Difference between revisions of "Blending functions"

Revision as of 14:22, 10 September 2015

Here are all the Photoshop Blending Modes, in easy to use functions for various languages.

HLSL

Category 1 (Darker)

float4 Darken (float4 cBase, float4 cBlend)
{
	float4 cNew;
	cNew.rgb = min(cBase.rgb, cBlend.rgb);
	cNew.a = 1.0;
	return cNew;
}

float4 Multiply (float4 cBase, float4 cBlend)
{
	return (cBase * cBlend);
}

float4 ColorBurn (float4 cBase, float4 cBlend)
{
	return (1 - (1 - cBase) / cBlend);
}

float4 LinearBurn (float4 cBase, float4 CBlend)
{
	return (cBase + cBlend - 1);
}

Category 2 (Lighter)

float4 Lighten (float4 cBase, float4 cBlend)
{
	float4 cNew;
	cNew.rgb = max(cBase.rgb, cBlend.rgb);
	cNew.a = 1.0;
	return cNew;
}

float4 Screen (float4 cBase, float4 cBlend)
{
	return (1 - (1 - cBase) * (1 - cBlend));
}

float4 ColorDodge (float4 cBase, float4 cBlend)
{
	return (cBase / (1 - cBlend));
}

float4 LinearDodge (float4 cBase, float4 cBlend)
{
	return (cBase + cBlend);
}

Category 3 (Complex)

Condintional statements in shaders

GPUs hate branching, so it's better to avoid if..else in your shaders. The promblem here is that GPU will execute both blocks of code anyway - for then and else respectively - eventually a part of calculations will be thrown away (the more unnecessary instructions executed, the lower shader performance will be). In most of the cases conditional statements could be substituted with a help of HLSL's intrinsic functions.

Let's optimize the following piece of code:

	float4 cNew;
	if (cBase.r > .5) { cNew.r = 1 - (1 - 2 * (cBase.r - .5)) * (1 - cBlend.r); }
	else { cNew.r = (2 * cBase.r) * cBlend.r; }
	
	if (cBase.g > .5) { cNew.g = 1 - (1 - 2 * (cBase.g - .5)) * (1 - cBlend.g); }
	else { cNew.g = (2 * cBase.g) * cBlend.g; }
	
	if (cBase.b > .5) { cNew.b = 1 - (1 - 2 * (cBase.b - .5)) * (1 - cBlend.b); }
	else { cNew.b = (2 * cBase.b) * cBlend.b; }

As we see the result values of R, G and B components (except for alpha) for this blending mode depends on it's base layer initial values

	if (baseColor > .5) then {
	    resultColor = A; 
	} else {
	    resultColor = B;
	}

Using of lerp(A, B, interpolation_value) - linear interpolation - function makes it simpler to switch between two branches.

	float4 newColor = lerp(A, B, interpolation_value); // = A*(1 - interpolation_value) + B*interpolation_value

Actually we don't want newColor to gradually change between A and B, but only switch A <-> B, so interpolation_value should be strictly equal to 0 or 1. And here is where the step(x, y) function come in. The equivalent for step functionality is:

	float step(x, y) {
	    if (x > y) then {
	        return 0;
	    } else {
	        return 1;
	    }
	}

Apparently, if it's required to invert output values, just change the order of step input arguments. Let's assembly all the information we've found out and re-write our shader:

	float isLessOrEq = step(cBase, 0.5);
	float4 cNew = lerp(1. - (1. - 2.*(cBase - .5))*(1. - cBlend), 2*cBase*cBlend, isLessOrEq);	// get rid of per component calculations and using vectorization
													// for performance optimization sake
	cNew.a = 1.;	// restore alpha

P.S. In a special case, when one of the possible values of cNew is 0, there is no more need in lerp

	float isLessOrEq = step(cBase, 0.5);
	float4 cNew = isLessOrEq * 2*cBase*cBlend;	// = (x <= 0.5) ? 2*cBase*cBlend : 0;

P.P.S. Although compiler will do this work for you at the optimization stage, keep in mind that explicit declaration variables as float (1.0, 1. or 1.f insted of 1) is more preferable then writing expressions with variables of different types and, in result, mixing float with int. It causes implicit type conversion, which sometimes leads to unforeseen consequences.

Shaders implementation

float4 Overlay (float4 cBase, float4 cBlend)
{
	float isLessOrEq = step(cBase, .5);
	float4 cNew = lerp(2*cBlend*cBase, 1 - (1 - 2*(Base - .5))*(1 - Blend), isLessOrEq);
	cNew.a = 1.0;
	return cNew;
}

float4 SoftLight (float4 cBase, float4 cBlend)
{
	float isLessOrEq = step(cBlend, .5);
	float4 cNew = lerp(1 - (1 - cBase)*(1 - 2*cBlend*cBase), cBase*(1 - (1 - cBase)*(1 - 2*cBlend)), isLessOrEq);
	cNew.a = 1.0;
	return cNew;
}

float4 HardLight (float4 cBase, float4 cBlend)
{
	float isLessOrEq = step(cBlend, .5);
	float4 cNew = lerp(1 - (1 - cBase)*(1 - 2*cBlend), 2*cBlend*cBase, isLessOrEq);
	cNew.a = 1.0;
	return cNew;
}

float4 VividLight (float4 cBase, float4 cBlend)
{
	float isLessOrEq = step(cBlend, .5);
	float4 cNew = lerp(1 - (1 - cBase)/(2*(cBlend - .5)), cBase/(1 - 2*cBlend), isGreaterThen);
	cNew.a = 1.0;
	return cNew;
}

float4 LinearLight (float4 cBase, float4 cBlend)
{
	float isLessOrEq = step(cBlend, .5);
	float4 cNew = lerp(cBase + 2*(cBlend - .5), cBase + 2*cBlend - 1., isLessOrEq);
	cNew.a = 1.0;
	return cNew;
}

float4 PinLight (float4 cBase, float4 cBlend)
{
	float isLessOrEq = step(cBlend, .5);
	float4 cNew = lerp(max(cBase, 2*(cBlend - .5)), min(cBase, 2*cBlend), isLessOrEq);
	cNew.a = 1.0;
	return cNew;
}

Category 4 (Comparison)

float4 Difference (float4 cBase, float4 cBlend)
{
	return (abs(cBase - cBlend));
}

float4 Exclusion (float4 cBase, float4 cBlend)
{
	return (.5 - 2 * (cBase - .5) * (cBlend - .5));
}