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/math.h>
16
17
xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__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)])18 void xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const uint8_t** restrict a,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
30 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 5);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (5 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(uint8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 4 * sizeof(uint8_t));
44 uint8_t* c0 = c;
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr < 2) {
47 c1 = c0;
48 }
49 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 c2 = c1;
52 }
53 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
54 if XNN_UNPREDICTABLE(mr < 4) {
55 c3 = c2;
56 }
57 uint8_t* c4 = (uint8_t*) ((uintptr_t) c3 + cm_stride);
58 if XNN_UNPREDICTABLE(mr <= 4) {
59 c4 = c3;
60 }
61
62 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
63
64 do {
65 // Initialize accumulators with bias. 8 bias values are loaded from the
66 // weight matrix, at the start of the group of 8 columns.
67 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
68 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
69 uint32x4_t vpacc1x0123 = vpacc0x0123;
70 uint32x4_t vpacc1x4567 = vpacc0x4567;
71 uint32x4_t vpacc2x0123 = vpacc0x0123;
72 uint32x4_t vpacc2x4567 = vpacc0x4567;
73 uint32x4_t vpacc3x0123 = vpacc0x0123;
74 uint32x4_t vpacc3x4567 = vpacc0x4567;
75 uint32x4_t vpacc4x0123 = vpacc0x0123;
76 uint32x4_t vpacc4x4567 = vpacc0x4567;
77 uint32x2_t vnacc0 = vmov_n_u32(0);
78 uint32x2_t vnacc1 = vmov_n_u32(0);
79 uint32x2_t vnacc2 = vmov_n_u32(0);
80 uint32x2_t vnacc3 = vmov_n_u32(0);
81 uint32x2_t vnacc4 = vmov_n_u32(0);
82
83 size_t p = ks;
84 do {
85 const uint8_t* restrict a0 = a[0];
86 if XNN_UNPREDICTABLE(a0 != zero) {
87 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
88 }
89 const uint8_t* restrict a1 = a[1];
90 if XNN_UNPREDICTABLE(a1 != zero) {
91 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
92 }
93 const uint8_t* restrict a2 = a[2];
94 if XNN_UNPREDICTABLE(a2 != zero) {
95 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
96 }
97 const uint8_t* restrict a3 = a[3];
98 if XNN_UNPREDICTABLE(a3 != zero) {
99 a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
100 }
101 const uint8_t* restrict a4 = a[4];
102 if XNN_UNPREDICTABLE(a4 != zero) {
103 a4 = (const uint8_t*) ((uintptr_t) a4 + a_offset);
104 }
105 a += 5;
106
107 // Inner accumulation loop along the 8 columns.
108 size_t k = kc;
109 // 2x partial unrolled loop to load 8 bytes at a time.
110 while (k >= 8 * sizeof(uint8_t)) {
111 // Load a 5x8 block of activations.
112 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
113 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
114 const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
115 const uint8x8_t va3x01234567 = vld1_u8(a3); a3 += 8;
116 const uint8x8_t va4x01234567 = vld1_u8(a4); a4 += 8;
117
118 // Load a 8x8 block of weights.
119 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
120 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
121 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
122 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
123
124 // Multiply-accumulate: 5x8 * 8x8 --> 5x8.
125 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
126 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
127 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
128 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
129 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, 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 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
134 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
135 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
136 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
137 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
138 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
139 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
140 vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
141 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
142 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
143 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb4567x0123, va3x01234567, 1);
144 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb4567x4567, va3x01234567, 1);
145 vnacc4 = vdot_u32(vnacc4, va_zero_point, va4x01234567);
146 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
147 vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
148 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb4567x0123, va4x01234567, 1);
149 vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb4567x4567, va4x01234567, 1);
150
151 k -= 8 * sizeof(uint8_t);
152 }
153 // Handle up to 4 final positions of `k`
154 if XNN_UNLIKELY(k != 0) {
155 // Load a 5x4 block of activations.
156 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
157 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
158 const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
159 const uint8x8_t va3x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a3, vmov_n_u32(0), 0)); a3 += 4;
160 const uint8x8_t va4x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a4, vmov_n_u32(0), 0)); a4 += 4;
161
162 // Load a 4x8 block of weights.
163 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
164 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
165
166 // Multiply-accumulate: 5x4 * 4x8 --> 5x8.
167 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
168 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
169 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
170 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
171 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
172 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
173 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
174 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
175 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
176 vnacc3 = vdot_u32(vnacc3, va_zero_point, va3x01234567);
177 vpacc3x0123 = vdotq_lane_u32(vpacc3x0123, vb0123x0123, va3x01234567, 0);
178 vpacc3x4567 = vdotq_lane_u32(vpacc3x4567, vb0123x4567, va3x01234567, 0);
179 vnacc4 = vdot_u32(vnacc4, va_zero_point, va4x01234567);
180 vpacc4x0123 = vdotq_lane_u32(vpacc4x0123, vb0123x0123, va4x01234567, 0);
181 vpacc4x4567 = vdotq_lane_u32(vpacc4x4567, vb0123x4567, va4x01234567, 0);
182 }
183 p -= 5 * sizeof(void*);
184 } while (p != 0);
185
186 // Subtract zero point from accumulators.
187 vnacc0 = vpadd_u32(vnacc0, vnacc0);
188 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
189 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
190 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
191 vnacc1 = vpadd_u32(vnacc1, vnacc1);
192 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
193 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
194 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
195 vnacc2 = vpadd_u32(vnacc2, vnacc2);
196 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
197 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
198 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
199 vnacc3 = vpadd_u32(vnacc3, vnacc3);
200 const uint32x4_t vnacc3x0123 = vcombine_u32(vnacc3, vnacc3);
201 int32x4_t vacc3x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x0123, vnacc3x0123));
202 int32x4_t vacc3x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc3x4567, vnacc3x0123));
203 vnacc4 = vpadd_u32(vnacc4, vnacc4);
204 const uint32x4_t vnacc4x0123 = vcombine_u32(vnacc4, vnacc4);
205 int32x4_t vacc4x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x0123, vnacc4x0123));
206 int32x4_t vacc4x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc4x4567, vnacc4x0123));
207
208 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
209 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
210 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
211
212 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
213 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
214 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
215 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
216 vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
217 vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
218 vacc3x0123 = vshlq_s32(vacc3x0123, vright_pre_shift);
219 vacc3x4567 = vshlq_s32(vacc3x4567, vright_pre_shift);
220 vacc4x0123 = vshlq_s32(vacc4x0123, vright_pre_shift);
221 vacc4x4567 = vshlq_s32(vacc4x4567, vright_pre_shift);
222
223 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
224 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
225 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
226 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
227 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
228 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
229 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
230 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
231 vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
232 vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
233
234 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
235 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
236 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
237 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
238 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
239 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
240 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
241 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
242 vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
243 vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
244
245 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
246 #if XNN_ARCH_ARM64
247 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
248 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
249 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
250 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
251 const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
252
253 uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
254 uint8x16_t vout2x01234567_3x01234567 = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc3x01234567);
255 uint8x8_t vout4x01234567 = vqmovun_s16(vacc4x01234567);
256 #else
257 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
258 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
259 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
260 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
261 const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
262
263 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
264 uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc3x01234567));
265 uint8x8_t vout4x01234567 = vqmovun_s16(vacc4x01234567);
266 #endif
267 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
268 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
269
270 vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
271 vout2x01234567_3x01234567 = vmaxq_u8(vout2x01234567_3x01234567, voutput_min);
272 vout4x01234567 = vmax_u8(vout4x01234567, vget_low_u8(voutput_min));
273
274 vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
275 vout2x01234567_3x01234567 = vminq_u8(vout2x01234567_3x01234567, voutput_max);
276 vout4x01234567 = vmin_u8(vout4x01234567, vget_low_u8(voutput_max));
277
278 if (nc >= 8) {
279 vst1_u8(c4 + 0, vout4x01234567);
280 vst1_u8(c3 + 0, vget_high_u8(vout2x01234567_3x01234567));
281 vst1_u8(c2 + 0, vget_low_u8(vout2x01234567_3x01234567));
282 vst1_u8(c1 + 0, vget_high_u8(vout0x01234567_1x01234567));
283 vst1_u8(c0 + 0, vget_low_u8(vout0x01234567_1x01234567));
284
285 c4 = (uint8_t*) ((uintptr_t) c4 + cn_stride);
286 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
287 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
288 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
289 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
290
291 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
292
293 nc -= 8;
294 } else {
295 if (nc & 4) {
296 vst1_lane_u32((void*) c4, vreinterpret_u32_u8(vout4x01234567), 0); c4 += 4;
297 vst1q_lane_u32((void*) c3, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
298 vst1q_lane_u32((void*) c2, vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
299 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
300 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
301 vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 4);
302 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
303 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
304 }
305 if (nc & 2) {
306 vst1_lane_u16((void*) c4, vreinterpret_u16_u8(vout4x01234567), 0); c4 += 2;
307 vst1q_lane_u16((void*) c3, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
308 vst1q_lane_u16((void*) c2, vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
309 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
310 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
311 vout4x01234567 = vext_u8(vout4x01234567, vout4x01234567, 2);
312 vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
313 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
314 }
315 if (nc & 1) {
316 vst1_lane_u8(c4, vout4x01234567, 0);
317 vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
318 vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
319 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
320 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
321 }
322
323 nc = 0;
324 }
325 } while (nc != 0);
326 }
327