System and Platform Information

Table of Contents

• GPU Enumeration and Selection
• DLA Engine Selection (Xavier SoC)

GPU Enumeration and Selection

On NVIDIA DRIVE^™ systems, multiple GPUs are present and available for the SDK.

• Integrated GPU (iGPU) is a GPU that shares the same die as the ARM CPU.
• Descrete GPU (dGPU) is a separate unit connected over a PCIEX bus with the main CPU.

At runtime, user applications can enumerate and select these GPUs. Most of the NVIDIA^® DriveWorks modules use the currently selected GPU.

#include <dw/core/context/Context.h>
 
...
 
// enumerate all available GPUs in the system and select first available dGPU
int32_t numGPUs = 0;
dwContext_getGPUCount(&numGPUs, sdkContext);
for (int32_t i=0; i < numGPUsl; i++)
{
    dwGPUDeviceType type;
    dwContext_getGPUDeviceType(&type, i, sdkContext);
    switch (type)
    {
        case DW_GPU_DEVICE_DISCRETE:
            printf("GPU %d is a dGPU, set it as current GPU\n", i);
 
            dwContext_selectGPUDevice(i, sdkContext);
 
            break;
 
        case DW_GPU_DEVICE_INTEGRATED:
            printf("GPU %d is an iGPU\n", i);
            break;
 
        default:
    }
}

Note

For applications that require access to GPU memory, ensure that such applications use the correct GPU.

DLA Engine Selection (Xavier SoC)

NVIDIA DRIVE^™ platforms with the NVIDIA Xavier SoC provide a hardware-accelerated deep learning accelerator (DLA). DLA accelerates inferencing of the networks, which frees up NVIDIA^® CUDA^® units to perform other tasks.

The following snippet shows how to activate a specific DLA engine for inferencing of DriveNet. For a full sample, see DriveNet Sample.

// check if current platform supports DLA engine
int32_t numDLAs = 0;
dwContext_getDLAEngineCount(&numDLAs, sdkContext);
if (numDLAs == 0)
{
    printf("The platform does not support DLA engine\n");
    return 0;
}
 
// setup DriveNet to use DLA engine for inferencing
dwDriveNetParams driveNetParams = {};
...
driveNetParams.processorType = DW_PROCESSOR_TYPE_DLA_0;
 
// DLA supports only FP16 precision
driveNetParams.networkPrecision = DW_PRECISION_FP16;
 
dwDriveNet_initialize(&driveNet, ..., &driveNetParams, sdkContext);