xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/2x16c4-minmax-rndnu-neon-mlal-dup.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_dup(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_dup(
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 int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
84         const int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
85         const int8x8_t va1x0 = vld1_s8(a1); a1 += 8;
86         const int8x8_t va1x1 = vld1_s8(a1); a1 += 8;
87 
88         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
89         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
92         const int8x8_t vb89c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93         const int8x8_t vbABc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94         const int8x8_t vbCDc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95         const int8x8_t vbEFc0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
96         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
98         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
99         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
100         const int8x8_t vb89c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101         const int8x8_t vbABc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102         const int8x8_t vbCDc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
103         const int8x8_t vbEFc1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
104 
105         const int8x8_t va0c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 0));
106         const int8x8_t va0c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 0));
107         const int8x8_t va1c0x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 0));
108         const int8x8_t va1c0x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 0));
109 
110         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
111         int16x8_t vprod1x01c0 = vmull_s8(vb01c0x0, va1c0x0);
112         const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
113         vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
114         vprod1x01c0 = vmlal_s8(vprod1x01c0, vb01c0x1, va1c0x1);
115         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
116         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
117         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
118         int16x8_t vprod1x23c0 = vmull_s8(vb23c0x0, va1c0x0);
119         const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120         vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
121         vprod1x23c0 = vmlal_s8(vprod1x23c0, vb23c0x1, va1c0x1);
122         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
123         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
124         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
125         int16x8_t vprod1x45c0 = vmull_s8(vb45c0x0, va1c0x0);
126         const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127         vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
128         vprod1x45c0 = vmlal_s8(vprod1x45c0, vb45c0x1, va1c0x1);
129         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
130         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
131         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
132         int16x8_t vprod1x67c0 = vmull_s8(vb67c0x0, va1c0x0);
133         const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134         vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
135         vprod1x67c0 = vmlal_s8(vprod1x67c0, vb67c0x1, va1c0x1);
136         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
137         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
138         int16x8_t vprod0x89c0 = vmull_s8(vb89c0x0, va0c0x0);
139         int16x8_t vprod1x89c0 = vmull_s8(vb89c0x0, va1c0x0);
140         const int8x8_t vb89c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
141         vprod0x89c0 = vmlal_s8(vprod0x89c0, vb89c0x1, va0c0x1);
142         vprod1x89c0 = vmlal_s8(vprod1x89c0, vb89c0x1, va1c0x1);
143         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
144         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
145         int16x8_t vprod0xABc0 = vmull_s8(vbABc0x0, va0c0x0);
146         int16x8_t vprod1xABc0 = vmull_s8(vbABc0x0, va1c0x0);
147         const int8x8_t vbABc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
148         vprod0xABc0 = vmlal_s8(vprod0xABc0, vbABc0x1, va0c0x1);
149         vprod1xABc0 = vmlal_s8(vprod1xABc0, vbABc0x1, va1c0x1);
150         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
151         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
152         int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0x0, va0c0x0);
153         int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0x0, va1c0x0);
154         const int8x8_t vbCDc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
155         vprod0xCDc0 = vmlal_s8(vprod0xCDc0, vbCDc0x1, va0c0x1);
156         vprod1xCDc0 = vmlal_s8(vprod1xCDc0, vbCDc0x1, va1c0x1);
157         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
158         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
159         int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0x0, va0c0x0);
160         int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0x0, va1c0x0);
161         const int8x8_t vbEFc0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162         vprod0xEFc0 = vmlal_s8(vprod0xEFc0, vbEFc0x1, va0c0x1);
163         vprod1xEFc0 = vmlal_s8(vprod1xEFc0, vbEFc0x1, va1c0x1);
164         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
165         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
166         const int8x8_t va0c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x0), 1));
167         const int8x8_t va0c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0x1), 1));
168         const int8x8_t va1c1x0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x0), 1));
169         const int8x8_t va1c1x1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1x1), 1));
170 
171         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
172         int16x8_t vprod1x01c1 = vmull_s8(vb01c1x0, va1c1x0);
173         const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
174         vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
175         vprod1x01c1 = vmlal_s8(vprod1x01c1, vb01c1x1, va1c1x1);
176         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
177         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
178         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
179         int16x8_t vprod1x23c1 = vmull_s8(vb23c1x0, va1c1x0);
180         const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
181         vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
182         vprod1x23c1 = vmlal_s8(vprod1x23c1, vb23c1x1, va1c1x1);
183         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
184         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
185         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
186         int16x8_t vprod1x45c1 = vmull_s8(vb45c1x0, va1c1x0);
187         const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
188         vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
189         vprod1x45c1 = vmlal_s8(vprod1x45c1, vb45c1x1, va1c1x1);
190         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
191         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
192         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
193         int16x8_t vprod1x67c1 = vmull_s8(vb67c1x0, va1c1x0);
194         const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
195         vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
196         vprod1x67c1 = vmlal_s8(vprod1x67c1, vb67c1x1, va1c1x1);
197         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
198         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
199         int16x8_t vprod0x89c1 = vmull_s8(vb89c1x0, va0c1x0);
200         int16x8_t vprod1x89c1 = vmull_s8(vb89c1x0, va1c1x0);
201         const int8x8_t vb89c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
202         vprod0x89c1 = vmlal_s8(vprod0x89c1, vb89c1x1, va0c1x1);
203         vprod1x89c1 = vmlal_s8(vprod1x89c1, vb89c1x1, va1c1x1);
204         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
205         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
206         int16x8_t vprod0xABc1 = vmull_s8(vbABc1x0, va0c1x0);
207         int16x8_t vprod1xABc1 = vmull_s8(vbABc1x0, va1c1x0);
208         const int8x8_t vbABc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
209         vprod0xABc1 = vmlal_s8(vprod0xABc1, vbABc1x1, va0c1x1);
210         vprod1xABc1 = vmlal_s8(vprod1xABc1, vbABc1x1, va1c1x1);
211         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
212         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
213         int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1x0, va0c1x0);
214         int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1x0, va1c1x0);
215         const int8x8_t vbCDc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
216         vprod0xCDc1 = vmlal_s8(vprod0xCDc1, vbCDc1x1, va0c1x1);
217         vprod1xCDc1 = vmlal_s8(vprod1xCDc1, vbCDc1x1, va1c1x1);
218         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
219         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
220         int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1x0, va0c1x0);
221         int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1x0, va1c1x0);
222         const int8x8_t vbEFc1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
223         vprod0xEFc1 = vmlal_s8(vprod0xEFc1, vbEFc1x1, va0c1x1);
224         vprod1xEFc1 = vmlal_s8(vprod1xEFc1, vbEFc1x1, va1c1x1);
225         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
226         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
227 
228         k -= 16 * sizeof(int8_t);
229       }
230 
231       if (k >= 8 * sizeof(int8_t)) {
232         const int8x8_t va0 = vld1_s8(a0); a0 += 8;
233         const int8x8_t va1 = vld1_s8(a1); a1 += 8;
234 
235         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
236         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
237         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
238         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
239         const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
240         const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
241         const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
242         const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
243         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
244         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
245         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
246         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
247         const int8x8_t vb89c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
248         const int8x8_t vbABc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
249         const int8x8_t vbCDc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
250         const int8x8_t vbEFc1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
251 
252         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
253         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
254 
255         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
256         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
257         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
258         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
259         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
260         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
261         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
262         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
263         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
264         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
265         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
266         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
267         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
268         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
269         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
270         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
271         const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
272         const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
273         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
274         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
275         const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
276         const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
277         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
278         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
279         const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
280         const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
281         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
282         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
283         const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
284         const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
285         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
286         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
287         const int8x8_t va0c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 1));
288         const int8x8_t va1c1 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 1));
289 
290         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
291         const int16x8_t vprod1x01c1 = vmull_s8(vb01c1, va1c1);
292         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
293         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c1);
294         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
295         const int16x8_t vprod1x23c1 = vmull_s8(vb23c1, va1c1);
296         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
297         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c1);
298         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
299         const int16x8_t vprod1x45c1 = vmull_s8(vb45c1, va1c1);
300         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
301         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c1);
302         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
303         const int16x8_t vprod1x67c1 = vmull_s8(vb67c1, va1c1);
304         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
305         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c1);
306         const int16x8_t vprod0x89c1 = vmull_s8(vb89c1, va0c1);
307         const int16x8_t vprod1x89c1 = vmull_s8(vb89c1, va1c1);
308         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c1);
309         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c1);
310         const int16x8_t vprod0xABc1 = vmull_s8(vbABc1, va0c1);
311         const int16x8_t vprod1xABc1 = vmull_s8(vbABc1, va1c1);
312         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc1);
313         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc1);
314         const int16x8_t vprod0xCDc1 = vmull_s8(vbCDc1, va0c1);
315         const int16x8_t vprod1xCDc1 = vmull_s8(vbCDc1, va1c1);
316         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc1);
317         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc1);
318         const int16x8_t vprod0xEFc1 = vmull_s8(vbEFc1, va0c1);
319         const int16x8_t vprod1xEFc1 = vmull_s8(vbEFc1, va1c1);
320         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc1);
321         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc1);
322 
323         k -= 8 * sizeof(int8_t);
324       }
325 
326       if XNN_UNLIKELY(k != 0) {
327         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
328         const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
329 
330         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
331         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
332         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
333         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
334         const int8x8_t vb89c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
335         const int8x8_t vbABc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
336         const int8x8_t vbCDc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
337         const int8x8_t vbEFc0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
338 
339         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
340         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
341         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
342         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
343         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
344         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
345         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
346         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
347         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
348         const int16x8_t vprod0x89c0 = vmull_s8(vb89c0, va0c0);
349         vacc0x89 = vpadalq_s16(vacc0x89, vprod0x89c0);
350         const int16x8_t vprod0xABc0 = vmull_s8(vbABc0, va0c0);
351         vacc0xAB = vpadalq_s16(vacc0xAB, vprod0xABc0);
352         const int16x8_t vprod0xCDc0 = vmull_s8(vbCDc0, va0c0);
353         vacc0xCD = vpadalq_s16(vacc0xCD, vprod0xCDc0);
354         const int16x8_t vprod0xEFc0 = vmull_s8(vbEFc0, va0c0);
355         vacc0xEF = vpadalq_s16(vacc0xEF, vprod0xEFc0);
356         const int8x8_t va1c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va1), 0));
357         const int16x8_t vprod1x01c0 = vmull_s8(vb01c0, va1c0);
358         vacc1x01 = vpadalq_s16(vacc1x01, vprod1x01c0);
359         const int16x8_t vprod1x23c0 = vmull_s8(vb23c0, va1c0);
360         vacc1x23 = vpadalq_s16(vacc1x23, vprod1x23c0);
361         const int16x8_t vprod1x45c0 = vmull_s8(vb45c0, va1c0);
362         vacc1x45 = vpadalq_s16(vacc1x45, vprod1x45c0);
363         const int16x8_t vprod1x67c0 = vmull_s8(vb67c0, va1c0);
364         vacc1x67 = vpadalq_s16(vacc1x67, vprod1x67c0);
365         const int16x8_t vprod1x89c0 = vmull_s8(vb89c0, va1c0);
366         vacc1x89 = vpadalq_s16(vacc1x89, vprod1x89c0);
367         const int16x8_t vprod1xABc0 = vmull_s8(vbABc0, va1c0);
368         vacc1xAB = vpadalq_s16(vacc1xAB, vprod1xABc0);
369         const int16x8_t vprod1xCDc0 = vmull_s8(vbCDc0, va1c0);
370         vacc1xCD = vpadalq_s16(vacc1xCD, vprod1xCDc0);
371         const int16x8_t vprod1xEFc0 = vmull_s8(vbEFc0, va1c0);
372         vacc1xEF = vpadalq_s16(vacc1xEF, vprod1xEFc0);
373       }
374       p -= 2 * sizeof(void*);
375     } while (p != 0);
376 
377 #if XNN_ARCH_ARM64
378     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
379     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
380     int32x4_t vacc0x89AB = vpaddq_s32(vacc0x89, vacc0xAB);
381     int32x4_t vacc0xCDEF = vpaddq_s32(vacc0xCD, vacc0xEF);
382     int32x4_t vacc1x0123 = vpaddq_s32(vacc1x01, vacc1x23);
383     int32x4_t vacc1x4567 = vpaddq_s32(vacc1x45, vacc1x67);
384     int32x4_t vacc1x89AB = vpaddq_s32(vacc1x89, vacc1xAB);
385     int32x4_t vacc1xCDEF = vpaddq_s32(vacc1xCD, vacc1xEF);
386 #else
387     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
388     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
389     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
390     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
391     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
392     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
393     const int32x2_t vsum0x89 = vpadd_s32(vget_low_s32(vacc0x89), vget_high_s32(vacc0x89));
394     const int32x2_t vsum0xAB = vpadd_s32(vget_low_s32(vacc0xAB), vget_high_s32(vacc0xAB));
395     int32x4_t vacc0x89AB = vcombine_s32(vsum0x89, vsum0xAB);
396     const int32x2_t vsum0xCD = vpadd_s32(vget_low_s32(vacc0xCD), vget_high_s32(vacc0xCD));
397     const int32x2_t vsum0xEF = vpadd_s32(vget_low_s32(vacc0xEF), vget_high_s32(vacc0xEF));
398     int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF);
399     const int32x2_t vsum1x01 = vpadd_s32(vget_low_s32(vacc1x01), vget_high_s32(vacc1x01));
400     const int32x2_t vsum1x23 = vpadd_s32(vget_low_s32(vacc1x23), vget_high_s32(vacc1x23));
401     int32x4_t vacc1x0123 = vcombine_s32(vsum1x01, vsum1x23);
402     const int32x2_t vsum1x45 = vpadd_s32(vget_low_s32(vacc1x45), vget_high_s32(vacc1x45));
403     const int32x2_t vsum1x67 = vpadd_s32(vget_low_s32(vacc1x67), vget_high_s32(vacc1x67));
404     int32x4_t vacc1x4567 = vcombine_s32(vsum1x45, vsum1x67);
405     const int32x2_t vsum1x89 = vpadd_s32(vget_low_s32(vacc1x89), vget_high_s32(vacc1x89));
406     const int32x2_t vsum1xAB = vpadd_s32(vget_low_s32(vacc1xAB), vget_high_s32(vacc1xAB));
407     int32x4_t vacc1x89AB = vcombine_s32(vsum1x89, vsum1xAB);
408     const int32x2_t vsum1xCD = vpadd_s32(vget_low_s32(vacc1xCD), vget_high_s32(vacc1xCD));
409     const int32x2_t vsum1xEF = vpadd_s32(vget_low_s32(vacc1xEF), vget_high_s32(vacc1xEF));
410     int32x4_t vacc1xCDEF = vcombine_s32(vsum1xCD, vsum1xEF);
411 #endif
412 
413     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
414     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
415     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
416 
417     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
418     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
419     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
420     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
421     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
422     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
423     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
424     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
425 
426     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
427     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
428     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
429     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
430     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
431     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
432     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
433     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
434 
435     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
436     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
437     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
438     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
439     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
440     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
441     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
442     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
443 
444     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
445 #if XNN_ARCH_ARM64
446     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
447     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
448     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
449     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
450 
451     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
452     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
453     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
454     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
455 
456     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
457     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
458 #else
459     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
460     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
461     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
462     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
463 
464     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
465     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
466     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
467     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
468 
469     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
470     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
471 #endif
472 
473     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
474     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
475     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
476 
477     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
478     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
479     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
480 
481     if (nc >= 16) {
482       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
483       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
484 
485       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
486       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
487 
488       a = (const int8_t**restrict) ((uintptr_t) a - ks);
489 
490       nc -= 16;
491     } else {
492       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
493       if (nc & 8) {
494         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
495         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
496         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
497       }
498       if (nc & 4) {
499         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
500         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
501         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
502       }
503       if (nc & 2) {
504         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
505         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
506         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
507       }
508       if (nc & 1) {
509         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
510         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
511       }
512 
513       nc = 0;
514     }
515   } while (nc != 0);
516 }
517