# TODO - v0.1.0
## Current Status
This crate is part of the sklears v0.1.0 initial release.
## Future Enhancements
- Performance optimizations
- Additional scikit-learn API coverage
- Enhanced documentation and examples
## OxiCUDA Migration (v0.2.0)
Migration status (2026-07-06): done except one explicitly deferred sub-item.
`gpu_computing` no longer fabricates any GPU hardware; device enumeration and
the core array ops (`add`/`mul`/`matmul`) route through the oxicuda-backed
`sklears_core::gpu` module (plus a direct optional `oxicuda-driver` dep for
device metadata `sklears_core::gpu` doesn't expose, e.g. SM count) behind a
new `gpu` feature. Default build stays CPU-only Pure Rust; `cargo check -p
sklears-utils` and `--features gpu` both pass warning-free, and all 491 tests
(29 of them in `gpu_computing`) pass in both configurations.
- [x] (S) Decide `gpu_computing.rs` strategy: thin layer over `sklears_core::gpu` (preferred) or direct oxicuda deps. Implemented the preferred thin-layer approach: `Cargo.toml` adds `gpu = ["sklears-core/gpu_support", "dep:oxicuda-driver"]` (the direct `oxicuda-driver` dep is only for device metadata -- SM count, clock rate, memory bus width -- that `sklears_core::gpu::GpuDeviceProperties` doesn't expose; all actual GPU *compute* goes through `sklears_core::gpu::{GpuBackend, GpuArray, GpuMatrixOps}`). `gpu_computing` module docs now state the honesty contract explicitly. Files: `Cargo.toml`, `src/gpu_computing.rs`, `src/lib.rs` (no lib.rs changes were needed -- the public re-export list at `src/lib.rs:250-252` still resolves, `GpuDevice`'s field set changed but the type name didn't).
- [x] (M) Replace fabricated GPU device list with real oxicuda-driver enumeration. `GpuUtils::init_devices` now calls `enumerate_real_devices()`, which (behind feature `gpu`) walks `oxicuda_driver::Device::count()`/`Device::get(i).info()` for real `name`/`total_memory`/`free_memory`/`compute_capability`/`multiprocessor_count` (SM count, not the old fake "cores")/`clock_rate_mhz`, plus a real-hardware-derived `memory_bandwidth_bytes_per_sec` (`2 * mem_clock_hz * bus_width_bytes`, the standard deviceQuery formula -- documented as a derived estimate, not a measurement). Without the `gpu` feature, or with no CUDA driver/device (this dev machine), it returns an empty `Vec`, never fake hardware. Dropped the unfillable `is_integrated` field (no such flag in `oxicuda-driver`'s `DeviceInfo`) rather than guess at it; `get_best_device` no longer filters on it. Files: `src/gpu_computing.rs`.
- [x] (M, was scoped (L)) Back `GpuArrayOps::{add_arrays,multiply_arrays,matrix_multiply}` with oxicuda, keep CPU fallback. Each now tries `sklears_core::gpu::{GpuBackend::with_device_id, GpuArray, GpuMatrixOps}` (real oxicuda-blas GEMM/elementwise add/mul) first; `Ok(None)` from the GPU helper (no such device / `gpu` feature off) transparently falls back to the existing CPU code, matching the `sklears-svm` `DeviceNotAvailable`-skip pattern. A genuine GPU-side error (as opposed to "no GPU here") is surfaced as `GpuError::InitializationFailed`, not silently swallowed.
- (deferred 2026-07-06: `apply_activation`/`reduce_sum`/`reduce_max` stay CPU-only -- there is no `oxicuda-primitives` elementwise/reduction kernel to wire them to yet, and building one is a materially larger task than this pass. `execute_kernel`'s `thread::sleep(1ms)` timing mock, the "optimized executor" variant, and the rest of the scheduling scaffolding (`MultiGpuCoordinator`, `AsyncGpuOps`, `GpuOptimizationAdvisor`, `GpuMemoryPool`) are unchanged: they have no real kernel payload attached anywhere in this abstraction (kernels are identified only by a name string + grid/block metadata, with no dispatch table), so there is nothing concrete to back with a device launch without a larger redesign. Doc comments were updated to say plainly that these remain CPU-side bookkeeping rather than implying real device execution.) The new doc comments/helpers pushed the file to 2110 lines, over the 2000-line policy limit, so it was split with `splitrs --split-test-modules` into `src/gpu_computing/mod.rs` (1323 lines, production code) and `src/gpu_computing/tests.rs` (455 lines); `src/gpu_computing.rs` no longer exists as a single file. `pub mod gpu_computing;` in `src/lib.rs` needed no change (directory modules resolve the same way).
- [x] (S) Record `distributed_computing` GPU fields as out of scope. Added module-doc notes (2026-07-06) to `src/distributed_computing/types.rs` and `src/distributed_computing/functions.rs` stating that `gpu_count`/`gpu_usage`/`min_gpu_count`/`gpu_time`/`total_gpu_count` are pure scheduling/capacity metadata with no GPU API calls, so future GPU audits don't re-flag them. Files: `src/distributed_computing/types.rs`, `src/distributed_computing/functions.rs`, `TODO.md`.
## Correctness Fixes (v0.2.0)
- [x] (S) Fixed a Miri-detected Stacked-Borrows violation in `memory::MemoryPool::add_block`.
Raw pointers into a freshly-allocated block were taken via `block.iter_mut()` *before*
the block was moved into `self.blocks`, but passing a `Box<[T]>` by value into
`Vec::push` forms a fresh `noalias` reborrow over the box's entire pointee, retroactively
invalidating any pointers taken beforehand. The block is now pushed into its final
storage first, and pointers are derived from the stored copy instead. Files: `src/memory.rs`.
See the main [workspace TODO](../../TODO.md) for overall project roadmap.