Multi-Target High-Performance Compute Library

---

trueno (Spanish: "thunder") provides unified compute primitives across CPU SIMD, GPU, and WebAssembly.

Features

• CPU SIMD: x86 (SSE2/AVX/AVX2/AVX-512), ARM (NEON), WASM (SIMD128)
• GPU: Pure Rust PTX generation via trueno-gpu (no nvcc required)
• Cross-platform GPU: Vulkan/Metal/DX12/WebGPU via wgpu
• Auto-dispatch: Runtime selection of optimal backend
• Zero unsafe in public API: Safety via type system

Installation

[dependencies]
trueno = "0.11"

# Optional: GPU support for large matrices
trueno = { version = "0.11", features = ["gpu"] }

# Optional: Pure Rust CUDA PTX generation
trueno-gpu = "0.4"

Quick Start

use trueno::{Vector, Matrix, SymmetricEigen};

// Vector operations - auto-selects best SIMD backend
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]);

let sum = a.add(&b).unwrap();           // [6.0, 8.0, 10.0, 12.0]
let dot = a.dot(&b).unwrap();           // 70.0
let activated = a.relu().unwrap();      // ReLU activation

// Matrix operations
let m = Matrix::from_vec(2, 2, vec![1.0, 2.0, 3.0, 4.0]).unwrap();
let product = m.matmul(&m).unwrap();    // Matrix multiplication
let transposed = m.transpose();          // Transpose

// Batched matmul for transformers (Q @ K^T pattern)
let batch = 2; let heads = 4; let seq = 8; let dim = 64;
let q: Vec<f32> = vec![0.1; batch * heads * seq * dim];
let kt: Vec<f32> = vec![0.1; batch * heads * dim * seq];
let attn = Matrix::batched_matmul_4d(&q, &kt, batch, heads, seq, dim, seq).unwrap();

// Eigendecomposition (PCA, spectral analysis)
let cov = Matrix::from_vec(2, 2, vec![3.0, 1.0, 1.0, 3.0]).unwrap();
let eigen = SymmetricEigen::new(&cov).unwrap();
let eigenvalues = eigen.eigenvalues();  // [4.0, 2.0]

Performance

Operation	SIMD Speedup	Notes
Dot product	6-17x	AVX-512 for compute-bound
Matrix multiply	2-10x	GPU for 500x500+
Reductions (sum, max, min)	3-12x	AVX-512 optimal
Element-wise (add, mul)	1-2x	Memory-bound
Convolution 2D	5-8x	AVX2/AVX-512 optimized

Benchmark Results (AMD Ryzen 9 7950X)

Benchmark	Throughput
Vector recip (AVX-512, 10K)	10.0 Gelem/s
Vector recip (AVX2, 10K)	9.7 Gelem/s
PTX module emit	3.1 µs
PTX kernel build	81 ns
Launch config	1.7 ns

GPU Note: GPU acceleration benefits matrix multiply only. Element-wise operations use CPU SIMD (GPU transfer overhead exceeds compute time).

trueno-gpu: Pure Rust CUDA

Generate CUDA PTX kernels without nvcc, LLVM, or external toolchains:

use trueno_gpu::kernels::{GemmKernel, Kernel, SoftmaxKernel};

// Generate optimized GEMM kernel
let gemm = GemmKernel::tensor_core(1024, 1024, 1024);
let ptx = gemm.emit_ptx();  // Pure Rust PTX generation

// Generate softmax with warp shuffle reduction
let softmax = SoftmaxKernel::new(4096);
let ptx = softmax.emit_ptx();

// Available kernels: GEMM, Softmax, LayerNorm, Attention, Quantize (Q4K/Q5K/Q6K)

Operations

Vector: add, sub, mul, div, dot, sum, min, max, argmin, argmax, norm_l1, norm_l2, normalize, recip, sqrt, abs, clamp

Activations: relu, leaky_relu, elu, sigmoid, tanh, gelu, swish, softmax, log_softmax, silu

Matrix: matmul, batched_matmul, batched_matmul_4d, transpose, matvec, convolve2d, pooling (max/avg), topk, gather, pad

Statistics: mean, variance, stddev, covariance, correlation, zscore

Eigen: symmetric eigendecomposition (Jacobi algorithm)

GPU Kernels: GEMM (naive/tiled/tensor core), Softmax, LayerNorm, RMSNorm, Attention, GEMV, Quantization

Development

cargo test                  # Run tests
cargo bench                 # Run benchmarks
make coverage              # Coverage report (requires cargo-llvm-cov)
cargo run --example backend_detection  # Check available backends

Ecosystem

Part of the Pragmatic AI Labs stack:

• trueno-gpu - Pure Rust PTX generation (no nvcc)
• trueno-db - GPU-first analytics database
• trueno-graph - Graph algorithms
• trueno-rag - RAG pipeline

License

MIT - see LICENSE