Add a library node for np.transpose for >2 dims and a cuTENSOR implementation #1303
Conversation
| cmake_variables = {} | ||
| cmake_includes = [] | ||
| cmake_libraries = ["cutensor"] | ||
| cmake_compile_flags = ["-L/users/jbazinsk/libcutensor-linux-x86_64-1.7.0.1-archive/lib/11"] |
There was a problem hiding this comment.
maybe better to set up the LIBRARY_PATH envvar locally.
| @@ -0,0 +1,54 @@ | |||
| # Copyright 2019-2021 ETH Zurich and the DaCe authors. All rights reserved. | |||
There was a problem hiding this comment.
| # Copyright 2019-2021 ETH Zurich and the DaCe authors. All rights reserved. | |
| # Copyright 2019-2023 ETH Zurich and the DaCe authors. All rights reserved. |
| @@ -0,0 +1,67 @@ | |||
| // Copyright 2019-2022 ETH Zurich and the DaCe authors. All rights reserved. | |||
There was a problem hiding this comment.
| // Copyright 2019-2022 ETH Zurich and the DaCe authors. All rights reserved. | |
| // Copyright 2019-2023 ETH Zurich and the DaCe authors. All rights reserved. |
|
|
||
| static void CheckCutensorError(cutensorStatus_t const& status) { | ||
| if (status != CUTENSOR_STATUS_SUCCESS) { | ||
| throw std::runtime_error("cuSPARSE failed with error code: " + std::to_string(status)); |
There was a problem hiding this comment.
isn't there a cutensorGetErrorString or something similar?
| @@ -0,0 +1,203 @@ | |||
| # Copyright 2019-2021 ETH Zurich and the DaCe authors. All rights reserved. | |||
There was a problem hiding this comment.
| # Copyright 2019-2021 ETH Zurich and the DaCe authors. All rights reserved. | |
| # Copyright 2019-2023 ETH Zurich and the DaCe authors. All rights reserved. |
| @@ -37,8 +41,39 @@ def test_transpose(): | |||
| assert rel_error <= 1e-5 | |||
|
|
|||
|
|
|||
| @pytest.mark.parametrize('implementation', ['pure', 'cuTENSOR']) | |||
There was a problem hiding this comment.
cutensor should be marked as pytest.mark.gpu, because it won't be available on the standard machines. See pytest.param and the mark kwarg for more info.
|
It would be best if you used the TensorTranspose library node in the ttranspose library, which includes an expansion for the HPTT library. You can also find in the same branch tests and a working cuTENSOR environment. |
|
What's the status here? Is this related to #1309 ("Support for tensor linear algebra (transpose, dot products)")? |
Adds
What I don't know: