Christoph Gohlke <[email protected]> writes:
> <[email protected]> bounced. Resending to <[email protected]>.
>
>
> Hi Andreas,
>
> sorry for the delay.
>
> I tried building PyOpenCL 2015.2.x on Windows and noticed several issues:
>
>
> 1. The PyPI source distribution is missing src/c_wrapper/mingw-std-threads.
>
> Using git master for now.
Fixed. [ea3d7a6]
> 2. Setup.py fails if git not found.
>
> Using this workaround:
>
> diff --git a/aksetup_helper.py b/aksetup_helper.py
> index b7378b5..29ed40a 100644
> --- a/aksetup_helper.py
> +++ b/aksetup_helper.py
> @@ -703,6 +703,7 @@ def substitute(substitutions, fname):
> def _run_git_command(cmd):
> git_error = None
> from subprocess import Popen, PIPE
> + stdout = None
> try:
> popen = Popen(["git"] + cmd, stdout=PIPE)
> stdout, stderr = popen.communicate()
> @@ -722,8 +723,11 @@ def _run_git_command(cmd):
> print(git_error)
> print("Hit Ctrl-C now if you'd like to think about the
> situation.")
> print(DASH_SEPARATOR)
> - count_down_delay(delay=5)
> - return stdout.decode("utf-8"), git_error
> + count_down_delay(delay=0)
> + if stdout:
> + return stdout.decode("utf-8"), git_error
> + else:
> + return '', None
>
>
> def check_git_submodules():
Merged, thanks. [e6ee3ca]
> 3. PyOpenCL 2015.2.x was apparently rewritten specifically for gcc
> compilers. The default platform compilers for Python on Windows, Visual
> Studio 2008, 2010, and 2015 no longer compile the code. Same for Intel
> Compilers 2013 and 2016. Too many different errors to report.
It uses a healthy dose of C++11. Clang and g++ are fine, FWIW, but I can
see why MSVC and Intel are unhappy. :/
I don't know how ready MS's (recent) Clang frontend for their compilers
is, but that may be a viable solution.
In the meantime, I've mentioned mingwpy in the README [5f08e13].
> The workaround is to use mingwpy compilers. However, they do not support
> Python 3.5, I can not debug the builds, and cffi might have issues:
>
> pyopencl._cffi.cpp:365:68: warning: unknown conversion type character
> 'z' in format
This is spurious. The code causing this looks like this:
_CFFI_UNUSED_FN
static PyObject **_cffi_unpack_args(PyObject *args_tuple, Py_ssize_t
expected,
const char *fnname)
{
if (PyTuple_GET_SIZE(args_tuple) != expected) {
PyErr_Format(PyExc_TypeError,
"%.150s() takes exactly %zd arguments (%zd given)",
fnname, expected, PyTuple_GET_SIZE(args_tuple));
return NULL;
}
return &PyTuple_GET_ITEM(args_tuple, 0); /* pointer to the first item,
the others follow */
}
and %zd is allowed in PyErr_Format:
https://docs.python.org/2/c-api/string.html#c.PyString_FromFormat
> 4. strdup failed to link:
>
> wrap_cl.cpp:(.text+0x42f): undefined reference to `strdup'
>
> Using _strdup instead:
>
> diff --git a/src/c_wrapper/wrap_cl.h b/src/c_wrapper/wrap_cl.h
> index 41e57c2..576de6c 100644
> --- a/src/c_wrapper/wrap_cl.h
> +++ b/src/c_wrapper/wrap_cl.h
> @@ -31,6 +31,7 @@
> #if defined(_WIN32)
> #define NOMINMAX
> #include <windows.h>
> +#define strdup _strdup
> #else
> #include <unistd.h>
> #endif
Merged, thanks. [971ad78]
> 5. CL_KERNEL_ARG_TYPE_const' was not declared in this scope:
>
> src/c_wrapper/wrap_constants.cpp:8:24: error: 'CL_KERNEL_ARG_TYPE_const'
> was not declared in this scope
>
> Disabling CL_KERNEL_ARG_TYPE_const for now:
>
> diff --git a/src/c_wrapper/wrap_constants.cpp
> b/src/c_wrapper/wrap_constants.cpp
> index 0685b1f..661d7e8 100644
> --- a/src/c_wrapper/wrap_constants.cpp
> +++ b/src/c_wrapper/wrap_constants.cpp
> @@ -613,7 +614,7 @@ void populate_constants(void(*add)(const char*,
> const char*, int64_t value))
> // kernel_arg_type_qualifier
> #if PYOPENCL_CL_VERSION >= 0x1020
> ADD_ATTR("kernel_arg_type_qualifier", KERNEL_ARG_TYPE_, NONE);
> - ADD_ATTR("kernel_arg_type_qualifier", KERNEL_ARG_TYPE_, CONST);
> + //ADD_ATTR("kernel_arg_type_qualifier", KERNEL_ARG_TYPE_, CONST);
> ADD_ATTR("kernel_arg_type_qualifier", KERNEL_ARG_TYPE_, RESTRICT);
> ADD_ATTR("kernel_arg_type_qualifier", KERNEL_ARG_TYPE_, VOLATILE);
> #endif
It's weird that that would be missing while the others are there, but so
be it. I've made an #ifdef gate to check whether it's there. [7a9da07]
> 6. 32 bit build fails with:
>
> src/c_wrapper/error.h:255:35: error: invalid conversion from 'cl_int
> (__attribute__((__stdcall__)) *)(cl_mem) {aka int
> (__attribute__((__stdcall__)) *)(_cl_mem*)}' to 'cl_int (*)(_cl_mem*)
> {aka int (*)(_cl_mem*)}' [-fpermissive]
> call_guarded(func, #func, args)
Missing __stdcall in function pointer. Worked around. [d19d94cf]
> 7. 64 bit builds succeed and pyopencl can be imported, but running any
> test results in immediate segfaults in _cffi.pyd.
>
> I used latest CUDA and Intel OpenCL SDKs. CFFI is version 1.3.1,
> compiled with Visual Studio. I'll try to debug the crash when I find the
> time.
Hard for me to investigate without a Windows build environment. If I can help
somehow, please let me know.
I'll probably push out a release of these fixes just to prevent some
grief for folks trying to build on Windows.
> Greetings to Urbana-Champaign. I worked at Loomis Lab for ~8 years.
:) I can see the place from my office in CS right now. Hope you enjoyed
it here!
Andreas
_______________________________________________
PyOpenCL mailing list
[email protected]
http://lists.tiker.net/listinfo/pyopencl
