Namespace DotCompute.Backends.CUDA.Memory
Classes
- CacheLine
Cache-line padded wrappers for atomic counters in lock-free queues.
- CudaAsyncMemoryManagerAdapter
Adapter that wraps CudaMemoryManager for async operations. Bridges the CUDA memory manager with the unified memory interface.
- CudaContextMemoryManager
CUDA context-specific memory manager wrapping CudaMemoryManager
- CudaMemoryBuffer
Represents a CUDA memory buffer allocated on the GPU device.
- CudaMemoryBuffer<T>
Represents a generic CUDA memory buffer allocated on the GPU device.
- CudaMemoryManager
High-level analytics and cleanup helpers for CudaMemoryManager. These methods were previously exposed by a parallel facade (Integration/Components/CudaMemoryManager) that has since been removed as part of the v1.0 deduplication pass. They are implemented here so the single canonical
CudaMemoryManagerremains the only public CUDA memory-manager type.
- CudaMemoryOrderingProvider
CUDA-specific implementation of memory ordering primitives.
- CudaMemoryPoolManager
Manages memory pools for efficient allocation and reuse of CUDA memory. Reduces allocation overhead and memory fragmentation.
- CudaMemoryPrefetcher
Manages memory prefetching for unified memory to optimize data movement. Uses cudaMemPrefetchAsync to proactively move data between host and device.
- CudaPinnedMemoryAllocator
Manages pinned (page-locked) host memory for high-bandwidth transfers. Pinned memory provides up to 10x bandwidth improvement (20GB/s vs 2GB/s).
- CudaRawMemoryBuffer
Raw untyped CUDA memory buffer for byte-level operations.
- MemoryPoolStatistics
Statistics for memory pool usage.
- OptimizedCudaMemoryPrefetcher
Advanced CUDA memory prefetcher with intelligent pattern recognition:
- Predictive prefetching based on access patterns
- Multi-level prefetch strategies (L1, L2, global memory)
- Adaptive prefetch distance based on bandwidth utilization
- NUMA-aware prefetching for multi-GPU systems
- Asynchronous prefetch operations with minimal overhead
- Cache pollution avoidance with smart eviction policies Target: 30-50% improvement in memory-bound kernel performance
- PinnedMemoryStatistics
Statistics for pinned memory usage.
- PoolSizeStatistics
Statistics for a specific pool size.
- PrefetchRequest
Request for batch prefetch operation.
- PrefetchStatistics
Statistics for prefetch operations.
- PrefetcherConfiguration
Configuration for the memory prefetcher.
- SimpleCudaUnifiedMemoryBuffer<T>
Simple CUDA unified memory buffer implementation for the memory adapter. This is a lightweight version that doesn't depend on CudaUnifiedMemoryManagerProduction.
Structs
- PaddedInt
A 32-bit counter padded to its own cache line.
- PaddedLong
A 64-bit counter padded to its own cache line to prevent false sharing. Pass
ref counter.ValuetoInterlocked.*APIs.
- PrefetcherStatistics
Performance statistics for the prefetcher.
Interfaces
- IPinnedMemoryBuffer<T>
Interface for pinned memory buffers.
- IPinnedMemoryRegistration
Interface for pinned memory registration.
- IPooledMemoryBuffer
Interface for pooled memory buffers.
Enums
- CacheLevel
An cache level enumeration.
- CudaHostAllocFlags
Flags for pinned memory allocation.
- CudaHostRegisterFlags
Flags for host memory registration.
- MemoryAccessHint
An memory access hint enumeration.
- MemoryAccessType
An memory access type enumeration.
- PrefetchPriority
An prefetch priority enumeration.
- PrefetchStrategy
An prefetch strategy enumeration.
- PrefetchTarget
Target location for prefetch operation.