rccl
rccl copied to clipboard
ROCmSoftwarePlatform
Add API for non-caching-load
RCCL provides "low latency" protocols for communication between agents, where the entire message consisting of data and flags is packed into a single L2 cache line. This is usually accomplished using atomic relaxed instructions in LLVM. But the 128-byte version of this protocol (LL-128) requires 128-bit load or store instructions that bypass the cache and are not broken up into multiple instructions. The nontemporal builtin is not always suitable for this use case.
The proposed approach is to provide a C++ function template that encapsulates an inline assembly call. This asm is intended to use the appropriate load/store parameters for each combination of data size and architecture.
Have you actually verified that the byte two-byte load instructions you are using exist on the ISAs you expect them to exist? If you have, perhaps you want to check again, carefully, if they exist on GFX9 and GFX10? Has this been tested at all? Should it not have some unit tests in tow?
Have you actually verified that the byte two-byte load instructions you are using exist on the ISAs you expect them to exist? If you have, perhaps you want to check again, carefully, if they exist on GFX9 and GFX10? Has this been tested at all? Should it not have some unit tests in tow?
Updated the patch with byte, 2-byte load and added test cases.
ping.
Can you try to integrate this into https://github.com/ROCm/rccl/blob/develop/tools/p2p-latency-test/ll_latency_test.cpp?
