Bring support for wasm and mac

.github/workflows/ci.yml

@@ -47,6 +47,22 @@ jobs:
 
       - run: cargo doc --workspace --no-deps
 
+  wasm:
+    name: Check WASM
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+
+      - uses: cwfitzgerald/repo-common/.github/actions/install-rust@trunk
+        with:
+          version: ${{ env.CI_RUST_VERSION }}
+          components: clippy
+          targets: wasm32-unknown-unknown
+
+      - run: cargo clippy -p wgpu-ffx --target wasm32-unknown-unknown --features wgpu/webgpu
+
+      - run: cargo doc -p wgpu-ffx --target wasm32-unknown-unknown --features wgpu/webgpu --no-deps
+
   test:
     name: Test (${{ matrix.os }})
     strategy:

CHANGELOG.md

@@ -20,6 +20,107 @@ Per Keep a Changelog there are 6 main categories of changes:
 
 ## Unreleased
 
+FSR3 no longer requires `TEXTURE_ADAPTER_SPECIFIC_FORMAT_FEATURES`. Devices
+are classified into a `FormatProfile` (`Core` | `Tier2` | `Native`),
+auto-detected at context creation, and the upscaler selects storage formats
+and shader variants the device actually supports:
+
+- **`Core`** — baseline WebGPU (`core-features-and-limits`). Works in the
+  browser; the crate now builds for `wasm32-unknown-unknown` (CI-verified).
+- **`Tier2`** — the `texture-formats-tier2` capability set. Auto-detected on
+  adapters whose reported capabilities cover the tier-2 set but not
+  `rg16float` read-write storage — notably Metal.
+- **`Native`** — adapter-specific format features with full reported
+  capabilities; unchanged behavior.
+
+Detection does not trust `TEXTURE_ADAPTER_SPECIFIC_FORMAT_FEATURES` alone:
+each profile's required per-format capabilities are verified through
+`Adapter::get_texture_format_features` before the profile is selected.
+
+### Added
+
+- `FormatProfile` with `FormatProfile::from_adapter` detection and
+  `FormatProfile::supported_by`.
+- `FsrFormats` — the formats the caller must use for each texture it
+  provides. Obtainable before a context exists via `FormatProfile::formats()`,
+  or from a live context via `FsrContext::formats()`.
+- `FsrContext::format_profile()` — the profile the context was created with.
+
+### Changed
+
+- Breaking: `FsrContextInfo` gained an `adapter: wgpu::Adapter` field (used to
+  query per-format capabilities) and a `format_profile: Option<FormatProfile>`
+  field. `None` auto-detects; `Some` forces a profile (e.g. to exercise
+  `Core` on a desktop adapter). `FsrContext::new` panics if a forced profile
+  is unsupported, naming the missing capability.
+
+  ```diff
+   FsrContextInfo {
+  +    adapter,
+       device,
+       flags,
+  +    format_profile: None,
+   }
+  ```
+
+- Breaking: caller texture formats are profile-dependent and validated.
+  `dilated_motion_vectors` is `Rg32Float` on `Core` (vs `Rg16Float`
+  elsewhere). Allocate your textures from `FsrFormats` instead of
+  hardcoding — it covers all six caller-provided textures (`color`, `depth`,
+  `motion_vectors`, `output`, `dilated_depth`, `dilated_motion_vectors`).
+  `FsrContext::dispatch` now validates every caller-provided texture against
+  the active profile before recording any GPU work.
+
+  ```diff
+  +let formats = FormatProfile::from_adapter(&adapter, &device).formats();
+  +// or, with a live context: context.formats()
+   device.create_texture(&wgpu::TextureDescriptor {
+  -    format: wgpu::TextureFormat::Rg16Float,
+  +    format: formats.dilated_motion_vectors,
+       ..
+   })
+  ```
+
+- The internal Lanczos2 LUT uses `r16snorm` only when the device has the
+  `TEXTURE_FORMAT_16BIT_NORM` feature enabled (independent of the format
+  profile); otherwise it widens to filterable `r16float`.
+
+- Breaking: `FsrDispatchError` is now `#[non_exhaustive]` and gained a
+  `TextureFormatMismatch { texture, expected, actual }` variant. Exhaustive
+  matches need a wildcard arm.
+
+  ```diff
+   match err {
+  +    FsrDispatchError::TextureFormatMismatch { texture, expected, actual } => ..,
+  +    _ => ..,
+       ..
+   }
+  ```
+
+- Internal (no action needed):
+  - Per-frame metadata moved from a 1×1 read-write storage texture to a
+    storage buffer.
+  - On `Core`, the luma / shading-change pyramids run as a write-only
+    multi-pass chain instead of single-pass SPD; scene-average luma and
+    auto-exposure match Native closely (covered by comparison tests).
+  - Store-only storage images are now declared `writeonly` (a correctness
+    improvement on all profiles).
+
+### Fixed
+
+- `FsrContext::dispatch` now runs on wgpu's `webgpu` backend (Chromium WebGPU),
+  fixing two runtime panics that only surfaced in the browser:
+  - Internal textures are zeroed with render-pass clears (`LoadOp::Clear`)
+    instead of `CommandEncoder::clear_texture`, which is unimplemented on the
+    `webgpu` backend and required `Features::CLEAR_TEXTURE` natively. The SPD
+    mips texture gained `RENDER_ATTACHMENT` usage; the clear value is unchanged
+    (all-zero) and `Features::CLEAR_TEXTURE` is no longer needed.
+  - The validation error scope around the clears is now popped synchronously
+    only on native (it surfaces clearing bugs via a panic there). The previous
+    `pollster::block_on` cannot run on wasm, where error scopes resolve through
+    JS promises; on wasm the scope is skipped and errors surface through wgpu's
+    default uncaptured-error handler.
+
 ## v0.1.0
 
 Released 2026-04-21