Add XLABackendIntf test (#3697)

* Add tensor transfer test * Scalar transfer test * add e2e test for xlabckend intf * initXlaBackend in cppTest
pytorch · Jul 19, 2022 · 75ac08b · 75ac08b
1 parent b2f9d01
commit 75ac08b
Show file tree

Hide file tree

Showing 4 changed files with 95 additions and 6 deletions.
diff --git a/test/cpp/CMakeLists.txt b/test/cpp/CMakeLists.txt
@@ -58,6 +58,7 @@ set(TORCH_XLA_TEST_SOURCES
   test_tensor.cpp
   test_xla_util_cache.cpp
   torch_xla_test.cpp
+  test_xla_backend_intf.cpp
 )
 
 add_executable(test_ptxla ${TORCH_XLA_TEST_SOURCES})

diff --git a/test/cpp/test_xla_backend_intf.cpp b/test/cpp/test_xla_backend_intf.cpp
@@ -0,0 +1,86 @@
+#include <vector>
+
+#include "cpp_test_util.h"
+#include "torch_xla/csrc/tensor_util.h"
+
+namespace torch_xla {
+namespace cpp_test {
+
+TEST(XLABackendTest, TestTensorTransfer) {
+  torch::lazy::BackendImplInterface* impl = GetXlaBackendImpl();
+  at::Tensor input = at::randint(std::numeric_limits<uint8_t>::min(),
+                                 std::numeric_limits<uint8_t>::max(), {2, 2},
+                                 at::TensorOptions(at::kByte));
+  ForEachDevice([&](const torch::lazy::BackendDevice& device) {
+    torch::lazy::BackendDataPtr data = impl->MakeComputationDataFromTensor(
+        input, torch::lazy::Shape(input.scalar_type(), input.sizes()), device);
+    at::Tensor res = impl->MakeTensorFromComputationData(data, at::kByte);
+    AllClose(input, res);
+  });
+}
+
+TEST(XLABackendTest, TestScalarTransfer) {
+  torch::lazy::BackendImplInterface* impl = GetXlaBackendImpl();
+  at::Scalar input = at::Scalar(int64_t(1));
+  ForEachDevice([&](const torch::lazy::BackendDevice& device) {
+    torch::lazy::BackendDataPtr data =
+        impl->MakeComputationDataFromScalar(input, device);
+    at::Tensor res = impl->MakeTensorFromComputationData(data, at::kByte);
+    AllClose(at::ones({}, at::TensorOptions(at::kByte)), res);
+  });
+}
+
+TEST(XLABackendTest, TestPlaceholder) {
+  torch::lazy::BackendImplInterface* impl = GetXlaBackendImpl();
+  torch::lazy::Shape shape(at::kFloat, {10, 10});
+  ForEachDevice([&](const torch::lazy::BackendDevice& device) {
+    torch::lazy::BackendDataPtr data =
+        impl->CreateDataPlaceholder(device, shape);
+    xla::ComputationClient::DataPtr computation_data = UnwrapXlaData(data);
+    EXPECT_EQ(computation_data->device(), device.toString());
+    EXPECT_EQ(computation_data->shape(),
+              MakeXlaShapeFromLazyShape(shape, device));
+  });
+}
+
+xla::XlaComputation CreateAddComputation(const xla::Shape& shape) {
+  xla::XlaBuilder builder("AddComputation");
+  xla::XlaOp x = xla::Parameter(&builder, 0, shape, "x");
+  xla::XlaOp y = xla::Parameter(&builder, 1, shape, "y");
+  xla::XlaOp sum = xla::Add(x, y);
+  return ConsumeValue(builder.Build());
+}
+
+TEST(XLABackendTest, TestE2E) {
+  torch::lazy::BackendImplInterface* impl = GetXlaBackendImpl();
+  xla::Shape input_shape =
+      xla::ShapeUtil::MakeShape(xla::PrimitiveType::F32, {8, 8});
+  at::Tensor one = at::ones({8, 8}, at::TensorOptions(at::kFloat));
+  std::vector<at::Tensor> tensors = {one, one};
+
+  ForEachDevice([&](const torch::lazy::BackendDevice& device) {
+    xla::XlaComputation xla_computation = CreateAddComputation(input_shape);
+    torch::lazy::ComputationPtr computation =
+        std::make_shared<torch_xla::Computation>(
+            "test", std::move(xla_computation), device);
+    std::vector<torch::lazy::ComputationPtr> compiled_programs =
+        impl->Compile({computation});
+    EXPECT_EQ(compiled_programs.size(), 1);
+
+    std::vector<torch::lazy::BackendDataPtr> parameters;
+    for (auto& tensor : tensors) {
+      parameters.push_back(impl->MakeComputationDataFromTensor(
+          tensor, torch::lazy::Shape(tensor.scalar_type(), tensor.sizes()),
+          device));
+    }
+    std::vector<torch::lazy::BackendDataPtr> res =
+        impl->ExecuteComputation(compiled_programs[0], parameters, device);
+    EXPECT_EQ(res.size(), 1);
+    at::Tensor res_tensor =
+        impl->MakeTensorFromComputationData(res[0], at::kFloat);
+    AllClose(one + one, res_tensor);
+  });
+}
+
+}  // namespace cpp_test
+}  // namespace torch_xla
diff --git a/torch_xla/csrc/xla_backend_impl.cpp b/torch_xla/csrc/xla_backend_impl.cpp
@@ -105,6 +105,7 @@ class XlaBackendImpl : public torch::lazy::BackendImplInterface {
     // c10::ArrayRef<torch::lazy::Node*> instead of
     // c10::ArrayRef<const torch::lazy::Node*> since c10::ArrayRef already
     // provided const for its member.
+    XLA_ERROR() << "Need to handle post_order";
     return std::make_unique<LoweringContext>(name, device);
   }
 
@@ -131,15 +132,16 @@ class XlaBackendImpl : public torch::lazy::BackendImplInterface {
     std::vector<torch::lazy::ComputationPtr> res;
     std::vector<xla::ComputationClient::CompileInstance> compile_instances;
     torch::lazy::BackendDevice current_device = GetCurrentDevice();
+    std::vector<xla::Shape> output_shapes;
 
     for (const torch::lazy::ComputationPtr instance : instances) {
       // TODO(JackCaoG): device is missing in instance, use CurrentDevice for
       // now
       const Computation* torch_xla_computation =
           dynamic_cast<Computation*>(instance.get());
-      xla::Shape shape = MakeShapeWithDeviceLayout(
+      output_shapes.push_back(MakeShapeWithDeviceLayout(
           torch_xla_computation->program_shape().result(),
-          static_cast<XlaDeviceType>(current_device.type()));
+          static_cast<XlaDeviceType>(current_device.type())));
 
       // Call GetCompilationDevices and passes all device here if needed.
       // Currently on TPU we always have 1 replica per device and one process
@@ -152,7 +154,7 @@ class XlaBackendImpl : public torch::lazy::BackendImplInterface {
       compile_instances.push_back(xla::ComputationClient::CompileInstance(
           torch_xla_computation->move_computation(),
           torch_xla_computation->get_device_string(),
-          {current_device.toString()}, &shape));
+          {current_device.toString()}, &output_shapes.back()));
     }
     std::vector<std::shared_ptr<xla::ComputationClient::Computation>>
         client_computations = xla::ComputationClient::Get()->Compile(
@@ -238,4 +240,4 @@ void InitXlaBackend() {
       std::make_unique<torch::lazy::BackendRegistrar>(GetXlaBackendImpl());
 };
 
-}  // namespace torch_xla
+}  // namespace torch_xla
diff --git a/torch_xla/csrc/xla_backend_impl.h b/torch_xla/csrc/xla_backend_impl.h
@@ -39,8 +39,8 @@ class XLAData : public torch::lazy::BackendData {
   xla::ComputationClient::DataPtr xla_data_;
 };
 
-// torch::lazy::BackendImplInterface* GetXlaBackendImpl();
+torch::lazy::BackendImplInterface* GetXlaBackendImpl();
 
 void InitXlaBackend();
 
-}  // namespace torch_xla
+}  // namespace torch_xla