gbuffer

assemble_trace.sh

@@ -1,6 +1,6 @@
 glslangValidator --spirv-val --spirv-dis -g0 -S frag --target-env vulkan1.4 --quiet src/trace.frag.glsl > src/trace.spv-dis
-#sed -i '/%11 =/,/OpFunctionEnd/d' src/trace.spv-dis
+sed -i '/%11 =/,/OpFunctionEnd/d' src/trace.spv-dis
-#sed -i 's/%11\b/%1000/g' src/trace.spv-dis
+sed -i 's/%11\b/%1000/g' src/trace.spv-dis
 spirv-as --preserve-numeric-ids --target-env vulkan1.4 -o src/recompiled.spv src/trace.spv-dis
 rm frag.spv
 # rm src/trace.spv-dis

src/lighting.frag.glsl

@@ -4,23 +4,39 @@
 
 layout(input_attachment_index = 0, set = 1, binding = 0) uniform subpassInput u_color;
 layout(input_attachment_index = 1, set = 1, binding = 1) uniform subpassInput u_normals;
+layout(input_attachment_index = 2, set = 1, binding = 2) uniform subpassInput u_depth;
 
+layout(location = 0) in vec2 v_pos;
 layout(location = 0) out vec4 f_color;
 
-vec3 shading(vec3 normal, vec4 position)
+vec3 WorldPosFromDepth(float depth) {
+    vec4 clipSpacePosition = vec4(v_pos, depth, 1.0);
+
+    vec4 worldSpacePosition = inverse(camera_uniforms.proj * camera_uniforms.view) * clipSpacePosition;
+
+    return worldSpacePosition.xyz /= worldSpacePosition.w;
+}
+
+vec3 shading(vec3 normal, vec3 position)
 {
     vec3 accum = vec3(0., 0., 0.);
-    mat3 rotation = mat3(pc.world[0].xyz, pc.world[1].xyz, pc.world[2].xyz);
-    //vec3 position=pc.world[3].xyz;
 
     for (int i = 0; (i < light_uniforms.light_count) && (i < 32); i++)
     {
-        accum += light_uniforms.col[i].xyz * ((dot(normalize(rotation * normal), normalize(light_uniforms.pos[i].xyz - position.xyz)) * .5) + .5);
+        accum += light_uniforms.col[i].xyz * ((dot(normal, normalize(light_uniforms.pos[i].xyz - position)) * .5) + .5);
     }
 
+    //return vec3(int(position.x * 256.0) ^ int(position.y * 256.0) ^ int(position.z * 256.0));
     return accum;
 }
 
 void main() {
-    f_color = vec4(subpassLoad(u_color).rgb, 1.0);
+    vec3 col = subpassLoad(u_color).xyz;
+    vec3 normal = subpassLoad(u_normals).xyz;
+    float depth = subpassLoad(u_depth).x;
+    if (depth == 1.0) {
+        discard;
+    }
+    vec3 position = WorldPosFromDepth(depth);
+    f_color = vec4(col * shading(normal, position), 1.0);
 }

src/lighting.vert.glsl

@@ -2,8 +2,16 @@
 #extension GL_GOOGLE_include_directive:require
 #include "include.glsl"
 
-const vec2 full_screen_triangle[3] = {vec2(-1., -1.), vec2(1., -1.), vec2(-1., 1.)};
+layout(location = 0) out vec2 v_pos;
+
+const vec2 full_screen_triangle[3] = {
+        vec2(0., 0.),
+        vec2(0., 2.),
+        vec2(2, 0.)
+    };
 
 void main() {
-    gl_Position = vec4(full_screen_triangle[gl_VertexIndex], 0.0, 1.0);
+    vec2 tri = (full_screen_triangle[gl_VertexIndex] * 2.0) - 1.0;
-}
+    v_pos = tri;
+    gl_Position = vec4(v_pos, 0.5, 1.0);
+}

src/main.rs

@@ -82,6 +82,7 @@ use winit::{
     event::{DeviceEvent, ElementState, MouseButton, WindowEvent},
     event_loop::{ActiveEventLoop, EventLoop},
     keyboard::{KeyCode, PhysicalKey},
+    platform::{wayland::EventLoopBuilderExtWayland, x11::EventLoopBuilderExtX11},
     window::Window,
 };
 
@@ -116,7 +117,7 @@ fn main() -> Result<(), impl Error> {
     ])
     .unwrap();
 
-    let event_loop = EventLoop::new().unwrap();
+    let event_loop = EventLoop::builder().with_x11().build().unwrap();
     let mut app = App::new(&event_loop);
 
     event_loop.run_app(&mut app)
@@ -209,6 +210,7 @@ struct RenderContext {
     framebuffers:       Vec<Arc<Framebuffer>>,
     color_buffer:       Arc<ImageView>,
     normal_buffer:      Arc<ImageView>,
+    depth_buffer:       Arc<ImageView>,
     shader_modules:     ShaderModules,
     implicit_pipeline:  Arc<GraphicsPipeline>,
     lighting_pipeline:  Arc<GraphicsPipeline>,
@@ -790,7 +792,7 @@ impl ApplicationHandler for App {
             {
                 color: [final_color],
                 depth_stencil: {},
-                input: [color, normals],
+                input: [color, normals, depth],
             },
             {
                 color: [final_color],
@@ -806,7 +808,7 @@ impl ApplicationHandler for App {
             depth_range: 0.0..=1.0,
         };
 
-        let (framebuffers, color_buffer, normal_buffer) =
+        let (framebuffers, color_buffer, normal_buffer, depth_buffer) =
             framebuffer_generation(&images, &render_pass, &self.memory_allocator);
 
         let (implicit_pipeline, lighting_pipeline) = deferred_pipeline_recompile(
@@ -840,6 +842,7 @@ impl ApplicationHandler for App {
             framebuffers,
             color_buffer,
             normal_buffer,
+            depth_buffer,
             shader_modules,
             implicit_pipeline,
             lighting_pipeline,
@@ -991,8 +994,12 @@ impl App {
                 .expect("Failed to recreate swapchain");
 
             rcx.swapchain = new_swapchain;
-            (rcx.framebuffers, rcx.color_buffer, rcx.normal_buffer) =
+            (
-                framebuffer_generation(&new_images, &rcx.render_pass, &self.memory_allocator);
+                rcx.framebuffers,
+                rcx.color_buffer,
+                rcx.normal_buffer,
+                rcx.depth_buffer,
+            ) = framebuffer_generation(&new_images, &rcx.render_pass, &self.memory_allocator);
 
             rcx.viewport.extent[0] = window_size.width as f32;
             rcx.viewport.extent[1] = window_size.height as f32 * -1.;
@@ -1175,9 +1182,16 @@ impl App {
         )
         .unwrap();
 
-        assert!(rcx.lighting_pipeline.layout().set_layouts().len() == 2);
+        //assert!(rcx.lighting_pipeline.layout().set_layouts().len() == 2);
 
         let lighting_layout1 = &rcx.lighting_pipeline.layout().set_layouts()[0];
+        //println!(
+        //    "Descriptor counts: {:?}",
+        //    lighting_layout1
+        //        .descriptor_counts()
+        //        .iter()
+        //        .collect::<Vec<_>>()
+        //);
         let lighting_set1 = DescriptorSet::new(
             self.descriptor_set_allocator.clone(),
             lighting_layout1.clone(),
@@ -1196,6 +1210,7 @@ impl App {
             [
                 WriteDescriptorSet::image_view(0, rcx.color_buffer.clone()),
                 WriteDescriptorSet::image_view(1, rcx.normal_buffer.clone()),
+                WriteDescriptorSet::image_view(2, rcx.depth_buffer.clone()),
             ],
             [],
         )
@@ -1314,6 +1329,8 @@ impl App {
             )
             .unwrap();
 
+        unsafe { builder.draw(3, 1, 0, 0) }.unwrap();
+
         builder
             .next_subpass(
                 Default::default(),
@@ -1368,7 +1385,12 @@ fn framebuffer_generation(
     images: &[Arc<Image>],
     render_pass: &Arc<RenderPass>,
     allocator: &Arc<StandardMemoryAllocator>,
-) -> (Vec<Arc<Framebuffer>>, Arc<ImageView>, Arc<ImageView>) {
+) -> (
+    Vec<Arc<Framebuffer>>,
+    Arc<ImageView>,
+    Arc<ImageView>,
+    Arc<ImageView>,
+) {
     let depth_buffer = ImageView::new_default(
         Image::new(
             allocator.clone(),
@@ -1376,7 +1398,9 @@ fn framebuffer_generation(
                 image_type: ImageType::Dim2d,
                 format: Format::D16_UNORM,
                 extent: images[0].extent(),
-                usage: ImageUsage::DEPTH_STENCIL_ATTACHMENT | ImageUsage::TRANSIENT_ATTACHMENT,
+                usage: ImageUsage::DEPTH_STENCIL_ATTACHMENT
+                    | ImageUsage::INPUT_ATTACHMENT
+                    | ImageUsage::TRANSIENT_ATTACHMENT,
                 ..Default::default()
             },
             AllocationCreateInfo::default(),
@@ -1441,7 +1465,7 @@ fn framebuffer_generation(
         })
         .collect::<Vec<_>>();
 
-    (framebuffers, color_buffer, normal_buffer)
+    (framebuffers, color_buffer, normal_buffer, depth_buffer)
 }
 
 fn deferred_pipeline_recompile(

src/trace.frag.glsl

@@ -57,14 +57,15 @@ void main() {
         vec3 n = getNormal(p, td.y);
 
         f_color = vec4(0.5, 0.5, 0.5, 1.0);
-        f_normal = vec4(n, 1.0);
+        mat3 rotation = mat3(pc.world[0].xyz, pc.world[1].xyz, pc.world[2].xyz);
+
+        f_normal = vec4(normalize(rotation * n), 1.0);
 
         vec4 tpoint = camera_uniforms.proj * camera_uniforms.view * pc.world * vec4(p, 1);
         gl_FragDepth = (tpoint.z / tpoint.w);
     }
     else
     {
-        //f_color = vec4(raydir, 1.0);
         discard;
     }
 

src/trace.spv-dis

@@ -1,12 +1,12 @@
 ; SPIR-V
 ; Version: 1.6
 ; Generator: Khronos Glslang Reference Front End; 11
-; Bound: 260
+; Bound: 289
 ; Schema: 0
                OpCapability Shader
           %1 = OpExtInstImport "GLSL.std.450"
                OpMemoryModel Logical GLSL450
-               OpEntryPoint Fragment %4 "main" %27 %35 %171 %182 %217 %220 %gl_FragDepth %259
+               OpEntryPoint Fragment %4 "main" %27 %35 %171 %182 %217 %246 %gl_FragDepth %288
                OpExecutionMode %4 OriginUpperLeft
                OpExecutionMode %4 DepthReplacing
                OpExecutionMode %4 DepthGreater
@@ -30,16 +30,16 @@
                OpDecorate %182 Binding 1
                OpDecorate %182 DescriptorSet 0
                OpDecorate %217 Location 0
-               OpDecorate %220 Location 1
+               OpDecorate %246 Location 1
                OpDecorate %gl_FragDepth BuiltIn FragDepth
                OpDecorate %_arr_v4float_uint_32 ArrayStride 16
                OpDecorate %_arr_v4float_uint_32_0 ArrayStride 16
-               OpDecorate %_struct_257 Block
+               OpDecorate %_struct_286 Block
-               OpMemberDecorate %_struct_257 0 Offset 0
+               OpMemberDecorate %_struct_286 0 Offset 0
-               OpMemberDecorate %_struct_257 1 Offset 512
+               OpMemberDecorate %_struct_286 1 Offset 512
-               OpMemberDecorate %_struct_257 2 Offset 1024
+               OpMemberDecorate %_struct_286 2 Offset 1024
-               OpDecorate %259 Binding 0
+               OpDecorate %288 Binding 0
-               OpDecorate %259 DescriptorSet 0
+               OpDecorate %288 DescriptorSet 0
        %void = OpTypeVoid
           %3 = OpTypeFunction %void
       %float = OpTypeFloat 32
@@ -89,7 +89,10 @@
         %217 = OpVariable %_ptr_Output_v4float Output
   %float_0_5 = OpConstant %float 0.5
         %219 = OpConstantComposite %v4float %float_0_5 %float_0_5 %float_0_5 %float_1
-        %220 = OpVariable %_ptr_Output_v4float Output
+%mat3v3float = OpTypeMatrix %v3float 3
+%_ptr_Function_mat3v3float = OpTypePointer Function %mat3v3float
+%_ptr_PushConstant_v4float = OpTypePointer PushConstant %v4float
+        %246 = OpVariable %_ptr_Output_v4float Output
 %_ptr_Function_v4float = OpTypePointer Function %v4float
 %_ptr_Uniform_mat4v4float = OpTypePointer Uniform %mat4v4float
 %_ptr_Output_float = OpTypePointer Output %float
@@ -98,9 +101,9 @@
     %uint_32 = OpConstant %uint 32
 %_arr_v4float_uint_32 = OpTypeArray %v4float %uint_32
 %_arr_v4float_uint_32_0 = OpTypeArray %v4float %uint_32
-%_struct_257 = OpTypeStruct %_arr_v4float_uint_32 %_arr_v4float_uint_32_0 %uint
+%_struct_286 = OpTypeStruct %_arr_v4float_uint_32 %_arr_v4float_uint_32_0 %uint
-%_ptr_Uniform__struct_257 = OpTypePointer Uniform %_struct_257
+%_ptr_Uniform__struct_286 = OpTypePointer Uniform %_struct_286
-        %259 = OpVariable %_ptr_Uniform__struct_257 Uniform
+        %288 = OpVariable %_ptr_Uniform__struct_286 Uniform
           %4 = OpFunction %void None %3
           %5 = OpLabel
         %169 = OpVariable %_ptr_Function_v3float Function
@@ -113,7 +116,8 @@
         %209 = OpVariable %_ptr_Function_v3float Function
         %210 = OpVariable %_ptr_Function_v3float Function
         %212 = OpVariable %_ptr_Function_float Function
-        %227 = OpVariable %_ptr_Function_v4float Function
+        %222 = OpVariable %_ptr_Function_mat3v3float Function
+        %256 = OpVariable %_ptr_Function_v4float Function
          %39 = OpAccessChain %_ptr_PushConstant_v3float %35 %int_1
          %40 = OpLoad %v3float %39
          %42 = OpCompositeConstruct %v3float %float_8 %float_8 %float_8
@@ -150,7 +154,7 @@
         %205 = OpLoad %float %204
         %206 = OpFOrdLessThan %bool %205 %float_9_99999975en05
                OpSelectionMerge %208 None
-               OpBranchConditional %206 %207 %251
+               OpBranchConditional %206 %207 %280
         %207 = OpLabel
         %211 = OpLoad %v3float %196
                OpStore %210 %211
@@ -160,47 +164,65 @@
         %215 = OpFunctionCall %v3float %17 %210 %212
                OpStore %209 %215
                OpStore %217 %219
-        %221 = OpLoad %v3float %209
+        %224 = OpAccessChain %_ptr_PushConstant_v4float %35 %int_0 %int_0
-        %222 = OpCompositeExtract %float %221 0
+        %225 = OpLoad %v4float %224
-        %223 = OpCompositeExtract %float %221 1
+        %226 = OpVectorShuffle %v3float %225 %225 0 1 2
-        %224 = OpCompositeExtract %float %221 2
+        %227 = OpAccessChain %_ptr_PushConstant_v4float %35 %int_0 %int_1
-        %225 = OpCompositeConstruct %v4float %222 %223 %224 %float_1
+        %228 = OpLoad %v4float %227
-               OpStore %220 %225
+        %229 = OpVectorShuffle %v3float %228 %228 0 1 2
-        %229 = OpAccessChain %_ptr_Uniform_mat4v4float %182 %int_1
+        %230 = OpAccessChain %_ptr_PushConstant_v4float %35 %int_0 %int_2
-        %230 = OpLoad %mat4v4float %229
+        %231 = OpLoad %v4float %230
-        %231 = OpAccessChain %_ptr_Uniform_mat4v4float %182 %int_0
+        %232 = OpVectorShuffle %v3float %231 %231 0 1 2
-        %232 = OpLoad %mat4v4float %231
+        %233 = OpCompositeExtract %float %226 0
-        %233 = OpMatrixTimesMatrix %mat4v4float %230 %232
+        %234 = OpCompositeExtract %float %226 1
-        %234 = OpAccessChain %_ptr_PushConstant_mat4v4float %35 %int_0
+        %235 = OpCompositeExtract %float %226 2
-        %235 = OpLoad %mat4v4float %234
+        %236 = OpCompositeExtract %float %229 0
-        %236 = OpMatrixTimesMatrix %mat4v4float %233 %235
+        %237 = OpCompositeExtract %float %229 1
-        %237 = OpLoad %v3float %196
+        %238 = OpCompositeExtract %float %229 2
-        %238 = OpCompositeExtract %float %237 0
+        %239 = OpCompositeExtract %float %232 0
-        %239 = OpCompositeExtract %float %237 1
+        %240 = OpCompositeExtract %float %232 1
-        %240 = OpCompositeExtract %float %237 2
+        %241 = OpCompositeExtract %float %232 2
-        %241 = OpCompositeConstruct %v4float %238 %239 %240 %float_1
+        %242 = OpCompositeConstruct %v3float %233 %234 %235
-        %242 = OpMatrixTimesVector %v4float %236 %241
+        %243 = OpCompositeConstruct %v3float %236 %237 %238
-               OpStore %227 %242
+        %244 = OpCompositeConstruct %v3float %239 %240 %241
-        %245 = OpAccessChain %_ptr_Function_float %227 %uint_2
+        %245 = OpCompositeConstruct %mat3v3float %242 %243 %244
-        %246 = OpLoad %float %245
+               OpStore %222 %245
-        %248 = OpAccessChain %_ptr_Function_float %227 %uint_3
+        %247 = OpLoad %mat3v3float %222
-        %249 = OpLoad %float %248
+        %248 = OpLoad %v3float %209
-        %250 = OpFDiv %float %246 %249
+        %249 = OpMatrixTimesVector %v3float %247 %248
-               OpStore %gl_FragDepth %250
+        %250 = OpExtInst %v3float %1 Normalize %249
+        %251 = OpCompositeExtract %float %250 0
+        %252 = OpCompositeExtract %float %250 1
+        %253 = OpCompositeExtract %float %250 2
+        %254 = OpCompositeConstruct %v4float %251 %252 %253 %float_1
+               OpStore %246 %254
+        %258 = OpAccessChain %_ptr_Uniform_mat4v4float %182 %int_1
+        %259 = OpLoad %mat4v4float %258
+        %260 = OpAccessChain %_ptr_Uniform_mat4v4float %182 %int_0
+        %261 = OpLoad %mat4v4float %260
+        %262 = OpMatrixTimesMatrix %mat4v4float %259 %261
+        %263 = OpAccessChain %_ptr_PushConstant_mat4v4float %35 %int_0
+        %264 = OpLoad %mat4v4float %263
+        %265 = OpMatrixTimesMatrix %mat4v4float %262 %264
+        %266 = OpLoad %v3float %196
+        %267 = OpCompositeExtract %float %266 0
+        %268 = OpCompositeExtract %float %266 1
+        %269 = OpCompositeExtract %float %266 2
+        %270 = OpCompositeConstruct %v4float %267 %268 %269 %float_1
+        %271 = OpMatrixTimesVector %v4float %265 %270
+               OpStore %256 %271
+        %274 = OpAccessChain %_ptr_Function_float %256 %uint_2
+        %275 = OpLoad %float %274
+        %277 = OpAccessChain %_ptr_Function_float %256 %uint_3
+        %278 = OpLoad %float %277
+        %279 = OpFDiv %float %275 %278
+               OpStore %gl_FragDepth %279
                OpBranch %208
-        %251 = OpLabel
+        %280 = OpLabel
                OpTerminateInvocation
         %208 = OpLabel
                OpReturn
                OpFunctionEnd
-         %11 = OpFunction %float None %9
-         %10 = OpFunctionParameter %_ptr_Function_v3float
-         %12 = OpLabel
-         %45 = OpLoad %v3float %10
-         %46 = OpExtInst %float %1 Length %45
-         %48 = OpFSub %float %46 %float_1
-               OpReturnValue %48
-               OpFunctionEnd
          %17 = OpFunction %v3float None %14
          %15 = OpFunctionParameter %_ptr_Function_v3float
          %16 = OpFunctionParameter %_ptr_Function_float
@@ -219,7 +241,7 @@
          %62 = OpLoad %float %61
          %63 = OpCompositeConstruct %v3float %56 %59 %62
                OpStore %64 %63
-         %65 = OpFunctionCall %float %11 %64
+         %65 = OpFunctionCall %float %1000 %64
          %66 = OpAccessChain %_ptr_Function_float %51 %uint_0
                OpStore %66 %65
          %67 = OpAccessChain %_ptr_Function_float %15 %uint_0
@@ -231,7 +253,7 @@
          %73 = OpLoad %float %72
          %74 = OpCompositeConstruct %v3float %68 %71 %73
                OpStore %75 %74
-         %76 = OpFunctionCall %float %11 %75
+         %76 = OpFunctionCall %float %1000 %75
          %77 = OpAccessChain %_ptr_Function_float %51 %uint_1
                OpStore %77 %76
          %78 = OpAccessChain %_ptr_Function_float %15 %uint_0
@@ -243,13 +265,13 @@
          %84 = OpFAdd %float %83 %float_9_99999975en05
          %85 = OpCompositeConstruct %v3float %79 %81 %84
                OpStore %86 %85
-         %87 = OpFunctionCall %float %11 %86
+         %87 = OpFunctionCall %float %1000 %86
          %88 = OpAccessChain %_ptr_Function_float %51 %uint_2
                OpStore %88 %87
          %89 = OpLoad %v3float %51
          %91 = OpLoad %v3float %15
                OpStore %90 %91
-         %92 = OpFunctionCall %float %11 %90
+         %92 = OpFunctionCall %float %1000 %90
          %93 = OpCompositeConstruct %v3float %92 %92 %92
          %94 = OpFSub %v3float %89 %93
          %95 = OpExtInst %v3float %1 Normalize %94
@@ -269,7 +291,7 @@
                OpStore %23 %102
         %104 = OpLoad %v3float %23
                OpStore %103 %104
-        %105 = OpFunctionCall %float %11 %103
+        %105 = OpFunctionCall %float %1000 %103
         %107 = OpFMul %float %105 %float_0_899999976
         %108 = OpAccessChain %_ptr_Function_float %99 %uint_1
                OpStore %108 %107
@@ -319,7 +341,7 @@
         %125 = OpLabel
         %148 = OpLoad %v3float %23
                OpStore %147 %148
-        %149 = OpFunctionCall %float %11 %147
+        %149 = OpFunctionCall %float %1000 %147
         %150 = OpFMul %float %149 %float_0_899999976
         %151 = OpAccessChain %_ptr_Function_float %99 %uint_1
                OpStore %151 %150