flutter3d_conformance 0.8.1

The behaviour a flutter3d_hardware backend has to have, as tests it can run against itself. Flat: no Flutter import of its own, and flutter3d_shaders (its one dependency that used to carry the SDK in) [...]

0.8.1

The linking check names ContactShadowResolve, the stage flutter3d_core 0.8.1 draws, so a backend whose bundle lacks it fails the check rather than the first frame with contact shadows and no temporal resolve.

It asks for flutter3d_shaders ^0.8.1.

0.8.0

Four new checks, forty-one in all, and thirty-one of them need the shader pipeline. A backend written against 0.7 meets each as a failure or a decline until it implements the members they ask about.

• a texture bound to a slot the stage lacks is false: a name no stage declares, and a fragment sampler bound through the vertex stage's handle, on every backend that reflects its stages.
• a device that filters 32-bit floats filters and renders them holds a device answering supportsFloat32Filtering with true to both halves of it: a bilinear tap between float texels, and a float target that renders and samples back. A device answering false declines.
• a field steps in a float target and reads back through a vertex steps a decay kernel three times from a start of two in each float format the device renders to, and reads the result through a vertex stage: 0.425 is right, and 0.3 means the target clamped to one. The upload check never covered rendering into a float target.
• a compute pass scans a buffer in place, the one entry of the new computeChecks, which conformanceChecks now includes: a prefix sum over 1024 integers, read back and held to i(i+1)/2. It declines where supportsCompute is false, which is Impeller and WebGL2.

The link check covers the stages 0.8.0 added. PbrLayered links against every mesh vertex stage; the three velocity vertex stages against Velocity and Reactive; the full-screen stage against DepthPyramid, CameraVelocity, TemporalResolve, TemporalAccumulate, VolumetricFog, VolumetricFogUpsample, VelocityTileMax, VelocityNeighborMax, MotionBlur, FieldDecay, IrradianceConvolve, EvsmFilter, WboitResolve and SceneColourCopy; and four pairs are new: ImpostorVertex with Impostor, and ParticleVertex with ParticleSixWay, ReactiveSprite and SplatHashed.

Fuzzing, new. generateFuzzProgram turns a seed into a program of draws over the probe stages every bundle ships, with viewports, scissors, blending and depth chosen at random, and drawFuzzProgram draws it on a fresh device from any DeviceFactory. The oracle is the program against itself: each FuzzTransform (SplitDraw, SwapDisjointDraws, FoldPowerOfTwo) rewrites it into one that must draw the same bytes, exactly in IEEE 754. fuzzSeed draws a seed's program and its rewrites and returns the ones that came out different, and shrinkFinding cuts a finding down to the draws that still show it. fuzzDifference compares a device against a reference device instead, as the share of pixels more than a few steps apart. FuzzRandom is Park and Miller's generator, so a seed is the same program on the VM and in a browser. Two hundred seeds run on the software rasteriser, and fifty on each of WebGL2 and WebGPU against it, held to 2% of pixels. The first run found two bugs in the software rasteriser's scissor and blending, fixed in flutter3d_cpu. H4

Its flutter3d_* dependencies ask for ^0.8.0.

0.7.1

Released with the rest of the stack at 0.7.1. Nothing in this package changed. The release it resolves against builds from pub.dev again and no longer crashes Metal on the first unlit draw.

Its flutter3d_* dependencies ask for ^0.7.1, and it asks for vector_math ^2.4.3.

0.7.0

Breaking. markTestSkipped and test, this package's one Flutter import, come from package:test instead of flutter_test — the same two symbols, so no check changes. Nothing else here ever named Flutter directly; the import outlived the reason for it, from before flutter3d_hardware had a dart:ui-free GraphicsDevice of its own to test. flutter3d_shaders dropped its own flutter: sdk dependency in the same release, so this package no longer resolves the Flutter SDK through it either, and a backend's conformance_test.dart runs under dart test. Floors to flutter3d_hardware ^0.7.0 and flutter3d_shaders ^0.7.0.

Four new checks, thirty-seven in all, and twenty-eight of them need the shader pipeline. a texture region overwrite lands only where it was aimed patches a 2x2 corner of a 4x4 texture through overwriteTexture and reads all sixteen texels back. a geometry overwrite draws what a fresh upload draws and a geometry overwrite leaves its neighbours untouched hold overwriteGeometry to the picture and to the bytes either side of the write, and expect a write past the end to be refused. the colour attachment limit is honoured in both directions asks for a second attachment that receives its own output and for a pass past maxColorAttachments that throws. A backend that answers one declines the first half and is still held to the second. A backend written against 0.6.0 meets all four as failures until it implements the three new GraphicsDevice members.

The link check covers the stages the engine added. ShadowDepthMasked and ShadowDistanceMasked link against each of the four mesh vertex stages, SsaoBlur, ContactShadow, LightShafts, DepthOfField and ViewportShade against the fullscreen one, and two pairs are new: PolylineVertex with Unlit, and ParticleVertex with Splat. Fxaa is in kRequiredShaders and is not in this list.

A device factory may decline. A ConformanceDeclined thrown while makeDevice runs skips the check with the sentence it carries, the way a declined check already did. Headless Chrome on a CI runner has navigator.gpu and hands out no adapter, which is nothing a check could have been asked about. Any other exception from the factory is still a failure.

The archive carries a skill for a coding agent, skills/flutter3d-conformance-running-the-suite/, about the suite that says a backend is finished and how declining differs from passing. dart run skills@ get installs it.

0.6.0

• A device factory may now answer with a Future. DeviceFactory returns FutureOr<GraphicsDevice> instead of GraphicsDevice, and the runner awaits it. The fourth backend cannot be built by a constructor — requestAdapter() and requestDevice() are both promises — while Impeller, WebGL2 and the software rasteriser are opened with a call. Nothing in flutter3d_hardware says how a device is made; the contract starts once one exists, which is why the widening costs the three backends that were here first exactly nothing and changed none of their call sites.
• Breaking only for code that stores the typedef. A factory that returns a device is still a valid FutureOr factory, so passing one is unchanged; a caller that held a DeviceFactory and used the result without awaiting is the one that has to add the await. That is the whole of the API change.
• The checks themselves are unchanged in number and in what they assert. What a backend declines it still declines by name.

0.5.1

• A new check: a float texture uploads as floats. Two of the three backends answered no and neither said so — one refused sixteen bytes a texel as the wrong size, the other filled float storage with bytes. Both looked finished. The check uploads a one-texel r32g32b32a32Float holding three quarters and samples it in a vertex stage, which is where morph deltas are actually read, and fails with a sentence naming the likely cause.
• Thirty-three checks in all, twenty-five of which link stages and draw.

0.5.0

• Follows the hardware layer's version. The checks are unchanged; the device suite now passes all thirty-one on Metal, including the blend constant, the multisample resolve and the object id, which had never been run on hardware.

0.4.2

• A block missing a member the caller named is refused. The one half of bindUniformBlock this suite named as deliberately outside it, and the half the two hardware backends had quietly disagreed about. It asks the backend which kind it is before holding it to the rule: a backend that answers false for a block no shader declares is one that can see inside a shader, and is then required to throw for a member the block does not have. One that accepts any block name has no reflection, and the check says so and stops rather than passing by construction.
• A sampler asking for more anisotropy than there is is accepted. Bound on a trilinear sampler, drawn through the textured particle stage and read back at a texel centre — a check that the bind lands on a device that allows fewer taps, which is three different clamps behind one promise: flutter_gpu clamps inside its bind, WebGL2 has to be clamped before texParameterf, and the software rasteriser ignores the field. Asked twice: at sixteen, the number the engine's documentation names, and at maxAnisotropy * 2, which is above the ceiling on any device — sixteen is also what every desktop context answers, so it alone would never reach a clamp. The capability check asks maxAnisotropy and requires at least one.
• a library loaded from bytes answers to the same names. The backend's own shaders, packed as a bundle by the harness — OwnShaderSection is the section id, the bytes and the SDK to stamp, or null from a backend that compiles nothing — are loaded through loadShaders and held to answering every name in kRequiredShaders, linking MeshVertex + Pbr through the loaded handles, keeping those handles' identity across a refresh, refusing bytes that are not a bundle with ShaderBundleRefused, and never answering a claimed stage no section holds with a handle. runDeviceConformance takes ownShaders and conformanceChecksWith builds the list for a harness that is not a test runner. The same check then refreshes with a bundle that no longer names Pbr while the Pbr handle is in use, and requires a ShaderBundleRefused naming the bundle and the stage with the library left as it was — the half of the identity promise the three backends had been keeping three different ways.
• a readback returns the frame before, in the core tier: red is cleared, a readback is asked for and not awaited, blue is cleared over it, and the answer has to be red. The same check holds a two-by-three region to two-by-three pixels and refuses a deviceTransient texture, a region past the edge and a texture in the device's own hdrColorFormat with an ArgumentError rather than an answer — the last because a half-float readback was three different answers on three backends, one of them a picture of zeros with no error.
• a readback of a region reads that region, in the shader tier: the top half of a picture is painted, a region in the top-left quarter has to be painted and one in the bottom-left not, and a region straddling the edge has its painted rows first — the check that found a backend measuring a region's y from the wrong edge.
• The link checks pair ObjectId with the three vertex stages the picking pass draws through and Luminance with the full-screen one.
• a stencil test keeps what it should. A mark written where a mesh is, through BlendState.keepDestination so the picture is left alone, then equal landing only on the mark and notEqual only off it — three pixels read back, and each of the three ways to be wrong named in its failure. A backend whose supportsStencil is false is asked nothing, as with a compressed format it does not sample. The capability check reads supportsStencil with the rest.
• The link checks pair Xray with all three mesh vertex stages. It declares one output where every other lighting entry point declares two, which is the pairing least like the rest of the table.
• a pass renders into a cube face and a mip. Three clears into three subresources of one cube — a face's base, another face's base, and the first face's second level where the device says it can — read back through the probe prefilter stage with a single tap, so a backend that cleared the whole cube, the base level, or face zero ends with the wrong colour in the wrong place. The capability check reads supportsRenderToMip; the link checks pair FullscreenVertex with ProbePrefilter.
• Three checks, and a way for a backend to decline one. A blend constant reaches the blend, or is refused draws through BlendFactor.blendColor against a known constant and demands the two refusals from a backend that has none. A multisample resolve resolves is the first check to carry RenderTargetSpec.sampleCount, ColorTarget.resolveTexture and StoreAction.multisampleResolve at all. An object id survives the draw and the readback binds IdInfo with a known id, draws through the standard five-attribute layout — which nothing else here does — and decodes r + g·256 + b·65536.
• ConformanceDeclined and decline. A check the backend cannot be asked used to return, which every harness reported as a pass. It is now a skip that names the backend and what it answered, in the test runner and in the application harness's log and tally alike.

0.4.1

• The link checks pair MeshLightmappedVertex with the lit models.
• A compressed format the backend claims is drawn and read back. One hand-built BC1 block and one ETC2 block, each uploaded where supportsTextureFormat says yes and sampled at the centre of a quad; a backend that answers no to both runs nothing, which is the honest outcome for the software rasteriser. The capability check asks supportsTextureFormat for every TextureFormat value.

0.4.0

• No changes of its own; the version moves with the workspace, whose sibling constraints name a single release. The README's closing section now says what the engine around this package is.

0.3.0

• No changes of its own. The workspace is released as a set, in the order ARCHITECTURE.md §16 gives, so this package's version moves with the rest and its constraints on its siblings move with it.

0.2.0

• The behaviour flutter3d_hardware requires, as a suite a backend runs against itself: formats that must be renderable, pixels that must keep their row order, and every stage pair the engine actually links.

• Split into coreChecks and shaderChecks, because a suite that claimed to need no shaders met a new backend with five failures it could do nothing about.

• A sampler asking for more anisotropy than there is is accepted. Bound on a trilinear sampler, drawn through the textured particle stage and read back at a texel centre — a check that the bind lands on a device that allows fewer taps, which is three different clamps behind one promise: flutter_gpu clamps inside its bind, WebGL2 has to be clamped before texParameterf, and the software rasteriser ignores the field. Asked twice: at sixteen, the number the engine's documentation names, and at maxAnisotropy * 2, which is above the ceiling on any device — sixteen is also what every desktop context answers, so it alone would never reach a clamp. The capability check asks maxAnisotropy and requires at least one.

• a library loaded from bytes answers to the same names. The backend's own shaders, packed as a bundle by the harness — OwnShaderSection is the section id, the bytes and the SDK to stamp, or null from a backend that compiles nothing — are loaded through loadShaders and held to answering every name in kRequiredShaders, linking MeshVertex + Pbr through the loaded handles, keeping those handles' identity across a refresh, refusing bytes that are not a bundle with ShaderBundleRefused, and never answering a claimed stage no section holds with a handle. runDeviceConformance takes ownShaders and conformanceChecksWith builds the list for a harness that is not a test runner. The same check then refreshes with a bundle that no longer names Pbr while the Pbr handle is in use, and requires a ShaderBundleRefused naming the bundle and the stage with the library left as it was — the half of the identity promise the three backends had been keeping three different ways.

• a readback returns the frame before, in the core tier: red is cleared, a readback is asked for and not awaited, blue is cleared over it, and the answer has to be red. The same check holds a two-by-three region to two-by-three pixels and refuses a deviceTransient texture, a region past the edge and a texture in the device's own hdrColorFormat with an ArgumentError rather than an answer — the last because a half-float readback was three different answers on three backends, one of them a picture of zeros with no error.

• a readback of a region reads that region, in the shader tier: the top half of a picture is painted, a region in the top-left quarter has to be painted and one in the bottom-left not, and a region straddling the edge has its painted rows first — the check that found a backend measuring a region's y from the wrong edge.

• The link checks pair ObjectId with the three vertex stages the picking pass draws through and Luminance with the full-screen one.

• a stencil test keeps what it should. A mark written where a mesh is, through BlendState.keepDestination so the picture is left alone, then equal landing only on the mark and notEqual only off it — three pixels read back, and each of the three ways to be wrong named in its failure. A backend whose supportsStencil is false is asked nothing, as with a compressed format it does not sample. The capability check reads supportsStencil with the rest.

• The link checks pair Xray with all three mesh vertex stages. It declares one output where every other lighting entry point declares two, which is the pairing least like the rest of the table.

• a pass renders into a cube face and a mip. Three clears into three subresources of one cube — a face's base, another face's base, and the first face's second level where the device says it can — read back through the probe prefilter stage with a single tap, so a backend that cleared the whole cube, the base level, or face zero ends with the wrong colour in the wrong place. The capability check reads supportsRenderToMip; the link checks pair FullscreenVertex with ProbePrefilter.

0.4.1

• The link checks pair MeshLightmappedVertex with the lit models.
• A compressed format the backend claims is drawn and read back. One hand-built BC1 block and one ETC2 block, each uploaded where supportsTextureFormat says yes and sampled at the centre of a quad; a backend that answers no to both runs nothing, which is the honest outcome for the software rasteriser. The capability check asks supportsTextureFormat for every TextureFormat value.

0.4.0

• No changes of its own; the version moves with the workspace, whose sibling constraints name a single release. The README's closing section now says what the engine around this package is.

0.3.0

• No changes of its own. The workspace is released as a set, in the order ARCHITECTURE.md §16 gives, so this package's version moves with the rest and its constraints on its siblings move with it.

0.2.0

• The behaviour flutter3d_hardware requires, as a suite a backend runs against itself: formats that must be renderable, pixels that must keep their row order, and every stage pair the engine actually links.
• Split into coreChecks and shaderChecks, because a suite that claimed to need no shaders met a new backend with five failures it could do nothing about.