SWDEV-339113 - Update sampl codes with correct kernel coordinate (#2704)

Change-Id: Ibfc0fd285441cd3d79b312d2b739729a039a6f84

samples/0_Intro/bit_extract/bit_extract.cpp

@@ -35,8 +35,8 @@ THE SOFTWARE.
     }
 
 __global__ void bit_extract_kernel(uint32_t* C_d, const uint32_t* A_d, size_t N) {
-    size_t offset = (hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x);
-    size_t stride = hipBlockDim_x * hipGridDim_x;
+    size_t offset = (blockIdx.x * blockDim.x + threadIdx.x);
+    size_t stride = blockDim.x * gridDim.x;
 
     for (size_t i = offset; i < N; i += stride) {
 #ifdef __HIP_PLATFORM_AMD__

samples/0_Intro/module_api/vcpy_kernel.cpp

@@ -23,6 +23,6 @@ THE SOFTWARE.
 #include "hip/hip_runtime.h"
 
 extern "C" __global__ void hello_world(float* a, float* b) {
-    int tx = hipThreadIdx_x;
+    int tx = threadIdx.x;
     b[tx] = a[tx];
 }

samples/0_Intro/module_api_global/vcpy_kernel.cpp

@@ -28,11 +28,11 @@ __device__ float myDeviceGlobal;
 __device__ float myDeviceGlobalArray[16];
 
 extern "C" __global__ void hello_world(const float* a, float* b) {
-    int tx = hipThreadIdx_x;
+    int tx = threadIdx.x;
     b[tx] = a[tx];
 }
 
 extern "C" __global__ void test_globals(const float* a, float* b) {
-    int tx = hipThreadIdx_x;
+    int tx = threadIdx.x;
     b[tx] = a[tx] + myDeviceGlobal + myDeviceGlobalArray[tx % ARRAY_SIZE];
 }

samples/0_Intro/square/square.hipref.cpp

@@ -38,8 +38,8 @@ THE SOFTWARE.
  */
 template <typename T>
 __global__ void vector_square(T* C_d, const T* A_d, size_t N) {
-    size_t offset = (hipBlockIdx_x * hipBlockDim_x + hipThreadIdx_x);
-    size_t stride = hipBlockDim_x * hipGridDim_x;
+    size_t offset = (blockIdx.x * blockDim_x + threadIdx.x);
+    size_t stride = blockDim.x * gridDim.x;
 
     for (size_t i = offset; i < N; i += stride) {
         C_d[i] = A_d[i] * A_d[i];