UCLA-VAST / HT-Deflate-FPGA

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

High-Throughput Compression

Table of Contents

Restrictions

Compression Format

The compression format here we use exactly follows RFC1951: each block of compressed data begins with 3 header bits and ends with code "7'b0000000". Since RFC1951 doesn't limit the compressed block size, we take the whole file as a single block to avoid extra overhead.

Deflate Kernel

We provide two kinds of kernels: one is for the Xilinx-AWS platform and the other is for the Intel-Altera HARPv2 platform. The HARPv2 kernel support a hash-chained depth of 3, while the Xilinx-AWS kernel has only one-level depth. To increase hash chain depth, you can refer to the corresponding hash_match.v in the HAPR2 kernel and custimize the depth. The trade-off here is the maximum clock frequency might be degraded.

To make the multiple rom files work, you have to modify the corresponding rom.v files to include the right path of .mif files. For example, in the line 90 of file "huffman_translate_d_extra_bits_rom.v", it should include its corresponding "huffman_translate_d_extra_bits_rom.mif". Assuming "huffman_translate_d_extra_bits_rom.mif" is in the directory of "C:/Users/Weikang/Desktop/work/design/xilinx/Compression_Vec16_V32/src/", you should change the path correspondingly.

Interface

The original design was implemented on Intel-Altera HARP and HARPv2 platforms, which have over 15GB/s CPU-FPGA communication bandwidth, to get a maximum end-to-end throughput of ~12GB/s. However, due to copyright issue we will not provide the communication interface code here. But it should be fairly easy for you to wrap the kernel with your own interface if you are a HARP/HARPv2 user and follow what our paper described.

You can also integrate this kernel into SDAccel flow on Xilinx Platforms, i.e. AWS F1 instance. However, as already discussed in our paper, the actual read/write PCIe bandwidth of SDAccel is very limited (less than 3GB/s). So we recommend you using AWS hdk flow to directly feed data from PCIe to the kernel and bypassing the FPGA DRAM access. You can find the relavent information from https://vast.cs.ucla.edu/sites/default/files/publications/st-accel-high.pdf.

If you don't know how to do it, no worries! We will release the interface which bypasses the FPGA DRAM and achieve an end-to-end throughput of >10 GB/s on AWS F1 instance. Please check back later.

Verification

Requirement

The provided exection files for decompression are generated through SDACCEL 2017.4. They are basically software emulation files and should be performed on the "xinlinx_vcu1525_dynamic_5_0" platform.

Execution

The verification procedure is similar to software emulation in xocc, you can refer to UG1023(v2017.4) Page 47-49 for detailed information.

First, create the emulation configuration file.

$ emconfigutil --platform xilinx_vcu1525_dynamic_5_0

Then, set your emulation mode to csim.

export XCL_EMULATION_MODE=sw_emu

Now you can execute the host application with proper arguments. For example, assume you have a compressed file "bib.comp" under the folder ${IN_DIR} and want to decompress it and store it as ${OUT_DIR}/bib.decomp. You need to type in the following command:

./host ${IN_DIR}/bib.comp ${OUT_DIR}/bib.decomp

About


Languages

Language:Verilog 96.8%Language:SystemVerilog 1.2%Language:HTML 1.1%Language:C 0.5%Language:Roff 0.2%Language:Common Lisp 0.2%Language:Coq 0.0%