Test - Use compute shader for combine stage

test/demo.srd

@@ -17,6 +17,7 @@
 
 	(program prg-fullscreen "fullscreen.v.glsl")
 
+	(set use-compute 1)
 	(call scene-init)
 	(call dof-init)
 	(call bloom-init)
@@ -410,9 +411,15 @@
 		(lod 0 100))
 	(texture tx-combined rgba-nu8 $vp-width $vp-height)
 	(odbg tx-combined ldr-alpha "Combined output")
-	(framebuffer rt-combined tx-combined)
+	(if $use-compute (
-	(program prg-combine "combine.f.glsl")
+		(program prg-combine "combine.c.glsl")
-	(pipeline pl-combine prg-fullscreen prg-combine)
+		(pipeline pl-combine prg-combine)
+		(set combine-compute-size 16)
+	)(
+		(framebuffer rt-combined tx-combined)
+		(program prg-combine "combine.f.glsl")
+		(pipeline pl-combine prg-fullscreen prg-combine)
+	))
 
 	# Bloom parameters
 	(input bloom-strength .45)
@@ -472,7 +479,9 @@
 		(uniforms prg-combine 2 $vp-width $vp-height)
 
 		(use-pipeline pl-combine)
-		(use-framebuffer rt-combined)
+		(if (not $use-compute) (
+			(use-framebuffer rt-combined)
+		))
 		(use-texture 0 in-main smp-combine-input)
 		(use-texture 1 in-bloom smp-combine-bloom)
 		(uniforms prg-combine 3
@@ -498,8 +507,17 @@
 			(get-input cg-gamma-r)
 			(get-input cg-gamma-g)
 			(get-input cg-gamma-b))
-		(viewport 0 0 $vp-width $vp-height)
+
-		(fullscreen)
+		(if $use-compute (
+			(image 0 tx-combined 0)
+			(compute
+				(add 1 (div $vp-width $combine-compute-size))
+				(add 1 (div $vp-height $combine-compute-size))
+				1)
+		)(
+			(viewport 0 0 $vp-width $vp-height)
+			(fullscreen)
+		))
 	)
 )
 

test/shaders/combine.c.glsl

@@ -0,0 +1,63 @@
+#version 450 core
+
+layout( local_size_x = 16 , local_size_y = 16 ) in;
+
+//! type compute
+//! include lib/utils.glsl
+
+layout( location = 0 ) uniform sampler2D u_MainInput;
+layout( location = 1 ) uniform sampler2D u_BlurInput;
+layout( location = 2 ) uniform vec2 u_OutputSize;
+layout( location = 3 ) uniform vec2 u_Bloom;
+layout( location = 4 ) uniform vec4 u_Vignette1;
+layout( location = 5 ) uniform vec2 u_Vignette2;
+layout( location = 6 ) uniform vec3 u_ColorLift;
+layout( location = 7 ) uniform vec3 u_ColorGain;
+layout( location = 8 ) uniform vec3 u_ColorGamma;
+
+layout( binding = 0 , rgba8 ) writeonly uniform image2D ub_Output;
+
+#define uVigShapeMul (u_Vignette1.x)
+#define uVigShapePow (u_Vignette1.y)
+#define uVigAspect (u_Vignette1.z)
+#define uVigShapeReverse (u_Vignette1.w)
+#define uVigApplyBase (u_Vignette2.x)
+#define uVigApplyMax (u_Vignette2.y)
+
+void main( void )
+{
+	ivec2 coords = ivec2( gl_GlobalInvocationID.xy );
+	vec2 pos = vec2( coords ) / u_OutputSize;
+	float f = u_Bloom.x;
+	vec3 color = textureLod( u_MainInput , pos , 0 ).rgb;
+
+	// Bloom
+	vec3 bloom = vec3( 0 );
+	for ( int i = 0 ; i < 6 ; i ++ ) {
+		bloom += f * textureLod( u_BlurInput , pos , i ).rgb;
+		f = pow( f , u_Bloom.y + 1 );
+	}
+	color = color + bloom / 2.2;
+
+	// Color grading
+	color = ( color * ( u_ColorGain - u_ColorLift ) ) + u_ColorLift;
+
+	// Vignette
+	vec2 vShape = pow( abs( pos * 2 - 1 ) * uVigShapeMul ,
+		vec2( uVigShapePow ) );
+	float vignette = 1 - pow(
+		2 * ( uVigAspect * vShape.x + ( 1 - uVigAspect) * vShape.y ) ,
+		1 / ( uVigShapePow * uVigShapeReverse ) );
+	color *= uVigApplyBase + smoothstep(
+		uVigApplyBase , uVigApplyBase + uVigApplyMax , vignette );
+
+	// Reinhart
+	color = color / ( color + 1. );
+
+	// Gamma
+	color = pow( color , vec3( 1 ) / ( u_ColorGamma + 2.2 ) );
+
+	// Write it
+	imageStore( ub_Output , coords ,
+			vec4( color , M_Luminosity( color ) ) );
+}