Fix partial any/all reduction

[simonbyrne]

Fixes #931.

[cliffburdick]

/build

[cliffburdick]

Looks good! Did you happen to reproduce the original bug?

[simonbyrne]

Yes, it occurs when you do a partial reduction of a tensor rank greater than 2 (basically when CUB isn't used).

[cliffburdick]

Yes, it occurs when you do a partial reduction of a tensor rank greater than 2 (basically when CUB isn't used).

Since this code hasn't been touched in a while, is there a reason we can't use CUB in that case? We used to not use CUB in that situation before we had operators as input into CUB, now using ReduceInput we should support all ranks.

[simonbyrne]

ah, I thought CUB required contiguous segments. Let me see if I can figure it out.

[simonbyrne]

Okay, it took me a bit to figure out how it all worked, but my reading of it is that we can now use CUB for all reductions.

[simonbyrne]

Okay, I've deleted all the non-CUB reduction code.

[cliffburdick]

/build

[cliffburdick]

Looks like docs are failing:

/home/jenkins/workspace/unit-tests/docs_input/api/manipulation/selecting/reduce.rst:8: WARNING: doxygenfunction: Unable to resolve function "reduce" with arguments (OutType, TensorIndexType, const InType&, ReduceOp, cudaStream_t, bool) in doxygen xml output for project "MatX" from directory: /home/jenkins/workspace/unit-tests/build/docs_input/doxygen/xml.
Potential matches:
- template<typename InType, int D, typename ReduceOp> __MATX_INLINE__ auto reduce(const InType &in, const int (&dims)[D], ReduceOp op, bool init = true)
- template<typename InType, typename ReduceOp> __MATX_INLINE__ auto reduce(const InType &in, ReduceOp op, bool init = true)
- template<typename OutType, typename InType, typename ReduceOp> void __MATX_INLINE__ reduce(OutType dest, const InType &in, ReduceOp op, cudaStream_t stream = 0, bool init = true) [docutils]
- ```

[simonbyrne]

Ah, I see what happened. So it turns out there are two different reduce methods

1. the MatX operators in operators/reduce.h

   • they don't appear in the current documentation

2. the mutating ones in transforms/reduce.h where the output tensor is provided as an argument to the function

   • are not actually MatX operators: e.g. they accept a stream as an argument, instead of returning an operator and calling run().
   • the first of these is mentioned in the docs
   • are removed by this PR

I can either remove 2 completely, or simply make it fall back on calling the operator?

[cliffburdick]

Ah, I see what happened. So it turns out there are two different reduce methods

1. the MatX operators in operators/reduce.h

   • they don't appear in the current documentation

2. the mutating ones in transforms/reduce.h where the output tensor is provided as an argument to the function

   • are not actually MatX operators: e.g. they accept a stream as an argument, instead of returning an operator and calling run().
   • the first of these is mentioned in the docs
   • are removed by this PR

I can either remove 2 completely, or simply make it fall back on calling the operator?

Let's remove it from the docs and only have the operator

[cliffburdick]

/build