Bxdf fixes cook torrance #930

devshgraphicsprogramming · 2025-09-16T14:03:10Z

Description

Continues #899 , #916 and #919

Testing

TODO list:

…eturn quants

include/nbl/builtin/hlsl/bxdf/common.hlsl

include/nbl/builtin/hlsl/math/functions.hlsl

include/nbl/builtin/hlsl/cpp_compat/vector.hlsl

include/nbl/builtin/hlsl/bxdf/fresnel.hlsl

include/nbl/builtin/hlsl/bxdf/cook_torrance_base.hlsl

include/nbl/builtin/hlsl/bxdf/ndf/ggx.hlsl

…ng invalid

include/nbl/builtin/hlsl/bxdf/base/cook_torrance_base.hlsl

devshgraphicsprogramming · 2025-11-03T11:44:29Z

include/nbl/builtin/hlsl/bxdf/base/cook_torrance_base.hlsl

+        if (isInfinity)
+            return hlsl::promote<spectral_type>(0.0);


it needs to be commented and documented why you don't hoist above line 176 (my previous review comment)

devshgraphicsprogramming · 2025-11-03T12:50:56Z

include/nbl/builtin/hlsl/bxdf/ndf/beckmann.hlsl

+            isInfinity = true;
            return bit_cast<scalar_type>(numeric_limits<scalar_type>::infinity);
+        }
+        isInfinity = false;


not that easy!

your exp2 will spit out inf waay before a2<Threshold it also depends on NdotH2

so you need to replace isInfinity = false with isInfinity = isinf(retval) after the multiplication by pi and /denom (or you compute nom/denom and check that value is <Max/Pi)

devshgraphicsprogramming · 2025-11-03T12:51:12Z

include/nbl/builtin/hlsl/bxdf/ndf/beckmann.hlsl

+            isInfinity = true;
            return bit_cast<scalar_type>(numeric_limits<scalar_type>::infinity);
+        }
+        isInfinity = false;


again, same thing as https://github.com/Devsh-Graphics-Programming/Nabla/pull/930/files/309e216f2e40e7fb84b77ae76717e82dc8c1bf17..18b66bed5bbe07a3682a5408b42e7c05941dd609#r2486385513

devshgraphicsprogramming · 2025-11-03T12:52:54Z

include/nbl/builtin/hlsl/bxdf/ndf/beckmann.hlsl

+            const scalar_type VdotH = cache.getVdotH();
+            const scalar_type VdotH_etaLdotH = hlsl::mix(VdotH + orientedEta * cache.getLdotH(),
+                                                        VdotH / orientedEta + cache.getLdotH(),
+                                                        interaction.getPathOrigin() == PathOrigin::PO_SENSOR);
+            quant_query.neg_rcp2_refractionDenom = scalar_type(-1.0) / (VdotH_etaLdotH * VdotH_etaLdotH);


why have we got code for this and not using somethign from microfacet_to_light_transform.hlsl ?

devshgraphicsprogramming · 2025-11-03T12:54:46Z

include/nbl/builtin/hlsl/bxdf/ndf/beckmann.hlsl

    {
+        scalar_type D = query.getNdf();
        scalar_type dg1 = query.getNdf() / (scalar_type(1.0) + query.getLambdaV());
+        isInfinity = D == bit_cast<scalar_type>(numeric_limits<scalar_type>::infinity);


the isInfinity is supposed to be an "early out" , it should even be called isNDFInfinity to make things clear.

You want to know if the NDF is infinite, sicne G2 and G2 can only be in [0,1]

devshgraphicsprogramming · 2025-11-03T12:55:32Z

include/nbl/builtin/hlsl/bxdf/ndf/ggx.hlsl

+            isInfinity = true;
            return bit_cast<scalar_type>(numeric_limits<scalar_type>::infinity);
+        }
+        isInfinity = false;


same as beckmann

devshgraphicsprogramming · 2025-11-03T12:55:39Z

include/nbl/builtin/hlsl/bxdf/ndf/ggx.hlsl

+            isInfinity = true;
            return bit_cast<scalar_type>(numeric_limits<scalar_type>::infinity);
+        }
+        isInfinity = false;


same as beckmann comment

devshgraphicsprogramming · 2025-11-03T12:55:56Z

include/nbl/builtin/hlsl/bxdf/ndf/ggx.hlsl

-            const scalar_type VdotH_etaLdotH = cache.getVdotH() + orientedEta * cache.getLdotH();
-            quant_query.neg_rcp2_VdotH_etaLdotH = scalar_type(-1.0) / (VdotH_etaLdotH * VdotH_etaLdotH);
+            const scalar_type VdotH = cache.getVdotH();
+            const scalar_type VdotH_etaLdotH = hlsl::mix(VdotH + orientedEta * cache.getLdotH(),
+                                                        VdotH / orientedEta + cache.getLdotH(),
+                                                        interaction.getPathOrigin() == PathOrigin::PO_SENSOR);
+            quant_query.neg_rcp2_refractionDenom = scalar_type(-1.0) / (VdotH_etaLdotH * VdotH_etaLdotH);


again, why duplicate code with Beckmann ?

devshgraphicsprogramming · 2025-11-03T12:57:09Z

include/nbl/builtin/hlsl/bxdf/ndf/ggx.hlsl

    {
+        scalar_type D = query.getNdfwoNumerator();
        scalar_type dg1_over_2NdotV = query.getNdfwoNumerator() * query.getG1over2NdotV();
+        isInfinity = D == bit_cast<scalar_type>(numeric_limits<scalar_type>::infinity);


you got the right thing this time, but I'd hoist to be right after D and use isinf or <Max for setting the bool

devshgraphicsprogramming · 2025-11-03T13:00:36Z

include/nbl/builtin/hlsl/math/angle_adding.hlsl

+    void addCosine(T cosA, T biasA)
    {
-        const bool AltminusB = tmp0 < (-tmp1);
-        const float cosSumAB = tmp0 * tmp1 - tmp3 * tmp4;
-        const bool ABltminusC = cosSumAB < (-tmp2);
-        const bool ABltC = cosSumAB < tmp2;
-        // apply triple angle formula
-        const float absArccosSumABC = acos<float>(clamp<float>(cosSumAB * tmp2 - (tmp0 * tmp4 + tmp3 * tmp1) * tmp5, -1.f, 1.f));
-        return ((AltminusB ? ABltC : ABltminusC) ? (-absArccosSumABC) : absArccosSumABC) + ((AltminusB || ABltminusC) ? numbers::pi<float> : (-numbers::pi<float>));
+        const T bias = biasA + runningSum.imag();


what's biasA ? how does it work ?

Let me recap some assumptions:

you initialize to an angle thats between 0 and 2pi (you need to handle the case where the initialization is with sin<0 so you get >pi when querying for the arccos)

each time you add an angle you add a positive angle (so again, handle case where new angle sine is negative)

Actually you can assume the initial and each added angle is between 0 and pi because a larger angle would mean a concave (not convex) spherical polygon

devshgraphicsprogramming · 2025-11-03T13:02:36Z

include/nbl/builtin/hlsl/math/angle_adding.hlsl

    }

-    float getArccosSumofABC_minus_PI()
+    void addCosine(T cosA, T biasA)


where's a addAngle(const complex_t<T>) if I don't want to reconstruct the sine of the angle I'm adding from the cosine?

devshgraphicsprogramming · 2025-11-03T13:04:11Z

include/nbl/builtin/hlsl/math/angle_adding.hlsl

+    static T getArccosSumofABC_minus_PI(T cosA, T cosB, T cosC, T sinA, T sinB, T sinC)
+    {
+        const bool AltminusB = cosA < (-cosB);
+        const T cosSumAB = cosA * cosB - sinA * sinB;
+        const bool ABltminusC = cosSumAB < (-cosC);
+        const bool ABltC = cosSumAB < cosC;
+        // apply triple angle formula
+        const T absArccosSumABC = acos<T>(clamp<T>(cosSumAB * cosC - (cosA * sinB + sinA * cosB) * sinC, T(-1.0), T(1.0)));
+        return ((AltminusB ? ABltC : ABltminusC) ? (-absArccosSumABC) : absArccosSumABC) + ((AltminusB || ABltminusC) ? numbers::pi<T> : (-numbers::pi<T>));
    }


this should disappear and everything that uses it, should just do

accumulator = sincos_accumulator::create(cosA,sinA) accumulator.addAngle(cosB,sinB); accumulator.addAngle(cosC,sinC); accumulator.getSumofArccos();

devshgraphicsprogramming · 2025-11-03T13:05:24Z

include/nbl/builtin/hlsl/math/angle_adding.hlsl

+        const T a = cosA;
+        const T b = runningSum.real();
+        const bool reverse = abs<T>(min<T>(a, b)) > max<T>(a, b);
+        const T c = a * b - sqrt<T>((T(1.0) - a * a) * (T(1.0) - b * b));


I shouldn't need to reconstruct the sine of the angle I've accumulated

devshgraphicsprogramming · 2025-11-03T13:06:14Z

include/nbl/builtin/hlsl/math/angle_adding.hlsl

-        retval.tmp3 = sinA;
-        retval.tmp4 = sinB;
-        retval.tmp5 = sinC;
+        retval.runningSum = complex_t<T>::create(cosA, T(0));


if you're using complex for storage, then you should initialize to cosA,sinA

what I meant by #899 (comment)

store the wraparound in integer part of sine

Is that a sine is [-1,1] and therefore you can store the wraparound in int(abs(complex.y)) or a similar encoding which retains the info about the fractional part of the sine or a separate uint16_t (if you see its faster despite using more regs- test with spherical triangle and rectangle sampling)

devshgraphicsprogramming · 2025-11-03T13:14:54Z

include/nbl/builtin/hlsl/bxdf/base/cook_torrance_base.hlsl

+        const vector3_type H = hlsl::mul(interaction.getFromTangentSpace(), localH);
+        assert(hlsl::abs(hlsl::length(H) - scalar_type(1.0)) < scalar_type(1e-4));


btw assertion for orthonormality of matrix is just checking the dot products between all pairs of rows are 1.0

You could have a general utility function bool dotIsUnity(A,B); then this becomes

// tangent frame orthonormality assert(dotIsUnity(fromTangent[0],fromTangent[1])); assert(dotIsUnity(fromTangent[1],fromTangent[2])); assert(dotIsUnity(fromTangent[2],fromTangent[0])); // NDF sampling produced a unit length direction assert(dotIsUnity(localH,localH));

you don't want to assert stuff on the output H

keptsecret added 23 commits September 4, 2025 10:46

changes to dual measure quant, old stuff commented out for now

5218545

create query structs in beckmann ndf, removed query concepts, funcs r…

f7525af

…eturn quants

did the same for ggx ndf, some fixes to beckmann

c50db68

Merge branch 'master' into bxdf_fixes_cook_torrance

c49bedb

fixes to ndf concept

319c954

moved eval, quotient/pdf into cook torrance base

4eeacf1

moved generate H into ndfs, generate impl in cook torrance base

e7bc784

use new cook torrance base in microfacet bxdfs

72226bb

Merge branch 'master' into bxdf_fixes_cook_torrance

f34b348

numerous typo bug fixes

ad13044

fixes to ggx ndf

d8d2116

fixed conductor fresnel naming ior as eta

238b08e

minor fixes to beckmann ndf, removed obsolete comments

8bc707f

added luminosity contribution hint to cook torrance bsdfs

9fb28cf

added aniso overloads to isotropic bxdf methods

4cb8a0a

checks for invalid generate sample

392dc31

removed obsolete commented out stuff

02de86e

added cartesian-polar conversions from unit tests

4a7f532

Merge branch 'fix_rotation_mat' into bxdf_fixes_cook_torrance

80c4f67

Merge branch 'master' into bxdf_fixes_cook_torrance

6233bd1

vector hashes

ce5fbac

Merge branch 'fix_rotation_mat' into bxdf_fixes_cook_torrance

47e814b

Merge branch 'master' into bxdf_fixes_cook_torrance

c983975

devshgraphicsprogramming commented Sep 16, 2025

View reviewed changes

include/nbl/builtin/hlsl/bxdf/common.hlsl Outdated Show resolved Hide resolved

devshgraphicsprogramming commented Sep 16, 2025

View reviewed changes

include/nbl/builtin/hlsl/math/functions.hlsl Outdated Show resolved Hide resolved

devshgraphicsprogramming commented Sep 16, 2025

View reviewed changes

include/nbl/builtin/hlsl/cpp_compat/vector.hlsl Outdated Show resolved Hide resolved

keptsecret added 2 commits September 17, 2025 15:26

fix fresnel usage in bxdf pdf

0f2ee0b

fix F calc in generate

b5f02e6

devshgraphicsprogramming commented Sep 17, 2025

View reviewed changes

include/nbl/builtin/hlsl/bxdf/fresnel.hlsl Outdated Show resolved Hide resolved

devshgraphicsprogramming commented Sep 17, 2025

View reviewed changes

include/nbl/builtin/hlsl/bxdf/cook_torrance_base.hlsl Outdated Show resolved Hide resolved

devshgraphicsprogramming commented Oct 20, 2025

View reviewed changes

include/nbl/builtin/hlsl/bxdf/ndf/ggx.hlsl Outdated Show resolved Hide resolved

fixes to VdotL calculation, contigency for beckmann vndf sampling bei…

309e216

…ng invalid

devshgraphicsprogramming commented Oct 25, 2025

View reviewed changes

include/nbl/builtin/hlsl/bxdf/base/cook_torrance_base.hlsl Outdated Show resolved Hide resolved

devshgraphicsprogramming commented Oct 25, 2025

View reviewed changes

include/nbl/builtin/hlsl/bxdf/base/cook_torrance_base.hlsl Outdated Show resolved Hide resolved

keptsecret added 8 commits October 27, 2025 11:04

use beta function in G2 correlated as well

c172b19

fixes ggx G2 using beta

ee864a6

pdf should also depend on surface interaction

b01ebf7

account for non-symmetry in transmission, added PathOrigin

36c1f4b

fixes calculations to NdotL, restore quick dot products in bsdf generate

79c80cd

moved around some redundant calculations in generate for optimization

589b7df

add sidechannel bool if D is infinity

36cd283

make angle adding stuff work better

18b66be

devshgraphicsprogramming commented Nov 3, 2025

View reviewed changes

devshgraphicsprogramming merged commit ba500ab into master Nov 3, 2025
19 of 20 checks passed

devshgraphicsprogramming deleted the bxdf_fixes_cook_torrance branch November 3, 2025 13:24

devshgraphicsprogramming mentioned this pull request Nov 3, 2025

Iridescent fresnel, bxdfs #918

Open

		const vector3_type H = hlsl::mul(interaction.getFromTangentSpace(), localH);
		assert(hlsl::abs(hlsl::length(H) - scalar_type(1.0)) < scalar_type(1e-4));

Bxdf fixes cook torrance #930

Bxdf fixes cook torrance #930

Uh oh!

Conversation

devshgraphicsprogramming commented Sep 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Testing

TODO list:

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

devshgraphicsprogramming Nov 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

devshgraphicsprogramming commented Sep 16, 2025 •

edited

Loading

devshgraphicsprogramming Nov 3, 2025 •

edited

Loading