A high-performance inference engine for AI models
use std::{mem::ManuallyDrop, os::raw::c_void, ptr::NonNull};
use crate::{
backends::{
common::{Buffer, CpuBuffer, GlobalBufferCpuGpu, GlobalGpuBuffer, GpuBuffer, allocator::bump::BumpAllocation},
webgpu::{WebGPU, context::WebGPUContext, error::WebGPUError},
},
utils::pointers::SendPtrMut,
};
#[derive(Debug)]
pub struct WebGPUDenseBufferGpu {
buffer: wgpu::Buffer,
size: usize,
}
impl WebGPUDenseBufferGpu {
pub(in crate::backends::webgpu) fn new(
context: &WebGPUContext,
size: usize,
) -> Result<Self, WebGPUError> {
let buffer = context.device.create_buffer(&wgpu::BufferDescriptor {
label: None,
size: size.next_multiple_of(wgpu::COPY_BUFFER_ALIGNMENT as usize) as u64,
usage: wgpu::BufferUsages::UNIFORM
| wgpu::BufferUsages::STORAGE
| wgpu::BufferUsages::INDIRECT
| wgpu::BufferUsages::COPY_SRC
| wgpu::BufferUsages::COPY_DST,
mapped_at_creation: false,
});
Ok(Self {
buffer,
size,
})
}
pub(in crate::backends::webgpu) fn wgpu_buffer(&self) -> &wgpu::Buffer {
&self.buffer
}
}
impl Buffer for WebGPUDenseBufferGpu {
type Backend = WebGPU;
fn size(&self) -> usize {
self.size
}
}
impl GpuBuffer for WebGPUDenseBufferGpu {}
impl GlobalGpuBuffer for WebGPUDenseBufferGpu {}
#[derive(Debug)]
pub struct WebGPUDenseBufferCpuGpu {
cpu_ptr: SendPtrMut<c_void>,
view: ManuallyDrop<wgpu::BufferViewMut>,
buffer: wgpu::Buffer,
size: usize,
}
impl WebGPUDenseBufferCpuGpu {
pub(in crate::backends::webgpu) fn new(
context: &WebGPUContext,
size: usize,
) -> Result<Self, WebGPUError> {
let buffer = context.device.create_buffer(&wgpu::BufferDescriptor {
label: None,
size: size.next_multiple_of(wgpu::COPY_BUFFER_ALIGNMENT as usize) as u64,
usage: wgpu::BufferUsages::UNIFORM
| wgpu::BufferUsages::STORAGE
| wgpu::BufferUsages::INDIRECT
| wgpu::BufferUsages::COPY_SRC
| wgpu::BufferUsages::COPY_DST,
mapped_at_creation: true,
});
let mut view = buffer.get_mapped_range_mut(..)?;
let cpu_ptr = SendPtrMut(view.slice(..).as_raw_element_ptr().cast().as_ptr());
Ok(Self {
cpu_ptr,
view: ManuallyDrop::new(view),
buffer,
size,
})
}
pub(in crate::backends::webgpu) fn wgpu_buffer(&self) -> &wgpu::Buffer {
&self.buffer
}
}
impl Drop for WebGPUDenseBufferCpuGpu {
fn drop(&mut self) {
unsafe { ManuallyDrop::drop(&mut self.view) };
self.buffer.unmap();
}
}
impl Buffer for WebGPUDenseBufferCpuGpu {
type Backend = WebGPU;
fn size(&self) -> usize {
self.size
}
}
impl GpuBuffer for BumpAllocation<WebGPUDenseBufferCpuGpu> {}
impl CpuBuffer for WebGPUDenseBufferCpuGpu {
fn cpu_ptr(&self) -> NonNull<c_void> {
unsafe { NonNull::new_unchecked(self.cpu_ptr.as_ptr()) }
}
}
impl GlobalBufferCpuGpu for WebGPUDenseBufferCpuGpu {
fn handoff(self) -> Result<WebGPUDenseBufferGpu, WebGPUError> {
let gpu = WebGPUDenseBufferGpu {
buffer: self.buffer.clone(),
size: self.size,
};
drop(self);
Ok(gpu)
}
}