Use radix sort for sorting draw commands, add ambient light, tweak lighting

assets/shaders/mesh.glsl

@@ -299,7 +299,7 @@ vec3 microfacetModel(EvalMaterial mat, int light_idx, vec3 P, vec3 N) {
 
     shadow_mult = sum / 16.0;
   }
-  shadow_mult = clamp(shadow_mult, 0.3, 1.0);
+  shadow_mult = clamp(shadow_mult, 0.0, 1.0);
 
   vec3 specBrdf = 0.25 * ggxDistribution(mat, NDotH) * schlickFresnel(mat, LDotH) * geomSmith(mat, NDotL) * geomSmith(mat, NDotV);
 
@@ -327,6 +327,11 @@ void main() {
     finalColor += microfacetModel(material, i, VertexOut.vPos, N);
   }
 
+  float ambient_a = dot(VertexOut.wNormal, vec3(0, 1, 0)) * 0.5 + 0.5;
+
+  // ambient
+  finalColor += material.albedo.rgb * mix(vec3(0.0116, 0.0127, 0.0200), vec3(0.185, 0.198, 0.250), ambient_a);
+
   FragColor = vec4(finalColor, material.albedo.a);
 }
 

src/Render.zig

@@ -428,6 +428,12 @@ pub const LightCommand = union(LightKind) {
     point: PointLight,
 };
 
+const DrawCommandKey = packed struct {
+    transparent: u1 = 0,
+    distance: u15 = 0,
+    mesh: u16 = 0,
+};
+
 pub fn drawLight(self: *Render, cmd: LightCommand) void {
     self.lights[self.light_count] = cmd;
     self.light_count += 1;
@@ -435,9 +441,99 @@ pub fn drawLight(self: *Render, cmd: LightCommand) void {
 
 pub fn draw(self: *Render, cmd: DrawCommand) void {
     self.command_buffer[self.command_count] = cmd;
+    // TODO: don't load the whole mesh here
+    const mesh = self.assetman.resolveMesh(cmd.mesh);
+    const material: Material = if (cmd.material_override) |mat| mat else mesh.material;
+    const view_origin = self.camera.view_mat.extractTranslation();
+    const max_value = @as(f32, @floatFromInt(std.math.maxInt(u15)));
+    const dist: u15 = @intFromFloat(std.math.clamp(view_origin.distance(cmd.transform.extractTranslation()) / max_value, 0.0, max_value));
+    const key = DrawCommandKey{
+        .transparent = if (material.blend_mode == .AlphaBlend) 1 else 0,
+        .distance = if (material.blend_mode == .AlphaBlend) ~dist else dist, // TODO: calculate distance. Opaque should be front to back, transparent back to front
+        .mesh = @intCast(cmd.mesh.id % std.math.maxInt(u16)),
+    };
+    self.command_buffer[self.command_count].key = key;
     self.command_count += 1;
 }
 
+// Multipass radix sort for u32
+fn sortCommands(self: *Render, in_cmds: []DrawCommand) void {
+    var cmds = in_cmds;
+    var aux = self.frame_arena.alloc(DrawCommand, cmds.len) catch @panic("OOM");
+
+    var cnt1: [256]usize = std.mem.zeroes([256]usize);
+    var cnt2: [256]usize = std.mem.zeroes([256]usize);
+    var cnt3: [256]usize = std.mem.zeroes([256]usize);
+    var cnt4: [256]usize = std.mem.zeroes([256]usize);
+
+    // Find counts
+    for (cmds) |*cmd| {
+        const key: u32 = @bitCast(cmd.key);
+
+        cnt1[(key >> 0) & 0xFF] += 1;
+        cnt2[(key >> 8) & 0xFF] += 1;
+        cnt3[(key >> 16) & 0xFF] += 1;
+        cnt4[(key >> 24) & 0xFF] += 1;
+    }
+
+    var a1: usize = 0;
+    var a2: usize = 0;
+    var a3: usize = 0;
+    var a4: usize = 0;
+    for (0..256) |i| {
+        const b1 = cnt1[i];
+        const b2 = cnt2[i];
+        const b3 = cnt3[i];
+        const b4 = cnt4[i];
+
+        cnt1[i] = a1;
+        cnt2[i] = a2;
+        cnt3[i] = a3;
+        cnt4[i] = a4;
+
+        a1 += b1;
+        a2 += b2;
+        a3 += b3;
+        a4 += b4;
+    }
+
+    for (0..cmds.len) |i| {
+        const key: u32 = @bitCast(cmds[i].key);
+        const k = (key >> 0) & 0xFF;
+        const dst = cnt1[k];
+        cnt1[k] += 1;
+        aux[dst] = cmds[i];
+    }
+    std.mem.swap([]DrawCommand, &cmds, &aux);
+
+    for (0..cmds.len) |i| {
+        const key: u32 = @bitCast(cmds[i].key);
+        const k = (key >> 8) & 0xFF;
+        const dst = cnt2[k];
+        cnt2[k] += 1;
+        aux[dst] = cmds[i];
+    }
+    std.mem.swap([]DrawCommand, &cmds, &aux);
+
+    for (0..cmds.len) |i| {
+        const key: u32 = @bitCast(cmds[i].key);
+        const k = (key >> 16) & 0xFF;
+        const dst = cnt3[k];
+        cnt3[k] += 1;
+        aux[dst] = cmds[i];
+    }
+    std.mem.swap([]DrawCommand, &cmds, &aux);
+
+    for (0..cmds.len) |i| {
+        const key: u32 = @bitCast(cmds[i].key);
+        const k = (key >> 24) & 0xFF;
+        const dst = cnt4[k];
+        cnt4[k] += 1;
+        aux[dst] = cmds[i];
+    }
+    std.mem.swap([]DrawCommand, &cmds, &aux);
+}
+
 pub fn finish(self: *Render) void {
     const zone = tracy.initZone(@src(), .{ .name = "Render.finish" });
     defer zone.deinit();
@@ -450,32 +546,7 @@ pub fn finish(self: *Render) void {
         const zoneSort = tracy.initZone(@src(), .{ .name = "Render.finish_sortDraws" });
         defer zoneSort.deinit();
 
-        const cmds = self.command_buffer[0..self.command_count];
+        self.sortCommands(self.command_buffer[0..self.command_count]);
-        std.mem.sortUnstable(DrawCommand, cmds, self, struct {
-            pub fn lessThan(render: *const Render, lhs: DrawCommand, rhs: DrawCommand) bool {
-                const lhs_mesh = render.assetman.resolveMesh(lhs.mesh);
-                const rhs_mesh = render.assetman.resolveMesh(rhs.mesh);
-                const lhs_material: Material = if (lhs.material_override) |mat| mat else lhs_mesh.material;
-                const rhs_material: Material = if (rhs.material_override) |mat| mat else rhs_mesh.material;
-
-                const lhs_blend_num = @intFromEnum(lhs_material.blend_mode);
-                const rhs_blend_num = @intFromEnum(rhs_material.blend_mode);
-
-                if (lhs_material.blend_mode == .Opaque and rhs_material.blend_mode == .Opaque) {
-                    return lhs.mesh.id < rhs.mesh.id;
-                }
-
-                if (lhs_material.blend_mode == .AlphaBlend and rhs_material.blend_mode == .AlphaBlend) {
-                    const lhs_view_pos = render.camera.view_mat.mulByVec4(lhs.transform.extractTranslation().toVec4(1));
-                    const rhs_view_pos = render.camera.view_mat.mulByVec4(rhs.transform.extractTranslation().toVec4(1));
-
-                    // Back to front sorting. View pos has negative Z
-                    return lhs_view_pos.z() < rhs_view_pos.z();
-                }
-
-                return lhs_blend_num < rhs_blend_num;
-            }
-        }.lessThan);
     }
 
     if (self.update_view_frustum) {
@@ -1129,6 +1200,7 @@ pub fn checkGLError() void {
 }
 
 pub const DrawCommand = struct {
+    key: DrawCommandKey = .{},
     mesh: AssetManager.Handle.Mesh,
     material_override: ?Material,
     transform: Mat4,

src/game.zig

@@ -191,7 +191,7 @@ export fn game_init(global_allocator: *std.mem.Allocator) void {
     _ = globals.g_mem.world.addEntity(.{
         .flags = .{ .dir_light = true, .rotate = true },
         .transform = .{ .rot = Quat.fromEulerAngles(Vec3.new(70, 0, 0)) },
-        .light = .{ .color_intensity = Vec4.new(std.math.pow(f32, 1, 2.2), std.math.pow(f32, 0.9568627450980393, 2.2), std.math.pow(f32, 0.9176470588235294, 2.2), 1.0) },
+        .light = .{ .color_intensity = Vec4.new(1.00, 0.885, 0.570, 1.0) },
         .rotate = .{ .axis = Vec3.up(), .rate = -10 },
     });