Tensor contractions #105
Draft
Tensor contractions #105
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Can you rebase against |
wardvermeulen
force-pushed
the
TensorContractions
branch
from
110754b to
841a2ca
Compare
|
See JuliaGPU/CUDA.jl#1960, both for CI and for any impact it may have on your work. |
|
If you want to use a Manifest, it'll have to be one generated by the oldest version of Julia you want to test, i.e., 1.6 (should be easy enough using |
thomasfaingnaert
force-pushed
the
TensorContractions
branch
from
aacf071 to
31156dc
Compare
maleadt
reviewed
Jun 28, 2023
src/GettContractions.jl
Outdated
| @@ -0,0 +1,56 @@ | |||
| export GETT | |||
There was a problem hiding this comment.
What is this file for? It's included nowhere.
maleadt
reviewed
Jun 28, 2023
| extent::Vector{Int} | ||
| stride::Vector{Int} | ||
| dataType::DataType | ||
| unaryOp |
There was a problem hiding this comment.
Why is this part of the tensor, and not the contraction plan?
wardvermeulen
force-pushed
the
TensorContractions
branch
3 times, most recently
from
f9266e8 to
6a7af93
Compare
wardvermeulen
force-pushed
the
TensorContractions
branch
3 times, most recently
from
582366c to
cf04fa2
Compare
wardvermeulen
force-pushed
the
TensorContractions
branch
from
d75016c to
d3d5b28
Compare
wardvermeulen
force-pushed
the
TensorContractions
branch
from
d3d5b28 to
700f50a
Compare
wardvermeulen
force-pushed
the
TensorContractions
branch
from
7f06ac8 to
2202f41
Compare
E.g. in case we do not find a configuration for a certain TC, the ratio is cuTENOR_time / Inf = 0, leading to a 0 geomean. While we should fix such cases, let's calculate the geomean w/o taking these into account to get an idea of the average performance for now.
thomasfaingnaert
force-pushed
the
TensorContractions
branch
from
968d1c2 to
5c50b5f
Compare
thomasfaingnaert
force-pushed
the
TensorContractions
branch
from
5c50b5f to
ab3d38c
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR adds Tensor Contraction functionality to GemmKernels.jl using the GEMM-like Tensor Tensor (GETT) multiplication algorithm. The API mimics the cuTENSOR API. It is still a draft, the benchmark scripts need to be refined further. Because it is a draft, I disabled the other tests temporarily. I also had to revert the CUDA runtime to v11.8 for cuTENSOR to work.
The following function of cuTENSOR is implemented:
contraction!This works both with
WMMAOpas withFPUOpof #101. As far as I can tell, cuTENSOR does not support different operations than multiplication and addition for its contraction. Because of the FPU operator, that is a possibility here.The following functions are not implemented:
permutation!reduction!elementwiseTrinary!elementwiseBinary!I think it could be interesting future work to create a permutation (i.e. transposition) kernel and reduction kernel through the reuse of GemmKernels.jl building blocks. The same goes for the elementwise functions. You could probably do something extremely similar to GETT, but the kernel would need changing since contractions are not allowed.
The
contraction!functionality is tested against the TCCG benchmark suite, using cuTENSOR to verify the results. Benchmarks for the Tesla P100, GeForce RTX 2080 Ti and Tesla V100 will be added later.