1 /***********************************************************************
2 Copyright (C) 2014 Vidyo
3 Redistribution and use in source and binary forms, with or without
4 modification, are permitted provided that the following conditions
5 are met:
6 - Redistributions of source code must retain the above copyright notice,
7 this list of conditions and the following disclaimer.
8 - Redistributions in binary form must reproduce the above copyright
9 notice, this list of conditions and the following disclaimer in the
10 documentation and/or other materials provided with the distribution.
11 - Neither the name of Internet Society, IETF or IETF Trust, nor the
12 names of specific contributors, may be used to endorse or promote
13 products derived from this software without specific prior written
14 permission.
15 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20 CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21 SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22 INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23 CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24 ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25 POSSIBILITY 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
silk_noise_shape_quantizer_short_prediction_neon(const opus_int32 * buf32,const opus_int32 * coef32,opus_int order)38 opus_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
silk_NSQ_noise_shape_feedback_loop_neon(const opus_int32 * data0,opus_int32 * data1,const opus_int16 * coef,opus_int order)72 opus_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