A high-performance inference engine for AI models
use std::{os::raw::c_void, ptr::NonNull};
use crate::{
backends::{
common::{Backend, CpuBuffer, GpuBuffer},
cpu::Cpu,
},
utils::downcast::downcast_ref,
};
pub mod dense;
pub mod sparse;
pub(super) trait CpuBufferExt: GpuBuffer<Backend = Cpu> {
fn cpu_address(&self) -> NonNull<c_void> {
if let Some(buffer) = downcast_ref::<<Cpu as Backend>::GlobalGpuBuffer>(self) {
buffer.cpu_ptr()
} else if let Some(buffer) = downcast_ref::<<Cpu as Backend>::ConstantBuffer>(self) {
buffer.cpu_ptr()
} else if let Some(buffer) = downcast_ref::<<Cpu as Backend>::ScratchBuffer>(self) {
buffer.cpu_ptr()
} else {
unreachable!("Unsupported Cpu buffer type")
}
}
}
impl<T: GpuBuffer<Backend = Cpu> + ?Sized> CpuBufferExt for T {}