Dynamically select unroll factor to build for when targeting local arch (#1371)

* Dynamically select unroll factor to build for when targeting local arch only
Dieser Commit ist enthalten in:
Bertan Dogancay
2024-10-21 10:53:11 -04:00
committet von GitHub
Ursprung 7c077db307
Commit 373f113524
9 geänderte Dateien mit 124 neuen und 125 gelöschten Zeilen
+2 -8
Datei anzeigen
@@ -18,17 +18,11 @@ struct RunWorkNop {
};
__launch_bounds__(NCCL_MAX_NTHREADS, 1) __global__ void ncclDevKernel_Generic(struct ncclDevComm* comm, struct channelMasks channelMask, struct ncclWork* workHead) {
ncclKernelMain<-1, RunWorkNop, false, 2>(comm, channelMask, workHead);
}
__launch_bounds__(NCCL_MAX_NTHREADS, 1) __global__ void ncclDevKernel_Generic_4(struct ncclDevComm* comm, struct channelMasks channelMask, struct ncclWork* workHead) {
ncclKernelMain<-1, RunWorkNop, false, 4>(comm, channelMask, workHead);
ncclKernelMain<-1, RunWorkNop, false>(comm, channelMask, workHead);
}
#ifdef ENABLE_COLLTRACE
__launch_bounds__(NCCL_MAX_NTHREADS, 1) __global__ void ncclDevKernelDebug_Generic(struct ncclDevComm* comm, struct channelMasks channelMask, struct ncclWork* workHead) {
ncclKernelMain<-1, RunWorkNop, true, 2>(comm, channelMask, workHead);
}
__launch_bounds__(NCCL_MAX_NTHREADS, 1) __global__ void ncclDevKernelDebug_Generic_4(struct ncclDevComm* comm, struct channelMasks channelMask, struct ncclWork* workHead) {
ncclKernelMain<-1, RunWorkNop, true, 4>(comm, channelMask, workHead);
ncclKernelMain<-1, RunWorkNop, true>(comm, channelMask, workHead);
}
#endif