LPC_inv_pred_gain_neon_intr.c (13314B)
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) */ 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 */ 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 if ( order > 16 ) { 214 t2_s16x8 = vld1q_s16( A_Q12 + 16 ); 215 } 216 t0_s32x4 = vpaddlq_s16( t0_s16x8 ); 217 218 switch( order - leftover ) 219 { 220 case 24: 221 t0_s32x4 = vpadalq_s16( t0_s32x4, t2_s16x8 ); 222 vst1q_s32( Atmp_QA + 16, vshll_n_s16( vget_low_s16 ( t2_s16x8 ), QA - 12 ) ); 223 vst1q_s32( Atmp_QA + 20, vshll_n_s16( vget_high_s16( t2_s16x8 ), QA - 12 ) ); 224 /* FALLTHROUGH */ 225 226 case 16: 227 t0_s32x4 = vpadalq_s16( t0_s32x4, t1_s16x8 ); 228 vst1q_s32( Atmp_QA + 8, vshll_n_s16( vget_low_s16 ( t1_s16x8 ), QA - 12 ) ); 229 vst1q_s32( Atmp_QA + 12, vshll_n_s16( vget_high_s16( t1_s16x8 ), QA - 12 ) ); 230 /* FALLTHROUGH */ 231 232 case 8: 233 { 234 const int32x2_t t_s32x2 = vpadd_s32( vget_low_s32( t0_s32x4 ), vget_high_s32( t0_s32x4 ) ); 235 const int64x1_t t_s64x1 = vpaddl_s32( t_s32x2 ); 236 DC_resp = vget_lane_s32( vreinterpret_s32_s64( t_s64x1 ), 0 ); 237 vst1q_s32( Atmp_QA + 0, vshll_n_s16( vget_low_s16 ( t0_s16x8 ), QA - 12 ) ); 238 vst1q_s32( Atmp_QA + 4, vshll_n_s16( vget_high_s16( t0_s16x8 ), QA - 12 ) ); 239 } 240 break; 241 242 default: 243 DC_resp = 0; 244 break; 245 } 246 A_Q12 += order - leftover; 247 248 switch( leftover ) 249 { 250 case 6: 251 DC_resp += (opus_int32)A_Q12[ 5 ]; 252 DC_resp += (opus_int32)A_Q12[ 4 ]; 253 Atmp_QA[ order - leftover + 5 ] = silk_LSHIFT32( (opus_int32)A_Q12[ 5 ], QA - 12 ); 254 Atmp_QA[ order - leftover + 4 ] = silk_LSHIFT32( (opus_int32)A_Q12[ 4 ], QA - 12 ); 255 /* FALLTHROUGH */ 256 257 case 4: 258 DC_resp += (opus_int32)A_Q12[ 3 ]; 259 DC_resp += (opus_int32)A_Q12[ 2 ]; 260 Atmp_QA[ order - leftover + 3 ] = silk_LSHIFT32( (opus_int32)A_Q12[ 3 ], QA - 12 ); 261 Atmp_QA[ order - leftover + 2 ] = silk_LSHIFT32( (opus_int32)A_Q12[ 2 ], QA - 12 ); 262 /* FALLTHROUGH */ 263 264 case 2: 265 DC_resp += (opus_int32)A_Q12[ 1 ]; 266 DC_resp += (opus_int32)A_Q12[ 0 ]; 267 Atmp_QA[ order - leftover + 1 ] = silk_LSHIFT32( (opus_int32)A_Q12[ 1 ], QA - 12 ); 268 Atmp_QA[ order - leftover + 0 ] = silk_LSHIFT32( (opus_int32)A_Q12[ 0 ], QA - 12 ); 269 /* FALLTHROUGH */ 270 271 default: 272 break; 273 } 274 275 /* If the DC is unstable, we don't even need to do the full calculations */ 276 if( DC_resp >= 4096 ) { 277 invGain_Q30 = 0; 278 } else { 279 invGain_Q30 = LPC_inverse_pred_gain_QA_neon( Atmp_QA, order ); 280 } 281 } 282 283 #ifdef OPUS_CHECK_ASM 284 silk_assert( invGain_Q30_c == invGain_Q30 ); 285 #endif 286 287 return invGain_Q30; 288 }