Merge pull request #8542 from panyx0718/test

Extend current profiler for timeline and more features.

CMakeLists.txt

@@ -146,6 +146,7 @@ include(external/cares)
 include(external/grpc)
 
 include(cudnn)              # set cudnn libraries, must before configure
+include(cupti)
 include(configure)          # add paddle env configuration
 include(generic)            # simplify cmake module
 include(package)            # set paddle packages

cmake/configure.cmake

@@ -59,6 +59,7 @@ endif(NOT WITH_GOLANG)
 
 if(NOT WITH_GPU)
     add_definitions(-DHPPL_STUB_FUNC)
+    add_definitions("-DCUPTI_LIB_PATH=\"\"")
 
     list(APPEND CMAKE_CXX_SOURCE_FILE_EXTENSIONS cu)
 else()
@@ -73,7 +74,14 @@ else()
     if(NOT CUDNN_FOUND)
         message(FATAL_ERROR "Paddle needs cudnn to compile")
     endif()
-
+    if(CUPTI_FOUND)
+        include_directories(${CUPTI_INCLUDE_DIR})
+        add_definitions(-DPADDLE_WITH_CUPTI)
+        add_definitions("-DCUPTI_LIB_PATH=\"${CUPTI_LIBRARY_PATH}\"")
+    else()
+        add_definitions("-DCUPTI_LIB_PATH=\"\"")
+        message(STATUS "Cannot find CUPTI, GPU Profiling is incorrect.")
+    endif()
     set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-Xcompiler ${SIMD_FLAG}")
 
     # Include cuda and cudnn

cmake/cuda.cmake

@@ -155,7 +155,8 @@ endif()
 include_directories(${CUDA_INCLUDE_DIRS})
 list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
 if(NOT WITH_DSO)
-    list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY})
+    # TODO(panyx0718): CUPTI only allows DSO?
+    list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUPTI_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY})
 endif(NOT WITH_DSO)
 
 # setting nvcc arch flags

cmake/cupti.cmake

+if(NOT WITH_GPU)
+    return()
+endif()
+
+
+set(CUPTI_ROOT "/usr" CACHE PATH "CUPTI ROOT")
+find_path(CUPTI_INCLUDE_DIR cupti.h
+        PATHS ${CUPTI_ROOT} ${CUPTI_ROOT}/include
+        $ENV{CUPTI_ROOT} $ENV{CUPTI_ROOT}/include
+        ${CUDA_TOOLKIT_ROOT_DIR}/extras/CUPTI/include
+        NO_DEFAULT_PATH
+        )
+
+get_filename_component(__libpath_hist ${CUDA_CUDART_LIBRARY} PATH)
+
+set(TARGET_ARCH "x86_64")
+if(NOT ${CMAKE_SYSTEM_PROCESSOR})
+    set(TARGET_ARCH ${CMAKE_SYSTEM_PROCESSOR})
+endif()
+
+list(APPEND CUPTI_CHECK_LIBRARY_DIRS
+        ${CUPTI_ROOT}
+        ${CUPTI_ROOT}/lib64
+        ${CUPTI_ROOT}/lib
+        ${CUPTI_ROOT}/lib/${TARGET_ARCH}-linux-gnu
+        $ENV{CUPTI_ROOT}
+        $ENV{CUPTI_ROOT}/lib64
+        $ENV{CUPTI_ROOT}/lib
+        /usr/lib
+        ${CUDA_TOOLKIT_ROOT_DIR}/extras/CUPTI/lib64)
+find_library(CUPTI_LIBRARY NAMES libcupti.so libcupti.dylib # libcupti_static.a
+       PATHS ${CUPTI_CHECK_LIBRARY_DIRS} ${CUPTI_INCLUDE_DIR} ${__libpath_hist}
+       NO_DEFAULT_PATH
+       DOC "Path to cuPTI library.")
+
+get_filename_component(CUPTI_LIBRARY_PATH ${CUPTI_LIBRARY} DIRECTORY)
+if(CUPTI_INCLUDE_DIR AND CUPTI_LIBRARY)
+    set(CUPTI_FOUND ON)
+else()
+    set(CUPTI_FOUND OFF)
+endif()

paddle/fluid/framework/executor.cc

@@ -127,7 +127,9 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
     auto op = paddle::framework::OpRegistry::CreateOp(*op_desc);
 
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    platform::RecordEvent record_event(op->Type(), pool.Get(place_));
+    // TODO(panyx0718): Need a program id to distinguish programs.
+    platform::RecordEvent record_event(op->Type(), pool.Get(place_),
+                                       op_desc->Block()->ID());
 
     VLOG(3) << place_ << " " << op->DebugStringEx(local_scope);
     op->Run(*local_scope, place_);

paddle/fluid/framework/framework.proto

@@ -167,4 +167,6 @@ message BlockDesc {
 // Please refer to
 // https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md
 // for more details.
+// TODO(panyx0718): A model can have multiple programs. Need a
+// way to distinguish them. Maybe ID or name?
 message ProgramDesc { repeated BlockDesc blocks = 1; }

paddle/fluid/framework/op_desc.h

@@ -125,6 +125,8 @@ class OpDesc {
 
   BlockDesc *Block() { return this->block_; }
 
+  const BlockDesc &BlockRef() const { return *this->block_; }
+
   void SetBlock(BlockDesc *block) { this->block_ = block; }
 
  private:

paddle/fluid/platform/CMakeLists.txt

+proto_library(profiler_proto SRCS profiler.proto)
+
 if(WITH_GPU)
   cc_library(enforce SRCS enforce.cc DEPS)
 else()
@@ -37,7 +39,8 @@ nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
 nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place device_context)
 nv_test(nccl_test SRCS nccl_test.cu DEPS dynload_cuda gpu_info device_context)
 
-cc_library(profiler SRCS profiler.cc DEPS device_context)
+cc_library(device_tracer SRCS device_tracer.cc DEPS profiler_proto ${GPU_CTX_DEPS})
+cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer)
 cc_test(profiler_test SRCS profiler_test.cc DEPS profiler)
 
 nv_test(float16_gpu_test SRCS float16_test.cu)

paddle/fluid/platform/device_tracer.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/platform/device_tracer.h"
+#include <map>
+#include <mutex>
+#include "glog/logging.h"
+#include "paddle/fluid/framework/block_desc.h"
+#include "paddle/fluid/string/printf.h"
+
+namespace paddle {
+namespace platform {
+namespace {
+
+thread_local const char *cur_annotation = nullptr;
+std::once_flag tracer_once_flag;
+DeviceTracer *tracer = nullptr;
+}  // namespace
+#ifdef PADDLE_WITH_CUPTI
+
+namespace {
+// TODO(panyx0718): Revisit the buffer size here.
+uint64_t kBufSize = 32 * 1024;
+uint64_t kAlignSize = 8;
+
+#define ALIGN_BUFFER(buffer, align)                                 \
+  (((uintptr_t)(buffer) & ((align)-1))                              \
+       ? ((buffer) + (align) - ((uintptr_t)(buffer) & ((align)-1))) \
+       : (buffer))
+
+#define CUPTI_CALL(call)                                                   \
+  do {                                                                     \
+    CUptiResult _status = call;                                            \
+    if (_status != CUPTI_SUCCESS) {                                        \
+      const char *errstr;                                                  \
+      dynload::cuptiGetResultString(_status, &errstr);                     \
+      fprintf(stderr, "%s:%d: error: function %s failed with error %s.\n", \
+              __FILE__, __LINE__, #call, errstr);                          \
+      exit(-1);                                                            \
+    }                                                                      \
+  } while (0)
+
+void EnableActivity() {
+  // Device activity record is created when CUDA initializes, so we
+  // want to enable it before cuInit() or any CUDA runtime call.
+  CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMCPY));
+  CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_KERNEL));
+  CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DEVICE));
+  CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMSET));
+  CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_OVERHEAD));
+  // We don't track these activities for now.
+  // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_CONTEXT));
+  // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DRIVER));
+  // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_RUNTIME));
+  // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_NAME));
+  // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MARKER));
+}
+
+void DisableActivity() {
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MEMCPY));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_KERNEL));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_DEVICE));
+  // Disable all other activity record kinds.
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_CONTEXT));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_DRIVER));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_RUNTIME));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MEMSET));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_NAME));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MARKER));
+  CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_OVERHEAD));
+}
+
+void CUPTIAPI bufferRequested(uint8_t **buffer, size_t *size,
+                              size_t *maxNumRecords) {
+  uint8_t *buf = (uint8_t *)malloc(kBufSize + kAlignSize);
+  *size = kBufSize;
+  *buffer = ALIGN_BUFFER(buf, kAlignSize);
+  *maxNumRecords = 0;
+}
+
+void CUPTIAPI bufferCompleted(CUcontext ctx, uint32_t streamId, uint8_t *buffer,
+                              size_t size, size_t validSize) {
+  CUptiResult status;
+  CUpti_Activity *record = NULL;
+  if (validSize > 0) {
+    do {
+      status = dynload::cuptiActivityGetNextRecord(buffer, validSize, &record);
+      if (status == CUPTI_SUCCESS) {
+        switch (record->kind) {
+          case CUPTI_ACTIVITY_KIND_KERNEL:
+          case CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL: {
+            auto *kernel =
+                reinterpret_cast<const CUpti_ActivityKernel3 *>(record);
+            tracer->AddKernelRecords(kernel->start, kernel->end,
+                                     kernel->deviceId, kernel->streamId,
+                                     kernel->correlationId);
+            break;
+          }
+          default: { break; }
+        }
+      } else if (status == CUPTI_ERROR_MAX_LIMIT_REACHED) {
+        // Seems not an error in this case.
+        break;
+      } else {
+        CUPTI_CALL(status);
+      }
+    } while (1);
+
+    size_t dropped;
+    CUPTI_CALL(
+        dynload::cuptiActivityGetNumDroppedRecords(ctx, streamId, &dropped));
+    if (dropped != 0) {
+      fprintf(stderr, "Dropped %u activity records\n", (unsigned int)dropped);
+    }
+  }
+  free(buffer);
+}
+}  // namespace
+
+class DeviceTracerImpl : public DeviceTracer {
+ public:
+  DeviceTracerImpl() : enabled_(false) {}
+
+  void AddAnnotation(uint64_t id, const std::string &anno) {
+    std::lock_guard<std::mutex> l(trace_mu_);
+    correlations_[id] = anno;
+  }
+
+  void AddKernelRecords(uint64_t start, uint64_t end, uint32_t device_id,
+                        uint32_t stream_id, uint32_t correlation_id) {
+    std::lock_guard<std::mutex> l(trace_mu_);
+    kernel_records_.push_back(
+        KernelRecord{start, end, device_id, stream_id, correlation_id});
+  }
+
+  bool IsEnabled() {
+    std::lock_guard<std::mutex> l(trace_mu_);
+    return enabled_;
+  }
+
+  void Enable() {
+    std::lock_guard<std::mutex> l(trace_mu_);
+    if (enabled_) {
+      fprintf(stderr, "DeviceTracer already enabled\n");
+      return;
+    }
+    EnableActivity();
+
+    // Register callbacks for buffer requests and completed by CUPTI.
+    CUPTI_CALL(dynload::cuptiActivityRegisterCallbacks(bufferRequested,
+                                                       bufferCompleted));
+
+    CUptiResult ret;
+    ret = dynload::cuptiSubscribe(
+        &subscriber_, static_cast<CUpti_CallbackFunc>(ApiCallback), this);
+    if (ret == CUPTI_ERROR_MAX_LIMIT_REACHED) {
+      fprintf(stderr, "CUPTI subcriber limit reached.\n");
+    } else if (ret != CUPTI_SUCCESS) {
+      fprintf(stderr, "Failed to create CUPTI subscriber.\n");
+    }
+    CUPTI_CALL(
+        dynload::cuptiEnableCallback(1, subscriber_, CUPTI_CB_DOMAIN_DRIVER_API,
+                                     CUPTI_DRIVER_TRACE_CBID_cuLaunchKernel));
+
+    CUPTI_CALL(dynload::cuptiGetTimestamp(&start_ns_));
+    enabled_ = true;
+  }
+
+  proto::Profile GenProfile() {
+    std::lock_guard<std::mutex> l(trace_mu_);
+    proto::Profile profile_pb;
+    profile_pb.set_start_ns(start_ns_);
+    profile_pb.set_end_ns(end_ns_);
+    std::map<std::string, std::vector<uint64_t>> event_times;
+    for (const KernelRecord &r : kernel_records_) {
+      if (correlations_.find(r.correlation_id) == correlations_.end()) {
+        fprintf(stderr, "cannot relate a kernel activity\n");
+        continue;
+      }
+      auto *event = profile_pb.add_events();
+      event->set_name(correlations_.at(r.correlation_id));
+      event->set_start_ns(r.start_ns);
+      event->set_end_ns(r.end_ns);
+      event->set_stream_id(r.stream_id);
+      event->set_device_id(r.device_id);
+      event_times[event->name()].push_back(r.end_ns - r.start_ns);
+    }
+    for (const auto &et : event_times) {
+      fprintf(
+          stderr, "%s: total: %fms invoked cuda kernels: %lu\n",
+          et.first.c_str(),
+          std::accumulate(et.second.begin(), et.second.end(), 0) / 1000000.0,
+          et.second.size());
+    }
+    return profile_pb;
+  }
+
+  void Disable() {
+    // flush might cause additional calls to DeviceTracker.
+    dynload::cuptiActivityFlushAll(CUPTI_ACTIVITY_FLAG_FLUSH_FORCED);
+    std::lock_guard<std::mutex> l(trace_mu_);
+    DisableActivity();
+    dynload::cuptiUnsubscribe(subscriber_);
+    CUPTI_CALL(dynload::cuptiGetTimestamp(&end_ns_));
+    PADDLE_ENFORCE(dynload::cuptiFinalize());
+    enabled_ = false;
+  }
+
+ private:
+  static void CUPTIAPI ApiCallback(void *userdata, CUpti_CallbackDomain domain,
+                                   CUpti_CallbackId cbid, const void *cbdata) {
+    auto *cbInfo = reinterpret_cast<const CUpti_CallbackData *>(cbdata);
+    DeviceTracer *tracer = reinterpret_cast<DeviceTracer *>(userdata);
+
+    if ((domain == CUPTI_CB_DOMAIN_DRIVER_API) &&
+        (cbid == CUPTI_DRIVER_TRACE_CBID_cuLaunchKernel)) {
+      if (cbInfo->callbackSite == CUPTI_API_ENTER) {
+        const std::string anno =
+            cur_annotation ? cur_annotation : cbInfo->symbolName;
+        tracer->AddAnnotation(cbInfo->correlationId, anno);
+      }
+    } else {
+      VLOG(1) << "Unhandled API Callback for " << domain << " " << cbid;
+    }
+  }
+
+  std::mutex trace_mu_;
+  bool enabled_;
+  uint64_t start_ns_;
+  uint64_t end_ns_;
+  std::vector<KernelRecord> kernel_records_;
+  std::unordered_map<uint32_t, std::string> correlations_;
+  CUpti_SubscriberHandle subscriber_;
+};
+
+#endif  // PADDLE_WITH_CUPTI
+
+class DeviceTracerDummy : public DeviceTracer {
+ public:
+  DeviceTracerDummy() {}
+
+  void AddAnnotation(uint64_t id, const std::string &anno) {}
+
+  void AddKernelRecords(uint64_t start, uint64_t end, uint32_t device_id,
+                        uint32_t stream_id, uint32_t correlation_id) {}
+
+  bool IsEnabled() { return false; }
+
+  void Enable() {}
+
+  proto::Profile GenProfile() { return proto::Profile(); }
+
+  void Disable() {}
+};
+
+void CreateTracer(DeviceTracer **t) {
+#ifdef PADDLE_WITH_CUPTI
+  *t = new DeviceTracerImpl();
+#else
+  *t = new DeviceTracerDummy();
+#endif  // PADDLE_WITH_CUPTI
+}
+
+DeviceTracer *GetDeviceTracer() {
+  std::call_once(tracer_once_flag, CreateTracer, &tracer);
+  return tracer;
+}
+
+void SetCurAnnotation(const char *anno) { cur_annotation = anno; }
+
+void ClearCurAnnotation() { cur_annotation = nullptr; }
+
+}  // namespace platform
+}  // namespace paddle

paddle/fluid/platform/device_tracer.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/fluid/platform/dynload/cupti.h"
+#include "paddle/fluid/platform/profiler.pb.h"
+
+namespace paddle {
+namespace platform {
+
+///////////////////////
+// WARN: Under Development. Don't depend on it yet.
+//////////////////////
+
+// DeviceTracer performs the following tasks:
+// 1. Register cuda callbacks for various events: kernel, memcpy, etc.
+// 2. Collect cuda statistics: start/end ts, memory, etc.
+// 3. Generate a protobuf for further analysis.
+class DeviceTracer {
+ public:
+  struct KernelRecord {
+    uint64_t start_ns;
+    uint64_t end_ns;
+    uint32_t device_id;
+    uint32_t stream_id;
+    uint32_t correlation_id;
+  };
+
+  virtual ~DeviceTracer() {}
+  // Needs to be called once before use.
+  virtual void Enable() = 0;
+  // Needs to be called once after use.
+  virtual void Disable() = 0;
+
+  // Add a pair to correlate internal cuda id with high level
+  // annotation (string). So cuda statistics can be represented by
+  // human-readable annotations.
+  virtual void AddAnnotation(uint64_t id, const std::string& anno) = 0;
+
+  // Add a cuda kernel stats. `correlation_id` will be mapped to annotation
+  // added before for human readability.
+  virtual void AddKernelRecords(uint64_t start, uint64_t end,
+                                uint32_t device_id, uint32_t stream_id,
+                                uint32_t correlation_id) = 0;
+
+  // Generate a proto after done (Disabled).
+  virtual proto::Profile GenProfile() = 0;
+
+  virtual bool IsEnabled() = 0;
+};
+
+// Get a DeviceTracer.
+DeviceTracer* GetDeviceTracer();
+
+// Set a name for the cuda kernel operation being launched by the thread.
+void SetCurAnnotation(const char* anno);
+// Clear the name after the operation is done.
+void ClearCurAnnotation();
+
+}  // namespace platform
+}  // namespace paddle

paddle/fluid/platform/dynload/CMakeLists.txt

 cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags enforce)
-nv_library(dynload_cuda SRCS cublas.cc cudnn.cc curand.cc nccl.cc
-        DEPS dynamic_loader)
+
+list(APPEND CUDA_SRCS cublas.cc cudnn.cc curand.cc nccl.cc)
+if (CUPTI_FOUND)
+    list(APPEND CUDA_SRCS cupti.cc)
+endif(CUPTI_FOUND)
+nv_library(dynload_cuda SRCS ${CUDA_SRCS} DEPS dynamic_loader)
 cc_library(dynload_warpctc SRCS warpctc.cc DEPS dynamic_loader warpctc)

paddle/fluid/platform/dynload/cupti.cc

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#ifdef PADDLE_WITH_CUPTI
+
+#include "paddle/fluid/platform/dynload/cupti.h"
+#include "paddle/fluid/platform/enforce.h"
+
+namespace paddle {
+namespace platform {
+namespace dynload {
+
+std::once_flag cupti_dso_flag;
+void *cupti_dso_handle = nullptr;
+
+#define DEFINE_WRAP(__name) DynLoad__##__name __name
+
+CUPTI_ROUTINE_EACH(DEFINE_WRAP);
+
+}  // namespace dynload
+}  // namespace platform
+}  // namespace paddle
+
+#endif  // PADDLE_WITH_CUPTI

paddle/fluid/platform/dynload/cupti.h

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#ifdef PADDLE_WITH_CUPTI
+#include <cuda.h>
+#include <cupti.h>
+#include <dlfcn.h>
+#include <mutex>
+#include "paddle/fluid/platform/dynload/dynamic_loader.h"
+
+namespace paddle {
+namespace platform {
+namespace dynload {
+
+extern std::once_flag cupti_dso_flag;
+extern void *cupti_dso_handle;
+
+/**
+ * The following macro definition can generate structs
+ * (for each function) to dynamic load cupti routine
+ * via operator overloading.
+ *
+ * note: default dynamic linked libs
+ */
+#ifdef PADDLE_USE_DSO
+#define DECLARE_DYNAMIC_LOAD_CUPTI_WRAP(__name)                    \
+  struct DynLoad__##__name {                                       \
+    template <typename... Args>                                    \
+    inline CUptiResult CUPTIAPI operator()(Args... args) {         \
+      typedef CUptiResult CUPTIAPI (*cuptiFunc)(Args...);          \
+      std::call_once(cupti_dso_flag,                               \
+                     paddle::platform::dynload::GetCUPTIDsoHandle, \
+                     &cupti_dso_handle);                           \
+      void *p_##__name = dlsym(cupti_dso_handle, #__name);         \
+      return reinterpret_cast<cuptiFunc>(p_##__name)(args...);     \
+    }                                                              \
+  };                                                               \
+  extern DynLoad__##__name __name
+#else
+#define DECLARE_DYNAMIC_LOAD_CUPTI_WRAP(__name)            \
+  struct DynLoad__##__name {                               \
+    template <typename... Args>                            \
+    inline CUptiResult CUPTIAPI operator()(Args... args) { \
+      return __name(args...);                              \
+    }                                                      \
+  };                                                       \
+  extern DynLoad__##__name __name
+#endif
+
+#define CUPTI_ROUTINE_EACH(__macro)           \
+  __macro(cuptiActivityEnable);               \
+  __macro(cuptiActivityDisable);              \
+  __macro(cuptiActivityRegisterCallbacks);    \
+  __macro(cuptiActivityGetAttribute);         \
+  __macro(cuptiActivitySetAttribute);         \
+  __macro(cuptiGetTimestamp);                 \
+  __macro(cuptiActivityGetNextRecord);        \
+  __macro(cuptiGetResultString);              \
+  __macro(cuptiActivityGetNumDroppedRecords); \
+  __macro(cuptiActivityFlushAll);             \
+  __macro(cuptiFinalize);                     \
+  __macro(cuptiSubscribe);                    \
+  __macro(cuptiUnsubscribe);                  \
+  __macro(cuptiEnableCallback);
+
+CUPTI_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUPTI_WRAP);
+
+#undef DECLARE_DYNAMIC_LOAD_CUPTI_WRAP
+}  // namespace dynload
+}  // namespace platform
+}  // namespace paddle
+
+#endif  // PADDLE_WITH_CUPTI

paddle/fluid/platform/dynload/dynamic_loader.cc

@@ -40,10 +40,14 @@ DEFINE_string(nccl_dir, "",
               "libcurand. For instance, /usr/local/cuda/lib64. If default, "
               "dlopen will search cuda from LD_LIBRARY_PATH");
 
+DEFINE_string(cupti_dir, "", "Specify path for loading cupti.so.");
+
 namespace paddle {
 namespace platform {
 namespace dynload {
 
+static const char* cupti_lib_path = CUPTI_LIB_PATH;
+
 static inline std::string join(const std::string& part1,
                                const std::string& part2) {
   // directory separator
@@ -143,6 +147,18 @@ void GetCUDNNDsoHandle(void** dso_handle) {
 #endif
 }
 
+void GetCUPTIDsoHandle(void** dso_handle) {
+  std::string cupti_path = cupti_lib_path;
+  if (!FLAGS_cupti_dir.empty()) {
+    cupti_path = FLAGS_cupti_dir;
+  }
+#if defined(__APPLE__) || defined(__OSX__)
+  GetDsoHandleFromSearchPath(cupti_path, "libcupti.dylib", dso_handle, false);
+#else
+  GetDsoHandleFromSearchPath(cupti_path, "libcupti.so", dso_handle, false);
+#endif
+}
+
 void GetCurandDsoHandle(void** dso_handle) {
 #if defined(__APPLE__) || defined(__OSX__)
   GetDsoHandleFromSearchPath(FLAGS_cuda_dir, "libcurand.dylib", dso_handle);

paddle/fluid/platform/dynload/dynamic_loader.h

@@ -34,6 +34,8 @@ void GetCublasDsoHandle(void** dso_handle);
  */
 void GetCUDNNDsoHandle(void** dso_handle);
 
+void GetCUPTIDsoHandle(void** dso_handle);
+
 /**
  * @brief    load the DSO of CURAND
  *

paddle/fluid/platform/profiler.cc

@@ -15,7 +15,13 @@ limitations under the License. */
 #include "paddle/fluid/platform/profiler.h"
 #include <iomanip>
 #include <map>
+#ifdef PADDLE_WITH_CUDA
+#include <cuda.h>
+#endif  // PADDLE_WITH_CUDA
 #include "glog/logging.h"
+#include "paddle/fluid/framework/block_desc.h"
+#include "paddle/fluid/platform/device_tracer.h"
+#include "paddle/fluid/string/printf.h"
 
 namespace paddle {
 namespace platform {
@@ -126,15 +132,20 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx) {
   GetEventList().Record(EventKind::kPopRange, name, g_thread_id, dev_ctx);
 }
 
-RecordEvent::RecordEvent(const std::string& name,
-                         const DeviceContext* dev_ctx) {
+RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx,
+                         int32_t block_id) {
   if (g_state == ProfilerState::kDisabled) return;
   dev_ctx_ = dev_ctx;
   name_ = name;
   PushEvent(name_, dev_ctx_);
+
+  full_name_ = string::Sprintf("%s_b%d", name, block_id);
+  // Maybe need the same push/pop behavior.
+  SetCurAnnotation(full_name_.c_str());
 }
 
 RecordEvent::~RecordEvent() {
+  ClearCurAnnotation();
   if (g_state == ProfilerState::kDisabled) return;
   PopEvent(name_, dev_ctx_);
 }
@@ -147,7 +158,14 @@ void EnableProfiler(ProfilerState state) {
                  "The profiling state should be disabled when calling ",
                  "EnableProfiler.");
   g_state = state;
-  g_profiler_place = (g_state == ProfilerState::kCUDA) ? "CUDA" : "CPU";
+  if (g_state == ProfilerState::kCUDA) {
+    g_profiler_place = "CUDA";
+  } else if (g_state == ProfilerState::kCPU) {
+    g_profiler_place = "CPU";
+  } else {
+    g_profiler_place = "All";
+    GetDeviceTracer()->Enable();
+  }
 #ifdef PADDLE_WITH_CUDA
   if (g_state == ProfilerState::kCUDA) {
     // Generate some dummy evenets first to reduce the startup overhead.
@@ -190,6 +208,12 @@ void DisableProfiler(EventSortingKey sorted_key) {
   Mark("_stop_profiler_", nullptr);
   g_state = ProfilerState::kDisabled;
 
+  DeviceTracer* tracer = GetDeviceTracer();
+  if (g_profiler_place == "All" && tracer && tracer->IsEnabled()) {
+    tracer->Disable();
+    tracer->GenProfile();
+  }
+
   std::vector<std::vector<Event>> all_events = GetAllEvents();
   ParseEvents(all_events, sorted_key);
   ResetProfiler();
@@ -254,9 +278,11 @@ void ParseEvents(std::vector<std::vector<Event>>& events,
         }
 
         if (rit != pushed_events.rend()) {
-          double event_time = (g_profiler_place == "CUDA")
+          double event_time =
+              (g_profiler_place == "CUDA" || g_profiler_place == "All")
                   ? rit->CudaElapsedMs(events[i][j])
                   : rit->CpuElapsedMs(events[i][j]);
+
           std::string event_name =
               "thread" + std::to_string(rit->thread_id()) + "::" + rit->name();
           max_name_width = std::max(max_name_width, event_name.size());

paddle/fluid/platform/profiler.h

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <mutex>
 #include <vector>
 #include "paddle/fluid/platform/device_context.h"
+#include "paddle/fluid/platform/profiler.pb.h"
 
 namespace paddle {
 namespace platform {
@@ -93,6 +94,7 @@ enum ProfilerState {
   kDisabled,  // disabled state
   kCPU,       // CPU profiling state
   kCUDA,      // GPU profiling state
+  kAll,       // Profile both CPU and GPU. (Currently experimental).
 };
 
 void Mark(const std::string& name, const DeviceContext* dev_ctx);
@@ -102,7 +104,8 @@ void PushEvent(const std::string& name, const DeviceContext* dev_ctx);
 void PopEvent(const std::string& name, const DeviceContext* dev_ctx);
 
 struct RecordEvent {
-  explicit RecordEvent(const std::string& name, const DeviceContext* dev_ctx);
+  RecordEvent(const std::string& name, const DeviceContext* dev_ctx,
+              int32_t block_id);
 
   ~RecordEvent();
 
@@ -110,9 +113,12 @@ struct RecordEvent {
   const DeviceContext* dev_ctx_;
   // Event name
   std::string name_;
+  // Need to distinguish name by op type, block_id, program_id and perhaps
+  // different kernel invocations within an op.
+  std::string full_name_;
 };
 
-// Return the event list of all threads. Asummed the returned value calls
+// Return the event list of all threads. Assumed the returned value calls
 // event_lists, event_lists[i][j] represents the j-th Event of i-th thread.
 std::vector<std::vector<Event>> GetAllEvents();
 

paddle/fluid/platform/profiler.proto

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+syntax = "proto2";
+package paddle.platform.proto;
+
+message Event {
+  optional string name = 1;
+  optional uint64 start_ns = 2;
+  optional uint64 end_ns = 3;
+  optional uint32 device_id = 5;
+  optional uint32 stream_id = 6;
+}
+
+message Profile {
+  repeated Event events = 1;
+  optional uint64 start_ns = 2;
+  optional uint64 end_ns = 3;
+}
\ No newline at end of file

paddle/fluid/platform/profiler_test.cc

@@ -95,7 +95,7 @@ TEST(RecordEvent, RecordEvent) {
    */
   for (int i = 1; i < 5; ++i) {
     std::string name = "evs_op_" + std::to_string(i);
-    RecordEvent record_event(name, dev_ctx);
+    RecordEvent record_event(name, dev_ctx, 0);
     int counter = 1;
     while (counter != i * 1000) counter++;
   }

paddle/fluid/pybind/pybind.cc

@@ -459,6 +459,7 @@ All parameter, weight, gradient are variables in Paddle.
       .value("kDisabled", platform::ProfilerState::kDisabled)
       .value("kCPU", platform::ProfilerState::kCPU)
       .value("kCUDA", platform::ProfilerState::kCUDA)
+      .value("kAll", platform::ProfilerState::kAll)
       .export_values();
 
   py::enum_<platform::EventSortingKey>(m, "EventSortingKey", py::arithmetic())

python/paddle/fluid/profiler.py

@@ -97,9 +97,14 @@ def profiler(state, sorted_key=None):
             The `ave` means sorting by the average execution time.
     """
 
-    if state not in ['CPU', 'GPU']:
-        raise ValueError("The state must be 'CPU' or 'GPU'.")
-    prof_state = core.ProfilerState.kCUDA if state == "GPU" else core.ProfilerState.kCPU
+    if state not in ['CPU', 'GPU', "All"]:
+        raise ValueError("The state must be 'CPU' or 'GPU' or 'All'.")
+    if state == "GPU":
+        prof_state = core.ProfilerState.kCUDA
+    elif state == "CPU":
+        prof_state = core.ProfilerState.kCPU
+    else:
+        prof_state = core.ProfilerState.kAll
     core.enable_profiler(prof_state)
     yield
 

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -41,6 +41,7 @@ list(REMOVE_ITEM TEST_OPS test_while_op)
 list(REMOVE_ITEM TEST_OPS test_lod_array_length_op)
 list(REMOVE_ITEM TEST_OPS test_reorder_lod_tensor)
 list(REMOVE_ITEM TEST_OPS test_profiler)
+list(REMOVE_ITEM TEST_OPS test_nvprof)
 list(REMOVE_ITEM TEST_OPS test_normalization_wrapper)
 list(REMOVE_ITEM TEST_OPS test_executor_and_mul)
 list(REMOVE_ITEM TEST_OPS test_assign_value_op)
@@ -75,6 +76,7 @@ py_test_modules(test_while_op MODULES test_while_op)
 py_test_modules(test_lod_array_length_op MODULES test_lod_array_length_op)
 py_test_modules(test_reorder_lod_tensor MODULES test_reorder_lod_tensor)
 py_test_modules(test_profiler MODULES test_profiler)
+py_test_modules(test_nvprof MODULES test_nvprof)
 py_test_modules(test_normalization_wrapper MODULES test_normalization_wrapper)
 py_test_modules(test_executor_and_mul MODULES test_executor_and_mul)
 py_test_modules(test_assign_value_op MODULES test_assign_value_op)

python/paddle/fluid/tests/unittests/test_nvprof.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+import os
+import numpy as np
+import paddle.fluid as fluid
+import paddle.fluid.profiler as profiler
+import paddle.fluid.layers as layers
+import paddle.fluid.core as core
+
+
+class TestNVProf(unittest.TestCase):
+    def test_nvprof(self):
+        if not fluid.core.is_compiled_with_cuda():
+            return
+        epoc = 8
+        dshape = [4, 3, 28, 28]
+        data = layers.data(name='data', shape=[3, 28, 28], dtype='float32')
+        conv = layers.conv2d(data, 20, 3, stride=[1, 1], padding=[1, 1])
+
+        place = fluid.CUDAPlace(0)
+        exe = fluid.Executor(place)
+        exe.run(fluid.default_startup_program())
+
+        output_file = 'cuda_profiler.txt'
+        with profiler.cuda_profiler(output_file, 'csv') as nvprof:
+            for i in range(epoc):
+                input = np.random.random(dshape).astype('float32')
+                exe.run(fluid.default_main_program(), feed={'data': input})
+        os.remove(output_file)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_profiler.py

@@ -22,27 +22,9 @@ import paddle.fluid.core as core
 
 
 class TestProfiler(unittest.TestCase):
-    def test_nvprof(self):
-        if not fluid.core.is_compiled_with_cuda():
-            return
-        epoc = 8
-        dshape = [4, 3, 28, 28]
-        data = layers.data(name='data', shape=[3, 28, 28], dtype='float32')
-        conv = layers.conv2d(data, 20, 3, stride=[1, 1], padding=[1, 1])
-
-        place = fluid.CUDAPlace(0)
-        exe = fluid.Executor(place)
-        exe.run(fluid.default_startup_program())
-
-        output_file = 'cuda_profiler.txt'
-        with profiler.cuda_profiler(output_file, 'csv') as nvprof:
-            for i in range(epoc):
-                input = np.random.random(dshape).astype('float32')
-                exe.run(fluid.default_main_program(), feed={'data': input})
-        os.remove(output_file)
-
     def net_profiler(self, state):
-        if state == 'GPU' and not core.is_compiled_with_cuda():
+        enable_if_gpu = state == 'GPU' or state == "All"
+        if enable_if_gpu and not core.is_compiled_with_cuda():
             return
         startup_program = fluid.Program()
         main_program = fluid.Program()
@@ -85,6 +67,9 @@ class TestProfiler(unittest.TestCase):
     def test_cuda_profiler(self):
         self.net_profiler('GPU')
 
+    def test_all_profiler(self):
+        self.net_profiler('All')
+
 
 if __name__ == '__main__':
     unittest.main()