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/arm/transpose_neon.h"
15 #include "vpx_dsp/txfm_common.h"
17 void vpx_idct16x16_256_add_neon_pass1(const int16_t *in, int16_t *out) {
18 int16x4_t d0s16, d1s16, d2s16, d3s16;
19 int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
20 int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
21 int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
22 int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
23 int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
24 int32x4_t q0s32, q1s32, q2s32, q3s32, q5s32, q6s32, q9s32;
25 int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
28 q0x2s16 = vld2q_s16(in);
29 q8s16 = q0x2s16.val[0];
31 q0x2s16 = vld2q_s16(in);
32 q9s16 = q0x2s16.val[0];
34 q0x2s16 = vld2q_s16(in);
35 q10s16 = q0x2s16.val[0];
37 q0x2s16 = vld2q_s16(in);
38 q11s16 = q0x2s16.val[0];
40 q0x2s16 = vld2q_s16(in);
41 q12s16 = q0x2s16.val[0];
43 q0x2s16 = vld2q_s16(in);
44 q13s16 = q0x2s16.val[0];
46 q0x2s16 = vld2q_s16(in);
47 q14s16 = q0x2s16.val[0];
49 q0x2s16 = vld2q_s16(in);
50 q15s16 = q0x2s16.val[0];
52 transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
55 d16s16 = vget_low_s16(q8s16);
56 d17s16 = vget_high_s16(q8s16);
57 d18s16 = vget_low_s16(q9s16);
58 d19s16 = vget_high_s16(q9s16);
59 d20s16 = vget_low_s16(q10s16);
60 d21s16 = vget_high_s16(q10s16);
61 d22s16 = vget_low_s16(q11s16);
62 d23s16 = vget_high_s16(q11s16);
63 d24s16 = vget_low_s16(q12s16);
64 d25s16 = vget_high_s16(q12s16);
65 d26s16 = vget_low_s16(q13s16);
66 d27s16 = vget_high_s16(q13s16);
67 d28s16 = vget_low_s16(q14s16);
68 d29s16 = vget_high_s16(q14s16);
69 d30s16 = vget_low_s16(q15s16);
70 d31s16 = vget_high_s16(q15s16);
73 d0s16 = vdup_n_s16((int16_t)cospi_28_64);
74 d1s16 = vdup_n_s16((int16_t)cospi_4_64);
76 q2s32 = vmull_s16(d18s16, d0s16);
77 q3s32 = vmull_s16(d19s16, d0s16);
78 q5s32 = vmull_s16(d18s16, d1s16);
79 q6s32 = vmull_s16(d19s16, d1s16);
81 q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
82 q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
83 q5s32 = vmlal_s16(q5s32, d30s16, d0s16);
84 q6s32 = vmlal_s16(q6s32, d31s16, d0s16);
86 d2s16 = vdup_n_s16((int16_t)cospi_12_64);
87 d3s16 = vdup_n_s16((int16_t)cospi_20_64);
89 d8s16 = vqrshrn_n_s32(q2s32, 14);
90 d9s16 = vqrshrn_n_s32(q3s32, 14);
91 d14s16 = vqrshrn_n_s32(q5s32, 14);
92 d15s16 = vqrshrn_n_s32(q6s32, 14);
93 q4s16 = vcombine_s16(d8s16, d9s16);
94 q7s16 = vcombine_s16(d14s16, d15s16);
96 q2s32 = vmull_s16(d26s16, d2s16);
97 q3s32 = vmull_s16(d27s16, d2s16);
98 q9s32 = vmull_s16(d26s16, d3s16);
99 q15s32 = vmull_s16(d27s16, d3s16);
101 q2s32 = vmlsl_s16(q2s32, d22s16, d3s16);
102 q3s32 = vmlsl_s16(q3s32, d23s16, d3s16);
103 q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
104 q15s32 = vmlal_s16(q15s32, d23s16, d2s16);
106 d10s16 = vqrshrn_n_s32(q2s32, 14);
107 d11s16 = vqrshrn_n_s32(q3s32, 14);
108 d12s16 = vqrshrn_n_s32(q9s32, 14);
109 d13s16 = vqrshrn_n_s32(q15s32, 14);
110 q5s16 = vcombine_s16(d10s16, d11s16);
111 q6s16 = vcombine_s16(d12s16, d13s16);
114 d30s16 = vdup_n_s16((int16_t)cospi_16_64);
116 q2s32 = vmull_s16(d16s16, d30s16);
117 q11s32 = vmull_s16(d17s16, d30s16);
118 q0s32 = vmull_s16(d24s16, d30s16);
119 q1s32 = vmull_s16(d25s16, d30s16);
121 d30s16 = vdup_n_s16((int16_t)cospi_24_64);
122 d31s16 = vdup_n_s16((int16_t)cospi_8_64);
124 q3s32 = vaddq_s32(q2s32, q0s32);
125 q12s32 = vaddq_s32(q11s32, q1s32);
126 q13s32 = vsubq_s32(q2s32, q0s32);
127 q1s32 = vsubq_s32(q11s32, q1s32);
129 d16s16 = vqrshrn_n_s32(q3s32, 14);
130 d17s16 = vqrshrn_n_s32(q12s32, 14);
131 d18s16 = vqrshrn_n_s32(q13s32, 14);
132 d19s16 = vqrshrn_n_s32(q1s32, 14);
133 q8s16 = vcombine_s16(d16s16, d17s16);
134 q9s16 = vcombine_s16(d18s16, d19s16);
136 q0s32 = vmull_s16(d20s16, d31s16);
137 q1s32 = vmull_s16(d21s16, d31s16);
138 q12s32 = vmull_s16(d20s16, d30s16);
139 q13s32 = vmull_s16(d21s16, d30s16);
141 q0s32 = vmlal_s16(q0s32, d28s16, d30s16);
142 q1s32 = vmlal_s16(q1s32, d29s16, d30s16);
143 q12s32 = vmlsl_s16(q12s32, d28s16, d31s16);
144 q13s32 = vmlsl_s16(q13s32, d29s16, d31s16);
146 d22s16 = vqrshrn_n_s32(q0s32, 14);
147 d23s16 = vqrshrn_n_s32(q1s32, 14);
148 d20s16 = vqrshrn_n_s32(q12s32, 14);
149 d21s16 = vqrshrn_n_s32(q13s32, 14);
150 q10s16 = vcombine_s16(d20s16, d21s16);
151 q11s16 = vcombine_s16(d22s16, d23s16);
153 q13s16 = vsubq_s16(q4s16, q5s16);
154 q4s16 = vaddq_s16(q4s16, q5s16);
155 q14s16 = vsubq_s16(q7s16, q6s16);
156 q15s16 = vaddq_s16(q6s16, q7s16);
157 d26s16 = vget_low_s16(q13s16);
158 d27s16 = vget_high_s16(q13s16);
159 d28s16 = vget_low_s16(q14s16);
160 d29s16 = vget_high_s16(q14s16);
163 q0s16 = vaddq_s16(q8s16, q11s16);
164 q1s16 = vaddq_s16(q9s16, q10s16);
165 q2s16 = vsubq_s16(q9s16, q10s16);
166 q3s16 = vsubq_s16(q8s16, q11s16);
168 d16s16 = vdup_n_s16((int16_t)cospi_16_64);
170 q11s32 = vmull_s16(d26s16, d16s16);
171 q12s32 = vmull_s16(d27s16, d16s16);
172 q9s32 = vmull_s16(d28s16, d16s16);
173 q10s32 = vmull_s16(d29s16, d16s16);
175 q6s32 = vsubq_s32(q9s32, q11s32);
176 q13s32 = vsubq_s32(q10s32, q12s32);
177 q9s32 = vaddq_s32(q9s32, q11s32);
178 q10s32 = vaddq_s32(q10s32, q12s32);
180 d10s16 = vqrshrn_n_s32(q6s32, 14);
181 d11s16 = vqrshrn_n_s32(q13s32, 14);
182 d12s16 = vqrshrn_n_s32(q9s32, 14);
183 d13s16 = vqrshrn_n_s32(q10s32, 14);
184 q5s16 = vcombine_s16(d10s16, d11s16);
185 q6s16 = vcombine_s16(d12s16, d13s16);
188 q8s16 = vaddq_s16(q0s16, q15s16);
189 q9s16 = vaddq_s16(q1s16, q6s16);
190 q10s16 = vaddq_s16(q2s16, q5s16);
191 q11s16 = vaddq_s16(q3s16, q4s16);
192 q12s16 = vsubq_s16(q3s16, q4s16);
193 q13s16 = vsubq_s16(q2s16, q5s16);
194 q14s16 = vsubq_s16(q1s16, q6s16);
195 q15s16 = vsubq_s16(q0s16, q15s16);
198 vst1q_s16(out, q8s16);
200 vst1q_s16(out, q9s16);
202 vst1q_s16(out, q10s16);
204 vst1q_s16(out, q11s16);
206 vst1q_s16(out, q12s16);
208 vst1q_s16(out, q13s16);
210 vst1q_s16(out, q14s16);
212 vst1q_s16(out, q15s16);
215 void vpx_idct16x16_256_add_neon_pass2(const int16_t *src, int16_t *out,
216 int16_t *pass1_output,
217 int16_t skip_adding, uint8_t *dest,
220 uint8x8_t d12u8, d13u8;
221 int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
222 int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
223 int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
224 int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
225 uint64x1_t d24u64, d25u64, d26u64, d27u64;
226 int64x1_t d12s64, d13s64;
227 uint16x8_t q2u16, q3u16, q4u16, q5u16, q8u16;
228 uint16x8_t q9u16, q12u16, q13u16, q14u16, q15u16;
229 int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
230 int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
231 int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q8s32, q9s32;
232 int32x4_t q10s32, q11s32, q12s32, q13s32;
235 q0x2s16 = vld2q_s16(src);
236 q8s16 = q0x2s16.val[0];
238 q0x2s16 = vld2q_s16(src);
239 q9s16 = q0x2s16.val[0];
241 q0x2s16 = vld2q_s16(src);
242 q10s16 = q0x2s16.val[0];
244 q0x2s16 = vld2q_s16(src);
245 q11s16 = q0x2s16.val[0];
247 q0x2s16 = vld2q_s16(src);
248 q12s16 = q0x2s16.val[0];
250 q0x2s16 = vld2q_s16(src);
251 q13s16 = q0x2s16.val[0];
253 q0x2s16 = vld2q_s16(src);
254 q14s16 = q0x2s16.val[0];
256 q0x2s16 = vld2q_s16(src);
257 q15s16 = q0x2s16.val[0];
259 transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
262 d16s16 = vget_low_s16(q8s16);
263 d17s16 = vget_high_s16(q8s16);
264 d18s16 = vget_low_s16(q9s16);
265 d19s16 = vget_high_s16(q9s16);
266 d20s16 = vget_low_s16(q10s16);
267 d21s16 = vget_high_s16(q10s16);
268 d22s16 = vget_low_s16(q11s16);
269 d23s16 = vget_high_s16(q11s16);
270 d24s16 = vget_low_s16(q12s16);
271 d25s16 = vget_high_s16(q12s16);
272 d26s16 = vget_low_s16(q13s16);
273 d27s16 = vget_high_s16(q13s16);
274 d28s16 = vget_low_s16(q14s16);
275 d29s16 = vget_high_s16(q14s16);
276 d30s16 = vget_low_s16(q15s16);
277 d31s16 = vget_high_s16(q15s16);
280 d12s16 = vdup_n_s16((int16_t)cospi_30_64);
281 d13s16 = vdup_n_s16((int16_t)cospi_2_64);
283 q2s32 = vmull_s16(d16s16, d12s16);
284 q3s32 = vmull_s16(d17s16, d12s16);
285 q1s32 = vmull_s16(d16s16, d13s16);
286 q4s32 = vmull_s16(d17s16, d13s16);
288 q2s32 = vmlsl_s16(q2s32, d30s16, d13s16);
289 q3s32 = vmlsl_s16(q3s32, d31s16, d13s16);
290 q1s32 = vmlal_s16(q1s32, d30s16, d12s16);
291 q4s32 = vmlal_s16(q4s32, d31s16, d12s16);
293 d0s16 = vqrshrn_n_s32(q2s32, 14);
294 d1s16 = vqrshrn_n_s32(q3s32, 14);
295 d14s16 = vqrshrn_n_s32(q1s32, 14);
296 d15s16 = vqrshrn_n_s32(q4s32, 14);
297 q0s16 = vcombine_s16(d0s16, d1s16);
298 q7s16 = vcombine_s16(d14s16, d15s16);
300 d30s16 = vdup_n_s16((int16_t)cospi_14_64);
301 d31s16 = vdup_n_s16((int16_t)cospi_18_64);
303 q2s32 = vmull_s16(d24s16, d30s16);
304 q3s32 = vmull_s16(d25s16, d30s16);
305 q4s32 = vmull_s16(d24s16, d31s16);
306 q5s32 = vmull_s16(d25s16, d31s16);
308 q2s32 = vmlsl_s16(q2s32, d22s16, d31s16);
309 q3s32 = vmlsl_s16(q3s32, d23s16, d31s16);
310 q4s32 = vmlal_s16(q4s32, d22s16, d30s16);
311 q5s32 = vmlal_s16(q5s32, d23s16, d30s16);
313 d2s16 = vqrshrn_n_s32(q2s32, 14);
314 d3s16 = vqrshrn_n_s32(q3s32, 14);
315 d12s16 = vqrshrn_n_s32(q4s32, 14);
316 d13s16 = vqrshrn_n_s32(q5s32, 14);
317 q1s16 = vcombine_s16(d2s16, d3s16);
318 q6s16 = vcombine_s16(d12s16, d13s16);
320 d30s16 = vdup_n_s16((int16_t)cospi_22_64);
321 d31s16 = vdup_n_s16((int16_t)cospi_10_64);
323 q11s32 = vmull_s16(d20s16, d30s16);
324 q12s32 = vmull_s16(d21s16, d30s16);
325 q4s32 = vmull_s16(d20s16, d31s16);
326 q5s32 = vmull_s16(d21s16, d31s16);
328 q11s32 = vmlsl_s16(q11s32, d26s16, d31s16);
329 q12s32 = vmlsl_s16(q12s32, d27s16, d31s16);
330 q4s32 = vmlal_s16(q4s32, d26s16, d30s16);
331 q5s32 = vmlal_s16(q5s32, d27s16, d30s16);
333 d4s16 = vqrshrn_n_s32(q11s32, 14);
334 d5s16 = vqrshrn_n_s32(q12s32, 14);
335 d11s16 = vqrshrn_n_s32(q5s32, 14);
336 d10s16 = vqrshrn_n_s32(q4s32, 14);
337 q2s16 = vcombine_s16(d4s16, d5s16);
338 q5s16 = vcombine_s16(d10s16, d11s16);
340 d30s16 = vdup_n_s16((int16_t)cospi_6_64);
341 d31s16 = vdup_n_s16((int16_t)cospi_26_64);
343 q10s32 = vmull_s16(d28s16, d30s16);
344 q11s32 = vmull_s16(d29s16, d30s16);
345 q12s32 = vmull_s16(d28s16, d31s16);
346 q13s32 = vmull_s16(d29s16, d31s16);
348 q10s32 = vmlsl_s16(q10s32, d18s16, d31s16);
349 q11s32 = vmlsl_s16(q11s32, d19s16, d31s16);
350 q12s32 = vmlal_s16(q12s32, d18s16, d30s16);
351 q13s32 = vmlal_s16(q13s32, d19s16, d30s16);
353 d6s16 = vqrshrn_n_s32(q10s32, 14);
354 d7s16 = vqrshrn_n_s32(q11s32, 14);
355 d8s16 = vqrshrn_n_s32(q12s32, 14);
356 d9s16 = vqrshrn_n_s32(q13s32, 14);
357 q3s16 = vcombine_s16(d6s16, d7s16);
358 q4s16 = vcombine_s16(d8s16, d9s16);
361 q9s16 = vsubq_s16(q0s16, q1s16);
362 q0s16 = vaddq_s16(q0s16, q1s16);
363 q10s16 = vsubq_s16(q3s16, q2s16);
364 q11s16 = vaddq_s16(q2s16, q3s16);
365 q12s16 = vaddq_s16(q4s16, q5s16);
366 q13s16 = vsubq_s16(q4s16, q5s16);
367 q14s16 = vsubq_s16(q7s16, q6s16);
368 q7s16 = vaddq_s16(q6s16, q7s16);
371 d18s16 = vget_low_s16(q9s16);
372 d19s16 = vget_high_s16(q9s16);
373 d20s16 = vget_low_s16(q10s16);
374 d21s16 = vget_high_s16(q10s16);
375 d26s16 = vget_low_s16(q13s16);
376 d27s16 = vget_high_s16(q13s16);
377 d28s16 = vget_low_s16(q14s16);
378 d29s16 = vget_high_s16(q14s16);
380 d30s16 = vdup_n_s16((int16_t)cospi_8_64);
381 d31s16 = vdup_n_s16((int16_t)cospi_24_64);
383 q2s32 = vmull_s16(d18s16, d31s16);
384 q3s32 = vmull_s16(d19s16, d31s16);
385 q4s32 = vmull_s16(d28s16, d31s16);
386 q5s32 = vmull_s16(d29s16, d31s16);
388 q2s32 = vmlal_s16(q2s32, d28s16, d30s16);
389 q3s32 = vmlal_s16(q3s32, d29s16, d30s16);
390 q4s32 = vmlsl_s16(q4s32, d18s16, d30s16);
391 q5s32 = vmlsl_s16(q5s32, d19s16, d30s16);
393 d12s16 = vqrshrn_n_s32(q2s32, 14);
394 d13s16 = vqrshrn_n_s32(q3s32, 14);
395 d2s16 = vqrshrn_n_s32(q4s32, 14);
396 d3s16 = vqrshrn_n_s32(q5s32, 14);
397 q1s16 = vcombine_s16(d2s16, d3s16);
398 q6s16 = vcombine_s16(d12s16, d13s16);
403 d30s16 = vdup_n_s16(-cospi_8_64);
404 q11s32 = vmull_s16(d26s16, d30s16);
405 q12s32 = vmull_s16(d27s16, d30s16);
406 q8s32 = vmull_s16(d20s16, d30s16);
407 q9s32 = vmull_s16(d21s16, d30s16);
409 q11s32 = vmlsl_s16(q11s32, d20s16, d31s16);
410 q12s32 = vmlsl_s16(q12s32, d21s16, d31s16);
411 q8s32 = vmlal_s16(q8s32, d26s16, d31s16);
412 q9s32 = vmlal_s16(q9s32, d27s16, d31s16);
414 d4s16 = vqrshrn_n_s32(q11s32, 14);
415 d5s16 = vqrshrn_n_s32(q12s32, 14);
416 d10s16 = vqrshrn_n_s32(q8s32, 14);
417 d11s16 = vqrshrn_n_s32(q9s32, 14);
418 q2s16 = vcombine_s16(d4s16, d5s16);
419 q5s16 = vcombine_s16(d10s16, d11s16);
422 q8s16 = vaddq_s16(q0s16, q3s16);
423 q9s16 = vaddq_s16(q1s16, q2s16);
424 q10s16 = vsubq_s16(q1s16, q2s16);
425 q11s16 = vsubq_s16(q0s16, q3s16);
426 q12s16 = vsubq_s16(q7s16, q4s16);
427 q13s16 = vsubq_s16(q6s16, q5s16);
428 q14s16 = vaddq_s16(q6s16, q5s16);
429 q15s16 = vaddq_s16(q7s16, q4s16);
432 d20s16 = vget_low_s16(q10s16);
433 d21s16 = vget_high_s16(q10s16);
434 d22s16 = vget_low_s16(q11s16);
435 d23s16 = vget_high_s16(q11s16);
436 d24s16 = vget_low_s16(q12s16);
437 d25s16 = vget_high_s16(q12s16);
438 d26s16 = vget_low_s16(q13s16);
439 d27s16 = vget_high_s16(q13s16);
441 d14s16 = vdup_n_s16((int16_t)cospi_16_64);
443 q3s32 = vmull_s16(d26s16, d14s16);
444 q4s32 = vmull_s16(d27s16, d14s16);
445 q0s32 = vmull_s16(d20s16, d14s16);
446 q1s32 = vmull_s16(d21s16, d14s16);
448 q5s32 = vsubq_s32(q3s32, q0s32);
449 q6s32 = vsubq_s32(q4s32, q1s32);
450 q10s32 = vaddq_s32(q3s32, q0s32);
451 q4s32 = vaddq_s32(q4s32, q1s32);
453 d4s16 = vqrshrn_n_s32(q5s32, 14);
454 d5s16 = vqrshrn_n_s32(q6s32, 14);
455 d10s16 = vqrshrn_n_s32(q10s32, 14);
456 d11s16 = vqrshrn_n_s32(q4s32, 14);
457 q2s16 = vcombine_s16(d4s16, d5s16);
458 q5s16 = vcombine_s16(d10s16, d11s16);
460 q0s32 = vmull_s16(d22s16, d14s16);
461 q1s32 = vmull_s16(d23s16, d14s16);
462 q13s32 = vmull_s16(d24s16, d14s16);
463 q6s32 = vmull_s16(d25s16, d14s16);
465 q10s32 = vsubq_s32(q13s32, q0s32);
466 q4s32 = vsubq_s32(q6s32, q1s32);
467 q13s32 = vaddq_s32(q13s32, q0s32);
468 q6s32 = vaddq_s32(q6s32, q1s32);
470 d6s16 = vqrshrn_n_s32(q10s32, 14);
471 d7s16 = vqrshrn_n_s32(q4s32, 14);
472 d8s16 = vqrshrn_n_s32(q13s32, 14);
473 d9s16 = vqrshrn_n_s32(q6s32, 14);
474 q3s16 = vcombine_s16(d6s16, d7s16);
475 q4s16 = vcombine_s16(d8s16, d9s16);
478 if (skip_adding != 0) {
480 // load the data in pass1
481 q0s16 = vld1q_s16(pass1_output);
483 q1s16 = vld1q_s16(pass1_output);
485 d12s64 = vld1_s64((int64_t *)dest);
487 d13s64 = vld1_s64((int64_t *)dest);
490 q12s16 = vaddq_s16(q0s16, q15s16);
491 q13s16 = vaddq_s16(q1s16, q14s16);
492 q12s16 = vrshrq_n_s16(q12s16, 6);
493 q13s16 = vrshrq_n_s16(q13s16, 6);
495 vaddw_u8(vreinterpretq_u16_s16(q12s16), vreinterpret_u8_s64(d12s64));
497 vaddw_u8(vreinterpretq_u16_s16(q13s16), vreinterpret_u8_s64(d13s64));
498 d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
499 d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
500 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
502 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
504 q14s16 = vsubq_s16(q1s16, q14s16);
505 q15s16 = vsubq_s16(q0s16, q15s16);
507 q10s16 = vld1q_s16(pass1_output);
509 q11s16 = vld1q_s16(pass1_output);
511 d12s64 = vld1_s64((int64_t *)dest);
513 d13s64 = vld1_s64((int64_t *)dest);
515 q12s16 = vaddq_s16(q10s16, q5s16);
516 q13s16 = vaddq_s16(q11s16, q4s16);
517 q12s16 = vrshrq_n_s16(q12s16, 6);
518 q13s16 = vrshrq_n_s16(q13s16, 6);
520 vaddw_u8(vreinterpretq_u16_s16(q12s16), vreinterpret_u8_s64(d12s64));
522 vaddw_u8(vreinterpretq_u16_s16(q13s16), vreinterpret_u8_s64(d13s64));
523 d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
524 d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
525 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
527 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
529 q4s16 = vsubq_s16(q11s16, q4s16);
530 q5s16 = vsubq_s16(q10s16, q5s16);
532 q0s16 = vld1q_s16(pass1_output);
534 q1s16 = vld1q_s16(pass1_output);
536 d12s64 = vld1_s64((int64_t *)dest);
538 d13s64 = vld1_s64((int64_t *)dest);
540 q12s16 = vaddq_s16(q0s16, q3s16);
541 q13s16 = vaddq_s16(q1s16, q2s16);
542 q12s16 = vrshrq_n_s16(q12s16, 6);
543 q13s16 = vrshrq_n_s16(q13s16, 6);
545 vaddw_u8(vreinterpretq_u16_s16(q12s16), vreinterpret_u8_s64(d12s64));
547 vaddw_u8(vreinterpretq_u16_s16(q13s16), vreinterpret_u8_s64(d13s64));
548 d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
549 d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
550 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
552 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
554 q2s16 = vsubq_s16(q1s16, q2s16);
555 q3s16 = vsubq_s16(q0s16, q3s16);
557 q10s16 = vld1q_s16(pass1_output);
559 q11s16 = vld1q_s16(pass1_output);
560 d12s64 = vld1_s64((int64_t *)dest);
562 d13s64 = vld1_s64((int64_t *)dest);
564 q12s16 = vaddq_s16(q10s16, q9s16);
565 q13s16 = vaddq_s16(q11s16, q8s16);
566 q12s16 = vrshrq_n_s16(q12s16, 6);
567 q13s16 = vrshrq_n_s16(q13s16, 6);
569 vaddw_u8(vreinterpretq_u16_s16(q12s16), vreinterpret_u8_s64(d12s64));
571 vaddw_u8(vreinterpretq_u16_s16(q13s16), vreinterpret_u8_s64(d13s64));
572 d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
573 d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
574 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
576 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
578 q8s16 = vsubq_s16(q11s16, q8s16);
579 q9s16 = vsubq_s16(q10s16, q9s16);
581 // store the data out 8,9,10,11,12,13,14,15
582 d12s64 = vld1_s64((int64_t *)dest);
584 q8s16 = vrshrq_n_s16(q8s16, 6);
585 q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_s64(d12s64));
586 d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
587 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
590 d12s64 = vld1_s64((int64_t *)dest);
592 q9s16 = vrshrq_n_s16(q9s16, 6);
593 q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_s64(d12s64));
594 d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
595 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
598 d12s64 = vld1_s64((int64_t *)dest);
600 q2s16 = vrshrq_n_s16(q2s16, 6);
601 q2u16 = vaddw_u8(vreinterpretq_u16_s16(q2s16), vreinterpret_u8_s64(d12s64));
602 d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q2u16));
603 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
606 d12s64 = vld1_s64((int64_t *)dest);
608 q3s16 = vrshrq_n_s16(q3s16, 6);
609 q3u16 = vaddw_u8(vreinterpretq_u16_s16(q3s16), vreinterpret_u8_s64(d12s64));
610 d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q3u16));
611 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
614 d12s64 = vld1_s64((int64_t *)dest);
616 q4s16 = vrshrq_n_s16(q4s16, 6);
617 q4u16 = vaddw_u8(vreinterpretq_u16_s16(q4s16), vreinterpret_u8_s64(d12s64));
618 d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q4u16));
619 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
622 d12s64 = vld1_s64((int64_t *)dest);
624 q5s16 = vrshrq_n_s16(q5s16, 6);
625 q5u16 = vaddw_u8(vreinterpretq_u16_s16(q5s16), vreinterpret_u8_s64(d12s64));
626 d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q5u16));
627 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
630 d12s64 = vld1_s64((int64_t *)dest);
632 q14s16 = vrshrq_n_s16(q14s16, 6);
634 vaddw_u8(vreinterpretq_u16_s16(q14s16), vreinterpret_u8_s64(d12s64));
635 d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q14u16));
636 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
639 d12s64 = vld1_s64((int64_t *)dest);
640 q15s16 = vrshrq_n_s16(q15s16, 6);
642 vaddw_u8(vreinterpretq_u16_s16(q15s16), vreinterpret_u8_s64(d12s64));
643 d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q15u16));
644 vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
645 } else { // skip_adding_dest
646 q0s16 = vld1q_s16(pass1_output);
648 q1s16 = vld1q_s16(pass1_output);
650 q12s16 = vaddq_s16(q0s16, q15s16);
651 q13s16 = vaddq_s16(q1s16, q14s16);
652 d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
653 d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
654 d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
655 d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
656 vst1_u64((uint64_t *)out, d24u64);
658 vst1_u64((uint64_t *)out, d25u64);
660 vst1_u64((uint64_t *)out, d26u64);
662 vst1_u64((uint64_t *)out, d27u64);
664 q14s16 = vsubq_s16(q1s16, q14s16);
665 q15s16 = vsubq_s16(q0s16, q15s16);
667 q10s16 = vld1q_s16(pass1_output);
669 q11s16 = vld1q_s16(pass1_output);
671 q12s16 = vaddq_s16(q10s16, q5s16);
672 q13s16 = vaddq_s16(q11s16, q4s16);
673 d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
674 d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
675 d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
676 d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
677 vst1_u64((uint64_t *)out, d24u64);
679 vst1_u64((uint64_t *)out, d25u64);
681 vst1_u64((uint64_t *)out, d26u64);
683 vst1_u64((uint64_t *)out, d27u64);
685 q4s16 = vsubq_s16(q11s16, q4s16);
686 q5s16 = vsubq_s16(q10s16, q5s16);
688 q0s16 = vld1q_s16(pass1_output);
690 q1s16 = vld1q_s16(pass1_output);
692 q12s16 = vaddq_s16(q0s16, q3s16);
693 q13s16 = vaddq_s16(q1s16, q2s16);
694 d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
695 d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
696 d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
697 d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
698 vst1_u64((uint64_t *)out, d24u64);
700 vst1_u64((uint64_t *)out, d25u64);
702 vst1_u64((uint64_t *)out, d26u64);
704 vst1_u64((uint64_t *)out, d27u64);
706 q2s16 = vsubq_s16(q1s16, q2s16);
707 q3s16 = vsubq_s16(q0s16, q3s16);
709 q10s16 = vld1q_s16(pass1_output);
711 q11s16 = vld1q_s16(pass1_output);
713 q12s16 = vaddq_s16(q10s16, q9s16);
714 q13s16 = vaddq_s16(q11s16, q8s16);
715 d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
716 d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
717 d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
718 d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
719 vst1_u64((uint64_t *)out, d24u64);
721 vst1_u64((uint64_t *)out, d25u64);
723 vst1_u64((uint64_t *)out, d26u64);
725 vst1_u64((uint64_t *)out, d27u64);
727 q8s16 = vsubq_s16(q11s16, q8s16);
728 q9s16 = vsubq_s16(q10s16, q9s16);
730 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q8s16)));
732 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q8s16)));
734 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q9s16)));
736 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q9s16)));
738 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q2s16)));
740 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q2s16)));
742 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q3s16)));
744 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q3s16)));
746 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q4s16)));
748 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q4s16)));
750 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q5s16)));
752 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q5s16)));
754 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q14s16)));
756 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q14s16)));
758 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q15s16)));
760 vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q15s16)));
764 void vpx_idct16x16_10_add_neon_pass1(const int16_t *in, int16_t *out) {
766 int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
767 int16x8_t q0s16, q1s16, q2s16, q4s16, q5s16, q6s16, q7s16;
768 int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
769 int32x4_t q6s32, q9s32;
770 int32x4_t q10s32, q11s32, q12s32, q15s32;
773 q0x2s16 = vld2q_s16(in);
774 q8s16 = q0x2s16.val[0];
776 q0x2s16 = vld2q_s16(in);
777 q9s16 = q0x2s16.val[0];
779 q0x2s16 = vld2q_s16(in);
780 q10s16 = q0x2s16.val[0];
782 q0x2s16 = vld2q_s16(in);
783 q11s16 = q0x2s16.val[0];
785 q0x2s16 = vld2q_s16(in);
786 q12s16 = q0x2s16.val[0];
788 q0x2s16 = vld2q_s16(in);
789 q13s16 = q0x2s16.val[0];
791 q0x2s16 = vld2q_s16(in);
792 q14s16 = q0x2s16.val[0];
794 q0x2s16 = vld2q_s16(in);
795 q15s16 = q0x2s16.val[0];
797 transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
801 q0s16 = vdupq_n_s16((int16_t)cospi_28_64 * 2);
802 q1s16 = vdupq_n_s16((int16_t)cospi_4_64 * 2);
804 q4s16 = vqrdmulhq_s16(q9s16, q0s16);
805 q7s16 = vqrdmulhq_s16(q9s16, q1s16);
808 q1s16 = vdupq_n_s16((int16_t)cospi_16_64 * 2);
809 d4s16 = vdup_n_s16((int16_t)cospi_16_64);
811 q8s16 = vqrdmulhq_s16(q8s16, q1s16);
813 d8s16 = vget_low_s16(q4s16);
814 d9s16 = vget_high_s16(q4s16);
815 d14s16 = vget_low_s16(q7s16);
816 d15s16 = vget_high_s16(q7s16);
817 q9s32 = vmull_s16(d14s16, d4s16);
818 q10s32 = vmull_s16(d15s16, d4s16);
819 q12s32 = vmull_s16(d9s16, d4s16);
820 q11s32 = vmull_s16(d8s16, d4s16);
822 q15s32 = vsubq_s32(q10s32, q12s32);
823 q6s32 = vsubq_s32(q9s32, q11s32);
824 q9s32 = vaddq_s32(q9s32, q11s32);
825 q10s32 = vaddq_s32(q10s32, q12s32);
827 d11s16 = vqrshrn_n_s32(q15s32, 14);
828 d10s16 = vqrshrn_n_s32(q6s32, 14);
829 d12s16 = vqrshrn_n_s32(q9s32, 14);
830 d13s16 = vqrshrn_n_s32(q10s32, 14);
831 q5s16 = vcombine_s16(d10s16, d11s16);
832 q6s16 = vcombine_s16(d12s16, d13s16);
835 q2s16 = vaddq_s16(q8s16, q7s16);
836 q9s16 = vaddq_s16(q8s16, q6s16);
837 q10s16 = vaddq_s16(q8s16, q5s16);
838 q11s16 = vaddq_s16(q8s16, q4s16);
839 q12s16 = vsubq_s16(q8s16, q4s16);
840 q13s16 = vsubq_s16(q8s16, q5s16);
841 q14s16 = vsubq_s16(q8s16, q6s16);
842 q15s16 = vsubq_s16(q8s16, q7s16);
845 vst1q_s16(out, q2s16);
847 vst1q_s16(out, q9s16);
849 vst1q_s16(out, q10s16);
851 vst1q_s16(out, q11s16);
853 vst1q_s16(out, q12s16);
855 vst1q_s16(out, q13s16);
857 vst1q_s16(out, q14s16);
859 vst1q_s16(out, q15s16);
862 void vpx_idct16x16_10_add_neon_pass2(const int16_t *src, int16_t *out,
863 int16_t *pass1_output) {
864 int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
865 int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
866 int16x4_t d20s16, d21s16, d22s16, d23s16;
867 int16x4_t d24s16, d25s16, d26s16, d27s16, d30s16, d31s16;
868 uint64x1_t d4u64, d5u64, d6u64, d7u64, d8u64, d9u64, d10u64, d11u64;
869 uint64x1_t d16u64, d17u64, d18u64, d19u64;
870 uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64;
871 int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
872 int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
873 int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q8s32, q9s32;
874 int32x4_t q10s32, q11s32, q12s32, q13s32;
877 q0x2s16 = vld2q_s16(src);
878 q8s16 = q0x2s16.val[0];
880 q0x2s16 = vld2q_s16(src);
881 q9s16 = q0x2s16.val[0];
883 q0x2s16 = vld2q_s16(src);
884 q10s16 = q0x2s16.val[0];
886 q0x2s16 = vld2q_s16(src);
887 q11s16 = q0x2s16.val[0];
889 q0x2s16 = vld2q_s16(src);
890 q12s16 = q0x2s16.val[0];
892 q0x2s16 = vld2q_s16(src);
893 q13s16 = q0x2s16.val[0];
895 q0x2s16 = vld2q_s16(src);
896 q14s16 = q0x2s16.val[0];
898 q0x2s16 = vld2q_s16(src);
899 q15s16 = q0x2s16.val[0];
901 transpose_s16_8x8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
905 q6s16 = vdupq_n_s16((int16_t)cospi_30_64 * 2);
906 q0s16 = vqrdmulhq_s16(q8s16, q6s16);
907 q6s16 = vdupq_n_s16((int16_t)cospi_2_64 * 2);
908 q7s16 = vqrdmulhq_s16(q8s16, q6s16);
910 q15s16 = vdupq_n_s16((int16_t)-cospi_26_64 * 2);
911 q14s16 = vdupq_n_s16((int16_t)cospi_6_64 * 2);
912 q3s16 = vqrdmulhq_s16(q9s16, q15s16);
913 q4s16 = vqrdmulhq_s16(q9s16, q14s16);
916 d0s16 = vget_low_s16(q0s16);
917 d1s16 = vget_high_s16(q0s16);
918 d6s16 = vget_low_s16(q3s16);
919 d7s16 = vget_high_s16(q3s16);
920 d8s16 = vget_low_s16(q4s16);
921 d9s16 = vget_high_s16(q4s16);
922 d14s16 = vget_low_s16(q7s16);
923 d15s16 = vget_high_s16(q7s16);
925 d30s16 = vdup_n_s16((int16_t)cospi_8_64);
926 d31s16 = vdup_n_s16((int16_t)cospi_24_64);
928 q12s32 = vmull_s16(d14s16, d31s16);
929 q5s32 = vmull_s16(d15s16, d31s16);
930 q2s32 = vmull_s16(d0s16, d31s16);
931 q11s32 = vmull_s16(d1s16, d31s16);
933 q12s32 = vmlsl_s16(q12s32, d0s16, d30s16);
934 q5s32 = vmlsl_s16(q5s32, d1s16, d30s16);
935 q2s32 = vmlal_s16(q2s32, d14s16, d30s16);
936 q11s32 = vmlal_s16(q11s32, d15s16, d30s16);
938 d2s16 = vqrshrn_n_s32(q12s32, 14);
939 d3s16 = vqrshrn_n_s32(q5s32, 14);
940 d12s16 = vqrshrn_n_s32(q2s32, 14);
941 d13s16 = vqrshrn_n_s32(q11s32, 14);
942 q1s16 = vcombine_s16(d2s16, d3s16);
943 q6s16 = vcombine_s16(d12s16, d13s16);
945 d30s16 = vdup_n_s16(-cospi_8_64);
946 q10s32 = vmull_s16(d8s16, d30s16);
947 q13s32 = vmull_s16(d9s16, d30s16);
948 q8s32 = vmull_s16(d6s16, d30s16);
949 q9s32 = vmull_s16(d7s16, d30s16);
951 q10s32 = vmlsl_s16(q10s32, d6s16, d31s16);
952 q13s32 = vmlsl_s16(q13s32, d7s16, d31s16);
953 q8s32 = vmlal_s16(q8s32, d8s16, d31s16);
954 q9s32 = vmlal_s16(q9s32, d9s16, d31s16);
956 d4s16 = vqrshrn_n_s32(q10s32, 14);
957 d5s16 = vqrshrn_n_s32(q13s32, 14);
958 d10s16 = vqrshrn_n_s32(q8s32, 14);
959 d11s16 = vqrshrn_n_s32(q9s32, 14);
960 q2s16 = vcombine_s16(d4s16, d5s16);
961 q5s16 = vcombine_s16(d10s16, d11s16);
964 q8s16 = vaddq_s16(q0s16, q3s16);
965 q9s16 = vaddq_s16(q1s16, q2s16);
966 q10s16 = vsubq_s16(q1s16, q2s16);
967 q11s16 = vsubq_s16(q0s16, q3s16);
968 q12s16 = vsubq_s16(q7s16, q4s16);
969 q13s16 = vsubq_s16(q6s16, q5s16);
970 q14s16 = vaddq_s16(q6s16, q5s16);
971 q15s16 = vaddq_s16(q7s16, q4s16);
974 d20s16 = vget_low_s16(q10s16);
975 d21s16 = vget_high_s16(q10s16);
976 d22s16 = vget_low_s16(q11s16);
977 d23s16 = vget_high_s16(q11s16);
978 d24s16 = vget_low_s16(q12s16);
979 d25s16 = vget_high_s16(q12s16);
980 d26s16 = vget_low_s16(q13s16);
981 d27s16 = vget_high_s16(q13s16);
983 d14s16 = vdup_n_s16((int16_t)cospi_16_64);
984 q3s32 = vmull_s16(d26s16, d14s16);
985 q4s32 = vmull_s16(d27s16, d14s16);
986 q0s32 = vmull_s16(d20s16, d14s16);
987 q1s32 = vmull_s16(d21s16, d14s16);
989 q5s32 = vsubq_s32(q3s32, q0s32);
990 q6s32 = vsubq_s32(q4s32, q1s32);
991 q0s32 = vaddq_s32(q3s32, q0s32);
992 q4s32 = vaddq_s32(q4s32, q1s32);
994 d4s16 = vqrshrn_n_s32(q5s32, 14);
995 d5s16 = vqrshrn_n_s32(q6s32, 14);
996 d10s16 = vqrshrn_n_s32(q0s32, 14);
997 d11s16 = vqrshrn_n_s32(q4s32, 14);
998 q2s16 = vcombine_s16(d4s16, d5s16);
999 q5s16 = vcombine_s16(d10s16, d11s16);
1001 q0s32 = vmull_s16(d22s16, d14s16);
1002 q1s32 = vmull_s16(d23s16, d14s16);
1003 q13s32 = vmull_s16(d24s16, d14s16);
1004 q6s32 = vmull_s16(d25s16, d14s16);
1006 q10s32 = vsubq_s32(q13s32, q0s32);
1007 q4s32 = vsubq_s32(q6s32, q1s32);
1008 q13s32 = vaddq_s32(q13s32, q0s32);
1009 q6s32 = vaddq_s32(q6s32, q1s32);
1011 d6s16 = vqrshrn_n_s32(q10s32, 14);
1012 d7s16 = vqrshrn_n_s32(q4s32, 14);
1013 d8s16 = vqrshrn_n_s32(q13s32, 14);
1014 d9s16 = vqrshrn_n_s32(q6s32, 14);
1015 q3s16 = vcombine_s16(d6s16, d7s16);
1016 q4s16 = vcombine_s16(d8s16, d9s16);
1019 q0s16 = vld1q_s16(pass1_output);
1021 q1s16 = vld1q_s16(pass1_output);
1023 q12s16 = vaddq_s16(q0s16, q15s16);
1024 q13s16 = vaddq_s16(q1s16, q14s16);
1025 d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
1026 d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
1027 d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
1028 d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
1029 vst1_u64((uint64_t *)out, d24u64);
1031 vst1_u64((uint64_t *)out, d25u64);
1033 vst1_u64((uint64_t *)out, d26u64);
1035 vst1_u64((uint64_t *)out, d27u64);
1037 q14s16 = vsubq_s16(q1s16, q14s16);
1038 q15s16 = vsubq_s16(q0s16, q15s16);
1040 q10s16 = vld1q_s16(pass1_output);
1042 q11s16 = vld1q_s16(pass1_output);
1044 q12s16 = vaddq_s16(q10s16, q5s16);
1045 q13s16 = vaddq_s16(q11s16, q4s16);
1046 d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
1047 d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
1048 d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
1049 d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
1050 vst1_u64((uint64_t *)out, d24u64);
1052 vst1_u64((uint64_t *)out, d25u64);
1054 vst1_u64((uint64_t *)out, d26u64);
1056 vst1_u64((uint64_t *)out, d27u64);
1058 q4s16 = vsubq_s16(q11s16, q4s16);
1059 q5s16 = vsubq_s16(q10s16, q5s16);
1061 q0s16 = vld1q_s16(pass1_output);
1063 q1s16 = vld1q_s16(pass1_output);
1065 q12s16 = vaddq_s16(q0s16, q3s16);
1066 q13s16 = vaddq_s16(q1s16, q2s16);
1067 d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
1068 d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
1069 d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
1070 d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
1071 vst1_u64((uint64_t *)out, d24u64);
1073 vst1_u64((uint64_t *)out, d25u64);
1075 vst1_u64((uint64_t *)out, d26u64);
1077 vst1_u64((uint64_t *)out, d27u64);
1079 q2s16 = vsubq_s16(q1s16, q2s16);
1080 q3s16 = vsubq_s16(q0s16, q3s16);
1082 q10s16 = vld1q_s16(pass1_output);
1084 q11s16 = vld1q_s16(pass1_output);
1085 q12s16 = vaddq_s16(q10s16, q9s16);
1086 q13s16 = vaddq_s16(q11s16, q8s16);
1087 d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
1088 d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
1089 d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
1090 d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
1091 vst1_u64((uint64_t *)out, d24u64);
1093 vst1_u64((uint64_t *)out, d25u64);
1095 vst1_u64((uint64_t *)out, d26u64);
1097 vst1_u64((uint64_t *)out, d27u64);
1099 q8s16 = vsubq_s16(q11s16, q8s16);
1100 q9s16 = vsubq_s16(q10s16, q9s16);
1102 d4u64 = vreinterpret_u64_s16(vget_low_s16(q2s16));
1103 d5u64 = vreinterpret_u64_s16(vget_high_s16(q2s16));
1104 d6u64 = vreinterpret_u64_s16(vget_low_s16(q3s16));
1105 d7u64 = vreinterpret_u64_s16(vget_high_s16(q3s16));
1106 d8u64 = vreinterpret_u64_s16(vget_low_s16(q4s16));
1107 d9u64 = vreinterpret_u64_s16(vget_high_s16(q4s16));
1108 d10u64 = vreinterpret_u64_s16(vget_low_s16(q5s16));
1109 d11u64 = vreinterpret_u64_s16(vget_high_s16(q5s16));
1110 d16u64 = vreinterpret_u64_s16(vget_low_s16(q8s16));
1111 d17u64 = vreinterpret_u64_s16(vget_high_s16(q8s16));
1112 d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16));
1113 d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16));
1114 d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16));
1115 d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16));
1116 d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16));
1117 d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16));
1119 vst1_u64((uint64_t *)out, d16u64);
1121 vst1_u64((uint64_t *)out, d17u64);
1123 vst1_u64((uint64_t *)out, d18u64);
1125 vst1_u64((uint64_t *)out, d19u64);
1127 vst1_u64((uint64_t *)out, d4u64);
1129 vst1_u64((uint64_t *)out, d5u64);
1131 vst1_u64((uint64_t *)out, d6u64);
1133 vst1_u64((uint64_t *)out, d7u64);
1135 vst1_u64((uint64_t *)out, d8u64);
1137 vst1_u64((uint64_t *)out, d9u64);
1139 vst1_u64((uint64_t *)out, d10u64);
1141 vst1_u64((uint64_t *)out, d11u64);
1143 vst1_u64((uint64_t *)out, d28u64);
1145 vst1_u64((uint64_t *)out, d29u64);
1147 vst1_u64((uint64_t *)out, d30u64);
1149 vst1_u64((uint64_t *)out, d31u64);