Created a new example to test the compilation and concepts of samplers by karimsayedre · Pull Request #252 · Devsh-Graphics-Programming/Nabla-Examples-and-Tests

karimsayedre · 2026-02-18T20:09:58Z

No description provided.

…hader permutation), configurable work group size in `ITester`

…omplete tests

devshgraphicsprogramming · 2026-03-23T13:12:02Z

Everything looks good, although it would be nice to try the CDF and Alias table samples at several different PoT sizes from 8 to 128*1024^2 to find the perf crossover point across sizes

Also do a 1 Reservoir sampling from size of 2 to 32 because that will likely be faster than CDF and Alias at really small sizes (same way linear search is vs binary and hashmap due to caching effects)

devshgraphicsprogramming · 2026-03-23T13:20:44Z

Also see if you can address any of the open comments in the BxDF testing PR
#236

devshgraphicsprogramming · 2026-03-27T21:31:33Z


-        pdf = rcpPdf > numeric_limits<scalar_type>::min ? (1.0 / rcpPdf) : numeric_limits<scalar_type>::max;
+        const scalar_type fwdPdf = pst.forwardPdf(pstCache);
+        pdf = fwdPdf < numeric_limits<scalar_type>::max ? fwdPdf : numeric_limits<scalar_type>::max;


this does nothing

devshgraphicsprogramming · 2026-03-27T21:32:20Z

+	using value_type = T;
+	template<typename V, typename I>
+	void get(I i, NBL_REF_ARG(V) val) NBL_CONST_MEMBER_FUNC { val = V(data[i]); }
+	T operator[](uint32_t i) NBL_CONST_MEMBER_FUNC { return data[i]; }


what do you need operator[] for ?

@karimsayedre still a TODO

devshgraphicsprogramming · 2026-03-27T21:34:26Z

-
-		float32_t2 diff = input.u - output.inverted;
-		output.roundtripError = nbl::hlsl::length(diff);
-		output.jacobianProduct = (float32_t(1.0) / output.forwardPdf) * output.backwardPdf;


even if you're no longer 100% bijective in all cases, you can perturb input.u by epsillon and work out the Jacobian * forwardPdf and compare against 1

@karimsayedre add this test back in modified form, its very important

devshgraphicsprogramming · 2026-03-27T21:35:42Z

-			sampling::Linear<float32_t>::cache_type cache;
-			output.generateInversed = _sampler.generateInverse(output.generated);
 			output.backwardPdf = _sampler.backwardPdf(output.generated);
 		}
-		output.roundtripError = abs(input.u - output.generateInversed);
-		output.jacobianProduct = (float32_t(1.0) / output.forwardPdf) * output.backwardPdf;


same as bilinear

@karimsayedre add this test back in modified form, its very important

devshgraphicsprogramming · 2026-03-27T21:36:21Z

-		{
-			sampling::ProjectedSphere<float32_t>::cache_type cache;
-			output.inverted = sampler.generateInverse(output.generated);
-			output.backwardPdf = sampler.backwardPdf(output.generated);
-		}
-		float32_t2 xyDiff = output.modifiedU.xy - output.inverted.xy;
-		output.roundtripError = nbl::hlsl::length(xyDiff);
-		output.jacobianProduct = (float32_t(1.0) / output.forwardPdf) * output.backwardPdf;


you can still work out a Jacobian from perturbing u

only for the hemisphere I guess

@karimsayedre add this test back in modified form, its very important

devshgraphicsprogramming · 2026-03-27T21:37:42Z

 			sampling::ProjectedSphericalTriangle<float32_t>::cache_type cache;
 			output.generated = sampler.generate(input.u, cache);
-			output.cachedPdf = cache.pdf;
 			output.forwardPdf = sampler.forwardPdf(cache);


you can do a JAcobian test, since its bijective if you wanted to

@karimsayedre still a TODO

devshgraphicsprogramming · 2026-03-27T21:39:25Z

 			output.inverted = sampler.generateInverse(output.generated);
 			// Backward: evaluate pdf at generated point (no cache needed)
 			output.backwardPdf = sampler.backwardPdf(output.generated);


differentiated generateInverse Jacobian determinant should equal the PDF I think

while differentiated generate multiplied with PDF comes out to 1

devshgraphicsprogramming · 2026-03-27T21:40:37Z

 		}

 		output.backwardPdf = sampler.backwardPdf(output.generated);
 		output.separateBackwardPdf = sampler.separateBackwardPdf(output.generated);


you can test a lot of things, like:

separate/independent PDF multiplying together to give the scalar PDF
2 Jacobian test follows PDF

devshgraphicsprogramming · 2026-03-27T21:44:04Z

+   const float32_t toFloat = asfloat(0x2f800004u);
+   uint32_t2 acc = (uint32_t2)0;
+   uint32_t accPdf = 0;
+   for (uint32_t i = 0u; i < uint32_t(BENCH_ITERS); i++)
+   {
+      float32_t2 u = float32_t2(rng(), rng()) * toFloat;
+      sampling::SphericalRectangle<float32_t>::cache_type cache;
+      acc ^= asuint(sampler.generate(u, cache));
+      accPdf ^= asuint(sampler.forwardPdf(cache));
+   }


you need two benchmark modes, multiple samples from same create and multiple samples with create (to measure set-up overhead)

first one matters only for rejection sampling or multisampling, second one is more real use case when you only draw 1 NEE sample per path vertex so the create to generate ratio is 1:1

# Conflicts: # common/include/nbl/examples/Tester/ITester.h

- lots of testers for all samplers with more focus on spherical ones, then made sure they *mostly pass* - `isValidSphericalTriangle()` is a lot more tolerant, from ~17 degrees to 1.7 degrees, which means more small triangles are tested - pdf and weight consistency missing tests

…ngle and tests for it

# Conflicts: # 31_HLSLPathTracer/CMakeLists.txt # 31_HLSLPathTracer/app_resources/hlsl/spirv/pt.compute.rectangle.rwmc.linear.proxy.hlsl # 31_HLSLPathTracer/app_resources/hlsl/spirv/pt.compute.variant.shared.hlsl # 31_HLSLPathTracer/include/nbl/this_example/render_variant_info.hpp # 31_HLSLPathTracer/main.cpp

… adapt everything tests to that, and: - Make sure `generete()` and `generateSurfaceOffset()` are on the same page, that way we don't have to do separate tests for it - No performance regression

- Benchmarks are done before tests - Reduced some CPU sampler test iterations because it was too slow - Workgroup size is now shared between tests and benchmarks, also reduced from 96 to 64

devshgraphicsprogramming · 2026-04-15T19:02:10Z

There's a fundamental problem in this PR, none of the continuous 2D samplers have Jacobian testing of generate (via numerical differentiation) which is why I think a bug on the Uniform Hemisphere sampler is not caught

I need in the next PR:

selective JAcobian tests (when forward PDF is not INF) always on all 2D continuous samplers (special handling of discontinuities like projected sphere and hemisphere is needed as @AnastaZIuk pointed out)
benchmark of 1:1 creation and sample drawing vs 1:16

Created an new example to test the compilation and concepts of samplers

ebf25f4

karimsayedre mentioned this pull request Feb 18, 2026

Created concepts for samplers, added quotient_and_pdf variants to satisfy the concepts Devsh-Graphics-Programming/Nabla#1001

Open

devshgraphicsprogramming reviewed Feb 19, 2026

View reviewed changes

Comment thread 37_HLSLSamplingTests/app_resources/test_compile.comp.hlsl Outdated

devshgraphicsprogramming requested a review from AnastaZIuk February 19, 2026 14:12

devshgraphicsprogramming reviewed Feb 19, 2026

View reviewed changes

Comment thread CMakeLists.txt Outdated

karimsayedre added 11 commits February 25, 2026 09:14

changed example number to 37

18fe5eb

Tests using ITester, Benchmarking using the same testing shaders (s…

d1fa476

…hader permutation), configurable work group size in `ITester`

Merge branch 'master' into sampler-concepts

d8a6e9b

Merge branch 'master' into sampler-concepts

6740883

fixes after merge

4a45531

Samplers now conform to concepts, benchmarks for all samplers, more c…

7931826

…omplete tests

added alias table and cumulative probablility testing and benchmarks

d9f4907

Merge branch 'master' into sampler-concepts

69fe759

minor formatting changes

8f3c1c9

addressed comments in concepts.hlsl, making sure 37 and 66 work

24baa87

renamed InvertableSampler to BackwardTracableSampler

54acb0c

karimsayedre added 3 commits March 23, 2026 17:36

Merge branch 'master' into sampler-concepts-upd

f94133f

formatting to 3 spaces, added pipeline info report

60e0a35

addressed more comments, made benchmarks more fair

b33f66a

devshgraphicsprogramming reviewed Mar 27, 2026

View reviewed changes

karimsayedre added 12 commits March 28, 2026 18:00

Merge branch 'master' into sampler-concepts

f76b427

# Conflicts: # common/include/nbl/examples/Tester/ITester.h

changed forwardPdf to take the codomain as an arg

ff397d0

test code reduction in main.cpp

cfefa02

Merge branch 'master' into sampler-concepts

d93a0f8

forwardPdf should take domain not codomain, also corrected some comments

eb10cf2

reverted projectedSolidAngle changes, added projected spherical recta…

00ed770

…ngle and tests for it

fixed a logging bug

dd53be0

changes after spherical rect generate() now returning a 3d direction,…

ad48fa1

… adapt everything tests to that, and: - Make sure `generete()` and `generateSurfaceOffset()` are on the same page, that way we don't have to do separate tests for it - No performance regression

benchmark logging improvements, also:

04d61dc

- Benchmarks are done before tests - Reduced some CPU sampler test iterations because it was too slow - Workgroup size is now shared between tests and benchmarks, also reduced from 96 to 64

Merge branch 'master' into sampler-concepts

19122d7

Conversation

karimsayedre commented Feb 18, 2026

Uh oh!

Uh oh!

Uh oh!

devshgraphicsprogramming commented Mar 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

devshgraphicsprogramming commented Mar 23, 2026

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 commented Apr 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

devshgraphicsprogramming commented Mar 23, 2026 •

edited

Loading

devshgraphicsprogramming commented Apr 15, 2026 •

edited

Loading