/**
* Intialize lrthresh data
*
* @param dims_decom - dimensions with levels at LEVEL_DIMS
* @param randshift - randshift boolean
* @param mflags - selects which dimensions gets reshaped as the first dimension in matrix
* @param blkdims - contains block dimensions for all levels
*
*/
static struct lrthresh_data_s* lrthresh_create_data(const long dims_decom[DIMS], bool randshift, unsigned long mflags, const long blkdims[MAX_LEV][DIMS], float lambda, bool noise, int remove_mean)
{
PTR_ALLOC(struct lrthresh_data_s, data);
SET_TYPEID(lrthresh_data_s, data);
data->randshift = randshift;
data->mflags = mflags;
data->lambda = lambda;
data->noise = noise;
data->remove_mean = remove_mean;
// level dimensions
md_copy_dims(DIMS, data->dims_decom, dims_decom);
md_calc_strides(DIMS, data->strs_lev, dims_decom, CFL_SIZE);
// image dimensions
data->levels = dims_decom[LEVEL_DIM];
md_select_dims(DIMS, ~LEVEL_FLAG, data->dims, dims_decom);
md_calc_strides(DIMS, data->strs, data->dims, CFL_SIZE);
// blkdims
for(long l = 0; l < data->levels; l++) {
for (long i = 0; i < DIMS; i++)
data->blkdims[l][i] = blkdims[l][i];
}
return PTR_PASS(data);
}
/**
* Proximal operator for l1-norm with unitary transform: f(x) = lambda || T x ||_1
*
* @param D number of dimensions
* @param dim dimensions of x
* @param lambda threshold parameter
* @param unitary_op unitary linear operator
* @param flags bitmask for joint soft-thresholding
*/
extern const struct operator_p_s* prox_unithresh_create(unsigned int D, const struct linop_s* unitary_op, const float lambda, const unsigned long flags)
{
PTR_ALLOC(struct thresh_s, data);
SET_TYPEID(thresh_s, data);
data->lambda = lambda;
data->D = D;
data->flags = flags;
data->unitary_op = unitary_op;
const long* dims = linop_domain(unitary_op)->dims;
PTR_ALLOC(long[D], ndim);
md_copy_dims(D, *ndim, dims);
data->dim = *PTR_PASS(ndim);
PTR_ALLOC(long[D], nstr);
md_calc_strides(D, *nstr, data->dim, CFL_SIZE);
data->str = *PTR_PASS(nstr);
// norm dimensions are the flagged transform dimensions
// FIXME should use linop_codomain(unitary_op)->N
PTR_ALLOC(long[D], norm_dim);
md_select_dims(D, ~flags, *norm_dim, linop_codomain(unitary_op)->dims);
data->norm_dim = *PTR_PASS(norm_dim);
return operator_p_create(D, dims, D, dims, CAST_UP(PTR_PASS(data)), unisoftthresh_apply, thresh_del);
}
static struct ufft_data* ufft_create_data(const long ksp_dims[DIMS], const long pat_dims[DIMS], const complex float* pat, unsigned int flags, bool use_gpu)
{
PTR_ALLOC(struct ufft_data, data);
SET_TYPEID(ufft_data, data);
data->flags = flags;
data->use_gpu = use_gpu;
md_copy_dims(DIMS, data->pat_dims, pat_dims);
md_copy_dims(DIMS, data->ksp_dims, ksp_dims);
md_calc_strides(DIMS, data->pat_strs, pat_dims, CFL_SIZE);
md_calc_strides(DIMS, data->ksp_strs, ksp_dims, CFL_SIZE);
#ifdef USE_CUDA
data->pat = (use_gpu ? md_alloc_gpu : md_alloc)(DIMS, data->pat_dims, CFL_SIZE);
#else
data->pat = md_alloc(DIMS, data->pat_dims, CFL_SIZE);
#endif
md_copy(DIMS, data->pat_dims, data->pat, pat, CFL_SIZE);
data->fft_op = linop_fftc_create(DIMS, ksp_dims, flags);
return PTR_PASS(data);
}
const struct operator_s* fft_create2(unsigned int D, const long dimensions[D], unsigned long flags, const long ostrides[D], complex float* dst, const long istrides[D], const complex float* src, bool backwards)
{
PTR_ALLOC(struct fft_plan_s, plan);
SET_TYPEID(fft_plan_s, plan);
plan->fftw = fft_fftwf_plan(D, dimensions, flags, ostrides, dst, istrides, src, backwards, false);
#ifdef USE_CUDA
plan->cuplan = NULL;
#ifndef LAZY_CUDA
if (cuda_ondevice(src))
plan->cuplan = fft_cuda_plan(D, dimensions, flags, ostrides, istrides, backwards);
#else
plan->D = D;
plan->flags = flags;
plan->backwards = backwards;
PTR_ALLOC(long[D], dims);
md_copy_dims(D, *dims, dimensions);
plan->dims = *PTR_PASS(dims);
PTR_ALLOC(long[D], istrs);
md_copy_strides(D, *istrs, istrides);
plan->istrs = *PTR_PASS(istrs);
PTR_ALLOC(long[D], ostrs);
md_copy_strides(D, *ostrs, ostrides);
plan->ostrs = *PTR_PASS(ostrs);
#endif
#endif
return operator_create2(D, dimensions, ostrides, D, dimensions, istrides, CAST_UP(PTR_PASS(plan)), fft_apply, fft_free_plan);
}
const struct operator_s* nufft_precond_create(const struct linop_s* nufft_op)
{
const auto data = CAST_DOWN(nufft_data, linop_get_data(nufft_op));
PTR_ALLOC(struct nufft_precond_data, pdata);
SET_TYPEID(nufft_precond_data, pdata);
assert(data->conf.toeplitz);
int N = data->N;
int ND = N + 1;
pdata->N = N;
pdata->cim_dims = *TYPE_ALLOC(long[ND]);
pdata->pre_dims = *TYPE_ALLOC(long[ND]);
pdata->cim_strs = *TYPE_ALLOC(long[ND]);
pdata->pre_strs = *TYPE_ALLOC(long[ND]);
md_copy_dims(ND, pdata->cim_dims, data->cim_dims);
md_select_dims(ND, data->flags, pdata->pre_dims, pdata->cim_dims);
md_calc_strides(ND, pdata->cim_strs, pdata->cim_dims, CFL_SIZE);
md_calc_strides(ND, pdata->pre_strs, pdata->pre_dims, CFL_SIZE);
pdata->pre = compute_precond(pdata->N, pdata->pre_dims, pdata->pre_strs, data->psf_dims, data->psf_strs, data->psf, data->linphase);
pdata->fft_op = linop_fft_create(pdata->N, pdata->cim_dims, data->flags);
const long* cim_dims = pdata->cim_dims; // need to dereference pdata before PTR_PASS
return operator_create(N, cim_dims, N, cim_dims, CAST_UP(PTR_PASS(pdata)), nufft_precond_apply, nufft_precond_del);
}
const struct linop_s* linop_zfinitediff_create(unsigned int D, const long dims[D], long diffdim, bool circular)
{
PTR_ALLOC(struct zfinitediff_data, data);
SET_TYPEID(zfinitediff_data, data);
data->D = D;
data->dim_diff = diffdim;
data->do_circdiff = circular;
data->dims_in = *TYPE_ALLOC(long[D]);
data->dims_adj = *TYPE_ALLOC(long[D]);
data->strides_in = *TYPE_ALLOC(long[D]);
data->strides_adj = *TYPE_ALLOC(long[D]);
md_copy_dims(D, data->dims_in, dims);
md_copy_dims(D, data->dims_adj, dims);
md_calc_strides(D, data->strides_in, data->dims_in, CFL_SIZE);
if (!data->do_circdiff)
data->dims_adj[data->dim_diff] -= 1;
md_calc_strides(D, data->strides_adj, data->dims_adj, CFL_SIZE);
const long* dims_adj = data->dims_adj;
const long* dims_in = data->dims_in;
return linop_create(D, dims_adj, D, dims_in, CAST_UP(PTR_PASS(data)),
zfinitediff_apply, zfinitediff_adjoint,
zfinitediff_normal, NULL, zfinitediff_del);
}
/**
* Thresholding operator for l0-norm: f(x) = || x ||_0 <= k, as used in NIHT algorithm.
* y = HT(x, k) (hard thresholding, ie keeping the k largest elements).
*
* @param D number of dimensions
* @param dim dimensions of x
* @param k threshold parameter (non-zero elements to keep)
* @param flags bitmask for joint thresholding
*/
const struct operator_p_s* prox_niht_thresh_create(unsigned int D, const long dim[D], const unsigned int k, const unsigned long flags)
{
PTR_ALLOC(struct thresh_s, data);
SET_TYPEID(thresh_s, data);
data->lambda = 0.;
data->k = k;
data->D = D;
data->flags = flags;
data->unitary_op = NULL;
PTR_ALLOC(long[D], ndim);
md_copy_dims(D, *ndim, dim);
data->dim = *PTR_PASS(ndim);
// norm dimensions are the flagged input dimensions
PTR_ALLOC(long[D], norm_dim);
md_select_dims(D, ~flags, *norm_dim, data->dim);
data->norm_dim = *PTR_PASS(norm_dim);
PTR_ALLOC(long[D], nstr);
md_calc_strides(D, *nstr, data->dim, CFL_SIZE);
data->str = *PTR_PASS(nstr);
return operator_p_create(D, dim, D, dim, CAST_UP(PTR_PASS(data)), hardthresh_apply, thresh_del);
}
static struct linop_s* linop_gdiag_create(unsigned int N, const long dims[N], unsigned int flags, const complex float* diag, bool rdiag)
{
PTR_ALLOC(struct cdiag_s, data);
SET_TYPEID(cdiag_s, data);
data->rmul = rdiag;
data->N = N;
PTR_ALLOC(long[N], dims2);
PTR_ALLOC(long[N], dstrs);
PTR_ALLOC(long[N], strs);
long ddims[N];
md_select_dims(N, flags, ddims, dims);
md_copy_dims(N, *dims2, dims);
md_calc_strides(N, *strs, dims, CFL_SIZE);
md_calc_strides(N, *dstrs, ddims, CFL_SIZE);
data->dims = *PTR_PASS(dims2);
data->strs = *PTR_PASS(strs);
data->dstrs = *PTR_PASS(dstrs);
data->diag = diag; // make a copy?
#ifdef USE_CUDA
data->gpu_diag = NULL;
#endif
return linop_create(N, dims, N, dims, CAST_UP(PTR_PASS(data)), cdiag_apply, cdiag_adjoint, cdiag_normal, NULL, cdiag_free);
}
/*
* Proximal function for real-value constraint
*/
const struct operator_p_s* prox_rvc_create(unsigned int N, const long dims[N])
{
PTR_ALLOC(struct prox_rvc_data, pdata);
SET_TYPEID(prox_rvc_data, pdata);
pdata->size = md_calc_size(N, dims);
return operator_p_create(N, dims, N, dims, CAST_UP(PTR_PASS(pdata)), prox_rvc_apply, prox_rvc_del);
}
static struct linop_s* linop_fft_create_priv(int N, const long dims[N], unsigned int flags, bool forward, bool center)
{
const struct operator_s* plan = fft_measure_create(N, dims, flags, true, false);
const struct operator_s* iplan = fft_measure_create(N, dims, flags, true, true);
PTR_ALLOC(struct fft_linop_s, data);
SET_TYPEID(fft_linop_s, data);
data->frw = plan;
data->adj = iplan;
data->N = N;
data->center = center;
data->dims = *TYPE_ALLOC(long[N]);
md_copy_dims(N, data->dims, dims);
data->strs = *TYPE_ALLOC(long[N]);
md_calc_strides(N, data->strs, data->dims, CFL_SIZE);
long fft_dims[N];
md_select_dims(N, flags, fft_dims, dims);
data->nscale = (float)md_calc_size(N, fft_dims);
lop_fun_t apply = forward ? fft_linop_apply : fft_linop_adjoint;
lop_fun_t adjoint = forward ? fft_linop_adjoint : fft_linop_apply;
struct linop_s* lop = linop_create(N, dims, N, dims, CAST_UP(PTR_PASS(data)), apply, adjoint, fft_linop_normal, NULL, fft_linop_free);
if (center) {
// FIXME: should only allocate flagged dims
complex float* fftmod_mat = md_alloc(N, dims, CFL_SIZE);
complex float* fftmodk_mat = md_alloc(N, dims, CFL_SIZE);
// we need fftmodk only because we want to apply scaling only once
complex float one[1] = { 1. };
md_fill(N, dims, fftmod_mat, one, CFL_SIZE);
fftmod(N, dims, flags, fftmodk_mat, fftmod_mat);
fftscale(N, dims, flags, fftmod_mat, fftmodk_mat);
struct linop_s* mod = linop_cdiag_create(N, dims, ~0u, fftmod_mat);
struct linop_s* modk = linop_cdiag_create(N, dims, ~0u, fftmodk_mat);
struct linop_s* tmp = linop_chain(mod, lop);
tmp = linop_chain(tmp, modk);
linop_free(lop);
linop_free(mod);
linop_free(modk);
lop = tmp;
}
return lop;
}
/**
* Create an Identity linear operator: I x
* @param N number of dimensions
* @param dims dimensions of input (domain)
*/
struct linop_s* linop_identity_create(unsigned int N, const long dims[N])
{
PTR_ALLOC(struct identity_data_s, data);
SET_TYPEID(identity_data_s, data);
data->domain = iovec_create(N, dims, CFL_SIZE);
return linop_create(N, dims, N, dims, CAST_UP(PTR_PASS(data)), identity_apply, identity_apply, identity_apply, NULL, identity_free);
}
const struct operator_p_s* prox_l2norm_create(unsigned int N, const long dims[N], float lambda)
{
PTR_ALLOC(struct prox_l2norm_data, pdata);
SET_TYPEID(prox_l2norm_data, pdata);
pdata->lambda = lambda;
pdata->size = md_calc_size(N, dims) * 2;
return operator_p_create(N, dims, N, dims, CAST_UP(PTR_PASS(pdata)), prox_l2norm_apply, prox_l2norm_del);
}
/**
* Convolution operator
*
* @param N number of dimensions
* @param flags bitmask of the dimensions to apply convolution
* @param ctype
* @param cmode
* @param odims output dimensions
* @param idims input dimensions
* @param kdims kernel dimensions
* @param krn convolution kernel
*/
struct linop_s* linop_conv_create(int N, unsigned int flags, enum conv_type ctype, enum conv_mode cmode, const long odims[N],
const long idims[N], const long kdims[N], const complex float* krn)
{
PTR_ALLOC(struct conv_data_s, data);
SET_TYPEID(conv_data_s, data);
data->plan = conv_plan(N, flags, ctype, cmode, odims, idims, kdims, krn);
return linop_create(N, odims, N, idims, CAST_UP(PTR_PASS(data)), linop_conv_forward, linop_conv_adjoint, NULL, NULL, linop_conv_free);
}
static const struct operator_p_s* prox_ineq_create(unsigned int N, const long dims[N], const complex float* b, bool positive)
{
PTR_ALLOC(struct prox_ineq_data, pdata);
SET_TYPEID(prox_ineq_data, pdata);
pdata->size = md_calc_size(N, dims) * 2;
pdata->b = (const float*)b;
pdata->positive = positive;
return operator_p_create(N, dims, N, dims, CAST_UP(PTR_PASS(pdata)), prox_ineq_apply, prox_ineq_del);
}
const struct operator_p_s* prox_l2ball_create(unsigned int N, const long dims[N], float eps, const complex float* center)
{
PTR_ALLOC(struct prox_l2ball_data, pdata);
SET_TYPEID(prox_l2ball_data, pdata);
pdata->center = (const float*)center;
pdata->eps = eps;
pdata->size = md_calc_size(N, dims) * 2;
return operator_p_create(N, dims, N, dims, CAST_UP(PTR_PASS(pdata)), prox_l2ball_apply, prox_l2ball_del);
}
const struct operator_p_s* prox_leastsquares_create(unsigned int N, const long dims[N], float lambda, const complex float* y)
{
PTR_ALLOC(struct prox_leastsquares_data, pdata);
SET_TYPEID(prox_leastsquares_data, pdata);
pdata->y = (const float*)y;
pdata->lambda = lambda;
pdata->size = md_calc_size(N, dims) * 2;
return operator_p_create(N, dims, N, dims, CAST_UP(PTR_PASS(pdata)), prox_leastsquares_apply, prox_leastsquares_del);
}
/**
* Proximal operator for l1-norm with Wavelet transform: f(x) = lambda || W x ||_1
*
* @param numdims number of dimensions
* @param imSize dimensions of x
* @param wave_flags bitmask for Wavelet transform
* @param minSize minimium size of coarse Wavelet scale
* @param lambda threshold parameter
* @param randshift apply random shift before Wavelet transforming
* @param use_gpu true if using gpu
*/
const struct operator_p_s* prox_wavethresh_create(int numdims, const long imSize[numdims], unsigned int wave_flags, const long minSize[numdims], float lambda, bool randshift, bool use_gpu)
{
PTR_ALLOC(struct wave_prox_s, data);
SET_TYPEID(wave_prox_s, data);
data->plan = prepare_wavelet_plan(numdims, imSize, wave_flags, minSize, use_gpu);
data->plan->randshift = randshift;
data->plan->lambda = lambda;
return operator_p_create(numdims, imSize, numdims, imSize, CAST_UP(PTR_PASS(data)), wavelet_thresh, wavelet_prox_del);
}
struct linop_s* linop_realval_create(unsigned int N, const long dims[N])
{
PTR_ALLOC(struct rvc_s, data);
SET_TYPEID(rvc_s, data);
PTR_ALLOC(long[N], dims2);
md_copy_dims(N, *dims2, dims);
data->N = N;
data->dims = *PTR_PASS(dims2);
return linop_create(N, dims, N, dims, CAST_UP(PTR_PASS(data)), rvc_apply, rvc_apply, rvc_apply, NULL, rvc_free);
}
/**
* Wavelet CFD9/7 transform operator
*
* @param N number of dimensions
* @param dims dimensions of input
* @param flags bitmask of the dimensions to apply the Fourier transform
*/
struct linop_s* linop_cdf97_create(int N, const long dims[N], unsigned int flags)
{
PTR_ALLOC(struct linop_cdf97_s, data);
SET_TYPEID(linop_cdf97_s, data);
PTR_ALLOC(long[N], ndims);
md_copy_dims(N, *ndims, dims);
data->N = N;
data->dims = *ndims;
data->flags = flags;
return linop_create(N, dims, N, dims, CAST_UP(PTR_PASS(data)), linop_cdf97_apply, linop_cdf97_adjoint, linop_cdf97_normal, NULL, linop_cdf97_free);
}
/**
* Create a resize linear operator: y = M x,
* where M either crops or expands the the input dimensions to match the output dimensions.
* Uses centered zero-padding and centered cropping
*
* @param N number of dimensions
* @param out_dims output dimensions
* @param in_dims input dimensions
*/
struct linop_s* linop_resize_create(unsigned int N, const long out_dims[N], const long in_dims[N])
{
PTR_ALLOC(struct resize_op_s, data);
SET_TYPEID(resize_op_s, data);
data->N = N;
data->out_dims = *TYPE_ALLOC(long[N]);
data->in_dims = *TYPE_ALLOC(long[N]);
md_copy_dims(N, (long*)data->out_dims, out_dims);
md_copy_dims(N, (long*)data->in_dims, in_dims);
return linop_create(N, out_dims, N, in_dims, CAST_UP(PTR_PASS(data)), resize_forward, resize_adjoint, resize_normal, NULL, resize_free);
}
/**
* Wavelet linear operator
*
* @param numdims number of dimensions
* @param imSize dimensions of x
* @param wave_flags bitmask for Wavelet transform
* @param minSize minimium size of coarse Wavelet scale
* @param randshift apply random shift before Wavelet transforming
* @param use_gpu true if using gpu
*/
const struct linop_s* wavelet_create(int numdims, const long imSize[numdims], unsigned int wave_flags, const long minSize[numdims], bool randshift, bool use_gpu)
{
PTR_ALLOC(struct wavelet_data_s, data);
SET_TYPEID(wavelet_data_s, data);
data->plan = prepare_wavelet_plan(numdims, imSize, wave_flags, minSize, use_gpu);
data->plan->randshift = randshift;
long coeff_dims[numdims];
md_select_dims(numdims, ~wave_flags, coeff_dims, imSize);
coeff_dims[0] = data->plan->numCoeff_tr;
coeff_dims[1] = 1;
coeff_dims[2] = 1;
return linop_create(numdims, coeff_dims, numdims, imSize, CAST_UP(PTR_PASS(data)), wavelet_forward, wavelet_inverse, wavelet_normal, NULL, wavelet_del);
}
struct linop_s* linop_grad_create(long N, const long dims[N], unsigned int flags)
{
PTR_ALLOC(struct grad_s, data);
SET_TYPEID(grad_s, data);
long dims2[N + 1];
grad_dims(N, dims2, flags, dims);
data->N = N + 1;
data->flags = flags;
data->dims = *TYPE_ALLOC(long[N + 1]);
md_copy_dims(N + 1, data->dims, dims2);
return linop_create(N + 1, dims2, N, dims, CAST_UP(PTR_PASS(data)), grad_op_apply, grad_op_adjoint, grad_op_normal, NULL, grad_op_free);
}
struct linop_s* linop_sampling_create(const long dims[DIMS], const long pat_dims[DIMS], const complex float* pattern)
{
PTR_ALLOC(struct sampling_data_s, data);
SET_TYPEID(sampling_data_s, data);
md_copy_dims(DIMS, data->pat_dims, pat_dims);
md_select_dims(DIMS, ~MAPS_FLAG, data->dims, dims); // dimensions of kspace
md_calc_strides(DIMS, data->strs, data->dims, CFL_SIZE);
md_calc_strides(DIMS, data->pat_strs, data->pat_dims, CFL_SIZE);
data->pattern = (complex float*)pattern;
#ifdef USE_CUDA
data->gpu_pattern = NULL;
#endif
const long* dims2 = data->dims;
return linop_create(DIMS, dims2, DIMS, dims2, CAST_UP(PTR_PASS(data)), sampling_apply, sampling_apply, sampling_apply, NULL, sampling_free);
}
/**
* Initialize finite difference operator
*
* @param D number of dimensions
* @param dim input dimensions
* @param flags bitmask for applying operator
* @param snip true: clear initial entry (i.c.); false: keep initial entry (i.c.)
*
* Returns a pointer to the finite difference operator
*/
extern const struct linop_s* linop_finitediff_create(unsigned int D, const long dim[D], const unsigned long flags, bool snip)
{
PTR_ALLOC(struct fdiff_s, data);
SET_TYPEID(fdiff_s, data);
data->D = D;
data->flags = flags;
data->order = 1;
data->snip = snip;
data->dims = *TYPE_ALLOC(long[D]);
md_copy_dims(D, data->dims, dim);
data->str = *TYPE_ALLOC(long[D]);
md_calc_strides(D, data->str, data->dims, CFL_SIZE);
return linop_create(D, dim, D, dim, CAST_UP(PTR_PASS(data)), fdiff_apply, fdiff_apply_adjoint, NULL, cumsum_apply, finite_diff_del);
}
const struct linop_s* linop_fmac_create(unsigned int N, const long dims[N],
unsigned int oflags, unsigned int iflags, unsigned int tflags, const complex float* tensor)
{
PTR_ALLOC(struct fmac_data, data);
SET_TYPEID(fmac_data, data);
data->N = N;
data->dims = *TYPE_ALLOC(long[N]);
md_copy_dims(N, data->dims, dims);
data->idims = *TYPE_ALLOC(long[N]);
data->istrs = *TYPE_ALLOC(long[N]);
md_select_dims(N, ~iflags, data->idims, dims);
md_calc_strides(N, data->istrs, data->idims, CFL_SIZE);
data->odims = *TYPE_ALLOC(long[N]);
data->ostrs = *TYPE_ALLOC(long[N]);
md_select_dims(N, ~oflags, data->odims, dims);
md_calc_strides(N, data->ostrs, data->odims, CFL_SIZE);
data->tstrs = *TYPE_ALLOC(long[N]);
data->tdims = *TYPE_ALLOC(long[N]);
md_select_dims(N, ~tflags, data->tdims, dims);
md_calc_strides(N, data->tstrs, data->tdims, CFL_SIZE);
data->tensor = tensor;
#ifdef USE_CUDA
data->gpu_tensor = NULL;
#endif
long odims[N];
md_copy_dims(N, odims, data->odims);
long idims[N];
md_copy_dims(N, idims, data->idims);
return linop_create(N, odims, N, idims,
CAST_UP(PTR_PASS(data)), fmac_apply, fmac_adjoint, NULL,
NULL, fmac_free_data);
}
const struct operator_s* fft_measure_create(unsigned int D, const long dimensions[D], unsigned long flags, bool inplace, bool backwards)
{
PTR_ALLOC(struct fft_plan_s, plan);
SET_TYPEID(fft_plan_s, plan);
complex float* src = md_alloc(D, dimensions, CFL_SIZE);
complex float* dst = inplace ? src : md_alloc(D, dimensions, CFL_SIZE);
long strides[D];
md_calc_strides(D, strides, dimensions, CFL_SIZE);
plan->fftw = fft_fftwf_plan(D, dimensions, flags, strides, dst, strides, src, backwards, true);
md_free(src);
if (!inplace)
md_free(dst);
#ifdef USE_CUDA
plan->cuplan = NULL;
#ifndef LAZY_CUDA
if (cuda_ondevice(src))
plan->cuplan = fft_cuda_plan(D, dimensions, flags, strides, strides, backwards);
#else
plan->D = D;
plan->flags = flags;
plan->backwards = backwards;
PTR_ALLOC(long[D], dims);
md_copy_dims(D, *dims, dimensions);
plan->dims = *PTR_PASS(dims);
PTR_ALLOC(long[D], istrs);
md_copy_strides(D, *istrs, strides);
plan->istrs = *PTR_PASS(istrs);
PTR_ALLOC(long[D], ostrs);
md_copy_strides(D, *ostrs, strides);
plan->ostrs = *PTR_PASS(ostrs);
#endif
#endif
return operator_create2(D, dimensions, strides, D, dimensions, strides, CAST_UP(PTR_PASS(plan)), fft_apply, fft_free_plan);
}
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);
}
/**
* Proximal operator for l1-norm with Wavelet transform: f(x) = lambda || W x ||_1
*
* @param N number of dimensions
* @param dims dimensions of x
* @param flags bitmask for Wavelet transform
* @param minsize minimium size of coarse Wavelet scale
* @param lambda threshold parameter
* @param randshift random shifting
*/
const struct operator_p_s* prox_wavelet3_thresh_create(unsigned int N, const long dims[N], unsigned int flags, const long minsize[N], float lambda, bool randshift)
{
PTR_ALLOC(struct wavelet3_thresh_s, data);
SET_TYPEID(wavelet3_thresh_s, data);
data->N = N;
long (*ndims)[N] = TYPE_ALLOC(long[N]);
md_copy_dims(N, (*ndims), dims);
data->dims = *ndims;
long (*nminsize)[N] = TYPE_ALLOC(long[N]);
md_copy_dims(N, (*nminsize), minsize);
data->minsize = *nminsize;
data->flags = flags;
data->lambda = lambda;
data->randshift = randshift;
data->rand_state = 1;
return operator_p_create(N, dims, N, dims, CAST_UP(PTR_PASS(data)), wavelet3_thresh_apply, wavelet3_thresh_del);
}
/**
* Operator interface for a true matrix:
* out = mat * in
* in: [x x x x 1 x x K x x]
* mat: [x x x x T x x K x x]
* out: [x x x x T x x 1 x x]
* where the x's are arbitrary dimensions and T and K may be transposed
*
* use this interface if K == 1 or T == 1
*
* @param N number of dimensions
* @param out_dims output dimensions after applying the matrix (codomain)
* @param in_dims input dimensions to apply the matrix (domain)
* @param T_dim dimension corresponding to the rows of A
* @param K_dim dimension corresponding to the columns of A
* @param matrix matrix data
*/
struct linop_s* linop_matrix_altcreate(unsigned int N, const long out_dims[N], const long in_dims[N], const unsigned int T_dim, const unsigned int K_dim, const complex float* matrix)
{
long matrix_dims[N];
md_singleton_dims(N, matrix_dims);
matrix_dims[K_dim] = in_dims[K_dim];
matrix_dims[T_dim] = out_dims[T_dim];
unsigned int T = out_dims[T_dim];
unsigned int K = in_dims[K_dim];
PTR_ALLOC(long[N], max_dims);
for (unsigned int i = 0; i < N; i++) {
if ((in_dims[i] > 1) && (out_dims[i] == 1)) {
(*max_dims)[i] = in_dims[i];
}
else if ((in_dims[i] == 1) && (out_dims[i] > 1)) {
(*max_dims)[i] = out_dims[i];
}
else {
assert(in_dims[i] == out_dims[i]);
(*max_dims)[i] = in_dims[i];
}
}
complex float* mat = md_alloc_sameplace(N, matrix_dims, CFL_SIZE, matrix);
complex float* matc = md_alloc_sameplace(N, matrix_dims, CFL_SIZE, matrix);
md_copy(N, matrix_dims, mat, matrix, CFL_SIZE);
md_zconj(N, matrix_dims, matc, mat);
complex float* gram = NULL;
const struct iovec_s* gram_iovec = compute_gram_matrix(N, T_dim, T, K_dim, K, &gram, matrix_dims, matrix);
PTR_ALLOC(struct operator_matrix_s, data);
SET_TYPEID(operator_matrix_s, data);
data->mat_iovec = iovec_create(N, matrix_dims, CFL_SIZE);
data->mat_gram_iovec = gram_iovec;
data->max_dims = *max_dims;
data->mat = mat;
data->mat_conj = matc;
data->mat_gram = gram;
data->K_dim = K_dim;
data->T_dim = T_dim;
data->K = K;
data->T = T;
data->domain_iovec = iovec_create(N, in_dims, CFL_SIZE);
data->codomain_iovec = iovec_create(N, out_dims, CFL_SIZE);
return linop_create(N, out_dims, N, in_dims, CAST_UP(PTR_PASS(data)), linop_matrix_apply, linop_matrix_apply_adjoint, linop_matrix_apply_normal, NULL, linop_matrix_del);
}
/* 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);
}
