hyfshishen
commented
3 months ago Hi, sorry for this late reply, and thanks for your interest.
I just pushed the code of GSZ in this repo [LINK]. Also, the name "GSZ" is a temporary name used in paper submission (due to the SC24 double-blind policy). And after discussion, we decided to name this compressor still as cuSZp (and its name in SC24 paper will be modified into cuSZp2 in the camera-ready version to distinguish it from other baseline compressors). As a result, a more complete version including README and explanations of cuSZp2/GSZ code will be updated in this repository.
Here is some information for repository: [LINK].
Differences between cuSZp2/GSZ and cuSZp
- Higher Throughput: cuSZp2/GSZ has much higher throughput compared with cuSZp. Especially compressing floating-point data that are greater than 1 GB, cuSZp/GSZ can achieve around 300~500 GB/s throughput for compression and decompression. The reasons are due to optimized memory access patterns and latency control. BTW, based on my testing and algorithm cuSZp2/GSZ will have very promising results for sparse datasets (lots of 0 data points).
- Higher Compression Ratio: cuSZp2/GSZ supports two lossless encoding modes: plain-fixed-length-encoding and outlier-fixed-length-encoding, where as cuSZp only supports plain-fixed-length-encoding.
- For plain-fixed-length-encoding, it will have the same compression ratios as cuSZp.
- For outlier-fixed-length-encoding. it will have higher throughputs than plain-fixed-length-encoding but also with a slight decrease in throughput (but still much faster than cuSZp). Based on my testing, the compression ratio gain will be higher if the datasets are smooth (i.e. lots of values are consecutive, such as an array with value [..., 1.4, 1.6, 1.7, 2.5, 1.8, ....]).
To compile and use cuSZp2/GSZ as executable binary
This can be found in README in this repo [LINK]. The compression/decompression kernel implementations can be found in file /src/GSZ.cu.
To use cuSZp2/GSZ as an internal API.
- Include header file
./include/GSZ_entry.h - Compression and Decompression function:
void GSZ_compress_deviceptr(float* d_oriData, unsigned char* d_cmpBytes, size_t nbEle, size_t* cmpSize, float errorBound, cudaStream_t stream = 0); void GSZ_decompress_deviceptr(float* d_decData, unsigned char* d_cmpBytes, size_t nbEle, size_t cmpSize, float errorBound, cudaStream_t stream = 0);Note that these two functions assume original and compressed data arrays are both on GPU.
- An example of using them can be found in file
./examples/example_gpu_api.cpp, and the key part can be found below (where timer_GPU denotes a strict end-to-end measurement by CUDA EVENT):// GSZ compression. timer_GPU.StartCounter(); // set timer GSZ_compress_deviceptr(d_oriData, d_cmpBytes, nbEle, &cmpSize, errorBound, stream); float cmpTime = timer_GPU.GetCounter();
// GSZ decompression. timer_GPU.StartCounter(); // set timer GSZ_decompress_deviceptr(d_decData, d_cmpBytes, nbEle, cmpSize, errorBound, stream); float decTime = timer_GPU.GetCounter();
### How to use outlier-fixed-length-encoding mode.
In this repo [\[LINK\]](https://github.com/hyfshishen/SC24-cuSZp2), the README part only executes program with plain-fixed-length-encoding. If you'd like to try outlier-fixed-length-encoding, what you can do is:
1. Go to file: ```/src/GSZ_entry.cu```.
5. Replace line 162 with ```GSZ_compress_kernel_outlier<<<gridSize, blockSize, sizeof(unsigned int)*2, stream>>>(d_oriData, d_cmpBytes, d_cmpOffset, d_locOffset, d_flag, errorBound, nbEle);```. (just replace name in function from "plain" to "outlier")
6. Replace line 209 with ```GSZ_decompress_kernel_outlier<<<gridSize, blockSize, sizeof(unsigned int)*2, stream>>>(d_decData, d_cmpBytes, d_cmpOffset, d_locOffset, d_flag, errorBound, nbEle);```. (just replace name in function from "plain" to "outlier")
7. Redo compilation and execute binary ```gpu_api_GSZ``` based on information provided in README.
### Later actions about cuSZp2/GSZ code.
Right now cuSZp2/GSZ is in SC24-AD-AE procedure, so this repository may be updated from time-to-time. If this repo [\[LINK\]](https://github.com/hyfshishen/SC24-cuSZp2) is updated with future commits, the instructions mentioned above will work for the [07.21.2024 commit](https://github.com/hyfshishen/SC24-cuSZp2/tree/ef09e38e3600ff12e3bd2b3e049771ba05e0e29a). In all, what I will do for cuSZp2/GSZ in the future are listed above:
- Within July this year, more detailed descriptions and execution supports (e.g. double-precision supports) will be updated in repository [\[LINK\]](https://github.com/hyfshishen/SC24-cuSZp2).
- Within Aug this year, a complete code with a detailed manual will be updated in the repo (i.e. cuSZp). In other words, the compressor cuSZp2/GSZ will be merged into cuSZp.
Thanks again for your interests and sorry for the confusion in "compressor name" I made lol. Please let me know if there are more questions/bugs while using cuSZp2/GSZ :)
lessw2020
I saw the link regarding the new GSZ that will be presented at SC24. Looks very useful and thus wanted to see if the code will be posted here for use and testing? I'm working on compressing llm activations and seems GSZ might be an ideal fit. Thanks for any info!