/* return the sum of all elements in an array. This works by calculating 4 totals (one for each lane) and adding those at the end to get the final total */
int sum_array(int16_t *array, int size)
{
/* initialize the accumulator vector to zero */
int16x4_t acc = vdup_n_s16(0);
int32x2_t acc1;
int64x1_t acc2;
/* this implementation assumes the size of the array is a multiple of 4 */
assert((size % 4) == 0);
/* counting backwards gives better code */
for (; size != 0; size -= 4)
{
int16x4_t vec;
/* load 4 values in parallel from the array */
vec = vld1_s16(array);
/* increment the array pointer to the next element */
array += 4;
/* add the vector to the accumulator vector */
acc = vadd_s16(acc, vec);
}
/* calculate the total */
acc1 = vpaddl_s16(acc);
acc2 = vpaddl_s32(acc1);
/* return the total as an integer */
return (int)vget_lane_s64(acc2, 0);
}
int64_t test_vget_lane_s64(int64x1_t v1) {
// CHECK: test_vget_lane_s64
return vget_lane_s64(v1, 0);
// CHECK: fmov {{x[0-9]+}}, {{d[0-9]+}}
}
int64_t test_vget_lane_s64(int64x1_t a) {
// CHECK-LABEL: test_vget_lane_s64:
// CHECK-NEXT: fmov x0, d0
// CHECK-NEXT: ret
return vget_lane_s64(a, 0);
}
void silk_warped_autocorrelation_FIX_neon(
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) */
)
{
if( ( MAX_SHAPE_LPC_ORDER > 24 ) || ( order < 6 ) ) {
silk_warped_autocorrelation_FIX_c( corr, scale, input, warping_Q16, length, order );
} else {
opus_int n, i, lsh;
opus_int64 corr_QC[ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 }; /* In reverse order */
opus_int64 corr_QC_orderT;
int64x2_t lsh_s64x2;
const opus_int orderT = ( order + 3 ) & ~3;
opus_int64 *corr_QCT;
opus_int32 *input_QS;
VARDECL( opus_int32, input_QST );
VARDECL( opus_int32, state );
SAVE_STACK;
/* Order must be even */
silk_assert( ( order & 1 ) == 0 );
silk_assert( 2 * QS - QC >= 0 );
ALLOC( input_QST, length + 2 * MAX_SHAPE_LPC_ORDER, opus_int32 );
input_QS = input_QST;
/* input_QS has zero paddings in the beginning and end. */
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
/* Loop over samples */
for( n = 0; n < length - 7; n += 8, input_QS += 8 ) {
const int16x8_t t0_s16x4 = vld1q_s16( input + n );
vst1q_s32( input_QS + 0, vshll_n_s16( vget_low_s16( t0_s16x4 ), QS ) );
vst1q_s32( input_QS + 4, vshll_n_s16( vget_high_s16( t0_s16x4 ), QS ) );
}
for( ; n < length; n++, input_QS++ ) {
input_QS[ 0 ] = silk_LSHIFT32( (opus_int32)input[ n ], QS );
}
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS += 4;
vst1q_s32( input_QS, vdupq_n_s32( 0 ) );
input_QS = input_QST + MAX_SHAPE_LPC_ORDER - orderT;
/* The following loop runs ( length + order ) times, with ( order ) extra epilogues. */
/* The zero paddings in input_QS guarantee corr_QC's correctness even with the extra epilogues. */
/* The values of state_QS will be polluted by the extra epilogues, however they are temporary values. */
/* Keep the C code here to help understand the intrinsics optimization. */
/*
{
opus_int32 state_QS[ 2 ][ MAX_SHAPE_LPC_ORDER + 1 ] = { 0 };
opus_int32 *state_QST[ 3 ];
state_QST[ 0 ] = state_QS[ 0 ];
state_QST[ 1 ] = state_QS[ 1 ];
for( n = 0; n < length + order; n++, input_QS++ ) {
state_QST[ 0 ][ orderT ] = input_QS[ orderT ];
for( i = 0; i < orderT; i++ ) {
corr_QC[ i ] += silk_RSHIFT64( silk_SMULL( state_QST[ 0 ][ i ], input_QS[ i ] ), 2 * QS - QC );
state_QST[ 1 ][ i ] = silk_SMLAWB( state_QST[ 1 ][ i + 1 ], state_QST[ 0 ][ i ] - state_QST[ 0 ][ i + 1 ], warping_Q16 );
}
state_QST[ 2 ] = state_QST[ 0 ];
state_QST[ 0 ] = state_QST[ 1 ];
state_QST[ 1 ] = state_QST[ 2 ];
}
}
*/
{
const int32x4_t warping_Q16_s32x4 = vdupq_n_s32( warping_Q16 << 15 );
const opus_int32 *in = input_QS + orderT;
opus_int o = orderT;
int32x4_t state_QS_s32x4[ 3 ][ 2 ];
ALLOC( state, length + orderT, opus_int32 );
state_QS_s32x4[ 2 ][ 1 ] = vdupq_n_s32( 0 );
/* Calculate 8 taps of all inputs in each loop. */
do {
state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 0 ][ 1 ] =
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 1 ][ 1 ] = vdupq_n_s32( 0 );
n = 0;
do {
calc_corr( input_QS + n, corr_QC, o - 8, state_QS_s32x4[ 0 ][ 0 ] );
calc_corr( input_QS + n, corr_QC, o - 4, state_QS_s32x4[ 0 ][ 1 ] );
state_QS_s32x4[ 2 ][ 1 ] = vld1q_s32( in + n );
vst1q_lane_s32( state + n, state_QS_s32x4[ 0 ][ 0 ], 0 );
state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 0 ][ 1 ], 1 );
state_QS_s32x4[ 2 ][ 1 ] = vextq_s32( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], 1 );
state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 );
state_QS_s32x4[ 0 ][ 1 ] = calc_state( state_QS_s32x4[ 0 ][ 1 ], state_QS_s32x4[ 2 ][ 1 ], state_QS_s32x4[ 1 ][ 1 ], warping_Q16_s32x4 );
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ];
state_QS_s32x4[ 1 ][ 1 ] = state_QS_s32x4[ 2 ][ 1 ];
} while( ++n < ( length + order ) );
in = state;
o -= 8;
} while( o > 4 );
if( o ) {
/* Calculate the last 4 taps of all inputs. */
opus_int32 *stateT = state;
silk_assert( o == 4 );
state_QS_s32x4[ 0 ][ 0 ] = state_QS_s32x4[ 1 ][ 0 ] = vdupq_n_s32( 0 );
n = length + order;
do {
calc_corr( input_QS, corr_QC, 0, state_QS_s32x4[ 0 ][ 0 ] );
state_QS_s32x4[ 2 ][ 0 ] = vld1q_s32( stateT );
vst1q_lane_s32( stateT, state_QS_s32x4[ 0 ][ 0 ], 0 );
state_QS_s32x4[ 2 ][ 0 ] = vextq_s32( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], 1 );
state_QS_s32x4[ 0 ][ 0 ] = calc_state( state_QS_s32x4[ 0 ][ 0 ], state_QS_s32x4[ 2 ][ 0 ], state_QS_s32x4[ 1 ][ 0 ], warping_Q16_s32x4 );
state_QS_s32x4[ 1 ][ 0 ] = state_QS_s32x4[ 2 ][ 0 ];
input_QS++;
stateT++;
} while( --n );
}
}
{
const opus_int16 *inputT = input;
int32x4_t t_s32x4;
int64x1_t t_s64x1;
int64x2_t t_s64x2 = vdupq_n_s64( 0 );
for( n = 0; n <= length - 8; n += 8 ) {
int16x8_t input_s16x8 = vld1q_s16( inputT );
t_s32x4 = vmull_s16( vget_low_s16( input_s16x8 ), vget_low_s16( input_s16x8 ) );
t_s32x4 = vmlal_s16( t_s32x4, vget_high_s16( input_s16x8 ), vget_high_s16( input_s16x8 ) );
t_s64x2 = vaddw_s32( t_s64x2, vget_low_s32( t_s32x4 ) );
t_s64x2 = vaddw_s32( t_s64x2, vget_high_s32( t_s32x4 ) );
inputT += 8;
}
t_s64x1 = vadd_s64( vget_low_s64( t_s64x2 ), vget_high_s64( t_s64x2 ) );
corr_QC_orderT = vget_lane_s64( t_s64x1, 0 );
for( ; n < length; n++ ) {
corr_QC_orderT += silk_SMULL( input[ n ], input[ n ] );
}
corr_QC_orderT = silk_LSHIFT64( corr_QC_orderT, QC );
corr_QC[ orderT ] = corr_QC_orderT;
}
corr_QCT = corr_QC + orderT - order;
lsh = silk_CLZ64( corr_QC_orderT ) - 35;
lsh = silk_LIMIT( lsh, -12 - QC, 30 - QC );
*scale = -( QC + lsh );
silk_assert( *scale >= -30 && *scale <= 12 );
lsh_s64x2 = vdupq_n_s64( lsh );
for( i = 0; i <= order - 3; i += 4 ) {
int32x4_t corr_s32x4;
int64x2_t corr_QC0_s64x2, corr_QC1_s64x2;
corr_QC0_s64x2 = vld1q_s64( corr_QCT + i );
corr_QC1_s64x2 = vld1q_s64( corr_QCT + i + 2 );
corr_QC0_s64x2 = vshlq_s64( corr_QC0_s64x2, lsh_s64x2 );
corr_QC1_s64x2 = vshlq_s64( corr_QC1_s64x2, lsh_s64x2 );
corr_s32x4 = vcombine_s32( vmovn_s64( corr_QC1_s64x2 ), vmovn_s64( corr_QC0_s64x2 ) );
corr_s32x4 = vrev64q_s32( corr_s32x4 );
vst1q_s32( corr + order - i - 3, corr_s32x4 );
}
if( lsh >= 0 ) {
for( ; i < order + 1; i++ ) {
corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_LSHIFT64( corr_QCT[ i ], lsh ) );
}
} else {
for( ; i < order + 1; i++ ) {
corr[ order - i ] = (opus_int32)silk_CHECK_FIT32( silk_RSHIFT64( corr_QCT[ i ], -lsh ) );
}
}
silk_assert( corr_QCT[ order ] >= 0 ); /* If breaking, decrease QC*/
RESTORE_STACK;
}
#ifdef OPUS_CHECK_ASM
{
opus_int32 corr_c[ MAX_SHAPE_LPC_ORDER + 1 ];
opus_int scale_c;
silk_warped_autocorrelation_FIX_c( corr_c, &scale_c, input, warping_Q16, length, order );
silk_assert( !memcmp( corr_c, corr, sizeof( corr_c[ 0 ] ) * ( order + 1 ) ) );
silk_assert( scale_c == *scale );
}
#endif
}
void mdrc5b_apply_limiter(MDRC5B_LOCAL_STRUCT_T *HeapPtr)
{
unsigned int LaIdx;
unsigned int NumMainCh;
unsigned int Samples;
unsigned int ch, k, n;
MMlong *Ptr;
MMlong *Ptr2;
MMlong *MemOutPtr;
MMshort PeakdB;
MMlong PeakMax;
int RmsMeasure;
MMshort LimiterAtCoef;
MMshort LimiterReCoef;
MMshort LimiterGainMant[MDRC5B_BLOCK_SIZE + 1];
MMshort LimiterGainExp;
MMshort LimiterTargetGaindB;
unsigned int LimiterHoldRem;
unsigned int LimiterHtSamp;
MMshort Exp, TargetGain;
MMshort MaxShiftBits;
unsigned int lookahead_len = (unsigned int) HeapPtr->LimiterLALen;
unsigned int cpt1, cpt2;
uint32x2x2_t Temp_u32x2x2;
uint32x2_t Ldbits_u32x2, Ldbits2_u32x2;
uint32x2_t bsl_u32x2;
int32x2_t LimGainMant_32x2;
int64x2_t TempX_64x2, MemOut_64x2;
int64x2_t Tmp_64x2;
int64x2_t LimiterGainExp_64x2, sample_64x2;
int64x1_t TempX_64x1, sample_64x1;
int32_t *LimiterGainMant_ptr;
int32x2_t Tmp_32x2, Ldbits_32x2, n_32x2;
int32x2_t TempX_low_32x2, TempX_high_32x2;
int32x2x2_t Tmp_32x2x2;
int64x1_t Peak_64x1, PeakMax_64x1, Tmp_64x1, diffX_64x1;
int64x1_t Peak_scale_pow_64x1, Peak_scale_64x1, Zero_s64x1;
int64x1_t MaxShiftBits_neg_64x1, MaxShiftBits_hd_64x1;
int64x2_t diffX_64x2;
uint64x1_t bsl_u64x1;
int32x2_t LimiterPeakCoef_32x2, diffX_low_32x2, diffX_high_32x2;
int32x2_t TargetGain_32x2;
uint32x2x2_t Peak_u32x2x2;
uint32x2_t Peak_exp_u32x2, Peak_exp2_u32x2, Peak_mant_u32x2;
int32x2_t x_32x2, xn_32x2, PeakdB_32x2, Peak_exp_32x2;
int32x2_t LimiterTargetGaindB_32x2, Exp_32x2, LimiterCoef_32x2;
int32x4_t Tmp_32x4;
START_PMU_MEASURE(PMU_MEASURE_MRDC5B_APPLY_LIMITER)
START_PMU_MEASURE(PMU_MEASURE_MRDC5B_LIMITER_COMPUTE_MAX_SHIFT_LEFT)
Samples = (unsigned int) HeapPtr->BlockSize;
NumMainCh = (unsigned int) HeapPtr->NumMainCh;
TempX_64x2 = vdupq_n_s64(0);
for(ch = 0; ch < NumMainCh; ch++)
{
Ptr = HeapPtr->MainInBuf[ch];
// compute the number of bits needs to be shifted to avoid overflow
for(k = (Samples >> 1); k > 0; k--)
{
sample_64x2 = vld1q_s64(Ptr);
Ptr +=2;
sample_64x2 = veorq_s64(sample_64x2, vshrq_n_s64(sample_64x2, 63));
TempX_64x2 = vorrq_s64(TempX_64x2, sample_64x2);
}
if(Samples & 1)
{
sample_64x1 = vld1_s64(Ptr);
sample_64x1 = veor_s64(sample_64x1, vshr_n_s64(sample_64x1, 63));
TempX_64x2 = vorrq_s64(TempX_64x2, vcombine_s64(sample_64x1, sample_64x1));
}
}
TempX_64x1 = vorr_s64(vget_low_s64(TempX_64x2), vget_high_s64(TempX_64x2));
Temp_u32x2x2 = vuzp_u32(vreinterpret_u32_s64(TempX_64x1), vreinterpret_u32_s64(TempX_64x1));
bsl_u32x2 = vceq_u32(Temp_u32x2x2.val[1], vdup_n_u32(0)); // MSB == 0 ?
// use clz instead of cls because we are sure that input value is positive
// and because cls(LSB) could be wrong (if MSB is equal to 0 and bit 31 of LSL is 1)
// thus clz result will be 1 more than cls result (that's why you may see (Ldbits - 1)
// instead of Ldbits below)
Ldbits_u32x2 = vadd_u32(vclz_u32(Temp_u32x2x2.val[0]), vdup_n_u32(32)); // clz(LSB)+32
Ldbits2_u32x2 = vclz_u32(Temp_u32x2x2.val[1]); // clz(MSB)
Ldbits_u32x2 = vbsl_u32(bsl_u32x2, Ldbits_u32x2, Ldbits2_u32x2); // MSB == 0 ? clz(LSB)+32 : clz(MSB)
bsl_u32x2 = vceq_u32(Ldbits_u32x2, vdup_n_u32(64)); // Ldbits == 64 ? (i.e. TempX == 0 ?)
// the aim of MaxShiftBits is that sample will be shifted so that it occupies
// 24 significant bits for 24 bits samples or 32 significant bits for 32 bits samples
// but we are in 64 bits architecture on CA9/NEON
// so we must right shift of ((64 - 24) - (Ldbits - 1)) bits for 24 bits samples
// or of ((64 - 32) - (Ldbits - 1)) bits for 32 bits samples
// and we add 1 because it was done this way on MMDSP (I don't know why !)
#ifdef SAMPLES_24_BITS
// MaxShiftBits = ((64 - 24) - (Ldbits - 1)) + 1
// = 42 - Ldbits
Ldbits_32x2 = vsub_s32(vdup_n_s32(42), vreinterpret_s32_u32(Ldbits_u32x2));
#else // SAMPLES_24_BITS
// MaxShiftBits = ((64 - 32) - (Ldbits - 1)) + 1
// = 34 - Ldbits
Ldbits_32x2 = vsub_s32(vdup_n_s32(34), vreinterpret_s32_u32(Ldbits_u32x2));
#endif // SAMPLES_24_BITS
Ldbits_32x2 = vmax_s32(vdup_n_s32(1), Ldbits_32x2);
Ldbits_32x2 = vbsl_s32(bsl_u32x2, vdup_n_s32(1), Ldbits_32x2); // if(TempX == 0) Ldbits = 1
MaxShiftBits = vget_lane_s32(Ldbits_32x2, 0);
STOP_PMU_MEASURE(PMU_MEASURE_MRDC5B_LIMITER_COMPUTE_MAX_SHIFT_LEFT)
#ifdef DEBUG_LIMITER_OUTPUT
if((debug_cpt_samples >= DEBUG_CPT_MIN) && (debug_cpt_samples <= DEBUG_CPT_MAX))
{
char string[100];
debug_write_string("MRDC5B_LIMITER_COMPUTE_MAX_SHIFT_LEFT\n");
sprintf(string, "MaxShiftBits=%d\n", MaxShiftBits);
debug_write_string(string);
}
#endif // DEBUG_LIMITER_OUTPUT
START_PMU_MEASURE(PMU_MEASURE_MRDC5B_LIMITER_INSERT_NEW_SUBBAND)
// insert the new subband samples into the lookahead buffers
RmsMeasure = HeapPtr->Limiter.RmsMeasure;
LaIdx = (unsigned int) HeapPtr->LimiterLaIdx;
if(LaIdx + Samples >= lookahead_len)
{
cpt1 = lookahead_len - LaIdx;
cpt2 = Samples - cpt1;
// update index
HeapPtr->LimiterLaIdx = (int) cpt2;
}
else
{
cpt1 = Samples;
cpt2 = 0;
// update index
HeapPtr->LimiterLaIdx = (int) (LaIdx + Samples);
}
LimiterPeakCoef_32x2 = vdup_n_s32(HeapPtr->LimiterPeakAtCoef); // LimiterPeakAtCoef, LimiterPeakAtCoef
LimiterPeakCoef_32x2 = vset_lane_s32(HeapPtr->LimiterPeakReCoef, LimiterPeakCoef_32x2, 1); // LimiterPeakAtCoef, LimiterPeakReCoef
Peak_scale_64x1 = vdup_n_s64(HeapPtr->PrevShiftBits - MaxShiftBits);
Peak_scale_pow_64x1 = vshl_n_s64(Peak_scale_64x1, 1);
MaxShiftBits_neg_64x1 = vdup_n_s64(-MaxShiftBits);
#ifdef SAMPLES_24_BITS
MaxShiftBits_hd_64x1 = vdup_n_s64(24 - MaxShiftBits);
#else // SAMPLES_24_BITS
MaxShiftBits_hd_64x1 = vdup_n_s64(32 - MaxShiftBits);
#endif // SAMPLES_24_BITS
PeakMax_64x1 = vdup_n_s64(0);
for(ch = 0; ch < NumMainCh; ch++)
{
Ptr = HeapPtr->MainInBuf[ch];
Ptr2 = HeapPtr->LimiterLABuf[ch] + LaIdx; // go to the first valid sample
Peak_64x1 = vdup_n_s64(HeapPtr->LimiterPeak[ch]);
if(RmsMeasure)
{
// compensate Peak according to the previous shift bits
Peak_64x1 = vqrshl_s64(Peak_64x1, Peak_scale_pow_64x1); // neg value => shift right rounding
// rms measure
for(k = cpt1; k > 0; k--)
{
Tmp_64x1 = vld1_s64(Ptr);
Ptr++;
vst1_s64(Ptr2, Tmp_64x1);
Ptr2++;
Tmp_64x1 = vqrshl_s64(Tmp_64x1, MaxShiftBits_neg_64x1);
Tmp_64x2 = vcombine_s64(Tmp_64x1, Tmp_64x1);
Tmp_32x2x2 = vuzp_s32(vget_low_s32(vreinterpretq_s32_s64(Tmp_64x2)), vget_high_s32(vreinterpretq_s32_s64(Tmp_64x2)));
Tmp_32x2 = Tmp_32x2x2.val[0]; // LSB of Tmp_64x2 (MSB is dummy)
TempX_64x2 = vqdmull_s32(Tmp_32x2, Tmp_32x2);
TempX_64x1 = vget_low_s64(TempX_64x2);
diffX_64x1 = vqsub_s64(Peak_64x1, TempX_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(diffX_64x1, 63)); // sign(diffX)
diffX_64x2 = vcombine_s64(diffX_64x1, diffX_64x1);
diffX_low_32x2 = vshrn_n_s64(vshlq_n_s64(diffX_64x2, 32), 32); // wextract_l(diffX), wextract_l(diffX)
diffX_high_32x2 = vrshrn_n_s64(diffX_64x2, 32); // wround_L(diffX), wround_L(diffX)
Tmp_64x2 = vmovl_s32(vqrdmulh_s32(LimiterPeakCoef_32x2, diffX_low_32x2)); // (MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), (MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef)
Tmp_64x2 = vqdmlal_s32(Tmp_64x2, LimiterPeakCoef_32x2, diffX_high_32x2); // wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), wL_fmul(wround_L(diffX), LimiterPeakAtCoef)), wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef), wL_fmul(wround_L(diffX), LimiterPeakReCoef))
Tmp_64x2 = vqaddq_s64(TempX_64x2, Tmp_64x2);
Peak_64x1 = vbsl_s64(bsl_u64x1, vget_low_s64(Tmp_64x2), vget_high_s64(Tmp_64x2));
Tmp_64x1 = vqsub_s64(Peak_64x1, PeakMax_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Peak_64x1 - PeakMax_64x1)
PeakMax_64x1 = vbsl_s64(bsl_u64x1, PeakMax_64x1, Peak_64x1);
}
Ptr2 = HeapPtr->LimiterLABuf[ch];
for(k = cpt2; k > 0; k--)
{
Tmp_64x1 = vld1_s64(Ptr);
Ptr++;
vst1_s64(Ptr2, Tmp_64x1);
Ptr2++;
Tmp_64x1 = vqrshl_s64(Tmp_64x1, MaxShiftBits_neg_64x1);
Tmp_64x2 = vcombine_s64(Tmp_64x1, Tmp_64x1);
Tmp_32x2x2 = vuzp_s32(vget_low_s32(vreinterpretq_s32_s64(Tmp_64x2)), vget_high_s32(vreinterpretq_s32_s64(Tmp_64x2)));
Tmp_32x2 = Tmp_32x2x2.val[0]; // LSB of Tmp_64x2 (MSB is dummy)
TempX_64x2 = vqdmull_s32(Tmp_32x2, Tmp_32x2);
TempX_64x1 = vget_low_s64(TempX_64x2);
diffX_64x1 = vqsub_s64(Peak_64x1, TempX_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(diffX_64x1, 63)); // sign(diffX)
diffX_64x2 = vcombine_s64(diffX_64x1, diffX_64x1);
diffX_low_32x2 = vshrn_n_s64(vshlq_n_s64(diffX_64x2, 32), 32); // wextract_l(diffX), wextract_l(diffX)
diffX_high_32x2 = vrshrn_n_s64(diffX_64x2, 32); // wround_L(diffX), wround_L(diffX)
Tmp_64x2 = vmovl_s32(vqrdmulh_s32(LimiterPeakCoef_32x2, diffX_low_32x2)); // (MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), (MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef)
Tmp_64x2 = vqdmlal_s32(Tmp_64x2, LimiterPeakCoef_32x2, diffX_high_32x2); // wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), wL_fmul(wround_L(diffX), LimiterPeakAtCoef)), wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef), wL_fmul(wround_L(diffX), LimiterPeakReCoef))
Tmp_64x2 = vqaddq_s64(TempX_64x2, Tmp_64x2);
Peak_64x1 = vbsl_s64(bsl_u64x1, vget_low_s64(Tmp_64x2), vget_high_s64(Tmp_64x2));
Tmp_64x1 = vqsub_s64(Peak_64x1, PeakMax_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Peak_64x1 - PeakMax_64x1)
PeakMax_64x1 = vbsl_s64(bsl_u64x1, PeakMax_64x1, Peak_64x1);
}
}
else
{
// compensate Peak according to the previous shift bits
Peak_64x1 = vqrshl_s64(Peak_64x1, Peak_scale_64x1);
// amplitude measure
Zero_s64x1 = vdup_n_s64(0);
for(k = cpt1; k > 0; k--)
{
Tmp_64x1 = vld1_s64(Ptr);
Ptr++;
vst1_s64(Ptr2, Tmp_64x1);
Ptr2++;
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Tmp_64x1)
TempX_64x1 = vqsub_s64(Zero_s64x1, Tmp_64x1); // -Tmp_64x1
TempX_64x1 = vbsl_s64(bsl_u64x1, TempX_64x1, Tmp_64x1);
TempX_64x1 = vqrshl_s64(TempX_64x1, MaxShiftBits_hd_64x1);
TempX_64x2 = vcombine_s64(TempX_64x1, TempX_64x1);
diffX_64x1 = vqsub_s64(Peak_64x1, TempX_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(diffX_64x1, 63)); // sign(diffX)
diffX_64x2 = vcombine_s64(diffX_64x1, diffX_64x1);
diffX_low_32x2 = vshrn_n_s64(vshlq_n_s64(diffX_64x2, 32), 32); // wextract_l(diffX), wextract_l(diffX)
diffX_high_32x2 = vrshrn_n_s64(diffX_64x2, 32); // wround_L(diffX), wround_L(diffX)
Tmp_64x2 = vmovl_s32(vqrdmulh_s32(LimiterPeakCoef_32x2, diffX_low_32x2)); // (MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), (MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef)
Tmp_64x2 = vqdmlal_s32(Tmp_64x2, LimiterPeakCoef_32x2, diffX_high_32x2); // wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), wL_fmul(wround_L(diffX), LimiterPeakAtCoef)), wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef), wL_fmul(wround_L(diffX), LimiterPeakReCoef))
Tmp_64x2 = vqaddq_s64(TempX_64x2, Tmp_64x2);
Peak_64x1 = vbsl_s64(bsl_u64x1, vget_low_s64(Tmp_64x2), vget_high_s64(Tmp_64x2));
Tmp_64x1 = vqsub_s64(Peak_64x1, PeakMax_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Peak_64x1 - PeakMax_64x1)
PeakMax_64x1 = vbsl_s64(bsl_u64x1, PeakMax_64x1, Peak_64x1);
}
Ptr2 = HeapPtr->LimiterLABuf[ch];
for(k = cpt2; k > 0; k--)
{
Tmp_64x1 = vld1_s64(Ptr);
Ptr++;
vst1_s64(Ptr2, Tmp_64x1);
Ptr2++;
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Tmp_64x1)
TempX_64x1 = vqsub_s64(Zero_s64x1, Tmp_64x1); // -Tmp_64x1
TempX_64x1 = vbsl_s64(bsl_u64x1, TempX_64x1, Tmp_64x1);
TempX_64x1 = vqrshl_s64(TempX_64x1, MaxShiftBits_hd_64x1);
TempX_64x2 = vcombine_s64(TempX_64x1, TempX_64x1);
diffX_64x1 = vqsub_s64(Peak_64x1, TempX_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(diffX_64x1, 63)); // sign(diffX)
diffX_64x2 = vcombine_s64(diffX_64x1, diffX_64x1);
diffX_low_32x2 = vshrn_n_s64(vshlq_n_s64(diffX_64x2, 32), 32); // wextract_l(diffX), wextract_l(diffX)
diffX_high_32x2 = vrshrn_n_s64(diffX_64x2, 32); // wround_L(diffX), wround_L(diffX)
Tmp_64x2 = vmovl_s32(vqrdmulh_s32(LimiterPeakCoef_32x2, diffX_low_32x2)); // (MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), (MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef)
Tmp_64x2 = vqdmlal_s32(Tmp_64x2, LimiterPeakCoef_32x2, diffX_high_32x2); // wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakAtCoef), wL_fmul(wround_L(diffX), LimiterPeakAtCoef)), wL_addsat((MMlong) wfmulr(wextract_l(diffX), LimiterPeakReCoef), wL_fmul(wround_L(diffX), LimiterPeakReCoef))
Tmp_64x2 = vqaddq_s64(TempX_64x2, Tmp_64x2);
Peak_64x1 = vbsl_s64(bsl_u64x1, vget_low_s64(Tmp_64x2), vget_high_s64(Tmp_64x2));
Tmp_64x1 = vqsub_s64(Peak_64x1, PeakMax_64x1);
bsl_u64x1 = vreinterpret_u64_s64(vshr_n_s64(Tmp_64x1, 63)); // sign(Peak_64x1 - PeakMax_64x1)
PeakMax_64x1 = vbsl_s64(bsl_u64x1, PeakMax_64x1, Peak_64x1);
}
}
HeapPtr->LimiterPeak[ch] = vget_lane_s64(Peak_64x1, 0); // save history
} // for(ch = 0...)
PeakMax = vget_lane_s64(PeakMax_64x1, 0);
HeapPtr->PrevShiftBits = MaxShiftBits;
STOP_PMU_MEASURE(PMU_MEASURE_MRDC5B_LIMITER_INSERT_NEW_SUBBAND)
if(PeakMax < MDRC5B_ALMOST_ZERO_THRESH)
{
PeakdB = (MDRC5B_POWER_DB_MINUS_INF << 16); // 8.16, [-128.0, 127.0] dB
}
else
{
Peak_u32x2x2 = vuzp_u32(vreinterpret_u32_s64(PeakMax_64x1), vreinterpret_u32_s64(PeakMax_64x1));
bsl_u32x2 = vceq_u32(Peak_u32x2x2.val[1], vdup_n_u32(0));
Peak_exp_u32x2 = vadd_u32(vclz_u32(Peak_u32x2x2.val[0]), vdup_n_u32(32));
Peak_exp2_u32x2 = vclz_u32(Peak_u32x2x2.val[1]);
Peak_exp_u32x2 = vbsl_u32(bsl_u32x2, Peak_exp_u32x2, Peak_exp2_u32x2);
Peak_mant_u32x2 = vrshrn_n_u64(vshlq_u64(vreinterpretq_u64_s64(vcombine_s64(PeakMax_64x1, PeakMax_64x1)), vreinterpretq_s64_u64(vmovl_u32(Peak_exp_u32x2))), 32);
// if(Peak_mant >= sqrt(0.5))
// {
// Peak_exp--;
// Peak_mant >>= 1;
// }
bsl_u32x2 = vcge_u32(Peak_mant_u32x2, vdup_n_u32(0xB504F334));
Peak_exp_u32x2 = vbsl_u32(bsl_u32x2, vsub_u32(Peak_exp_u32x2, vdup_n_u32(1)), Peak_exp_u32x2);
Peak_mant_u32x2 = vbsl_u32(bsl_u32x2, vrshr_n_u32(Peak_mant_u32x2, 1), Peak_mant_u32x2);
Peak_exp_32x2 = vreinterpret_s32_u32(Peak_exp_u32x2);
#ifdef SAMPLES_24_BITS
// correction of 16 bits if input samples are 24 bits
Peak_exp_32x2 = vsub_s32(Peak_exp_32x2, vdup_n_s32(16));
#endif // SAMPLES_24_BITS
// at this point : sqrt(0.5)/2 <= Peak_mant < sqrt(0.5)
//
// ln(1+x) = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7 - x^8/8 + x^9/9 - x^10/10 ... accuracy OK if |x| < 0.5
// sqrt(0.5)/2 <= Peak_mant < sqrt(0.5) => sqrt(0.5)-1 <= 2*Peak_mant-1 < 2*sqrt(0.5)-1
// => ln(Peak_mant) = ln(1+x)-ln(2) with x=2*Peak_mant-1, i.e. |x| < 0.414214...
// x=2*PeakMax_mant-1 in Q31
// => sqrt(0.5)-1 <= x < 2*sqrt(0.5)-1
x_32x2 = vreinterpret_s32_u32(vsub_u32(Peak_mant_u32x2, vdup_n_u32(0x80000000)));
PeakdB_32x2 = x_32x2; // PeakdB = x
xn_32x2 = vqrdmulh_s32(x_32x2, x_32x2); // xn = x^2
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vrshr_n_s32(xn_32x2, 1)); // PeakdB = x - x^2/2
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^3
PeakdB_32x2 = vqadd_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x2AAAAAAB))); // PeakdB = x - x^2/2 + x^3/3
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^4
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vrshr_n_s32(xn_32x2, 2)); // PeakdB = x - x^2/2 + x^3/3 - x^4/4
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^5
PeakdB_32x2 = vqadd_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x1999999A))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^6
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x15555555))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^7
PeakdB_32x2 = vqadd_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x12492492))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^8
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vrshr_n_s32(xn_32x2, 3)); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7 - x^8/8
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^9
PeakdB_32x2 = vqadd_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x0E38E38E))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7 - x^8/8 + x^9/9
xn_32x2 = vqrdmulh_s32(xn_32x2, x_32x2); // xn = x^10
PeakdB_32x2 = vqsub_s32(PeakdB_32x2, vqrdmulh_s32(xn_32x2, vdup_n_s32(0x0CCCCCCD))); // PeakdB = x - x^2/2 + x^3/3 - x^4/4 + x^5/5 - x^6/6 + x^7/7 - x^8/8 + x^9/9 - x^10/10
// at this point : PeakMaxdB contains ln(1+x) in Q31
if(RmsMeasure)
{
// dB(power) = 10*log10(power)
// PeakMaxdB = 10*log10(PeakMax)+20*log10(2)*(HEADROOM+MaxShiftBits)
// = 10*ln(PeakMax)/ln(10)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 10/ln(10)*ln(PeakMax_mant*2^(-PeakMax_exp))+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 10/ln(10)*(ln(PeakMax_mant)-PeakMax_exp*ln(2))+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 10/ln(10)*ln(PeakMax_mant)-PeakMax_exp*10*ln(2)/ln(10)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 10/ln(10)*ln(PeakMax_mant)+10*ln(2)/ln(10)*(2*(HEADROOM+MaxShiftBits)-PeakMax_exp)
//
// => RmsdB = 10/ln(10)*ln(1+x)+10*ln(2)/ln(10)*(2*(HEADROOM+MaxShiftBits)-PeakMax_exp)
// => RmsdB (Q16) = 0x457CB*ln(1+x)+0x302A3*(2*(HEADROOM+MaxShiftBits)-PeakMax_exp)
// fractional mutiply 0x457CB*ln(1+x) in Q16
PeakdB_32x2 = vqrdmulh_s32(PeakdB_32x2, vdup_n_s32(0x457CB));
// PeakdB_exp = 2*(HEADROOM+MaxShiftBits)-PeakdB_exp
Peak_exp_32x2 = vsub_s32(vdup_n_s32(2 * (HEADROOM + MaxShiftBits)), Peak_exp_32x2);
// PeakMaxdB final value (integer mac 0x302A3*PeakdB_exp)
PeakdB_32x2 = vmla_s32(PeakdB_32x2, Peak_exp_32x2, vdup_n_s32(0x302A3));
}
else
{
// dB(power) = 20*log10(abs)
// PeakMaxdB = 20*log10(PeakMax)+20*log10(2)*(HEADROOM+MaxShiftBits)
// = 20*ln(PeakMax)/ln(10)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 20/ln(10)*ln(PeakMax_mant*2^(-PeakMax_exp))+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 20/ln(10)*(ln(PeakMax_mant)-PeakMax_exp*ln(2))+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 20/ln(10)*ln(PeakMax_mant)-PeakMax_exp*20*ln(2)/ln(10)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits)
// = 20/ln(10)*ln(PeakMax_mant)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits-PeakMax_exp)
//
// => RmsdB = 20/ln(10)*ln(1+x)+20*ln(2)/ln(10)*(HEADROOM+MaxShiftBits-PeakMax_exp)
// => RmsdB (Q16) = 0x8AF96*ln(1+x)+0x60546*(HEADROOM+MaxShiftBits-PeakMax_exp)
// fractional mutiply 0x8AF96*ln(1+x) in Q16
PeakdB_32x2 = vqrdmulh_s32(PeakdB_32x2, vdup_n_s32(0x8AF96));
// PeakdB_exp = HEADROOM+MaxShiftBits-PeakdB_exp
Peak_exp_32x2 = vsub_s32(vdup_n_s32(HEADROOM + MaxShiftBits), Peak_exp_32x2);
// PeakMaxdB final value (integer mac 0x60546*PeakdB_exp)
PeakdB_32x2 = vmla_s32(PeakdB_32x2, Peak_exp_32x2, vdup_n_s32(0x60546));
}
PeakdB = vget_lane_s32(PeakdB_32x2, 0);
}
#ifdef DEBUG_LIMITER_OUTPUT
if((debug_cpt_samples >= DEBUG_CPT_MIN) && (debug_cpt_samples <= DEBUG_CPT_MAX))
{
char string[100];
debug_write_string("MRDC5B_LIMITER_PEAKMAX_PEAKDB\n");
sprintf(string, "PeakMax=0x%012llX, HEADROOM+MaxShiftBits=%d => PeakdB=0x%06X\n",
#ifdef SAMPLES_24_BITS
PeakMax & 0xFFFFFFFFFFFFLL,
#else // SAMPLES_24_BITS
(PeakMax >> 16) & 0xFFFFFFFFFFFFLL,
#endif // SAMPLES_24_BITS
HEADROOM + MaxShiftBits,
PeakdB & 0xFFFFFF);
debug_write_string(string);
}
// CHECK-LABEL: define i64 @test_vget_lane_s64(<1 x i64> %a) #0 {
// CHECK: [[TMP0:%.*]] = bitcast <1 x i64> %a to <8 x i8>
// CHECK: [[TMP1:%.*]] = bitcast <8 x i8> [[TMP0]] to <1 x i64>
// CHECK: [[VGET_LANE:%.*]] = extractelement <1 x i64> [[TMP1]], i32 0
// CHECK: ret i64 [[VGET_LANE]]
int64_t test_vget_lane_s64(int64x1_t a) {
return vget_lane_s64(a, 0);
}
f64 dotProduct(const Size2D &_size,
const s8 * src0Base, ptrdiff_t src0Stride,
const s8 * src1Base, ptrdiff_t src1Stride)
{
internal::assertSupportedConfiguration();
#ifdef CAROTENE_NEON
Size2D size(_size);
if (src0Stride == src1Stride &&
src0Stride == (ptrdiff_t)(size.width))
{
size.width *= size.height;
size.height = 1;
}
// It is possible to accumulate up to 131071 schar multiplication results in sint32 without overflow
// We process 16 elements and accumulate two new elements per step. So we could handle 131071/2*16 elements
#define DOT_INT_BLOCKSIZE 131070*8
f64 result = 0.0;
for (size_t row = 0; row < size.height; ++row)
{
const s8 * src0 = internal::getRowPtr(src0Base, src0Stride, row);
const s8 * src1 = internal::getRowPtr(src1Base, src1Stride, row);
size_t i = 0;
int64x2_t ws = vmovq_n_s64(0);
while(i + 16 <= size.width)
{
size_t lim = std::min(i + DOT_UINT_BLOCKSIZE, size.width) - 16;
int32x4_t s1 = vmovq_n_s32(0);
int32x4_t s2 = vmovq_n_s32(0);
for (; i <= lim; i += 16)
{
internal::prefetch(src0 + i);
internal::prefetch(src1 + i);
int8x16_t vs1 = vld1q_s8(src0 + i);
int8x16_t vs2 = vld1q_s8(src1 + i);
int16x8_t vdot1 = vmull_s8(vget_low_s8(vs1), vget_low_s8(vs2));
int16x8_t vdot2 = vmull_s8(vget_high_s8(vs1), vget_high_s8(vs2));
s1 = vpadalq_s16(s1, vdot1);
s2 = vpadalq_s16(s2, vdot2);
}
ws = vpadalq_s32(ws, s1);
ws = vpadalq_s32(ws, s2);
}
if(i + 8 <= size.width)
{
int8x8_t vs1 = vld1_s8(src0 + i);
int8x8_t vs2 = vld1_s8(src1 + i);
ws = vpadalq_s32(ws, vpaddlq_s16(vmull_s8(vs1, vs2)));
i += 8;
}
result += (double)vget_lane_s64(vadd_s64(vget_low_s64(ws), vget_high_s64(ws)), 0);
for (; i < size.width; ++i)
result += s32(src0[i]) * s32(src1[i]);
}
return result;
#else
(void)_size;
(void)src0Base;
(void)src0Stride;
(void)src1Base;
(void)src1Stride;
return 0;
#endif
}
