xref: /aosp_15_r20/external/XNNPACK/src/qu8-gemm/gen/2x16-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_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const uint8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     uint8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 2);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(uint8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   const uint8_t* a0 = a;
40   uint8_t* c0 = c;
41   const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
42   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr != 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47 
48   const uint8x8_t vb_zero_point = vld1_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
49   do {
50     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
51     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
52     int32x4_t vacc0x89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
53     int32x4_t vacc0xCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
54     int32x4_t vacc1x0123 = vacc0x0123;
55     int32x4_t vacc1x4567 = vacc0x4567;
56     int32x4_t vacc1x89AB = vacc0x89AB;
57     int32x4_t vacc1xCDEF = vacc0xCDEF;
58 
59     size_t k = kc;
60     while (k >= 8 * sizeof(uint8_t)) {
61       const uint8x8_t va0 = vld1_u8(a0); a0 += 8;
62       const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
63       const uint8x8_t va1 = vld1_u8(a1); a1 += 8;
64       const int16x8_t vxa1 = vreinterpretq_s16_u16(vmovl_u8(va1));
65 
66       const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
67       const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
68 
69       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
70       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
71       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
72       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
73       const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
74       const int16x8_t vxb89ABCDEFc0 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc0, vb_zero_point));
75 
76       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
77       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
78       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
79       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
80       const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
81       const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
82 
83       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
84       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
85       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
86       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
87       const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
88       const int16x8_t vxb89ABCDEFc1 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc1, vb_zero_point));
89 
90       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
91       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
92       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
93       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
94       const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
95       const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
96 
97       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
98       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
99       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
100       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
101       const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
102       const int16x8_t vxb89ABCDEFc2 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc2, vb_zero_point));
103 
104       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
105       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
106       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
107       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
108       const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
109       const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
110 
111       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
112       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
113       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
114       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
115       const uint8x8_t vb89ABCDEFc3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
116       const int16x8_t vxb89ABCDEFc3 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc3, vb_zero_point));
117 
118       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
119       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
120       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
121       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
122 
123 
124       const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
125       const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
126 
127       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
128       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
129       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
130       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
131       const uint8x8_t vb89ABCDEFc4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
132       const int16x8_t vxb89ABCDEFc4 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc4, vb_zero_point));
133 
134       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
135       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
136       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
137       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
138       const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
139       const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
140 
141       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
142       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
143       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
144       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
145       const uint8x8_t vb89ABCDEFc5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
146       const int16x8_t vxb89ABCDEFc5 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc5, vb_zero_point));
147 
148       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
149       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
150       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
151       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
152       const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
153       const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
154 
155       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
156       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
157       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
158       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
159       const uint8x8_t vb89ABCDEFc6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
160       const int16x8_t vxb89ABCDEFc6 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc6, vb_zero_point));
161 
162       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
163       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
164       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
165       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
166       const uint8x8_t vb01234567c7 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
167       const int16x8_t vxb01234567c7 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c7, vb_zero_point));
168 
169       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
170       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3);
171       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
172       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3);
173       const uint8x8_t vb89ABCDEFc7 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
174       const int16x8_t vxb89ABCDEFc7 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc7, vb_zero_point));
175 
176       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
177       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa0), 3);
178       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
179       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc7), vget_high_s16(vxa1), 3);
180 
181       k -= 8 * sizeof(uint8_t);
182     }
183     if XNN_UNLIKELY(k != 0) {
184       const uint8x8_t va0 = vld1_u8(a0); a0 = (const uint8_t*) ((uintptr_t) a0 + k);
185       const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
186       const uint8x8_t va1 = vld1_u8(a1); a1 = (const uint8_t*) ((uintptr_t) a1 + k);
187       const int16x8_t vxa1 = vreinterpretq_s16_u16(vmovl_u8(va1));
188 
189       const uint8x8_t vb01234567c0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
190       const int16x8_t vxb01234567c0 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c0, vb_zero_point));
191       const uint8x8_t vb89ABCDEFc0 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
192       const int16x8_t vxb89ABCDEFc0 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc0, vb_zero_point));
193 
194       vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
195       vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0);
196       vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
197       vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa0), 0);
198       vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
199       vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0);
200       vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
201       vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc0), vget_low_s16(vxa1), 0);
202 
203       if (k >= 2 * sizeof(uint8_t)) {
204         const uint8x8_t vb01234567c1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
205         const int16x8_t vxb01234567c1 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c1, vb_zero_point));
206         const uint8x8_t vb89ABCDEFc1 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
207         const int16x8_t vxb89ABCDEFc1 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc1, vb_zero_point));
208 
209         vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
210         vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1);
211         vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
212         vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa0), 1);
213         vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
214         vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1);
215         vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
216         vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc1), vget_low_s16(vxa1), 1);
217 
218         if (k > 2 * sizeof(uint8_t)) {
219           const uint8x8_t vb01234567c2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
220           const int16x8_t vxb01234567c2 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c2, vb_zero_point));
221           const uint8x8_t vb89ABCDEFc2 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
222           const int16x8_t vxb89ABCDEFc2 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc2, vb_zero_point));
223 
224           vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
225           vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2);
226           vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
227           vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa0), 2);
228           vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
229           vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2);
230           vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
231           vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc2), vget_low_s16(vxa1), 2);
232 
233           if (k >= 4 * sizeof(uint8_t)) {
234             const uint8x8_t vb01234567c3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
235             const int16x8_t vxb01234567c3 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c3, vb_zero_point));
236             const uint8x8_t vb89ABCDEFc3 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
237             const int16x8_t vxb89ABCDEFc3 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc3, vb_zero_point));
238 
239             vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
240             vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3);
241             vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
242             vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa0), 3);
243             vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
244             vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3);
245             vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
246             vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc3), vget_low_s16(vxa1), 3);
247 
248             if (k > 4 * sizeof(uint8_t)) {
249               const uint8x8_t vb01234567c4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
250               const int16x8_t vxb01234567c4 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c4, vb_zero_point));
251               const uint8x8_t vb89ABCDEFc4 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
252               const int16x8_t vxb89ABCDEFc4 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc4, vb_zero_point));
253 
254               vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
255               vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0);
256               vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
257               vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa0), 0);
258               vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
259               vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0);
260               vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
261               vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc4), vget_high_s16(vxa1), 0);
262 
263               if (k >= 6 * sizeof(uint8_t)) {
264                 const uint8x8_t vb01234567c5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
265                 const int16x8_t vxb01234567c5 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c5, vb_zero_point));
266                 const uint8x8_t vb89ABCDEFc5 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
267                 const int16x8_t vxb89ABCDEFc5 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc5, vb_zero_point));
268 
269                 vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
270                 vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1);
271                 vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
272                 vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa0), 1);
273                 vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
274                 vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1);
275                 vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
276                 vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc5), vget_high_s16(vxa1), 1);
277 
278                 if (k > 6 * sizeof(uint8_t)) {
279                   const uint8x8_t vb01234567c6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
280                   const int16x8_t vxb01234567c6 = vreinterpretq_s16_u16(vsubl_u8(vb01234567c6, vb_zero_point));
281                   const uint8x8_t vb89ABCDEFc6 = vld1_u8(w); w = (const void*) ((const uint8_t*) w + 8);
282                   const int16x8_t vxb89ABCDEFc6 = vreinterpretq_s16_u16(vsubl_u8(vb89ABCDEFc6, vb_zero_point));
283 
284                   vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
285                   vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2);
286                   vacc0x89AB = vmlal_lane_s16(vacc0x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
287                   vacc0xCDEF = vmlal_lane_s16(vacc0xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa0), 2);
288                   vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
289                   vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2);
290                   vacc1x89AB = vmlal_lane_s16(vacc1x89AB, vget_low_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
291                   vacc1xCDEF = vmlal_lane_s16(vacc1xCDEF, vget_high_s16(vxb89ABCDEFc6), vget_high_s16(vxa1), 2);
292                 }
293               }
294             }
295           }
296         }
297       }
298     }
299 
300     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
301     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
302     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
303 
304     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
305     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
306     vacc0x89AB = vqshlq_s32(vacc0x89AB, vright_pre_shift);
307     vacc0xCDEF = vqshlq_s32(vacc0xCDEF, vright_pre_shift);
308     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
309     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
310     vacc1x89AB = vqshlq_s32(vacc1x89AB, vright_pre_shift);
311     vacc1xCDEF = vqshlq_s32(vacc1xCDEF, vright_pre_shift);
312 
313     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
314     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
315     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
316     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
317     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
318     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
319     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
320     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
321 
322     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
323     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
324     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
325     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
326     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
327     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
328     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
329     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
330 
331     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
332 #if XNN_ARCH_ARM64
333     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
334     int16x8_t vacc0x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF);
335     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
336     int16x8_t vacc1x89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF);
337 
338     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
339     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
340     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
341     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
342 
343     uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
344     uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
345 #else
346     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
347     int16x8_t vacc0x89ABCDEF = vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF));
348     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
349     int16x8_t vacc1x89ABCDEF = vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF));
350 
351     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
352     vacc0x89ABCDEF = vqaddq_s16(vacc0x89ABCDEF, voutput_zero_point);
353     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
354     vacc1x89ABCDEF = vqaddq_s16(vacc1x89ABCDEF, voutput_zero_point);
355 
356     uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
357     uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
358 #endif
359 
360     const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
361     vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
362     vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
363 
364     const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
365     vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
366     vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
367 
368     if (nc >= 16) {
369       vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
370       vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
371 
372       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
373       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
374 
375       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
376       a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
377 
378       nc -= 16;
379     } else {
380       uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
381       if (nc & 8) {
382         vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
383         vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
384         vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
385       }
386       if (nc & 4) {
387         vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
388         vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
389         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
390       }
391       if (nc & 2) {
392         vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
393         vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
394         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
395       }
396       if (nc & 1) {
397         vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
398         vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
399       }
400 
401       nc = 0;
402     }
403   } while (nc != 0);
404 }
405