xref: /aosp_15_r20/external/XNNPACK/src/qu8-gemm/gen/8x16c4-minmax-rndnu-neondot.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qu8-gemm/c4-neondot.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/gemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qu8_gemm_minmax_rndnu_ukernel_8x16c4__neondot(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_8x16c4__neondot(
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 <= 8);
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   kc = round_up_po2(kc, 4 * sizeof(uint8_t));
40   const uint8_t* a0 = a;
41   uint8_t* c0 = c;
42   const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
43   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr < 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48   const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
49   uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     a2 = a1;
52     c2 = c1;
53   }
54   const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
55   uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
56   if XNN_UNPREDICTABLE(mr < 4) {
57     a3 = a2;
58     c3 = c2;
59   }
60   const uint8_t* a4 = (const uint8_t*) ((uintptr_t) a3 + a_stride);
61   uint8_t* c4 = (uint8_t*) ((uintptr_t) c3 + cm_stride);
62   if XNN_UNPREDICTABLE(mr <= 4) {
63     a4 = a3;
64     c4 = c3;
65   }
66   const uint8_t* a5 = (const uint8_t*) ((uintptr_t) a4 + a_stride);
67   uint8_t* c5 = (uint8_t*) ((uintptr_t) c4 + cm_stride);
68   if XNN_UNPREDICTABLE(mr < 6) {
69     a5 = a4;
70     c5 = c4;
71   }
72   const uint8_t* a6 = (const uint8_t*) ((uintptr_t) a5 + a_stride);
73   uint8_t* c6 = (uint8_t*) ((uintptr_t) c5 + cm_stride);
74   if XNN_UNPREDICTABLE(mr <= 6) {
75     a6 = a5;
76     c6 = c5;
77   }
78   const uint8_t* a7 = (const uint8_t*) ((uintptr_t) a6 + a_stride);
79   uint8_t* c7 = (uint8_t*) ((uintptr_t) c6 + cm_stride);
80   if XNN_UNPREDICTABLE(mr != 8) {
81     a7 = a6;
82     c7 = c6;
83   }
84 
85   const uint8x8_t va_zero_point = vld1_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
86 
87   // Loop over groups of 16 columns.
88   do {
89     // Initialize accumulators with bias. 16 bias values are loaded from the
90     // weight matrix, at the start of the group of 16 columns.
91     uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
92     uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
93     uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
94     uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
95     uint32x4_t vpacc1x0123 = vpacc0x0123;
96     uint32x4_t vpacc1x4567 = vpacc0x4567;
97     uint32x4_t vpacc1x89AB = vpacc0x89AB;
98     uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
99     uint32x4_t vpacc2x0123 = vpacc0x0123;
100     uint32x4_t vpacc2x4567 = vpacc0x4567;
101     uint32x4_t vpacc2x89AB = vpacc0x89AB;
102     uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
103     uint32x4_t vpacc3x0123 = vpacc0x0123;
104     uint32x4_t vpacc3x4567 = vpacc0x4567;
105     uint32x4_t vpacc3x89AB = vpacc0x89AB;
106     uint32x4_t vpacc3xCDEF = vpacc0xCDEF;
107     uint32x4_t vpacc4x0123 = vpacc0x0123;
108     uint32x4_t vpacc4x4567 = vpacc0x4567;
109     uint32x4_t vpacc4x89AB = vpacc0x89AB;
110     uint32x4_t vpacc4xCDEF = vpacc0xCDEF;
111     uint32x4_t vpacc5x0123 = vpacc0x0123;
112     uint32x4_t vpacc5x4567 = vpacc0x4567;
113     uint32x4_t vpacc5x89AB = vpacc0x89AB;
114     uint32x4_t vpacc5xCDEF = vpacc0xCDEF;
115     uint32x4_t vpacc6x0123 = vpacc0x0123;
116     uint32x4_t vpacc6x4567 = vpacc0x4567;
117     uint32x4_t vpacc6x89AB = vpacc0x89AB;
118     uint32x4_t vpacc6xCDEF = vpacc0xCDEF;
119     uint32x4_t vpacc7x0123 = vpacc0x0123;
120     uint32x4_t vpacc7x4567 = vpacc0x4567;
121     uint32x4_t vpacc7x89AB = vpacc0x89AB;
122     uint32x4_t vpacc7xCDEF = vpacc0xCDEF;
123     uint32x2_t vnacc0 = vmov_n_u32(0);
124     uint32x2_t vnacc1 = vmov_n_u32(0);
125     uint32x2_t vnacc2 = vmov_n_u32(0);
126     uint32x2_t vnacc3 = vmov_n_u32(0);
127     uint32x2_t vnacc4 = vmov_n_u32(0);
128     uint32x2_t vnacc5 = vmov_n_u32(0);
129     uint32x2_t vnacc6 = vmov_n_u32(0);
130     uint32x2_t vnacc7 = vmov_n_u32(0);
131 
132     // Inner accumulation loop along the 16 columns.
133     size_t k = kc;
134     // 2x partial unrolled loop to load 8 bytes at a time.
135     while (k >= 8 * sizeof(uint8_t)) {
136       // Load a 8x8 block of activations.
137       const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
138       const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
139       const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
140       const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
141       const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 8;
142       const uint8x8_t va5x01234567 = vld1_u8(a5); a5 += 8;
143       const uint8x8_t va6x01234567 = vld1_u8(a6); a6 += 8;
144       const uint8x8_t va7x01234567 = vld1_u8(a7); a7 += 8;
145 
146       // Load a 8x16 block of weights.
147       const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
148       const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
149       const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
150       const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
151       const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
152       const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
153       const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
154       const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
155 
156       // Multiply-accumulate: 8x8 * 8x16 --> 8x16.
157       vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
158       vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
159       vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
160       vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
161       vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
162       vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
163       vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
164       vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
165       vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
166       vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
167       vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
168       vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
169       vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
170       vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
171       vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
172       vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
173       vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
174       vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
175       vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
176       vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
177       vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
178       vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
179       vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
180       vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
181       vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
182       vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
183       vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
184       vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
185       vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
186       vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
187       vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
188       vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
189       vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
190       vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
191       vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb4567x89AB, va3x01234567, 1);
192       vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
193       vnacc4 = vdot_u32(vnacc4, va_zero_point, va4x01234567);
194       vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
195       vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
196       vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb0123x89AB, va4x01234567, 0);
197       vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
198       vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1);
199       vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb4567x4567, va4x01234567, 1);
200       vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb4567x89AB, va4x01234567, 1);
201       vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
202       vnacc5 = vdot_u32(vnacc5, va_zero_point, va5x01234567);
203       vpacc5x0123 = vdotq_lane_u32(vpacc5x0123, vb0123x0123, va5x01234567, 0);
204       vpacc5x4567 = vdotq_lane_u32(vpacc5x4567, vb0123x4567, va5x01234567, 0);
205       vpacc5x89AB = vdotq_lane_u32(vpacc5x89AB, vb0123x89AB, va5x01234567, 0);
206       vpacc5xCDEF = vdotq_lane_u32(vpacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
207       vpacc5x0123 = vdotq_lane_u32(vpacc5x0123, vb4567x0123, va5x01234567, 1);
208       vpacc5x4567 = vdotq_lane_u32(vpacc5x4567, vb4567x4567, va5x01234567, 1);
209       vpacc5x89AB = vdotq_lane_u32(vpacc5x89AB, vb4567x89AB, va5x01234567, 1);
210       vpacc5xCDEF = vdotq_lane_u32(vpacc5xCDEF, vb4567xCDEF, va5x01234567, 1);
211       vnacc6 = vdot_u32(vnacc6, va_zero_point, va6x01234567);
212       vpacc6x0123 = vdotq_lane_u32(vpacc6x0123, vb0123x0123, va6x01234567, 0);
213       vpacc6x4567 = vdotq_lane_u32(vpacc6x4567, vb0123x4567, va6x01234567, 0);
214       vpacc6x89AB = vdotq_lane_u32(vpacc6x89AB, vb0123x89AB, va6x01234567, 0);
215       vpacc6xCDEF = vdotq_lane_u32(vpacc6xCDEF, vb0123xCDEF, va6x01234567, 0);
216       vpacc6x0123 = vdotq_lane_u32(vpacc6x0123, vb4567x0123, va6x01234567, 1);
217       vpacc6x4567 = vdotq_lane_u32(vpacc6x4567, vb4567x4567, va6x01234567, 1);
218       vpacc6x89AB = vdotq_lane_u32(vpacc6x89AB, vb4567x89AB, va6x01234567, 1);
219       vpacc6xCDEF = vdotq_lane_u32(vpacc6xCDEF, vb4567xCDEF, va6x01234567, 1);
220       vnacc7 = vdot_u32(vnacc7, va_zero_point, va7x01234567);
221       vpacc7x0123 = vdotq_lane_u32(vpacc7x0123, vb0123x0123, va7x01234567, 0);
222       vpacc7x4567 = vdotq_lane_u32(vpacc7x4567, vb0123x4567, va7x01234567, 0);
223       vpacc7x89AB = vdotq_lane_u32(vpacc7x89AB, vb0123x89AB, va7x01234567, 0);
224       vpacc7xCDEF = vdotq_lane_u32(vpacc7xCDEF, vb0123xCDEF, va7x01234567, 0);
225       vpacc7x0123 = vdotq_lane_u32(vpacc7x0123, vb4567x0123, va7x01234567, 1);
226       vpacc7x4567 = vdotq_lane_u32(vpacc7x4567, vb4567x4567, va7x01234567, 1);
227       vpacc7x89AB = vdotq_lane_u32(vpacc7x89AB, vb4567x89AB, va7x01234567, 1);
228       vpacc7xCDEF = vdotq_lane_u32(vpacc7xCDEF, vb4567xCDEF, va7x01234567, 1);
229 
230       k -= 8 * sizeof(uint8_t);
231     }
232     // Handle up to 4 final positions of `k`
233     if XNN_UNLIKELY(k != 0) {
234       // Load a 8x4 block of activations.
235       const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
236       const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
237       const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
238       const uint8x8_t va3x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a3, vmov_n_u32(0), 0)); a3 += 4;
239       const uint8x8_t va4x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a4, vmov_n_u32(0), 0)); a4 += 4;
240       const uint8x8_t va5x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a5, vmov_n_u32(0), 0)); a5 += 4;
241       const uint8x8_t va6x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a6, vmov_n_u32(0), 0)); a6 += 4;
242       const uint8x8_t va7x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a7, vmov_n_u32(0), 0)); a7 += 4;
243 
244       // Load a 4x16 block of weights.
245       const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
246       const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
247       const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
248       const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
249 
250       // Multiply-accumulate: 8x4 * 4x16 --> 8x16.
251       vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
252       vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
253       vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
254       vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
255       vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
256       vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
257       vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
258       vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
259       vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
260       vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
261       vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
262       vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
263       vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
264       vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
265       vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
266       vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
267       vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
268       vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
269       vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
270       vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
271       vnacc4 = vdot_u32(vnacc4, va_zero_point, va4x01234567);
272       vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
273       vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
274       vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb0123x89AB, va4x01234567, 0);
275       vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
276       vnacc5 = vdot_u32(vnacc5, va_zero_point, va5x01234567);
277       vpacc5x0123 = vdotq_lane_u32(vpacc5x0123, vb0123x0123, va5x01234567, 0);
278       vpacc5x4567 = vdotq_lane_u32(vpacc5x4567, vb0123x4567, va5x01234567, 0);
279       vpacc5x89AB = vdotq_lane_u32(vpacc5x89AB, vb0123x89AB, va5x01234567, 0);
280       vpacc5xCDEF = vdotq_lane_u32(vpacc5xCDEF, vb0123xCDEF, va5x01234567, 0);
281       vnacc6 = vdot_u32(vnacc6, va_zero_point, va6x01234567);
282       vpacc6x0123 = vdotq_lane_u32(vpacc6x0123, vb0123x0123, va6x01234567, 0);
283       vpacc6x4567 = vdotq_lane_u32(vpacc6x4567, vb0123x4567, va6x01234567, 0);
284       vpacc6x89AB = vdotq_lane_u32(vpacc6x89AB, vb0123x89AB, va6x01234567, 0);
285       vpacc6xCDEF = vdotq_lane_u32(vpacc6xCDEF, vb0123xCDEF, va6x01234567, 0);
286       vnacc7 = vdot_u32(vnacc7, va_zero_point, va7x01234567);
287       vpacc7x0123 = vdotq_lane_u32(vpacc7x0123, vb0123x0123, va7x01234567, 0);
288       vpacc7x4567 = vdotq_lane_u32(vpacc7x4567, vb0123x4567, va7x01234567, 0);
289       vpacc7x89AB = vdotq_lane_u32(vpacc7x89AB, vb0123x89AB, va7x01234567, 0);
290       vpacc7xCDEF = vdotq_lane_u32(vpacc7xCDEF, vb0123xCDEF, va7x01234567, 0);
291     }
292 
293     // Subtract zero point from accumulators.
294     vnacc0 = vpadd_u32(vnacc0, vnacc0);
295     const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
296     int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
297     int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
298     int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
299     int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
300     vnacc1 = vpadd_u32(vnacc1, vnacc1);
301     const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
302     int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
303     int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
304     int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
305     int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
306     vnacc2 = vpadd_u32(vnacc2, vnacc2);
307     const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
308     int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
309     int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
310     int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123));
311     int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123));
312     vnacc3 = vpadd_u32(vnacc3, vnacc3);
313     const uint32x4_t vnacc3x0123 = vcombine_u32(vnacc3, vnacc3);
314     int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
315     int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x0123));
316     int32x4_t vacc3x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc3x89AB, vnacc3x0123));
317     int32x4_t vacc3xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc3xCDEF, vnacc3x0123));
318     vnacc4 = vpadd_u32(vnacc4, vnacc4);
319     const uint32x4_t vnacc4x0123 = vcombine_u32(vnacc4, vnacc4);
320     int32x4_t vacc4x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x0123, vnacc4x0123));
321     int32x4_t vacc4x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x4567, vnacc4x0123));
322     int32x4_t vacc4x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc4x89AB, vnacc4x0123));
323     int32x4_t vacc4xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc4xCDEF, vnacc4x0123));
324     vnacc5 = vpadd_u32(vnacc5, vnacc5);
325     const uint32x4_t vnacc5x0123 = vcombine_u32(vnacc5, vnacc5);
326     int32x4_t vacc5x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc5x0123, vnacc5x0123));
327     int32x4_t vacc5x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc5x4567, vnacc5x0123));
328     int32x4_t vacc5x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc5x89AB, vnacc5x0123));
329     int32x4_t vacc5xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc5xCDEF, vnacc5x0123));
330     vnacc6 = vpadd_u32(vnacc6, vnacc6);
331     const uint32x4_t vnacc6x0123 = vcombine_u32(vnacc6, vnacc6);
332     int32x4_t vacc6x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc6x0123, vnacc6x0123));
333     int32x4_t vacc6x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc6x4567, vnacc6x0123));
334     int32x4_t vacc6x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc6x89AB, vnacc6x0123));
335     int32x4_t vacc6xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc6xCDEF, vnacc6x0123));
336     vnacc7 = vpadd_u32(vnacc7, vnacc7);
337     const uint32x4_t vnacc7x0123 = vcombine_u32(vnacc7, vnacc7);
338     int32x4_t vacc7x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc7x0123, vnacc7x0123));
339     int32x4_t vacc7x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc7x4567, vnacc7x0123));
340     int32x4_t vacc7x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc7x89AB, vnacc7x0123));
341     int32x4_t vacc7xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc7xCDEF, vnacc7x0123));
342 
343     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
344     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
345     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
346 
347     vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
348     vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
349     vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
350     vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
351     vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
352     vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
353     vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
354     vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
355     vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
356     vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
357     vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
358     vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
359     vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
360     vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
361     vacc3x89AB = vshlq_s32(vacc3x89AB, vright_pre_shift);
362     vacc3xCDEF = vshlq_s32(vacc3xCDEF, vright_pre_shift);
363     vacc4x0123 = vshlq_s32(vacc4x0123, vright_pre_shift);
364     vacc4x4567 = vshlq_s32(vacc4x4567, vright_pre_shift);
365     vacc4x89AB = vshlq_s32(vacc4x89AB, vright_pre_shift);
366     vacc4xCDEF = vshlq_s32(vacc4xCDEF, vright_pre_shift);
367     vacc5x0123 = vshlq_s32(vacc5x0123, vright_pre_shift);
368     vacc5x4567 = vshlq_s32(vacc5x4567, vright_pre_shift);
369     vacc5x89AB = vshlq_s32(vacc5x89AB, vright_pre_shift);
370     vacc5xCDEF = vshlq_s32(vacc5xCDEF, vright_pre_shift);
371     vacc6x0123 = vshlq_s32(vacc6x0123, vright_pre_shift);
372     vacc6x4567 = vshlq_s32(vacc6x4567, vright_pre_shift);
373     vacc6x89AB = vshlq_s32(vacc6x89AB, vright_pre_shift);
374     vacc6xCDEF = vshlq_s32(vacc6xCDEF, vright_pre_shift);
375     vacc7x0123 = vshlq_s32(vacc7x0123, vright_pre_shift);
376     vacc7x4567 = vshlq_s32(vacc7x4567, vright_pre_shift);
377     vacc7x89AB = vshlq_s32(vacc7x89AB, vright_pre_shift);
378     vacc7xCDEF = vshlq_s32(vacc7xCDEF, vright_pre_shift);
379 
380     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
381     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
382     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
383     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
384     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
385     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
386     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
387     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
388     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
389     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
390     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
391     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
392     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
393     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
394     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
395     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
396     vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
397     vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
398     vacc4x89AB = vqdmulhq_s32(vacc4x89AB, vmultiplier);
399     vacc4xCDEF = vqdmulhq_s32(vacc4xCDEF, vmultiplier);
400     vacc5x0123 = vqdmulhq_s32(vacc5x0123, vmultiplier);
401     vacc5x4567 = vqdmulhq_s32(vacc5x4567, vmultiplier);
402     vacc5x89AB = vqdmulhq_s32(vacc5x89AB, vmultiplier);
403     vacc5xCDEF = vqdmulhq_s32(vacc5xCDEF, vmultiplier);
404     vacc6x0123 = vqdmulhq_s32(vacc6x0123, vmultiplier);
405     vacc6x4567 = vqdmulhq_s32(vacc6x4567, vmultiplier);
406     vacc6x89AB = vqdmulhq_s32(vacc6x89AB, vmultiplier);
407     vacc6xCDEF = vqdmulhq_s32(vacc6xCDEF, vmultiplier);
408     vacc7x0123 = vqdmulhq_s32(vacc7x0123, vmultiplier);
409     vacc7x4567 = vqdmulhq_s32(vacc7x4567, vmultiplier);
410     vacc7x89AB = vqdmulhq_s32(vacc7x89AB, vmultiplier);
411     vacc7xCDEF = vqdmulhq_s32(vacc7xCDEF, vmultiplier);
412 
413     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
414     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
415     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
416     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
417     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
418     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
419     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
420     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
421     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
422     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
423     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
424     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
425     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
426     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
427     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
428     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
429     vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
430     vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
431     vacc4x89AB = vrshlq_s32(vacc4x89AB, vright_post_shift);
432     vacc4xCDEF = vrshlq_s32(vacc4xCDEF, vright_post_shift);
433     vacc5x0123 = vrshlq_s32(vacc5x0123, vright_post_shift);
434     vacc5x4567 = vrshlq_s32(vacc5x4567, vright_post_shift);
435     vacc5x89AB = vrshlq_s32(vacc5x89AB, vright_post_shift);
436     vacc5xCDEF = vrshlq_s32(vacc5xCDEF, vright_post_shift);
437     vacc6x0123 = vrshlq_s32(vacc6x0123, vright_post_shift);
438     vacc6x4567 = vrshlq_s32(vacc6x4567, vright_post_shift);
439     vacc6x89AB = vrshlq_s32(vacc6x89AB, vright_post_shift);
440     vacc6xCDEF = vrshlq_s32(vacc6xCDEF, vright_post_shift);
441     vacc7x0123 = vrshlq_s32(vacc7x0123, vright_post_shift);
442     vacc7x4567 = vrshlq_s32(vacc7x4567, vright_post_shift);
443     vacc7x89AB = vrshlq_s32(vacc7x89AB, vright_post_shift);
444     vacc7xCDEF = vrshlq_s32(vacc7xCDEF, vright_post_shift);
445 
446     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
447 #if XNN_ARCH_ARM64
448     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
449     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
450     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
451     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
452     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
453     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
454     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
455     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
456     const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
457     const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), voutput_zero_point);
458     const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
459     const int16x8_t vacc5x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x89AB), vacc5xCDEF), voutput_zero_point);
460     const int16x8_t vacc6x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x0123), vacc6x4567), voutput_zero_point);
461     const int16x8_t vacc6x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x89AB), vacc6xCDEF), voutput_zero_point);
462     const int16x8_t vacc7x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x0123), vacc7x4567), voutput_zero_point);
463     const int16x8_t vacc7x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x89AB), vacc7xCDEF), voutput_zero_point);
464 
465     uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
466     uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
467     uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
468     uint8x16_t vout3x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc3x01234567), vacc3x89ABCDEF);
469     uint8x16_t vout4x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc4x01234567), vacc4x89ABCDEF);
470     uint8x16_t vout5x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc5x01234567), vacc5x89ABCDEF);
471     uint8x16_t vout6x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc6x01234567), vacc6x89ABCDEF);
472     uint8x16_t vout7x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc7x01234567), vacc7x89ABCDEF);
473 #else
474     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
475     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
476     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
477     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
478     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
479     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
480     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
481     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
482     const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
483     const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF)), voutput_zero_point);
484     const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
485     const int16x8_t vacc5x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x89AB), vqmovn_s32(vacc5xCDEF)), voutput_zero_point);
486     const int16x8_t vacc6x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x0123), vqmovn_s32(vacc6x4567)), voutput_zero_point);
487     const int16x8_t vacc6x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x89AB), vqmovn_s32(vacc6xCDEF)), voutput_zero_point);
488     const int16x8_t vacc7x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc7x0123), vqmovn_s32(vacc7x4567)), voutput_zero_point);
489     const int16x8_t vacc7x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc7x89AB), vqmovn_s32(vacc7xCDEF)), voutput_zero_point);
490 
491     uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
492     uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
493     uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
494     uint8x16_t vout3x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc3x01234567), vqmovun_s16(vacc3x89ABCDEF));
495     uint8x16_t vout4x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc4x01234567), vqmovun_s16(vacc4x89ABCDEF));
496     uint8x16_t vout5x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc5x01234567), vqmovun_s16(vacc5x89ABCDEF));
497     uint8x16_t vout6x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc6x01234567), vqmovun_s16(vacc6x89ABCDEF));
498     uint8x16_t vout7x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc7x01234567), vqmovun_s16(vacc7x89ABCDEF));
499 #endif
500     const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
501     const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
502 
503     vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
504     vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
505     vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
506     vout3x0123456789ABCDEF = vmaxq_u8(vout3x0123456789ABCDEF, voutput_min);
507     vout4x0123456789ABCDEF = vmaxq_u8(vout4x0123456789ABCDEF, voutput_min);
508     vout5x0123456789ABCDEF = vmaxq_u8(vout5x0123456789ABCDEF, voutput_min);
509     vout6x0123456789ABCDEF = vmaxq_u8(vout6x0123456789ABCDEF, voutput_min);
510     vout7x0123456789ABCDEF = vmaxq_u8(vout7x0123456789ABCDEF, voutput_min);
511 
512     vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
513     vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
514     vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
515     vout3x0123456789ABCDEF = vminq_u8(vout3x0123456789ABCDEF, voutput_max);
516     vout4x0123456789ABCDEF = vminq_u8(vout4x0123456789ABCDEF, voutput_max);
517     vout5x0123456789ABCDEF = vminq_u8(vout5x0123456789ABCDEF, voutput_max);
518     vout6x0123456789ABCDEF = vminq_u8(vout6x0123456789ABCDEF, voutput_max);
519     vout7x0123456789ABCDEF = vminq_u8(vout7x0123456789ABCDEF, voutput_max);
520 
521     if (nc >= 16) {
522       vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
523       vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
524       vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
525       vst1q_u8(c3 + 0, vout3x0123456789ABCDEF);
526       vst1q_u8(c4 + 0, vout4x0123456789ABCDEF);
527       vst1q_u8(c5 + 0, vout5x0123456789ABCDEF);
528       vst1q_u8(c6 + 0, vout6x0123456789ABCDEF);
529       vst1q_u8(c7 + 0, vout7x0123456789ABCDEF);
530 
531       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
532       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
533       c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
534       c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
535       c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
536       c5 = (uint8_t*) ((uintptr_t) c5 + cn_stride);
537       c6 = (uint8_t*) ((uintptr_t) c6 + cn_stride);
538       c7 = (uint8_t*) ((uintptr_t) c7 + cn_stride);
539 
540       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
541       a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
542       a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
543       a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
544       a4 = (const uint8_t*) ((uintptr_t) a4 - kc);
545       a5 = (const uint8_t*) ((uintptr_t) a5 - kc);
546       a6 = (const uint8_t*) ((uintptr_t) a6 - kc);
547       a7 = (const uint8_t*) ((uintptr_t) a7 - kc);
548 
549       nc -= 16;
550     } else {
551       uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
552       uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vget_low_u8(vout2x0123456789ABCDEF), vget_low_u8(vout3x0123456789ABCDEF));
553       uint8x16_t vout4x01234567_5x01234567 = vcombine_u8(vget_low_u8(vout4x0123456789ABCDEF), vget_low_u8(vout5x0123456789ABCDEF));
554       uint8x16_t vout6x01234567_7x01234567 = vcombine_u8(vget_low_u8(vout6x0123456789ABCDEF), vget_low_u8(vout7x0123456789ABCDEF));
555       if (nc & 8) {
556         vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
557         vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
558         vst1_u8(c2, vget_low_u8(vout2x01234567_3x01234567)); c2 += 8;
559         vst1_u8(c3, vget_high_u8(vout2x01234567_3x01234567)); c3 += 8;
560         vst1_u8(c4, vget_low_u8(vout4x01234567_5x01234567)); c4 += 8;
561         vst1_u8(c5, vget_high_u8(vout4x01234567_5x01234567)); c5 += 8;
562         vst1_u8(c6, vget_low_u8(vout6x01234567_7x01234567)); c6 += 8;
563         vst1_u8(c7, vget_high_u8(vout6x01234567_7x01234567)); c7 += 8;
564         vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
565         vout2x01234567_3x01234567 = vcombine_u8(vget_high_u8(vout2x0123456789ABCDEF), vget_high_u8(vout3x0123456789ABCDEF));
566         vout4x01234567_5x01234567 = vcombine_u8(vget_high_u8(vout4x0123456789ABCDEF), vget_high_u8(vout5x0123456789ABCDEF));
567         vout6x01234567_7x01234567 = vcombine_u8(vget_high_u8(vout6x0123456789ABCDEF), vget_high_u8(vout7x0123456789ABCDEF));
568       }
569       if (nc & 4) {
570         vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
571         vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
572         vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
573         vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
574         vst1q_lane_u32((void*) c4, vreinterpretq_u32_u8(vout4x01234567_5x01234567), 0); c4 += 4;
575         vst1q_lane_u32((void*) c5, vreinterpretq_u32_u8(vout4x01234567_5x01234567), 2); c5 += 4;
576         vst1q_lane_u32((void*) c6, vreinterpretq_u32_u8(vout6x01234567_7x01234567), 0); c6 += 4;
577         vst1q_lane_u32((void*) c7, vreinterpretq_u32_u8(vout6x01234567_7x01234567), 2); c7 += 4;
578         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
579         vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
580         vout4x01234567_5x01234567 = vextq_u8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
581         vout6x01234567_7x01234567 = vextq_u8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 4);
582       }
583       if (nc & 2) {
584         vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
585         vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
586         vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
587         vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
588         vst1q_lane_u16((void*) c4, vreinterpretq_u16_u8(vout4x01234567_5x01234567), 0); c4 += 2;
589         vst1q_lane_u16((void*) c5, vreinterpretq_u16_u8(vout4x01234567_5x01234567), 4); c5 += 2;
590         vst1q_lane_u16((void*) c6, vreinterpretq_u16_u8(vout6x01234567_7x01234567), 0); c6 += 2;
591         vst1q_lane_u16((void*) c7, vreinterpretq_u16_u8(vout6x01234567_7x01234567), 4); c7 += 2;
592         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
593         vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
594         vout4x01234567_5x01234567 = vextq_u8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
595         vout6x01234567_7x01234567 = vextq_u8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 2);
596       }
597       if (nc & 1) {
598         vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
599         vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
600         vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
601         vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
602         vst1q_lane_u8(c4, vout4x01234567_5x01234567, 0);
603         vst1q_lane_u8(c5, vout4x01234567_5x01234567, 8);
604         vst1q_lane_u8(c6, vout6x01234567_7x01234567, 0);
605         vst1q_lane_u8(c7, vout6x01234567_7x01234567, 8);
606       }
607 
608       nc = 0;
609     }
610   } while (nc != 0);
611 }
612