Parallel Computing Final Project FineTunedHashML Spring 2020

The final repo is stored at gitlab, added some hash function to the original algorithm.

Dear Professor,
I’m Yiwei Victor Yang 2018533218 from your parallel computing class.
My teammate and I are struggling in selecting the project. I proposed to set up a projects transplant of a CPU LSH-Deep learning (, which is a technique newly proposed by Rice University and Intel co, ltd to utilize the avx512 instruction set and huge pages intrinsic in locality sensitive hashing which is the main overhead of the project. (

SLIDE is a project to reimplement BP on the CPU. After the LSH pre-processing, the paper proposed to utilize maximum inner product search (MIPS) which mainly do the Sparse matrix operations- Sparse Backpropagation or Gradient Update.

After the update of weight stored in CSR format Spmv, the paper do the Batch computing by OpenMP.

As research in the paper, we can see that the main overhead is pre-processing on CPU avx512 with huge pages cache-heap optimization on, which gained around 30% faster than the raw GPU Tesla V100. If we transplant that part(dynamic hashing) in GPU which is said not applicable by the author from intel, I think it may have a 50 percent chance to be faster than the current CPU version.
As the paper’s reference paper puts it: it gains something like 2 times faster after porting to Cuda, which triggers conflicts.

Our proposal is to make a comparison amid the raw Tensorflow on GPU, SLIDE on avx512, and SLIDE on GPU using the dataset mentioned in the paper.

My question is whether the transplant of the CPU dynamic cuckoo in the SLIDE do you think can have real speed up than the original version and if we attempted to transplant the program and get little speed up, will we eventually get the score?

Thanks a lot!