Faster coalescing

This commit is contained in:
2026-09-13 15:48:27 +02:00
parent 9015ed85d6
commit 59e3ed5f87
5 changed files with 119 additions and 22 deletions
+22 -12
View File
@@ -62,6 +62,7 @@ use crate::voxel_cache::data::CacheResponse;
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::producer_interface::CacheProducerInterface;
use crate::voxel_cache::producer_interface::CacheRequest;
@@ -142,7 +143,8 @@ impl State
.request_device(&wgpu::DeviceDescriptor {
required_features: Features::IMMEDIATES
| Features::SHADER_EARLY_DEPTH_TEST
| Features::TIMESTAMP_QUERY,
| Features::TIMESTAMP_QUERY
| Features::SHADER_INT64,
required_limits: wgpu::Limits {
max_immediate_size: 112,
max_storage_buffers_per_shader_stage: 16,
@@ -161,22 +163,22 @@ impl State
let egui_renderer = EguiRenderer::new(&device, surface_format, &window);
let mut voxel_cache = VoxelCache::<4>::new(200_000, device.clone(), queue.clone());
let mut voxel_cache = VoxelCache::<4>::new(100_000, device.clone(), queue.clone());
let cache_interface = CacheProducerInterface::new(1024, &device);
let terrain_generator = TerrainGenerator::<4>::new(5, "vxls_height.tif", 0.2, "img.jpg");
//let terrain_generator = TerrainGenerator::<4>::new(5, "vxls_height.tif", 0.2, "img.jpg");
// let terrain_generator = TerrainGenerator::<4>::new(
// 5,
// "./pointe_percee/height.tif",
// 0.2,
// "./pointe_percee/ortho.jpg",
// );
// let terrain_generator = TerrainGenerator::<4>::new(
// 5,
// "/home/albin/Documents/vxls_maps/lapiz/height.tif",
// 0.2,
// "/home/albin/Documents/vxls_maps/lapiz/ortho.jpg",
// );
let terrain_generator = TerrainGenerator::<4>::new(
5,
"/home/albin/Documents/vxls_maps/lapiz/height.tif",
0.2,
"/home/albin/Documents/vxls_maps/lapiz/ortho.jpg",
);
let mut chunk_pos_map = HashMap::new();
let chunk_instances = (0..terrain_generator.chunk_width)
@@ -270,7 +272,7 @@ impl State
depth_stencil: Some(wgpu::DepthStencilState {
format: wgpu::TextureFormat::Depth24PlusStencil8,
depth_write_enabled: Some(true),
depth_compare: Some(wgpu::CompareFunction::Less),
depth_compare: Some(wgpu::CompareFunction::LessEqual),
stencil: wgpu::StencilState::default(),
bias: wgpu::DepthBiasState::default(),
}),
@@ -645,7 +647,7 @@ impl State
let generator = cloned_generator;
let gen_test = SineGenerator::<4>::new(5);
let mut structure_nodes = vec![];
let mut structure_nodes: Vec<StructurePoolElement<4>> = vec![];
let mut color_nodes: Vec<[ColorBytes; 64]> = vec![];
let mut location_nodes = vec![];
@@ -704,7 +706,15 @@ impl State
};
}
(node.structure, node.colors, location)
(
StructurePoolElement {
occupancy_low: 0,
occupancy_high: 0,
pointers: node.structure,
},
node.colors,
location,
)
})
.collect::<Vec<_>>()
.into_iter()
+34 -1
View File
@@ -346,6 +346,8 @@ where
format!("
struct StructurePoolElement
{{
occupancy_low: atomic<u32>,
occupancy_high: atomic<u32>,
pointers: array<u32, {children_count}>
}}
@@ -565,15 +567,37 @@ where
{{
// Phase 1
// Copy into cache page
var occupancy_high = u32(0);
var occupancy_low = u32(0);
for(var i = 0; i < {children_count}; i++)
{{
structure_pool[overwritten_element].pointers[i] = select(u32(0), u32(1)<<31, structure_nodes[index].pointers[i] != 0);
let turn_on_bit =
select(
u32(0),
u32(u32(1) << u32(i % ({children_count} / 2))),
((color_nodes[index].colors[i] >> 24) != 0) ||
structure_nodes[index].pointers[i] != 0
);
if(i > {children_count} / 2)
{{
occupancy_high |= turn_on_bit;
}}else
{{
occupancy_low |= turn_on_bit;
}}
}}
atomicStore(&structure_pool[overwritten_element].occupancy_low, occupancy_low);
atomicStore(&structure_pool[overwritten_element].occupancy_high, occupancy_high);
color_pool[overwritten_element] = color_nodes[index];
location_pool[overwritten_element] = locations[index];
// Mark dirty/correct timestamp
usage_buffer[overwritten_element] = parameters.frame_timestamp + 1;
usage_buffer[overwritten_element] = parameters.frame_timestamp + 1;
}}
if(parameters.write_pointers != 0 && usage_buffer[overwritten_element] != parameters.frame_timestamp)
@@ -588,6 +612,15 @@ where
}}else if usage_buffer[requests_wb[index].node_index] != parameters.frame_timestamp + 1
{{
structure_pool[requests_wb[index].node_index].pointers[requests_wb[index].child_index] = new_pointer;
let turn_on_bit = u32(1 << (requests_wb[index].child_index % ({children_count} / 2)));
if(requests_wb[index].child_index > {children_count} / 2)
{{
//atomicOr(&structure_pool[requests_wb[index].node_index].occupancy_high, turn_on_bit);
}}else
{{
//atomicOr(&structure_pool[requests_wb[index].node_index].occupancy_low, turn_on_bit);
}}
}}
}}
+4 -1
View File
@@ -24,11 +24,14 @@ where
request_count: [u32; N * N * N],
}
#[repr(C)]
pub struct StructurePoolElement<const N: usize>
where
[(); N * N * N]:,
{
pointers: [StructurePointer; N * N * N],
pub occupancy_low: u32,
pub occupancy_high: u32,
pub pointers: [StructurePointer; N * N * N],
}
pub struct DestinationElement