Error: Missing DSA or platform repo. Stop.

[liyb13]

when I try to compile the gemm project, the following error happens: /tmp/cc8fJfho.s: Assembler messages: /tmp/cc8fJfho.s:318: Error: expecting string instruction after rep' /tmp/cc8fJfho.s:386: Error: expecting string instruction afterrep' /tmp/cc8fJfho.s:6845: Error: expecting string instruction after rep' /tmp/cc8fJfho.s:16931: Error: expecting string instruction afterrep' /tmp/cc8fJfho.s:21658: Error: expecting string instruction after `rep' make: *** [out_host/gemx_host.exe] Error 1

[lisaliu1]

Which version of SDx are you using? If you compile for vu9pf1, you should use SDx2017.1

[lisaliu1]

string instruction afterrep' /tmp/cc8fJfho.s:6845: Error: expecting string instruction after rep' /tmp/cc8fJfho.s:16931: Error:

This error is caused by the lower version of binutils you are using. Can you run following command to check which version of binutils you are using. ld -v

Currently we use 2.26, if your version number is lower than that, you might need to upgrade your binutils.

[liyb13]

Thanks for your help. I have successfully compile the project. I run it on AWS F1 and then get some results. It reports that kernel result does not match the reference.

[lisaliu1]

Which command did you use to build .xclbin? And how do you test the .xclbin on F1? I will try to see if I can reproduce your results on my side.

[liyb13]

Thanks for your reply. I did as the README.md (https://github.com/Xilinx/gemx):

• Compiling and building the FPGA image: make run_hw SDA_FLOW=hw GEMX_ddrWidth=32 GEMX_gemmMBlocks=8 GEMX_gemmKBlocks=8 GEMX_gemmNBlocks=8 GEMX_numKernels=4 GEMX_runGemv=0 GEMX_runTransp=0 GEMX_runGemm=1 GEMX_part=vu9pf1 GEMX_kernelHlsFreq=250 GEMX_kernelVivadoFreq=300 GEMX_useURAM=1 GEMX_vivadoFlow=EXP

• Compiling host code: make GEMX_ddrWidth=32 GEMX_gemmMBlocks=8 GEMX_gemmKBlocks=8 GEMX_gemmNBlocks=8 GEMX_numKernels=4 out_host/gemx_api_gemm.exe

• AWS F1 test: source /opt/Xilinx/SDx/2017.1.rte/setup.sh gemx_api_gemm.exe gemx.awsxclbin 512 512 512

• Report: Compared 262144 values: exact match 259072 within tolerance 135 mismatch 2937 INFO: accelerator kernel 0 result does NOT match the reference Compared 262144 values: exact match 241760 within tolerance 136 mismatch 20248 INFO: accelerator kernel 1 result does NOT match the reference Compared 262144 values: exact match 0 within tolerance 0 mismatch 262144 INFO: accelerator kernel 2 result does NOT match the reference Compared 262144 values: exact match 0 within tolerance 0 mismatch 262144 INFO: accelerator kernel 3 result does NOT match the reference

Another question: I think "gemx_api_gemm.exe gemx.awsxclbin 512 512 512" means the buffer size of A is 512*512. Then buffer A could store the 512*512*16bit. While the buffer size of A is GEMX_gemmMBlocks*GEMX_ddrWidth x GEMX_gemmKBlocks*GEMX_ddrWidth = 256*256. Should I modify the default values of variables?

[lisaliu1]

I tried the steps you mentioned, and it works on F1 for me. Can you check following items:

1. if you copied the generated out_hw/xbinst to the out_hw/ directory on F1?
2. if you have setup the env correctly? Below is the content of the run.sh I used to run this command source /opt/Xilinx/SDx/2017.1.rte/setup.sh pushd out_hw/xbinst export SDACCEL_TIMELINE_REPORT=true export SDACCEL_DEVICE_PROFILE=true export LD_LIBRARY_PATH=$PWD/../../lib/opencv:$PWD/../../lib:$LD_LIBRARY_PATH GEMX_HOST_DIR=../../out_host GEMX_HW_DIR=../ ${GEMX_HOST_DIR}/gemx_api_gemm.exe ${GEMX_HW_DIR}/gemx.awsxclbin 512 512 512 popd

regarding your question: the 512 512 and 512 mean the value for dimention M, K and N. the matrix multiplication is doing C = A * B, where A has dimension M x K, B has dimension K x N and C has dimension M x N the local buffer size for A, B and C is controlled by the setup of GEMX_gemmMBlocks, GEMX_gemmKBlocks ang GEMX_gemmNBlocks and GEMX_ddrWidth.

[liyb13]

Thanks for your reply. But I can not find /lib/opencv on aws f1 instance

[lisaliu1]

you can remove /lib/opencv from the LD_LIBRARY_PATH, it should still work.

[liyb13]

Thanks. It is declaired that the GEMX supports sparse matrix multiplication. Would you please tell me the detailed procedures to run the simulation and implementation for the sparse mode. Many thanks.

[lisaliu1]

please sync your master branch with the repository first. After that, please follow the steps below to run cpu emulation for SPMV implementation.

1. run settings64.sh / settings64.csh to set up the SDx 2017.4 environment
2. navigate to gemx/
3. run

make clean

4. run

make run_cpu_em GEMX_ddrWidth=16 GEMX_argInstrWidth=1 GEMX_numKernels=1 GEMX_runGemv=0 GEMX_runGemm=0 GEMX_runTransp=0 GEMX_runSpmv=1 GEMX_dataType=float GEMX_part=ku115 GEN_BIN_PROGRAM="spmv 96 128 256 none A0 B0 C0 spmv 0 0 0 data/spmv/Rucci1.mtx.gz A1 B1 C1"

if you change "run_cpu_em" to "run_hw_em" and add "HWEMUGUI=1" then you can run hw emulation with GUI to deplay the waveform of the SPMV simulation results.