#define VC_EXTRALEAN
#define WIN32_LEAN_AND_MEAN
#include <Windows.h>
#include <cstdlib>

#define D3D11_NO_HELPERS
#include <d3d11.h>
#include <dxgi.h>
#include <d3dcompiler.h>

#define ATOM_EDIT 0xC018
#define ATOM_STATIC 0xC019
#define WIDTH 1280
#define HEIGHT 720
#define _FULLSCREEN

typedef struct {
	// DXGI_MODE_DESC BufferDesc;
	UINT Width;
	UINT Height;
	//DXGI_RATIONAL RefreshRate;
	UINT Numerator;
	UINT Denominator;
	DXGI_FORMAT Format;
	DXGI_MODE_SCANLINE_ORDER ScanlineOrdering;
	DXGI_MODE_SCALING Scaling;

	// DXGI_SAMPLE_DESC SampleDesc;
	UINT Count;
	UINT Quality;

	DXGI_USAGE BufferUsage;
	UINT BufferCount;
	HWND OutputWindow;
	BOOL Windowed;
	DXGI_SWAP_EFFECT SwapEffect;
	UINT Flags;
} HG_DXGI_SWAP_CHAIN_DESC;

struct ConstantBufferType
{
	float demoTime;
	float deltaTime;
	float dummy1;
	float dummy2;
};

#pragma bss_seg(".bss")
static HWND hWnd;
static ID3D11Device* device;
static ID3D11DeviceContext* context;
static IDXGISwapChain* swapChain;
static ID3D11Texture2D* backBufferTexture;
static ID3D11RenderTargetView* renderTargetView;
static ID3D11UnorderedAccessView* particleBuffer1UAV;
static ID3D11ShaderResourceView* particleBuffer1SRV;
static ID3D11UnorderedAccessView* particleBuffer2UAV;
static ID3D11ShaderResourceView* particleBuffer2SRV;
static ID3D11Buffer* particleBuffer1;
static ID3D11Buffer* particleBuffer2;
static ID3D11Texture2D* depthBuffer;
static ID3D11Buffer* constantBuffer;
static ID3D11ComputeShader* initShader;
static ID3D11ComputeShader* updateShader;
static ID3D11ComputeShader* drawShader;
static ID3D11VertexShader* particleVertexShader;
static ID3D11GeometryShader* particleGeometryShader;
static ID3D11PixelShader* particlePixelShader;
static ID3D11InputLayout* particleInputLayout;
static ID3D11RenderTargetView* backBufferRtv;
static ID3D11DepthStencilState* depthStencilState;
static ID3D11DepthStencilView* depthStencilView;
static ID3D11BlendState* alphaBlendState;
static pD3DCompile d3dCompile;
static ID3DBlob* shaderBlob;
static void* shaderCode;
static SIZE_T shaderSize;
static D3D11_MAPPED_SUBRESOURCE mappedResource;
static float zero[4];

#pragma data_seg(".d3dcompiler")
static char* d3dcompiler = "d3dcompiler_47.dll";

#pragma data_seg(".swapChainDesc")
#ifdef FULLSCREEN
HG_DXGI_SWAP_CHAIN_DESC swapChainDesc = {
	WIDTH, HEIGHT, 0, 1, DXGI_FORMAT_R16G16B16A16_FLOAT,
	DXGI_MODE_SCANLINE_ORDER_UNSPECIFIED, DXGI_MODE_SCALING_UNSPECIFIED,
	4, 0, DXGI_USAGE_RENDER_TARGET_OUTPUT | DXGI_USAGE_SHADER_INPUT, 1, 0, 0, DXGI_SWAP_EFFECT_DISCARD, 0
};
#else
HG_DXGI_SWAP_CHAIN_DESC swapChainDesc = {
	WIDTH, HEIGHT, 0, 1, DXGI_FORMAT_R16G16B16A16_FLOAT,
	DXGI_MODE_SCANLINE_ORDER_UNSPECIFIED, DXGI_MODE_SCALING_UNSPECIFIED,
	4, 0, DXGI_USAGE_RENDER_TARGET_OUTPUT | DXGI_USAGE_SHADER_INPUT, 1, 0, -1, DXGI_SWAP_EFFECT_DISCARD, 0
};
#endif

#pragma data_seg(".depthBufferDesc")
static D3D11_TEXTURE2D_DESC depthBufferDesc =
{
	WIDTH,
	HEIGHT,
	1,
	1,
	DXGI_FORMAT_D32_FLOAT,
	{
		1, 0
	},
	D3D11_USAGE_DEFAULT,
	D3D11_BIND_DEPTH_STENCIL,
	0,
	0
};

#pragma data_seg(".constantBufferDesc")
static D3D11_BUFFER_DESC constantBufferDesc =
{
	16,
	D3D11_USAGE_DYNAMIC,
	D3D11_BIND_CONSTANT_BUFFER,
	D3D11_CPU_ACCESS_WRITE,
	0,
	0
};

#pragma data_seg(".particleBufferDesc")
static D3D11_BUFFER_DESC particleBufferDesc =
{
	1024 * 1024 * 48,
	D3D11_USAGE_DEFAULT,
	D3D11_BIND_UNORDERED_ACCESS | D3D11_BIND_SHADER_RESOURCE,
	0,
	D3D11_RESOURCE_MISC_BUFFER_STRUCTURED,
	48
};

#pragma data_seg(".particleBufferDesc")
static D3D11_SHADER_RESOURCE_VIEW_DESC particleSrvDesc =
{
	DXGI_FORMAT_UNKNOWN,
	D3D11_SRV_DIMENSION_BUFFER,
	{
		0,
		1024 * 1024
	}
};

#pragma data_seg(".particleUavDesc")
static D3D11_UNORDERED_ACCESS_VIEW_DESC particleUavDesc =
{
	DXGI_FORMAT_UNKNOWN,
	D3D11_UAV_DIMENSION_BUFFER,
	0,
	1024 * 1024,
	D3D11_BUFFER_UAV_FLAG_APPEND
};

#pragma data_seg(".vertexUavDesc")
static D3D11_UNORDERED_ACCESS_VIEW_DESC vertexUavDesc =
{
	DXGI_FORMAT_UNKNOWN,
	D3D11_UAV_DIMENSION_BUFFER,
	0,
	1024 * 1024,
	0
};

#pragma data_seg(".particleInputElementDescs")
static D3D11_INPUT_ELEMENT_DESC particleInputElementDescs[] =
{
	{
		"SV_POSITION",
		0,
		DXGI_FORMAT_R32G32B32A32_FLOAT,
		0,
		0,
		D3D11_INPUT_PER_VERTEX_DATA,
		0
	},
	{
		"LENGTH",
		0,
		DXGI_FORMAT_R32_FLOAT,
		0,
		16,
		D3D11_INPUT_PER_VERTEX_DATA,
		0
	},
	{
		"AGE",
		0,
		DXGI_FORMAT_R32_FLOAT,
		0,
		20,
		D3D11_INPUT_PER_VERTEX_DATA,
		0
	},
};

#pragma data_seg(".backBufferRtvDesc")
static D3D11_RENDER_TARGET_VIEW_DESC backBufferRtvDesc =
{
	DXGI_FORMAT_R16G16B16A16_FLOAT,
	D3D11_RTV_DIMENSION_TEXTURE2DMS,
	{
		0
	}
};

#pragma data_seg(".depthStencilStateDesc")
static D3D11_DEPTH_STENCIL_DESC depthStencilStateDesc =
{
	FALSE, // TRUE
	D3D11_DEPTH_WRITE_MASK_ALL,
	D3D11_COMPARISON_LESS,
	FALSE,
	0,
	0,
	{},
	{}
};

#pragma data_seg(".swapChainViewport")
static D3D11_VIEWPORT swapChainViewport =
{
	0,
	0,
	WIDTH,
	HEIGHT,
	0.0f,
	1.0f,
};

#pragma data_seg(".depthStencilViewDesc")
static D3D11_DEPTH_STENCIL_VIEW_DESC depthStencilViewDesc =
{
	DXGI_FORMAT_D32_FLOAT,
	D3D11_DSV_DIMENSION_TEXTURE2D,
	0,
	{
		0
	}
};

#pragma data_seg(".alphaBlendDesc")
static D3D11_BLEND_DESC alphaBlendDesc = 
{
	FALSE,
	FALSE,
	{
		TRUE,
		D3D11_BLEND_SRC_ALPHA,
		D3D11_BLEND_INV_SRC_ALPHA,
		D3D11_BLEND_OP_ADD,
		D3D11_BLEND_ONE,
		D3D11_BLEND_ONE,
		D3D11_BLEND_OP_ADD,
		D3D11_COLOR_WRITE_ENABLE_ALL,
	}
};