Data-Aware Compression of Neural Networks

Falahati, Hajar; Peyro, Masoud; Amini, Hossein; Taghian, Mehran; Sadrosadati, Mohammad; Lotfi-Kamran, Pejman; Sarbazi-Azad, Hamid

doi:10.1109/lca.2021.3096191

Cited by 3 publications

(2 citation statements)

References 2 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Step 3: Let's assume that = [ ] { } W 5 2, 5, 3, 8, 10 and = [ ] { } binary_seq1 5 0, 1, 1, 0, 0 . Based on the values in the array binary_seq1, W [1] and W [2] are confirmed to belong to the array x_seq. Therefore, based on the three sets of data, bianry_seq1 and W, binary_seq2 and V, and bianry_seq3 and U, all the data in x_seq, y_seq, and z_seq can be recovered, respectively.…”

Section: Decompression Algorithmmentioning

confidence: 99%

“…Data compression has always been one of the key issues in physics research. As far as today's data compression techniques are concerned, such as deep learning compression models [1][2][3][4][5][6][7][8][9], they mainly rely on relational data rather than random data with little correlation. However, many data in the real world are often not only weakly correlated with each other, but also discontinuous.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

A lossless compression method for random integer based on three-dimensional product- triangular chaotic iterative sequences

Zhang

2023

Phys. Scr.

View full text Add to dashboard Cite

Using the ergodicity of chaotic iterative sequences to realize data compression is a new research perspective. We find that, under suitable initial conditions, one or more local segments that are numerically identical to random integer sequences appear on the iterative sequence, which is a prerequisite for compression. Based on this, this paper designs a random integer lossless compression method based on three-dimensional product-triangular chaotic iterative sequences. The method proposed only needs to input a small amount of iterative initial information to compress a large amount of data through an iterative sequence of limited length. The key lies in three aspects. First, according to the characteristics of the data to be compressed, the iterative initial conditions suitable for compression are obtained by screening. Secondly, map the traversal results into a sequence of binary integers to complete the recording of key information. Finally, this binary integer sequence is rapidly compressed using a designed parity symmetric transformation algorithm, and decompression is achieved in its reverse process. As a new way to achieve compression, this approach is not only simple, but also requires less computation time. The experimental results show that the compression effect achieved by this method has obvious advantages in terms of compression ratio, data reconstruction quality, and compression and decompression speed. Keywords: trigonometric function, chaos, data compression

show abstract

Section: Decompression Algorithmmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

A lossless compression method for random integer based on three-dimensional product- triangular chaotic iterative sequences

Zhang

2023

Phys. Scr.

View full text Add to dashboard Cite

show abstract

Snake: A Variable-length Chain-based Prefetching for GPUs

Mostofi,

Falahati,

Mahani

et al. 2023

56th Annual IEEE/ACM International Symposium on Microarchitecture

View full text Add to dashboard Cite

Cross-core Data Sharing for Energy-efficient GPUs

Falahati,

Sadrosadati,

et al. 2024

ACM Trans. Archit. Code Optim.

View full text Add to dashboard Cite

Graphics Processing Units (GPUs) are the accelerator of choice in a variety of application domains because they can accelerate massively parallel workloads and can be easily programmed using general-purpose programming frameworks such as CUDA and OpenCL. Each Streaming Multiprocessor (SM) contains an L1 data cache (L1D) to exploit the locality in data accesses. L1D misses are costly for GPUs due to two reasons. First, L1D misses consume a lot of energy as they need to access the L2 cache (L2) via an on-chip network and the off-chip DRAM in case of L2 misses. Second, L1D misses impose performance overhead if the GPU does not have enough active warps to hide the long memory access latency. We observe that threads running on different SMs share 55% of the data they read from the memory. Unfortunately, as the L1Ds are in the non-coherent memory domain, each SM independently fetches data from the L2 or the off-chip memory into its L1D, even though the data may be currently available in the L1D of another SM. Our goal is to service L1D read misses via other SMs, as much as possible, to cut down costly accesses to the L2 or the off-chip DRAM. To this end, we propose a new data sharing mechanism, called Cross-Core Data Sharing (CCDS) . CCDS employs a predictor to estimate whether or not the required cache block exists in another SM. If the block is predicted to exist in another SM’s L1D, CCDS fetches the data from the L1D that contains the block. Our experiments on a suite of 26 workloads show that CCDS improves average energy and performance by 1.30 × and 1.20 ×, respectively, compared to the baseline GPU. Compared to the state-of-the-art data-sharing mechanism, CCDS improves average energy and performance by 1.37 × and 1.11 ×, respectively.

show abstract

Data-Aware Compression of Neural Networks

Cited by 3 publications

References 2 publications

A lossless compression method for random integer based on three-dimensional product- triangular chaotic iterative sequences

A lossless compression method for random integer based on three-dimensional product- triangular chaotic iterative sequences

Snake: A Variable-length Chain-based Prefetching for GPUs

Cross-core Data Sharing for Energy-efficient GPUs

Contact Info

Product

Resources

About