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_inv_walsh4x4_neon(int16_t * input,int16_t * mb_dqcoeff)13 void vp8_short_inv_walsh4x4_neon(
14         int16_t *input,
15         int16_t *mb_dqcoeff) {
16     int16x8_t q0s16, q1s16, q2s16, q3s16;
17     int16x4_t d4s16, d5s16, d6s16, d7s16;
18     int16x4x2_t v2tmp0, v2tmp1;
19     int32x2x2_t v2tmp2, v2tmp3;
20     int16x8_t qAdd3;
21 
22     q0s16 = vld1q_s16(input);
23     q1s16 = vld1q_s16(input + 8);
24 
25     // 1st for loop
26     d4s16 = vadd_s16(vget_low_s16(q0s16), vget_high_s16(q1s16));
27     d6s16 = vadd_s16(vget_high_s16(q0s16), vget_low_s16(q1s16));
28     d5s16 = vsub_s16(vget_low_s16(q0s16), vget_high_s16(q1s16));
29     d7s16 = vsub_s16(vget_high_s16(q0s16), vget_low_s16(q1s16));
30 
31     q2s16 = vcombine_s16(d4s16, d5s16);
32     q3s16 = vcombine_s16(d6s16, d7s16);
33 
34     q0s16 = vaddq_s16(q2s16, q3s16);
35     q1s16 = vsubq_s16(q2s16, q3s16);
36 
37     v2tmp2 = vtrn_s32(vreinterpret_s32_s16(vget_low_s16(q0s16)),
38                       vreinterpret_s32_s16(vget_low_s16(q1s16)));
39     v2tmp3 = vtrn_s32(vreinterpret_s32_s16(vget_high_s16(q0s16)),
40                       vreinterpret_s32_s16(vget_high_s16(q1s16)));
41     v2tmp0 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[0]),
42                       vreinterpret_s16_s32(v2tmp3.val[0]));
43     v2tmp1 = vtrn_s16(vreinterpret_s16_s32(v2tmp2.val[1]),
44                       vreinterpret_s16_s32(v2tmp3.val[1]));
45 
46     // 2nd for loop
47     d4s16 = vadd_s16(v2tmp0.val[0], v2tmp1.val[1]);
48     d6s16 = vadd_s16(v2tmp0.val[1], v2tmp1.val[0]);
49     d5s16 = vsub_s16(v2tmp0.val[0], v2tmp1.val[1]);
50     d7s16 = vsub_s16(v2tmp0.val[1], v2tmp1.val[0]);
51     q2s16 = vcombine_s16(d4s16, d5s16);
52     q3s16 = vcombine_s16(d6s16, d7s16);
53 
54     qAdd3 = vdupq_n_s16(3);
55 
56     q0s16 = vaddq_s16(q2s16, q3s16);
57     q1s16 = vsubq_s16(q2s16, q3s16);
58 
59     q0s16 = vaddq_s16(q0s16, qAdd3);
60     q1s16 = vaddq_s16(q1s16, qAdd3);
61 
62     q0s16 = vshrq_n_s16(q0s16, 3);
63     q1s16 = vshrq_n_s16(q1s16, 3);
64 
65     // store
66     vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16),  0);
67     mb_dqcoeff += 16;
68     vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 0);
69     mb_dqcoeff += 16;
70     vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16),  0);
71     mb_dqcoeff += 16;
72     vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 0);
73     mb_dqcoeff += 16;
74 
75     vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16),  1);
76     mb_dqcoeff += 16;
77     vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 1);
78     mb_dqcoeff += 16;
79     vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16),  1);
80     mb_dqcoeff += 16;
81     vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 1);
82     mb_dqcoeff += 16;
83 
84     vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16),  2);
85     mb_dqcoeff += 16;
86     vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 2);
87     mb_dqcoeff += 16;
88     vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16),  2);
89     mb_dqcoeff += 16;
90     vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 2);
91     mb_dqcoeff += 16;
92 
93     vst1_lane_s16(mb_dqcoeff, vget_low_s16(q0s16),  3);
94     mb_dqcoeff += 16;
95     vst1_lane_s16(mb_dqcoeff, vget_high_s16(q0s16), 3);
96     mb_dqcoeff += 16;
97     vst1_lane_s16(mb_dqcoeff, vget_low_s16(q1s16),  3);
98     mb_dqcoeff += 16;
99     vst1_lane_s16(mb_dqcoeff, vget_high_s16(q1s16), 3);
100     mb_dqcoeff += 16;
101     return;
102 }
103