xref: /aosp_15_r20/external/XNNPACK/src/qu8-gemm/gen/5x16c4-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_5x16c4__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_5x16c4__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 <= 5);
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 
67   const uint8x8_t va_zero_point = vld1_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
68 
69   // Loop over groups of 16 columns.
70   do {
71     // Initialize accumulators with bias. 16 bias values are loaded from the
72     // weight matrix, at the start of the group of 16 columns.
73     uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
74     uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
75     uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
76     uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
77     uint32x4_t vpacc1x0123 = vpacc0x0123;
78     uint32x4_t vpacc1x4567 = vpacc0x4567;
79     uint32x4_t vpacc1x89AB = vpacc0x89AB;
80     uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
81     uint32x4_t vpacc2x0123 = vpacc0x0123;
82     uint32x4_t vpacc2x4567 = vpacc0x4567;
83     uint32x4_t vpacc2x89AB = vpacc0x89AB;
84     uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
85     uint32x4_t vpacc3x0123 = vpacc0x0123;
86     uint32x4_t vpacc3x4567 = vpacc0x4567;
87     uint32x4_t vpacc3x89AB = vpacc0x89AB;
88     uint32x4_t vpacc3xCDEF = vpacc0xCDEF;
89     uint32x4_t vpacc4x0123 = vpacc0x0123;
90     uint32x4_t vpacc4x4567 = vpacc0x4567;
91     uint32x4_t vpacc4x89AB = vpacc0x89AB;
92     uint32x4_t vpacc4xCDEF = vpacc0xCDEF;
93     uint32x2_t vnacc0 = vmov_n_u32(0);
94     uint32x2_t vnacc1 = vmov_n_u32(0);
95     uint32x2_t vnacc2 = vmov_n_u32(0);
96     uint32x2_t vnacc3 = vmov_n_u32(0);
97     uint32x2_t vnacc4 = vmov_n_u32(0);
98 
99     // Inner accumulation loop along the 16 columns.
100     size_t k = kc;
101     // 2x partial unrolled loop to load 8 bytes at a time.
102     while (k >= 8 * sizeof(uint8_t)) {
103       // Load a 5x8 block of activations.
104       const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
105       const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
106       const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
107       const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
108       const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 8;
109 
110       // Load a 8x16 block of weights.
111       const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
112       const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
113       const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
114       const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
115       const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
116       const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
117       const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
118       const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
119 
120       // Multiply-accumulate: 5x8 * 8x16 --> 5x16.
121       vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
122       vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
123       vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
124       vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
125       vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
126       vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
127       vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
128       vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
129       vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
130       vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
131       vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
132       vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
133       vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
134       vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
135       vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
136       vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
137       vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
138       vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
139       vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
140       vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
141       vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
142       vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
143       vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
144       vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
145       vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
146       vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
147       vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
148       vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
149       vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
150       vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
151       vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
152       vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
153       vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
154       vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
155       vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb4567x89AB, va3x01234567, 1);
156       vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
157       vnacc4 = vdot_u32(vnacc4, va_zero_point, va4x01234567);
158       vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
159       vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
160       vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb0123x89AB, va4x01234567, 0);
161       vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
162       vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1);
163       vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb4567x4567, va4x01234567, 1);
164       vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb4567x89AB, va4x01234567, 1);
165       vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb4567xCDEF, va4x01234567, 1);
166 
167       k -= 8 * sizeof(uint8_t);
168     }
169     // Handle up to 4 final positions of `k`
170     if XNN_UNLIKELY(k != 0) {
171       // Load a 5x4 block of activations.
172       const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
173       const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
174       const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
175       const uint8x8_t va3x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a3, vmov_n_u32(0), 0)); a3 += 4;
176       const uint8x8_t va4x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a4, vmov_n_u32(0), 0)); a4 += 4;
177 
178       // Load a 4x16 block of weights.
179       const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
180       const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
181       const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
182       const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
183 
184       // Multiply-accumulate: 5x4 * 4x16 --> 5x16.
185       vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
186       vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
187       vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
188       vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
189       vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
190       vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
191       vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
192       vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
193       vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
194       vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
195       vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
196       vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
197       vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
198       vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
199       vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
200       vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
201       vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
202       vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
203       vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
204       vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
205       vnacc4 = vdot_u32(vnacc4, va_zero_point, va4x01234567);
206       vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
207       vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
208       vpacc4x89AB = vdotq_lane_u32(vpacc4x89AB, vb0123x89AB, va4x01234567, 0);
209       vpacc4xCDEF = vdotq_lane_u32(vpacc4xCDEF, vb0123xCDEF, va4x01234567, 0);
210     }
211 
212     // Subtract zero point from accumulators.
213     vnacc0 = vpadd_u32(vnacc0, vnacc0);
214     const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
215     int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
216     int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
217     int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
218     int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
219     vnacc1 = vpadd_u32(vnacc1, vnacc1);
220     const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
221     int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
222     int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
223     int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
224     int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
225     vnacc2 = vpadd_u32(vnacc2, vnacc2);
226     const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
227     int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
228     int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
229     int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123));
230     int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123));
231     vnacc3 = vpadd_u32(vnacc3, vnacc3);
232     const uint32x4_t vnacc3x0123 = vcombine_u32(vnacc3, vnacc3);
233     int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
234     int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x0123));
235     int32x4_t vacc3x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc3x89AB, vnacc3x0123));
236     int32x4_t vacc3xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc3xCDEF, vnacc3x0123));
237     vnacc4 = vpadd_u32(vnacc4, vnacc4);
238     const uint32x4_t vnacc4x0123 = vcombine_u32(vnacc4, vnacc4);
239     int32x4_t vacc4x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x0123, vnacc4x0123));
240     int32x4_t vacc4x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x4567, vnacc4x0123));
241     int32x4_t vacc4x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc4x89AB, vnacc4x0123));
242     int32x4_t vacc4xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc4xCDEF, vnacc4x0123));
243 
244     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
245     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
246     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
247 
248     vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
249     vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
250     vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
251     vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
252     vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
253     vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
254     vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
255     vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
256     vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
257     vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
258     vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
259     vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
260     vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
261     vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
262     vacc3x89AB = vshlq_s32(vacc3x89AB, vright_pre_shift);
263     vacc3xCDEF = vshlq_s32(vacc3xCDEF, vright_pre_shift);
264     vacc4x0123 = vshlq_s32(vacc4x0123, vright_pre_shift);
265     vacc4x4567 = vshlq_s32(vacc4x4567, vright_pre_shift);
266     vacc4x89AB = vshlq_s32(vacc4x89AB, vright_pre_shift);
267     vacc4xCDEF = vshlq_s32(vacc4xCDEF, vright_pre_shift);
268 
269     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
270     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
271     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
272     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
273     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
274     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
275     vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
276     vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
277     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
278     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
279     vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
280     vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
281     vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
282     vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
283     vacc3x89AB = vqdmulhq_s32(vacc3x89AB, vmultiplier);
284     vacc3xCDEF = vqdmulhq_s32(vacc3xCDEF, vmultiplier);
285     vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
286     vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
287     vacc4x89AB = vqdmulhq_s32(vacc4x89AB, vmultiplier);
288     vacc4xCDEF = vqdmulhq_s32(vacc4xCDEF, vmultiplier);
289 
290     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
291     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
292     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
293     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
294     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
295     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
296     vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
297     vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
298     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
299     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
300     vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
301     vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
302     vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
303     vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
304     vacc3x89AB = vrshlq_s32(vacc3x89AB, vright_post_shift);
305     vacc3xCDEF = vrshlq_s32(vacc3xCDEF, vright_post_shift);
306     vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
307     vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
308     vacc4x89AB = vrshlq_s32(vacc4x89AB, vright_post_shift);
309     vacc4xCDEF = vrshlq_s32(vacc4xCDEF, vright_post_shift);
310 
311     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
312 #if XNN_ARCH_ARM64
313     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
314     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
315     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
316     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
317     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
318     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
319     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
320     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
321     const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
322     const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x89AB), vacc4xCDEF), voutput_zero_point);
323 
324     uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
325     uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
326     uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
327     uint8x16_t vout3x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc3x01234567), vacc3x89ABCDEF);
328     uint8x16_t vout4x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc4x01234567), vacc4x89ABCDEF);
329 #else
330     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
331     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
332     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
333     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
334     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
335     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
336     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
337     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
338     const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
339     const int16x8_t vacc4x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x89AB), vqmovn_s32(vacc4xCDEF)), voutput_zero_point);
340 
341     uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
342     uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
343     uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
344     uint8x16_t vout3x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc3x01234567), vqmovun_s16(vacc3x89ABCDEF));
345     uint8x16_t vout4x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc4x01234567), vqmovun_s16(vacc4x89ABCDEF));
346 #endif
347     const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
348     const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
349 
350     vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
351     vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
352     vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
353     vout3x0123456789ABCDEF = vmaxq_u8(vout3x0123456789ABCDEF, voutput_min);
354     vout4x0123456789ABCDEF = vmaxq_u8(vout4x0123456789ABCDEF, voutput_min);
355 
356     vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
357     vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
358     vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
359     vout3x0123456789ABCDEF = vminq_u8(vout3x0123456789ABCDEF, voutput_max);
360     vout4x0123456789ABCDEF = vminq_u8(vout4x0123456789ABCDEF, voutput_max);
361 
362     if (nc >= 16) {
363       vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
364       vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
365       vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
366       vst1q_u8(c3 + 0, vout3x0123456789ABCDEF);
367       vst1q_u8(c4 + 0, vout4x0123456789ABCDEF);
368 
369       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
370       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
371       c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
372       c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
373       c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
374 
375       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
376       a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
377       a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
378       a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
379       a4 = (const uint8_t*) ((uintptr_t) a4 - kc);
380 
381       nc -= 16;
382     } else {
383       uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
384       uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vget_low_u8(vout2x0123456789ABCDEF), vget_low_u8(vout3x0123456789ABCDEF));
385       uint8x8_t vout4x01234567 = vget_low_u8(vout4x0123456789ABCDEF);
386       if (nc & 8) {
387         vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
388         vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
389         vst1_u8(c2, vget_low_u8(vout2x01234567_3x01234567)); c2 += 8;
390         vst1_u8(c3, vget_high_u8(vout2x01234567_3x01234567)); c3 += 8;
391         vst1_u8(c4, vout4x01234567); c4 += 8;
392         vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
393         vout2x01234567_3x01234567 = vcombine_u8(vget_high_u8(vout2x0123456789ABCDEF), vget_high_u8(vout3x0123456789ABCDEF));
394         vout4x01234567 = vget_high_u8(vout4x0123456789ABCDEF);
395       }
396       if (nc & 4) {
397         vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
398         vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
399         vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
400         vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
401         vst1_lane_u32((void*) c4, vreinterpret_u32_u8(vout4x01234567), 0); c4 += 4;
402         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
403         vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
404         vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 4);
405       }
406       if (nc & 2) {
407         vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
408         vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
409         vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
410         vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
411         vst1_lane_u16((void*) c4, vreinterpret_u16_u8(vout4x01234567), 0); c4 += 2;
412         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
413         vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
414         vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 2);
415       }
416       if (nc & 1) {
417         vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
418         vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
419         vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
420         vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
421         vst1_lane_u8(c4, vout4x01234567, 0);
422       }
423 
424       nc = 0;
425     }
426   } while (nc != 0);
427 }
428