Fast Gaussian blur in real time on GPU

Gaussian filtering is widely used standard algorithm which is a must in many applications, starting from Sharp/USM to SIFT/SURF. Gauss filter is isotropic and separable. These properties are very important for fast and efficient image processing. Gaussian filtering usually is time-consuming task, that's why it's a good idea try to accelerate it on GPU.

Standard features for Gaussian filtering on NVIDIA GPUs

• Data input: 8/16 or 24/48-bit images in CPU or GPU memory
• Data output: final image in CPU or GPU memory
• Parameters: sigma (radius blur)
• Optimized for the latest NVIDIA GPUs
• Compatible with Windows-10/11 and Linux Ubuntu/L4T

Benchmarks for Gaussian blur on GeForce GTX 1080 (Windows-10 and CUDA-10, 64-bit)

Now we need just ~8 ms for Gaussian blur (sigma ~1, window 5×5) of 24-bit color image with 3840×2160 resolution. These are benchmarks for 2K / 4K images, 24-bit (computations on GPU, without DeviceIO latency)

• Full HD (2K, 1920×1080) ~ 2.4 GByte/s
• 4K (3840×2160) ~ 3 GByte/s