Introducing 3 new APIs:
ncclResult_t ncclGather(const void* sendbuff, void* recvbuff, size_t sendcount,
ncclDataType_t datatype, int root, ncclComm_t comm, hipStream_t stream);
ncclResult_t ncclScatter(const void* sendbuff, void* recvbuff,
size_t recvcount, ncclDataType_t datatype, int root, ncclComm_t comm,
hipStream_t stream);
ncclResult_t ncclAllToAll(const void* sendbuff, void* recvbuff, size_t count,
ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream);
Only out of place operation is supported.
Preprocessor symbol RCCL_GATHER_SCATTER=1 indicates API availibility.
By default the APIs launche RCCL kernel implementation, which can be disabled by
RCCL_ALLTOALL_KERNEL_DISABLE=1. Then the APIs use wrapper around ncclSend and ncclRecv.
For gfx908, support simple detection of ring topology.
Call ReduceOrCopyMulti directly from kernel.
Also simplify code by removing kernel start synchronization option
which has no effect on throughput measurements.
1. Fix RCCL unit test
2. Add ROME detection and tuning
3. Change default P2P level
4. Fix search algorithm for XGMI
5. Remove explicit channel duplication with implicit by using half of link speed
6. Add collective trace support
7. Correct Intel Skylake CPU detection and bandwidth
8. Fix topo connect function
9. Disable GDR read and remove unreachable code
10. Disable LL128 kernels
11. Add tuning parameters
12. Use original clock64() implementation which returns RTC counter value
13. Print out timestamp of collective trace
14. Do not use struct ncclColl in kernel launch parameter
15. Fix abort handling and add tracing
17. Add __launch_bounds__ to kernel functions
18. Remove unused abortCount
19. Unset default MIN_NRINGS and MIN_NCHANNELS
20. Do not allocate shared memory when not using LL128 kernels
21. Correct time print out in tuning log
Allow user specified ring topology from command line and duplicated
to requested number of workgroups:
./rccl_prim_test -w 12 -p copy -r "0 1 2 3|3 2 1 0|0 2 1 3|3 1 2 0|0 2 3 1|1 3 2 0"
* Refactor primitive test to support multiple GPUs in rings
* Make GPUs sync before transfer optional
* Use same ring format as RCCL
* Extend to 8 GPUs and report errors if there is no P2P access
* Control GPUs sync before ops from command line with "-s" option
* Change buffer size through command line option "-n"
Rename iterations command line option to "-i"
Wenkai Du
Wenkai Du
Saad Rahim
Gilbert Lee
Aaron Enye Shi
Stanley Tsang
gilbertlee-amd
rohit pathania