heterogeneous - Module

`enum Engine`

Source: heterogeneous.joule:18

`fn is_single(&self) -> bool`

Is this a single engine or cooperative?

Source: heterogeneous.joule:37

`fn components(&self) -> Vec<Engine>`

Get component engines for cooperative modes

Source: heterogeneous.joule:42

`enum ComputeDType`

Source: heterogeneous.joule:55

`fn bits(&self) -> usize`

Size in bits

Source: heterogeneous.joule:71

`fn bytes(&self) -> usize`

Size in bytes (rounded up)

Source: heterogeneous.joule:84

`fn is_float(&self) -> bool`

Is this a floating point type?

Source: heterogeneous.joule:89

`fn is_quantized(&self) -> bool`

Is this a quantized/integer type?

Source: heterogeneous.joule:95

`fn preferred_engine(&self) -> Engine`

Preferred engine for this dtype

Source: heterogeneous.joule:101

`enum OpClass`

Source: heterogeneous.joule:121

`fn classify(op: &str, m: usize, n: usize, k: usize, batch: usize) -> Self`

Classify operation from name and dimensions

Source: heterogeneous.joule:154

`fn preferred_engine(&self, dtype: ComputeDType) -> Engine`

Get preferred engine for this operation class and dtype

Source: heterogeneous.joule:182

`fn offload_threshold(&self) -> u64`

Minimum FLOPS to justify offload to accelerator

Source: heterogeneous.joule:231

`struct EngineCapabilities`

Source: heterogeneous.joule:249

`fn peak_ops_per_sec(&self, dtype: ComputeDType) -> f64`

Estimate peak performance for a given dtype

Source: heterogeneous.joule:278

`fn efficiency_tops_per_watt(&self, dtype: ComputeDType) -> f64`

Energy efficiency in TOPS/Watt

Source: heterogeneous.joule:291

`fn arithmetic_intensity_threshold(&self, dtype: ComputeDType) -> f64`

Arithmetic intensity threshold (roofline model) Operations with AI below this are memory-bound

Source: heterogeneous.joule:297

`fn apple_m5_cpu() -> Self`

Apple M5 CPU (estimated specs)

Source: heterogeneous.joule:307

`fn apple_m5_gpu() -> Self`

Apple M5 GPU (estimated specs)

Source: heterogeneous.joule:328

`fn apple_m5_npu() -> Self`

Apple M5 Neural Engine (estimated specs)

Source: heterogeneous.joule:349

`fn apple_m5_all() -> Vec<Self>`

Get all Apple M5 engines

Source: heterogeneous.joule:369

`struct ScheduleOp`

Source: heterogeneous.joule:384

`fn new(`

Create from operation parameters

Source: heterogeneous.joule:409

`fn with_dep(mut self, dep_id: usize) -> Self`

Add dependency

Source: heterogeneous.joule:437

`fn with_deps(mut self, deps: &[usize]) -> Self`

Add multiple dependencies

Source: heterogeneous.joule:443

`fn arithmetic_intensity(&self) -> f64`

Arithmetic intensity (FLOPS / bytes)

Source: heterogeneous.joule:449

`fn is_compute_bound(&self, caps: &EngineCapabilities) -> bool`

Is this operation compute-bound on given engine?

Source: heterogeneous.joule:458

`fn should_offload(&self) -> bool`

Should this operation be offloaded from CPU?

Source: heterogeneous.joule:464

`struct ScheduleDecision`

Source: heterogeneous.joule:471

`struct TransformerSchedule`

Source: heterogeneous.joule:492

`enum BufferLocation`

Source: heterogeneous.joule:519

`struct Buffer`

Unified memory buffer that can be used by any engine On unified memory systems (Apple Silicon), no copies needed

Source: heterogeneous.joule:532

`fn zeros(shape: &[usize], dtype: ComputeDType) -> Self`

Create new zero-initialized buffer

Source: heterogeneous.joule:542

`fn from_vec(data: Vec<T>, shape: &[usize], dtype: ComputeDType) -> Self`

Create from existing data

Source: heterogeneous.joule:556

`fn compute_strides(shape: &[usize]) -> Vec<usize>`

Compute row-major strides

Source: heterogeneous.joule:571

`fn shape(&self) -> &[usize]`

Get shape

Source: heterogeneous.joule:580

`fn strides(&self) -> &[usize]`

Get strides

Source: heterogeneous.joule:585

`fn len(&self) -> usize`

Get total element count

Source: heterogeneous.joule:590

`fn size_bytes(&self) -> usize`

Get size in bytes

Source: heterogeneous.joule:595

`fn dtype(&self) -> ComputeDType`

Get dtype

Source: heterogeneous.joule:600

`fn location(&self) -> BufferLocation`

Get current location

Source: heterogeneous.joule:605

`fn ensure_on(&mut self, engine: Engine)`

Ensure buffer is accessible on engine (no-op for unified memory)

Source: heterogeneous.joule:610

`fn as_ptr(&self) -> *const T`

Get raw data pointer

Source: heterogeneous.joule:616

`fn as_mut_ptr(&mut self) -> *mut T`

Get mutable raw data pointer

Source: heterogeneous.joule:621

`fn as_slice(&self) -> &[T]`

Get data slice

Source: heterogeneous.joule:626

`fn as_mut_slice(&mut self) -> &mut [T]`

Get mutable data slice

Source: heterogeneous.joule:631

`struct Gemm`

General Matrix Multiply: C = α * A @ B + β * C

Source: heterogeneous.joule:641

`fn default() -> Self`

Source: heterogeneous.joule:649

`fn new() -> Self`

Source: heterogeneous.joule:660

`fn transposed_a(mut self) -> Self`

Source: heterogeneous.joule:664

`fn transposed_b(mut self) -> Self`

Source: heterogeneous.joule:669

`fn with_alpha(mut self, alpha: f32) -> Self`

Source: heterogeneous.joule:674

`fn with_beta(mut self, beta: f32) -> Self`

Source: heterogeneous.joule:679

`fn execute_f32(`

Execute GEMM on the optimal engine

Source: heterogeneous.joule:685

`fn execute_cpu_f32(&self, a: &Buffer<f32>, b: &Buffer<f32>, c: &mut Buffer<f32>, m: usize, n: usize, k: usize)`

Source: heterogeneous.joule:718

`fn execute_gpu_f32(&self, a: &Buffer<f32>, b: &Buffer<f32>, c: &mut Buffer<f32>, m: usize, n: usize, k: usize)`

Source: heterogeneous.joule:745

`fn execute_cpu_f32(&self, _a: &Buffer<f32>, _b: &Buffer<f32>, _c: &mut Buffer<f32>, _m: usize, _n: usize, _k: usize)`

Source: heterogeneous.joule:752

`fn execute_gpu_f32(&self, _a: &Buffer<f32>, _b: &Buffer<f32>, _c: &mut Buffer<f32>, _m: usize, _n: usize, _k: usize)`

Source: heterogeneous.joule:757

`struct GemmResult`

Source: heterogeneous.joule:764

`fn scheduler() -> &'static Scheduler`

Get the global heterogeneous scheduler

Source: heterogeneous.joule:779

`fn init()`

Initialize the compute subsystem

Source: heterogeneous.joule:786