Compile nntrainer on windows

meson.build

@@ -134,6 +134,11 @@ if get_option('enable-fp16')
    endif  
 endif
 
+if get_option('enable-mmap')
+  message ('MMAP enabled')
+  extra_defines += '-DUSE_MMAP=1'
+endif
+
 if get_option('enable-opencl')
   message ('OpenCL build is enabled. Will work only if OpenCL supported GPU is available.')
   extra_defines += '-DENABLE_OPENCL=1'

meson_options.txt

@@ -17,6 +17,7 @@ option('enable-memory-swap', type: 'boolean', value: false)
 option('memory-swap-path', type: 'string', value: '')
 option('test-timeout', type: 'integer', value: 60)
 option('opencl-kernel-path', type: 'string', value: 'nntrainer_opencl_kernels')
+option('enable-mmap', type: 'boolean', value: true)
 
 # dependency conflict resolution
 option('capi-ml-inference-actual', type: 'string', value: 'capi-ml-inference',

nntrainer/app_context.cpp

@@ -222,14 +222,6 @@ const std::string getFullPath(const std::string &path,
 
 std::mutex factory_mutex;
 
-/**
- * @brief finalize global context
- *
- */
-static void fini_global_context_nntrainer(void) __attribute__((destructor));
-
-static void fini_global_context_nntrainer(void) {}
-
 std::once_flag global_app_context_init_flag;
 
 static void add_default_object(AppContext &ac) {

nntrainer/compiler/remap_realizer.h

@@ -14,6 +14,7 @@
 
 #include <functional>
 #include <memory>
+#include <string>
 #include <vector>
 
 #include <realizer.h>

nntrainer/dataset/databuffer.cpp

@@ -33,6 +33,7 @@
 #include <nntrainer_error.h>
 #include <nntrainer_log.h>
 #include <node_exporter.h>
+#include <numeric>
 #include <sstream>
 #include <stdexcept>
 #include <stdio.h>
@@ -61,9 +62,7 @@ class PropsBufferSize : public nntrainer::PositiveIntegerProperty {
 constexpr char USER_DATA[] = "user_data";
 
 DataBuffer::DataBuffer(std::unique_ptr<DataProducer> &&producer_) :
-  producer(std::move(producer_)),
-  db_props(new Props()),
-  user_data(nullptr) {
+  producer(std::move(producer_)), db_props(new Props()), user_data(nullptr) {
   rng.seed(0);
 }
 

nntrainer/dataset/dir_data_producers.cpp

@@ -104,7 +104,7 @@ DirDataProducer::finalize(const std::vector<TensorDim> &input_dims,
   const auto &dir_path = std::get<props::DirPath>(*dir_data_props).get();
 
   for (const auto &entry : std::filesystem::directory_iterator(dir_path))
-    class_names.push_back(entry.path());
+    class_names.push_back(entry.path().string());
 
   num_class = class_names.size();
 

nntrainer/layers/acti_func.h

@@ -19,6 +19,11 @@
 #include <blas_interface.h>
 #include <common_properties.h>
 
+#if defined(_WIN32)
+#define _USE_MATH_DEFINES
+#include <math.h>
+#endif
+
 namespace nntrainer {
 
 class Tensor;

nntrainer/layers/dropout.h

@@ -29,8 +29,8 @@ class DropOutLayer : public Layer {
   /**
    * @brief     Constructor of DropOut Layer
    */
-  DropOutLayer(float dropout = 0.0) :
-    Layer(), dropout_rate(props::DropOutRate(dropout)), epsilon(1e-3) {}
+  DropOutLayer(float dropout = 0.0f) :
+    Layer(), dropout_rate(props::DropOutRate(dropout)), epsilon(1e-3f) {}
 
   /**
    * @brief     Destructor of DropOut Layer

nntrainer/layers/rnn.cpp

@@ -44,7 +44,7 @@ RNNLayer::RNNLayer() :
     props::Unit(), props::HiddenStateActivation() = ActivationType::ACT_TANH,
     props::ReturnSequences(), props::DropOutRate(), props::IntegrateBias()),
   acti_func(ActivationType::ACT_NONE, true),
-  epsilon(1e-3) {
+  epsilon(1e-3f) {
   wt_idx.fill(std::numeric_limits<unsigned>::max());
 }
 

nntrainer/layers/rnncell.cpp

@@ -45,7 +45,7 @@ RNNCellLayer::RNNCellLayer() :
                 props::HiddenStateActivation() = ActivationType::ACT_TANH,
                 props::DropOutRate()),
   acti_func(ActivationType::ACT_NONE, true),
-  epsilon(1e-3) {
+  epsilon(1e-3f) {
   wt_idx.fill(std::numeric_limits<unsigned>::max());
 }
 

nntrainer/models/dynamic_training_optimization.cpp

@@ -11,6 +11,7 @@
  *
  */
 
+#include <numeric>
 #include <random>
 #include <vector>
 

nntrainer/nntrainer_log.h

@@ -56,27 +56,27 @@
 #include <nntrainer_logger.h>
 
 #if !defined(ml_logi)
-#define ml_logi(format, args...)                                            \
+#define ml_logi(format, ...)                                                \
   __nntrainer_log_print(NNTRAINER_LOG_INFO, "(%s:%s:%d) " format, __FILE__, \
-                        __func__, __LINE__, ##args)
+                        __func__, __LINE__, ##__VA_ARGS__)
 #endif
 
 #if !defined(ml_logw)
-#define ml_logw(format, args...)                                            \
+#define ml_logw(format, ...)                                                \
   __nntrainer_log_print(NNTRAINER_LOG_WARN, "(%s:%s:%d) " format, __FILE__, \
-                        __func__, __LINE__, ##args)
+                        __func__, __LINE__, ##__VA_ARGS__)
 #endif
 
 #if !defined(ml_loge)
-#define ml_loge(format, args...)                                             \
+#define ml_loge(format, ...)                                                 \
   __nntrainer_log_print(NNTRAINER_LOG_ERROR, "(%s:%s:%d) " format, __FILE__, \
-                        __func__, __LINE__, ##args)
+                        __func__, __LINE__, ##__VA_ARGS__)
 #endif
 
 #if !defined(ml_logd)
-#define ml_logd(format, args...)                                             \
+#define ml_logd(format, ...)                                                 \
   __nntrainer_log_print(NNTRAINER_LOG_DEBUG, "(%s:%s:%d) " format, __FILE__, \
-                        __func__, __LINE__, ##args)
+                        __func__, __LINE__, ##__VA_ARGS__)
 #endif
 
 #endif

nntrainer/nntrainer_logger.cpp

@@ -30,7 +30,6 @@
 #include <sstream>
 #include <stdarg.h>
 #include <stdexcept>
-#include <unistd.h>
 #include <util_func.h>
 
 namespace nntrainer {
@@ -78,7 +77,9 @@ Logger::Logger() : ts_type(NNTRAINER_LOG_TIMESTAMP_SEC) {
      << std::setw(2) << now.tm_sec << ".out";
   outputstream.open(ss.str(), std::ios_base::app);
   if (!outputstream.good()) {
-    char buf[256] = {0,};
+    char buf[256] = {
+      0,
+    };
     std::string cur_path = std::string(buf);
     std::string err_msg =
       "Unable to initialize the Logger on path(" + cur_path + ")";

nntrainer/optimizers/lr_scheduler_cosine.cpp

@@ -11,7 +11,12 @@
  *
  */
 
+#if defined(_WIN32)
+#define _USE_MATH_DEFINES
+#include <math.h>
+#else
 #include <cmath>
+#endif
 
 #include <common_properties.h>
 #include <lr_scheduler_cosine.h>

nntrainer/tensor/float_tensor.cpp

@@ -11,6 +11,7 @@
 
 #include <iomanip>
 #include <iostream>
+#include <numeric>
 
 #include <blas_interface.h>
 #include <float_tensor.h>

nntrainer/tensor/manager.cpp

@@ -24,11 +24,14 @@
 #include <functional>
 #include <limits>
 #include <stdexcept>
-#include <sys/mman.h>
 #include <sys/stat.h>
-#include <unistd.h>
 #include <vector>
 
+#if !defined(_WIN32)
+#include <sys/mman.h>
+#include <unistd.h>
+#endif
+
 #include <activation_layer.h>
 #include <basic_planner.h>
 #include <bn_layer.h>
@@ -51,6 +54,8 @@
 #include <var_grad.h>
 
 namespace nntrainer {
+
+#if !defined(_WIN32)
 MMapedMemory::MMapedMemory(size_t size, bool allocate_fd_) :
   fd(-1), buf(nullptr), buf_size(0), allocate_fd(allocate_fd_) {
 
@@ -136,6 +141,7 @@ MMapedMemory::~MMapedMemory() noexcept {
   buf_size = 0;
   ml_logd("[MMapedMemory] buf released");
 }
+#endif
 
 void Manager::reinitialize() {
   inputs_v2.clear();

nntrainer/tensor/manager.h

@@ -38,6 +38,8 @@
 
 namespace nntrainer {
 using ExecutionMode = ml::train::ExecutionMode;
+
+#if !defined(_WIN32)
 /**
  * @class MMappedMemory
  * @brief Memory Handler, that has mmaped memory with a file descriptor
@@ -104,6 +106,7 @@ class MMapedMemory {
   size_t buf_size;  /**< buffer size */
   bool allocate_fd; /**< option to choose to allocate an fd */
 };
+#endif
 
 /**
  * @class   Manager

nntrainer/tensor/swap_device.cpp

@@ -16,9 +16,7 @@
 #include <malloc.h>
 #include <profiler.h>
 #include <stdlib.h>
-#include <sys/mman.h>
 #include <sys/types.h>
-#include <unistd.h>
 
 #include <nntrainer_error.h>
 #include <nntrainer_log.h>
@@ -30,8 +28,7 @@ void SwapDevice::start(size_t size) {
   if (fd > 0)
     return;
 
-  fd =
-    open(dev_path.c_str(), O_RDWR | O_CREAT | O_TRUNC | O_SYNC, (mode_t)0666);
+  fd = open(dev_path.c_str(), O_RDWR | O_CREAT | O_TRUNC | O_SYNC, 0666UL);
   NNTR_THROW_IF(fd < 0, std::runtime_error)
     << "SwapDevice: open file: " << dev_path;
 
@@ -153,7 +150,7 @@ void SwapDevice::putBuffer(void *ptr, bool dealloc_only) {
   free(ptr);
   allocated.erase(ptr);
 
-#ifndef __ANDROID__
+#if !defined(__ANDROID__) && !defined(_WIN32)
   malloc_trim(0);
 #endif
 

nntrainer/tensor/swap_device.h

@@ -18,15 +18,22 @@
 #include <map>
 #include <memory>
 #include <string>
-#include <sys/mman.h>
 #include <sys/stat.h>
 #include <sys/types.h>
 #include <system_error>
-#include <unistd.h>
 #include <utility>
 
-/* Uncomment this to use mmap for swap data */
-#define USE_MMAP
+#if defined(_WIN32)
+#include <io.h>
+#define O_SYNC 0UL
+#else
+#include <sys/mman.h>
+#include <unistd.h>
+#endif
+
+#if defined(_WIN32)
+using ssize_t = std::make_signed_t<size_t>;
+#endif
 
 namespace nntrainer {
 
@@ -47,16 +54,14 @@ class SwapDevice {
    *
    */
   explicit SwapDevice(const std::string &name) :
-    dev_path(swap_device_default_path + name),
-    fd(-1) {}
+    dev_path(swap_device_default_path + name), fd(-1) {}
 
   /**
    * @brief SwapDevice default constructor
    *
    */
   explicit SwapDevice(const std::string &path, const std::string &name) :
-    dev_path(path + "/" + name),
-    fd(-1) {}
+    dev_path(path + "/" + name), fd(-1) {}
 
   /**
    * @brief SwapDevice destructor

nntrainer/tensor/task_executor.h

@@ -22,7 +22,6 @@
 #include <memory>
 #include <mutex>
 #include <thread>
-#include <unistd.h>
 
 #include <task.h>
 

nntrainer/tensor/tensor.cpp

@@ -9,6 +9,8 @@
  * @bug		No known bugs except for NYI items
  */
 
+#include <numeric>
+
 #include <char_tensor.h>
 #include <float_tensor.h>
 #include <lazy_tensor.h>

nntrainer/tensor/tensor.h

@@ -1745,7 +1745,7 @@ class Tensor {
    */
   bool isValid() const { return itensor->isValid(); };
 
-  static constexpr float epsilon = 1e-5;
+  static constexpr float epsilon = 1e-5f;
 
 private:
   std::shared_ptr<TensorBase> itensor;

nntrainer/tensor/tensor_base.h

@@ -704,7 +704,7 @@ class TensorBase {
    */
   virtual bool isValid() const = 0;
 
-  static constexpr float epsilon = 1e-5;
+  static constexpr float epsilon = 1e-5f;
 
 protected:
   TensorDim dim;

nntrainer/tensor/weight.h

@@ -364,9 +364,9 @@ class Weight : public Var_Grad {
   const float getLossScale() { return loss_scale; };
 
 private:
-  static constexpr float epsilon = 1e-6; /**< epsilon for zero comparison */
+  static constexpr float epsilon = 1e-6f; /**< epsilon for zero comparison */
   static constexpr float epsilon_decay =
-    1e-8; /**< epsilon for zero comparison */
+    1e-8f; /**< epsilon for zero comparison */
 
   WeightRegularizer regularizer; /**< regularizer for this variable */
   float regularizer_constant;    /**< constant factor for regularization */