The configuration and function handling have been improved.
The configuration now accepts the function signature as input, which will automatically call the correct kernel implementation.
Furthermore, a new gemm_args_t struct is defined holding all necessary layer parameters. These parameters are transferred from DRAM to TCDM to avoid significant function setup overheads.
The function signatures have been unified across all implementations.
[ ] Align other kernels with GEMM kernel approach.
[ ] Make sure GEMM is called correctly from all other kernels (FA-2, FusedConcatLinear)
[ ] Ensure that the correct impl symbol is called (currently not the case!)
The configuration and function handling have been improved.
gemm_args_tstruct is defined holding all necessary layer parameters. These parameters are transferred from DRAM to TCDM to avoid significant function setup overheads.implsymbol is called (currently not the case!)