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

Merge remote-tracking branch 'nccl/master' into develop

Browse Source
This commit is contained in:
Wenkai Du
2021-04-30 16:57:36 -07:00
parent 4f8e788a61 a46ea10583
commit a4ea1fed5b
80 changed files with 3136 additions and 1675 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/include/proxy.h
+56 -30
View File
@@ -1,5 +1,6 @@
/*************************************************************************
* Copyright (c) 2016-2020, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2016-2021, NVIDIA CORPORATION. All rights reserved.
* Modifications Copyright (c) 2019-2021 Advanced Micro Devices, Inc. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
@@ -14,48 +15,67 @@ enum ncclProxyOpState { ncclProxyOpNone, ncclProxyOpReady, ncclProxyOpProgress }
struct ncclProxyArgs;
typedef ncclResult_t (*proxyProgressFunc_t)(struct ncclProxyArgs*);
struct ncclProxyArgs {
proxyProgressFunc_t progress;
#define NCCL_PROXY_MAX_SUBS MAXCHANNELS
static_assert(NCCL_MAX_WORK_ELEMENTS <= MAXCHANNELS, "Not enough sub space for max work elements");
struct ncclProxySubArgs {
struct ncclChannel* channel;
struct ncclConnector* connector;
size_t sendbytes;
size_t recvbytes;
int sliceSteps;
int chunkSteps;
int nsteps;
uint64_t opCount;
int protocol;
int segment; // Only for profiling
ncclDataType_t dtype;
ncclRedOp_t redOp;
int state; // add component before this line -- it is left out during initialization
ssize_t sendbytes;
ssize_t recvbytes;
int sendChunkSize;
int recvChunkSize;
int delta;
// Internal state
uint64_t base;
uint64_t posted;
uint64_t received; // Only used by recv proxy to wait for flush.
uint64_t received;
uint64_t flushed;
uint64_t transmitted;
uint64_t done;
uint64_t end;
uint64_t hdp_flushed;
void* requests[NCCL_STEPS];
};
struct ncclProxyArgs {
proxyProgressFunc_t progress;
struct ncclProxySubArgs subs[NCCL_PROXY_MAX_SUBS];
int nsubs;
int done;
int sliceSteps;
int chunkSteps;
int chunkSize;
uint64_t opCount;
uint64_t commOpCount;
int protocol;
ncclDataType_t dtype;
ncclRedOp_t redOp;
ncclPattern_t pattern;
int root;
int state;
char* sharedBuff[NCCL_STEPS];
int sharedSize[NCCL_STEPS];
int idle;
uint64_t hdp_flushed;
// Element linking
pthread_mutex_t mutex;
struct ncclProxyArgs* next;
struct ncclProxyArgs* nextPeer;
struct ncclProxyArgs* nextGroup;
struct ncclProxyArgs** proxyAppendPtr;
};
struct ncclProxySharedBuffers {
int nslots;
int slotSize;
char* cudaBuff[2*MAXCHANNELS];
int* cudaUsed[2*MAXCHANNELS];
char* hostBuff[2*MAXCHANNELS];
int* hostUsed[2*MAXCHANNELS];
int size;
char* cudaBuff;
char* hostBuff;
struct ncclProxyArgs* proxyAppend[2*MAXCHANNELS]; // Separate send and recv
// Collnet sharing is technically per device, but for now MAXDEVICES == MAXCHANNELS.
struct ncclProxyArgs* proxyAppendCollNet[2*MAXCHANNELS];
void* collNetResources;
};
struct ncclProxyPool;
@@ -64,11 +84,16 @@ struct ncclProxyState {
pthread_mutex_t opsMutex;
pthread_mutex_t poolMutex;
bool stop;
struct ncclProxySharedBuffers* sharedBuffs;
struct ncclProxyArgs* ops;
struct ncclProxyArgs* nextOps;
struct ncclProxySharedBuffers sharedBuffs;
struct ncclProxyArgs* ops; // Running operations, used by proxy thread
struct ncclProxyArgs* postedOps; // Posted operations, shared between proxy and main thread, locked with opsMutex
struct ncclProxyArgs* postedOpsEnd;
struct ncclProxyArgs* nextOps; // Pending operations, used by main thread (could still be cancelled)
struct ncclProxyArgs* nextOpsEnd;
struct ncclProxyArgs* pool;
struct ncclProxyArgs* pool; // Free operations for main thread
struct ncclProxyArgs* poolFreed; // Freed operations by the progress thread
struct ncclProxyArgs* poolReturned; // Shared between main and progress thread, lock with poolMutex
struct ncclProxyPool* pools;
};
@@ -80,15 +105,16 @@ enum proxyMode {
proxyTo = 2
};
ncclResult_t ncclProxySaveColl(struct ncclProxyArgs* args, int pattern, int root, int nranks);
ncclResult_t ncclProxySaveP2p(struct ncclInfo* info, struct ncclChannel* channel, int segment);
ncclResult_t ncclProxySaveColl(struct ncclProxyArgs* args, int nranks);
ncclResult_t ncclProxyComputeP2p(struct ncclInfo* info, struct ncclProxyArgs* args);
ncclResult_t ncclProxySaveP2p(struct ncclComm* comm, struct ncclProxyArgs* args);
ncclResult_t ncclProxyStart(struct ncclComm* comm);
ncclResult_t ncclProxyCreate(struct ncclComm* comm);
ncclResult_t ncclProxyDestroy(struct ncclComm* comm);
ncclResult_t ncclProxySharedBuffersInit(struct ncclComm* comm, int cuda, int* size, char** ptr);
ncclResult_t ncclProxySharedBuffersAlloc(struct ncclComm* comm, int cuda, int type, int channel, int size, char** ptr);
ncclResult_t ncclProxySharedBuffersFree(struct ncclComm* comm, int cuda, int type, int channel, int size, char* ptr);
ncclResult_t ncclProxySharedBuffersGetP2p(struct ncclComm* comm, int cuda, int type, int channel, int slot, int index, char** ptr);
ncclResult_t ncclProxySharedBuffersGetCollNet(struct ncclComm* comm, int cuda, int type, int slot, int channel, char** ptr);
ncclResult_t ncclProxySharedBuffersDestroy(struct ncclComm* comm);
#include <unistd.h>