array constant(cfloat val, const dim4 &dims)
{
af_array res;
AF_THROW(af_constant_complex(&res, real(val), imag(val),
dims.ndims(), dims.get(), c32));
return array(res);
}
array
constant(T val, const dim4 &dims, const af::dtype type)
{
af_array res;
if (type != s64 && type != u64) {
AF_THROW(af_constant(&res, (double)val,
dims.ndims(), dims.get(), type));
}
else if (type == s64) {
AF_THROW(af_constant_long (&res, ( intl)val,
dims.ndims(),
dims.get()));
} else {
AF_THROW(af_constant_ulong(&res, (uintl)val,
dims.ndims(),
dims.get()));
}
return array(res);
}
dim4 calcStrides(const dim4 &parentDim)
{
dim4 out(1, 1, 1, 1);
dim_t *out_dims = out.get();
const dim_t *parent_dims = parentDim.get();
for (dim_t i=1; i < 4; i++) {
out_dims[i] = out_dims[i - 1] * parent_dims[i-1];
}
return out;
}
AFAPI array constant(cdouble val, const dim4 &dims, const af::dtype type)
{
if (type != c32 && type != c64) {
return constant(real(val), dims, type);
}
af_array res;
AF_THROW(af_constant_complex(&res,
real(val),
imag(val),
dims.ndims(),
dims.get(), type));
return array(res);
}
void fast_pyramid(std::vector<unsigned>& feat_pyr,
std::vector<float*>& d_x_pyr,
std::vector<float*>& d_y_pyr,
std::vector<unsigned>& lvl_best,
std::vector<float>& lvl_scl,
std::vector<CParam<T> >& img_pyr,
CParam<T> in,
const float fast_thr,
const unsigned max_feat,
const float scl_fctr,
const unsigned levels,
const unsigned patch_size)
{
unsigned min_side = std::min(in.dims[0], in.dims[1]);
unsigned max_levels = 0;
float scl_sum = 0.f;
for (unsigned i = 0; i < levels; i++) {
min_side /= scl_fctr;
// Minimum image side for a descriptor to be computed
if (min_side < patch_size || max_levels == levels) break;
max_levels++;
scl_sum += 1.f / (float)std::pow(scl_fctr,(float)i);
}
// Compute number of features to keep for each level
lvl_best.resize(max_levels);
lvl_scl.resize(max_levels);
unsigned feat_sum = 0;
for (unsigned i = 0; i < max_levels-1; i++) {
float scl = (float)std::pow(scl_fctr,(float)i);
lvl_scl[i] = scl;
lvl_best[i] = ceil((max_feat / scl_sum) / lvl_scl[i]);
feat_sum += lvl_best[i];
}
lvl_scl[max_levels-1] = (float)std::pow(scl_fctr,(float)max_levels-1);
lvl_best[max_levels-1] = max_feat - feat_sum;
// Hold multi-scale image pyramids
static const dim4 dims0;
static const CParam<T> emptyCParam(NULL, dims0.get(), dims0.get());
// Need to do this as CParam does not have a default constructor
// And resize needs a default constructor or default value prior to C++11
img_pyr.resize(max_levels, emptyCParam);
// Create multi-scale image pyramid
for (unsigned i = 0; i < max_levels; i++) {
if (i == 0) {
// First level is used in its original size
img_pyr[i].ptr = in.ptr;
for (int k = 0; k < 4; k++) {
img_pyr[i].dims[k] = in.dims[k];
img_pyr[i].strides[k] = in.strides[k];
}
}
else {
// Resize previous level image to current level dimensions
Param<T> lvl_img;
lvl_img.dims[0] = round(in.dims[0] / lvl_scl[i]);
lvl_img.dims[1] = round(in.dims[1] / lvl_scl[i]);
lvl_img.strides[0] = 1;
lvl_img.strides[1] = lvl_img.dims[0] * lvl_img.strides[0];
for (int k = 2; k < 4; k++) {
lvl_img.dims[k] = 1;
lvl_img.strides[k] = lvl_img.dims[k - 1] * lvl_img.strides[k - 1];
}
int lvl_elem = lvl_img.strides[3] * lvl_img.dims[3];
lvl_img.ptr = memAlloc<T>(lvl_elem);
resize<T, AF_INTERP_BILINEAR>(lvl_img, img_pyr[i-1]);
img_pyr[i].ptr = lvl_img.ptr;
for (int k = 0; k < 4; k++) {
img_pyr[i].dims[k] = lvl_img.dims[k];
img_pyr[i].strides[k] = lvl_img.strides[k];
}
}
}
feat_pyr.resize(max_levels);
d_x_pyr.resize(max_levels);
d_y_pyr.resize(max_levels);
for (unsigned i = 0; i < max_levels; i++) {
unsigned lvl_feat = 0;
float* d_x_feat = NULL;
float* d_y_feat = NULL;
float* d_score_feat = NULL;
// Round feature size to nearest odd integer
float size = 2.f * floor(patch_size / 2.f) + 1.f;
// Avoid keeping features that are too wide and might not fit the image,
// sqrt(2.f) is the radius when angle is 45 degrees and represents
// widest case possible
unsigned edge = ceil(size * sqrt(2.f) / 2.f);
// Detects FAST features
fast(&lvl_feat, &d_x_feat, &d_y_feat, &d_score_feat,
img_pyr[i], fast_thr, 9, 1, 0.15f, edge);
// FAST score is not used
memFree(d_score_feat);
if (lvl_feat == 0) {
feat_pyr[i] = 0;
d_x_pyr[i] = NULL;
d_x_pyr[i] = NULL;
}
else {
feat_pyr[i] = lvl_feat;
d_x_pyr[i] = d_x_feat;
d_y_pyr[i] = d_y_feat;
}
}
}
// Assign values to an array
array::array_proxy&
af::array::array_proxy::operator=(const array &other)
{
unsigned nd = numDims(impl->parent_->get());
const dim4 this_dims = getDims(impl->parent_->get());
const dim4 other_dims = other.dims();
int dim = gforDim(impl->indices_);
af_array other_arr = other.get();
bool batch_assign = false;
bool is_reordered = false;
if (dim >= 0) {
//FIXME: Figure out a faster, cleaner way to do this
dim4 out_dims = seqToDims(impl->indices_, this_dims, false);
batch_assign = true;
for (int i = 0; i < AF_MAX_DIMS; i++) {
if (this->impl->indices_[i].isBatch) batch_assign &= (other_dims[i] == 1);
else batch_assign &= (other_dims[i] == out_dims[i]);
}
if (batch_assign) {
af_array out;
AF_THROW(af_tile(&out, other_arr,
out_dims[0] / other_dims[0],
out_dims[1] / other_dims[1],
out_dims[2] / other_dims[2],
out_dims[3] / other_dims[3]));
other_arr = out;
} else if (out_dims != other_dims) {
// HACK: This is a quick check to see if other has been reordered inside gfor
// TODO: Figure out if this breaks and implement a cleaner method
other_arr = gforReorder(other_arr, dim);
is_reordered = true;
}
}
af_array par_arr = 0;
if (impl->is_linear_) {
AF_THROW(af_flat(&par_arr, impl->parent_->get()));
nd = 1;
} else {
par_arr = impl->parent_->get();
}
af_array tmp = 0;
AF_THROW(af_assign_gen(&tmp, par_arr, nd, impl->indices_, other_arr));
af_array res = 0;
if (impl->is_linear_) {
AF_THROW(af_moddims(&res, tmp, this_dims.ndims(), this_dims.get()));
AF_THROW(af_release_array(par_arr));
AF_THROW(af_release_array(tmp));
} else {
res = tmp;
}
impl->parent_->set(res);
if (dim >= 0 && (is_reordered || batch_assign)) {
if (other_arr) AF_THROW(af_release_array(other_arr));
}
return *this;
}
array::array(const array& input, const dim4& dims) : arr(0)
{
AF_THROW(af_moddims(&arr, input.get(), AF_MAX_DIMS, dims.get()));
}
array moddims(const array& in, const dim4& dims)
{
return af::moddims(in, dims.ndims(), dims.get());
}
array identity(const dim4 &dims, const af::dtype type)
{
af_array res;
AF_THROW(af_identity(&res, dims.ndims(), dims.get(), type));
return array(res);
}
array iota(const dim4 &dims, const dim4 &tile_dims, const af::dtype ty)
{
af_array out;
AF_THROW(af_iota(&out, dims.ndims(), dims.get(), tile_dims.ndims(), tile_dims.get(), ty));
return array(out);
}
array range(const dim4 &dims, const int seq_dim, const af::dtype ty)
{
af_array out;
AF_THROW(af_range(&out, dims.ndims(), dims.get(), seq_dim, ty));
return array(out);
}
array randn(const dim4 &dims, const af::dtype type)
{
af_array res;
AF_THROW(af_randn(&res, dims.ndims(), dims.get(), type));
return array(res);
}
array iota(const dim4 &dims, const unsigned rep, af_dtype ty)
{
af_array out;
AF_THROW(af_iota(&out, dims.ndims(), dims.get(), rep, ty));
return array(out);
}
array constant(double val, const dim4 &dims, af_dtype type)
{
af_array res;
AF_THROW(af_constant(&res, val, dims.ndims(), dims.get(), type));
return array(res);
}
array randn(const dim4 &dims, const dtype ty, randomEngine &r)
{
af_array out;
AF_THROW(af_random_normal(&out, dims.ndims(), dims.get(), ty, r.get()));
return array(out);
}