2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
13 #include "./vpx_config.h"
14 #include "vpx_dsp/txfm_common.h"
16 static INLINE void TRANSPOSE8X8(int16x8_t *q8s16, int16x8_t *q9s16,
17 int16x8_t *q10s16, int16x8_t *q11s16,
18 int16x8_t *q12s16, int16x8_t *q13s16,
19 int16x8_t *q14s16, int16x8_t *q15s16) {
20 int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
21 int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
22 int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
23 int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
25 d16s16 = vget_low_s16(*q8s16);
26 d17s16 = vget_high_s16(*q8s16);
27 d18s16 = vget_low_s16(*q9s16);
28 d19s16 = vget_high_s16(*q9s16);
29 d20s16 = vget_low_s16(*q10s16);
30 d21s16 = vget_high_s16(*q10s16);
31 d22s16 = vget_low_s16(*q11s16);
32 d23s16 = vget_high_s16(*q11s16);
33 d24s16 = vget_low_s16(*q12s16);
34 d25s16 = vget_high_s16(*q12s16);
35 d26s16 = vget_low_s16(*q13s16);
36 d27s16 = vget_high_s16(*q13s16);
37 d28s16 = vget_low_s16(*q14s16);
38 d29s16 = vget_high_s16(*q14s16);
39 d30s16 = vget_low_s16(*q15s16);
40 d31s16 = vget_high_s16(*q15s16);
42 *q8s16 = vcombine_s16(d16s16, d24s16); // vswp d17, d24
43 *q9s16 = vcombine_s16(d18s16, d26s16); // vswp d19, d26
44 *q10s16 = vcombine_s16(d20s16, d28s16); // vswp d21, d28
45 *q11s16 = vcombine_s16(d22s16, d30s16); // vswp d23, d30
46 *q12s16 = vcombine_s16(d17s16, d25s16);
47 *q13s16 = vcombine_s16(d19s16, d27s16);
48 *q14s16 = vcombine_s16(d21s16, d29s16);
49 *q15s16 = vcombine_s16(d23s16, d31s16);
52 vtrnq_s32(vreinterpretq_s32_s16(*q8s16), vreinterpretq_s32_s16(*q10s16));
54 vtrnq_s32(vreinterpretq_s32_s16(*q9s16), vreinterpretq_s32_s16(*q11s16));
56 vtrnq_s32(vreinterpretq_s32_s16(*q12s16), vreinterpretq_s32_s16(*q14s16));
58 vtrnq_s32(vreinterpretq_s32_s16(*q13s16), vreinterpretq_s32_s16(*q15s16));
60 q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]), // q8
61 vreinterpretq_s16_s32(q1x2s32.val[0])); // q9
62 q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]), // q10
63 vreinterpretq_s16_s32(q1x2s32.val[1])); // q11
64 q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]), // q12
65 vreinterpretq_s16_s32(q3x2s32.val[0])); // q13
66 q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]), // q14
67 vreinterpretq_s16_s32(q3x2s32.val[1])); // q15
69 *q8s16 = q0x2s16.val[0];
70 *q9s16 = q0x2s16.val[1];
71 *q10s16 = q1x2s16.val[0];
72 *q11s16 = q1x2s16.val[1];
73 *q12s16 = q2x2s16.val[0];
74 *q13s16 = q2x2s16.val[1];
75 *q14s16 = q3x2s16.val[0];
76 *q15s16 = q3x2s16.val[1];
80 static INLINE void IDCT8x8_1D(int16x8_t *q8s16, int16x8_t *q9s16,
81 int16x8_t *q10s16, int16x8_t *q11s16,
82 int16x8_t *q12s16, int16x8_t *q13s16,
83 int16x8_t *q14s16, int16x8_t *q15s16) {
84 int16x4_t d0s16, d1s16, d2s16, d3s16;
85 int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
86 int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
87 int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
88 int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
89 int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32;
90 int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
92 d0s16 = vdup_n_s16(cospi_28_64);
93 d1s16 = vdup_n_s16(cospi_4_64);
94 d2s16 = vdup_n_s16(cospi_12_64);
95 d3s16 = vdup_n_s16(cospi_20_64);
97 d16s16 = vget_low_s16(*q8s16);
98 d17s16 = vget_high_s16(*q8s16);
99 d18s16 = vget_low_s16(*q9s16);
100 d19s16 = vget_high_s16(*q9s16);
101 d20s16 = vget_low_s16(*q10s16);
102 d21s16 = vget_high_s16(*q10s16);
103 d22s16 = vget_low_s16(*q11s16);
104 d23s16 = vget_high_s16(*q11s16);
105 d24s16 = vget_low_s16(*q12s16);
106 d25s16 = vget_high_s16(*q12s16);
107 d26s16 = vget_low_s16(*q13s16);
108 d27s16 = vget_high_s16(*q13s16);
109 d28s16 = vget_low_s16(*q14s16);
110 d29s16 = vget_high_s16(*q14s16);
111 d30s16 = vget_low_s16(*q15s16);
112 d31s16 = vget_high_s16(*q15s16);
114 q2s32 = vmull_s16(d18s16, d0s16);
115 q3s32 = vmull_s16(d19s16, d0s16);
116 q5s32 = vmull_s16(d26s16, d2s16);
117 q6s32 = vmull_s16(d27s16, d2s16);
119 q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
120 q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
121 q5s32 = vmlsl_s16(q5s32, d22s16, d3s16);
122 q6s32 = vmlsl_s16(q6s32, d23s16, d3s16);
124 d8s16 = vqrshrn_n_s32(q2s32, 14);
125 d9s16 = vqrshrn_n_s32(q3s32, 14);
126 d10s16 = vqrshrn_n_s32(q5s32, 14);
127 d11s16 = vqrshrn_n_s32(q6s32, 14);
128 q4s16 = vcombine_s16(d8s16, d9s16);
129 q5s16 = vcombine_s16(d10s16, d11s16);
131 q2s32 = vmull_s16(d18s16, d1s16);
132 q3s32 = vmull_s16(d19s16, d1s16);
133 q9s32 = vmull_s16(d26s16, d3s16);
134 q13s32 = vmull_s16(d27s16, d3s16);
136 q2s32 = vmlal_s16(q2s32, d30s16, d0s16);
137 q3s32 = vmlal_s16(q3s32, d31s16, d0s16);
138 q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
139 q13s32 = vmlal_s16(q13s32, d23s16, d2s16);
141 d14s16 = vqrshrn_n_s32(q2s32, 14);
142 d15s16 = vqrshrn_n_s32(q3s32, 14);
143 d12s16 = vqrshrn_n_s32(q9s32, 14);
144 d13s16 = vqrshrn_n_s32(q13s32, 14);
145 q6s16 = vcombine_s16(d12s16, d13s16);
146 q7s16 = vcombine_s16(d14s16, d15s16);
148 d0s16 = vdup_n_s16(cospi_16_64);
150 q2s32 = vmull_s16(d16s16, d0s16);
151 q3s32 = vmull_s16(d17s16, d0s16);
152 q13s32 = vmull_s16(d16s16, d0s16);
153 q15s32 = vmull_s16(d17s16, d0s16);
155 q2s32 = vmlal_s16(q2s32, d24s16, d0s16);
156 q3s32 = vmlal_s16(q3s32, d25s16, d0s16);
157 q13s32 = vmlsl_s16(q13s32, d24s16, d0s16);
158 q15s32 = vmlsl_s16(q15s32, d25s16, d0s16);
160 d0s16 = vdup_n_s16(cospi_24_64);
161 d1s16 = vdup_n_s16(cospi_8_64);
163 d18s16 = vqrshrn_n_s32(q2s32, 14);
164 d19s16 = vqrshrn_n_s32(q3s32, 14);
165 d22s16 = vqrshrn_n_s32(q13s32, 14);
166 d23s16 = vqrshrn_n_s32(q15s32, 14);
167 *q9s16 = vcombine_s16(d18s16, d19s16);
168 *q11s16 = vcombine_s16(d22s16, d23s16);
170 q2s32 = vmull_s16(d20s16, d0s16);
171 q3s32 = vmull_s16(d21s16, d0s16);
172 q8s32 = vmull_s16(d20s16, d1s16);
173 q12s32 = vmull_s16(d21s16, d1s16);
175 q2s32 = vmlsl_s16(q2s32, d28s16, d1s16);
176 q3s32 = vmlsl_s16(q3s32, d29s16, d1s16);
177 q8s32 = vmlal_s16(q8s32, d28s16, d0s16);
178 q12s32 = vmlal_s16(q12s32, d29s16, d0s16);
180 d26s16 = vqrshrn_n_s32(q2s32, 14);
181 d27s16 = vqrshrn_n_s32(q3s32, 14);
182 d30s16 = vqrshrn_n_s32(q8s32, 14);
183 d31s16 = vqrshrn_n_s32(q12s32, 14);
184 *q13s16 = vcombine_s16(d26s16, d27s16);
185 *q15s16 = vcombine_s16(d30s16, d31s16);
187 q0s16 = vaddq_s16(*q9s16, *q15s16);
188 q1s16 = vaddq_s16(*q11s16, *q13s16);
189 q2s16 = vsubq_s16(*q11s16, *q13s16);
190 q3s16 = vsubq_s16(*q9s16, *q15s16);
192 *q13s16 = vsubq_s16(q4s16, q5s16);
193 q4s16 = vaddq_s16(q4s16, q5s16);
194 *q14s16 = vsubq_s16(q7s16, q6s16);
195 q7s16 = vaddq_s16(q7s16, q6s16);
196 d26s16 = vget_low_s16(*q13s16);
197 d27s16 = vget_high_s16(*q13s16);
198 d28s16 = vget_low_s16(*q14s16);
199 d29s16 = vget_high_s16(*q14s16);
201 d16s16 = vdup_n_s16(cospi_16_64);
203 q9s32 = vmull_s16(d28s16, d16s16);
204 q10s32 = vmull_s16(d29s16, d16s16);
205 q11s32 = vmull_s16(d28s16, d16s16);
206 q12s32 = vmull_s16(d29s16, d16s16);
208 q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
209 q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
210 q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
211 q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
213 d10s16 = vqrshrn_n_s32(q9s32, 14);
214 d11s16 = vqrshrn_n_s32(q10s32, 14);
215 d12s16 = vqrshrn_n_s32(q11s32, 14);
216 d13s16 = vqrshrn_n_s32(q12s32, 14);
217 q5s16 = vcombine_s16(d10s16, d11s16);
218 q6s16 = vcombine_s16(d12s16, d13s16);
220 *q8s16 = vaddq_s16(q0s16, q7s16);
221 *q9s16 = vaddq_s16(q1s16, q6s16);
222 *q10s16 = vaddq_s16(q2s16, q5s16);
223 *q11s16 = vaddq_s16(q3s16, q4s16);
224 *q12s16 = vsubq_s16(q3s16, q4s16);
225 *q13s16 = vsubq_s16(q2s16, q5s16);
226 *q14s16 = vsubq_s16(q1s16, q6s16);
227 *q15s16 = vsubq_s16(q0s16, q7s16);
231 void vpx_idct8x8_64_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
233 uint8x8_t d0u8, d1u8, d2u8, d3u8;
234 uint64x1_t d0u64, d1u64, d2u64, d3u64;
235 int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
236 uint16x8_t q8u16, q9u16, q10u16, q11u16;
238 q8s16 = vld1q_s16(input);
239 q9s16 = vld1q_s16(input + 8);
240 q10s16 = vld1q_s16(input + 16);
241 q11s16 = vld1q_s16(input + 24);
242 q12s16 = vld1q_s16(input + 32);
243 q13s16 = vld1q_s16(input + 40);
244 q14s16 = vld1q_s16(input + 48);
245 q15s16 = vld1q_s16(input + 56);
247 TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
250 IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
253 TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
256 IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
259 q8s16 = vrshrq_n_s16(q8s16, 5);
260 q9s16 = vrshrq_n_s16(q9s16, 5);
261 q10s16 = vrshrq_n_s16(q10s16, 5);
262 q11s16 = vrshrq_n_s16(q11s16, 5);
263 q12s16 = vrshrq_n_s16(q12s16, 5);
264 q13s16 = vrshrq_n_s16(q13s16, 5);
265 q14s16 = vrshrq_n_s16(q14s16, 5);
266 q15s16 = vrshrq_n_s16(q15s16, 5);
270 d0u64 = vld1_u64((uint64_t *)d1);
272 d1u64 = vld1_u64((uint64_t *)d1);
274 d2u64 = vld1_u64((uint64_t *)d1);
276 d3u64 = vld1_u64((uint64_t *)d1);
279 q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
280 q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
281 q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
282 q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
284 d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
285 d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
286 d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
287 d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
289 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
291 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
293 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
295 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
303 d0u64 = vld1_u64((uint64_t *)d1);
305 d1u64 = vld1_u64((uint64_t *)d1);
307 d2u64 = vld1_u64((uint64_t *)d1);
309 d3u64 = vld1_u64((uint64_t *)d1);
312 q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
313 q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
314 q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
315 q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
317 d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
318 d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
319 d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
320 d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
322 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
324 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
326 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
328 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
333 void vpx_idct8x8_12_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
335 uint8x8_t d0u8, d1u8, d2u8, d3u8;
336 int16x4_t d10s16, d11s16, d12s16, d13s16, d16s16;
337 int16x4_t d26s16, d27s16, d28s16, d29s16;
338 uint64x1_t d0u64, d1u64, d2u64, d3u64;
339 int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
340 int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
341 uint16x8_t q8u16, q9u16, q10u16, q11u16;
342 int32x4_t q9s32, q10s32, q11s32, q12s32;
344 q8s16 = vld1q_s16(input);
345 q9s16 = vld1q_s16(input + 8);
346 q10s16 = vld1q_s16(input + 16);
347 q11s16 = vld1q_s16(input + 24);
348 q12s16 = vld1q_s16(input + 32);
349 q13s16 = vld1q_s16(input + 40);
350 q14s16 = vld1q_s16(input + 48);
351 q15s16 = vld1q_s16(input + 56);
353 TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
356 // First transform rows
358 q0s16 = vdupq_n_s16(cospi_28_64 * 2);
359 q1s16 = vdupq_n_s16(cospi_4_64 * 2);
361 q4s16 = vqrdmulhq_s16(q9s16, q0s16);
363 q0s16 = vdupq_n_s16(-cospi_20_64 * 2);
365 q7s16 = vqrdmulhq_s16(q9s16, q1s16);
367 q1s16 = vdupq_n_s16(cospi_12_64 * 2);
369 q5s16 = vqrdmulhq_s16(q11s16, q0s16);
371 q0s16 = vdupq_n_s16(cospi_16_64 * 2);
373 q6s16 = vqrdmulhq_s16(q11s16, q1s16);
375 // stage 2 & stage 3 - even half
376 q1s16 = vdupq_n_s16(cospi_24_64 * 2);
378 q9s16 = vqrdmulhq_s16(q8s16, q0s16);
380 q0s16 = vdupq_n_s16(cospi_8_64 * 2);
382 q13s16 = vqrdmulhq_s16(q10s16, q1s16);
384 q15s16 = vqrdmulhq_s16(q10s16, q0s16);
387 q0s16 = vaddq_s16(q9s16, q15s16);
388 q1s16 = vaddq_s16(q9s16, q13s16);
389 q2s16 = vsubq_s16(q9s16, q13s16);
390 q3s16 = vsubq_s16(q9s16, q15s16);
392 // stage 2 - odd half
393 q13s16 = vsubq_s16(q4s16, q5s16);
394 q4s16 = vaddq_s16(q4s16, q5s16);
395 q14s16 = vsubq_s16(q7s16, q6s16);
396 q7s16 = vaddq_s16(q7s16, q6s16);
397 d26s16 = vget_low_s16(q13s16);
398 d27s16 = vget_high_s16(q13s16);
399 d28s16 = vget_low_s16(q14s16);
400 d29s16 = vget_high_s16(q14s16);
402 d16s16 = vdup_n_s16(cospi_16_64);
403 q9s32 = vmull_s16(d28s16, d16s16);
404 q10s32 = vmull_s16(d29s16, d16s16);
405 q11s32 = vmull_s16(d28s16, d16s16);
406 q12s32 = vmull_s16(d29s16, d16s16);
408 q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
409 q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
410 q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
411 q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
413 d10s16 = vqrshrn_n_s32(q9s32, 14);
414 d11s16 = vqrshrn_n_s32(q10s32, 14);
415 d12s16 = vqrshrn_n_s32(q11s32, 14);
416 d13s16 = vqrshrn_n_s32(q12s32, 14);
417 q5s16 = vcombine_s16(d10s16, d11s16);
418 q6s16 = vcombine_s16(d12s16, d13s16);
421 q8s16 = vaddq_s16(q0s16, q7s16);
422 q9s16 = vaddq_s16(q1s16, q6s16);
423 q10s16 = vaddq_s16(q2s16, q5s16);
424 q11s16 = vaddq_s16(q3s16, q4s16);
425 q12s16 = vsubq_s16(q3s16, q4s16);
426 q13s16 = vsubq_s16(q2s16, q5s16);
427 q14s16 = vsubq_s16(q1s16, q6s16);
428 q15s16 = vsubq_s16(q0s16, q7s16);
430 TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
433 IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
436 q8s16 = vrshrq_n_s16(q8s16, 5);
437 q9s16 = vrshrq_n_s16(q9s16, 5);
438 q10s16 = vrshrq_n_s16(q10s16, 5);
439 q11s16 = vrshrq_n_s16(q11s16, 5);
440 q12s16 = vrshrq_n_s16(q12s16, 5);
441 q13s16 = vrshrq_n_s16(q13s16, 5);
442 q14s16 = vrshrq_n_s16(q14s16, 5);
443 q15s16 = vrshrq_n_s16(q15s16, 5);
447 d0u64 = vld1_u64((uint64_t *)d1);
449 d1u64 = vld1_u64((uint64_t *)d1);
451 d2u64 = vld1_u64((uint64_t *)d1);
453 d3u64 = vld1_u64((uint64_t *)d1);
456 q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
457 q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
458 q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
459 q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
461 d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
462 d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
463 d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
464 d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
466 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
468 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
470 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
472 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
480 d0u64 = vld1_u64((uint64_t *)d1);
482 d1u64 = vld1_u64((uint64_t *)d1);
484 d2u64 = vld1_u64((uint64_t *)d1);
486 d3u64 = vld1_u64((uint64_t *)d1);
489 q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
490 q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
491 q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
492 q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
494 d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
495 d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
496 d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
497 d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
499 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
501 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
503 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
505 vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));