Struct KernelParamStorage 

pub struct KernelParamStorage { /* private fields */ }

Storage backing the void** array cuLaunchKernel expects.

CUDA’s cuLaunchKernel takes a *mut *mut c_void whose elements each point to the value bytes of a single argument. The lifetimes are awkward: the value bytes must outlive the launch call, and so must the pointer array. KernelParamStorage owns both so the launch site can take a *mut *mut c_void into it and pass that to cuLaunchKernel without leaking.

Currently only used on CUDA builds; the no-CUDA host path records the parsed Vec<LoweredArg> directly. The struct is still defined non-conditionally so the type checker can prove the no-CUDA build keeps the lowering free of CUDA-specific types.

Layout (post v0.3.4)

A single contiguous Vec<u8> (backing) holds every scalar value in declaration order, aligned per its native type. slots is the pointer table CUDA reads through — each *mut c_void points to the start of one slot inside backing. A 16-arg kernel costs one Vec<u8> allocation plus one Vec<*mut c_void> allocation rather than 16 small Box<[u8]> allocations.

The single-buffer layout is sound because cuLaunchKernel reads each parameter through kernelParams[i] independently; the driver has no per-slot alignment requirement that depends on slots being distinct allocations. We still align each slot to std::mem::align_of::<T>() so a misaligned f64 cannot read garbage on architectures that fault on unaligned loads.

Implementations

impl KernelParamStorage

pub fn as_ptr(&mut self) -> *mut *mut c_void

Borrow the pointer slots as a *mut *mut c_void suitable for cuLaunchKernel.

pub fn len(&self) -> usize

Number of arguments in this storage.

pub fn is_empty(&self) -> bool

True when the storage carries no arguments.

Trait Implementations

impl Send for KernelParamStorage