Fix CUDA linkage issues with bitstream

Author: lindstro

src/cuda/interface.cu

@@ -1,5 +1,4 @@
 #include <iostream>
-#include "interface.h"
 #include "../share/device.h"
 #include "shared.cuh"
 #include "error.cuh"
@@ -16,6 +15,7 @@
 #include "decode1.cuh"
 #include "decode2.cuh"
 #include "decode3.cuh"
+#include "interface.h"
 
 zfp_bool
 zfp_internal_cuda_init(zfp_exec_params_cuda* params)

src/cuda/shared.cuh

@@ -4,22 +4,21 @@
 // report throughput; set via CMake
 // #define ZFP_WITH_CUDA_PROFILE 1
 
+// we need to know about bitstream, but we don't want duplicate symbols
+#ifndef inline_
+  #define inline_ static inline
+#endif
+#include "zfp/bitstream.inl"
+#include "zfp.h"
+
 #include <cmath>
 #include <cstdio>
-#include "zfp.h"
 #include "traits.cuh"
 #include "constants.cuh"
 #ifdef ZFP_WITH_CUDA_PROFILE
   #include "timer.cuh"
 #endif
 
-// we need to know about bitstream, but we don't want duplicate symbols
-#ifndef inline_
-  #define inline_ static inline
-#endif
-
-#include "zfp/bitstream.inl"
-
 // bit stream word/buffer type; granularity of stream I/O operations
 typedef unsigned long long Word;