I guess the parallel actually works on modules.
```
int TVMBackendParallelLaunch(
FTVMParallelLambda flambda,
void* cdata,
int num_task) {
#if !TVM_THREADPOOL_USE_OPENMP
int res = tvm::runtime::ThreadPool::ThreadLocal()->Launch(
flambda, cdata, num_task, 1);
return res;
#else
int num_workers = tvm::runtime::threading::MaxConcurrency();
if (num_task == 0) num_task = num_workers;
omp_set_num_threads(num_workers);
#pragma omp parallel num_threads(num_workers)
{
TVMParallelGroupEnv env;
env.num_task = num_task;
(*flambda)(omp_get_thread_num(), &env, cdata);
}
return 0;
#endif
}
```
This function launches the thread. And it is invoked by module building
process. Graph runtime may have nothing to do with paralelism.
---
[Visit
Topic](https://discuss.tvm.ai/t/execution-order-of-operators-at-runtime-in-tvm/6572/6)
to respond.
You are receiving this because you enabled mailing list mode.
To unsubscribe from these emails, [click
here](https://discuss.tvm.ai/email/unsubscribe/9642d0c1515d9cf8dd6a35c8cff6c90d3fb305017fb0f43434e6a12c62c4101c).
