I'm having trouble with the shape of a struct to pass to an NVML Library function via pinvoke. The struct contains a fixed size array and some unsigned long long's I'm not encountering any compiler errors but the data returned seems to be corrupted. (An int isRunning should only be a zero or a 1, I'm getting a 30, a 3...) Here is it's declaration: Accounting struct
Public Variables
unsigned int gpuUtilization
unsigned int isRunning //Flag to represent if the process is running (1 for running, 0 for terminated).
unsigned long long maxMemoryUsage
unsigned int memoryUtilization
unsigned int reserved[5] //Reserved for future use.
unsigned long long startTime //CPU Timestamp in usec representing start time for the process.
unsigned long long time
My code(minimal repeatable version, console c# app):
// c# version of accounting struct:
public struct nvmlAccountingStats_t
{
public uint gpuUtilization;
public uint isRunning;
public ulong maxMemoryUsage;
public uint memoryUtilization;
[MarshalAs(UnmanagedType.ByValArray, SizeConst = 5)]
public uint[] reserved;
public ulong startTime;
public ulong time;
}
public enum nvmlReturn_t
{
NVML_SUCCESS = 0, // The operation was successful.
NVML_ERROR_UNINITIALIZED = 1, // NVML was not first initialized with nvmlInit().
NVML_ERROR_INVALID_ARGUMENT = 2, // A supplied argument is invalid.
NVML_ERROR_NOT_SUPPORTED = 3, // The requested operation is not available on target device.
NVML_ERROR_NO_PERMISSION = 4, // The current user does not have permission for operation.
NVML_ERROR_ALREADY_INITIALIZED = 5, // Deprecated: Multiple initializations are now allowed through ref counting.
NVML_ERROR_NOT_FOUND = 6, // A query to find an object was unsuccessful.
NVML_ERROR_INSUFFICIENT_SIZE = 7, // An input argument is not large enough.
NVML_ERROR_INSUFFICIENT_POWER = 8, // A device's external power cables are not properly attached.
NVML_ERROR_DRIVER_NOT_LOADED = 9, // NVIDIA driver is not loaded.
NVML_ERROR_TIMEOUT = 10, // User provided timeout passed.
NVML_ERROR_IRQ_ISSUE = 11, // NVIDIA Kernel detected an interrupt issue with a GPU.
NVML_ERROR_LIBRARY_NOT_FOUND = 12, // NVML Shared Library couldn't be found or loaded.
NVML_ERROR_FUNCTION_NOT_FOUND = 13, // Local version of NVML doesn't implement this function.
NVML_ERROR_CORRUPTED_INFOROM = 14, // infoROM is corrupted
NVML_ERROR_GPU_IS_LOST = 15, // The GPU has fallen off the bus or has otherwise become inaccessible.
NVML_ERROR_RESET_REQUIRED = 16, // The GPU requires a reset before it can be used again.
NVML_ERROR_OPERATING_SYSTEM = 17, // The GPU control device has been blocked by the operating system/cgroups.
NVML_ERROR_LIB_RM_VERSION_MISMATCH = 18, // RM detects a driver/library version mismatch.
NVML_ERROR_IN_USE = 19, // An operation cannot be performed because the GPU is currently in use.
NVML_ERROR_MEMORY = 20, // Insufficient memory.
NVML_ERROR_NO_DATA = 21, // No data.
NVML_ERROR_VGPU_ECC_NOT_SUPPORTED = 22, // The requested vgpu operation is not available on target device, becasue ECC is enabled.
NVML_ERROR_INSUFFICIENT_RESOURCES = 23, // Ran out of critical resources, other than memory.
NVML_ERROR_FREQ_NOT_SUPPORTED = 24, // Ran out of critical resources, other than memory.
NVML_ERROR_UNKNOWN = 999 // An internal driver error occurred.
}
[DllImport(NVMLDLL)]
private static extern nvmlReturn_t nvmlInit_v2();
[DllImport(NVMLDLL)]
private static extern nvmlReturn_t nvmlShutdown();
[DllImport(NVMLDLL)]
private static extern nvmlReturn_t nvmlDeviceGetCount_v2(out uint deviceCount);
[DllImport(NVMLDLL)]
private static extern nvmlReturn_t nvmlDeviceGetHandleByIndex_v2(int index, out IntPtr handle);
[DllImport(NVMLDLL)]
private static extern nvmlReturn_t nvmlDeviceGetAccountingPids(IntPtr device, ref uint count, uint[]? pids);
[DllImport(NVMLDLL)]
private static extern nvmlReturn_t nvmlDeviceGetAccountingStats(IntPtr device, uint pid, out nvmlAccountingStats_t stats);
static void Main(string[] args)
{
if (nvmlInit_v2() == nvmlReturn_t.NVML_SUCCESS)
{
nvmlDeviceGetCount_v2(out uint deviceCount);
for (int i = 0; i < deviceCount; i )
{
nvmlDeviceGetHandleByIndex_v2(i, out IntPtr _device);
// get the pids
uint count = 0;
nvmlDeviceGetAccountingPids(_device, ref count, null);
uint[] pids = new uint[count];
if (nvmlDeviceGetAccountingPids(_device, ref count, pids) == nvmlReturn_t.NVML_SUCCESS)
{
foreach (uint pid in pids)
{
nvmlDeviceGetAccountingStats(_device, pid, out nvmlAccountingStats_t stats);
Console.WriteLine($"pid: {pid}, isRunning: {stats.isRunning}");
}
}
}
nvmlShutdown();
}
}
}
at first I saw I had some data back and was excited but I checked the isRunning results and they were thus:
pid: 6560, isRunning: 2
pid: 31636, isRunning: 0
pid: 30000, isRunning: 0
pid: 4164, isRunning: 0
pid: 3368, isRunning: 0
pid: 30420, isRunning: 0
pid: 732, isRunning: 0
pid: 14576, isRunning: 0
pid: 27352, isRunning: 0
pid: 25632, isRunning: 0
pid: 7352, isRunning: 0
pid: 32960, isRunning: 0
pid: 25380, isRunning: 30
pid: 21376, isRunning: 0
pid: 21396, isRunning: 0
pid: 7488, isRunning: 0
pid: 39200, isRunning: 0
pid: 15536, isRunning: 3
and for a value that's supposed to be only a 1 or 0, seeing a 30 and a 3 made me scratch my head ^_^ The gpuutilization values were also off but the isRunning is more noticeable lol
So if anyone knows the correct form that the struct should take, or if I'm not passing it correctly to the function I'd sure appreciate any help ^_^ thank you very much
Updated with answer thanks to @Swift - Friday Pie: I learned a lesson that I should read the implementation file itself if possible rather than referring to the documentation and assuming it's showing the correct order =)
[StructLayout(LayoutKind.Sequential)]
public struct nvmlAccountingStats_t
{
public uint gpuUtilization;
public uint memoryUtilization;
public ulong maxMemoryUsage;
public ulong time;
public ulong startTime;
public uint isRunning;
[MarshalAs(UnmanagedType.ByValArray, SizeConst = 5)]
public uint[] reserved;
}
the updated c# struct shape is returning the information expected that I can see so far(isRunning is only 1 or 0!), thank you for the help!
CodePudding user response：
Documentation doesn't describe actual layout of structure (blame NVIDIA), it's an alphabetically sorted list. The real implementation in nvml.h looks like:
typedef struct nvmlAccountingStats_st {
unsigned int gpuUtilization; //!< Percent of time over the process's lifetime during which one or more kernels was executing on the GPU.
//! Utilization stats just like returned by \ref nvmlDeviceGetUtilizationRates but for the life time of a
//! process (not just the last sample period).
//! Set to NVML_VALUE_NOT_AVAILABLE if nvmlDeviceGetUtilizationRates is not supported
unsigned int memoryUtilization; //!< Percent of time over the process's lifetime during which global (device) memory was being read or written.
//! Set to NVML_VALUE_NOT_AVAILABLE if nvmlDeviceGetUtilizationRates is not supported
unsigned long long maxMemoryUsage; //!< Maximum total memory in bytes that was ever allocated by the process.
//! Set to NVML_VALUE_NOT_AVAILABLE if nvmlProcessInfo_t->usedGpuMemory is not supported
unsigned long long time; //!< Amount of time in ms during which the compute context was active. The time is reported as 0 if
//!< the process is not terminated
unsigned long long startTime; //!< CPU Timestamp in usec representing start time for the process
unsigned int isRunning; //!< Flag to represent if the process is running (1 for running, 0 for terminated)
unsigned int reserved[5]; //!< Reserved for future use
} nvmlAccountingStats_t;
That's without accounting that C# might have different memory layout with same declaration, it's implementation-defined.