singe-cuda 0.1.0-alpha.3

#[allow(unused_imports)]
use crate::error::ErrorCode;

use std::{ffi::CString, ptr, sync::Arc};

use singe_cuda_sys::driver;

use crate::{
    context::Context,
    error::{Error, Result},
    graph::{ExecutableGraph, Graph, GraphNode},
    kernel::{self, LibraryKernelHandle},
    module::{KernelFunction, KernelParameters, LaunchConfig, Module},
    try_cuda,
    types::{DeviceFunction, FunctionAttribute, FunctionCache},
};

#[derive(Debug)]
pub struct Library {
    handle: driver::CUlibrary,
    ctx: Arc<Context>,
}

#[derive(Debug, Clone, Copy)]
pub struct LibraryGlobal<'a> {
    ptr: *mut (),
    size: usize,
    _library: &'a Library,
}

#[derive(Debug, Clone, Copy)]
pub struct LibraryKernel<'a> {
    handle: driver::CUkernel,
    library: &'a Library,
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct KernelParamInfo {
    pub offset: usize,
    pub size: usize,
}

impl Library {
    pub const unsafe fn from_raw(handle: driver::CUlibrary, ctx: Arc<Context>) -> Self {
        Self { handle, ctx }
    }

    /// Returns the handle of the kernel with the given name located in this library.
    /// If kernel handle is not found, the call returns [`ErrorCode::NotFound`].
    pub fn kernel(&self, name: &str) -> Result<LibraryKernel<'_>> {
        let c_name = CString::new(name)?;
        let mut handle = ptr::null_mut();
        self.ctx.bind()?;
        unsafe {
            try_cuda!(driver::cuLibraryGetKernel(
                &raw mut handle,
                self.handle,
                c_name.as_ptr(),
            ))?;
        }
        if handle.is_null() {
            return Err(Error::NullHandle);
        }
        Ok(LibraryKernel {
            handle,
            library: self,
        })
    }

    /// Returns the number of kernels in this library.
    pub fn kernel_count(&self) -> Result<usize> {
        let mut count = 0;
        self.ctx.bind()?;
        unsafe {
            try_cuda!(driver::cuLibraryGetKernelCount(&raw mut count, self.handle))?;
        }
        Ok(count as usize)
    }

    /// Returns the module handle associated with the current context located in this library.
    /// If module handle is not found, the call returns [`ErrorCode::NotFound`].
    pub fn module(&self) -> Result<Module> {
        let mut handle = ptr::null_mut();
        self.ctx.bind()?;
        unsafe {
            try_cuda!(driver::cuLibraryGetModule(&raw mut handle, self.handle))?;
        }
        if handle.is_null() {
            return Err(Error::NullHandle);
        }
        Ok(unsafe { Module::from_borrowed_raw(handle, Arc::clone(&self.ctx)) })
    }

    /// Returns the base pointer and size of the global with the given name for the requested library and the current context.
    /// If no global for the requested name exists, the call returns [`ErrorCode::NotFound`].
    pub fn global(&self, name: &str) -> Result<LibraryGlobal<'_>> {
        let c_name = CString::new(name)?;
        let mut ptr = 0;
        let mut size = 0;
        self.ctx.bind()?;
        unsafe {
            try_cuda!(driver::cuLibraryGetGlobal(
                &raw mut ptr,
                &raw mut size,
                self.handle,
                c_name.as_ptr(),
            ))?;
        }
        Ok(LibraryGlobal {
            ptr: ptr as *mut (),
            size: size as usize,
            _library: self,
        })
    }

    /// Returns the base pointer and size of the managed memory with the given name for the requested library.
    /// If no managed memory with the requested name exists, the call returns [`ErrorCode::NotFound`].
    /// Note that managed memory for library library is shared across devices and is registered when the library is loaded into atleast one context.
    pub fn managed(&self, name: &str) -> Result<LibraryGlobal<'_>> {
        let c_name = CString::new(name)?;
        let mut ptr = 0;
        let mut size = 0;
        self.ctx.bind()?;
        unsafe {
            try_cuda!(driver::cuLibraryGetManaged(
                &raw mut ptr,
                &raw mut size,
                self.handle,
                c_name.as_ptr(),
            ))?;
        }
        Ok(LibraryGlobal {
            ptr: ptr as *mut (),
            size: size as usize,
            _library: self,
        })
    }

    /// Returns the function pointer to a unified function denoted by symbol.
    /// If no unified function with name symbol exists, the call returns [`ErrorCode::NotFound`].
    /// If no device in the system supports unified function pointers, the call may return [`ErrorCode::NotFound`].
    pub fn unified_function(&self, symbol: &str) -> Result<*mut ()> {
        let c_symbol = CString::new(symbol)?;
        let mut ptr = ptr::null_mut();
        self.ctx.bind()?;
        unsafe {
            try_cuda!(driver::cuLibraryGetUnifiedFunction(
                &raw mut ptr,
                self.handle,
                c_symbol.as_ptr(),
            ))?;
        }
        if ptr.is_null() {
            return Err(Error::NullHandle);
        }
        Ok(ptr.cast())
    }

    pub const unsafe fn as_raw(&self) -> driver::CUlibrary {
        self.handle
    }
}

impl Drop for Library {
    fn drop(&mut self) {
        if let Err(err) = self.ctx.bind() {
            #[cfg(debug_assertions)]
            eprintln!("failed to bind context before unloading library: {err}");
            return;
        }

        unsafe {
            if let Err(err) = try_cuda!(driver::cuLibraryUnload(self.handle)) {
                #[cfg(debug_assertions)]
                eprintln!("failed to unload cuda library: {err}");
            }
        }
    }
}

impl LibraryGlobal<'_> {
    pub const fn as_ptr(&self) -> *mut () {
        self.ptr
    }

    pub const fn size(&self) -> usize {
        self.size
    }
}

impl LibraryKernel<'_> {
    pub fn name(&self) -> Result<String> {
        kernel::name::<LibraryKernelHandle>(self.library.ctx.as_ref(), self.handle)
    }

    /// Returns the handle of the function for this kernel and the current context.
    /// If function handle is not found, the call returns [`ErrorCode::NotFound`].
    pub fn function(&self) -> Result<DeviceFunction> {
        self.library.ctx.bind()?;
        let mut handle = ptr::null_mut();
        unsafe {
            try_cuda!(driver::cuKernelGetFunction(&raw mut handle, self.handle))?;
        }
        if handle.is_null() {
            return Err(Error::NullHandle);
        }
        Ok(handle.into())
    }

    pub fn add_to_graph(
        &self,
        graph: &mut Graph,
        dependencies: &[GraphNode],
        config: &LaunchConfig,
        params: &mut KernelParameters,
    ) -> Result<GraphNode> {
        let function = self.function()?;
        let module = self.library.module()?;
        let function = unsafe { KernelFunction::from_raw(function, &module) };
        function.add_to_graph(graph, dependencies, config, params)
    }

    pub fn set_graph_node_params(
        &self,
        executable: &mut ExecutableGraph,
        node: GraphNode,
        config: &LaunchConfig,
        params: &mut KernelParameters,
    ) -> Result<()> {
        let function = self.function()?;
        let module = self.library.module()?;
        let function = unsafe { KernelFunction::from_raw(function, &module) };
        function.set_graph_node_params(executable, node, config, params)
    }

    pub fn attribute(&self, attribute: FunctionAttribute) -> Result<i32> {
        kernel::attribute::<LibraryKernelHandle>(self.library.ctx.as_ref(), self.handle, attribute)
    }

    pub fn set_attribute(&self, attribute: FunctionAttribute, value: i32) -> Result<()> {
        kernel::set_attribute::<LibraryKernelHandle>(
            self.library.ctx.as_ref(),
            self.handle,
            attribute,
            value,
        )
    }

    /// On devices where the L1 cache and shared memory use the same hardware resources, this sets through config the preferred cache configuration for the device kernel kernel on the requested device dev.
    /// This is only a preference.
    /// The driver will use the requested configuration if possible, but it is free to choose a different configuration if required to execute kernel.
    /// Any context-wide preference set via [`sys::cuCtxSetCacheConfig`](singe_cuda_sys::driver::cuCtxSetCacheConfig) will be overridden by this per-kernel setting.
    ///
    /// Note that attributes set using [`sys::cuFuncSetCacheConfig`](singe_cuda_sys::driver::cuFuncSetCacheConfig) will override the attribute set by this API irrespective of whether the call to [`sys::cuFuncSetCacheConfig`](singe_cuda_sys::driver::cuFuncSetCacheConfig) is made before or after this API call.
    ///
    /// This setting does nothing on devices where the size of the L1 cache and shared memory are fixed.
    ///
    /// Launching a kernel with a different preference than the most recent preference setting may insert a device-side synchronization point.
    ///
    /// The supported cache configurations are:
    ///
    /// * [`FunctionCache::PreferNone`]: no preference for shared memory or L1 (default)
    /// * [`FunctionCache::PreferShared`]: prefer larger shared memory and smaller L1 cache
    /// * [`FunctionCache::PreferL1`]: prefer larger L1 cache and smaller shared memory
    /// * [`FunctionCache::PreferEqual`]: prefer equal sized L1 cache and shared memory
    ///
    /// Note:
    ///
    /// The API has stricter locking requirements in comparison to its legacy counterpart [`sys::cuFuncSetCacheConfig`](singe_cuda_sys::driver::cuFuncSetCacheConfig) due to device-wide semantics.
    /// If multiple threads are trying to set a config on the same device simultaneously, the cache config setting will depend on the interleavings chosen by the OS scheduler and memory consistency.
    pub fn set_cache_config(&self, config: FunctionCache) -> Result<()> {
        self.library.ctx.bind()?;
        unsafe {
            try_cuda!(driver::cuKernelSetCacheConfig(
                self.handle,
                config.into(),
                self.library.ctx.device().id() as _,
            ))?;
        }
        Ok(())
    }

    /// Queries the kernel parameter at the given index, returning the offset and size where the parameter will reside in the device-side parameter layout.
    /// This information can be used to update kernel node parameters from the device. The index must be less than the number of parameters that the kernel takes.
    ///
    /// Note:
    ///
    /// Note that this function may also return error codes from previous, asynchronous launches.
    pub fn param_info(&self, index: usize) -> Result<KernelParamInfo> {
        self.library.ctx.bind()?;
        let mut offset = 0;
        let mut size = 0;
        unsafe {
            try_cuda!(driver::cuKernelGetParamInfo(
                self.handle,
                index as _,
                &raw mut offset,
                &raw mut size,
            ))?;
        }
        Ok(KernelParamInfo {
            offset: offset as usize,
            size: size as usize,
        })
    }

    pub const unsafe fn as_raw(&self) -> driver::CUkernel {
        self.handle
    }
}