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

[ROCm/rccl-tests commit: 5625599dda]
This commit is contained in:
nileshnegi
2025-04-11 12:00:15 -05:00
24 zmienionych plików z 766 dodań i 271 usunięć
+5 -5
Wyświetl plik
@@ -8,12 +8,12 @@
#include "cuda_runtime.h"
#include "common.h"
void AlltoAllGetCollByteCount(size_t *sendcount, size_t *recvcount, size_t *paramcount, size_t *sendInplaceOffset, size_t *recvInplaceOffset, size_t count, int nranks) {
*sendcount = (count/nranks)*nranks;
*recvcount = (count/nranks)*nranks;
void AlltoAllGetCollByteCount(size_t *sendcount, size_t *recvcount, size_t *paramcount, size_t *sendInplaceOffset, size_t *recvInplaceOffset, size_t count, size_t eltSize, int nranks) {
*paramcount = (count/nranks) & -(16/eltSize);
*sendcount = nranks*(*paramcount);
*recvcount = *sendcount;
*sendInplaceOffset = 0;
*recvInplaceOffset = 0;
*paramcount = count/nranks;
}
testResult_t AlltoAllInitData(struct threadArgs* args, ncclDataType_t type, ncclRedOp_t op, int root, int rep, int in_place) {
@@ -61,7 +61,7 @@ struct testColl alltoAllTest = {
void AlltoAllGetBuffSize(size_t *sendcount, size_t *recvcount, size_t count, int nranks) {
size_t paramcount, sendInplaceOffset, recvInplaceOffset;
AlltoAllGetCollByteCount(sendcount, recvcount, &paramcount, &sendInplaceOffset, &recvInplaceOffset, count, nranks);
AlltoAllGetCollByteCount(sendcount, recvcount, &paramcount, &sendInplaceOffset, &recvInplaceOffset, count, /*eltSize=*/1, nranks);
}
testResult_t AlltoAllRunTest(struct threadArgs* args, int root, ncclDataType_t type, const char* typeName, ncclRedOp_t op, const char* opName) {