Xilinx
vectorized kernel is currently erroneous in matrix-vector
Could you please inform me when the issue noted in # FIXME (the vectorized kernel is currently erroneous) (in matrix-vector) will be resolved, or provide insights into how to address it?
I have fixed this by set
[ (k // 2 , 2), (m, k), (2, 1), ]
Running Kernel (iteration 0). Running Kernel (iteration 1). Verifying against reference matmul ... Verify time: 0.00 s.
Avg NPU matmul time: 1948.00us. Avg NPU gflops: 17.23
Min NPU matmul time: 1948.00us. Max NPU gflops: 17.23
Max NPU matmul time: 1948.00us. Min NPU gflops: 17.23
PASS!
Please refer to the matrix-vector kernel I linked in your other issue for a working implementation. It should be pretty simple to port that fixed kernel to this repository and we welcome and appreciate PRs, if you would like to take it on. Unfortunately, my to-do list is quite long and this is not highest priority at the moment.
