Build in public
A physically based renderer and glTF 2.0 viewer, on a functional core, with a pipeline you can see.
A renderer in C# and Vulkan that aims to open any glTF 2.0 asset and render it the way the spec says it should look. The frame is a pure function, so the pipeline is data, which is why it can be drawn as a graph rather than guessed at. Every technique ships as working code and a companion post that explains why it exists.
Every confirmed post, in reading order. Status reflects the current state of the draft, not a promise of when it will ship.
From a blank window to a composed frame.
Motivation and context. Limitations of existing engines, friction with OOP, rendering as data transformation, and the Functional Core + Imperative Shell pattern.
Read post →A minimal modern rendering pipeline in C# and Vulkan. Device, swapchain, shaders, handles, descriptors, and the first render command, explained from the ground up.
Read post →The G-Buffer and the geometry pass. Why deferred over forward, multi-render-target setup, G-Buffer layout, and writing scene data without a single light.
Read post →The lighting pass, the tonemap pass, and the render graph. Composing passes as pure data, compiling order and barriers from declarations, and testing a frame without a GPU. Closes the introduction.
Read draft →A plausible guess gives way to a physical model.
Swapping the Blinn-Phong lobe for a Cook-Torrance metallic-roughness BRDF: GGX distribution, Smith geometry, Schlick Fresnel, and what energy conservation actually buys you. The G-Buffer's two spare scalar slots become roughness and metallic, so the whole change is a shader plus a material record. Ends on a live A/B against Blinn-Phong in the same scene.
Read draft →Per-pixel normals perturbed from a texture, faking high-frequency surface detail without geometry. Tangent space and the TBN basis, why this is the change that finally grows the vertex layout, and deriving tangents at import when the file does not ship them. The lighting pass does not change by one line. The derived bitangent pointed the wrong way through three milestones, and the lab's own tests could not see it, because the fixture had been written to match the bug.
Read draft →Retiring the hemispheric ambient constant for a real environment. The split-sum approximation: irradiance convolution, a prefiltered mip chain, and a BRDF integration LUT, plus the render graph's first precompute passes.
A specification decides what correct means.
Why the lab handed correctness to a specification. What a personal renderer gets wrong when the only test is the author's eye, what glTF 2.0 actually requires, and what the importer was quietly dropping along the way. Also the scope decision taken before the work rather than during it: animation, skinning and morph targets sit outside a named static profile.
Building the ruler before the thing it measures: a device with no window, a checked-in golden, and a comparison metric that is pure and tested without a GPU. The argument is the ordering, because a milestone of specification work judged by eye means finding out at the end. What made it cheap is that rendering into a viewport, a thumbnail or a file was already the same code.
Reading the texture slots the loader skipped, and the consequence: the G-Buffer runs out of room and grows a fourth attachment, because emissive is a colour rather than a scalar. Plus sampler state, the half of importing a texture that is easy to forget, and the material leaving push constants for a buffer, because five texture slots with a sampler and a transform each do not fit in 128 bytes.
Alpha masking is a discard and costs nothing. Alpha blending is a contradiction: a G-Buffer stores one surface per pixel, and a blended surface is by definition not the only one there. The forward transparent pass this forces, and why the predicted second copy of the shading code never arrived.
The phase where reading the specification properly changed the data model. A negative determinant reverses winding, so the front face becomes dynamic state per draw rather than another axis on the pipeline key. A rotation between two non-uniform scales composes to a shear, and no TRS spells shear, so a drawable gives up its transform for a world matrix while the node keeps its TRS.
Pointing the harness at the Khronos sample assets and publishing what comes back: 67 pass, 0 fail, 3 outside the profile, of 70 models. The metric is LDR-FLIP, checked bit for bit against the reference implementation. The claim is a reviewed golden, which catches every regression and no misreading of the specification that the review shared.
A conformance table's real output is a work list. Twenty failing rows named three missing features and nothing else: flat normals for a primitive that ships none, glTF's default material for a primitive with no material, and the primitive modes that are not triangle lists. Ends on the run that moved exactly the twenty rows it meant to and left the other forty-nine byte-identical.
The frame stops being a sequence of calls and becomes a picture.
Drawing the pipeline: layering passes by longest path, ordering within a layer to reduce crossings, and weighting nodes by measured GPU cost. The argument underneath it is that the layout is a pure function of the compiled graph, and an imperative renderer has nothing to hand such a function. Passes and resources are both nodes, which is the decision the post has to defend.
Resource lifetimes as a timeline under the picture: one bar per resource, from the pass that brings it into existence to the last pass that still needs it, on the same horizontal axis as the graph above it. This is the part of a frame a dependency picture cannot show, that four G-Buffer images are live at once and three are dead before the forward pass runs. Aliasing is what the lifetimes make possible, not something the compiler does.
Infrastructure decisions and architecture records.
Field note. Why the engine centralized its own allocator instead of adopting VMA. Only one of VMA's three benefits was actually about allocation, and that was the one with no evidence behind it.
Read draft →Field note. Why the lab uses FIFO_RELAXED with MinImageCount instead of the AAA Mailbox default: honest frame-time signal over smoothed presentation.
Read draft →Field note. The pipeline shift where the scene stops being the swapchain target and becomes an image a panel draws. Covers the UI backend you have to write yourself, input-capture gating, and why swapping the UI library later cost so little.
Field note. Watching shader sources, recompiling to SPIR-V, and swapping pipelines without restarting. The decision worth recording is how a failed compile keeps the previous pipeline live instead of dropping the frame. Includes are watched too, so editing one recompiles everything that reads it.
Field note. The editor's Elm-style Model/Msg/Update layer: how a pure reducer lets panels edit transforms, lights, and materials live while the renderer stays a pure consumer of the scene snapshot.
Field note. A texture sampled inside divergent control flow takes an implicit level of detail, and the derivatives that needs are undefined, whatever the source said before the driver lowered it. It turned up three times during the compliance work, once as a ternary that moved 107 pixels of a golden containing nothing the branch was written for.
Field note. One SPIR-V module and eight G-Buffer pipelines, with no variant table in the build. Specializing alpha masking and topology rather than branching on a uniform keeps the discard statically unreachable, so early depth testing survives on the opaque variant, and a topology is baked into a pipeline and cannot be told at record time.
