1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297
//! A very simple compute shader that updates a gpu buffer.
//! That buffer is then copied to the cpu and sent to the main world.
//!
//! This example is not meant to teach compute shaders.
//! It is only meant to explain how to read a gpu buffer on the cpu and then use it in the main world.
//!
//! The code is based on this wgpu example:
//! <https://github.com/gfx-rs/wgpu/blob/fb305b85f692f3fbbd9509b648dfbc97072f7465/examples/src/repeated_compute/mod.rs>
use bevy::{
prelude::*,
render::{
render_graph::{self, RenderGraph, RenderLabel},
render_resource::{binding_types::storage_buffer, *},
renderer::{RenderContext, RenderDevice},
Render, RenderApp, RenderSet,
},
};
use crossbeam_channel::{Receiver, Sender};
// The length of the buffer sent to the gpu
const BUFFER_LEN: usize = 16;
// To communicate between the main world and the render world we need a channel.
// Since the main world and render world run in parallel, there will always be a frame of latency
// between the data sent from the render world and the data received in the main world
//
// frame n => render world sends data through the channel at the end of the frame
// frame n + 1 => main world receives the data
/// This will receive asynchronously any data sent from the render world
#[derive(Resource, Deref)]
struct MainWorldReceiver(Receiver<Vec<u32>>);
/// This will send asynchronously any data to the main world
#[derive(Resource, Deref)]
struct RenderWorldSender(Sender<Vec<u32>>);
fn main() {
App::new()
.insert_resource(ClearColor(Color::BLACK))
.add_plugins((DefaultPlugins, GpuReadbackPlugin))
.add_systems(Update, receive)
.run();
}
/// This system will poll the channel and try to get the data sent from the render world
fn receive(receiver: Res<MainWorldReceiver>) {
// We don't want to block the main world on this,
// so we use try_recv which attempts to receive without blocking
if let Ok(data) = receiver.try_recv() {
println!("Received data from render world: {data:?}");
}
}
// We need a plugin to organize all the systems and render node required for this example
struct GpuReadbackPlugin;
impl Plugin for GpuReadbackPlugin {
fn build(&self, _app: &mut App) {}
// The render device is only accessible inside finish().
// So we need to initialize render resources here.
fn finish(&self, app: &mut App) {
let (s, r) = crossbeam_channel::unbounded();
app.insert_resource(MainWorldReceiver(r));
let render_app = app.sub_app_mut(RenderApp);
render_app
.insert_resource(RenderWorldSender(s))
.init_resource::<ComputePipeline>()
.init_resource::<Buffers>()
.add_systems(
Render,
(
prepare_bind_group
.in_set(RenderSet::PrepareBindGroups)
// We don't need to recreate the bind group every frame
.run_if(not(resource_exists::<GpuBufferBindGroup>)),
// We need to run it after the render graph is done
// because this needs to happen after submit()
map_and_read_buffer.after(RenderSet::Render),
),
);
// Add the compute node as a top level node to the render graph
// This means it will only execute once per frame
render_app
.world_mut()
.resource_mut::<RenderGraph>()
.add_node(ComputeNodeLabel, ComputeNode::default());
}
}
#[derive(Resource)]
struct Buffers {
// The buffer that will be used by the compute shader
gpu_buffer: Buffer,
// The buffer that will be read on the cpu.
// The `gpu_buffer` will be copied to this buffer every frame
cpu_buffer: Buffer,
}
impl FromWorld for Buffers {
fn from_world(world: &mut World) -> Self {
let render_device = world.resource::<RenderDevice>();
let mut init_data = encase::StorageBuffer::new(Vec::new());
// Init the buffer with 0
let data = vec![0; BUFFER_LEN];
init_data.write(&data).expect("Failed to write buffer");
// The buffer that will be accessed by the gpu
let gpu_buffer = render_device.create_buffer_with_data(&BufferInitDescriptor {
label: Some("gpu_buffer"),
contents: init_data.as_ref(),
usage: BufferUsages::STORAGE | BufferUsages::COPY_SRC,
});
// For portability reasons, WebGPU draws a distinction between memory that is
// accessible by the CPU and memory that is accessible by the GPU. Only
// buffers accessible by the CPU can be mapped and accessed by the CPU and
// only buffers visible to the GPU can be used in shaders. In order to get
// data from the GPU, we need to use `CommandEncoder::copy_buffer_to_buffer` to
// copy the buffer modified by the GPU into a mappable, CPU-accessible buffer
let cpu_buffer = render_device.create_buffer(&BufferDescriptor {
label: Some("readback_buffer"),
size: (BUFFER_LEN * std::mem::size_of::<u32>()) as u64,
usage: BufferUsages::MAP_READ | BufferUsages::COPY_DST,
mapped_at_creation: false,
});
Self {
gpu_buffer,
cpu_buffer,
}
}
}
#[derive(Resource)]
struct GpuBufferBindGroup(BindGroup);
fn prepare_bind_group(
mut commands: Commands,
pipeline: Res<ComputePipeline>,
render_device: Res<RenderDevice>,
buffers: Res<Buffers>,
) {
let bind_group = render_device.create_bind_group(
None,
&pipeline.layout,
&BindGroupEntries::single(buffers.gpu_buffer.as_entire_binding()),
);
commands.insert_resource(GpuBufferBindGroup(bind_group));
}
#[derive(Resource)]
struct ComputePipeline {
layout: BindGroupLayout,
pipeline: CachedComputePipelineId,
}
impl FromWorld for ComputePipeline {
fn from_world(world: &mut World) -> Self {
let render_device = world.resource::<RenderDevice>();
let layout = render_device.create_bind_group_layout(
None,
&BindGroupLayoutEntries::single(
ShaderStages::COMPUTE,
storage_buffer::<Vec<u32>>(false),
),
);
let shader = world.load_asset("shaders/gpu_readback.wgsl");
let pipeline_cache = world.resource::<PipelineCache>();
let pipeline = pipeline_cache.queue_compute_pipeline(ComputePipelineDescriptor {
label: Some("GPU readback compute shader".into()),
layout: vec![layout.clone()],
push_constant_ranges: Vec::new(),
shader: shader.clone(),
shader_defs: Vec::new(),
entry_point: "main".into(),
});
ComputePipeline { layout, pipeline }
}
}
fn map_and_read_buffer(
render_device: Res<RenderDevice>,
buffers: Res<Buffers>,
sender: Res<RenderWorldSender>,
) {
// Finally time to get our data back from the gpu.
// First we get a buffer slice which represents a chunk of the buffer (which we
// can't access yet).
// We want the whole thing so use unbounded range.
let buffer_slice = buffers.cpu_buffer.slice(..);
// Now things get complicated. WebGPU, for safety reasons, only allows either the GPU
// or CPU to access a buffer's contents at a time. We need to "map" the buffer which means
// flipping ownership of the buffer over to the CPU and making access legal. We do this
// with `BufferSlice::map_async`.
//
// The problem is that map_async is not an async function so we can't await it. What
// we need to do instead is pass in a closure that will be executed when the slice is
// either mapped or the mapping has failed.
//
// The problem with this is that we don't have a reliable way to wait in the main
// code for the buffer to be mapped and even worse, calling get_mapped_range or
// get_mapped_range_mut prematurely will cause a panic, not return an error.
//
// Using channels solves this as awaiting the receiving of a message from
// the passed closure will force the outside code to wait. It also doesn't hurt
// if the closure finishes before the outside code catches up as the message is
// buffered and receiving will just pick that up.
//
// It may also be worth noting that although on native, the usage of asynchronous
// channels is wholly unnecessary, for the sake of portability to WASM
// we'll use async channels that work on both native and WASM.
let (s, r) = crossbeam_channel::unbounded::<()>();
// Maps the buffer so it can be read on the cpu
buffer_slice.map_async(MapMode::Read, move |r| match r {
// This will execute once the gpu is ready, so after the call to poll()
Ok(_) => s.send(()).expect("Failed to send map update"),
Err(err) => panic!("Failed to map buffer {err}"),
});
// In order for the mapping to be completed, one of three things must happen.
// One of those can be calling `Device::poll`. This isn't necessary on the web as devices
// are polled automatically but natively, we need to make sure this happens manually.
// `Maintain::Wait` will cause the thread to wait on native but not on WebGpu.
// This blocks until the gpu is done executing everything
render_device.poll(Maintain::wait()).panic_on_timeout();
// This blocks until the buffer is mapped
r.recv().expect("Failed to receive the map_async message");
{
let buffer_view = buffer_slice.get_mapped_range();
let data = buffer_view
.chunks(std::mem::size_of::<u32>())
.map(|chunk| u32::from_ne_bytes(chunk.try_into().expect("should be a u32")))
.collect::<Vec<u32>>();
sender
.send(data)
.expect("Failed to send data to main world");
}
// We need to make sure all `BufferView`'s are dropped before we do what we're about
// to do.
// Unmap so that we can copy to the staging buffer in the next iteration.
buffers.cpu_buffer.unmap();
}
/// Label to identify the node in the render graph
#[derive(Debug, Hash, PartialEq, Eq, Clone, RenderLabel)]
struct ComputeNodeLabel;
/// The node that will execute the compute shader
#[derive(Default)]
struct ComputeNode {}
impl render_graph::Node for ComputeNode {
fn run(
&self,
_graph: &mut render_graph::RenderGraphContext,
render_context: &mut RenderContext,
world: &World,
) -> Result<(), render_graph::NodeRunError> {
let pipeline_cache = world.resource::<PipelineCache>();
let pipeline = world.resource::<ComputePipeline>();
let bind_group = world.resource::<GpuBufferBindGroup>();
if let Some(init_pipeline) = pipeline_cache.get_compute_pipeline(pipeline.pipeline) {
let mut pass =
render_context
.command_encoder()
.begin_compute_pass(&ComputePassDescriptor {
label: Some("GPU readback compute pass"),
..default()
});
pass.set_bind_group(0, &bind_group.0, &[]);
pass.set_pipeline(init_pipeline);
pass.dispatch_workgroups(BUFFER_LEN as u32, 1, 1);
}
// Copy the gpu accessible buffer to the cpu accessible buffer
let buffers = world.resource::<Buffers>();
render_context.command_encoder().copy_buffer_to_buffer(
&buffers.gpu_buffer,
0,
&buffers.cpu_buffer,
0,
(BUFFER_LEN * std::mem::size_of::<u32>()) as u64,
);
Ok(())
}
}