[REVIEW] Implement type_dispatcher

[jrhemstad]

The purpose of this new function is to alleviate some of the pain of having to reconstruct the type of the type-erased data buffer from a gdf_column.

In more detail, the problem is that a gdf_column is type-erased, i.e., in the struct:

struct gdf_column{
void* data;
gdf_dtype dtype;
}
Where data points to a contiguous data buffer on the device, and the gdf_dtype enum tells us what type is actually contained in the buffer.

This requires a great deal of code of the form:

switch(gdf_column.dtype)
{
case GDF_INT: // cast gdf_column.data to int
case GDF_DOUBLE: // cast gdf_column.data to double
case GDF_FLOAT: // cast gdf_column.data to float
... 15 and counting total cases
}
Currently this particular switch logic is repeated all over the source of libgdf. Apart from being an eye-sore, it is also a nightmare for maintainability. Any time a new type is added (or removed), then a great deal of source code will need to be changed, which is prone to introduce bugs.

The new gdf_type_dispatcher function will centralize this switch logic to exactly one place. This will be a boon to both maintainability as well as code readability.

This PR introduces the gdf_type_dispatcher as well as updates some of the existing code to make use of this function, including:

• get_column_byte_width
• gdf_table::copy_row
• gdf_table::hash_row
• gdf_table::rows_equal.

It is not the goal of this PR to update all of the code-base to use the gdf_type_dispatcher, but rather to introduce it and show how it may be used.

This PR also updates the C++ standard used to C++14. This is required for the automatic return type deduction of the gdf_type_dispatcher.

[harrism]

Overall looks good. A bunch of minor comments. One concern about testing.

[harrism]

This is the next PR I would like to merge, because other PRs (I count at least 3 currently) can then be updated to use its functionality. @jrhemstad, is it ready once the conflicts are resolved?

[GPUtester]

Can one of the admins verify this patch?

[harrism]

add to whitelist

[eyalroz]

These changes are independent of the type_dispatcher, and I suggest they be merged separately to prevent confusion / for orderliness' sake.

[jrhemstad]

@eyalroz not sure what changes you're referring to, but there's still some issues I'm ironing out with the submodules after merging in master after the refactor.

[eyalroz]

@eyalroz not sure what changes you're referring to, but there's still some issues I'm ironing out with the submodules after merging in master after the refactor.

I was sure I saw changes in Rename libgdf namespace to cudf which were in CMakeLists.txt or other such files... maybe they disappeared somehow. Never mind, ignore my last comment.

[jrhemstad]

This is the next PR I would like to merge, because other PRs (I count at least 3 currently) can then be updated to use its functionality. @jrhemstad, is it ready once the conflicts are resolved?

@harrism conflicts are resolved and this is ready for final review/merge.

[eyalroz]

Congratulations @jrhemstad for a well-deserved merge :-)