Crate trueno_gpu 

trueno-gpu: Pure Rust PTX Generation for NVIDIA CUDA

Generate PTX assembly directly from Rust - no LLVM, no nvcc, no external dependencies.

Philosophy

Own the Stack - Build everything from first principles for complete control, auditability, and reproducibility.

Quick Start

use trueno_gpu::ptx::{PtxModule, PtxKernel, PtxType};

// Build a vector addition kernel
let module = PtxModule::new()
    .version(8, 0)
    .target("sm_70")
    .address_size(64);

let ptx_source = module.emit();
assert!(ptx_source.contains(".version 8.0"));

Modules

• [ptx] - PTX code generation (builder pattern)
• [driver] - CUDA driver API (minimal FFI, optional)
• [kernels] - Hand-optimized GPU kernels
• [memory] - GPU memory management
• backend - Multi-backend abstraction

Re-exports

pub use error::GpuError;

pub use error::Result;

pub use monitor::cuda_device_count;

pub use monitor::cuda_monitoring_available;

pub use monitor::CudaDeviceInfo;

pub use monitor::CudaMemoryInfo;

Modules

backend

Multi-Backend Abstraction

error

Error types for trueno-gpu operations Error types for trueno-gpu operations

graph

PMAT-291: Tensor compute graph for GPU inference (reduces 430 dispatches to ~15) PMAT-291: Tensor Compute Graph for GPU Inference

monitor

TUI Compute Mode Flow: CPU, GPU, and Memory Monitoring

testing

E2E visual testing framework for GPU kernels E2E Visual Testing Framework for GPU Kernels