tcew / occa2 Goto Github PK

How does OCCA do in the presence of multiple host threads?

One obvious thing I spotted is the stateful kernel launch interface. (setWorkDimSize or so)

Does not build. Inspecting output occa file shows:

Every self-respecting parallel programming API has one.

I am looking at the addVectors example in fortran. I've changed the occaBuildKernel line to

  addVectors = occaBuildKernelFromSource(device, "addVectors.ofl", "addVectors")

to try to engage the OFL kernel. The kernel compilation fails at run-time. When I inspect the code OCCA is trying to compile, I see that the contents of addVectors.ofl have been included verbatim:

...
#endif
#undef OCCA_USING_OPENMP
#define OCCA_USING_OPENMP 1
#undef OCCA_USING_CPU
#define OCCA_USING_CPU 1
kernel subroutine addVectors(entries, a, b, ab)
  implicit none

  integer(4), intent(in)  :: entries
  real(4)   , intent(in)  :: a(:), b(:)
  real(4)   , intent(out) :: ab(:)

  integer(4), shared    :: sharedVar(16,30)
  integer(4), exclusive :: exclusiveVar

  integer :: group, item, N

  do group = 1, entries, 16, outer0
     do item = 1, 16, inner0
        N = (item + (16 * (group - 1)))

        if (N < entries) then
           ab(i) = a(i) + b(i)
     end do
  end do

end subroutine addVectors
^@

FYI

maxhutch@edoras:~/src/OCCA2/examples/addVectors$ git log |head -3
commit 16ef4d4d8dd26c62d8a952faa9e98a5549133665
Author: Tim Warburton <[email protected]>
Date:   Mon Dec 8 15:26:35 2014 -0600

In the C interface occaKernelRun there is a memory leak when "casted" occa types are used (such as occaInt).

For instance the following call

occaKernelRun(some_kernel, occaInt(a), d_ptr);

will have a leak because 'occaInt' has a new without an associated delete.

From other places in the library, such as occaKernelInfoAddDefine the delete is issued in the call, so I assume that it make sense for the delete to be issued at the end of occaKernelRun (as opposed to a separate all by the user).

Hi,
I tried building OCCA on Windows using MINGW, but it fails as it cannot find sys/sysctl.h .
I had to hack around a bit to run it on windows, by changing HOME/scripts/makefile at around line 110 to put bash instead of /bin/bash . However, I got stuck at the error mentioned as sys/sysctl.h is not available in Windows apparently. Is there any alternative ?

Subject says it. Imagine two codes running on a node and trampling on each other's cached binaries. (Or is that what the "salt" parameter is about?)

I am in need of an occaKernelInfoAddDefine that is able to accept a string (character(len=*)) as its third argument, i.e for something like occaKernelInfoAddDefine(info,"datafloat","double"). The current fortran interface can only accept a "character" as its third argument, hence the above definition will result in "datafloat" being defined as just "d".
Thanks

Here's a log of running gNUMA with a (f)loopy kernel:

Found cached binary of [occa/clusterVerticalStrongVolumeKernel.occa] in [/exthome/andreas/._occa/f51421d852343e72]
floopy --lang=floopy --target=cl:0,0  --occa-defines=/exthome/andreas/._occa/loopy1_7d27af49ef7c8cc7.defs  --occa-add-dummy-arg occa/verticalStrongVolumeKernel.floopy /exthome/andreas/._occa/loopy2_7d27af49ef7c8cc7.ocl
sh: 1: floopy: not found
Found cached binary of [/exthome/andreas/._occa/loopy2_7d27af49ef7c8cc7.ocl] in [/exthome/andreas/._occa/9dc467c8d2d13cc6]
...

I'd argue it should stop when it fails to execute floopy.

The floopy transform executable is now just called loopy, not floopy. It also no longer takes a --target flag.

If I define a double precision kernel constant defined as :

gamma = 1.40023693379791_r8
call occaKernelInfoAddDefine(info, "p_Gamma", gamma)

Inside an occa kernel the constant is truncated to
gamma = 1.400240000000

I see that both versions with val = real(4) and real(8) are defined in occaF.f90 so there may be an issue in how the constant is stored for the later.

I think that free template should have a delete command for the data
delete (OpenCLKernelData_t*) this->data;

I am also getting some memory leaks in occaParser.cpp (from the new on line 3522) and occaParserNodes.cpp (from the new on line 216), but I am not sure where in my code these functions are getting called so I haven't been able to track down the leak fully.

I will give screenshots of the affected lines.

With these lines code compiles and my tests all pass.

However if I add a C-style comment like so:

I get the following errors

Looking at the relevant cached OCCA file I find:

There is Windows specific code in addVectors C++ example which prevents it to run on Linux and Mac OS X.