short DSP_maxval (
const short *x, /* x[nx] = input vector */
int nx /* nx = number of elements */
)
{
int i;
const long long *xll;
double x0123, x4567;
int max01, max23, max45, max67;
/* Set all 8 intermediate max values to most negative */
/* Each 32bit var contains two shorts */
max01 = 0x80008000;
max23 = 0x80008000;
max45 = 0x80008000;
max67 = 0x80008000;
/* Convert the short pointer to a 64bit long long pointer */
xll = (const long long *)x;
/* In each loop iteration we will load 8 short values from the array. */
/* On the C64x+ we can do 4 max2 operations in one cycle. This will */
/* give us 8 results, that we keep seperated. Outside the loop we'll */
/* find the max out of these 8 intermediate values. */
_nassert((int)(xll) % 8 == 0);
#pragma MUST_ITERATE(1,,1);
for (i = 0; i < nx; i += 8) {
x0123 = _amemd8((void *)xll++); /* Use LDDW to load 4 shorts */
x4567 = _amemd8((void *)xll++); /* Use LDDW to load 4 shorts */
max01 = _max2(max01, _lo(x0123));
max23 = _max2(max23, _hi(x0123));
max45 = _max2(max45, _lo(x4567));
max67 = _max2(max67, _hi(x4567));
}
max01 = _max2(max01, max23); /* Calculate 2 maximums of max01 and max23 */
max45 = _max2(max45, max67); /* Calculate 2 maximums of max45 and max67 */
max01 = _max2(max01, max45); /* Get the 2 max values of the remaining 4 */
max45 = _rotl(max01, 16); /* Swap lower and higher 16 bit */
/* Find the final maximum value (will be in higher and lower part) */
max01 = _max2(max01, max45);
/* max01 is a 32-bit value with the result in the upper and lower 16 bit */
/* Use an AND operation to only return the lower 16 bit to the caller. */
return (max01 & 0xFFFF);
}
void FilterSlot::set(int slot_nr, NRPixBlock *pb)
{
/* Unnamed slot is for saving filter primitive results, when parameter
* 'result' is not set. Only the filter immediately after this one
* can access unnamed results, so we don't have to worry about overwriting
* previous results in filter chain. On the other hand, we may not
* overwrite any other image with this one, because they might be
* accessed later on. */
int index = ((slot_nr != NR_FILTER_SLOT_NOT_SET)
? _get_index(slot_nr)
: _get_index(NR_FILTER_UNNAMED_SLOT));
assert(index >= 0);
// Unnamed slot is only for Inkscape::Filters::FilterSlot internal use.
assert(slot_nr != NR_FILTER_UNNAMED_SLOT);
assert(slot_nr == NR_FILTER_SLOT_NOT_SET ||_slot_number[index] == slot_nr);
if (slot_nr == NR_FILTER_SOURCEGRAPHIC || slot_nr == NR_FILTER_BACKGROUNDIMAGE) {
Geom::Matrix trans = units.get_matrix_display2pb();
if (fabs(trans[1]) > 1e-6 || fabs(trans[2]) > 1e-6) {
NRPixBlock *trans_pb = new NRPixBlock;
int x0 = pb->area.x0;
int y0 = pb->area.y0;
int x1 = pb->area.x1;
int y1 = pb->area.y1;
int min_x = _min4(trans[0] * x0 + trans[2] * y0 + trans[4],
trans[0] * x0 + trans[2] * y1 + trans[4],
trans[0] * x1 + trans[2] * y0 + trans[4],
trans[0] * x1 + trans[2] * y1 + trans[4]);
int max_x = _max4(trans[0] * x0 + trans[2] * y0 + trans[4],
trans[0] * x0 + trans[2] * y1 + trans[4],
trans[0] * x1 + trans[2] * y0 + trans[4],
trans[0] * x1 + trans[2] * y1 + trans[4]);
int min_y = _min4(trans[1] * x0 + trans[3] * y0 + trans[5],
trans[1] * x0 + trans[3] * y1 + trans[5],
trans[1] * x1 + trans[3] * y0 + trans[5],
trans[1] * x1 + trans[3] * y1 + trans[5]);
int max_y = _max4(trans[1] * x0 + trans[3] * y0 + trans[5],
trans[1] * x0 + trans[3] * y1 + trans[5],
trans[1] * x1 + trans[3] * y0 + trans[5],
trans[1] * x1 + trans[3] * y1 + trans[5]);
nr_pixblock_setup_fast(trans_pb, pb->mode,
min_x, min_y,
max_x, max_y, true);
if (trans_pb->size != NR_PIXBLOCK_SIZE_TINY && trans_pb->data.px == NULL) {
/* TODO: this gets hit occasionally. Worst case scenario:
* images are exported in horizontal stripes. One stripe
* is not too high, but can get thousands of pixels wide.
* Rotate this 45 degrees -> _huge_ image */
g_warning("Memory allocation failed in Inkscape::Filters::FilterSlot::set (transform)");
return;
}
if (filterquality == FILTER_QUALITY_BEST) {
NR::transform_bicubic(trans_pb, pb, trans);
} else {
NR::transform_nearest(trans_pb, pb, trans);
}
nr_pixblock_release(pb);
delete pb;
pb = trans_pb;
} else if (fabs(trans[0] - 1) > 1e-6 || fabs(trans[3] - 1) > 1e-6) {
NRPixBlock *trans_pb = new NRPixBlock;
int x0 = pb->area.x0;
int y0 = pb->area.y0;
int x1 = pb->area.x1;
int y1 = pb->area.y1;
int min_x = _min2(trans[0] * x0 + trans[4],
trans[0] * x1 + trans[4]);
int max_x = _max2(trans[0] * x0 + trans[4],
trans[0] * x1 + trans[4]);
int min_y = _min2(trans[3] * y0 + trans[5],
trans[3] * y1 + trans[5]);
int max_y = _max2(trans[3] * y0 + trans[5],
trans[3] * y1 + trans[5]);
nr_pixblock_setup_fast(trans_pb, pb->mode,
min_x, min_y, max_x, max_y, true);
if (trans_pb->size != NR_PIXBLOCK_SIZE_TINY && trans_pb->data.px == NULL) {
g_warning("Memory allocation failed in Inkscape::Filters::FilterSlot::set (scaling)");
return;
}
NR::scale_bicubic(trans_pb, pb);
nr_pixblock_release(pb);
delete pb;
pb = trans_pb;
}
}
if(_slot[index]) {
nr_pixblock_release(_slot[index]);
delete _slot[index];
}
_slot[index] = pb;
_last_out = index;
}
void Calc_Exc_Rand(Word16 curGain, Word16 *PrevExc, Word16 *DataExc,
Word16 *nRandom, LINEDEF *Line,eG723Mode mode)
{
int i, i_subfr, iblk;
Word16 temp, temp2;
Word16 j;
Word16 TabPos[2*NbPulsBlk], TabSign[2*NbPulsBlk];
Word16 *ptr_TabPos, *ptr_TabSign;
Word16 *ptr1, *curExc;
Word16 sh1, x1, x2, inter_exc, delta, b0;
Word32 L_acc, L_c, L_temp;
Word16 tmp[SubFrLen/Sgrid];
Word16 offset[SubFrames];
Word16 tempExc[SubFrLenD];
/*
* generate LTP codes
*/
Line->Olp[0] = random_number(21, nRandom) + (Word16)123;
Line->Olp[1] = random_number(21, nRandom) + (Word16)123;
for(i_subfr=0; i_subfr<SubFrames; i_subfr++) { /* in [1, NbFilt] */
Line->Sfs[i_subfr].AcGn = random_number(NbFilt, nRandom) + (Word16)1;
}
Line->Sfs[0].AcLg = 1;
Line->Sfs[1].AcLg = 0;
Line->Sfs[2].AcLg = 1;
Line->Sfs[3].AcLg = 3;
/*
* Random innovation :
* Selection of the grids, signs and pulse positions
*/
/* Signs and Grids */
ptr_TabSign = TabSign;
ptr1 = offset;
for(iblk=0; iblk<SubFrames/2; iblk++) {
temp = random_number((1 << (NbPulsBlk+2)), nRandom);
*ptr1++ = temp & (Word16)0x0001;
temp = shr(temp, 1);
*ptr1++ = add( (Word16) SubFrLen, (Word16) (temp & 0x0001) );
for(i=0; i<NbPulsBlk; i++) {
*ptr_TabSign++= shl(sub((temp & (Word16)0x0002), 1), 14);
temp = shr(temp, 1);
}
}
/* Positions */
ptr_TabPos = TabPos;
for(i_subfr=0; i_subfr<SubFrames; i_subfr++) {
for(i=0; i<(SubFrLen/Sgrid); i++) tmp[i] = (Word16)i;
temp = (SubFrLen/Sgrid);
for(i=0; i<Nb_puls[i_subfr]; i++) {
j = random_number(temp, nRandom);
*ptr_TabPos++ = add(shl(tmp[(int)j],1), offset[i_subfr]);
temp = sub(temp, 1);
tmp[(int)j] = tmp[(int)temp];
}
}
/*
* Compute fixed codebook gains
*/
ptr_TabPos = TabPos;
ptr_TabSign = TabSign;
curExc = DataExc;
i_subfr = 0;
for(iblk=0; iblk<SubFrames/2; iblk++) {
/* decode LTP only */
Decod_Acbk(curExc, &PrevExc[0], Line->Olp[iblk],
Line->Sfs[i_subfr].AcLg, Line->Sfs[i_subfr].AcGn,mode);
Decod_Acbk(&curExc[SubFrLen], &PrevExc[SubFrLen], Line->Olp[iblk],
Line->Sfs[i_subfr+1].AcLg, Line->Sfs[i_subfr+1].AcGn,mode);
temp2 = 0;
for(i=0; i<SubFrLenD; i++) {
temp = abs_s(curExc[i]);
#ifdef VC
if(temp > temp2) temp2 = temp;
#endif
#ifdef CCS
temp2=_max2(temp2,temp);
#endif
}
if(temp2 == 0) sh1 = 0;
else {
sh1 = sub(4,norm_s(temp2)); /* 4 bits of margin */
if(sh1 < -2) sh1 = -2;
}
L_temp = 0L;
for(i=0; i<SubFrLenD; i++) {
temp = shr(curExc[i], sh1); /* left if sh1 < 0 */
L_temp = L_mac(L_temp, temp, temp);
tempExc[i] = temp;
} /* ener_ltp x 2**(-2sh1+1) */
L_acc = 0L;
for(i=0; i<NbPulsBlk; i++) {
L_acc = L_mac(L_acc, tempExc[(int)ptr_TabPos[i]], ptr_TabSign[i]);
}
inter_exc = extract_h(L_shl(L_acc, 1)); /* inter_exc x 2-sh1 */
/* compute SubFrLenD x curGain**2 x 2**(-2sh1+1) */
/* curGain input = 2**5 curGain */
L_acc = L_mult(curGain, SubFrLen);
L_acc = L_shr(L_acc, 6);
temp = extract_l(L_acc); /* SubFrLen x curGain : avoids overflows */
L_acc = L_mult(temp, curGain);
temp = shl(sh1, 1);
temp = add(temp, 4);
L_acc = L_shr(L_acc, temp); /* SubFrLenD x curGain**2 x 2**(-2sh1+1) */
/* c = (ener_ltp - SubFrLenD x curGain**2)/nb_pulses_blk */
/* compute L_c = c >> 2sh1-1 */
L_acc = L_sub(L_temp, L_acc);
/* x 1/nb_pulses_blk */
L_c = L_mls(L_acc, InvNbPulsBlk);
/*
* Solve EQ(X) = X**2 + 2 b0 X + c
*/
/* delta = b0 x b0 - c */
b0 = mult_r(inter_exc, InvNbPulsBlk); /* b0 >> sh1 */
L_acc = L_msu(L_c, b0, b0); /* (c - b0**2) >> 2sh1-1 */
L_acc = L_negate(L_acc); /* delta x 2**(-2sh1+1) */
/* Case delta <= 0 */
if(L_acc <= 0) { /* delta <= 0 */
x1 = negate(b0); /* sh1 */
}
/* Case delta > 0 */
else {
delta = Sqrt_lbc(L_acc); /* >> sh1 */
x1 = sub(delta, b0); /* x1 >> sh1 */
x2 = add(b0, delta); /* (-x2) >> sh1 */
if(abs_s(x2) < abs_s(x1)) {
x1 = negate(x2);
}
}
/* Update DataExc */
sh1 = add(sh1, 1);
temp = shl(x1, sh1);
#ifdef VC
if(temp > (2*Gexc_Max)) temp = (2*Gexc_Max);
if(temp < -(2*Gexc_Max)) temp = -(2*Gexc_Max);
#endif
#ifdef CCS
temp=_min2(temp,(2*Gexc_Max));
temp=_max2(temp,-(2*Gexc_Max));
#endif
for(i=0; i<NbPulsBlk; i++) {
j = *ptr_TabPos++;
curExc[(int)j] = add(curExc[(int)j], mult(temp,
(*ptr_TabSign++)) );
}
/* update PrevExc */
ptr1 = PrevExc;
for(i=SubFrLenD; i<PitchMax; i++) *ptr1++ = PrevExc[i];
for(i=0; i<SubFrLenD; i++) *ptr1++ = curExc[i];
curExc += SubFrLenD;
i_subfr += 2;
} /* end of loop on LTP blocks */
return;
}
