1/*
2 *  Copyright (c) 2015 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#include "./vpx_config.h"
14#include "vpx_dsp/txfm_common.h"
15
16void vpx_fdct8x8_neon(const int16_t *input, int16_t *final_output, int stride) {
17  int i;
18  // stage 1
19  int16x8_t input_0 = vshlq_n_s16(vld1q_s16(&input[0 * stride]), 2);
20  int16x8_t input_1 = vshlq_n_s16(vld1q_s16(&input[1 * stride]), 2);
21  int16x8_t input_2 = vshlq_n_s16(vld1q_s16(&input[2 * stride]), 2);
22  int16x8_t input_3 = vshlq_n_s16(vld1q_s16(&input[3 * stride]), 2);
23  int16x8_t input_4 = vshlq_n_s16(vld1q_s16(&input[4 * stride]), 2);
24  int16x8_t input_5 = vshlq_n_s16(vld1q_s16(&input[5 * stride]), 2);
25  int16x8_t input_6 = vshlq_n_s16(vld1q_s16(&input[6 * stride]), 2);
26  int16x8_t input_7 = vshlq_n_s16(vld1q_s16(&input[7 * stride]), 2);
27  for (i = 0; i < 2; ++i) {
28    int16x8_t out_0, out_1, out_2, out_3, out_4, out_5, out_6, out_7;
29    const int16x8_t v_s0 = vaddq_s16(input_0, input_7);
30    const int16x8_t v_s1 = vaddq_s16(input_1, input_6);
31    const int16x8_t v_s2 = vaddq_s16(input_2, input_5);
32    const int16x8_t v_s3 = vaddq_s16(input_3, input_4);
33    const int16x8_t v_s4 = vsubq_s16(input_3, input_4);
34    const int16x8_t v_s5 = vsubq_s16(input_2, input_5);
35    const int16x8_t v_s6 = vsubq_s16(input_1, input_6);
36    const int16x8_t v_s7 = vsubq_s16(input_0, input_7);
37    // fdct4(step, step);
38    int16x8_t v_x0 = vaddq_s16(v_s0, v_s3);
39    int16x8_t v_x1 = vaddq_s16(v_s1, v_s2);
40    int16x8_t v_x2 = vsubq_s16(v_s1, v_s2);
41    int16x8_t v_x3 = vsubq_s16(v_s0, v_s3);
42    // fdct4(step, step);
43    int32x4_t v_t0_lo = vaddl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
44    int32x4_t v_t0_hi = vaddl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
45    int32x4_t v_t1_lo = vsubl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
46    int32x4_t v_t1_hi = vsubl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
47    int32x4_t v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_24_64);
48    int32x4_t v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_24_64);
49    int32x4_t v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_24_64);
50    int32x4_t v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_24_64);
51    v_t2_lo = vmlal_n_s16(v_t2_lo, vget_low_s16(v_x3), (int16_t)cospi_8_64);
52    v_t2_hi = vmlal_n_s16(v_t2_hi, vget_high_s16(v_x3), (int16_t)cospi_8_64);
53    v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x2), (int16_t)cospi_8_64);
54    v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x2), (int16_t)cospi_8_64);
55    v_t0_lo = vmulq_n_s32(v_t0_lo, cospi_16_64);
56    v_t0_hi = vmulq_n_s32(v_t0_hi, cospi_16_64);
57    v_t1_lo = vmulq_n_s32(v_t1_lo, cospi_16_64);
58    v_t1_hi = vmulq_n_s32(v_t1_hi, cospi_16_64);
59    {
60      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
61      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
62      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
63      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
64      const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
65      const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
66      const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
67      const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
68      out_0 = vcombine_s16(a, c);  // 00 01 02 03 40 41 42 43
69      out_2 = vcombine_s16(e, g);  // 20 21 22 23 60 61 62 63
70      out_4 = vcombine_s16(b, d);  // 04 05 06 07 44 45 46 47
71      out_6 = vcombine_s16(f, h);  // 24 25 26 27 64 65 66 67
72    }
73    // Stage 2
74    v_x0 = vsubq_s16(v_s6, v_s5);
75    v_x1 = vaddq_s16(v_s6, v_s5);
76    v_t0_lo = vmull_n_s16(vget_low_s16(v_x0), (int16_t)cospi_16_64);
77    v_t0_hi = vmull_n_s16(vget_high_s16(v_x0), (int16_t)cospi_16_64);
78    v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_16_64);
79    v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_16_64);
80    {
81      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
82      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
83      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
84      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
85      const int16x8_t ab = vcombine_s16(a, b);
86      const int16x8_t cd = vcombine_s16(c, d);
87      // Stage 3
88      v_x0 = vaddq_s16(v_s4, ab);
89      v_x1 = vsubq_s16(v_s4, ab);
90      v_x2 = vsubq_s16(v_s7, cd);
91      v_x3 = vaddq_s16(v_s7, cd);
92    }
93    // Stage 4
94    v_t0_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_4_64);
95    v_t0_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_4_64);
96    v_t0_lo = vmlal_n_s16(v_t0_lo, vget_low_s16(v_x0), (int16_t)cospi_28_64);
97    v_t0_hi = vmlal_n_s16(v_t0_hi, vget_high_s16(v_x0), (int16_t)cospi_28_64);
98    v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_12_64);
99    v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_12_64);
100    v_t1_lo = vmlal_n_s16(v_t1_lo, vget_low_s16(v_x2), (int16_t)cospi_20_64);
101    v_t1_hi = vmlal_n_s16(v_t1_hi, vget_high_s16(v_x2), (int16_t)cospi_20_64);
102    v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_12_64);
103    v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_12_64);
104    v_t2_lo = vmlsl_n_s16(v_t2_lo, vget_low_s16(v_x1), (int16_t)cospi_20_64);
105    v_t2_hi = vmlsl_n_s16(v_t2_hi, vget_high_s16(v_x1), (int16_t)cospi_20_64);
106    v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_28_64);
107    v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_28_64);
108    v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x0), (int16_t)cospi_4_64);
109    v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x0), (int16_t)cospi_4_64);
110    {
111      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
112      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
113      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
114      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
115      const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
116      const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
117      const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
118      const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
119      out_1 = vcombine_s16(a, c);  // 10 11 12 13 50 51 52 53
120      out_3 = vcombine_s16(e, g);  // 30 31 32 33 70 71 72 73
121      out_5 = vcombine_s16(b, d);  // 14 15 16 17 54 55 56 57
122      out_7 = vcombine_s16(f, h);  // 34 35 36 37 74 75 76 77
123    }
124    // transpose 8x8
125    {
126      // 00 01 02 03 40 41 42 43
127      // 10 11 12 13 50 51 52 53
128      // 20 21 22 23 60 61 62 63
129      // 30 31 32 33 70 71 72 73
130      // 04 05 06 07 44 45 46 47
131      // 14 15 16 17 54 55 56 57
132      // 24 25 26 27 64 65 66 67
133      // 34 35 36 37 74 75 76 77
134      const int32x4x2_t r02_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_0),
135                                            vreinterpretq_s32_s16(out_2));
136      const int32x4x2_t r13_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_1),
137                                            vreinterpretq_s32_s16(out_3));
138      const int32x4x2_t r46_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_4),
139                                            vreinterpretq_s32_s16(out_6));
140      const int32x4x2_t r57_s32 = vtrnq_s32(vreinterpretq_s32_s16(out_5),
141                                            vreinterpretq_s32_s16(out_7));
142      const int16x8x2_t r01_s16 =
143          vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[0]),
144                    vreinterpretq_s16_s32(r13_s32.val[0]));
145      const int16x8x2_t r23_s16 =
146          vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[1]),
147                    vreinterpretq_s16_s32(r13_s32.val[1]));
148      const int16x8x2_t r45_s16 =
149          vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[0]),
150                    vreinterpretq_s16_s32(r57_s32.val[0]));
151      const int16x8x2_t r67_s16 =
152          vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[1]),
153                    vreinterpretq_s16_s32(r57_s32.val[1]));
154      input_0 = r01_s16.val[0];
155      input_1 = r01_s16.val[1];
156      input_2 = r23_s16.val[0];
157      input_3 = r23_s16.val[1];
158      input_4 = r45_s16.val[0];
159      input_5 = r45_s16.val[1];
160      input_6 = r67_s16.val[0];
161      input_7 = r67_s16.val[1];
162      // 00 10 20 30 40 50 60 70
163      // 01 11 21 31 41 51 61 71
164      // 02 12 22 32 42 52 62 72
165      // 03 13 23 33 43 53 63 73
166      // 04 14 24 34 44 54 64 74
167      // 05 15 25 35 45 55 65 75
168      // 06 16 26 36 46 56 66 76
169      // 07 17 27 37 47 57 67 77
170    }
171  }  // for
172  {
173    // from vpx_dct_sse2.c
174    // Post-condition (division by two)
175    //    division of two 16 bits signed numbers using shifts
176    //    n / 2 = (n - (n >> 15)) >> 1
177    const int16x8_t sign_in0 = vshrq_n_s16(input_0, 15);
178    const int16x8_t sign_in1 = vshrq_n_s16(input_1, 15);
179    const int16x8_t sign_in2 = vshrq_n_s16(input_2, 15);
180    const int16x8_t sign_in3 = vshrq_n_s16(input_3, 15);
181    const int16x8_t sign_in4 = vshrq_n_s16(input_4, 15);
182    const int16x8_t sign_in5 = vshrq_n_s16(input_5, 15);
183    const int16x8_t sign_in6 = vshrq_n_s16(input_6, 15);
184    const int16x8_t sign_in7 = vshrq_n_s16(input_7, 15);
185    input_0 = vhsubq_s16(input_0, sign_in0);
186    input_1 = vhsubq_s16(input_1, sign_in1);
187    input_2 = vhsubq_s16(input_2, sign_in2);
188    input_3 = vhsubq_s16(input_3, sign_in3);
189    input_4 = vhsubq_s16(input_4, sign_in4);
190    input_5 = vhsubq_s16(input_5, sign_in5);
191    input_6 = vhsubq_s16(input_6, sign_in6);
192    input_7 = vhsubq_s16(input_7, sign_in7);
193    // store results
194    vst1q_s16(&final_output[0 * 8], input_0);
195    vst1q_s16(&final_output[1 * 8], input_1);
196    vst1q_s16(&final_output[2 * 8], input_2);
197    vst1q_s16(&final_output[3 * 8], input_3);
198    vst1q_s16(&final_output[4 * 8], input_4);
199    vst1q_s16(&final_output[5 * 8], input_5);
200    vst1q_s16(&final_output[6 * 8], input_6);
201    vst1q_s16(&final_output[7 * 8], input_7);
202  }
203}
204
205void vpx_fdct8x8_1_neon(const int16_t *input, int16_t *output, int stride) {
206  int r;
207  int16x8_t sum = vld1q_s16(&input[0]);
208  for (r = 1; r < 8; ++r) {
209    const int16x8_t input_00 = vld1q_s16(&input[r * stride]);
210    sum = vaddq_s16(sum, input_00);
211  }
212  {
213    const int32x4_t a = vpaddlq_s16(sum);
214    const int64x2_t b = vpaddlq_s32(a);
215    const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
216                                 vreinterpret_s32_s64(vget_high_s64(b)));
217    output[0] = vget_lane_s16(vreinterpret_s16_s32(c), 0);
218    output[1] = 0;
219  }
220}
221