Skip to content

CHANGELOG.md

Latest commit

 

History

History
126 lines (106 loc) · 12.5 KB

CHANGELOG.md

File metadata and controls

126 lines (106 loc) · 12.5 KB
Raw

Changelog

All notable changes to OxiONNX will be documented in this file.

The format is based on Keep a Changelog.

0.1.4 - 2026-06-08

Added

  • feat: Subgraph parsing + end-to-end execution for If, Loop, Scan operators — ONNX models containing control-flow nodes now parse and execute correctly from real .onnx files. Previously the If/Loop/Scan kernels existed and passed in-memory unit tests but were never reachable from a loaded model due to two disconnects:

    1. oxionnx-proto: AttributeValue now carries g: Option<Box<GraphProto>> for graph-valued attributes (wire field 6); parse_attribute parses nested subgraphs via the existing parse_graph recursively (both eager and streaming parser paths fixed simultaneously). build_subgraph in model.rs converts a nested GraphProto into a runtime Graph, synthesising Constant nodes for subgraph-local initializers without touching the 48-literal Graph struct.
    2. oxionnx-core: OpContext gains a weights: Option<&'a HashMap<String, Tensor>> field so model weights can be accessed by subgraph nodes without cloning. Session::dispatch_node now populates outer_scope, weights, and registry in all three dispatch paths (slot-write, inplace, default-execute) and the parallel rayon path.
    3. oxionnx-ops control-flow kernels (IfOp, LoopOp, ScanOp) pass ctx.weights through to execute_subgraph, giving subgraph nodes access to outer-model initializers at zero allocation cost.

  • feat: 9 new end-to-end integration tests (tests/control_flow_e2e.rs): If true/false branch, Loop accumulation, Loop zero-iteration, Loop with outer-scope weight, Scan element-wise map, Scan with running-state, If with sequential subgraph ops, If with subgraph-local initializer.

  • feat: 2 new oxionnx-proto unit tests: test_subgraph_attribute_parsed (parser round-trip), test_subgraph_attribute_wired_into_graph (build_subgraph populates Attributes.graphs).

  • perf: Phase F.13 complete — zero-copy execute_into_slots for AttentionOp and MultiHeadAttentionOp, the last two ops in the Phase F sweep. Backed by sdpa_into / sdpa_output_shape / reshape_from_heads_into / multi_head_attention_into in attention/core.rs (allocation-free kernels; existing public APIs unchanged). The no-outproj MHA path now skips the concat allocation entirely via reshape_from_heads_into; the with-outproj path writes the projected output directly into the caller's pooled buffer. Phase F slot-write sweep closed at 58 of 121 ops.

  • fix: SIMD path in scaled_dot_product_attention did not zero row_out between query rows — weighted_sum_v accumulates, so for seq_q > 1 with the simd feature enabled the result was incorrect. The extraction into sdpa_into naturally fixed this by writing directly into per-row slices of the caller buffer (no shared row_out temp at all).

  • feat: 21 new tests in oxionnx-ops/tests/output_slots_attention_test.rs: AttentionOp and MultiHeadAttentionOp slot-write correctness vs execute, broadcast mask, custom scale, seq_k ≠ seq_q, 4D Q, second-call resize (grow and shrink), qkv-projection, out-projection with/without bias, error parity, batch > 1.

  • perf: Phase F.14 complete — zero-copy execute_into_slots for 47 additional operators, bringing the total to 105 of 121 ops with hand-coded slot-write bodies. True zero-copy (no output Vec allocation): LayerNorm/GroupNorm/BatchNorm/RmsNorm/InstanceNorm/Softmax/LogSoftmax via new _into kernel variants in nn/normalization.rs; Not/IsInf/IsNaN/BitwiseNot/Shape/Size/Constant/ConstantOfShape inline into slot; CumSum inline (same-size as input). Reduced-allocation: ReduceSum/Mean/Max/Min/Prod/L1/L2/LogSum/LogSumExp/SumSquare via new reduce_with_into+reduce_output_shape in math/reduce.rs; ArgMax/ArgMin via arg_reduce_into; TopK (2 outputs) via top_k_into; variadic ops and comparison/bitwise binary ops via macro update (call+move avoiding Vec wrapper). 41 new tests in output_slots_f14_test.rs.

  • feat: NodeInfo and Session::nodes(): public per-node graph introspection — enumerate a model's compute nodes (name, op type, inputs, outputs, and a deterministic attribute summary) in topological order; NodeInfo re-exported as oxionnx::NodeInfo (#2).

  • ReshapeOp now reads the ONNX allowzero attribute: when allowzero=0 (default) a 0 in the target shape copies the corresponding input dimension; when allowzero=1 a 0 is a literal zero-size dimension. A new resolve_reshape(input_dims, numel, shape, allowzero) helper is re-exported from oxionnx-ops::shape. 7 new unit tests cover all allowzero code paths (test_reshape_allowzero_* in oxionnx-ops/src/shape/tests.rs).

  • console_error_panic_hook = "0.1.7" added as an optional workspace dependency wired into the wasm feature for improved panic messages in WebAssembly environments.

Changed

  • Updated CUDA / OxiCUDA dependencies to 0.1.8.
  • oxifft updated from 0.3 to 0.3.2.
  • wasm32 compatibility: thread-pool initialization in src/session/loading.rs is now skipped under #[cfg(target_arch = "wasm32")]; all rayon imports and parallel-execution functions in src/session/run/parallel.rs are likewise gated with #[cfg(not(target_arch = "wasm32"))]. WASM sessions now build cleanly without a thread pool.

Fixed

  • SplitOp now honors the opset-1..12 split attribute for chunk sizes when no split input tensor is present, fixing YOLO11 (opset 11) models that failed with reshape: element count mismatch (33600 vs [1, 128, 20, 20]) in the C2PSA attention block (#1).

0.1.3 - 2026-05-16

Added

  • feat: ProviderKind::DirectMl variant (behind directml feature) — ProviderKind enum now covers CPU, GPU (wgpu), CUDA, and DirectML
  • feat: SessionBuilder::with_provider_kinds() — non-trivial ordered execution provider list that routes dispatch at runtime; CPU is always the implicit terminal fallback
  • feat: SessionBuilder::provider_kinds() accessor — returns the configured provider list for introspection
  • feat: Session::provider_kinds() accessor — returns the provider list stored in the session
  • feat: provider-list dispatch loop in run_sequential_inner — when providers is non-empty, each node is attempted through the list in order; falls back to legacy heuristic path when list is empty (backward compatible)
  • feat: try_provider_list_dispatch helper — encapsulates CUDA / DirectML / GPU / CPU provider dispatch for the provider-list path; all providers return None-graceful results with CPU fallback guaranteed
  • ProviderKind re-exported from oxionnx crate root

0.1.2 - 2026-04-19

Added

  • Phase D — Operator-Native TypedTensor Dispatch: native_dtypes() + execute_typed() opt-in hooks on the Operator trait; 40 pilot operators (23 math, 14 NN, Identity, Cast, Reshape) execute natively without f32 round-trips; all other operators fall back transparently via a correct default implementation
  • Phase E — DirectML Execution Provider (Windows): new oxionnx-directml subcrate with DirectMLContext::try_new() cross-platform shim; Windows D3D12 context skeleton; HLSL compute shader scaffolds for MatMul, Add, Mul, Relu, Sigmoid; feature-gated behind directml; CPU fallback on non-Windows and unsupported ops
  • Phase F — Operator-Level IOBinding Reuse: supports_output_slots() + execute_into_slots() opt-in hooks on the Operator trait; SizeClassPool::acquire() integrated on the slot path; pilot implementations for 40 operators; IoBinding::take_output_buffer / put_output_buffer helpers for caller-owned buffer management
  • OnnxError::DTypeMismatch — new error variant for dtype validation in the typed dispatch path
  • TypedOpContext<'a> — parallel context struct for typed operator dispatch
  • Phase F complete — F.10 remainder sweep: all 121 operators opt into supports_output_slots = true; full slot-write path with pointer-identity pool reuse across the entire operator set
  • Phase F.12 — zero-copy execute_into_slots hand-coded bodies for 30 operators: 22 pilot ops (Phase D/F pilot), plus 9 shape ops (Squeeze/Unsqueeze/Flatten/Expand/Split/Tile/DepthToSpace/SpaceToDepth/ReverseSequence), 12 elementwise NN ops (Clip/LeakyRelu/PRelu/HardSigmoid/Celu/Elu/Selu/ThresholdedRelu/LpNorm/MeanVarianceNorm/Hardmax/Shrink), 6 conv/pool ops (MaxPool/AveragePool/GlobalAveragePool/GlobalMaxPool/Pad/Resize), and Gather/ScatterND/ScatterElements
  • D.3 scoped slice — MatMulOp::execute_typed with native dispatch for F32, F16, BF16, I8→I32, I32 dtypes; native_dtypes() declared; INT8×INT8→I32 triple-loop GEMM kernel; F16/BF16 kernels with f32 accumulator — no round-trip through f32 for quantized or half-precision matmul
  • oxionnx-ops/src/conv.rs refactored into six focused submodules (conv2d, im2col, winograd, pooling, tests) to clear headroom for D.3 Conv native typed kernels (v0.1.9+)
  • Integration test full_slot_coverage_smoke.rs: pointer-identity across 100 iterations + pool reuse assertions for the slot-write infrastructure

Improved

  • 100 new tests across typed dispatch, IOBinding, DirectML EP, and session slot paths (1074 total)
  • Operator trait is fully backward-compatible: all four new methods carry correct default implementations

0.1.1 - 2026-04-14

Added

  • RNN operators: GRU, LSTM, RNN, and sequence processing ops (oxionnx-ops/src/rnn.rs)
  • Extended ML operators: SVM, tree ensemble, decision tree classifiers/regressors (oxionnx-ops/src/ml_svm.rs, ml_tree.rs)
  • Attention operators: expanded multi-head attention and transformer support (oxionnx-ops/src/attention.rs)
  • Indexing operators: additional gather/scatter/slice variants (oxionnx-ops/src/indexing.rs)
  • Quantized inference ops: extended quantized operator coverage (oxionnx-ops/src/quantized.rs)
  • Spatial operators: additional convolution and pooling variants (oxionnx-ops/src/spatial.rs)

Improved

  • Fusion optimizer: major expansion of graph fusion rules and pattern matching (src/optimizer/fusion.rs)
  • Session API: improved execution flow and error handling (src/session.rs)
  • GPU compute context: refined WebGPU context management and kernel dispatch (oxionnx-gpu/src/context.rs)
  • GPU shaders: additional WGSL shader kernels (oxionnx-gpu/src/shaders.rs)
  • CUDA backend: conv, elementwise, and quantized op refinements (oxionnx-cuda/src/conv.rs, elementwise.rs)
  • Proto streaming parser and mmap loader improvements (oxionnx-proto/src/streaming_parser.rs, mmap_loader.rs)
  • SIMD ops: extended SIMD-accelerated operator paths (oxionnx-ops/src/simd_ops.rs)
  • Math and NN ops: expanded coverage (oxionnx-ops/src/math.rs, nn.rs)

0.1.0 - 2026-03-26

Added

  • Initial release of OxiONNX pure Rust ONNX inference engine
  • 147 ONNX operator implementations
  • GPU acceleration via wgpu (MatMul, Softmax, ReLU, Sigmoid, GELU, Reduce)
  • SIMD optimization (NEON aarch64, AVX2 x86_64)
  • Graph optimization: constant folding, operator fusion, CSE, dead code elimination
  • Multi-dtype support (DType enum, TypedTensor, TensorStorage)
  • INT8 quantized MatMul with per-channel scale/zero-point
  • Mixed-precision inference (f16 activations, f32 accumulation)
  • Control flow operators: If, Loop, Scan with nested subgraphs
  • Streaming token generation for autoregressive models
  • Async inference API (run_async)
  • Session serialization (save/load pre-optimized graphs)
  • Memory-mapped weight loading (mmap feature)
  • Tensor arena allocator with buffer pooling
  • Strided tensor views (zero-copy transpose, slice, squeeze)
  • Broadcasting iterator (zero-allocation)
  • AES-GCM model encryption (encryption feature)
  • WebAssembly support via wasm-bindgen
  • no_std support for oxionnx-core
  • Execution provider abstraction with CPU/GPU fallback chain
  • ONNX-ML operators: Linear, TreeEnsemble, SVM, Normalizer, Scaler, LabelEncoder, TfIdf
  • Model metadata extraction API
  • Model validation and schema checking
  • Graph diff utility for debugging optimizations
  • Model pruning (strip unused weights)
  • Symbolic shape propagation
  • Opset coverage report generator
  • Benchmark-based CPU/GPU path selection
  • Numerical tolerance validation
  • CI pipeline (GitHub Actions)
  • Property-based testing with proptest
  • Fuzz testing for protobuf parser
  • 595 tests, 0 clippy warnings