Add support for gl_light_raytrace for cards that do not support rayquery

src/common/rendering/vulkan/renderer/vk_descriptorset.cpp

@@ -104,7 +104,15 @@ void VkDescriptorSetManager::UpdateFixedSet()
 	update.AddCombinedImageSampler(FixedSet.get(), 0, fb->GetTextureManager()->Shadowmap.View.get(), fb->GetSamplerManager()->ShadowmapSampler.get(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
 	update.AddCombinedImageSampler(FixedSet.get(), 1, fb->GetTextureManager()->Lightmap.View.get(), fb->GetSamplerManager()->LightmapSampler.get(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
 	if (fb->device->SupportsExtension(VK_KHR_RAY_QUERY_EXTENSION_NAME))
+	{
 		update.AddAccelerationStructure(FixedSet.get(), 2, fb->GetRaytrace()->GetAccelStruct());
+	}
+	else
+	{
+		update.AddBuffer(FixedSet.get(), 2, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, fb->GetRaytrace()->GetNodeBuffer());
+		update.AddBuffer(FixedSet.get(), 3, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, fb->GetRaytrace()->GetVertexBuffer());
+		update.AddBuffer(FixedSet.get(), 4, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, fb->GetRaytrace()->GetIndexBuffer());
+	}
 	update.Execute(fb->device.get());
 }
 
@@ -264,7 +272,15 @@ void VkDescriptorSetManager::CreateFixedSetLayout()
 	builder.AddBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 1, VK_SHADER_STAGE_FRAGMENT_BIT);
 	builder.AddBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 1, VK_SHADER_STAGE_FRAGMENT_BIT);
 	if (fb->device->SupportsExtension(VK_KHR_RAY_QUERY_EXTENSION_NAME))
+	{
 		builder.AddBinding(2, VK_DESCRIPTOR_TYPE_ACCELERATION_STRUCTURE_KHR, 1, VK_SHADER_STAGE_FRAGMENT_BIT);
+	}
+	else
+	{
+		builder.AddBinding(2, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, 1, VK_SHADER_STAGE_FRAGMENT_BIT);
+		builder.AddBinding(3, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, 1, VK_SHADER_STAGE_FRAGMENT_BIT);
+		builder.AddBinding(4, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, 1, VK_SHADER_STAGE_FRAGMENT_BIT);
+	}
 	builder.DebugName("VkDescriptorSetManager.FixedSetLayout");
 	FixedSetLayout = builder.Create(fb->device.get());
 }
@@ -284,7 +300,13 @@ void VkDescriptorSetManager::CreateFixedSetPool()
 	DescriptorPoolBuilder poolbuilder;
 	poolbuilder.AddPoolSize(VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 2 * maxSets);
 	if (fb->device->SupportsExtension(VK_KHR_RAY_QUERY_EXTENSION_NAME))
+	{
 		poolbuilder.AddPoolSize(VK_DESCRIPTOR_TYPE_ACCELERATION_STRUCTURE_KHR, 1 * maxSets);
+	}
+	else
+	{
+		poolbuilder.AddPoolSize(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, 3 * maxSets);
+	}
 	poolbuilder.MaxSets(maxSets);
 	poolbuilder.DebugName("VkDescriptorSetManager.FixedDescriptorPool");
 	FixedDescriptorPool = poolbuilder.Create(fb->device.get());

src/common/rendering/vulkan/renderer/vk_raytrace.cpp

@@ -28,6 +28,8 @@
 
 VkRaytrace::VkRaytrace(VulkanRenderDevice* fb) : fb(fb)
 {
+	useRayQuery = fb->device->SupportsExtension(VK_KHR_RAY_QUERY_EXTENSION_NAME);
+
 	NullMesh.MeshVertices.Push({ -1.0f, -1.0f, -1.0f });
 	NullMesh.MeshVertices.Push({ 1.0f, -1.0f, -1.0f });
 	NullMesh.MeshVertices.Push({ 1.0f,  1.0f, -1.0f });
@@ -43,6 +45,8 @@ VkRaytrace::VkRaytrace(VulkanRenderDevice* fb) : fb(fb)
 	for (int i = 0; i < 3 * 4; i++)
 		NullMesh.MeshElements.Push(i);
 
+	NullMesh.Collision = std::make_unique<TriangleMeshShape>(NullMesh.MeshVertices.Data(), NullMesh.MeshVertices.Size(), NullMesh.MeshElements.Data(), NullMesh.MeshElements.Size());
+
 	SetLevelMesh(nullptr);
 }
 
@@ -55,10 +59,7 @@ void VkRaytrace::SetLevelMesh(hwrenderer::LevelMesh* mesh)
 	{
 		Reset();
 		Mesh = mesh;
-		if (fb->device->SupportsExtension(VK_KHR_RAY_QUERY_EXTENSION_NAME))
+		CreateVulkanObjects();
-		{
-			CreateVulkanObjects();
-		}
 	}
 }
 
@@ -68,6 +69,7 @@ void VkRaytrace::Reset()
 	deletelist->Add(std::move(vertexBuffer));
 	deletelist->Add(std::move(indexBuffer));
 	deletelist->Add(std::move(transferBuffer));
+	deletelist->Add(std::move(nodesBuffer));
 	deletelist->Add(std::move(blScratchBuffer));
 	deletelist->Add(std::move(blAccelStructBuffer));
 	deletelist->Add(std::move(blAccelStruct));
@@ -81,38 +83,35 @@ void VkRaytrace::Reset()
 void VkRaytrace::CreateVulkanObjects()
 {
 	CreateVertexAndIndexBuffers();
-	CreateBottomLevelAccelerationStructure();
+	if (useRayQuery)
-	CreateTopLevelAccelerationStructure();
+	{
+		CreateBottomLevelAccelerationStructure();
+		CreateTopLevelAccelerationStructure();
+	}
 }
 
 void VkRaytrace::CreateVertexAndIndexBuffers()
 {
-	static_assert(sizeof(FVector3) == 3 * 4, "sizeof(FVector3) is not 12 bytes!");
+	std::vector<CollisionNode> nodes = CreateCollisionNodes();
 
-	size_t vertexbuffersize = (size_t)Mesh->MeshVertices.Size() * sizeof(FVector3);
+	// std430 alignment rules forces us to convert the vec3 to a vec4
-	size_t indexbuffersize = (size_t)Mesh->MeshElements.Size() * sizeof(uint32_t);
+	std::vector<FVector4> vertices;
-	size_t transferbuffersize = vertexbuffersize + indexbuffersize;
+	vertices.reserve(Mesh->MeshVertices.Size());
-	size_t vertexoffset = 0;
+	for (const FVector3& v : Mesh->MeshVertices)
-	size_t indexoffset = vertexoffset + vertexbuffersize;
+		vertices.push_back({ v, 1.0f });
 
-	transferBuffer = BufferBuilder()
+	CollisionNodeBufferHeader nodesHeader;
-		.Usage(VK_BUFFER_USAGE_TRANSFER_SRC_BIT, VMA_MEMORY_USAGE_CPU_ONLY)
+	nodesHeader.root = Mesh->Collision->get_root();
-		.Size(transferbuffersize)
-		.DebugName("transferBuffer")
-		.Create(fb->device.get());
-
-	uint8_t* data = (uint8_t*)transferBuffer->Map(0, transferbuffersize);
-	memcpy(data + vertexoffset, Mesh->MeshVertices.Data(), vertexbuffersize);
-	memcpy(data + indexoffset, Mesh->MeshElements.Data(), indexbuffersize);
-	transferBuffer->Unmap();
 
 	vertexBuffer = BufferBuilder()
 		.Usage(
 			VK_BUFFER_USAGE_VERTEX_BUFFER_BIT |
 			VK_BUFFER_USAGE_TRANSFER_DST_BIT |
-			VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT |
+			(useRayQuery ?
-			VK_BUFFER_USAGE_ACCELERATION_STRUCTURE_BUILD_INPUT_READ_ONLY_BIT_KHR)
+				VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT |
-		.Size(vertexbuffersize)
+				VK_BUFFER_USAGE_ACCELERATION_STRUCTURE_BUILD_INPUT_READ_ONLY_BIT_KHR : 0) |
+			VK_BUFFER_USAGE_STORAGE_BUFFER_BIT)
+		.Size(vertices.size() * sizeof(FVector4))
 		.DebugName("vertexBuffer")
 		.Create(fb->device.get());
 
@@ -120,19 +119,29 @@ void VkRaytrace::CreateVertexAndIndexBuffers()
 		.Usage(
 			VK_BUFFER_USAGE_INDEX_BUFFER_BIT |
 			VK_BUFFER_USAGE_TRANSFER_DST_BIT |
-			VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT |
+			(useRayQuery ?
-			VK_BUFFER_USAGE_ACCELERATION_STRUCTURE_BUILD_INPUT_READ_ONLY_BIT_KHR)
+				VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT |
-		.Size(indexbuffersize)
+				VK_BUFFER_USAGE_ACCELERATION_STRUCTURE_BUILD_INPUT_READ_ONLY_BIT_KHR : 0) |
+			VK_BUFFER_USAGE_STORAGE_BUFFER_BIT)
+		.Size((size_t)Mesh->MeshElements.Size() * sizeof(uint32_t))
 		.DebugName("indexBuffer")
 		.Create(fb->device.get());
 
-	fb->GetCommands()->GetTransferCommands()->copyBuffer(transferBuffer.get(), vertexBuffer.get(), vertexoffset);
+	nodesBuffer = BufferBuilder()
-	fb->GetCommands()->GetTransferCommands()->copyBuffer(transferBuffer.get(), indexBuffer.get(), indexoffset);
+		.Usage(VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT)
+		.Size(sizeof(CollisionNodeBufferHeader) + nodes.size() * sizeof(CollisionNode))
+		.DebugName("nodesBuffer")
+		.Create(fb->device.get());
+
+	transferBuffer = BufferTransfer()
+		.AddBuffer(vertexBuffer.get(), vertices.data(), vertices.size() * sizeof(FVector4))
+		.AddBuffer(indexBuffer.get(), Mesh->MeshElements.Data(), (size_t)Mesh->MeshElements.Size() * sizeof(uint32_t))
+		.AddBuffer(nodesBuffer.get(), &nodesHeader, sizeof(CollisionNodeBufferHeader), nodes.data(), nodes.size() * sizeof(CollisionNode))
+		.Execute(fb->device.get(), fb->GetCommands()->GetTransferCommands());
 
-	// Finish transfer before using it for building
 	PipelineBarrier()
 		.AddMemory(VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT)
-		.Execute(fb->GetCommands()->GetTransferCommands(), VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ACCELERATION_STRUCTURE_BUILD_BIT_KHR);
+		.Execute(fb->GetCommands()->GetTransferCommands(), VK_PIPELINE_STAGE_TRANSFER_BIT, useRayQuery ? VK_PIPELINE_STAGE_ACCELERATION_STRUCTURE_BUILD_BIT_KHR : VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT);
 }
 
 void VkRaytrace::CreateBottomLevelAccelerationStructure()
@@ -146,9 +155,9 @@ void VkRaytrace::CreateBottomLevelAccelerationStructure()
 	accelStructBLDesc.geometryType = VK_GEOMETRY_TYPE_TRIANGLES_KHR;
 	accelStructBLDesc.flags = VK_GEOMETRY_OPAQUE_BIT_KHR;
 	accelStructBLDesc.geometry.triangles = { VK_STRUCTURE_TYPE_ACCELERATION_STRUCTURE_GEOMETRY_TRIANGLES_DATA_KHR };
-	accelStructBLDesc.geometry.triangles.vertexFormat = VK_FORMAT_R32G32B32_SFLOAT;
+	accelStructBLDesc.geometry.triangles.vertexFormat = VK_FORMAT_R32G32B32A32_SFLOAT;
 	accelStructBLDesc.geometry.triangles.vertexData.deviceAddress = vertexBuffer->GetDeviceAddress();
-	accelStructBLDesc.geometry.triangles.vertexStride = sizeof(FVector3);
+	accelStructBLDesc.geometry.triangles.vertexStride = sizeof(FVector4);
 	accelStructBLDesc.geometry.triangles.indexType = VK_INDEX_TYPE_UINT32;
 	accelStructBLDesc.geometry.triangles.indexData.deviceAddress = indexBuffer->GetDeviceAddress();
 	accelStructBLDesc.geometry.triangles.maxVertex = Mesh->MeshVertices.Size() - 1;
@@ -277,3 +286,78 @@ void VkRaytrace::CreateTopLevelAccelerationStructure()
 		.AddMemory(VK_ACCESS_ACCELERATION_STRUCTURE_WRITE_BIT_KHR, VK_ACCESS_SHADER_READ_BIT)
 		.Execute(fb->GetCommands()->GetTransferCommands(), VK_PIPELINE_STAGE_ACCELERATION_STRUCTURE_BUILD_BIT_KHR, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT);
 }
+
+std::vector<CollisionNode> VkRaytrace::CreateCollisionNodes()
+{
+	std::vector<CollisionNode> nodes;
+	nodes.reserve(Mesh->Collision->get_nodes().size());
+	for (const auto& node : Mesh->Collision->get_nodes())
+	{
+		CollisionNode info;
+		info.center = node.aabb.Center;
+		info.extents = node.aabb.Extents;
+		info.left = node.left;
+		info.right = node.right;
+		info.element_index = node.element_index;
+		nodes.push_back(info);
+	}
+	if (nodes.empty()) // vulkan doesn't support zero byte buffers
+		nodes.push_back(CollisionNode());
+	return nodes;
+}
+
+
+/////////////////////////////////////////////////////////////////////////////
+
+BufferTransfer& BufferTransfer::AddBuffer(VulkanBuffer* buffer, const void* data, size_t size)
+{
+	bufferCopies.push_back({ buffer, data, size, nullptr, 0 });
+	return *this;
+}
+
+BufferTransfer& BufferTransfer::AddBuffer(VulkanBuffer* buffer, const void* data0, size_t size0, const void* data1, size_t size1)
+{
+	bufferCopies.push_back({ buffer, data0, size0, data1, size1 });
+	return *this;
+}
+
+std::unique_ptr<VulkanBuffer> BufferTransfer::Execute(VulkanDevice* device, VulkanCommandBuffer* cmdbuffer)
+{
+	size_t transferbuffersize = 0;
+	for (const auto& copy : bufferCopies)
+		transferbuffersize += copy.size0 + copy.size1;
+
+	if (transferbuffersize == 0)
+		return nullptr;
+
+	auto transferBuffer = BufferBuilder()
+		.Usage(VK_BUFFER_USAGE_TRANSFER_SRC_BIT, VMA_MEMORY_USAGE_CPU_ONLY)
+		.Size(transferbuffersize)
+		.DebugName("BufferTransfer.transferBuffer")
+		.Create(device);
+
+	uint8_t* data = (uint8_t*)transferBuffer->Map(0, transferbuffersize);
+	size_t pos = 0;
+	for (const auto& copy : bufferCopies)
+	{
+		memcpy(data + pos, copy.data0, copy.size0);
+		pos += copy.size0;
+		memcpy(data + pos, copy.data1, copy.size1);
+		pos += copy.size1;
+	}
+	transferBuffer->Unmap();
+
+	pos = 0;
+	for (const auto& copy : bufferCopies)
+	{
+		if (copy.size0 > 0)
+			cmdbuffer->copyBuffer(transferBuffer.get(), copy.buffer, pos, 0, copy.size0);
+		pos += copy.size0;
+
+		if (copy.size1 > 0)
+			cmdbuffer->copyBuffer(transferBuffer.get(), copy.buffer, pos, copy.size0, copy.size1);
+		pos += copy.size1;
+	}
+
+	return transferBuffer;
+}

src/common/rendering/vulkan/renderer/vk_raytrace.h

@@ -6,6 +6,26 @@
 
 class VulkanRenderDevice;
 
+struct CollisionNodeBufferHeader
+{
+	int root;
+	int padding1;
+	int padding2;
+	int padding3;
+};
+
+struct CollisionNode
+{
+	FVector3 center;
+	float padding1;
+	FVector3 extents;
+	float padding2;
+	int left;
+	int right;
+	int element_index;
+	int padding3;
+};
+
 class VkRaytrace
 {
 public:
@@ -14,6 +34,9 @@ public:
 	void SetLevelMesh(hwrenderer::LevelMesh* mesh);
 
 	VulkanAccelerationStructure* GetAccelStruct() { return tlAccelStruct.get(); }
+	VulkanBuffer* GetVertexBuffer() { return vertexBuffer.get(); }
+	VulkanBuffer* GetIndexBuffer() { return indexBuffer.get(); }
+	VulkanBuffer* GetNodeBuffer() { return nodesBuffer.get(); }
 
 private:
 	void Reset();
@@ -22,14 +45,19 @@ private:
 	void CreateBottomLevelAccelerationStructure();
 	void CreateTopLevelAccelerationStructure();
 
+	std::vector<CollisionNode> CreateCollisionNodes();
+
 	VulkanRenderDevice* fb = nullptr;
 
+	bool useRayQuery = true;
+
 	hwrenderer::LevelMesh NullMesh;
 	hwrenderer::LevelMesh* Mesh = nullptr;
 
 	std::unique_ptr<VulkanBuffer> vertexBuffer;
 	std::unique_ptr<VulkanBuffer> indexBuffer;
 	std::unique_ptr<VulkanBuffer> transferBuffer;
+	std::unique_ptr<VulkanBuffer> nodesBuffer;
 
 	std::unique_ptr<VulkanBuffer> blScratchBuffer;
 	std::unique_ptr<VulkanBuffer> blAccelStructBuffer;
@@ -41,3 +69,22 @@ private:
 	std::unique_ptr<VulkanBuffer> tlAccelStructBuffer;
 	std::unique_ptr<VulkanAccelerationStructure> tlAccelStruct;
 };
+
+class BufferTransfer
+{
+public:
+	BufferTransfer& AddBuffer(VulkanBuffer* buffer, const void* data, size_t size);
+	BufferTransfer& AddBuffer(VulkanBuffer* buffer, const void* data0, size_t size0, const void* data1, size_t size1);
+	std::unique_ptr<VulkanBuffer> Execute(VulkanDevice* device, VulkanCommandBuffer* cmdbuffer);
+
+private:
+	struct BufferCopy
+	{
+		VulkanBuffer* buffer;
+		const void* data0;
+		size_t size0;
+		const void* data1;
+		size_t size1;
+	};
+	std::vector<BufferCopy> bufferCopies;
+};

wadsrc/static/shaders/scene/layout_shared.glsl

@@ -1,8 +1,32 @@
 
 layout(set = 0, binding = 0) uniform sampler2D ShadowMap;
 layout(set = 0, binding = 1) uniform sampler2DArray LightMap;
-#if defined(USE_RAYTRACE) && defined(SUPPORTS_RAYQUERY)
+#if defined(USE_RAYTRACE)
+#if defined(SUPPORTS_RAYQUERY)
 layout(set = 0, binding = 2) uniform accelerationStructureEXT TopLevelAS;
+#else
+struct CollisionNode
+{
+	vec3 center;
+	float padding1;
+	vec3 extents;
+	float padding2;
+	int left;
+	int right;
+	int element_index;
+	int padding3;
+};
+layout(std430, set = 0, binding = 2) buffer NodeBuffer
+{
+	int nodesRoot;
+	int nodebufferPadding1;
+	int nodebufferPadding2;
+	int nodebufferPadding3;
+	CollisionNode nodes[];
+};
+layout(std430, set = 0, binding = 3) buffer VertexBuffer { vec4 vertices[]; };
+layout(std430, set = 0, binding = 4) buffer ElementBuffer { int elements[]; };
+#endif
 #endif
 
 // This must match the HWViewpointUniforms struct

wadsrc/static/shaders/scene/light_shadow.glsl

@@ -1,7 +1,9 @@
 
 // Check if light is in shadow
 
-#if defined(USE_RAYTRACE) && defined(SUPPORTS_RAYQUERY)
+#if defined(USE_RAYTRACE)
+
+#if defined(SUPPORTS_RAYQUERY)
 
 bool traceHit(vec3 origin, vec3 direction, float dist)
 {
@@ -11,6 +13,160 @@ bool traceHit(vec3 origin, vec3 direction, float dist)
 	return rayQueryGetIntersectionTypeEXT(rayQuery, true) != gl_RayQueryCommittedIntersectionNoneEXT;
 }
 
+#else
+
+struct RayBBox
+{
+	vec3 start, end;
+	vec3 c, w, v;
+};
+
+RayBBox create_ray(vec3 ray_start, vec3 ray_end)
+{
+	RayBBox ray;
+	ray.start = ray_start;
+	ray.end = ray_end;
+	ray.c = (ray_start + ray_end) * 0.5;
+	ray.w = ray_end - ray.c;
+	ray.v = abs(ray.w);
+	return ray;
+}
+
+bool overlap_bv_ray(RayBBox ray, int a)
+{
+	vec3 v = ray.v;
+	vec3 w = ray.w;
+	vec3 h = nodes[a].extents;
+	vec3 c = ray.c - nodes[a].center;
+
+	if (abs(c.x) > v.x + h.x ||
+		abs(c.y) > v.y + h.y ||
+		abs(c.z) > v.z + h.z)
+	{
+		return false;
+	}
+
+	if (abs(c.y * w.z - c.z * w.y) > h.y * v.z + h.z * v.y ||
+		abs(c.x * w.z - c.z * w.x) > h.x * v.z + h.z * v.x ||
+		abs(c.x * w.y - c.y * w.x) > h.x * v.y + h.y * v.x)
+	{
+		return false;
+	}
+
+	return true;
+}
+
+#define FLT_EPSILON 1.192092896e-07F // smallest such that 1.0+FLT_EPSILON != 1.0
+
+float intersect_triangle_ray(RayBBox ray, int a, out float barycentricB, out float barycentricC)
+{
+	int start_element = nodes[a].element_index;
+
+	vec3 p[3];
+	p[0] = vertices[elements[start_element]].xyz;
+	p[1] = vertices[elements[start_element + 1]].xyz;
+	p[2] = vertices[elements[start_element + 2]].xyz;
+
+	// Moeller-Trumbore ray-triangle intersection algorithm:
+
+	vec3 D = ray.end - ray.start;
+
+	// Find vectors for two edges sharing p[0]
+	vec3 e1 = p[1] - p[0];
+	vec3 e2 = p[2] - p[0];
+
+	// Begin calculating determinant - also used to calculate u parameter
+	vec3 P = cross(D, e2);
+	float det = dot(e1, P);
+
+	// Backface check
+	//if (det < 0.0f)
+	//	return 1.0f;
+
+	// If determinant is near zero, ray lies in plane of triangle
+	if (det > -FLT_EPSILON && det < FLT_EPSILON)
+		return 1.0f;
+
+	float inv_det = 1.0f / det;
+
+	// Calculate distance from p[0] to ray origin
+	vec3 T = ray.start - p[0];
+
+	// Calculate u parameter and test bound
+	float u = dot(T, P) * inv_det;
+
+	// Check if the intersection lies outside of the triangle
+	if (u < 0.f || u > 1.f)
+		return 1.0f;
+
+	// Prepare to test v parameter
+	vec3 Q = cross(T, e1);
+
+	// Calculate V parameter and test bound
+	float v = dot(D, Q) * inv_det;
+
+	// The intersection lies outside of the triangle
+	if (v < 0.f || u + v  > 1.f)
+		return 1.0f;
+
+	float t = dot(e2, Q) * inv_det;
+	if (t <= FLT_EPSILON)
+		return 1.0f;
+
+	// Return hit location on triangle in barycentric coordinates
+	barycentricB = u;
+	barycentricC = v;
+	
+	return t;
+}
+
+bool is_leaf(int node_index)
+{
+	return nodes[node_index].element_index != -1;
+}
+
+bool TraceAnyHit(vec3 origin, float tmin, vec3 dir, float tmax)
+{
+	if (tmax <= 0.0f)
+		return false;
+
+	RayBBox ray = create_ray(origin, origin + dir * tmax);
+	tmin /= tmax;
+
+	int stack[64];
+	int stackIndex = 0;
+	stack[stackIndex++] = nodesRoot;
+	do
+	{
+		int a = stack[--stackIndex];
+		if (overlap_bv_ray(ray, a))
+		{
+			if (is_leaf(a))
+			{
+				float baryB, baryC;
+				float t = intersect_triangle_ray(ray, a, baryB, baryC);
+				if (t >= tmin && t < 1.0)
+				{
+					return true;
+				}
+			}
+			else
+			{
+				stack[stackIndex++] = nodes[a].right;
+				stack[stackIndex++] = nodes[a].left;
+			}
+		}
+	} while (stackIndex > 0);
+	return false;
+}
+
+bool traceHit(vec3 origin, vec3 direction, float dist)
+{
+	return TraceAnyHit(origin, 0.01f, direction, dist);
+}
+
+#endif
+
 vec2 softshadow[9 * 3] = vec2[](
 	vec2( 0.0, 0.0),
 	vec2(-2.0,-2.0),