xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/4x16-minmax-rndnu-neon-mlal-lane.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/neon-mlal-lane.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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/common.h>
15 #include <xnnpack/gemm.h>
16 
17 
xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 4);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   const int8_t* a0 = a;
40   int8_t* c0 = c;
41   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr < 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     a2 = a1;
51     c2 = c1;
52   }
53   const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
54   int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
55   if XNN_UNPREDICTABLE(mr != 4) {
56     a3 = a2;
57     c3 = c2;
58   }
59 
60   do {
61     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
62     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
63     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
64     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
65     int32x4_t vacc1x0123 = vacc0x0123;
66     int32x4_t vacc1x4567 = vacc0x4567;
67     int32x4_t vacc1x89AB = vacc0x89AB;
68     int32x4_t vacc1xCDEF = vacc0xCDEF;
69     int32x4_t vacc2x0123 = vacc0x0123;
70     int32x4_t vacc2x4567 = vacc0x4567;
71     int32x4_t vacc2x89AB = vacc0x89AB;
72     int32x4_t vacc2xCDEF = vacc0xCDEF;
73     int32x4_t vacc3x0123 = vacc0x0123;
74     int32x4_t vacc3x4567 = vacc0x4567;
75     int32x4_t vacc3x89AB = vacc0x89AB;
76     int32x4_t vacc3xCDEF = vacc0xCDEF;
77 
78     size_t k = kc;
79     while (k >= 8 * sizeof(int8_t)) {
80       const int8x8_t va0 = vld1_s8(a0); a0 += 8;
81       const int16x8_t vxa0 = vmovl_s8(va0);
82       const int8x8_t va1 = vld1_s8(a1); a1 += 8;
83       const int16x8_t vxa1 = vmovl_s8(va1);
84       const int8x8_t va2 = vld1_s8(a2); a2 += 8;
85       const int16x8_t vxa2 = vmovl_s8(va2);
86       const int8x8_t va3 = vld1_s8(a3); a3 += 8;
87       const int16x8_t vxa3 = vmovl_s8(va3);
88 
89       const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
90       const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
91 
92       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
93       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
94       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
95       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
96       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
97       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
98       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
99       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
100       const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
101       const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
102 
103       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
104       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
105       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
106       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
107       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
108       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
109       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
110       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
111       const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
112       const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
113 
114       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
115       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
116       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
117       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
118       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
119       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
120       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
121       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
122       const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
123       const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
124 
125       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
126       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
127       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
128       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
129       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
130       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
131       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
132       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
133       const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
134       const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
135 
136       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
137       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
138       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
139       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
140       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
141       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
142       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
143       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
144       const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
145       const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
146 
147       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
148       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
149       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
150       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
151       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
152       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
153       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
154       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
155       const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
156       const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
157 
158       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
159       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
160       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
161       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
162       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
163       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
164       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
165       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
166       const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
167       const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
168 
169       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
170       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
171       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
172       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
173       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
174       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
175       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
176       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
177 
178 
179       const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
180       const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
181 
182       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
183       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
184       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
185       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
186       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
187       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
188       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
189       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
190       const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
191       const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
192 
193       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
194       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
195       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
196       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
197       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
198       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
199       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
200       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
201       const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
202       const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
203 
204       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
205       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
206       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
207       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
208       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
209       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
210       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
211       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
212       const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
213       const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
214 
215       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
216       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
217       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
218       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
219       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
220       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
221       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
222       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
223       const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
224       const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
225 
226       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
227       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
228       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
229       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
230       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
231       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
232       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
233       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
234       const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
235       const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
236 
237       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
238       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
239       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
240       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
241       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
242       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
243       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
244       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
245       const int8x8_t vb01234567c7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
246       const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7);
247 
248       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
249       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
250       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
251       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
252       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
253       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3);
254       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
255       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3);
256       const int8x8_t vb89ABCDEFc7 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
257       const int16x8_t vxb89ABCDEFc7 = vmovl_s8(vb89ABCDEFc7);
258 
259       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
260       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
261       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
262       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
263       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);
264       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa2), 3);
265       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);
266       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa3), 3);
267 
268       k -= 8 * sizeof(int8_t);
269     }
270     if XNN_UNLIKELY(k != 0) {
271       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
272       const int16x8_t vxa0 = vmovl_s8(va0);
273       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
274       const int16x8_t vxa1 = vmovl_s8(va1);
275       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
276       const int16x8_t vxa2 = vmovl_s8(va2);
277       const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k);
278       const int16x8_t vxa3 = vmovl_s8(va3);
279 
280       const int8x8_t vb01234567c0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
281       const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0);
282       const int8x8_t vb89ABCDEFc0 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
283       const int16x8_t vxb89ABCDEFc0 = vmovl_s8(vb89ABCDEFc0);
284 
285       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
286       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
287       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
288       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
289       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
290       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
291       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
292       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
293       vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
294       vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0);
295       vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
296       vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa2), 0);
297       vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
298       vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0);
299       vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
300       vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa3), 0);
301 
302       if (k >= 2 * sizeof(int8_t)) {
303         const int8x8_t vb01234567c1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
304         const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1);
305         const int8x8_t vb89ABCDEFc1 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
306         const int16x8_t vxb89ABCDEFc1 = vmovl_s8(vb89ABCDEFc1);
307 
308         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
309         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
310         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
311         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
312         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
313         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
314         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
315         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
316         vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
317         vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1);
318         vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
319         vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa2), 1);
320         vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
321         vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1);
322         vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
323         vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa3), 1);
324 
325         if (k > 2 * sizeof(int8_t)) {
326           const int8x8_t vb01234567c2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
327           const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2);
328           const int8x8_t vb89ABCDEFc2 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
329           const int16x8_t vxb89ABCDEFc2 = vmovl_s8(vb89ABCDEFc2);
330 
331           vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
332           vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
333           vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
334           vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
335           vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
336           vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
337           vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
338           vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
339           vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
340           vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2);
341           vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
342           vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa2), 2);
343           vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
344           vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2);
345           vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
346           vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa3), 2);
347 
348           if (k >= 4 * sizeof(int8_t)) {
349             const int8x8_t vb01234567c3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
350             const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3);
351             const int8x8_t vb89ABCDEFc3 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
352             const int16x8_t vxb89ABCDEFc3 = vmovl_s8(vb89ABCDEFc3);
353 
354             vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
355             vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
356             vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
357             vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
358             vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
359             vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
360             vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
361             vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
362             vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
363             vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3);
364             vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
365             vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa2), 3);
366             vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
367             vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3);
368             vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
369             vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa3), 3);
370 
371             if (k > 4 * sizeof(int8_t)) {
372               const int8x8_t vb01234567c4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
373               const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4);
374               const int8x8_t vb89ABCDEFc4 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
375               const int16x8_t vxb89ABCDEFc4 = vmovl_s8(vb89ABCDEFc4);
376 
377               vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
378               vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
379               vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
380               vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
381               vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
382               vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
383               vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
384               vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
385               vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
386               vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0);
387               vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
388               vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa2), 0);
389               vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
390               vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0);
391               vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
392               vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa3), 0);
393 
394               if (k >= 6 * sizeof(int8_t)) {
395                 const int8x8_t vb01234567c5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
396                 const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5);
397                 const int8x8_t vb89ABCDEFc5 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
398                 const int16x8_t vxb89ABCDEFc5 = vmovl_s8(vb89ABCDEFc5);
399 
400                 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
401                 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
402                 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
403                 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
404                 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
405                 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
406                 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
407                 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
408                 vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
409                 vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1);
410                 vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
411                 vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa2), 1);
412                 vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
413                 vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1);
414                 vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
415                 vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa3), 1);
416 
417                 if (k > 6 * sizeof(int8_t)) {
418                   const int8x8_t vb01234567c6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
419                   const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6);
420                   const int8x8_t vb89ABCDEFc6 = vld1_s8(w); w = (const void*) ((const int8_t*) w + 8);
421                   const int16x8_t vxb89ABCDEFc6 = vmovl_s8(vb89ABCDEFc6);
422 
423                   vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
424                   vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
425                   vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
426                   vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
427                   vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
428                   vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
429                   vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
430                   vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
431                   vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
432                   vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2);
433                   vacc2x89AB = vmlal_lane_s16(vacc2x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
434                   vacc2xCDEF = vmlal_lane_s16(vacc2xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa2), 2);
435                   vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
436                   vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2);
437                   vacc3x89AB = vmlal_lane_s16(vacc3x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
438                   vacc3xCDEF = vmlal_lane_s16(vacc3xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa3), 2);
439                 }
440               }
441             }
442           }
443         }
444       }
445     }
446 
447     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
448     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
449     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
450 
451     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
452     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
453     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
454     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
455     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
456     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
457     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
458     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
459     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
460     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
461     vacc2x89AB = vqshlq_s32(vacc2x89AB, vright_pre_shift);
462     vacc2xCDEF = vqshlq_s32(vacc2xCDEF, vright_pre_shift);
463     vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
464     vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
465     vacc3x89AB = vqshlq_s32(vacc3x89AB, vright_pre_shift);
466     vacc3xCDEF = vqshlq_s32(vacc3xCDEF, vright_pre_shift);
467 
468     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
469     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
470     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
471     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
472     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
473     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
474     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
475     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
476     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
477     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
478     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
479     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
480     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
481     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
482     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
483     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
484 
485     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
486     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
487     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
488     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
489     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
490     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
491     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
492     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
493     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
494     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
495     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
496     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
497     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
498     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
499     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
500     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
501 
502     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
503 #if XNN_ARCH_ARM64
504     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
505     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
506     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
507     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
508     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
509     int16x8_t vacc2x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF);
510     int16x8_t vacc3x01234567 = vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567);
511     int16x8_t vacc3x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF);
512 
513     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
514     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
515     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
516     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
517     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
518     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
519     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
520     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
521 
522     int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
523     int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF);
524     int8x16_t vout2x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc2x89ABCDEF);
525     int8x16_t vout3x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc3x01234567), vacc3x89ABCDEF);
526 #else
527     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
528     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
529     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
530     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
531     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
532     int16x8_t vacc2x89ABCDEF = vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF));
533     int16x8_t vacc3x01234567 = vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567));
534     int16x8_t vacc3x89ABCDEF = vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF));
535 
536     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
537     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
538     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
539     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
540     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
541     vacc2x89ABCDEF = vqaddq_s16(vacc2x89ABCDEF, voutput_zero_point);
542     vacc3x01234567 = vqaddq_s16(vacc3x01234567, voutput_zero_point);
543     vacc3x89ABCDEF = vqaddq_s16(vacc3x89ABCDEF, voutput_zero_point);
544 
545     int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
546     int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF));
547     int8x16_t vout2x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc2x89ABCDEF));
548     int8x16_t vout3x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc3x01234567), vqmovn_s16(vacc3x89ABCDEF));
549 #endif
550 
551     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
552     vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
553     vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min);
554     vout2x0123456789ABCDEF = vmaxq_s8(vout2x0123456789ABCDEF, voutput_min);
555     vout3x0123456789ABCDEF = vmaxq_s8(vout3x0123456789ABCDEF, voutput_min);
556 
557     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
558     vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
559     vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max);
560     vout2x0123456789ABCDEF = vminq_s8(vout2x0123456789ABCDEF, voutput_max);
561     vout3x0123456789ABCDEF = vminq_s8(vout3x0123456789ABCDEF, voutput_max);
562 
563     if (nc >= 16) {
564       vst1q_s8(c0 + 0, vout0x0123456789ABCDEF);
565       vst1q_s8(c1 + 0, vout1x0123456789ABCDEF);
566       vst1q_s8(c2 + 0, vout2x0123456789ABCDEF);
567       vst1q_s8(c3 + 0, vout3x0123456789ABCDEF);
568 
569       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
570       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
571       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
572       c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
573 
574       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
575       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
576       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
577       a3 = (const int8_t*) ((uintptr_t) a3 - kc);
578 
579       nc -= 16;
580     } else {
581       int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF));
582       int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vget_low_s8(vout2x0123456789ABCDEF), vget_low_s8(vout3x0123456789ABCDEF));
583       if (nc & 8) {
584         vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8;
585         vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8;
586         vst1_s8(c2, vget_low_s8(vout2x01234567_3x01234567)); c2 += 8;
587         vst1_s8(c3, vget_high_s8(vout2x01234567_3x01234567)); c3 += 8;
588         vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF));
589         vout2x01234567_3x01234567 = vcombine_s8(vget_high_s8(vout2x0123456789ABCDEF), vget_high_s8(vout3x0123456789ABCDEF));
590       }
591       if (nc & 4) {
592         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
593         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
594         vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
595         vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
596         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
597         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
598       }
599       if (nc & 2) {
600         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
601         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
602         vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
603         vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
604         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
605         vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
606       }
607       if (nc & 1) {
608         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
609         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
610         vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
611         vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
612       }
613 
614       nc = 0;
615     }
616   } while (nc != 0);
617 }
618