#define WIDTH 512
#define HEIGHT 424

#pragma kernel CSMain

#include "UnityCG.cginc"

RWStructuredBuffer<float4> KinectPositionBuffer;
RWStructuredBuffer<float4> KinectVelocityBuffer;
RWStructuredBuffer<float4> KinectGridBuffer;
RWStructuredBuffer<float4> FinalPositionBuffer;
RWStructuredBuffer<float4> FinalVelocityBuffer;

AppendStructuredBuffer<float4> HomePositionBuffer;
AppendStructuredBuffer<float4> HomeVelocityBuffer;

float4 ClipMin;
float4 ClipMax;
float4x4 ModelMatrix;
float4 FluidSize;
float4 FluidRoot;
float4 FluidDim;
int flipX;
int NumKinects;

[numthreads(512, 1, 1)]
void CSMain (uint3 id : SV_DispatchThreadID)
{
	//uint2 tex_coord = uint2(id.x % WIDTH, id.x / WIDTH);
	 
	float4 pos = FinalPositionBuffer[id.x];
	//pos.y *= -1;
	pos.z *= -1;
	pos.x = lerp(pos.x, -pos.x, flipX);

	float pos_mag = length(pos.xyz);

	{
		pos = mul(pos, ModelMatrix);

		float4 vel = FinalVelocityBuffer[id.x];
		//vel = float4(0, 0, 0.01, 0);
		//vel.y *= -1;
		vel.z *= -1;
		vel.x = lerp(vel.x, -vel.x, flipX);

		float3 grid_pos = (pos.xyz - FluidRoot.xyz) / FluidSize.xyz;
		int grid_index = -1;
		if (grid_pos.x > 0 && grid_pos.x < 1 &&
			grid_pos.y > 0 && grid_pos.y < 1 &&
			grid_pos.z > 0 && grid_pos.z < 1 &&
			pos_mag > 0.0f)// && length(pos) > 10.0f)
		{
			grid_pos *= FluidDim.xyz;
			grid_index = (int)grid_pos.x + (int)grid_pos.y * FluidDim.x + (int)grid_pos.z * FluidDim.x * FluidDim.y;
			KinectGridBuffer[grid_index] += float4(vel.xyz, 1);

			if (pos_mag > 0.1f)// && pos.y < 0.5f)
			{
				//pos.y = 0.0f;
				//vel.xyz = float3(0, 0, 0);

				HomePositionBuffer.Append(float4(pos.xyz, 0));
				HomeVelocityBuffer.Append(float4(vel.xyz, 0));
			}

		}

		KinectPositionBuffer[id.x] = float4(pos.xyz, grid_index);
		KinectVelocityBuffer[id.x] = float4((pos_mag == 0.0f) ? float3(0, 0, 0) : vel.xyz, 0);

	}

}