Faster raytracing

Shaders/raytrace/src/raytrace_bake_ao.hlsl

@@ -3,9 +3,10 @@
 #include "std/math.hlsl"
 
 struct Vertex {
-	float3 position;
-	float3 normal;
-	float2 tex;
+	uint posxy;
+	uint poszw;
+	uint nor;
+	uint tex;
 };
 
 struct RayGenConstantBuffer {

Shaders/raytrace/src/raytrace_bake_bent.hlsl

@@ -3,9 +3,10 @@
 #include "std/math.hlsl"
 
 struct Vertex {
-	float3 position;
-	float3 normal;
-	float2 tex;
+	uint posxy;
+	uint poszw;
+	uint nor;
+	uint tex;
 };
 
 struct RayGenConstantBuffer {

Shaders/raytrace/src/raytrace_bake_light.hlsl

@@ -4,9 +4,10 @@
 #include "std/attrib.hlsl"
 
 struct Vertex {
-	float3 position;
-	float3 normal;
-	float2 tex;
+	uint posxy;
+	uint poszw;
+	uint nor;
+	uint tex;
 };
 
 struct RayGenConstantBuffer {
@@ -91,16 +92,16 @@ void closesthit(inout RayPayload payload, in BuiltInTriangleIntersectionAttribut
 	uint3 indices_sample = indices.Load3(base_index);
 
 	float3 vertex_normals[3] = {
-		float3(vertices[indices_sample[0]].normal),
-		float3(vertices[indices_sample[1]].normal),
-		float3(vertices[indices_sample[2]].normal)
+		float3(S16toF32(vertices[indices_sample[0]].nor), S16toF32(vertices[indices_sample[0]].poszw).y),
+		float3(S16toF32(vertices[indices_sample[1]].nor), S16toF32(vertices[indices_sample[1]].poszw).y),
+		float3(S16toF32(vertices[indices_sample[2]].nor), S16toF32(vertices[indices_sample[2]].poszw).y)
 	};
 	float3 n = normalize(hit_attribute(vertex_normals, attr));
 
 	float2 vertex_uvs[3] = {
-		float2(vertices[indices_sample[0]].tex),
-		float2(vertices[indices_sample[1]].tex),
-		float2(vertices[indices_sample[2]].tex)
+		S16toF32(vertices[indices_sample[0]].tex),
+		S16toF32(vertices[indices_sample[1]].tex),
+		S16toF32(vertices[indices_sample[2]].tex)
 	};
 	float2 tex_coord = hit_attribute2d(vertex_uvs, attr);
 

Shaders/raytrace/src/raytrace_bake_thick.hlsl

@@ -3,9 +3,10 @@
 #include "std/math.hlsl"
 
 struct Vertex {
-	float3 position;
-	float3 normal;
-	float2 tex;
+	uint posxy;
+	uint poszw;
+	uint nor;
+	uint tex;
 };
 
 struct RayGenConstantBuffer {

Shaders/raytrace/src/raytrace_brute.hlsl

@@ -4,9 +4,10 @@
 #include "std/math.hlsl"
 
 struct Vertex {
-	float3 position;
-	float3 normal;
-	float2 tex;
+	uint posxy;
+	uint poszw;
+	uint nor;
+	uint tex;
 };
 
 struct RayGenConstantBuffer {
@@ -122,9 +123,9 @@ void closesthit(inout RayPayload payload, in BuiltInTriangleIntersectionAttribut
 	uint3 indices_sample = indices.Load3(base_index);
 
 	float2 vertex_uvs[3] = {
-		float2(vertices[indices_sample[0]].tex),
-		float2(vertices[indices_sample[1]].tex),
-		float2(vertices[indices_sample[2]].tex)
+		S16toF32(vertices[indices_sample[0]].tex),
+		S16toF32(vertices[indices_sample[1]].tex),
+		S16toF32(vertices[indices_sample[2]].tex)
 	};
 	float2 tex_coord = hit_attribute2d(vertex_uvs, attr);
 
@@ -140,9 +141,9 @@ void closesthit(inout RayPayload payload, in BuiltInTriangleIntersectionAttribut
 	}
 
 	float3 vertex_normals[3] = {
-		float3(vertices[indices_sample[0]].normal),
-		float3(vertices[indices_sample[1]].normal),
-		float3(vertices[indices_sample[2]].normal)
+		float3(S16toF32(vertices[indices_sample[0]].nor), S16toF32(vertices[indices_sample[0]].poszw).y),
+		float3(S16toF32(vertices[indices_sample[1]].nor), S16toF32(vertices[indices_sample[1]].poszw).y),
+		float3(S16toF32(vertices[indices_sample[2]].nor), S16toF32(vertices[indices_sample[2]].poszw).y)
 	};
 	float3 n = normalize(hit_attribute(vertex_normals, attr));
 

Shaders/raytrace/src/std/attrib.hlsl

@@ -2,6 +2,12 @@
 #ifndef _ATTRIB_HLSL_
 #define _ATTRIB_HLSL_
 
+float2 S16toF32(uint val) {
+	int a = (int)(val << 16) >> 16;
+	int b = (int)(val & 0xffff0000) >> 16;
+	return float2(a, b) / 32767.0f;
+}
+
 float3 hit_world_position() {
 	return WorldRayOrigin() + RayTCurrent() * WorldRayDirection();
 }

Sources/arm/render/RenderPathRaytrace.hx

@@ -17,8 +17,9 @@ class RenderPathRaytrace {
 	static var first = true;
 	static var f32 = new kha.arrays.Float32Array(24);
 	static var helpMat = iron.math.Mat4.identity();
-	static var vb: kha.arrays.Float32Array;
-	static var ib: kha.arrays.Uint32Array;
+	static var vb_scale = 1.0;
+	static var vb: kha.graphics4.VertexBuffer;
+	static var ib: kha.graphics4.IndexBuffer;
 	static var raysTimer = 0.0;
 	static var raysCounter = 0;
 	static var lastLayer: kha.Image = null;
@@ -218,32 +219,18 @@ class RenderPathRaytrace {
 			var bnoise_sobol = Scene.active.embedded.get("bnoise_sobol.k");
 			var bnoise_scramble = Scene.active.embedded.get("bnoise_scramble.k");
 			var bnoise_rank = Scene.active.embedded.get("bnoise_rank.k");
-			Krom.raytraceInit(shader.bytes.getData(), untyped vb.buffer, untyped ib.buffer);
+			Krom.raytraceInit(shader.bytes.getData(), untyped vb.buffer, untyped ib.buffer, vb_scale);
 		});
 	}
 
 	static function buildData() {
 		if (Context.mergedObject == null) arm.util.MeshUtil.mergeMesh();
 		var mo = Context.mergedObject;
-		var sc = mo.parent.transform.scale.x * mo.data.scalePos;
 		var md = mo.data;
 		var geom = md.geom;
-		var count = Std.int(geom.positions.length / 4);
-		vb = new kha.arrays.Float32Array(count * 8);
-		for (i in 0...count) {
-			vb[i * 8    ] = (geom.positions[i * 4    ] * sc / 32767);
-			vb[i * 8 + 1] = (geom.positions[i * 4 + 1] * sc / 32767);
-			vb[i * 8 + 2] = (geom.positions[i * 4 + 2] * sc / 32767);
-			vb[i * 8 + 3] =  geom.normals  [i * 2    ] / 32767;
-			vb[i * 8 + 4] =  geom.normals  [i * 2 + 1] / 32767;
-			vb[i * 8 + 5] =  geom.positions[i * 4 + 3] / 32767;
-			vb[i * 8 + 6] = (geom.uvs[i * 2    ] / 32767);
-			vb[i * 8 + 7] = (geom.uvs[i * 2 + 1] / 32767);
-		}
-
-		var indices = geom.indices[0];
-		ib = new kha.arrays.Uint32Array(indices.length);
-		for (i in 0...indices.length) ib[i] = indices[i];
+		vb_scale = mo.parent.transform.scale.x * md.scalePos;
+		vb = geom.vertexBuffer;
+		ib = geom.indexBuffers[0];
 	}
 
 	static function getBakeShaderName(): String {