Documentation of CUDA MSM limitations about sppark HOT 2 CLOSED

Greetings!

I met the same problem as you mentioned. Our team tried it for several times to locate the specific problem.

Here's what we found: for function "invoke" at msm/pippenger.cuh, there is slightly difference between variable "N" and "this->N". When they are not equal (our tests shows that this->N equals 5, and N could be 1,2,3,4,5 when points.len()<=512), the result of GPU will differ from that of CPU.

To solve the problem, you may adjust function "collect" at msm/pippenger.cuh to something like this:

And also, you should pay attention to function "invoke" who calls "collect" at msm/pippenger.cuh.

Hope this could help you.

from sppark.

There was a major MSM makeover and it's no longer a problem. Well, at least for example env TEST_NPOW=7 cargo test --release --features=bls12_381 works if you pull from github.

from sppark.