(Noticed whilst reviewing #6695)

From the docs for numba.cuda.atomic.compare_and_swap:

class compare_and_swap(ary, old, val)

Conditionally assign val to the first element of an 1D array ary if the current value matches old.
Returns the current value as if it is loaded atomically.

It seem

Problem: the approximate method can still be slow for many trees
catboost version: master
Operating System: ubuntu 18.04
CPU: i9
GPU: RTX2080

Would be good to be able to specify how many trees to use for shapley. The model.predict and prediction_type versions allow this. lgbm/xgb allow this.

The constant array

__constant__ char max_string_sentinel[5]{"\xF7\xBF\xBF\xBF"};

is defined in a header file (device_operators.cuh). Therefore, when that header is included multiple times in multiple, separate compilation units, we will have compiling error saying that there are multiple definitions of max_string_sentinel.

Current implementation of join can be improved by performing the operation in a single call to the backend kernel instead of multiple calls.

This is a fairly easy kernel and may be a good issue for someone getting to know CUDA/ArrayFire internals. Ping me if you want additional info.

PR NVIDIA/cub#218 fixes this CUB's radix sort. We should:

• Check whether Thrust's other backends handle this case correctly.
• Provide a guarantee of this in the stable_sort documentation.
• Add regression tests to enforce this on all backends.

Describe the bug
We should be able to compile the googletests of the libcuml++ algorithms without needing to compile the C wrappers (i.e. libcuml.so).

Steps/Code to reproduce bug
Configure the compilation of libcuml++.so and test/ml with:

cmake .. -DBUILD_CUML_TESTS=ON -DBUILD_CUML_CPP_LIBRARY=ON -DBUILD_CUML_C_LIBRARY=OFF

This leads to:

/usr/bin/ld:

I often use -v just to see that something is going on, but a progress bar (enabled by default) would serve the same purpose and be more concise.

We can just factor out the code from futhark bench for this.

Thank you for this fantastic work!

Could it be possible the fit_transform() method returns the KL divergence of the run?

Thx!