Function csr2coo

src/backend/opencl/kernel/sparse.hpp:166–202

template<typename T>
void csr2coo(Param ovalues, Param orowIdx, Param ocolIdx, const Param ivalues,
             const Param irowIdx, const Param icolIdx, Param index) {
    std::vector<TemplateArg> tmpltArgs = {
        TemplateTypename<T>(),
    };
    std::vector<std::string> compileOpts = {
        DefineKeyValue(T, dtype_traits<T>::getName()),
    };
    compileOpts.emplace_back(getTypeBuildDefinition<T>());

    auto csr2coo = common::getKernel("csr2Coo", {{csr2coo_cl_src}}, tmpltArgs,
                                     compileOpts);

    const int MAX_GROUPS = 4096;
    int M                = irowIdx.info.dims[0] - 1;
    // FIXME: This needs to be based non nonzeros per row
    int threads = 64;

    cl::Buffer *scratch = bufferAlloc(orowIdx.info.dims[0] * sizeof(int));

    cl::NDRange local(threads, 1);
    int groups_x = std::min((int)(divup(M, local[0])), MAX_GROUPS);
    cl::NDRange global(local[0] * groups_x, 1);

    csr2coo(cl::EnqueueArgs(getQueue(), global, local), *scratch, *ocolIdx.data,
            *irowIdx.data, *icolIdx.data, M);

    // Now we need to sort this into column major
    kernel::sort0ByKeyIterative<int, int>(ocolIdx, index, true);

    // Now use index to sort values and rows
    kernel::swapIndex<T>(ovalues, orowIdx, ivalues, scratch, index);

    CL_DEBUG_FINISH(getQueue());

    bufferFree(scratch);
}

template<typename T>
void coo2csr(Param ovalues, Param orowIdx, Param ocolIdx, const Param ivalues,