Add GpuBuffer class

.azure-pipelines/ut.yml

@@ -35,7 +35,7 @@ jobs:
         cmake -DCMAKE_BUILD_TYPE=Release -DMSCCLPP_BYPASS_GPU_CHECK=ON -DMSCCLPP_USE_CUDA=ON ..
         make -j
       workingDirectory: '$(System.DefaultWorkingDirectory)'
-  
+
   - task: DownloadSecureFile@1
     name: SshKeyFile
     displayName: Download key file

apps/nccl/src/broadcast.hpp

@@ -15,7 +15,7 @@
 template <bool IsOutOfPlace>
 __global__ void __launch_bounds__(1024, 1)
     broadcast6(void* sendbuff, void* scratchbuff, void* recvbuff, mscclpp::DeviceHandle<mscclpp::SmChannel>* smChannels,
-               size_t channelOutOffset, size_t rank, [[maybe_unused]] size_t worldSize, size_t root,
+               [[maybe_unused]] size_t channelOutOffset, size_t rank, [[maybe_unused]] size_t worldSize, size_t root,
                size_t nRanksPerNode, size_t nelemsPerGPU) {
   const size_t nThread = blockDim.x * gridDim.x;
   const size_t nPeer = nRanksPerNode - 1;

apps/nccl/src/nccl.cu

@@ -176,9 +176,9 @@ static std::shared_ptr<mscclpp::DeviceHandle<mscclpp::SmChannel>> setupSmChannel
   std::transform(smChannels.begin(), smChannels.end(), std::back_inserter(smChannelDeviceHandles),
                  [](const mscclpp::SmChannel& smChannel) { return mscclpp::deviceHandle(smChannel); });
   std::shared_ptr<mscclpp::DeviceHandle<mscclpp::SmChannel>> ptr =
-      mscclpp::allocSharedCuda<mscclpp::DeviceHandle<mscclpp::SmChannel>>(smChannelDeviceHandles.size());
-  mscclpp::memcpyCuda<mscclpp::DeviceHandle<mscclpp::SmChannel>>(ptr.get(), smChannelDeviceHandles.data(),
-                                                                 smChannelDeviceHandles.size(), cudaMemcpyHostToDevice);
+      mscclpp::detail::gpuCallocShared<mscclpp::DeviceHandle<mscclpp::SmChannel>>(smChannelDeviceHandles.size());
+  mscclpp::gpuMemcpy<mscclpp::DeviceHandle<mscclpp::SmChannel>>(ptr.get(), smChannelDeviceHandles.data(),
+                                                                smChannelDeviceHandles.size(), cudaMemcpyHostToDevice);
   return ptr;
 }
 
@@ -360,7 +360,7 @@ static void ncclCommInitRankFallbackSingleNode(ncclComm* commPtr, std::shared_pt
   commPtr->smSemaphores = std::move(smSemaphores);
   commPtr->buffFlag = 0;
   commPtr->numScratchBuff = 2;
-  commPtr->scratchBuff = mscclpp::allocExtSharedCuda<char>(SCRATCH_SIZE);
+  commPtr->scratchBuff = mscclpp::GpuBuffer(SCRATCH_SIZE).memory();
   commPtr->remoteScratchRegMemories =
       setupRemoteMemories(commPtr->comm, rank, commPtr->scratchBuff.get(), SCRATCH_SIZE, mscclpp::Transport::CudaIpc);
 }
@@ -624,7 +624,6 @@ NCCL_API ncclResult_t ncclBroadcast(const void* sendbuff, void* recvbuff, size_t
   }
 
   int rank = comm->comm->bootstrap()->getRank();
-  int nRank = comm->comm->bootstrap()->getNranks();
 
   std::vector<executionPlanInstance>& plans = comm->executionPlans["broadcast"];
   std::shared_ptr<mscclpp::ExecutionPlan> plan;
@@ -817,18 +816,13 @@ NCCL_API ncclResult_t ncclCommDeregister(const ncclComm_t, void*) {
 }
 
 ncclResult_t ncclMemAlloc(void** ptr, size_t size) {
-  // Allocate memory using mscclpp::allocSharedPhysicalCuda
   if (ptr == nullptr || size == 0) {
     WARN("ptr is nullptr or size is 0");
     return ncclInvalidArgument;
   }
   std::shared_ptr<char> sharedPtr;
   try {
-    if (mscclpp::isNvlsSupported()) {
-      sharedPtr = mscclpp::allocSharedPhysicalCuda<char>(size);
-    } else {
-      sharedPtr = mscclpp::allocExtSharedCuda<char>(size);
-    }
+    sharedPtr = mscclpp::GpuBuffer(size).memory();
     if (sharedPtr == nullptr) {
       INFO(MSCCLPP_ALLOC, "Failed to allocate memory");
       return ncclSystemError;

docs/getting-started/tutorials/python-api.md

@@ -14,6 +14,7 @@ from mscclpp import (
     ProxyService,
     Transport,
 )
+from mscclpp.utils import GpuBuffer
 import mscclpp.comm as mscclpp_comm
 
 def create_connection(group: mscclpp_comm.CommGroup, transport: str):
@@ -32,7 +33,7 @@ if __name__ == "__main__":
     mscclpp_group = mscclpp_comm.CommGroup(MPI.COMM_WORLD)
     connections = create_connection(mscclpp_group, "NVLink")
     nelems = 1024
-    memory = cp.zeros(nelem, dtype=cp.int32)
+    memory = GpuBuffer(nelem, dtype=cp.int32)
     proxy_service = ProxyService()
     simple_channels = group.make_proxy_channels(proxy_service, memory, connections)
     proxy_service.start_proxy()