• 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_ld4r(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_ld4r(
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 int16x4x4_t va0 = vld4_dup_s16((const void*)a0); a0 += 8;
73       const int16x4x4_t va1 = vld4_dup_s16((const void*)a1); a1 += 8;
74       const int16x4x4_t va2 = vld4_dup_s16((const void*)a2); a2 += 8;
75 
76       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82       const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83       const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86       const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87       const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91       const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92 
93       const int8x8_t va0c0 = vreinterpret_s8_s16(va0.val[0]);
94       const int8x8_t va1c0 = vreinterpret_s8_s16(va1.val[0]);
95       const int8x8_t va2c0 = vreinterpret_s8_s16(va2.val[0]);
96 
97       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
98       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
99       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
100       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
101       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
102       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
103       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
104       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
105       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
106       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
107       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
108       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
109       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
110       const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
111       const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
112       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
113       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
114       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
115       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
116       const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
117       const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
118       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
119       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
120       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
121       const int8x8_t va0c1 = vreinterpret_s8_s16(va0.val[1]);
122       const int8x8_t va1c1 = vreinterpret_s8_s16(va1.val[1]);
123       const int8x8_t va2c1 = vreinterpret_s8_s16(va2.val[1]);
124 
125       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
126       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
127       const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
128       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
129       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
130       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
131       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
132       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
133       const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
134       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
135       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
136       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
137       const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
138       const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
139       const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
140       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
141       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
142       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
143       const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
144       const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
145       const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
146       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
147       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
148       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
149       const int8x8_t va0c2 = vreinterpret_s8_s16(va0.val[2]);
150       const int8x8_t va1c2 = vreinterpret_s8_s16(va1.val[2]);
151       const int8x8_t va2c2 = vreinterpret_s8_s16(va2.val[2]);
152 
153       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
154       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
155       const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
156       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
157       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
158       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
159       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
160       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
161       const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
162       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
163       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
164       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
165       const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
166       const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
167       const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
168       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
169       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
170       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
171       const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
172       const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
173       const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
174       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
175       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
176       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
177       const int8x8_t va0c3 = vreinterpret_s8_s16(va0.val[3]);
178       const int8x8_t va1c3 = vreinterpret_s8_s16(va1.val[3]);
179       const int8x8_t va2c3 = vreinterpret_s8_s16(va2.val[3]);
180 
181       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
182       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
183       const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
184       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
185       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
186       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
187       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
188       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
189       const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
190       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
191       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
192       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
193       const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
194       const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
195       const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
196       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
197       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
198       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
199       const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
200       const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
201       const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
202       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
203       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
204       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
205 
206       k -= 8 * sizeof(int8_t);
207     }
208 
209     if XNN_UNLIKELY(k != 0) {
210       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
211       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
212       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
213 
214       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
215       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
216       const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
217       const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
218 
219       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
220       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
221       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
222       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
223       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
224       const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
225       vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
226       const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
227       vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
228       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
229       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
230       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
231       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
232       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
233       const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
234       vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
235       const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
236       vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
237       const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
238       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
239       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
240       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
241       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
242       const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
243       vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
244       const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
245       vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
246 
247       if (k > 2 * sizeof(int8_t)) {
248         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
249         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
252 
253         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
254         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
255         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
256         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
257         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
258         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
259         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
260         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
261         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
262         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
263         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
264         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
265         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
266         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
267         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
268         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
269         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
270         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
271         const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
272         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
273         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
274         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
275         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
276         const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
277         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
278         const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
279         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
280 
281         if (k > 4 * sizeof(int8_t)) {
282           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
283           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
284           const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
285           const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
286 
287           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
288           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
289           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
290           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
291           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
292           const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
293           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
294           const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
295           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
296           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
297           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
298           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
299           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
300           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
301           const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
302           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
303           const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
304           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
305           const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
306           const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
307           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
308           const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
309           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
310           const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
311           vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
312           const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
313           vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
314         }
315       }
316     }
317 
318     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
319     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
320     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
321 
322     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
323     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
324     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
325     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
326     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
327     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
328     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
329     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
330     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
331     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
332     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
333     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
334 
335     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
336     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
337     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
338     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
339     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
340     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
341     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
342     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
343     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
344     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
345     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
346     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
347 
348     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
349     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
350     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
351     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
352     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
353     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
354     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
355     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
356     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
357     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
358     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
359     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
360 
361     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
362 #if XNN_ARCH_ARM64
363     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
364     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
365     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
366     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
367     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
368     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
369 
370     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
371     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
372     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
373     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
374     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
375     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
376 
377     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
378     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
379     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
380 #else
381     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
382     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
383     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
384     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
385     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
386     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
387 
388     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
389     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
390     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
391     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
392     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
393     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
394 
395     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
396     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
397     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
398 #endif
399 
400     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
401     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
402     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
403     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
404 
405     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
406     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
407     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
408     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
409 
410     if (nc >= 16) {
411       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
412       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
413       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
414 
415       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
416       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
417       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
418 
419       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
420       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
421       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
422 
423       nc -= 16;
424     } else {
425       // Final case where not all of the 16 columns fit in the destination.
426       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
427       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
428       if (nc & 8) {
429         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
430         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
431         vst1_s8(c2, vout2x01234567); c2 += 8;
432         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
433         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
434       }
435       if (nc & 4) {
436         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
437         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
438         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
439         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
440         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
441       }
442       if (nc & 2) {
443         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
444         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
445         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
446         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
447         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
448       }
449       if (nc & 1) {
450         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
451         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
452         vst1_lane_s8(c2, vout2x01234567, 0);
453       }
454 
455       nc = 0;
456     }
457   } while (nc != 0);
458 }
459