bgfx/examples/39-assao/cs_assao_smart_blur_wide.sc

/*
 * Copyright 2018 Attila Kocsis. All rights reserved.
 * License: https://github.com/bkaradzic/bgfx/blob/master/LICENSE
 */

#include "bgfx_compute.sh" 
#include "uniforms.sh"

IMAGE2D_ARRAY_WR(s_target, rg8, 0);
SAMPLER2DARRAY(s_blurInput, 1);

// unpacking for edges; 2 bits per edge mean 4 gradient values (0, 0.33, 0.66, 1) for smoother transitions!
vec4 UnpackEdges( float _packedVal )
{
    uint packedVal = uint(_packedVal * 255.5);
    vec4 edgesLRTB;
    edgesLRTB.x = float((packedVal >> 6) & 0x03) / 3.0;          // there's really no need for mask (as it's an 8 bit input) but I'll leave it in so it doesn't cause any trouble in the future
    edgesLRTB.y = float((packedVal >> 4) & 0x03) / 3.0;
    edgesLRTB.z = float((packedVal >> 2) & 0x03) / 3.0;
    edgesLRTB.w = float((packedVal >> 0) & 0x03) / 3.0;

    return saturate( edgesLRTB + u_invSharpness );
}

// ********************************************************************************************************
// Pixel shader that does smart blurring (to avoid bleeding)

void AddSample( float ssaoValue, float edgeValue, inout float sum, inout float sumWeight )
{
    float weight = edgeValue;    

    sum += (weight * ssaoValue);
    sumWeight += weight;
}

vec2 SampleBlurredWide(vec3 coord)
{
	vec2 vC = texture2DArrayLodOffset(s_blurInput, coord, 0.0, ivec2(0, 0)).xy;
	vec2 vL = texture2DArrayLodOffset(s_blurInput, coord, 0.0, ivec2(-2, 0)).xy;
	vec2 vT = texture2DArrayLodOffset(s_blurInput, coord, 0.0, ivec2(0, -2)).xy;
	vec2 vR = texture2DArrayLodOffset(s_blurInput, coord, 0.0, ivec2(2, 0)).xy;
	vec2 vB = texture2DArrayLodOffset(s_blurInput, coord, 0.0, ivec2(0, 2)).xy;

	float packedEdges = vC.y;
	vec4 edgesLRTB = UnpackEdges(packedEdges);
	edgesLRTB.x *= UnpackEdges(vL.y).y;
	edgesLRTB.z *= UnpackEdges(vT.y).w;
	edgesLRTB.y *= UnpackEdges(vR.y).x;
	edgesLRTB.w *= UnpackEdges(vB.y).z;

	float ssaoValue = vC.x;
	float ssaoValueL = vL.x;
	float ssaoValueT = vT.x;
	float ssaoValueR = vR.x;
	float ssaoValueB = vB.x;

	float sumWeight = 0.8f;
	float sum = ssaoValue * sumWeight;

	AddSample(ssaoValueL, edgesLRTB.x, sum, sumWeight);
	AddSample(ssaoValueR, edgesLRTB.y, sum, sumWeight);
	AddSample(ssaoValueT, edgesLRTB.z, sum, sumWeight);
	AddSample(ssaoValueB, edgesLRTB.w, sum, sumWeight);

	float ssaoAvg = sum / sumWeight;

	ssaoValue = ssaoAvg; //min( ssaoValue, ssaoAvg ) * 0.2 + ssaoAvg * 0.8;

	return vec2(ssaoValue, packedEdges);
}

// edge-sensitive blur (wider kernel)
NUM_THREADS(8, 8, 1)
void main() 
{
	uvec2 dtID = uvec2(gl_GlobalInvocationID.xy) + uvec2(u_rect.xy);
	if (all(lessThan(dtID.xy, u_rect.zw) ) )
	{
		vec2 inUV = (dtID.xy+vec2(0.5,0.5)) * u_halfViewportPixelSize;
		imageStore(s_target, ivec3(dtID.xy,u_layer), SampleBlurredWide(vec3(inUV,0.0)).xyyy);
	}
}