1/*
2 *  Copyright (c) 2017 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#include "vpx_dsp/vpx_dsp_common.h"
16#include "vpx_dsp/arm/idct_neon.h"
17#include "vpx_dsp/arm/mem_neon.h"
18#include "vpx_dsp/arm/transpose_neon.h"
19
20void vpx_fdct4x4_neon(const int16_t *input, tran_low_t *final_output,
21                      int stride) {
22  int i;
23  // input[M * stride] * 16
24  int16x4_t input_0 = vshl_n_s16(vld1_s16(input + 0 * stride), 4);
25  int16x4_t input_1 = vshl_n_s16(vld1_s16(input + 1 * stride), 4);
26  int16x4_t input_2 = vshl_n_s16(vld1_s16(input + 2 * stride), 4);
27  int16x4_t input_3 = vshl_n_s16(vld1_s16(input + 3 * stride), 4);
28
29  // If the very first value != 0, then add 1.
30  if (input[0] != 0) {
31    const int16x4_t one = vreinterpret_s16_s64(vdup_n_s64(1));
32    input_0 = vadd_s16(input_0, one);
33  }
34
35  for (i = 0; i < 2; ++i) {
36    const int16x8_t input_01 = vcombine_s16(input_0, input_1);
37    const int16x8_t input_32 = vcombine_s16(input_3, input_2);
38
39    // in_0 +/- in_3, in_1 +/- in_2
40    const int16x8_t s_01 = vaddq_s16(input_01, input_32);
41    const int16x8_t s_32 = vsubq_s16(input_01, input_32);
42
43    // step_0 +/- step_1, step_2 +/- step_3
44    const int16x4_t s_0 = vget_low_s16(s_01);
45    const int16x4_t s_1 = vget_high_s16(s_01);
46    const int16x4_t s_2 = vget_high_s16(s_32);
47    const int16x4_t s_3 = vget_low_s16(s_32);
48
49    // (s_0 +/- s_1) * cospi_16_64
50    // Must expand all elements to s32. See 'needs32' comment in fwd_txfm.c.
51    const int32x4_t s_0_p_s_1 = vaddl_s16(s_0, s_1);
52    const int32x4_t s_0_m_s_1 = vsubl_s16(s_0, s_1);
53    const int32x4_t temp1 = vmulq_n_s32(s_0_p_s_1, (int16_t)cospi_16_64);
54    const int32x4_t temp2 = vmulq_n_s32(s_0_m_s_1, (int16_t)cospi_16_64);
55
56    // fdct_round_shift
57    int16x4_t out_0 = vrshrn_n_s32(temp1, DCT_CONST_BITS);
58    int16x4_t out_2 = vrshrn_n_s32(temp2, DCT_CONST_BITS);
59
60    // s_3 * cospi_8_64 + s_2 * cospi_24_64
61    // s_3 * cospi_24_64 - s_2 * cospi_8_64
62    const int32x4_t s_3_cospi_8_64 = vmull_n_s16(s_3, (int16_t)cospi_8_64);
63    const int32x4_t s_3_cospi_24_64 = vmull_n_s16(s_3, (int16_t)cospi_24_64);
64
65    const int32x4_t temp3 =
66        vmlal_n_s16(s_3_cospi_8_64, s_2, (int16_t)cospi_24_64);
67    const int32x4_t temp4 =
68        vmlsl_n_s16(s_3_cospi_24_64, s_2, (int16_t)cospi_8_64);
69
70    // fdct_round_shift
71    int16x4_t out_1 = vrshrn_n_s32(temp3, DCT_CONST_BITS);
72    int16x4_t out_3 = vrshrn_n_s32(temp4, DCT_CONST_BITS);
73
74    transpose_s16_4x4d(&out_0, &out_1, &out_2, &out_3);
75
76    input_0 = out_0;
77    input_1 = out_1;
78    input_2 = out_2;
79    input_3 = out_3;
80  }
81
82  {
83    // Not quite a rounding shift. Only add 1 despite shifting by 2.
84    const int16x8_t one = vdupq_n_s16(1);
85    int16x8_t out_01 = vcombine_s16(input_0, input_1);
86    int16x8_t out_23 = vcombine_s16(input_2, input_3);
87    out_01 = vshrq_n_s16(vaddq_s16(out_01, one), 2);
88    out_23 = vshrq_n_s16(vaddq_s16(out_23, one), 2);
89    store_s16q_to_tran_low(final_output + 0 * 8, out_01);
90    store_s16q_to_tran_low(final_output + 1 * 8, out_23);
91  }
92}
93