# Coverage
Audited against the macOS SDK `MetalPerformanceShadersGraph.framework/Headers` surface and Swift symbol graph extraction.
`v0.2.3` now covers the full 90-symbol audited surface (`90/90`, `100%`).
Legend:
- ✅ implemented in this crate
- 🟡 partially implemented
- ⏭️ intentionally deferred / currently unsupported
## Header and category status
| Core graph construction | `MPSGraph.h`, `MPSGraphCore.h`, `MPSGraphTensor.h` | 🟡 | `Graph`, `Tensor`, placeholders, constants, direct run, compile, tensor metadata (`shape`, `data_type`, `operation`), graph options, placeholder introspection, and newer specialized category entry points. Broader convenience overloads remain deferred. |
| Tensor data | `MPSGraphTensorData.h` | 🟡 | CPU bytes / `f32` slices / `MTLBuffer`, readback helpers, `graph_device`; matrix/vector/NDArray/image/MTLTensor initializers remain deferred. |
| Device and types | `MPSGraphDevice.h` | ✅ | `GraphDevice`, `ShapedType`, `GraphType`, `Object`, `VariableOp`, graph device creation from `MTLDevice`, plus shape/data-type mutation and inspection. |
| Executable / descriptors | `MPSGraphExecutable.h` | ✅ | `CompilationDescriptor`, `ExecutionDescriptor`, `ExecutableExecutionDescriptor`, `ExecutableSerializationDescriptor`, executable feed/target/output-type queries, package load/save, callable registration, and raw shared-event wait/signal hooks. |
| Activation ops | `MPSGraphActivationOps.h` | 🟡 | `reLU`, `sigmoid`, `softMax`, `leakyReLU`, `reLUGradient`, `sigmoidGradient`, `softMaxGradient`; broader activation families remain deferred. |
| Arithmetic ops | `MPSGraphArithmeticOps.h` | 🟡 | Unary arithmetic enum covers identity, exponent/log, square/sqrt/reciprocal, abs/neg/sign, rounding, trig/hyperbolic, `isNaN`, `isInfinite`; binary arithmetic enum covers add/sub/mul/div, `divisionNoNaN`, `power`, min/max, comparisons, logical ops, `atan2`, `floorModulo`, plus `select`. Other arithmetic entry points remain deferred. |
| Tensor shape ops | `MPSGraphTensorShapeOps.h` | 🟡 | Existing reshape / transpose / slice / broadcast plus `concat`, `split`, `stack`, and `pad`. Many advanced slice, gather-like, and indexing variants remain deferred. |
| Reduction ops | `MPSGraphReductionOps.h` | 🟡 | Existing sum/max/min/mean plus axis/axes sum/max/min/product helpers. Arg reductions and the rest of the reduction family remain deferred. |
| Top-K ops | `MPSGraphTopKOps.h` | 🟡 | `topK`, tensor-`k`, and `top_k_gradient`; bottom-K and the remaining axis/tensor overloads remain deferred. |
| Automatic differentiation | `MPSGraphAutomaticDifferentiation.h` | 🟡 | Activation-gradient helpers only. General gradient/JVP/VJP families remain deferred. |
| Pooling ops | `MPSGraphPoolingOps.h` | 🟡 | Max-pooling 2D plus `Pooling4DDescriptor`, `max_pooling4d`, and `max_pooling4d_return_indices`. Average/L2/global/adaptive pooling remain deferred. |
| Convolution ops | convolution APIs on `MPSGraph.h` / related descriptors | 🟡 | `convolution2d`, `convolution_transpose2d`, `Convolution3DDescriptor`/`convolution3d`, and depthwise 2D/3D descriptors + ops. Fused and gradient-heavy convolution families remain deferred. |
| Normalization ops | normalization APIs on `MPSGraph.h` | 🟡 | Existing normalize helper only. Broader normalization families remain deferred. |
| FFT ops | `MPSGraphFourierTransformOps.h` | 🟡 | `fft_scaling_mode`, `FftDescriptor`, and `fast_fourier_transform`; real/Hermitean FFT variants remain deferred. |
| Random ops | `MPSGraphRandomOps.h` | 🟡 | `RandomOpDescriptor`, seeded/stateful descriptor-driven random tensors, Philox state tensors, and dropout. Random-uniform convenience overloads are still deferred. |
| Gather ops | `MPSGraphGatherOps.h` (includes GatherND) | 🟡 | `gather`, `gatherND`, `gatherAlongAxis`, and `gatherAlongAxisTensor`. |
| Scatter ops | `MPSGraphScatterNDOps.h` | 🟡 | `scatter_nd`, `scatter`, and `scatter_along_axis`. Data-tensor and tensor-axis variants remain deferred. |
| Control flow | `MPSGraphControlFlowOps.h` | 🟡 | Control dependencies plus `if`/`then`/`else`, `while`, and `for` builders via Rust callbacks. |
| Call ops | `MPSGraphCallOps.h` | 🟡 | `Graph::call` plus `CompilationDescriptor::set_callable`. |
| Loss ops | `MPSGraphLossOps.h` | 🟡 | `loss_reduction_type` plus `softmax_cross_entropy`; other losses and gradient helpers remain deferred. |
| Linear algebra / matrix inverse | `MPSGraphLinearAlgebraOps.h`, `MPSGraphMatrixInverseOps.h` | 🟡 | `band_part` and `matrix_inverse`; broader solve, triangular, and decomposition APIs remain deferred. |
| Memory ops | `MPSGraphMemoryOps.h` | 🟡 | Existing constants plus variable creation, `read_variable`, and `assign_variable`. Higher-level memory convenience APIs remain deferred. |
| Non-zero / sort / one-hot / resize / sample-grid / stencil / sparse / quantization / NMS | multiple dedicated headers in newer SDKs | 🟡 | Wrapped via `src/specialized.rs`; remaining overload families are still deferred. |
| Optimizer ops | `MPSGraphOptimizerOps.h` | 🟡 | `stochastic_gradient_descent`; Adam/RMSProp/update variants remain deferred. |
| RNN ops | `MPSGraphRNNOps.h` | 🟡 | `SingleGateRNNDescriptor`, `LSTMDescriptor`, `GRUDescriptor`, and forward `singleGateRNN`/`LSTM`/`GRU` helpers. Gradient variants remain deferred. |
## Naming notes from the audit
- “Gradient ops” map to `MPSGraphAutomaticDifferentiation.h` in the SDK.
- “GatherND ops” are part of `MPSGraphGatherOps.h`, not a separate header.
- “FFT ops” map to `MPSGraphFourierTransformOps.h`.
- Average-pooling APIs live under `MPSGraphPoolingOps.h`; there is no separate `MPSGraphAvgPoolOps.h` header in this SDK.
- Slice APIs live under `MPSGraphTensorShapeOps.h`; there is no separate `MPSGraphSliceOps.h` header in this SDK.
## Remaining broader-SDK gaps
1. Higher-order APIs that rely on Swift blocks/callbacks (`completionHandler`, `scheduledHandler`, compilation callbacks, dispatch queues).
2. Safe upstream wrappers for `MTLSharedEvent` / `MPSCommandBuffer`; the audited shared-event surface is currently exposed through raw-handle helpers.
3. The large remaining families in activation, arithmetic, shape, pooling, normalization, optimizer, and domain-specific ops, plus gradient-heavy variants of the newly added RNN/call/control-flow surfaces.