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 
vp8_short_fdct4x4_neon(int16_t * input,int16_t * output,int pitch)13 void vp8_short_fdct4x4_neon(
14         int16_t *input,
15         int16_t *output,
16         int pitch) {
17     int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
18     int16x4_t d16s16, d17s16, d26s16, dEmptys16;
19     uint16x4_t d4u16;
20     int16x8_t q0s16, q1s16;
21     int32x4_t q9s32, q10s32, q11s32, q12s32;
22     int16x4x2_t v2tmp0, v2tmp1;
23     int32x2x2_t v2tmp2, v2tmp3;
24 
25     d16s16 = vdup_n_s16(5352);
26     d17s16 = vdup_n_s16(2217);
27     q9s32 = vdupq_n_s32(14500);
28     q10s32 = vdupq_n_s32(7500);
29     q11s32 = vdupq_n_s32(12000);
30     q12s32 = vdupq_n_s32(51000);
31 
32     // Part one
33     pitch >>= 1;
34     d0s16 = vld1_s16(input);
35     input += pitch;
36     d1s16 = vld1_s16(input);
37     input += pitch;
38     d2s16 = vld1_s16(input);
39     input += pitch;
40     d3s16 = vld1_s16(input);
41 
42     v2tmp2 = vtrn_s32(vreinterpret_s32_s16(d0s16),
43                       vreinterpret_s32_s16(d2s16));
44     v2tmp3 = vtrn_s32(vreinterpret_s32_s16(d1s16),
45                       vreinterpret_s32_s16(d3s16));
46     v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]),   // d0
47                       vreinterpret_s16_s32(v2tmp3.val[0]));  // d1
48     v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]),   // d2
49                       vreinterpret_s16_s32(v2tmp3.val[1]));  // d3
50 
51     d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[1]);
52     d5s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[0]);
53     d6s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[0]);
54     d7s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[1]);
55 
56     d4s16 = vshl_n_s16(d4s16, 3);
57     d5s16 = vshl_n_s16(d5s16, 3);
58     d6s16 = vshl_n_s16(d6s16, 3);
59     d7s16 = vshl_n_s16(d7s16, 3);
60 
61     d0s16 = vadd_s16(d4s16, d5s16);
62     d2s16 = vsub_s16(d4s16, d5s16);
63 
64     q9s32 = vmlal_s16(q9s32, d7s16, d16s16);
65     q10s32 = vmlal_s16(q10s32, d7s16, d17s16);
66     q9s32 = vmlal_s16(q9s32, d6s16, d17s16);
67     q10s32 = vmlsl_s16(q10s32, d6s16, d16s16);
68 
69     d1s16 = vshrn_n_s32(q9s32, 12);
70     d3s16 = vshrn_n_s32(q10s32, 12);
71 
72     // Part two
73     v2tmp2 = vtrn_s32(vreinterpret_s32_s16(d0s16),
74                       vreinterpret_s32_s16(d2s16));
75     v2tmp3 = vtrn_s32(vreinterpret_s32_s16(d1s16),
76                       vreinterpret_s32_s16(d3s16));
77     v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]),   // d0
78                       vreinterpret_s16_s32(v2tmp3.val[0]));  // d1
79     v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]),   // d2
80                       vreinterpret_s16_s32(v2tmp3.val[1]));  // d3
81 
82     d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[1]);
83     d5s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[0]);
84     d6s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[0]);
85     d7s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[1]);
86 
87     d26s16 = vdup_n_s16(7);
88     d4s16 = vadd_s16(d4s16, d26s16);
89 
90     d0s16 = vadd_s16(d4s16, d5s16);
91     d2s16 = vsub_s16(d4s16, d5s16);
92 
93     q11s32 = vmlal_s16(q11s32, d7s16, d16s16);
94     q12s32 = vmlal_s16(q12s32, d7s16, d17s16);
95 
96     dEmptys16 = vdup_n_s16(0);
97     d4u16 = vceq_s16(d7s16, dEmptys16);
98 
99     d0s16 = vshr_n_s16(d0s16, 4);
100     d2s16 = vshr_n_s16(d2s16, 4);
101 
102     q11s32 = vmlal_s16(q11s32, d6s16, d17s16);
103     q12s32 = vmlsl_s16(q12s32, d6s16, d16s16);
104 
105     d4u16 = vmvn_u16(d4u16);
106     d1s16 = vshrn_n_s32(q11s32, 16);
107     d1s16 = vsub_s16(d1s16, vreinterpret_s16_u16(d4u16));
108     d3s16 = vshrn_n_s32(q12s32, 16);
109 
110     q0s16 = vcombine_s16(d0s16, d1s16);
111     q1s16 = vcombine_s16(d2s16, d3s16);
112 
113     vst1q_s16(output, q0s16);
114     vst1q_s16(output + 8, q1s16);
115     return;
116 }
117 
vp8_short_fdct8x4_neon(int16_t * input,int16_t * output,int pitch)118 void vp8_short_fdct8x4_neon(
119         int16_t *input,
120         int16_t *output,
121         int pitch) {
122     int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
123     int16x4_t d16s16, d17s16, d26s16, d27s16, d28s16, d29s16;
124     uint16x4_t d28u16, d29u16;
125     uint16x8_t q14u16;
126     int16x8_t q0s16, q1s16, q2s16, q3s16;
127     int16x8_t q11s16, q12s16, q13s16, q14s16, q15s16, qEmptys16;
128     int32x4_t q9s32, q10s32, q11s32, q12s32;
129     int16x8x2_t v2tmp0, v2tmp1;
130     int32x4x2_t v2tmp2, v2tmp3;
131 
132     d16s16 = vdup_n_s16(5352);
133     d17s16 = vdup_n_s16(2217);
134     q9s32 = vdupq_n_s32(14500);
135     q10s32 = vdupq_n_s32(7500);
136 
137     // Part one
138     pitch >>= 1;
139     q0s16 = vld1q_s16(input);
140     input += pitch;
141     q1s16 = vld1q_s16(input);
142     input += pitch;
143     q2s16 = vld1q_s16(input);
144     input += pitch;
145     q3s16 = vld1q_s16(input);
146 
147     v2tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q0s16),
148                        vreinterpretq_s32_s16(q2s16));
149     v2tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q1s16),
150                        vreinterpretq_s32_s16(q3s16));
151     v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]),   // q0
152                        vreinterpretq_s16_s32(v2tmp3.val[0]));  // q1
153     v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]),   // q2
154                        vreinterpretq_s16_s32(v2tmp3.val[1]));  // q3
155 
156     q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]);
157     q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]);
158     q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]);
159     q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]);
160 
161     q11s16 = vshlq_n_s16(q11s16, 3);
162     q12s16 = vshlq_n_s16(q12s16, 3);
163     q13s16 = vshlq_n_s16(q13s16, 3);
164     q14s16 = vshlq_n_s16(q14s16, 3);
165 
166     q0s16 = vaddq_s16(q11s16, q12s16);
167     q2s16 = vsubq_s16(q11s16, q12s16);
168 
169     q11s32 = q9s32;
170     q12s32 = q10s32;
171 
172     d26s16 = vget_low_s16(q13s16);
173     d27s16 = vget_high_s16(q13s16);
174     d28s16 = vget_low_s16(q14s16);
175     d29s16 = vget_high_s16(q14s16);
176 
177     q9s32 = vmlal_s16(q9s32, d28s16, d16s16);
178     q10s32 = vmlal_s16(q10s32, d28s16, d17s16);
179     q11s32 = vmlal_s16(q11s32, d29s16, d16s16);
180     q12s32 = vmlal_s16(q12s32, d29s16, d17s16);
181 
182     q9s32 = vmlal_s16(q9s32, d26s16, d17s16);
183     q10s32 = vmlsl_s16(q10s32, d26s16, d16s16);
184     q11s32 = vmlal_s16(q11s32, d27s16, d17s16);
185     q12s32 = vmlsl_s16(q12s32, d27s16, d16s16);
186 
187     d2s16 = vshrn_n_s32(q9s32, 12);
188     d6s16 = vshrn_n_s32(q10s32, 12);
189     d3s16 = vshrn_n_s32(q11s32, 12);
190     d7s16 = vshrn_n_s32(q12s32, 12);
191     q1s16 = vcombine_s16(d2s16, d3s16);
192     q3s16 = vcombine_s16(d6s16, d7s16);
193 
194     // Part two
195     q9s32 = vdupq_n_s32(12000);
196     q10s32 = vdupq_n_s32(51000);
197 
198     v2tmp2 = vtrnq_s32(vreinterpretq_s32_s16(q0s16),
199                        vreinterpretq_s32_s16(q2s16));
200     v2tmp3 = vtrnq_s32(vreinterpretq_s32_s16(q1s16),
201                        vreinterpretq_s32_s16(q3s16));
202     v2tmp0 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[0]),   // q0
203                        vreinterpretq_s16_s32(v2tmp3.val[0]));  // q1
204     v2tmp1 = vtrnq_s16(vreinterpretq_s16_s32(v2tmp2.val[1]),   // q2
205                        vreinterpretq_s16_s32(v2tmp3.val[1]));  // q3
206 
207     q11s16 = vaddq_s16(v2tmp0.val[0], v2tmp1.val[1]);
208     q12s16 = vaddq_s16(v2tmp0.val[1], v2tmp1.val[0]);
209     q13s16 = vsubq_s16(v2tmp0.val[1], v2tmp1.val[0]);
210     q14s16 = vsubq_s16(v2tmp0.val[0], v2tmp1.val[1]);
211 
212     q15s16 = vdupq_n_s16(7);
213     q11s16 = vaddq_s16(q11s16, q15s16);
214     q0s16 = vaddq_s16(q11s16, q12s16);
215     q1s16 = vsubq_s16(q11s16, q12s16);
216 
217     q11s32 = q9s32;
218     q12s32 = q10s32;
219 
220     d0s16 = vget_low_s16(q0s16);
221     d1s16 = vget_high_s16(q0s16);
222     d2s16 = vget_low_s16(q1s16);
223     d3s16 = vget_high_s16(q1s16);
224 
225     d0s16 = vshr_n_s16(d0s16, 4);
226     d4s16 = vshr_n_s16(d1s16, 4);
227     d2s16 = vshr_n_s16(d2s16, 4);
228     d6s16 = vshr_n_s16(d3s16, 4);
229 
230     d26s16 = vget_low_s16(q13s16);
231     d27s16 = vget_high_s16(q13s16);
232     d28s16 = vget_low_s16(q14s16);
233     d29s16 = vget_high_s16(q14s16);
234 
235     q9s32 = vmlal_s16(q9s32, d28s16, d16s16);
236     q10s32 = vmlal_s16(q10s32, d28s16, d17s16);
237     q11s32 = vmlal_s16(q11s32, d29s16, d16s16);
238     q12s32 = vmlal_s16(q12s32, d29s16, d17s16);
239 
240     q9s32 = vmlal_s16(q9s32, d26s16, d17s16);
241     q10s32 = vmlsl_s16(q10s32, d26s16, d16s16);
242     q11s32 = vmlal_s16(q11s32, d27s16, d17s16);
243     q12s32 = vmlsl_s16(q12s32, d27s16, d16s16);
244 
245     d1s16 = vshrn_n_s32(q9s32, 16);
246     d3s16 = vshrn_n_s32(q10s32, 16);
247     d5s16 = vshrn_n_s32(q11s32, 16);
248     d7s16 = vshrn_n_s32(q12s32, 16);
249 
250     qEmptys16 = vdupq_n_s16(0);
251     q14u16 = vceqq_s16(q14s16, qEmptys16);
252     q14u16 = vmvnq_u16(q14u16);
253 
254     d28u16 = vget_low_u16(q14u16);
255     d29u16 = vget_high_u16(q14u16);
256     d1s16 = vsub_s16(d1s16, vreinterpret_s16_u16(d28u16));
257     d5s16 = vsub_s16(d5s16, vreinterpret_s16_u16(d29u16));
258 
259     q0s16 = vcombine_s16(d0s16, d1s16);
260     q1s16 = vcombine_s16(d2s16, d3s16);
261     q2s16 = vcombine_s16(d4s16, d5s16);
262     q3s16 = vcombine_s16(d6s16, d7s16);
263 
264     vst1q_s16(output, q0s16);
265     vst1q_s16(output + 8, q1s16);
266     vst1q_s16(output + 16, q2s16);
267     vst1q_s16(output + 24, q3s16);
268     return;
269 }
270