twang006/tw

Hybrid CPU and GPU real-time dynamic digital image correlation engine and application

TW

Parallel DIC Engine and its application research codes.

Introduction

This repository provides a pipelined real-time DIC system implementation that unifying the computation capabilities of both CPU and GPU. The system framework is based on the figure below:

1. TW_Engine: The revised implementations of the paDIC algorithm to make thame more suitable for real-time systems and applications
2. TW_EngineTester: Use Google Test to perform the unit test of the TW_Engine
3. TW_Core: The implementation of the proposed real-time DIC system.

Dependencies

*Note: Please make sure you have at least one camera connected to the computer before you start.

1. Intel Math Kernel Library (MKL): using fftw3 to do fast Fourier transform (FFT) and LAPACK routine to solve linear system in parlalel on CPU.
2. CUDA 8.0+: for parallel computing on NVIDIA GPUs.
3. CUFFT: associated with CUDA, for perform parallel FFT on GPU.
4. Qt 5.5+ with OpenGL Integration: for GUI and multi-media used in App_DPRA.
5. OpenCV 3.1+: for fast and convenient image I/O.

References

[1] Zhang, L., Wang, T., Jiang, Z., Kemao, Q., Liu, Y., Liu, Z., ... & Dong, S. (2015). High accuracy digital image correlation powered by GPU-based parallel computing. Optics and Lasers in Engineering, 69, 7-12.

[2] Wang, T., Jiang, Z., Kemao, Q., Lin, F., & Soon, S. H. (2016). GPU accelerated digital volume correlation. Experimental Mechanics, 56(2), 297-309.

[3] Wang, T., Kemao, Q., Seah, H. S., & Lin, F. (2018). A flexible heterogeneous real-time digital image correlation system. Optics and Lasers in Engineering, 110, 7-17.

tw's Issues

cuda执效率

感谢您的分享，我们基于您的相关计算实现了较好的任务效果。但仍然存在一些问题，望您给出一些建议。
我们的硬件是intel i7处理器（13代）gpu是nvidia3050。在实际运行中我们发现gpuuu并行加速运算效率并不比多核CPU快多少。
甚至多核cpu在执行ICGN计算时还比gpu端快0.333秒。我们将gpu端的线程数设置为您原代码的4倍，但执行速度的提升很少。
实在找不到其他优化的方法，还望您能给些建议。十分感谢🙏

please help me

Hello author, I would like to ask where is the include "ui_glcanvas.h" file in your code? I don't have this file in the archive I clone.

CUDA error at D: TW-master TW TW Engine TW-paDIC_CuICGN2D,cu:1294 code=700

感谢您分享的工作，有一些问题想向您请教。
我们的GPU型号是RTX 3050，cuda版本为12.1
1.在调用TW_paDIC_cuFFTCC2D.cu时，我们发现全局图像计算的结果更为准确（我们用了两张v方向上偏移-4的图进行测试，u=0，v=-4、-3以及zncc=0.8xx），但采用兴趣区域（从上述图像上截取了一个小的区域，但实际偏移应该不变，可计算结果却变成了，u=0,v=-14，zncc=0.4xx）计算时结果确比实际值偏移了不少？
2.在调用TW_paDIC_cuICGN2D.cu时，最后一个核函数ICGN_Computation_Kernel之前的核函数计算结果似乎都没问题，也不存在线程访问冲突之类的。但进入到ICGN_Computation_Kernel的计算后就会报错CUDA error at D:Tw-master TW TW_Engine TW-paDIC_CuICGN2D.cu:1285 code=700(cudaErrocuHandleICGN.m_d_fSubsetT, sizeof(real_t)* m_ipoINumber* (m isubsetsize + 1), cuda
并且发现和函数的输入targetimg的灰度信息（我们输出了一下单个兴趣点的对应值）也被改变了（从142变为156）。

3.在统计FFT和ICGN运行时间时存在统计时间不稳定，有时极大有时极小。
上述问题是因为线程或者资源分配导致了内存访问冲突所导致么？不知道您是否提供些修改或者优化建议？