Skip to content

Latest commit

 

History

History
45 lines (35 loc) · 2.94 KB

README.md

File metadata and controls

45 lines (35 loc) · 2.94 KB

backup_1

Config files for my GitHub profile.

更新

此项目移动到另一个仓库: https://github.com/param0037/DECX

This repository is moved and kept updated on my another repository: https://github.com/param0037/DECX

宗旨

本项目本来是我学习CUDA和cpu并发编程的记录。如果您是正在做毕设的大学生，不想配置主流复杂的BLAS库，并且在C++的层面编程，不妨试一下我写的BLAS库，希望可以帮到您。

警告

目前只适配了Windows端

做了以下更新：

1. 重写了方阵矩阵乘的CUDA内核，速度大幅度提升。

2. 重写了二维滑窗卷积CUDA内核，优化了速度。

3. 增加了基于im2col方法的卷积，适用于卷积神经网络。

4. 增加了三个内存池，优化了内存管理。

5. 增加了简易的线程池，基于c++11的future和function实现，目前的缺点：在低频（2.2GHz）下cpu负荷只有80%。

6. 增加了cpu端的矩阵乘算子。

7. 重写了fft1D和fft2D的CUDA内核，向量化访存，并尽可能地利用共享内存提速。

8. 增加了Tensor，TensorArray类，并给每个类都增加了其GPU版本，避免了在主机端分配内存并拷贝。

9. 重写了NLM（非局部均值滤波）内核，直接读取uchar，并应用共享内存，速度提升

若有留言请通过[email protected]联系我。

purpose

Originally, I create this project to store some of my notes when I was learning CUDA and CPU concurrency programming. If you are a collage student who is busy with your graduation projects, and you are somehow not allowed to use or tired to config the mainstream BLAS library, maybe it's a good way to have a try on my library. I hope I can help you.

WARNINGS

Currently the source codes only work on Windows.

updates

1. I rewrote the CUDA kernel used in matrix multiplication. The speed is boost dramatically.

2. I rewrote the CUDA kernel used in 2D convolution. The speed is boost dramatically.

3. A new method to improve performance in 2D convolution called "im2col" is implemented, which is suitable for convolution neuro networks.

4. Three memory pools is added, which optimizes the memory management.

5. I also create a thread pool based on the new features in c++11. But it has some drabacks, like, the occupancy only reaches to 80% of CPU when the frequency is relatively low (2.2GHz).

6. New GEMM operators on CPU is added.

7. I rewrote CUDA kernels used in fft1D and fft2D by vectorizing memory access, and using shared memory when it is possible.

8. I increase Tensor and TensorArray classes. At the meantime, I created the corresponding classes operates on GPU to avoid some unnecessary allocation on host and memory copying between host and device.

9. I rewrote NLM (non-local means) kernels, which can read data in uchar directly, and makes full use of shared memory.

Do not hesitate to contact me via [email protected] if you have any comment.