Faster shader based dxDrawCircle

dxDrawCircle.fx

float4x4 gWorldViewProjection : WORLDVIEWPROJECTION;
struct VSInput
{
    float3 Position : POSITION0;
    float4 Diffuse : COLOR0;
    float2 TexCoord : TEXCOORD0;
};
struct PSInput
{
    float4 Position : POSITION0;
    float4 Diffuse : COLOR0;
    float4 Diffuse2 : COLOR1;
    float2 TexCoord : TEXCOORD0;
    float2 Settings : TEXCOORD1;
};
PSInput VertexShaderFunction(VSInput VS)
{
    PSInput PS = (PSInput)0;
    PS.Position = mul(float4(VS.Position, 1), gWorldViewProjection);

    int bAntialias = floor(VS.Diffuse.a*255+0.5)%2;
    int bTwoColors = floor(VS.Diffuse.a*127+0.5)%2;
    [branch] if(bTwoColors == 1){
        float aComp = floor(VS.Diffuse.a*63+0.5)%64;;
        PS.Diffuse = float4((floor(VS.Diffuse.rgb*15+0.5)%16)/15, (aComp%8)/7);
        PS.Diffuse2 = float4(floor((VS.Diffuse.rgb*255+0.5)%16)/15, ((aComp/8)%8)/7);
    } else {
        PS.Diffuse = float4(VS.Diffuse.rgb, floor(VS.Diffuse.a*63+0.5)%64/63);
        PS.Diffuse2 = PS.Diffuse;
    }

    PS.TexCoord = VS.TexCoord;
    PS.Settings = float2(bTwoColors,bAntialias);

    return PS;
}


float4 PixelShaderFunction(PSInput PS) : COLOR0
{
    const float dist = length(PS.TexCoord*2-1);
    float4 color = PS.Settings.x > 0.5 ? lerp(PS.Diffuse, PS.Diffuse2, dist) : PS.Diffuse;

    if (PS.Settings.y < 0.5){
        return dist <= 1 ? color : float4(0,0,0,0);
    }

    return float4(color.rgb, smoothstep(1, 1 - fwidth(dist) * 2, dist)*color.a);
}
technique Technique1
{
    pass Pass1
    {
        SrcBlend = SrcAlpha;
        DestBlend = InvSrcAlpha;
        VertexShader = compile vs_3_0 VertexShaderFunction();
        PixelShader = compile ps_3_0 PixelShaderFunction();
    }
}

dxDrawCircle.lua

local floor = math.floor
function fromColor(col)
	local r,g,b,a
	b = col%256
	g = floor(col/0x100)%256
	r = floor(col/0x10000)%256
	a = floor(col/0x1000000)%256
	return r,g,b,a
end

local shader = dxCreateShader( "dxDrawCircle.fx" )
local shaderColCache = {}
--dxDrawCircleShader(posX, posY, radius, innerColor, outerColor, antialiasing)
function dxDrawCircleShader(posX,posY,radius,color1,color2,antialias)
	assert(type(color1) == "number", "Expected number as color argument to dxDrawCircleShader")
	if (not color2) or (color2 == color1) then color2 = "def" end

	local finColor
	local col1 = shaderColCache[color1]
	if col1 then
		local col2 = col1[color2]
		if col2 then
			finColor = col2
		end
	end

	if not finColor then
		shaderColCache[color1] = shaderColCache[color1] or {}
		local twoColors = (color2 ~= "def")
		if twoColors then
			local r1,g1,b1,a1 = fromColor(color1)
			local r2,g2,b2,a2 = fromColor(color2)
			local finR = floor((r1/256)*16)*16 + floor((r2/256)*16)
			local finG = floor((g1/256)*16)*16 + floor((g2/256)*16)
			local finB = floor((b1/256)*16)*16 + floor((b2/256)*16)
			local finA = floor((a1/256)*8)*32 + floor((a2/256)*8)*4 + 2
			finColor = tocolor( finR, finG, finB, finA )
		else
			local r,g,b,a = fromColor(color1)
			local finA = (floor((a/256)*64)*4)%256
			finColor = tocolor( r,g,b, finA )
		end
		shaderColCache[color1][color2] = finColor
	end

	local antialiasComp = 16777216
	if not antialias then antialiasComp = 0 end
	finColor = finColor + antialiasComp
	dxDrawImage( posX-radius, posY-radius, radius*2, radius*2, shader, nil, nil,  finColor )
end