AnchorWave uses the SIMD instructions to speed up the dynamic programming algorithm. Specific functions have been implemented for SSE2, SSE4.1, AVX2 and AVX512 instruction sets.
Empirically, the time cost: SSE2 > SSE4.1 > AVX2 > AVX512.
The conda version compile all of them and pick the fastest version to run.
To check what CPU instructions are supported by your machine with Linux system, you could run this command:
cat /proc/cpuinfo | grep "flags" | uniq
By default, we assume the machine supports SSE4.1.
Clone the repository, and replace the default CMakeLists.txt with the avx512 one.
git clone https://github.com/baoxingsong/anchorwave.git
cd anchorwave
mv CMakeLists_avx512.txt CMakeLists.txt
cmake ./
make
Clone the repository, and replace the default CMakeLists.txt with the avx2 one.
git clone https://github.com/baoxingsong/anchorwave.git
cd anchorwave
mv CMakeLists_avx2.txt CMakeLists.txt
cmake ./
make
Clone the repository, and replace the default CMakeLists.txt with the ARM one.
git clone https://github.com/baoxingsong/anchorwave.git
cd anchorwave
mv CMakeLists_arm.txt CMakeLists.txt
cmake ./
make
Clone the repository, and replace the default CMakeLists.txt with the sse2 one.
git clone https://github.com/baoxingsong/anchorwave.git
cd anchorwave
mv CMakeLists_sse2.txt CMakeLists.txt
cmake ./
make
AnchorWave was not tested on that kind of platforms. High likely would not work properly.
We compiled AnchorWave in 4 different ways on a machine that having two Intel(R) Xeon(R) W-2295 CPUs and 128Gb RAM.
We aligned the maize Mo17 CAU assembly against the B73 V4 reference assembly. Input data were prepared using the following commands:
wget ftp://ftp.ensemblgenomes.org/pub/plants/release-34/gff3/zea_mays/Zea_mays.AGPv4.34.gff3.gz
gunzip Zea_mays.AGPv4.34.gff3.gz
wget ftp://ftp.ensemblgenomes.org/pub/plants/release-34/fasta/zea_mays/dna/Zea_mays.AGPv4.dna.toplevel.fa.gz
gunzip Zea_mays.AGPv4.dna.toplevel.fa.gz
wget https://download.maizegdb.org/Zm-Mo17-REFERENCE-CAU-1.0/Zm-Mo17-REFERENCE-CAU-1.0.fa.gz
gunzip Zm-Mo17-REFERENCE-CAU-1.0.fa.gz
sed -i 's/>chr/>/g' Zm-Mo17-REFERENCE-CAU-1.0.fa
./anchorwave_avx512 gff2seq -r Zea_mays.AGPv4.dna.toplevel.fa -i Zea_mays.AGPv4.34.gff3 -o cds.fa
minimap2 -x splice -t 80 -k 12 -a -p 0.4 -N 20 Zm-Mo17-REFERENCE-CAU-1.0.fa cds.fa > cds.sam
minimap2 -x splice -t 80 -k 12 -a -p 0.4 -N 20 Zea_mays.AGPv4.dna.toplevel.fa cds.fa > ref.sam
Then we run the following 4 commands on the same computer simultaneously:
/usr/bin/time ./anchorwave_sse2 genoAli -i Zea_mays.AGPv4.34.gff3 -as cds.fa -r Zea_mays.AGPv4.dna.toplevel.fa -a cds.sam -ar ref.sam -s Zm-Mo17-REFERENCE-CAU-1.0.fa -n sse2.anchors -o sse2.anchorwave.maf -f sse2.anchorwave.f.maf -t 1 -IV >sse2.log 2>&1
/usr/bin/time ./anchorwave_sse4.1 genoAli -i Zea_mays.AGPv4.34.gff3 -as cds.fa -r Zea_mays.AGPv4.dna.toplevel.fa -a cds.sam -ar ref.sam -s Zm-Mo17-REFERENCE-CAU-1.0.fa -n sse4.1.anchors -o sse4.1.anchorwave.maf -f sse4.1.anchorwave.f.maf -t 1 -IV >sse4.1.log 2>&1
/usr/bin/time ./anchorwave_avx2 genoAli -i Zea_mays.AGPv4.34.gff3 -as cds.fa -r Zea_mays.AGPv4.dna.toplevel.fa -a cds.sam -ar ref.sam -s Zm-Mo17-REFERENCE-CAU-1.0.fa -n avx2.anchors -o avx2.anchorwave.maf -f avx2.anchorwave.f.maf -t 1 -IV >avx2.log 2>&1
/usr/bin/time ./anchorwave_avx512 genoAli -i Zea_mays.AGPv4.34.gff3 -as cds.fa -r Zea_mays.AGPv4.dna.toplevel.fa -a cds.sam -ar ref.sam -s Zm-Mo17-REFERENCE-CAU-1.0.fa -n avx512.anchors -o avx512.anchorwave.maf -f avx512.anchorwave.f.maf -t 1 -IV >avx512.log 2>&1
gtime -l anchorwave_arm genoAli -i Zea_mays.AGPv4.34.gff3 -as cds.fa -r Zea_mays.AGPv4.dna.toplevel.fa -a cds.sam -ar ref.sam -s Zm-Mo17-REFERENCE-CAU-1.0.fa -n arm.anchors -o arm.anchorwave.maf -f arm.anchorwave.f.maf -t 1 -IV >arm.log 2>&1
Compare the output alignments:
maf-convert sam sse2.anchorwave.maf > sse2.anchorwave.sam
maf-convert sam sse4.1.anchorwave.maf > sse4.1.anchorwave.sam
maf-convert sam avx2.anchorwave.maf > avx2.anchorwave.sam
maf-convert sam avx512.anchorwave.maf > avx512.anchorwave.sam
maf-convert sam arm.anchorwave.maf > arm.anchorwave.sam
sed -i -r 's/[0-9]+H//g' sse2.anchorwave.sam
sed -i -r 's/[0-9]+H//g' sse4.1.anchorwave.sam
sed -i -r 's/[0-9]+H//g' avx2.anchorwave.sam
sed -i -r 's/[0-9]+H//g' avx512.anchorwave.sam
sed -i -r 's/[0-9]+H//g' arm.anchorwave.sam
samtools view -O CRAM --reference Zea_mays.AGPv4.dna.toplevel.fa sse2.anchorwave.sam | samtools sort -O CRAM - > sse2.anchorwave.cram
samtools view -O CRAM --reference Zea_mays.AGPv4.dna.toplevel.fa sse4.1.anchorwave.sam | samtools sort -O CRAM - > sse4.1.anchorwave.cram
samtools view -O CRAM --reference Zea_mays.AGPv4.dna.toplevel.fa avx2.anchorwave.sam | samtools sort -O CRAM - > avx2.anchorwave.cram
samtools view -O CRAM --reference Zea_mays.AGPv4.dna.toplevel.fa avx512.anchorwave.sam | samtools sort -O CRAM - > avx512.anchorwave.cram
samtools view -O CRAM --reference Zea_mays.AGPv4.dna.toplevel.fa arm.anchorwave.sam | samtools sort -O CRAM - > arm.anchorwave.cram
samtools view -O SAM -o sse2.anchorwave.sorted.sam sse2.anchorwave.cram
samtools view -O SAM -o sse4.1.anchorwave.sorted.sam sse4.1.anchorwave.cram
samtools view -O SAM -o avx2.anchorwave.sorted.sam avx2.anchorwave.cram
samtools view -O SAM -o avx512.anchorwave.sorted.sam avx512.anchorwave.cram
samtools view -O SAM -o arm.anchorwave.sorted.sam arm.anchorwave.cram
diff sse4.1.anchorwave.sorted.sam avx2.anchorwave.sorted.sam
diff sse4.1.anchorwave.sorted.sam avx512.anchorwave.sorted.sam
diff sse2.anchorwave.sorted.sam avx512.anchorwave.sorted.sam
diff arm.anchorwave.sorted.sam avx512.anchorwave.sorted.sam
Different compiling ways generated identical alignments
| peak memory (Gb) | wall time | |
|---|---|---|
| SSE2 | 20.1 | 26:47:17 |
| SSE4.1 | 20.6 | 24:05:07 |
| AVX2 | 20.1 | 21:40:00 |
| AVX512 | 20.1 | 18:31:39 |
| ARM | 34.2 | 18:08:57 |
*the cost comparison was performed for once (did not perform multiple times and calculate the average)
*the test of SSE2, SSE4.1, AVX2 and AVX512 were conduction on a machine with Intel Xeon W-2295 CPU
*the test of ARM was conduction on a MacBook Pro with M2 MAX chip