# About the additional memory allocation that appears in the OMEinsum package

**URL:** <https://discourse.julialang.org/t/about-the-additional-memory-allocation-that-appears-in-the-omeinsum-package/87938>\
**Category:** Performance\
**Tags:** package, memory, omeinsum\
**Created:** [September 28, 2022, 3:08pm UTC](https://discourse.julialang.org/t/about-the-additional-memory-allocation-that-appears-in-the-omeinsum-package/87938 "2022-09-28T15:08:13Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![F-YF](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/f-yf/32/17363_2.png) [@F-YF](https://discourse.julialang.org/u/F-YF)\
**Post date:** [September 28, 2022, 3:08pm UTC](https://discourse.julialang.org/t/about-the-additional-memory-allocation-that-appears-in-the-omeinsum-package/87938/1 "2022-09-28T15:08:13Z")

</div>

```julia
julia> using OMEinsum

julia> a=rand(1024,1024);

julia> b=rand(1024,1024,2);

julia> using BenchmarkTools

julia> @btime ein"xy,yzw->xzw"(a,b);
  23.790 ms (77 allocations: 16.00 MiB)

julia> @btime ein"yx,yzw->xzw"(a,b);
  28.217 ms (101 allocations: 24.01 MiB)

```

As the example above shows, when I try to do this kind of index summation: ein"yx,yzw-\>xzw"(a,b), there’s always an extra memory allocation. This is equivalent to doing an index swap of the A matrix first using permutedims(), but I can always use permutedims!() to avoid additional memory allocation.

It seems that OMEinsum does not save this memory. When I actually use it, I get a lot of index summation operations, and I do a permutedims! every time I call ein operation can be tiring. My question is, does OMEinsum have any parameter settings that will save memory in this case?

---

<div class="post-metadata">

**Author:** ![mcabbott](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/mcabbott/32/6603_2.png) [@mcabbott](https://discourse.julialang.org/u/mcabbott)\
**Post date:** [September 28, 2022, 3:17pm UTC](https://discourse.julialang.org/t/about-the-additional-memory-allocation-that-appears-in-the-omeinsum-package/87938/2 "2022-09-28T15:17:15Z")

</div>

Ideally this would reshape `b` and call `a*b_mat` or `a'*b_mat`, both of which are handled by BLAS without further copies. But it seems not to figure this out, and inserts `permutedims(a)`. TensorOperations manages not to:

```julia
julia> using TensorOperations

julia> ein"xy,yzw->xzw"(a,b) ≈ @tensor c[x,z,w] := a[x,y] * b[y,z,w]
true

julia> @btime @tensor c[x,z,w] := a[x,y] * b[y,z,w];
  min 17.247 ms, mean 18.096 ms (3 allocations, 16.00 MiB)

julia> @btime @tensor c[x,z,w] := a[y,x] * b[y,z,w];
  min 16.403 ms, mean 17.084 ms (3 allocations, 16.00 MiB)

# transpose by hand -- probably misses BLAS?
julia> ein"yx,yzw->xzw"(a,b) ≈ @btime ein"xy,yzw->xzw"(a',b)
  min 681.191 ms, mean 681.278 ms (82 allocations, 16.03 MiB)
true

```

---

<div class="post-metadata">

**Author:** ![F-YF](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/f-yf/32/17363_2.png) [@F-YF](https://discourse.julialang.org/u/F-YF)\
**Post date:** [September 29, 2022, 2:40am UTC](https://discourse.julialang.org/t/about-the-additional-memory-allocation-that-appears-in-the-omeinsum-package/87938/3 "2022-09-29T02:40:20Z")

</div>

Okay,thank you very much!