Merge branch 'cmakelist_rewrite_and_C_API_conformity_07-26' into node_device_interface_revision_07-23

src/core/CMakeLists.txt

@@ -12,8 +12,7 @@ set(CUDA_ARCH_FLAGS -gencode arch=compute_37,code=sm_37
                     -gencode arch=compute_61,code=sm_61
                     -lineinfo
                     -ftz=true # Flush denormalized floats to zero
-                    -std=c++11
-                    --compiler-options -march=native) # Native host machine code
+                    -std=c++11)
                     #--maxrregcount=255
                     # -Xptxas -dlcm=ca opt-in to cache all global loads to L1/texture cache
                     # =cg to opt out

src/core/kernels/kernels.cuh

@@ -659,7 +659,7 @@ read_out(const int idx, AcReal* __restrict__ field[], const int3 handle)
 }
 
 #define WRITE_OUT(handle, value) (write(buffer.out, handle, idx, value))
-#define READ(handle) (read_data(vertexIdx, buffer.in, handle))
+#define READ(handle) (read_data(vertexIdx, globalVertexIdx, buffer.in, handle))
 #define READ_OUT(handle) (read_out(idx, buffer.out, handle))
 
 // also write for clarity here also, not for the DSL