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_3x16c4__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_3x16c4__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 <= 3);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (3 * 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
54 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->rndnu_neon.kernel_zero_point[0]);
55
56 do {
57 // Initialize accumulators with bias. 16 bias values are loaded from the
58 // weight matrix, at the start of the group of 16 columns.
59 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
60 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
61 uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
62 uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
63 uint32x4_t vpacc1x0123 = vpacc0x0123;
64 uint32x4_t vpacc1x4567 = vpacc0x4567;
65 uint32x4_t vpacc1x89AB = vpacc0x89AB;
66 uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
67 uint32x4_t vpacc2x0123 = vpacc0x0123;
68 uint32x4_t vpacc2x4567 = vpacc0x4567;
69 uint32x4_t vpacc2x89AB = vpacc0x89AB;
70 uint32x4_t vpacc2xCDEF = vpacc0xCDEF;
71 uint32x2_t vnacc0 = vmov_n_u32(0);
72 uint32x2_t vnacc1 = vmov_n_u32(0);
73 uint32x2_t vnacc2 = vmov_n_u32(0);
74
75 size_t p = ks;
76 do {
77 const uint8_t* restrict a0 = a[0];
78 if XNN_UNPREDICTABLE(a0 != zero) {
79 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
80 }
81 const uint8_t* restrict a1 = a[1];
82 if XNN_UNPREDICTABLE(a1 != zero) {
83 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
84 }
85 const uint8_t* restrict a2 = a[2];
86 if XNN_UNPREDICTABLE(a2 != zero) {
87 a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
88 }
89 a += 3;
90
91 // Inner accumulation loop along the 16 columns.
92 size_t k = kc;
93 // 2x partial unrolled loop to load 8 bytes at a time.
94 while (k >= 8 * sizeof(uint8_t)) {
95 // Load a 3x8 block of activations.
96 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
97 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
98 const uint8x8_t va2x01234567 = vld1_u8(a2); a2 += 8;
99
100 // Load a 8x16 block of weights.
101 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
102 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
103 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
104 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
105 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
106 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
107 const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
108 const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
109
110 // Multiply-accumulate: 3x8 * 8x16 --> 3x16.
111 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
112 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
113 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
114 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
115 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
116 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
117 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
118 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
119 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
120 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
121 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
122 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
123 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
124 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
125 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
126 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
127 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
128 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
129 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
130 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
131 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
132 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
133 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
134 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb4567x0123, va2x01234567, 1);
135 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb4567x4567, va2x01234567, 1);
136 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb4567x89AB, va2x01234567, 1);
137 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
138
139 k -= 8 * sizeof(uint8_t);
140 }
141 // Handle up to 4 final positions of `k`
142 if XNN_UNLIKELY(k != 0) {
143 // Load a 3x4 block of activations.
144 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
145 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
146 const uint8x8_t va2x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a2, vmov_n_u32(0), 0)); a2 += 4;
147
148 // Load a 4x16 block of weights.
149 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
150 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
151 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
152 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
153
154 // Multiply-accumulate: 3x4 * 4x16 --> 3x16.
155 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
156 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
157 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
158 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
159 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
160 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
161 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
162 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
163 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
164 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
165 vnacc2 = vdot_u32(vnacc2, va_zero_point, va2x01234567);
166 vpacc2x0123 = vdotq_lane_u32(vpacc2x0123, vb0123x0123, va2x01234567, 0);
167 vpacc2x4567 = vdotq_lane_u32(vpacc2x4567, vb0123x4567, va2x01234567, 0);
168 vpacc2x89AB = vdotq_lane_u32(vpacc2x89AB, vb0123x89AB, va2x01234567, 0);
169 vpacc2xCDEF = vdotq_lane_u32(vpacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
170 }
171 p -= 3 * sizeof(void*);
172 } while (p != 0);
173
174 // Subtract zero point from accumulators.
175 vnacc0 = vpadd_u32(vnacc0, vnacc0);
176 const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
177 int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
178 int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
179 int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
180 int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
181 vnacc1 = vpadd_u32(vnacc1, vnacc1);
182 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
183 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
184 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
185 int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
186 int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
187 vnacc2 = vpadd_u32(vnacc2, vnacc2);
188 const uint32x4_t vnacc2x0123 = vcombine_u32(vnacc2, vnacc2);
189 int32x4_t vacc2x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x0123, vnacc2x0123));
190 int32x4_t vacc2x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc2x4567, vnacc2x0123));
191 int32x4_t vacc2x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc2x89AB, vnacc2x0123));
192 int32x4_t vacc2xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc2xCDEF, vnacc2x0123));
193
194 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
195 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
196 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
197
198 vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
199 vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
200 vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
201 vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
202 vacc1x0123 = vshlq_s32(vacc1x0123, vright_pre_shift);
203 vacc1x4567 = vshlq_s32(vacc1x4567, vright_pre_shift);
204 vacc1x89AB = vshlq_s32(vacc1x89AB, vright_pre_shift);
205 vacc1xCDEF = vshlq_s32(vacc1xCDEF, vright_pre_shift);
206 vacc2x0123 = vshlq_s32(vacc2x0123, vright_pre_shift);
207 vacc2x4567 = vshlq_s32(vacc2x4567, vright_pre_shift);
208 vacc2x89AB = vshlq_s32(vacc2x89AB, vright_pre_shift);
209 vacc2xCDEF = vshlq_s32(vacc2xCDEF, vright_pre_shift);
210
211 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
212 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
213 vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
214 vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
215 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
216 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
217 vacc1x89AB = vqdmulhq_s32(vacc1x89AB, vmultiplier);
218 vacc1xCDEF = vqdmulhq_s32(vacc1xCDEF, vmultiplier);
219 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
220 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
221 vacc2x89AB = vqdmulhq_s32(vacc2x89AB, vmultiplier);
222 vacc2xCDEF = vqdmulhq_s32(vacc2xCDEF, vmultiplier);
223
224 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
225 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
226 vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
227 vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
228 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
229 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
230 vacc1x89AB = vrshlq_s32(vacc1x89AB, vright_post_shift);
231 vacc1xCDEF = vrshlq_s32(vacc1xCDEF, vright_post_shift);
232 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
233 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
234 vacc2x89AB = vrshlq_s32(vacc2x89AB, vright_post_shift);
235 vacc2xCDEF = vrshlq_s32(vacc2xCDEF, vright_post_shift);
236
237 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
238 #if XNN_ARCH_ARM64
239 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
240 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
241 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
242 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
243 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
244 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x89AB), vacc2xCDEF), voutput_zero_point);
245
246 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
247 uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
248 uint8x16_t vout2x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc2x89ABCDEF);
249 #else
250 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
251 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
252 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
253 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
254 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
255 const int16x8_t vacc2x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x89AB), vqmovn_s32(vacc2xCDEF)), voutput_zero_point);
256
257 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
258 uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
259 uint8x16_t vout2x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc2x89ABCDEF));
260 #endif
261 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->rndnu_neon.output_min);
262 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->rndnu_neon.output_max);
263
264 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
265 vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
266 vout2x0123456789ABCDEF = vmaxq_u8(vout2x0123456789ABCDEF, voutput_min);
267
268 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
269 vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
270 vout2x0123456789ABCDEF = vminq_u8(vout2x0123456789ABCDEF, voutput_max);
271
272 if (nc >= 16) {
273 vst1q_u8(c2 + 0, vout2x0123456789ABCDEF);
274 vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
275 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
276
277 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
278 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
279 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
280
281 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
282
283 nc -= 16;
284 } else {
285 uint8x8_t vout2x01234567 = vget_low_u8(vout2x0123456789ABCDEF);
286 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
287 if (nc & 8) {
288 vst1_u8(c2, vout2x01234567); c2 += 8; // This line
289 vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
290 vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
291 vout2x01234567 = vget_high_u8(vout2x0123456789ABCDEF);
292 vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
293 }
294 if (nc & 4) {
295 vst1_lane_u32((void*) c2, vreinterpret_u32_u8(vout2x01234567), 0); c2 += 4;
296 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
297 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
298 vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 4);
299 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
300 }
301 if (nc & 2) {
302 vst1_lane_u16((void*) c2, vreinterpret_u16_u8(vout2x01234567), 0); c2 += 2;
303 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
304 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
305 vout2x01234567 = vext_u8(vout2x01234567, vout2x01234567, 2);
306 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
307 }
308 if (nc & 1) {
309 vst1_lane_u8(c2, vout2x01234567, 0);
310 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
311 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
312 }
313
314 nc = 0;
315 }
316 } while (nc != 0);
317 }
318