starting compaction stuff

.gitattributes

@@ -0,0 +1,3 @@
+img.jpg filter=lfs diff=lfs merge=lfs -text
+img_low.jpg filter=lfs diff=lfs merge=lfs -text
+vxls_height.tif filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -1,11 +1,7 @@
 /target
 Cargo.lock
 
-img.jpg
-img.png
-img_low.jpg
 imgs.tar.gz
-vxls_height.tif
 
 # Added by cargo
 #

Cargo.toml

@@ -20,5 +20,5 @@ pollster = "1.0.1"
 rand = "0.10.2"
 rayon = "1.12.0"
 tiff = "0.11.3"
-wgpu = "30"
+wgpu = {version = "30", features = ["spirv"]}
 winit = "0.30.13"

shaders/Makefile

@@ -0,0 +1,4 @@
+all: voxel.spv
+
+%.spv: %.slang
+	slangc $< -O3 -fvk-use-entrypoint-name -target spirv -o $@

shaders/compaction.slang

@@ -0,0 +1,124 @@
+
+[[vk::binding(0, 0)]]
+RWStructuredBuffer<uint32_t> count_buffer;
+
+[[vk::binding(0, 1)]]
+RWStructuredBuffer<uint32_t> reduced_buffer;
+[[vk::binding(1, 1)]]
+RWStructuredBuffer<uint32_t> sum_buffer;
+[[vk::binding(2, 1)]]
+RWStructuredBuffer<uint32_t> compaction_buffer;
+
+groupshared uint32_t local_data[256 * 2];
+static uint32_t      THREAD_WIDTH = 256;
+static uint32_t      DATA_WIDTH = THREAD_WIDTH * 2;
+
+[numthreads(256, 1, 1)]
+[shader("compute")]
+void block_sum(
+    uint32_t3 workgroup_id: SV_GroupID,
+    uint32_t3 local_thread_id: SV_GroupThreadID,
+    uint32_t3 global_thread_id: SV_DispatchThreadID)
+{
+    // Perform sum in current block
+
+    // Copy local_datainto LDS with predicate
+    let thread_index = global_thread_id.x;
+    let local_thread_index = local_thread_id.x;
+    let total = count_buffer.getCount();
+
+    if (thread_index * 2 < total)
+    {
+        local_data[local_thread_index * 2] = select(count_buffer[thread_index * 2] != 0, 1, 0);
+    }
+    else
+    {
+        local_data[local_thread_index * 2] = 0;
+    }
+
+    if (thread_index * 2 + 1 < total)
+    {
+        local_data[local_thread_index * 2 + 1] = select(count_buffer[thread_index * 2 + 1] != 0, 1, 0);
+    }
+    else
+    {
+        local_data[local_thread_index * 2 + 1] = 0;
+    }
+
+    GroupMemoryBarrierWithGroupSync();
+
+    var width : uint32_t = 2;
+    while (width <= DATA_WIDTH)
+    {
+        let dest_index = width * (thread_index + 1) - 1;
+        let get_index = dest_index - (width / 2);
+        // println!("{}, {}", get_index, dest_index);
+        if (dest_index < DATA_WIDTH)
+        {
+            local_data[dest_index] += local_data[get_index];
+        }
+        width *= 2;
+        GroupMemoryBarrierWithGroupSync();
+    }
+
+    local_data[DATA_WIDTH - 1] = 0;
+    while (width >= 2)
+    {
+        let dest_index = width * (thread_index + 1) - 1;
+        let get_index = dest_index - (width / 2);
+        // println!("{}, {}", get_index, dest_index);
+        if (dest_index < DATA_WIDTH)
+        {
+            let self_data = local_data[dest_index];
+            local_data[dest_index] += local_data[get_index];
+            local_data[get_index] = self_data;
+        }
+        width /= 2;
+        GroupMemoryBarrierWithGroupSync();
+    }
+
+    // Block now contains running local sum
+    // Dump back to sum buffer
+    sum_buffer[2 * thread_index] = local_data[2 * local_thread_index];
+    sum_buffer[2 * thread_index + 1] = local_data[2 * local_thread_index + 1];
+
+    // Write to reduced buffer
+    reduced_buffer[workgroup_id.x] = local_data[DATA_WIDTH - 1];
+}
+
+[numthreads(1, 1, 1)]
+[shader("compute")]
+void linear_reduced_sum(
+    uint32_t3 workgroup_id: SV_GroupID,
+    uint32_t3 local_thread_id: SV_GroupThreadID,
+    uint32_t3 global_thread_id: SV_DispatchThreadID)
+{
+    let size = reduced_buffer.getCount();
+
+    // Perform exclusive sum
+    var running_sum : uint32_t = 0;
+    for (uint32_t i = 0; i < size; i++)
+    {
+        let value = reduced_buffer[i];
+        reduced_buffer[i] = running_sum;
+        running_sum += value;
+    }
+}
+
+[numthreads(256, 1, 1)]
+[shader("compute")]
+void uniform_add(
+    uint32_t3 workgroup_id: SV_GroupID,
+    uint32_t3 local_thread_id: SV_GroupThreadID,
+    uint32_t3 global_thread_id: SV_DispatchThreadID)
+{
+    let thread_index = global_thread_id.x;
+    let local_thread_index = local_thread_id.x;
+
+    // Gather
+    let reduced_value = reduced_buffer[global_thread_id.x];
+    // Apply
+    sum_buffer[thread_index * 2] += reduced_value;
+    sum_buffer[thread_index * 2 + 1] += reduced_value;
+}
+

shaders/voxel.slang

@@ -0,0 +1,471 @@
+struct PushConstants
+{
+    float4x4 view_proj;
+    float3   cam_pos;
+    uint32_t frame_timestamp;
+    uint32_t width;
+    uint32_t height;
+
+    uint32_t chunk_width;
+    uint32_t chunk_height;
+    uint32_t chunk_alt;
+}
+
+public struct VertexOutput
+{
+    public float4 position : SV_Position;
+
+    [vk::location(0)]
+    public float3 world_position;
+
+    [vk::location(1)]
+    public nointerpolation uint32_t structure_id;
+
+    [vk::location(2)]
+    public float3 cam_position;
+
+    [vk::location(3)]
+    public float3 chunk_position;
+}
+
+[[vk::push_constant]]
+uniform PushConstants constants;
+
+[shader("vertex")]
+VertexOutput chunk(
+    uint                     index: SV_VulkanVertexID,
+    [vk::location(0)] float3 chunk_position,
+    [vk::location(1)] uint   id)
+{
+    let cube_vertices : float3[8] =
+                            float3[](
+                                float3(0., 0., 0.),
+                                float3(0., 0., 1.),
+                                float3(1., 0., 1.),
+                                float3(1., 0., 0.),
+
+                                float3(0., 1., 0.),
+                                float3(0., 1., 1.),
+                                float3(1., 1., 1.),
+                                float3(1., 1., 0.), );
+    // clang-format off
+    let cube_faces: int[24] = int[](
+        // Bottom face
+        1, 0, 2, 3,
+
+        // Top face
+        4, 5, 7, 6,
+
+        // Side faces
+        0, 1, 4, 5,
+        1, 2, 5, 6,
+        2, 3, 6, 7,
+        3, 0, 7, 4,
+    );
+
+    let quad_index = index / (3 * 2);
+    let triangle_index = index % (3 * 2);
+    let triangle_map: int[6] = int[](
+        0, 1, 2, 1, 3, 2
+    );
+
+
+    let vertex = cube_vertices[cube_faces[quad_index * 4 + triangle_map[triangle_index]]];
+    let output_vertex = mul(constants.view_proj, float4(vertex + chunk_position, 1.0f));
+
+    VertexOutput vertex_output;
+    vertex_output.position = output_vertex;
+    vertex_output.world_position = vertex + chunk_position;
+    vertex_output.structure_id = id;
+    vertex_output.cam_position = constants.cam_pos;
+    vertex_output.chunk_position = chunk_position;
+
+    return vertex_output;
+}
+
+struct StructurePointer
+{
+    uint32_t value;
+    bool subdivided()
+    {
+        return (this.value & 0x80000000) != 0;
+    }
+
+    bool pointer_valid()
+    {
+        return (this.value & 0x40000000) != 0;
+    }
+
+    bool subdivided_valid()
+    {
+        return (this.value & 0xC0000000) == 0xC0000000;
+    }
+
+    uint32_t pointer()
+    {
+        return this.value & 0x3FFFFFFF;
+    }
+}
+
+struct ByteColor
+{
+    uint32_t byte_color;
+    
+    property uint32_t byte_r {
+        get {return byte_color & 0xFF;}
+    }
+
+    property uint32_t byte_g {
+        get {return (byte_color >> 8) & 0xFF;}
+    }
+
+    property uint32_t byte_b {
+        get {return (byte_color >> 16) & 0xFF;}
+    }
+
+    property uint32_t byte_a {
+        get {return byte_color >> 24;}
+    }
+
+    property float4 float_color {
+        get {return float4(
+                float(byte_r) / 255.,
+                float(byte_g) / 255.,
+                float(byte_b) / 255.,
+                float(byte_a) / 255.
+        ); }
+    }
+}
+
+struct StructurePoolElement
+{
+    uint32_t occupancy_low;
+    uint32_t occupancy_high;
+    StructurePointer pointers[64];
+}
+
+struct RequestBufferElement
+{
+    Atomic<uint32_t> requests[64];
+}
+
+struct ColorPoolElement
+{
+    ByteColor colors[64];
+}
+
+struct LocationPoolElement
+{
+    uint32_t structure_id;
+    uint32_t structure_locator;
+}
+
+[[vk::binding(0, 0)]] RWStructuredBuffer<StructurePoolElement> structure_pool;
+[[vk::binding(1, 0)]] RWStructuredBuffer<ColorPoolElement> color_pool;
+[[vk::binding(2, 0)]] RWStructuredBuffer<LocationPoolElement> location_pool;
+[[vk::binding(3, 0)]] RWStructuredBuffer<RequestBufferElement> request_buffer;
+[[vk::binding(4, 0)]] RWStructuredBuffer<uint32_t> usage_buffer;
+[[vk::binding(5, 0)]] RWStructuredBuffer<StructurePointer> structure_table_pointer;
+[[vk::binding(6, 0)]] RWStructuredBuffer<Atomic<uint32_t>> structure_table_request_buffer;
+
+uint32_t3 get_children_pos(float3 position, uint32_t scale_exp)
+{
+    return (asuint(position) >> scale_exp) & 3;
+}
+
+uint32_t get_children_index(float3 position, uint32_t scale_exp)
+{
+    // Get mantissa bits for this scale exp an retain bits for the specific children
+    uint32_t3 cell_position = (asuint(position) >> scale_exp) & 3;
+    return cell_position.x + cell_position.y * 4 + cell_position.z * 4 * 4;
+}
+
+uint64_t get_child_mask(uint32_t low, uint32_t high)
+{
+    return ((uint64_t)high << 32) | (uint64_t)low;
+}
+
+float3 floor_scale(float3 position, uint32_t scale_exp)
+{
+    uint32_t mask = ~0u << scale_exp;
+    return asfloat(asuint(position) & mask);
+}
+
+struct HitInformation
+{
+    bool hit;
+    float3 hit_pos;
+    float4 color;
+}
+
+// struct NodeStack
+// {
+//     uint32_t node_stack[5];
+// }
+
+groupshared uint32_t stack[256 * 5];
+
+HitInformation ray_march(float3 ray_direction, float3 ray_origin, uint32_t root_id, float dist_offset, uint32_t stack_index)
+{
+    float fov_deg = 100. / 1920.;
+    float fov_rad = (float.getPi() * fov_deg) / 180.;
+    float cone_factor = tan(fov_rad / 2.) * 2; // Horizontal size of pixel
+
+    let st_pointer = structure_table_pointer[root_id];
+    if(!st_pointer.subdivided())
+    {
+        var hit: HitInformation;
+        hit.hit = false;
+        return hit;
+    }
+
+    if(!st_pointer.pointer_valid())
+    {
+        // Record request
+        structure_table_request_buffer[root_id].add(1);
+        var hit: HitInformation;
+        hit.hit = false;
+        return hit;
+    }
+
+    ray_origin += float3(1.);
+    ray_origin = clamp(ray_origin , float(1.), asfloat(0x3fffffff));
+    ray_origin = select(ray_direction > 0., asfloat(asuint(ray_origin) ^ 0x007fffff), ray_origin);
+    uint32_t child_mirror = 0;
+    if(ray_direction.x > 0.) {child_mirror |= 3;}
+    if(ray_direction.y > 0.) {child_mirror |= 3 << 2;}
+    if(ray_direction.z > 0.) {child_mirror |= 3 << 4;}
+    ray_direction = -abs(ray_direction);
+
+    float3 pos = ray_origin;
+
+    uint32_t scale_exp = 23 - 2;
+    uint32_t node_stack[5] = 
+    {
+        0
+    };
+
+    uint32_t current_node_index = structure_table_pointer[root_id].pointer();
+    usage_buffer[current_node_index] = constants.frame_timestamp;
+    stack[stack_index * 5 + 10 - scale_exp / 2] = current_node_index;
+    //node_stack[10 - scale_exp / 2] = current_node_index;
+
+    [loop]
+    for(uint32_t iter = 0; iter < 500; iter ++)
+    {
+        uint32_t child_index = get_children_index(pos, scale_exp) ^ child_mirror;
+        StructurePointer current_node = structure_pool[current_node_index].pointers[child_index];
+
+        // Scale computations
+        let cone_size = (length(ray_origin - pos) + dist_offset) * cone_factor;
+        let exponent = 
+            select(
+                cone_size == 0.,
+                0,
+                23 - (127 - (asuint(cone_size) >> 23))
+            );
+
+        while(
+            current_node.subdivided_valid() && 
+            scale_exp - 2 > exponent
+            )
+        {
+            scale_exp -= 2;
+            current_node_index = current_node.pointer();
+            stack[stack_index * 5 + 10 - scale_exp / 2] = current_node_index;
+            //node_stack[10 - scale_exp / 2] = current_node_index;
+            child_index = get_children_index(pos, scale_exp) ^ child_mirror;
+            current_node = structure_pool[current_node_index].pointers[child_index];
+            // Write usage
+            usage_buffer[current_node_index] = constants.frame_timestamp;
+        }
+
+        // Request subdiv
+        if(current_node.subdivided() && !current_node.pointer_valid())
+        {
+            request_buffer[current_node_index].requests[child_index].add(1);
+        }
+
+        if(color_pool[current_node_index].colors[child_index].byte_a != 0)
+        {
+            var hit: HitInformation;
+            hit.hit = true;
+            hit.hit_pos = pos - float3(1.);
+            hit.color = color_pool[current_node_index].colors[child_index].float_color;
+            return hit;
+        }
+
+        uint64_t occupancy = get_child_mask(structure_pool[current_node_index].occupancy_low, structure_pool[current_node_index].occupancy_high);
+        uint32_t adv_scale_exp = scale_exp;
+        if(((occupancy >> (child_index & 0b101010)) & 0x00330033) == 0)
+        {
+            adv_scale_exp ++;
+        }
+
+        // Perform dda
+        // Compute correct exponent, and shift it into the exponent part of floatt
+        let child_pos : float3 = floor_scale(pos, adv_scale_exp);
+        // Intersection t
+        let inter_ts : float3 = (child_pos - ray_origin) / ray_direction;
+        float inter_t = min(inter_ts.x, min(inter_ts.y, inter_ts.z));
+        //return float4(inter_t);
+
+        // Perform dda step
+        //let neighbor_max = asint(child_pos) + select(inter_t == inter_ts, -1, (1 << adv_scale_exp) - 1);
+        let neighbor_max = asint(child_pos) + select(inter_t == inter_ts, -1, (1 << adv_scale_exp) - 1);
+        pos = min(ray_origin + ray_direction * inter_t, asfloat(neighbor_max));
+
+        // Find most common ancestor
+        uint32_t3 diffs = asuint(child_pos) ^ asuint(pos);
+        uint32_t diff = (diffs.x | diffs.y | diffs.z);
+        
+        int32_t common_depth = (1 + (22 - firstbithigh(diff)) / 2) * 2;
+        if(common_depth <= 0)
+        {
+            break;
+        }
+
+        scale_exp = 23 - common_depth;
+        current_node_index = stack[stack_index * 5 + 10 - scale_exp / 2];
+        //current_node_index = node_stack[10 - scale_exp / 2];
+    }
+
+
+    var hit: HitInformation;
+    hit.hit = false;
+    return hit;
+}
+
+struct FragmentOutput
+{
+    float depth : SV_Depth;
+    float4 color : SV_Target<0>;
+}
+
+/*
+//[earlydepthstencil]
+[shader("fragment")]
+FragmentOutput fragment(VertexOutput vertex_out)
+{
+    let ray_direction = normalize(vertex_out.world_position - vertex_out.cam_position);
+    let intersection_t = box_intersect(vertex_out.cam_position, ray_direction, vertex_out.chunk_position, vertex_out.chunk_position + float3(1.));
+    let local_ray_origin = max(intersection_t.x, 0.) * ray_direction + vertex_out.cam_position - vertex_out.chunk_position;
+    // Figure out intersection
+    let hit = ray_march(ray_direction, local_ray_origin, vertex_out.structure_id, max(0., intersection_t.x));
+    let world_hit_pos = hit.hit_pos + vertex_out.chunk_position;
+    let clip = mul(constants.view_proj, float4(world_hit_pos, 1.));
+    let depth = clip.z / clip.w;
+
+    var frag_out : FragmentOutput;
+    frag_out.depth = depth;
+    frag_out.color = hit.color;
+
+    return frag_out;
+}
+*/
+
+bool3 min_mask(float3 val)
+{
+    let min_val = min(val.x, min(val.y, val.z));
+    return val == min_val;
+}
+
+[[vk::binding(0, 1)]] 
+[[format("rgba32f")]]
+WTexture2D<float4> output_texture;
+
+float3 get_ray_direction(uint32_t2 pixel_loc)
+{
+    let ndc_loc_x = (float)pixel_loc.x / (float)constants.width * 2. - 1.;
+    let ndc_loc_y = 1. - (float)pixel_loc.y / (float)constants.height * 2.;
+    var world_loc = mul(constants.view_proj, float4(ndc_loc_x, ndc_loc_y, 1., 1.));
+    world_loc /= world_loc.w;
+    return normalize(world_loc.xyz - constants.cam_pos);
+}
+
+[shader("compute")]
+[numthreads(16, 16, 1)]
+void ray_march_compute(uint32_t3 location : SV_DispatchThreadID, uint32_t3 local_location: SV_GroupThreadID)
+{
+    if(location.x >= constants.width || location.y >= constants.height)
+    {
+        return;
+    }
+
+    let stack_index = local_location.x + local_location.y * 16;
+    let ray_direction = get_ray_direction(location.xy);
+    var inter = box_intersect(constants.cam_pos, ray_direction, float3(0.), float3(constants.chunk_width, constants.chunk_alt, constants.chunk_height));
+
+    // Clear if out of box
+    if(inter.y <= inter.x || inter.y <= 0.)
+    {
+        output_texture.Store(location.xy, float4(0.));
+        return;
+    }
+    inter.x = max(0., inter.x);
+    // float3 position = constants.cam_pos + ray_direction * inter.x;
+    // int32_t3 current_voxel = clamp(
+    //     int32_t3(floor(position)), 
+    //     int32_t3(0),
+    //     int32_t3(constants.chunk_width - 1, constants.chunk_alt - 1, constants.chunk_height - 1)
+    // );
+    // let chunk_index = current_voxel.y + current_voxel.z * constants.chunk_alt + current_voxel.x * constants.chunk_alt * constants.chunk_height;
+    // let hit = ray_march(ray_direction, position - float3(current_voxel), chunk_index, length(position - constants.cam_pos), stack_index);
+    // if(hit.hit)
+    // {
+    //     output_texture.Store(location.xy, float4(hit.color));
+    // }
+
+
+    let start_position = inter.x * ray_direction + constants.cam_pos;
+    // FVT
+    int32_t3 current_voxel = clamp(
+        int32_t3(floor(start_position)), 
+        int32_t3(0),
+        int32_t3(constants.chunk_width - 1, constants.chunk_alt - 1, constants.chunk_height - 1)
+    );
+    //int32_t3 offset = int32_t3(sign(ray_direction));
+    //float3 delta = abs(1. / ray_direction);
+    float3 t = select(ray_direction > 0., current_voxel + int32_t3(1) - start_position, start_position - current_voxel) / abs(ray_direction);
+    t += inter.x;
+
+    [loop]
+    while(true)
+    {
+        {
+            let off = t - abs(1. / ray_direction);
+            let t_adv = max(max(off.x, max(off.y, off.z)), 0.);
+            let chunk_index = current_voxel.y + current_voxel.z * constants.chunk_alt + current_voxel.x * constants.chunk_alt * constants.chunk_height;
+            let position = constants.cam_pos + ray_direction * t_adv;
+            let hit = ray_march(ray_direction, position - float3(current_voxel), chunk_index, t_adv, stack_index);
+            if(hit.hit)
+            {
+                output_texture.Store(location.xy, hit.color);
+                return;
+            }
+        }
+
+        let min_mask = min_mask(t);    
+        t += select(min_mask, abs(1. / ray_direction), float3(0.));
+        current_voxel += select(min_mask, int32_t3(sign(ray_direction)), int32_t3(0));
+
+        if(any(current_voxel < 0) || any(current_voxel >= int32_t3(constants.chunk_width, constants.chunk_alt, constants.chunk_height)))
+        {
+            output_texture.Store(location.xy, float4(0.));
+            return;
+        }
+    }
+}
+
+float2 box_intersect(float3 origin, float3 ray_direction, float3 box_min, float3 box_max)
+{
+    let min_ts = (box_min - origin) / ray_direction;
+    let max_ts = (box_max - origin) / ray_direction;
+
+    let far_ts = max(min_ts, max_ts);
+    let near_ts = min(min_ts, max_ts);
+
+    let far_t = min(far_ts.x, min(far_ts.y, far_ts.z));
+    let near_t = max(near_ts.x, max(near_ts.y, near_ts.z));
+    return float2(near_t, far_t);
+}

shaders/voxel.temp.wgsl

@@ -0,0 +1,399 @@
+struct VertexOutput
+{
+    @builtin(position) postion: vec4<f32>,
+    @location(0) @interpolate(flat) chunk_index: u32,
+    @location(1) color: vec4<f32>,
+    @location(2) cam_pos: vec3<f32>,
+    @location(3) world_pos: vec3<f32>,
+    @location(4) @interpolate(flat) structure_id: u32,
+    @location(5) chunk_position: vec3<f32>
+}
+
+struct ChunkImmediate
+{
+    view_proj: mat4x4<f32>,
+    cam_pos: vec3<f32>,
+    frame_timestamp: u32,
+}
+
+var<immediate> constants: ChunkImmediate;
+//var<push_constant> constants: ChunkInfo;
+
+struct CacheChunkObject
+{
+    transform: mat4x4<f32>,
+    color: vec4<f32>,
+    id: u32,
+    pointer: u32
+}
+
+
+struct StructurePoolElement
+{
+    pointers: array<u32, 64>
+}
+
+struct RequestBufferElement
+{
+    requests: array<atomic<u32>, 64>
+}
+
+struct ColorPoolElement
+{
+    colors: array<u32, 64>
+}
+
+struct LocationPoolElement
+{
+    structure_id: u32,
+    structure_locator: u32
+}
+
+struct SortedRequestsElement
+{
+    node: u32,
+    child: u32
+}
+
+fn unpack_color(color: u32) -> vec4<f32>
+{
+    return vec4<f32>(
+            f32(color & 0xFF) / 255.,
+            f32((color >> 8) & 0xFF) / 255.,
+            f32((color >> 16) & 0xFF) / 255.,
+        f32((color >> 24) & 0xFF) / 255.
+    );
+}
+
+@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_write> request_buffer: array<RequestBufferElement>;
+@group(0) @binding(4) var<storage, read_write> usage_buffer: array<atomic<u32>>;
+@group(0) @binding(5) var<storage, read_write> structure_table_pointer: array<u32>;
+@group(0) @binding(6) var<storage, read_write> structure_table_request_buffer: array<atomic<u32>>;
+
+struct FragmentOutput {
+    @location(0) color: vec4<f32>,
+    @builtin(frag_depth) depth: f32, // Equivalent to gl_FragDepth
+}
+
+@vertex
+fn chunk(@builtin(vertex_index) index: u32, @location(0) position: vec3<f32>, @location(1) id: u32) -> @builtin(position) vec4<f32>
+{
+    let cube_vertices = array<vec3<f32>, 8>(
+        vec3<f32>(0., 0., 0.),
+        vec3<f32>(0., 0., 1.),
+        vec3<f32>(1., 0., 1.),
+        vec3<f32>(1., 0., 0.),
+
+        vec3<f32>(0., 1., 0.),
+        vec3<f32>(0., 1., 1.),
+        vec3<f32>(1., 1., 1.),
+        vec3<f32>(1., 1., 0.),
+    );
+
+    let cube_faces = array<u32, 24>(
+        // Bottom face
+        1, 0, 2, 3,
+
+        // Top face
+        4, 5, 7, 6,
+
+        // Side faces
+        0, 1, 4, 5,
+        1, 2, 5, 6,
+        2, 3, 6, 7,
+        3, 0, 7, 4,
+    );
+
+    let quad_index = index / (3 * 2);
+    let triangle_index = index % (3 * 2);
+    let triangle_map = array<u32, 6>(
+        0, 1, 2, 1, 3, 2
+    );
+
+
+    let vertex = cube_vertices[cube_faces[quad_index * 4 + triangle_map[triangle_index]]];
+    let output_vertex = constants.view_proj * vec4<f32>(vertex + position, 1.0f);
+
+    return output_vertex;
+}
+
+
+struct StructureElement
+{
+    children: array<u32, 64> 
+}
+
+struct ColorElement 
+{
+    children: array<vec4<f32>, 64> 
+}
+
+struct LocationElement 
+{
+    children: array<vec4<f32>, 64> 
+}
+
+struct RequestElement
+{
+    children: array<atomic<u32>, 64> 
+}
+
+fn box_inter(pos: vec3<f32>, ray_dir: vec3<f32>, box_min: vec3<f32>, box_max: vec3<f32>) -> vec2<f32>
+{
+    let box_min_t = (box_min - pos) / ray_dir;
+    let box_max_t = (box_max - pos) / ray_dir;
+
+    let near_ts = min(box_min_t, box_max_t);
+    let far_ts = max(box_min_t, box_max_t);
+
+    let far_t = min(min(far_ts.x, far_ts.y), far_ts.z);
+    let near_t = max(max(near_ts.x, near_ts.y), near_ts.z);
+
+    return vec2(near_t, far_t);
+}
+
+fn sdf(voxel: vec3<i32>) -> bool
+{
+    let len = length(vec3<f32>(voxel) - vec3(128)) / 128.;
+    return len <= 1.;
+}
+
+fn min_vec(x: vec3<f32>) -> f32
+{
+    return min(x.x, min(x.y, x.z));
+}
+
+fn min_mask(x: vec3<f32>) -> vec3<bool>
+{
+    let min = min(x.x, min(x.y, x.z));
+
+    return vec3<bool>(min == x.x, min == x.y, min == x.z);
+}
+
+fn node_subdivided(node: u32) -> bool
+{
+    return ((node >> 31) & 1) != 0;
+}
+
+fn node_pointer_valid(node: u32) -> bool
+{
+    return ((node >> 30) & 1) != 0;
+}
+
+fn node_pointer(node: u32) -> u32
+{
+    return node & 0x3FFFFFFF;
+}
+
+fn voxel_from_wall(position: vec3<f32>, ray_dir: vec3<f32>) -> vec3<i32>
+{
+    let integers = round(position);
+    let wall_mask = min_mask(abs(position - vec3<f32>(integers)));
+    let offsets = select(vec3<f32>(-0.5), vec3<f32>(0.5), ray_dir > vec3(0.));
+    return vec3<i32>(floor(position + select(vec3<f32>(0.), offsets, wall_mask)));
+}
+
+struct HitResult
+{
+    color: vec4<f32>,
+    hit_pos: vec3<f32>
+}
+
+fn new_traverse(ray_dir: vec3<f32>, ray_origin: vec3<f32>, root_id: u32, dist_offset: f32) -> HitResult
+{
+    let max_depth = 5;
+    let dist_offset_voxel = dist_offset * f32(1 << u32(max_depth * 2));
+    let fovy_deg = 100. / 1920.;
+    let fovy_rad = (fovy_deg * 3.14) / 180.;
+    let cone_factor = tan(fovy_rad / 2.) * 2.;
+
+    let st_pointer = structure_table_pointer[root_id];
+
+    
+    if (!node_subdivided(st_pointer))
+    {
+        discard;
+        var result: HitResult;
+        result.color = vec4(0., 1., 0., 1.);
+        result.hit_pos = ray_origin;
+        return result;
+    }
+    if(!node_pointer_valid(st_pointer))
+    {
+        // Node is subdivided, but not valid
+        // Send request on structure table
+        atomicAdd(&structure_table_request_buffer[root_id], 1); 
+
+        discard;
+        var result: HitResult;
+        result.color = vec4(0., 1., 0., 1.);
+        result.hit_pos = ray_origin;
+        return result;
+    }
+    //var current_node = node_pointer(st_pointer);
+
+    var dfs_stack = array<u32, 6>(node_pointer(st_pointer), 0, 0, 0, 0, 0);
+    var current_depth = 0;
+    var current_node = dfs_stack[current_depth];
+
+    usage_buffer[current_node] = constants.frame_timestamp;
+
+    // Start location
+    //let voxel_dir = select(vec3(-1), vec3(1), ray_dir >= vec3(0.));
+    var node_shift = (max_depth - current_depth) * 2;
+
+    var child_size = 1 << u32(node_shift - 2);
+    var node_size = 1 << u32(node_shift);
+
+    var pos_origin = clamp(ray_origin * f32(1 << u32(max_depth * 2)), vec3(0.), vec3(f32(node_size) - 1.));
+    var voxel = vec3<i32>(pos_origin);
+    var far_t = 0.;
+    var inv_ray_dir = 1. / ray_dir;
+    var ray_positive = ray_dir > vec3(0.);
+    var step_dir = select(vec3(-1), vec3(1), ray_positive);
+
+    for(var iter = 0; iter < 400; iter ++)
+    {
+        // Shift into voxel position
+        node_shift = (max_depth - current_depth) * 2;
+        child_size = 1 << u32(node_shift - 2);
+
+        // Compute child position position from voxel position
+        var child_pos = (voxel >> vec3(u32(node_shift - 2))) & vec3(3);
+        // Compute child index in pointers
+        var child_index = child_pos.x + child_pos.y * 4 + child_pos.z * 4 * 4;
+        // Candidate child pointer
+        var pointer = structure_pool[current_node].pointers[child_index];
+        
+        // Descent loop
+        let min_child_size = (length(vec3<f32>(voxel) - pos_origin) + dist_offset_voxel) * cone_factor;
+        while(node_subdivided(pointer) && node_pointer_valid(pointer) &&
+            f32(child_size / 4) >= min_child_size
+            )
+        {
+
+            // Descend 
+            current_depth += 1;
+
+            // Try to descend again
+            node_shift = (max_depth - current_depth) * 2;
+            child_size = 1 << u32(node_shift - 2);
+            child_pos = (voxel >> vec3(u32(node_shift - 2))) & vec3(3);
+            current_node = node_pointer(pointer);
+            dfs_stack[current_depth] = current_node;
+            child_index = child_pos.x + child_pos.y * 4 + child_pos.z * 4 * 4;
+
+            pointer = structure_pool[current_node].pointers[child_index];
+
+            // Record usage in usage buffer
+            usage_buffer[current_node] = constants.frame_timestamp;
+        }
+
+        // If we could not descencd, request the child
+        if(node_subdivided(pointer) && !node_pointer_valid(pointer) && 
+            f32(child_size / 4) >= min_child_size)
+        {
+            // Record request
+            atomicAdd(&request_buffer[dfs_stack[current_depth]].requests[child_index], 1);
+        }
+
+        // Check color
+        let color = color_pool[current_node].colors[child_index];
+        if(((color >> 24) & 0xFF) != 0)
+        {
+            var result: HitResult;
+            result.color = unpack_color(color);
+            result.hit_pos = (far_t / f32(1 << u32(max_depth * 2))) * ray_dir + ray_origin;
+            return result;
+        }
+
+        // Advance
+        child_pos = voxel & vec3(i32(0xFFFFFFFF << u32(node_shift - 2)));
+        let far_wall = child_pos + select(vec3(0), vec3(child_size), ray_positive);
+        let far_wall_inter = (vec3<f32>(far_wall) - pos_origin) * inv_ray_dir;
+        far_t = min(min(far_wall_inter.x, far_wall_inter.y), far_wall_inter.z);
+
+        // Perform dda step on the children scale
+        let next_child = select(child_pos, child_pos + step_dir * vec3(child_size), vec3(far_t) == far_wall_inter);
+        
+        let previous_voxel = voxel;
+        voxel = clamp(vec3<i32>(pos_origin + far_t * ray_dir), next_child, next_child + vec3(child_size) - vec3(1));
+
+        if any(voxel < vec3(0)) || any(voxel >= vec3(1 << u32((max_depth * 2))))
+        {
+            discard;
+        }
+        
+        // We touched a voxel as if we explored blocks sized by the child size of the current node.
+        // But we might have exited the current node.
+        
+        // If this is the case we have to walk back up the tree
+        // And then back down to the next node over
+
+        // As such we find the lowest ancestor that can contain both the privous voxel (in node) and the new voxel (out of node)
+        let bit_diffs = voxel ^ previous_voxel;
+        let bit_diffs_lowest = bit_diffs.x | bit_diffs.y | bit_diffs.z;
+
+        let common_depth = ((countLeadingZeros(bit_diffs_lowest) - i32(32 - max_depth * 2)) / 2);
+
+        current_depth = common_depth;
+        current_node = dfs_stack[current_depth];
+    }
+
+    // Iter max color
+    var result: HitResult;
+    result.color = vec4(1., 0., 1., 1.);
+    result.hit_pos = (far_t / f32(1 << u32(max_depth * 2))) * ray_dir + ray_origin;
+    return result;
+}
+
+@fragment
+fn fragment() -> @location(0) vec4<f32>
+{
+    return vec4(1., 0., 0., 1.) ;
+}
+
+@early_depth_test(less_equal)
+@fragment
+fn _fragment(in: VertexOutput) -> FragmentOutput
+{
+    //frag_out.color = vec4<f32>(2 * 0.01 / (100. + 0.01 - depth * (100. - 0.01)));
+    let ray_dir = normalize(in.world_pos - in.cam_pos);
+    let interp = box_inter(in.cam_pos - in.chunk_position, ray_dir, vec3(0.), vec3(1));
+    let ray_origin = (in.cam_pos - in.chunk_position) + ray_dir * (max(0., interp.x));
+
+
+    let result = new_traverse(ray_dir, ray_origin, in.structure_id, length(in.cam_pos - (ray_origin + in.chunk_position)));
+    let clip_pos = constants.view_proj * vec4(result.hit_pos + in.chunk_position, 1.);
+    let depth = clip_pos.z / clip_pos.w;
+    var frag_out: FragmentOutput;
+    //frag_out.color = result.color;
+    frag_out.color = result.color;
+    frag_out.depth = depth;
+    return frag_out;
+
+    //return vec4<f32>(ray_origin, 1.);
+    //return frag_out;
+    //return vec4(interp.y / 10.);
+}
+
+/*
+@fragment
+fn fragment(in: VertexOutput) -> @location(0) vec4<f32>
+{
+    let st = structure_table_pointer[0];
+    let subdivided = ((st >> 31) & 1) != 0;
+    let pointer_valid = ((st >> 30) & 1) != 0;
+    // Request stuff
+    atomicAdd(&structure_table_request_buffer[0], 1); 
+    if(subdivided && !pointer_valid)
+    {
+        return vec4(0., 1., 0., 1.);
+    }
+    return vec4(1., 0., 0., 1.);
+}
+*/
+

shaders/voxel.wgsl

@@ -225,6 +225,7 @@ fn new_traverse(ray_dir: vec3<f32>, ray_origin: vec3<f32>, root_id: u32, dist_of
     
     if (!node_subdivided(st_pointer))
     {
+        discard;
         var result: HitResult;
         result.color = vec4(0., 1., 0., 1.);
         result.hit_pos = ray_origin;
@@ -236,6 +237,7 @@ fn new_traverse(ray_dir: vec3<f32>, ray_origin: vec3<f32>, root_id: u32, dist_of
         // Send request on structure table
         atomicAdd(&structure_table_request_buffer[root_id], 1); 
 
+        discard;
         var result: HitResult;
         result.color = vec4(0., 1., 0., 1.);
         result.hit_pos = ray_origin;
@@ -243,11 +245,11 @@ fn new_traverse(ray_dir: vec3<f32>, ray_origin: vec3<f32>, root_id: u32, dist_of
     }
     //var current_node = node_pointer(st_pointer);
 
-    // Record usage
     var dfs_stack = array<u32, 6>(node_pointer(st_pointer), 0, 0, 0, 0, 0);
     var current_depth = 0;
+    var current_node = dfs_stack[current_depth];
 
-    usage_buffer[dfs_stack[current_depth]] = constants.frame_timestamp;
+    usage_buffer[current_node] = constants.frame_timestamp;
 
     // Start location
     //let voxel_dir = select(vec3(-1), vec3(1), ray_dir >= vec3(0.));
@@ -270,7 +272,8 @@ fn new_traverse(ray_dir: vec3<f32>, ray_origin: vec3<f32>, root_id: u32, dist_of
         child_size = 1 << u32(node_shift - 2);
 
         var child_pos = (voxel >> vec3(u32(node_shift - 2))) & vec3(3);
-        var pointer = structure_pool[dfs_stack[current_depth]].pointers[child_pos.x + child_pos.y * 4 + child_pos.z * 4 * 4];
+        var child_index = child_pos.x + child_pos.y * 4 + child_pos.z * 4 * 4;
+        var pointer = structure_pool[current_node].pointers[child_index];
         
         let min_child_size = (length(vec3<f32>(voxel) - pos_origin) + dist_offset_voxel) * cone_factor;
         while(node_subdivided(pointer) && 
@@ -280,7 +283,7 @@ fn new_traverse(ray_dir: vec3<f32>, ray_origin: vec3<f32>, root_id: u32, dist_of
             if(!node_pointer_valid(pointer) && node_subdivided(pointer))
             {
                 // Record request
-                atomicAdd(&request_buffer[dfs_stack[current_depth]].requests[child_pos.x + child_pos.y * 4 + child_pos.z * 4 * 4], 1);
+                atomicAdd(&request_buffer[dfs_stack[current_depth]].requests[child_index], 1);
                 break;
             }
 
@@ -290,17 +293,19 @@ fn new_traverse(ray_dir: vec3<f32>, ray_origin: vec3<f32>, root_id: u32, dist_of
             node_shift = (max_depth - current_depth) * 2;
             child_size = 1 << u32(node_shift - 2);
             child_pos = (voxel >> vec3(u32(node_shift - 2))) & vec3(3);
-            dfs_stack[current_depth] = node_pointer(pointer);
+            current_node = node_pointer(pointer);
+            dfs_stack[current_depth] = current_node;
+            child_index = child_pos.x + child_pos.y * 4 + child_pos.z * 4 * 4;
 
-            pointer = structure_pool[dfs_stack[current_depth]].pointers[child_pos.x + child_pos.y * 4 + child_pos.z * 4 * 4];
+            pointer = structure_pool[current_node].pointers[child_index];
 
             // Record usage
-            usage_buffer[dfs_stack[current_depth]] = constants.frame_timestamp;
+            usage_buffer[current_node] = constants.frame_timestamp;
         }
 
 
         // Check color
-        let color = color_pool[dfs_stack[current_depth]].colors[child_pos.x + child_pos.y * 4 + child_pos.z * 4 * 4];
+        let color = color_pool[current_node].colors[child_index];
         if(((color >> 24) & 0xFF) != 0)
         {
             var result: HitResult;
@@ -340,7 +345,7 @@ fn new_traverse(ray_dir: vec3<f32>, ray_origin: vec3<f32>, root_id: u32, dist_of
         let common_depth = ((countLeadingZeros(bit_diffs_lowest) - i32(32 - max_depth * 2)) / 2);
 
         current_depth = common_depth;
-        //current_node = dfs_stack[current_depth];
+        current_node = dfs_stack[current_depth];
     }
 
     // Iter max color
@@ -350,204 +355,6 @@ fn new_traverse(ray_dir: vec3<f32>, ray_origin: vec3<f32>, root_id: u32, dist_of
     return result;
 }
 
-fn traverse(ray_dir: vec3<f32>, ray_origin: vec3<f32>, root_id: u32, dist_offset: f32) -> vec4<f32>
-{
-    let st_pointer = structure_table_pointer[root_id];
-
-    if (!node_subdivided(st_pointer))
-    {
-        return vec4(0., 1., 0., 1.);
-    }
-    if(!node_pointer_valid(st_pointer))
-    {
-        atomicAdd(&structure_table_request_buffer[root_id], 1); 
-        return vec4(0., 1., 0., 1.);
-    }
-
-    let fovy_deg = 100.;
-    let fovy = 3.14159 * (fovy_deg / 180.);
-    let definition = 1920.;
-    let cone_fovy = fovy / definition; 
-
-    let factor = 1.;
-    let cone_size_factor = 2. * tan(cone_fovy) * factor;
-
-    // Current depth of the node we are exploring
-    var current_depth = 0;
-
-    // Index of the current node's data
-    var current_node = u32(st_pointer & 0x3FFFFFFF);
-    usage_buffer[current_node] = constants.frame_timestamp;
-    var dfs_stack = array<u32, 6>(current_node, 0, 0, 0, 0, 0);
-
-
-
-    // Lut of the node_size per depth
-    var node_size_lut = array<i32, 6>(
-            4 * 4 * 4 * 4 * 4,
-            4 * 4 * 4 * 4,
-            4 * 4 * 4,
-            4 * 4,
-            4,
-            1,
-            );
-
-    let local_dist_offset = dist_offset * f32(node_size_lut[0]);
-
-
-    // Current node size
-    var node_size = node_size_lut[0]; // 128
-
-    // Size of a child of this node
-    var child_size = node_size / 4;
-
-    // Simple FVT
-    let t_off = abs(1. / ray_dir);
-
-    // Start location
-    let voxel_dir = select(vec3(-1), vec3(1), ray_dir >= vec3(0.));
-    var pos_origin = clamp(ray_origin * f32(node_size), vec3(0.), vec3(f32(node_size) - 1.));
-    var voxel = vec3<i32>(pos_origin);
-    var last_voxel = voxel;
-
-    let wall_offset = select(vec3(0), vec3(1), ray_dir > vec3(0.));
-
-    let max_depth = u32(5);
-    var adaptive_depth = i32(max_depth);
-    var far_t = 0.;
-
-    let ray_dir_inv = 1. / ray_dir;
-    let fma_offset = - pos_origin * ray_dir_inv;
-
-    //let depth_limit = 3;
-    for(var iter = 0; iter < 400; iter ++)
-    {
-
-        // Our ray is currently touching a voxel.
-        // Descend to the lowest node that contains this voxel
-
-        // Position of the child we are in
-        var child_pos = (vec3<u32>(voxel) >> vec3<u32>((max_depth - u32(current_depth + 1)) * 2)) & vec3<u32>(3); // Hardcode for 4-tree
-        var child_index = child_pos.x + child_pos.y * 4 + child_pos.z * 4 * 4;
-
-        // Current node has been used. report
-        usage_buffer[current_node] = constants.frame_timestamp;
-        
-        while(
-            node_subdivided(structure_pool[current_node].pointers[child_index]) &&
-            (local_dist_offset + far_t) * cone_size_factor < f32(node_size_lut[current_depth])
-            )
-        {
-            if(!node_pointer_valid(structure_pool[current_node].pointers[child_index]))
-            {
-                atomicAdd(&request_buffer[current_node].requests[child_index], 1);
-                break;
-            }
-            // Child node is subdivided, we go in, save position in stack
-            current_node = node_pointer(structure_pool[current_node].pointers[child_index]);
-            usage_buffer[current_node] = constants.frame_timestamp;
-            current_depth += 1;
-            dfs_stack[current_depth] = current_node;
-            node_size = node_size_lut[current_depth];
-
-            child_pos = (vec3<u32>(voxel) >> vec3<u32>((max_depth - u32(current_depth + 1)) * 2)) & vec3<u32>(3); // Hardcode for 4-tree
-            child_index = child_pos.x + child_pos.y * 4 + child_pos.z * 4 * 4;
-            child_size = node_size / 4;
-        }
-        usage_buffer[current_node] = constants.frame_timestamp;
-
-
-
-
-        // At this point current_depth is the depth of the node that contains the voxel
-        // child_pos and child_index relate to the specific child of the node that contains this voxel
-        // It is guaranteed that the child is leave
-
-        // Check current leave's color
-        let color = unpack_color(color_pool[current_node].colors[child_index]);
-        if(color.w != 0.) // Not transparent
-        {
-            /*
-            let k = child_pos.x + child_pos.y + child_pos.z;
-            let w = voxel.x + voxel.y + voxel.z;
-            let x = select(0.5, 1., k % 2 == 0) * select(0.8, 1., w % 2 == 0);
-
-            var div = 1;
-            var overlay = 1.;
-            for(var i = 1; i <= 5; i++)
-            {
-                let x = (voxel.x / div + voxel.y / div + voxel.z / div) % 2 == 0;
-                overlay -= select(0., 1. / (f32(i) * 2.5), x);
-                div *= 4;
-            }
-            */
-
-            return color;
-            //return overlay * color;
-        }
-
-        // Voxel and whole child containing it is empty
-        
-        // Perform a step through the children of the node
-        let child_position = (voxel / child_size) * child_size;
-        let far_corner = child_position + wall_offset * child_size;
-        //let far_ts = (vec3<f32>(far_corner) - pos_origin) / ray_dir; // TODO: Turn into fma
-        let far_ts = fma(vec3<f32>(far_corner), ray_dir_inv, fma_offset);
-        far_t = min(min(far_ts.x, far_ts.y), far_ts.z);
-
-        let next_child_min = select(child_position, child_position + voxel_dir * child_size, vec3(far_t) == far_ts);
-        let next_child_max = next_child_min + vec3(child_size) - vec3(1);
-        
-        // The ray (far_t) is now touching the new child to explore
-        // Find out which actual voxel we are touching
-        let previous_voxel = voxel;
-        let float_voxel = clamp(vec3<i32>(pos_origin + far_t * ray_dir), next_child_min, next_child_max);
-        /*
-        voxel = vec3<i32>(
-            floor(
-            select(
-                float_voxel - vec3(0.5),
-                float_voxel + vec3(0.5),
-                ray_dir > vec3(0.)
-            ))
-        );
-        */
-        //voxel = vec3<i32>(round(float_voxel));
-        //voxel = voxel_from_wall(float_voxel, ray_dir);
-        //voxel = voxel_from_wall(float_voxel, ray_dir);
-        voxel = float_voxel;
-        if(any(voxel < vec3(0)) || any(voxel >= vec3(node_size_lut[0])))
-        {
-            //return vec4(f32(iter) / 100.);
-            discard;
-        }
-
-        // We touched a voxel as if we explored blocks sized by the child size of the current node.
-        // But we might have exited the current node.
-        
-        // If this is the case we have to walk back up the tree
-        // And then back down to the next node over
-
-        // As such we find the lowest ancestor that can contain both the privous voxel (in node) and the new voxel (out of node)
-        let bit_diffs = voxel ^ previous_voxel;
-        let bit_diffs_lowest = bit_diffs.x | bit_diffs.y | bit_diffs.z;
-
-        let flb = ((countLeadingZeros(bit_diffs_lowest) - i32(32 - max_depth * 2)) / 2);
-        let common_depth = flb;
-
-        current_depth = common_depth;
-        node_size = node_size_lut[current_depth];
-        child_size = node_size / 4;
-        current_node = dfs_stack[current_depth];
-        
-        // Figure out current voxel position
-        //voxel = vec3<i32>(ray_origin + ray_dir * t); 
-
-    }
-    return vec4<f32>(1., 0., 1., 1.);
-
-}
-
 @early_depth_test(less_equal)
 @fragment
 fn fragment(in: VertexOutput) -> FragmentOutput

src/host_production.rs

@@ -0,0 +1,15 @@
+// Contains useful facilities to render data streamed in from the host
+
+// Can produce a voxel given
+// - Its depth
+// - Its position within the chunk
+// - The chunks position
+pub trait ChunkVoxelProducer
+{
+    fn produce_voxel(
+        &mut self,
+        depth: usize,
+        chunk_position: (usize, usize, usize),
+        voxel_position: (usize, usize, usize),
+    );
+}

src/main.rs

@@ -2,6 +2,7 @@
 #![feature(float_algebraic)]
 
 use std::collections::HashMap;
+use std::ops::Div;
 use std::sync::Arc;
 use std::sync::mpsc::sync_channel;
 
@@ -19,8 +20,10 @@ use itertools::Itertools;
 use rayon::iter::IndexedParallelIterator;
 use rayon::iter::IntoParallelRefIterator;
 use rayon::iter::ParallelIterator;
+use wgpu::BindGroupLayout;
 use wgpu::Buffer;
 use wgpu::BufferUsages;
+use wgpu::ComputePipeline;
 use wgpu::Device;
 use wgpu::Extent3d;
 use wgpu::Features;
@@ -28,10 +31,14 @@ use wgpu::InstanceDescriptor;
 use wgpu::InstanceFlags;
 use wgpu::MemoryBudgetThresholds;
 use wgpu::Operations;
+use wgpu::Origin3d;
 use wgpu::RenderPipeline;
+use wgpu::ShaderStages;
 use wgpu::Texture;
 use wgpu::TextureUsages;
 use wgpu::TextureView;
+use wgpu::include_spirv;
+use wgpu::include_wgsl;
 use wgpu::util::BufferInitDescriptor;
 use wgpu::util::DeviceExt;
 use wgpu::util::DownloadBuffer;
@@ -44,6 +51,7 @@ use winit::event_loop::ActiveEventLoop;
 use winit::event_loop::ControlFlow;
 use winit::event_loop::EventLoop;
 use winit::event_loop::OwnedDisplayHandle;
+use winit::platform::x11::EventLoopBuilderExtX11;
 use winit::window::Window;
 use winit::window::WindowId;
 
@@ -59,6 +67,7 @@ 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;
 
@@ -78,10 +87,14 @@ struct State
     size: winit::dpi::PhysicalSize<u32>,
     surface: wgpu::Surface<'static>,
     depth_buffer: (wgpu::Texture, wgpu::TextureView),
+    target_texture: (wgpu::Texture, wgpu::TextureView),
     surface_format: wgpu::TextureFormat,
+    target_blitter: wgpu::util::TextureBlitter,
     egui_renderer: EguiRenderer,
 
-    pipeline: RenderPipeline,
+    //pipeline: RenderPipeline,
+    pipeline: ComputePipeline,
+    surface_bg_layout: BindGroupLayout,
     voxel_cache: Arc<Mutex<VoxelCache<4>>>,
     cache_interface: Arc<CacheProducerInterface<4>>,
     terrain_generator: Arc<TerrainGenerator<4>>,
@@ -107,6 +120,12 @@ struct Immediates
     view_proj: Mat4,
     cam_pos: Vec3,
     frame_timestamp: u32,
+    width: u32,
+    height: u32,
+
+    chunk_width: u32,
+    chunk_height: u32,
+    chunk_alt: u32,
 }
 
 #[derive(Debug, Clone, Copy, Zeroable, Pod)]
@@ -127,6 +146,7 @@ impl State
             backends: wgpu::Backends::VULKAN,
             display: Some(Box::new(display)),
 
+            //flags: InstanceFlags::default() | InstanceFlags::debugging(),
             flags: InstanceFlags::default(),
             memory_budget_thresholds: MemoryBudgetThresholds::default(),
             backend_options: Default::default(),
@@ -139,7 +159,10 @@ impl State
             .request_device(&wgpu::DeviceDescriptor {
                 required_features: Features::IMMEDIATES
                     | Features::SHADER_EARLY_DEPTH_TEST
-                    | Features::TIMESTAMP_QUERY,
+                    | Features::TIMESTAMP_QUERY
+                    | Features::SHADER_I16
+                    | Features::SHADER_F16
+                    | Features::SHADER_INT64,
                 required_limits: wgpu::Limits {
                     max_immediate_size: 112,
                     max_storage_buffers_per_shader_stage: 16,
@@ -158,8 +181,8 @@ 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 cache_interface = CacheProducerInterface::new(1024, &device);
+        let mut voxel_cache = VoxelCache::<4>::new(100_000, device.clone(), queue.clone());
+        let cache_interface = CacheProducerInterface::new(256, &device);
 
         let terrain_generator = TerrainGenerator::<4>::new(5, "vxls_height.tif", 0.2, "img.jpg");
         // let terrain_generator = TerrainGenerator::<4>::new(
@@ -197,25 +220,85 @@ impl State
         let instance_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
             label: Some("Instance buffer"),
             contents: bytemuck::cast_slice(chunk_instances.as_slice()),
-            usage: BufferUsages::COPY_DST | BufferUsages::VERTEX,
+            usage: BufferUsages::COPY_DST | BufferUsages::VERTEX | BufferUsages::STORAGE,
         });
 
-        let shader_module = device.create_shader_module(wgpu::ShaderModuleDescriptor {
-            label: Some("Main shader module"),
-            source: wgpu::ShaderSource::Wgsl(
-                std::fs::read_to_string("shaders/voxel.wgsl")
-                    .unwrap()
-                    .into(),
-            ),
-        });
+        // let shader_module = unsafe {
+        //     device.create_shader_module_trusted(
+        //         wgpu::ShaderModuleDescriptor {
+        //             label: Some("Main shader module"),
+        //             source: wgpu::ShaderSource::Wgsl(
+        //                 std::fs::read_to_string("shaders/voxel.wgsl")
+        //                     .unwrap()
+        //                     .into(),
+        //             ),
+        //         },
+        //         wgpu::ShaderRuntimeChecks {
+        //             bounds_checks: false,
+        //             force_loop_bounding: false,
+        //             ray_query_initialization_tracking: false,
+        //             task_shader_dispatch_tracking: false,
+        //             mesh_shader_primitive_indices_clamp: false,
+        //             int_div_checks: false,
+        //         },
+        //     )
+        // };
+
+        //let shader_module = device.create_shader_module(include_wgsl!("../shaders/voxel.wgsl"));
+        let shader_module = unsafe {
+            device.create_shader_module_trusted(
+                wgpu::ShaderModuleDescriptor {
+                    label: Some("../shaders/voxel.spv"),
+                    source: wgpu::ShaderSource::SpirV(wgpu::__macro_helpers::Cow::Borrowed(
+                        wgpu::include_spirv_source!("../shaders/voxel.spv"),
+                    )),
+                },
+                wgpu::ShaderRuntimeChecks {
+                    bounds_checks: false,
+                    force_loop_bounding: false,
+                    ray_query_initialization_tracking: false,
+                    task_shader_dispatch_tracking: false,
+                    mesh_shader_primitive_indices_clamp: false,
+                    int_div_checks: false,
+                },
+            )
+        };
+
+        let surface_bind_group_layout =
+            device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
+                label: Some("surface_bg_layout"),
+                entries: &[wgpu::BindGroupLayoutEntry {
+                    binding: 0,
+                    visibility: ShaderStages::COMPUTE,
+                    ty: wgpu::BindingType::StorageTexture {
+                        access: wgpu::StorageTextureAccess::WriteOnly,
+                        format: wgpu::TextureFormat::Rgba32Float,
+                        view_dimension: wgpu::TextureViewDimension::D2,
+                    },
+                    count: None,
+                }],
+            });
 
         let pipeline_layout = device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
             label: Some("Voxel pipeline layout"),
 
-            bind_group_layouts: &[Some(&voxel_cache.bind_group_layout())],
+            bind_group_layouts: &[
+                Some(&voxel_cache.bind_group_layout()),
+                Some(&surface_bind_group_layout),
+            ],
             immediate_size: size_of::<Immediates>() as u32,
         });
 
+        let chunk_pipeline = device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
+            label: Some("Compute render"),
+            layout: Some(&pipeline_layout),
+            module: &shader_module,
+            entry_point: Some("ray_march_compute"),
+            compilation_options: wgpu::PipelineCompilationOptions::default(),
+            cache: None,
+        });
+
+        /*
         let chunk_pipeline = device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
             label: Some("Render pipeline"),
             layout: Some(&pipeline_layout),
@@ -252,7 +335,7 @@ 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(),
             }),
@@ -270,6 +353,7 @@ impl State
             multiview_mask: None,
             cache: None,
         });
+        */
 
         let state = State {
             instance,
@@ -279,8 +363,13 @@ impl State
             surface_format,
             egui_renderer,
             depth_buffer: Self::create_depth_buffer(&device, size.width, size.height),
+            target_texture: Self::create_target_texture(&device, size.width, size.height),
+            target_blitter: wgpu::util::TextureBlitterBuilder::new(&device, surface_format)
+                .sample_type(wgpu::FilterMode::Nearest)
+                .build(),
             queue,
             device,
+            surface_bg_layout: surface_bind_group_layout,
             usage_vec: Arc::new(Mutex::new(vec![])),
             pipeline: chunk_pipeline,
             voxel_cache: Arc::new(Mutex::new(voxel_cache)),
@@ -300,6 +389,27 @@ impl State
         state
     }
 
+    fn create_target_texture(device: &Device, width: u32, height: u32) -> (Texture, TextureView)
+    {
+        let texture = device.create_texture(&wgpu::TextureDescriptor {
+            label: Some("Target texture"),
+            size: Extent3d {
+                width,
+                height,
+                depth_or_array_layers: 1,
+            },
+            mip_level_count: 1,
+            sample_count: 1,
+            dimension: wgpu::TextureDimension::D2,
+            format: wgpu::TextureFormat::Rgba32Float,
+            usage: TextureUsages::STORAGE_BINDING | TextureUsages::TEXTURE_BINDING,
+            view_formats: &[],
+        });
+
+        let texture_view = texture.create_view(&wgpu::wgt::TextureViewDescriptor::default());
+        (texture, texture_view)
+    }
+
     fn get_window(&self) -> &Window
     {
         &self.window
@@ -375,6 +485,8 @@ impl State
         self.configure_surface();
         self.depth_buffer =
             Self::create_depth_buffer(&self.device, new_size.width, new_size.height);
+        self.target_texture =
+            Self::create_target_texture(&self.device, new_size.width, new_size.height);
     }
 
     fn render(&mut self)
@@ -422,6 +534,15 @@ impl State
                 ..Default::default()
             });
 
+        let surface_bg = self.device.create_bind_group(&wgpu::BindGroupDescriptor {
+            label: Some("surface_bg"),
+            layout: &self.surface_bg_layout,
+            entries: &[wgpu::BindGroupEntry {
+                binding: 0,
+                resource: wgpu::BindingResource::TextureView(&self.target_texture.1),
+            }],
+        });
+
         // ~~ Ray-marching timestamp query setup ~~
         let timestamp_query = self.device.create_query_set(&wgpu::QuerySetDescriptor {
             label: Some("timestamp_query_set"),
@@ -438,6 +559,48 @@ impl State
 
         // ~~ Main render pass ~~
         let mut encoder = self.device.create_command_encoder(&Default::default());
+        {
+            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),
+                }),
+            });
+
+            renderpass.set_pipeline(&self.pipeline);
+            renderpass.set_bind_group(0, Some(&self.voxel_cache.lock().bind_group()), &[]);
+            renderpass.set_bind_group(1, Some(&surface_bg), &[]);
+            let imm = [Immediates {
+                view_proj: self.camera.view_proj().inverse(),
+                cam_pos: self.camera.position,
+                frame_timestamp: self.voxel_cache.lock().current_timestamp(),
+                width: self.size.width,
+                height: self.size.height,
+
+                chunk_width: self.terrain_generator.chunk_width as u32,
+                chunk_height: self.terrain_generator.chunk_height as u32,
+                chunk_alt: self.terrain_generator.chunk_alt as u32,
+            }];
+            renderpass.set_immediates(0, unsafe { as_raw_bytes(&imm) });
+            renderpass.dispatch_workgroups(
+                self.size.width.div_ceil(16),
+                self.size.height.div_ceil(16),
+                1,
+            );
+            drop(renderpass);
+
+            encoder.resolve_query_set(&timestamp_query, 0..2, &timestamp_buffer, 0);
+        }
+
+        self.target_blitter.copy(
+            &self.device,
+            &mut encoder,
+            &self.target_texture.1,
+            &texture_view,
+        );
+        /*
         {
             let mut renderpass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
                 label: None,
@@ -488,6 +651,7 @@ impl State
 
             encoder.resolve_query_set(&timestamp_query, 0..2, &timestamp_buffer, 0);
         }
+        */
 
         // ~~ EGUI Render pass ~~
         {
@@ -561,21 +725,6 @@ impl State
             },
         );
 
-        // ~~ Get Ray-marching timestamps, report time ~~
-        let cloned_rm_time = self.rm_time.clone();
-        let cloned_queue = self.queue.clone();
-        DownloadBuffer::read_buffer(
-            &self.device,
-            &self.queue,
-            &timestamp_buffer.slice(..),
-            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;
-            },
-        );
-
         // ~~ Do cache managment
         if !self
             .camera
@@ -592,6 +741,21 @@ impl State
         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_queue = self.queue.clone();
+        DownloadBuffer::read_buffer(
+            &self.device,
+            &self.queue,
+            &timestamp_buffer.slice(..),
+            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;
+            },
+        );
+
         // ~~ Do cache managment
         if !self
             .camera
@@ -627,7 +791,7 @@ 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![];
 
@@ -686,7 +850,15 @@ impl State
                                 };
                             }
 
-                            (node.structure, node.colors, location)
+                            (
+                                StructurePoolElement {
+                                    occupancy_low: 0,
+                                    occupancy_high: 0,
+                                    pointers: node.structure,
+                                },
+                                node.colors,
+                                location,
+                            )
                         })
                         .collect::<Vec<_>>()
                         .into_iter()

src/voxel_cache.rs

@@ -8,6 +8,7 @@ pub mod request_buffer;
 pub mod structure_table;
 pub mod usage_buffer;
 pub mod producer_interface;
+pub mod indirect_buffer;
 
 pub mod data;
 
@@ -95,7 +96,7 @@ where
                     // Structure pool
                     wgpu::BindGroupLayoutEntry {
                         binding: 0,
-                        visibility: ShaderStages::FRAGMENT,
+                        visibility: ShaderStages::COMPUTE,
                         ty: wgpu::BindingType::Buffer {
                             ty: wgpu::BufferBindingType::Storage { read_only: false },
                             has_dynamic_offset: false,
@@ -106,7 +107,7 @@ where
                     // Color pool
                     wgpu::BindGroupLayoutEntry {
                         binding: 1,
-                        visibility: ShaderStages::FRAGMENT,
+                        visibility: ShaderStages::COMPUTE,
                         ty: wgpu::BindingType::Buffer {
                             ty: wgpu::BufferBindingType::Storage { read_only: false },
                             has_dynamic_offset: false,
@@ -117,7 +118,7 @@ where
                     // Location pool
                     wgpu::BindGroupLayoutEntry {
                         binding: 2,
-                        visibility: ShaderStages::FRAGMENT,
+                        visibility: ShaderStages::COMPUTE,
                         ty: wgpu::BindingType::Buffer {
                             ty: wgpu::BufferBindingType::Storage { read_only: false },
                             has_dynamic_offset: false,
@@ -128,7 +129,7 @@ where
                     // Request buffer
                     wgpu::BindGroupLayoutEntry {
                         binding: 3,
-                        visibility: ShaderStages::FRAGMENT,
+                        visibility: ShaderStages::COMPUTE,
                         ty: wgpu::BindingType::Buffer {
                             ty: wgpu::BufferBindingType::Storage { read_only: false },
                             has_dynamic_offset: false,
@@ -139,7 +140,7 @@ where
                     // Usage buffer
                     wgpu::BindGroupLayoutEntry {
                         binding: 4,
-                        visibility: ShaderStages::FRAGMENT,
+                        visibility: ShaderStages::COMPUTE,
                         ty: wgpu::BindingType::Buffer {
                             ty: wgpu::BufferBindingType::Storage { read_only: false },
                             has_dynamic_offset: false,
@@ -152,7 +153,7 @@ where
                     // Structure table pointers
                     wgpu::BindGroupLayoutEntry {
                         binding: 5,
-                        visibility: ShaderStages::FRAGMENT,
+                        visibility: ShaderStages::COMPUTE,
                         ty: wgpu::BindingType::Buffer {
                             ty: wgpu::BufferBindingType::Storage { read_only: false },
                             has_dynamic_offset: false,
@@ -164,7 +165,7 @@ where
                     // Structure table request buffer
                     wgpu::BindGroupLayoutEntry {
                         binding: 6,
-                        visibility: ShaderStages::FRAGMENT,
+                        visibility: ShaderStages::COMPUTE,
                         ty: wgpu::BindingType::Buffer {
                             ty: wgpu::BufferBindingType::Storage { read_only: false },
                             has_dynamic_offset: false,
@@ -346,6 +347,8 @@ where
             format!("
                 struct StructurePoolElement
                 {{
+                     occupancy_low: atomic<u32>,
+                     occupancy_high: atomic<u32>,
                      pointers: array<u32, {children_count}>
                 }}
 
@@ -565,15 +568,37 @@ 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 +613,15 @@ 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);
+                                            }}
                                         }}
                                     }}
 

src/voxel_cache/data.rs

@@ -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

src/voxel_cache/indirect_buffer.rs

@@ -0,0 +1,44 @@
+use wgpu::{Buffer, BufferUsages, Device, util::DeviceExt};
+
+pub struct BufferCompactor
+{
+    size: usize,
+    block_count: usize,
+    reduced_buffer: Buffer,
+    sum_buffer: Buffer,
+    compaction_buffer: Buffer,
+}
+
+impl BufferCompactor
+{
+    const THREAD_COUNT: usize = 256;
+    pub fn new(device: &Device, size: usize) -> Self
+    {
+        let block_count = size.div_ceil(size);
+        let reduced_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+            label: Some("indirect_buffer_reduced"),
+            contents: bytemuck::cast_slice(vec![0; block_count].as_slice()),
+            usage: BufferUsages::STORAGE,
+        });
+
+        let sum_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+            label: Some("indirect_buffer_sum"),
+            contents: bytemuck::cast_slice(vec![0; size].as_slice()),
+            usage: BufferUsages::STORAGE,
+        });
+
+        let compaction_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+            label: Some("indirect_buffer_compaction"),
+            contents: bytemuck::cast_slice(vec![0; size].as_slice()),
+            usage: BufferUsages::STORAGE,
+        });
+
+        BufferCompactor {
+            size,
+            block_count,
+            reduced_buffer,
+            sum_buffer,
+            compaction_buffer,
+        }
+    }
+}

src/voxel_cache/request_buffer.rs

@@ -249,17 +249,14 @@ where
 
 
                                  // Check if phase is last
-                                 /*
                                  if(phase == sub_phase)
                                  {{
                                       let next_phase_width = phase_total_width * 2;
-                                      if(next_phase_width >= element_sort_count && phase_total_width >= element_sort_count)
+                                      if(next_phase_width >= (element_sort_count * 2) && phase_total_width >= (element_sort_count * 2))
                                       {{
-                                        // This was the final phase, stop
                                         indirect_count = vec3<u32>(0);
                                       }}
                                  }}
-                                 */
 
                                  if(phase == 0)
                                  {{