Added proper device data types

Change-Id: I42029635ff68c3c13a764a3eda6447e6c77878c6


[ROCm/hip commit: 39910029a6]

projects/hip/include/hip/hcc_detail/device_functions.h

@@ -20,13 +20,18 @@ THE SOFTWARE.
 #ifndef HIP_HCC_DETAIL_DEVICE_FUNCTIONS_H
 #define HIP_HCC_DETAIL_DEVICE_FUNCTIONS_H
 
-#include "hip_runtime.h"
+#include <hip/hip_runtime.h>
+#include <hip/hip_vector_types.h>
 
 __device__ float __int_as_float (int x);
 
 __device__ double __hiloint2double (int hi, int lo);
 
-extern __HIP_DEVICE__ double  __longlong_as_double(long long int x);
-extern __HIP_DEVICE__ long long int __double_as_longlong(double x);
+__device__ char4 __hip_hc_add8pk(char4, char4);
+__device__ char4 __hip_hc_sub8pk(char4, char4);
+__device__ char4 __hip_hc_mul8pk(char4, char4);
+
+extern __device__ double  __longlong_as_double(long long int x);
+extern __device__ long long int __double_as_longlong(double x);
 
 #endif

projects/hip/include/hip/hcc_detail/hip_ldg.h

@@ -25,8 +25,8 @@ THE SOFTWARE.
 
 #if __HCC__
 #if __hcc_workweek__ >= 16164
-#include "hip/hip_vector_types.h"
-#include "hip/hcc_detail/host_defines.h"
+#include "hip_vector_types.h"
+#include "host_defines.h"
 
 
 __device__ char                 __ldg(const char* );
@@ -75,4 +75,3 @@ __device__ double2              __ldg(const double2* );
 #endif  // __HCC__
 
 #endif  // HIP_LDG_H
-

projects/hip/include/hip/hcc_detail/hip_runtime.h

@@ -46,6 +46,7 @@ THE SOFTWARE.
 #define CUDA_SUCCESS hipSuccess
 
 #include <hip/hip_runtime_api.h>
+
 //#include "hip/hcc_detail/hip_hcc.h"
 //---
 // Remainder of this file only compiles with HCC
@@ -815,9 +816,6 @@ extern "C" __device__ void* __hip_hc_free(void *ptr);
 //extern "C" __device__ void* malloc(size_t size);
 //extern "C" __device__ void* free(void *ptr);
 
-extern "C" __device__ char4 __hip_hc_add8pk(char4, char4);
-extern "C" __device__ char4 __hip_hc_sub8pk(char4, char4);
-extern "C" __device__ char4 __hip_hc_mul8pk(char4, char4);
 
 #define __syncthreads() hc_barrier(CLK_LOCAL_MEM_FENCE)