oraluben opened a new issue, #292:
URL: https://github.com/apache/tvm-ffi/issues/292

   tvm-ffi just introduced [cubin 
launcher](https://github.com/apache/tvm-ffi/pull/283) to allow embed one 
cuda/cubin file into c++ source code, and call kernel via ffi. This feature 
allows a more flexible integration of cuda kernel.
   
   However, while investigating the feature, we've found it not that close to 
the simplest cuda-development workflow (which requires more efforts while 
migrating from existing kernels), it also doesn't reuse the existing CUDA utils 
in cmake. Also, it's rely on cuda runtime library and sometimes this introduces 
unexpedted constraints to CUDA runtime version.
   
   We've implemented a POC to resolve such quirks 
(https://github.com/oraluben/kernel-lib-template/commit/e973a2e256f46fb08d108d520e9797499d9c45b8)
 and would like to propose some changes to the current cubin launcher APIs, and 
would like to discuss with the team before coding.
   
   Those changes might include:
   1. cmake utils to allow user to generate header from cuda/cubin files
   2. tvm-ffi macros to load libraries from generated header
   3. switch between cudart api and cuda driver api, and necessary checks to 
detect major version mismatch.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to