refactor build, clean up

[ROCm/composable_kernel commit: e80fbbdd71]

src/include/device.hpp

@@ -0,0 +1,30 @@
+#pragma once
+#include <memory>
+
+struct DeviceMem
+{
+    DeviceMem() = delete;
+    DeviceMem(std::size_t mem_size);
+    void* GetDeviceBuffer();
+    void ToDevice(const void* p);
+    void FromDevice(void* p);
+    ~DeviceMem();
+
+    void* mpDeviceBuf;
+    std::size_t mMemSize;
+};
+
+struct KernelTimerImpl;
+
+struct KernelTimer
+{
+    KernelTimer();
+    ~KernelTimer();
+    void Start();
+    void End();
+    float GetElapsedTime() const;
+
+    std::unique_ptr<KernelTimerImpl> impl;
+};
+
+void launch_kernel(const void* func, dim3 grid_dim, dim3 block_dim, void** args, float& time);