Difference between revisions of "CUDA ISP for NVIDIA Jetson/Performance/Library"
Line 28: | Line 28: | ||
For each system, we also used <code>jetson_clocks</code> to maximize the device clock frequency and thus the performance. | For each system, we also used <code>jetson_clocks</code> to maximize the device clock frequency and thus the performance. | ||
− | The following table summarizes CUDA ISP's performance results. | + | The following table summarizes CUDA ISP's performance results. |
+ | |||
<center> | <center> | ||
{| class="wikitable" style="text-align:center;" | {| class="wikitable" style="text-align:center;" |
Revision as of 12:24, 26 April 2023
CUDA ISP for NVIDIA Jetson | |
---|---|
![]() | |
CUDA ISP for NVIDIA Jetson Basics | |
|
|
Getting Started | |
|
|
User Manual | |
|
|
GStreamer | |
|
|
Examples | |
|
|
Performance | |
|
|
Contact Us | |
|
This page is still under development. Please, do not take the performance data as definitive.
Library API performance
To measure the CUDA ISP API performance, we built a simple example (provided upon request) that iterates over the Apply
methods for each algorithm and records performance metrics for each iteration. We measured the duration of each algorithm's Apply
method. We also measured CPU, CPU RAM, GPU, and GPU RAM usage for the complete processing pipeline iterating at 30fps. We ran the experiments on both 1080p and 4K buffers. We also ran the experiments on the Jetson Nano, Jetson Xavier NX, Jetson Xavier AGX, and Jetson AGX Orin.
- We measured the duration of each
Apply
method separately using thechrono
library. - We mainly used
sys/times.h
library to obtain the CPU usage. However, we used theproc/status
to create a secondary measure in order to verify. - We read the
/proc/self/status
file to obtain the CPU RAM usage. - We used jtop to measure GPU usage on the Jetson Nano and Jetson Xavier NX. We use jetson-stats to measure GPU usage on the Jetson Xavier AGX and the Jetson AGX Orin.
- We used
cudaMemGetInfo
from CUDA to measure GPU RAM usage.
This is the hardware setup we used:
- On the Jetson Nano, we used Jetpack 4.5.3 and 10W 4 Core MAXN Power Mode (NVP model 0)
- On the Jetson Xavier NX, we used Jetpack 4.5.3 and 20W 6 Core Power Mode (NVP model 8)
- On the Jetson Xavier AGX, we used Jetpack 4.5.1 and 30W 8 Core Power Mode (NVP model 3)
- On the Jetson AGX Orin, we used Jetpack 5.0.2 and 50W 12 Core Power Mode (NVP model 3)
For each system, we also used jetson_clocks
to maximize the device clock frequency and thus the performance.
The following table summarizes CUDA ISP's performance results.
Algorithm | Jetson AGX Orin | Jetson Xavier AGX | Jetson Xavier NX | Jetson Nano | ||||
---|---|---|---|---|---|---|---|---|
Buffer size | 1080p | 4K | 1080p | 4K | 1080p | 4K | 1080p | 4K |
Duration (ms) | ||||||||
CudaShift | 0.82 | 1.52 | 1.56 | 4.18 | 0.71 | 1.82 | 2.10 | 7.52 |
CudaDebayer | 0.68 | 1.30 | 1.93 | 5.74 | 0.79 | 2.13 | 2.40 | 8.75 |
CudaWhiteBalancer (Gray World Algorithm) | 0.84 | 1.66 | 1.94 | 5.21 | 0.99 | 2215 | 2.51 | 8.51 |
CudaWhiteBalancer (Histogram Stretch Algorithm) | 1.24 | 1.91 | 2.33 | 6.87 | 1.11 | 2.71 | 3.18 | 11.06 |
CudaColorSpaceConverter | 0.91 | 1.60 | 1.23 | 3.13 | 0.59 | 1.31 | 2.05 | 7.70 |
Framerate (fps) | ||||||||
CudaShift | 1216 | 660 | 641 | 239 | 1408 | 550 | 475 | 132 |
CudaDebayer | 1479 | 771 | 519 | 174 | 1259 | 469 | 415 | 114 |
CudaWhiteBalancer (Gray World Algorithm) | 1197 | 603 | 515 | 191 | 1011 | 451 | 398 | 117 |
CudaWhiteBalancer (Histogram Stretch Algorithm) | 807 | 522 | 429 | 145 | 902 | 368 | 314 | 90 |
CudaColorSpaceConverter | 1104 | 623 | 814 | 319 | 1697 | 761 | 487 | 129 |
CPU usage (%) | ||||||||
CudaShift | 0.198 | 0.218 | 0.285 | 0.255 | 0.287 | 0.356 | 0.800 | 0.817 |
CudaDebayer | 0.121 | 0.161 | 0.238 | 0.237 | 0.263 | 0.280 | 0.873 | 0.665 |
CudaWhiteBalancer (Gray World Algorithm) | 0.201 | 0.277 | 0.338 | 0.316 | 0.443 | 0.471 | 1.286 | 1.299 |
CudaWhiteBalancer (Histogram Stretch Algorithm) | 0.260 | 0.280 | 0.351 | 0.341 | 0.527 | 0.442 | 1.569 | 1.454 |
CudaColorSpaceConverter | 0.172 | 0.201 | 0.246 | 0.237 | 0.278 | 0.251 | 0.647 | 0.680 |
CPU RAM (MB) | ||||||||
CudaShift | 89.7 | 90.8 | 90.0 | 91.5 | 89.9 | 90.0 | 77.5 | 88.7 |
CudaDebayer | 44.4 | 44.5 | 41.8 | 42.2 | 41.7 | 42.9 | 33.0 | 33.4 |
CudaWhiteBalancer (Gray World Algorithm) | 99.8 | 99.1 | 108.5 | 107.4 | 107.6 | 108.1 | 40.6 | 77.5 |
CudaWhiteBalancer (Histogram Stretch Algorithm) | 99.2 | 99.3 | 107.7 | 108.6 | 108.1 | 107.5 | 52.1 | 33.0 |
CudaColorSpaceConverter | 44.1 | 44.3 | 42.0 | 42.1 | 42.1 | 42.0 | 33.2 | 89.1 |
GPU usage (%) | ||||||||
CudaShift | 12.29 | 7.81 | 18.03 | 13.00 | 11.59 | 4.23 | 85.32 | 25.96 |
CudaDebayer | 11.60 | 13.04 | 28.42 | 29.00 | 7.28 | 10.38 | 67.27 | 42.53 |
CudaWhiteBalancer (Gray World Algorithm) | 9.75 | 19.27 | 17.22 | 25.54 | 4.70 | 15.15 | 20.24 | 75.14 |
CudaWhiteBalancer (Histogram Stretch Algorithm) | 8.89 | 24.42 | 17.00 | 24.56 | 5.36 | 17.11 | 26.84 | 83.35 |
CudaColorSpaceConverter | 4.85 | 9.14 | 13.47 | 24.68 | 4.64 | 15.36 | 20.11 | 81.19 |
GPU RAM (MB) | ||||||||
CudaShift | 26.8 | 28.4 | 43.0 | 85.7 | 42.9 | 63.7 | 41.6 | 43.3 |
CudaDebayer | 6.0 | 6.1 | 10.7 | 10.3 | 9.8 | 11.2 | 11.9 | 11.8 |
CudaWhiteBalancer (Gray World Algorithm) | 33.7 | 32.7 | 57.5 | 61.6 | 56.8 | 59.6 | 40.6 | 31.5 |
CudaWhiteBalancer (Histogram Stretch Algorithm) | 33.2 | 33.4 | 56.8 | 66.3 | 57.1 | 62.2 | 52.1 | 52.4 |
CudaColorSpaceConverter | 5.4 | 6.0 | 9.9 | 10.4 | 10.0 | 10.3 | 11.2 | 11.8 |