Performance & Benchmark

From CUVI Wiki

Measured with NVIDIA's Performance tools for Windows and Linux. Timing figure represents time of kernel/function in milliseconds (rounded) on a single GPU. The benchmarks are performed on color images with 8-bits per channel except where mentioned otherwise. The list below is a small subset of 100+ features in CUVI.

Kernel Time in milliseconds (ms) with CUVI v1.7.0 on Jetson Nano having 128 CUDA Cores.
Algorithm / Image Size 720p 1080p 4k (3840x2160) 8k (7680x4320)
add - 2 Images 2.99 8.38 15.63 50.27
channelMix 4.09 7.42 15.70 53.35
demosaic 8.11 11.77 42.99 172.40
demosaicDFPD 12.6 23.86 88.87 357.94
gammaCorrect 3.12 5.69 14.13 45.80
histEq - Single Channel 5.29 7.88 20.53 61.52
LUT 1.89 2.77 11.26 25.42
blackGammaLUT 4.08 7.52 18.36 62.27
rgb2gray 1.95 2.67 10.64 20.89
focusStack - Stacking 5 Images 252.52 452.81 1830.54 7320.52
bitConversion - From 8 to 16 bits 3.74 8.46 15.71 63.33
crop 1.67 4.76 9.04 28.93
resize - Scale=2.0 9.29 18.43 55.58 222.41
resize - Scale=0.5 2.33 5.23 7.01 17.08
rotate - Non Cropping, Angle = -3.76f 3.43 7.73 26.00 58.34
warpPerspective 5.09 11.48 19.86 81.36
imageFilter - 5x5 floating point window 18.44 26.80 89.58 358.29
underwaterFilter 29.55 50.74 79.09 332.95
haarFwd 10.27 18.60 40.45 130.86
Kernel Time in milliseconds (ms) with CUVI v1.8.0 on GTX 1080 having 2,560 CUDA Cores.
Algorithm / Image Size 720p 1080p 4k (3840x2160) 8k (7680x4320)
add - 2 Images 0.05 0.10 0.42 1.69
channelMix 0.04 0.08 0.34 1.33
demosaic 0.12 0.26 1.01 4.04
demosaicDFPD 0.31 0.69 2.77 10.98
gammaCorrect 0.04 0.10 0.40 1.61
histEq - Single Channel 0.08 0.18 0.61 2.18
LUT 0.05 0.10 0.35 1.25
blackGammaLUT 0.99 0.21 0.74 2.73
rgb2gray 0.02 0.05 0.21 0.83
focusStack - Stacking 5 Images 8.66 14.44 65.14 270.59
bitConversion - From 8 to 16 bits 0.06 0.14 0.58 2.30
crop 0.03 0.07 0.23 0.93
resize - Scale=2.0 0.19 0.41 1.70 6.83
resize - Scale=0.5 0.02 0.04 0.14 0.58
rotate - Non Cropping, Angle = -3.76f 0.08 0.16 0.66 2.69
warpPerspective 0.08 0.22 0.79 3.21
imageFilter - 5x5 floating point window 0.30 0.66 2.63 9.18
underwaterFilter 0.45 0.96 3.39 11.62
haarFwd 0.14 0.34 1.35 5.10
Kernel Time in milliseconds (ms) with CUVI v1.8.0 on Jetson Xavier NX having 384 CUDA Cores.
Algorithm / Image Size 720p 1080p 4k (3840x2160) 8k (7680x4320)
add - 2 Images 0.29 0.61 2.04 8.61
channelMix 0.27 0.61 2.31 9.02
demosaic 1.87 2.3 9.17 36.74
demosaicDFPD 2.33 4.96 19.07 77.75
gammaCorrect 0.22 0.48 1.89 7.47
histEq - Single Channel 0.68 0.92 3.24 9.20
LUT 0.10 0.30 0.86 3.28
blackGammaLUT 0.36 0.68 1.86 7.29
rgb2gray 0.14 0.25 0.96 3.83
focusStack - Stacking 5 Images 142.56 285.95 1103.14 4399.84
bitConversion - From 8 to 16 bits 0.38 0.77 3.12 12.34
crop 0.13 0.48 2.05 6.05
resize - Scale=2.0 0.85 1.90 7.57 30.32
resize - Scale=0.5 0.08 0.33 0.82 2.89
rotate - Non Cropping, Angle = -3.76f 0.23 0.49 1.90 7.64
warpPerspective 0.24 0.68 2.26 9.38
imageFilter - 5x5 floating point window 2.97 7.89 23.76 108.21
underwaterFilter 1.57 3.49 13.6 47.39
haarFwd 1.07 2.39 6.47 25.70
Kernel Time in milliseconds (ms) with CUVI v1.8.0 on GTX 1650 having 896 CUDA Cores.
Algorithm / Image Size 720p 1080p 4k (3840x2160) 8k (7680x4320)
add - 2 Images 0.08 0.18 0.72 2.92
channelMix 0.09 0.21 0.85 3.41
demosaic 0.35 0.78 3.53 13.1
demosaicDFPD 0.75 1.69 6.74 27.1
gammaCorrect 0.18 0.41 1.60 6.34
histEq - Single Channel 0.15 0.32 1.21 9.44
LUT 0.05 0.11 0.42 1.74
blackGammaLUT 0.09 0.22 0.90 3.66
rgb2gray 0.06 0.12 0.49 2.01
focusStack - Stacking 5 Images 46.10 97.24 257.62 1180.50
bitConversion - From 8 to 16 bits 0.15 0.35 1.40 5.63
crop 0.06 0.18 0.61 2.49
resize - Scale=2.0 0.36 0.80 3.22 12.88
resize - Scale=0.5 0.03 0.06 0.23 0.93
rotate - Non Cropping, Angle = -3.76f 0.14 0.33 1.30 5.16
warpPerspective 0.12 0.29 1.14 4.68
imageFilter - 5x5 floating point window 0.97 2.17 8.66 34.64
underwaterFilter 0.66 1.22 4.59 18.61
haarFwd 0.19 0.43 1.77 6.84
Kernel Time in milliseconds (ms) with CUVI v1.8.0 on RTX 2060 Mobile having 1,920 CUDA Cores.
Algorithm / Image Size 720p 1080p 4k (3840x2160) 8k (7680x4320)
add - 2 Images 0.06 0.14 0.51 2.01
channelMix 0.07 0.14 0.55 2.25
demosaic 0.24 0.53 2.10 8.10
demosaicDFPD 0.52 1.22 4.53 18.1
gammaCorrect 0.12 0.28 1.02 4.30
histEq - Single Channel 0.21 0.24 0.84 3.10
LUT 0.03 0.08 0.29 1.20
blackGammaLUT 0.069 0.16 0.61 2.50
rgb2gray 0.04 0.09 0.34 1.43
focusStack - Stacking 5 Images 25.77 55.86 221.60 605.53
bitConversion - From 8 to 16 bits 0.01 0.24 0.95 3.81
crop 0.04 0.12 0.41 1.70
resize - Scale=2.0 0.25 0.55 2.21 8.70
resize - Scale=0.5 0.02 0.05 0.16 0.64
rotate - Non Cropping, Angle = -3.76f 0.04 0.09 0.36 1.11
warpPerspective 0.08 0.20 0.77 3.10
imageFilter - 5x5 floating point window 0.65 1.56 5.81 13.7
underwaterFilter 0.53 1.10 4.00 15.2
haarFwd 0.14 0.30 1.21 4.90

Color Pipeline

Let's take a typical color pipeline and measure its performance on one of the entry level GPUs. Any color pipeline almost always starts with the Raw image. Before converting to RGB, you might want to do some processing on the raw which may include applying LUTs (look up tables), FPN (fixed point noise) removal and fixing white balance. Next comes demosaic/debayer followed by several further enhancement functions and a color space conversion into the desired format. This pipeline can perform in real-time on a decent entry level GPU on an 8k images and at over 100 FPS on a 2k image:

Color pipeline where each box represents one or more functions.

Performance

  • Image size: 8k
  • Debayer method: DFPD
  • RAW Size: 59.9 MB
  • Codec: JPEG2000
  • Sharpening: 7x7
  • GPU: GTX 1080
  • FPS: 26 FPS