Removed gpu.release_gil config flag again, releasing GIL by default now

theano/sandbox/cuda/cuda_ndarray.cu

@@ -26,6 +26,21 @@
 //if you want this to work.
 #define PRECHECK_ERROR 0
 
+//If true, we release the GIL around blocking GPU calls, to allow other Python
+//threads to run in the meantime. For a single-threaded program, the overhead
+//is neglectible (about 20ms for 1 million GIL release/reclaim cycles). Can
+//still be overridden on compilation with -DRELEASE_GIL=0 in nvcc.flags.
+#ifndef RELEASE_GIL
+#define RELEASE_GIL 1
+#endif
+#if RELEASE_GIL
+#define CNDA_BEGIN_ALLOW_THREADS Py_BEGIN_ALLOW_THREADS
+#define CNDA_END_ALLOW_THREADS Py_END_ALLOW_THREADS
+#else
+#define CNDA_BEGIN_ALLOW_THREADS
+#define CNDA_END_ALLOW_THREADS
+#endif
+
 /////////////////////////
 // Alloc and Free
 /////////////////////////

theano/sandbox/cuda/cuda_ndarray.cuh

@@ -68,14 +68,6 @@ typedef float real;
 #define CNDA_THREAD_SYNC cudaThreadSynchronize();
 #endif
 
-// Define shortcuts to implement the config.gpu.release_gil flag
-#ifdef RELEASE_GIL
-#define CNDA_BEGIN_ALLOW_THREADS Py_BEGIN_ALLOW_THREADS
-#define CNDA_END_ALLOW_THREADS Py_END_ALLOW_THREADS
-#else
-#define CNDA_BEGIN_ALLOW_THREADS
-#define CNDA_END_ALLOW_THREADS
-#endif
 
 #ifndef SHARED_SIZE
 #define SHARED_SIZE (16*1024)

theano/sandbox/cuda/nvcc_compiler.py

@@ -30,12 +30,6 @@ AddConfigVar('nvcc.compiler_bindir',
              StrParam(""),
              in_c_key=False)
 
-AddConfigVar('gpu.release_gil',
-             "If True, theano will release the GIL when waiting for "
-             "GPU operations, allowing other Python threads to run",
-             BoolParam(False),
-             in_c_key=True)
-
 user_provided_cuda_root = True
 
 
@@ -159,8 +153,6 @@ class NVCC_compiler(object):
         flags = [flag for flag in config.nvcc.flags.split(' ') if flag]
         if config.nvcc.fastmath:
             flags.append('-use_fast_math')
-        if config.gpu.release_gil:
-            flags.append('-DRELEASE_GIL')
         cuda_ndarray_cuh_hash = hash_from_file(
             os.path.join(os.path.split(__file__)[0], 'cuda_ndarray.cuh'))
         flags.append('-DCUDA_NDARRAY_CUH=' + cuda_ndarray_cuh_hash)