#include "ecs.hpp"

#include "game-settings.hpp"
#include "math-helpers.hpp"
#include "memory.hpp"
#include "physics.hpp"
#include "vendor/glm_include.hpp"
#include "window.hpp"

#include <btBulletDynamicsCommon.h>
#include <BulletCollision/CollisionShapes/btConvexHullShape.h>
#include <glm/gtc/type_ptr.hpp>

TypeSafeInt_B(EntityHandle);
const uint64_t maxBucketItemCount = 256;

struct EntityBucket {
	Entity entities[maxBucketItemCount];
};
struct GrBindsBucket {
	CompGrBinds compGrBinds[maxBucketItemCount];
};
struct InstanceBucket {
	CompInstance instances[maxBucketItemCount];
};

/*
 * Entities that have been marked for removal by calling ECS_MarkForRemoval
 *
 * Used to build the other "removal" lists.
 */
DynArray<EntityHandle> entitiesMarkedForRemoval{16};

/*
 * Public list of entities that will be removed next tick
 *
 * Entity or child of entity that had ECS_MarkForRemoval called
 */
DynArray<EntityHandle> EntitiesToBeRemoved{16}; // public

/*
 * The entities being removed this tick
 *
 * Each of these entities has gone a full tick in the "to be removed" state
 */
DynArray<Entity> entitiesYetToBeRemoved{0, nullptr};

DynArray<EntityHandle> EntitiesWithExcessInstances{16};

BucketContainer<EntityBucket, EntityHandle_T> Entities_BucketContainer{};
BucketContainer<GrBindsBucket, GrBindsHandle_T> Comp_GrBinds_BucketContainer{};
BucketContainer<InstanceBucket, InstanceHandle_T> Comp_Instance_BucketContainer{};

void ECS_GetEntity_Inner(EntityHandle entHandle, Entity*& ent) {
	EntityHandle_T entHandle_t{static_cast<EntityHandle_T>(entHandle)};
	assert(entHandle_t != EntityHandle_T_MAX && "Unknown entity handle");
	auto b = Buckets_GetBucketIndex(static_cast<EntityHandle_T>(entHandle));
	auto e = Buckets_GetItemIndex(static_cast<EntityHandle_T>(entHandle));
	ent = &Entities_BucketContainer.buckets[b].entities[e];
}

void ECS_Init() {
	Buckets_Init(Entities_BucketContainer);
	Buckets_Init(Comp_GrBinds_BucketContainer);
	Buckets_Init(Comp_Instance_BucketContainer);
}

uint64_t ECS_GetEntities_BucketCount() {
	return Entities_BucketContainer.bucketCounter + 1;
}
Entity *ECS_GetEntities(uint64_t bucketIndex, uint64_t &itemCount) {
	assert(bucketIndex <= Entities_BucketContainer.bucketCounter);
	itemCount = bucketIndex == Entities_BucketContainer.bucketCounter ? Entities_BucketContainer.itemCounter >> 32 : maxBucketItemCount;
	return Entities_BucketContainer.buckets[bucketIndex].entities;
}

EntityHandle ECS_CreateEntity_Inner(EntityHandle parentEntHandle) {
	EntityHandle_T entityHandle_T{Buckets_NewHandle<EntityBucket>(maxBucketItemCount, Entities_BucketContainer)};
	EntityHandle entityHandle{entityHandle_T};

	auto b = Buckets_GetBucketIndex(static_cast<EntityHandle_T>(entityHandle));
	auto e = Buckets_GetItemIndex(static_cast<EntityHandle_T>(entityHandle));
	Entity *entity = &Entities_BucketContainer.buckets[b].entities[e];
	entity = new (entity) Entity{};
	entity->handle = entityHandle;
	entity->parentHandle = parentEntHandle;

	return entityHandle;
}

EntityHandle ECS_CreateEntity(EntityHandle parentEntHandle) {
	return ECS_CreateEntity_Inner(parentEntHandle);
}

void ECS_MarkForRemoval(EntityHandle entityHandle) {
	auto b = Buckets_GetBucketIndex(static_cast<EntityHandle_T>(entityHandle));
	auto e = Buckets_GetItemIndex(static_cast<EntityHandle_T>(entityHandle));
	const Entity *ent = &Entities_BucketContainer.buckets[b].entities[e];
	assert(ent->isMarkedForRemoval == false && "Entity already marked for removal");
	entitiesMarkedForRemoval.Push(entityHandle);
}

void ECS_Tick_Early(double delta) {
	// these reserves might happen 1 tick early, but that's fine
	entitiesYetToBeRemoved.Reserve(entitiesMarkedForRemoval.Count());
	EntitiesToBeRemoved.Resize(entitiesMarkedForRemoval.Count());
	memcpy(EntitiesToBeRemoved.GetPtr(), entitiesMarkedForRemoval.GetPtr(), sizeof(EntityHandle) * entitiesMarkedForRemoval.Count());
	entitiesYetToBeRemoved.Resize(0);
	for (long b = 0; b <= Entities_BucketContainer.bucketCounter; ++b) {
		uint64_t entCount = b == Entities_BucketContainer.bucketCounter ? Entities_BucketContainer.itemCounter >> 32 : maxBucketItemCount;
		for (long e  = 0; e < entCount; ++e) {
			Entity *ent = &Entities_BucketContainer.buckets[b].entities[e];
			if (ent->handle == EntityHandle_MAX) continue;
			if (ent->isMarkedForRemoval) {
				entitiesYetToBeRemoved.Push(*ent);
				ent->handle = EntityHandle_MAX;
				ent->parentHandle = EntityHandle_MAX;
				ent->isMarkedForRemoval = false;
			} else if (EntitiesToBeRemoved.Has(ent->handle)) {
				ent->isMarkedForRemoval = true;
		  } else if (EntitiesToBeRemoved.Has(ent->parentHandle)) {
				ent->isMarkedForRemoval = true;
				EntitiesToBeRemoved.Push(ent->handle);
			}
		}
	}
	for (long e = 0; e < entitiesYetToBeRemoved.Count(); ++e) {
		Entity &clonedEnt = entitiesYetToBeRemoved[e];
		CompGrBinds *grBinds = nullptr;
		InstanceBucket *instBucket = nullptr;
		CompInstance *inst = nullptr;
		uint64_t instBucketIndex = 0;
		if (clonedEnt.instanceHandle != InstanceHandle_MAX) {
			InstanceHandle_T instanceHandle_t{static_cast<InstanceHandle_T>(clonedEnt.instanceHandle)};
			auto b = Buckets_GetBucketIndex(instanceHandle_t);
			auto i = Buckets_GetItemIndex(instanceHandle_t);
			instBucketIndex = b;
			instBucket = &Comp_Instance_BucketContainer.buckets[b];
			inst = &instBucket->instances[i];
			GrBindsHandle_T grBindsHandle_t = static_cast<GrBindsHandle_T>(inst->grBindsHandle);
			b = Buckets_GetBucketIndex(grBindsHandle_t);
			i = Buckets_GetItemIndex(grBindsHandle_t);
			grBinds = &Comp_GrBinds_BucketContainer.buckets[b].compGrBinds[i];
		}
		if (grBinds == nullptr && clonedEnt.grBindsHandle != GrBindsHandle_MAX) {
			GrBindsHandle_T bindsHandle_t{static_cast<GrBindsHandle_T>(clonedEnt.grBindsHandle)};
			auto b = Buckets_GetBucketIndex(bindsHandle_t);
			auto i = Buckets_GetItemIndex(bindsHandle_t);
			grBinds = &Comp_GrBinds_BucketContainer.buckets[b].compGrBinds[i];
		}
		if (inst != nullptr) {
			assert(grBinds != nullptr);
			for (long bi = instBucketIndex; bi <= Comp_Instance_BucketContainer.bucketCounter; ++bi) {
				uint64_t instCounter = 0;
				if (bi == Comp_Instance_BucketContainer.bucketCounter) {
					instCounter = Comp_Instance_BucketContainer.itemCounter >> 32;
				} else {
					instCounter = maxBucketItemCount;
				}
				auto &bucket = Comp_Instance_BucketContainer.buckets[bi];
				for (long ii = 0; ii < instCounter; ++ii) {
					if (bucket.instances[ii].entHandle == EntityHandle_MAX) {
						continue;
					}
					if (bucket.instances[ii].grBindsHandle != clonedEnt.grBindsHandle) {
						continue;
					}
					if (bucket.instances[ii].instanceHandle < clonedEnt.instanceHandle) {
						continue;
					}
					bucket.instances[ii].index -= 1;
					bucket.instances[ii].isNeedingUpdated = true;
				}
			}
			grBinds->instanceCounter -= 1;
			inst->entHandle = EntityHandle_MAX;
			inst->grBindsHandle = GrBindsHandle_MAX;
			inst->index = ECS_UNSET_VAL;
			inst->instanceHandle = InstanceHandle_MAX;
			inst->isNeedingUpdated = false;
			BtDynamicsWorld->removeRigidBody(inst->bt.rigidBody);
			Pke_Delete<btDefaultMotionState>(inst->bt.motionState, MemBkt_Bullet);
			Pke_Delete<btRigidBody>(inst->bt.rigidBody, MemBkt_Bullet);
		} else if (grBinds != nullptr) {
			/*
			 * 2023-09-05 JB note - the Vulkan assets (device memory, buffers,
			 *  pipeline layout, and descriptor set) are unloaded elsewhere, just,
			 *  as they were created elsewhere.
			 */
		}
	}
	entitiesMarkedForRemoval.Resize(0);
}

void ECS_Tick(double delta) {
	int32_t physicsTickCount = Physics_Tick(delta);
	if (physicsTickCount != 0) {
		for (long b = 0; b <= Comp_Instance_BucketContainer.bucketCounter; ++b) {
			auto &bkt = Comp_Instance_BucketContainer.buckets[b];
			long count = Comp_Instance_BucketContainer.bucketCounter == b ? Comp_Instance_BucketContainer.itemCounter >> 32 : maxBucketItemCount;
			for (uint32_t i = 0; i < count; ++i) {
				auto &inst = bkt.instances[i];
				if (inst.entHandle == EntityHandle_MAX)
					continue;
				auto activationState = inst.bt.rigidBody->getActivationState();
				if (activationState == ISLAND_SLEEPING || activationState == DISABLE_SIMULATION || activationState == WANTS_DEACTIVATION) {
					continue;
				}
				inst.isNeedingUpdated = true;
			}
		}
	}
}

struct InstanceBufferCopyChunk {
	uint64_t startingIndex;
	uint64_t endingIndex;
	DynArray<glm::mat4> mats{0, nullptr};
	VkBufferCopy dstBufferCopy;
};
struct InstanceBufferCopy {
	CompGrBinds *grBinds = nullptr;
	VkDeviceSize runningSize = 0;
	DynArray<InstanceBufferCopyChunk> chunks{4};
};
void ECS_Tick_Late(double delta) {
	DynArray<InstanceBufferCopy> bufferUpdates{0, nullptr};
	for (long b = 0; b <= Comp_Instance_BucketContainer.bucketCounter; ++b) {
		auto &bkt = Comp_Instance_BucketContainer.buckets[b];
		long count = Comp_Instance_BucketContainer.bucketCounter == b ? Comp_Instance_BucketContainer.itemCounter >> 32 : maxBucketItemCount;
		for (uint32_t i = 0; i < count; ++i) {
			auto &inst = bkt.instances[i];
			if (inst.isNeedingUpdated == false)
				continue;
			if (inst.entHandle == EntityHandle_MAX)
				continue;

			GrBindsHandle_T grBindsHandle_t{static_cast<GrBindsHandle_T>(inst.grBindsHandle)};
			auto b_inner = Buckets_GetBucketIndex(grBindsHandle_t);
			auto bi_inner = Buckets_GetItemIndex(grBindsHandle_t);
			auto &grBinds = Comp_GrBinds_BucketContainer.buckets[b_inner].compGrBinds[bi_inner];

			InstanceBufferCopy *bfrUpdate = nullptr;
			for (long u = 0; u < bufferUpdates.Count(); ++u) {
				if (bufferUpdates[u].grBinds->grBindsHandle == inst.grBindsHandle) {
					bfrUpdate = &bufferUpdates[u];
				}
			}
			if (bfrUpdate == nullptr) {
				bufferUpdates.Push({
					.grBinds = &grBinds,
				});
				bfrUpdate = &bufferUpdates[bufferUpdates.Count() - 1];
			}

			InstanceBufferCopyChunk *chunk = nullptr;
			for (long ii = 0; ii < bfrUpdate->chunks.Count(); ++ii) {
				if (bfrUpdate->chunks[ii].endingIndex == inst.index - 1) {
					chunk = &bfrUpdate->chunks[ii];
					chunk->endingIndex += 1;
					break;
				}
			}
			if (chunk == nullptr) {
				bfrUpdate->chunks.Push({
					.startingIndex = inst.index,
					.endingIndex = inst.index,
				});
				chunk = &bfrUpdate->chunks[bfrUpdate->chunks.Count() - 1];
				chunk->dstBufferCopy.dstOffset = sizeof(glm::mat4) * inst.index;
			}

			btTransform btMatrix_posRot;
			inst.bt.motionState->getWorldTransform(btMatrix_posRot);
			float openglMatrix[16];
			btMatrix_posRot.getOpenGLMatrix(openglMatrix);
			glm::mat4 glmMat_posRot = glm::make_mat4(openglMatrix);

			glm::vec3 scale;
			BulletToGlm(inst.bt.rigidBody->getCollisionShape()->getLocalScaling(), scale);

			chunk->mats.Push(glm::scale(glmMat_posRot, scale));
			bfrUpdate->runningSize += sizeof(glm::mat4);
			inst.isNeedingUpdated = false;
		}
	}
	while (bufferUpdates.Count() > 0) {
		InstanceBufferCopy ibc = bufferUpdates.Pop();

		VkDeviceSize instanceBytes = sizeof(glm::mat4);
		VkDeviceSize byteCount = ibc.runningSize;
		VkBuffer transferBuffer;
		VkDeviceMemory transferMemory;
		void *deviceData;
		BeginTransferBuffer(byteCount, transferBuffer, transferMemory, deviceData);
		VkDeviceSize runningOffset = 0;
		for (long i = 0; i < ibc.chunks.Count(); ++i) {
			auto *chunk = &ibc.chunks[i];
			memcpy(static_cast<char *>(deviceData) + runningOffset, chunk->mats.GetPtr(), byteCount);
			chunk->dstBufferCopy.srcOffset = runningOffset;
			chunk->dstBufferCopy.size = instanceBytes * (chunk->endingIndex - chunk->startingIndex + 1);
			runningOffset += chunk->dstBufferCopy.size;
		}
		assert(runningOffset == ibc.runningSize);

		{
			VkCommandBufferBeginInfo cbbi;
			cbbi.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
			cbbi.pNext = nullptr;
			cbbi.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
			cbbi.pInheritanceInfo = nullptr;

			vkBeginCommandBuffer(transferCommandBuffer, &cbbi);

			for (long i = 0; i < ibc.chunks.Count(); ++i) {
				vkCmdCopyBuffer(transferCommandBuffer, transferBuffer, ibc.grBinds->instanceBuffer, 1, &ibc.chunks[i].dstBufferCopy);
			}

			vkEndCommandBuffer(transferCommandBuffer);

			VkSubmitInfo vkSubmitInfo{};
			vkSubmitInfo.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
			vkSubmitInfo.pNext = nullptr;
			vkSubmitInfo.waitSemaphoreCount = 0;
			vkSubmitInfo.pWaitSemaphores = nullptr;
			vkSubmitInfo.pWaitDstStageMask = nullptr;
			vkSubmitInfo.commandBufferCount = 1;
			vkSubmitInfo.pCommandBuffers = &transferCommandBuffer;
			vkSubmitInfo.signalSemaphoreCount = 0;
			vkSubmitInfo.pSignalSemaphores = nullptr;
			vkQueueSubmit(transferQueue, 1, &vkSubmitInfo, nullptr);
			vkQueueWaitIdle(transferQueue);
			vkResetCommandBuffer(transferCommandBuffer, 0);
		}
		EndTransferBuffer(transferBuffer, transferMemory);
	}
}

CompGrBinds &ECS_CreateGrBinds(EntityHandle entHandle) {
	assert(entHandle != EntityHandle_MAX);

	Entity *ent = nullptr;
	ECS_GetEntity_Inner(entHandle, ent);

	assert(ent->grBindsHandle == GrBindsHandle_MAX);

	GrBindsHandle_T newHandle{Buckets_NewHandle(maxBucketItemCount, Comp_GrBinds_BucketContainer)};
	ent->grBindsHandle = GrBindsHandle{newHandle};
	auto b = Buckets_GetBucketIndex(newHandle);
	auto i = Buckets_GetItemIndex(newHandle);
	auto *comp = &Comp_GrBinds_BucketContainer.buckets[b].compGrBinds[i];
	comp = new (comp) CompGrBinds{};
	comp->entHandle = entHandle;
	comp->grBindsHandle = ent->grBindsHandle;
	return *comp;
}

CompGrBinds *ECS_GetGrBinds(EntityHandle entHandle) {
	Entity *ent = nullptr;
	ECS_GetEntity_Inner(entHandle, ent);
	if (ent->grBindsHandle == GrBindsHandle_MAX) return nullptr;
	GrBindsHandle_T grBindsHandle_t{static_cast<GrBindsHandle_T>(ent->grBindsHandle)};
	auto b = Buckets_GetBucketIndex(grBindsHandle_t);
	auto i = Buckets_GetItemIndex(grBindsHandle_t);
	return &Comp_GrBinds_BucketContainer.buckets[b].compGrBinds[i];
}
CompGrBinds *ECS_GetGrBinds(GrBindsHandle grBindsHandle) {
	if (grBindsHandle == GrBindsHandle_MAX) return nullptr;
	GrBindsHandle_T grBindsHandle_t{static_cast<GrBindsHandle_T>(grBindsHandle)};
	auto b = Buckets_GetBucketIndex(grBindsHandle_t);
	auto i = Buckets_GetItemIndex(grBindsHandle_t);
	return &Comp_GrBinds_BucketContainer.buckets[b].compGrBinds[i];
}

uint64_t ECS_GetGrBinds_BucketCount() {
	return Comp_GrBinds_BucketContainer.bucketCounter + 1;
}

CompGrBinds *ECS_GetGrBinds(uint64_t bucketIndex, uint64_t &itemCount) {
	if (bucketIndex == Comp_GrBinds_BucketContainer.bucketCounter) {
		itemCount = Comp_GrBinds_BucketContainer.itemCounter >> 32;
	} else {
		itemCount = maxBucketItemCount;
	}
	return Comp_GrBinds_BucketContainer.buckets[bucketIndex].compGrBinds;
}

CompInstance &ECS_CreateInstance(EntityHandle entHandle, EntityHandle entityTypeEntityHandle) {
	assert(entHandle != EntityHandle_MAX);
	assert(entityTypeEntityHandle != EntityHandle_MAX);

	Entity *ent = nullptr;
	Entity *entTypeEnt = nullptr;
	ECS_GetEntity_Inner(entHandle, ent);
	ECS_GetEntity_Inner(entityTypeEntityHandle, entTypeEnt);

	assert(ent->grBindsHandle == GrBindsHandle_MAX);
	assert(entTypeEnt->grBindsHandle != GrBindsHandle_MAX);

	InstanceHandle_T newHandle{Buckets_NewHandle(maxBucketItemCount, Comp_Instance_BucketContainer)};
	ent->instanceHandle = InstanceHandle{newHandle};
	GrBindsHandle_T grBindsHandle_t{static_cast<GrBindsHandle_T>(entTypeEnt->grBindsHandle)};

	auto b = Buckets_GetBucketIndex(newHandle);
	auto i = Buckets_GetItemIndex(newHandle);
	auto *instBkt = &Comp_Instance_BucketContainer.buckets[b];
	auto *comp = &instBkt->instances[i];
	comp = new (comp) CompInstance{};
	comp->entHandle = entHandle;

	auto b2 = Buckets_GetBucketIndex(grBindsHandle_t);
	auto i2 = Buckets_GetItemIndex(grBindsHandle_t);
	auto *grBinds = &Comp_GrBinds_BucketContainer.buckets[b2].compGrBinds[i2];

	comp->grBindsHandle = grBinds->grBindsHandle;
	comp->instanceHandle = ent->instanceHandle;
	comp->index = grBinds->instanceCounter++;
	comp->isNeedingUpdated = true;

	if (grBinds->instanceCounter > grBinds->instanceBufferMaxCount) {
		EntitiesWithExcessInstances.Push(grBinds->entHandle);
	}

	return *comp;
}

CompInstance *ECS_GetInstance(EntityHandle entHandle) {
	Entity *ent = nullptr;
	ECS_GetEntity_Inner(entHandle, ent);
	
	if (ent->instanceHandle == InstanceHandle_MAX)
		return nullptr;

	InstanceHandle_T instanceHandle_t{static_cast<InstanceHandle_T>(ent->instanceHandle)};
	auto b = Buckets_GetBucketIndex(instanceHandle_t);
	auto i = Buckets_GetItemIndex(instanceHandle_t);
	auto *inst = &Comp_Instance_BucketContainer.buckets[b].instances[i];

	return inst;
}

void ECS_UpdateInstance(EntityHandle entHandle, const InstPos &instPos, bool overridePhysics) {
	Entity *ent = nullptr;
	ECS_GetEntity_Inner(entHandle, ent);

	if (ent->instanceHandle == InstanceHandle_MAX)
		return;

	InstanceHandle_T instanceHandle_t{static_cast<InstanceHandle_T>(ent->instanceHandle)};
	auto b = Buckets_GetBucketIndex(instanceHandle_t);
	auto i = Buckets_GetItemIndex(instanceHandle_t);
	auto *inst = &Comp_Instance_BucketContainer.buckets[b].instances[i];

	if (BtDynamicsWorld && overridePhysics) {
		btVector3 localInertia(0, 0, 0);
		inst->bt.rigidBody->getCollisionShape()->calculateLocalInertia(instPos.mass, localInertia);
		inst->bt.rigidBody->setMassProps(instPos.mass, localInertia);
		inst->bt.rigidBody->getMotionState()->setWorldTransform(instPos.posRot);
		inst->bt.rigidBody->setWorldTransform(instPos.posRot);
		inst->bt.rigidBody->getCollisionShape()->setLocalScaling(instPos.scale);
		inst->bt.rigidBody->setLinearVelocity(btVector3(0,0,0));
		inst->bt.rigidBody->setAngularVelocity(btVector3(0,0,0));
		inst->bt.rigidBody->activate();
		inst->isNeedingUpdated = true;
	}
}

uint64_t ECS_GetInstances_BucketCount() {
	return Comp_Instance_BucketContainer.bucketCounter + 1;
}

CompInstance *ECS_GetInstances(uint64_t bucketIndex, uint64_t &itemCount) {
	if (bucketIndex == Comp_Instance_BucketContainer.bucketCounter) {
		itemCount = Comp_Instance_BucketContainer.itemCounter >> 32;
	} else {
		itemCount = maxBucketItemCount;
	}
	return Comp_Instance_BucketContainer.buckets[bucketIndex].instances;
}

void ECS_Teardown() {
	EntitiesWithExcessInstances.~DynArray();
	entitiesYetToBeRemoved.~DynArray();
	EntitiesToBeRemoved.~DynArray();
	entitiesMarkedForRemoval.~DynArray();
	Buckets_Destroy(Entities_BucketContainer);
	Buckets_Destroy(Comp_GrBinds_BucketContainer);
	Buckets_Destroy(Comp_Instance_BucketContainer);
}