mirror of
https://github.com/gfx-rs/wgpu.git
synced 2024-11-22 06:44:14 +00:00
[wgpu] Clarify documentation for Queue
transfer methods.
Explain more clearly that the `write_buffer`, `write_buffer_with`, and `write_texture` methods on `wgpu::Queue` do not immediately submit the transfers for execution. Provide sample code for flushing them.
This commit is contained in:
parent
e5201a7267
commit
fb3b33d092
@ -4991,11 +4991,24 @@ impl<'a> Drop for QueueWriteBufferView<'a> {
|
||||
impl Queue {
|
||||
/// Schedule a data write into `buffer` starting at `offset`.
|
||||
///
|
||||
/// This method is intended to have low performance costs.
|
||||
/// As such, the write is not immediately submitted, and instead enqueued
|
||||
/// internally to happen at the start of the next `submit()` call.
|
||||
///
|
||||
/// This method fails if `data` overruns the size of `buffer` starting at `offset`.
|
||||
///
|
||||
/// This does *not* submit the transfer to the GPU immediately. Calls to
|
||||
/// `write_buffer` begin execution only on the next call to
|
||||
/// [`Queue::submit`]. To get a set of scheduled transfers started
|
||||
/// immediately, it's fine to call `submit` with no command buffers at all:
|
||||
///
|
||||
/// ```no_run
|
||||
/// # let queue: wgpu::Queue = todo!();
|
||||
/// queue.submit([]);
|
||||
/// ```
|
||||
///
|
||||
/// However, `data` will be immediately copied into staging memory, so the
|
||||
/// caller may discard it any time after this call completes.
|
||||
///
|
||||
/// If possible, consider using [`Queue::write_buffer_with`] instead. That
|
||||
/// method avoids an intermediate copy and is often able to transfer data
|
||||
/// more efficiently than this one.
|
||||
pub fn write_buffer(&self, buffer: &Buffer, offset: BufferAddress, data: &[u8]) {
|
||||
DynContext::queue_write_buffer(
|
||||
&*self.context,
|
||||
@ -5008,14 +5021,32 @@ impl Queue {
|
||||
)
|
||||
}
|
||||
|
||||
/// Schedule a data write into `buffer` starting at `offset` via the returned
|
||||
/// [`QueueWriteBufferView`].
|
||||
/// Write to a buffer via a directly mapped staging buffer.
|
||||
///
|
||||
/// Reading from this buffer is slow and will not yield the actual contents of the buffer.
|
||||
/// Return a [`QueueWriteBufferView`] which, when dropped, schedules a copy
|
||||
/// of its contents into `buffer` at `offset`. The returned view
|
||||
/// dereferences to a `size`-byte long `&mut [u8]`, in which you should
|
||||
/// store the data you would like written to `buffer`.
|
||||
///
|
||||
/// This method is intended to have low performance costs.
|
||||
/// As such, the write is not immediately submitted, and instead enqueued
|
||||
/// internally to happen at the start of the next `submit()` call.
|
||||
/// This method may perform transfers faster than [`Queue::write_buffer`],
|
||||
/// because the returned [`QueueWriteBufferView`] is actually the staging
|
||||
/// buffer for the write, mapped into the caller's address space. Writing
|
||||
/// your data directly into this staging buffer avoids the temporary
|
||||
/// CPU-side buffer needed by `write_buffer`.
|
||||
///
|
||||
/// Reading from the returned view is slow, and will not yield the current
|
||||
/// contents of `buffer`.
|
||||
///
|
||||
/// Note that dropping the [`QueueWriteBufferView`] does *not* submit the
|
||||
/// transfer to the GPU immediately. The transfer begins only on the next
|
||||
/// call to [`Queue::submit`] after the view is dropped. To get a set of
|
||||
/// scheduled transfers started immediately, it's fine to call `submit` with
|
||||
/// no command buffers at all:
|
||||
///
|
||||
/// ```no_run
|
||||
/// # let queue: wgpu::Queue = todo!();
|
||||
/// queue.submit([]);
|
||||
/// ```
|
||||
///
|
||||
/// This method fails if `size` is greater than the size of `buffer` starting at `offset`.
|
||||
#[must_use]
|
||||
@ -5059,13 +5090,20 @@ impl Queue {
|
||||
/// texture (coordinate offset, mip level) that will be overwritten.
|
||||
/// * `size` is the size, in texels, of the region to be written.
|
||||
///
|
||||
/// This method is intended to have low performance costs.
|
||||
/// As such, the write is not immediately submitted, and instead enqueued
|
||||
/// internally to happen at the start of the next `submit()` call.
|
||||
/// However, `data` will be immediately copied into staging memory; so the caller may
|
||||
/// discard it any time after this call completes.
|
||||
///
|
||||
/// This method fails if `size` overruns the size of `texture`, or if `data` is too short.
|
||||
///
|
||||
/// This does *not* submit the transfer to the GPU immediately. Calls to
|
||||
/// `write_texture` begin execution only on the next call to
|
||||
/// [`Queue::submit`]. To get a set of scheduled transfers started
|
||||
/// immediately, it's fine to call `submit` with no command buffers at all:
|
||||
///
|
||||
/// ```no_run
|
||||
/// # let queue: wgpu::Queue = todo!();
|
||||
/// queue.submit([]);
|
||||
/// ```
|
||||
///
|
||||
/// However, `data` will be immediately copied into staging memory, so the
|
||||
/// caller may discard it any time after this call completes.
|
||||
pub fn write_texture(
|
||||
&self,
|
||||
texture: ImageCopyTexture<'_>,
|
||||
|
Loading…
Reference in New Issue
Block a user