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       = 17734;
15 // because the lowest bit in 0x8a8c is 0, we can pre-shift this
16 
idct_dequant_full_2x_neon(int16_t * q,int16_t * dq,unsigned char * dst,int stride)17 void idct_dequant_full_2x_neon(
18         int16_t *q,
19         int16_t *dq,
20         unsigned char *dst,
21         int stride) {
22     unsigned char *dst0, *dst1;
23     int32x2_t d28, d29, d30, d31;
24     int16x8_t q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11;
25     int16x8_t qEmpty = vdupq_n_s16(0);
26     int32x4x2_t q2tmp0, q2tmp1;
27     int16x8x2_t q2tmp2, q2tmp3;
28     int16x4_t dLow0, dLow1, dHigh0, dHigh1;
29 
30     d28 = d29 = d30 = d31 = vdup_n_s32(0);
31 
32     // load dq
33     q0 = vld1q_s16(dq);
34     dq += 8;
35     q1 = vld1q_s16(dq);
36 
37     // load q
38     q2 = vld1q_s16(q);
39     vst1q_s16(q, qEmpty);
40     q += 8;
41     q3 = vld1q_s16(q);
42     vst1q_s16(q, qEmpty);
43     q += 8;
44     q4 = vld1q_s16(q);
45     vst1q_s16(q, qEmpty);
46     q += 8;
47     q5 = vld1q_s16(q);
48     vst1q_s16(q, qEmpty);
49 
50     // load src from dst
51     dst0 = dst;
52     dst1 = dst + 4;
53     d28 = vld1_lane_s32((const int32_t *)dst0, d28, 0);
54     dst0 += stride;
55     d28 = vld1_lane_s32((const int32_t *)dst1, d28, 1);
56     dst1 += stride;
57     d29 = vld1_lane_s32((const int32_t *)dst0, d29, 0);
58     dst0 += stride;
59     d29 = vld1_lane_s32((const int32_t *)dst1, d29, 1);
60     dst1 += stride;
61 
62     d30 = vld1_lane_s32((const int32_t *)dst0, d30, 0);
63     dst0 += stride;
64     d30 = vld1_lane_s32((const int32_t *)dst1, d30, 1);
65     dst1 += stride;
66     d31 = vld1_lane_s32((const int32_t *)dst0, d31, 0);
67     d31 = vld1_lane_s32((const int32_t *)dst1, d31, 1);
68 
69     q2 = vmulq_s16(q2, q0);
70     q3 = vmulq_s16(q3, q1);
71     q4 = vmulq_s16(q4, q0);
72     q5 = vmulq_s16(q5, q1);
73 
74     // vswp
75     dLow0 = vget_low_s16(q2);
76     dHigh0 = vget_high_s16(q2);
77     dLow1 = vget_low_s16(q4);
78     dHigh1 = vget_high_s16(q4);
79     q2 = vcombine_s16(dLow0, dLow1);
80     q4 = vcombine_s16(dHigh0, dHigh1);
81 
82     dLow0 = vget_low_s16(q3);
83     dHigh0 = vget_high_s16(q3);
84     dLow1 = vget_low_s16(q5);
85     dHigh1 = vget_high_s16(q5);
86     q3 = vcombine_s16(dLow0, dLow1);
87     q5 = vcombine_s16(dHigh0, dHigh1);
88 
89     q6 = vqdmulhq_n_s16(q4, sinpi8sqrt2);
90     q7 = vqdmulhq_n_s16(q5, sinpi8sqrt2);
91     q8 = vqdmulhq_n_s16(q4, cospi8sqrt2minus1);
92     q9 = vqdmulhq_n_s16(q5, cospi8sqrt2minus1);
93 
94     q10 = vqaddq_s16(q2, q3);
95     q11 = vqsubq_s16(q2, q3);
96 
97     q8 = vshrq_n_s16(q8, 1);
98     q9 = vshrq_n_s16(q9, 1);
99 
100     q4 = vqaddq_s16(q4, q8);
101     q5 = vqaddq_s16(q5, q9);
102 
103     q2 = vqsubq_s16(q6, q5);
104     q3 = vqaddq_s16(q7, q4);
105 
106     q4 = vqaddq_s16(q10, q3);
107     q5 = vqaddq_s16(q11, q2);
108     q6 = vqsubq_s16(q11, q2);
109     q7 = vqsubq_s16(q10, q3);
110 
111     q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6));
112     q2tmp1 = vtrnq_s32(vreinterpretq_s32_s16(q5), vreinterpretq_s32_s16(q7));
113     q2tmp2 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[0]),
114                        vreinterpretq_s16_s32(q2tmp1.val[0]));
115     q2tmp3 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[1]),
116                        vreinterpretq_s16_s32(q2tmp1.val[1]));
117 
118     // loop 2
119     q8  = vqdmulhq_n_s16(q2tmp2.val[1], sinpi8sqrt2);
120     q9  = vqdmulhq_n_s16(q2tmp3.val[1], sinpi8sqrt2);
121     q10 = vqdmulhq_n_s16(q2tmp2.val[1], cospi8sqrt2minus1);
122     q11 = vqdmulhq_n_s16(q2tmp3.val[1], cospi8sqrt2minus1);
123 
124     q2 = vqaddq_s16(q2tmp2.val[0], q2tmp3.val[0]);
125     q3 = vqsubq_s16(q2tmp2.val[0], q2tmp3.val[0]);
126 
127     q10 = vshrq_n_s16(q10, 1);
128     q11 = vshrq_n_s16(q11, 1);
129 
130     q10 = vqaddq_s16(q2tmp2.val[1], q10);
131     q11 = vqaddq_s16(q2tmp3.val[1], q11);
132 
133     q8 = vqsubq_s16(q8, q11);
134     q9 = vqaddq_s16(q9, q10);
135 
136     q4 = vqaddq_s16(q2, q9);
137     q5 = vqaddq_s16(q3, q8);
138     q6 = vqsubq_s16(q3, q8);
139     q7 = vqsubq_s16(q2, q9);
140 
141     q4 = vrshrq_n_s16(q4, 3);
142     q5 = vrshrq_n_s16(q5, 3);
143     q6 = vrshrq_n_s16(q6, 3);
144     q7 = vrshrq_n_s16(q7, 3);
145 
146     q2tmp0 = vtrnq_s32(vreinterpretq_s32_s16(q4), vreinterpretq_s32_s16(q6));
147     q2tmp1 = vtrnq_s32(vreinterpretq_s32_s16(q5), vreinterpretq_s32_s16(q7));
148     q2tmp2 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[0]),
149                        vreinterpretq_s16_s32(q2tmp1.val[0]));
150     q2tmp3 = vtrnq_s16(vreinterpretq_s16_s32(q2tmp0.val[1]),
151                        vreinterpretq_s16_s32(q2tmp1.val[1]));
152 
153     q4 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp2.val[0]),
154                                           vreinterpret_u8_s32(d28)));
155     q5 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp2.val[1]),
156                                           vreinterpret_u8_s32(d29)));
157     q6 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp3.val[0]),
158                                           vreinterpret_u8_s32(d30)));
159     q7 = vreinterpretq_s16_u16(vaddw_u8(vreinterpretq_u16_s16(q2tmp3.val[1]),
160                                           vreinterpret_u8_s32(d31)));
161 
162     d28 = vreinterpret_s32_u8(vqmovun_s16(q4));
163     d29 = vreinterpret_s32_u8(vqmovun_s16(q5));
164     d30 = vreinterpret_s32_u8(vqmovun_s16(q6));
165     d31 = vreinterpret_s32_u8(vqmovun_s16(q7));
166 
167     dst0 = dst;
168     dst1 = dst + 4;
169     vst1_lane_s32((int32_t *)dst0, d28, 0);
170     dst0 += stride;
171     vst1_lane_s32((int32_t *)dst1, d28, 1);
172     dst1 += stride;
173     vst1_lane_s32((int32_t *)dst0, d29, 0);
174     dst0 += stride;
175     vst1_lane_s32((int32_t *)dst1, d29, 1);
176     dst1 += stride;
177 
178     vst1_lane_s32((int32_t *)dst0, d30, 0);
179     dst0 += stride;
180     vst1_lane_s32((int32_t *)dst1, d30, 1);
181     dst1 += stride;
182     vst1_lane_s32((int32_t *)dst0, d31, 0);
183     vst1_lane_s32((int32_t *)dst1, d31, 1);
184     return;
185 }
186