merge w conflict in cmodule

theano/gof/cmodule.py

@@ -1258,7 +1258,7 @@ dummy_in = open(os.devnull)
 dummy_err = open(os.devnull, 'w')
 p = None
 try:
-    p = subprocess.Popen(['gcc', '-dumpversion'], stdout=subprocess.PIPE,
+    p = subprocess.Popen(['g++', '-dumpversion'], stdout=subprocess.PIPE,
                          stdin=dummy_in.fileno(), stderr=dummy_err.fileno())
     p.wait()
     gcc_version_str = p.stdout.readline().strip()

theano/sandbox/cuda/__init__.py

 import atexit, os, stat, sys
+import logging
+
 from theano.compile import optdb
 from theano import config
 from theano.gof.cmodule import get_lib_extension
+import nvcc_compiler
 
-import logging
 _logger_name = 'theano.sandbox.cuda'
 _logger = logging.getLogger(_logger_name)
 _logger.setLevel(logging.WARNING)
@@ -69,6 +71,17 @@ date = max(stat_times)
 cuda_ndarray_loc = os.path.join(config.compiledir, 'cuda_ndarray')
 cuda_ndarray_so = os.path.join(cuda_ndarray_loc,
                                'cuda_ndarray.' + get_lib_extension())
+libcuda_ndarray_so = os.path.join(cuda_ndarray_loc,
+                               'libcuda_ndarray.' + get_lib_extension())
+
+
+# Add the theano cache directory's cuda_ndarray subdirectory to the list of
+# places that are hard-coded into compiled modules' runtime library search
+# list.  This works in conjunction with nvcc_compiler.nvcc_module_compile_str
+# which adds this folder during compilation with -L and also adds -lcuda_ndarray
+# when compiling modules.
+nvcc_compiler.add_standard_rpath(cuda_ndarray_loc)
+
 compile_cuda_ndarray = True
 
 if os.path.exists(cuda_ndarray_so):
@@ -85,7 +98,6 @@ if not compile_cuda_ndarray:
 
 try:
     if compile_cuda_ndarray:
-        import nvcc_compiler
         if not nvcc_compiler.is_nvcc_available():
             # Hide the error message if the user manually unset the
             # config.cuda.root configuration variable.
@@ -105,7 +117,14 @@ try:
                     code,
                     location=cuda_ndarray_loc,
                     include_dirs=[cuda_path], libs=['cublas'])
-
+            # If necessary, 
+            # create a symlink called libcuda_ndarray.so
+            # which nvcc_module_compile_str uses when linking
+            # any module except "cuda_ndarray" itself.
+            try:
+                open(libcuda_ndarray_so).close()
+            except IOError:
+                os.symlink(cuda_ndarray_so, libcuda_ndarray_so)
             from cuda_ndarray.cuda_ndarray import *
 except Exception, e:
     error( "Failed to compile cuda_ndarray.cu: %s" % str(e))

theano/sandbox/cuda/nvcc_compiler.py

@@ -71,9 +71,14 @@ if config.cuda.root == "AUTO":
 
 is_nvcc_available()#to set nvcc_path correctly and get the version
 
+rpath_defaults = []
+def add_standard_rpath(rpath):
+    rpath_defaults.append(rpath)
+
 def nvcc_module_compile_str(
         module_name, src_code,
-        location=None, include_dirs=[], lib_dirs=[], libs=[], preargs=[]):
+        location=None, include_dirs=[], lib_dirs=[], libs=[], preargs=[],
+        rpaths=rpath_defaults):
     """
     :param module_name: string (this has been embedded in the src_code
     :param src_code: a complete c or c++ source listing for the module
@@ -82,6 +87,7 @@ def nvcc_module_compile_str(
     :param lib_dirs: a list of library search path directory names (each gets prefixed with -L)
     :param libs: a list of libraries to link with (each gets prefixed with -l)
     :param preargs: a list of extra compiler arguments
+    :param rpaths: list of rpaths to use with Xlinker. Defaults to `rpath_defaults`.
 
     :returns: dynamically-imported python module of the compiled code.
 
@@ -89,6 +95,8 @@ def nvcc_module_compile_str(
              Otherwise nvcc never finish.
     """
 
+    rpaths = list(rpaths)
+
     if sys.platform=="win32":
         # Remove some compilation args that cl.exe does not understand.
         # cl.exe is the compiler used by nvcc on Windows.
@@ -171,19 +179,22 @@ def nvcc_module_compile_str(
         cmd.extend(['-Xcompiler', ','.join(preargs2)])
 
     if config.cuda.root and os.path.exists(os.path.join(config.cuda.root,'lib')):
-        cmd.extend(['-Xlinker',','.join(['-rpath',os.path.join(config.cuda.root,'lib')])])
+        rpaths.append(os.path.join(config.cuda.root,'lib'))
         if sys.platform != 'darwin':
             # the 64bit CUDA libs are in the same files as are named by the function above
-            cmd.extend(['-Xlinker',','.join(['-rpath',os.path.join(config.cuda.root,'lib64')])])
+            rpaths.append(os.path.join(config.cuda.root,'lib64'))
+
+    for rpath in rpaths:
+        cmd.extend(['-Xlinker',','.join(['-rpath',rpath])])
     nvccflags = [flag for flag in config.cuda.nvccflags.split(' ') if flag]
     cmd.extend(nvccflags)
     cmd.extend('-I%s'%idir for idir in include_dirs)
     cmd.extend(['-o',lib_filename])
     cmd.append(os.path.split(cppfilename)[-1])
-    if module_name != 'cuda_ndarray':
-        cmd.append(os.path.join(os.path.split(cppfilename)[0],'..','cuda_ndarray','cuda_ndarray.'+get_lib_extension()))
     cmd.extend(['-L%s'%ldir for ldir in lib_dirs])
     cmd.extend(['-l%s'%l for l in libs])
+    if module_name != 'cuda_ndarray':
+        cmd.append("-lcuda_ndarray")
     if sys.platform == 'darwin':
         cmd.extend(darwin_python_lib.split())