Smart compaction goes brrrrr

This commit is contained in:
2026-09-17 23:01:29 +02:00
parent f34550e692
commit 306b01b44c
5 changed files with 180 additions and 401 deletions
+3 -2
View File
@@ -113,6 +113,7 @@ void scatter(
uint32_t3 local_thread_id: SV_GroupThreadID,
uint32_t3 global_thread_id: SV_DispatchThreadID)
{
let total = index_buffer.getCount();
let thread_index = global_thread_id.x;
let local_thread_index = local_thread_id.x;
@@ -123,12 +124,12 @@ void scatter(
let sum2 = sum_buffer[thread_index * 2 + 1] + reduced_value;
// Scatter index to compact index buffer if predicate is true
if (count_buffer[thread_index * 2] != 0)
if (count_buffer[thread_index * 2] != 0 && sum1 < total)
{
index_buffer[sum1] = thread_index * 2;
}
if (count_buffer[thread_index * 2 + 1] != 0)
if (count_buffer[thread_index * 2 + 1] != 0 && sum2 < total)
{
index_buffer[sum2] = thread_index * 2 + 1;
}
Binary file not shown.
+20 -97
View File
@@ -69,7 +69,7 @@ use crate::voxel_cache::data::ColorBytes;
use crate::voxel_cache::data::DestinationElement;
use crate::voxel_cache::data::LocationPoolElement;
use crate::voxel_cache::data::StructurePoolElement;
use crate::voxel_cache::indirect_buffer::BufferCompactor;
use crate::voxel_cache::indirect_buffer::NonZeroBufferCompactor;
use crate::voxel_cache::producer_interface::CacheProducerInterface;
use crate::voxel_cache::producer_interface::CacheRequest;
@@ -105,6 +105,7 @@ struct State
instance_count: usize,
usage_vec: Arc<Mutex<Vec<usize>>>,
rm_time: Arc<Mutex<f32>>,
frame_time: Arc<Mutex<f32>>,
insertion_debounce: bool,
camera: Camera,
@@ -164,6 +165,7 @@ impl State
| Features::TIMESTAMP_QUERY
| Features::SHADER_I16
| Features::SHADER_F16
| Features::TIMESTAMP_QUERY_INSIDE_ENCODERS
| Features::SHADER_INT64,
required_limits: wgpu::Limits {
max_immediate_size: 112,
@@ -176,94 +178,6 @@ impl State
.unwrap();
let mut encoder = device.create_command_encoder(&Default::default());
// --------------------------
let mut count_buffer = vec![0u32; 65536];
let mut check_buffer = vec![];
// Spread random numbers throughout the buffer
let mut picked = HashSet::new();
for i in 1..=1000
{
let mut r = rand::random_range(0..(count_buffer.len()));
while picked.contains(&r)
{
r = rand::random_range(0..(count_buffer.len()));
}
picked.insert(r);
check_buffer.push(r);
count_buffer[r] = i;
}
check_buffer.sort();
count_buffer
.iter()
.enumerate()
.filter(|(_, x)| **x != 0)
.take(100)
.for_each(|(x, y)| println!("{}, {}", x, y));
let buffer_compactor = BufferCompactor::new(&device, count_buffer.len());
let gpu_count_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: None,
contents: bytemuck::cast_slice(count_buffer.as_slice()),
usage: BufferUsages::STORAGE,
});
let gpu_index_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: None,
contents: bytemuck::cast_slice(vec![0u32; count_buffer.len()].as_slice()),
usage: BufferUsages::STORAGE | BufferUsages::COPY_SRC,
});
buffer_compactor.compact_buffer(
&device,
&mut encoder,
&gpu_count_buffer,
&gpu_index_buffer,
);
queue.submit([encoder.finish()]);
// --------------------------
wgpu::util::DownloadBuffer::read_buffer(
&device,
&queue,
&buffer_compactor.reduced_buffer.slice(..),
//&buffer_compactor.total_sum_buffer(),
|result| {
let result = result.unwrap();
let indices: Vec<u32> = result
.to_vec()
.chunks(4)
.map(|x| u32::from_ne_bytes([x[0], x[1], x[2], x[3]]))
.collect();
println!("Total sum {:?}", indices);
},
);
wgpu::util::DownloadBuffer::read_buffer(
&device,
&queue,
&gpu_index_buffer.slice(..),
|result| {
let result = result.unwrap();
let indices: Vec<u32> = result
.to_vec()
.chunks(4)
.map(|x| u32::from_ne_bytes([x[0], x[1], x[2], x[3]]))
.collect();
println!("{:?}", &indices[..1010]);
},
);
for _ in 0..16
{
device.poll(wgpu::wgt::PollType::Poll);
}
println!("Check buffer : {:?}", check_buffer);
panic!();
let size = window.inner_size();
@@ -473,6 +387,7 @@ impl State
terrain_generator: Arc::new(terrain_generator),
chunk_pos_map: chunk_pos_map.into(),
rm_time: Arc::new(Mutex::new(0.)),
frame_time: Arc::new(Mutex::new(0.)),
};
// Configure surface for the first time
@@ -639,25 +554,26 @@ impl State
let timestamp_query = self.device.create_query_set(&wgpu::QuerySetDescriptor {
label: Some("timestamp_query_set"),
ty: wgpu::QueryType::Timestamp,
count: 2,
count: 4,
});
let timestamp_buffer = self.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("timestamp_buffer"),
size: (size_of::<u64>() * 2) as u64,
size: (size_of::<u64>() * 4) as u64,
usage: BufferUsages::QUERY_RESOLVE | BufferUsages::COPY_SRC,
mapped_at_creation: false,
});
// ~~ Main render pass ~~
let mut encoder = self.device.create_command_encoder(&Default::default());
encoder.write_timestamp(&timestamp_query, 0);
{
let mut renderpass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("compute_render_pass"),
timestamp_writes: Some(wgpu::ComputePassTimestampWrites {
query_set: &timestamp_query,
beginning_of_pass_write_index: Some(0),
end_of_pass_write_index: Some(1),
beginning_of_pass_write_index: Some(2),
end_of_pass_write_index: Some(3),
}),
});
@@ -682,8 +598,6 @@ impl State
1,
);
drop(renderpass);
encoder.resolve_query_set(&timestamp_query, 0..2, &timestamp_buffer, 0);
}
self.target_blitter.copy(
@@ -763,6 +677,10 @@ impl State
);
}
ui.label(format!(
"frame time: {}",
*self.frame_time.lock() / 1_000_000.
));
ui.label(format!(
"Ray-marching time: {}",
*self.rm_time.lock() / 1_000_000.
@@ -829,12 +747,15 @@ impl State
}
// ~~ Submit command buffer ~~
encoder.write_timestamp(&timestamp_query, 1);
encoder.resolve_query_set(&timestamp_query, 0..4, &timestamp_buffer, 0);
self.queue.submit([encoder.finish()]);
self.window.pre_present_notify();
self.queue.present(surface_texture);
// ~~ Get Ray-marching timestamps, report time ~~
let cloned_rm_time = self.rm_time.clone();
let cloned_frame_time = self.frame_time.clone();
let cloned_queue = self.queue.clone();
DownloadBuffer::read_buffer(
&self.device,
@@ -843,8 +764,10 @@ impl State
move |buffer| {
let buffer_slice = buffer.unwrap();
let slice: &[u64] = cast_slice(&buffer_slice);
let time = (slice[1] - slice[0]) as f32 * cloned_queue.get_timestamp_period();
*cloned_rm_time.lock() = time;
let frame_time = (slice[1] - slice[0]) as f32 * cloned_queue.get_timestamp_period();
let rm_time = (slice[3] - slice[2]) as f32 * cloned_queue.get_timestamp_period();
*cloned_rm_time.lock() = rm_time;
*cloned_frame_time.lock() = frame_time;
},
);
+14 -11
View File
@@ -282,7 +282,7 @@ where
// Sorted requests
wgpu::BindGroupEntry {
binding: 3,
resource: request_buffer.sort_buffer().as_entire_binding(),
resource: request_buffer.index_buffer().as_entire_binding(),
},
// LRU list
wgpu::BindGroupEntry {
@@ -372,13 +372,13 @@ where
@group(0) @binding(0) var<storage, read_write> structure_pool: array<StructurePoolElement>;
@group(0) @binding(1) var<storage, read_write> color_pool: array<ColorPoolElement>;
@group(0) @binding(2) var<storage, read_write> location_pool: array<LocationPoolElement>;
@group(0) @binding(3) var<storage, read> sorted_requests: array<SortedRequestsElement>;
@group(0) @binding(3) var<storage, read> sorted_requests: array<u32>;
@group(0) @binding(4) var<storage, read> lru_list: array<u32>;
@group(0) @binding(5) var<storage, read_write> usage_buffer: array<u32>;
@group(0) @binding(6) var<storage, read> pools_request_count: u32;
@group(1) @binding(0) var<storage, read_write> structure_table_pointers: array<u32>;
@group(1) @binding(1) var<storage, read_write> structure_table_sorted_requests: array<SortedRequestsElement>;
@group(1) @binding(1) var<storage, read_write> structure_table_sorted_requests: array<u32>;
@group(1) @binding(2) var<storage, read_write> structure_table_request_count: u32;
");
@@ -445,7 +445,7 @@ where
{{
var request: CacheInterfaceRequest;
// Index in structure table IS structure id
request.structure_id = structure_table_sorted_requests[index].node;
request.structure_id = structure_table_sorted_requests[index];
request.locator = 0; // Root request -> locator 0
request.child_index = 0xFFFFFFFF; // child index u32::MAX ->
@@ -460,16 +460,19 @@ where
}}
let pool_index = index - structure_table_request_count;
let sorted_requests_node = sorted_requests[pool_index] / 64;
let sorted_requests_child = sorted_requests[pool_index] % 64;
// Request location comes from location pool
var request: CacheInterfaceRequest;
request.structure_id = location_pool[sorted_requests[pool_index].node].structure_id;
request.locator = location_pool[sorted_requests[pool_index].node].structure_locator;
request.child_index = sorted_requests[pool_index].child;
request.structure_id = location_pool[sorted_requests_node].structure_id;
request.locator = location_pool[sorted_requests_node].structure_locator;
request.child_index = sorted_requests_child;
var request_wb: CacheInterfaceRequestWb;
request_wb.node_index = sorted_requests[pool_index].node;
request_wb.child_index = sorted_requests[pool_index].child;
request_wb.node_index = sorted_requests_node;
request_wb.child_index = sorted_requests_child;
// Write request
requests[index] = request;
@@ -809,7 +812,7 @@ where
wgpu::BindGroupEntry
{
binding: 1,
resource: self.structure_table.request_buffer.sort_buffer().as_entire_binding(),
resource: self.structure_table.request_buffer.index_buffer().as_entire_binding(),
},
wgpu::BindGroupEntry
{
+143 -291
View File
@@ -9,23 +9,27 @@ use wgpu::Device;
use wgpu::ShaderStages;
use crate::voxel_cache::data::RequestBufferElement;
use crate::voxel_cache::indirect_buffer::NonZeroBufferCompactor;
pub struct RequestBuffer<const N: usize>
{
pub device: Device,
pub cache_size: usize,
pub request_buffer: Buffer,
pub request_count_buffer: Buffer,
pub indirect_count_storage: Buffer,
pub element_sort_count_buffer: Buffer,
pub indirect_count: Buffer,
pub sort_buffer: Buffer,
pub index_buffer: Buffer,
pub indirect_count_storage: Buffer,
pub indirect_count: Buffer,
pub request_count: Buffer,
pub buffer_compactor: NonZeroBufferCompactor,
pub bindgroup: BindGroup,
pub compute_indirect_pipeline: ComputePipeline,
pub reset_pipeline: ComputePipeline,
pub sort_pipeline: ComputePipeline,
pub bindgroup: BindGroup,
pub device: Device,
pub compaction_pipeline: ComputePipeline,
pub running_sum_pipeline: ComputePipeline,
}
impl<const N: usize> RequestBuffer<N>
@@ -41,13 +45,6 @@ where
mapped_at_creation: false,
});
let request_count_buffer = device.create_buffer(&wgpu::BufferDescriptor {
label: Some("Request_count_buffer"),
size: size_of::<u32>() as u64,
usage: BufferUsages::STORAGE | BufferUsages::COPY_SRC,
mapped_at_creation: false,
});
let indirect_count = device.create_buffer(&wgpu::BufferDescriptor {
label: Some("indirect_count_buffer"),
usage: BufferUsages::INDIRECT | BufferUsages::COPY_DST,
@@ -62,33 +59,29 @@ where
mapped_at_creation: false,
});
let element_sort_count_buffer = device.create_buffer(&wgpu::BufferDescriptor {
label: Some("indirect_count_storage_buffer"),
usage: BufferUsages::STORAGE,
let request_count = device.create_buffer(&wgpu::BufferDescriptor {
label: Some("request_count_storage_buffer"),
usage: BufferUsages::STORAGE | BufferUsages::COPY_SRC | BufferUsages::COPY_DST,
size: size_of::<u32>() as u64,
mapped_at_creation: false,
});
// One element number per child entry
// One number to identify node, one to identify sub child
let sort_buffer_count = cache_size * N * N * N;
let sort_buffer = device.create_buffer(&wgpu::BufferDescriptor {
label: Some(format!("request_sort_buffer_{N}").as_str()),
size: (size_of::<(u32, u32)>() * sort_buffer_count) as u64,
let index_buffer_count = cache_size * N * N * N;
let index_buffer = device.create_buffer(&wgpu::BufferDescriptor {
label: Some(format!("request_index_buffer_{N}").as_str()),
size: (size_of::<u32>() * index_buffer_count) as u64,
usage: BufferUsages::STORAGE | BufferUsages::COPY_SRC,
mapped_at_creation: true,
mapped_at_creation: false,
});
let init_data = (0..(cache_size as u32))
.flat_map(|i| (0..((N * N * N) as u32)).map(move |j| (i, j)))
.flat_map(|(i, j)| [i, j]) // Because bytemuck does not like tuples ...
.collect::<Vec<_>>();
sort_buffer
.get_mapped_range_mut(0..)
.unwrap()
.copy_from_slice(bytemuck::cast_slice(&init_data));
sort_buffer.unmap();
let sort_buffer = device.create_buffer(&wgpu::BufferDescriptor {
label: Some(format!("request_sort_buffer_{N}").as_str()),
size: (size_of::<u32>() * index_buffer_count) as u64,
usage: BufferUsages::STORAGE | BufferUsages::COPY_SRC,
mapped_at_creation: false,
});
let bindgroup_layout = device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("request_sort_pipeline_bing_group_layout"),
@@ -160,7 +153,7 @@ where
},
BindGroupEntry {
binding: 2,
resource: request_count_buffer.as_entire_binding(),
resource: index_buffer.as_entire_binding(),
},
BindGroupEntry {
binding: 3,
@@ -168,7 +161,7 @@ where
},
BindGroupEntry {
binding: 4,
resource: element_sort_count_buffer.as_entire_binding(),
resource: request_count.as_entire_binding(),
},
],
});
@@ -188,24 +181,13 @@ where
label: Some("clear_chunk_requests shader module"),
source: wgpu::ShaderSource::Wgsl(
format!("
struct RequestElement
{{
children: array<u32, {children_count}>
}}
struct SortElement
{{
node: u32,
child: u32
}}
var<immediate> both_phase: u32;
@group(0) @binding(0) var<storage, read_write> request_buffer: array<RequestElement>;
@group(0) @binding(1) var<storage, read_write> sort_indirection: array<SortElement>;
@group(0) @binding(2) var<storage, read_write> request_count: u32;
@group(0) @binding(3) var<storage, read_write> indirect_count: vec3<u32>;
@group(0) @binding(4) var<storage, read_write> element_sort_count: u32;
@group(0) @binding(0) var<storage, read_write> request_buffer: array<u32>;
@group(0) @binding(1) var<storage, read_write> sort_buffer: array<u32>;
@group(0) @binding(2) var<storage, read_write> index_buffer: array<u32>;
@group(0) @binding(3) var<storage, read_write> indirect_count: vec3<u32>;
@group(0) @binding(4) var<storage, read_write> request_count: u32;
fn big_fusion(index: u32, phase_size: u32) -> vec2<u32>
{{
@@ -234,14 +216,14 @@ where
@compute
@workgroup_size(64)
@workgroup_size(64, 1, 1)
fn main(
@builtin(global_invocation_id) global_invocation_id: vec3<u32>
)
{{
let index = global_invocation_id.x;
let total = arrayLength(&sort_indirection);
//let total = arrayLength(&sort_indirection);
let sub_phase = (both_phase >> 16) & 0xFFFF;
let phase = both_phase & 0xFFFF;
@@ -252,18 +234,30 @@ where
if(phase == sub_phase)
{{
let next_phase_width = phase_total_width * 2;
if(next_phase_width >= (element_sort_count * 2) && phase_total_width >= (element_sort_count * 2))
if(next_phase_width >= (request_count * 2) && phase_total_width >= (request_count * 2))
{{
indirect_count = vec3<u32>(0);
//indirect_count = vec3<u32>(0);
}}
}}
if(phase == 0)
{{
// Phase 0. Index buffer contains indicies of compacted buffer.
// We can perform compaction on the data
let a = index * 2;
let b = a + 1;
sort_indirection[a].child = a % {children_count};
sort_indirection[b].child = b % {children_count};
if(a < request_count)
{{
let compact_index_a = index_buffer[a];
sort_buffer[a] = request_buffer[compact_index_a];
}}
if(b < request_count)
{{
let compact_index_b = index_buffer[b];
sort_buffer[b] = request_buffer[compact_index_b];
}}
}}
var swap_indices = vec2<u32>(0, 0);
@@ -280,180 +274,24 @@ where
// Do swap
//if(swap_indices.y >= request_count_round_up)
if(swap_indices.y >= element_sort_count)
if(swap_indices.y >= request_count)
{{
// Suppose that swap_indices.y is -inf, dont swap
return;
}}
let av = request_buffer[sort_indirection[swap_indices.x].node].children[sort_indirection[swap_indices.x].child];
let bv = request_buffer[sort_indirection[swap_indices.y].node].children[sort_indirection[swap_indices.y].child];
let av = sort_buffer[swap_indices.x];
let bv = sort_buffer[swap_indices.y];
if(bv > av)
{{
let temp = sort_indirection[swap_indices.x];
sort_indirection[swap_indices.x] = sort_indirection[swap_indices.y];
sort_indirection[swap_indices.y] = temp;
}}
}}
")
.into(),
),
}),
entry_point: Some("main"),
compilation_options: Default::default(),
cache: None,
});
let sort_temp = sort_buffer[swap_indices.x];
sort_buffer[swap_indices.x] = sort_buffer[swap_indices.y];
sort_buffer[swap_indices.y] = sort_temp;
let running_sum_pipeline =
device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("running_sum_pipeline"),
layout: Some(pipeline_layouts),
module: &device.create_shader_module(wgpu::ShaderModuleDescriptor {
label: Some("running_sum shader module"),
source: wgpu::ShaderSource::Wgsl(
format!("
struct RequestElement
{{
children: array<u32, {children_count}>
}}
struct SortElement
{{
node: u32,
child: u32
}}
var<immediate> phase: u32;
@group(0) @binding(0) var<storage, read_write> request_buffer: array<RequestElement>;
@group(0) @binding(1) var<storage, read_write> sort_indirection: array<SortElement>;
@group(0) @binding(2) var<storage, read_write> request_counts: atomic<u32>;
@group(0) @binding(3) var<storage, read_write> indirect_count: u32;
@group(0) @binding(4) var<storage, read_write> node_sort_count: u32;
@compute
@workgroup_size(64)
fn main(
@builtin(global_invocation_id) global_invocation_id: vec3<u32>
)
{{
let index = global_invocation_id.x;
let len = arrayLength(&request_buffer);
if(index > len) {{ return; }}
let sindex = index * {children_count};
if(phase == 0)
{{
var count = 0;
for(var i = 0; i < {children_count}; i++)
{{
count += select(0, 1, request_buffer[index].children[i] != 0);
}}
sort_indirection[sindex].child = select(u32(0), u32(1), count != 0);
if count != 0
{{
atomicAdd(&request_counts, 1);
}}
return;
}}
if(phase == 0xFFFFFFFF)
{{
// Double buffering bring back
sort_indirection[sindex].child = sort_indirection[sindex].node;
}}
// Phase is not zero, running sum part
let running_sum_phase = phase - 1;
let running_sum_offset = u32((1 << running_sum_phase) * {children_count});
var add = u32(0);
// Double buffering
if(running_sum_phase % 2 == 0)
{{
if(sindex >= running_sum_offset)
{{
add = sort_indirection[sindex - running_sum_offset].child;
}}
sort_indirection[sindex].node = sort_indirection[sindex].child + add;
}}
else
{{
if(sindex >= running_sum_offset)
{{
add = sort_indirection[sindex - running_sum_offset].node;
}}
sort_indirection[sindex].child = sort_indirection[sindex].node + add;
}}
}}
")
.into(),
),
}),
entry_point: Some("main"),
compilation_options: Default::default(),
cache: None,
});
let compaction_pipeline =
device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("compaction_pipeline"),
layout: Some(pipeline_layouts),
module: &device.create_shader_module(wgpu::ShaderModuleDescriptor {
label: Some("compaction shader module"),
source: wgpu::ShaderSource::Wgsl(
format!("
struct RequestElement
{{
children: array<u32, {children_count}>
}}
struct SortElement
{{
node: u32,
child: u32
}}
var<immediate> phase: u32;
@group(0) @binding(0) var<storage, read_write> request_buffer: array<RequestElement>;
@group(0) @binding(1) var<storage, read_write> sort_indirection: array<SortElement>;
@group(0) @binding(2) var<storage, read_write> request_counts: atomic<u32>;
@group(0) @binding(3) var<storage, read_write> indirect_count: vec3<u32>;
@group(0) @binding(4) var<storage, read_write> element_sort_count: u32;
@compute
@workgroup_size(64)
fn main(
@builtin(global_invocation_id) global_invocation_id: vec3<u32>
)
{{
let index = global_invocation_id.x;
let sindex = index * {children_count};
let len = arrayLength(&request_buffer);
element_sort_count = sort_indirection[(len - 1) * {children_count}].child * {children_count};
let invocation_count = (element_sort_count / 2) + select(u32(0), u32(1), element_sort_count % 2 != 0);
indirect_count = vec3(
(invocation_count / 64) + select(u32(0), u32(1), invocation_count % 64 != 0),
1, 1
);
let destination_index = sort_indirection[sindex].child - 1;
var count = u32(0);
for(var i = 0; i < {children_count}; i++)
{{
count += request_buffer[index].children[i];
}}
if(count != 0) // Keep ?
{{
for(var i = u32(0); i < u32({children_count}); i++)
{{
sort_indirection[destination_index * u32({children_count}) + i].node = index;
}}
let index_temp = index_buffer[swap_indices.x];
index_buffer[swap_indices.x] = index_buffer[swap_indices.y];
index_buffer[swap_indices.y] = index_temp;
}}
}}
")
@@ -478,13 +316,14 @@ where
children: array<u32, {children_count}>
}}
@group(0) @binding(0) var<storage, read_write> request_buffer: array<RequestElement>;
@group(0) @binding(1) var<storage, read_write> _ignore: array<u32>;
@group(0) @binding(2) var<storage, read_write> request_count: u32;
@group(0) @binding(4) var<storage, read_write> node_sort_count: u32;
@group(0) @binding(0) var<storage, read_write> request_buffer: array<u32>;
@group(0) @binding(1) var<storage, read_write> sort_buffer: array<u32>;
@group(0) @binding(2) var<storage, read_write> index_buffer: array<u32>;
@group(0) @binding(3) var<storage, read_write> indirect_count: vec3<u32>;
@group(0) @binding(4) var<storage, read_write> request_count: u32;
@compute
@workgroup_size(64)
@workgroup_size(256, 1, 1)
fn main(
@builtin(global_invocation_id) global_invocation_id: vec3<u32>
)
@@ -494,20 +333,44 @@ where
if(index >= total)
{{ return; }}
if(index < total)
{{
for(var i = 0; i < {children_count}; i += 1)
{{
/*
if(request_buffer[index].children[i] != 0)
{{
atomicAdd(&request_count, 1);
}}
*/
request_buffer[index].children[i] = 0;
}}
request_count = 0;
}}
request_buffer[index] = 0;
}}
")
.into(),
),
}),
entry_point: Some("main"),
compilation_options: Default::default(),
cache: None,
});
let compute_indirect_pipeline =
device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("compute_indirect_pipeline"),
layout: Some(pipeline_layouts),
module: &device.create_shader_module(wgpu::ShaderModuleDescriptor {
label: Some("clear_chunk_requests shader module"),
source: wgpu::ShaderSource::Wgsl(
format!("
@group(0) @binding(0) var<storage, read_write> request_buffer: array<u32>;
@group(0) @binding(1) var<storage, read_write> sort_buffer: array<u32>;
@group(0) @binding(2) var<storage, read_write> index_buffer: array<u32>;
@group(0) @binding(3) var<storage, read_write> indirect_count: vec3<u32>;
@group(0) @binding(4) var<storage, read_write> request_count: u32;
@compute
@workgroup_size(1, 1, 1)
fn main(
@builtin(global_invocation_id) global_invocation_id: vec3<u32>
)
{{
let invocation_count = (request_count / 2)
+ select(u32(0), u32(1), request_count % 2 != 0);
let workgroup_count = (invocation_count / 64)
+ select(u32(0), u32(1), invocation_count % 64 != 0);
indirect_count = vec3<u32>(workgroup_count, 1, 1);
}}
")
.into(),
@@ -519,17 +382,17 @@ where
});
Self {
buffer_compactor: NonZeroBufferCompactor::new(&device, index_buffer_count),
cache_size,
request_buffer,
request_count_buffer,
indirect_count,
indirect_count_storage,
element_sort_count_buffer,
running_sum_pipeline,
index_buffer,
sort_buffer,
indirect_count,
request_count,
indirect_count_storage,
compute_indirect_pipeline,
sort_pipeline,
reset_pipeline,
compaction_pipeline,
bindgroup,
device,
}
@@ -542,12 +405,12 @@ where
pub fn request_count_buffer(&self) -> &Buffer
{
&self.request_count_buffer
&self.request_count
}
pub fn sort_buffer(&self) -> &Buffer
pub fn index_buffer(&self) -> &Buffer
{
&self.sort_buffer
&self.index_buffer
}
pub fn reset_requests(&self, encoder: &mut CommandEncoder)
@@ -560,57 +423,46 @@ where
compute_pass.set_bind_group(0, Some(&self.bindgroup), &[]);
compute_pass.set_pipeline(&self.reset_pipeline);
let shader_invocations = self.cache_size; // one invocation per element
let workgroup_invocations = shader_invocations.div_ceil(64);
let shader_invocations = self.cache_size * N * N * N; // one invocation per element
let workgroup_invocations = shader_invocations.div_ceil(256);
compute_pass.dispatch_workgroups(workgroup_invocations as u32, 1, 1);
}
pub fn sort_requests(&self, encoder: &mut CommandEncoder)
{
let mut compaction_compute_pass =
encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("request_buffer_compaction_compute_pass"),
timestamp_writes: None,
});
// Compact list
self.buffer_compactor.compact_buffer(
&self.device,
encoder,
&self.request_buffer,
&self.index_buffer,
);
compaction_compute_pass.set_bind_group(0, Some(&self.bindgroup), &[]);
let request_element_count = self.cache_size; // Each child slot is sorted
let workgroups_invocations = request_element_count.div_ceil(64);
self.buffer_compactor
.copy_count_to_buffer(encoder, &self.request_count);
// = Perform list compaction
let mut compute_indirect_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("compute_indirect_pass"),
timestamp_writes: None,
});
// == Running sum
compaction_compute_pass.set_pipeline(&self.running_sum_pipeline);
compute_indirect_pass.set_bind_group(0, Some(&self.bindgroup), &[]);
// Phase 0: Put ones in correct location
compaction_compute_pass.set_immediates(0, bytes_of(&0));
compaction_compute_pass.dispatch_workgroups(workgroups_invocations as u32, 1, 1);
// Phase _: running sum
// Running sum phase
let running_sum_steps = request_element_count.next_power_of_two().ilog2();
for i in 1..=running_sum_steps
{
compaction_compute_pass.set_immediates(0, bytes_of(&i));
compaction_compute_pass.dispatch_workgroups(workgroups_invocations as u32, 1, 1);
}
compute_indirect_pass.set_pipeline(&self.compute_indirect_pipeline);
compute_indirect_pass.dispatch_workgroups(1, 1, 1);
drop(compute_indirect_pass);
if !running_sum_steps.is_multiple_of(2)
{
// Bring back double buffer
compaction_compute_pass.set_immediates(0, bytes_of(&0xFFFFFFFF_u32));
compaction_compute_pass.dispatch_workgroups(workgroups_invocations as u32, 1, 1);
}
// Now the indirect count storate buffer contains the workgroup count for the sort
// Copy that to the indirect buffer
encoder.copy_buffer_to_buffer(
&self.indirect_count_storage,
0,
&self.indirect_count,
0,
Some(size_of::<wgpu::util::DispatchIndirectArgs>() as u64),
);
// == Stream compaction
compaction_compute_pass.set_pipeline(&self.compaction_pipeline);
compaction_compute_pass.dispatch_workgroups(workgroups_invocations as u32, 1, 1);
drop(compaction_compute_pass);
// == Copy count into indirect buffer
// = Sort
// == Bitonic sort
// Sort
let sort_element_count = self.cache_size * N * N * N;
let phases_upper_bound = sort_element_count.next_power_of_two().ilog2();