[QST] Avoiding fp16 errors during GEMM from CUTE example

[SamKG]

What is your question?
In the CUTE GEMM Example, the accumulation buffer is fp16. However, I noticed a rather annoying downstream effect: due to rounding errors, the accumulation will silently ignore small deltas if the exponent is large (e.g. 2048 + 1 = 2048, due to precision issues).

I have two (related) questions:

1. Is there a way to error or warn if the accumulator is silently dropping bits?
2. Is there a way (aside from setting to fp32 or bf16) to avoid this issue?

[hwu36]

if this is an issue to you, it means fp16 is not enough for you. you need to use fp32 accumulation.

in fp16 math, 2048 + 1 = 2048. wikipedia explains it in detail.

[mnicely]

I believe this has been resolved. Feel free to reopen if you disagree.