Mercurial > hg > sv-dependency-builds

annotate src/opus-1.3/silk/arm/biquad_alt_neon_intr.c @ 78:7ea7031c0e5c pa_catalina

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Apply patch from Tim Bunnell on PortAudio mailing list (2016-12-28, Mac 10.11 deprecation warning)
author Chris Cannam
date Wed, 30 Oct 2019 11:28:45 +0000
parents 7aeed7906520
children
rev   line source
Chris@69 1 /***********************************************************************
Chris@69 2 Copyright (c) 2017 Google Inc.
Chris@69 3 Redistribution and use in source and binary forms, with or without
Chris@69 4 modification, are permitted provided that the following conditions
Chris@69 5 are met:
Chris@69 6 - Redistributions of source code must retain the above copyright notice,
Chris@69 7 this list of conditions and the following disclaimer.
Chris@69 8 - Redistributions in binary form must reproduce the above copyright
Chris@69 9 notice, this list of conditions and the following disclaimer in the
Chris@69 10 documentation and/or other materials provided with the distribution.
Chris@69 11 - Neither the name of Internet Society, IETF or IETF Trust, nor the
Chris@69 12 names of specific contributors, may be used to endorse or promote
Chris@69 13 products derived from this software without specific prior written
Chris@69 14 permission.
Chris@69 15 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
Chris@69 16 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
Chris@69 17 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
Chris@69 18 ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
Chris@69 19 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
Chris@69 20 CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
Chris@69 21 SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
Chris@69 22 INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
Chris@69 23 CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
Chris@69 24 ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
Chris@69 25 POSSIBILITY OF SUCH DAMAGE.
Chris@69 26 ***********************************************************************/
Chris@69 27
Chris@69 28 #ifdef HAVE_CONFIG_H
Chris@69 29 #include "config.h"
Chris@69 30 #endif
Chris@69 31
Chris@69 32 #include <arm_neon.h>
Chris@69 33 #ifdef OPUS_CHECK_ASM
Chris@69 34 # include <string.h>
Chris@69 35 # include "stack_alloc.h"
Chris@69 36 #endif
Chris@69 37 #include "SigProc_FIX.h"
Chris@69 38
Chris@69 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 )
Chris@69 40 {
Chris@69 41 int32x4_t t_s32x4, out32_Q14_s32x4;
Chris@69 42
Chris@69 43 *out32_Q14_s32x2 = vadd_s32( vget_low_s32( *S_s32x4 ), t_s32x2 ); /* silk_SMLAWB( S{0,1}, B_Q28[ 0 ], in{0,1} ) */
Chris@69 44 *S_s32x4 = vcombine_s32( vget_high_s32( *S_s32x4 ), vdup_n_s32( 0 ) ); /* S{0,1} = S{2,3}; S{2,3} = 0; */
Chris@69 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 ); */
Chris@69 46 out32_Q14_s32x4 = vcombine_s32( *out32_Q14_s32x2, *out32_Q14_s32x2 ); /* out32_Q14_{0,1,0,1} */
Chris@69 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 ) */
Chris@69 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(); */
Chris@69 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 ) */
Chris@69 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 ); */
Chris@69 51 t_s32x4 = vqdmulhq_s32( in_s32x4, B_Q28_s32x4 ); /* silk_SMULWB( B_Q28[ {1,1,2,2} ], in{0,1,0,1} ) */
Chris@69 52 *S_s32x4 = vaddq_s32( *S_s32x4, t_s32x4 ); /* S0 = silk_SMLAWB( S0, B_Q28[ {1,1,2,2} ], in{0,1,0,1} ); */
Chris@69 53 }
Chris@69 54
Chris@69 55 void silk_biquad_alt_stride2_neon(
Chris@69 56 const opus_int16 *in, /* I input signal */
Chris@69 57 const opus_int32 *B_Q28, /* I MA coefficients [3] */
Chris@69 58 const opus_int32 *A_Q28, /* I AR coefficients [2] */
Chris@69 59 opus_int32 *S, /* I/O State vector [4] */
Chris@69 60 opus_int16 *out, /* O output signal */
Chris@69 61 const opus_int32 len /* I signal length (must be even) */
Chris@69 62 )
Chris@69 63 {
Chris@69 64 /* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */
Chris@69 65 opus_int k = 0;
Chris@69 66 const int32x2_t offset_s32x2 = vdup_n_s32( (1<<14) - 1 );
Chris@69 67 const int32x4_t offset_s32x4 = vcombine_s32( offset_s32x2, offset_s32x2 );
Chris@69 68 int16x4_t in_s16x4 = vdup_n_s16( 0 );
Chris@69 69 int16x4_t out_s16x4;
Chris@69 70 int32x2_t A_Q28_s32x2, A_L_s32x2, A_U_s32x2, B_Q28_s32x2, t_s32x2;
Chris@69 71 int32x4_t A_L_s32x4, A_U_s32x4, B_Q28_s32x4, S_s32x4, out32_Q14_s32x4;
Chris@69 72 int32x2x2_t t0_s32x2x2, t1_s32x2x2, t2_s32x2x2, S_s32x2x2;
Chris@69 73
Chris@69 74 #ifdef OPUS_CHECK_ASM
Chris@69 75 opus_int32 S_c[ 4 ];
Chris@69 76 VARDECL( opus_int16, out_c );
Chris@69 77 SAVE_STACK;
Chris@69 78 ALLOC( out_c, 2 * len, opus_int16 );
Chris@69 79
Chris@69 80 silk_memcpy( &S_c, S, sizeof( S_c ) );
Chris@69 81 silk_biquad_alt_stride2_c( in, B_Q28, A_Q28, S_c, out_c, len );
Chris@69 82 #endif
Chris@69 83
Chris@69 84 /* Negate A_Q28 values and split in two parts */
Chris@69 85 A_Q28_s32x2 = vld1_s32( A_Q28 );
Chris@69 86 A_Q28_s32x2 = vneg_s32( A_Q28_s32x2 );
Chris@69 87 A_L_s32x2 = vshl_n_s32( A_Q28_s32x2, 18 ); /* ( -A_Q28[] & 0x00003FFF ) << 18 */
Chris@69 88 A_L_s32x2 = vreinterpret_s32_u32( vshr_n_u32( vreinterpret_u32_s32( A_L_s32x2 ), 3 ) ); /* ( -A_Q28[] & 0x00003FFF ) << 15 */
Chris@69 89 A_U_s32x2 = vshr_n_s32( A_Q28_s32x2, 14 ); /* silk_RSHIFT( -A_Q28[], 14 ) */
Chris@69 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.) */
Chris@69 91 A_U_s32x2 = vshr_n_s32( A_U_s32x2, 1 ); /* silk_RSHIFT( -A_Q28[], 14 ) << 15 */
Chris@69 92
Chris@69 93 B_Q28_s32x2 = vld1_s32( B_Q28 );
Chris@69 94 t_s32x2 = vld1_s32( B_Q28 + 1 );
Chris@69 95 t0_s32x2x2 = vzip_s32( A_L_s32x2, A_L_s32x2 );
Chris@69 96 t1_s32x2x2 = vzip_s32( A_U_s32x2, A_U_s32x2 );
Chris@69 97 t2_s32x2x2 = vzip_s32( t_s32x2, t_s32x2 );
Chris@69 98 A_L_s32x4 = vcombine_s32( t0_s32x2x2.val[ 0 ], t0_s32x2x2.val[ 1 ] ); /* A{0,0,1,1}_L_Q28 */
Chris@69 99 A_U_s32x4 = vcombine_s32( t1_s32x2x2.val[ 0 ], t1_s32x2x2.val[ 1 ] ); /* A{0,0,1,1}_U_Q28 */
Chris@69 100 B_Q28_s32x4 = vcombine_s32( t2_s32x2x2.val[ 0 ], t2_s32x2x2.val[ 1 ] ); /* B_Q28[ {1,1,2,2} ] */
Chris@69 101 S_s32x4 = vld1q_s32( S ); /* S0 = S[ 0 ]; S3 = S[ 3 ]; */
Chris@69 102 S_s32x2x2 = vtrn_s32( vget_low_s32( S_s32x4 ), vget_high_s32( S_s32x4 ) ); /* S2 = S[ 1 ]; S1 = S[ 2 ]; */
Chris@69 103 S_s32x4 = vcombine_s32( S_s32x2x2.val[ 0 ], S_s32x2x2.val[ 1 ] );
Chris@69 104
Chris@69 105 for( ; k < len - 1; k += 2 ) {
Chris@69 106 int32x4_t in_s32x4[ 2 ], t_s32x4;
Chris@69 107 int32x2_t out32_Q14_s32x2[ 2 ];
Chris@69 108
Chris@69 109 /* S[ 2 * i + 0 ], S[ 2 * i + 1 ], S[ 2 * i + 2 ], S[ 2 * i + 3 ]: Q12 */
Chris@69 110 in_s16x4 = vld1_s16( &in[ 2 * k ] ); /* in{0,1,2,3} = in[ 2 * k + {0,1,2,3} ]; */
Chris@69 111 in_s32x4[ 0 ] = vshll_n_s16( in_s16x4, 15 ); /* in{0,1,2,3} << 15 */
Chris@69 112 t_s32x4 = vqdmulhq_lane_s32( in_s32x4[ 0 ], B_Q28_s32x2, 0 ); /* silk_SMULWB( B_Q28[ 0 ], in{0,1,2,3} ) */
Chris@69 113 in_s32x4[ 1 ] = vcombine_s32( vget_high_s32( in_s32x4[ 0 ] ), vget_high_s32( in_s32x4[ 0 ] ) ); /* in{2,3,2,3} << 15 */
Chris@69 114 in_s32x4[ 0 ] = vcombine_s32( vget_low_s32 ( in_s32x4[ 0 ] ), vget_low_s32 ( in_s32x4[ 0 ] ) ); /* in{0,1,0,1} << 15 */
Chris@69 115 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 ] );
Chris@69 116 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 ] );
Chris@69 117
Chris@69 118 /* Scale back to Q0 and saturate */
Chris@69 119 out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2[ 0 ], out32_Q14_s32x2[ 1 ] ); /* out32_Q14_{0,1,2,3} */
Chris@69 120 out32_Q14_s32x4 = vaddq_s32( out32_Q14_s32x4, offset_s32x4 ); /* out32_Q14_{0,1,2,3} + (1<<14) - 1 */
Chris@69 121 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 ) ) */
Chris@69 122 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 ) ); */
Chris@69 123 }
Chris@69 124
Chris@69 125 /* Process leftover. */
Chris@69 126 if( k < len ) {
Chris@69 127 int32x4_t in_s32x4;
Chris@69 128 int32x2_t out32_Q14_s32x2;
Chris@69 129
Chris@69 130 /* S[ 2 * i + 0 ], S[ 2 * i + 1 ]: Q12 */
Chris@69 131 in_s16x4 = vld1_lane_s16( &in[ 2 * k + 0 ], in_s16x4, 0 ); /* in{0,1} = in[ 2 * k + {0,1} ]; */
Chris@69 132 in_s16x4 = vld1_lane_s16( &in[ 2 * k + 1 ], in_s16x4, 1 ); /* in{0,1} = in[ 2 * k + {0,1} ]; */
Chris@69 133 in_s32x4 = vshll_n_s16( in_s16x4, 15 ); /* in{0,1} << 15 */
Chris@69 134 t_s32x2 = vqdmulh_lane_s32( vget_low_s32( in_s32x4 ), B_Q28_s32x2, 0 ); /* silk_SMULWB( B_Q28[ 0 ], in{0,1} ) */
Chris@69 135 in_s32x4 = vcombine_s32( vget_low_s32( in_s32x4 ), vget_low_s32( in_s32x4 ) ); /* in{0,1,0,1} << 15 */
Chris@69 136 silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, t_s32x2, in_s32x4, &S_s32x4, &out32_Q14_s32x2 );
Chris@69 137
Chris@69 138 /* Scale back to Q0 and saturate */
Chris@69 139 out32_Q14_s32x2 = vadd_s32( out32_Q14_s32x2, offset_s32x2 ); /* out32_Q14_{0,1} + (1<<14) - 1 */
Chris@69 140 out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2, out32_Q14_s32x2 ); /* out32_Q14_{0,1,0,1} + (1<<14) - 1 */
Chris@69 141 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 ) ) */
Chris@69 142 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 ) ); */
Chris@69 143 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 ) ); */
Chris@69 144 }
Chris@69 145
Chris@69 146 vst1q_lane_s32( &S[ 0 ], S_s32x4, 0 ); /* S[ 0 ] = S0; */
Chris@69 147 vst1q_lane_s32( &S[ 1 ], S_s32x4, 2 ); /* S[ 1 ] = S2; */
Chris@69 148 vst1q_lane_s32( &S[ 2 ], S_s32x4, 1 ); /* S[ 2 ] = S1; */
Chris@69 149 vst1q_lane_s32( &S[ 3 ], S_s32x4, 3 ); /* S[ 3 ] = S3; */
Chris@69 150
Chris@69 151 #ifdef OPUS_CHECK_ASM
Chris@69 152 silk_assert( !memcmp( S_c, S, sizeof( S_c ) ) );
Chris@69 153 silk_assert( !memcmp( out_c, out, 2 * len * sizeof( opus_int16 ) ) );
Chris@69 154 RESTORE_STACK;
Chris@69 155 #endif
Chris@69 156 }