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

Merge remote-tracking branch 'nccl/master' into no-target-id

Browse Source
This commit is contained in:
Wenkai Du
2020-12-01 11:33:47 -05:00
parent 2e8b3a0857 920dbe5b35
commit d469947641
106 changed files with 11943 additions and 4104 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/group.cc
+115 -79
View File
@@ -35,7 +35,6 @@ struct ncclInitArgs {
};
struct ncclCollArgs {
ncclComm_t comm;
int connect;
};
enum ncclAsyncFuncType {
@@ -110,6 +109,7 @@ ncclResult_t ncclAsyncColl(ncclComm_t comm) {
NCCL_API(ncclResult_t, ncclGroupStart);
ncclResult_t ncclGroupStart() {
NVTX3_FUNC_RANGE_IN(nccl_domain);
if (ncclGroupMode == 0) {
memset(ncclGroupArgs, 0, sizeof(struct ncclAsyncArgs)*MAX_ASYNC_OPS);
}
@@ -118,7 +118,7 @@ ncclResult_t ncclGroupStart() {
}
static ncclResult_t scheduleSendRecv(struct ncclComm* comm, int delta, int channelId, ssize_t recvbytes, void* recvbuff, ssize_t sendbytes, const void* sendbuff) {
struct ncclInfo info = { ncclCollSendRecv, "SendRecv",
struct ncclInfo info = { ncclFuncSendRecv, "SendRecv",
sendbuff, recvbuff, (size_t)std::max<ssize_t>(sendbytes,recvbytes), ncclInt8, ncclSum, -1, comm, comm->userStream, /* Args */
1, 1 };
info.delta = delta;
@@ -126,26 +126,32 @@ static ncclResult_t scheduleSendRecv(struct ncclComm* comm, int delta, int chann
info.sendbytes = sendbytes;
info.recvbytes = recvbytes;
if (delta == 0 && sendbytes != recvbytes) return ncclInvalidUsage;
NCCLCHECK(ncclSaveKernel(&info));
NCCLCHECK(ncclSaveP2pKernel(&info));
return ncclSuccess;
}
void* ncclAsyncThreadPreconnect(void* args_) {
struct ncclAsyncArgs* args = (struct ncclAsyncArgs*)args_;
CUDACHECKTHREAD(hipSetDevice(args->coll.comm->cudaDev));
for (int c=0; c<args->coll.comm->p2pnChannels; c++) {
struct ncclComm* comm = args->coll.comm;
struct ncclChannel* channel = comm->channels+c;
struct ncclP2PConnect* connect = &comm->p2plist.connect;
NCCLCHECKTHREAD(ncclTransportP2pSetup(comm, NULL, channel, connect->nrecv[c], connect->recv+c*comm->nRanks, connect->nsend[c], connect->send+c*comm->nRanks));
connect->nrecv[c] = 0;
connect->nsend[c] = 0;
}
struct ncclComm* comm = args->coll.comm;
CUDACHECKTHREAD(hipSetDevice(comm->cudaDev));
NCCLCHECKTHREAD(ncclTransportP2pSetup(comm, NULL));
return args;
}
static size_t getP2pChunkSize(size_t totalSize, int minChannels, int maxChannels, size_t minSize, size_t maxSize) {
size_t size = std::max(minSize, DIVUP(totalSize, minChannels));
int nChannels = minChannels;
while (size > maxSize && nChannels <= maxChannels/2) {
nChannels *= 2;
size = DIVUP(totalSize, nChannels);
}
ALIGN_SIZE(size, minSize);
return size;
}
NCCL_API(ncclResult_t, ncclGroupEnd);
ncclResult_t ncclGroupEnd() {
NVTX3_FUNC_RANGE_IN(nccl_domain);
if (ncclGroupMode == 0) {
WARN("ncclGroupEnd: not in a group call.");
return ncclInvalidUsage;
@@ -186,29 +192,21 @@ ncclResult_t ncclGroupEnd() {
for (int i=0; i<ncclGroupIndex; i++) {
struct ncclAsyncArgs* args = ncclGroupArgs+i;
if (args->funcType == ASYNC_FUNC_COLL) {
struct ncclP2Plist* p2plist = &args->coll.comm->p2plist;
if (p2plist->count != 0) {
struct ncclComm* comm = args->coll.comm;
args->coll.connect = 0;
for (int c=0; c<comm->p2pnChannels; c++)
args->coll.connect += comm->p2plist.connect.nsend[c] + comm->p2plist.connect.nrecv[c];
if (args->coll.connect) {
pthread_create(ncclGroupThreads+i, NULL, ncclAsyncThreadPreconnect, args);
}
}
if (args->funcType == ASYNC_FUNC_COLL && args->coll.comm->connect) {
pthread_create(ncclGroupThreads+i, NULL, ncclAsyncThreadPreconnect, args);
}
}
for (int i=0; i<ncclGroupIndex; i++) {
struct ncclAsyncArgs* args = ncclGroupArgs+i;
if (args->funcType == ASYNC_FUNC_COLL && (args->coll.connect)) {
if (args->funcType == ASYNC_FUNC_COLL && args->coll.comm->connect) {
int err = pthread_join(ncclGroupThreads[i], NULL);
if (err != 0) {
WARN("Error waiting for pthread_join : %s\n", strerror(errno));
return ncclSystemError;
}
NCCLCHECKGOTO(args->ret, ret, end);
args->coll.comm->connect = 0;
}
}
@@ -218,56 +216,98 @@ ncclResult_t ncclGroupEnd() {
struct ncclComm* comm = args->coll.comm;
int rank = comm->rank;
int nRanks = comm->nRanks;
struct ncclP2Plist* p2plist = &args->coll.comm->p2plist;
if (p2plist->count) {
for (int delta=0; delta<nRanks; delta++) {
struct ncclP2Plist* p2pSends = comm->p2pSends;
struct ncclP2Plist* p2pRecvs = comm->p2pRecvs;
// Compute how much to split operations
// Natural step size matching buffer steps.
ssize_t stepSize = comm->buffSizes[NCCL_PROTO_SIMPLE] / NCCL_STEPS;
// Try to use all channels
int nChannelsMax = comm->p2pnChannelsPerPeer;
int nChannelsMin = nChannelsMax;
// Try to use all channels, but one channel per operation.
while (nChannelsMin*comm->nRanks > comm->p2pnChannels && nChannelsMin > 1) nChannelsMin /= 2;
// Avoid overloading channels with 8+ operations as we loose the sync warp, hence a bit of bandwidth.
while (nChannelsMax*comm->nRanks > comm->p2pnChannels*4 && nChannelsMax > 1) nChannelsMax /= 2;
while (comm->p2pSendCount > 0 || comm->p2pRecvCount > 0) {
// schedule delta 0, +1, -1, +2, -2, ...
// also make sure we don't do 0 twice, nor +n/2 and -n/2 if n is even.
for (int d=0; d<=nRanks/4; d++) {
int deltas[4] = { d, (nRanks-d)%nRanks, nRanks/2-d, nRanks-(nRanks/2-d) };
int index = 0;
int delta = deltas[index];
sched_delta:
uint32_t from = (rank+nRanks-delta)%nRanks;
uint32_t to = (rank+delta)%nRanks;
struct ncclP2Pinfo* recv = p2pRecvs[from].head;
struct ncclP2Pinfo* send = p2pSends[to].head;
if (recv != NULL || send != NULL) {
ssize_t totRecvBytes = -1, totSendBytes = -1;
if (recv != NULL) totRecvBytes = recv->nbytes;
if (send != NULL) totSendBytes = send->nbytes;
ssize_t recvChunkSize = getP2pChunkSize(totRecvBytes, nChannelsMin, nChannelsMax, stepSize, SENDRECV_SLICEFACTOR*stepSize);
ssize_t sendChunkSize = getP2pChunkSize(totSendBytes, nChannelsMin, nChannelsMax, stepSize, SENDRECV_SLICEFACTOR*stepSize);
// Compute how much to split operations
// Natural step size matching buffer steps.
ssize_t stepSize = 4*comm->buffSizes[NCCL_PROTO_SIMPLE] / NCCL_STEPS;
// Split each operation on p2pnChannelsPerPeer max.
ssize_t recvChunkSize = DIVUP(p2plist->peerlist[from].recvbytes, comm->p2pnChannelsPerPeer);
ssize_t sendChunkSize = DIVUP(p2plist->peerlist[to].sendbytes, comm->p2pnChannelsPerPeer);
recvChunkSize = std::max((ssize_t)1, DIVUP(recvChunkSize, stepSize)) * stepSize;
sendChunkSize = std::max((ssize_t)1, DIVUP(sendChunkSize, stepSize)) * stepSize;
ssize_t sendOffset = 0;
ssize_t recvOffset = 0;
int remaining = 1;
int chunk = 0;
while (remaining) {
int channelId = (delta+comm->p2pChannels[chunk%comm->p2pnChannelsPerPeer]) % comm->p2pnChannels;
remaining = 0;
ssize_t recvbytes = p2plist->peerlist[from].recvbytes-recvOffset;
ssize_t sendbytes = p2plist->peerlist[to].sendbytes-sendOffset;
if (recvbytes > recvChunkSize) { remaining = 1; recvbytes = recvChunkSize; } else p2plist->peerlist[from].recvbytes = -1;
if (sendbytes > sendChunkSize) { remaining = 1; sendbytes = sendChunkSize; } else p2plist->peerlist[to].sendbytes = -1;
if (sendbytes >= 0 || recvbytes >= 0) {
NCCLCHECKGOTO(scheduleSendRecv(comm, delta, channelId,
recvbytes, ((char*)(p2plist->peerlist[from].recvbuff)) + recvOffset,
sendbytes, ((const char*)(p2plist->peerlist[to].sendbuff)) + sendOffset), ret, end);
ssize_t sendOffset = 0;
ssize_t recvOffset = 0;
int sendRemaining = 1, recvRemaining = 1;
int chunk = 0;
do {
int channelId = (delta+comm->p2pChannels[chunk%comm->p2pnChannelsPerPeer]) % comm->p2pnChannels;
ssize_t recvbytes = totRecvBytes-recvOffset;
ssize_t sendbytes = totSendBytes-sendOffset;
if (recvbytes > recvChunkSize) { recvbytes = recvChunkSize; } else { recvRemaining = 0; }
if (sendbytes > sendChunkSize) { sendbytes = sendChunkSize; } else { sendRemaining = 0; }
if (sendbytes >= 0 || recvbytes >= 0) {
NCCLCHECKGOTO(scheduleSendRecv(comm, delta, channelId,
recvbytes, recv ? ((char*)(recv->buff)) + recvOffset : NULL,
sendbytes, send ? ((const char*)(send->buff)) + sendOffset : NULL), ret, group_cleanup);
}
recvOffset += recvChunkSize;
sendOffset += sendChunkSize;
chunk++;
} while (sendRemaining || recvRemaining);
if (recv) {
NCCLCHECKGOTO(dequeueP2pInfo(p2pRecvs+from), ret, group_cleanup);
comm->p2pRecvCount--;
}
recvOffset += recvChunkSize;
sendOffset += sendChunkSize;
chunk++;
if (send) {
NCCLCHECKGOTO(dequeueP2pInfo(p2pSends+to), ret, group_cleanup);
comm->p2pSendCount--;
}
}
index++;
if (index == 1 && deltas[1] == deltas[0]) index++;
if (index == 2 && deltas[2] == deltas[0]) index++;
if (index == 3 && deltas[3] == deltas[2]) index++;
if (index == 3 && deltas[3] == deltas[1]) index++;
if (index < 4) {
delta = deltas[index];
goto sched_delta;
}
}
p2plist->count = 0;
}
}
}
/* Collectives are done in three steps :
* 0. Save kernels previously enqueued. Compute channel, algo, proto, etc.
* 1. Barrier Check In. Only the last call may call cudaLaunchKernel[cooperative]
* 2. Barrier Wait. No CUDA call is permitted
* 3. Enqueue Events. CUDA event wait/enqueue.
* This is needed because step 2 cannot call any CUDA primitive, otherwise if
* hipFree happens between 1 and 3, it could block that CUDA call and
* cudaFree happens between 1 and 3, it could block that CUDA call and
* prevent some ranks from launching their network threads, which would
* prevent the NCCL call from completing, blocking the hipFree call.
* prevent the NCCL call from completing, blocking the cudaFree call.
*/
for (int i=0; i<ncclGroupIndex; i++) {
struct ncclAsyncArgs* args = ncclGroupArgs+i;
if (args->funcType == ASYNC_FUNC_COLL) {
ncclComm_t comm = args->coll.comm;
NCCLCHECKGOTO(ncclSaveCommKernels(comm), ret, group_cleanup);
}
}
for (int i=0; i<ncclGroupIndex; i++) {
struct ncclAsyncArgs* args = ncclGroupArgs+i;
if (args->funcType == ASYNC_FUNC_COLL) {
@@ -304,32 +344,28 @@ group_cleanup:
*args->init.newcomm = NULL;
} else {
struct ncclComm* comm = args->coll.comm;
for (int c=0; c<std::max(comm->nChannels, comm->p2pnChannels); c++) {
struct ncclChannel* channel = comm->channels+c;
for (int i=0; i<channel->collCount; i++) {
channel->collectives[(channel->collStart + i)%NCCL_MAX_OPS].active = 0;
// Reset aggregation counters
comm->asyncOpCount = 0;
comm->asyncTotalSize = 0;
// Dequeue p2p lists
if (comm->p2pSendCount > 0 || comm->p2pRecvCount > 0) {
struct ncclP2Plist* p2pSends = comm->p2pSends;
struct ncclP2Plist* p2pRecvs = comm->p2pRecvs;
for (int peer=0; peer<comm->nRanks; peer++) {
while (p2pSends[peer].head != NULL) dequeueP2pInfo(p2pSends+peer);
while (p2pRecvs[peer].head != NULL) dequeueP2pInfo(p2pRecvs+peer);
}
channel->collFifoTail = channel->collStart;
channel->collCount = 0;
comm->p2pSendCount = comm->p2pRecvCount = 0;
}
/* Cancel all proxy ops : mark them as ncclProxyOpNone and they should be freed later on */
/* Free all proxy ops in state->nextOps */
struct ncclProxyState* state = &comm->proxyState;
struct ncclProxyArgs *op, *start;
pthread_mutex_lock(&state->mutex);
op = start = state->ops;
while (op) {
if (op->opCount >= comm->lastOpCount) op->state = ncclProxyOpNone;
struct ncclProxyArgs* peerOp = op->nextPeer;
while (peerOp) {
if (peerOp->opCount >= comm->lastOpCount) peerOp->state = ncclProxyOpNone;
peerOp = peerOp->nextPeer;
}
op = op->next;
if (op == start) break;
pthread_mutex_lock(&state->poolMutex);
for (struct ncclProxyArgs *op = state->nextOps; op; op = op->next) {
op->next = state->pool;
state->pool = op;
}
comm->opCount = comm->lastOpCount;
pthread_cond_signal(&state->cond);
pthread_mutex_unlock(&state->mutex);
pthread_mutex_unlock(&state->poolMutex);
state->nextOps = NULL;
comm->myParams->gridDim.x = comm->myParams->blockDim.x = 0;
comm->userStreamSet = false;