Include common uarch.h in all backends (fixes include issue)

src/cuda/uarch.cpp

@@ -5,10 +5,10 @@
 #include <cstdio>
 #include <cstring>
 
-#include "../common/uarch.hpp"
 #include "../common/global.hpp"
 #include "../common/gpu.hpp"
 #include "chips.hpp"
+#include "uarch.hpp"
 
 // Any clock multiplier
 #define CM_ANY               -1

src/cuda/uarch.hpp

@@ -2,6 +2,7 @@
 #define __CUDA_UARCH__
 
 #include "../common/gpu.hpp"
+#include "../common/uarch.hpp"
 
 struct uarch;
 

src/hsa/uarch.cpp

@@ -2,9 +2,9 @@
 #include <cstdint>
 #include <cstring>
 
-#include "../common/uarch.hpp"
 #include "../common/global.hpp"
 #include "../common/gpu.hpp"
+#include "uarch.hpp"
 #include "chips.hpp"
 
 // MICROARCH values

src/hsa/uarch.hpp

@@ -2,6 +2,7 @@
 #define __HSA_UARCH__
 
 #include "../common/gpu.hpp"
+#include "../common/uarch.hpp"
 
 struct uarch;
 

src/intel/uarch.cpp

@@ -3,12 +3,12 @@
 #include <cstring>
 #include <cstdio>
 
-#include "../common/uarch.hpp"
 #include "../common/global.hpp"
 #include "../common/gpu.hpp"
 #include "chips.hpp"
 #include "pci.hpp"
 #include "cpuid.hpp"
+#include "uarch.hpp"
 
 // Data not available
 #define NA                   -1

src/intel/uarch.hpp

@@ -2,6 +2,7 @@
 #define __INTEL_UARCH__
 
 #include "../common/gpu.hpp"
+#include "../common/uarch.hpp"
 
 struct uarch;