xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c4-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 
17 
xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_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)])18 void xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld1r(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const int8_t** restrict a,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const int8_t* zero,
30     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32   assert(mr != 0);
33   assert(mr <= 2);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (2 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(int8_t) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   kc = round_up_po2(kc, 4 * sizeof(int8_t));
44   int8_t* c0 = c;
45   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr != 2) {
47     c1 = c0;
48   }
49 
50   do {
51     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
53     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
54     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
55     int32x4_t vacc0x89 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
56     int32x4_t vacc0xAB = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
57     int32x4_t vacc0xCD = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
58     int32x4_t vacc0xEF = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
59     int32x4_t vacc1x01 = vacc0x01;
60     int32x4_t vacc1x23 = vacc0x23;
61     int32x4_t vacc1x45 = vacc0x45;
62     int32x4_t vacc1x67 = vacc0x67;
63     int32x4_t vacc1x89 = vacc0x89;
64     int32x4_t vacc1xAB = vacc0xAB;
65     int32x4_t vacc1xCD = vacc0xCD;
66     int32x4_t vacc1xEF = vacc0xEF;
67 
68     size_t p = ks;
69     do {
70       const int8_t* restrict a0 = a[0];
71       if XNN_UNPREDICTABLE(a0 != zero) {
72         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
73       }
74       const int8_t* restrict a1 = a[1];
75       if XNN_UNPREDICTABLE(a1 != zero) {
76         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
77       }
78       a += 2;
79 
80       size_t k = kc;
81 
82       while (k >= 16 * sizeof(int8_t)) {
83         const int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
84         const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
85         const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
86         const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
87         const int32x2_t va10x0 = vld1_dup_s32((const void*)a1);
88         const int32x2_t va11x0 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
89         const int32x2_t va10x1 = vld1_dup_s32((const void*)a1);
90         const int32x2_t va11x1 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
91 
92         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96         const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97         const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98         const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99         const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104         const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105         const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106         const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107         const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
108 
109         const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
110         const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
111         const int8x8_t va1c0x0 = vreinterpret_s8_s32(va10x0);
112         const int8x8_t va1c0x1 = vreinterpret_s8_s32(va10x1);
113 
114         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
115         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
116         const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117         vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
118         vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
119         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
120         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
121         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
122         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
123         const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124         vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
125         vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
126         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
127         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
128         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
129         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
130         const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131         vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
132         vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
133         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
134         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
135         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
136         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
137         const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
138         vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
139         vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
140         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
141         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
142         int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0);
143         int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0);
144         const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
145         vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1);
146         vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1);
147         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
148         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
149         int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0);
150         int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1c0x0);
151         const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
152         vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1);
153         vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1c0x1);
154         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
155         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
156         int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0);
157         int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1c0x0);
158         const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
159         vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1);
160         vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1c0x1);
161         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
162         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
163         int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0);
164         int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1c0x0);
165         const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166         vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1);
167         vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1c0x1);
168         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
169         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
170         const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
171         const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
172         const int8x8_t va1c1x0 = vreinterpret_s8_s32(va11x0);
173         const int8x8_t va1c1x1 = vreinterpret_s8_s32(va11x1);
174 
175         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
176         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
177         const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178         vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
179         vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
180         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
181         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
182         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
183         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
184         const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
185         vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
186         vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
187         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
188         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
189         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
190         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
191         const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
192         vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
193         vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
194         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
195         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
196         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
197         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
198         const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
199         vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
200         vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
201         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
202         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
203         int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0);
204         int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0);
205         const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
206         vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1);
207         vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1);
208         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
209         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
210         int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0);
211         int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0);
212         const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
213         vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1);
214         vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1);
215         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
216         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
217         int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0);
218         int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0);
219         const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
220         vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1);
221         vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1c1x1);
222         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
223         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
224         int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0);
225         int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1c1x0);
226         const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
227         vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1);
228         vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1c1x1);
229         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
230         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
231 
232         k -= 16 * sizeof(int8_t);
233       }
234 
235       if (k >= 8 * sizeof(int8_t)) {
236         const int32x2_t va00 = vld1_dup_s32((const void*)a0);
237         const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
238         const int32x2_t va10 = vld1_dup_s32((const void*)a1);
239         const int32x2_t va11 = vld1_dup_s32((const void*)(a1 + 4)); a1 += 8;
240 
241         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
242         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
243         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
244         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
245         const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246         const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
247         const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
248         const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
249         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
252         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
253         const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
254         const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
255         const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
256         const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
257 
258         const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
259         const int8x8_t va1c0 = vreinterpret_s8_s32(va10);
260 
261         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
262         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
263         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
264         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
265         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
266         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
267         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
268         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
269         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
270         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
271         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
272         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
273         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
274         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
275         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
276         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
277         const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
278         const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
279         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
280         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
281         const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
282         const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
283         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
284         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
285         const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
286         const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
287         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
288         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
289         const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
290         const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
291         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
292         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
293         const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
294         const int8x8_t va1c1 = vreinterpret_s8_s32(va11);
295 
296         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
297         const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
298         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
299         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
300         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
301         const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
302         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
303         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
304         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
305         const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
306         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
307         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
308         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
309         const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
310         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
311         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
312         const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
313         const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
314         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
315         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
316         const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
317         const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
318         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
319         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
320         const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
321         const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
322         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
323         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
324         const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
325         const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
326         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
327         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
328 
329         k -= 8 * sizeof(int8_t);
330       }
331 
332       if XNN_UNLIKELY(k != 0) {
333         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
334         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
335 
336         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
337         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
338         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
339         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
340         const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
341         const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
342         const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
343         const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
344 
345         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
346         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
347         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
348         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
349         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
350         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
351         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
352         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
353         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
354         const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
355         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
356         const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
357         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
358         const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
359         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
360         const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
361         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
362         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
363         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
364         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
365         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
366         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
367         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
368         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
369         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
370         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
371         const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
372         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
373         const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
374         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
375         const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
376         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
377         const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
378         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
379       }
380       p -= 2 * sizeof(void*);
381     } while (p != 0);
382 
383 #if XNN_ARCH_ARM64
384     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
385     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
386     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
387     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
388     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
389     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
390     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
391     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
392 #else
393     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
394     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
395     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
396     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
397     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
398     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
399     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
400     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
401     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
402     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
403     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
404     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
405     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
406     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
407     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
408     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
409     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
410     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
411     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
412     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
413     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
414     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
415     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
416     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
417 #endif
418 
419     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
420     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
421     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
422 
423     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
424     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
425     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
426     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
427     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
428     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
429     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
430     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
431 
432     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
433     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
434     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
435     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
436     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
437     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
438     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
439     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
440 
441     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
442     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
443     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
444     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
445     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
446     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
447     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
448     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
449 
450     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
451 #if XNN_ARCH_ARM64
452     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
453     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
454     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
455     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
456 
457     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
458     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
459     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
460     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
461 
462     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
463     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
464 #else
465     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
466     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
467     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
468     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
469 
470     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
471     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
472     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
473     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
474 
475     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
476     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
477 #endif
478 
479     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
480     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
481     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
482 
483     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
484     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
485     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
486 
487     if (nc >= 16) {
488       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
489       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
490 
491       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
492       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
493 
494       a = (const int8_t**restrict) ((uintptr_t) a - ks);
495 
496       nc -= 16;
497     } else {
498       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
499       if (nc & 8) {
500         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
501         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
502         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
503       }
504       if (nc & 4) {
505         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
506         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
507         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
508       }
509       if (nc & 2) {
510         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
511         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
512         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
513       }
514       if (nc & 1) {
515         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
516         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
517       }
518 
519       nc = 0;
520     }
521   } while (nc != 0);
522 }
523