Module ml::sparse

sparse tensors (CSR) with autograd and devices.

A SparseTensor is a rows x cols matrix that stores only its nonzeros, in compressed sparse row form: for row r, entries rowPtr[r] .. rowPtr[r+1]-1 of colIdx / values. The values are an ordinary FloatTensor of length nnz, so a sparse matrix can be a trained parameter (a sparse layer, a graph's learned edge weights): gradients reach the values and nothing else -- the sparsity pattern is fixed. The index arrays are float32 tensors too (indices exact up to 2^24), so the whole matrix moves to a device like any tensor and its ops run there.

struct SparseTensor A = SparseTensor::from_dense(adj, (float)0.0, 1);   // learned edge weights
&FloatTensor H = tensor_spmm(&A, X);           // [rows, n] = A [rows, cols] . X [cols, n]
... loss ... loss.backward();                   // A.values.grad, X.grad
A.place(DEVICE_ROCM);                          // the values and indices on the GPU

Conforms to trait Tensor (element access, fill, backward, zero_grad, free) and trait DeviceTensor (device, place, sync), so generic code written against those traits takes it next to a DenseTensor.