Resource Pools

class ResourcePools

Host-side allocation and reference accounting for all context-cache storage types.

Each ResourceType has an independent fixed-capacity pool and circular free list. ResourcePools manages only IDs and reference counts; the corresponding GPU buffers are allocated and bound by the runtime. An active reference pins a resource for an in-flight request, while a cache reference retains it for a published CacheRecord. A resource returns to its free list only when both counts reach zero.

Multi-pool allocation is all-or-nothing. Valid releases do not allocate and cannot fail after their preconditions have been checked.

Public Functions

explicit ResourcePools(ResourceDemand capacities)

std::optional<std::vector<ResourceId>> allocate(

ResourceDemand const &demand

)

Atomically allocate the full typed demand with active=1 and cache=0, or return nullopt without mutation.

bool allocateInto(

ResourceDemand const &demand,

std::vector<ResourceId> &resources

)

Fill caller-owned storage and commit without dynamic memory allocation. The output vector must already match the total demand; insufficient capacity returns false without mutation.

void addActiveRef(ResourceId const &resource)

Pin an allocated resource for an active request.

void releaseActiveRef(ResourceId const &resource)

Drop an active pin, returning the resource to the free list only when both reference counts reach zero.

void addCacheRef(ResourceId const &resource)

Add published-record ownership to an allocated resource.

void releaseCacheRef(ResourceId const &resource)

Drop published ownership, returning the resource to the free list only when both reference counts reach zero.

int32_t activeRefCount(ResourceId const &resource) const

int32_t cacheRefCount(ResourceId const &resource) const

int32_t freeCount(ResourceType type) const noexcept

int32_t capacity(ResourceType type) const noexcept

bool canAllocate(ResourceDemand const &demand) const noexcept