• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include "SigProc_FIX.h"
34 #include "define.h"
35 
36 #define QA                          24
37 #define A_LIMIT                     SILK_FIX_CONST( 0.99975, QA )
38 
39 #define MUL32_FRAC_Q(a32, b32, Q)   ((opus_int32)(silk_RSHIFT_ROUND64(silk_SMULL(a32, b32), Q)))
40 
41 /* The difficulty is how to judge a 64-bit signed integer tmp64 is 32-bit overflowed,
42  * since NEON has no 64-bit min, max or comparison instructions.
43  * A failed idea is to compare the results of vmovn(tmp64) and vqmovn(tmp64) whether they are equal or not.
44  * However, this idea fails when the tmp64 is something like 0xFFFFFFF980000000.
45  * Here we know that mult2Q >= 1, so the highest bit (bit 63, sign bit) of tmp64 must equal to bit 62.
46  * tmp64 was shifted left by 1 and we got tmp64'. If high_half(tmp64') != 0 and high_half(tmp64') != -1,
47  * then we know that bit 31 to bit 63 of tmp64 can not all be the sign bit, and therefore tmp64 is 32-bit overflowed.
48  * That is, we judge if tmp64' > 0x00000000FFFFFFFF, or tmp64' <= 0xFFFFFFFF00000000.
49  * We use narrowing shift right 31 bits to tmp32' to save data bandwidth and instructions.
50  * That is, we judge if tmp32' > 0x00000000, or tmp32' <= 0xFFFFFFFF.
51  */
52 
53 /* Compute inverse of LPC prediction gain, and                          */
54 /* test if LPC coefficients are stable (all poles within unit circle)   */
LPC_inverse_pred_gain_QA_neon(opus_int32 A_QA[SILK_MAX_ORDER_LPC],const opus_int order)55 static OPUS_INLINE opus_int32 LPC_inverse_pred_gain_QA_neon( /* O   Returns inverse prediction gain in energy domain, Q30    */
56     opus_int32           A_QA[ SILK_MAX_ORDER_LPC ],         /* I   Prediction coefficients                                  */
57     const opus_int       order                               /* I   Prediction order                                         */
58 )
59 {
60     opus_int   k, n, mult2Q;
61     opus_int32 invGain_Q30, rc_Q31, rc_mult1_Q30, rc_mult2, tmp1, tmp2;
62     opus_int32 max, min;
63     int32x4_t  max_s32x4, min_s32x4;
64     int32x2_t  max_s32x2, min_s32x2;
65 
66     max_s32x4 = vdupq_n_s32( silk_int32_MIN );
67     min_s32x4 = vdupq_n_s32( silk_int32_MAX );
68     invGain_Q30 = SILK_FIX_CONST( 1, 30 );
69     for( k = order - 1; k > 0; k-- ) {
70         int32x2_t rc_Q31_s32x2, rc_mult2_s32x2;
71         int64x2_t mult2Q_s64x2;
72 
73         /* Check for stability */
74         if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) {
75             return 0;
76         }
77 
78         /* Set RC equal to negated AR coef */
79         rc_Q31 = -silk_LSHIFT( A_QA[ k ], 31 - QA );
80 
81         /* rc_mult1_Q30 range: [ 1 : 2^30 ] */
82         rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) );
83         silk_assert( rc_mult1_Q30 > ( 1 << 15 ) );                   /* reduce A_LIMIT if fails */
84         silk_assert( rc_mult1_Q30 <= ( 1 << 30 ) );
85 
86         /* Update inverse gain */
87         /* invGain_Q30 range: [ 0 : 2^30 ] */
88         invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 );
89         silk_assert( invGain_Q30 >= 0           );
90         silk_assert( invGain_Q30 <= ( 1 << 30 ) );
91         if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) {
92             return 0;
93         }
94 
95         /* rc_mult2 range: [ 2^30 : silk_int32_MAX ] */
96         mult2Q = 32 - silk_CLZ32( silk_abs( rc_mult1_Q30 ) );
97         rc_mult2 = silk_INVERSE32_varQ( rc_mult1_Q30, mult2Q + 30 );
98 
99         /* Update AR coefficient */
100         rc_Q31_s32x2   = vdup_n_s32( rc_Q31 );
101         mult2Q_s64x2   = vdupq_n_s64( -mult2Q );
102         rc_mult2_s32x2 = vdup_n_s32( rc_mult2 );
103 
104         for( n = 0; n < ( ( k + 1 ) >> 1 ) - 3; n += 4 ) {
105             /* We always calculate extra elements of A_QA buffer when ( k % 4 ) != 0, to take the advantage of SIMD parallelization. */
106             int32x4_t tmp1_s32x4, tmp2_s32x4, t0_s32x4, t1_s32x4, s0_s32x4, s1_s32x4, t_QA0_s32x4, t_QA1_s32x4;
107             int64x2_t t0_s64x2, t1_s64x2, t2_s64x2, t3_s64x2;
108             tmp1_s32x4  = vld1q_s32( A_QA + n );
109             tmp2_s32x4  = vld1q_s32( A_QA + k - n - 4 );
110             tmp2_s32x4  = vrev64q_s32( tmp2_s32x4 );
111             tmp2_s32x4  = vcombine_s32( vget_high_s32( tmp2_s32x4 ), vget_low_s32( tmp2_s32x4 ) );
112             t0_s32x4    = vqrdmulhq_lane_s32( tmp2_s32x4, rc_Q31_s32x2, 0 );
113             t1_s32x4    = vqrdmulhq_lane_s32( tmp1_s32x4, rc_Q31_s32x2, 0 );
114             t_QA0_s32x4 = vqsubq_s32( tmp1_s32x4, t0_s32x4 );
115             t_QA1_s32x4 = vqsubq_s32( tmp2_s32x4, t1_s32x4 );
116             t0_s64x2    = vmull_s32( vget_low_s32 ( t_QA0_s32x4 ), rc_mult2_s32x2 );
117             t1_s64x2    = vmull_s32( vget_high_s32( t_QA0_s32x4 ), rc_mult2_s32x2 );
118             t2_s64x2    = vmull_s32( vget_low_s32 ( t_QA1_s32x4 ), rc_mult2_s32x2 );
119             t3_s64x2    = vmull_s32( vget_high_s32( t_QA1_s32x4 ), rc_mult2_s32x2 );
120             t0_s64x2    = vrshlq_s64( t0_s64x2, mult2Q_s64x2 );
121             t1_s64x2    = vrshlq_s64( t1_s64x2, mult2Q_s64x2 );
122             t2_s64x2    = vrshlq_s64( t2_s64x2, mult2Q_s64x2 );
123             t3_s64x2    = vrshlq_s64( t3_s64x2, mult2Q_s64x2 );
124             t0_s32x4    = vcombine_s32( vmovn_s64( t0_s64x2 ), vmovn_s64( t1_s64x2 ) );
125             t1_s32x4    = vcombine_s32( vmovn_s64( t2_s64x2 ), vmovn_s64( t3_s64x2 ) );
126             s0_s32x4    = vcombine_s32( vshrn_n_s64( t0_s64x2, 31 ), vshrn_n_s64( t1_s64x2, 31 ) );
127             s1_s32x4    = vcombine_s32( vshrn_n_s64( t2_s64x2, 31 ), vshrn_n_s64( t3_s64x2, 31 ) );
128             max_s32x4   = vmaxq_s32( max_s32x4, s0_s32x4 );
129             min_s32x4   = vminq_s32( min_s32x4, s0_s32x4 );
130             max_s32x4   = vmaxq_s32( max_s32x4, s1_s32x4 );
131             min_s32x4   = vminq_s32( min_s32x4, s1_s32x4 );
132             t1_s32x4    = vrev64q_s32( t1_s32x4 );
133             t1_s32x4    = vcombine_s32( vget_high_s32( t1_s32x4 ), vget_low_s32( t1_s32x4 ) );
134             vst1q_s32( A_QA + n,         t0_s32x4 );
135             vst1q_s32( A_QA + k - n - 4, t1_s32x4 );
136         }
137         for( ; n < (k + 1) >> 1; n++ ) {
138             opus_int64 tmp64;
139             tmp1 = A_QA[ n ];
140             tmp2 = A_QA[ k - n - 1 ];
141             tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp1,
142                   MUL32_FRAC_Q( tmp2, rc_Q31, 31 ) ), rc_mult2 ), mult2Q);
143             if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) {
144                return 0;
145             }
146             A_QA[ n ] = ( opus_int32 )tmp64;
147             tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp2,
148                   MUL32_FRAC_Q( tmp1, rc_Q31, 31 ) ), rc_mult2), mult2Q);
149             if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) {
150                return 0;
151             }
152             A_QA[ k - n - 1 ] = ( opus_int32 )tmp64;
153         }
154     }
155 
156     /* Check for stability */
157     if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) {
158         return 0;
159     }
160 
161     max_s32x2 = vmax_s32( vget_low_s32( max_s32x4 ), vget_high_s32( max_s32x4 ) );
162     min_s32x2 = vmin_s32( vget_low_s32( min_s32x4 ), vget_high_s32( min_s32x4 ) );
163     max_s32x2 = vmax_s32( max_s32x2, vreinterpret_s32_s64( vshr_n_s64( vreinterpret_s64_s32( max_s32x2 ), 32 ) ) );
164     min_s32x2 = vmin_s32( min_s32x2, vreinterpret_s32_s64( vshr_n_s64( vreinterpret_s64_s32( min_s32x2 ), 32 ) ) );
165     max = vget_lane_s32( max_s32x2, 0 );
166     min = vget_lane_s32( min_s32x2, 0 );
167     if( ( max > 0 ) || ( min < -1 ) ) {
168         return 0;
169     }
170 
171     /* Set RC equal to negated AR coef */
172     rc_Q31 = -silk_LSHIFT( A_QA[ 0 ], 31 - QA );
173 
174     /* Range: [ 1 : 2^30 ] */
175     rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) );
176 
177     /* Update inverse gain */
178     /* Range: [ 0 : 2^30 ] */
179     invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 );
180     silk_assert( invGain_Q30 >= 0           );
181     silk_assert( invGain_Q30 <= ( 1 << 30 ) );
182     if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) {
183         return 0;
184     }
185 
186     return invGain_Q30;
187 }
188 
189 /* For input in Q12 domain */
silk_LPC_inverse_pred_gain_neon(const opus_int16 * A_Q12,const opus_int order)190 opus_int32 silk_LPC_inverse_pred_gain_neon(         /* O   Returns inverse prediction gain in energy domain, Q30        */
191     const opus_int16            *A_Q12,             /* I   Prediction coefficients, Q12 [order]                         */
192     const opus_int              order               /* I   Prediction order                                             */
193 )
194 {
195 #ifdef OPUS_CHECK_ASM
196     const opus_int32 invGain_Q30_c = silk_LPC_inverse_pred_gain_c( A_Q12, order );
197 #endif
198 
199     opus_int32 invGain_Q30;
200     if( ( SILK_MAX_ORDER_LPC != 24 ) || ( order & 1 )) {
201         invGain_Q30 = silk_LPC_inverse_pred_gain_c( A_Q12, order );
202     }
203     else {
204         opus_int32 Atmp_QA[ SILK_MAX_ORDER_LPC ];
205         opus_int32 DC_resp;
206         int16x8_t  t0_s16x8, t1_s16x8, t2_s16x8;
207         int32x4_t  t0_s32x4;
208         const opus_int leftover = order & 7;
209 
210         /* Increase Q domain of the AR coefficients */
211         t0_s16x8 = vld1q_s16( A_Q12 +  0 );
212         t1_s16x8 = vld1q_s16( A_Q12 +  8 );
213         t2_s16x8 = vld1q_s16( A_Q12 + 16 );
214         t0_s32x4 = vpaddlq_s16( t0_s16x8 );
215 
216         switch( order - leftover )
217         {
218         case 24:
219             t0_s32x4 = vpadalq_s16( t0_s32x4, t2_s16x8 );
220             /* FALLTHROUGH */
221 
222         case 16:
223             t0_s32x4 = vpadalq_s16( t0_s32x4, t1_s16x8 );
224             vst1q_s32( Atmp_QA + 16, vshll_n_s16( vget_low_s16 ( t2_s16x8 ), QA - 12 ) );
225             vst1q_s32( Atmp_QA + 20, vshll_n_s16( vget_high_s16( t2_s16x8 ), QA - 12 ) );
226             /* FALLTHROUGH */
227 
228         case 8:
229         {
230             const int32x2_t t_s32x2 = vpadd_s32( vget_low_s32( t0_s32x4 ), vget_high_s32( t0_s32x4 ) );
231             const int64x1_t t_s64x1 = vpaddl_s32( t_s32x2 );
232             DC_resp = vget_lane_s32( vreinterpret_s32_s64( t_s64x1 ), 0 );
233             vst1q_s32( Atmp_QA +  8, vshll_n_s16( vget_low_s16 ( t1_s16x8 ), QA - 12 ) );
234             vst1q_s32( Atmp_QA + 12, vshll_n_s16( vget_high_s16( t1_s16x8 ), QA - 12 ) );
235         }
236         break;
237 
238         default:
239             DC_resp = 0;
240             break;
241         }
242         A_Q12 += order - leftover;
243 
244         switch( leftover )
245         {
246         case 6:
247             DC_resp += (opus_int32)A_Q12[ 5 ];
248             DC_resp += (opus_int32)A_Q12[ 4 ];
249             /* FALLTHROUGH */
250 
251         case 4:
252             DC_resp += (opus_int32)A_Q12[ 3 ];
253             DC_resp += (opus_int32)A_Q12[ 2 ];
254             /* FALLTHROUGH */
255 
256         case 2:
257             DC_resp += (opus_int32)A_Q12[ 1 ];
258             DC_resp += (opus_int32)A_Q12[ 0 ];
259             /* FALLTHROUGH */
260 
261         default:
262             break;
263         }
264 
265         /* If the DC is unstable, we don't even need to do the full calculations */
266         if( DC_resp >= 4096 ) {
267             invGain_Q30 = 0;
268         } else {
269             vst1q_s32( Atmp_QA + 0, vshll_n_s16( vget_low_s16 ( t0_s16x8 ), QA - 12 ) );
270             vst1q_s32( Atmp_QA + 4, vshll_n_s16( vget_high_s16( t0_s16x8 ), QA - 12 ) );
271             invGain_Q30 = LPC_inverse_pred_gain_QA_neon( Atmp_QA, order );
272         }
273     }
274 
275 #ifdef OPUS_CHECK_ASM
276     silk_assert( invGain_Q30_c == invGain_Q30 );
277 #endif
278 
279     return invGain_Q30;
280 }
281