xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/3x16c2-minmax-rndnu-neon-mull-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/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_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld1r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const int8_t** restrict a,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const int8_t* zero,
29     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 3);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (3 * sizeof(void*)) == 0);
37   assert(a_offset % sizeof(int8_t) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   kc = round_up_po2(kc, 2 * sizeof(int8_t));
43   int8_t* c0 = c;
44   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr < 2) {
46     c1 = c0;
47   }
48   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     c2 = c1;
51   }
52 
53   do {
54     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
55     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
56     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
58     int32x4_t vacc1x0123 = vacc0x0123;
59     int32x4_t vacc1x4567 = vacc0x4567;
60     int32x4_t vacc1x89AB = vacc0x89AB;
61     int32x4_t vacc1xCDEF = vacc0xCDEF;
62     int32x4_t vacc2x0123 = vacc0x0123;
63     int32x4_t vacc2x4567 = vacc0x4567;
64     int32x4_t vacc2x89AB = vacc0x89AB;
65     int32x4_t vacc2xCDEF = vacc0xCDEF;
66 
67     size_t p = ks;
68     do {
69       const int8_t* restrict a0 = a[0];
70       if XNN_UNPREDICTABLE(a0 != zero) {
71         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
72       }
73       const int8_t* restrict a1 = a[1];
74       if XNN_UNPREDICTABLE(a1 != zero) {
75         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
76       }
77       const int8_t* restrict a2 = a[2];
78       if XNN_UNPREDICTABLE(a2 != zero) {
79         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
80       }
81       a += 3;
82 
83       size_t k = kc;
84 
85 
86       while (k >= 8 * sizeof(int8_t)) {
87         const int16x4_t va00 = vld1_dup_s16((const void*)a0);
88         const int16x4_t va01 = vld1_dup_s16((const void*)(a0 + 2));
89         const int16x4_t va02 = vld1_dup_s16((const void*)(a0 + 4));
90         const int16x4_t va03 = vld1_dup_s16((const void*)(a0 + 6)); a0 += 8;
91         const int16x4_t va10 = vld1_dup_s16((const void*)a1);
92         const int16x4_t va11 = vld1_dup_s16((const void*)(a1 + 2));
93         const int16x4_t va12 = vld1_dup_s16((const void*)(a1 + 4));
94         const int16x4_t va13 = vld1_dup_s16((const void*)(a1 + 6)); a1 += 8;
95         const int16x4_t va20 = vld1_dup_s16((const void*)a2);
96         const int16x4_t va21 = vld1_dup_s16((const void*)(a2 + 2));
97         const int16x4_t va22 = vld1_dup_s16((const void*)(a2 + 4));
98         const int16x4_t va23 = vld1_dup_s16((const void*)(a2 + 6)); a2 += 8;
99 
100         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
109         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110         const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
111         const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114         const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115         const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116 
117         const int8x8_t va0c0 = vreinterpret_s8_s16(va00);
118         const int8x8_t va1c0 = vreinterpret_s8_s16(va10);
119         const int8x8_t va2c0 = vreinterpret_s8_s16(va20);
120 
121         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
122         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
123         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
124         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
125         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
126         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
127         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
128         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
129         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
130         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
131         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
132         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
133         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
134         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
135         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
136         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
137         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
138         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
139         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
140         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
141         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
142         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
143         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
144         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
145         const int8x8_t va0c1 = vreinterpret_s8_s16(va01);
146         const int8x8_t va1c1 = vreinterpret_s8_s16(va11);
147         const int8x8_t va2c1 = vreinterpret_s8_s16(va21);
148 
149         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
150         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
151         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
152         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
153         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
154         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
155         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
156         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
157         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
158         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
159         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
160         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
161         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
162         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
163         const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
164         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
165         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
166         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
167         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
168         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
169         const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
170         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
171         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
172         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
173         const int8x8_t va0c2 = vreinterpret_s8_s16(va02);
174         const int8x8_t va1c2 = vreinterpret_s8_s16(va12);
175         const int8x8_t va2c2 = vreinterpret_s8_s16(va22);
176 
177         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
178         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
179         const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
180         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
181         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
182         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
183         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
184         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
185         const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
186         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
187         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
188         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
189         const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
190         const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
191         const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
192         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
193         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
194         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
195         const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
196         const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
197         const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
198         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
199         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
200         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
201         const int8x8_t va0c3 = vreinterpret_s8_s16(va03);
202         const int8x8_t va1c3 = vreinterpret_s8_s16(va13);
203         const int8x8_t va2c3 = vreinterpret_s8_s16(va23);
204 
205         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
206         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
207         const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
208         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
209         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
210         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
211         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
212         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
213         const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
214         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
215         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
216         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
217         const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
218         const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
219         const int16x8_t vprod2x89ABc3 = vmull_s8(vb89ABc3, va2c3);
220         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
221         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
222         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc3);
223         const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
224         const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
225         const int16x8_t vprod2xCDEFc3 = vmull_s8(vbCDEFc3, va2c3);
226         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
227         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
228         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc3);
229 
230         k -= 8 * sizeof(int8_t);
231       }
232 
233       if XNN_UNLIKELY(k != 0) {
234         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
235         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
236         const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
237 
238         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
239         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
240         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
241         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
242 
243         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
244         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
245         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
246         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
247         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
248         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
249         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
250         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
251         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
252         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
253         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
254         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
255         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
256         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
257         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
258         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
259         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
260         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
261         const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
262         const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
263         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
264         const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
265         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
266         const int16x8_t vprod2x89ABc0 = vmull_s8(vb89ABc0, va2c0);
267         vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc0);
268         const int16x8_t vprod2xCDEFc0 = vmull_s8(vbCDEFc0, va2c0);
269         vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc0);
270 
271         if (k > 2 * sizeof(int8_t)) {
272           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
273           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
274           const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
275           const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
276 
277           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
278           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
279           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
280           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
281           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
282           const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
283           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
284           const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
285           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
286           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
287           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
288           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
289           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
290           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
291           const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
292           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
293           const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
294           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
295           const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
296           const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
297           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
298           const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
299           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
300           const int16x8_t vprod2x89ABc1 = vmull_s8(vb89ABc1, va2c1);
301           vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc1);
302           const int16x8_t vprod2xCDEFc1 = vmull_s8(vbCDEFc1, va2c1);
303           vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc1);
304 
305           if (k > 4 * sizeof(int8_t)) {
306             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
307             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
308             const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
309             const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
310 
311             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
312             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
313             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
314             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
315             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
316             const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
317             vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
318             const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
319             vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
320             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
321             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
322             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
323             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
324             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
325             const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
326             vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
327             const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
328             vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
329             const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
330             const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
331             vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
332             const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
333             vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
334             const int16x8_t vprod2x89ABc2 = vmull_s8(vb89ABc2, va2c2);
335             vacc2x89AB = vpadalq_s16(vacc2x89AB, vprod2x89ABc2);
336             const int16x8_t vprod2xCDEFc2 = vmull_s8(vbCDEFc2, va2c2);
337             vacc2xCDEF = vpadalq_s16(vacc2xCDEF, vprod2xCDEFc2);
338           }
339         }
340       }
341       p -= 3 * sizeof(void*);
342     } while (p != 0);
343 
344     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
345     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
346     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
347 
348     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
349     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
350     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
351     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
352     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
353     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
354     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
355     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
356     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
357     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
358     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
359     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
360 
361     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
362     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
363     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
364     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
365     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
366     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
367     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
368     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
369     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
370     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
371     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
372     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
373 
374     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
375     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
376     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
377     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
378     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
379     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
380     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
381     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
382     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
383     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
384     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
385     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
386 
387     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
388 #if XNN_ARCH_ARM64
389     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
390     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
391     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
392     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
393     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
394     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
395 
396     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
397     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
398     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
399     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
400     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
401     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
402 
403     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
404     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
405     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
406 #else
407     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
408     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
409     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
410     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
411     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
412     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
413 
414     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
415     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
416     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
417     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
418     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
419     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
420 
421     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
422     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
423     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
424 #endif
425 
426     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
427     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
428     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
429     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
430 
431     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
432     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
433     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
434     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
435 
436     if (nc >= 16) {
437       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
438       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
439       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
440 
441       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
442       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
443       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
444 
445       a = (const int8_t**restrict) ((uintptr_t) a - ks);
446 
447       nc -= 16;
448     } else {
449       int8x8_t vout2x01234567 = vget_low_s8(vout2x0123456789ABCDEF);
450       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
451       if (nc & 8) {
452         vst1_s8(c2, vout2x01234567); c2 += 8;
453         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
454         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
455         vout2x01234567 = vget_high_s8(vout2x0123456789ABCDEF);
456         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
457       }
458       if (nc & 4) {
459         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
460         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
461         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
462         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
463         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
464       }
465       if (nc & 2) {
466         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
467         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
468         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
469         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
470         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
471       }
472       if (nc & 1) {
473         vst1_lane_s8(c2, vout2x01234567, 0);
474         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
475         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
476       }
477 
478       nc = 0;
479     }
480   } while (nc != 0);
481 }
482