_s16((const void*)a2);
82       const int16x4_t va21x0 = vld1_dup_s16((const void*)(a2 + 2));
83       const int16x4_t va22x0 = vld1_dup_s16((const void*)(a2 + 4));
84       const int16x4_t va23x0 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
85       const int16x4_t va20x1 = vld1_dup_s16((const void*)a2);
86       const int16x4_t va21x1 = vld1_dup_s16((const void*)(a2 + 2));
87       const int16x4_t va22x1 = vld1_dup_s16((const void*)(a2 + 4));
88       const int16x4_t va23x1 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
89       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97 
98       const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0);
99       const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1);
100       const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0);
101       const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1);
102       const int8x8_t va2c0x0 = vreinterpret_s8_s16(va20x0);
103       const int8x8_t va2c0x1 = vreinterpret_s8_s16(va20x1);
104 
105       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
106       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
107       int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0);
108       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
110       vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
111       vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2c0x1);
112       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
113       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
114       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
115       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
116       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
117       int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0);
118       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
120       vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
121       vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2c0x1);
122       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
123       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
124       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
125       const int8x8_t va0c1x0 = vreinterpret_s8_s16(va01x0);
126       const int8x8_t va0c1x1 = vreinterpret_s8_s16(va01x1);
127       const int8x8_t va1c1x0 = vreinterpret_s8_s16(va11x0);
128       const int8x8_t va1c1x1 = vreinterpret_s8_s16(va11x1);
129       const int8x8_t va2c1x0 = vreinterpret_s8_s16(va21x0);
130       const int8x8_t va2c1x1 = vreinterpret_s8_s16(va21x1);
131 
132       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
133       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
134       int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0);
135       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
136       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
137       vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
138       vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1);
139       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
140       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
141       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
142       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
143       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
144       int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2c1x0);
145       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
146       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
147       vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
148       vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2c1x1);
149       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
150       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
151       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
152       const int8x8_t va0c2x0 = vreinterpret_s8_s16(va02x0);
153       const int8x8_t va0c2x1 = vreinterpret_s8_s16(va02x1);
154       const int8x8_t va1c2x0 = vreinterpret_s8_s16(va12x0);
155       const int8x8_t va1c2x1 = vreinterpret_s8_s16(va12x1);
156       const int8x8_t va2c2x0 = vreinterpret_s8_s16(va22x0);
157       const int8x8_t va2c2x1 = vreinterpret_s8_s16(va22x1);
158 
159       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
160       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
161       int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2c2x0);
162       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
164       vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
165       vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2c2x1);
166       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
167       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
168       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
169       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
170       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
171       int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2c2x0);
172       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
173       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
174       vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
175       vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2c2x1);
176       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
177       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
178       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
179       const int8x8_t va0c3x0 = vreinterpret_s8_s16(va03x0);
180       const int8x8_t va0c3x1 = vreinterpret_s8_s16(va03x1);
181       const int8x8_t va1c3x0 = vreinterpret_s8_s16(va13x0);
182       const int8x8_t va1c3x1 = vreinterpret_s8_s16(va13x1);
183       const int8x8_t va2c3x0 = vreinterpret_s8_s16(va23x0);
184       const int8x8_t va2c3x1 = vreinterpret_s8_s16(va23x1);
185 
186       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
187       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
188       int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2c3x0);
189       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
190       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
191       vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
192       vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2c3x1);
193       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
194       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
195       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
196       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
197       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
198       int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2c3x0);
199       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
200       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
201       vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
202       vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2c3x1);
203       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
204       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
205       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
206 
207       k -= 16 * sizeof(int8_t);
208     }
209 
210     if (k >= 8 * sizeof(int8_t)) {
211       const int16x4_t va00 = vld1_dup_s16((const void*)a0);
212       const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
213       const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
214       const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
215       const int16x4_t va10 = vld1_dup_s16((const void*)a1);
216       const int16x4_t va11 = vld1_dup_s16((const void*)(a1 + 2));
217       const int16x4_t va12 = vld1_dup_s16((const void*)(a1 + 4));
218       const int16x4_t va13 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
219       const int16x4_t va20 = vld1_dup_s16((const void*)a2);
220       const int16x4_t va21 = vld1_dup_s16((const void*)(a2 + 2));
221       const int16x4_t va22 = vld1_dup_s16((const void*)(a2 + 4));
222       const int16x4_t va23 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
223 
224       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
225       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
226       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
227       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
228       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
229       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
230       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
231       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
232 
233       const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
234       const int8x8_t va1c0 = vreinterpret_s8_s16(va10);
235       const int8x8_t va2c0 = vreinterpret_s8_s16(va20);
236 
237       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
238       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
239       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
240       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
241       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
242       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
243       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
244       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
245       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
246       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
247       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
248       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
249       const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
250       const int8x8_t va1c1 = vreinterpret_s8_s16(va11);
251       const int8x8_t va2c1 = vreinterpret_s8_s16(va21);
252 
253       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
254       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
255       const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
256       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
257       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
258       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
259       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
260       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
261       const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
262       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
263       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
264       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
265       const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
266       const int8x8_t va1c2 = vreinterpret_s8_s16(va12);
267       const int8x8_t va2c2 = vreinterpret_s8_s16(va22);
268 
269       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
270       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
271       const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
272       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
273       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
274       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
275       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
276       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
277       const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
278       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
279       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
280       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
281       const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
282       const int8x8_t va1c3 = vreinterpret_s8_s16(va13);
283       const int8x8_t va2c3 = vreinterpret_s8_s16(va23);
284 
285       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
286       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
287       const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
288       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
289       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
290       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
291       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
292       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
293       const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
294       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
295       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
296       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
297 
298       k -= 8 * sizeof(int8_t);
299     }
300 
301     if XNN_UNLIKELY(k != 0) {
302       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
303       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
304       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
305 
306       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
307       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
308 
309       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
310       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
311       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
312       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
313       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
314       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
315       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
316       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
317       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
318       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
319       const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
320       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
321       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
322       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
323       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
324 
325       if (k > 2 * sizeof(int8_t)) {
326         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
327         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
328 
329         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
330         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
331         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
332         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
333         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
334         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
335         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
336         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
337         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
338         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
339         const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
340         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
341         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
342         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
343         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
344 
345         if (k > 4 * sizeof(int8_t)) {
346           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
347           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
348 
349           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
350           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
351           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
352           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
353           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
354           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
355           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
356           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
357           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
358           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
359           const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
360           const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
361           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
362           const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
363           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
364         }
365       }
366     }
367 
368     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
369     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
370     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
371 
372     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
373     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
374     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
375     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
376     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
377     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
378 
379     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
380     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
381     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
382     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
383     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
384     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
385 
386     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
387     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
388     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
389     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
390     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
391     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
392 
393     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
394 #if XNN_ARCH_ARM64
395     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
396     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
397     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
398 
399     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
400     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
401     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
402 
403     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
404     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
405 #else
406     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
407     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
408     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
409 
410     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
411     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
412     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
413 
414     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
415     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
416 #endif
417 
418     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
419     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
420     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
421 
422     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
423     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
424     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
425 
426     if (nc >= 8) {
427       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
428       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
429       vst1_s8(c2 + 0, vout2x01234567);
430 
431       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
432       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
433       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
434 
435       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
436       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
437       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
438 
439       nc -= 8;
440     } else {
441       // Final case where not all of the 8 columns fit in the destination.
442       if (nc & 4) {
443         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
444         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
445         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
446         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
447         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
448       }
449       if (nc & 2) {
450         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
451         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
452         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
453         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
454         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
455       }
456       if (nc & 1) {
457         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
458         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
459         vst1_lane_s8(c2, vout2x01234567, 0);
460       }
461 
462       nc = 0;
463     }
464   } while (nc != 0);
465 }
466