Module kernel_tuning 

Automatic kernel tuning for hardware adaptation

This module provides automatic performance tuning for kernel operations across different hardware backends. It profiles kernel execution times and adaptively selects optimal parameters (block sizes, thread counts, memory layouts) for the specific hardware being used.

Features

• Auto-tuning: Automatic parameter selection through benchmarking
• Hardware Detection: Platform capability detection and profiling
• Caching: Persistent tuning results for faster subsequent runs
• Multi-Backend: Support for CUDA, ROCm, Metal, CPU, and more
• Adaptive: Dynamic adjustment based on tensor sizes and operations

Examples

use trustformers_core::kernel_tuning::{KernelTuner, TuningConfig, Operation};

// Create tuner with default configuration
let mut tuner = KernelTuner::new(TuningConfig::default())?;

// Auto-tune matrix multiplication parameters for 1024x768 * 768x512
let params = tuner.tune_matmul(1024, 512, 768)?;
println!("Optimal block size: {:?}", params.block_size);

Structs

KernelParams

Tuned kernel parameters

KernelTuner

Automatic kernel tuner

PlatformInfo

Platform characteristics for tuning decisions

TuningConfig

Tuning configuration

TuningStatistics

Statistics about tuning results

Enums

Backend

Hardware backend types

Operation

Kernel operation types for tuning

Functions

get_kernel_tuner

Get or initialize the global kernel tuner