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 
13 static const int16_t cospi8sqrt2minus1 = 20091;
14 static const int16_t sinpi8sqrt2       = 35468;
15 
vp8_short_idct4x4llm_neon(int16_t * input,unsigned char * pred_ptr,int pred_stride,unsigned char * dst_ptr,int dst_stride)16 void vp8_short_idct4x4llm_neon(
17         int16_t *input,
18         unsigned char *pred_ptr,
19         int pred_stride,
20         unsigned char *dst_ptr,
21         int dst_stride) {
22     int i;
23     uint32x2_t d6u32 = vdup_n_u32(0);
24     uint8x8_t d1u8;
25     int16x4_t d2, d3, d4, d5, d10, d11, d12, d13;
26     uint16x8_t q1u16;
27     int16x8_t q1s16, q2s16, q3s16, q4s16;
28     int32x2x2_t v2tmp0, v2tmp1;
29     int16x4x2_t v2tmp2, v2tmp3;
30 
31     d2 = vld1_s16(input);
32     d3 = vld1_s16(input + 4);
33     d4 = vld1_s16(input + 8);
34     d5 = vld1_s16(input + 12);
35 
36     // 1st for loop
37     q1s16 = vcombine_s16(d2, d4);  // Swap d3 d4 here
38     q2s16 = vcombine_s16(d3, d5);
39 
40     q3s16 = vqdmulhq_n_s16(q2s16, sinpi8sqrt2);
41     q4s16 = vqdmulhq_n_s16(q2s16, cospi8sqrt2minus1);
42 
43     d12 = vqadd_s16(vget_low_s16(q1s16), vget_high_s16(q1s16));  // a1
44     d13 = vqsub_s16(vget_low_s16(q1s16), vget_high_s16(q1s16));  // b1
45 
46     q3s16 = vshrq_n_s16(q3s16, 1);
47     q4s16 = vshrq_n_s16(q4s16, 1);
48 
49     q3s16 = vqaddq_s16(q3s16, q2s16);
50     q4s16 = vqaddq_s16(q4s16, q2s16);
51 
52     d10 = vqsub_s16(vget_low_s16(q3s16), vget_high_s16(q4s16));  // c1
53     d11 = vqadd_s16(vget_high_s16(q3s16), vget_low_s16(q4s16));  // d1
54 
55     d2 = vqadd_s16(d12, d11);
56     d3 = vqadd_s16(d13, d10);
57     d4 = vqsub_s16(d13, d10);
58     d5 = vqsub_s16(d12, d11);
59 
60     v2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
61     v2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
62     v2tmp2 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[0]),
63                       vreinterpret_s16_s32(v2tmp1.val[0]));
64     v2tmp3 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[1]),
65                       vreinterpret_s16_s32(v2tmp1.val[1]));
66 
67     // 2nd for loop
68     q1s16 = vcombine_s16(v2tmp2.val[0], v2tmp3.val[0]);
69     q2s16 = vcombine_s16(v2tmp2.val[1], v2tmp3.val[1]);
70 
71     q3s16 = vqdmulhq_n_s16(q2s16, sinpi8sqrt2);
72     q4s16 = vqdmulhq_n_s16(q2s16, cospi8sqrt2minus1);
73 
74     d12 = vqadd_s16(vget_low_s16(q1s16), vget_high_s16(q1s16));  // a1
75     d13 = vqsub_s16(vget_low_s16(q1s16), vget_high_s16(q1s16));  // b1
76 
77     q3s16 = vshrq_n_s16(q3s16, 1);
78     q4s16 = vshrq_n_s16(q4s16, 1);
79 
80     q3s16 = vqaddq_s16(q3s16, q2s16);
81     q4s16 = vqaddq_s16(q4s16, q2s16);
82 
83     d10 = vqsub_s16(vget_low_s16(q3s16), vget_high_s16(q4s16));  // c1
84     d11 = vqadd_s16(vget_high_s16(q3s16), vget_low_s16(q4s16));  // d1
85 
86     d2 = vqadd_s16(d12, d11);
87     d3 = vqadd_s16(d13, d10);
88     d4 = vqsub_s16(d13, d10);
89     d5 = vqsub_s16(d12, d11);
90 
91     d2 = vrshr_n_s16(d2, 3);
92     d3 = vrshr_n_s16(d3, 3);
93     d4 = vrshr_n_s16(d4, 3);
94     d5 = vrshr_n_s16(d5, 3);
95 
96     v2tmp0 = vtrn_s32(vreinterpret_s32_s16(d2), vreinterpret_s32_s16(d4));
97     v2tmp1 = vtrn_s32(vreinterpret_s32_s16(d3), vreinterpret_s32_s16(d5));
98     v2tmp2 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[0]),
99                       vreinterpret_s16_s32(v2tmp1.val[0]));
100     v2tmp3 = vtrn_s16(vreinterpret_s16_s32(v2tmp0.val[1]),
101                       vreinterpret_s16_s32(v2tmp1.val[1]));
102 
103     q1s16 = vcombine_s16(v2tmp2.val[0], v2tmp2.val[1]);
104     q2s16 = vcombine_s16(v2tmp3.val[0], v2tmp3.val[1]);
105 
106     // dc_only_idct_add
107     for (i = 0; i < 2; i++, q1s16 = q2s16) {
108         d6u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d6u32, 0);
109         pred_ptr += pred_stride;
110         d6u32 = vld1_lane_u32((const uint32_t *)pred_ptr, d6u32, 1);
111         pred_ptr += pred_stride;
112 
113         q1u16 = vaddw_u8(vreinterpretq_u16_s16(q1s16),
114                          vreinterpret_u8_u32(d6u32));
115         d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q1u16));
116 
117         vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d1u8), 0);
118         dst_ptr += dst_stride;
119         vst1_lane_u32((uint32_t *)dst_ptr, vreinterpret_u32_u8(d1u8), 1);
120         dst_ptr += dst_stride;
121     }
122     return;
123 }
124