Merge pull request BVLC#82 from borisfom/caffe-0.14

Removing CNMEM switches

CMakeLists.txt

@@ -22,7 +22,6 @@ include(cmake/ConfigGen.cmake)
 # ---[ Options
 caffe_option(CPU_ONLY  "Build Caffe without CUDA support" OFF) # TODO: rename to USE_CUDA
 caffe_option(USE_CUDNN "Build Caffe with cuDNN library support" ON IF NOT CPU_ONLY)
-caffe_option(USE_CNMEM "Build Caffe with CNMeM memory pool support" OFF)
 caffe_option(BUILD_SHARED_LIBS "Build shared libraries" ON)
 caffe_option(BUILD_python "Build Python wrapper" ON)
 set(python_version "2" CACHE STRING "Specify which Python version to use")
@@ -42,11 +41,6 @@ if(UNIX OR APPLE)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fPIC -Wall")
 endif()
 
-if(USE_CNMEM)
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DUSE_CNMEM -lcnmem")
-  message("-- Info: Using CNMEM memory pool configuration") 
-endif()
-
 if(USE_libstdcpp)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -stdlib=libstdc++")
   message("-- Warning: forcing libstdc++ (controlled by USE_libstdcpp option in cmake)")

Makefile

@@ -325,12 +325,6 @@ ifeq ($(USE_CUDNN), 1)
 	COMMON_FLAGS += -DUSE_CUDNN
 endif
 
-# CNMEM integration
-ifeq ($(USE_CNMEM), 1)
-	LIBRARIES += cnmem
-	COMMON_FLAGS += -DUSE_CNMEM
-endif
-
 # configure IO libraries
 ifeq ($(USE_OPENCV), 1)
 	COMMON_FLAGS += -DUSE_OPENCV

Makefile.config.example

@@ -4,9 +4,6 @@
 # cuDNN acceleration switch (uncomment to build with cuDNN).
 # USE_CUDNN := 1
 
-# CNMeM memory pool switch (experimental, may go away)
-# USE_CNMEM := 1
-
 # CPU-only switch (uncomment to build without GPU support).
 # CPU_ONLY := 1
 

cmake/ConfigGen.cmake

@@ -77,12 +77,6 @@ function(caffe_generate_export_configs)
     list(APPEND DEFINITIONS -DUSE_CUDNN)
   endif()
 
-  if(NOT HAVE_CNMEM)
-    set(HAVE_CNMEM FALSE)
-  else()
-    list(APPEND DEFINITIONS -DUSE_CNMEM)
-  endif()
-
   if(BLAS STREQUAL "MKL" OR BLAS STREQUAL "mkl")
     list(APPEND Caffe_DEFINITIONS -DUSE_MKL)
   endif()

cmake/Cuda.cmake

@@ -188,33 +188,6 @@ function(detect_cuDNN)
   endif()
 endfunction()
 
-################################################################################################
-# Short command for cuDNN detection. Believe it soon will be a part of CUDA toolkit distribution.
-# That's why not FindcuDNN.cmake file, but just the macro
-# Usage:
-#   detect_cuDNN()
-function(detect_CNMeM)
-  set(CNMEM_ROOT "" CACHE PATH "CNMeM root folder")
-
-  find_path(CNMEM_INCLUDE cnmem.h
-    PATHS ${CNMEM_ROOT} $ENV{CNMEM_ROOT} ${CUDA_TOOLKIT_INCLUDE}
-    DOC "Path to CNMeM include directory." )
-
-  get_filename_component(__libpath_hist ${CUDA_CUDART_LIBRARY} PATH)
-  find_library(CNMEM_LIBRARY NAMES libcnmem.so # libcudnn_static.a
-                             PATHS ${CNMEM_ROOT} $ENV{CNMEM_ROOT} ${CNMEM_INCLUDE} ${__libpath_hist}
-                             DOC "Path to CNMeM library.")
-
-  if(CNMEM_INCLUDE AND CNMEM_LIBRARY)
-    set(HAVE_CNMEM  TRUE PARENT_SCOPE)
-    set(CNMEM_FOUND TRUE PARENT_SCOPE)
-
-    mark_as_advanced(CNMEM_INCLUDE CNMEM_LIBRARY CNMEM_ROOT)
-    message(STATUS "Found CNMeM (include: ${CNMEM_INCLUDE}, library: ${CNMEM_LIBRARY})")
-  endif()
-endfunction()
-
-
 ################################################################################################
 ###  Non macro section
 ################################################################################################
@@ -242,15 +215,6 @@ if(USE_CUDNN)
   endif()
 endif()
 
-if(USE_CNMEM)
-  detect_CNMeM()
-  if(HAVE_CNMEM)
-    add_definitions(-DUSE_CNMEM)
-    include_directories(SYSTEM ${CNMEM_INCLUDE})
-    list(APPEND Caffe_LINKER_LIBS ${CNMEM_LIBRARY})
-  endif()
-endif()
-
 # setting nvcc arch flags
 caffe_select_nvcc_arch_flags(NVCC_FLAGS_EXTRA)
 list(APPEND CUDA_NVCC_FLAGS ${NVCC_FLAGS_EXTRA})

cmake/Summary.cmake

@@ -146,11 +146,6 @@ function(caffe_print_configuration_summary)
     else()
       caffe_status("  cuDNN             :   Disabled")
     endif()
-    if (USE_CNMEM)
-      caffe_status("  CNMeM             : " HAVE_CNMEM THEN "Yes" ELSE "Not found")
-    else()
-      caffe_status("  CNMeM             :   Disabled")
-    endif()
     caffe_status("")
   endif()
   if(HAVE_PYTHON)

cmake/Templates/CaffeConfig.cmake.in

@@ -15,7 +15,6 @@
 #
 #   Caffe_HAVE_CUDA    - signals about CUDA support
 #   Caffe_HAVE_CUDNN   - signals about cuDNN support
-#   Caffe_HAVE_CNMEM   - signals about CNMeM support
 # 
 #
 #
@@ -60,4 +59,4 @@ set(Caffe_DEFINITIONS "@Caffe_DEFINITIONS@")
 set(Caffe_CPU_ONLY @CPU_ONLY@)
 set(Caffe_HAVE_CUDA @HAVE_CUDA@)
 set(Caffe_HAVE_CUDNN @HAVE_CUDNN@)
-set(Caffe_HAVE_CNMEM @HAVE_CNMEM@)
+

cmake/Templates/caffe_config.h.in

@@ -10,10 +10,6 @@
 /* NVIDA cuDNN */
 #cmakedefine HAVE_CUDNN
 #cmakedefine USE_CUDNN
-
-/* NVIDIA CNMeM */
-#cmakedefine HAVE_CNMEM
-#cmakedefine USE_CNMEM
 
 /* NVIDA cuDNN */
 #cmakedefine CPU_ONLY

include/caffe/util/device_alternate.hpp

@@ -66,18 +66,6 @@ void classname<Dtype>::funcname##_##gpu(const vector<Blob<Dtype>*>& top, \
       << caffe::curandGetErrorString(status); \
   } while (0)
 
-#ifdef USE_CNMEM
-
-#define CNMEM_CHECK(condition) \
-  do { \
-    cnmemStatus_t status = condition; \
-    CHECK_EQ(status, CNMEM_STATUS_SUCCESS) << " " \
-      << cnmemGetErrorString(status); \
-  } while (0)
-#else
-#define CNMEM_CHECK(condition)
-#endif
-
 // CUDA: grid stride looping
 #define CUDA_KERNEL_LOOP(i, n) \
   for (int i = blockIdx.x * blockDim.x + threadIdx.x; \

include/caffe/util/gpu_memory.hpp

@@ -10,16 +10,11 @@ class gpu_memory {
  public:
   enum PoolMode {
     NoPool,     // Straight CUDA malllc/free. May be very expensive
-    CnMemPool,  // CNMEM arena allocator
     CubPool,     // CUB caching allocator
 #ifdef CPU_ONLY
     DefaultPool = NoPool
 #else
-#  if (USE_CNMEM) && !defined (__arm__)
-    DefaultPool = CnMemPool   // CNMEM pool only uses dedicated video memory.
-# else
     DefaultPool = CubPool     // CUB pool is able to use unified memory properly
-# endif
 #endif
   };
 

src/caffe/util/gpu_memory.cpp

@@ -1,15 +1,9 @@
 #include <algorithm>
 #include <vector>
 #include "caffe/common.hpp"
-
 #include "caffe/util/gpu_memory.hpp"
 
 
-#ifdef USE_CNMEM
-// CNMEM integration
-#include "cnmem.h"
-#endif
-
 #ifndef CPU_ONLY
 #include "cub/cub/util_allocator.cuh"
 #endif
@@ -42,7 +36,6 @@ namespace caffe {
 
     switch (m) {
     case CubPool:
-    case CnMemPool:
       initMEM(gpus, m);
       break;
     default:
@@ -56,9 +49,6 @@ namespace caffe {
 
   void gpu_memory::destroy() {
     switch (mode_) {
-    case CnMemPool:
-      CNMEM_CHECK(cnmemFinalize());
-      break;
     case CubPool:
       delete cubAlloc;
       cubAlloc = NULL;
@@ -73,9 +63,6 @@ namespace caffe {
   void gpu_memory::allocate(void **ptr, size_t size, cudaStream_t stream) {
     CHECK((ptr) != NULL);
     switch (mode_) {
-    case CnMemPool:
-      CNMEM_CHECK(cnmemMalloc(ptr, size, stream));
-      break;
     case CubPool:
       CUDA_CHECK(cubAlloc->DeviceAllocate(ptr, size, stream));
       break;
@@ -90,9 +77,6 @@ namespace caffe {
     if (!ptr)
       return;
     switch (mode_) {
-    case CnMemPool:
-      CNMEM_CHECK(cnmemFree(ptr, stream));
-      break;
     case CubPool:
       CUDA_CHECK(cubAlloc->DeviceFree(ptr));
       break;
@@ -104,9 +88,6 @@ namespace caffe {
 
   void gpu_memory::registerStream(cudaStream_t stream) {
     switch (mode_) {
-    case CnMemPool:
-      CNMEM_CHECK(cnmemRegisterStream(stream));
-      break;
     case CubPool:
     default:
       break;
@@ -116,9 +97,6 @@ namespace caffe {
   void gpu_memory::initMEM(const std::vector<int>& gpus, PoolMode m) {
     mode_ = m;
     int initial_device;
-#if USE_CNMEM
-    cnmemDevice_t* devs = new cnmemDevice_t[gpus.size()];
-#endif
 
     CUDA_CHECK(cudaGetDevice(&initial_device));
 
@@ -142,21 +120,10 @@ namespace caffe {
       // find out the smallest GPU size
       if (poolsize_ == 0 || poolsize_ > free_mem)
         poolsize_ = free_mem;
-#if USE_CNMEM
-      devs[i].device = gpus[i];
-      devs[i].size = free_mem;
-      devs[i].numStreams = 0;
-      devs[i].streams = NULL;
-#endif
     }
 
 
     switch ( mode_ ) {
-      case CnMemPool:
-#if USE_CNMEM
-        CNMEM_CHECK(cnmemInit(gpus.size(), devs, CNMEM_FLAGS_DEFAULT));
-#endif
-        break;
       case CubPool:
         try {
           // if you are paranoid, that doesn't mean they are not after you :)
@@ -177,15 +144,10 @@ namespace caffe {
       }
 
     CUDA_CHECK(cudaSetDevice(initial_device));
-#if USE_CNMEM
-    delete [] devs;
-#endif
   }
 
   const char* gpu_memory::getPoolName()  {
     switch (mode_) {
-    case CnMemPool:
-      return "CNMEM Pool";
     case CubPool:
       return "CUB Pool";
     default:
@@ -195,9 +157,6 @@ namespace caffe {
 
   void gpu_memory::getInfo(size_t *free_mem, size_t *total_mem) {
     switch (mode_) {
-    case CnMemPool:
-      CNMEM_CHECK(cnmemMemGetInfo(free_mem, total_mem, cudaStreamDefault));
-      break;
     case CubPool:
       int cur_device;
       CUDA_CHECK(cudaGetDevice(&cur_device));