Crate trueno 

Trueno: Multi-Target High-Performance Compute Library

Trueno (Spanish: “thunder”) provides unified, high-performance compute primitives across three execution targets:

1. CPU SIMD - x86 (SSE2/AVX/AVX2/AVX-512), ARM (NEON), WASM (SIMD128)
2. GPU - Vulkan/Metal/DX12/WebGPU via wgpu
3. WebAssembly - Portable SIMD128 for browser/edge deployment

Design Principles

• Write once, optimize everywhere: Single algorithm, multiple backends
• Runtime dispatch: Auto-select best implementation based on CPU features
• Zero unsafe in public API: Safety via type system, unsafe isolated in backends
• Benchmarked performance: Every optimization must prove ≥10% speedup
• Extreme TDD: >90% test coverage, mutation testing, property-based tests

Quick Start

use trueno::Vector;

let a = Vector::from_slice(&[1.0, 2.0, 3.0, 4.0]);
let b = Vector::from_slice(&[5.0, 6.0, 7.0, 8.0]);

// Auto-selects best backend (AVX2/GPU/WASM)
let result = a.add(&b).unwrap();
assert_eq!(result.as_slice(), &[6.0, 8.0, 10.0, 12.0]);

Re-exports

pub use eigen::SymmetricEigen;

pub use error::Result;

pub use error::TruenoError;

pub use hash::hash_bytes;

pub use hash::hash_key;

pub use hash::hash_keys_batch;

pub use hash::hash_keys_batch_with_backend;

pub use matrix::Matrix;

pub use monitor::cuda_monitor_available;

pub use monitor::GpuBackend;

pub use monitor::GpuClockMetrics;

pub use monitor::GpuDeviceInfo;

pub use monitor::GpuMemoryMetrics;

pub use monitor::GpuMetrics;

pub use monitor::GpuMonitor;

pub use monitor::GpuPcieMetrics;

pub use monitor::GpuPowerMetrics;

pub use monitor::GpuThermalMetrics;

pub use monitor::GpuUtilization;

pub use monitor::GpuVendor;

pub use monitor::MonitorConfig;

pub use monitor::MonitorError;

pub use vector::Vector;

pub use brick::fnv1a_f32_checksum;

pub use brick::AddOp;

pub use brick::AssertionResult;

pub use brick::AttentionOp;

pub use brick::BlockQ5K;

pub use brick::BlockQ6K;

pub use brick::BrickBottleneck;

pub use brick::BrickCategory;

pub use brick::BrickError;

pub use brick::BrickId;

pub use brick::BrickIdTimer;

pub use brick::BrickLayer;

pub use brick::BrickProfiler;

pub use brick::BrickSample;

pub use brick::BrickStats;

pub use brick::BrickTimer;

pub use brick::BrickVerification;

pub use brick::ByteBudget;

pub use brick::CategoryStats;

pub use brick::ComputeAssertion;

pub use brick::ComputeBackend;

pub use brick::ComputeBrick;

pub use brick::ComputeOp;

pub use brick::DivergenceInfo;

pub use brick::DotOp;

pub use brick::DotQ5KOp;

pub use brick::DotQ6KOp;

pub use brick::EdgeType;

pub use brick::ExecutionEdge;

pub use brick::ExecutionGraph;

pub use brick::ExecutionNode;

pub use brick::ExecutionNodeId;

pub use brick::FusedGateUpOp;

pub use brick::FusedGateUpWeights;

pub use brick::FusedQKVOp;

pub use brick::FusedQKVWeights;

pub use brick::KernelChecksum;

pub use brick::MatmulOp;

pub use brick::PtxRegistry;

pub use brick::SoftmaxOp;

pub use brick::SyncMode;

pub use brick::TileLevel;

pub use brick::TileStats;

pub use brick::TileTimer;

pub use brick::TokenBudget;

pub use brick::TokenResult;

pub use hardware::default_hardware_path;

pub use hardware::Bottleneck;

pub use hardware::CpuCapability;

pub use hardware::GpuBackend as HardwareGpuBackend;

pub use hardware::GpuCapability;

pub use hardware::HardwareCapability;

pub use hardware::RooflineParams;

pub use hardware::SimdWidth;

pub use tuner::BottleneckClass;

pub use tuner::BottleneckPrediction;

pub use tuner::BrickTuner;

pub use tuner::ConceptDriftStatus;

pub use tuner::ExperimentSuggestion;

pub use tuner::FeatureExtractor;

pub use tuner::KernelClassifier;

pub use tuner::KernelRecommendation;

pub use tuner::KernelType;

pub use tuner::QuantType;

pub use tuner::RunConfig;

pub use tuner::ThroughputPrediction;

pub use tuner::ThroughputRegressor;

pub use tuner::TrainingSample;

pub use tuner::TrainingStats;

pub use tuner::TunerDataCollector;

pub use tuner::TunerError;

pub use tuner::TunerFeatures;

pub use tuner::TunerRecommendation;

pub use tuner::UserFeedback;

pub use tiling::optimal_prefetch_distance;

pub use tiling::pack_a_index;

pub use tiling::pack_b_index;

pub use tiling::swizzle_index;

pub use tiling::PackingLayout;

pub use tiling::PrefetchLocality;

pub use tiling::TcbGeometry;

pub use tiling::TcbIndexCalculator;

pub use tiling::TcbLevel;

pub use tiling::TiledQ4KMatvec;

pub use tiling::TilingBackend;

pub use tiling::TilingConfig;

pub use tiling::TilingError;

pub use tiling::TilingStats;

pub use tiling::Q4K_SUPERBLOCK_BYTES;

pub use tiling::Q4K_SUPERBLOCK_SIZE;

Modules

backends

Backend implementations for different SIMD instruction sets

blis

BLIS-Style Matrix Multiplication

brick

ComputeBrick: Token-Centric Compute Units

chaos

Chaos Engineering Configuration

eigen

Eigendecomposition for symmetric matrices

error

Error types for Trueno operations

hardware

Hardware Capability Detection (PMAT-447)

hash

SIMD-optimized hash functions for key-value store operations.

matrix

Matrix operations for Trueno

monitor

GPU Monitoring, Tracing, and Visualization (TRUENO-SPEC-010)

simulation

Simulation Testing Framework (TRUENO-SPEC-012)

tiling

Tiling Compute Blocks (TCB) - Work Partitioning for High-Performance Kernels

tuner

ML-Based ComputeBrick Tuner

vector

Vector type with multi-backend support

Macros

dispatch_binary_op

Macro to dispatch binary operations to appropriate backend

dispatch_reduction

Macro to dispatch reduction operations (return f32)

dispatch_unary_op

Macro to dispatch unary operations (a -> result)

time_brick

Macro for convenient brick timing with automatic sync.

Enums

Backend

Backend execution target

OpComplexity

Operation complexity for GPU dispatch eligibility

OperationType

Operation type for SIMD backend selection

Functions

select_backend_for_operation

Select the optimal backend for a specific operation type

select_best_available_backend

Select the best available backend for the current platform