int unpackdouble_array(double **valp, uint32_t* size_val, Buf buffer)
{
uint32_t i = 0;
if (unpack32(size_val, buffer))
return SLURM_ERROR;
*valp = xmalloc_nz((*size_val) * sizeof(double));
for (i = 0; i < *size_val; i++) {
if (unpackdouble((*valp) + i, buffer))
return SLURM_ERROR;
}
return SLURM_SUCCESS;
}
/**
* @brief Perform inverse DWT (periodic boundaries) on 3D data.
*
* @param wc Wavelet presentation of 3D data.
* @param img Reconstructed signal.
*/
void
idpwt3 (const Matrix <T> & wc, Matrix <T> & img)
{
// assign dwt to result image
img = wc;
# pragma omp parallel default (shared) num_threads (_num_threads)
{
T * wcplo, * wcphi, * templo, * temphi, * temptop, * tmp;
size_t stride;
int sl1 = _sl1_scale,
sl2 = _sl2_scale,
sl3 = _sl3_scale;
const int t_num = omp_get_thread_num ();
// loop over levels of backwards DWT
for (int j = _min_level; j < _max_level; j++)
{
// update stride
stride = 6 * sl3 * t_num;
tmp = & _temp [stride];
templo = & _temp [2 * sl3 + stride];
temphi = & _temp [3 * sl3 + stride];
temptop = & _temp [4 * sl3 + stride];
# pragma omp for schedule (OMP_SCHEDULE)
// loop over lines along third dimension ('third') of result image
for (int c1_loc = 0; c1_loc < 2 * sl1 * 2 * sl2; c1_loc++)
{
int c1_glob = (c1_loc % (2 * sl1)) + (c1_loc / (2 * sl1)) * _sl1;
// copy lowpass part of current line to temporary memory
unpackdouble (& img [c1_glob], sl3, _ld12, 0, templo);
// copy highpass part of current line to temporary memory
unpackdouble (& img [c1_glob + sl3 * _ld12], sl3, _ld12, 0, temphi);
// perform lowpass reconstruction
uplo (templo, sl3, tmp);
// perform highpass reconstruction
uphi (temphi, sl3, temptop);
// fusion of reconstruction parts
adddouble (tmp, temptop, sl3 * 2, tmp);
// write back reconstructed line
packdouble (tmp, sl3 * 2, _ld12, 0, & img [c1_glob]);
} // loop over lines along third dimension of result image
// update stride
stride = 6 * sl2 * t_num;
tmp = & _temp [stride];
templo = & _temp [2 * sl2 + stride];
temphi = & _temp [3 * sl2 + stride];
temptop = & _temp [4 * sl2 + stride];
# pragma omp for schedule (OMP_SCHEDULE)
// loop over lines along second dimension ('rows') of result image
for (int c1_loc = 0; c1_loc < 2 * sl1 * 2 * sl3; c1_loc++)
{
int c1_glob = (c1_loc / (2 * sl1)) * _sl1 * _sl2;
// copy lowpass part of current line to temporary memory
unpackdouble (& img [c1_glob], sl2, _sl1, c1_loc % (2 * sl1), templo);
// copy highpass part of current line to temporary memory
unpackdouble (& img [c1_glob + sl2 * _sl1], sl2, _sl1, c1_loc % (2 * sl1), temphi);
// perform lowpass reconstruction
uplo (templo, sl2, tmp);
// perform highpass reconstruction
uphi (temphi, sl2, temptop);
// fusion of reconstruction parts
adddouble (tmp, temptop, sl2 * 2, tmp);
// write back reconstructed line
packdouble (tmp, sl2 * 2, _sl1, c1_loc % (2 * sl1), & img [c1_glob]);
} // loop over lines along second dimension of result image
// update stride
stride = 5 * sl1 * t_num;
tmp = & _temp [stride];
templo = & _temp [ sl1 + stride];
temphi = & _temp [3 * sl1 + stride];
# pragma omp for schedule (OMP_SCHEDULE)
// loop over lines along first dimension ('columns') of result image
for (int c2_loc = 0; c2_loc < 2 * sl2 * 2 * sl3; c2_loc++)
{
int c2_glob = (c2_loc / (2 * sl2)) * _sl2 * _sl1 + (c2_loc % (2 * sl2)) * _sl1;
// assign address of current line's lowpass part
wcplo = & img [c2_glob];
// assign address of current line's highpass part
wcphi = & img [c2_glob + sl1];
// copy lowpass part to temporary memory
copydouble (wcplo, tmp, sl1);
// perform lowpass reconstruction
uplo (wcplo, sl1, templo);
// perform highpass reconstruction
uphi (wcphi, sl1, temphi);
// combine reconstructed parts and write back to current line
adddouble (templo, temphi, sl1 * 2, wcplo);
} // loop over lines along first dimension ('columns') of result image
// update current row / column size
sl2 *= 2;
sl1 *= 2;
sl3 *= 2;
} // loop over levels of backwards DWT
} // omp parallel
}
void
_fp_pack(
fp_simd_type *pfpsd, /* Pointer to simulator data */
unpacked *pu, /* unpacked operand */
uint_t n, /* register where datum starts */
enum fp_op_type type) /* type of datum */
{
switch (type) {
case fp_op_int32:
{
int32_t x;
packint32(pfpsd, pu, &x);
if (!(pfpsd->fp_current_exceptions & pfpsd->fp_fsrtem))
pfpsd->fp_current_write_freg(&x, n, pfpsd);
break;
}
case fp_op_int64:
{
int64_t x;
packint64(pfpsd, pu, &x);
if ((n & 0x1) == 1) /* fix register encoding */
n = (n & 0x1e) | 0x20;
if (!(pfpsd->fp_current_exceptions & pfpsd->fp_fsrtem))
pfpsd->fp_current_write_dreg(&x, DOUBLE(n), pfpsd);
break;
}
case fp_op_single:
{
single_type x;
packsingle(pfpsd, pu, &x);
if (!(pfpsd->fp_current_exceptions & pfpsd->fp_fsrtem))
pfpsd->fp_current_write_freg(&x, n, pfpsd);
break;
}
case fp_op_double:
{
union {
double_type x[2];
uint32_t y[2];
uint64_t ll;
} db;
packdouble(pfpsd, pu, &db.x[0], &db.y[1]);
if (!(pfpsd->fp_current_exceptions &
pfpsd->fp_fsrtem)) {
if ((n & 0x1) == 1) /* fix register encoding */
n = (n & 0x1e) | 0x20;
pfpsd->fp_current_write_dreg(&db.ll, DOUBLE(n),
pfpsd);
}
break;
}
case fp_op_extended:
{
union {
extended_type x;
uint32_t y[4];
uint64_t ll[2];
} ex;
unpacked U;
int k;
switch (pfpsd->fp_precision) {
/*
* Implement extended
* rounding precision
* mode.
*/
case fp_single:
{
single_type tx;
packsingle(pfpsd, pu, &tx);
pu = &U;
unpacksingle(pfpsd, pu, tx);
break;
}
case fp_double:
{
double_type tx;
uint_t ty;
packdouble(pfpsd, pu, &tx, &ty);
pu = &U;
unpackdouble(pfpsd, pu, tx, ty);
break;
}
case fp_precision_3: /* rounded to 64 bits */
{
k = pu->exponent + EXTENDED_BIAS;
if (k >= 0) k = 113-64;
else k = 113-64-k;
fpu_rightshift(pu, 113-64);
round(pfpsd, pu);
pu->sticky = pu->rounded = 0;
pu->exponent += k;
fpu_normalize(pu);
break;
}
}
packextended(pfpsd, pu, &ex.x, &ex.y[1],
&ex.y[2], &ex.y[3]);
if (!(pfpsd->fp_current_exceptions &
pfpsd->fp_fsrtem)) {
if ((n & 0x1) == 1) /* fix register encoding */
n = (n & 0x1e) | 0x20;
pfpsd->fp_current_write_dreg(&ex.ll[0],
QUAD_E(n), pfpsd);
pfpsd->fp_current_write_dreg(&ex.ll[1],
QUAD_F(n), pfpsd);
}
break;
}
}
}
/**
* @brief Perform forward DWT (periodic boundaries) on 3D data.
*
* @param sig Signal to be transformed.
* @param res Decomposed signal.
*/
void
dpwt3 (const Matrix <T> & sig, Matrix <T> & res)
{
// assign signal to result matrix
res = sig;
# pragma omp parallel default (shared), num_threads (_num_threads)
{
T * wcplo, * wcphi, * templo, * temphi, * tmp;
size_t stride;
int sl1 = _sl1,
sl2 = _sl2,
sl3 = _sl3;
const int t_num = omp_get_thread_num ();
// loop over levels of DWT
for (int j = (_max_level-1); j >= _min_level; --j)
{
// update stride
stride = sl1 * t_num;
// update thread's temporary memory address
tmp = & _temp [stride];
# pragma omp for schedule (OMP_SCHEDULE)
// loop over lines along first dimension ('columns') of image
for (int c2_loc = 0; c2_loc < sl2 * sl3; c2_loc++)
{
int c2_glob = (c2_loc / sl2) * _sl1 * _sl2 + (c2_loc % sl2) * _sl1;
// access to lowpass part of DWT
wcplo = & res [c2_glob /** _sl1*/];
// access to highpass part of DWT
wcphi = & res [c2_glob /** _sl1*/ + sl1 / 2];
// copy part of image to _temp memory
copydouble (wcplo, tmp, sl1);
// apply low pass filter on current line and write to result matrix
downlo (tmp, sl1, wcplo);
// apply high pass filter on current line and write to result matrix
downhi (tmp, sl1, wcphi);
} // loop over lines along first dimension
// update stride
stride = 2 * sl2 * t_num;
// update thread's temporary memory address
tmp = & _temp [stride];
templo = & _temp [ sl2 + stride];
temphi = & _temp [1.5 * sl2 + stride];
# pragma omp for schedule (OMP_SCHEDULE)
// loop over lines along second dimension ('rows') of image
for (int c1_loc = 0; c1_loc < sl1 * sl3; c1_loc++)
{
int c1_glob = (c1_loc / sl1) * _sl1 * _sl2;
// copy c1-th line of image to temp_mem
unpackdouble (& res [c1_glob], sl2, _sl1, c1_loc % sl1, tmp);
// apply low pass filter on current line and write to _temp mem
downlo (tmp, sl2, templo);
// apply high pass filter on current line and write to _temp mem
downhi (tmp, sl2, temphi);
// write temp lowpass result to result matrix
packdouble (templo, sl2 / 2, _sl1, c1_loc % sl1, & res [c1_glob]);
// write temp highpass result to result matrix
packdouble (temphi, sl2 / 2, _sl1, c1_loc % sl1, & res [c1_glob + sl2 / 2 * _sl1]);
} // loop over lines along second dimension
// update stride
stride = 2 * sl3 * t_num;
// update thread's temporary memory address
tmp = & _temp [stride];
templo = & _temp [ sl3 + stride];
temphi = & _temp [1.5 * sl3 + stride];
# pragma omp for schedule (OMP_SCHEDULE)
// loop over lines along third dimension ('third') of image
for (int c1_loc = 0; c1_loc < sl1 * sl2; c1_loc++)
{
int c1_glob = (c1_loc % sl1) + (c1_loc / sl1) * _sl1;
// copy c2-th line of image to temp_mem
unpackdouble (& res [c1_glob], sl3, _ld12, 0, tmp);
// apply low pass filter on current line and write to _temp mem
downlo (tmp, sl3, templo);
// apply high pass filter on current line and write to _temp mem
downhi (tmp, sl3, temphi);
// write temp lowpass result to result matrix
packdouble (templo, sl3 / 2, _ld12, 0, & res [c1_glob]);
// write temp highpass result to result matrix
packdouble (temphi, sl3 / 2, _ld12, 0, & res [c1_glob + sl3 / 2 * _ld12]);
} // loop over lines along third dimension
// reduce dimensions for next level
sl1 /= 2;
sl2 /= 2;
sl3 /= 2;
} // loop over levels of DWT
} // omp parallel
}
