Merge remote-tracking branch 'nccl/master' into 2.8.3

[ROCm/rccl commit: c985358e11]
Этот коммит содержится в:
Wenkai Du
2021-02-15 19:29:45 +00:00
родитель ea2e665834 fc7bdb38a5
Коммит ab71643c99
23 изменённых файлов: 82 добавлений и 52 удалений
+5 -5
Просмотреть файл
@@ -72,7 +72,7 @@ ncclResult_t dumpProxyState(struct ncclProxyState* state) {
struct ncclProxyArgs* op = state->ops;
while (op) {
if (op->idle & OP_SEEN) {
WARN("Active list loop at element %ld\n", OP_INDEX(op));
WARN("Active list loop at element %ld", OP_INDEX(op));
}
op->idle |= OP_SEEN;
printf("[%ld]", OP_INDEX(op));
@@ -98,7 +98,7 @@ ncclResult_t dumpProxyState(struct ncclProxyState* state) {
struct ncclProxyArgs* free = state->pool;
while (free) {
if (free->idle & OP_SEEN) {
WARN("Free list loop at element %ld\n", OP_INDEX(free));
WARN("Free list loop at element %ld", OP_INDEX(free));
}
free->idle |= OP_SEEN;
free = free->next;
@@ -109,7 +109,7 @@ ncclResult_t dumpProxyState(struct ncclProxyState* state) {
while (p) {
for (int e=0; e<PROXYARGS_ALLOCATE_SIZE; e++) {
if ((p->elems[e].idle & OP_SEEN) == 0) {
WARN("Element %d of pool %d has been lost\n", e, i);
WARN("Element %d of pool %d has been lost", e, i);
struct ncclProxyArgs* free = state->pool;
printf("Free list ");
while (free) {
@@ -164,7 +164,7 @@ static ncclResult_t SaveProxy(int type, int peer, struct ncclProxyArgs* args) {
struct ncclPeer* peerComm = args->channel->peers+peer;
struct ncclConnector* connector = type == proxyRecv ? &peerComm->recv : &peerComm->send;
if (connector->transportComm == NULL) {
WARN("[%d] Error no transport for %s peer %d on channel %d\n", connector->comm->rank,
WARN("[%d] Error no transport for %s peer %d on channel %d", connector->comm->rank,
type == proxyRecv ? "recv" : "send", peer, args->channel->id);
return ncclInternalError;
}
@@ -480,7 +480,7 @@ ncclResult_t ncclProxySharedBuffersFree(struct ncclComm* comm, int cuda, int typ
while (nslots*state->slotSize < size) nslots *= 2;
int s = (ptr-buff)/state->slotSize;
if (s < 0 || s+nslots > state->nslots) {
WARN("Error freeing shared buffer : freeing ptr %p size %d (start %p slot size %d nslots %d)\n", ptr, size, buff, state->slotSize, state->nslots);
WARN("Error freeing shared buffer : freeing ptr %p size %d (start %p slot size %d nslots %d)", ptr, size, buff, state->slotSize, state->nslots);
return ncclInternalError;
}
for (int i=0; i<nslots; i++) used[s+i] = 0;