Enabled performance benchmark tests for Eager Dygraph (#37653)

* Enabled performance benchmark tests for Eager Dygraph

* Protected CUDA tests with macro

* Fixed dependency issues for windows-ci

paddle/fluid/eager/CMakeLists.txt

@@ -2,7 +2,7 @@ set(eager_deps pten pten_api hook_utils tensor_utils utils global_utils backward
 set(fluid_deps tracer layer proto_desc operator op_registry variable_helper memcpy)
 set(generated_deps dygraph_function dygraph_node)
 
-if(NOT DEFINED ON_INFER)
+if(NOT ON_INFER)
     message("Performing Eager Dygraph Auto Code Generation")
     add_subdirectory(auto_code_generator)
 endif()

paddle/fluid/eager/api/generated/CMakeLists.txt

 add_subdirectory(eager_generated)
 
-if(NOT DEFINED ON_INFER)
+if(NOT ON_INFER)
     add_subdirectory(fluid_generated)
 endif()

paddle/fluid/eager/auto_code_generator/CMakeLists.txt

@@ -17,9 +17,38 @@ execute_process(
 )
 
 if(WIN32)
+    set(EAGER_CODEGEN_DEPS eager_generator)
+    if("${CMAKE_GENERATOR}" STREQUAL "Ninja")
+      set(eager_generator_path "${CMAKE_CURRENT_BINARY_DIR}")
+    else()
+      set(eager_generator_path "${CMAKE_CURRENT_BINARY_DIR}/${CMAKE_BUILD_TYPE}")
+    endif()
+    
+    if(${CBLAS_PROVIDER} STREQUAL MKLML)
+      message("Copied libiomp5md.dll for Eager AutoCodeGen")
+      ADD_CUSTOM_COMMAND(OUTPUT ${eager_generator_path}/libiomp5md.dll
+        COMMAND ${CMAKE_COMMAND} -E copy ${MKLML_SHARED_IOMP_LIB} ${eager_generator_path}
+        DEPENDS mklml)
+      list(APPEND EAGER_CODEGEN_DEPS ${eager_generator_path}/libiomp5md.dll)
+    else(${CBLAS_PROVIDER} STREQUAL EXTERN_OPENBLAS)
+      message("Copied openblas.dll for Eager AutoCodeGen")
+      ADD_CUSTOM_COMMAND(OUTPUT ${eager_generator_path}/openblas.dll
+        COMMAND ${CMAKE_COMMAND} -E copy ${OPENBLAS_SHARED_LIB} ${eager_generator_path}
+        DEPENDS extern_openblas)
+      list(APPEND EAGER_CODEGEN_DEPS ${eager_generator_path}/openblas.dll)
+    endif()
+
+    if(WITH_MKLDNN)
+      message("Copied mkldnn.dll for Eager AutoCodeGen")
+      ADD_CUSTOM_COMMAND(OUTPUT ${eager_generator_path}/mkldnn.dll
+        COMMAND ${CMAKE_COMMAND} -E copy ${MKLDNN_SHARED_LIB} ${eager_generator_path}
+        DEPENDS mkldnn)
+        list(APPEND EAGER_CODEGEN_DEPS ${eager_generator_path}/mkldnn.dll)
+    endif()
+
     add_custom_target(eager_codegen
-      COMMAND "${CMAKE_CURRENT_BINARY_DIR}/eager_generator.exe" "${PADDLE_SOURCE_DIR}/paddle/fluid/eager/api/generated/fluid_generated" 
-      DEPENDS eager_generator
+      COMMAND "${eager_generator_path}/eager_generator.exe" "${PADDLE_SOURCE_DIR}/paddle/fluid/eager/api/generated/fluid_generated" 
+      DEPENDS ${EAGER_CODEGEN_DEPS}
       VERBATIM)
 else()
     add_custom_target(eager_codegen

paddle/fluid/eager/tests/CMakeLists.txt

 add_subdirectory(data_structure_tests)
 add_subdirectory(task_tests)
+
+if(NOT ON_INFER)
+    add_subdirectory(performance_tests)
+endif()

paddle/fluid/eager/tests/performance_tests/CMakeLists.txt

+cc_library(performance_benchmark_utils SRCS benchmark_utils.cc DEPS ${eager_deps} ${fluid_deps} ${generated_deps} eager_scale scale_node scale_op matmul_v2_op)
+
+cc_test(test_egr_performance_benchmark_eager_cpu SRCS benchmark_eager_cpu.cc DEPS performance_benchmark_utils ${eager_deps} ${fluid_deps})
+cc_test(test_egr_performance_benchmark_fluid_cpu SRCS benchmark_fluid_cpu.cc DEPS performance_benchmark_utils ${eager_deps} ${fluid_deps})
+
+cc_test(test_egr_performance_benchmark_eager_cuda SRCS benchmark_eager_cuda.cc DEPS performance_benchmark_utils ${eager_deps} ${fluid_deps})
+cc_test(test_egr_performance_benchmark_fluid_cuda SRCS benchmark_fluid_cuda.cc DEPS performance_benchmark_utils ${eager_deps} ${fluid_deps})

paddle/fluid/eager/tests/performance_tests/benchmark_eager_cpu.cc

@@ -25,7 +25,7 @@
 
 #include "paddle/fluid/imperative/tracer.h"
 
-#include "paddle/fluid/eager/tests/benchmark/benchmark_utils.h"
+#include "paddle/fluid/eager/tests/performance_tests/benchmark_utils.h"
 #include "paddle/fluid/eager/tests/test_utils.h"
 
 #ifdef WITH_GPERFTOOLS
@@ -42,11 +42,11 @@ TEST(Benchmark, Init) { FLAGS_run_pten_kernel = false; }
 
 TEST(Benchmark, EagerScaleCPU) {
   // Prepare Device Contexts
-  egr::InitEnv(paddle::platform::CPUPlace());
+  eager_test::InitEnv(paddle::platform::CPUPlace());
 
   for (const std::string& mode : {"Accuracy", "Performance"}) {
     paddle::framework::DDim ddim = paddle::framework::make_ddim({2, 4, 4, 4});
-    egr::EagerTensor tensor = EagerUtils::CreateTensorWithValue(
+    egr::EagerTensor tensor = CreateTensorWithValue(
         ddim, paddle::platform::CPUPlace(), pten::DataType::FLOAT32,
         pten::DataLayout::NCHW, 5.0, true);
     RetainGradForTensor(tensor);
@@ -78,20 +78,20 @@ TEST(Benchmark, EagerScaleCPU) {
 
 TEST(Benchmark, EagerIntermediateMatmulCPU) {
   // Prepare Device Contexts
-  InitEnv(paddle::platform::CPUPlace());
+  eager_test::InitEnv(paddle::platform::CPUPlace());
 
   auto tracer = std::make_shared<paddle::imperative::Tracer>();
   paddle::imperative::SetCurrentTracer(tracer);
 
   for (const std::string& mode : {"Accuracy", "Performance"}) {
     paddle::framework::DDim ddimX = paddle::framework::make_ddim({2, 2});
-    egr::EagerTensor X = EagerUtils::CreateTensorWithValue(
+    egr::EagerTensor X = CreateTensorWithValue(
         ddimX, paddle::platform::CPUPlace(), pten::DataType::FLOAT32,
         pten::DataLayout::NCHW, 1.0, true);
     RetainGradForTensor(X);
 
     paddle::framework::DDim ddimY = paddle::framework::make_ddim({2, 2});
-    egr::EagerTensor Y = EagerUtils::CreateTensorWithValue(
+    egr::EagerTensor Y = CreateTensorWithValue(
         ddimY, paddle::platform::CPUPlace(), pten::DataType::FLOAT32,
         pten::DataLayout::NCHW, 2.0, true);
     RetainGradForTensor(Y);
@@ -122,7 +122,7 @@ TEST(Benchmark, EagerIntermediateMatmulCPU) {
 
 TEST(Benchmark, EagerIntermediateMLPCPU) {
   // Prepare Device Contexts
-  InitEnv(paddle::platform::CPUPlace());
+  eager_test::InitEnv(paddle::platform::CPUPlace());
 
   auto tracer = std::make_shared<paddle::imperative::Tracer>();
   paddle::imperative::SetCurrentTracer(tracer);
@@ -130,7 +130,7 @@ TEST(Benchmark, EagerIntermediateMLPCPU) {
   for (const std::string& mode : {"Accuracy", "Performance"}) {
     paddle::framework::DDim ddimX =
         paddle::framework::make_ddim({MLP_M, MLP_N});
-    egr::EagerTensor X = EagerUtils::CreateTensorWithValue(
+    egr::EagerTensor X = CreateTensorWithValue(
         ddimX, paddle::platform::CPUPlace(), pten::DataType::FLOAT32,
         pten::DataLayout::NCHW, MLP_X_VAL, true);
     RetainGradForTensor(X);
@@ -140,13 +140,13 @@ TEST(Benchmark, EagerIntermediateMLPCPU) {
     for (size_t i = 0; i < MLP_NUM_LINEAR; i++) {
       paddle::framework::DDim ddimW =
           paddle::framework::make_ddim({MLP_N, MLP_K});
-      egr::EagerTensor W = EagerUtils::CreateTensorWithValue(
+      egr::EagerTensor W = CreateTensorWithValue(
           ddimW, paddle::platform::CPUPlace(), pten::DataType::FLOAT32,
           pten::DataLayout::NCHW, MLP_W_VAL, true);
       RetainGradForTensor(W);
 
       paddle::framework::DDim ddimB = paddle::framework::make_ddim({MLP_K});
-      egr::EagerTensor B = EagerUtils::CreateTensorWithValue(
+      egr::EagerTensor B = CreateTensorWithValue(
           ddimB, paddle::platform::CPUPlace(), pten::DataType::FLOAT32,
           pten::DataLayout::NCHW, MLP_B_VAL, true);
       RetainGradForTensor(B);

paddle/fluid/eager/tests/performance_tests/benchmark_eager_cuda.cc

@@ -24,7 +24,7 @@
 
 #include "paddle/fluid/imperative/tracer.h"
 
-#include "paddle/fluid/eager/tests/benchmark/benchmark_utils.h"
+#include "paddle/fluid/eager/tests/performance_tests/benchmark_utils.h"
 #include "paddle/fluid/eager/tests/test_utils.h"
 
 #ifdef WITH_GPERFTOOLS
@@ -38,12 +38,14 @@ DECLARE_bool(run_pten_kernel);
 
 TEST(Benchmark, Init) { FLAGS_run_pten_kernel = false; }
 
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+
 TEST(Benchmark, EagerScaleCUDA) {
-  egr::InitEnv(paddle::platform::CUDAPlace());
+  eager_test::InitEnv(paddle::platform::CUDAPlace());
 
   for (const std::string& mode : {"Accuracy", "WarmUp", "Performance"}) {
     paddle::framework::DDim ddim = paddle::framework::make_ddim({2, 4, 4, 4});
-    egr::EagerTensor tensor = EagerUtils::CreateTensorWithValue(
+    egr::EagerTensor tensor = CreateTensorWithValue(
         ddim, paddle::platform::CUDAPlace(), pten::DataType::FLOAT32,
         pten::DataLayout::NCHW, 5.0 /*value*/, true /*is_leaf*/);
     RetainGradForTensor(tensor);
@@ -77,7 +79,7 @@ TEST(Benchmark, EagerScaleCUDA) {
 
 TEST(Benchmark, EagerIntermediateMatmulCUDA) {
   paddle::platform::CUDAPlace place;
-  egr::InitEnv(place);
+  eager_test::InitEnv(place);
 
   auto tracer = std::make_shared<paddle::imperative::Tracer>();
   tracer->SetExpectedPlace(place);
@@ -85,13 +87,13 @@ TEST(Benchmark, EagerIntermediateMatmulCUDA) {
 
   for (const std::string& mode : {"Accuracy", "WarmUp", "Performance"}) {
     paddle::framework::DDim ddimX = paddle::framework::make_ddim({2, 2});
-    egr::EagerTensor X = EagerUtils::CreateTensorWithValue(
+    egr::EagerTensor X = CreateTensorWithValue(
         ddimX, paddle::platform::CUDAPlace(), pten::DataType::FLOAT32,
         pten::DataLayout::NCHW, 1.0, true);
     RetainGradForTensor(X);
 
     paddle::framework::DDim ddimY = paddle::framework::make_ddim({2, 2});
-    egr::EagerTensor Y = EagerUtils::CreateTensorWithValue(
+    egr::EagerTensor Y = CreateTensorWithValue(
         ddimY, paddle::platform::CUDAPlace(), pten::DataType::FLOAT32,
         pten::DataLayout::NCHW, 2.0, true);
     RetainGradForTensor(Y);
@@ -125,7 +127,7 @@ TEST(Benchmark, EagerIntermediateMatmulCUDA) {
 
 TEST(Benchmark, EagerIntermediateMLPCUDA) {
   paddle::platform::CUDAPlace place;
-  egr::InitEnv(place);
+  eager_test::InitEnv(place);
 
   auto tracer = std::make_shared<paddle::imperative::Tracer>();
   tracer->SetExpectedPlace(place);
@@ -134,7 +136,7 @@ TEST(Benchmark, EagerIntermediateMLPCUDA) {
   for (const std::string& mode : {"Accuracy", "WarmUp", "Performance"}) {
     paddle::framework::DDim ddimX =
         paddle::framework::make_ddim({MLP_M, MLP_N});
-    egr::EagerTensor X = EagerUtils::CreateTensorWithValue(
+    egr::EagerTensor X = CreateTensorWithValue(
         ddimX, paddle::platform::CUDAPlace(), pten::DataType::FLOAT32,
         pten::DataLayout::NCHW, MLP_X_VAL, true);
     RetainGradForTensor(X);
@@ -144,13 +146,13 @@ TEST(Benchmark, EagerIntermediateMLPCUDA) {
     for (size_t i = 0; i < MLP_NUM_LINEAR; i++) {
       paddle::framework::DDim ddimW =
           paddle::framework::make_ddim({MLP_N, MLP_K});
-      egr::EagerTensor W = EagerUtils::CreateTensorWithValue(
+      egr::EagerTensor W = CreateTensorWithValue(
           ddimW, paddle::platform::CUDAPlace(), pten::DataType::FLOAT32,
           pten::DataLayout::NCHW, MLP_W_VAL, true);
       RetainGradForTensor(W);
 
       paddle::framework::DDim ddimB = paddle::framework::make_ddim({MLP_K});
-      egr::EagerTensor B = EagerUtils::CreateTensorWithValue(
+      egr::EagerTensor B = CreateTensorWithValue(
           ddimB, paddle::platform::CUDAPlace(), pten::DataType::FLOAT32,
           pten::DataLayout::NCHW, MLP_B_VAL, true);
       RetainGradForTensor(B);
@@ -185,3 +187,5 @@ TEST(Benchmark, EagerIntermediateMLPCUDA) {
     }
   }
 }
+
+#endif  // PADDLE_WITH_CUDA || PADDLE_WITH_HIP

paddle/fluid/eager/tests/performance_tests/benchmark_fluid_cpu.cc

@@ -24,7 +24,7 @@
 #include "glog/logging.h"
 #include "gtest/gtest.h"
 
-#include "paddle/fluid/eager/tests/benchmark/benchmark_utils.h"
+#include "paddle/fluid/eager/tests/performance_tests/benchmark_utils.h"
 #include "paddle/fluid/eager/tests/test_utils.h"
 #include "paddle/fluid/imperative/basic_engine.h"
 #include "paddle/fluid/imperative/tracer.h"
@@ -45,7 +45,7 @@ namespace imperative {
 TEST(Benchmark, FluidScaleCPU) {
   // Prepare Device Contexts
   platform::CPUPlace place;
-  egr::InitEnv(place);
+  eager_test::InitEnv(place);
 
   for (const std::string& mode : {"Accuracy", "Performance"}) {
     std::shared_ptr<imperative::VarBase> X(new imperative::VarBase(true, "X"));
@@ -88,7 +88,7 @@ TEST(Benchmark, FluidScaleCPU) {
 TEST(Benchmark, FluidMatmulCPU) {
   // Prepare Device Contexts
   platform::CPUPlace place;
-  egr::InitEnv(place);
+  eager_test::InitEnv(place);
 
   for (const std::string& mode : {"Accuracy", "Performance"}) {
     std::shared_ptr<imperative::VarBase> X(new imperative::VarBase(true, "X"));
@@ -141,7 +141,7 @@ TEST(Benchmark, FluidMatmulCPU) {
 TEST(Benchmark, FluidMLPCPU) {
   // Prepare Device Contexts
   platform::CPUPlace place;
-  egr::InitEnv(place);
+  eager_test::InitEnv(place);
 
   for (const std::string& mode : {"Accuracy", "Performance"}) {
     std::vector<float> x_src_data(MLP_M * MLP_N, MLP_X_VAL);

paddle/fluid/eager/tests/performance_tests/benchmark_fluid_cuda.cc

@@ -24,7 +24,7 @@
 #include "glog/logging.h"
 #include "gtest/gtest.h"
 
-#include "paddle/fluid/eager/tests/benchmark/benchmark_utils.h"
+#include "paddle/fluid/eager/tests/performance_tests/benchmark_utils.h"
 #include "paddle/fluid/eager/tests/test_utils.h"
 #include "paddle/fluid/imperative/basic_engine.h"
 #include "paddle/fluid/imperative/tracer.h"
@@ -39,13 +39,15 @@ DECLARE_bool(run_pten_kernel);
 
 TEST(Benchmark, Init) { FLAGS_run_pten_kernel = false; }
 
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+
 namespace paddle {
 namespace imperative {
 
 TEST(Benchmark, FluidScaleCUDA) {
   // Prepare Device Contexts
   platform::CUDAPlace place;
-  egr::InitEnv(place);
+  eager_test::InitEnv(place);
 
   for (const std::string& mode : {"Accuracy", "WarmUp", "Performance"}) {
     std::shared_ptr<imperative::VarBase> X(new imperative::VarBase(true, "X"));
@@ -98,7 +100,7 @@ TEST(Benchmark, FluidScaleCUDA) {
 TEST(Benchmark, FluidMatmulCUDA) {
   // Prepare Device Contexts
   platform::CUDAPlace place;
-  egr::InitEnv(place);
+  eager_test::InitEnv(place);
 
   for (const std::string& mode : {"Accuracy", "WarmUp", "Performance"}) {
     std::shared_ptr<imperative::VarBase> X(new imperative::VarBase(true, "X"));
@@ -161,7 +163,7 @@ TEST(Benchmark, FluidMatmulCUDA) {
 TEST(Benchmark, FluidMLPCUDA) {
   // Prepare Device Contexts
   platform::CUDAPlace place;
-  egr::InitEnv(place);
+  eager_test::InitEnv(place);
 
   for (const std::string& mode : {"Accuracy", "WarmUp", "Performance"}) {
     paddle::platform::DeviceContextPool& pool =
@@ -252,3 +254,5 @@ USE_OP(scale);
 USE_OP(matmul_v2);
 USE_OP(reduce_sum);
 USE_OP(reduce_sum_grad);
+
+#endif  // PADDLE_WITH_CUDA || PADDLE_WITH_HIP

paddle/fluid/eager/tests/performance_tests/benchmark_utils.cc

@@ -36,10 +36,6 @@
 #include "paddle/fluid/imperative/tracer.h"
 #include "paddle/fluid/memory/memcpy.h"
 
-#include "paddle/fluid/eager/tests/benchmark/benchmark_utils.h"
-
-#include "paddle/pten/core/kernel_registry.h"
-
 static size_t max_num_benchmark_runs = 5000;
 
 namespace egr {
@@ -64,9 +60,9 @@ void benchmark_eager_scale(const EagerTensor& tensor, bool accuracy_check) {
 
   if (accuracy_check) {
     // Examine Forward Grad (w.r.t max_num_runs = 10)
-    CompareTensorWithValue<float>(input_tensor, 8189.0);
+    eager_test::CompareTensorWithValue<float>(input_tensor, 8189.0);
     // Examine Backward Grad (w.r.t max_num_runs = 10)
-    CompareGradTensorWithValue<float>(tensor, 1024.0);
+    eager_test::CompareGradTensorWithValue<float>(tensor, 1024.0);
   }
 }
 
@@ -89,10 +85,10 @@ void benchmark_eager_intermediate_matmul(const EagerTensor& X,
 
   if (accuracy_check) {
     // Examine Forward Grad (w.r.t max_num_runs = 2)
-    CompareVariableWithValue<float>(input_tensor0, 16);
+    eager_test::CompareVariableWithValue<float>(input_tensor0, 16);
     // Examine Backward Grad (w.r.t max_num_runs = 2)
-    CompareGradVariableWithValue<float>(X, 16);
-    CompareGradVariableWithValue<float>(Y, 16);
+    eager_test::CompareGradVariableWithValue<float>(X, 16);
+    eager_test::CompareGradVariableWithValue<float>(Y, 16);
   }
 }
 
@@ -122,11 +118,11 @@ void benchmark_eager_intermediate_mlp(const EagerTensor& X,
         compute_mlp_expected_results();
 
     // Examine Forward Grad (w.r.t max_num_runs = 2)
-    CompareVariableWithValue<float>(Out, result["Out"]);
+    eager_test::CompareVariableWithValue<float>(Out, result["Out"]);
 
     // Examine Backward Grad (w.r.t max_num_runs = 2)
-    CompareGradVariableWithValue<float>(X, result["GradX"]);
-    CompareGradVariableWithValue<float>(Ws[0], result["GradW"]);
+    eager_test::CompareGradVariableWithValue<float>(X, result["GradX"]);
+    eager_test::CompareGradVariableWithValue<float>(Ws[0], result["GradW"]);
   }
 }
 
@@ -141,6 +137,8 @@ static void FluidCheckTensorValue(const std::shared_ptr<imperative::VarBase>& X,
   auto* tensor = X->MutableVar()->GetMutable<framework::LoDTensor>();
   float* t_ptr = tensor->mutable_data<float>(place);
   std::vector<float> host_data(tensor->numel());
+
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
   if (place == paddle::platform::CUDAPlace()) {
     paddle::platform::DeviceContextPool& pool =
         paddle::platform::DeviceContextPool::Instance();
@@ -153,6 +151,8 @@ static void FluidCheckTensorValue(const std::shared_ptr<imperative::VarBase>& X,
                          sizeof(float) * tensor->numel(), stream);
     t_ptr = host_data.data();
   }
+#endif
+
   VLOG(6) << "Tensor Value: " << t_ptr[0] << ", Expected Value: " << value;
   PADDLE_ENFORCE(
       t_ptr[0] == value,
@@ -166,6 +166,8 @@ static void FluidCheckGradTensorValue(
   auto* grad_tensor = X->MutableGradVar()->GetMutable<framework::LoDTensor>();
   float* g_ptr = grad_tensor->mutable_data<float>(place);
   std::vector<float> g_host_data(grad_tensor->numel());
+
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
   if (place == paddle::platform::CUDAPlace()) {
     paddle::platform::DeviceContextPool& pool =
         paddle::platform::DeviceContextPool::Instance();
@@ -178,6 +180,8 @@ static void FluidCheckGradTensorValue(
                          sizeof(float) * grad_tensor->numel(), stream);
     g_ptr = g_host_data.data();
   }
+#endif
+
   VLOG(6) << "Tensor Value: " << g_ptr[0] << ", Expected Value: " << value;
   PADDLE_ENFORCE(
       g_ptr[0] == value,

paddle/fluid/eager/tests/task_tests/CMakeLists.txt

@@ -6,6 +6,6 @@ cc_test(test_egr_task_hook SRCS hook_test.cc DEPS ${eager_deps} ${fluid_deps} ea
 cc_test(test_egr_task_cross_batch SRCS cross_batch_accumulation_test.cc DEPS ${eager_deps} ${fluid_deps} eager_scale scale_node)
 cc_test(test_egr_task_fwd_bwd_joint SRCS fwd_bwd_joint_test.cc DEPS ${eager_deps} ${fluid_deps} eager_scale scale_node)
 
-if(NOT DEFINED ON_INFER)
+if(NOT ON_INFER)
     cc_test(test_egr_task_autocodegen SRCS generated_test.cc DEPS ${eager_deps} ${fluid_deps} ${generated_deps})
 endif()