• Note that this change is not kernel change, it reorganized and/or removed some unnecessary code which brought about 6x speed improvement in OpenCL backend's for sparse-matrix--dense-vector matmulNN operation.
• There has been no change in speed for matmulTN operation though.

I am going to look into kernel improvements of NN and TN ops now - differed for later.

* Note that this change is not kernel change, it reorganized and/or removed some unnecessary code
   which brought  about 6x speed improvement in OpenCL backend's for sparse-matrix--dense-vector
   matmulNN operation.

* There has been no change in speed for matmulTN operation though.

I am going to look into kernel improvements of NN and TN ops now - differed for later.

@umar456

I was unable to reproduce the 6x performance difference with this code reorg on lastest master with latest versions of dependencies. It seems like some sort of driver or upstream fix addressed the problem of optimizing out the unused global counter being passed for use_greedy code-path.

Nevertheless, the cleanup of the wrapper code is still valid. We don't need to pass the global counter when not used and we most definitely don't need to compile an additional kernel when not using it at all. I am going to reword the PR title to appropriately reflect the change.