Function block_reduce_sum

src/backend/cuda/kernel/orb.hpp:50–65

}

inline __device__ float block_reduce_sum(float val) {
    __shared__ float data[THREADS_X * THREADS_Y];

    unsigned idx = threadIdx.x * blockDim.x + threadIdx.y;

    data[idx] = val;
    __syncthreads();

    for (unsigned i = blockDim.y / 2; i > 0; i >>= 1) {
        if (threadIdx.y < i) { data[idx] += data[idx + i]; }

        __syncthreads();
    }

    return data[threadIdx.x * blockDim.x];
}

template<typename T>
__global__ void keep_features(float* x_out, float* y_out, float* score_out,