-
Notifications
You must be signed in to change notification settings - Fork 2
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add scripts for benchmarking opt_einsum as well as manyinds benchmarks #3
base: master
Are you sure you want to change the base?
Conversation
Thanks for the PR! julia> code = ein"abcdefghijklmnop,flnqrcipstujvgamdwxyz->bcdeghkmnopqrstuvwxyz"
abcdefghijklmnop, flnqrcipstujvgamdwxyz -> bcdeghkmnopqrstuvwxyz
julia> OMEinsum.timespace_complexity(code, uniformsize(code, 2))
(26.0, 21.0)
julia> @btime code(x, y);
24.924 ms (111 allocations: 48.51 MiB) This is because OMEinsum uses the permutedims + reshape + matmul to perform tensor contraction. While in this pattern, the julia> using LinearAlgebra, TensorOperations
julia> LinearAlgebra.permutedims(a::Array{T,N}, perm::NTuple{N}) where {T,N} = (TensorOperations.tensorcopy(a, collect(1:ndims(a)), perm))
julia> LinearAlgebra.permutedims!(o::Array{T}, a::Array{T}, perm::Vector) where T = (TensorOperations.tensorcopy!(a, collect(1:ndims(a)), o, perm))
julia> @btime code(x, y);
13.353 ms (398 allocations: 48.54 MiB) I think the best way to sovle this issue is to remove |
Great job, wondering how the And unfortunately, I do not have the write access to this repo. @under-Peter Can you please give me a write access or help merge this PR. |
@GiggleLiu is it enough to add you as a collaborator or do i have to give you write access separately? Thanks so much and sorry for being short on time atm. |
Now I have the write access, thanks for reacting so fast, @under-Peter. |
Is this good to be merged? |
It looks good, but I want to make sure the benchmark on GPU is correct (especially the manyindex case), it is different from the running the case on my own host. Can you please show me the script generating the result? After that, it should be good to merge. |
Currently we don't benchmark computations like this with many indices. I add benchmarks for these (see also the PR to OMEinsum.jl) and scripts for benchmarking the Python library opt_einsum.