float* create_coo(const char* name, unsigned int D, const long dims[D])
{
int ofd;
if (-1 == (ofd = open(name, O_RDWR|O_CREAT, S_IRUSR|S_IWUSR)))
io_error("Creating coo file %s", name);
if (-1 == write_coo(ofd, D, dims))
io_error("Creating coo file %s", name);
long T = md_calc_size(D, dims) * sizeof(float);
void* addr;
if (NULL == (addr = create_data(ofd, 4096, T)))
io_error("Creating coo file %s", name);
if (-1 == close(ofd))
io_error("Creating coo file %s", name);
return (float*)addr;
}
const struct operator_p_s* prox_normaleq_create(const struct linop_s* op, const complex float* y)
{
PTR_ALLOC(struct prox_normaleq_data, pdata);
SET_TYPEID(prox_normaleq_data, pdata);
PTR_ALLOC(struct iter_conjgrad_conf, cgconf);
*cgconf = iter_conjgrad_defaults;
cgconf->maxiter = 10;
cgconf->l2lambda = 0;
pdata->cgconf = PTR_PASS(cgconf);
pdata->op = op;
pdata->size = 2 * md_calc_size(linop_domain(op)->N, linop_domain(op)->dims);
pdata->adj = md_alloc_sameplace(1, &(pdata->size), FL_SIZE, y);
linop_adjoint_iter((struct linop_s*)op, pdata->adj, (const float*)y);
return operator_p_create(linop_domain(op)->N, linop_domain(op)->dims,
linop_domain(op)->N, linop_domain(op)->dims,
CAST_UP(PTR_PASS(pdata)), prox_normaleq_apply, prox_normaleq_del);
}
extern gboolean refresh_callback(GtkWidget *widget, gpointer data)
{
struct view_s* v = data;
v->invalid = true;
long size = md_calc_size(DIMS, v->dims);
double max = 0.;
for (long j = 0; j < size; j++)
if (max < cabsf(v->data[j]))
max = cabsf(v->data[j]);
if (0. == max)
max = 1.;
v->max = max;
update_view(v);
return FALSE;
}
void noir_adj(struct noir_data* data, complex float* dst, const complex float* src)
{
long split = md_calc_size(DIMS, data->imgs_dims);
md_zmulc2(DIMS, data->sign_dims, data->sign_strs, data->tmp, data->data_strs, src, data->ptrn_strs, data->pattern);
ifft(DIMS, data->sign_dims, FFT_FLAGS, data->tmp, data->tmp);
// we should move it to the end, but fft scaling is applied so this would be need to moved into data->xn or weights maybe?
md_zmulc2(DIMS, data->sign_dims, data->sign_strs, data->tmp, data->sign_strs, data->tmp, data->mask_strs, data->mask);
md_clear(DIMS, data->coil_dims, dst + split, CFL_SIZE);
md_zfmacc2(DIMS, data->sign_dims, data->coil_strs, dst + split, data->sign_strs, data->tmp, data->imgs_strs, data->xn);
noir_back_coils(data, dst + split, dst + split);
md_clear(DIMS, data->imgs_dims, dst, CFL_SIZE);
md_zfmacc2(DIMS, data->sign_dims, data->imgs_strs, dst, data->sign_strs, data->tmp, data->coil_strs, data->sens);
if (data->rvc)
md_zreal(DIMS, data->imgs_dims, dst, dst);
}
static void linop_matrix_apply_normal(const linop_data_t* _data, complex float* dst, const complex float* src)
{
const struct operator_matrix_s* data = CAST_DOWN(operator_matrix_s, _data);
unsigned int N = data->mat_iovec->N;
// FIXME check all the cases where computation can be done with blas
//debug_printf(DP_DEBUG1, "compute normal\n");
if (cgemm_forward_standard(data)) {
long max_dims_gram[N];
md_copy_dims(N, max_dims_gram, data->domain_iovec->dims);
max_dims_gram[data->T_dim] = data->K;
long tmp_dims[N];
long tmp_str[N];
md_copy_dims(N, tmp_dims, max_dims_gram);
tmp_dims[data->K_dim] = 1;
md_calc_strides(N, tmp_str, tmp_dims, CFL_SIZE);
complex float* tmp = md_alloc_sameplace(N, data->domain_iovec->dims, CFL_SIZE, dst);
md_clear(N, data->domain_iovec->dims, tmp, CFL_SIZE);
md_zfmac2(N, max_dims_gram, tmp_str, tmp, data->domain_iovec->strs, src, data->mat_gram_iovec->strs, data->mat_gram);
md_transpose(N, data->T_dim, data->K_dim, data->domain_iovec->dims, dst, tmp_dims, tmp, CFL_SIZE);
md_free(tmp);
} else {
long L = md_calc_size(data->T_dim, data->domain_iovec->dims);
blas_cgemm('N', 'T', L, data->K, data->K, 1.,
L, (const complex float (*)[])src,
data->K, (const complex float (*)[])data->mat_gram,
0., L, (complex float (*)[])dst);
}
}
float nucnorm_blockproc( const void* _data, const long blkdims[DIMS], complex float* dst, const complex float* src )
{
UNUSED(dst);
const struct svthresh_blockproc_data* data = (const struct svthresh_blockproc_data*) _data;
long M = 1;
long N = md_calc_size( DIMS, blkdims );
for ( unsigned int i = 0; i < DIMS; i++ )
{
if (MD_IS_SET(data->mflags, i))
{
M *= blkdims[i];
N /= blkdims[i];
}
}
float G = sqrtf(M) + sqrtf(N);
return G * nuclearnorm(M, N, src);
}
static void fftmod2_r(unsigned int N, const long dims[N], unsigned long flags, const long ostrs[N], complex float* dst, const long istrs[N], const complex float* src, bool inv, double phase)
{
if (0 == flags) {
md_zsmul2(N, dims, ostrs, dst, istrs, src, cexp(M_PI * 2.i * (inv ? -phase : phase)));
return;
}
/* this will also currently be slow on the GPU because we do not
* support strides there on the lowest level */
unsigned int i = N - 1;
while (!MD_IS_SET(flags, i))
i--;
#if 1
// If there is only one dimensions left and it is the innermost
// which is contiguous optimize using md_zfftmod2
if ((0u == MD_CLEAR(flags, i)) && (1 == md_calc_size(i, dims))
&& (CFL_SIZE == ostrs[i]) && (CFL_SIZE == istrs[i])) {
md_zfftmod2(N - i, dims + i, ostrs + i, dst, istrs + i, src, inv, phase);
return;
}
#endif
long tdims[N];
md_select_dims(N, ~MD_BIT(i), tdims, dims);
#pragma omp parallel for
for (int j = 0; j < dims[i]; j++)
fftmod2_r(N, tdims, MD_CLEAR(flags, i),
ostrs, (void*)dst + j * ostrs[i], istrs, (void*)src + j * istrs[i],
inv, phase + fftmod_phase(dims[i], j));
}
void direct_calib(const long dims[5], complex float* sens, const long caldims[5], const complex float* data)
{
complex float* tmp = md_alloc(5, caldims, CFL_SIZE);
assert(1 == caldims[4]);
assert(1 == dims[4]);
md_copy(5, caldims, tmp, data, CFL_SIZE);
// apply Kaiser-Bessel Window beta=4
for (int z = 0; z < caldims[2]; z++)
for (int y = 0; y < caldims[1]; y++)
for (int x = 0; x < caldims[0]; x++)
for (int c = 0; c < caldims[3]; c++)
tmp[((c * caldims[2] + z) * caldims[1] + y) * caldims[0] + x]
*= kaiser(4., caldims[2], z)
* kaiser(4., caldims[1], y)
* kaiser(4., caldims[0], x);
md_resize_center(5, dims, sens, caldims, tmp, CFL_SIZE);
ifftc(5, dims, 7, sens, sens);
long dims1[5];
md_select_dims(5, ~MD_BIT(COIL_DIM), dims1, dims);
complex float* img = md_alloc(5, dims1, CFL_SIZE);
md_zrss(5, dims, COIL_FLAG, img, sens);
#if 1
long T = md_calc_size(5, dims1);
for (int i = 0; i < T; i++)
for (int j = 0; j < dims[COIL_DIM]; j++)
sens[j * T + i] *= (cabs(img[i]) == 0.) ? 0. : (1. / cabs(img[i]));
#endif
md_free(img);
}
void iwt(unsigned int N, unsigned int flags, const long shifts[N], const long dims[N], const long ostr[N], complex float* out, const complex float* in, const long minsize[N], const long flen, const float filter[2][2][flen])
{
if (0 == flags) {
if (out != in)
md_copy2(N, dims, ostr, out, ostr, in, CFL_SIZE);
return;
}
unsigned long coeffs = wavelet_coeffs(N, flags, dims, minsize, flen);
long wdims[2 * N];
wavelet_dims(N, flags, wdims, dims, flen);
long istr[2 * N];
md_calc_strides(2 * N, istr, wdims, CFL_SIZE);
long offset = coeffs - md_calc_size(2 * N, wdims);
debug_printf(DP_DEBUG4, "%d %ld %ld\n", flags, coeffs, offset);
complex float* tmp = md_alloc_sameplace(2 * N, wdims, CFL_SIZE, out);
md_copy(2 * N, wdims, tmp, in + offset, CFL_SIZE);
long shifts0[N];
for (unsigned int i = 0; i < N; i++)
shifts0[i] = 0;
// fix me we need temp storage
iwt(N, wavelet_filter_flags(N, flags, wdims, minsize), shifts0, wdims, istr, tmp, in, minsize, flen, filter);
iwtN(N, flags, shifts, dims, ostr, out, istr, tmp, flen, filter);
md_free(tmp);
}
int main(int argc, char *argv[])
{
MATFile *mat;
const char* name = NULL;
const mwSize* dims;
if (argc != 2) {
fprintf(stderr, "Usage: %s file.mat\n", argv[0]);
exit(1);
}
if (NULL == (mat = matOpen(argv[1], "r")))
exit(1);
mxArray* ar;
while (NULL != (ar = matGetNextVariable(mat, &name))) {
int ndim = (int)mxGetNumberOfDimensions(ar);
dims = mxGetDimensions(ar);
bool cmp = mxIsComplex(ar);
bool dbl = mxIsDouble(ar);
printf("%s: [ ", name);
if ((!cmp) || (!dbl)) {
printf("not complex double\n");
mxDestroyArray(ar);
continue;
}
long ldims[ndim];
for (int i = 0; i < ndim; i++)
ldims[i] = dims[i];
for (int i = 0; i < ndim; i++)
printf("%ld ", ldims[i]);
char outname[256];
snprintf(outname, 256, "%s_%s", strtok(argv[1], "."), name);
complex float* buf = create_cfl(outname, ndim, ldims);
double* re = mxGetPr(ar);
double* im = mxGetPi(ar);
size_t size = md_calc_size(ndim, ldims);
for (unsigned long i = 0; i < size; i++)
buf[i] = re[i] + 1.i * im[i];
printf("] -> %s\n", outname);
unmap_cfl(ndim, ldims, buf);
mxDestroyArray(ar);
}
matClose(mat);
}
void overlapandsave2HB(const struct vec_ops* ops, int N, unsigned int flags, const long blk[N], const long dims1[N], complex float* dst, const long odims[N], const complex float* src1, const long dims2[N], const complex float* src2, const long mdims[N], const complex float* msk)
{
long dims1B[N];
long tdims[2 * N];
long nodims[2 * N];
long ndims2[2 * N];
long nmdims[2 * N];
int e = N;
for (int i = 0; i < N; i++) {
if (MD_IS_SET(flags, i)) {
assert(1 == dims2[i] % 2);
assert(0 == blk[i] % 2);
assert(0 == dims1[i] % 2);
assert(0 == odims[i] % blk[i]);
assert(0 == dims1[i] % blk[i]);
assert(dims1[i] == odims[i]);
assert(dims2[i] <= blk[i]);
assert(dims1[i] >= dims2[i]);
assert((1 == mdims[i]) || (mdims[i] == dims1[i]));
// blocked output
nodims[e] = odims[i] / blk[i];
nodims[i] = blk[i];
// expanded temporary storage
tdims[e] = dims1[i] / blk[i];
tdims[i] = blk[i] + dims2[i] - 1;
// blocked input
// ---|---,---,---|---
// + +++ +
// + +++ +
if (1 == mdims[i]) {
nmdims[2 * i + 1] = 1;
nmdims[2 * i + 1] = 1;
} else {
nmdims[2 * i + 1] = mdims[i] / blk[i];
nmdims[2 * i + 0] = blk[i];
}
// resized input
// minimal padding
dims1B[i] = dims1[i] + (dims2[i] - 1);
// kernel
ndims2[e] = 1;
ndims2[i] = dims2[i];
e++;
} else {
nodims[i] = odims[i];
tdims[i] = dims1[i];
nmdims[2 * i + 1] = 1;
nmdims[2 * i + 0] = mdims[i];
dims1B[i] = dims1[i];
ndims2[i] = dims2[i];
}
}
int NE = e;
// long S = md_calc_size(N, dims1B, 1);
long str1[NE];
long str1B[N];
md_calc_strides(N, str1B, dims1B, sizeof(complex float));
e = N;
for (int i = 0; i < N; i++) {
str1[i] = str1B[i];
if (MD_IS_SET(flags, i))
str1[e++] = str1B[i] * blk[i];
}
assert(NE == e);
long str2[NE];
md_calc_strides(NE, str2, tdims, sizeof(complex float));
long ostr[NE];
long mstr[NE];
long mstrB[2 * N];
md_calc_strides(NE, ostr, nodims, sizeof(complex float));
md_calc_strides(2 * N, mstrB, nmdims, sizeof(complex float));
e = N;
for (int i = 0; i < N; i++) {
mstr[i] = mstrB[2 * i + 0];
if (MD_IS_SET(flags, i))
mstr[e++] = mstrB[2 * i + 1];
}
assert(NE == e);
// we can loop here
assert(NE == N + 3);
assert(1 == ndims2[N + 0]);
assert(1 == ndims2[N + 1]);
assert(1 == ndims2[N + 2]);
assert(tdims[N + 0] == nodims[N + 0]);
assert(tdims[N + 1] == nodims[N + 1]);
assert(tdims[N + 2] == nodims[N + 2]);
long R = md_calc_size(N, nodims);
long T = md_calc_size(N, tdims);
//complex float* src1C = xmalloc(S * sizeof(complex float));
complex float* src1C = dst;
md_clear(N, dims1B, src1C, CFL_SIZE); // must be done here
#pragma omp parallel for collapse(3)
for (int k = 0; k < nodims[N + 2]; k++) {
for (int j = 0; j < nodims[N + 1]; j++) {
for (int i = 0; i < nodims[N + 0]; i++) {
complex float* tmp = (complex float*)ops->allocate(2 * T);
complex float* tmpX = (complex float*)ops->allocate(2 * R);
long off1 = str1[N + 0] * i + str1[N + 1] * j + str1[N + 2] * k;
long off2 = mstr[N + 0] * i + mstr[N + 1] * j + mstr[N + 2] * k;
long off3 = ostr[N + 0] * i + ostr[N + 1] * j + ostr[N + 2] * k;
md_zmul2(N, nodims, ostr, tmpX, ostr, ((const void*)src1) + off3, mstr, ((const void*)msk) + off2);
convH(N, flags, CONV_VALID, CONV_SYMMETRIC, tdims, tmp, nodims, tmpX, ndims2, src2);
#pragma omp critical
md_zadd2(N, tdims, str1, ((void*)src1C) + off1, str1, ((void*)src1C) + off1, str2, tmp);
ops->del((void*)tmpX);
ops->del((void*)tmp);
}}}
}
/* calculate point-wise maps
*
*/
void eigenmaps(const long out_dims[DIMS], complex float* optr, complex float* eptr, const complex float* imgcov2, const long msk_dims[3], const bool* msk, bool orthiter, bool ecal_usegpu)
{
#ifdef USE_CUDA
if (ecal_usegpu) {
//FIXME cuda version should be able to return sensitivities for a subset of image-space points
assert(!msk);
eigenmapscu(out_dims, optr, eptr, imgcov2);
return;
}
#else
assert(!ecal_usegpu);
#endif
long channels = out_dims[3];
long maps = out_dims[4];
assert(DIMS >= 5);
assert(1 == md_calc_size(DIMS - 5, out_dims + 5));
assert(maps <= channels);
long xx = out_dims[0];
long yy = out_dims[1];
long zz = out_dims[2];
float scale = 1.; // for some reason, not
if (msk_dims) {
assert(msk_dims[0] == xx);
assert(msk_dims[1] == yy);
assert(msk_dims[2] == zz);
}
md_clear(5, out_dims, optr, CFL_SIZE);
#pragma omp parallel for collapse(3)
for (long k = 0; k < zz; k++) {
for (long j = 0; j < yy; j++) {
for (long i = 0; i < xx; i++) {
if (!msk || msk[i + xx * (j + yy * k)]) {
float val[channels];
complex float cov[channels][channels];
complex float tmp[channels * (channels + 1) / 2];
for (long l = 0; l < channels * (channels + 1) / 2; l++)
tmp[l] = imgcov2[((l * zz + k) * yy + j) * xx + i] / scale;
unpack_tri_matrix(channels, cov, tmp);
if (orthiter)
eigen_herm3(maps, channels, val, cov);
else
lapack_eig(channels, val, cov);
for (long u = 0; u < maps; u++) {
long ru = (orthiter ? maps : channels) - 1 - u;
for (long v = 0; v < channels; v++)
optr[((((u * channels + v) * zz + k) * yy + j) * xx + i)] = cov[ru][v];
if (NULL != eptr)
eptr[((u * zz + k) * yy + j) * xx + i] = val[ru];
}
}
}
}
}
}
int main_pics(int argc, char* argv[])
{
// Initialize default parameters
struct sense_conf conf = sense_defaults;
bool use_gpu = false;
bool randshift = true;
unsigned int maxiter = 30;
float step = -1.;
// Start time count
double start_time = timestamp();
// Read input options
struct nufft_conf_s nuconf = nufft_conf_defaults;
nuconf.toeplitz = false;
float restrict_fov = -1.;
const char* pat_file = NULL;
const char* traj_file = NULL;
bool scale_im = false;
bool eigen = false;
float scaling = 0.;
unsigned int llr_blk = 8;
const char* image_truth_file = NULL;
bool im_truth = false;
const char* image_start_file = NULL;
bool warm_start = false;
bool hogwild = false;
bool fast = false;
float admm_rho = iter_admm_defaults.rho;
unsigned int admm_maxitercg = iter_admm_defaults.maxitercg;
struct opt_reg_s ropts;
ropts.r = 0;
ropts.algo = CG;
ropts.lambda = -1.;
const struct opt_s opts[] = {
{ 'l', true, opt_reg, &ropts, "1/-l2\t\ttoggle l1-wavelet or l2 regularization." },
OPT_FLOAT('r', &ropts.lambda, "lambda", "regularization parameter"),
{ 'R', true, opt_reg, &ropts, " <T>:A:B:C\tgeneralized regularization options (-Rh for help)" },
OPT_SET('c', &conf.rvc, "real-value constraint"),
OPT_FLOAT('s', &step, "step", "iteration stepsize"),
OPT_UINT('i', &maxiter, "iter", "max. number of iterations"),
OPT_STRING('t', &traj_file, "file", "k-space trajectory"),
OPT_CLEAR('n', &randshift, "disable random wavelet cycle spinning"),
OPT_SET('g', &use_gpu, "use GPU"),
OPT_STRING('p', &pat_file, "file", "pattern or weights"),
OPT_SELECT('I', enum algo_t, &ropts.algo, IST, "(select IST)"),
OPT_UINT('b', &llr_blk, "blk", "Lowrank block size"),
OPT_SET('e', &eigen, "Scale stepsize based on max. eigenvalue"),
OPT_SET('H', &hogwild, "(hogwild)"),
OPT_SET('F', &fast, "(fast)"),
OPT_STRING('T', &image_truth_file, "file", "(truth file)"),
OPT_STRING('W', &image_start_file, "<img>", "Warm start with <img>"),
OPT_INT('d', &debug_level, "level", "Debug level"),
OPT_INT('O', &conf.rwiter, "rwiter", "(reweighting)"),
OPT_FLOAT('o', &conf.gamma, "gamma", "(reweighting)"),
OPT_FLOAT('u', &admm_rho, "rho", "ADMM rho"),
OPT_UINT('C', &admm_maxitercg, "iter", "ADMM max. CG iterations"),
OPT_FLOAT('q', &conf.cclambda, "cclambda", "(cclambda)"),
OPT_FLOAT('f', &restrict_fov, "rfov", "restrict FOV"),
OPT_SELECT('m', enum algo_t, &ropts.algo, ADMM, "Select ADMM"),
OPT_FLOAT('w', &scaling, "val", "scaling"),
OPT_SET('S', &scale_im, "Re-scale the image after reconstruction"),
};
cmdline(&argc, argv, 3, 3, usage_str, help_str, ARRAY_SIZE(opts), opts);
if (NULL != image_truth_file)
im_truth = true;
if (NULL != image_start_file)
warm_start = true;
long max_dims[DIMS];
long map_dims[DIMS];
long pat_dims[DIMS];
long img_dims[DIMS];
long coilim_dims[DIMS];
long ksp_dims[DIMS];
long traj_dims[DIMS];
// load kspace and maps and get dimensions
complex float* kspace = load_cfl(argv[1], DIMS, ksp_dims);
complex float* maps = load_cfl(argv[2], DIMS, map_dims);
complex float* traj = NULL;
if (NULL != traj_file)
traj = load_cfl(traj_file, DIMS, traj_dims);
md_copy_dims(DIMS, max_dims, ksp_dims);
md_copy_dims(5, max_dims, map_dims);
md_select_dims(DIMS, ~COIL_FLAG, img_dims, max_dims);
md_select_dims(DIMS, ~MAPS_FLAG, coilim_dims, max_dims);
if (!md_check_compat(DIMS, ~(MD_BIT(MAPS_DIM)|FFT_FLAGS), img_dims, map_dims))
error("Dimensions of image and sensitivities do not match!\n");
assert(1 == ksp_dims[MAPS_DIM]);
(use_gpu ? num_init_gpu : num_init)();
// print options
if (use_gpu)
debug_printf(DP_INFO, "GPU reconstruction\n");
if (map_dims[MAPS_DIM] > 1)
debug_printf(DP_INFO, "%ld maps.\nESPIRiT reconstruction.\n", map_dims[MAPS_DIM]);
if (hogwild)
debug_printf(DP_INFO, "Hogwild stepsize\n");
if (im_truth)
debug_printf(DP_INFO, "Compare to truth\n");
// initialize sampling pattern
complex float* pattern = NULL;
if (NULL != pat_file) {
pattern = load_cfl(pat_file, DIMS, pat_dims);
assert(md_check_compat(DIMS, COIL_FLAG, ksp_dims, pat_dims));
} else {
md_select_dims(DIMS, ~COIL_FLAG, pat_dims, ksp_dims);
pattern = md_alloc(DIMS, pat_dims, CFL_SIZE);
estimate_pattern(DIMS, ksp_dims, COIL_DIM, pattern, kspace);
}
if ((NULL != traj_file) && (NULL == pat_file)) {
md_free(pattern);
pattern = NULL;
nuconf.toeplitz = true;
} else {
// print some statistics
long T = md_calc_size(DIMS, pat_dims);
long samples = (long)pow(md_znorm(DIMS, pat_dims, pattern), 2.);
debug_printf(DP_INFO, "Size: %ld Samples: %ld Acc: %.2f\n", T, samples, (float)T / (float)samples);
}
if (NULL == traj_file) {
fftmod(DIMS, ksp_dims, FFT_FLAGS, kspace, kspace);
fftmod(DIMS, map_dims, FFT_FLAGS, maps, maps);
}
// apply fov mask to sensitivities
if (-1. != restrict_fov) {
float restrict_dims[DIMS] = { [0 ... DIMS - 1] = 1. };
restrict_dims[0] = restrict_fov;
restrict_dims[1] = restrict_fov;
restrict_dims[2] = restrict_fov;
apply_mask(DIMS, map_dims, maps, restrict_dims);
}
void calib2(const struct ecalib_conf* conf, const long out_dims[DIMS], complex float* out_data, complex float* eptr, unsigned int SN, float svals[SN], const long calreg_dims[DIMS], const complex float* data, const long msk_dims[3], const bool* msk)
{
long channels = calreg_dims[3];
long maps = out_dims[4];
assert(calreg_dims[3] == out_dims[3]);
assert(maps <= channels);
assert(1 == md_calc_size(DIMS - 5, out_dims + 5));
assert(1 == md_calc_size(DIMS - 5, calreg_dims + 5));
complex float rot[channels][channels];
if (conf->rotphase) {
long scc_dims[DIMS] = MD_INIT_ARRAY(DIMS, 1);
scc_dims[COIL_DIM] = channels;
scc_dims[MAPS_DIM] = channels;
scc(scc_dims, &rot[0][0], calreg_dims, data);
} else {
for (unsigned int i = 0; i < channels; i++)
for (unsigned int j = 0; j < channels; j++)
rot[i][j] = (i == j) ? 1. : 0.;
}
long cov_dims[4];
calone_dims(conf, cov_dims, channels);
complex float* imgcov = md_alloc(4, cov_dims, CFL_SIZE);
calone(conf, cov_dims, imgcov, SN, svals, calreg_dims, data);
caltwo(conf, out_dims, out_data, eptr, cov_dims, imgcov, msk_dims, msk);
/* Intensity and phase normalization similar as proposed
* for adaptive combine (Walsh's method) in
* Griswold et al., ISMRM 10:2410 (2002)
*/
if (conf->intensity) {
debug_printf(DP_DEBUG1, "Normalize...\n");
/* I think the reason this works is because inhomogeneity usually
* comes from only a few coil elements which are close. The l1-norm
* is more resilient against such outliers. -- Martin
*/
normalizel1(DIMS, COIL_FLAG, out_dims, out_data);
md_zsmul(DIMS, out_dims, out_data, out_data, sqrtf((float)channels));
}
debug_printf(DP_DEBUG1, "Crop maps... (%.2f)\n", conf->crop);
crop_sens(out_dims, out_data, conf->softcrop, conf->crop, eptr);
debug_printf(DP_DEBUG1, "Fix phase...\n");
// rotate the the phase with respect to the first principle component
fixphase2(DIMS, out_dims, COIL_DIM, rot[0], out_data, out_data);
md_free(imgcov);
}
void iter2_niht(iter_conf* _conf,
const struct operator_s* normaleq_op,
unsigned int D,
const struct operator_p_s* prox_ops[D],
const struct linop_s* ops[D],
const float* biases[D],
const struct operator_p_s* xupdate_op,
long size, float* image, const float* image_adj,
struct iter_monitor_s* monitor)
{
UNUSED(xupdate_op);
UNUSED(biases);
assert(D == 1);
auto conf = CAST_DOWN(iter_niht_conf, _conf);
struct niht_conf_s niht_conf = {
.maxiter = conf->maxiter,
.N = size,
.trans = 0,
.do_warmstart = conf->do_warmstart,
};
struct niht_transop trans;
if (NULL != ops) {
trans.forward = OPERATOR2ITOP(ops[0]->forward);
trans.adjoint = OPERATOR2ITOP(ops[0]->adjoint);
trans.N = 2 * md_calc_size(linop_codomain(ops[0])->N, linop_codomain(ops[0])->dims);
niht_conf.trans = 1;
}
float eps = md_norm(1, MD_DIMS(size), image_adj);
if (checkeps(eps))
goto cleanup;
niht_conf.epsilon = eps * conf->tol;
niht(&niht_conf, &trans, select_vecops(image_adj), OPERATOR2ITOP(normaleq_op), OPERATOR_P2ITOP(prox_ops[0]), image, image_adj, monitor);
cleanup:
;
}
void iter2_call_iter(iter_conf* _conf,
const struct operator_s* normaleq_op,
unsigned int D,
const struct operator_p_s* prox_ops[D],
const struct linop_s* ops[D],
const float* biases[D],
const struct operator_p_s* xupdate_op,
long size, float* image, const float* image_adj,
struct iter_monitor_s* monitor)
{
assert(D <= 1);
assert(NULL == ops);
assert(NULL == biases);
UNUSED(xupdate_op);
auto it = CAST_DOWN(iter_call_s, _conf);
it->fun(it->_conf, normaleq_op, (1 == D) ? prox_ops[0] : NULL,
size, image, image_adj, monitor);
}
int main_sense(int argc, char* argv[])
{
struct sense_conf conf = sense_defaults;
double start_time = timestamp();
bool admm = false;
bool ist = false;
bool use_gpu = false;
bool l1wav = false;
bool lowrank = false;
bool randshift = true;
int maxiter = 30;
float step = 0.95;
float lambda = 0.;
float restrict_fov = -1.;
const char* pat_file = NULL;
const char* traj_file = NULL;
const char* image_truth_file = NULL;
bool im_truth = false;
bool scale_im = false;
bool hogwild = false;
bool fast = false;
float admm_rho = iter_admm_defaults.rho;
int c;
while (-1 != (c = getopt(argc, argv, "Fq:l:r:s:i:u:o:O:f:t:cT:Imghp:Sd:H"))) {
switch(c) {
case 'H':
hogwild = true;
break;
case 'F':
fast = true;
break;
case 'I':
ist = true;
break;
case 'T':
im_truth = true;
image_truth_file = strdup(optarg);
assert(NULL != image_truth_file);
break;
case 'd':
debug_level = atoi(optarg);
break;
case 'r':
lambda = atof(optarg);
break;
case 'O':
conf.rwiter = atoi(optarg);
break;
case 'o':
conf.gamma = atof(optarg);
break;
case 's':
step = atof(optarg);
break;
case 'i':
maxiter = atoi(optarg);
break;
case 'l':
if (1 == atoi(optarg)) {
l1wav = true;
lowrank = false;
} else
if (2 == atoi(optarg)) {
l1wav = false;
lowrank = false;
} else
if (3 == atoi(optarg)) {
lowrank = true;
l1wav = false;
} else {
usage(argv[0], stderr);
exit(1);
}
break;
case 'q':
conf.cclambda = atof(optarg);
break;
case 'c':
conf.rvc = true;
break;
case 'f':
restrict_fov = atof(optarg);
break;
case 'm':
admm = true;
break;
case 'u':
admm_rho = atof(optarg);
break;
case 'g':
use_gpu = true;
break;
case 'p':
pat_file = strdup(optarg);
break;
case 't':
assert(0);
break;
case 'S':
scale_im = true;
break;
case 'h':
usage(argv[0], stdout);
help();
exit(0);
default:
usage(argv[0], stderr);
exit(1);
}
}
if (argc - optind != 3) {
usage(argv[0], stderr);
exit(1);
}
long map_dims[DIMS];
long pat_dims[DIMS];
long img_dims[DIMS];
long ksp_dims[DIMS];
long max_dims[DIMS];
// load kspace and maps and get dimensions
complex float* kspace = load_cfl(argv[optind + 0], DIMS, ksp_dims);
complex float* maps = load_cfl(argv[optind + 1], DIMS, map_dims);
md_copy_dims(DIMS, max_dims, ksp_dims);
max_dims[MAPS_DIM] = map_dims[MAPS_DIM];
md_select_dims(DIMS, ~COIL_FLAG, pat_dims, ksp_dims);
md_select_dims(DIMS, ~COIL_FLAG, img_dims, max_dims);
for (int i = 0; i < 4; i++) { // sizes2[4] may be > 1
if (ksp_dims[i] != map_dims[i]) {
fprintf(stderr, "Dimensions of kspace and sensitivities do not match!\n");
exit(1);
}
}
assert(1 == ksp_dims[MAPS_DIM]);
(use_gpu ? num_init_gpu : num_init)();
// print options
if (use_gpu)
debug_printf(DP_INFO, "GPU reconstruction\n");
if (map_dims[MAPS_DIM] > 1)
debug_printf(DP_INFO, "%ld maps.\nESPIRiT reconstruction.\n", map_dims[MAPS_DIM]);
if (l1wav)
debug_printf(DP_INFO, "l1-wavelet regularization\n");
if (ist)
debug_printf(DP_INFO, "Use IST\n");
if (im_truth)
debug_printf(DP_INFO, "Compare to truth\n");
// initialize sampling pattern
complex float* pattern = NULL;
long pat_dims2[DIMS];
if (NULL != pat_file) {
pattern = load_cfl(pat_file, DIMS, pat_dims2);
// FIXME: check compatibility
} else {
pattern = md_alloc(DIMS, pat_dims, CFL_SIZE);
estimate_pattern(DIMS, ksp_dims, COIL_DIM, pattern, kspace);
}
// print some statistics
size_t T = md_calc_size(DIMS, pat_dims);
long samples = (long)pow(md_znorm(DIMS, pat_dims, pattern), 2.);
debug_printf(DP_INFO, "Size: %ld Samples: %ld Acc: %.2f\n", T, samples, (float)T / (float)samples);
fftmod(DIMS, ksp_dims, FFT_FLAGS, kspace, kspace);
fftmod(DIMS, map_dims, FFT_FLAGS, maps, maps);
// apply fov mask to sensitivities
if (-1. != restrict_fov) {
float restrict_dims[DIMS] = { [0 ... DIMS - 1] = 1. };
restrict_dims[0] = restrict_fov;
restrict_dims[1] = restrict_fov;
restrict_dims[2] = restrict_fov;
apply_mask(DIMS, map_dims, maps, restrict_dims);
}
/* O I M G
* 1 1 1 1 - not used
* 1 1 A ! - forbidden
* 1 A 1 ! - forbidden
* A 1 1 ! - forbidden
* A A 1 1 - replicated
* A 1 A 1 - output
* 1 A A A/A - input
* A A A A - batch
*/
static struct operator_matrix_s* linop_matrix_priv2(unsigned int N, const long out_dims[N], const long in_dims[N], const long matrix_dims[N], const complex float* matrix)
{
// to get assertions and cost estimate
long max_dims[N];
md_tenmul_dims(N, max_dims, out_dims, in_dims, matrix_dims);
PTR_ALLOC(struct operator_matrix_s, data);
SET_TYPEID(operator_matrix_s, data);
data->N = N;
PTR_ALLOC(long[N], out_dims1);
md_copy_dims(N, *out_dims1, out_dims);
data->out_dims = *PTR_PASS(out_dims1);
PTR_ALLOC(long[N], mat_dims1);
md_copy_dims(N, *mat_dims1, matrix_dims);
data->mat_dims = *PTR_PASS(mat_dims1);
PTR_ALLOC(long[N], in_dims1);
md_copy_dims(N, *in_dims1, in_dims);
data->in_dims = *PTR_PASS(in_dims1);
complex float* mat = md_alloc(N, matrix_dims, CFL_SIZE);
md_copy(N, matrix_dims, mat, matrix, CFL_SIZE);
data->mat = mat;
data->mat_gram = NULL;
#ifdef USE_CUDA
data->mat_gpu = NULL;
data->mat_gram_gpu = NULL;
#endif
#if 1
// pre-multiply gram matrix (if there is a cost reduction)
unsigned long out_flags = md_nontriv_dims(N, out_dims);
unsigned long in_flags = md_nontriv_dims(N, in_dims);
unsigned long del_flags = in_flags & ~out_flags;
unsigned long new_flags = out_flags & ~in_flags;
/* we double (again) for the gram matrix
*/
PTR_ALLOC(long[2 * N], mat_dims2);
PTR_ALLOC(long[2 * N], in_dims2);
PTR_ALLOC(long[2 * N], gmt_dims2);
PTR_ALLOC(long[2 * N], gin_dims2);
PTR_ALLOC(long[2 * N], grm_dims2);
PTR_ALLOC(long[2 * N], gout_dims2);
shadow_dims(N, *gmt_dims2, matrix_dims);
shadow_dims(N, *mat_dims2, matrix_dims);
shadow_dims(N, *in_dims2, in_dims);
shadow_dims(N, *gout_dims2, in_dims);
shadow_dims(N, *gin_dims2, in_dims);
shadow_dims(N, *grm_dims2, matrix_dims);
/* move removed input dims into shadow position
* for the gram matrix can have an output there
*/
for (unsigned int i = 0; i < N; i++) {
if (MD_IS_SET(del_flags, i)) {
assert((*mat_dims2)[2 * i + 0] == (*in_dims2)[2 * i + 0]);
(*mat_dims2)[2 * i + 1] = (*mat_dims2)[2 * i + 0];
(*mat_dims2)[2 * i + 0] = 1;
(*in_dims2)[2 * i + 1] = (*gin_dims2)[2 * i + 0];
(*in_dims2)[2 * i + 0] = 1;
}
}
for (unsigned int i = 0; i < N; i++) {
if (MD_IS_SET(new_flags, i)) {
(*grm_dims2)[2 * i + 0] = 1;
(*grm_dims2)[2 * i + 1] = 1;
}
if (MD_IS_SET(del_flags, i)) {
(*gout_dims2)[2 * i + 1] = (*gin_dims2)[2 * i + 0];
(*gout_dims2)[2 * i + 0] = 1;
(*grm_dims2)[2 * i + 0] = in_dims[i];
(*grm_dims2)[2 * i + 1] = in_dims[i];
}
}
long gmx_dims[2 * N];
md_tenmul_dims(2 * N, gmx_dims, *gout_dims2, *gin_dims2, *grm_dims2);
long mult_mat = md_calc_size(N, max_dims);
long mult_gram = md_calc_size(2 * N, gmx_dims);
if (mult_gram < 2 * mult_mat) { // FIXME: rethink
debug_printf(DP_DEBUG2, "Gram matrix: 2x %ld vs %ld\n", mult_mat, mult_gram);
complex float* mat_gram = md_alloc(2 * N, *grm_dims2, CFL_SIZE);
md_ztenmulc(2 * N, *grm_dims2, mat_gram, *gmt_dims2, matrix, *mat_dims2, matrix);
data->mat_gram = mat_gram;
}
PTR_FREE(gmt_dims2);
PTR_FREE(mat_dims2);
PTR_FREE(in_dims2);
data->gin_dims = *PTR_PASS(gin_dims2);
data->gout_dims = *PTR_PASS(gout_dims2);
data->grm_dims = *PTR_PASS(grm_dims2);
#else
data->gin_dims = NULL;
data->gout_dims = NULL;
data->grm_dims = NULL;
#endif
return PTR_PASS(data);
}
int main_reshape(int argc, char* argv[])
{
int c;
while (-1 != (c = getopt(argc, argv, "h"))) {
switch (c) {
case 'h':
usage(argv[0], stdout);
help();
exit(0);
default:
usage(argv[0], stderr);
exit(1);
}
}
if (argc - optind < 3) {
usage(argv[0], stderr);
exit(1);
}
unsigned int flags = atoi(argv[optind + 0]);
unsigned int n = bitcount(flags);
assert((int)n + 3 == argc - optind);
long in_dims[DIMS];
long in_strs[DIMS];
long out_dims[DIMS];
long out_strs[DIMS];
complex float* in_data = load_cfl(argv[optind + n + 1], DIMS, in_dims);
md_calc_strides(DIMS, in_strs, in_dims, CFL_SIZE);
md_copy_dims(DIMS, out_dims, in_dims);
unsigned int j = 0;
for (unsigned int i = 0; i < DIMS; i++)
if (MD_IS_SET(flags, i))
out_dims[i] = atoi(argv[optind + j++ + 1]);
assert(j == n);
assert(md_calc_size(DIMS, in_dims) == md_calc_size(DIMS, out_dims));
md_calc_strides(DIMS, out_strs, out_dims, CFL_SIZE);
for (unsigned int i = 0; i < DIMS; i++)
if (!(MD_IS_SET(flags, i) || (in_strs[i] == out_strs[i])))
error("Dimensions are not consistent at index %d.\n");
complex float* out_data = create_cfl(argv[optind + n + 2], DIMS, out_dims);
md_copy(DIMS, in_dims, out_data, in_data, CFL_SIZE);
unmap_cfl(DIMS, in_dims, in_data);
unmap_cfl(DIMS, out_dims, out_data);
exit(0);
}
/*
* Low rank threhsolding for arbitrary block sizes
*/
static void lrthresh_apply(const operator_data_t* _data, float mu, complex float* dst, const complex float* src)
{
struct lrthresh_data_s* data = CAST_DOWN(lrthresh_data_s, _data);
float lambda = mu * data->lambda;
long strs1[DIMS];
md_calc_strides(DIMS, strs1, data->dims_decom, 1);
//#pragma omp parallel for
for (int l = 0; l < data->levels; l++) {
complex float* dstl = dst + l * strs1[LEVEL_DIM];
const complex float* srcl = src + l * strs1[LEVEL_DIM];
long blkdims[DIMS];
long shifts[DIMS];
long unshifts[DIMS];
long zpad_dims[DIMS];
long M = 1;
for (unsigned int i = 0; i < DIMS; i++) {
blkdims[i] = data->blkdims[l][i];
zpad_dims[i] = (data->dims[i] + blkdims[i] - 1) / blkdims[i];
zpad_dims[i] *= blkdims[i];
if (MD_IS_SET(data->mflags, i))
M *= blkdims[i];
if (data->randshift)
shifts[i] = rand_lim(MIN(blkdims[i] - 1, zpad_dims[i] - blkdims[i]));
else
shifts[i] = 0;
unshifts[i] = -shifts[i];
}
long zpad_strs[DIMS];
md_calc_strides(DIMS, zpad_strs, zpad_dims, CFL_SIZE);
long blk_size = md_calc_size(DIMS, blkdims);
long img_size = md_calc_size(DIMS, zpad_dims);
long N = blk_size / M;
long B = img_size / blk_size;
if (data->noise && (l == data->levels - 1)) {
M = img_size;
N = 1;
B = 1;
}
complex float* tmp = md_alloc_sameplace(DIMS, zpad_dims, CFL_SIZE, dst);
md_circ_ext(DIMS, zpad_dims, tmp, data->dims, srcl, CFL_SIZE);
md_circ_shift(DIMS, zpad_dims, shifts, tmp, tmp, CFL_SIZE);
long mat_dims[2];
basorati_dims(DIMS, mat_dims, blkdims, zpad_dims);
complex float* tmp_mat = md_alloc_sameplace(2, mat_dims, CFL_SIZE, dst);
// Reshape image into a blk_size x number of blocks matrix
basorati_matrix(DIMS, blkdims, mat_dims, tmp_mat, zpad_dims, zpad_strs, tmp);
batch_svthresh(M, N, mat_dims[1], lambda * GWIDTH(M, N, B), *(complex float (*)[mat_dims[1]][M][N])tmp_mat);
// for ( int b = 0; b < mat_dims[1]; b++ )
// svthresh(M, N, lambda * GWIDTH(M, N, B), tmp_mat, tmp_mat);
basorati_matrixH(DIMS, blkdims, zpad_dims, zpad_strs, tmp, mat_dims, tmp_mat);
md_circ_shift(DIMS, zpad_dims, unshifts, tmp, tmp, CFL_SIZE);
md_resize(DIMS, data->dims, dstl, zpad_dims, tmp, CFL_SIZE);
md_free(tmp);
md_free(tmp_mat);
}
}
int main_nufft(int argc, char* argv[])
{
bool adjoint = false;
bool inverse = false;
bool use_gpu = false;
bool precond = false;
bool dft = false;
struct nufft_conf_s conf = nufft_conf_defaults;
struct iter_conjgrad_conf cgconf = iter_conjgrad_defaults;
long coilim_vec[3] = { 0 };
float lambda = 0.;
const struct opt_s opts[] = {
OPT_SET('a', &adjoint, "adjoint"),
OPT_SET('i', &inverse, "inverse"),
OPT_VEC3('d', &coilim_vec, "x:y:z", "dimensions"),
OPT_VEC3('D', &coilim_vec, "", "()"),
OPT_SET('t', &conf.toeplitz, "Toeplitz embedding for inverse NUFFT"),
OPT_SET('c', &precond, "Preconditioning for inverse NUFFT"),
OPT_FLOAT('l', &lambda, "lambda", "l2 regularization"),
OPT_UINT('m', &cgconf.maxiter, "", "()"),
OPT_SET('s', &dft, "DFT"),
};
cmdline(&argc, argv, 3, 3, usage_str, help_str, ARRAY_SIZE(opts), opts);
long coilim_dims[DIMS] = { 0 };
md_copy_dims(3, coilim_dims, coilim_vec);
// Read trajectory
long traj_dims[DIMS];
complex float* traj = load_cfl(argv[1], DIMS, traj_dims);
assert(3 == traj_dims[0]);
num_init();
if (inverse || adjoint) {
long ksp_dims[DIMS];
const complex float* ksp = load_cfl(argv[2], DIMS, ksp_dims);
assert(1 == ksp_dims[0]);
assert(md_check_compat(DIMS, ~(PHS1_FLAG|PHS2_FLAG), ksp_dims, traj_dims));
md_copy_dims(DIMS - 3, coilim_dims + 3, ksp_dims + 3);
if (0 == md_calc_size(DIMS, coilim_dims)) {
estimate_im_dims(DIMS, coilim_dims, traj_dims, traj);
debug_printf(DP_INFO, "Est. image size: %ld %ld %ld\n", coilim_dims[0], coilim_dims[1], coilim_dims[2]);
}
complex float* img = create_cfl(argv[3], DIMS, coilim_dims);
md_clear(DIMS, coilim_dims, img, CFL_SIZE);
const struct linop_s* nufft_op;
if (!dft)
nufft_op = nufft_create(DIMS, ksp_dims, coilim_dims, traj_dims, traj, NULL, conf, use_gpu);
else
nufft_op = nudft_create(DIMS, FFT_FLAGS, ksp_dims, coilim_dims, traj_dims, traj);
if (inverse) {
const struct operator_s* precond_op = NULL;
if (conf.toeplitz && precond)
precond_op = nufft_precond_create(nufft_op);
lsqr(DIMS, &(struct lsqr_conf){ lambda }, iter_conjgrad, CAST_UP(&cgconf),
nufft_op, NULL, coilim_dims, img, ksp_dims, ksp, precond_op);
if (conf.toeplitz && precond)
operator_free(precond_op);
} else {
void iter2_admm(void* _conf,
const struct operator_s* normaleq_op,
unsigned int D,
const struct operator_p_s** prox_ops,
const struct linop_s** ops,
const struct operator_p_s* xupdate_op,
long size, float* image, const float* image_adj,
const float* image_truth,
void* obj_eval_data,
float (*obj_eval)(const void*, const float*))
{
struct iter_admm_conf* conf = _conf;
struct admm_plan_s admm_plan = {
.maxiter = conf->maxiter,
.maxitercg = conf->maxitercg,
.rho = conf->rho,
.image_truth = image_truth,
.num_funs = D,
.do_warmstart = conf->do_warmstart,
.dynamic_rho = conf->dynamic_rho,
.hogwild = conf->hogwild,
.ABSTOL = conf->ABSTOL,
.RELTOL = conf->RELTOL,
.alpha = conf->alpha,
.tau = conf->tau,
.mu = conf->mu,
.fast = conf->fast,
};
struct admm_op a_ops[D];
struct admm_prox_op a_prox_ops[D];
for (unsigned int i = 0; i < D; i++) {
a_ops[i].forward = linop_forward_iter;
a_ops[i].normal = linop_normal_iter;
a_ops[i].adjoint = linop_adjoint_iter;
a_ops[i].data = (void*)ops[i];
a_prox_ops[i].prox_fun = operator_p_iter;
a_prox_ops[i].data = (void*)prox_ops[i];
}
admm_plan.ops = a_ops;
admm_plan.prox_ops = a_prox_ops;
admm_plan.xupdate_fun = operator_p_iter;
admm_plan.xupdate_data = (void*)xupdate_op;
struct admm_history_s admm_history;
long z_dims[D];
for (unsigned int i = 0; i < D; i++)
z_dims[i] = 2 * md_calc_size(linop_codomain(ops[i])->N, linop_codomain(ops[i])->dims);
if (NULL != image_adj) {
float eps = md_norm(1, MD_DIMS(size), image_adj);
if (checkeps(eps))
goto cleanup;
}
admm(&admm_history, &admm_plan, admm_plan.num_funs, z_dims, size, (float*)image, image_adj, select_vecops(image), operator_iter, (void*)normaleq_op, obj_eval_data, obj_eval);
cleanup:
;
}
void iter2_pocs(void* _conf,
const struct operator_s* normaleq_op,
unsigned int D,
const struct operator_p_s** prox_ops,
const struct linop_s** ops,
const struct operator_p_s* xupdate_op,
long size, float* image, const float* image_adj,
const float* image_truth,
void* obj_eval_data,
float (*obj_eval)(const void*, const float*))
{
const struct iter_pocs_conf* conf = _conf;
assert(NULL == normaleq_op);
assert(NULL == ops);
assert(NULL == image_adj);
UNUSED(xupdate_op);
UNUSED(image_adj);
struct pocs_proj_op proj_ops[D];
for (unsigned int i = 0; i < D; i++) {
proj_ops[i].proj_fun = operator_p_iter;
proj_ops[i].data = (void*)prox_ops[i];
}
pocs(conf->maxiter, D, proj_ops, select_vecops(image), size, image, image_truth, obj_eval_data, obj_eval);
}
void iter2_call_iter(void* _conf,
const struct operator_s* normaleq_op,
unsigned int D,
const struct operator_p_s** prox_ops,
const struct linop_s** ops,
const struct operator_p_s* xupdate_op,
long size, float* image, const float* image_adj,
const float* image_truth,
void* obj_eval_data,
float (*obj_eval)(const void*, const float*))
{
assert(D <= 1);
assert(NULL == ops);
UNUSED(xupdate_op);
struct iter_call_s* it = _conf;
it->fun(it->_conf, normaleq_op, (1 == D) ? prox_ops[0] : NULL,
size, image, image_adj,
image_truth, obj_eval_data, obj_eval);
}
/*
* Low rank threhsolding for arbitrary block sizes
*/
static void lrthresh_apply(const void* _data, float mu, complex float* dst, const complex float* src)
{
struct lrthresh_data_s* data = (struct lrthresh_data_s*)_data;
float lambda = mu * data->lambda;
long strs1[DIMS];
md_calc_strides(DIMS, strs1, data->dims_decom, 1);
//#pragma omp parallel for
for (int l = 0; l < data->levels; l++) {
complex float* dstl = dst + l * strs1[LEVEL_DIM];
const complex float* srcl = src + l * strs1[LEVEL_DIM];
// Initialize
long blkdims[DIMS];
long shifts[DIMS];
long unshifts[DIMS];
long zpad_dims[DIMS];
long M = 1;
for (unsigned int i = 0; i < DIMS; i++) {
blkdims[i] = data->blkdims[l][i];
zpad_dims[i] = (data->dims[i] + blkdims[i] - 1) / blkdims[i];
zpad_dims[i] *= blkdims[i];
if (MD_IS_SET(data->mflags, i))
M *= blkdims[i];
if (data->randshift)
shifts[i] = rand_lim(MIN(blkdims[i] - 1, zpad_dims[i] - blkdims[i]));
else
shifts[i] = 0;
unshifts[i] = -shifts[i];
}
long zpad_strs[DIMS];
md_calc_strides(DIMS, zpad_strs, zpad_dims, CFL_SIZE);
long blk_size = md_calc_size( DIMS, blkdims );
long img_size = md_calc_size( DIMS, zpad_dims );
long N = blk_size / M;
long B = img_size / blk_size;
if (data->noise && (l == data->levels - 1)) {
M = img_size;
N = 1;
B = 1;
}
// Initialize tmp
complex float* tmp_ext;
#ifdef USE_CUDA
tmp_ext = (data->use_gpu ? md_alloc_gpu : md_alloc)(DIMS, zpad_dims, CFL_SIZE);
#else
tmp_ext = md_alloc(DIMS, zpad_dims, CFL_SIZE);
#endif
complex float* tmp;
#ifdef USE_CUDA
tmp = (data->use_gpu ? md_alloc_gpu : md_alloc)(DIMS, zpad_dims, CFL_SIZE);
#else
tmp = md_alloc(DIMS, zpad_dims, CFL_SIZE);
#endif
// Copy to tmp
md_circ_ext(DIMS, zpad_dims, tmp_ext, data->dims, srcl, CFL_SIZE);
if (data->randshift)
md_circ_shift(DIMS, zpad_dims, shifts, tmp, tmp_ext, CFL_SIZE);
// Initialize tmp_mat
long mat_dims[2];
basorati_dims(DIMS, mat_dims, blkdims, zpad_dims);
complex float* tmp_mat;
#ifdef USE_CUDA
tmp_mat = (data->use_gpu ? md_alloc_gpu : md_alloc)(2, mat_dims, CFL_SIZE);
#else
tmp_mat = md_alloc(2, mat_dims, CFL_SIZE);
#endif
// Reshape image into a blk_size x number of blocks matrix
basorati_matrix(DIMS, blkdims, mat_dims, tmp_mat, zpad_dims, zpad_strs, tmp);
batch_svthresh(M, N, mat_dims[1], lambda * GWIDTH(M, N, B), tmp_mat, tmp_mat);
// for ( int b = 0; b < mat_dims[1]; b++ )
// svthresh(M, N, lambda * GWIDTH(M, N, B), tmp_mat, tmp_mat);
basorati_matrixH(DIMS, blkdims, zpad_dims, zpad_strs, tmp, mat_dims, tmp_mat);
// Copy to tmp
if (data->randshift)
md_circ_shift(DIMS, zpad_dims, unshifts, tmp_ext, tmp, CFL_SIZE);
md_resize(DIMS, data->dims, dstl, zpad_dims, tmp_ext, CFL_SIZE);
// Free data
md_free(tmp);
md_free(tmp_ext);
md_free(tmp_mat);
}
}
void overlapandsave2NE(int N, unsigned int flags, const long blk[N], const long odims[N], complex float* dst, const long dims1[N], complex float* src1, const long dims2[N], complex float* src2, const long mdims[N], complex float* msk)
{
long dims1B[N];
long tdims[2 * N];
long nodims[2 * N];
long ndims1[2 * N];
long ndims2[2 * N];
long shift[2 * N];
unsigned int nflags = 0;
for (int i = 0; i < N; i++) {
if (MD_IS_SET(flags, i)) {
nflags = MD_SET(nflags, 2 * i);
assert(1 == dims2[i] % 2);
assert(0 == blk[i] % 2);
assert(0 == dims1[i] % 2);
assert(0 == odims[i] % blk[i]);
assert(0 == dims1[i] % blk[i]);
assert(dims1[i] == odims[i]);
assert(dims2[i] <= blk[i]);
assert(dims1[i] >= dims2[i]);
// blocked output
nodims[i * 2 + 1] = odims[i] / blk[i];
nodims[i * 2 + 0] = blk[i];
// expanded temporary storage
tdims[i * 2 + 1] = dims1[i] / blk[i];
tdims[i * 2 + 0] = blk[i] + dims2[i] - 1;
// blocked input
// ---|---,---,---|---
// + +++ +
// + +++ +
// resized input
dims1B[i] = dims1[i] + 2 * blk[i];
ndims1[i * 2 + 1] = dims1[i] / blk[i] + 2; // do we need two full blocks?
ndims1[i * 2 + 0] = blk[i];
shift[i * 2 + 1] = 0;
shift[i * 2 + 0] = blk[i] - (dims2[i] - 1) / 2;
// kernel
ndims2[i * 2 + 1] = 1;
ndims2[i * 2 + 0] = dims2[i];
} else {
nodims[i * 2 + 1] = 1;
nodims[i * 2 + 0] = odims[i];
tdims[i * 2 + 1] = 1;
tdims[i * 2 + 0] = dims1[i];
ndims1[i * 2 + 1] = 1;
ndims1[i * 2 + 0] = dims1[i];
shift[i * 2 + 1] = 0;
shift[i * 2 + 0] = 0;
dims1B[i] = dims1[i];
ndims2[i * 2 + 1] = 1;
ndims2[i * 2 + 0] = dims2[i];
}
}
long R = md_calc_size(N, odims);
long T = md_calc_size(2 * N, tdims);
long S = md_calc_size(N, dims1B);
complex float* src1B = xmalloc(S * sizeof(complex float));
complex float* tmp = xmalloc(T * sizeof(complex float));
complex float* tmpX = xmalloc(R * sizeof(complex float));
long str1[2 * N];
long str2[2 * N];
md_calc_strides(2 * N, str1, ndims1, sizeof(complex float));
md_calc_strides(2 * N, str2, tdims, sizeof(complex float));
long off = md_calc_offset(2 * N, str1, shift);
md_resize_center(N, dims1B, src1B, dims1, src1, sizeof(complex float));
// we can loop here
md_copy2(2 * N, tdims, str2, tmp, str1, ((void*)src1B) + off, sizeof(complex float));
conv(2 * N, nflags, CONV_VALID, CONV_SYMMETRIC, nodims, tmpX, tdims, tmp, ndims2, src2);
long ostr[N];
long mstr[N];
md_calc_strides(N, ostr, odims, sizeof(complex float));
md_calc_strides(N, mstr, mdims, sizeof(complex float));
md_zmul2(N, odims, ostr, tmpX, ostr, tmpX, mstr, msk);
convH(2 * N, nflags, CONV_VALID, CONV_SYMMETRIC, tdims, tmp, nodims, tmpX, ndims2, src2);
md_clear(N, dims1B, src1B, sizeof(complex float));
md_zadd2(2 * N, tdims, str1, ((void*)src1B) + off, str1, ((void*)src1B) + off, str2, tmp);
//
md_resize_center(N, dims1, dst, dims1B, src1B, sizeof(complex float));
free(src1B);
free(tmpX);
free(tmp);
}
/**
* Efficiently chain two matrix linops by multiplying the actual matrices together.
* Stores a copy of the new matrix.
* Returns: C = B A
*
* @param a first matrix (applied to input)
* @param b second matrix (applied to output of first matrix)
*/
struct linop_s* linop_matrix_chain(const struct linop_s* a, const struct linop_s* b)
{
const struct operator_matrix_s* a_data = CAST_DOWN(operator_matrix_s, linop_get_data(a));
const struct operator_matrix_s* b_data = CAST_DOWN(operator_matrix_s, linop_get_data(b));
// check compatibility
assert(linop_codomain(a)->N == linop_domain(b)->N);
assert(md_check_compat(linop_codomain(a)->N, 0u, linop_codomain(a)->dims, linop_domain(b)->dims));
unsigned int D = linop_domain(a)->N;
unsigned long outB_flags = md_nontriv_dims(D, linop_codomain(b)->dims);
unsigned long inB_flags = md_nontriv_dims(D, linop_domain(b)->dims);
unsigned long delB_flags = inB_flags & ~outB_flags;
unsigned int N = a_data->N;
assert(N == 2 * D);
long in_dims[N];
md_copy_dims(N, in_dims, a_data->in_dims);
long matA_dims[N];
md_copy_dims(N, matA_dims, a_data->mat_dims);
long matB_dims[N];
md_copy_dims(N, matB_dims, b_data->mat_dims);
long out_dims[N];
md_copy_dims(N, out_dims, b_data->out_dims);
for (unsigned int i = 0; i < D; i++) {
if (MD_IS_SET(delB_flags, i)) {
matA_dims[2 * i + 0] = a_data->mat_dims[2 * i + 1];
matA_dims[2 * i + 1] = a_data->mat_dims[2 * i + 0];
in_dims[2 * i + 0] = a_data->in_dims[2 * i + 1];
in_dims[2 * i + 1] = a_data->in_dims[2 * i + 0];
}
}
long matrix_dims[N];
md_singleton_dims(N, matrix_dims);
unsigned long iflags = md_nontriv_dims(N, in_dims);
unsigned long oflags = md_nontriv_dims(N, out_dims);
unsigned long flags = iflags | oflags;
// we combine a and b and sum over dims not in input or output
md_max_dims(N, flags, matrix_dims, matA_dims, matB_dims);
debug_printf(DP_DEBUG1, "tensor chain: %ld x %ld -> %ld\n",
md_calc_size(N, matA_dims), md_calc_size(N, matB_dims), md_calc_size(N, matrix_dims));
complex float* matrix = md_alloc(N, matrix_dims, CFL_SIZE);
debug_print_dims(DP_DEBUG2, N, matrix_dims);
debug_print_dims(DP_DEBUG2, N, in_dims);
debug_print_dims(DP_DEBUG2, N, matA_dims);
debug_print_dims(DP_DEBUG2, N, matB_dims);
debug_print_dims(DP_DEBUG2, N, out_dims);
md_ztenmul(N, matrix_dims, matrix, matA_dims, a_data->mat, matB_dims, b_data->mat);
// priv2 takes our doubled dimensions
struct operator_matrix_s* data = linop_matrix_priv2(N, out_dims, in_dims, matrix_dims, matrix);
/* although we internally use different dimensions we define the
* correct interface
*/
struct linop_s* c = linop_create(linop_codomain(b)->N, linop_codomain(b)->dims,
linop_domain(a)->N, linop_domain(a)->dims, CAST_UP(data),
linop_matrix_apply, linop_matrix_apply_adjoint,
linop_matrix_apply_normal, NULL, linop_matrix_del);
md_free(matrix);
return c;
}
int main_nlinv(int argc, char* argv[])
{
int iter = 8;
float l1 = -1.;
bool waterfat = false;
bool rvc = false;
bool normalize = true;
float restrict_fov = -1.;
float csh[3] = { 0., 0., 0. };
bool usegpu = false;
const char* psf = NULL;
const struct opt_s opts[] = {
{ 'l', true, opt_float, &l1, NULL },
{ 'i', true, opt_int, &iter, NULL },
{ 'c', false, opt_set, &rvc, NULL },
{ 'N', false, opt_clear, &normalize, NULL },
{ 'f', true, opt_float, &restrict_fov, NULL },
{ 'p', true, opt_string, &psf, NULL },
{ 'g', false, opt_set, &usegpu, NULL },
};
cmdline(&argc, argv, 2, 3, usage_str, help_str, ARRAY_SIZE(opts), opts);
num_init();
assert(iter > 0);
long ksp_dims[DIMS];
complex float* kspace_data = load_cfl(argv[1], DIMS, ksp_dims);
long dims[DIMS];
md_copy_dims(DIMS, dims, ksp_dims);
if (waterfat)
dims[CSHIFT_DIM] = 2;
long img_dims[DIMS];
md_select_dims(DIMS, FFT_FLAGS|CSHIFT_FLAG, img_dims, dims);
long img_strs[DIMS];
md_calc_strides(DIMS, img_strs, img_dims, CFL_SIZE);
complex float* image = create_cfl(argv[2], DIMS, img_dims);
long msk_dims[DIMS];
md_select_dims(DIMS, FFT_FLAGS, msk_dims, dims);
long msk_strs[DIMS];
md_calc_strides(DIMS, msk_strs, msk_dims, CFL_SIZE);
complex float* mask;
complex float* norm = md_alloc(DIMS, msk_dims, CFL_SIZE);
complex float* sens;
if (4 == argc) {
sens = create_cfl(argv[3], DIMS, ksp_dims);
} else {
sens = md_alloc(DIMS, ksp_dims, CFL_SIZE);
}
complex float* pattern = NULL;
long pat_dims[DIMS];
if (NULL != psf) {
pattern = load_cfl(psf, DIMS, pat_dims);
// FIXME: check compatibility
} else {
pattern = md_alloc(DIMS, img_dims, CFL_SIZE);
estimate_pattern(DIMS, ksp_dims, COIL_DIM, pattern, kspace_data);
}
if (waterfat) {
size_t size = md_calc_size(DIMS, msk_dims);
md_copy(DIMS, msk_dims, pattern + size, pattern, CFL_SIZE);
long shift_dims[DIMS];
md_select_dims(DIMS, FFT_FLAGS, shift_dims, msk_dims);
long shift_strs[DIMS];
md_calc_strides(DIMS, shift_strs, shift_dims, CFL_SIZE);
complex float* shift = md_alloc(DIMS, shift_dims, CFL_SIZE);
unsigned int X = shift_dims[READ_DIM];
unsigned int Y = shift_dims[PHS1_DIM];
unsigned int Z = shift_dims[PHS2_DIM];
for (unsigned int x = 0; x < X; x++)
for (unsigned int y = 0; y < Y; y++)
for (unsigned int z = 0; z < Z; z++)
shift[(z * Z + y) * Y + x] = cexp(2.i * M_PI * ((csh[0] * x) / X + (csh[1] * y) / Y + (csh[2] * z) / Z));
md_zmul2(DIMS, msk_dims, msk_strs, pattern + size, msk_strs, pattern + size, shift_strs, shift);
md_free(shift);
}
#if 0
float scaling = 1. / estimate_scaling(ksp_dims, NULL, kspace_data);
#else
float scaling = 100. / md_znorm(DIMS, ksp_dims, kspace_data);
#endif
debug_printf(DP_INFO, "Scaling: %f\n", scaling);
md_zsmul(DIMS, ksp_dims, kspace_data, kspace_data, scaling);
if (-1. == restrict_fov) {
mask = md_alloc(DIMS, msk_dims, CFL_SIZE);
md_zfill(DIMS, msk_dims, mask, 1.);
} else {
float restrict_dims[DIMS] = { [0 ... DIMS - 1] = 1. };
restrict_dims[0] = restrict_fov;
restrict_dims[1] = restrict_fov;
restrict_dims[2] = restrict_fov;
mask = compute_mask(DIMS, msk_dims, restrict_dims);
}
#ifdef USE_CUDA
if (usegpu) {
complex float* kspace_gpu = md_alloc_gpu(DIMS, ksp_dims, CFL_SIZE);
md_copy(DIMS, ksp_dims, kspace_gpu, kspace_data, CFL_SIZE);
noir_recon(dims, iter, l1, image, NULL, pattern, mask, kspace_gpu, rvc, usegpu);
md_free(kspace_gpu);
md_zfill(DIMS, ksp_dims, sens, 1.);
} else
#endif
noir_recon(dims, iter, l1, image, sens, pattern, mask, kspace_data, rvc, usegpu);
if (normalize) {
md_zrss(DIMS, ksp_dims, COIL_FLAG, norm, sens);
md_zmul2(DIMS, img_dims, img_strs, image, img_strs, image, msk_strs, norm);
}
if (4 == argc) {
long strs[DIMS];
md_calc_strides(DIMS, strs, ksp_dims, CFL_SIZE);
if (norm)
md_zdiv2(DIMS, ksp_dims, strs, sens, strs, sens, img_strs, norm);
fftmod(DIMS, ksp_dims, FFT_FLAGS, sens, sens);
unmap_cfl(DIMS, ksp_dims, sens);
} else {
md_free(sens);
}
md_free(norm);
md_free(mask);
if (NULL != psf)
unmap_cfl(DIMS, pat_dims, pattern);
else
md_free(pattern);
unmap_cfl(DIMS, img_dims, image);
unmap_cfl(DIMS, ksp_dims, kspace_data);
exit(0);
}
void compute_kernels(const struct ecalib_conf* conf, long nskerns_dims[5], complex float** nskerns_ptr, unsigned int SN, float svals[SN], const long caldims[DIMS], const complex float* caldata)
{
assert(1 == md_calc_size(DIMS - 5, caldims + 5));
nskerns_dims[0] = conf->kdims[0];
nskerns_dims[1] = conf->kdims[1];
nskerns_dims[2] = conf->kdims[2];
nskerns_dims[3] = caldims[3];
long N = md_calc_size(4, nskerns_dims);
assert(N > 0);
nskerns_dims[4] = N;
complex float* nskerns = md_alloc(5, nskerns_dims, CFL_SIZE);
*nskerns_ptr = nskerns;
complex float (*vec)[N] = xmalloc(N * N * sizeof(complex float));
assert((NULL == svals) || (SN == N));
float* val = (NULL != svals) ? svals : xmalloc(N * FL_SIZE);
debug_printf(DP_DEBUG1, "Build calibration matrix and SVD...\n");
#ifdef CALMAT_SVD
calmat_svd(conf->kdims, N, vec, val, caldims, caldata);
for (int i = 0; i < N; i++)
for (int j = 0; j < N; j++)
#ifndef FLIP
nskerns[i * N + j] = (vec[j][i]) * (conf->weighting ? val[i] : 1.);
#else
nskerns[i * N + j] = (vec[j][N - 1 - i]) * (conf->weighting ? val[N - 1 - i] : 1.);
#endif
#else
covariance_function(conf->kdims, N, vec, caldims, caldata);
debug_printf(DP_DEBUG1, "Eigen decomposition... (size: %ld)\n", N);
// we could apply Nystroem method here to speed it up
float tmp_val[N];
lapack_eig(N, tmp_val, vec);
for (int i = 0; i < N; i++)
val[i] = sqrtf(tmp_val[N - 1 - i]);
for (int i = 0; i < N; i++)
for (int j = 0; j < N; j++)
#ifndef FLIP
nskerns[i * N + j] = vec[N - 1 - i][j] * (conf->weighting ? val[i] : 1.); // flip
#else
nskerns[i * N + j] = vec[i][j] * (conf->weighting ? val[N - 1 - i] : 1.); // flip
#endif
#endif
if (conf->perturb > 0.) {
long dims[2] = { N, N };
perturb(dims, nskerns, conf->perturb);
}
#ifndef FLIP
nskerns_dims[4] = number_of_kernels(conf, N, val);
#else
nskerns_dims[4] = N - number_of_kernels(conf, N, val);
#endif
if (NULL == svals)
free(val);
free(vec);
}
static void prox_4pt_dfwavelet_thresh(const operator_data_t* _data, float thresh, complex float* out, const complex float* in)
{
struct prox_4pt_dfwavelet_data* data = CONTAINER_OF(_data, struct prox_4pt_dfwavelet_data, base);
bool done = false;
long pos[DIMS];
md_set_dims(DIMS, pos, 0);
while (!done) {
// copy pc
md_slice(DIMS, data->slice_flag, pos, data->im_dims, data->pc0, in, CFL_SIZE);
pos[data->flow_dim]++;
md_slice(DIMS, data->slice_flag, pos, data->im_dims, data->pc1, in, CFL_SIZE);
pos[data->flow_dim]++;
md_slice(DIMS, data->slice_flag, pos, data->im_dims, data->pc2, in, CFL_SIZE);
pos[data->flow_dim]++;
md_slice(DIMS, data->slice_flag, pos, data->im_dims, data->pc3, in, CFL_SIZE);
pos[data->flow_dim] = 0;
// pc to velocity
// TODO: Make gpu
for (int i = 0; i < md_calc_size(DIMS, data->tim_dims); i++) {
data->vx[i] = (data->pc1[i] - data->pc0[i]) / 2;
data->vy[i] = (data->pc2[i] - data->pc1[i]) / 2;
data->vz[i] = (data->pc3[i] - data->pc2[i]) / 2;
data->ph0[i] = (data->pc0[i] + data->pc3[i]) / 2;
}
// threshold
dfwavelet_thresh(data->plan, thresh * data->lambda, thresh* data->lambda, data->vx, data->vy, data->vz, data->vx, data->vy, data->vz);
operator_p_apply(data->wthresh_op, thresh, DIMS, data->tim_dims, data->ph0, DIMS, data->tim_dims, data->ph0);
// velocity to pc
for (int i = 0; i < md_calc_size(DIMS, data->tim_dims ); i++) {
data->pc0[i] = (- data->vx[i] - data->vy[i] - data->vz[i] + data->ph0[i]);
data->pc1[i] = (+ data->vx[i] - data->vy[i] - data->vz[i] + data->ph0[i]);
data->pc2[i] = (+ data->vx[i] + data->vy[i] - data->vz[i] + data->ph0[i]);
data->pc3[i] = (+ data->vx[i] + data->vy[i] + data->vz[i] + data->ph0[i]);
}
// copy pc
md_copy_block(DIMS, pos, data->im_dims, out, data->tim_dims, data->pc0, CFL_SIZE);
pos[data->flow_dim]++;
md_copy_block(DIMS, pos, data->im_dims, out, data->tim_dims, data->pc1, CFL_SIZE);
pos[data->flow_dim]++;
md_copy_block( DIMS, pos, data->im_dims, out, data->tim_dims, data->pc2, CFL_SIZE );
pos[data->flow_dim]++;
md_copy_block( DIMS, pos, data->im_dims, out, data->tim_dims, data->pc3, CFL_SIZE );
pos[data->flow_dim] = 0;
// increment pos
long carryon = 1;
for(unsigned int i = 0; i < DIMS; i++) {
if (MD_IS_SET(data->slice_flag & ~MD_BIT(data->flow_dim), i)) {
pos[i] += carryon;
if (pos[i] < data->im_dims[i]) {
carryon = 0;
break;
} else {
carryon = 1;
pos[i] = 0;
}
}
}
done = carryon;
}
}
void iter2_admm(iter_conf* _conf,
const struct operator_s* normaleq_op,
unsigned int D,
const struct operator_p_s* prox_ops[D],
const struct linop_s* ops[D],
const float* biases[D],
const struct operator_p_s* xupdate_op,
long size, float* image, const float* image_adj,
struct iter_monitor_s* monitor)
{
auto conf = CAST_DOWN(iter_admm_conf, _conf);
struct admm_plan_s admm_plan = {
.maxiter = conf->maxiter,
.maxitercg = conf->maxitercg,
.cg_eps = conf->cg_eps,
.rho = conf->rho,
.num_funs = D,
.do_warmstart = conf->do_warmstart,
.dynamic_rho = conf->dynamic_rho,
.dynamic_tau = conf->dynamic_tau,
.relative_norm = conf->relative_norm,
.hogwild = conf->hogwild,
.ABSTOL = conf->ABSTOL,
.RELTOL = conf->RELTOL,
.alpha = conf->alpha,
.tau = conf->tau,
.tau_max = conf->tau_max,
.mu = conf->mu,
.fast = conf->fast,
.biases = biases,
};
struct admm_op a_ops[D ?:1];
struct iter_op_p_s a_prox_ops[D ?:1];
for (unsigned int i = 0; i < D; i++) {
a_ops[i].forward = OPERATOR2ITOP(ops[i]->forward),
a_ops[i].normal = OPERATOR2ITOP(ops[i]->normal);
a_ops[i].adjoint = OPERATOR2ITOP(ops[i]->adjoint);
a_prox_ops[i] = OPERATOR_P2ITOP(prox_ops[i]);
}
admm_plan.ops = a_ops;
admm_plan.prox_ops = a_prox_ops;
admm_plan.xupdate = OPERATOR_P2ITOP(xupdate_op);
long z_dims[D ?: 1];
for (unsigned int i = 0; i < D; i++)
z_dims[i] = 2 * md_calc_size(linop_codomain(ops[i])->N, linop_codomain(ops[i])->dims);
if (NULL != image_adj) {
float eps = md_norm(1, MD_DIMS(size), image_adj);
if (checkeps(eps))
goto cleanup;
}
admm(&admm_plan, admm_plan.num_funs, z_dims, size, (float*)image, image_adj, select_vecops(image), OPERATOR2ITOP(normaleq_op), monitor);
cleanup:
;
}
void iter2_pocs(iter_conf* _conf,
const struct operator_s* normaleq_op,
unsigned int D,
const struct operator_p_s* prox_ops[D],
const struct linop_s* ops[D],
const float* biases[D],
const struct operator_p_s* xupdate_op,
long size, float* image, const float* image_adj,
struct iter_monitor_s* monitor)
{
auto conf = CAST_DOWN(iter_pocs_conf, _conf);
assert(NULL == normaleq_op);
assert(NULL == ops);
assert(NULL == biases);
assert(NULL == image_adj);
UNUSED(xupdate_op);
UNUSED(image_adj);
struct iter_op_p_s proj_ops[D];
for (unsigned int i = 0; i < D; i++)
proj_ops[i] = OPERATOR_P2ITOP(prox_ops[i]);
pocs(conf->maxiter, D, proj_ops, select_vecops(image), size, image, monitor);
}
int main_bpsense(int argc, char* argv[])
{
// -----------------------------------------------------------
// set up conf and option parser
struct bpsense_conf conf = bpsense_defaults;
struct iter_admm_conf iconf = iter_admm_defaults;
conf.iconf = &iconf;
conf.iconf->rho = 10; // more sensibile default
bool usegpu = false;
const char* psf = NULL;
const char* image_truth_fname = NULL;
bool im_truth = false;
bool use_tvnorm = false;
double start_time = timestamp();
const struct opt_s opts[] = {
OPT_FLOAT('e', &conf.eps, "eps", "data consistency error"),
OPT_FLOAT('r', &conf.lambda, "lambda", "l2 regularization parameter"),
OPT_FLOAT('u', &conf.iconf->rho, "rho", "ADMM penalty parameter"),
OPT_SET('c', &conf.rvc, "real-value constraint"),
OPT_SET('t', &use_tvnorm, "use TV norm"),
OPT_STRING('T', &image_truth_fname, "file", "compare to truth image"),
OPT_UINT('i', &conf.iconf->maxiter, "iter", "max. iterations"),
OPT_SET('g', &usegpu, "(use gpu)"),
OPT_STRING('p', &psf, "file", "point-spread function"),
};
cmdline(&argc, argv, 3, 3, usage_str, help_str, ARRAY_SIZE(opts), opts);
if (NULL != image_truth_fname)
im_truth = true;
// -----------------------------------------------------------
// load data and print some info about the recon
int N = DIMS;
long dims[N];
long dims1[N];
long img_dims[N];
long ksp_dims[N];
complex float* kspace_data = load_cfl(argv[1], N, ksp_dims);
complex float* sens_maps = load_cfl(argv[2], N, dims);
for (int i = 0; i < 4; i++) // sizes2[4] may be > 1
if (ksp_dims[i] != dims[i])
error("Dimensions of kspace and sensitivities do not match!\n");
assert(1 == ksp_dims[MAPS_DIM]);
(usegpu ? num_init_gpu : num_init)();
if (dims[MAPS_DIM] > 1)
debug_printf(DP_INFO, "%ld maps.\nESPIRiT reconstruction.\n", dims[4]);
if (conf.lambda > 0.)
debug_printf(DP_INFO, "l2 regularization: %f\n", conf.lambda);
if (use_tvnorm)
debug_printf(DP_INFO, "use Total Variation\n");
else
debug_printf(DP_INFO, "use Wavelets\n");
if (im_truth)
debug_printf(DP_INFO, "Compare to truth\n");
md_select_dims(N, ~(COIL_FLAG | MAPS_FLAG), dims1, dims);
md_select_dims(N, ~COIL_FLAG, img_dims, dims);
// -----------------------------------------------------------
// initialize sampling pattern
complex float* pattern = NULL;
long pat_dims[N];
if (NULL != psf) {
pattern = load_cfl(psf, N, pat_dims);
// FIXME: check compatibility
} else {
pattern = md_alloc(N, dims1, CFL_SIZE);
estimate_pattern(N, ksp_dims, COIL_DIM, pattern, kspace_data);
}
// -----------------------------------------------------------
// print some statistics
size_t T = md_calc_size(N, dims1);
long samples = (long)pow(md_znorm(N, dims1, pattern), 2.);
debug_printf(DP_INFO, "Size: %ld Samples: %ld Acc: %.2f\n", T, samples, (float)T/(float)samples);
// -----------------------------------------------------------
// fftmod to un-center data
fftmod(N, ksp_dims, FFT_FLAGS, kspace_data, kspace_data);
fftmod(N, dims, FFT_FLAGS, sens_maps, sens_maps);
// -----------------------------------------------------------
// apply scaling
float scaling = estimate_scaling(ksp_dims, NULL, kspace_data);
debug_printf(DP_INFO, "Scaling: %f\n", scaling);
if (scaling != 0.)
md_zsmul(N, ksp_dims, kspace_data, kspace_data, 1. / scaling);
// -----------------------------------------------------------
// create l1 prox operator and transform
long minsize[DIMS] = { [0 ... DIMS - 1] = 1 };
minsize[0] = MIN(img_dims[0], 16);
minsize[1] = MIN(img_dims[1], 16);
minsize[2] = MIN(img_dims[2], 16);
const struct linop_s* l1op = NULL;
const struct operator_p_s* l1prox = NULL;
if (use_tvnorm) {
l1op = grad_init(DIMS, img_dims, FFT_FLAGS);
l1prox = prox_thresh_create(DIMS + 1, linop_codomain(l1op)->dims, 1., 0u, usegpu);
conf.l1op_obj = l1op;
} else {
bool randshift = true;
l1op = linop_identity_create(DIMS, img_dims);
conf.l1op_obj = wavelet_create(DIMS, img_dims, FFT_FLAGS, minsize, false, usegpu);
l1prox = prox_wavethresh_create(DIMS, img_dims, FFT_FLAGS, minsize, 1., randshift, usegpu);
}
// -----------------------------------------------------------
// create image and load truth image
complex float* image = create_cfl(argv[3], N, img_dims);
md_clear(N, img_dims, image, CFL_SIZE);
long img_truth_dims[DIMS];
complex float* image_truth = NULL;
if (im_truth)
image_truth = load_cfl(image_truth_fname, DIMS, img_truth_dims);
// -----------------------------------------------------------
// call recon
if (usegpu)
#ifdef USE_CUDA
bpsense_recon_gpu(&conf, dims, image, sens_maps, dims1, pattern, l1op, l1prox, ksp_dims, kspace_data, image_truth);
#else
assert(0);
#endif
else
void iwt2(unsigned int N, unsigned int flags, const long shifts[N], const long odims[N], const long ostr[N], complex float* out, const long idims[N], const long istr[N], const complex float* in, const long minsize[N], const long flen, const float filter[2][2][flen])
{
assert(wavelet_check_dims(N, flags, odims, minsize));
if (0 == flags) { // note: recursion does *not* end here
assert(md_check_compat(N, 0u, odims, idims));
md_copy2(N, idims, ostr, out, istr, in, CFL_SIZE);
return;
}
// check input dimensions
long idims2[N];
wavelet_coeffs2(N, flags, idims2, odims, minsize, flen);
assert(md_check_compat(N, 0u, idims2, idims));
long wdims2[2 * N];
wavelet_dims(N, flags, wdims2, odims, flen);
// only consider transform dims...
long dims1[N];
md_select_dims(N, flags, dims1, odims);
long wdims[2 * N];
wavelet_dims(N, flags, wdims, dims1, flen);
long level_coeffs = md_calc_size(2 * N, wdims);
// ... which get embedded in dimension b
unsigned int b = ffs(flags) - 1;
long istr2[2 * N];
md_calc_strides(2 * N, istr2, wdims, istr[b]);
// merge with original strides
for (unsigned int i = 0; i < N; i++)
if (!MD_IS_SET(flags, i))
istr2[i] = istr[i];
assert(idims[b] >= level_coeffs);
long offset = (idims[b] - level_coeffs) * (istr[b] / CFL_SIZE);
long bands = md_calc_size(N, wdims + N);
long coeffs = md_calc_size(N, wdims + 0);
// subtract coefficients in high band
idims2[b] -= (bands - 1) * coeffs;
assert(idims2[b] > 0);
debug_printf(DP_DEBUG4, "ifwt2: flags:%d lcoeffs:%ld coeffs:%ld (space:%ld) bands:%ld str:%ld off:%ld\n", flags, level_coeffs, coeffs, idims2[b], bands, istr[b], offset / ostr[b]);
// fix me we need temp storage
complex float* tmp = md_alloc_sameplace(2 * N, wdims2, CFL_SIZE, out);
long tstr[2 * N];
md_calc_strides(2 * N, tstr, wdims2, CFL_SIZE);
md_copy2(2 * N, wdims2, tstr, tmp, istr2, in + offset, CFL_SIZE);
long shifts0[N];
for (unsigned int i = 0; i < N; i++)
shifts0[i] = 0;
unsigned int flags2 = wavelet_filter_flags(N, flags, wdims, minsize);
assert((0 == offset) == (0u == flags2));
if (0u != flags2) {
long idims3[N];
wavelet_coeffs2(N, flags2, idims3, wdims2, minsize, flen);
long istr3[N];
embed(N, flags, istr3, idims3, istr);
iwt2(N, flags2, shifts0, wdims2, tstr, tmp, idims3, istr3, in, minsize, flen, filter);
}
iwtN(N, flags, shifts, odims, ostr, out, tstr, tmp, flen, filter);
md_free(tmp);
}
