Struct WebGpuMemoryManager 

pub struct WebGpuMemoryManager { /* private fields */ }

Manages a pool of device-resident wgpu::Buffer objects, returning opaque u64 handles to callers.

All public methods are &self to allow shared references from the backend.

Implementations

impl WebGpuMemoryManager

pub fn new(device: Arc<WebGpuDevice>) -> Self

Create a new memory manager backed by device.

pub fn alloc(&self, bytes: usize) -> WebGpuResult<u64>

Allocate a new device buffer of at least bytes bytes.

The physical buffer size is rounded up to wgpu::COPY_BUFFER_ALIGNMENT (4 bytes): WebGPU requires copy_buffer_to_buffer sizes, map_async ranges, and STORAGE-bound bindings to be multiples of 4, so an odd-sized allocation (3 bytes, or an odd count of 2-byte f16 elements) would otherwise pass alloc() cleanly and only fail later — fatally, with no handler installed — on the first readback or bind. The rounded-up size is what later copy_to_device/copy_from_device calls validate against, which is strictly more permissive than the caller’s requested size, never less.

Returns WebGpuError::InvalidArgument for a zero-byte request and WebGpuError::OutOfMemory if the (rounded) size exceeds what this device can bind — checked against the adapter-derived limits resolved in WebGpuDevice::new, instead of discovered via a fatal wgpu validation abort inside create_buffer.

pub fn free(&self, handle: u64) -> WebGpuResult<()>

Release the buffer associated with handle.

The handle is silently ignored if it is unknown (already freed).

pub fn copy_to_device(&self, handle: u64, src: &[u8]) -> WebGpuResult<()>

Upload src (host bytes) into the device buffer identified by handle.

pub fn copy_from_device(&self, dst: &mut [u8], handle: u64) -> WebGpuResult<()>

Download the device buffer identified by handle into dst (host bytes).

Only the bytes actually requested (dst.len(), rounded up to wgpu::COPY_BUFFER_ALIGNMENT) are staged and copied — previously this always staged and DMA’d the entire source buffer regardless of how much the caller asked for, so e.g. reading a single scalar out of a multi-MiB reduction output moved the whole buffer across the copy engine to deliver 4 bytes.

Uses a temporary MAP_READ staging buffer and blocks — bounded by a generous internal timeout, see WebGpuError::Timeout — until the GPU work completes.

Trait Implementations

impl Debug for WebGpuMemoryManager

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.