Rewrote all CMakeLists. Now much cleaner and there's a clear separation during compilation between the core and standalone modules.

src/core/CMakeLists.txt

@@ -2,59 +2,30 @@
 ##  CMakeLists.txt for Astaroth Core  ##
 ########################################
 
-#----------------------Find CUDA-----------------------------------------------#
-
+## Find packages
 find_package(CUDA 9 REQUIRED)
 
-#----------------------CUDA settings-------------------------------------------#
-
-set(CUDA_SEPARABLE_COMPILATION OFF)
-set(CUDA_PROPAGATE_HOST_FLAGS ON)
-
-#----------------------Setup CUDA compilation flags----------------------------#
-
-# Generate code for the default architecture (Pascal)
+## Architecture and optimization flags
 set(CUDA_ARCH_FLAGS -gencode arch=compute_37,code=sm_37
                     -gencode arch=compute_50,code=sm_50
                     -gencode arch=compute_60,code=sm_60
                     -gencode arch=compute_61,code=sm_61
                     -lineinfo
-                    -ftz=true
-                    -std=c++11) #--maxrregcount=255 -ftz=true #ftz = flush denormalized floats to zero
-# -Xptxas -dlcm=ca opt-in to cache all global loads to L1/texture cache
-# =cg to opt out
-
-# Additional CUDA optimization flags
-if (CMAKE_BUILD_TYPE MATCHES RELEASE)
-    # Doesn't set any additional flags, see CUDA_NVCC_FLAGS_DEBUG below on how
-    # to add more
-    set(CUDA_NVCC_FLAGS_RELEASE ${CUDA_NVCC_FLAGS_RELEASE})
-endif()
-
-# Additional CUDA debug flags
-if (CMAKE_BUILD_TYPE MATCHES DEBUG)
-    # The debug flags must be set inside this if clause, since either CMake 3.5
-    # or nvcc 7.5 is bugged:
-    # CMake converts these into empty strings when doing RELEASE build, but nvcc
-    # 7.5 fails to parse empty flags.
-    set(CUDA_NVCC_FLAGS_DEBUG ${CUDA_NVCC_FLAGS_DEBUG};
-                               --device-debug;
-                               --generate-line-info;
-                               --ptxas-options=-v)
-endif()
-
-set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS};${CUDA_ARCH_FLAGS}")
+                    -ftz=true # Flush denormalized floats to zero
+                    -std=c++11,
+                    --compiler-options -march=native) # Native host machine code
+                    #--maxrregcount=255
+                    # -Xptxas -dlcm=ca opt-in to cache all global loads to L1/texture cache
+                    # =cg to opt out
 
 
-message("CUDA_NVCC_FLAGS: " ${CUDA_NVCC_FLAGS})
+set(CUDA_WARNING_FLAGS --compiler-options -Wall,-Wextra,-Werror,-Wdouble-promotion,-Wfloat-conversion) # -Wshadow
 
+set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} ${CUDA_ARCH_FLAGS} ${CUDA_WARNING_FLAGS})
+set(CUDA_NVCC_FLAGS_RELEASE)
+set(CUDA_NVCC_FLAGS_DEBUG --device-debug --generate-line-info --ptxas-options=-v)
 
-#------------------Compile and create a static library-------------------------#
-file(GLOB CUDA_SOURCES "*.cu" "kernels/*.cu")
-
-# Use -fPIC if -fpic not supported. Some quick non-scientific tests:
-# Without fpic: 4.94 user, 4.04 system, 0:09.88 elapsed
-# With fpic: 4.96 user, 4.02 system, 0:09.90 elapsed
-# With fPIC: 4.94 user, 4.05 system, 0:10.23 elapsed
-CUDA_ADD_LIBRARY(astaroth_core STATIC ${CUDA_SOURCES} OPTIONS --compiler-options "-fpic")
+## Create and link the library
+include_directories(.)
+cuda_add_library(astaroth_core STATIC astaroth.cu device.cu)
 target_link_libraries(astaroth_core m)

src/core/kernels/kernels.cuh

@@ -708,7 +708,7 @@ read_out(const int idx, AcReal* __restrict__ field[], const int3 handle)
  */
 
 ////////////////REDUCE///////////////////////////
-#include "src/core/math_utils.h" // is_power_of_two
+#include "math_utils.h" // is_power_of_two
 
 /*
 Reduction steps: