Potentially incorrect `hypre_assert` about hypre HOT 3 OPEN

@liruipeng, should we update the assertion to test for the actual memory location instead?

from hypre.

Hi @v-dobrev Thank you for reporting this issue. This is a code design philosophy we have had from the beginning of the GPU work. When hypre is not configured with GPU, we treat HOST to be the "DEVICE", so there is no #ifdef of setting the default memory space to HYPRE_MEMORY_DEVICE, regardless of GPU or not. And the assert (among many others in the same form I guess) basically just say A and G lives in the same "device" space. Apparently, MFEM treat this differently. Does it work in the GPU case?

from hypre.

When hypre is built for GPU, there's no issue -- MFEM sets the memory location to HYPRE_MEMORY_DEVICE.

I guees the question is: do you consider it an error if a user provides to hypre_AMESetup (or other similar functions) a matrix that has memory location HYPRE_MEMORY_HOST when hypre is built without GPU support? If you consider it a user error, we can close this issue, and we can make adjustments in MFEM to address the problem on our side. If you don't consider it a user error, then you probably will want to change the hypre_assert in some way to allow the case I ran into.

from hypre.