cOMS/asset/AssetManagementSystem.h

/**
 * Jingga
 *
 * @copyright Jingga
 * @license   OMS License 2.0
 * @version   1.0.0
 * @link      https://jingga.app
 */
#ifndef COMS_ASSET_MANAGEMENT_SYSTEM_H
#define COMS_ASSET_MANAGEMENT_SYSTEM_H

#include <string.h>
#include "../stdlib/Types.h"
#include "Asset.h"
#include "../memory/ChunkMemory.h"
#include "../utils/TestUtils.h"
#include "../utils/BitUtils.h"
#include "../stdlib/HashMap.h"
#include "../log/DebugMemory.h"
#include "../thread/Atomic.h"

// The major asset types should have their own asset component system
// All other entities are grouped together in one asset component system
struct AssetComponent {
    ChunkMemory asset_memory;

    uint64 ram_size;
    uint64 vram_size;
    uint64 asset_count;

    // @question Do we want to add a mutex to assets. This way we don't have to lock the entire ams.
    coms_pthread_mutex_t mutex;
};

struct AssetManagementSystem {
    HashMap hash_map;

    int32 asset_component_count;
    AssetComponent* asset_components;
};

inline
void ams_create(AssetManagementSystem* ams, BufferMemory* buf, int32 asset_component_count, int32 count)
{
    LOG_1("Create AMS for %n assets", {{LOG_DATA_INT32, &count}});
    hashmap_create(&ams->hash_map, count, sizeof(HashEntry) + sizeof(Asset), buf);
    ams->asset_component_count = asset_component_count;
    ams->asset_components = (AssetComponent *) buffer_get_memory(buf, asset_component_count * sizeof(AssetComponent), 64, true);
}

inline
void ams_component_create(AssetComponent* ac, BufferMemory* buf, int32 chunk_size, int32 count)
{
    ASSERT_SIMPLE(chunk_size);
    LOG_1("Create AMS Component for %n assets and %n B", {{LOG_DATA_INT32, &count}, {LOG_DATA_UINT32, &chunk_size}});

    chunk_init(&ac->asset_memory, buf, count, chunk_size, 64);
    coms_pthread_mutex_init(&ac->mutex, NULL);
}

inline
void ams_component_create(AssetComponent* ac, byte* buf, int32 chunk_size, int32 count)
{
    ASSERT_SIMPLE(chunk_size);
    LOG_1("Create AMS Component for %n assets and %n B", {{LOG_DATA_INT32, &count}, {LOG_DATA_UINT32, &chunk_size}});

    ac->asset_memory.count = count;
    ac->asset_memory.chunk_size = chunk_size;
    ac->asset_memory.last_pos = 0;
    ac->asset_memory.alignment = 64;
    ac->asset_memory.memory = buf;
    ac->asset_memory.free = (uint64 *) (ac->asset_memory.memory + ac->asset_memory.chunk_size * count);

    coms_pthread_mutex_init(&ac->mutex, NULL);
}

inline
void ams_component_free(AssetComponent* ac)
{
    coms_pthread_mutex_destroy(&ac->mutex);
}

inline
void ams_free(AssetManagementSystem* ams)
{
    for (int32 i = 0; i < ams->asset_component_count; ++i) {
        ams_component_free(&ams->asset_components[i]);
    }
}

inline
uint16 ams_calculate_chunks(const AssetComponent* ac, int32 byte_size, int32 overhead)
{
    return (uint16) CEIL_DIV(byte_size + overhead, ac->asset_memory.chunk_size);
}

inline
void thrd_ams_set_loaded(Asset* asset)
{
    atomic_set_release(&asset->is_loaded, 1);
}

inline
bool thrd_ams_is_loaded(Asset* asset)
{
    return asset && atomic_get_acquire(&asset->is_loaded) > 0;
}

inline
bool thrd_ams_is_in_vram(Asset* asset)
{
    return asset && atomic_get_acquire(&asset->is_loaded)
        && (asset->state & ASSET_STATE_IN_VRAM);
}

inline
Asset* ams_get_asset(AssetManagementSystem* ams, const char* key)
{
    HashEntry* entry = hashmap_get_entry(&ams->hash_map, key);

    DEBUG_MEMORY_READ(
        (uint64) (entry ? ((Asset *) entry->value)->self : 0),
        entry ? ((Asset *) entry->value)->ram_size : 0
    );

    return entry ? (Asset *) entry->value : NULL;
}

// @performance We could probably avoid locking by adding a atomic flag to indicate if the value is valid
inline
Asset* thrd_ams_get_asset(AssetManagementSystem* ams, const char* key) {
    HashEntry* entry = hashmap_get_entry(&ams->hash_map, key);

    if (!entry || atomic_get_acquire(&((Asset *) entry->value)->is_loaded) <= 0) {
        return NULL;
    }

    DEBUG_MEMORY_READ(
        (uint64) (entry ? ((Asset *) entry->value)->self : 0),
        entry ? ((Asset *) entry->value)->ram_size : 0
    );

    return (Asset *) entry->value;
}

inline
Asset* thrd_ams_get_asset_wait(AssetManagementSystem* ams, const char* key) {
    HashEntry* entry = hashmap_get_entry(&ams->hash_map, key);

    if (!entry) {
        return NULL;
    }

    int32 state = 0;
    while (!(state = atomic_get_acquire(&((Asset *) entry->value)->is_loaded))) {}
    if (state < 0) {
        // Marked for removal
        return NULL;
    }

    DEBUG_MEMORY_READ(
        (uint64) (entry ? ((Asset *) entry->value)->self : 0),
        entry ? ((Asset *) entry->value)->ram_size : 0
    );

    return (Asset *) entry->value;
}

inline
Asset* thrd_ams_get_asset_wait(AssetManagementSystem* ams, const char* key, uint64 hash) {
    HashEntry* entry = hashmap_get_entry(&ams->hash_map, key, hash);

    if (!entry) {
        return NULL;
    }

    int32 state = 0;
    while (!(state = atomic_get_acquire(&((Asset *) entry->value)->is_loaded))) {}
    if (state < 0) {
        // Marked for removal
        return NULL;
    }

    DEBUG_MEMORY_READ(
        (uint64) (entry ? ((Asset *) entry->value)->self : 0),
        entry ? ((Asset *) entry->value)->ram_size : 0
    );

    return (Asset *) entry->value;
}

inline
Asset* thrd_ams_get_reserve_asset_wait(AssetManagementSystem* ams, byte type, const char* name, uint32 size, uint32 overhead = 0)
{
    // @bug Isn't hashmap_get_reserve unsafe for threading?
    HashEntry* entry = hashmap_get_reserve(&ams->hash_map, name);
    Asset* asset = (Asset *) entry->value;

    if (asset->self) {
        int32 state = 0;
        while (!(state = atomic_get_acquire(&((Asset *) entry->value)->is_loaded))) {}
        if (state > 0) {
            return asset;
        }
    }

    AssetComponent* ac = &ams->asset_components[type];
    uint16 elements = ams_calculate_chunks(ac, size, overhead);
    int32 free_data = chunk_reserve(&ac->asset_memory, elements);

    byte* data = chunk_get_element(&ac->asset_memory, free_data, true);

    asset->component_id = type;
    asset->self = data;
    asset->size = elements; // Crucial for freeing
    asset->ram_size = ac->asset_memory.chunk_size * elements;

    ac->vram_size += asset->vram_size;
    ac->ram_size += asset->ram_size;
    ++ac->asset_count;

    DEBUG_MEMORY_RESERVE((uintptr_t) asset, asset->ram_size);

    return asset;
}

inline
void ams_remove_asset(AssetManagementSystem* ams, AssetComponent* ac, Asset* asset, const char* name)
{
    // @todo remove from vram

    asset->is_loaded = 0;
    ac->vram_size -= asset->vram_size;
    ac->ram_size -= asset->ram_size;
    --ac->asset_count;

    hashmap_remove(&ams->hash_map, name);
    chunk_free_elements(
        &ac->asset_memory,
        chunk_id_from_memory(&ac->asset_memory, asset->self),
        asset->size
    );
}

inline
void ams_remove_asset_ram(AssetComponent* ac, Asset* asset)
{
    ac->ram_size -= asset->ram_size;
    chunk_free_elements(
        &ac->asset_memory,
        chunk_id_from_memory(&ac->asset_memory, asset->self),
        asset->size
    );
}

// @todo It would be nice if we could remove the asset by passing it as a parameter instead of the name
// The problem is there is no correlation between asset data (e.g. internal_id) and global hashmap (e.g. element_id)
// This means we would have to iterate all hashmap entries and remove it this way, which is very slow
inline
void ams_remove_asset(AssetManagementSystem* ams, const char* name)
{
    // @todo remove from vram

    Asset* asset = ams_get_asset(ams, name);
    AssetComponent* ac = &ams->asset_components[asset->component_id];

    asset->is_loaded = 0;
    ac->vram_size -= asset->vram_size;
    ac->ram_size -= asset->ram_size;
    --ac->asset_count;

    hashmap_remove(&ams->hash_map, name);
    chunk_free_elements(
        &ac->asset_memory,
        chunk_id_from_memory(&ac->asset_memory, asset->self),
        asset->size
    );
}

inline
void ams_remove_asset(AssetManagementSystem* ams, Asset* asset, const char* name)
{
    // @todo remove from vram

    AssetComponent* ac = &ams->asset_components[asset->component_id];

    asset->is_loaded = 0;
    ac->vram_size -= asset->vram_size;
    ac->ram_size -= asset->ram_size;
    --ac->asset_count;

    hashmap_remove(&ams->hash_map, name);
    chunk_free_elements(
        &ac->asset_memory,
        chunk_id_from_memory(&ac->asset_memory, asset->self),
        asset->size
    );
}

inline
void ams_remove_asset_ram(AssetManagementSystem* ams, Asset* asset)
{
    AssetComponent* ac = &ams->asset_components[asset->component_id];
    ac->ram_size -= asset->ram_size;

    chunk_free_elements(
        &ac->asset_memory,
        chunk_id_from_memory(&ac->asset_memory, asset->self),
        asset->size
    );
}

inline
void thrd_ams_remove_asset(AssetManagementSystem* ams, AssetComponent* ac, Asset* asset, const char* name)
{
    // @todo remove from vram

    asset->is_loaded = 0;
    ac->vram_size -= asset->vram_size;
    ac->ram_size -= asset->ram_size;
    --ac->asset_count;

    atomic_set_release(&asset->is_loaded, 0);
    hashmap_remove(&ams->hash_map, name);
    chunk_free_elements(
        &ac->asset_memory,
        chunk_id_from_memory(&ac->asset_memory, asset->self),
        asset->size
    );
}

void thrd_ams_remove_asset(AssetManagementSystem* ams, const char* name)
{
    HashEntry* entry = hashmap_get_entry(&ams->hash_map, name);
    Asset* asset = (Asset *) entry->value;
    atomic_set_release(&asset->is_loaded, -1);
    hashmap_remove(&ams->hash_map, name);

    AssetComponent* ac = &ams->asset_components[asset->component_id];
    chunk_free_elements(
        &ac->asset_memory,
        chunk_id_from_memory(&ac->asset_memory, asset->self),
        asset->size
    );

    ac->vram_size -= asset->vram_size;
    ac->ram_size -= asset->ram_size;
    --ac->asset_count;
}

void thrd_ams_remove_asset(AssetManagementSystem* ams, const char* name, Asset* asset)
{
    atomic_set_release(&asset->is_loaded, -1);
    hashmap_remove(&ams->hash_map, name);

    AssetComponent* ac = &ams->asset_components[asset->component_id];
    chunk_free_elements(
        &ac->asset_memory,
        chunk_id_from_memory(&ac->asset_memory, asset->self),
        asset->size
    );

    ac->vram_size -= asset->vram_size;
    ac->ram_size -= asset->ram_size;
    --ac->asset_count;
}

// @todo implement defragment command to optimize memory layout since the memory layout will become fragmented over time

Asset* ams_reserve_asset(AssetManagementSystem* ams, byte type, const char* name, uint32 size, uint32 overhead = 0)
{
    AssetComponent* ac = &ams->asset_components[type];
    uint16 elements = ams_calculate_chunks(ac, size, overhead);

    int32 free_data = chunk_reserve(&ac->asset_memory, elements);
    if (free_data < 0) {
        ASSERT_SIMPLE(free_data >= 0);
        return NULL;
    }

    byte* asset_data = chunk_get_element(&ac->asset_memory, free_data, true);
    Asset* asset = (Asset *) hashmap_reserve(&ams->hash_map, name)->value;

    asset->component_id = type;
    asset->self = asset_data;
    asset->size = elements; // Crucial for freeing
    asset->ram_size = ac->asset_memory.chunk_size * elements;

    ac->vram_size += asset->vram_size;
    ac->ram_size += asset->ram_size;
    ++ac->asset_count;

    DEBUG_MEMORY_RESERVE((uintptr_t) asset, asset->ram_size);

    return asset;
}

inline
Asset* thrd_ams_reserve_asset(AssetManagementSystem* ams, byte type, const char* name, uint32 size, uint32 overhead = 0) {
    AssetComponent* ac = &ams->asset_components[type];
    uint16 elements = ams_calculate_chunks(ac, size, overhead);

    coms_pthread_mutex_lock(&ams->asset_components[type].mutex);
    int32 free_data = chunk_reserve(&ac->asset_memory, elements);
    if (free_data < 0) {
        coms_pthread_mutex_unlock(&ams->asset_components[type].mutex);
        ASSERT_SIMPLE(free_data >= 0);

        return NULL;
    }
    coms_pthread_mutex_unlock(&ams->asset_components[type].mutex);

    byte* asset_data = chunk_get_element(&ac->asset_memory, free_data, true);

    Asset asset = {};

    asset.component_id = type;
    asset.self = asset_data;
    asset.size = elements; // Crucial for freeing
    asset.ram_size = ac->asset_memory.chunk_size * elements;

    ac->vram_size += asset.vram_size;
    ac->ram_size += asset.ram_size;
    ++ac->asset_count;

    DEBUG_MEMORY_RESERVE((uintptr_t) asset_data, asset.ram_size);

    return (Asset *) hashmap_insert(&ams->hash_map, name, (byte *) &asset)->value;
}

// @todo Find a way to handle manual ram/vram changes
// Either implement a ams_update(AssetManagementSystem* ams, Asset* asset) function
// Or set .has_changed = true (even if garbage collection gets set) and call this func somewhere (maybe thread?)
// Perform general ams update (stats and garbage collection)
// We perform multiple things in one iteration to reduce the iteration costs
// @todo don't use uint64 for time, use uint32 and use relative time to start of program
void thrd_ams_update(AssetManagementSystem* ams, uint64 time, uint64 dt)
{
    PROFILE(PROFILE_AMS_UPDATE);
    for (int32 i = 0; i < ams->asset_component_count; ++i) {
        ams->asset_components[i].vram_size = 0;
        ams->asset_components[i].ram_size = 0;
        ams->asset_components[i].asset_count = 0;
    }

    // Iterate the hash map to find all assets
    uint32 chunk_id = 0;
    chunk_iterate_start(&ams->hash_map.buf, chunk_id) {
        HashEntry* entry = (HashEntry *) chunk_get_element(&ams->hash_map.buf, chunk_id);
        Asset* asset = (Asset *) entry->value;

        if (!thrd_ams_is_loaded(asset)) {
            continue;
        }

        ams->asset_components[asset->component_id].vram_size += asset->vram_size;
        ams->asset_components[asset->component_id].ram_size += asset->ram_size;
        ++ams->asset_components[asset->component_id].asset_count;

        if ((asset->state & ASSET_STATE_RAM_GC) || (asset->state & ASSET_STATE_VRAM_GC)) {
            // @todo Currently we cannot really delete based on last access since we are not updating the last_access reliably
            // This is usually the case for global/static assets such as font, ui_asset = ui vertices, ...
            // The reason for this is that we store a reference to those assets in a global struct
            // One solution could be to manually update the last_access at the end of every frame (shouldn't be THAT many assets)?
            // Maybe we can even implement a scene specific post_scene() function that does this for scene specific assets e.g. post_scene_scene1();
            if ((asset->state & ASSET_STATE_RAM_GC)
                && (asset->state & ASSET_STATE_VRAM_GC)
                && time - asset->last_access <= dt
            ) {
                // @performance Ideally we would like to pass the entry to delete
                // The problem is the hashmap_delete function can't work with entries directly since it is not a doubly linked list
                thrd_ams_remove_asset(ams, &ams->asset_components[asset->component_id], asset, entry->key);
            } else if ((asset->state & ASSET_STATE_RAM_GC)
                && time - asset->last_access <= dt
            ) {
                ams_remove_asset_ram(&ams->asset_components[asset->component_id], asset);
            } else if ((asset->state & ASSET_STATE_VRAM_GC)
                && time - asset->last_access <= dt
            ) {
                ams->asset_components[asset->component_id].vram_size -= asset->vram_size;
            }
        }
    } chunk_iterate_end;
}

Asset* ams_insert_asset(AssetManagementSystem* ams, Asset* asset_temp, const char* name)
{
    AssetComponent* ac = &ams->asset_components[asset_temp->component_id];

    int32 free_data = chunk_reserve(&ac->asset_memory, asset_temp->size);
    if (free_data < 0) {
        ASSERT_SIMPLE(free_data >= 0);
        return NULL;
    }

    byte* asset_data = chunk_get_element(&ac->asset_memory, free_data);

    asset_temp->self = asset_data;
    asset_temp->size = asset_temp->size; // Crucial for freeing
    asset_temp->ram_size = ac->asset_memory.chunk_size * asset_temp->size;

    ac->vram_size += asset_temp->vram_size;
    ac->ram_size += asset_temp->ram_size;
    ++ac->asset_count;

    Asset* asset = (Asset *) hashmap_insert(&ams->hash_map, name, (byte *) asset_temp)->value;
    DEBUG_MEMORY_RESERVE((uintptr_t) asset->self, asset->ram_size);

    return asset;
}

inline
Asset* thrd_ams_insert_asset(AssetManagementSystem* ams, Asset* asset_temp, const char* name)
{
    AssetComponent* ac = &ams->asset_components[asset_temp->component_id];

    coms_pthread_mutex_lock(&ams->asset_components[asset_temp->component_id].mutex);
    int32 free_data = chunk_reserve(&ac->asset_memory, asset_temp->size);
    if (free_data < 0) {
        coms_pthread_mutex_unlock(&ams->asset_components[asset_temp->component_id].mutex);
        ASSERT_SIMPLE(free_data >= 0);

        return NULL;
    }
    coms_pthread_mutex_unlock(&ams->asset_components[asset_temp->component_id].mutex);

    byte* asset_data = chunk_get_element(&ac->asset_memory, free_data);
    memcpy(asset_data, asset_temp->self, sizeof(Asset));

    asset_temp->self = asset_data;
    asset_temp->ram_size = ac->asset_memory.chunk_size * asset_temp->size;

    ac->vram_size += asset_temp->vram_size;
    ac->ram_size += asset_temp->ram_size;
    ++ac->asset_count;

    Asset* asset = (Asset *) hashmap_insert(&ams->hash_map, name, (byte *) asset_temp)->value;
    DEBUG_MEMORY_RESERVE((uintptr_t) asset->self, asset->ram_size);

    atomic_set_release(&asset->is_loaded, 1);

    return asset;
}

#endif