xavierandrade / fortrancl Goto Github PK

I try to compile fortrancl on a Xeon server, OS: Ubuntu server, 12.04.4. Intel 
OpenCL SDK 2013 and AMD OpenCL SDK have been tried. Fortrancl library can be 
compiled correctly while examples in package could not generated correctly.


What steps will reproduce the problem?
1. Platform: Intel Xeon 2653, with Intel OpenCL SDK 2013, AMD SDK gives same 
error
2. Compiler: ifort + gcc, ifort + icc give same error
3. Make fortrancl library works fine. But compile and link examples in 
fortrancl failed.
4. Noticed warning messages likes

checking CL/cl.h usability... yes
checking CL/cl.h presence... no
configure: WARNING: CL/cl.h: accepted by the compiler, rejected by the 
preprocessor!
configure: WARNING: CL/cl.h: proceeding with the compiler's result
checking for CL/cl.h... yes

When I try to compile and link examples, it will failed like

Furthermore, I have tried to use fortrancl on my Mac and a personal computer 
with AMD SDK and ubuntu 12.04.4 and without any problem and no above warning 
message given.

The "fortrancl" is a cute tools.

install cygwin on windows7 64 bit
add fortran xe studio 2013 for ifort intel compiler
add path to cygwin

start 
./configure FC=ifort

later it will claim that it is unable to compile .f90 files

is this a problem with the current version of ifort?

config.log is attached

After compiling with Intel Fortran (latest version)
, the example code "sum" run as following :

-----------------------------------------------------
CL device: 
 Caicos                                                                         

 "/tmp/OCLc88xhl.cl", line 1: error: expected a declaration
  /oi  1 r
  ^

At e
 nd of source: warning: parsing restarts here after previous syntax error

1 err
 or detected in the compilation of "/tmp/OCLc88xhl.cl".

Internal error: clc com
 piler invocation failed.

Error: program build failed.
------------------------------------------------------

My OS is scientific linux 63

Thanks

Looks like to many C&P for me ;)

The code will not work if the Fortran compiler uses 8 bytes integers by default.

What steps will reproduce the problem?
1. Create a command queue with CL_QUEUE_PROFILING_ENABLE
2. Call a C function who uses it. "call cfunction(command_queue)" "void 
cfuction(cl_command_queue *cmd_queue)"
3. call clGetEventProfilingInfo to get info about the event who used the 
*cmd_queue

What is the expected output? What do you see instead?
Info about the event. CL_PROFILING_INFO_NOT_AVAILABLE.

What version of the product are you using? On what operating system?
Running on Debian 6 kernel 2.6.32, ifort and icpc as compilers.

Please provide any additional information below.

There are multiple definitions in fortrancl.pc.in. The multiple definitions of 
prefix cause an error with pkg-config --all. See the patch.

At line 91 of cl_kernel_low.c is:

void FC_FUNC_(clgetkernelworkgroupinfo_int64, CLGETKERNELWORKGROUPINFO_int64)

The second argument should be all uppercase.

I would like to propose a few enhancement to the build system:
 1) allow to build both shared and static libs, (while defaulting to shared) [1]
 2) install .pc file
 3) don't install module to flat namespace.

Reasoning behind 1) is mostly described in provided link, 2) is just for 
convience, 3) will allow in the future to install simultaneously multiple 
modules for multiple compilers.
Please note that supplied patch also fixes possible parallel build issues.

The implementation of CLGetPlatformIDs isn't consistent with the OpenCL API and 
fails if there is more than one platform. The _single entry passes a 1:1 array 
but allows clgetplatformids_listall to overwrite num_platforms with the actual 
number of platforms. The subsequent DO loop in clgetplatforms_list then tries 
to write to elements past 1, either corrupting memory or causing a run-time 
error.

I note that the Fortran interface of CLGetPlatformIDs doesn't correspond to the 
spec in the way that num_entries is used. Ideally, the Fortran interface would 
mimic the C interface as close as possible, using OPTIONAL arguments to allow 
for platforms to be omitted. Having a signature that allows a scalar platform 
is a bit odd and encourages improper application programming.

I've got a temporary fix for this that works, but I'd be happier if it were 
restructured to match the API for ease of programming and translation. Perhaps 
it can be done without an incompatible change on the Fortran side. I will be 
playing with this more over the coming months.

That said, I did get it to build and run the SUM sample (with modifications of 
the sample as well) using Intel Visual Fortran on Windows and the Intel OpenCL 
SDK.

It is very useful! And I want to thank you for your time and energy.

It appears that clSetKernelArgLocal  is not added to the library when i build 
it. 
I am using clSetKernelArgLocal_low now, but then the next iteration fails when 
calling clEnqueueWriteBuffer with a CL_MEM_OBJECT_ALLOCATION_FAILURE.

Is clSetKernelArgLocal supposed to be not included?

The only types accepted when compiling the kernel are int and float. The wiki 
suggests double may be used for real(8). Here is the error message:

"/tmp/OCLkFtnN3.cl", line 2: error: identifier "double" is undefined