ref: a56c34f3816ece0cd46338c04af97cd715da9bc0
parent: 25aeed74f17bd0c99f549e91e117e3dd1d908099
author: Jean-Marc Valin <[email protected]>
date: Wed Mar 26 10:36:07 EDT 2008
optimisations: faster handling of the zero for compute_band_energies() and normalise_bands(). Also, another bunch of restrict qualifiers -- mainly to tell the compiler there's no aliasing between the array bounds and the data.
--- a/libcelt/bands.c
+++ b/libcelt/bands.c
@@ -94,14 +94,14 @@
maxval = MAX32(maxval, ABS32(X[j*C+c]));
if (maxval > 0)
{- int shift = celt_zlog2(maxval)-10;
+ int shift = celt_ilog2(maxval)-10;
for (j=B*eBands[i];j<B*eBands[i+1];j++)
sum += VSHR32(X[j*C+c],shift)*VSHR32(X[j*C+c],shift);
/* We're adding one here to make damn sure we never end up with a pitch vector that's
larger than unity norm */
- bank[i*C+c] = 1+VSHR32(EXTEND32(celt_sqrt(sum)),-shift);
+ bank[i*C+c] = EPSILON+VSHR32(EXTEND32(celt_sqrt(sum)),-shift);
} else {- bank[i*C+c] = 0;
+ bank[i*C+c] = EPSILON;
}
/*printf ("%f ", bank[i*C+c]);*/}
@@ -110,7 +110,7 @@
}
/* Normalise each band such that the energy is one. */
-void normalise_bands(const CELTMode *m, const celt_sig_t *freq, celt_norm_t *X, const celt_ener_t *bank)
+void normalise_bands(const CELTMode *m, const celt_sig_t * restrict freq, celt_norm_t * restrict X, const celt_ener_t *bank)
{int i, c, B, C;
const celt_int16_t *eBands = m->eBands;
@@ -125,10 +125,7 @@
celt_word16_t E;
shift = celt_zlog2(bank[i*C+c])-13;
E = VSHR32(bank[i*C+c], shift);
- if (E>0)
- g = EXTRACT16(celt_rcp(SHR32(MULT16_16(E,sqrtC_1[C-1]),11)));
- else
- g = 0;
+ g = EXTRACT16(celt_rcp(SHR32(MULT16_16(E,sqrtC_1[C-1]),11)));
for (j=B*eBands[i];j<B*eBands[i+1];j++)
X[j*C+c] = MULT16_16_Q14(VSHR32(freq[j*C+c],shift),g);
}
@@ -137,7 +134,7 @@
X[i] = 0;
}
-void renormalise_bands(const CELTMode *m, celt_norm_t *X)
+void renormalise_bands(const CELTMode *m, celt_norm_t * restrict X)
{int i;
VARDECL(celt_ener_t, tmpE);
@@ -175,7 +172,7 @@
}
/* Normalise each band such that the energy is one. */
-void normalise_bands(const CELTMode *m, const celt_sig_t *freq, celt_norm_t *X, const celt_ener_t *bank)
+void normalise_bands(const CELTMode *m, const celt_sig_t * restrict freq, celt_norm_t * restrict X, const celt_ener_t *bank)
{int i, c, B, C;
const celt_int16_t *eBands = m->eBands;
@@ -195,7 +192,7 @@
X[i] = 0;
}
-void renormalise_bands(const CELTMode *m, celt_norm_t *X)
+void renormalise_bands(const CELTMode *m, celt_norm_t * restrict X)
{VARDECL(celt_ener_t, tmpE);
SAVE_STACK;
@@ -207,7 +204,7 @@
#endif
/* De-normalise the energy to produce the synthesis from the unit-energy bands */
-void denormalise_bands(const CELTMode *m, const celt_norm_t *X, celt_sig_t *freq, const celt_ener_t *bank)
+void denormalise_bands(const CELTMode *m, const celt_norm_t * restrict X, celt_sig_t * restrict freq, const celt_ener_t *bank)
{int i, c, B, C;
const celt_int16_t *eBands = m->eBands;
@@ -291,7 +288,8 @@
{int i, j, B, bits;
const celt_int16_t *eBands = m->eBands;
- VARDECL(celt_norm_t, norm);
+ celt_norm_t * restrict norm;
+ VARDECL(celt_norm_t, _norm);
VARDECL(int, pulses);
VARDECL(int, offsets);
SAVE_STACK;
@@ -298,9 +296,10 @@
B = m->nbMdctBlocks*m->nbChannels;
- ALLOC(norm, B*eBands[m->nbEBands+1], celt_norm_t);
+ ALLOC(_norm, B*eBands[m->nbEBands+1], celt_norm_t);
ALLOC(pulses, m->nbEBands, int);
ALLOC(offsets, m->nbEBands, int);
+ norm = _norm;
for (i=0;i<m->nbEBands;i++)
offsets[i] = 0;
@@ -351,7 +350,8 @@
{int i, j, B, bits;
const celt_int16_t *eBands = m->eBands;
- VARDECL(celt_norm_t, norm);
+ celt_norm_t * restrict norm;
+ VARDECL(celt_norm_t, _norm);
VARDECL(int, pulses);
VARDECL(int, offsets);
SAVE_STACK;
@@ -358,9 +358,10 @@
B = m->nbMdctBlocks*m->nbChannels;
- ALLOC(norm, B*eBands[m->nbEBands+1], celt_norm_t);
+ ALLOC(_norm, B*eBands[m->nbEBands+1], celt_norm_t);
ALLOC(pulses, m->nbEBands, int);
ALLOC(offsets, m->nbEBands, int);
+ norm = _norm;
for (i=0;i<m->nbEBands;i++)
offsets[i] = 0;
--- a/libcelt/bands.h
+++ b/libcelt/bands.h
@@ -56,9 +56,9 @@
* @param X Spectrum (returned normalised)
* @param bands Square root of the energy for each band
*/
-void normalise_bands(const CELTMode *m, const celt_sig_t *freq, celt_norm_t *X, const celt_ener_t *bands);
+void normalise_bands(const CELTMode *m, const celt_sig_t * restrict freq, celt_norm_t * restrict X, const celt_ener_t *bands);
-void renormalise_bands(const CELTMode *m, celt_norm_t *X);
+void renormalise_bands(const CELTMode *m, celt_norm_t * restrict X);
/** Denormalise each band of X to restore full amplitude
* @param m Mode data
@@ -65,7 +65,7 @@
* @param X Spectrum (returned de-normalised)
* @param bands Square root of the energy for each band
*/
-void denormalise_bands(const CELTMode *m, const celt_norm_t *X, celt_sig_t *freq, const celt_ener_t *bands);
+void denormalise_bands(const CELTMode *m, const celt_norm_t * restrict X, celt_sig_t * restrict freq, const celt_ener_t *bands);
/** Compute the pitch predictor gain for each pitch band
* @param m Mode data
@@ -76,7 +76,7 @@
*/
void compute_pitch_gain(const CELTMode *m, const celt_norm_t *X, const celt_norm_t *P, celt_pgain_t *gains);
-void pitch_quant_bands(const CELTMode *m, celt_norm_t *P, const celt_pgain_t *gains);
+void pitch_quant_bands(const CELTMode *m, celt_norm_t * restrict P, const celt_pgain_t * restrict gains);
/** Quantisation/encoding of the residual spectrum
* @param m Mode data
@@ -86,7 +86,7 @@
* @param total_bits Total number of bits that can be used for the frame (including the ones already spent)
* @param enc Entropy encoder
*/
-void quant_bands(const CELTMode *m, celt_norm_t *X, celt_norm_t *P, celt_mask_t *W, int total_bits, ec_enc *enc);
+void quant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, celt_mask_t *W, int total_bits, ec_enc *enc);
/** Decoding of the residual spectrum
* @param m Mode data
@@ -95,7 +95,7 @@
* @param total_bits Total number of bits that can be used for the frame (including the ones already spent)
* @param dec Entropy decoder
*/
-void unquant_bands(const CELTMode *m, celt_norm_t *X, celt_norm_t *P, int total_bits, ec_dec *dec);
+void unquant_bands(const CELTMode *m, celt_norm_t * restrict X, celt_norm_t *P, int total_bits, ec_dec *dec);
void stereo_mix(const CELTMode *m, celt_norm_t *X, const celt_ener_t *bank, int dir);