Performance Improvements

Comparison with VPI-0.1.0

VPI-0.2.0 brings several performance improvements in most of its algorithms across all backends with respect to previous VPI-0.1.0.

The charts below shows the performance increase per backend, per device. The benchmarking procedure used is described here.

Algorithm Parameters

Here is a list of the parameters used while benchmarking all above algorithms. This helps puts the values above into context.

• Gaussian Pyramid Generator
  • CPU/CUDA: 8-bit 1920x1080 input, 4 levels, 0.5x scale (each level is 4x smaller than previous)
  • PVA: 16-bit 3264x2448, 5 levels, 0.5x scale
• Gaussian Image Filter
  • CPU/CUDA: 8-bit 1920x1080 input, 3x3 kernel support size
  • PVA: 8-bit 3264x2448 input, 3x3 kernel support size.
• Image Convolver
  • CPU/CUDA: 8-bit 1920x1080 input, 5x5 kernel support size.
  • PVA: 8-bit 3264x2448 input, 5x5 kernel support size.
• Box Image Filter
  • CPU/CUDA: 8-bit 1920x1080 input, 3x3 kernel support, clamp boundary condition.
  • PVA: 8-bit 3264x2448 input, 3x3 kernel support, zero boundary condition.
• Image Resampler
  • CPU/CUDA: 8-bit 480x270 upscale to 1920x1080 with linear interpolation
• Stereo Disparity Estimator
  • CPU/CUDA/PVA 16-bit 480x270 input.
• Bilateral Image Filter
  • CPU/CUDA: 8-bit 3264x2448 input, 5x5 spatial kernel support size.
• Separable Image Convolver
  • CPU/CUDA: 16-bit 1920x1080 input, 5x5 kernel support size.
  • PVA: 16-bit 3264x2448 input, 5x5 kernel support size.
• Harris Keypoint Detector
  • CPU/CUDA: 16-bit 1920x1080 input, 3x3 gradient kernel size, 3x3 block size.