Resampled Importance Sampling bxdfs by devshgraphicsprogramming · Pull Request #1027 · Devsh-Graphics-Programming/Nabla

devshgraphicsprogramming · 2026-03-20T11:08:54Z

Description

Testing

TODO list:

…ession reciprocation (swapping of interface order) Unfortunately I realized our THINDIELECTRIC correction is badly designed

…a, won't implement right now

… mitsuba_xml_2_material_frontend

…ff last

@Themperror

Point Boost to the exact Wave one-line backport for emitted pragma newlines, remove the temporary local pragma workaround, and keep the remaining include-path fixes in Nabla. This leaves the Wave pragma issue fixed at the dependency level while preserving the Nabla-side fixes for Windows backslash includes and single-leading-slash virtual includes. Thanks to @Themperror for the additional pragma and include repros. Those made it straightforward to verify the dependency-level fix and drop the local workaround cleanly.

…dates Fix remaining preprocess bugs and backport Wave pragma fix

…te-nsc-windows-x64-release-master CI: Promote NSC channel to 66da590

…mmon bxdfs

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

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

devshgraphicsprogramming · 2026-03-20T11:21:06Z

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

+    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((bxdf.eval_and_weight(_sample, aniso, anisocache)), ::nbl::hlsl::is_same_v, typename T::quotient_pdf_type))
    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((bxdf.pdf(_sample, aniso, anisocache)), ::nbl::hlsl::is_same_v, typename T::scalar_type))
-    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((bxdf.quotient_and_pdf(_sample, aniso, anisocache)), ::nbl::hlsl::is_same_v, typename T::quotient_pdf_type))
+    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((bxdf.quotient_and_weight(_sample, aniso, anisocache)), ::nbl::hlsl::is_same_v, typename T::quotient_pdf_type))


lets make all BxDFs have a cache type which goes from generate to quotient_and_weight but then microfacet_bxdf_common would be the weird one that puts a further constraint that the anisocache_type needs to fit some concept for Cook Torrance to be useful

and allow a 2nd cache type just for eval_and_weight (the two can obviously be identical in practice, e.g. our base impl for now)

you could actually get rid of microfacex_bdf_common and the iso_microfacet concept (basically the entire block in namespace impl ehere)

and just check that {Isotropic}MicrofacetB{R|S}DF = {Isotropic}B{R|S}DF<T> && (Isotropic ? CreatableIsotropicMicrofacetCache<typedef T::cache_type>:AnisotropicMicrofacetCache<typedef T::cache_type>);

devshgraphicsprogramming · 2026-03-20T11:28:15Z

tools/nsc/main.cpp

+    if (!system || outputPath.empty())
+        return false;
+
+    const auto parent = outputPath.parent_path();
+    if (!parent.empty() && !system->isDirectory(parent))
+    {
+        if (!system->createDirectory(parent))
+            return false;
+    }
+
+    auto tempPath = outputPath;
+    tempPath += ".tmp";
+    system->deleteFile(tempPath);
+
+    smart_refctd_ptr<IFile> outputFile;
+    {
+        ISystem::future_t<smart_refctd_ptr<IFile>> future;
+        system->createFile(future, tempPath, IFileBase::ECF_WRITE);
+        if (!future.wait())
+            return false;
+
+        auto lock = future.acquire();
+        if (!lock)
+            return false;
+        lock.move_into(outputFile);
+    }
+    if (!outputFile)
+        return false;
+
+    if (!content.empty())
+    {
+        IFile::success_t success;
+        outputFile->write(success, content.data(), 0ull, content.size());
+        if (!success)
+        {
+            outputFile = nullptr;
+            system->deleteFile(tempPath);
+            return false;
+        }
+    }
+    outputFile = nullptr;
+
+    system->deleteFile(outputPath);
+    const auto moveError = system->moveFileOrDirectory(tempPath, outputPath);
+    if (moveError)
+    {
+        system->deleteFile(tempPath);
+        return false;
+    }
+
+    return true;


@AnastaZIuk do you think you could add an option to createFile flags (or pack the path+ IFileBase flaggs into a strcut) to create all the directories needed on the way to the path ?

cold be useful

…_2_material_frontend

…d weight is separate)

…d camelcase

…_2_material_frontend

…=3) for RIS sampling

include/nbl/builtin/hlsl/path_tracing/concepts.hlsl

devshgraphicsprogramming · 2026-03-23T10:33:07Z

include/nbl/builtin/hlsl/path_tracing/concepts.hlsl

+    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((matsys.evalAndWeight(matid, _sample, aniso_inter)), ::nbl::hlsl::is_same_v, typename T::quotient_pdf_type))
    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((matsys.generate(matid, aniso_inter, u, cache_)), ::nbl::hlsl::is_same_v, typename T::sample_type))
-    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((matsys.pdf(matid, _sample, aniso_inter)), ::nbl::hlsl::is_same_v, typename T::scalar_type))
-    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((matsys.quotient_and_pdf(matid, _sample, aniso_inter, cache_)), ::nbl::hlsl::is_same_v, typename T::quotient_pdf_type))
+    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((matsys.quotientAndWeight(matid, _sample, aniso_inter, cache_)), ::nbl::hlsl::is_same_v, typename T::quotient_pdf_type))


the typedefs returned should be the new value_and_weight types

Need a new file called value_and_weight which is essentially the same as value_and_pdf with similar typedefs
https://github.com/Devsh-Graphics-Programming/Nabla/blob/38f73af10a6047ede44d7822930c04ce2a395f9d/include/nbl/builtin/hlsl/sampling/value_and_pdf.hlsl

With the exception of providing the value() method!

Then implement the value_and_pdf (and rcpPdf) via composition (cause in HLSL inheritance is bound to fuck us over with templated functions) from _and_weight and only have it differ by providing a value() method

include/nbl/builtin/hlsl/path_tracing/unidirectional.hlsl

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

devshgraphicsprogramming · 2026-03-23T11:10:25Z

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

        using quant_query_type = typename ndf_type::quant_query_type;
        quant_query_type qq = impl::quant_query_helper<ndf_type, fresnel_type, IsBSDF>::template __call<Interaction, MicrofacetCache>(ndf, _f, interaction, cache);


you've already computed this in __forwardPdf

any way to making __forwardPdf spit this out for you?
Make make __forwardPdf<bool FromGenerator> returna struct

scalar_type pdf; frensel_type orientedFresenel; quant_query_type quantQuery; [optional only when `IsBSDF && not fresnel_type::ReturnsMonochrome`] spectral_type reflectance; [optional only when `IsBSDF && not fresnel_type::ReturnsMonochrome`] scalar_type scaled_reflectance ;

and doesn't abuse NBL_REF_ARG to return extra things

Also this way this whole block of code

retval.pdf = 0.f; retval.orientedFresenel = __getOrientedFresnel(fresnel, interaction.getNdotV()); const bool cacheIsValid = __checkValid<Interaction, MicrofacetCache>(retval.orientedFresenel, _sample, interaction, cache); assert(!FromGenerator || cacheIsValid); if (FromGenerator ? _sample.isValid():cacheIsValid) { ... fill the members with the old function body } return retval;

can replace the preludes in quotientAndWeight and evalAndWeight with just

if (pdfQuery.pdf==0.f) return ...;

up until the G2_over_G1 computation

or the __overwriteDG

the quotient... function can just not use quantQueryits the same as using a dummy on a NBL_REF_ARG

devshgraphicsprogramming · 2026-03-23T11:24:38Z

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

hmm we still get a pointless mul with 1.0 if the PDF is infinity, if you move this computation to after quo is constructed, we can turn the conditional G2_over_G1 = into quo *=

It does require though that you quo = promote<spectral_type>(1.0) for Monochrome BSDF, thats not a problem because you go contexpr expressions first, so they'll optimize out and the mul with unity will optimize out too

Just the way you have it now, won't

devshgraphicsprogramming · 2026-03-23T11:52:01Z

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

            {
                spectral_type reflectance;
                const scalar_type scaled_reflectance = __getScaledReflectance(_f, interaction, hlsl::abs(cache.getVdotH()), cache.isTransmission(), reflectance);
                quo = reflectance / scaled_reflectance * G2_over_G1;


__forwardPDF already computes these would be cool to return them

devshgraphicsprogramming · 2026-03-23T11:52:52Z

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

        bool isInfinity;
-        scalar_type _pdf = __pdf<Interaction, MicrofacetCache>(_sample, interaction, cache, isInfinity);
+        scalar_type _pdf = __forwardPdf<Interaction, MicrofacetCache>(_sample, interaction, cache, isInfinity);
        return hlsl::mix(_pdf, scalar_type(0.0), isInfinity);


this is incorrect, it should return INF for forward, it was only backward when it was not meant to (hence why I asked for the distinction)

devshgraphicsprogramming · 2026-03-23T12:03:44Z

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

        scalar_type clampedVdotH = cache.getVdotH();
        NBL_IF_CONSTEXPR(IsBSDF)
            clampedVdotH = hlsl::abs(clampedVdotH);


you only need this to call const spectral_type reflectance = impl::__implicit_promote<spectral_type, typename fresnel_type::vector_type>::__call(_f(clampedVdotH));

which we already would have had __forwardPdf return to us because it needed to compute it

it doesn't compute it for you in the BRDF case, but you don't ened to have this whole tempvar and clamp it then

devshgraphicsprogramming · 2026-03-23T12:11:21Z

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


+        spectral_type quo;
        NBL_IF_CONSTEXPR(IsBSDF)
        {
            const spectral_type reflectance = impl::__implicit_promote<spectral_type, typename fresnel_type::vector_type>::__call(_f(clampedVdotH));
-            return hlsl::mix(reflectance, hlsl::promote<spectral_type>(1.0) - reflectance, cache.isTransmission()) * DG;
+            quo = hlsl::mix(reflectance, hlsl::promote<spectral_type>(1.0) - reflectance, cache.isTransmission()) * DG;
        }
        else
-            return impl::__implicit_promote<spectral_type, typename fresnel_type::vector_type>::__call(_f(clampedVdotH)) * DG;
+            quo = impl::__implicit_promote<spectral_type, typename fresnel_type::vector_type>::__call(_f(clampedVdotH)) * DG;
+
+        return quotient_pdf_type::create(quo, _pdf);
    }



in combination with previous comment it would simplify to

NBL_IF_CONSTEXPR(fresnel_type::IsBSDF) { NBL_IF_CONSTEXPR(fresnel_type::ReturnsMonochrome) { quo = pdfRetval.reflectance; } else { quo = mix(pdfRetval.reflectance,hlsl::promote<spectral_type>(1.0) - pdfRetval.reflectance, cache.isTransmission(); } } else quo =impl::__implicit_promote<spectral_type, typename fresnel_type::vector_type>::__call( _f(cache.getVdotH())); quo *= DG;

with DG mul done outside

devshgraphicsprogramming · 2026-03-23T12:40:36Z

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

@@ -902,6 +902,7 @@ NBL_CONCEPT_END(
    ((NBL_CONCEPT_REQ_TYPE)(T::sample_type))
    ((NBL_CONCEPT_REQ_TYPE)(T::spectral_type))
    ((NBL_CONCEPT_REQ_TYPE)(T::quotient_pdf_type))


this now needs to be value_and_weight_type

also require a cache_type which will be passed from generate to quotientAndWeight

also require a cache_type which will be passed from generate to quotientAndWeight

done

devshgraphicsprogramming · 2026-03-23T12:45:23Z

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

you already check these elsewhere I think

…cet checks cache with concept

devshgraphicsprogramming · 2026-03-24T15:20:22Z

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

+    struct Cache {};
+    using isocache_type = Cache;
+    using anisocache_type = Cache;
+    using evalcache_type = Cache;


there should be no evalcache_type, its still same cache as for quotient and generate

But having an evalAndWeight with an overload that takes it should be a Cook Torrance / Microfacet thing only

devshgraphicsprogramming · 2026-03-24T15:21:40Z

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

    ((NBL_CONCEPT_REQ_TYPE)(T::anisotropic_interaction_type))
    ((NBL_CONCEPT_REQ_TYPE)(T::sample_type))
+    ((NBL_CONCEPT_REQ_TYPE)(T::anisocache_type))
+    ((NBL_CONCEPT_REQ_TYPE)(T::evalcache_type))


don't require a separate evaluation cache type

devshgraphicsprogramming · 2026-03-24T15:22:45Z

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

    ((NBL_CONCEPT_REQ_TYPE)(T::isotropic_interaction_type))
    ((NBL_CONCEPT_REQ_TYPE)(T::isocache_type))
-    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((bxdf.eval_and_weight(_sample, iso, isocache)), ::nbl::hlsl::is_same_v, typename T::quotient_pdf_type))
+    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((bxdf.evalAndWeight(_sample, iso, evalcache)), ::nbl::hlsl::is_same_v, typename T::quotient_pdf_type))


evalAndWeight taking the anisocache or isocache should be an extra overload in the imcrofacet BxDF concepts

kevyuu and others added 24 commits March 14, 2026 10:44

Fix checking valid jpeg file logic

5def0b4

get opacity going!

f93f430

really struggling with TWO_SIDED

1a7bf02

add CFrontendIR node copying, stack reversal and individual BxDF expr…

0dc31aa

…ession reciprocation (swapping of interface order) Unfortunately I realized our THINDIELECTRIC correction is badly designed

Beer node needs a thickness unfortunately.

8aff0f7

quick hack around normalmaps and bumpmaps, sketch out the general ide…

1b0b6f3

…a, won't implement right now

Merge remote-tracking branch 'remotes/origin/hotfix_jpeg_loader' into…

6a4e8df

… mitsuba_xml_2_material_frontend

move UV transform to front of SParameterSet, keep variable length stu…

0c1e045

…ff last

implement BLEND_BSDF

850536c

I forgot that for blends all layers need to be blended!

011e7f0

I don't want to admit how long it took me to debug this one

2b34047

fix the last thing in the material frontend

2d3719f

fix TLAS export when Polygon Geometry can't be made into a BLAS

ad0cf8d

merge master

5d2568b

Fix Wave include paths and failure dumps

c86f8e1

Clean up preprocess failure plumbing

4871e24

Merge pull request #1025 from Devsh-Graphics-Programming/preprocessUp…

e4df2b2

…dates Fix remaining preprocess bugs and backport Wave pragma fix

Use upstream gh-pages action for badges

66da590

Promote NSC channel 66da590

f830a47

Merge pull request #1026 from Devsh-Graphics-Programming/ci/nsc-promo…

00017ad

…te-nsc-windows-x64-release-master CI: Promote NSC channel to 66da590

Merge branch 'master' into ris_bxdfs

83e30e4

change concepts for bxdf to *_and_weight methods, make changes for co…

067d107

…mmon bxdfs

pdf method renamed to denominator, don't expect to be used

fc6deb8

devshgraphicsprogramming changed the title ~~Ris bxdfs~~ Resampled Importance Sampling bxdfs Mar 20, 2026

devshgraphicsprogramming commented Mar 20, 2026

View reviewed changes

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

devshgraphicsprogramming commented Mar 20, 2026

View reviewed changes

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

devshgraphicsprogramming commented Mar 20, 2026

View reviewed changes

Merge remote-tracking branch 'remotes/origin/master' into mitsuba_xml…

c292f41

…_2_material_frontend

keptsecret and others added 9 commits March 23, 2026 11:13

camelcase method names, no need for denominator method (pdf is pdf an…

f2e2265

…d weight is separate)

refactor cook torrance to fit concept

0708859

znowu w zyciu mi nie wyszlo

b609179

changes to method names in pt concept, use *_weight instead of pdf an…

63cd4df

…d camelcase

record my discoveries

59e090e

renamed pdf to forwardPdf

a877934

add tractable pdf trait to bxdfs

64ce3ba

Merge remote-tracking branch 'remotes/origin/master' into mitsuba_xml…

2539522

…_2_material_frontend

bxdf input random to generate has variable dimensions (brdf>=2; bsdf>…

148b9bd

…=3) for RIS sampling

devshgraphicsprogramming commented Mar 23, 2026

View reviewed changes

include/nbl/builtin/hlsl/path_tracing/concepts.hlsl Outdated Show resolved Hide resolved

devshgraphicsprogramming commented Mar 23, 2026

View reviewed changes

include/nbl/builtin/hlsl/path_tracing/unidirectional.hlsl Outdated Show resolved Hide resolved

devshgraphicsprogramming commented Mar 23, 2026

View reviewed changes

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

devshgraphicsprogramming commented Mar 23, 2026

View reviewed changes

devshgraphicsprogramming mentioned this pull request Mar 23, 2026

Address path tracer merge leftover comments #991

Merged

keptsecret added 5 commits March 24, 2026 10:03

merge master, fix conflicts

95a3862

pass ray to nee generate instead, store depth in ray

946c9c7

nee deferredPdf is deferredWeight

4f4421d

all bxdfs have a cache type from generate to quotient_weight, microfa…

a70f420

…cet checks cache with concept

eval should have its own cache type in concepts

dab4a18

devshgraphicsprogramming commented Mar 24, 2026

View reviewed changes

		using quant_query_type = typename ndf_type::quant_query_type;
		quant_query_type qq = impl::quant_query_helper<ndf_type, fresnel_type, IsBSDF>::template __call<Interaction, MicrofacetCache>(ndf, _f, interaction, cache);

Conversation

devshgraphicsprogramming commented Mar 20, 2026

Description

Testing

TODO list:

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!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

devshgraphicsprogramming Mar 23, 2026 • 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!

devshgraphicsprogramming Mar 23, 2026 • 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!

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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

devshgraphicsprogramming Mar 23, 2026 •

edited

Loading

devshgraphicsprogramming Mar 23, 2026 •

edited

Loading