refs #6: Fix a critical memory overlapping bug.

 * Yes, now CUDA (>= 6.5) supports C++11 and we can get rid of the
   macro value mismatch!

Snakefile

@@ -104,12 +104,12 @@ if v: CFLAGS += ' -DNBA_RANDOM_PORT_ACCESS'
 # NVIDIA CUDA configurations
 if USE_CUDA:
     CUDA_ARCHS    = compilelib.get_cuda_arch()
-    NVCFLAGS      = '-O2 -g --use_fast_math -I/usr/local/cuda/include'
+    NVCFLAGS      = '-O2 -g -std=c++11 --use_fast_math -I/usr/local/cuda/include'
     CFLAGS       += ' -I/usr/local/cuda/include'
     LIBS         += ' -L/usr/local/cuda/lib64 -lcudart' #' -lnvidia-ml'
     print(CUDA_ARCHS)
     if os.getenv('DEBUG', 0):
-        NVCFLAGS  = '-O0 --device-debug -g -G --use_fast_math -I/usr/local/cuda/include --ptxas-options=-v'
+        NVCFLAGS  = '-O0 --device-debug -g -G -std=c++11 --use_fast_math -I/usr/local/cuda/include --ptxas-options=-v'
     if len(CUDA_ARCHS) == 0:
         NVCFLAGS += ' -DMP_USE_64BIT=0' \
                   + ' -gencode arch=compute_10,code=sm_10' \

engines/cuda/compat.hh

@@ -1,7 +1,13 @@
 #ifndef __NBA_ENGINES_CUDA_COMPAT_HH__
 #define __NBA_ENGINES_CUDA_COMPAT_HH__
 
-#define NBA_MAX_COPROC_PPDEPTH (256u)
+/*
+ * This header is included by .cu sources.
+ * We put only relevant data structures here for use in CUDA codes.
+ * Note that the nvcc should support C++11 (CUDA v6.5 or higher).
+ */
+
+#include "../../lib/config.hh"
 
 struct datablock_kernel_arg {
     uint32_t total_item_count_in;