Fix result error when calling AllReduce serially.

This commit is contained in:
ZPaC 2020-05-25 14:46:08 +08:00
parent c8f69f5db2
commit d9bcdac3dc
3 changed files with 64 additions and 14 deletions

View File

@ -40,21 +40,24 @@ void AssignGpuStream(const std::shared_ptr<session::KernelGraph> &kernel_graph)
}
}
if (allreduce_kernels.size() > 1) {
DeviceStream comm_stream = nullptr;
GPUDeviceManager::GetInstance().CreateStream(&comm_stream);
std::transform(
allreduce_kernels.begin(), allreduce_kernels.end(), allreduce_kernels.begin(), [&](CNodePtr allreduce_kernel) {
AnfAlgo::SetNodeAttr("stream_id", MakeValue(reinterpret_cast<uintptr_t>(comm_stream)), allreduce_kernel);
return allreduce_kernel;
});
// Assign multiple streams only when there's Recv node for AllReduce.
std::vector<SendRecvPair> send_recv_pairs;
FindAllReduceStreamSwitchPos(kernel_graph, &send_recv_pairs);
InsertStreamSwitchNode(kernel_graph, send_recv_pairs);
if (FindAllReduceStreamSwitchPos(kernel_graph, &send_recv_pairs)) {
DeviceStream comm_stream = nullptr;
GPUDeviceManager::GetInstance().CreateStream(&comm_stream);
std::transform(
allreduce_kernels.begin(), allreduce_kernels.end(), allreduce_kernels.begin(), [&](CNodePtr allreduce_kernel) {
AnfAlgo::SetNodeAttr("stream_id", MakeValue(reinterpret_cast<uintptr_t>(comm_stream)), allreduce_kernel);
return allreduce_kernel;
});
InsertStreamSwitchNode(kernel_graph, send_recv_pairs);
} else {
return;
}
}
}
void FindAllReduceStreamSwitchPos(const std::shared_ptr<session::KernelGraph> &kernel_graph,
bool FindAllReduceStreamSwitchPos(const std::shared_ptr<session::KernelGraph> &kernel_graph,
std::vector<SendRecvPair> *send_recv_pairs) {
auto execution_kernels = kernel_graph->execution_order();
std::vector<CNodePtr>::iterator iter, iter_begin;
@ -77,14 +80,15 @@ void FindAllReduceStreamSwitchPos(const std::shared_ptr<session::KernelGraph> &k
std::vector<CNodePtr>::iterator mock_recv_node_iter =
FindRecvNodePos(iter, iter_end, *iter, kAllReduceStreamSwitch);
if (mock_recv_node_iter == iter_end) {
MS_LOG(WARNING) << "Can't find send node place before AllReduce node.";
continue;
MS_LOG(WARNING) << "Can't find recv node place after AllReduce node.";
return false;
}
SendRecvPair pair2 = {kAllReduceStreamSwitch, *iter, *mock_recv_node_iter, IntToSize(iter - iter_begin + 1),
IntToSize(mock_recv_node_iter - iter_begin)};
send_recv_pairs->push_back(pair2);
}
}
return true;
}
std::vector<CNodePtr>::iterator FindSendNodePos(std::vector<CNodePtr>::iterator begin,

View File

@ -48,7 +48,7 @@ struct StreamSwitchNode {
}
};
void AssignGpuStream(const std::shared_ptr<session::KernelGraph> &kernel_graph);
void FindAllReduceStreamSwitchPos(const std::shared_ptr<session::KernelGraph> &kernel_graph,
bool FindAllReduceStreamSwitchPos(const std::shared_ptr<session::KernelGraph> &kernel_graph,
std::vector<SendRecvPair> *send_recv_pairs);
// Find Send node position according to "mock" recv node.
// "mock" recv node is a gpu kernel node after a real Recv node, e.g. AllReduce node.

View File

@ -75,3 +75,49 @@ def test_AllReduce():
error2 = np.ones(shape=expect2.shape) * 1.0e-5
assert np.all(diff2 < error2)
assert output[2].shape() == expect2.shape
class Net2(nn.Cell):
def __init__(self):
super(Net2, self).__init__()
self.x1 = Parameter(initializer(Tensor(x), x.shape), name='x1')
self.op0 = "sum"
self.op1 = "sum"
self.op2 = "sum"
self.all_reduce1 = P.AllReduce(self.op0, group=NCCL_WORLD_COMM_GROUP)
self.all_reduce2 = P.AllReduce(self.op1, group=NCCL_WORLD_COMM_GROUP)
self.all_reduce3 = P.AllReduce(self.op2, group=NCCL_WORLD_COMM_GROUP)
def construct(self):
x = self.all_reduce1(self.x1)
y = self.all_reduce2(x)
z = self.all_reduce3(y)
return (x, y, z)
def test_AllReduce2():
all_reduce = Net2()
output = all_reduce()
expect0 = np.ones([3, 1, 3, 3]).astype(np.float32) * 0
for i in range(size):
part = np.ones([3, 1, 3, 3]).astype(np.float32) * 0.01 * (i + 1)
expect0 += part
diff0 = abs(output[0].asnumpy() - expect0)
error0 = np.ones(shape=expect0.shape) * 1.0e-5
assert np.all(diff0 < error0)
assert output[0].shape() == expect0.shape
expect1 = expect0 * size
diff1 = abs(output[1].asnumpy() - expect1)
error1 = np.ones(shape=expect1.shape) * 1.0e-5
assert np.all(diff1 < error1)
assert output[1].shape() == expect1.shape
expect2 = expect1 * size
diff2 = abs(output[2].asnumpy() - expect2)
error2 = np.ones(shape=expect2.shape) * 1.0e-5
assert np.all(diff2 < error2)
assert output[2].shape() == expect2.shape