AlphaBlend Formula/Algorithm

[fabiano.salles]

Hello, guys!

I'm using VCL's TBitmap with PixelFormat = pf32bit and using the 4st of the pixel byte as my alpha channel. Instead of using the TRGBQuad record of delphi, I'm using the following datatypes:
[pascal]type
PRGB32 = ^TRGB32;
TRGB32 = packed record
B, G, R, A : byte;
end;
[/pascal]
Ok. It woks just fine and I can load and save bitmaps with a valid alpha channel.
Now I'm trying to blend pixels using the alpha information, but the formula I'm usiging is producing a strange result. Could anyone, please, tell me a valid formula to blend pixels using alpha channel or say what I'm doing wrong? This is what my code looks like:

[pascal]procedure TfrmCGScreen.DrawScene;
var
x, y, m : integer;
p : PRGB32;
begin

//Just creating a background pattern...
for x:=0 to fBuffer.Width-1 do
for y:=0 to fBuffer.Height-1 do
with fBufferFisrtLine[Y*fBufferLineLength+X] do
begin
B := 0;
R := 0;
A := 0;
m := Tag div fBuffer.Height;
case m of
0 : B := ((x or y) + Tag) mod 255;
1 : G := ((x and y) + Tag) mod 255;
2 : R := ((x or y) + Tag) mod 255;
else
begin
G := ((x xor y) + Tag) mod 255;
tag := 0;
end;
end;
end;

//now I'm trying to blend the "fBlendImage" with the image in my backffuer
for y:=0 to fBlendImage.Height-1 do
begin
p := fBlendImage.ScanLine[y];
for x:= 0 to fBlendImage.Width -1 do
begin
with fBufferFisrtLine[Y*fBufferLineLength+ (X + (tag mod 600))] do
begin
B := B + p.B * p.A div 255;
G := G + p.G * p.A div 255;
R := R + p.R * p.A div 255;
end;
Inc(p);
end;
end;
Tag := Tag+1;
end;[/pascal]

Thanks a lot!

[chronozphere]

Let's say B is the background pixel and S is the sourcepixel and D is the destenation pixel.

I would try this:

[pascal]
D.R := B.R + (S.R - B.R)*(S.A/255);
D.G := B.G + (S.R - B.G)*(S.A/255);
D.B := B.B + (S.R - B.B)*(S.A/255);
[/pascal]

It's just weighted interpolation.
I've not tested this. Just made it up.

Hope this helps....

[fabiano.salles]

Hey man, thanks for the fast reply!

Your interpolation works only for the 100% black pixels on the alpha channel (A=255). Changing it for the code below, it works for the shades of gray too.

[pascal]D.R := B.R + (S.R - B.R) * S.A div 255;
D.G := B.G + (S.R - B.G) * S.A div 255;
D.B := B.B + (S.R - B.B) * S.A div 255;
[/pascal]

[chronozphere]

Hmm.. ofcourse. Using / turns the result into a float, and the components of D are bytes.

Does your sollution with DIV really work. DIV's result can only be an integer. :?
I assume that your alpha value is a byte (0-255). So dividing it by 255, will give you a result inside the 0..1 range. If you use DIV, i assume all results in this range will be rounded to zero.

I think this will be more accurate:

//Include the math unit to use "Floor"
D.R := Floor(B.R + (S.R - B.R)*(S.A/255));
D.G := Floor(B.G + (S.R - B.G)*(S.A/255));
D.B := Floor(B.B + (S.R - B.B)*(S.A/255));

Edit: I noticed you removed the brackets around "S.A div 255". I guess that'll make the difference. But could you try my snippet? I wonder wether it works.

[User137]

[pascal]D.R := B.R + (S.R - B.R) * S.A div 255;
D.G := B.G + (S.R - B.G) * S.A div 255;
D.B := B.B + (S.R - B.B) * S.A div 255;
[/pascal]
This code definitely should work. Division is last operation there so there will not be any math errors. Using function to round float to int makes it just slower.

[chronozphere]

Ok, thanks. Good to know.

[fabiano.salles]

User137 is rigth, chronozphere.

As you ask me, I tested your other suggestion and it worked too (but the speed difference is very big between the two approaches).
If you wanna see the code, you can download it here:
http://www.fabianosalles.eti.br/blog...endTeste01.rar

ps: use the '1' and '2' keys to switch between the calculations. Here in my machine it run with ~400fps, but as I'm using only GDI and VCL, the runtime speed may vary too much.

Thanks

[chronozphere]

Your demo looks really cool. I like the animated background.

Switching between the modes doesn't make a real difference for me. The FPSses are varying between 380 and 420 (for both modes). It might be, because things like rounding and floating-point divisions are alot faster on my CPU than on yours. :?

What CPU do you have and what kind of speed difference do you see in your FPSses? I have an AMD 2X 64 5400+.

Thanks.

[fabiano.salles]

I run this last code only on the work computer and here I have a variation beteewn 490 and 520 fps, but when I swicth the "blend mode" ( :lol: ), it drops to ~300fps.