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

Merge pull request #946 from nusislam/msccl-redop

Browse Source 
msccl: remove dereference of reduce args

[ROCm/rccl commit: 61aed56ca7]
This commit is contained in:
Nusrat Islam
2023-11-02 16:22:24 -05:00
committed by GitHub
parent 3eeaea3f00 88c8bb1495
commit 7c24d9970e
1 changed files with 0 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/collectives/device/msccl_kernel_impl.h
-19
View File
@@ -210,25 +210,6 @@ __device__ __forceinline__ void mscclRunInterpreter(
}
#endif
// Deference reduce args if required
if (tid == 0 && mscclShmem.work.hasReduce && mscclShmem.work.redOpArgIsPtr) {
switch (sizeof(T)) {
case 1:
mscclShmem.work.redOpArg = *reinterpret_cast<uint8_t*>(mscclShmem.work.redOpArg);
break;
case 2:
mscclShmem.work.redOpArg = *reinterpret_cast<uint16_t*>(mscclShmem.work.redOpArg);
break;
case 4:
mscclShmem.work.redOpArg = *reinterpret_cast<uint32_t*>(mscclShmem.work.redOpArg);
break;
case 8:
mscclShmem.work.redOpArg = *reinterpret_cast<uint64_t*>(mscclShmem.work.redOpArg);
break;
default:
break;
}
}
__synclds(); // publish shmem
// User pointers for primitives