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

Fix MSCCL work FIFO out-of-bound issue (#935)

Browse Source
This commit is contained in:
Ziyue Yang authored and GitHub committed 2023-10-27 11:24:52 -07:00
1 parent 96ec3ffe2e
commit 4c117e5335
3 files changed
+13 -13
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/include/msccl/msccl_struct.h
+2 -2
View File
@@ -15,7 +15,7 @@
#define MSCCL_MAX_NUM_STEPS 64
#define MSCCL_MAX_NUM_THREAD_BLOCKS_PER_CHANNEL 32
#define MSCCL_MAX_NUM_THREAD_BLOCKS (MSCCL_MAX_NUM_THREAD_BLOCKS_PER_CHANNEL * MAXCHANNELS)
#define MSCCL_MAX_NUM_THREAD_BLOCKS 64
#define MSCCL_MAX_COUNT 72 // max concurrent number of msccl chunk transmission
#define MSCCL_MAX_REDUCE_FUSION 16
#define MSCCL_MAX_NUM_ALGOS 1024
@@ -216,7 +216,7 @@ struct mscclStatus {
struct mscclWork* workFifo;
uint32_t* workFifoDone;
uint32_t workFifoSent;
uint32_t workFifoSentPerChannel[MAXCHANNELS];
uint32_t workFifoSentPerThreadBlock[MSCCL_MAX_NUM_THREAD_BLOCKS];
uint32_t workFifoAckdMin;
};