Function fast

src/backend/cuda/kernel/fast.hpp:323–443

template<typename T>
void fast(unsigned *out_feat, float **x_out, float **y_out, float **score_out,
          const Array<T> &in, const float thr, const unsigned arc_length,
          const unsigned nonmax, const float feature_ratio, const unsigned edge,
          const LookupTable1D<unsigned char> &luTable) {
    dim4 indims             = in.dims();
    const unsigned max_feat = ceil(indims[0] * indims[1] * feature_ratio);

    dim3 threads(16, 16);
    dim3 blocks(divup(indims[0] - edge * 2, threads.x),
                divup(indims[1] - edge * 2, threads.y));

    // Matrix containing scores for detected features, scores are stored in the
    // same coordinates as features, dimensions should be equal to in.
    auto d_score = memAlloc<float>(indims[0] * indims[1] + 1);

    float *d_flags = d_score.get();
    uptr<float> d_flags_alloc;
    if (nonmax) {
        d_flags_alloc = memAlloc<float>(indims[0] * indims[1]);
        d_flags       = d_flags_alloc.get();
    }

    // Shared memory size
    size_t shared_size = (threads.x + 6) * (threads.y + 6) * sizeof(T);

    switch (arc_length) {
        case 9:
            CUDA_LAUNCH_SMEM((locate_features<T, 9>), blocks, threads,
                             shared_size, in, d_score.get(), thr, edge,
                             luTable.get());
            break;
        case 10:
            CUDA_LAUNCH_SMEM((locate_features<T, 10>), blocks, threads,
                             shared_size, in, d_score.get(), thr, edge,
                             luTable.get());
            break;
        case 11:
            CUDA_LAUNCH_SMEM((locate_features<T, 11>), blocks, threads,
                             shared_size, in, d_score.get(), thr, edge,
                             luTable.get());
            break;
        case 12:
            CUDA_LAUNCH_SMEM((locate_features<T, 12>), blocks, threads,
                             shared_size, in, d_score.get(), thr, edge,
                             luTable.get());
            break;
        case 13:
            CUDA_LAUNCH_SMEM((locate_features<T, 13>), blocks, threads,
                             shared_size, in, d_score.get(), thr, edge,
                             luTable.get());
            break;
        case 14:
            CUDA_LAUNCH_SMEM((locate_features<T, 14>), blocks, threads,
                             shared_size, in, d_score.get(), thr, edge,
                             luTable.get());
            break;
        case 15:
            CUDA_LAUNCH_SMEM((locate_features<T, 15>), blocks, threads,