Corrupted rendering of specular roughness for shading modules on Android

[kwokcb]

Issue

On Android devices if you set the specular roughness low enough, you appear to get solid white pixels in the areas facing the primary light direction which I believe is computed based on the light environment map.

This occurs for all shading models so appears to be a common math issue.

This does not occur on Apple devices.

Triage

• Around 0.07 for glTF and 0.08 for OpenPBR and standard surface causes this with all other input values the default.
• Code is ESSL with mediump precision setting.

After hunting around a guess as to areas with issues are the following:

// FIX 1: Add safety to GGX Smith G1 function
float mx_ggx_smith_G1(float cosTheta, float alpha)
{
    float cosTheta2 = mx_square(cosTheta);
    float tanTheta2 = (1.0 - cosTheta2) / max(cosTheta2, 1e-6); // PREVENT DIVISION BY ZERO
    float alpha2 = mx_square(max(alpha, 0.001)); // ENFORCE MINIMUM ALPHA
    return 2.0 / (1.0 + sqrt(1.0 + alpha2 * tanTheta2));
}

// FIX 2: Add safety to GGX Smith G2 function  
float mx_ggx_smith_G2(float NdotL, float NdotV, float alpha)
{
    float alpha2 = mx_square(max(alpha, 0.001)); // ENFORCE MINIMUM ALPHA
    
    // ADD SAFETY CHECKS FOR NEAR-ZERO DOT PRODUCTS
    NdotL = max(NdotL, 1e-6);
    NdotV = max(NdotV, 1e-6);
    
    float lambdaL = sqrt(alpha2 + (1.0 - alpha2) * mx_square(NdotL));
    float lambdaV = sqrt(alpha2 + (1.0 - alpha2) * mx_square(NdotV));
    
    // PREVENT DIVISION BY ZERO
    float denominator = lambdaL * NdotV + lambdaV * NdotL;
    return 2.0 * NdotL * NdotV / max(denominator, 1e-8);
}

// FIX 3: Add safety to the main BSDF response calculation
// In mx_dielectric_bsdf and mx_generalized_schlick_bsdf, find this line:
bsdf.response = D * F * G * comp * safeTint * closureData.occlusion * weight / (4.0 * NdotV);

// CHANGE TO:
float denominator = 4.0 * max(NdotV, 1e-6);
bsdf.response = D * F * G * comp * safeTint * closureData.occlusion * weight / denominator;

[jstone-lucasfilm]

Instead of manually modifying alpha within each of our BSDF implementations, I wonder if we could provide a shader generation option that defines a setting for MIN_SAFE_ALPHA, and then apply this in our standard computations of safeAlpha:

https://github.com/AcademySoftwareFoundation/MaterialX/blob/main/libraries/pbrlib/genglsl/mx_conductor_bsdf.glsl#L21

This would allow each platform to make the decision as to what minimum alpha is safe for them, without hardcoding any of these decisions in our BSDF implementations.

[jstone-lucasfilm]

Thanks for this original report, @kwokcb, as well as for the fix in #2792!