1/*
2 *  Copyright (c) 2014 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
13void vp8_short_fdct4x4_neon(int16_t *input, int16_t *output, int pitch) {
14  int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
15  int16x4_t d16s16, d17s16, d26s16, dEmptys16;
16  uint16x4_t d4u16;
17  int16x8_t q0s16, q1s16;
18  int32x4_t q9s32, q10s32, q11s32, q12s32;
19  int16x4x2_t v2tmp0, v2tmp1;
20  int32x2x2_t v2tmp2, v2tmp3;
21
22  d16s16 = vdup_n_s16(5352);
23  d17s16 = vdup_n_s16(2217);
24  q9s32 = vdupq_n_s32(14500);
25  q10s32 = vdupq_n_s32(7500);
26  q11s32 = vdupq_n_s32(12000);
27  q12s32 = vdupq_n_s32(51000);
28
29  // Part one
30  pitch >>= 1;
31  d0s16 = vld1_s16(input);
32  input += pitch;
33  d1s16 = vld1_s16(input);
34  input += pitch;
35  d2s16 = vld1_s16(input);
36  input += pitch;
37  d3s16 = vld1_s16(input);
38
39  v2tmp2 = vtrn_s32(vreinterpret_s32_s16(d0s16), vreinterpret_s32_s16(d2s16));
40  v2tmp3 = vtrn_s32(vreinterpret_s32_s16(d1s16), vreinterpret_s32_s16(d3s16));
41  v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]),   // d0
42                    vreinterpret_s16_s32(v2tmp3.val[0]));  // d1
43  v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]),   // d2
44                    vreinterpret_s16_s32(v2tmp3.val[1]));  // d3
45
46  d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[1]);
47  d5s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[0]);
48  d6s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[0]);
49  d7s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[1]);
50
51  d4s16 = vshl_n_s16(d4s16, 3);
52  d5s16 = vshl_n_s16(d5s16, 3);
53  d6s16 = vshl_n_s16(d6s16, 3);
54  d7s16 = vshl_n_s16(d7s16, 3);
55
56  d0s16 = vadd_s16(d4s16, d5s16);
57  d2s16 = vsub_s16(d4s16, d5s16);
58
59  q9s32 = vmlal_s16(q9s32, d7s16, d16s16);
60  q10s32 = vmlal_s16(q10s32, d7s16, d17s16);
61  q9s32 = vmlal_s16(q9s32, d6s16, d17s16);
62  q10s32 = vmlsl_s16(q10s32, d6s16, d16s16);
63
64  d1s16 = vshrn_n_s32(q9s32, 12);
65  d3s16 = vshrn_n_s32(q10s32, 12);
66
67  // Part two
68  v2tmp2 = vtrn_s32(vreinterpret_s32_s16(d0s16), vreinterpret_s32_s16(d2s16));
69  v2tmp3 = vtrn_s32(vreinterpret_s32_s16(d1s16), vreinterpret_s32_s16(d3s16));
70  v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]),   // d0
71                    vreinterpret_s16_s32(v2tmp3.val[0]));  // d1
72  v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]),   // d2
73                    vreinterpret_s16_s32(v2tmp3.val[1]));  // d3
74
75  d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[1]);
76  d5s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[0]);
77  d6s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[0]);
78  d7s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[1]);
79
80  d26s16 = vdup_n_s16(7);
81  d4s16 = vadd_s16(d4s16, d26s16);
82
83  d0s16 = vadd_s16(d4s16, d5s16);
84  d2s16 = vsub_s16(d4s16, d5s16);
85
86  q11s32 = vmlal_s16(q11s32, d7s16, d16s16);
87  q12s32 = vmlal_s16(q12s32, d7s16, d17s16);
88
89  dEmptys16 = vdup_n_s16(0);
90  d4u16 = vceq_s16(d7s16, dEmptys16);
91
92  d0s16 = vshr_n_s16(d0s16, 4);
93  d2s16 = vshr_n_s16(d2s16, 4);
94
95  q11s32 = vmlal_s16(q11s32, d6s16, d17s16);
96  q12s32 = vmlsl_s16(q12s32, d6s16, d16s16);
97
98  d4u16 = vmvn_u16(d4u16);
99  d1s16 = vshrn_n_s32(q11s32, 16);
100  d1s16 = vsub_s16(d1s16, vreinterpret_s16_u16(d4u16));
101  d3s16 = vshrn_n_s32(q12s32, 16);
102
103  q0s16 = vcombine_s16(d0s16, d1s16);
104  q1s16 = vcombine_s16(d2s16, d3s16);
105
106  vst1q_s16(output, q0s16);
107  vst1q_s16(output + 8, q1s16);
108  return;
109}
110
111void vp8_short_fdct8x4_neon(int16_t *input, int16_t *output, int pitch) {
112  int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
113  int16x4_t d16s16, d17s16, d26s16, d27s16, d28s16, d29s16;
114  uint16x4_t d28u16, d29u16;
115  uint16x8_t q14u16;
116  int16x8_t q0s16, q1s16, q2s16, q3s16;
117  int16x8_t q11s16, q12s16, q13s16, q14s16, q15s16, qEmptys16;
118  int32x4_t q9s32, q10s32, q11s32, q12s32;
119  int16x8x2_t v2tmp0, v2tmp1;
120  int32x4x2_t v2tmp2, v2tmp3;
121
122  d16s16 = vdup_n_s16(5352);
123  d17s16 = vdup_n_s16(2217);
124  q9s32 = vdupq_n_s32(14500);
125  q10s32 = vdupq_n_s32(7500);
126
127  // Part one
128  pitch >>= 1;
129  q0s16 = vld1q_s16(input);
130  input += pitch;
131  q1s16 = vld1q_s16(input);
132  input += pitch;
133  q2s16 = vld1q_s16(input);
134  input += pitch;
135  q3s16 = vld1q_s16(input);
136
137  v2tmp2 =
138      vtrnq_s32(vreinterpretq_s32_s16(q0s16), vreinterpretq_s32_s16(q2s16));
139  v2tmp3 =
140      vtrnq_s32(vreinterpretq_s32_s16(q1s16), vreinterpretq_s32_s16(q3s16));
141  v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]),   // q0
142                     vreinterpretq_s16_s32(v2tmp3.val[0]));  // q1
143  v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]),   // q2
144                     vreinterpretq_s16_s32(v2tmp3.val[1]));  // q3
145
146  q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]);
147  q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]);
148  q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]);
149  q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]);
150
151  q11s16 = vshlq_n_s16(q11s16, 3);
152  q12s16 = vshlq_n_s16(q12s16, 3);
153  q13s16 = vshlq_n_s16(q13s16, 3);
154  q14s16 = vshlq_n_s16(q14s16, 3);
155
156  q0s16 = vaddq_s16(q11s16, q12s16);
157  q2s16 = vsubq_s16(q11s16, q12s16);
158
159  q11s32 = q9s32;
160  q12s32 = q10s32;
161
162  d26s16 = vget_low_s16(q13s16);
163  d27s16 = vget_high_s16(q13s16);
164  d28s16 = vget_low_s16(q14s16);
165  d29s16 = vget_high_s16(q14s16);
166
167  q9s32 = vmlal_s16(q9s32, d28s16, d16s16);
168  q10s32 = vmlal_s16(q10s32, d28s16, d17s16);
169  q11s32 = vmlal_s16(q11s32, d29s16, d16s16);
170  q12s32 = vmlal_s16(q12s32, d29s16, d17s16);
171
172  q9s32 = vmlal_s16(q9s32, d26s16, d17s16);
173  q10s32 = vmlsl_s16(q10s32, d26s16, d16s16);
174  q11s32 = vmlal_s16(q11s32, d27s16, d17s16);
175  q12s32 = vmlsl_s16(q12s32, d27s16, d16s16);
176
177  d2s16 = vshrn_n_s32(q9s32, 12);
178  d6s16 = vshrn_n_s32(q10s32, 12);
179  d3s16 = vshrn_n_s32(q11s32, 12);
180  d7s16 = vshrn_n_s32(q12s32, 12);
181  q1s16 = vcombine_s16(d2s16, d3s16);
182  q3s16 = vcombine_s16(d6s16, d7s16);
183
184  // Part two
185  q9s32 = vdupq_n_s32(12000);
186  q10s32 = vdupq_n_s32(51000);
187
188  v2tmp2 =
189      vtrnq_s32(vreinterpretq_s32_s16(q0s16), vreinterpretq_s32_s16(q2s16));
190  v2tmp3 =
191      vtrnq_s32(vreinterpretq_s32_s16(q1s16), vreinterpretq_s32_s16(q3s16));
192  v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]),   // q0
193                     vreinterpretq_s16_s32(v2tmp3.val[0]));  // q1
194  v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]),   // q2
195                     vreinterpretq_s16_s32(v2tmp3.val[1]));  // q3
196
197  q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]);
198  q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]);
199  q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]);
200  q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]);
201
202  q15s16 = vdupq_n_s16(7);
203  q11s16 = vaddq_s16(q11s16, q15s16);
204  q0s16 = vaddq_s16(q11s16, q12s16);
205  q1s16 = vsubq_s16(q11s16, q12s16);
206
207  q11s32 = q9s32;
208  q12s32 = q10s32;
209
210  d0s16 = vget_low_s16(q0s16);
211  d1s16 = vget_high_s16(q0s16);
212  d2s16 = vget_low_s16(q1s16);
213  d3s16 = vget_high_s16(q1s16);
214
215  d0s16 = vshr_n_s16(d0s16, 4);
216  d4s16 = vshr_n_s16(d1s16, 4);
217  d2s16 = vshr_n_s16(d2s16, 4);
218  d6s16 = vshr_n_s16(d3s16, 4);
219
220  d26s16 = vget_low_s16(q13s16);
221  d27s16 = vget_high_s16(q13s16);
222  d28s16 = vget_low_s16(q14s16);
223  d29s16 = vget_high_s16(q14s16);
224
225  q9s32 = vmlal_s16(q9s32, d28s16, d16s16);
226  q10s32 = vmlal_s16(q10s32, d28s16, d17s16);
227  q11s32 = vmlal_s16(q11s32, d29s16, d16s16);
228  q12s32 = vmlal_s16(q12s32, d29s16, d17s16);
229
230  q9s32 = vmlal_s16(q9s32, d26s16, d17s16);
231  q10s32 = vmlsl_s16(q10s32, d26s16, d16s16);
232  q11s32 = vmlal_s16(q11s32, d27s16, d17s16);
233  q12s32 = vmlsl_s16(q12s32, d27s16, d16s16);
234
235  d1s16 = vshrn_n_s32(q9s32, 16);
236  d3s16 = vshrn_n_s32(q10s32, 16);
237  d5s16 = vshrn_n_s32(q11s32, 16);
238  d7s16 = vshrn_n_s32(q12s32, 16);
239
240  qEmptys16 = vdupq_n_s16(0);
241  q14u16 = vceqq_s16(q14s16, qEmptys16);
242  q14u16 = vmvnq_u16(q14u16);
243
244  d28u16 = vget_low_u16(q14u16);
245  d29u16 = vget_high_u16(q14u16);
246  d1s16 = vsub_s16(d1s16, vreinterpret_s16_u16(d28u16));
247  d5s16 = vsub_s16(d5s16, vreinterpret_s16_u16(d29u16));
248
249  q0s16 = vcombine_s16(d0s16, d1s16);
250  q1s16 = vcombine_s16(d2s16, d3s16);
251  q2s16 = vcombine_s16(d4s16, d5s16);
252  q3s16 = vcombine_s16(d6s16, d7s16);
253
254  vst1q_s16(output, q0s16);
255  vst1q_s16(output + 8, q1s16);
256  vst1q_s16(output + 16, q2s16);
257  vst1q_s16(output + 24, q3s16);
258  return;
259}
260