xref: /aosp_15_r20/external/XNNPACK/src/qu8-igemm/gen/4x16c4-minmax-fp32-neondot.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Auto-generated file. Do not edit!
2 //   Template: src/qu8-igemm/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/igemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 
18 
xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     size_t ks,
24     const uint8_t** restrict a,
25     const void* restrict w,
26     uint8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     size_t a_offset,
30     const uint8_t* zero,
31     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33   assert(mr != 0);
34   assert(mr <= 4);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(ks != 0);
38   assert(ks % (4 * sizeof(void*)) == 0);
39   assert(a_offset % sizeof(uint8_t) == 0);
40   assert(a != NULL);
41   assert(w != NULL);
42   assert(c != NULL);
43 
44   kc = round_up_po2(kc, 4 * sizeof(uint8_t));
45   uint8_t* c0 = c;
46   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
47   if XNN_UNPREDICTABLE(mr < 2) {
48     c1 = c0;
49   }
50   uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
51   if XNN_UNPREDICTABLE(mr <= 2) {
52     c2 = c1;
53   }
54   uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
55   if XNN_UNPREDICTABLE(mr != 4) {
56     c3 = c2;
57   }
58 
59   const uint8x8_t va_zero_point = vld1_dup_u8(&params->fp32_neonv8.kernel_zero_point[0]);
60 
61   do {
62     // Initialize accumulators with bias. 16 bias values are loaded from the
63     // weight matrix, at the start of the group of 16 columns.
64     uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
65     uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
66     uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
67     uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
68     uint32x4_t vpacc1x0123 = vpacc0x0123;
69     uint32x4_t vpacc1x4567 = vpacc0x4567;
70     uint32x4_t vpacc1x89AB = vpacc0x89AB;
71     uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
72     uint32x4_t vpacc2x0123 = vpacc0x0123;
73     uint32x4_t vpacc2x4567 = vpacc0x4567;
74     uint32x4_t vpacc2x89AB = vpacc0x89AB;
75     uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
76     uint32x4_t vpacc3x0123 = vpacc0x0123;
77     uint32x4_t vpacc3x4567 = vpacc0x4567;
78     uint32x4_t vpacc3x89AB = vpacc0x89AB;
79     uint32x4_t vpacc3xCDEF = vpacc0xCDEF;
80     uint32x2_t vnacc0 = vmov_n_u32(0);
81     uint32x2_t vnacc1 = vmov_n_u32(0);
82     uint32x2_t vnacc2 = vmov_n_u32(0);
83     uint32x2_t vnacc3 = vmov_n_u32(0);
84 
85     size_t p = ks;
86     do {
87       const uint8_t* restrict a0 = a[0];
88       if XNN_UNPREDICTABLE(a0 != zero) {
89         a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
90       }
91       const uint8_t* restrict a1 = a[1];
92       if XNN_UNPREDICTABLE(a1 != zero) {
93         a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
94       }
95       const uint8_t* restrict a2 = a[2];
96       if XNN_UNPREDICTABLE(a2 != zero) {
97         a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
98       }
99       const uint8_t* restrict a3 = a[3];
100       if XNN_UNPREDICTABLE(a3 != zero) {
101         a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
102       }
103       a += 4;
104 
105       // Inner accumulation loop along the 16 columns.
106       size_t k = kc;
107       // 2x partial unrolled loop to load 8 bytes at a time.
108       while (k >= 8 * sizeof(uint8_t)) {
109         // Load a 4x8 block of activations.
110         const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
111         const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
112         const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
113         const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
114 
115         // Load a 8x16 block of weights.
116         const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
117         const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
118         const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
119         const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
120         const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
121         const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
122         const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
123         const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
124 
125         // Multiply-accumulate: 4x8 * 8x16 --> 4x16.
126         vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
127         vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
128         vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
129         vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
130         vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
131         vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
132         vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
133         vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
134         vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
135         vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
136         vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
137         vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
138         vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
139         vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
140         vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
141         vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
142         vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
143         vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
144         vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
145         vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
146         vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
147         vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
148         vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
149         vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
150         vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
151         vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
152         vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
153         vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
154         vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
155         vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
156         vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
157         vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
158         vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
159         vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
160         vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb4567x89AB, va3x01234567, 1);
161         vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb4567xCDEF, va3x01234567, 1);
162 
163         k -= 8 * sizeof(uint8_t);
164       }
165       // Handle up to 4 final positions of `k`
166       if XNN_UNLIKELY(k != 0) {
167         // Load a 4x4 block of activations.
168         const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
169         const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
170         const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
171         const uint8x8_t va3x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a3, vmov_n_u32(0), 0)); a3 += 4;
172 
173         // Load a 4x16 block of weights.
174         const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
175         const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
176         const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
177         const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
178 
179         // Multiply-accumulate: 4x4 * 4x16 --> 4x16.
180         vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
181         vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
182         vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
183         vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
184         vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
185         vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
186         vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
187         vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
188         vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
189         vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
190         vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
191         vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
192         vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
193         vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
194         vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
195         vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
196         vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
197         vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
198         vpacc3x89AB = vdotq_lane_u32(vpacc3x89AB, vb0123x89AB, va3x01234567, 0);
199         vpacc3xCDEF = vdotq_lane_u32(vpacc3xCDEF, vb0123xCDEF, va3x01234567, 0);
200       }
201       p -= 4 * sizeof(void*);
202     } while (p != 0);
203 
204     // Subtract zero point from accumulators.
205     vnacc0 = vpadd_u32(vnacc0, vnacc0);
206     const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
207     int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
208     int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
209     int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
210     int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
211     vnacc1 = vpadd_u32(vnacc1, vnacc1);
212     const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
213     int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
214     int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
215     int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
216     int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
217     vnacc2 = vpadd_u32(vnacc2, vnacc2);
218     const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
219     int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
220     int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
221     int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123));
222     int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123));
223     vnacc3 = vpadd_u32(vnacc3, vnacc3);
224     const uint32x4_t vnacc3x0123 = vcombine_u32(vnacc3, vnacc3);
225     int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
226     int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x0123));
227     int32x4_t vacc3x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc3x89AB, vnacc3x0123));
228     int32x4_t vacc3xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc3xCDEF, vnacc3x0123));
229 
230     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
231     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
232     float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
233     float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
234     float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
235     float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
236     float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB);
237     float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
238     float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
239     float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
240     float32x4_t vfpacc2x89AB = vcvtq_f32_s32(vacc2x89AB);
241     float32x4_t vfpacc2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
242     float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
243     float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
244     float32x4_t vfpacc3x89AB = vcvtq_f32_s32(vacc3x89AB);
245     float32x4_t vfpacc3xCDEF = vcvtq_f32_s32(vacc3xCDEF);
246 
247     const float32x4_t vscale = vld1q_dup_f32(&params->fp32_neonv8.scale);
248     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
249     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
250     vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale);
251     vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscale);
252     vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale);
253     vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale);
254     vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale);
255     vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscale);
256     vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale);
257     vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale);
258     vfpacc2x89AB = vmulq_f32(vfpacc2x89AB, vscale);
259     vfpacc2xCDEF = vmulq_f32(vfpacc2xCDEF, vscale);
260     vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale);
261     vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale);
262     vfpacc3x89AB = vmulq_f32(vfpacc3x89AB, vscale);
263     vfpacc3xCDEF = vmulq_f32(vfpacc3xCDEF, vscale);
264 
265     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
266     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
267     vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
268     vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
269     vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
270     vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
271     vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB);
272     vacc1xCDEF = vcvtnq_s32_f32(vfpacc1xCDEF);
273     vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
274     vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
275     vacc2x89AB = vcvtnq_s32_f32(vfpacc2x89AB);
276     vacc2xCDEF = vcvtnq_s32_f32(vfpacc2xCDEF);
277     vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
278     vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
279     vacc3x89AB = vcvtnq_s32_f32(vfpacc3x89AB);
280     vacc3xCDEF = vcvtnq_s32_f32(vfpacc3xCDEF);
281 
282     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
283 #if XNN_ARCH_ARM64
284     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
285     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
286     const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
287     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
288     const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
289     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
290     const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
291     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x89AB), vacc3xCDEF), voutput_zero_point);
292 
293     uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
294     uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
295     uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
296     uint8x16_t vout3x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc3x01234567), vacc3x89ABCDEF);
297 #else
298     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
299     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
300     const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
301     const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
302     const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
303     const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
304     const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
305     const int16x8_t vacc3x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x89AB), vqmovn_s32(vacc3xCDEF)), voutput_zero_point);
306 
307     uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
308     uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
309     uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
310     uint8x16_t vout3x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc3x01234567), vqmovun_s16(vacc3x89ABCDEF));
311 #endif
312     const uint8x16_t voutput_min = vld1q_dup_u8(&params->fp32_neonv8.output_min);
313     const uint8x16_t voutput_max = vld1q_dup_u8(&params->fp32_neonv8.output_max);
314 
315     vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
316     vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
317     vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
318     vout3x0123456789ABCDEF = vmaxq_u8(vout3x0123456789ABCDEF, voutput_min);
319 
320     vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
321     vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
322     vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
323     vout3x0123456789ABCDEF = vminq_u8(vout3x0123456789ABCDEF, voutput_max);
324 
325     if (nc >= 16) {
326       vst1q_u8(c3 + 0, vout3x0123456789ABCDEF);
327       vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
328       vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
329       vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
330 
331       c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
332       c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
333       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
334       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
335 
336       a = (const uint8_t**restrict) ((uintptr_t) a - ks);
337 
338       nc -= 16;
339     } else {
340       uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vget_low_u8(vout2x0123456789ABCDEF), vget_low_u8(vout3x0123456789ABCDEF));
341       uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
342       if (nc & 8) {
343         vst1_u8(c3, vget_high_u8(vout2x01234567_3x01234567)); c3 += 8;
344         vst1_u8(c2, vget_low_u8(vout2x01234567_3x01234567)); c2 += 8;
345         vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
346         vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
347         vout2x01234567_3x01234567 = vcombine_u8(vget_high_u8(vout2x0123456789ABCDEF), vget_high_u8(vout3x0123456789ABCDEF));
348         vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
349       }
350       if (nc & 4) {
351         vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
352         vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
353         vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
354         vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
355         vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
356         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
357       }
358       if (nc & 2) {
359         vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
360         vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
361         vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
362         vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
363         vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
364         vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
365       }
366       if (nc & 1) {
367         vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
368         vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
369         vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
370         vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
371       }
372 
373       nc = 0;
374     }
375   } while (nc != 0);
376 }
377