mganahl
commented
3 years ago Hi @saeedjahromi, thank you for the message! Great to hear that you are using the library for PEPS.
You are saying that the code is very slow. What is the baseline you are comparing it to (plain numpy, other symmetric tensor code)? Also, are you doing variational optimization, CTM, or something else? How large are your bond dimensions?
The functions you are mentioning are computing the symmetry blocks of a given tensor (when reshaped into a matrix). This step can indeed become a bottleneck if your bond dimensions are small and your tensor have only a few (~<=3) legs.
To give some context: Our code uses an approach that is somewhat different from many other publicly available packages. We don't store the individual symmetry blocks of a tensors separately. Instead, all non-zero elements are stored in a 1d data-array (numpy array). If we want to perform contractions, decompositions a.s.o. we use the charge information of each tensor leg to work out which elements of this data-array go into which symmetry block. Working out this mapping can become the bottleneck for small bond dimensions. That said, this approach can have significant advantages if the tensors have higher ranks (~>=4) and/or more than one simultaneous symmetry (e.g. two or more species for fermions, charge conservation + Z2 conservation a.s.o.).
One thing you can do to increase speed is to turn on a "caching" option within the block-sparse code. You can do this with tensornetwork.block_sparse.enable_caching(). If this is turned on, the functions you mentioned above are cached on the inputs. This works very well if there is no truncation step in your algorithm, e.g. via an SVD. If there is an SVD you can still use caching, but the chances of getting a cache hit are significantly reduced. This is because SVDs/truncations often involve more than one tensor in your network, and lead to redistribution of charges across the involved tensors. This makes caching less efficient. Furthermore, due to the constant charge redistribution, the cache may fill up and use a lot of memory. If this happens you can clear the cache using tensornetwork.block_sparse.clear_cache(). You can also just disable caching for the SVD with tensornetwork.block_sparse.disable_caching().
Let me know if this helps!
saeedjahromi
Hey dev team. I have been working on a PEPS algorithms for simulating 2D fermionic systems with the symmetric backend and it turned out the symmetric backend is very slow for PEPS tensors. I did some profiling and found that the bottleneck is the slow performance of the functions in the blocksparse_utils.py file which mainly contributes to the _find_diagonal_sparse_blocks() and _find_transposed_diagonal_sparse_blocks() functions.
I was wondering if there is any solution to speed up the symmetric backend or any advice or suggestion from your side?