[QST] Questions about correctness test and layout

Hello, I have several question using CUTLASS. It would be very much appreciated to be answered.

1. How to know if I am calling CUTLASS code correctly?
   I am using matrix A (M x K), matrix B (N x K), matrix C (M x N) and calling cutlass call like below. I transposed B and provided A, B^T and C as argument, which aligns with the layout in the picture. Each A, B, C are half precision, and accumulator is fp32.

I am also calling CUBLAS kernel, but this time providing A and B, making kernel transpose by itself.

When I compare the result for M = N = K = 4096, I see maximum error of 0.25, which seems to be above 0.05 default value given for CUTLASS profiler. I wonder whether changes in instruction shape, warp size or thread block size affects error value. I am using shapes like this picture.

Should I be getting 0 error to know whether I am calling the same kernel with same layout setting?
What is epilson and non-zero floor in this function used by CUTLASS profiler?

2. What is the correlation between thread block shape, warp shape and instruction shape?
   If I use any of the warp shapes given in documentation of picture below, it gives me error in the second picture.
   Thus, none of the warp shapes I can use to directly use to call GemmUniversal.

Thanks in advance!

Start with the CUTLASS profiler's epsilon and non-zero floor behavior, then trace the GemmUniversal usage and the reported matrix layouts against the CUBLAS comparison. A useful resolution should explain the observed numerical difference, how thread block, warp, and instruction shapes relate, and which documented warp shapes are valid.