easy-async-cl3

A high-level, async-first Rust wrapper for OpenCL with intelligent multi-device management and declarative task execution.

Overview

easy-async-cl3 provides a modern, ergonomic interface to OpenCL that embraces Rust's async/await paradigm. The library automatically manages resources, distributes work across multiple devices, and provides compile-time safety guarantees.

Key Features

• Async/Await Integration: All GPU operations return futures for seamless async workflows
• Automatic Multi-Device Support: Intelligent work distribution across multiple GPUs based on device capabilities
• Type-Safe API: Compile-time guarantees prevent common errors (e.g., using unbuilt programs)
• Declarative Task Building: Fluent builder pattern for constructing GPU tasks
• Zero-Cost Abstractions: RAII-based resource management with no runtime overhead
• Comprehensive OpenCL Support: Full support for OpenCL 1.1 through 3.0 features including Pipes, SVM, and Images
• Built-in Profiling: Optional performance measurement with negligible overhead

Installation

Add this to your Cargo.toml:

[dependencies]

easy-async-cl3 = "0.1"

tokio = { version = "1", features = ["macros", "rt-multi-thread"] }

Quick Start

use easy_async_cl3::{
    async_executor::AsyncExecutor,
    cl_types::memory_flags::MemoryFlags,
    error::ClError,
};

#[tokio::main]
async fn main() -> Result<(), ClError> {
    // Initialize executor with best available platform
    let executor = AsyncExecutor::new_best_platform()?;
    
    // Define and build kernel
    let source = r#"
        kernel void vector_add(global float* a, global const float* b) {
            size_t i = get_global_id(0);
            a[i] += b[i];
        }
    "#;
    let program = executor.build_program(source.to_string(), None)?;
    let kernel = executor.create_kernel(&program, "vector_add")?;
    
    // Prepare data
    let size = 1_000_000;
    let mut a = vec![1.0f32; size];
    let b = vec![2.0f32; size];
    
    // Create GPU buffers
    let buf_a = executor.create_buffer(
        &[MemoryFlags::ReadWrite, MemoryFlags::CopyHostPtr],
        size * std::mem::size_of::<f32>(),
        a.as_mut_ptr() as *mut _
    )?;
    let buf_b = executor.create_buffer(
        &[MemoryFlags::ReadOnly, MemoryFlags::CopyHostPtr],
        size * std::mem::size_of::<f32>(),
        b.as_ptr() as *mut _
    )?;
    
    // Execute task
    executor.create_task(kernel)
        .arg_buffer(0, &buf_a)
        .arg_buffer(1, &buf_b)
        .global_work_dims(size, 1, 1)
        .read_buffer(&buf_a, &mut a)
        .run()
        .await?;
    
    assert_eq!(a[0], 3.0);
    Ok(())
}

Advanced Features

Multi-Device Execution

The library automatically detects and utilizes all available compute devices:

let executor = AsyncExecutor::new_best_platform_with_options(true)?; // Enable profiling

let report = executor.create_task(kernel)
    .arg_buffer(0, &buffer)
    .global_work_dims(10_000_000, 1, 1)
    .run()
    .await?;

println!("Execution time: {} μs", report.total_kernel_duration_ns() / 1000);

Shared Virtual Memory (OpenCL 2.0+)

Zero-copy memory sharing between CPU and GPU:

let mut svm_buffer = executor.create_svm_buffer::<f32>(
    &[MemoryFlags::ReadWrite], 
    1024
)?;

executor.create_task(kernel)
    .arg_svm(0, &svm_buffer)
    .global_work_dims(1024, 1, 1)
    .run()
    .await?;

// Direct CPU access without explicit copy
let queue = &executor.get_queues()[0];
let mapped = svm_buffer.map_mut(queue, &vec![MemoryFlags::ReadWrite])?;
println!("Result: {}", mapped[0]);

Image Processing

Native support for OpenCL images with hardware-accelerated filtering:

use easy_async_cl3::cl_types::cl_image::{
    ClImageFormats, ClImageDesc, image_type::ClImageType
};

let format = ClImageFormats::rgba_unorm_int8();
let desc = ClImageDesc {
    image_type: ClImageType::Image2D,
    image_width: Some(1920),
    image_height: Some(1080),
    ..Default::default()
};

let image = executor.create_image(
    &[MemoryFlags::ReadWrite],
    &format,
    &desc,
    std::ptr::null_mut()
)?;

executor.create_task(kernel)
    .arg_image(0, &image)
    .global_work_dims(1920, 1080, 1)
    .run()
    .await?;

Pipes for Inter-Kernel Communication (OpenCL 2.0+)

Stream data between kernels without CPU involvement:

use easy_async_cl3::cl_types::cl_pipe::ClPipe;

let pipe = ClPipe::new(
    executor.get_context().as_ref(),
    &[MemoryFlags::ReadWrite],
    4,    // packet size (bytes)
    1024  // max packets
)?;

// Producer writes to pipe
executor.create_task(producer_kernel)
    .arg_pipe(0, &pipe)
    .global_work_dims(1024, 1, 1)
    .run()
    .await?;

// Consumer reads from pipe
executor.create_task(consumer_kernel)
    .arg_pipe(0, &pipe)
    .global_work_dims(1024, 1, 1)
    .run()
    .await?;

Architecture

The library is structured in three main layers:

1. AsyncExecutor: High-level interface managing platforms, devices, and command queues
2. TaskBuilder: Declarative API for constructing and executing GPU tasks
3. CL Types: Type-safe wrappers around OpenCL objects (buffers, images, kernels, etc.)

Work is automatically distributed across available devices based on their compute capabilities and memory capacity.

Documentation

• API Documentation - Complete API reference
• Examples - Comprehensive usage examples
• OpenCL Specification - Kernel programming reference

Requirements

• Rust 1.70 or later
• OpenCL runtime (provided by GPU vendor drivers)
• Tokio async runtime

Use Cases

• High-performance scientific computing
• Real-time image and video processing
• Machine learning inference and training
• Cryptographic operations
• Financial modeling and simulations
• Parallel data analytics

Contributing

Contributions are welcome. Please ensure all tests pass and follow the existing code style.

License

Licensed under either of:

• Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.

Acknowledgments

Built on the cl3 library.