"optimize" out matrix inverses and fix bugs

Author: devshgraphicsprogramming

examples_tests/22.RaytracedAO/closestHit.comp

@@ -52,21 +52,22 @@ void main()
 		if (hit)
 		{
 			const uint triangleID = intersection.primid;
-			const uvec3 indices = get_triangle_indices(batchInstanceGUID,triangleID);
-			
-			const mat4x3 batchWorldTform = InstData.data[batchInstanceGUID].tform;
+			const nbl_glsl_ext_Mitsuba_Loader_instance_data_t batchInstanceData = InstData.data[batchInstanceGUID];
 			// clear the hit success flag
 			intersections[vertex_depth_mod_2].data[gl_GlobalInvocationID.x].shapeid = -1;
+			const uvec3 indices = get_triangle_indices(batchInstanceGUID,triangleID);
+			
 			// obtain ray incoming direction
-			normalizedV = -inverse(mat3(batchWorldTform))*ray.direction;
+			normalizedV = -mat3(batchInstanceData.normalMatrixRow0,batchInstanceData.normalMatrixRow1,batchInstanceData.normalMatrixRow2)*ray.direction;
+			normalizedV = normalize(normalizedV);  // doesn't non-uniform scale screw up BxDF eval and generation?
 
 			// positions
 			const vec3 last_vx_pos = load_positions(indices,batchInstanceGUID);
 			const vec3 geomNormal = cross(dPdBary[0],dPdBary[1]);
 			const bool frontfacing = dot(geomNormal,normalizedV)>0.f;
 
 			// get material
-			const nbl_glsl_MC_oriented_material_t material = nbl_glsl_MC_material_data_t_getOriented(InstData.data[batchInstanceGUID].material,frontfacing);
+			const nbl_glsl_MC_oriented_material_t material = nbl_glsl_MC_material_data_t_getOriented(batchInstanceData.material,frontfacing);
 			emissive = nbl_glsl_MC_oriented_material_t_getEmissive(material);
 
 			const bool _continue = vertex_depth!=MAX_PATH_DEPTH && ray.maxT==FLT_MAX; // last vertex or was a NEE path
@@ -84,7 +85,7 @@ void main()
 				const vec3 hitWorldPos = dPdBary*compactBary.xy+last_vx_pos;
 
 				generate_next_rays(
-					MAX_RAYS_GENERATED,batchWorldTform,material,frontfacing,vertex_depth,
+					MAX_RAYS_GENERATED,batchInstanceData,material,frontfacing,vertex_depth,
 					scramble_start_state,sampleID,outPixelLocation,hitWorldPos,geomNormal,throughput
 				);
 			}

examples_tests/22.RaytracedAO/raygen.comp

@@ -37,15 +37,14 @@ void main()
 			const uvec4 visBuffer = texelFetch(frontFacingTriangleIDDrawID_unorm16Bary_dBarydScreenHalf2x2,ivec2(outPixelLocation),0);
 
 			// unproject (TODO: redo)
-			vec3 hitWorldPos;
+			vec3 hitWorldPos,V;
 			{
 				const vec3 NDC = vec3(vec2(outPixelLocation)*staticViewData.rcpPixelSize+staticViewData.rcpHalfPixelSize,1.0-revdepth);
 
 				const vec4 tmp = nbl_glsl_pseudoMul4x4with3x1(pc.cummon.inverseMVP,NDC);
 				hitWorldPos = tmp.xyz/tmp.w;
 
-				const vec3 V = nbl_glsl_pseudoMul3x4with3x1(pc.cummon.ndcToV,NDC);
-				normalizedV = normalize(V);
+				V = nbl_glsl_pseudoMul3x4with3x1(pc.cummon.ndcToV,NDC);
 			}
 
 			// vis buffer decode
@@ -58,15 +57,18 @@ void main()
 			const mat2 dBarydScreen = mat2(unpackHalf2x16(visBuffer[2]),unpackHalf2x16(visBuffer[3]));
 			#endif
 
-			// TODO: redo
-			const mat4x3 batchWorldTform = InstData.data[batchInstanceGUID].tform;
-			hitWorldPos = inverse(mat3(batchWorldTform))*(hitWorldPos-batchWorldTform[3]);
-			normalizedV = inverse(mat3(batchWorldTform))*normalizedV;
-
+			const nbl_glsl_ext_Mitsuba_Loader_instance_data_t batchInstanceData = InstData.data[batchInstanceGUID];
 			const uvec3 indices = get_triangle_indices(batchInstanceGUID,triangleID);
+
+			// TODO: redo
+			{
+				const mat3 batchScaleRotInv = mat3(batchInstanceData.normalMatrixRow0,batchInstanceData.normalMatrixRow1,batchInstanceData.normalMatrixRow2);
+				hitWorldPos = batchScaleRotInv*(hitWorldPos-batchInstanceData.tform[3]);
+				normalizedV = normalize(batchScaleRotInv*V); // doesn't non-uniform scale screw up BxDF eval and generation?
+			}
 
 			// get material while waiting for indices
-			const nbl_glsl_MC_oriented_material_t material = nbl_glsl_MC_material_data_t_getOriented(InstData.data[batchInstanceGUID].material,frontfacing);
+			const nbl_glsl_MC_oriented_material_t material = nbl_glsl_MC_material_data_t_getOriented(batchInstanceData.material,frontfacing);
 			emissive = nbl_glsl_MC_oriented_material_t_getEmissive(material);
 
 			const uint vertex_depth_mod_2 = 0x1u;
@@ -83,7 +85,7 @@ void main()
 			const vec3 geomNormal = cross(dPdBary[0],dPdBary[1]);
 			const uint vertex_depth = 1u;
 			generate_next_rays(
-				staticViewData.samplesPerPixelPerDispatch,batchWorldTform,material,frontfacing,vertex_depth,
+				staticViewData.samplesPerPixelPerDispatch,batchInstanceData,material,frontfacing,vertex_depth,
 				scramble_start_state,sampleID,outPixelLocation,hitWorldPos,geomNormal,vec3(1.0)
 			);
 		}

examples_tests/22.RaytracedAO/raytraceCommon.glsl

@@ -210,9 +210,10 @@ nbl_glsl_xoroshiro64star_state_t load_aux_vertex_attrs(
 }
 
 void generate_next_rays(
-	in uint maxRaysToGen, in mat4x3 batchWorldTform, in nbl_glsl_MC_oriented_material_t material, in bool frontfacing,
-	in uint vertex_depth, in nbl_glsl_xoroshiro64star_state_t scramble_start_state, in uint sampleID, in uvec2 outPixelLocation,
-	in vec3 origin, vec3 geomNormal, in vec3 prevThroughput)
+	in uint maxRaysToGen, in nbl_glsl_ext_Mitsuba_Loader_instance_data_t batchInstanceData,
+	in nbl_glsl_MC_oriented_material_t material, in bool frontfacing, in uint vertex_depth,
+	in nbl_glsl_xoroshiro64star_state_t scramble_start_state, in uint sampleID, in uvec2 outPixelLocation,
+	in vec3 pos, vec3 geomNormal, in vec3 prevThroughput)
 {
 	// get material streams as well
 	const nbl_glsl_MC_instr_stream_t gcs = nbl_glsl_MC_oriented_material_t_getGenChoiceStream(material);
@@ -231,6 +232,12 @@ void generate_next_rays(
 	// prepare rays
 	uint raysToAllocate = 0u;
 	float maxT[MAX_RAYS_GENERATED]; vec3 direction[MAX_RAYS_GENERATED]; vec3 nextThroughput[MAX_RAYS_GENERATED];	
+for (uint i=1u; i!=vertex_depth; i++)
+{
+	nbl_glsl_xoroshiro64star(scramble_start_state);
+	nbl_glsl_xoroshiro64star(scramble_start_state);
+	nbl_glsl_xoroshiro64star(scramble_start_state);
+}
 	for (uint i=0u; i<maxRaysToGen; i++)
 	{
 		nbl_glsl_xoroshiro64star_state_t scramble_state = scramble_start_state;
@@ -250,17 +257,17 @@ void generate_next_rays(
 	// set up dispatch indirect
 	atomicMax(traceIndirect[vertex_depth_mod_2_inv].params.num_groups_x,(baseOutputID+raysToAllocate-1u)/WORKGROUP_SIZE+1u);
 
-	const mat3 batchWorldScaleRot = mat3(batchWorldTform);
+	const mat4x3 batchWorldTform = batchInstanceData.tform;
 	uint offset = 0u;
 	for (uint i=0u; i<maxRaysToGen; i++)
 	if (maxT[i]!=0.f)
 	{
 		nbl_glsl_ext_RadeonRays_ray newRay;
 		// TODO: improve ray offsets
 		const float err = 1.f/96.f;
-		newRay.origin = origin+/*geomNormal/max(max(geomNormal.x,geomNormal.y),geomNormal.z)*sign(dot(geomNormal,direction[i]))*/direction[i]*err;
+		newRay.origin = mat3(batchWorldTform)*(pos+/*geomNormal/max(max(geomNormal.x,geomNormal.y),geomNormal.z)*sign(dot(geomNormal,direction[i]))*/direction[i]*err)+batchWorldTform[3];
 		newRay.maxT = maxT[i];
-		newRay.direction = batchWorldScaleRot*direction[i];
+		newRay.direction = mat3(batchWorldTform)*direction[i]; // normalize after ? (doesn't non-uniform scale screw up BxDF eval and generation?)
 		newRay.time = packOutPixelLocation(outPixelLocation);
 		newRay.mask = -1;
 		newRay._active = 1;