AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update mscclpp (#1488)

Browse Source 
* update commit hash for mscclpp submodule

* update mscclpp submodule

* remove print messages in cmake

* add back some print messages, update MSCLPP CMAKE_ARGS

* enable MSCCL++ patches regardless of finding mscclpp_nccl package
This commit is contained in:
isaki001
2025-01-20 08:06:43 -06:00
committed by GitHub
parent c68b558ed5
commit d89432e8c8
6 changed files with 81 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ext-src/mem-reg.patch
+16 -14
View File
@@ -1,10 +1,10 @@
diff --git a/apps/nccl/include/nccl.h b/apps/nccl/include/nccl.h
index 7f50792..b8b146d 100644
index bfdb226..7fd07a8 100644
--- a/apps/nccl/include/nccl.h
+++ b/apps/nccl/include/nccl.h
@@ -344,6 +344,13 @@ ncclResult_t ncclAllGather(const void* sendbuff, void* recvbuff, size_t sendcoun
ncclResult_t pncclAllGather(const void* sendbuff, void* recvbuff, size_t sendcount, ncclDataType_t datatype,
ncclComm_t comm, cudaStream_t stream);
@@ -167,6 +167,14 @@ ncclResult_t pncclCommCuDevice(const ncclComm_t comm, int* device);
ncclResult_t ncclCommUserRank(const ncclComm_t comm, int* rank);
ncclResult_t pncclCommUserRank(const ncclComm_t comm, int* rank);
+/*
+ * Register/Deregister
@@ -13,14 +13,15 @@ index 7f50792..b8b146d 100644
+ncclResult_t ncclCommDeregister(ncclComm_t comm, void* handle);
+bool mscclpp_BuffIsRegistered(ncclComm_t comm, const void* buff, size_t count);
+size_t mscclpp_BufferSize(ncclComm_t comm, void* handle);
/*
* Send
*
+
/* Reduction operation selector */
typedef enum { ncclNumOps_dummy = 5 } ncclRedOp_dummy_t;
typedef enum {
diff --git a/apps/nccl/src/nccl.cu b/apps/nccl/src/nccl.cu
index a697be2..1d4af61 100644
index 022d398..2a39643 100644
--- a/apps/nccl/src/nccl.cu
+++ b/apps/nccl/src/nccl.cu
@@ -65,6 +65,7 @@ struct ncclComm {
@@ -85,6 +85,7 @@ struct ncclComm {
std::unordered_map<channelKey, ChannelInfo> channelInInfos;
std::unordered_map<channelKey, ChannelInfo> channelOutInfos;
std::unordered_map<channelKey, ChannelInfo> channelScratchInfos;
@@ -28,7 +29,7 @@ index a697be2..1d4af61 100644
std::shared_ptr<char> scratchBuff;
std::vector<mscclpp::RegisteredMemory> remoteScratchRegMemories;
@@ -73,6 +74,11 @@ struct ncclComm {
@@ -92,6 +93,11 @@ struct ncclComm {
uint32_t buffFlag;
};
@@ -40,8 +41,8 @@ index a697be2..1d4af61 100644
static size_t ncclTypeSize(ncclDataType_t type) {
switch (type) {
case ncclInt8:
@@ -577,6 +583,104 @@ NCCL_API ncclResult_t ncclAllGather(const void* sendbuff, void* recvbuff, size_t
return ncclSuccess;
@@ -561,6 +567,105 @@ NCCL_API ncclResult_t ncclRedOpDestroy(ncclRedOp_t, ncclComm_t) {
return ncclInternalError;
}
+NCCL_API ncclResult_t ncclCommRegister(ncclComm_t comm, void* buff, size_t size, void** handle) {
@@ -142,6 +143,7 @@ index a697be2..1d4af61 100644
+ auto buffKeyIt = comm->handleKeys.find(handle);
+ return buffKeyIt != comm->handleKeys.end() ? buffKeyIt->second.bytes : 0;
+}
NCCL_API ncclResult_t ncclSend(const void*, size_t, ncclDataType_t, int, ncclComm_t, cudaStream_t) {
+
NCCL_API ncclResult_t ncclReduce(const void*, void*, size_t, ncclDataType_t, ncclRedOp_t, int, ncclComm_t,
cudaStream_t) {
// TODO: implement this function
return ncclInternalError;