xblas_gemv implementation

[ramcn]

Hello,

I am trying to implement the xblas_gemv similar to xblas_gemm and use it to accelerate a genomics tool. And was hoping to get the code changes reviewed as the test is not working as expected.

Here is the related code changes in my clone.

1. XBLAS_GEMV Implementation: ramcn/gemxx-2017@6295836
2. XBLAS_GEMV Test case: ramcn/gemxx-2017@899b54a

When I run the test case gemv_xblas_main.cpp, the resultant matrix C is all zeroes and I am unable to figure out what the issue is. It will be great if the gemx developers can spot anything obviously wrong with the above implementation.

Thank you,
Ram

[lisaliu1]

Hi Ram,

We have the latest FPGA accelerated gemm and gemv software API. See the example usage of gemv from the following link:
https://github.com/Xilinx/Vitis_Libraries/tree/master/blas/L3/examples/gemv

You can follow the instructions from the link below to try them on Nimbix with our Alvel U200 card.
https://xilinx.github.io/Vitis_Libraries/blas/user_guide/L3/L3.html

Thanks,
Lisa

[ramcn]

Hello Lisa,

Thanks for the response and pointers about the latest API. I will check it out. Meanwhile I am able to resolve the issue with my implementation of GEMV. (I had to explicitly pass GEMX_runGemv=1 GEMX_runGemm=0 to the make command during build). If you any recommendation about the optimal values of the block buffer size of the GEMV engine, kindly let me know.

One reason I wanted to stick to gemx-2017.4 was because it seems more suitable and easily extendable for my requirement. For my requirement, once I have the GEMV implementation tested, I want to extend it to implement the GRU kernel which the genomics tool uses. Please refer the grumod_step function this module which invokes GEMV. [link]

Hope this sounds feasible and in the right direction.

Cheers,
Ram