| 
    NVIDIA NvNeural SDK
    2022.1
    
   GPU inference framework for NVIDIA Nsight Deep Learning Designer 
   | 
 
Structure describing details of an object's memory allocation behavior. More...
#include <nvneural/CoreTypes.h>
Public Attributes | |
| std::size_t | peakCommittedMemory | 
| This is the high water mark of on device memory (committed memory).  | |
| std::size_t | peakRequestedMemory | 
| This is the high water mark of requested memory.  More... | |
| std::size_t | persistingCommittedMemory | 
| This is the high water mark of on device memory (committed memory) that persists after a network is launched.  More... | |
| std::size_t | persistingRequestedMemory | 
| This is the high water mark of requested memory that persists after a network finishes an inference run.  More... | |
| std::size_t | structSize | 
| This is the sizeof() this struct.  More... | |
Structure describing details of an object's memory allocation behavior.
Use INetworkBackend2::getMemoryTrackingData to obtain instances of this structure.
| std::size_t nvneural::MemoryTrackingData::peakRequestedMemory | 
This is the high water mark of requested memory.
Requested memory can differ from committed memory in that on device memory may be in preallocated buffers.
| std::size_t nvneural::MemoryTrackingData::persistingCommittedMemory | 
This is the high water mark of on device memory (committed memory) that persists after a network is launched.
Usually weight data or permanent layers.
| std::size_t nvneural::MemoryTrackingData::persistingRequestedMemory | 
This is the high water mark of requested memory that persists after a network finishes an inference run.
Usually weight data or permanent layers. Requested memory can differ from committed memory in that on device memory may be in preallocated buffers.
| std::size_t nvneural::MemoryTrackingData::structSize | 
This is the sizeof() this struct.
It is used for versioning. Future versions of MemoryTrackingData will be layout-compatible with this structure. Use structSize to check for v2, v3, etc., and then cast to the newer structure if appropriate.