Merge branch 'amd-develop' into amd-master

Change-Id: I53d5a8916d769c4f0fe60d2ee3b240551da80b4f
(cherry picked from commit 01c523f6c9)

tests/src/deviceLib/hipDeviceMemcpy.cpp

@@ -1,18 +1,29 @@
-#include<iostream>
+#include <iostream>
 #include "hip/hip_runtime.h"
 #include "hip/hip_runtime_api.h"
+#include "../test_common.h"
+
 
 #define LEN 1030
 #define SIZE LEN << 2
 
-__global__ void cpy(hipLaunchParm lp, uint32_t *Out, uint32_t *In, uint32_t *Vald)
+/* HIT_START
+ * BUILD: %t %s ../test_common.cpp
+ * RUN: %t
+ * HIT_END
+ */
+
+
+__global__ void cpy(hipLaunchParm lp, uint32_t *Out, uint32_t *In)
 {
-    memcpy(Out, In, SIZE, Vald);
+    int tx = hipThreadIdx_x;
+    memcpy(Out + tx, In + tx, SIZE/LEN);
 }
 
 __global__ void set(hipLaunchParm lp, uint32_t *ptr, uint8_t val, size_t size)
 {
-    memset(ptr, val, size);
+    int tx = hipThreadIdx_x;
+    memset(ptr + tx, val, size);
 }
 
 int main()
@@ -24,19 +35,29 @@ int main()
     Val = new uint32_t;
     *Val = 0;
     for(int i=0;i<LEN;i++){
-        A[i] = i *1.0f;
-        B[i] = 0.0f;
+        A[i] = i;
+        B[i] = 0;
     }
     hipMalloc((void**)&Ad, SIZE);
     hipMalloc((void**)&Bd, SIZE);
-    hipMalloc((void**)&Vald, sizeof(uint32_t));
     hipMemcpy(Ad, A, SIZE, hipMemcpyHostToDevice);
-    hipLaunchKernel(cpy, dim3(1), dim3(LEN/4), 0, 0, Bd, Ad, Vald);
-    hipLaunchKernel(set, dim3(1), dim3(LEN/4), 0, 0, Bd, 0x1, SIZE);
+
+    hipLaunchKernel(cpy, dim3(1), dim3(LEN), 0, 0, Bd, Ad);
+
     hipMemcpy(B, Bd, SIZE, hipMemcpyDeviceToHost);
-    hipMemcpy(Val, Vald, sizeof(uint32_t), hipMemcpyDeviceToHost);
     for(int i=LEN-16;i<LEN;i++){
-        std::cout<<A[i]<<" "<<B[i]<<std::endl;
+        if(A[i]!=B[i]){
+            return 0;
+        }
     }
-    std::cout<<*Val<<std::endl;
+    hipLaunchKernel(set, dim3(1), dim3(LEN), 0, 0, Bd, 0x1, LEN);
+
+    hipMemcpy(B, Bd, SIZE, hipMemcpyDeviceToHost);
+    for(int i=LEN-16;i<LEN;i++){
+        if(0x01010101!=B[i]){
+            return 0;
+        }
+    }
+ 
+    passed();
 }