Struct GpuContext 

pub struct GpuContext { /* private fields */ }

A device plus everything FlashRaw needs to run its kernels on it.

Create one with GpuContext::new (default adapter), or wrap an application’s existing device with GpuContext::from_device so that FlashRaw outputs live on the same device as the rest of the app (required for zero-copy interop).

Implementations

impl GpuContext

pub async fn new() -> Result<Self>

Request the default high-performance adapter and create a device.

pub fn from_device(device: Device, queue: Queue, info: &AdapterInfo) -> Self

Use an application-owned device (e.g. the one a wgpu/WebGPU renderer already uses) so that GPU-resident outputs can be consumed without copies. TIMESTAMP_QUERY is used when the device was created with it.

pub fn trim(&self)

Free all pooled buffers.

pub fn unified_memory(&self) -> bool

True when storage buffers are mapped directly (unified memory).

pub fn info(&self) -> &GpuAdapterInfo

pub fn device(&self) -> &Device

pub fn queue(&self) -> &Queue

pub fn adapter(&self) -> Option<&Adapter>

The adapter, when this context created its own device.

pub async fn wait_idle(&self) -> Result<()>

Wait until all submitted work has completed. On native this blocks in Device::poll; in the browser it yields to the event loop.

pub fn override_limits(&self, max_binding: Option<u64>, max_buffer: Option<u64>)

Cap the storage-binding size and/or the buffer size FlashRaw plans with, below the adapter’s limits (None restores the adapter limit). Frames are processed in bands that fit these caps, so this bounds the size of every buffer FlashRaw binds (useful to limit peak device memory, and to exercise tiling in tests). Values are rounded down to a multiple of 256 bytes, with a 4 KB minimum.

pub fn max_binding_bytes(&self) -> u64

Largest storage binding FlashRaw uses (adapter limit, caps, 32-bit addressing).

pub fn max_buffer_bytes(&self) -> u64

Largest single buffer FlashRaw allocates (adapter limit and cap).

pub fn estimate_bands(&self, input_bytes: u64, output_bytes: u64) -> u32

Number of bands a frame with input_bytes of upload and output_bytes of result needs on this device (an estimate for cost models; 1 = untiled).

impl GpuContext

pub async fn calibration(self: &Arc<Self>) -> GpuCalibration

Measured characteristics of this device (one-time micro-benchmark, ~2048x2048 synthetic frame; cached). Returns static defaults if the benchmark fails or FLASHRAW_GPU_CALIBRATION=static.

pub fn cached_calibration(&self) -> Option<GpuCalibration>

The calibration if it has already been measured (never runs the benchmark).

pub fn calibration_blocking(self: &Arc<Self>) -> GpuCalibration

Blocking GpuContext::calibration (native only).

impl GpuContext

pub async fn run( self: &Arc<Self>, frame: &GpuFrame<'_>, info: &FrameInfo, op: &GpuOp, ) -> Result<GpuRun>

Upload frame, run op and return the GPU-resident result.

Frames whose input or output exceed one storage binding (GpuContext::max_binding_bytes) are processed in horizontal bands: each band uploads only the rows it needs (plus a one-row halo for the demosaic), the upload of band N+1 overlaps the kernel of band N, and the output is either one buffer (bound per band) or a tiled GpuImage.

pub fn run_blocking( self: &Arc<Self>, frame: &GpuFrame<'_>, info: &FrameInfo, op: &GpuOp, ) -> Result<GpuRun>

Blocking GpuContext::run (native only).

Trait Implementations

impl Debug for GpuContext

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

Formats the value using the given formatter.