Better rate allocation for stereo SILK in hybrid mode
[opus.git] / silk / arm / biquad_alt_neon_intr.c
1 /***********************************************************************
2 Copyright (c) 2017 Google Inc.
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
28 #ifdef HAVE_CONFIG_H
29 #include "config.h"
30 #endif
31
32 #include <arm_neon.h>
33 #ifdef OPUS_CHECK_ASM
34 # include <string.h>
35 # include "stack_alloc.h"
36 #endif
37 #include "SigProc_FIX.h"
38
39 static inline void silk_biquad_alt_stride2_kernel( const int32x4_t A_L_s32x4, const int32x4_t A_U_s32x4, const int32x4_t B_Q28_s32x4, const int32x2_t t_s32x2, const int32x4_t in_s32x4, int32x4_t *S_s32x4, int32x2_t *out32_Q14_s32x2 )
40 {
41     int32x4_t t_s32x4, out32_Q14_s32x4;
42
43     *out32_Q14_s32x2 = vadd_s32( vget_low_s32( *S_s32x4 ), t_s32x2 );              /* silk_SMLAWB( S{0,1}, B_Q28[ 0 ], in{0,1} )                                      */
44     *S_s32x4         = vcombine_s32( vget_high_s32( *S_s32x4 ), vdup_n_s32( 0 ) ); /* S{0,1} = S{2,3}; S{2,3} = 0;                                                    */
45     *out32_Q14_s32x2 = vshl_n_s32( *out32_Q14_s32x2, 2 );                          /* out32_Q14_{0,1} = silk_LSHIFT( silk_SMLAWB( S{0,1}, B_Q28[ 0 ], in{0,1} ), 2 ); */
46     out32_Q14_s32x4  = vcombine_s32( *out32_Q14_s32x2, *out32_Q14_s32x2 );         /* out32_Q14_{0,1,0,1}                                                             */
47     t_s32x4          = vqdmulhq_s32( out32_Q14_s32x4, A_L_s32x4 );                 /* silk_SMULWB( out32_Q14_{0,1,0,1}, A{0,0,1,1}_L_Q28 )                            */
48     *S_s32x4         = vrsraq_n_s32( *S_s32x4, t_s32x4, 14 );                      /* S{0,1} = S{2,3} + silk_RSHIFT_ROUND();  S{2,3} = silk_RSHIFT_ROUND();           */
49     t_s32x4          = vqdmulhq_s32( out32_Q14_s32x4, A_U_s32x4 );                 /* silk_SMULWB( out32_Q14_{0,1,0,1}, A{0,0,1,1}_U_Q28 )                            */
50     *S_s32x4         = vaddq_s32( *S_s32x4, t_s32x4 );                             /* S0 = silk_SMLAWB( S{0,1,2,3}, out32_Q14_{0,1,0,1}, A{0,0,1,1}_U_Q28 );          */
51     t_s32x4          = vqdmulhq_s32( in_s32x4, B_Q28_s32x4 );                      /* silk_SMULWB( B_Q28[ {1,1,2,2} ], in{0,1,0,1} )                                  */
52     *S_s32x4         = vaddq_s32( *S_s32x4, t_s32x4 );                             /* S0 = silk_SMLAWB( S0, B_Q28[ {1,1,2,2} ], in{0,1,0,1} );                        */
53 }
54
55 void silk_biquad_alt_stride2_neon(
56     const opus_int16            *in,                /* I     input signal                                               */
57     const opus_int32            *B_Q28,             /* I     MA coefficients [3]                                        */
58     const opus_int32            *A_Q28,             /* I     AR coefficients [2]                                        */
59     opus_int32                  *S,                 /* I/O   State vector [4]                                           */
60     opus_int16                  *out,               /* O     output signal                                              */
61     const opus_int32            len                 /* I     signal length (must be even)                               */
62 )
63 {
64     /* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */
65     opus_int        k            = 0;
66     const int32x2_t offset_s32x2 = vdup_n_s32( (1<<14) - 1 );
67     const int32x4_t offset_s32x4 = vcombine_s32( offset_s32x2, offset_s32x2 );
68     int16x4_t       in_s16x4  = vdup_n_s16( 0 );
69     int16x4_t       out_s16x4;
70     int32x2_t       A_Q28_s32x2, A_L_s32x2, A_U_s32x2, B_Q28_s32x2, t_s32x2;
71     int32x4_t       A_L_s32x4, A_U_s32x4, B_Q28_s32x4, S_s32x4, out32_Q14_s32x4;
72     int32x2x2_t     t0_s32x2x2, t1_s32x2x2, t2_s32x2x2, S_s32x2x2;
73
74 #ifdef OPUS_CHECK_ASM
75     opus_int32 S_c[ 4 ];
76     VARDECL( opus_int16, out_c );
77     SAVE_STACK;
78     ALLOC( out_c, 2 * len, opus_int16 );
79
80     silk_memcpy( &S_c, S, sizeof( S_c ) );
81     silk_biquad_alt_stride2_c( in, B_Q28, A_Q28, S_c, out_c, len );
82 #endif
83
84     /* Negate A_Q28 values and split in two parts */
85     A_Q28_s32x2 = vld1_s32( A_Q28 );
86     A_Q28_s32x2 = vneg_s32( A_Q28_s32x2 );
87     A_L_s32x2   = vshl_n_s32( A_Q28_s32x2, 18 );                                                        /* ( -A_Q28[] & 0x00003FFF ) << 18                                                     */
88     A_L_s32x2   = vreinterpret_s32_u32( vshr_n_u32( vreinterpret_u32_s32( A_L_s32x2 ), 3 ) );           /* ( -A_Q28[] & 0x00003FFF ) << 15                                                     */
89     A_U_s32x2   = vshr_n_s32( A_Q28_s32x2, 14 );                                                        /* silk_RSHIFT( -A_Q28[], 14 )                                                         */
90     A_U_s32x2   = vshl_n_s32( A_U_s32x2, 16 );                                                          /* silk_RSHIFT( -A_Q28[], 14 ) << 16 (Clip two leading bits to conform to C function.) */
91     A_U_s32x2   = vshr_n_s32( A_U_s32x2, 1 );                                                           /* silk_RSHIFT( -A_Q28[], 14 ) << 15                                                   */
92
93     B_Q28_s32x2  = vld1_s32( B_Q28 );
94     t_s32x2      = vld1_s32( B_Q28 + 1 );
95     t0_s32x2x2   = vzip_s32( A_L_s32x2, A_L_s32x2 );
96     t1_s32x2x2   = vzip_s32( A_U_s32x2, A_U_s32x2 );
97     t2_s32x2x2   = vzip_s32( t_s32x2, t_s32x2 );
98     A_L_s32x4    = vcombine_s32( t0_s32x2x2.val[ 0 ], t0_s32x2x2.val[ 1 ] );                            /* A{0,0,1,1}_L_Q28          */
99     A_U_s32x4    = vcombine_s32( t1_s32x2x2.val[ 0 ], t1_s32x2x2.val[ 1 ] );                            /* A{0,0,1,1}_U_Q28          */
100     B_Q28_s32x4  = vcombine_s32( t2_s32x2x2.val[ 0 ], t2_s32x2x2.val[ 1 ] );                            /* B_Q28[ {1,1,2,2} ]        */
101     S_s32x4      = vld1q_s32( S );                                                                      /* S0 = S[ 0 ]; S3 = S[ 3 ]; */
102     S_s32x2x2    = vtrn_s32( vget_low_s32( S_s32x4 ), vget_high_s32( S_s32x4 ) );                       /* S2 = S[ 1 ]; S1 = S[ 2 ]; */
103     S_s32x4      = vcombine_s32( S_s32x2x2.val[ 0 ], S_s32x2x2.val[ 1 ] );
104
105     for( ; k < len - 1; k += 2 ) {
106         int32x4_t in_s32x4[ 2 ], t_s32x4;
107         int32x2_t out32_Q14_s32x2[ 2 ];
108
109         /* S[ 2 * i + 0 ], S[ 2 * i + 1 ], S[ 2 * i + 2 ], S[ 2 * i + 3 ]: Q12 */
110         in_s16x4      = vld1_s16( &in[ 2 * k ] );                                                       /* in{0,1,2,3} = in[ 2 * k + {0,1,2,3} ]; */
111         in_s32x4[ 0 ] = vshll_n_s16( in_s16x4, 15 );                                                    /* in{0,1,2,3} << 15                      */
112         t_s32x4       = vqdmulhq_lane_s32( in_s32x4[ 0 ], B_Q28_s32x2, 0 );                             /* silk_SMULWB( B_Q28[ 0 ], in{0,1,2,3} ) */
113         in_s32x4[ 1 ] = vcombine_s32( vget_high_s32( in_s32x4[ 0 ] ), vget_high_s32( in_s32x4[ 0 ] ) ); /* in{2,3,2,3} << 15                      */
114         in_s32x4[ 0 ] = vcombine_s32( vget_low_s32 ( in_s32x4[ 0 ] ), vget_low_s32 ( in_s32x4[ 0 ] ) ); /* in{0,1,0,1} << 15                      */
115         silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, vget_low_s32 ( t_s32x4 ), in_s32x4[ 0 ], &S_s32x4, &out32_Q14_s32x2[ 0 ] );
116         silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, vget_high_s32( t_s32x4 ), in_s32x4[ 1 ], &S_s32x4, &out32_Q14_s32x2[ 1 ] );
117
118         /* Scale back to Q0 and saturate */
119         out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2[ 0 ], out32_Q14_s32x2[ 1 ] );                   /* out32_Q14_{0,1,2,3}                                                                                        */
120         out32_Q14_s32x4 = vaddq_s32( out32_Q14_s32x4, offset_s32x4 );                                   /* out32_Q14_{0,1,2,3} + (1<<14) - 1                                                                          */
121         out_s16x4       = vqshrn_n_s32( out32_Q14_s32x4, 14 );                                          /* (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,2,3} + (1<<14) - 1, 14 ) )                             */
122         vst1_s16( &out[ 2 * k ], out_s16x4 );                                                           /* out[ 2 * k + {0,1,2,3} ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,2,3} + (1<<14) - 1, 14 ) ); */
123     }
124
125     /* Process leftover. */
126     if( k < len ) {
127         int32x4_t in_s32x4;
128         int32x2_t out32_Q14_s32x2;
129
130         /* S[ 2 * i + 0 ], S[ 2 * i + 1 ]: Q12 */
131         in_s16x4     = vld1_lane_s16( &in[ 2 * k + 0 ], in_s16x4, 0 );                                  /* in{0,1} = in[ 2 * k + {0,1} ];     */
132         in_s16x4     = vld1_lane_s16( &in[ 2 * k + 1 ], in_s16x4, 1 );                                  /* in{0,1} = in[ 2 * k + {0,1} ];     */
133         in_s32x4     = vshll_n_s16( in_s16x4, 15 );                                                     /* in{0,1} << 15                      */
134         t_s32x2      = vqdmulh_lane_s32( vget_low_s32( in_s32x4 ), B_Q28_s32x2, 0 );                    /* silk_SMULWB( B_Q28[ 0 ], in{0,1} ) */
135         in_s32x4     = vcombine_s32( vget_low_s32( in_s32x4 ), vget_low_s32( in_s32x4 ) );              /* in{0,1,0,1} << 15                  */
136         silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, t_s32x2, in_s32x4, &S_s32x4, &out32_Q14_s32x2 );
137
138         /* Scale back to Q0 and saturate */
139         out32_Q14_s32x2 = vadd_s32( out32_Q14_s32x2, offset_s32x2 );                                    /* out32_Q14_{0,1} + (1<<14) - 1                                                              */
140         out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2, out32_Q14_s32x2 );                             /* out32_Q14_{0,1,0,1} + (1<<14) - 1                                                          */
141         out_s16x4       = vqshrn_n_s32( out32_Q14_s32x4, 14 );                                          /* (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_{0,1,0,1} + (1<<14) - 1, 14 ) )             */
142         vst1_lane_s16( &out[ 2 * k + 0 ], out_s16x4, 0 );                                               /* out[ 2 * k + 0 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_0 + (1<<14) - 1, 14 ) ); */
143         vst1_lane_s16( &out[ 2 * k + 1 ], out_s16x4, 1 );                                               /* out[ 2 * k + 1 ] = (opus_int16)silk_SAT16( silk_RSHIFT( out32_Q14_1 + (1<<14) - 1, 14 ) ); */
144     }
145
146     vst1q_lane_s32( &S[ 0 ], S_s32x4, 0 );                                                              /* S[ 0 ] = S0; */
147     vst1q_lane_s32( &S[ 1 ], S_s32x4, 2 );                                                              /* S[ 1 ] = S2; */
148     vst1q_lane_s32( &S[ 2 ], S_s32x4, 1 );                                                              /* S[ 2 ] = S1; */
149     vst1q_lane_s32( &S[ 3 ], S_s32x4, 3 );                                                              /* S[ 3 ] = S3; */
150
151 #ifdef OPUS_CHECK_ASM
152     silk_assert( !memcmp( S_c, S, sizeof( S_c ) ) );
153     silk_assert( !memcmp( out_c, out, 2 * len * sizeof( opus_int16 ) ) );
154     RESTORE_STACK;
155 #endif
156 }