Crate nxpu_opt 

IR optimization passes for NxPU.

Provides a Pass trait, a PassManager with fixed-point iteration, and built-in optimization passes (constant folding, FMA fusion, dead code elimination).

Re-exports

pub use calibrate::CalibrationDataset;

pub use calibrate::CalibrationError;

pub use calibrate::CalibrationMethod;

pub use calibrate::CalibrationResult;

pub use calibrate::HistogramCollector;

pub use calibrate::TensorHistogram;

pub use calibrate::calibrate;

pub use calibrate::calibrate_kl_divergence;

pub use calibrate::calibrate_minmax;

pub use calibrate::calibrate_percentile;

pub use calibrate::per_channel_quantize;

pub use calibrate::run_calibration;

pub use fusion::OperatorFusion;

pub use layout::LayoutTransform;

pub use layout::TransposeInsertion;

pub use layout::TransposeRecord;

pub use layout::layout_permutation;

pub use layout::reorder_dims;

pub use memory::LiveInterval;

pub use memory::MemoryPlanning;

pub use memory::plan_memory;

pub use quantize::CalibrationData;

pub use quantize::F32ToBf16;

pub use quantize::F32ToF16;

pub use quantize::F32ToInt8;

pub use quantize::MixedPrecisionPass;

pub use quantize::MixedPrecisionPolicy;

pub use quantize::PerChannelQuantParams;

pub use quantize::QuantizationParams;

pub use schedule::Schedule;

pub use schedule::SchedulePass;

pub use schedule::ScheduleSlot;

pub use schedule::compute_schedules;

pub use schedule::format_schedule;

pub use shape::ShapeInference;

pub use tiling::Conv2DShape as TilingConv2DShape;

pub use tiling::MatMulShape as TilingMatMulShape;

pub use tiling::TileConfig;

pub use tiling::TilingDefaults;

pub use tiling::TilingPass;

pub use tiling::TilingPlan;

pub use tiling::tile_conv2d;

pub use tiling::tile_matmul;

pub use vectorize::VectorWidth;

pub use vectorize::VectorizationHint;

pub use vectorize::VectorizationPass;

pub use vectorize::analyze_vectorization;

pub use workgroup::OccupancyResult;

pub use workgroup::WorkgroupHwParams;

pub use workgroup::WorkgroupOptimization;

pub use workgroup::calculate_occupancy;

pub use workgroup::optimize_workgroup_size;

Modules

calibrate

Calibration pipeline for quantization.

fusion

Operator fusion pass.

layout

Memory layout optimization pass.

memory

Memory planning and buffer allocation pass.

quantize

Precision conversion passes for NPU quantization.

schedule

Operation scheduling pass.

shape

Shape inference pass.

tiling

Loop tiling and cache blocking.

vectorize

SIMD vectorization hints.

workgroup

Workgroup size optimization.

Structs

BufferAllocation

A buffer allocation for a single tensor.

CommonSubexprElimination

Common subexpression elimination pass.

ConstantFolding

Folds constant expressions (literal operands) at compile time.

DeadCodeElimination

Removes unused Emit statements and unreferenced local variables from function bodies.

FmaFusion

Fuses multiply-add patterns into fma math operations.

IrValidation

Validates IR structural invariants. Returns false (never modifies the module).

MemoryPlan

The result of memory planning: buffer assignments and peak usage.

PassManager

Runs passes in sequence with fixed-point iteration.

TensorId

A unique identifier for a tensor in the memory plan.

ValidationWarning

A validation warning describing a structural issue in the IR.

Enums

OptLevel

Optimization level.

Traits

Pass

An optimization pass that transforms an IR module.

Functions

collect_warnings

Collect all validation warnings for a module without logging.

optimize

Convenience function: runs O1 optimization passes on a module.