VPI - Vision Programming Interface

1.0 Release



The Convolution algorithm performs a 2D convolution operation on the input image with the provided 2D kernel. This is useful when the kernel isn't separable and its dimensions are smaller than 5x5. In other cases, it's usually preferable to use the Separable Convolution algorithm due to its speed.

Input Kernel Output

\[ \begin{bmatrix} 1 & 0 & -1 \\ 0 & 0 & 0 \\ -1 & 0 & 1 \end{bmatrix} \]


Discrete 2D convolution is implemented using the following discrete function:

\[ I'[x,y] = \sum_{m=0}^{k_h} \sum_{n=0}^{k_w} K[m,n] \times I[x-(n-\lfloor k_w/2 \rfloor), y-(m-\lfloor k_h/2 \rfloor) ] \]


  • \(I\) is the input image.
  • \(I'\) is the result image.
  • \(K\) is the convolution kernel.
  • \(k_w,k_h\) are the kernel's width and height, respectively.
Most computer vision libraries expect the kernel to be reversed before calling their convolution functions. Not so with VPI, we implement an actual convolution, not cross-correlation. Naturally, this is irrelevant if the kernel is symmetric.


  1. Initialization phase
    1. Include the header that defines the needed functions and structures.
      Declares functions to perform image filtering with convolution kernels.
    2. Define the input image object.
      VPIImage input = /*...*/;
      struct VPIImageImpl * VPIImage
      A handle to an image.
      Definition: Types.h:197
    3. Create the output image. It gets its dimensions and format from the input image.
      int32_t w, h;
      vpiImageGetSize(input, &w, &h);
      vpiImageGetFormat(input, &type);
      VPIImage output;
      vpiImageCreate(w, h, type, 0, &output);
      Pre-defined image formats.
      Definition: ImageFormat.h:99
      VPIStatus vpiImageCreate(int32_t width, int32_t height, VPIImageFormat fmt, uint32_t flags, VPIImage *img)
      Create an empty image instance with the specified flags.
      VPIStatus vpiImageGetFormat(VPIImage img, VPIImageFormat *format)
      Get the image format.
      VPIStatus vpiImageGetSize(VPIImage img, int32_t *width, int32_t *height)
      Get the image size in pixels.
    4. Create the stream where the algorithm will be submitted for execution.
      VPIStream stream;
      vpiStreamCreate(0, &stream);
      struct VPIStreamImpl * VPIStream
      A handle to a stream.
      Definition: Types.h:191
      VPIStatus vpiStreamCreate(uint32_t flags, VPIStream *stream)
      Create a stream instance.
  2. Processing phase
    1. Define the kernel to be used. In this case, a simple 3x3 edge detector.
      float kernel[3 * 3] = { 1, 0,-1,
      0, 0, 0,
      -1, 0, 1};
    2. Submit the algorithm to the stream, passing the kernel and other parameters. It'll be executed by the CPU backend.
      vpiSubmitConvolution(stream, VPI_BACKEND_CPU, input, output, kernel, 3, 3, VPI_BORDER_ZERO);
      VPIStatus vpiSubmitConvolution(VPIStream stream, uint32_t backend, VPIImage input, VPIImage output, const float *kernelData, int32_t kernelWidth, int32_t kernelHeight, VPIBorderExtension border)
      Runs a generic 2D convolution over an image.
      CPU backend.
      Definition: Types.h:91
      All pixels outside the image are considered to be zero.
      Definition: Types.h:219
    3. Optionally, wait until the processing is done.
      VPIStatus vpiStreamSync(VPIStream stream)
      Blocks the calling thread until all submitted commands in this stream queue are done (queue is empty)...
  3. Cleanup phase
    1. Free resources held by the stream and the input and output images.
      void vpiImageDestroy(VPIImage img)
      Destroy an image instance.
      void vpiStreamDestroy(VPIStream stream)
      Destroy a stream instance and deallocate all HW resources.

Consult the Image Convolution for a complete example.

For more information, see Convolution in the "API Reference" section of VPI - Vision Programming Interface.

Limitations and Constraints

Constraints for specific backends supersede the ones specified for all backends.

All Backends


  • Only available on Jetson Xavier devices.
  • Input and output dimensions must be between 65x33 and 3264x2448.
  • Minimum convolution kernel size is 2x2.
  • Maximum convolution kernel size is 11x11.
  • Kernel weights are restricted to \(|weight| < 1\)
  • Only VPI_BORDER_ZERO is accepted.


  • Not implemented.


For information on how to use the performance table below, see Algorithm Performance Tables.
Before comparing measurements, consult Comparing Algorithm Elapsed Times.
For further information on how performance was benchmarked, see Performance Benchmark.