zheng-da commented on a change in pull request #9947: [WIP] Performance optimization for dot(csr, rsp) on cpu URL: https://github.com/apache/incubator-mxnet/pull/9947#discussion_r173386908
########## File path: src/operator/tensor/dot-inl.h ########## @@ -793,16 +764,21 @@ inline void DotCsrRspDnsImpl(const OpContext& ctx, mxnet_op::Kernel<mxnet_op::set_zero, cpu>::Launch(s, num_threads, ret->dptr<DType>()); } - num_threads = mxnet_op::get_num_threads<cpu>(ret->shape_[0]); - dim_t seg_len = (ret->shape_[0] + num_threads - 1) / num_threads; + num_threads = ret->shape_[0]; if (trans_lhs) { LOG(FATAL) << "DotCsrRspDnsImpl has not implemented dot(csr.T, rsp) = dns yet"; } else { + const RType* row_idx_ptr = row_idx_r.dptr<RType>(); + std::unordered_map<RType, dim_t> row_idx_map; + row_idx_map.reserve(nnr * 7); + for (dim_t ind = 0; ind < nnr; ind++) { + row_idx_map.emplace(row_idx_ptr[ind], ind); + } Review comment: @eric-haibin-lin if we maintain an array for mapping and an array for recording the number of non-zero entries in the mapping array, the overhead of cleaning the mapping is O(nnr). @reminisce depending on how we use a tree to store the index. we can achieve O(log_k N) accesses. in practice, k is a large number (e.g., 128), so it's 3 or 4 memory accesses before finding the right location. memory overhead is O(min(N, k * nnr)). it might be better than using a single array or hashtable. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services