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_3x16c4__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_3x16c4__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 <= 3);
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
55 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
56
57 // Loop over groups of 16 columns.
58 do {
59 // Initialize accumulators with bias. 16 bias values are loaded from the
60 // weight matrix, at the start of the group of 16 columns.
61 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
62 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
63 uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
64 uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
65 uint32x4_t vpacc1x0123 = vpacc0x0123;
66 uint32x4_t vpacc1x4567 = vpacc0x4567;
67 uint32x4_t vpacc1x89AB = vpacc0x89AB;
68 uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
69 uint32x4_t vpacc2x0123 = vpacc0x0123;
70 uint32x4_t vpacc2x4567 = vpacc0x4567;
71 uint32x4_t vpacc2x89AB = vpacc0x89AB;
72 uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
73 uint32x2_t vnacc0 = vmov_n_u32(0);
74 uint32x2_t vnacc1 = vmov_n_u32(0);
75 uint32x2_t vnacc2 = vmov_n_u32(0);
76
77 // Inner accumulation loop along the 16 columns.
78 size_t k = kc;
79 // 2x partial unrolled loop to load 8 bytes at a time.
80 while (k >= 8 * sizeof(uint8_t)) {
81 // Load a 3x8 block of activations.
82 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
83 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
84 const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
85
86 // Load a 8x16 block of weights.
87 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
88 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
89 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
90 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
91 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
92 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
93 const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
94 const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
95
96 // Multiply-accumulate: 3x8 * 8x16 --> 3x16.
97 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
98 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
99 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
100 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
101 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
102 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
103 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
104 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
105 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
106 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
107 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
108 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
109 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
110 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
111 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
112 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
113 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
114 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
115 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
116 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
117 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
118 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
119 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
120 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
121 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
122 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
123 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
124
125 k -= 8 * sizeof(uint8_t);
126 }
127 // Handle up to 4 final positions of `k`
128 if XNN_UNLIKELY(k != 0) {
129 // Load a 3x4 block of activations.
130 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
131 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
132 const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
133
134 // Load a 4x16 block of weights.
135 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
136 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
137 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
138 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
139
140 // Multiply-accumulate: 3x4 * 4x16 --> 3x16.
141 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
142 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
143 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
144 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
145 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
146 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
147 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
148 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
149 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
150 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
151 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
152 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
153 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
154 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
155 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
156 }
157
158 // Subtract zero point from accumulators.
159 vnacc0 = vpadd_u32(vnacc0, vnacc0);
160 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
161 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
162 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
163 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
164 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
165 vnacc1 = vpadd_u32(vnacc1, vnacc1);
166 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
167 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
168 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
169 int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
170 int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
171 vnacc2 = vpadd_u32(vnacc2, vnacc2);
172 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
173 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
174 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
175 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123));
176 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123));
177
178 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
179 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
180 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
181
182 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
183 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
184 vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
185 vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
186 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
187 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
188 vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
189 vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
190 vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
191 vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
192 vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
193 vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
194
195 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
196 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
197 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
198 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
199 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
200 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
201 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
202 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
203 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
204 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
205 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
206 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
207
208 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
209 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
210 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
211 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
212 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
213 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
214 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
215 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
216 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
217 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
218 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
219 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
220
221 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
222 #if XNN_ARCH_ARM64
223 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
224 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
225 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
226 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
227 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
228 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
229
230 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
231 uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
232 uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
233 #else
234 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
235 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
236 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
237 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
238 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
239 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
240
241 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
242 uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
243 uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
244 #endif
245 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
246 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
247
248 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
249 vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
250 vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
251
252 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
253 vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
254 vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
255
256 if (nc >= 16) {
257 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
258 vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
259 vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
260
261 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
262 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
263 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
264
265 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
266 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
267 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
268
269 nc -= 16;
270 } else {
271 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
272 uint8x8_t vout2x01234567 = vget_low_u8(vout2x0123456789ABCDEF);
273 if (nc & 8) {
274 vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
275 vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
276 vst1_u8(c2, vout2x01234567); c2 += 8;
277 vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
278 vout2x01234567 = vget_high_u8(vout2x0123456789ABCDEF);
279 }
280 if (nc & 4) {
281 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
282 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
283 vst1_lane_u32((void*) c2, vreinterpret_u32_u8(vout2x01234567), 0); c2 += 4;
284 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
285 vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 4);
286 }
287 if (nc & 2) {
288 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
289 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
290 vst1_lane_u16((void*) c2, vreinterpret_u16_u8(vout2x01234567), 0); c2 += 2;
291 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
292 vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 2);
293 }
294 if (nc & 1) {
295 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
296 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
297 vst1_lane_u8(c2, vout2x01234567, 0);
298 }
299
300 nc = 0;
301 }
302 } while (nc != 0);
303 }
304