###Latest release v1.0.1 (Date: Feb 12th, 2016) ###Requirements

1. Operating system(s): Linux
2. Programming language(s): CUDA C/C++, PTX assembly
3. Other requirement(s): CUDA 7.0 or later, GCC/G++ 4.4.7 or later, CUDA- enabled GPUs with Compute Capability of 3.5 or higher
4. License: MIT License

###Run code

CUDAMPF [filterOpt] [modelOpt] [databaseOpt]

There are three filter options:

• "-ssv": Single Segment Viterbi filter (since HMMER3.1)
• "-msv": Multiple Segment Viterbi filter (since HMMER3.0)
• "-vit": Full P7Viterbi filter (since HMMER3.0)

Sequence database file must be FASTA format. An available protein sequence database is here: ftp://ftp.ebi.ac.uk/pub/databases/Pfam/releases/Pfam27.0/

Model files must be Profile HMM format (usually with ".hmm" extension), which looks like:

In folder query models, there are 24 HMM models with different length provided.

###Notes

1. Please make sure paths of library and compiler correct and detectable when using Makefile to compile whole project: make all.(i.e, CUDA_PATH, LIBRARIES, INCLUDES, HOST_COMPILER and NVCC)
2. Keep all kernel files (i.e, LMEM_MSV.cuh) in the same project folder. Otherwise, it will raise an error unable to open LMEM_MSV.cuh for reading!
3. Current release v1.0.1 only contains the code for one-batch benchmarking. Streaming benchmarking will be released in subsequent version.
4. Data preprocessing on host side are not optimized (i.e, sequence/model format), which are only plain implementations to guarantee all inputs correct.
5. kernel time: xxxx(ms) indicates the consuming time on device.
6. Final results show the number of sequence pass through filters, which is 100% same as hmmsearch . Anyone wants to check out P-value for each sequence may modify RTC_XXX.cpp and print out variable p_value.

###Contact Any questions about the project please send email to: hjiang5 AT stevens DOT edu

##Licence Please check the MIT License for CUDAMPF.