Struct Module

pub struct Module { /* private fields */ }

A compiled CUDA module, loaded into a context.

Implementations

impl Module

pub fn load_from_file(filename: &CStr) -> CudaResult<Module>

Load a module from the given file name into the current context.

The given file should be either a cubin file, a ptx file, or a fatbin file such as those produced by nvcc.

Example

use rustacuda::module::Module;
use std::ffi::CString;

let filename = CString::new("./resources/add.ptx")?;
let module = Module::load_from_file(&filename)?;

pub fn load_from_string(image: &CStr) -> CudaResult<Module>

Load a module from a CStr.

This is useful in combination with include_str!, to include the device code into the compiled executable.

The given CStr must contain the bytes of a cubin file, a ptx file or a fatbin file such as those produced by nvcc.

Example

use rustacuda::module::Module;
use std::ffi::CString;

let image = CString::new(include_str!("../resources/add.ptx"))?;
let module = Module::load_from_string(&image)?;

Examples found in repository

examples/launch.rs (line 15)

fn main() -> Result<(), Box<dyn Error>> {
    // Set up the context, load the module, and create a stream to run kernels in.
    rustacuda::init(CudaFlags::empty())?;
    let device = Device::get_device(0)?;
    let _ctx = Context::create_and_push(ContextFlags::MAP_HOST | ContextFlags::SCHED_AUTO, device)?;

    let ptx = CString::new(include_str!("../resources/add.ptx"))?;
    let module = Module::load_from_string(&ptx)?;
    let stream = Stream::new(StreamFlags::NON_BLOCKING, None)?;

    // Create buffers for data
    let mut in_x = DeviceBuffer::from_slice(&[1.0f32; 10])?;
    let mut in_y = DeviceBuffer::from_slice(&[2.0f32; 10])?;
    let mut out_1 = DeviceBuffer::from_slice(&[0.0f32; 10])?;
    let mut out_2 = DeviceBuffer::from_slice(&[0.0f32; 10])?;

    // This kernel adds each element in `in_x` and `in_y` and writes the result into `out`.
    unsafe {
        // Launch the kernel with one block of one thread, no dynamic shared memory on `stream`.
        let result = launch!(module.sum<<<1, 1, 0, stream>>>(
            in_x.as_device_ptr(),
            in_y.as_device_ptr(),
            out_1.as_device_ptr(),
            out_1.len()
        ));
        result?;

        // Launch the kernel again using the `function` form:
        let function_name = CString::new("sum")?;
        let sum = module.get_function(&function_name)?;
        // Launch with 1x1x1 (1) blocks of 10x1x1 (10) threads, to show that you can use tuples to
        // configure grid and block size.
        let result = launch!(sum<<<(1, 1, 1), (10, 1, 1), 0, stream>>>(
            in_x.as_device_ptr(),
            in_y.as_device_ptr(),
            out_2.as_device_ptr(),
            out_2.len()
        ));
        result?;
    }

    // Kernel launches are asynchronous, so we wait for the kernels to finish executing.
    stream.synchronize()?;

    // Copy the results back to host memory
    let mut out_host = [0.0f32; 20];
    out_1.copy_to(&mut out_host[0..10])?;
    out_2.copy_to(&mut out_host[10..20])?;

    for x in out_host.iter() {
        assert_eq!(3.0 as u32, *x as u32);
    }

    println!("Launched kernel successfully.");
    Ok(())
}

pub fn get_global<'a, T: DeviceCopy>( &'a self, name: &CStr, ) -> CudaResult<Symbol<'a, T>>

Get a reference to a global symbol, which can then be copied to/from.

Panics:

This function panics if the size of the symbol is not the same as the mem::sizeof<T>().

Examples

use rustacuda::module::Module;
use std::ffi::CString;

let ptx = CString::new(include_str!("../resources/add.ptx"))?;
let module = Module::load_from_string(&ptx)?;
let name = CString::new("my_constant")?;
let symbol = module.get_global::<u32>(&name)?;
let mut host_const = 0;
symbol.copy_to(&mut host_const)?;
assert_eq!(314, host_const);

pub fn get_function<'a>(&'a self, name: &CStr) -> CudaResult<Function<'a>>

Get a reference to a kernel function which can then be launched.

Examples

use rustacuda::module::Module;
use std::ffi::CString;

let ptx = CString::new(include_str!("../resources/add.ptx"))?;
let module = Module::load_from_string(&ptx)?;
let name = CString::new("sum")?;
let function = module.get_function(&name)?;

Examples found in repository

examples/launch.rs (line 37)

fn main() -> Result<(), Box<dyn Error>> {
    // Set up the context, load the module, and create a stream to run kernels in.
    rustacuda::init(CudaFlags::empty())?;
    let device = Device::get_device(0)?;
    let _ctx = Context::create_and_push(ContextFlags::MAP_HOST | ContextFlags::SCHED_AUTO, device)?;

    let ptx = CString::new(include_str!("../resources/add.ptx"))?;
    let module = Module::load_from_string(&ptx)?;
    let stream = Stream::new(StreamFlags::NON_BLOCKING, None)?;

    // Create buffers for data
    let mut in_x = DeviceBuffer::from_slice(&[1.0f32; 10])?;
    let mut in_y = DeviceBuffer::from_slice(&[2.0f32; 10])?;
    let mut out_1 = DeviceBuffer::from_slice(&[0.0f32; 10])?;
    let mut out_2 = DeviceBuffer::from_slice(&[0.0f32; 10])?;

    // This kernel adds each element in `in_x` and `in_y` and writes the result into `out`.
    unsafe {
        // Launch the kernel with one block of one thread, no dynamic shared memory on `stream`.
        let result = launch!(module.sum<<<1, 1, 0, stream>>>(
            in_x.as_device_ptr(),
            in_y.as_device_ptr(),
            out_1.as_device_ptr(),
            out_1.len()
        ));
        result?;

        // Launch the kernel again using the `function` form:
        let function_name = CString::new("sum")?;
        let sum = module.get_function(&function_name)?;
        // Launch with 1x1x1 (1) blocks of 10x1x1 (10) threads, to show that you can use tuples to
        // configure grid and block size.
        let result = launch!(sum<<<(1, 1, 1), (10, 1, 1), 0, stream>>>(
            in_x.as_device_ptr(),
            in_y.as_device_ptr(),
            out_2.as_device_ptr(),
            out_2.len()
        ));
        result?;
    }

    // Kernel launches are asynchronous, so we wait for the kernels to finish executing.
    stream.synchronize()?;

    // Copy the results back to host memory
    let mut out_host = [0.0f32; 20];
    out_1.copy_to(&mut out_host[0..10])?;
    out_2.copy_to(&mut out_host[10..20])?;

    for x in out_host.iter() {
        assert_eq!(3.0 as u32, *x as u32);
    }

    println!("Launched kernel successfully.");
    Ok(())
}

pub fn drop(module: Module) -> DropResult<Module>

Destroy a Module, returning an error.

Destroying a module can return errors from previous asynchronous work. This function destroys the given module and returns the error and the un-destroyed module on failure.

Example

use rustacuda::module::Module;
use std::ffi::CString;

let ptx = CString::new(include_str!("../resources/add.ptx"))?;
let module = Module::load_from_string(&ptx)?;
match Module::drop(module) {
    Ok(()) => println!("Successfully destroyed"),
    Err((e, module)) => {
        println!("Failed to destroy module: {:?}", e);
        // Do something with module
    },
}

Trait Implementations

impl Debug for Module

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Drop for Module

fn drop(&mut self)

Executes the destructor for this type.