1/*
2 *  Copyright (c) 2017 The WebM project authors. All Rights Reserved.
3 *
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.
9 */
10
11#include <arm_neon.h>
12
13#include "./vpx_config.h"
14#include "./vpx_dsp_rtcd.h"
15#include "vpx_dsp/arm/idct_neon.h"
16#include "vpx_dsp/arm/transpose_neon.h"
17#include "vpx_dsp/txfm_common.h"
18
19static INLINE void load_8x8_s32_dual(
20    const tran_low_t *input, int32x4x2_t *const in0, int32x4x2_t *const in1,
21    int32x4x2_t *const in2, int32x4x2_t *const in3, int32x4x2_t *const in4,
22    int32x4x2_t *const in5, int32x4x2_t *const in6, int32x4x2_t *const in7) {
23  in0->val[0] = vld1q_s32(input);
24  in0->val[1] = vld1q_s32(input + 4);
25  input += 32;
26  in1->val[0] = vld1q_s32(input);
27  in1->val[1] = vld1q_s32(input + 4);
28  input += 32;
29  in2->val[0] = vld1q_s32(input);
30  in2->val[1] = vld1q_s32(input + 4);
31  input += 32;
32  in3->val[0] = vld1q_s32(input);
33  in3->val[1] = vld1q_s32(input + 4);
34  input += 32;
35  in4->val[0] = vld1q_s32(input);
36  in4->val[1] = vld1q_s32(input + 4);
37  input += 32;
38  in5->val[0] = vld1q_s32(input);
39  in5->val[1] = vld1q_s32(input + 4);
40  input += 32;
41  in6->val[0] = vld1q_s32(input);
42  in6->val[1] = vld1q_s32(input + 4);
43  input += 32;
44  in7->val[0] = vld1q_s32(input);
45  in7->val[1] = vld1q_s32(input + 4);
46}
47
48static INLINE void load_4x8_s32_dual(const tran_low_t *input,
49                                     int32x4_t *const in0, int32x4_t *const in1,
50                                     int32x4_t *const in2, int32x4_t *const in3,
51                                     int32x4_t *const in4, int32x4_t *const in5,
52                                     int32x4_t *const in6,
53                                     int32x4_t *const in7) {
54  *in0 = vld1q_s32(input);
55  input += 32;
56  *in1 = vld1q_s32(input);
57  input += 32;
58  *in2 = vld1q_s32(input);
59  input += 32;
60  *in3 = vld1q_s32(input);
61  input += 32;
62  *in4 = vld1q_s32(input);
63  input += 32;
64  *in5 = vld1q_s32(input);
65  input += 32;
66  *in6 = vld1q_s32(input);
67  input += 32;
68  *in7 = vld1q_s32(input);
69}
70
71// Only for the first pass of the  _135_ variant. Since it only uses values from
72// the top left 16x16 it can safely assume all the remaining values are 0 and
73// skip an awful lot of calculations. In fact, only the first 12 columns make
74// the cut. None of the elements in the 13th, 14th, 15th or 16th columns are
75// used so it skips any calls to input[12|13|14|15] too.
76// In C this does a single row of 32 for each call. Here it transposes the top
77// left 12x8 to allow using SIMD.
78
79// vp9/common/vp9_scan.c:vp9_default_iscan_32x32 arranges the first 135 non-zero
80// coefficients as follows:
81//      0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
82//  0   0   2   5  10  17  25  38  47  62  83 101 121
83//  1   1   4   8  15  22  30  45  58  74  92 112 133
84//  2   3   7  12  18  28  36  52  64  82 102 118
85//  3   6  11  16  23  31  43  60  73  90 109 126
86//  4   9  14  19  29  37  50  65  78  98 116 134
87//  5  13  20  26  35  44  54  72  85 105 123
88//  6  21  27  33  42  53  63  80  94 113 132
89//  7  24  32  39  48  57  71  88 104 120
90//  8  34  40  46  56  68  81  96 111 130
91//  9  41  49  55  67  77  91 107 124
92// 10  51  59  66  76  89  99 119 131
93// 11  61  69  75  87 100 114 129
94// 12  70  79  86  97 108 122
95// 13  84  93 103 110 125
96// 14  98 106 115 127
97// 15 117 128
98static void vpx_highbd_idct32_12_neon(const tran_low_t *const input,
99                                      int32_t *output) {
100  int32x4x2_t in[12], s1[32], s2[32], s3[32], s4[32], s5[32], s6[32], s7[32],
101      s8[32];
102
103  load_8x8_s32_dual(input, &in[0], &in[1], &in[2], &in[3], &in[4], &in[5],
104                    &in[6], &in[7]);
105  transpose_s32_8x8(&in[0], &in[1], &in[2], &in[3], &in[4], &in[5], &in[6],
106                    &in[7]);
107
108  load_4x8_s32_dual(input + 8, &in[8].val[0], &in[8].val[1], &in[9].val[0],
109                    &in[9].val[1], &in[10].val[0], &in[10].val[1],
110                    &in[11].val[0], &in[11].val[1]);
111  transpose_s32_4x8(&in[8].val[0], &in[8].val[1], &in[9].val[0], &in[9].val[1],
112                    &in[10].val[0], &in[10].val[1], &in[11].val[0],
113                    &in[11].val[1]);
114
115  // stage 1
116  s1[16] = multiply_shift_and_narrow_s32_dual(in[1], cospi_31_64);
117  s1[31] = multiply_shift_and_narrow_s32_dual(in[1], cospi_1_64);
118
119  s1[18] = multiply_shift_and_narrow_s32_dual(in[9], cospi_23_64);
120  s1[29] = multiply_shift_and_narrow_s32_dual(in[9], cospi_9_64);
121
122  s1[19] = multiply_shift_and_narrow_s32_dual(in[7], -cospi_25_64);
123  s1[28] = multiply_shift_and_narrow_s32_dual(in[7], cospi_7_64);
124
125  s1[20] = multiply_shift_and_narrow_s32_dual(in[5], cospi_27_64);
126  s1[27] = multiply_shift_and_narrow_s32_dual(in[5], cospi_5_64);
127
128  s1[21] = multiply_shift_and_narrow_s32_dual(in[11], -cospi_21_64);
129  s1[26] = multiply_shift_and_narrow_s32_dual(in[11], cospi_11_64);
130
131  s1[23] = multiply_shift_and_narrow_s32_dual(in[3], -cospi_29_64);
132  s1[24] = multiply_shift_and_narrow_s32_dual(in[3], cospi_3_64);
133
134  // stage 2
135  s2[8] = multiply_shift_and_narrow_s32_dual(in[2], cospi_30_64);
136  s2[15] = multiply_shift_and_narrow_s32_dual(in[2], cospi_2_64);
137
138  s2[10] = multiply_shift_and_narrow_s32_dual(in[10], cospi_22_64);
139  s2[13] = multiply_shift_and_narrow_s32_dual(in[10], cospi_10_64);
140
141  s2[11] = multiply_shift_and_narrow_s32_dual(in[6], -cospi_26_64);
142  s2[12] = multiply_shift_and_narrow_s32_dual(in[6], cospi_6_64);
143
144  s2[18] = highbd_idct_sub_dual(s1[19], s1[18]);
145  s2[19] = highbd_idct_add_dual(s1[18], s1[19]);
146  s2[20] = highbd_idct_add_dual(s1[20], s1[21]);
147  s2[21] = highbd_idct_sub_dual(s1[20], s1[21]);
148  s2[26] = highbd_idct_sub_dual(s1[27], s1[26]);
149  s2[27] = highbd_idct_add_dual(s1[26], s1[27]);
150  s2[28] = highbd_idct_add_dual(s1[28], s1[29]);
151  s2[29] = highbd_idct_sub_dual(s1[28], s1[29]);
152
153  // stage 3
154  s3[4] = multiply_shift_and_narrow_s32_dual(in[4], cospi_28_64);
155  s3[7] = multiply_shift_and_narrow_s32_dual(in[4], cospi_4_64);
156
157  s3[10] = highbd_idct_sub_dual(s2[11], s2[10]);
158  s3[11] = highbd_idct_add_dual(s2[10], s2[11]);
159  s3[12] = highbd_idct_add_dual(s2[12], s2[13]);
160  s3[13] = highbd_idct_sub_dual(s2[12], s2[13]);
161
162  s3[17] = multiply_accumulate_shift_and_narrow_s32_dual(s1[16], -cospi_4_64,
163                                                         s1[31], cospi_28_64);
164  s3[30] = multiply_accumulate_shift_and_narrow_s32_dual(s1[16], cospi_28_64,
165                                                         s1[31], cospi_4_64);
166
167  s3[18] = multiply_accumulate_shift_and_narrow_s32_dual(s2[18], -cospi_28_64,
168                                                         s2[29], -cospi_4_64);
169  s3[29] = multiply_accumulate_shift_and_narrow_s32_dual(s2[18], -cospi_4_64,
170                                                         s2[29], cospi_28_64);
171
172  s3[21] = multiply_accumulate_shift_and_narrow_s32_dual(s2[21], -cospi_20_64,
173                                                         s2[26], cospi_12_64);
174  s3[26] = multiply_accumulate_shift_and_narrow_s32_dual(s2[21], cospi_12_64,
175                                                         s2[26], cospi_20_64);
176
177  s3[22] = multiply_accumulate_shift_and_narrow_s32_dual(s1[23], -cospi_12_64,
178                                                         s1[24], -cospi_20_64);
179  s3[25] = multiply_accumulate_shift_and_narrow_s32_dual(s1[23], -cospi_20_64,
180                                                         s1[24], cospi_12_64);
181
182  // stage 4
183  s4[0] = multiply_shift_and_narrow_s32_dual(in[0], cospi_16_64);
184  s4[2] = multiply_shift_and_narrow_s32_dual(in[8], cospi_24_64);
185  s4[3] = multiply_shift_and_narrow_s32_dual(in[8], cospi_8_64);
186
187  s4[9] = multiply_accumulate_shift_and_narrow_s32_dual(s2[8], -cospi_8_64,
188                                                        s2[15], cospi_24_64);
189  s4[14] = multiply_accumulate_shift_and_narrow_s32_dual(s2[8], cospi_24_64,
190                                                         s2[15], cospi_8_64);
191
192  s4[10] = multiply_accumulate_shift_and_narrow_s32_dual(s3[10], -cospi_24_64,
193                                                         s3[13], -cospi_8_64);
194  s4[13] = multiply_accumulate_shift_and_narrow_s32_dual(s3[10], -cospi_8_64,
195                                                         s3[13], cospi_24_64);
196
197  s4[16] = highbd_idct_add_dual(s1[16], s2[19]);
198  s4[17] = highbd_idct_add_dual(s3[17], s3[18]);
199  s4[18] = highbd_idct_sub_dual(s3[17], s3[18]);
200  s4[19] = highbd_idct_sub_dual(s1[16], s2[19]);
201  s4[20] = highbd_idct_sub_dual(s1[23], s2[20]);
202  s4[21] = highbd_idct_sub_dual(s3[22], s3[21]);
203  s4[22] = highbd_idct_add_dual(s3[21], s3[22]);
204  s4[23] = highbd_idct_add_dual(s2[20], s1[23]);
205  s4[24] = highbd_idct_add_dual(s1[24], s2[27]);
206  s4[25] = highbd_idct_add_dual(s3[25], s3[26]);
207  s4[26] = highbd_idct_sub_dual(s3[25], s3[26]);
208  s4[27] = highbd_idct_sub_dual(s1[24], s2[27]);
209  s4[28] = highbd_idct_sub_dual(s1[31], s2[28]);
210  s4[29] = highbd_idct_sub_dual(s3[30], s3[29]);
211  s4[30] = highbd_idct_add_dual(s3[29], s3[30]);
212  s4[31] = highbd_idct_add_dual(s2[28], s1[31]);
213
214  // stage 5
215  s5[0] = highbd_idct_add_dual(s4[0], s4[3]);
216  s5[1] = highbd_idct_add_dual(s4[0], s4[2]);
217  s5[2] = highbd_idct_sub_dual(s4[0], s4[2]);
218  s5[3] = highbd_idct_sub_dual(s4[0], s4[3]);
219
220  s5[5] = sub_multiply_shift_and_narrow_s32_dual(s3[7], s3[4], cospi_16_64);
221  s5[6] = add_multiply_shift_and_narrow_s32_dual(s3[4], s3[7], cospi_16_64);
222
223  s5[8] = highbd_idct_add_dual(s2[8], s3[11]);
224  s5[9] = highbd_idct_add_dual(s4[9], s4[10]);
225  s5[10] = highbd_idct_sub_dual(s4[9], s4[10]);
226  s5[11] = highbd_idct_sub_dual(s2[8], s3[11]);
227  s5[12] = highbd_idct_sub_dual(s2[15], s3[12]);
228  s5[13] = highbd_idct_sub_dual(s4[14], s4[13]);
229  s5[14] = highbd_idct_add_dual(s4[13], s4[14]);
230  s5[15] = highbd_idct_add_dual(s2[15], s3[12]);
231
232  s5[18] = multiply_accumulate_shift_and_narrow_s32_dual(s4[18], -cospi_8_64,
233                                                         s4[29], cospi_24_64);
234  s5[29] = multiply_accumulate_shift_and_narrow_s32_dual(s4[18], cospi_24_64,
235                                                         s4[29], cospi_8_64);
236
237  s5[19] = multiply_accumulate_shift_and_narrow_s32_dual(s4[19], -cospi_8_64,
238                                                         s4[28], cospi_24_64);
239  s5[28] = multiply_accumulate_shift_and_narrow_s32_dual(s4[19], cospi_24_64,
240                                                         s4[28], cospi_8_64);
241
242  s5[20] = multiply_accumulate_shift_and_narrow_s32_dual(s4[20], -cospi_24_64,
243                                                         s4[27], -cospi_8_64);
244  s5[27] = multiply_accumulate_shift_and_narrow_s32_dual(s4[20], -cospi_8_64,
245                                                         s4[27], cospi_24_64);
246
247  s5[21] = multiply_accumulate_shift_and_narrow_s32_dual(s4[21], -cospi_24_64,
248                                                         s4[26], -cospi_8_64);
249  s5[26] = multiply_accumulate_shift_and_narrow_s32_dual(s4[21], -cospi_8_64,
250                                                         s4[26], cospi_24_64);
251
252  // stage 6
253  s6[0] = highbd_idct_add_dual(s5[0], s3[7]);
254  s6[1] = highbd_idct_add_dual(s5[1], s5[6]);
255  s6[2] = highbd_idct_add_dual(s5[2], s5[5]);
256  s6[3] = highbd_idct_add_dual(s5[3], s3[4]);
257  s6[4] = highbd_idct_sub_dual(s5[3], s3[4]);
258  s6[5] = highbd_idct_sub_dual(s5[2], s5[5]);
259  s6[6] = highbd_idct_sub_dual(s5[1], s5[6]);
260  s6[7] = highbd_idct_sub_dual(s5[0], s3[7]);
261
262  s6[10] = sub_multiply_shift_and_narrow_s32_dual(s5[13], s5[10], cospi_16_64);
263  s6[13] = add_multiply_shift_and_narrow_s32_dual(s5[10], s5[13], cospi_16_64);
264
265  s6[11] = sub_multiply_shift_and_narrow_s32_dual(s5[12], s5[11], cospi_16_64);
266  s6[12] = add_multiply_shift_and_narrow_s32_dual(s5[11], s5[12], cospi_16_64);
267
268  s6[16] = highbd_idct_add_dual(s4[16], s4[23]);
269  s6[17] = highbd_idct_add_dual(s4[17], s4[22]);
270  s6[18] = highbd_idct_add_dual(s5[18], s5[21]);
271  s6[19] = highbd_idct_add_dual(s5[19], s5[20]);
272  s6[20] = highbd_idct_sub_dual(s5[19], s5[20]);
273  s6[21] = highbd_idct_sub_dual(s5[18], s5[21]);
274  s6[22] = highbd_idct_sub_dual(s4[17], s4[22]);
275  s6[23] = highbd_idct_sub_dual(s4[16], s4[23]);
276
277  s6[24] = highbd_idct_sub_dual(s4[31], s4[24]);
278  s6[25] = highbd_idct_sub_dual(s4[30], s4[25]);
279  s6[26] = highbd_idct_sub_dual(s5[29], s5[26]);
280  s6[27] = highbd_idct_sub_dual(s5[28], s5[27]);
281  s6[28] = highbd_idct_add_dual(s5[27], s5[28]);
282  s6[29] = highbd_idct_add_dual(s5[26], s5[29]);
283  s6[30] = highbd_idct_add_dual(s4[25], s4[30]);
284  s6[31] = highbd_idct_add_dual(s4[24], s4[31]);
285
286  // stage 7
287  s7[0] = highbd_idct_add_dual(s6[0], s5[15]);
288  s7[1] = highbd_idct_add_dual(s6[1], s5[14]);
289  s7[2] = highbd_idct_add_dual(s6[2], s6[13]);
290  s7[3] = highbd_idct_add_dual(s6[3], s6[12]);
291  s7[4] = highbd_idct_add_dual(s6[4], s6[11]);
292  s7[5] = highbd_idct_add_dual(s6[5], s6[10]);
293  s7[6] = highbd_idct_add_dual(s6[6], s5[9]);
294  s7[7] = highbd_idct_add_dual(s6[7], s5[8]);
295  s7[8] = highbd_idct_sub_dual(s6[7], s5[8]);
296  s7[9] = highbd_idct_sub_dual(s6[6], s5[9]);
297  s7[10] = highbd_idct_sub_dual(s6[5], s6[10]);
298  s7[11] = highbd_idct_sub_dual(s6[4], s6[11]);
299  s7[12] = highbd_idct_sub_dual(s6[3], s6[12]);
300  s7[13] = highbd_idct_sub_dual(s6[2], s6[13]);
301  s7[14] = highbd_idct_sub_dual(s6[1], s5[14]);
302  s7[15] = highbd_idct_sub_dual(s6[0], s5[15]);
303
304  s7[20] = sub_multiply_shift_and_narrow_s32_dual(s6[27], s6[20], cospi_16_64);
305  s7[27] = add_multiply_shift_and_narrow_s32_dual(s6[20], s6[27], cospi_16_64);
306
307  s7[21] = sub_multiply_shift_and_narrow_s32_dual(s6[26], s6[21], cospi_16_64);
308  s7[26] = add_multiply_shift_and_narrow_s32_dual(s6[21], s6[26], cospi_16_64);
309
310  s7[22] = sub_multiply_shift_and_narrow_s32_dual(s6[25], s6[22], cospi_16_64);
311  s7[25] = add_multiply_shift_and_narrow_s32_dual(s6[22], s6[25], cospi_16_64);
312
313  s7[23] = sub_multiply_shift_and_narrow_s32_dual(s6[24], s6[23], cospi_16_64);
314  s7[24] = add_multiply_shift_and_narrow_s32_dual(s6[23], s6[24], cospi_16_64);
315
316  // final stage
317  s8[0] = highbd_idct_add_dual(s7[0], s6[31]);
318  s8[1] = highbd_idct_add_dual(s7[1], s6[30]);
319  s8[2] = highbd_idct_add_dual(s7[2], s6[29]);
320  s8[3] = highbd_idct_add_dual(s7[3], s6[28]);
321  s8[4] = highbd_idct_add_dual(s7[4], s7[27]);
322  s8[5] = highbd_idct_add_dual(s7[5], s7[26]);
323  s8[6] = highbd_idct_add_dual(s7[6], s7[25]);
324  s8[7] = highbd_idct_add_dual(s7[7], s7[24]);
325  s8[8] = highbd_idct_add_dual(s7[8], s7[23]);
326  s8[9] = highbd_idct_add_dual(s7[9], s7[22]);
327  s8[10] = highbd_idct_add_dual(s7[10], s7[21]);
328  s8[11] = highbd_idct_add_dual(s7[11], s7[20]);
329  s8[12] = highbd_idct_add_dual(s7[12], s6[19]);
330  s8[13] = highbd_idct_add_dual(s7[13], s6[18]);
331  s8[14] = highbd_idct_add_dual(s7[14], s6[17]);
332  s8[15] = highbd_idct_add_dual(s7[15], s6[16]);
333  s8[16] = highbd_idct_sub_dual(s7[15], s6[16]);
334  s8[17] = highbd_idct_sub_dual(s7[14], s6[17]);
335  s8[18] = highbd_idct_sub_dual(s7[13], s6[18]);
336  s8[19] = highbd_idct_sub_dual(s7[12], s6[19]);
337  s8[20] = highbd_idct_sub_dual(s7[11], s7[20]);
338  s8[21] = highbd_idct_sub_dual(s7[10], s7[21]);
339  s8[22] = highbd_idct_sub_dual(s7[9], s7[22]);
340  s8[23] = highbd_idct_sub_dual(s7[8], s7[23]);
341  s8[24] = highbd_idct_sub_dual(s7[7], s7[24]);
342  s8[25] = highbd_idct_sub_dual(s7[6], s7[25]);
343  s8[26] = highbd_idct_sub_dual(s7[5], s7[26]);
344  s8[27] = highbd_idct_sub_dual(s7[4], s7[27]);
345  s8[28] = highbd_idct_sub_dual(s7[3], s6[28]);
346  s8[29] = highbd_idct_sub_dual(s7[2], s6[29]);
347  s8[30] = highbd_idct_sub_dual(s7[1], s6[30]);
348  s8[31] = highbd_idct_sub_dual(s7[0], s6[31]);
349
350  vst1q_s32(output + 0, s8[0].val[0]);
351  vst1q_s32(output + 4, s8[0].val[1]);
352  output += 16;
353  vst1q_s32(output + 0, s8[1].val[0]);
354  vst1q_s32(output + 4, s8[1].val[1]);
355  output += 16;
356  vst1q_s32(output + 0, s8[2].val[0]);
357  vst1q_s32(output + 4, s8[2].val[1]);
358  output += 16;
359  vst1q_s32(output + 0, s8[3].val[0]);
360  vst1q_s32(output + 4, s8[3].val[1]);
361  output += 16;
362  vst1q_s32(output + 0, s8[4].val[0]);
363  vst1q_s32(output + 4, s8[4].val[1]);
364  output += 16;
365  vst1q_s32(output + 0, s8[5].val[0]);
366  vst1q_s32(output + 4, s8[5].val[1]);
367  output += 16;
368  vst1q_s32(output + 0, s8[6].val[0]);
369  vst1q_s32(output + 4, s8[6].val[1]);
370  output += 16;
371  vst1q_s32(output + 0, s8[7].val[0]);
372  vst1q_s32(output + 4, s8[7].val[1]);
373  output += 16;
374
375  vst1q_s32(output + 0, s8[8].val[0]);
376  vst1q_s32(output + 4, s8[8].val[1]);
377  output += 16;
378  vst1q_s32(output + 0, s8[9].val[0]);
379  vst1q_s32(output + 4, s8[9].val[1]);
380  output += 16;
381  vst1q_s32(output + 0, s8[10].val[0]);
382  vst1q_s32(output + 4, s8[10].val[1]);
383  output += 16;
384  vst1q_s32(output + 0, s8[11].val[0]);
385  vst1q_s32(output + 4, s8[11].val[1]);
386  output += 16;
387  vst1q_s32(output + 0, s8[12].val[0]);
388  vst1q_s32(output + 4, s8[12].val[1]);
389  output += 16;
390  vst1q_s32(output + 0, s8[13].val[0]);
391  vst1q_s32(output + 4, s8[13].val[1]);
392  output += 16;
393  vst1q_s32(output + 0, s8[14].val[0]);
394  vst1q_s32(output + 4, s8[14].val[1]);
395  output += 16;
396  vst1q_s32(output + 0, s8[15].val[0]);
397  vst1q_s32(output + 4, s8[15].val[1]);
398  output += 16;
399
400  vst1q_s32(output + 0, s8[16].val[0]);
401  vst1q_s32(output + 4, s8[16].val[1]);
402  output += 16;
403  vst1q_s32(output + 0, s8[17].val[0]);
404  vst1q_s32(output + 4, s8[17].val[1]);
405  output += 16;
406  vst1q_s32(output + 0, s8[18].val[0]);
407  vst1q_s32(output + 4, s8[18].val[1]);
408  output += 16;
409  vst1q_s32(output + 0, s8[19].val[0]);
410  vst1q_s32(output + 4, s8[19].val[1]);
411  output += 16;
412  vst1q_s32(output + 0, s8[20].val[0]);
413  vst1q_s32(output + 4, s8[20].val[1]);
414  output += 16;
415  vst1q_s32(output + 0, s8[21].val[0]);
416  vst1q_s32(output + 4, s8[21].val[1]);
417  output += 16;
418  vst1q_s32(output + 0, s8[22].val[0]);
419  vst1q_s32(output + 4, s8[22].val[1]);
420  output += 16;
421  vst1q_s32(output + 0, s8[23].val[0]);
422  vst1q_s32(output + 4, s8[23].val[1]);
423  output += 16;
424
425  vst1q_s32(output + 0, s8[24].val[0]);
426  vst1q_s32(output + 4, s8[24].val[1]);
427  output += 16;
428  vst1q_s32(output + 0, s8[25].val[0]);
429  vst1q_s32(output + 4, s8[25].val[1]);
430  output += 16;
431  vst1q_s32(output + 0, s8[26].val[0]);
432  vst1q_s32(output + 4, s8[26].val[1]);
433  output += 16;
434  vst1q_s32(output + 0, s8[27].val[0]);
435  vst1q_s32(output + 4, s8[27].val[1]);
436  output += 16;
437  vst1q_s32(output + 0, s8[28].val[0]);
438  vst1q_s32(output + 4, s8[28].val[1]);
439  output += 16;
440  vst1q_s32(output + 0, s8[29].val[0]);
441  vst1q_s32(output + 4, s8[29].val[1]);
442  output += 16;
443  vst1q_s32(output + 0, s8[30].val[0]);
444  vst1q_s32(output + 4, s8[30].val[1]);
445  output += 16;
446  vst1q_s32(output + 0, s8[31].val[0]);
447  vst1q_s32(output + 4, s8[31].val[1]);
448}
449
450static void vpx_highbd_idct32_16_neon(const int32_t *const input,
451                                      uint16_t *const output, const int stride,
452                                      const int bd) {
453  int32x4x2_t in[16], s1[32], s2[32], s3[32], s4[32], s5[32], s6[32], s7[32],
454      out[32];
455
456  load_and_transpose_s32_8x8(input, 16, &in[0], &in[1], &in[2], &in[3], &in[4],
457                             &in[5], &in[6], &in[7]);
458
459  load_and_transpose_s32_8x8(input + 8, 16, &in[8], &in[9], &in[10], &in[11],
460                             &in[12], &in[13], &in[14], &in[15]);
461
462  // stage 1
463  s1[16] = multiply_shift_and_narrow_s32_dual(in[1], cospi_31_64);
464  s1[31] = multiply_shift_and_narrow_s32_dual(in[1], cospi_1_64);
465
466  s1[17] = multiply_shift_and_narrow_s32_dual(in[15], -cospi_17_64);
467  s1[30] = multiply_shift_and_narrow_s32_dual(in[15], cospi_15_64);
468
469  s1[18] = multiply_shift_and_narrow_s32_dual(in[9], cospi_23_64);
470  s1[29] = multiply_shift_and_narrow_s32_dual(in[9], cospi_9_64);
471
472  s1[19] = multiply_shift_and_narrow_s32_dual(in[7], -cospi_25_64);
473  s1[28] = multiply_shift_and_narrow_s32_dual(in[7], cospi_7_64);
474
475  s1[20] = multiply_shift_and_narrow_s32_dual(in[5], cospi_27_64);
476  s1[27] = multiply_shift_and_narrow_s32_dual(in[5], cospi_5_64);
477
478  s1[21] = multiply_shift_and_narrow_s32_dual(in[11], -cospi_21_64);
479  s1[26] = multiply_shift_and_narrow_s32_dual(in[11], cospi_11_64);
480
481  s1[22] = multiply_shift_and_narrow_s32_dual(in[13], cospi_19_64);
482  s1[25] = multiply_shift_and_narrow_s32_dual(in[13], cospi_13_64);
483
484  s1[23] = multiply_shift_and_narrow_s32_dual(in[3], -cospi_29_64);
485  s1[24] = multiply_shift_and_narrow_s32_dual(in[3], cospi_3_64);
486
487  // stage 2
488  s2[8] = multiply_shift_and_narrow_s32_dual(in[2], cospi_30_64);
489  s2[15] = multiply_shift_and_narrow_s32_dual(in[2], cospi_2_64);
490
491  s2[9] = multiply_shift_and_narrow_s32_dual(in[14], -cospi_18_64);
492  s2[14] = multiply_shift_and_narrow_s32_dual(in[14], cospi_14_64);
493
494  s2[10] = multiply_shift_and_narrow_s32_dual(in[10], cospi_22_64);
495  s2[13] = multiply_shift_and_narrow_s32_dual(in[10], cospi_10_64);
496
497  s2[11] = multiply_shift_and_narrow_s32_dual(in[6], -cospi_26_64);
498  s2[12] = multiply_shift_and_narrow_s32_dual(in[6], cospi_6_64);
499
500  s2[16] = highbd_idct_add_dual(s1[16], s1[17]);
501  s2[17] = highbd_idct_sub_dual(s1[16], s1[17]);
502  s2[18] = highbd_idct_sub_dual(s1[19], s1[18]);
503  s2[19] = highbd_idct_add_dual(s1[18], s1[19]);
504  s2[20] = highbd_idct_add_dual(s1[20], s1[21]);
505  s2[21] = highbd_idct_sub_dual(s1[20], s1[21]);
506  s2[22] = highbd_idct_sub_dual(s1[23], s1[22]);
507  s2[23] = highbd_idct_add_dual(s1[22], s1[23]);
508  s2[24] = highbd_idct_add_dual(s1[24], s1[25]);
509  s2[25] = highbd_idct_sub_dual(s1[24], s1[25]);
510  s2[26] = highbd_idct_sub_dual(s1[27], s1[26]);
511  s2[27] = highbd_idct_add_dual(s1[26], s1[27]);
512  s2[28] = highbd_idct_add_dual(s1[28], s1[29]);
513  s2[29] = highbd_idct_sub_dual(s1[28], s1[29]);
514  s2[30] = highbd_idct_sub_dual(s1[31], s1[30]);
515  s2[31] = highbd_idct_add_dual(s1[30], s1[31]);
516
517  // stage 3
518  s3[4] = multiply_shift_and_narrow_s32_dual(in[4], cospi_28_64);
519  s3[7] = multiply_shift_and_narrow_s32_dual(in[4], cospi_4_64);
520
521  s3[5] = multiply_shift_and_narrow_s32_dual(in[12], -cospi_20_64);
522  s3[6] = multiply_shift_and_narrow_s32_dual(in[12], cospi_12_64);
523
524  s3[8] = highbd_idct_add_dual(s2[8], s2[9]);
525  s3[9] = highbd_idct_sub_dual(s2[8], s2[9]);
526  s3[10] = highbd_idct_sub_dual(s2[11], s2[10]);
527  s3[11] = highbd_idct_add_dual(s2[10], s2[11]);
528  s3[12] = highbd_idct_add_dual(s2[12], s2[13]);
529  s3[13] = highbd_idct_sub_dual(s2[12], s2[13]);
530  s3[14] = highbd_idct_sub_dual(s2[15], s2[14]);
531  s3[15] = highbd_idct_add_dual(s2[14], s2[15]);
532
533  s3[17] = multiply_accumulate_shift_and_narrow_s32_dual(s2[17], -cospi_4_64,
534                                                         s2[30], cospi_28_64);
535  s3[30] = multiply_accumulate_shift_and_narrow_s32_dual(s2[17], cospi_28_64,
536                                                         s2[30], cospi_4_64);
537
538  s3[18] = multiply_accumulate_shift_and_narrow_s32_dual(s2[18], -cospi_28_64,
539                                                         s2[29], -cospi_4_64);
540  s3[29] = multiply_accumulate_shift_and_narrow_s32_dual(s2[18], -cospi_4_64,
541                                                         s2[29], cospi_28_64);
542
543  s3[21] = multiply_accumulate_shift_and_narrow_s32_dual(s2[21], -cospi_20_64,
544                                                         s2[26], cospi_12_64);
545  s3[26] = multiply_accumulate_shift_and_narrow_s32_dual(s2[21], cospi_12_64,
546                                                         s2[26], cospi_20_64);
547
548  s3[22] = multiply_accumulate_shift_and_narrow_s32_dual(s2[22], -cospi_12_64,
549                                                         s2[25], -cospi_20_64);
550  s3[25] = multiply_accumulate_shift_and_narrow_s32_dual(s2[22], -cospi_20_64,
551                                                         s2[25], cospi_12_64);
552
553  // stage 4
554  s4[0] = multiply_shift_and_narrow_s32_dual(in[0], cospi_16_64);
555  s4[2] = multiply_shift_and_narrow_s32_dual(in[8], cospi_24_64);
556  s4[3] = multiply_shift_and_narrow_s32_dual(in[8], cospi_8_64);
557
558  s4[4] = highbd_idct_add_dual(s3[4], s3[5]);
559  s4[5] = highbd_idct_sub_dual(s3[4], s3[5]);
560  s4[6] = highbd_idct_sub_dual(s3[7], s3[6]);
561  s4[7] = highbd_idct_add_dual(s3[6], s3[7]);
562
563  s4[9] = multiply_accumulate_shift_and_narrow_s32_dual(s3[9], -cospi_8_64,
564                                                        s3[14], cospi_24_64);
565  s4[14] = multiply_accumulate_shift_and_narrow_s32_dual(s3[9], cospi_24_64,
566                                                         s3[14], cospi_8_64);
567
568  s4[10] = multiply_accumulate_shift_and_narrow_s32_dual(s3[10], -cospi_24_64,
569                                                         s3[13], -cospi_8_64);
570  s4[13] = multiply_accumulate_shift_and_narrow_s32_dual(s3[10], -cospi_8_64,
571                                                         s3[13], cospi_24_64);
572
573  s4[16] = highbd_idct_add_dual(s2[16], s2[19]);
574  s4[17] = highbd_idct_add_dual(s3[17], s3[18]);
575  s4[18] = highbd_idct_sub_dual(s3[17], s3[18]);
576  s4[19] = highbd_idct_sub_dual(s2[16], s2[19]);
577  s4[20] = highbd_idct_sub_dual(s2[23], s2[20]);
578  s4[21] = highbd_idct_sub_dual(s3[22], s3[21]);
579  s4[22] = highbd_idct_add_dual(s3[21], s3[22]);
580  s4[23] = highbd_idct_add_dual(s2[20], s2[23]);
581  s4[24] = highbd_idct_add_dual(s2[24], s2[27]);
582  s4[25] = highbd_idct_add_dual(s3[25], s3[26]);
583  s4[26] = highbd_idct_sub_dual(s3[25], s3[26]);
584  s4[27] = highbd_idct_sub_dual(s2[24], s2[27]);
585  s4[28] = highbd_idct_sub_dual(s2[31], s2[28]);
586  s4[29] = highbd_idct_sub_dual(s3[30], s3[29]);
587  s4[30] = highbd_idct_add_dual(s3[29], s3[30]);
588  s4[31] = highbd_idct_add_dual(s2[28], s2[31]);
589
590  // stage 5
591  s5[0] = highbd_idct_add_dual(s4[0], s4[3]);
592  s5[1] = highbd_idct_add_dual(s4[0], s4[2]);
593  s5[2] = highbd_idct_sub_dual(s4[0], s4[2]);
594  s5[3] = highbd_idct_sub_dual(s4[0], s4[3]);
595
596  s5[5] = sub_multiply_shift_and_narrow_s32_dual(s4[6], s4[5], cospi_16_64);
597  s5[6] = add_multiply_shift_and_narrow_s32_dual(s4[5], s4[6], cospi_16_64);
598
599  s5[8] = highbd_idct_add_dual(s3[8], s3[11]);
600  s5[9] = highbd_idct_add_dual(s4[9], s4[10]);
601  s5[10] = highbd_idct_sub_dual(s4[9], s4[10]);
602  s5[11] = highbd_idct_sub_dual(s3[8], s3[11]);
603  s5[12] = highbd_idct_sub_dual(s3[15], s3[12]);
604  s5[13] = highbd_idct_sub_dual(s4[14], s4[13]);
605  s5[14] = highbd_idct_add_dual(s4[13], s4[14]);
606  s5[15] = highbd_idct_add_dual(s3[15], s3[12]);
607
608  s5[18] = multiply_accumulate_shift_and_narrow_s32_dual(s4[18], -cospi_8_64,
609                                                         s4[29], cospi_24_64);
610  s5[29] = multiply_accumulate_shift_and_narrow_s32_dual(s4[18], cospi_24_64,
611                                                         s4[29], cospi_8_64);
612
613  s5[19] = multiply_accumulate_shift_and_narrow_s32_dual(s4[19], -cospi_8_64,
614                                                         s4[28], cospi_24_64);
615  s5[28] = multiply_accumulate_shift_and_narrow_s32_dual(s4[19], cospi_24_64,
616                                                         s4[28], cospi_8_64);
617
618  s5[20] = multiply_accumulate_shift_and_narrow_s32_dual(s4[20], -cospi_24_64,
619                                                         s4[27], -cospi_8_64);
620  s5[27] = multiply_accumulate_shift_and_narrow_s32_dual(s4[20], -cospi_8_64,
621                                                         s4[27], cospi_24_64);
622
623  s5[21] = multiply_accumulate_shift_and_narrow_s32_dual(s4[21], -cospi_24_64,
624                                                         s4[26], -cospi_8_64);
625  s5[26] = multiply_accumulate_shift_and_narrow_s32_dual(s4[21], -cospi_8_64,
626                                                         s4[26], cospi_24_64);
627
628  // stage 6
629  s6[0] = highbd_idct_add_dual(s5[0], s4[7]);
630  s6[1] = highbd_idct_add_dual(s5[1], s5[6]);
631  s6[2] = highbd_idct_add_dual(s5[2], s5[5]);
632  s6[3] = highbd_idct_add_dual(s5[3], s4[4]);
633  s6[4] = highbd_idct_sub_dual(s5[3], s4[4]);
634  s6[5] = highbd_idct_sub_dual(s5[2], s5[5]);
635  s6[6] = highbd_idct_sub_dual(s5[1], s5[6]);
636  s6[7] = highbd_idct_sub_dual(s5[0], s4[7]);
637
638  s6[10] = sub_multiply_shift_and_narrow_s32_dual(s5[13], s5[10], cospi_16_64);
639  s6[13] = add_multiply_shift_and_narrow_s32_dual(s5[10], s5[13], cospi_16_64);
640
641  s6[11] = sub_multiply_shift_and_narrow_s32_dual(s5[12], s5[11], cospi_16_64);
642  s6[12] = add_multiply_shift_and_narrow_s32_dual(s5[11], s5[12], cospi_16_64);
643
644  s6[16] = highbd_idct_add_dual(s4[16], s4[23]);
645  s6[17] = highbd_idct_add_dual(s4[17], s4[22]);
646  s6[18] = highbd_idct_add_dual(s5[18], s5[21]);
647  s6[19] = highbd_idct_add_dual(s5[19], s5[20]);
648  s6[20] = highbd_idct_sub_dual(s5[19], s5[20]);
649  s6[21] = highbd_idct_sub_dual(s5[18], s5[21]);
650  s6[22] = highbd_idct_sub_dual(s4[17], s4[22]);
651  s6[23] = highbd_idct_sub_dual(s4[16], s4[23]);
652  s6[24] = highbd_idct_sub_dual(s4[31], s4[24]);
653  s6[25] = highbd_idct_sub_dual(s4[30], s4[25]);
654  s6[26] = highbd_idct_sub_dual(s5[29], s5[26]);
655  s6[27] = highbd_idct_sub_dual(s5[28], s5[27]);
656  s6[28] = highbd_idct_add_dual(s5[27], s5[28]);
657  s6[29] = highbd_idct_add_dual(s5[26], s5[29]);
658  s6[30] = highbd_idct_add_dual(s4[25], s4[30]);
659  s6[31] = highbd_idct_add_dual(s4[24], s4[31]);
660
661  // stage 7
662  s7[0] = highbd_idct_add_dual(s6[0], s5[15]);
663  s7[1] = highbd_idct_add_dual(s6[1], s5[14]);
664  s7[2] = highbd_idct_add_dual(s6[2], s6[13]);
665  s7[3] = highbd_idct_add_dual(s6[3], s6[12]);
666  s7[4] = highbd_idct_add_dual(s6[4], s6[11]);
667  s7[5] = highbd_idct_add_dual(s6[5], s6[10]);
668  s7[6] = highbd_idct_add_dual(s6[6], s5[9]);
669  s7[7] = highbd_idct_add_dual(s6[7], s5[8]);
670  s7[8] = highbd_idct_sub_dual(s6[7], s5[8]);
671  s7[9] = highbd_idct_sub_dual(s6[6], s5[9]);
672  s7[10] = highbd_idct_sub_dual(s6[5], s6[10]);
673  s7[11] = highbd_idct_sub_dual(s6[4], s6[11]);
674  s7[12] = highbd_idct_sub_dual(s6[3], s6[12]);
675  s7[13] = highbd_idct_sub_dual(s6[2], s6[13]);
676  s7[14] = highbd_idct_sub_dual(s6[1], s5[14]);
677  s7[15] = highbd_idct_sub_dual(s6[0], s5[15]);
678
679  s7[20] = sub_multiply_shift_and_narrow_s32_dual(s6[27], s6[20], cospi_16_64);
680  s7[27] = add_multiply_shift_and_narrow_s32_dual(s6[20], s6[27], cospi_16_64);
681
682  s7[21] = sub_multiply_shift_and_narrow_s32_dual(s6[26], s6[21], cospi_16_64);
683  s7[26] = add_multiply_shift_and_narrow_s32_dual(s6[21], s6[26], cospi_16_64);
684
685  s7[22] = sub_multiply_shift_and_narrow_s32_dual(s6[25], s6[22], cospi_16_64);
686  s7[25] = add_multiply_shift_and_narrow_s32_dual(s6[22], s6[25], cospi_16_64);
687
688  s7[23] = sub_multiply_shift_and_narrow_s32_dual(s6[24], s6[23], cospi_16_64);
689  s7[24] = add_multiply_shift_and_narrow_s32_dual(s6[23], s6[24], cospi_16_64);
690
691  // final stage
692  out[0] = highbd_idct_add_dual(s7[0], s6[31]);
693  out[1] = highbd_idct_add_dual(s7[1], s6[30]);
694  out[2] = highbd_idct_add_dual(s7[2], s6[29]);
695  out[3] = highbd_idct_add_dual(s7[3], s6[28]);
696  out[4] = highbd_idct_add_dual(s7[4], s7[27]);
697  out[5] = highbd_idct_add_dual(s7[5], s7[26]);
698  out[6] = highbd_idct_add_dual(s7[6], s7[25]);
699  out[7] = highbd_idct_add_dual(s7[7], s7[24]);
700  out[8] = highbd_idct_add_dual(s7[8], s7[23]);
701  out[9] = highbd_idct_add_dual(s7[9], s7[22]);
702  out[10] = highbd_idct_add_dual(s7[10], s7[21]);
703  out[11] = highbd_idct_add_dual(s7[11], s7[20]);
704  out[12] = highbd_idct_add_dual(s7[12], s6[19]);
705  out[13] = highbd_idct_add_dual(s7[13], s6[18]);
706  out[14] = highbd_idct_add_dual(s7[14], s6[17]);
707  out[15] = highbd_idct_add_dual(s7[15], s6[16]);
708  out[16] = highbd_idct_sub_dual(s7[15], s6[16]);
709  out[17] = highbd_idct_sub_dual(s7[14], s6[17]);
710  out[18] = highbd_idct_sub_dual(s7[13], s6[18]);
711  out[19] = highbd_idct_sub_dual(s7[12], s6[19]);
712  out[20] = highbd_idct_sub_dual(s7[11], s7[20]);
713  out[21] = highbd_idct_sub_dual(s7[10], s7[21]);
714  out[22] = highbd_idct_sub_dual(s7[9], s7[22]);
715  out[23] = highbd_idct_sub_dual(s7[8], s7[23]);
716  out[24] = highbd_idct_sub_dual(s7[7], s7[24]);
717  out[25] = highbd_idct_sub_dual(s7[6], s7[25]);
718  out[26] = highbd_idct_sub_dual(s7[5], s7[26]);
719  out[27] = highbd_idct_sub_dual(s7[4], s7[27]);
720  out[28] = highbd_idct_sub_dual(s7[3], s6[28]);
721  out[29] = highbd_idct_sub_dual(s7[2], s6[29]);
722  out[30] = highbd_idct_sub_dual(s7[1], s6[30]);
723  out[31] = highbd_idct_sub_dual(s7[0], s6[31]);
724
725  highbd_idct16x16_add_store(out, output, stride, bd);
726  highbd_idct16x16_add_store(out + 16, output + 16 * stride, stride, bd);
727}
728
729void vpx_highbd_idct32x32_135_add_neon(const tran_low_t *input, uint16_t *dest,
730                                       int stride, int bd) {
731  int i;
732
733  if (bd == 8) {
734    int16_t temp[32 * 16];
735    int16_t *t = temp;
736    vpx_idct32_12_neon(input, temp);
737    vpx_idct32_12_neon(input + 32 * 8, temp + 8);
738
739    for (i = 0; i < 32; i += 8) {
740      vpx_idct32_16_neon(t, dest, stride, 1);
741      t += (16 * 8);
742      dest += 8;
743    }
744  } else {
745    int32_t temp[32 * 16];
746    int32_t *t = temp;
747    vpx_highbd_idct32_12_neon(input, temp);
748    vpx_highbd_idct32_12_neon(input + 32 * 8, temp + 8);
749
750    for (i = 0; i < 32; i += 8) {
751      vpx_highbd_idct32_16_neon(t, dest, stride, bd);
752      t += (16 * 8);
753      dest += 8;
754    }
755  }
756}
757