1 | #include <gtest/gtest.h> |
2 | |
3 | #include <c10/util/irange.h> |
4 | #include <tensorpipe/common/cpu_buffer.h> |
5 | #include <tensorpipe/core/message.h> |
6 | #include <torch/csrc/distributed/rpc/tensorpipe_utils.h> |
7 | #include <torch/torch.h> |
8 | |
9 | #include <memory> |
10 | #include <string> |
11 | #include <vector> |
12 | |
13 | TEST(TensorpipeSerialize, Base) { |
14 | // Sender serializes |
15 | at::Tensor t1 = torch::ones({1024}, at::ScalarType::Int); |
16 | at::Tensor t2 = torch::ones({1024}, at::ScalarType::Float); |
17 | std::vector<at::Tensor> tensors{t1, t2}; |
18 | std::vector<char> payload = {'1', '2', '3'}; |
19 | std::vector<char> payloadCopy = payload; // for testing |
20 | torch::distributed::rpc::MessageType mtype = |
21 | torch::distributed::rpc::MessageType::UNKNOWN; |
22 | int64_t mId = 100; |
23 | auto sendingRpcMessage = |
24 | c10::make_intrusive<torch::distributed::rpc::Message>( |
25 | std::move(payload), std::move(tensors), mtype); |
26 | sendingRpcMessage->setId(mId); |
27 | tensorpipe::Message sendingTpMessage; |
28 | torch::distributed::rpc::TensorpipeWriteBuffers sendingTpBuffers; |
29 | std::tie(sendingTpMessage, sendingTpBuffers) = |
30 | torch::distributed::rpc::tensorpipeSerialize( |
31 | std::move(sendingRpcMessage), {}, {}); |
32 | |
33 | // Mimic receiving message descriptor: recvingTpDescriptor is a copy of |
34 | // sendingTpMessage except for the data pointers which are left null. |
35 | tensorpipe::Descriptor recvingTpDescriptor; |
36 | recvingTpDescriptor.metadata = sendingTpMessage.metadata; |
37 | recvingTpDescriptor.payloads.reserve(sendingTpMessage.payloads.size()); |
38 | for (auto& tpPayload : sendingTpMessage.payloads) { |
39 | tensorpipe::Descriptor::Payload p; |
40 | p.length = tpPayload.length; |
41 | p.metadata = tpPayload.metadata; |
42 | recvingTpDescriptor.payloads.push_back(std::move(p)); |
43 | } |
44 | EXPECT_EQ( |
45 | recvingTpDescriptor.payloads.size(), sendingTpMessage.payloads.size()); |
46 | recvingTpDescriptor.tensors.reserve(sendingTpMessage.tensors.size()); |
47 | for (auto& tpTensor : sendingTpMessage.tensors) { |
48 | tensorpipe::Descriptor::Tensor t; |
49 | t.length = tpTensor.length; |
50 | t.sourceDevice = tpTensor.buffer.device(); |
51 | t.targetDevice = tpTensor.targetDevice; |
52 | t.metadata = tpTensor.metadata; |
53 | recvingTpDescriptor.tensors.push_back(std::move(t)); |
54 | } |
55 | EXPECT_EQ( |
56 | recvingTpDescriptor.tensors.size(), sendingTpMessage.tensors.size()); |
57 | |
58 | // Mimic readDescriptor() callback: |
59 | // - Allocate buffers |
60 | // - Fill pointers in tensorpipe message |
61 | tensorpipe::Allocation recvingTpAllocation; |
62 | torch::distributed::rpc::TensorpipeReadBuffers recvingTpBuffers; |
63 | std::tie(recvingTpAllocation, recvingTpBuffers) = |
64 | torch::distributed::rpc::tensorpipeAllocate(recvingTpDescriptor, {}); |
65 | |
66 | // Mimic tensorpipe data transfer |
67 | EXPECT_EQ( |
68 | recvingTpAllocation.payloads.size(), sendingTpMessage.payloads.size()); |
69 | for (const auto i : c10::irange(recvingTpAllocation.payloads.size())) { |
70 | tensorpipe::Message::Payload& srcPayload = sendingTpMessage.payloads[i]; |
71 | tensorpipe::Allocation::Payload& dstPayload = |
72 | recvingTpAllocation.payloads[i]; |
73 | if (srcPayload.length) { |
74 | // Empty vector's data() can return nullptr, use the length to avoid |
75 | // coying into nullptr |
76 | memcpy(dstPayload.data, srcPayload.data, srcPayload.length); |
77 | } |
78 | } |
79 | EXPECT_EQ( |
80 | recvingTpAllocation.tensors.size(), sendingTpMessage.tensors.size()); |
81 | for (const auto i : c10::irange(recvingTpAllocation.tensors.size())) { |
82 | tensorpipe::Message::Tensor& srcTensor = sendingTpMessage.tensors[i]; |
83 | tensorpipe::Allocation::Tensor& dstTensor = recvingTpAllocation.tensors[i]; |
84 | memcpy( |
85 | dstTensor.buffer.unwrap<tensorpipe::CpuBuffer>().ptr, |
86 | srcTensor.buffer.unwrap<tensorpipe::CpuBuffer>().ptr, |
87 | srcTensor.length); |
88 | } |
89 | |
90 | // Mimic read() callback: |
91 | // - Unpickle |
92 | c10::intrusive_ptr<torch::distributed::rpc::Message> recvingRpcMessage = |
93 | torch::distributed::rpc::tensorpipeDeserialize( |
94 | std::move(recvingTpDescriptor), std::move(recvingTpBuffers)); |
95 | |
96 | // Data is ready |
97 | EXPECT_EQ(mtype, recvingRpcMessage->type()); |
98 | EXPECT_EQ(payloadCopy, recvingRpcMessage->payload()); |
99 | EXPECT_EQ(mId, recvingRpcMessage->id()); |
100 | EXPECT_TRUE(torch::equal(t1, recvingRpcMessage->tensors()[0])); |
101 | EXPECT_TRUE(torch::equal(t2, recvingRpcMessage->tensors()[1])); |
102 | } |
103 | |
104 | TEST(TensorpipeSerialize, RecopySparseTensors) { |
105 | // Take a 1K row of a 1M tensors, and make sure we don't send across 1M rows. |
106 | constexpr size_t k1K = 1024; |
107 | at::Tensor main = torch::randn({k1K, k1K}); |
108 | at::Tensor tiny = main.select(0, 2); // Select a row in the middle |
109 | EXPECT_EQ(tiny.numel(), k1K); |
110 | EXPECT_EQ(tiny.storage().nbytes() / tiny.itemsize(), k1K * k1K); |
111 | |
112 | std::vector<at::Tensor> tensors{main, tiny}; |
113 | std::vector<char> payload = {'1', '2', '3'}; |
114 | torch::distributed::rpc::MessageType mtype = |
115 | torch::distributed::rpc::MessageType::UNKNOWN; |
116 | auto sendingRpcMessage = |
117 | c10::make_intrusive<torch::distributed::rpc::Message>( |
118 | std::move(payload), std::move(tensors), mtype); |
119 | |
120 | tensorpipe::Message sendingTpMessage; |
121 | torch::distributed::rpc::TensorpipeWriteBuffers tpBuffers; |
122 | std::tie(sendingTpMessage, tpBuffers) = |
123 | torch::distributed::rpc::tensorpipeSerialize( |
124 | std::move(sendingRpcMessage), {}, {}); |
125 | |
126 | EXPECT_EQ(tpBuffers.tensors.size(), 2); |
127 | EXPECT_EQ(sendingTpMessage.tensors.size(), 2); |
128 | EXPECT_TRUE(torch::equal(main, tpBuffers.tensors[0])); |
129 | EXPECT_TRUE(torch::equal(tiny, tpBuffers.tensors[1])); |
130 | // Test cloned storage |
131 | EXPECT_EQ( |
132 | main.storage().data(), |
133 | sendingTpMessage.tensors[0].buffer.unwrap<tensorpipe::CpuBuffer>().ptr); |
134 | EXPECT_NE( |
135 | tiny.storage().data(), |
136 | sendingTpMessage.tensors[1].buffer.unwrap<tensorpipe::CpuBuffer>().ptr); |
137 | EXPECT_EQ(tiny.element_size() * k1K, sendingTpMessage.tensors[1].length); |
138 | } |
139 | |
140 | TEST(TensorpipeSerialize, NoDeleterTensors) { |
141 | std::vector<float> blob1{.8, .2}; |
142 | std::vector<float> blob2{.7, .5, .9}; |
143 | at::Tensor t1 = torch::from_blob((float*)(blob1.data()), blob1.size()); |
144 | at::Tensor t2 = torch::from_blob((float*)(blob2.data()), blob2.size()); |
145 | std::vector<at::Tensor> tensors{t1, t2}; |
146 | std::vector<char> payload = {'1', '2', '3'}; |
147 | torch::distributed::rpc::MessageType mtype = |
148 | torch::distributed::rpc::MessageType::UNKNOWN; |
149 | auto sendingRpcMessage = |
150 | c10::make_intrusive<torch::distributed::rpc::Message>( |
151 | std::move(payload), std::move(tensors), mtype); |
152 | |
153 | tensorpipe::Message sendingTpMessage; |
154 | torch::distributed::rpc::TensorpipeWriteBuffers tpBuffers; |
155 | std::tie(sendingTpMessage, tpBuffers) = |
156 | torch::distributed::rpc::tensorpipeSerialize( |
157 | std::move(sendingRpcMessage), {}, {}); |
158 | |
159 | EXPECT_EQ(tpBuffers.copiedTensors.size(), 2); |
160 | EXPECT_EQ(sendingTpMessage.tensors.size(), 2); |
161 | EXPECT_EQ( |
162 | tpBuffers.copiedTensors[0].size(), sendingTpMessage.tensors[0].length); |
163 | EXPECT_EQ( |
164 | tpBuffers.copiedTensors[1].size(), sendingTpMessage.tensors[1].length); |
165 | EXPECT_EQ( |
166 | tpBuffers.copiedTensors[0].data(), |
167 | sendingTpMessage.tensors[0].buffer.unwrap<tensorpipe::CpuBuffer>().ptr); |
168 | EXPECT_EQ( |
169 | tpBuffers.copiedTensors[1].data(), |
170 | sendingTpMessage.tensors[1].buffer.unwrap<tensorpipe::CpuBuffer>().ptr); |
171 | EXPECT_TRUE( |
172 | memcmp( |
173 | tpBuffers.copiedTensors[0].data(), |
174 | t1.storage().data(), |
175 | sendingTpMessage.tensors[0].length) == 0); |
176 | EXPECT_TRUE( |
177 | memcmp( |
178 | tpBuffers.copiedTensors[1].data(), |
179 | t2.storage().data(), |
180 | sendingTpMessage.tensors[1].length) == 0); |
181 | } |
182 | |