trace g traces GPU actions; currently only kernel launches will be reported.

The CUDA runtime system and driver provide very little in the
way of helpful or clear error messages. Especially in the case of
kernel dimensions being out-of-bound. Depending on what CUDA-CC
one uses (e.g. 2.0, 3.5, 5.0, etc.) there is a max-limit to how
large the X, Y, and Z dimensions can be for the kernel.

This patch adds a check which will error out if we generate a dimension
greater than permitted by the CUDA-CC.

Furthermore a test is provided.

Improved GPU error checking reporting. GPU errors are now specific to CUDA API calls and include the file name and line number.

We introduce DBUG_UNREACHABLE macro to mark the places in the code
that must not happen.  We produce an error message stating that we
hit such a place in the code.

    * debug.[ch]:    DBUG_UNREACHABLE macro defined for both DBUG_OFF
                     and !DBUG_OFF cases.

                     exit_function was moved outside of !DBUG_OFF
                     region.

    * ctinfo.[ch]:   Removed CTIterminateCompilationSilient in favour
                     of existing CTIexit and replaced occurences.

    * main.c         Adjusted comments to match doxygen style, set
                     exit_function unconditionally to DBUG_OFF.

Final changes which eliminate almost almost all statement-passing
macros, and replace long fprintf-s with `out` and `indout`.  There
are some more icm2c_xxx files, where the fprintf replacement can
be done.  Let's leave it for a future refactoring.

  scafolding work would be required otherwise, as they can only be compiled by 
  the NVidia compiler, and I don't have the time for that now.
Provided macros for cudahybrid related ICMs.
Expanded the number and arguments of cudahybrid primitive functions to better
  match the required ICMs, specifically distinctions between host and device
  transfers and arguments for setting ownership of data.
Created specific ICM for (de)allocation and reference counting of distributed
  variables.
Modified kernel launches to make use of streams on cudahybrid backend.

1) loopallocopt.c to lift allocation out of a for loop
2) reusewithregion.c to perform memory reuse base on polyhedral model.

a scalar. Withloop scalarizing is applied to the fold withloops (uncompleted)

No commit message

No commit message

back end, an architecture option must be specified.

No commit message

No commit message

to compilation error. improve tag cuda execution mode.

No commit message

No commit message