internal static int stereo_itheta(int[] X, int X_ptr, int[] Y, int Y_ptr, int stereo, int N)
{
int i;
int itheta;
int mid, side;
int Emid, Eside;
Emid = Eside = CeltConstants.EPSILON;
if (stereo != 0)
{
for (i = 0; i < N; i++)
{
int m, s;
m = Inlines.ADD16(Inlines.SHR16(X[X_ptr + i], 1), Inlines.SHR16(Y[Y_ptr + i], 1));
s = Inlines.SUB16(Inlines.SHR16(X[X_ptr + i], 1), Inlines.SHR16(Y[Y_ptr + i], 1));
Emid = Inlines.MAC16_16(Emid, m, m);
Eside = Inlines.MAC16_16(Eside, s, s);
}
}
else
{
Emid += Kernels.celt_inner_prod(X, X_ptr, X, X_ptr, N);
Eside += Kernels.celt_inner_prod(Y, Y_ptr, Y, Y_ptr, N);
}
mid = (Inlines.celt_sqrt(Emid));
side = (Inlines.celt_sqrt(Eside));
/* 0.63662 = 2/pi */
itheta = Inlines.MULT16_16_Q15(((short)(0.5 + (0.63662f) * (((int)1) << (15)))) /*Inlines.QCONST16(0.63662f, 15)*/, Inlines.celt_atan2p(side, mid));
return(itheta);
}
internal static uint alg_quant(int[] X, int X_ptr, int N, int K, int spread, int B, EntropyCoder enc
)
{
int[] y = new int[N];
int[] iy = new int[N];
int[] signx = new int[N];
int i, j;
int s;
int pulsesLeft;
int sum;
int xy;
int yy;
uint collapse_mask;
Inlines.OpusAssert(K > 0, "alg_quant() needs at least one pulse");
Inlines.OpusAssert(N > 1, "alg_quant() needs at least two dimensions");
exp_rotation(X, X_ptr, N, 1, B, K, spread);
/* Get rid of the sign */
sum = 0;
j = 0;
do
{
int xpj = X_ptr + j;
/* OPT: Make sure the following two lines result in conditional moves
* rather than branches. */
signx[j] = X[xpj] > 0 ? 1 : -1;
X[xpj] = Inlines.ABS16(X[xpj]);
iy[j] = 0;
y[j] = 0;
} while (++j < N);
xy = yy = 0;
pulsesLeft = K;
/* Do a pre-search by projecting on the pyramid */
if (K > (N >> 1))
{
int rcp;
j = 0; do
{
sum += X[X_ptr + j];
} while (++j < N);
/* If X is too small, just replace it with a pulse at 0 */
/* Prevents infinities and NaNs from causing too many pulses
* to be allocated. 64 is an approximation of infinity here. */
if (sum <= K)
{
X[X_ptr] = ((short)(0.5 + (1.0f) * (((int)1) << (14)))) /*Inlines.QCONST16(1.0f, 14)*/;
j = X_ptr + 1;
do
{
X[j] = 0;
} while (++j < N + X_ptr);
sum = ((short)(0.5 + (1.0f) * (((int)1) << (14)))) /*Inlines.QCONST16(1.0f, 14)*/;
}
rcp = Inlines.EXTRACT16(Inlines.MULT16_32_Q16((K - 1), Inlines.celt_rcp(sum)));
j = 0;
do
{
/* It's really important to round *towards zero* here */
iy[j] = Inlines.MULT16_16_Q15(X[X_ptr + j], rcp);
y[j] = (int)iy[j];
yy = (Inlines.MAC16_16(yy, y[j], y[j]));
xy = Inlines.MAC16_16(xy, X[X_ptr + j], y[j]);
y[j] *= 2;
pulsesLeft -= iy[j];
} while (++j < N);
}
Inlines.OpusAssert(pulsesLeft >= 1, "Allocated too many pulses in the quick pass");
/* This should never happen, but just in case it does (e.g. on silence)
* we fill the first bin with pulses. */
if (pulsesLeft > N + 3)
{
int tmp = (int)pulsesLeft;
yy = (Inlines.MAC16_16(yy, tmp, tmp));
yy = (Inlines.MAC16_16(yy, tmp, y[0]));
iy[0] += pulsesLeft;
pulsesLeft = 0;
}
s = 1;
for (i = 0; i < pulsesLeft; i++)
{
int best_id;
int best_num = 0 - CeltConstants.VERY_LARGE16;
int best_den = 0;
int rshift = 1 + Inlines.celt_ilog2(K - pulsesLeft + i + 1);
best_id = 0;
/* The squared magnitude term gets added anyway, so we might as well
* add it outside the loop */
yy = Inlines.ADD16(yy, 1); // opus bug - was add32
j = 0;
do
{
int Rxy, Ryy;
/* Temporary sums of the new pulse(s) */
Rxy = Inlines.EXTRACT16(Inlines.SHR32(Inlines.ADD32(xy, Inlines.EXTEND32(X[X_ptr + j])), rshift));
/* We're multiplying y[j] by two so we don't have to do it here */
Ryy = Inlines.ADD16(yy, y[j]);
/* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that
* Rxy is positive because the sign is pre-computed) */
Rxy = Inlines.MULT16_16_Q15(Rxy, Rxy);
/* The idea is to check for num/den >= best_num/best_den, but that way
* we can do it without any division */
/* OPT: Make sure to use conditional moves here */
if (Inlines.MULT16_16(best_den, Rxy) > Inlines.MULT16_16(Ryy, best_num))
{
best_den = Ryy;
best_num = Rxy;
best_id = j;
}
} while (++j < N);
/* Updating the sums of the new pulse(s) */
xy = Inlines.ADD32(xy, Inlines.EXTEND32(X[X_ptr + best_id]));
/* We're multiplying y[j] by two so we don't have to do it here */
yy = Inlines.ADD16(yy, y[best_id]);
/* Only now that we've made the final choice, update y/iy */
/* Multiplying y[j] by 2 so we don't have to do it everywhere else */
y[best_id] = (y[best_id] + (2 * s));
iy[best_id]++;
}
/* Put the original sign back */
j = 0;
do
{
X[X_ptr + j] = (Inlines.MULT16_16(signx[j], X[X_ptr + j]));
/* OPT: Make sure your compiler uses a conditional move here rather than
* a branch. */
iy[j] = signx[j] < 0 ? -iy[j] : iy[j];
} while (++j < N);
CWRS.encode_pulses(iy, N, K, enc);
collapse_mask = extract_collapse_mask(iy, N, B);
return(collapse_mask);
}
