Initial community commit

This commit is contained in:
Jef 2024-09-24 14:54:57 +02:00
parent 537bcbc862
commit fc06254474
16440 changed files with 4239995 additions and 2 deletions

View file

@ -0,0 +1,57 @@
/***********************************************************************
Copyright (c) 2017 Google Inc.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifndef SILK_LPC_INV_PRED_GAIN_ARM_H
# define SILK_LPC_INV_PRED_GAIN_ARM_H
# include "celt/arm/armcpu.h"
# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR)
opus_int32 silk_LPC_inverse_pred_gain_neon( /* O Returns inverse prediction gain in energy domain, Q30 */
const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */
const opus_int order /* I Prediction order */
);
# if !defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_PRESUME_NEON)
# define OVERRIDE_silk_LPC_inverse_pred_gain (1)
# define silk_LPC_inverse_pred_gain(A_Q12, order, arch) ((void)(arch), PRESUME_NEON(silk_LPC_inverse_pred_gain)(A_Q12, order))
# endif
# endif
# if !defined(OVERRIDE_silk_LPC_inverse_pred_gain)
/*Is run-time CPU detection enabled on this platform?*/
# if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR))
extern opus_int32 (*const SILK_LPC_INVERSE_PRED_GAIN_IMPL[OPUS_ARCHMASK+1])(const opus_int16 *A_Q12, const opus_int order);
# define OVERRIDE_silk_LPC_inverse_pred_gain (1)
# define silk_LPC_inverse_pred_gain(A_Q12, order, arch) ((*SILK_LPC_INVERSE_PRED_GAIN_IMPL[(arch)&OPUS_ARCHMASK])(A_Q12, order))
# elif defined(OPUS_ARM_PRESUME_NEON_INTR)
# define OVERRIDE_silk_LPC_inverse_pred_gain (1)
# define silk_LPC_inverse_pred_gain(A_Q12, order, arch) ((void)(arch), silk_LPC_inverse_pred_gain_neon(A_Q12, order))
# endif
# endif
#endif /* end SILK_LPC_INV_PRED_GAIN_ARM_H */

View file

@ -0,0 +1,280 @@
/***********************************************************************
Copyright (c) 2017 Google Inc.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include <arm_neon.h>
#include "SigProc_FIX.h"
#include "define.h"
#define QA 24
#define A_LIMIT SILK_FIX_CONST( 0.99975, QA )
#define MUL32_FRAC_Q(a32, b32, Q) ((opus_int32)(silk_RSHIFT_ROUND64(silk_SMULL(a32, b32), Q)))
/* The difficulty is how to judge a 64-bit signed integer tmp64 is 32-bit overflowed,
* since NEON has no 64-bit min, max or comparison instructions.
* A failed idea is to compare the results of vmovn(tmp64) and vqmovn(tmp64) whether they are equal or not.
* However, this idea fails when the tmp64 is something like 0xFFFFFFF980000000.
* Here we know that mult2Q >= 1, so the highest bit (bit 63, sign bit) of tmp64 must equal to bit 62.
* tmp64 was shifted left by 1 and we got tmp64'. If high_half(tmp64') != 0 and high_half(tmp64') != -1,
* 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.
* That is, we judge if tmp64' > 0x00000000FFFFFFFF, or tmp64' <= 0xFFFFFFFF00000000.
* We use narrowing shift right 31 bits to tmp32' to save data bandwidth and instructions.
* That is, we judge if tmp32' > 0x00000000, or tmp32' <= 0xFFFFFFFF.
*/
/* Compute inverse of LPC prediction gain, and */
/* test if LPC coefficients are stable (all poles within unit circle) */
static OPUS_INLINE opus_int32 LPC_inverse_pred_gain_QA_neon( /* O Returns inverse prediction gain in energy domain, Q30 */
opus_int32 A_QA[ SILK_MAX_ORDER_LPC ], /* I Prediction coefficients */
const opus_int order /* I Prediction order */
)
{
opus_int k, n, mult2Q;
opus_int32 invGain_Q30, rc_Q31, rc_mult1_Q30, rc_mult2, tmp1, tmp2;
opus_int32 max, min;
int32x4_t max_s32x4, min_s32x4;
int32x2_t max_s32x2, min_s32x2;
max_s32x4 = vdupq_n_s32( silk_int32_MIN );
min_s32x4 = vdupq_n_s32( silk_int32_MAX );
invGain_Q30 = SILK_FIX_CONST( 1, 30 );
for( k = order - 1; k > 0; k-- ) {
int32x2_t rc_Q31_s32x2, rc_mult2_s32x2;
int64x2_t mult2Q_s64x2;
/* Check for stability */
if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) {
return 0;
}
/* Set RC equal to negated AR coef */
rc_Q31 = -silk_LSHIFT( A_QA[ k ], 31 - QA );
/* rc_mult1_Q30 range: [ 1 : 2^30 ] */
rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) );
silk_assert( rc_mult1_Q30 > ( 1 << 15 ) ); /* reduce A_LIMIT if fails */
silk_assert( rc_mult1_Q30 <= ( 1 << 30 ) );
/* Update inverse gain */
/* invGain_Q30 range: [ 0 : 2^30 ] */
invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 );
silk_assert( invGain_Q30 >= 0 );
silk_assert( invGain_Q30 <= ( 1 << 30 ) );
if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) {
return 0;
}
/* rc_mult2 range: [ 2^30 : silk_int32_MAX ] */
mult2Q = 32 - silk_CLZ32( silk_abs( rc_mult1_Q30 ) );
rc_mult2 = silk_INVERSE32_varQ( rc_mult1_Q30, mult2Q + 30 );
/* Update AR coefficient */
rc_Q31_s32x2 = vdup_n_s32( rc_Q31 );
mult2Q_s64x2 = vdupq_n_s64( -mult2Q );
rc_mult2_s32x2 = vdup_n_s32( rc_mult2 );
for( n = 0; n < ( ( k + 1 ) >> 1 ) - 3; n += 4 ) {
/* We always calculate extra elements of A_QA buffer when ( k % 4 ) != 0, to take the advantage of SIMD parallelization. */
int32x4_t tmp1_s32x4, tmp2_s32x4, t0_s32x4, t1_s32x4, s0_s32x4, s1_s32x4, t_QA0_s32x4, t_QA1_s32x4;
int64x2_t t0_s64x2, t1_s64x2, t2_s64x2, t3_s64x2;
tmp1_s32x4 = vld1q_s32( A_QA + n );
tmp2_s32x4 = vld1q_s32( A_QA + k - n - 4 );
tmp2_s32x4 = vrev64q_s32( tmp2_s32x4 );
tmp2_s32x4 = vcombine_s32( vget_high_s32( tmp2_s32x4 ), vget_low_s32( tmp2_s32x4 ) );
t0_s32x4 = vqrdmulhq_lane_s32( tmp2_s32x4, rc_Q31_s32x2, 0 );
t1_s32x4 = vqrdmulhq_lane_s32( tmp1_s32x4, rc_Q31_s32x2, 0 );
t_QA0_s32x4 = vqsubq_s32( tmp1_s32x4, t0_s32x4 );
t_QA1_s32x4 = vqsubq_s32( tmp2_s32x4, t1_s32x4 );
t0_s64x2 = vmull_s32( vget_low_s32 ( t_QA0_s32x4 ), rc_mult2_s32x2 );
t1_s64x2 = vmull_s32( vget_high_s32( t_QA0_s32x4 ), rc_mult2_s32x2 );
t2_s64x2 = vmull_s32( vget_low_s32 ( t_QA1_s32x4 ), rc_mult2_s32x2 );
t3_s64x2 = vmull_s32( vget_high_s32( t_QA1_s32x4 ), rc_mult2_s32x2 );
t0_s64x2 = vrshlq_s64( t0_s64x2, mult2Q_s64x2 );
t1_s64x2 = vrshlq_s64( t1_s64x2, mult2Q_s64x2 );
t2_s64x2 = vrshlq_s64( t2_s64x2, mult2Q_s64x2 );
t3_s64x2 = vrshlq_s64( t3_s64x2, mult2Q_s64x2 );
t0_s32x4 = vcombine_s32( vmovn_s64( t0_s64x2 ), vmovn_s64( t1_s64x2 ) );
t1_s32x4 = vcombine_s32( vmovn_s64( t2_s64x2 ), vmovn_s64( t3_s64x2 ) );
s0_s32x4 = vcombine_s32( vshrn_n_s64( t0_s64x2, 31 ), vshrn_n_s64( t1_s64x2, 31 ) );
s1_s32x4 = vcombine_s32( vshrn_n_s64( t2_s64x2, 31 ), vshrn_n_s64( t3_s64x2, 31 ) );
max_s32x4 = vmaxq_s32( max_s32x4, s0_s32x4 );
min_s32x4 = vminq_s32( min_s32x4, s0_s32x4 );
max_s32x4 = vmaxq_s32( max_s32x4, s1_s32x4 );
min_s32x4 = vminq_s32( min_s32x4, s1_s32x4 );
t1_s32x4 = vrev64q_s32( t1_s32x4 );
t1_s32x4 = vcombine_s32( vget_high_s32( t1_s32x4 ), vget_low_s32( t1_s32x4 ) );
vst1q_s32( A_QA + n, t0_s32x4 );
vst1q_s32( A_QA + k - n - 4, t1_s32x4 );
}
for( ; n < (k + 1) >> 1; n++ ) {
opus_int64 tmp64;
tmp1 = A_QA[ n ];
tmp2 = A_QA[ k - n - 1 ];
tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp1,
MUL32_FRAC_Q( tmp2, rc_Q31, 31 ) ), rc_mult2 ), mult2Q);
if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) {
return 0;
}
A_QA[ n ] = ( opus_int32 )tmp64;
tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp2,
MUL32_FRAC_Q( tmp1, rc_Q31, 31 ) ), rc_mult2), mult2Q);
if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) {
return 0;
}
A_QA[ k - n - 1 ] = ( opus_int32 )tmp64;
}
}
/* Check for stability */
if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) {
return 0;
}
max_s32x2 = vmax_s32( vget_low_s32( max_s32x4 ), vget_high_s32( max_s32x4 ) );
min_s32x2 = vmin_s32( vget_low_s32( min_s32x4 ), vget_high_s32( min_s32x4 ) );
max_s32x2 = vmax_s32( max_s32x2, vreinterpret_s32_s64( vshr_n_s64( vreinterpret_s64_s32( max_s32x2 ), 32 ) ) );
min_s32x2 = vmin_s32( min_s32x2, vreinterpret_s32_s64( vshr_n_s64( vreinterpret_s64_s32( min_s32x2 ), 32 ) ) );
max = vget_lane_s32( max_s32x2, 0 );
min = vget_lane_s32( min_s32x2, 0 );
if( ( max > 0 ) || ( min < -1 ) ) {
return 0;
}
/* Set RC equal to negated AR coef */
rc_Q31 = -silk_LSHIFT( A_QA[ 0 ], 31 - QA );
/* Range: [ 1 : 2^30 ] */
rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) );
/* Update inverse gain */
/* Range: [ 0 : 2^30 ] */
invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 );
silk_assert( invGain_Q30 >= 0 );
silk_assert( invGain_Q30 <= ( 1 << 30 ) );
if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) {
return 0;
}
return invGain_Q30;
}
/* For input in Q12 domain */
opus_int32 silk_LPC_inverse_pred_gain_neon( /* O Returns inverse prediction gain in energy domain, Q30 */
const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */
const opus_int order /* I Prediction order */
)
{
#ifdef OPUS_CHECK_ASM
const opus_int32 invGain_Q30_c = silk_LPC_inverse_pred_gain_c( A_Q12, order );
#endif
opus_int32 invGain_Q30;
if( ( SILK_MAX_ORDER_LPC != 24 ) || ( order & 1 )) {
invGain_Q30 = silk_LPC_inverse_pred_gain_c( A_Q12, order );
}
else {
opus_int32 Atmp_QA[ SILK_MAX_ORDER_LPC ];
opus_int32 DC_resp;
int16x8_t t0_s16x8, t1_s16x8, t2_s16x8;
int32x4_t t0_s32x4;
const opus_int leftover = order & 7;
/* Increase Q domain of the AR coefficients */
t0_s16x8 = vld1q_s16( A_Q12 + 0 );
t1_s16x8 = vld1q_s16( A_Q12 + 8 );
t2_s16x8 = vld1q_s16( A_Q12 + 16 );
t0_s32x4 = vpaddlq_s16( t0_s16x8 );
switch( order - leftover )
{
case 24:
t0_s32x4 = vpadalq_s16( t0_s32x4, t2_s16x8 );
/* FALLTHROUGH */
case 16:
t0_s32x4 = vpadalq_s16( t0_s32x4, t1_s16x8 );
vst1q_s32( Atmp_QA + 16, vshll_n_s16( vget_low_s16 ( t2_s16x8 ), QA - 12 ) );
vst1q_s32( Atmp_QA + 20, vshll_n_s16( vget_high_s16( t2_s16x8 ), QA - 12 ) );
/* FALLTHROUGH */
case 8:
{
const int32x2_t t_s32x2 = vpadd_s32( vget_low_s32( t0_s32x4 ), vget_high_s32( t0_s32x4 ) );
const int64x1_t t_s64x1 = vpaddl_s32( t_s32x2 );
DC_resp = vget_lane_s32( vreinterpret_s32_s64( t_s64x1 ), 0 );
vst1q_s32( Atmp_QA + 8, vshll_n_s16( vget_low_s16 ( t1_s16x8 ), QA - 12 ) );
vst1q_s32( Atmp_QA + 12, vshll_n_s16( vget_high_s16( t1_s16x8 ), QA - 12 ) );
}
break;
default:
DC_resp = 0;
break;
}
A_Q12 += order - leftover;
switch( leftover )
{
case 6:
DC_resp += (opus_int32)A_Q12[ 5 ];
DC_resp += (opus_int32)A_Q12[ 4 ];
/* FALLTHROUGH */
case 4:
DC_resp += (opus_int32)A_Q12[ 3 ];
DC_resp += (opus_int32)A_Q12[ 2 ];
/* FALLTHROUGH */
case 2:
DC_resp += (opus_int32)A_Q12[ 1 ];
DC_resp += (opus_int32)A_Q12[ 0 ];
/* FALLTHROUGH */
default:
break;
}
/* If the DC is unstable, we don't even need to do the full calculations */
if( DC_resp >= 4096 ) {
invGain_Q30 = 0;
} else {
vst1q_s32( Atmp_QA + 0, vshll_n_s16( vget_low_s16 ( t0_s16x8 ), QA - 12 ) );
vst1q_s32( Atmp_QA + 4, vshll_n_s16( vget_high_s16( t0_s16x8 ), QA - 12 ) );
invGain_Q30 = LPC_inverse_pred_gain_QA_neon( Atmp_QA, order );
}
}
#ifdef OPUS_CHECK_ASM
silk_assert( invGain_Q30_c == invGain_Q30 );
#endif
return invGain_Q30;
}

View file

@ -0,0 +1,100 @@
/***********************************************************************
Copyright (c) 2017 Google Inc.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifndef SILK_NSQ_DEL_DEC_ARM_H
#define SILK_NSQ_DEL_DEC_ARM_H
#include "celt/arm/armcpu.h"
#if defined(OPUS_ARM_MAY_HAVE_NEON_INTR)
void silk_NSQ_del_dec_neon(
const silk_encoder_state *psEncC, silk_nsq_state *NSQ,
SideInfoIndices *psIndices, const opus_int16 x16[], opus_int8 pulses[],
const opus_int16 PredCoef_Q12[2 * MAX_LPC_ORDER],
const opus_int16 LTPCoef_Q14[LTP_ORDER * MAX_NB_SUBFR],
const opus_int16 AR_Q13[MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER],
const opus_int HarmShapeGain_Q14[MAX_NB_SUBFR],
const opus_int Tilt_Q14[MAX_NB_SUBFR],
const opus_int32 LF_shp_Q14[MAX_NB_SUBFR],
const opus_int32 Gains_Q16[MAX_NB_SUBFR],
const opus_int pitchL[MAX_NB_SUBFR], const opus_int Lambda_Q10,
const opus_int LTP_scale_Q14);
#if !defined(OPUS_HAVE_RTCD)
#define OVERRIDE_silk_NSQ_del_dec (1)
#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, \
LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, \
LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, \
LTP_scale_Q14, arch) \
((void)(arch), \
PRESUME_NEON(silk_NSQ_del_dec)( \
psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, \
AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, \
Lambda_Q10, LTP_scale_Q14))
#endif
#endif
#if !defined(OVERRIDE_silk_NSQ_del_dec)
/*Is run-time CPU detection enabled on this platform?*/
#if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && \
!defined(OPUS_ARM_PRESUME_NEON_INTR))
extern void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])(
const silk_encoder_state *psEncC, silk_nsq_state *NSQ,
SideInfoIndices *psIndices, const opus_int16 x16[], opus_int8 pulses[],
const opus_int16 PredCoef_Q12[2 * MAX_LPC_ORDER],
const opus_int16 LTPCoef_Q14[LTP_ORDER * MAX_NB_SUBFR],
const opus_int16 AR_Q13[MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER],
const opus_int HarmShapeGain_Q14[MAX_NB_SUBFR],
const opus_int Tilt_Q14[MAX_NB_SUBFR],
const opus_int32 LF_shp_Q14[MAX_NB_SUBFR],
const opus_int32 Gains_Q16[MAX_NB_SUBFR],
const opus_int pitchL[MAX_NB_SUBFR], const opus_int Lambda_Q10,
const opus_int LTP_scale_Q14);
#define OVERRIDE_silk_NSQ_del_dec (1)
#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, \
LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, \
LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, \
LTP_scale_Q14, arch) \
((*SILK_NSQ_DEL_DEC_IMPL[(arch)&OPUS_ARCHMASK])( \
psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, LTPCoef_Q14, \
AR_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, pitchL, \
Lambda_Q10, LTP_scale_Q14))
#elif defined(OPUS_ARM_PRESUME_NEON_INTR)
#define OVERRIDE_silk_NSQ_del_dec (1)
#define silk_NSQ_del_dec(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, \
LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, \
LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, \
LTP_scale_Q14, arch) \
((void)(arch), \
silk_NSQ_del_dec_neon(psEncC, NSQ, psIndices, x16, pulses, PredCoef_Q12, \
LTPCoef_Q14, AR_Q13, HarmShapeGain_Q14, Tilt_Q14, \
LF_shp_Q14, Gains_Q16, pitchL, Lambda_Q10, \
LTP_scale_Q14))
#endif
#endif
#endif /* end SILK_NSQ_DEL_DEC_ARM_H */

File diff suppressed because it is too large Load diff

View file

@ -0,0 +1,112 @@
/***********************************************************************
Copyright (C) 2014 Vidyo
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include <arm_neon.h>
#include "main.h"
#include "stack_alloc.h"
#include "NSQ.h"
#include "celt/cpu_support.h"
#include "celt/arm/armcpu.h"
opus_int32 silk_noise_shape_quantizer_short_prediction_neon(const opus_int32 *buf32, const opus_int32 *coef32, opus_int order)
{
int32x4_t coef0 = vld1q_s32(coef32);
int32x4_t coef1 = vld1q_s32(coef32 + 4);
int32x4_t coef2 = vld1q_s32(coef32 + 8);
int32x4_t coef3 = vld1q_s32(coef32 + 12);
int32x4_t a0 = vld1q_s32(buf32 - 15);
int32x4_t a1 = vld1q_s32(buf32 - 11);
int32x4_t a2 = vld1q_s32(buf32 - 7);
int32x4_t a3 = vld1q_s32(buf32 - 3);
int32x4_t b0 = vqdmulhq_s32(coef0, a0);
int32x4_t b1 = vqdmulhq_s32(coef1, a1);
int32x4_t b2 = vqdmulhq_s32(coef2, a2);
int32x4_t b3 = vqdmulhq_s32(coef3, a3);
int32x4_t c0 = vaddq_s32(b0, b1);
int32x4_t c1 = vaddq_s32(b2, b3);
int32x4_t d = vaddq_s32(c0, c1);
int64x2_t e = vpaddlq_s32(d);
int64x1_t f = vadd_s64(vget_low_s64(e), vget_high_s64(e));
opus_int32 out = vget_lane_s32(vreinterpret_s32_s64(f), 0);
out += silk_RSHIFT( order, 1 );
return out;
}
opus_int32 silk_NSQ_noise_shape_feedback_loop_neon(const opus_int32 *data0, opus_int32 *data1, const opus_int16 *coef, opus_int order)
{
opus_int32 out;
if (order == 8)
{
int32x4_t a00 = vdupq_n_s32(data0[0]);
int32x4_t a01 = vld1q_s32(data1); /* data1[0] ... [3] */
int32x4_t a0 = vextq_s32 (a00, a01, 3); /* data0[0] data1[0] ...[2] */
int32x4_t a1 = vld1q_s32(data1 + 3); /* data1[3] ... [6] */
/*TODO: Convert these once in advance instead of once per sample, like
silk_noise_shape_quantizer_short_prediction_neon() does.*/
int16x8_t coef16 = vld1q_s16(coef);
int32x4_t coef0 = vmovl_s16(vget_low_s16(coef16));
int32x4_t coef1 = vmovl_s16(vget_high_s16(coef16));
/*This is not bit-exact with the C version, since we do not drop the
lower 16 bits of each multiply, but wait until the end to truncate
precision. This is an encoder-specific calculation (and unlike
silk_noise_shape_quantizer_short_prediction_neon(), is not meant to
simulate what the decoder will do). We still could use vqdmulhq_s32()
like silk_noise_shape_quantizer_short_prediction_neon() and save
half the multiplies, but the speed difference is not large, since we
then need two extra adds.*/
int64x2_t b0 = vmull_s32(vget_low_s32(a0), vget_low_s32(coef0));
int64x2_t b1 = vmlal_s32(b0, vget_high_s32(a0), vget_high_s32(coef0));
int64x2_t b2 = vmlal_s32(b1, vget_low_s32(a1), vget_low_s32(coef1));
int64x2_t b3 = vmlal_s32(b2, vget_high_s32(a1), vget_high_s32(coef1));
int64x1_t c = vadd_s64(vget_low_s64(b3), vget_high_s64(b3));
int64x1_t cS = vrshr_n_s64(c, 15);
int32x2_t d = vreinterpret_s32_s64(cS);
out = vget_lane_s32(d, 0);
vst1q_s32(data1, a0);
vst1q_s32(data1 + 4, a1);
return out;
}
return silk_NSQ_noise_shape_feedback_loop_c(data0, data1, coef, order);
}

View file

@ -0,0 +1,114 @@
/***********************************************************************
Copyright (C) 2014 Vidyo
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifndef SILK_NSQ_NEON_H
#define SILK_NSQ_NEON_H
#include "cpu_support.h"
#include "SigProc_FIX.h"
#undef silk_short_prediction_create_arch_coef
/* For vectorized calc, reverse a_Q12 coefs, convert to 32-bit, and shift for vqdmulhq_s32. */
static OPUS_INLINE void silk_short_prediction_create_arch_coef_neon(opus_int32 *out, const opus_int16 *in, opus_int order)
{
out[15] = silk_LSHIFT32(in[0], 15);
out[14] = silk_LSHIFT32(in[1], 15);
out[13] = silk_LSHIFT32(in[2], 15);
out[12] = silk_LSHIFT32(in[3], 15);
out[11] = silk_LSHIFT32(in[4], 15);
out[10] = silk_LSHIFT32(in[5], 15);
out[9] = silk_LSHIFT32(in[6], 15);
out[8] = silk_LSHIFT32(in[7], 15);
out[7] = silk_LSHIFT32(in[8], 15);
out[6] = silk_LSHIFT32(in[9], 15);
if (order == 16)
{
out[5] = silk_LSHIFT32(in[10], 15);
out[4] = silk_LSHIFT32(in[11], 15);
out[3] = silk_LSHIFT32(in[12], 15);
out[2] = silk_LSHIFT32(in[13], 15);
out[1] = silk_LSHIFT32(in[14], 15);
out[0] = silk_LSHIFT32(in[15], 15);
}
else
{
out[5] = 0;
out[4] = 0;
out[3] = 0;
out[2] = 0;
out[1] = 0;
out[0] = 0;
}
}
#if defined(OPUS_ARM_PRESUME_NEON_INTR)
#define silk_short_prediction_create_arch_coef(out, in, order) \
(silk_short_prediction_create_arch_coef_neon(out, in, order))
#elif defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_MAY_HAVE_NEON_INTR)
#define silk_short_prediction_create_arch_coef(out, in, order) \
do { if (arch == OPUS_ARCH_ARM_NEON) { silk_short_prediction_create_arch_coef_neon(out, in, order); } } while (0)
#endif
opus_int32 silk_noise_shape_quantizer_short_prediction_neon(const opus_int32 *buf32, const opus_int32 *coef32, opus_int order);
opus_int32 silk_NSQ_noise_shape_feedback_loop_neon(const opus_int32 *data0, opus_int32 *data1, const opus_int16 *coef, opus_int order);
#if defined(OPUS_ARM_PRESUME_NEON_INTR)
#undef silk_noise_shape_quantizer_short_prediction
#define silk_noise_shape_quantizer_short_prediction(in, coef, coefRev, order, arch) \
((void)arch,silk_noise_shape_quantizer_short_prediction_neon(in, coefRev, order))
#undef silk_NSQ_noise_shape_feedback_loop
#define silk_NSQ_noise_shape_feedback_loop(data0, data1, coef, order, arch) ((void)arch,silk_NSQ_noise_shape_feedback_loop_neon(data0, data1, coef, order))
#elif defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_MAY_HAVE_NEON_INTR)
/* silk_noise_shape_quantizer_short_prediction implementations take different parameters based on arch
(coef vs. coefRev) so can't use the usual IMPL table implementation */
#undef silk_noise_shape_quantizer_short_prediction
#define silk_noise_shape_quantizer_short_prediction(in, coef, coefRev, order, arch) \
(arch == OPUS_ARCH_ARM_NEON ? \
silk_noise_shape_quantizer_short_prediction_neon(in, coefRev, order) : \
silk_noise_shape_quantizer_short_prediction_c(in, coef, order))
extern opus_int32
(*const SILK_NSQ_NOISE_SHAPE_FEEDBACK_LOOP_IMPL[OPUS_ARCHMASK+1])(
const opus_int32 *data0, opus_int32 *data1, const opus_int16 *coef,
opus_int order);
#undef silk_NSQ_noise_shape_feedback_loop
#define silk_NSQ_noise_shape_feedback_loop(data0, data1, coef, order, arch) \
(SILK_NSQ_NOISE_SHAPE_FEEDBACK_LOOP_IMPL[(arch)&OPUS_ARCHMASK](data0, data1, \
coef, order))
#endif
#endif /* SILK_NSQ_NEON_H */

View file

@ -0,0 +1,47 @@
/***********************************************************************
Copyright (C) 2013 Xiph.Org Foundation and contributors
Copyright (c) 2013 Parrot
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifndef SILK_SIGPROC_FIX_ARMv4_H
#define SILK_SIGPROC_FIX_ARMv4_H
#undef silk_MLA
static OPUS_INLINE opus_int32 silk_MLA_armv4(opus_int32 a, opus_int32 b,
opus_int32 c)
{
opus_int32 res;
__asm__(
"#silk_MLA\n\t"
"mla %0, %1, %2, %3\n\t"
: "=&r"(res)
: "r"(b), "r"(c), "r"(a)
);
return res;
}
#define silk_MLA(a, b, c) (silk_MLA_armv4(a, b, c))
#endif

View file

@ -0,0 +1,61 @@
/***********************************************************************
Copyright (c) 2006-2011, Skype Limited. All rights reserved.
Copyright (c) 2013 Parrot
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifndef SILK_SIGPROC_FIX_ARMv5E_H
#define SILK_SIGPROC_FIX_ARMv5E_H
#undef silk_SMULTT
static OPUS_INLINE opus_int32 silk_SMULTT_armv5e(opus_int32 a, opus_int32 b)
{
opus_int32 res;
__asm__(
"#silk_SMULTT\n\t"
"smultt %0, %1, %2\n\t"
: "=r"(res)
: "%r"(a), "r"(b)
);
return res;
}
#define silk_SMULTT(a, b) (silk_SMULTT_armv5e(a, b))
#undef silk_SMLATT
static OPUS_INLINE opus_int32 silk_SMLATT_armv5e(opus_int32 a, opus_int32 b,
opus_int32 c)
{
opus_int32 res;
__asm__(
"#silk_SMLATT\n\t"
"smlatt %0, %1, %2, %3\n\t"
: "=r"(res)
: "%r"(b), "r"(c), "r"(a)
);
return res;
}
#define silk_SMLATT(a, b, c) (silk_SMLATT_armv5e(a, b, c))
#endif

View file

@ -0,0 +1,123 @@
/***********************************************************************
Copyright (C) 2014 Vidyo
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifdef HAVE_CONFIG_H
# include "config.h"
#endif
#include "main_FIX.h"
#include "NSQ.h"
#include "SigProc_FIX.h"
#if defined(OPUS_HAVE_RTCD)
# if (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && \
!defined(OPUS_ARM_PRESUME_NEON_INTR))
void (*const SILK_BIQUAD_ALT_STRIDE2_IMPL[OPUS_ARCHMASK + 1])(
const opus_int16 *in, /* I input signal */
const opus_int32 *B_Q28, /* I MA coefficients [3] */
const opus_int32 *A_Q28, /* I AR coefficients [2] */
opus_int32 *S, /* I/O State vector [4] */
opus_int16 *out, /* O output signal */
const opus_int32 len /* I signal length (must be even) */
) = {
silk_biquad_alt_stride2_c, /* ARMv4 */
silk_biquad_alt_stride2_c, /* EDSP */
silk_biquad_alt_stride2_c, /* Media */
silk_biquad_alt_stride2_neon, /* Neon */
};
opus_int32 (*const SILK_LPC_INVERSE_PRED_GAIN_IMPL[OPUS_ARCHMASK + 1])( /* O Returns inverse prediction gain in energy domain, Q30 */
const opus_int16 *A_Q12, /* I Prediction coefficients, Q12 [order] */
const opus_int order /* I Prediction order */
) = {
silk_LPC_inverse_pred_gain_c, /* ARMv4 */
silk_LPC_inverse_pred_gain_c, /* EDSP */
silk_LPC_inverse_pred_gain_c, /* Media */
silk_LPC_inverse_pred_gain_neon, /* Neon */
};
void (*const SILK_NSQ_DEL_DEC_IMPL[OPUS_ARCHMASK + 1])(
const silk_encoder_state *psEncC, /* I Encoder State */
silk_nsq_state *NSQ, /* I/O NSQ state */
SideInfoIndices *psIndices, /* I/O Quantization Indices */
const opus_int16 x16[], /* I Input */
opus_int8 pulses[], /* O Quantized pulse signal */
const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */
const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */
const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */
const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */
const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */
const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */
const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */
const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */
const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */
const opus_int LTP_scale_Q14 /* I LTP state scaling */
) = {
silk_NSQ_del_dec_c, /* ARMv4 */
silk_NSQ_del_dec_c, /* EDSP */
silk_NSQ_del_dec_c, /* Media */
silk_NSQ_del_dec_neon, /* Neon */
};
/*There is no table for silk_noise_shape_quantizer_short_prediction because the
NEON version takes different parameters than the C version.
Instead RTCD is done via if statements at the call sites.
See NSQ_neon.h for details.*/
opus_int32
(*const SILK_NSQ_NOISE_SHAPE_FEEDBACK_LOOP_IMPL[OPUS_ARCHMASK+1])(
const opus_int32 *data0, opus_int32 *data1, const opus_int16 *coef,
opus_int order) = {
silk_NSQ_noise_shape_feedback_loop_c, /* ARMv4 */
silk_NSQ_noise_shape_feedback_loop_c, /* EDSP */
silk_NSQ_noise_shape_feedback_loop_c, /* Media */
silk_NSQ_noise_shape_feedback_loop_neon, /* NEON */
};
# endif
# if defined(FIXED_POINT) && \
defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR)
void (*const SILK_WARPED_AUTOCORRELATION_FIX_IMPL[OPUS_ARCHMASK + 1])(
opus_int32 *corr, /* O Result [order + 1] */
opus_int *scale, /* O Scaling of the correlation vector */
const opus_int16 *input, /* I Input data to correlate */
const opus_int warping_Q16, /* I Warping coefficient */
const opus_int length, /* I Length of input */
const opus_int order /* I Correlation order (even) */
) = {
silk_warped_autocorrelation_FIX_c, /* ARMv4 */
silk_warped_autocorrelation_FIX_c, /* EDSP */
silk_warped_autocorrelation_FIX_c, /* Media */
silk_warped_autocorrelation_FIX_neon, /* Neon */
};
# endif
#endif /* OPUS_HAVE_RTCD */

View file

@ -0,0 +1,68 @@
/***********************************************************************
Copyright (c) 2017 Google Inc.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifndef SILK_BIQUAD_ALT_ARM_H
# define SILK_BIQUAD_ALT_ARM_H
# include "celt/arm/armcpu.h"
# if defined(OPUS_ARM_MAY_HAVE_NEON_INTR)
void silk_biquad_alt_stride2_neon(
const opus_int16 *in, /* I input signal */
const opus_int32 *B_Q28, /* I MA coefficients [3] */
const opus_int32 *A_Q28, /* I AR coefficients [2] */
opus_int32 *S, /* I/O State vector [4] */
opus_int16 *out, /* O output signal */
const opus_int32 len /* I signal length (must be even) */
);
# if !defined(OPUS_HAVE_RTCD) && defined(OPUS_ARM_PRESUME_NEON)
# define OVERRIDE_silk_biquad_alt_stride2 (1)
# define silk_biquad_alt_stride2(in, B_Q28, A_Q28, S, out, len, arch) ((void)(arch), PRESUME_NEON(silk_biquad_alt_stride2)(in, B_Q28, A_Q28, S, out, len))
# endif
# endif
# if !defined(OVERRIDE_silk_biquad_alt_stride2)
/*Is run-time CPU detection enabled on this platform?*/
# if defined(OPUS_HAVE_RTCD) && (defined(OPUS_ARM_MAY_HAVE_NEON_INTR) && !defined(OPUS_ARM_PRESUME_NEON_INTR))
extern void (*const SILK_BIQUAD_ALT_STRIDE2_IMPL[OPUS_ARCHMASK+1])(
const opus_int16 *in, /* I input signal */
const opus_int32 *B_Q28, /* I MA coefficients [3] */
const opus_int32 *A_Q28, /* I AR coefficients [2] */
opus_int32 *S, /* I/O State vector [4] */
opus_int16 *out, /* O output signal */
const opus_int32 len /* I signal length (must be even) */
);
# define OVERRIDE_silk_biquad_alt_stride2 (1)
# define silk_biquad_alt_stride2(in, B_Q28, A_Q28, S, out, len, arch) ((*SILK_BIQUAD_ALT_STRIDE2_IMPL[(arch)&OPUS_ARCHMASK])(in, B_Q28, A_Q28, S, out, len))
# elif defined(OPUS_ARM_PRESUME_NEON_INTR)
# define OVERRIDE_silk_biquad_alt_stride2 (1)
# define silk_biquad_alt_stride2(in, B_Q28, A_Q28, S, out, len, arch) ((void)(arch), silk_biquad_alt_stride2_neon(in, B_Q28, A_Q28, S, out, len))
# endif
# endif
#endif /* end SILK_BIQUAD_ALT_ARM_H */

View file

@ -0,0 +1,156 @@
/***********************************************************************
Copyright (c) 2017 Google Inc.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include <arm_neon.h>
#ifdef OPUS_CHECK_ASM
# include <string.h>
# include "stack_alloc.h"
#endif
#include "SigProc_FIX.h"
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 )
{
int32x4_t t_s32x4, out32_Q14_s32x4;
*out32_Q14_s32x2 = vadd_s32( vget_low_s32( *S_s32x4 ), t_s32x2 ); /* silk_SMLAWB( S{0,1}, B_Q28[ 0 ], in{0,1} ) */
*S_s32x4 = vcombine_s32( vget_high_s32( *S_s32x4 ), vdup_n_s32( 0 ) ); /* S{0,1} = S{2,3}; S{2,3} = 0; */
*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 ); */
out32_Q14_s32x4 = vcombine_s32( *out32_Q14_s32x2, *out32_Q14_s32x2 ); /* out32_Q14_{0,1,0,1} */
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 ) */
*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(); */
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 ) */
*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 ); */
t_s32x4 = vqdmulhq_s32( in_s32x4, B_Q28_s32x4 ); /* silk_SMULWB( B_Q28[ {1,1,2,2} ], in{0,1,0,1} ) */
*S_s32x4 = vaddq_s32( *S_s32x4, t_s32x4 ); /* S0 = silk_SMLAWB( S0, B_Q28[ {1,1,2,2} ], in{0,1,0,1} ); */
}
void silk_biquad_alt_stride2_neon(
const opus_int16 *in, /* I input signal */
const opus_int32 *B_Q28, /* I MA coefficients [3] */
const opus_int32 *A_Q28, /* I AR coefficients [2] */
opus_int32 *S, /* I/O State vector [4] */
opus_int16 *out, /* O output signal */
const opus_int32 len /* I signal length (must be even) */
)
{
/* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */
opus_int k = 0;
const int32x2_t offset_s32x2 = vdup_n_s32( (1<<14) - 1 );
const int32x4_t offset_s32x4 = vcombine_s32( offset_s32x2, offset_s32x2 );
int16x4_t in_s16x4 = vdup_n_s16( 0 );
int16x4_t out_s16x4;
int32x2_t A_Q28_s32x2, A_L_s32x2, A_U_s32x2, B_Q28_s32x2, t_s32x2;
int32x4_t A_L_s32x4, A_U_s32x4, B_Q28_s32x4, S_s32x4, out32_Q14_s32x4;
int32x2x2_t t0_s32x2x2, t1_s32x2x2, t2_s32x2x2, S_s32x2x2;
#ifdef OPUS_CHECK_ASM
opus_int32 S_c[ 4 ];
VARDECL( opus_int16, out_c );
SAVE_STACK;
ALLOC( out_c, 2 * len, opus_int16 );
silk_memcpy( &S_c, S, sizeof( S_c ) );
silk_biquad_alt_stride2_c( in, B_Q28, A_Q28, S_c, out_c, len );
#endif
/* Negate A_Q28 values and split in two parts */
A_Q28_s32x2 = vld1_s32( A_Q28 );
A_Q28_s32x2 = vneg_s32( A_Q28_s32x2 );
A_L_s32x2 = vshl_n_s32( A_Q28_s32x2, 18 ); /* ( -A_Q28[] & 0x00003FFF ) << 18 */
A_L_s32x2 = vreinterpret_s32_u32( vshr_n_u32( vreinterpret_u32_s32( A_L_s32x2 ), 3 ) ); /* ( -A_Q28[] & 0x00003FFF ) << 15 */
A_U_s32x2 = vshr_n_s32( A_Q28_s32x2, 14 ); /* silk_RSHIFT( -A_Q28[], 14 ) */
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.) */
A_U_s32x2 = vshr_n_s32( A_U_s32x2, 1 ); /* silk_RSHIFT( -A_Q28[], 14 ) << 15 */
B_Q28_s32x2 = vld1_s32( B_Q28 );
t_s32x2 = vld1_s32( B_Q28 + 1 );
t0_s32x2x2 = vzip_s32( A_L_s32x2, A_L_s32x2 );
t1_s32x2x2 = vzip_s32( A_U_s32x2, A_U_s32x2 );
t2_s32x2x2 = vzip_s32( t_s32x2, t_s32x2 );
A_L_s32x4 = vcombine_s32( t0_s32x2x2.val[ 0 ], t0_s32x2x2.val[ 1 ] ); /* A{0,0,1,1}_L_Q28 */
A_U_s32x4 = vcombine_s32( t1_s32x2x2.val[ 0 ], t1_s32x2x2.val[ 1 ] ); /* A{0,0,1,1}_U_Q28 */
B_Q28_s32x4 = vcombine_s32( t2_s32x2x2.val[ 0 ], t2_s32x2x2.val[ 1 ] ); /* B_Q28[ {1,1,2,2} ] */
S_s32x4 = vld1q_s32( S ); /* S0 = S[ 0 ]; S3 = S[ 3 ]; */
S_s32x2x2 = vtrn_s32( vget_low_s32( S_s32x4 ), vget_high_s32( S_s32x4 ) ); /* S2 = S[ 1 ]; S1 = S[ 2 ]; */
S_s32x4 = vcombine_s32( S_s32x2x2.val[ 0 ], S_s32x2x2.val[ 1 ] );
for( ; k < len - 1; k += 2 ) {
int32x4_t in_s32x4[ 2 ], t_s32x4;
int32x2_t out32_Q14_s32x2[ 2 ];
/* S[ 2 * i + 0 ], S[ 2 * i + 1 ], S[ 2 * i + 2 ], S[ 2 * i + 3 ]: Q12 */
in_s16x4 = vld1_s16( &in[ 2 * k ] ); /* in{0,1,2,3} = in[ 2 * k + {0,1,2,3} ]; */
in_s32x4[ 0 ] = vshll_n_s16( in_s16x4, 15 ); /* in{0,1,2,3} << 15 */
t_s32x4 = vqdmulhq_lane_s32( in_s32x4[ 0 ], B_Q28_s32x2, 0 ); /* silk_SMULWB( B_Q28[ 0 ], in{0,1,2,3} ) */
in_s32x4[ 1 ] = vcombine_s32( vget_high_s32( in_s32x4[ 0 ] ), vget_high_s32( in_s32x4[ 0 ] ) ); /* in{2,3,2,3} << 15 */
in_s32x4[ 0 ] = vcombine_s32( vget_low_s32 ( in_s32x4[ 0 ] ), vget_low_s32 ( in_s32x4[ 0 ] ) ); /* in{0,1,0,1} << 15 */
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 ] );
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 ] );
/* Scale back to Q0 and saturate */
out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2[ 0 ], out32_Q14_s32x2[ 1 ] ); /* out32_Q14_{0,1,2,3} */
out32_Q14_s32x4 = vaddq_s32( out32_Q14_s32x4, offset_s32x4 ); /* out32_Q14_{0,1,2,3} + (1<<14) - 1 */
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 ) ) */
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 ) ); */
}
/* Process leftover. */
if( k < len ) {
int32x4_t in_s32x4;
int32x2_t out32_Q14_s32x2;
/* S[ 2 * i + 0 ], S[ 2 * i + 1 ]: Q12 */
in_s16x4 = vld1_lane_s16( &in[ 2 * k + 0 ], in_s16x4, 0 ); /* in{0,1} = in[ 2 * k + {0,1} ]; */
in_s16x4 = vld1_lane_s16( &in[ 2 * k + 1 ], in_s16x4, 1 ); /* in{0,1} = in[ 2 * k + {0,1} ]; */
in_s32x4 = vshll_n_s16( in_s16x4, 15 ); /* in{0,1} << 15 */
t_s32x2 = vqdmulh_lane_s32( vget_low_s32( in_s32x4 ), B_Q28_s32x2, 0 ); /* silk_SMULWB( B_Q28[ 0 ], in{0,1} ) */
in_s32x4 = vcombine_s32( vget_low_s32( in_s32x4 ), vget_low_s32( in_s32x4 ) ); /* in{0,1,0,1} << 15 */
silk_biquad_alt_stride2_kernel( A_L_s32x4, A_U_s32x4, B_Q28_s32x4, t_s32x2, in_s32x4, &S_s32x4, &out32_Q14_s32x2 );
/* Scale back to Q0 and saturate */
out32_Q14_s32x2 = vadd_s32( out32_Q14_s32x2, offset_s32x2 ); /* out32_Q14_{0,1} + (1<<14) - 1 */
out32_Q14_s32x4 = vcombine_s32( out32_Q14_s32x2, out32_Q14_s32x2 ); /* out32_Q14_{0,1,0,1} + (1<<14) - 1 */
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 ) ) */
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 ) ); */
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 ) ); */
}
vst1q_lane_s32( &S[ 0 ], S_s32x4, 0 ); /* S[ 0 ] = S0; */
vst1q_lane_s32( &S[ 1 ], S_s32x4, 2 ); /* S[ 1 ] = S2; */
vst1q_lane_s32( &S[ 2 ], S_s32x4, 1 ); /* S[ 2 ] = S1; */
vst1q_lane_s32( &S[ 3 ], S_s32x4, 3 ); /* S[ 3 ] = S3; */
#ifdef OPUS_CHECK_ASM
silk_assert( !memcmp( S_c, S, sizeof( S_c ) ) );
silk_assert( !memcmp( out_c, out, 2 * len * sizeof( opus_int16 ) ) );
RESTORE_STACK;
#endif
}

View file

@ -0,0 +1,39 @@
/***********************************************************************
Copyright (C) 2015 Vidyo
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifndef SILK_MACROS_ARM64_H
#define SILK_MACROS_ARM64_H
#include <arm_neon.h>
#undef silk_ADD_SAT32
#define silk_ADD_SAT32(a, b) (vqadds_s32((a), (b)))
#undef silk_SUB_SAT32
#define silk_SUB_SAT32(a, b) (vqsubs_s32((a), (b)))
#endif /* SILK_MACROS_ARM64_H */

View file

@ -0,0 +1,110 @@
/***********************************************************************
Copyright (C) 2013 Xiph.Org Foundation and contributors.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifndef SILK_MACROS_ARMv4_H
#define SILK_MACROS_ARMv4_H
/* This macro only avoids the undefined behaviour from a left shift of
a negative value. It should only be used in macros that can't include
SigProc_FIX.h. In other cases, use silk_LSHIFT32(). */
#define SAFE_SHL(a,b) ((opus_int32)((opus_uint32)(a) << (b)))
/* (a32 * (opus_int32)((opus_int16)(b32))) >> 16 output have to be 32bit int */
#undef silk_SMULWB
static OPUS_INLINE opus_int32 silk_SMULWB_armv4(opus_int32 a, opus_int16 b)
{
unsigned rd_lo;
int rd_hi;
__asm__(
"#silk_SMULWB\n\t"
"smull %0, %1, %2, %3\n\t"
: "=&r"(rd_lo), "=&r"(rd_hi)
: "%r"(a), "r"(SAFE_SHL(b,16))
);
return rd_hi;
}
#define silk_SMULWB(a, b) (silk_SMULWB_armv4(a, b))
/* a32 + (b32 * (opus_int32)((opus_int16)(c32))) >> 16 output have to be 32bit int */
#undef silk_SMLAWB
#define silk_SMLAWB(a, b, c) ((a) + silk_SMULWB(b, c))
/* (a32 * (b32 >> 16)) >> 16 */
#undef silk_SMULWT
static OPUS_INLINE opus_int32 silk_SMULWT_armv4(opus_int32 a, opus_int32 b)
{
unsigned rd_lo;
int rd_hi;
__asm__(
"#silk_SMULWT\n\t"
"smull %0, %1, %2, %3\n\t"
: "=&r"(rd_lo), "=&r"(rd_hi)
: "%r"(a), "r"(b&~0xFFFF)
);
return rd_hi;
}
#define silk_SMULWT(a, b) (silk_SMULWT_armv4(a, b))
/* a32 + (b32 * (c32 >> 16)) >> 16 */
#undef silk_SMLAWT
#define silk_SMLAWT(a, b, c) ((a) + silk_SMULWT(b, c))
/* (a32 * b32) >> 16 */
#undef silk_SMULWW
static OPUS_INLINE opus_int32 silk_SMULWW_armv4(opus_int32 a, opus_int32 b)
{
unsigned rd_lo;
int rd_hi;
__asm__(
"#silk_SMULWW\n\t"
"smull %0, %1, %2, %3\n\t"
: "=&r"(rd_lo), "=&r"(rd_hi)
: "%r"(a), "r"(b)
);
return SAFE_SHL(rd_hi,16)+(rd_lo>>16);
}
#define silk_SMULWW(a, b) (silk_SMULWW_armv4(a, b))
#undef silk_SMLAWW
static OPUS_INLINE opus_int32 silk_SMLAWW_armv4(opus_int32 a, opus_int32 b,
opus_int32 c)
{
unsigned rd_lo;
int rd_hi;
__asm__(
"#silk_SMLAWW\n\t"
"smull %0, %1, %2, %3\n\t"
: "=&r"(rd_lo), "=&r"(rd_hi)
: "%r"(b), "r"(c)
);
return a+SAFE_SHL(rd_hi,16)+(rd_lo>>16);
}
#define silk_SMLAWW(a, b, c) (silk_SMLAWW_armv4(a, b, c))
#undef SAFE_SHL
#endif /* SILK_MACROS_ARMv4_H */

View file

@ -0,0 +1,220 @@
/***********************************************************************
Copyright (c) 2006-2011, Skype Limited. All rights reserved.
Copyright (c) 2013 Parrot
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of Internet Society, IETF or IETF Trust, nor the
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifndef SILK_MACROS_ARMv5E_H
#define SILK_MACROS_ARMv5E_H
/* This macro only avoids the undefined behaviour from a left shift of
a negative value. It should only be used in macros that can't include
SigProc_FIX.h. In other cases, use silk_LSHIFT32(). */
#define SAFE_SHL(a,b) ((opus_int32)((opus_uint32)(a) << (b)))
/* (a32 * (opus_int32)((opus_int16)(b32))) >> 16 output have to be 32bit int */
#undef silk_SMULWB
static OPUS_INLINE opus_int32 silk_SMULWB_armv5e(opus_int32 a, opus_int16 b)
{
int res;
__asm__(
"#silk_SMULWB\n\t"
"smulwb %0, %1, %2\n\t"
: "=r"(res)
: "r"(a), "r"(b)
);
return res;
}
#define silk_SMULWB(a, b) (silk_SMULWB_armv5e(a, b))
/* a32 + (b32 * (opus_int32)((opus_int16)(c32))) >> 16 output have to be 32bit int */
#undef silk_SMLAWB
static OPUS_INLINE opus_int32 silk_SMLAWB_armv5e(opus_int32 a, opus_int32 b,
opus_int16 c)
{
int res;
__asm__(
"#silk_SMLAWB\n\t"
"smlawb %0, %1, %2, %3\n\t"
: "=r"(res)
: "r"(b), "r"(c), "r"(a)
);
return res;
}
#define silk_SMLAWB(a, b, c) (silk_SMLAWB_armv5e(a, b, c))
/* (a32 * (b32 >> 16)) >> 16 */
#undef silk_SMULWT
static OPUS_INLINE opus_int32 silk_SMULWT_armv5e(opus_int32 a, opus_int32 b)
{
int res;
__asm__(
"#silk_SMULWT\n\t"
"smulwt %0, %1, %2\n\t"
: "=r"(res)
: "r"(a), "r"(b)
);
return res;
}
#define silk_SMULWT(a, b) (silk_SMULWT_armv5e(a, b))
/* a32 + (b32 * (c32 >> 16)) >> 16 */
#undef silk_SMLAWT
static OPUS_INLINE opus_int32 silk_SMLAWT_armv5e(opus_int32 a, opus_int32 b,
opus_int32 c)
{
int res;
__asm__(
"#silk_SMLAWT\n\t"
"smlawt %0, %1, %2, %3\n\t"
: "=r"(res)
: "r"(b), "r"(c), "r"(a)
);
return res;
}
#define silk_SMLAWT(a, b, c) (silk_SMLAWT_armv5e(a, b, c))
/* (opus_int32)((opus_int16)(a3))) * (opus_int32)((opus_int16)(b32)) output have to be 32bit int */
#undef silk_SMULBB
static OPUS_INLINE opus_int32 silk_SMULBB_armv5e(opus_int32 a, opus_int32 b)
{
int res;
__asm__(
"#silk_SMULBB\n\t"
"smulbb %0, %1, %2\n\t"
: "=r"(res)
: "%r"(a), "r"(b)
);
return res;
}
#define silk_SMULBB(a, b) (silk_SMULBB_armv5e(a, b))
/* a32 + (opus_int32)((opus_int16)(b32)) * (opus_int32)((opus_int16)(c32)) output have to be 32bit int */
#undef silk_SMLABB
static OPUS_INLINE opus_int32 silk_SMLABB_armv5e(opus_int32 a, opus_int32 b,
opus_int32 c)
{
int res;
__asm__(
"#silk_SMLABB\n\t"
"smlabb %0, %1, %2, %3\n\t"
: "=r"(res)
: "%r"(b), "r"(c), "r"(a)
);
return res;
}
#define silk_SMLABB(a, b, c) (silk_SMLABB_armv5e(a, b, c))
/* (opus_int32)((opus_int16)(a32)) * (b32 >> 16) */
#undef silk_SMULBT
static OPUS_INLINE opus_int32 silk_SMULBT_armv5e(opus_int32 a, opus_int32 b)
{
int res;
__asm__(
"#silk_SMULBT\n\t"
"smulbt %0, %1, %2\n\t"
: "=r"(res)
: "r"(a), "r"(b)
);
return res;
}
#define silk_SMULBT(a, b) (silk_SMULBT_armv5e(a, b))
/* a32 + (opus_int32)((opus_int16)(b32)) * (c32 >> 16) */
#undef silk_SMLABT
static OPUS_INLINE opus_int32 silk_SMLABT_armv5e(opus_int32 a, opus_int32 b,
opus_int32 c)
{
int res;
__asm__(
"#silk_SMLABT\n\t"
"smlabt %0, %1, %2, %3\n\t"
: "=r"(res)
: "r"(b), "r"(c), "r"(a)
);
return res;
}
#define silk_SMLABT(a, b, c) (silk_SMLABT_armv5e(a, b, c))
/* add/subtract with output saturated */
#undef silk_ADD_SAT32
static OPUS_INLINE opus_int32 silk_ADD_SAT32_armv5e(opus_int32 a, opus_int32 b)
{
int res;
__asm__(
"#silk_ADD_SAT32\n\t"
"qadd %0, %1, %2\n\t"
: "=r"(res)
: "%r"(a), "r"(b)
);
return res;
}
#define silk_ADD_SAT32(a, b) (silk_ADD_SAT32_armv5e(a, b))
#undef silk_SUB_SAT32
static OPUS_INLINE opus_int32 silk_SUB_SAT32_armv5e(opus_int32 a, opus_int32 b)
{
int res;
__asm__(
"#silk_SUB_SAT32\n\t"
"qsub %0, %1, %2\n\t"
: "=r"(res)
: "r"(a), "r"(b)
);
return res;
}
#define silk_SUB_SAT32(a, b) (silk_SUB_SAT32_armv5e(a, b))
#undef silk_CLZ16
static OPUS_INLINE opus_int32 silk_CLZ16_armv5(opus_int16 in16)
{
int res;
__asm__(
"#silk_CLZ16\n\t"
"clz %0, %1;\n"
: "=r"(res)
: "r"(SAFE_SHL(in16,16)|0x8000)
);
return res;
}
#define silk_CLZ16(in16) (silk_CLZ16_armv5(in16))
#undef silk_CLZ32
static OPUS_INLINE opus_int32 silk_CLZ32_armv5(opus_int32 in32)
{
int res;
__asm__(
"#silk_CLZ32\n\t"
"clz %0, %1\n\t"
: "=r"(res)
: "r"(in32)
);
return res;
}
#define silk_CLZ32(in32) (silk_CLZ32_armv5(in32))
#undef SAFE_SHL
#endif /* SILK_MACROS_ARMv5E_H */