xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/2x16c2-minmax-rndnu-neon-mlal-ld2r.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_2x16c2__neon_mlal_ld2r(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_2x16c2__neon_mlal_ld2r(
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 <= 2);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(ks != 0);
36   assert(ks % (2 * 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 
49   do {
50     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
51     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
52     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
53     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
54     int32x4_t vacc1x0123 = vacc0x0123;
55     int32x4_t vacc1x4567 = vacc0x4567;
56     int32x4_t vacc1x89AB = vacc0x89AB;
57     int32x4_t vacc1xCDEF = vacc0xCDEF;
58 
59     size_t p = ks;
60     do {
61       const int8_t* restrict a0 = a[0];
62       if XNN_UNPREDICTABLE(a0 != zero) {
63         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
64       }
65       const int8_t* restrict a1 = a[1];
66       if XNN_UNPREDICTABLE(a1 != zero) {
67         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
68       }
69       a += 2;
70 
71       size_t k = kc;
72 
73       while (k >= 16 * sizeof(int8_t)) {
74         const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
75         const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
76         const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
77         const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
78         const int16x4x2_t va10x0 = vld2_dup_s16((const void*)a1);
79         const int16x4x2_t va11x0 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
80         const int16x4x2_t va10x1 = vld2_dup_s16((const void*)a1);
81         const int16x4x2_t va11x1 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
82 
83         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85         const int8x8_t vb89ABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
86         const int8x8_t vbCDEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
88         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89         const int8x8_t vb89ABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90         const int8x8_t vbCDEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93         const int8x8_t vb89ABc2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94         const int8x8_t vbCDEFc2x0 = 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         const int8x8_t vb89ABc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98         const int8x8_t vbCDEFc3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99 
100         const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
101         const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
102         const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0.val[0]);
103         const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1.val[0]);
104 
105         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
106         int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
107         const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108         vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
109         vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
110         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
111         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
112         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
113         int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
114         const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
115         vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
116         vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
117         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
118         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
119         int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0x0, va0c0x0);
120         int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0x0, va1c0x0);
121         const int8x8_t vb89ABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122         vprod0x89ABc0 = vmlal_s8(vprod0x89ABc0, vb89ABc0x1, va0c0x1);
123         vprod1x89ABc0 = vmlal_s8(vprod1x89ABc0, vb89ABc0x1, va1c0x1);
124         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
125         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
126         int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0x0, va0c0x0);
127         int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0x0, va1c0x0);
128         const int8x8_t vbCDEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129         vprod0xCDEFc0 = vmlal_s8(vprod0xCDEFc0, vbCDEFc0x1, va0c0x1);
130         vprod1xCDEFc0 = vmlal_s8(vprod1xCDEFc0, vbCDEFc0x1, va1c0x1);
131         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
132         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
133         const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
134         const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
135         const int8x8_t va1c1x0 = vreinterpret_s8_s16(va10x0.val[1]);
136         const int8x8_t va1c1x1 = vreinterpret_s8_s16(va10x1.val[1]);
137 
138         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
139         int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
140         const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141         vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
142         vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
143         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
144         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
145         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
146         int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
147         const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
148         vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
149         vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
150         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
151         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
152         int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1x0, va0c1x0);
153         int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1x0, va1c1x0);
154         const int8x8_t vb89ABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155         vprod0x89ABc1 = vmlal_s8(vprod0x89ABc1, vb89ABc1x1, va0c1x1);
156         vprod1x89ABc1 = vmlal_s8(vprod1x89ABc1, vb89ABc1x1, va1c1x1);
157         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
158         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
159         int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1x0, va0c1x0);
160         int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1x0, va1c1x0);
161         const int8x8_t vbCDEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162         vprod0xCDEFc1 = vmlal_s8(vprod0xCDEFc1, vbCDEFc1x1, va0c1x1);
163         vprod1xCDEFc1 = vmlal_s8(vprod1xCDEFc1, vbCDEFc1x1, va1c1x1);
164         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
165         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
166         const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
167         const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
168         const int8x8_t va1c2x0 = vreinterpret_s8_s16(va11x0.val[0]);
169         const int8x8_t va1c2x1 = vreinterpret_s8_s16(va11x1.val[0]);
170 
171         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
172         int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
173         const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174         vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
175         vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
176         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
177         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
178         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
179         int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
180         const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181         vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
182         vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
183         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
184         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
185         int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2x0, va0c2x0);
186         int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2x0, va1c2x0);
187         const int8x8_t vb89ABc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
188         vprod0x89ABc2 = vmlal_s8(vprod0x89ABc2, vb89ABc2x1, va0c2x1);
189         vprod1x89ABc2 = vmlal_s8(vprod1x89ABc2, vb89ABc2x1, va1c2x1);
190         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
191         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
192         int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2x0, va0c2x0);
193         int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2x0, va1c2x0);
194         const int8x8_t vbCDEFc2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
195         vprod0xCDEFc2 = vmlal_s8(vprod0xCDEFc2, vbCDEFc2x1, va0c2x1);
196         vprod1xCDEFc2 = vmlal_s8(vprod1xCDEFc2, vbCDEFc2x1, va1c2x1);
197         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
198         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
199         const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
200         const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
201         const int8x8_t va1c3x0 = vreinterpret_s8_s16(va11x0.val[1]);
202         const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]);
203 
204         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
205         int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
206         const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
207         vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
208         vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
209         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
210         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
211         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
212         int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
213         const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
214         vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
215         vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
216         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
217         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
218         int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3x0, va0c3x0);
219         int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3x0, va1c3x0);
220         const int8x8_t vb89ABc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
221         vprod0x89ABc3 = vmlal_s8(vprod0x89ABc3, vb89ABc3x1, va0c3x1);
222         vprod1x89ABc3 = vmlal_s8(vprod1x89ABc3, vb89ABc3x1, va1c3x1);
223         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
224         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
225         int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3x0, va0c3x0);
226         int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3x0, va1c3x0);
227         const int8x8_t vbCDEFc3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
228         vprod0xCDEFc3 = vmlal_s8(vprod0xCDEFc3, vbCDEFc3x1, va0c3x1);
229         vprod1xCDEFc3 = vmlal_s8(vprod1xCDEFc3, vbCDEFc3x1, va1c3x1);
230         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
231         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
232 
233         k -= 16 * sizeof(int8_t);
234       }
235 
236       if (k >= 8 * sizeof(int8_t)) {
237         const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
238         const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
239         const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
240         const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
241 
242         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
243         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
244         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
245         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
247         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
248         const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
249         const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
252         const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
253         const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
254         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
255         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
256         const int8x8_t vb89ABc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
257         const int8x8_t vbCDEFc3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
258 
259         const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
260         const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
261 
262         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
263         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
264         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
265         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
266         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
267         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
268         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
269         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
270         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
271         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
272         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
273         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
274         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
275         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
276         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
277         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
278         const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
279         const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
280 
281         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
282         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
283         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
284         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
285         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
286         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
287         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
288         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
289         const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
290         const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
291         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
292         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
293         const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
294         const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
295         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
296         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
297         const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
298         const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
299 
300         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
301         const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
302         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
303         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
304         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
305         const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
306         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
307         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
308         const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
309         const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
310         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
311         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
312         const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
313         const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
314         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
315         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
316         const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
317         const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
318 
319         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
320         const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
321         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
322         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
323         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
324         const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
325         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
326         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
327         const int16x8_t vprod0x89ABc3 = vmull_s8(vb89ABc3, va0c3);
328         const int16x8_t vprod1x89ABc3 = vmull_s8(vb89ABc3, va1c3);
329         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc3);
330         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc3);
331         const int16x8_t vprod0xCDEFc3 = vmull_s8(vbCDEFc3, va0c3);
332         const int16x8_t vprod1xCDEFc3 = vmull_s8(vbCDEFc3, va1c3);
333         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc3);
334         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc3);
335 
336         k -= 8 * sizeof(int8_t);
337       }
338 
339       if XNN_UNLIKELY(k != 0) {
340         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
341         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
342 
343         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
344         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
345         const int8x8_t vb89ABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
346         const int8x8_t vbCDEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
347 
348         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
349         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
350         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
351         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
352         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
353         const int16x8_t vprod0x89ABc0 = vmull_s8(vb89ABc0, va0c0);
354         vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc0);
355         const int16x8_t vprod0xCDEFc0 = vmull_s8(vbCDEFc0, va0c0);
356         vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc0);
357         const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
358         const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
359         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
360         const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
361         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
362         const int16x8_t vprod1x89ABc0 = vmull_s8(vb89ABc0, va1c0);
363         vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc0);
364         const int16x8_t vprod1xCDEFc0 = vmull_s8(vbCDEFc0, va1c0);
365         vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc0);
366 
367         if (k > 2 * sizeof(int8_t)) {
368           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
369           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
370           const int8x8_t vb89ABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
371           const int8x8_t vbCDEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
372 
373           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
374           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
375           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
376           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
377           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
378           const int16x8_t vprod0x89ABc1 = vmull_s8(vb89ABc1, va0c1);
379           vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc1);
380           const int16x8_t vprod0xCDEFc1 = vmull_s8(vbCDEFc1, va0c1);
381           vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc1);
382           const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
383           const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
384           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
385           const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
386           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
387           const int16x8_t vprod1x89ABc1 = vmull_s8(vb89ABc1, va1c1);
388           vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc1);
389           const int16x8_t vprod1xCDEFc1 = vmull_s8(vbCDEFc1, va1c1);
390           vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc1);
391 
392           if (k > 4 * sizeof(int8_t)) {
393             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
394             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
395             const int8x8_t vb89ABc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
396             const int8x8_t vbCDEFc2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
397 
398             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
399             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
400             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
401             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
402             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
403             const int16x8_t vprod0x89ABc2 = vmull_s8(vb89ABc2, va0c2);
404             vacc0x89AB = vpadalq_s16(vacc0x89AB, vprod0x89ABc2);
405             const int16x8_t vprod0xCDEFc2 = vmull_s8(vbCDEFc2, va0c2);
406             vacc0xCDEF = vpadalq_s16(vacc0xCDEF, vprod0xCDEFc2);
407             const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
408             const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
409             vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
410             const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
411             vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
412             const int16x8_t vprod1x89ABc2 = vmull_s8(vb89ABc2, va1c2);
413             vacc1x89AB = vpadalq_s16(vacc1x89AB, vprod1x89ABc2);
414             const int16x8_t vprod1xCDEFc2 = vmull_s8(vbCDEFc2, va1c2);
415             vacc1xCDEF = vpadalq_s16(vacc1xCDEF, vprod1xCDEFc2);
416           }
417         }
418       }
419       p -= 2 * sizeof(void*);
420     } while (p != 0);
421 
422     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
423     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
424     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
425 
426     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
427     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
428     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
429     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
430     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
431     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
432     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
433     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
434 
435     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
436     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
437     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
438     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
439     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
440     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
441     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
442     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
443 
444     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
445     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
446     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
447     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
448     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
449     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
450     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
451     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
452 
453     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
454 #if XNN_ARCH_ARM64
455     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
456     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
457     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
458     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
459 
460     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
461     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
462     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
463     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
464 
465     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
466     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
467 #else
468     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
469     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
470     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
471     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
472 
473     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
474     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
475     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
476     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
477 
478     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
479     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
480 #endif
481 
482     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
483     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
484     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
485 
486     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
487     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
488     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
489 
490     if (nc >= 16) {
491       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
492       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
493 
494       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
495       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
496 
497       a = (const int8_t**restrict) ((uintptr_t) a - ks);
498 
499       nc -= 16;
500     } else {
501       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
502       if (nc & 8) {
503         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
504         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
505         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
506       }
507       if (nc & 4) {
508         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
509         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
510         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
511       }
512       if (nc & 2) {
513         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
514         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
515         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
516       }
517       if (nc & 1) {
518         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
519         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
520       }
521 
522       nc = 0;
523     }
524   } while (nc != 0);
525 }
526