Struct VirtualMemoryManager 

pub struct VirtualMemoryManager;

Manager for GPU virtual memory operations.

Provides methods for reserving virtual address ranges, allocating physical memory, mapping/unmapping, and setting access permissions.

Status

The underlying CUDA virtual memory driver functions (cuMemAddressReserve, cuMemCreate, cuMemMap, cuMemUnmap, cuMemSetAccess, cuMemRelease, cuMemAddressFree) are wired through oxicuda-driver. On systems without a CUDA driver the calls fail with CudaError::NotInitialized; on systems with a driver that lacks a specific VMM symbol the calls fail with CudaError::NotSupported.

Implementations

impl VirtualMemoryManager

pub fn reserve(size: usize, alignment: usize) -> CudaResult<VirtualAddressRange>

Reserves a range of virtual addresses in the GPU address space.

The reserved range is not backed by physical memory until map is called.

Parameters

• size - Size of the virtual range to reserve in bytes. Must be a multiple of alignment.
• alignment - Alignment requirement in bytes. Must be a power of two and non-zero.

Errors

• CudaError::InvalidValue if size is zero, alignment is zero, alignment is not a power of two, or size is not a multiple of alignment.

pub fn release(va: VirtualAddressRange) -> CudaResult<()>

Releases a previously reserved virtual address range.

After this call, the virtual addresses are no longer reserved and may be reused by future reservations.

Errors

• CudaError::NotInitialized if no CUDA driver is available (e.g. on macOS).
• CudaError::NotSupported if the driver does not export cuMemAddressFree.
• Other CudaError variants on driver failure.

pub fn alloc_physical( size: usize, device_ordinal: i32, ) -> CudaResult<PhysicalAllocation>

Allocates physical memory on the specified device.

The allocated memory is not accessible until mapped into a virtual address range via map.

Parameters

• size - Size of the allocation in bytes. Must be non-zero.
• device_ordinal - Ordinal of the device to allocate on.

Errors

• CudaError::InvalidValue if size is zero or device_ordinal is negative.
• CudaError::NotInitialized if no CUDA driver is available (e.g. on macOS).
• CudaError::NotSupported if the driver does not export cuMemCreate.
• Other CudaError variants on driver failure.

pub fn free_physical(phys: PhysicalAllocation) -> CudaResult<()>

Frees a physical memory allocation.

The allocation must not be currently mapped to any virtual range.

Errors

• CudaError::NotInitialized if no CUDA driver is available (e.g. on macOS).
• CudaError::NotSupported if the driver does not export cuMemRelease.
• Other CudaError variants on driver failure.

pub fn map( va: &VirtualAddressRange, phys: &PhysicalAllocation, offset: usize, ) -> CudaResult<()>

Maps a physical allocation to a region of a virtual address range.

After mapping, GPU kernels can access the virtual addresses and reads/writes will be routed to the physical memory.

Parameters

• va - The virtual address range to map into.
• phys - The physical allocation to map.
• offset - Byte offset within the virtual range at which to start the mapping. Must be aligned to the VA’s alignment.

Errors

• CudaError::InvalidValue if offset is not aligned, or if the physical allocation would extend past the end of the virtual range.
• CudaError::NotInitialized if no CUDA driver is available (e.g. on macOS).
• CudaError::NotSupported if the driver does not export cuMemMap.
• Other CudaError variants on driver failure.

pub fn unmap( va: &VirtualAddressRange, offset: usize, size: usize, ) -> CudaResult<()>

Unmaps a region of a virtual address range.

After unmapping, accesses to the affected virtual addresses will fault. The physical memory is not freed — it can be remapped elsewhere.

Parameters

• va - The virtual address range to unmap from.
• offset - Byte offset within the range where unmapping starts.
• size - Number of bytes to unmap.

Errors

• CudaError::InvalidValue if the offset+size exceeds the virtual range bounds.
• CudaError::NotInitialized if no CUDA driver is available (e.g. on macOS).
• CudaError::NotSupported if the driver does not export cuMemUnmap.
• Other CudaError variants on driver failure.

pub fn set_access( va: &VirtualAddressRange, device_ordinal: i32, flags: AccessFlags, ) -> CudaResult<()>

Sets access permissions for a virtual address range on a device.

This controls whether the specified device can read and/or write to the mapped virtual addresses.

Parameters

• va - The virtual address range to set permissions on.
• device_ordinal - The device to grant/deny access for.
• flags - The access permission flags.

Errors

• CudaError::NotInitialized if no CUDA driver is available (e.g. on macOS).
• CudaError::NotSupported if the driver does not export cuMemSetAccess.
• Other CudaError variants on driver failure.