static int compute_mask_matrix(cl_mem cl_mask_matrix, int step_x, int step_y)
{
int i, j, ret = 0;
uint32_t *mask_matrix, *mask_x, *mask_y;
size_t size_matrix = sizeof(uint32_t) * (2 * step_x + 1) * (2 * step_y + 1);
mask_x = av_mallocz(sizeof(uint32_t) * (2 * step_x + 1));
if (!mask_x) {
ret = AVERROR(ENOMEM);
goto end;
}
mask_y = av_mallocz(sizeof(uint32_t) * (2 * step_y + 1));
if (!mask_y) {
ret = AVERROR(ENOMEM);
goto end;
}
mask_matrix = av_mallocz(size_matrix);
if (!mask_matrix) {
ret = AVERROR(ENOMEM);
goto end;
}
ret = compute_mask(step_x, mask_x);
if (ret < 0)
goto end;
ret = compute_mask(step_y, mask_y);
if (ret < 0)
goto end;
for (j = 0; j < 2 * step_y + 1; j++) {
for (i = 0; i < 2 * step_x + 1; i++) {
mask_matrix[i + j * (2 * step_x + 1)] = mask_y[j] * mask_x[i];
}
}
ret = av_opencl_buffer_write(cl_mask_matrix, (uint8_t *)mask_matrix, size_matrix);
end:
av_freep(&mask_x);
av_freep(&mask_y);
av_freep(&mask_matrix);
return ret;
}
hash_t *hash_init(hash_t *hash, hashcount_t maxcount,
hash_comp_t compfun, hash_fun_t hashfun, hnode_t **table,
hashcount_t nchains)
{
if (hash_val_t_bit == 0) /* 1 */
compute_bits();
assert (is_power_of_two(nchains));
hash->table = table; /* 2 */
hash->nchains = nchains;
hash->nodecount = 0;
hash->maxcount = maxcount;
hash->compare = compfun ? compfun : hash_comp_default;
hash->function = hashfun ? hashfun : hash_fun_default;
hash->dynamic = 0; /* 3 */
hash->mask = compute_mask(nchains); /* 4 */
clear_table(hash); /* 5 */
assert (hash_verify(hash));
return hash;
}
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);
}
// Caller is responsible to provide enough storage space in res
// Select bit range [start, end]
// NOTE: Bit indices are, as usual, zero-based
void get_bit_range(I start, I end, C *res) {
#ifdef EXPENSIVE_SANITY_CHECKS
if (end > n || start > n || start >= end) {
// TODO: Throw an exception instead
std::cerr << "Internal error: Bit index bogous" << std::endl;
std::cerr << "(start=" << start << ", end=" << end
<< ", n=" << n << ")" << std::endl;
std::exit(-1);
}
#endif
memset(res, 0, ceil(((double)(end - start + 1))/8));
// Compute chain elements and bit within this element for the
// start and end positions
I start_chunk = start/bits_per_type;
I end_chunk = end/bits_per_type;
unsigned short start_bit = start % bits_per_type;
unsigned short end_bit;
I dest_chunk = 0;
C chunk;
C mask;
if (start_chunk == end_chunk)
end_bit = end % bits_per_type;
else
end_bit = bits_per_type - 1;
mask = compute_mask(start_bit, end_bit);
res[dest_chunk] = (data[start_chunk] & mask) >> start_bit;
if (start_chunk == end_chunk)
return;
I dest_start_bit = end_bit - start_bit + 1;
if (dest_start_bit == bits_per_type) {
dest_chunk++;
dest_start_bit = 0;
}
I dest_bits;
for (I curr_chunk = start_chunk+1; curr_chunk < end_chunk; curr_chunk++) {
// For the inner chunks, we can always select the full chunk from
// the input data, but need to split it across the output
// data field
chunk = data[curr_chunk];
// How many bits remain in the destination chunk
dest_bits = bits_per_type - dest_start_bit;
// Fill up the current destination chunk
mask = compute_mask(0, dest_bits - 1);
res[dest_chunk] |= ((chunk & mask) << dest_start_bit);
dest_chunk++;
// ... and fill the next destination chunk as far as
// possible unless the previous chunk was completely
// drained and there is nothing left for the new chunk
if (dest_bits != bits_per_type) {
mask = compute_mask(dest_bits, bits_per_type - 1);
res[dest_chunk] = (chunk & mask) >> dest_bits;
}
// Compute new starting position in the destination chunk
dest_start_bit = bits_per_type - dest_bits;
}
// ... and fill the next destination chunk as far as
// possible unless the previous chunk was completely
// drained and there is nothing left for the new chunk
if (dest_bits != bits_per_type) {
mask = compute_mask(dest_bits, bits_per_type - 1);
res[dest_chunk] = (chunk & mask) >> dest_bits;
}
// Compute new starting position in the destination chunk
dest_start_bit = bits_per_type - dest_bits;
}
end_bit = end % bits_per_type;
dest_bits = bits_per_type - dest_start_bit;
mask = compute_mask(0, end_bit);
chunk = data[end_chunk] & mask;
#ifdef EXPENSIVE_SANITY_CHECKS
if (debug_level >= EXCESSIVE_INFO) {
std::cerr << "end_bit: " << end_bit << ", dest_bits: " << dest_bits
<< ", mask: " << std::bitset<32>(mask) << ", end_chunk: "
<< end_chunk << std::endl;
std::cerr << "data[end_chunk]: " << std::bitset<32>(data[end_chunk])
<< std::endl;
std::cerr << "masked data: " << std::bitset<32>(chunk)
<< std::endl;
std::cerr << "dest_start_bit: " << dest_start_bit << std::endl;
}
#endif
// Any excess bits that do not fit into the current result chunk