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

Two temporary workarounds for cuda-clang issues.

Browse Source
This commit is contained in:
Christian Sigg
2018-12-13 16:09:12 +01:00
committed by Sylvain Jeaugey
parent d08e9b5279
commit 346fc49514
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/collectives/device/functions.cu
+8
View File
@@ -56,9 +56,17 @@
// Must be consistent with the ncclFuncSet enum // Must be consistent with the ncclFuncSet enum
__device__ ncclKern_t ncclFuncs[ncclCollCount*ncclNumOps*ncclNumTypes*2] = { __device__ ncclKern_t ncclFuncs[ncclCollCount*ncclNumOps*ncclNumTypes*2] = {
// Don't try to initialize the host shadow copy of this device-side global
// variable. There is no host pointer to a device-side function, which
// confuses clang. This will be fixed in the next clang release.
#if __CUDA_ARCH__
NCCL_FUNCS2B(ncclBroadcast), NCCL_FUNCS2B(ncclBroadcast),
NCCL_FUNCS2A(ncclReduce), NCCL_FUNCS2A(ncclReduce),
NCCL_FUNCS2B(ncclAllGather), NCCL_FUNCS2B(ncclAllGather),
NCCL_FUNCS2A(ncclReduceScatter), NCCL_FUNCS2A(ncclReduceScatter),
NCCL_FUNCS2A(ncclAllReduce) NCCL_FUNCS2A(ncclAllReduce)
#endif
}; };
// Workaround for https://reviews.llvm.org/D55580
__device__ void ncclWorkaroundClangD55580() {}