A high-performance inference engine for AI models
use std::{
alloc::{Layout, alloc, dealloc, handle_alloc_error},
os::raw::c_void,
ptr::NonNull,
};
use crate::backends::{
common::{Buffer, CpuBuffer, GlobalBufferCpuGpu, GlobalGpuBuffer, GpuBuffer},
cpu::{Cpu, error::CpuError},
};
#[derive(Debug)]
pub struct CpuDenseBuffer {
ptr: NonNull<c_void>,
layout: Layout,
}
impl CpuDenseBuffer {
pub(in crate::backends::cpu) fn new(size: usize) -> Self {
assert!(size > 0, "CPU buffer size must be nonzero");
let layout = Layout::from_size_align(size, 64).expect("invalid CPU buffer layout");
let ptr = NonNull::new(unsafe { alloc(layout) }).unwrap_or_else(|| handle_alloc_error(layout)).cast();
Self {
ptr,
layout,
}
}
}
// The allocation has a stable address; callers synchronize access through buffer borrows and command buffers.
unsafe impl Send for CpuDenseBuffer {}
unsafe impl Sync for CpuDenseBuffer {}
impl Drop for CpuDenseBuffer {
fn drop(&mut self) {
unsafe { dealloc(self.ptr.as_ptr().cast(), self.layout) };
}
}
impl Buffer for CpuDenseBuffer {
type Backend = Cpu;
fn size(&self) -> usize {
self.layout.size()
}
}
impl GpuBuffer for CpuDenseBuffer {}
impl CpuBuffer for CpuDenseBuffer {
fn cpu_ptr(&self) -> NonNull<c_void> {
self.ptr
}
}
impl GlobalGpuBuffer for CpuDenseBuffer {}
impl GlobalBufferCpuGpu for CpuDenseBuffer {
fn handoff(self) -> Result<Self, CpuError> {
Ok(self)
}
}