Nice! Recently, I took a quick look at the code and the kernel implementation. Does Mori or NVSHMEM provide a method to allocate cross-device shared buffer space that can be exposed for use within the kernel (allocated in an uncached manner)? Within the kernel, can ordinary memory access instructions, such as buffer or flat instructions, be used to read from or write to a cross-device shared buffer? If this is correct, ASM comm+compute fusion kernel can be implemented in a much easier way!

I don't know much about NVSHMEM but MORI provides methods to manage both intra-node / inter-node cross device memory allocation, see this class

mori/include/mori/application/memory/symmetric_memory.hpp

Line 85 in 4284f28

class SymmMemManager {

With method ExtMallocWithFlags and flag= hipDeviceMallocUncached, you can allocate uncached memory.
For intra-node case, yes you can access memory of other devices just like accessing local memory through exchanged pointers. For inter-node you have to use RDMA to access them.
Absolutely! Fusion of communication and computation kernel is one of the most important techniques that MORI wants to enable.

Mori docs improvement #54

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions