wgpu: Switch to arena

This commit is contained in:
Aaron Hill 2022-09-06 23:32:15 -05:00 committed by Nathan Adams
parent c63aa2cfc8
commit b62c17577b
4 changed files with 111 additions and 37 deletions

47
Cargo.lock generated
View File

@ -2,6 +2,12 @@
# It is not intended for manual editing.
version = 3
[[package]]
name = "Inflector"
version = "0.11.4"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "fe438c63458706e03479442743baae6c88256498e6431708f6dfc520a26515d3"
[[package]]
name = "adler"
version = "1.0.2"
@ -34,6 +40,12 @@ dependencies = [
"memchr",
]
[[package]]
name = "aliasable"
version = "0.1.3"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "250f629c0161ad8107cf89319e990051fae62832fd343083bea452d93e2205fd"
[[package]]
name = "alsa"
version = "0.6.0"
@ -273,9 +285,9 @@ dependencies = [
[[package]]
name = "bumpalo"
version = "3.10.0"
version = "3.11.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "37ccbd214614c6783386c1af30caf03192f17891059cecc394b4fb119e363de3"
checksum = "c1ad822118d20d2c234f427000d5acc36eabe1e29a348c89b63dd60b13f28e5d"
[[package]]
name = "bytemuck"
@ -2603,6 +2615,29 @@ version = "6.2.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "648001efe5d5c0102d8cea768e348da85d90af8ba91f0bea908f157951493cd4"
[[package]]
name = "ouroboros"
version = "0.15.4"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "7f56a2b0aa5fc88687aaf63e85a7974422790ce3419a2e1a15870f8a55227822"
dependencies = [
"aliasable",
"ouroboros_macro",
]
[[package]]
name = "ouroboros_macro"
version = "0.15.4"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "6c40641e27d0eb38cae3dee081d920104d2db47a8e853c1a592ef68d33f5ebf4"
dependencies = [
"Inflector",
"proc-macro-error",
"proc-macro2",
"quote",
"syn",
]
[[package]]
name = "output_vt100"
version = "0.1.3"
@ -3181,9 +3216,11 @@ dependencies = [
"futures",
"image",
"log",
"ouroboros",
"raw-window-handle 0.4.3",
"ruffle_render",
"swf",
"typed-arena",
"web-sys",
"wgpu",
]
@ -3836,6 +3873,12 @@ dependencies = [
"strength_reduce",
]
[[package]]
name = "typed-arena"
version = "2.0.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "0685c84d5d54d1c26f7d3eb96cd41550adb97baed141a761cf335d3d33bcd0ae"
[[package]]
name = "typenum"
version = "1.15.0"

View File

@ -16,6 +16,8 @@ enum-map = "2.4.1"
fnv = "1.0.7"
swf = { path = "../../swf" }
image = { version = "0.24.2", default-features = false }
ouroboros = "0.15.4"
typed-arena = "2.0.1"
# desktop
[target.'cfg(not(target_family = "wasm"))'.dependencies.futures]

View File

@ -150,7 +150,7 @@ impl<T: RenderTarget> WgpuRenderBackend<T> {
globals.set_resolution(target.width(), target.height());
let uniform_buffers_storage =
BufferStorage::new(descriptors.limits.min_uniform_buffer_offset_alignment);
BufferStorage::from_alignment(descriptors.limits.min_uniform_buffer_offset_alignment);
Ok(Self {
descriptors,

View File

@ -1,23 +1,35 @@
use bytemuck::Pod;
use std::pin::Pin;
use ouroboros::self_referencing;
use std::cell::RefCell;
use std::{marker::PhantomData, mem};
use typed_arena::Arena;
use wgpu::util::StagingBelt;
/// A simple chunked bump allacator for managing dynamic uniforms that change per-draw.
/// Each draw call may use `UniformBuffer::write_uniforms` can be used to queue
/// the upload of uniform data to the GPU.
pub struct UniformBuffer<'a, T: Pod> {
buffers: &'a mut BufferStorage<T>,
buffers: &'a BufferStorage<T>,
cur_block: usize,
cur_offset: u32,
}
#[self_referencing]
pub struct BufferStorage<T: Pod> {
_phantom: PhantomData<T>,
blocks: Vec<Pin<Box<Block>>>,
staging_belt: StagingBelt,
phantom: PhantomData<T>,
arena: Arena<Block>,
#[borrows(arena)]
#[not_covariant]
allocator: RefCell<Allocator<'this>>,
staging_belt: RefCell<StagingBelt>,
aligned_uniforms_size: u32,
}
struct Allocator<'a> {
arena: &'a Arena<Block>,
blocks: Vec<&'a Block>,
}
impl<T: Pod> BufferStorage<T> {
/// The size of each block.
@ -28,20 +40,27 @@ impl<T: Pod> BufferStorage<T> {
/// The uniform data size for a single draw call.
pub const UNIFORMS_SIZE: u64 = mem::size_of::<T>() as u64;
pub fn new(uniform_alignment: u32) -> Self {
pub fn from_alignment(uniform_alignment: u32) -> Self {
// Calculate alignment of uniforms.
let align_mask = uniform_alignment - 1;
let aligned_uniforms_size = (Self::UNIFORMS_SIZE as u32 + align_mask) & !align_mask;
Self {
blocks: Vec::with_capacity(8),
staging_belt: StagingBelt::new(u64::from(Self::BLOCK_SIZE) / 2),
BufferStorageBuilder {
arena: Arena::with_capacity(8),
allocator_builder: |arena| {
RefCell::new(Allocator {
arena,
blocks: Vec::with_capacity(8),
})
},
staging_belt: RefCell::new(StagingBelt::new(u64::from(Self::BLOCK_SIZE) / 2)),
aligned_uniforms_size,
_phantom: PhantomData,
phantom: PhantomData,
}
.build()
}
/// Adds a newly allocated buffer to the block list, and returns it.
pub fn allocate_block(&mut self, device: &wgpu::Device, layout: &wgpu::BindGroupLayout) {
pub fn allocate_block(&self, device: &wgpu::Device, layout: &wgpu::BindGroupLayout) {
let buffer_label = create_debug_label!("Dynamic buffer");
let buffer = device.create_buffer(&wgpu::BufferDescriptor {
label: buffer_label.as_deref(),
@ -64,12 +83,15 @@ impl<T: Pod> BufferStorage<T> {
}],
});
self.blocks
.push(Pin::new(Box::new(Block { buffer, bind_group })));
self.with_allocator(|alloc| {
let mut alloc = alloc.borrow_mut();
let block = alloc.arena.alloc(Block { buffer, bind_group });
alloc.blocks.push(block);
});
}
pub fn recall(&mut self) {
self.staging_belt.recall();
self.with_staging_belt(|belt| belt.borrow_mut().recall());
}
}
@ -97,34 +119,40 @@ impl<'a, T: Pod> UniformBuffer<'a, T> {
'a: 'b,
{
// Allocate a new block if we've exceeded our capacity.
if self.cur_block >= self.buffers.blocks.len() {
if self.cur_block
>= self
.buffers
.with_allocator(|alloc| alloc.borrow().blocks.len())
{
self.buffers.allocate_block(device, layout);
}
let block = &self.buffers.blocks[self.cur_block];
let block: &'a Block = self
.buffers
.with_allocator(|alloc| alloc.borrow().blocks[self.cur_block]);
// Copy the data into the buffer via the staging belt.
self.buffers
.staging_belt
.write_buffer(
command_encoder,
&block.buffer,
self.cur_offset.into(),
BufferStorage::<T>::UNIFORMS_SIZE.try_into().unwrap(),
device,
)
.copy_from_slice(bytemuck::cast_slice(std::slice::from_ref(data)));
self.buffers.with_staging_belt(|belt| {
belt.borrow_mut()
.write_buffer(
command_encoder,
&block.buffer,
self.cur_offset.into(),
BufferStorage::<T>::UNIFORMS_SIZE.try_into().unwrap(),
device,
)
.copy_from_slice(bytemuck::cast_slice(std::slice::from_ref(data)));
});
// Set the bind group to the final uniform location.
render_pass.set_bind_group(
bind_group_index,
unsafe { mem::transmute::<_, &'a wgpu::BindGroup>(&block.bind_group) },
&[self.cur_offset],
);
render_pass.set_bind_group(bind_group_index, &block.bind_group, &[self.cur_offset]);
// Advance offset.
self.cur_offset += self.buffers.aligned_uniforms_size;
self.cur_offset += self.buffers.borrow_aligned_uniforms_size();
// Advance to next buffer if we are out of room in this buffer.
if BufferStorage::<T>::BLOCK_SIZE - self.cur_offset < self.buffers.aligned_uniforms_size {
if BufferStorage::<T>::BLOCK_SIZE - self.cur_offset
< *self.buffers.borrow_aligned_uniforms_size()
{
self.cur_block += 1;
self.cur_offset = 0;
}
@ -132,7 +160,8 @@ impl<'a, T: Pod> UniformBuffer<'a, T> {
/// Should be called at the end of a frame.
pub fn finish(self) {
self.buffers.staging_belt.finish();
self.buffers
.with_staging_belt(|belt| belt.borrow_mut().finish());
}
}