A high-performance inference engine for AI models
use std::range::Range;
use crate::{
backends::{
common::{Backend, GpuBuffer},
webgpu::WebGPU,
},
utils::downcast::downcast_ref,
};
pub mod dense;
pub mod sparse;
pub fn binding_parts(
buffer: &(impl GpuBuffer<Backend = WebGPU> + ?Sized),
range: Range<usize>,
) -> (&wgpu::Buffer, usize, usize) {
let (buffer, offset) = if let Some(buffer) = downcast_ref::<<WebGPU as Backend>::GlobalGpuBuffer>(buffer) {
(buffer.wgpu_buffer(), 0)
} else if let Some(buffer) = downcast_ref::<<WebGPU as Backend>::ConstantBuffer>(buffer) {
(buffer.page().wgpu_buffer(), buffer.range().start)
} else if let Some(buffer) = downcast_ref::<<WebGPU as Backend>::ScratchBuffer>(buffer) {
(buffer.page().wgpu_buffer(), buffer.range().start)
} else {
unreachable!()
};
(buffer, offset + range.start, range.iter().len())
}