[dnn/shift] now strictly only shifting the interior
This commit is contained in:
@@ -75,7 +75,7 @@ torch::Tensor shift_common(
|
||||
|
||||
triton::driver::cu_buffer c(ctx, (CUdeviceptr)torchc.storage().data(), false);
|
||||
// Enqueue
|
||||
shift.enqueue(&stream, {&a, &b, &c}, true);
|
||||
shift.enqueue(&stream, {&a, &b, &c}, false);
|
||||
return torchc;
|
||||
}
|
||||
|
||||
|
@@ -122,7 +122,7 @@ public:
|
||||
triton::driver::cu_buffer da(ctx, (CUdeviceptr)tf_a.flat<Eigen::half>().data(), false);
|
||||
triton::driver::cu_buffer db(ctx, (CUdeviceptr)tf_b.flat<Eigen::half>().data(), false);
|
||||
triton::driver::cu_buffer dc(ctx, (CUdeviceptr)tf_c->flat<Eigen::half>().data(), false);
|
||||
shift.enqueue(stream, {&da, &db, &dc});
|
||||
shift.enqueue(stream, {&da, &db, &dc}, false);
|
||||
}
|
||||
|
||||
private:
|
||||
|
Reference in New Issue
Block a user