void sort0_index(Param<T> val, Param<unsigned> idx)
{
thrust::device_ptr<T> val_ptr = thrust::device_pointer_cast(val.ptr);
thrust::device_ptr<unsigned> idx_ptr = thrust::device_pointer_cast(idx.ptr);
for(int w = 0; w < val.dims[3]; w++) {
int valW = w * val.strides[3];
int idxW = w * idx.strides[3];
for(int z = 0; z < val.dims[2]; z++) {
int valWZ = valW + z * val.strides[2];
int idxWZ = idxW + z * idx.strides[2];
for(int y = 0; y < val.dims[1]; y++) {
int valOffset = valWZ + y * val.strides[1];
int idxOffset = idxWZ + y * idx.strides[1];
THRUST_SELECT(thrust::sequence, idx_ptr + idxOffset, idx_ptr + idxOffset + idx.dims[0]);
if(isAscending) {
THRUST_SELECT(thrust::sort_by_key,
val_ptr + valOffset, val_ptr + valOffset + val.dims[0],
idx_ptr + idxOffset);
} else {
THRUST_SELECT(thrust::sort_by_key,
val_ptr + valOffset, val_ptr + valOffset + val.dims[0],
idx_ptr + idxOffset, thrust::greater<T>());
}
}
}
}
POST_LAUNCH_CHECK();
}
void regions(cuda::Param<T> out, cuda::CParam<char> in, cudaTextureObject_t tex)
{
const dim3 threads(THREADS_X, THREADS_Y);
const int blk_x = divup(in.dims[0], threads.x*2);
const int blk_y = divup(in.dims[1], threads.y*2);
const dim3 blocks(blk_x, blk_y);
CUDA_LAUNCH((initial_label<T,n_per_thread>), blocks, threads, out, in);
POST_LAUNCH_CHECK();
int h_continue = 1;
while (h_continue) {
h_continue = 0;
CUDA_CHECK(cudaMemcpyToSymbolAsync(continue_flag, &h_continue, sizeof(int),
0, cudaMemcpyHostToDevice,
cuda::getActiveStream()));
CUDA_LAUNCH((update_equiv<T, 16, n_per_thread, full_conn>), blocks, threads, out, tex);
POST_LAUNCH_CHECK();
CUDA_CHECK(cudaMemcpyFromSymbolAsync(&h_continue, continue_flag, sizeof(int),
0, cudaMemcpyDeviceToHost,
cuda::getActiveStream()));
CUDA_CHECK(cudaStreamSynchronize(cuda::getActiveStream()));
}
// Now, perform the final relabeling. This converts the equivalency
// map from having unique labels based on the lowest pixel in the
// component to being sequentially numbered components starting at
// 1.
int size = in.dims[0] * in.dims[1];
auto tmp = cuda::memAlloc<T>(size);
CUDA_CHECK(cudaMemcpyAsync(tmp.get(), out.ptr, size * sizeof(T),
cudaMemcpyDeviceToDevice,
cuda::getActiveStream()));
// Wrap raw device ptr
thrust::device_ptr<T> wrapped_tmp = thrust::device_pointer_cast(tmp.get());
// Sort the copy
THRUST_SELECT(thrust::sort, wrapped_tmp, wrapped_tmp + size);
// Take the max element which is the number
// of label assignments to compute.
const int num_bins = wrapped_tmp[size - 1] + 1;
// If the number of label assignments is two,
// then either the entire input image is one big
// component(1's) or it has only one component other than
// background(0's). Either way, no further
// post-processing of labels is required.
if (num_bins<=2)
return;
cuda::ThrustVector<T> labels(num_bins);
// Find the end of each section of values
thrust::counting_iterator<T> search_begin(0);
THRUST_SELECT(thrust::upper_bound, wrapped_tmp, wrapped_tmp + size,
search_begin, search_begin + num_bins,
labels.begin());
THRUST_SELECT(thrust::adjacent_difference, labels.begin(), labels.end(), labels.begin());
// Operators for the scan
clamp_to_one<T> clamp;
thrust::plus<T> add;
// Perform scan -- this computes the correct labels for each component
THRUST_SELECT(thrust::transform_exclusive_scan,
labels.begin(),
labels.end(),
labels.begin(),
clamp,
0,
add);
// Apply the correct labels to the equivalency map
CUDA_LAUNCH((final_relabel<T,n_per_thread>), blocks,threads,
out, in, thrust::raw_pointer_cast(&labels[0]));
POST_LAUNCH_CHECK();
}