[FEA] Implementation of MATLAB's find()

[akfite]

First of all, let me just say that I'm excited to see this project. I'm a long-time user of MATLAB but very new to CUDA, and I love that you are lowering the barrier of entry to GPU programming.

Is your feature request related to a problem? Please describe.
Slicing a matrix is often paired with MATLAB's built-in find function. I think it would be a valuable addition to this project.

Describe the solution you'd like
Similar to how you have implemented linspace, meshgrid, etc, it would be great to see a similar syntax & functionality to the above link.

Describe alternatives you've considered
I just stumbled upon this library today--I read through all the help docs and didn't see this anywhere. Hopefully I didn't miss it!

[cliffburdick]

Hi @akfite, you're in luck, and I just haven't updated the documentation yet. We do have a find function, and you can see how it's used in the unit tests here:

https://github.com/NVIDIA/MatX/blob/main/test/00_operators/ReductionTests.cu#L520

Because it's C++, it's slightly clunkier than MATLAB's find since we don't do any dynamic memory allocations for output tensors. Instead, you have to provide an output tensor that's at least as large as the most things you expect to find(), or worst case, the same size as the input tensor. Then the num_found output variable will tell you how many values in that tensor are populated for you to use. This API is subject to change to make it easier, but it likely won't be too much different, and you can use it now.

[akfite]

Awesome, thanks for the quick reply. I'll use your unit tests as a guide. Looking forward to using this code & I appreciate your help!

[akfite]

As far as I can tell, this implementation of find acts like a mask to return the values that match the selection operator. But what is really desired is the index to the values that were matched--that is how MATLAB's find works.

The desired functionality here is to get all the indices that match some selector, and then to be able to use those indices together with the input tensor to create a new tensor or view. That would open up all kinds of useful ways to slice & dice tensors.

[cliffburdick]

Hi @akfite, understood. This should be doable using different iterator types in CUB. We'll get back to you.

[cliffburdick]

@akfite I had previously opened this same issue, but I'll leave both open for tracking purposes: #135

[akfite]

@cliffburdick Sorry, you're right; I should have looked closer at the outstanding issues.

I will say that #135 describes returning the indices as an optional output, but I think it would be more general-form to only return the indices. If one desired the masked values (as it returns now), it would be easily obtained by applying the indices to the input matrix. find is such a foundational operator in MATLAB that it would be wise to align as closely as possible to its pattern of usage--especially given the parallels that already exist.

[cliffburdick]

Understood. I have it working and will post the code/examples shortly for you. I was going to default to calling the functions find and find_idx, but haven't settled on that yet.

[cliffburdick]

Hi @akfite , you can find the test for it here: https://github.com/NVIDIA/MatX/blob/main/test/00_operators/ReductionTests.cu#L553

The one caveat is we'll be adding a way to select these items shortly. It looks like MATLAB's find will return an absolute index regardless of the rank of the array, so it's reasonable for us to do that as well. For example, if you have a 10x5 tensor, index 49 will be the last element and not an array of (9, 4). If you need to translate an absolute index into a coordinate index, we have the utility function GetIdxFromAbs to return this. I'll post an example soon.

[akfite]

Excellent! I think returning an absolute index is perfectly reasonable, especially if you have a function to do the conversion. Thanks for turning this around so quickly.

[cliffburdick]

Hi @akfite, please see the unit test here:
https://github.com/NVIDIA/MatX/blob/main/test/00_operators/ReductionTests.cu#L588

This shows a complete example of both the find() and the find_idx() functions. find_idx() can be used with a new select() operator, which will index a tensor with a list of indices in another tensor:

    find_idx(t1o_idx, num_found, t1, GT{thresh});
    cudaStreamSynchronize(0);

    auto t1o_slice = t1o.Slice({0}, {num_found()});
    auto t1o_idx_slice = t1o_idx.Slice({0}, {num_found()});
    (t1o_slice = select(t1o_slice, t1o_idx_slice)).run();

The slices are here to only select num_found values. Remember that you need to synchronize the stream before the select() since Slice is done on the host and num_found is asynchronously computed on the device. It would be a race condition without that.

[akfite]

The example makes sense--thank you for that. I'm looking forward to trying it out. Unfortunately I still can't compile the unit tests, I've been at it for a few hours and at this point I'm just gonna have to pick it up again tomorrow. I'm trying to get it to work with VS2022 on Windows 10 but have some weird issues. I tried following the build instructions but it wasn't working for me--specifically, I used:

cmake -DMATX_BUILD_TESTS=ON -DMATX_BUILD_BENCHMARKS=ON -DMATX_BUILD_EXAMPLES=ON -DMATX_BUILD_DOCS=OFF ..

which fails in cmake because "No CMAKE_CUDA_COMPILER could be found", even though nvcc is on the path. If I specify the architecture:

cmake -DMATX_BUILD_TESTS=ON -DMATX_BUILD_BENCHMARKS=ON -DMATX_BUILD_EXAMPLES=ON -DMATX_BUILD_DOCS=OFF -DCMAKE_CUDA_ARCHITECTURES=52 ..

it now successfully finds the CUDA compiler, but now cmake gives another error that:

CMake Error at CMakeLists.txt:110 (if):
  if given arguments:

    "NOT" "CMAKE_BUILD_TYPE" "OR" "STREQUAL" "Debug"

  Unknown arguments specified

So I tried

cmake -DMATX_BUILD_TESTS=ON -DMATX_BUILD_BENCHMARKS=ON -DMATX_BUILD_EXAMPLES=ON -DMATX_BUILD_DOCS=OFF -DCMAKE_CUDA_ARCHITECTURES=52 -DCMAKE_BUILD_TYPE=Debug ..

and that got me a .sln, but ultimately only 9 of 20 projects compile in VS2022. It seems some projects are failing because of -fvisibility=hidden being passed to nvcc, which Google leads me to believe may be caused by a pybind11 config? I don't know if I'm doing something dumb or if something is weird in the CMake. I attached the output from VS2022 for reference.

Hopefully tomorrow I'll figure it out so I can actually test this out! 🤞

buildlog.txt