• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_3x16c2__neon_mull_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_3x16c2__neon_mull_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 vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
58     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
59     int32x4_t vacc1x0123 = vacc0x0123;
60     int32x4_t vacc1x4567 = vacc0x4567;
61     int32x4_t vacc1x89AB = vacc0x89AB;
62     int32x4_t vacc1xCDEF = vacc0xCDEF;
63     int32x4_t vacc2x0123 = vacc0x0123;
64     int32x4_t vacc2x4567 = vacc0x4567;
65     int32x4_t vacc2x89AB = vacc0x89AB;
66     int32x4_t vacc2xCDEF = vacc0xCDEF;
67 
68     size_t k = kc;
69 
70 
71     while (k >= 8 * sizeof(int8_t)) {
72       const int16x4_t va00 = vld1_dup_s16((const void*)a0);
73       const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
74       const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
75       const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
76       const int16x4_t va10 = vld1_dup_s16((const void*)a1);
77       const int16x4_t va11 = vld1_dup_s16((const void*)(a1 + 2));
78       const int16x4_t va12 = vld1_dup_s16((const void*)(a1 + 4));
79       const int16x4_t va13 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
80       const int16x4_t va20 = vld1_dup_s16((const void*)a2);
81       const int16x4_t va21 = vld1_dup_s16((const void*)(a2 + 2));
82       const int16x4_t va22 = vld1_dup_s16((const void*)(a2 + 4));
83       const int16x4_t va23 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
84 
85       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91       const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92       const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95       const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96       const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99       const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100       const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101 
102       const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
103       const int8x8_t va1c0 = vreinterpret_s8_s16(va10);
104       const int8x8_t va2c0 = vreinterpret_s8_s16(va20);
105 
106       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
107       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
108       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
109       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
110       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
111       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
112       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
113       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
114       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
115       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
116       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
117       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
118       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
119       const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
120       const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
121       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
122       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
123       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
124       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
125       const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
126       const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
127       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
128       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
129       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
130       const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
131       const int8x8_t va1c1 = vreinterpret_s8_s16(va11);
132       const int8x8_t va2c1 = vreinterpret_s8_s16(va21);
133 
134       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
135       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
136       const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
137       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
138       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
139       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
140       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
141       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
142       const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
143       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
144       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
145       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
146       const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
147       const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
148       const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
149       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
150       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
151       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
152       const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
153       const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
154       const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
155       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
156       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
157       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
158       const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
159       const int8x8_t va1c2 = vreinterpret_s8_s16(va12);
160       const int8x8_t va2c2 = vreinterpret_s8_s16(va22);
161 
162       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
163       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
164       const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
165       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
166       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
167       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
168       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
169       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
170       const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
171       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
172       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
173       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
174       const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
175       const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
176       const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
177       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
178       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
179       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
180       const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
181       const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
182       const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
183       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
184       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
185       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
186       const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
187       const int8x8_t va1c3 = vreinterpret_s8_s16(va13);
188       const int8x8_t va2c3 = vreinterpret_s8_s16(va23);
189 
190       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
191       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
192       const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
193       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
194       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
195       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
196       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
197       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
198       const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
199       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
200       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
201       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
202       const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
203       const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
204       const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
205       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
206       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
207       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
208       const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
209       const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
210       const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
211       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
212       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
213       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
214 
215       k -= 8 * sizeof(int8_t);
216     }
217 
218     if XNN_UNLIKELY(k != 0) {
219       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
220       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
221       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
222 
223       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
224       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
225       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
226       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
227 
228       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
229       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
230       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
231       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
232       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
233       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
234       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
235       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
236       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
237       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
238       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
239       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
240       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
241       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
242       const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
243       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
244       const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
245       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
246       const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
247       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
248       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
249       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
250       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
251       const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
252       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
253       const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
254       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
255 
256       if (k > 2 * sizeof(int8_t)) {
257         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
258         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
259         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
260         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
261 
262         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
263         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
264         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
265         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
266         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
267         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
268         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
269         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
270         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
271         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
272         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
273         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
274         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
275         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
276         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
277         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
278         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
279         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
280         const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
281         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
282         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
283         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
284         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
285         const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
286         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
287         const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
288         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
289 
290         if (k > 4 * sizeof(int8_t)) {
291           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
292           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
293           const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
294           const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
295 
296           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
297           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
298           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
299           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
300           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
301           const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
302           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
303           const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
304           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
305           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
306           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
307           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
308           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
309           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
310           const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
311           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
312           const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
313           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
314           const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
315           const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
316           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
317           const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
318           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
319           const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
320           vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
321           const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
322           vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
323         }
324       }
325     }
326 
327     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
328     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
329     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
330 
331     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
332     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
333     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
334     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
335     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
336     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
337     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
338     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
339     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
340     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
341     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
342     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
343 
344     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
345     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
346     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
347     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
348     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
349     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
350     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
351     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
352     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
353     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
354     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
355     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
356 
357     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
358     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
359     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
360     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
361     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
362     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
363     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
364     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
365     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
366     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
367     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
368     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
369 
370     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
371 #if XNN_ARCH_ARM64
372     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
373     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
374     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
375     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
376     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
377     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
378 
379     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
380     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
381     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
382     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
383     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
384     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
385 
386     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
387     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
388     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
389 #else
390     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
391     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
392     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
393     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
394     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
395     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
396 
397     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
398     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
399     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
400     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
401     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
402     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
403 
404     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
405     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
406     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
407 #endif
408 
409     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
410     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
411     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
412     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
413 
414     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
415     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
416     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
417     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
418 
419     if (nc >= 16) {
420       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
421       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
422       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
423 
424       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
425       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
426       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
427 
428       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
429       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
430       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
431 
432       nc -= 16;
433     } else {
434       // Final case where not all of the 16 columns fit in the destination.
435       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
436       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
437       if (nc & 8) {
438         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
439         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
440         vst1_s8(c2, vout2x01234567); c2 += 8;
441         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
442         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
443       }
444       if (nc & 4) {
445         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
446         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
447         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
448         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
449         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
450       }
451       if (nc & 2) {
452         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
453         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
454         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
455         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
456         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
457       }
458       if (nc & 1) {
459         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
460         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
461         vst1_lane_s8(c2, vout2x01234567, 0);
462       }
463 
464       nc = 0;
465     }
466   } while (nc != 0);
467 }
468