├── E_coli_K12
│ ├── benchmark # benchmark CPU && FPGA
│ ├── compute.xclbin # xclbin
│ ├── do_index # index E_coli_K12
│ ├── gatk_cpu # run gatk with cpu mode
│ ├── gatk_fpga # run gatk with fpga mode
│ ├── input # source file
│ │ ├── fasta
│ │ └── fastq
│ └── libcompute.so # jni impl by xilopencl
├── prebuilt
│ └── gatk.jar # gatk prebuit
├── README.md # Readme
└── src
├── gatk # gatk patch && build
│ ├── gatk.patch
├── jni
│ └── compute # libcompute source
└── vitis
├── compute_kernels # vitis kernel
├── gene_system
└── gene_system_hw_link
- Configurable compute units
- Configurable parallel scale
- Configurable computational refinement
- Variable input and output length
- Multi-core parallel computing
- Definable sequence length
- Flexible configuration according to hardware resource size
- Supports different scenarios such as Vitis CL/HLS/SD Accel
- Zynq/MP（with generic uio）
- Ultra96 6cu @ 500MHz
- 7 series / UltraScale / Plus / HBM (xdma)
- VCU1525 24cu @ 250 MHz
- Vitis (OpenCL)
- C1100 6cu @ 300 MHz
-
depend
- Ubuntu Linx
- Xilinx XRT for C1100
- cmake bwa samtools openjdk-11-jdk
-
run test prebuilt
git clone https://github.com/ChomperT/pairHMM sudo apt-get install cmake bwa samtools cd pairHMM/E_coli_K12 ./do_index # bwa index ./benchmark # benchmark... ./gatk_cpu # gatk CPU ./gatk_fpga # gatk FPGA
-
Build JNI Library libcompute.so
cd pariHMM && mkdir build cd build && cmake ../src/jni/compute make
-
Build Gatk.jar
cd pairHMM/src/gatk git clone https://github.com/broadinstitute/gatk cd gatk git apply ../gatk.patch ./gradlew cp build/libs/gatk.jar ../../prebuilt/
# benchmark C1100 cu: 6, group: 11
# cat /proc/cpuinfo | grep "model name" | head -n 1
model name : 12th Gen Intel(R) Core(TM) i7-12700K
# ./benchmark gatk.in
INFO: Reading compute.xclbin
Loading: 'compute.xclbin'
181217 pairs need compute!
..........................
109, -1.675846099853515625
109, -1.675838470458984375
109, -1.800930023193359375
109, -1.800930023193359375
109, -1.660717010498046875
109, -1.660717010498046875
fpga cost: 2653.1589ms
cpu cost: 90585.0078ms- FP32 calculation accuracy error (0.01%)
- Part of the data FP32 calculation overflow
- DSP48 over-occupancy causes wiring difficulties
- C1100 xdma driver does not work properly
- Testing Versal Series
- Hardware's DSP58 performance improvement is huge and requires VCK5000
- Improve Gatk data input efficiency
- Improve Gatk data input capability and increase parallel computing to facilitate the use of multiple FPGA computing units
- Break OpenCL frequency, scale limits
- C1100 300/500 Mhz limits the number and frequency of deployed cores
- XDMA/QDMA implementation performance would be better if the driver is stable
- Current area footprint is less