1/***********************************************************************
2Copyright (C) 2014 Vidyo
3Redistribution and use in source and binary forms, with or without
4modification, are permitted provided that the following conditions
5are met:
6- Redistributions of source code must retain the above copyright notice,
7this list of conditions and the following disclaimer.
8- Redistributions in binary form must reproduce the above copyright
9notice, this list of conditions and the following disclaimer in the
10documentation and/or other materials provided with the distribution.
11- Neither the name of Internet Society, IETF or IETF Trust, nor the
12names of specific contributors, may be used to endorse or promote
13products derived from this software without specific prior written
14permission.
15THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25POSSIBILITY OF SUCH DAMAGE.
26***********************************************************************/
27#ifdef HAVE_CONFIG_H
28#include "config.h"
29#endif
30
31#include <arm_neon.h>
32#include "main.h"
33#include "stack_alloc.h"
34#include "NSQ.h"
35#include "celt/cpu_support.h"
36#include "celt/arm/armcpu.h"
37
38opus_int32 silk_noise_shape_quantizer_short_prediction_neon(const opus_int32 *buf32, const opus_int32 *coef32, opus_int order)
39{
40    int32x4_t coef0 = vld1q_s32(coef32);
41    int32x4_t coef1 = vld1q_s32(coef32 + 4);
42    int32x4_t coef2 = vld1q_s32(coef32 + 8);
43    int32x4_t coef3 = vld1q_s32(coef32 + 12);
44
45    int32x4_t a0 = vld1q_s32(buf32 - 15);
46    int32x4_t a1 = vld1q_s32(buf32 - 11);
47    int32x4_t a2 = vld1q_s32(buf32 - 7);
48    int32x4_t a3 = vld1q_s32(buf32 - 3);
49
50    int32x4_t b0 = vqdmulhq_s32(coef0, a0);
51    int32x4_t b1 = vqdmulhq_s32(coef1, a1);
52    int32x4_t b2 = vqdmulhq_s32(coef2, a2);
53    int32x4_t b3 = vqdmulhq_s32(coef3, a3);
54
55    int32x4_t c0 = vaddq_s32(b0, b1);
56    int32x4_t c1 = vaddq_s32(b2, b3);
57
58    int32x4_t d = vaddq_s32(c0, c1);
59
60    int64x2_t e = vpaddlq_s32(d);
61
62    int64x1_t f = vadd_s64(vget_low_s64(e), vget_high_s64(e));
63
64    opus_int32 out = vget_lane_s32(vreinterpret_s32_s64(f), 0);
65
66    out += silk_RSHIFT( order, 1 );
67
68    return out;
69}
70
71
72opus_int32 silk_NSQ_noise_shape_feedback_loop_neon(const opus_int32 *data0, opus_int32 *data1, const opus_int16 *coef, opus_int order)
73{
74    opus_int32 out;
75    if (order == 8)
76    {
77        int32x4_t a00 = vdupq_n_s32(data0[0]);
78        int32x4_t a01 = vld1q_s32(data1);  /* data1[0] ... [3] */
79
80        int32x4_t a0 = vextq_s32 (a00, a01, 3); /* data0[0] data1[0] ...[2] */
81        int32x4_t a1 = vld1q_s32(data1 + 3);  /* data1[3] ... [6] */
82
83        /*TODO: Convert these once in advance instead of once per sample, like
84          silk_noise_shape_quantizer_short_prediction_neon() does.*/
85        int16x8_t coef16 = vld1q_s16(coef);
86        int32x4_t coef0 = vmovl_s16(vget_low_s16(coef16));
87        int32x4_t coef1 = vmovl_s16(vget_high_s16(coef16));
88
89        /*This is not bit-exact with the C version, since we do not drop the
90          lower 16 bits of each multiply, but wait until the end to truncate
91          precision. This is an encoder-specific calculation (and unlike
92          silk_noise_shape_quantizer_short_prediction_neon(), is not meant to
93          simulate what the decoder will do). We still could use vqdmulhq_s32()
94          like silk_noise_shape_quantizer_short_prediction_neon() and save
95          half the multiplies, but the speed difference is not large, since we
96          then need two extra adds.*/
97        int64x2_t b0 = vmull_s32(vget_low_s32(a0), vget_low_s32(coef0));
98        int64x2_t b1 = vmlal_s32(b0, vget_high_s32(a0), vget_high_s32(coef0));
99        int64x2_t b2 = vmlal_s32(b1, vget_low_s32(a1), vget_low_s32(coef1));
100        int64x2_t b3 = vmlal_s32(b2, vget_high_s32(a1), vget_high_s32(coef1));
101
102        int64x1_t c = vadd_s64(vget_low_s64(b3), vget_high_s64(b3));
103        int64x1_t cS = vrshr_n_s64(c, 15);
104        int32x2_t d = vreinterpret_s32_s64(cS);
105
106        out = vget_lane_s32(d, 0);
107        vst1q_s32(data1, a0);
108        vst1q_s32(data1 + 4, a1);
109        return out;
110    }
111    return silk_NSQ_noise_shape_feedback_loop_c(data0, data1, coef, order);
112}
113