PVA Pre-processing#

Introduction#

The PVA preprocessing function demonstrates how to use PVA to perform preprocessing tasks in car detection to off-load the GPU resources.

It includes instructions on how to compile and deploy the PVA operators and how to enable the PVA function.

Dataflow

image1

The data flow is similar to that of car detection. The main difference is that pre-processing is performed in the PVA engine. If the input image is in block linear memory layout, the conversion from block linear to pitch linear is necessary and can occur in the GPU by running a specific CUDA kernel. After processing by the resize operator and the remap operator, the result is fed into the GPU or DLA for inference.

To use the PVA preprocessing feature, complete the following steps:

  1. Confirm car detection and the PVA function are enabled in nvsipl_multicast.

    Set NV_BUILD_CARDETECT = 1 in the Makefile and build the application.

  2. Copy the following files to the target machine:

    boundingBox.map, XY_FixPtMap.raw, XY_UV_FixPtMap.raw,
XY_FixPtMap_Resize.raw, XY_UV_FixPtMap_Resize.raw,
libcupva_host.so.2.6, and libpva_pre_lib.so

  3. Ensure that libcupva_host.so.2.6, libpva_pre_lib.so, and the nvsipl_multicast binary are in the same folder.

  4. Enable PVA Authority:

    echo 0 | sudo tee /sys/kernel/debug/pva0/vpu_app_authentication

  5. Run the nvsipl_multicast sample.

Constraints#

  • PVA authority is required to enable this feature.

  • To be consistent with map file requirements, PVA requires input resolution of 3840x2160. For other resolution requirements, contact your NVIDIA representative.

  • PVA preprocessing in this pipeline can only process block linear layout data.

On this page