Pile of truly horrible hacks to *almost* get things building on windows

18 files changed, 76 insertions, 29 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 54716c0e..6e2d7a69 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -23,9 +23,12 @@ message(STATUS "Project name: ${PROJECT_NAME}")
 message(STATUS "Project version: ${PROJECT_VERSION_STRING_FULL}")
 
 # Set compilation flags
-set(CMAKE_CXX_FLAGS_RELEASE " -std=c++11 -g -O3 -Ofast -m64 -pthread -march=native -Wl,--no-as-needed -funroll-loops -ffinite-math-only -fPIC -Wno-unused-result -Wno-deprecated -Wno-deprecated-gpu-targets")
-set(CMAKE_CXX_FLAGS_DEBUG " -std=c++11 -g -O0 -pthread -fPIC -Wno-unused-result -Wno-deprecated -Wno-deprecated-gpu-targets")
-set(CMAKE_CXX_FLAGS_PROFILE "${CMAKE_CXX_FLAGS_RELEASE}  -pg -g")
+# -Wl,--no-as-needed -Wno-unused-result -Wno-deprecated -Wno-deprecated-gpu-targets -funroll-loops -ffinite-math-only 
+# -Wno-unused-result -Wno-deprecated -Wno-deprecated-gpu-targets
+# -g 
+set(CMAKE_CXX_FLAGS_RELEASE " -std=c++ -O3 -Ofast -m64 -pthread -march=native -fPIC -DBOOST_NO_UNREACHABLE_RETURN_DETECTION")
+set(CMAKE_CXX_FLAGS_DEBUG " -std=c++11 -O0 -pthread -fPIC")
+set(CMAKE_CXX_FLAGS_PROFILE "${CMAKE_CXX_FLAGS_RELEASE}  -pg")
 #set(CMAKE_CXX_FLAGS_PROFGEN "${CMAKE_CXX_FLAGS_RELEASE} -fprofile-generate -fprofile-correction")
 #set(CMAKE_CXX_FLAGS_PROFUSE "${CMAKE_CXX_FLAGS_RELEASE} -fprofile-use -fprofile-correction")
 set(CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS_RELEASE})
@@ -153,6 +156,9 @@ find_package(Boost COMPONENTS ${BOOST_COMPONENTS})
 if(Boost_FOUND)
   include_directories(${Boost_INCLUDE_DIRS})
   set(EXT_LIBS ${EXT_LIBS} ${Boost_LIBRARIES})
+  #message("Found boost with toolset compiler ${_boost_COMPILER}")
+  #set(_boost_COMPILER "-vc140")
+  #message("Overwrote compiler version as ${_boost_COMPILER}")
 else(Boost_FOUND)
   message(SEND_ERROR "Cannot find Boost libraries. Terminating.")
 endif(Boost_FOUND)
diff --git a/cmake/FindMKL.cmake b/cmake/FindMKL.cmake
index 09532baf..e8ce1ade 100644
--- a/cmake/FindMKL.cmake
+++ b/cmake/FindMKL.cmake
@@ -35,10 +35,17 @@ if (MKL_INCLUDE_DIRS AND MKL_LIBRARIES AND MKL_INTERFACE_LIBRARY AND
 endif()
 
 if(NOT BUILD_SHARED_LIBS)
-  set(INT_LIB "libmkl_intel_ilp64.a")
-  set(SEQ_LIB "libmkl_sequential.a")
-  set(THR_LIB "libmkl_intel_thread.a")
-  set(COR_LIB "libmkl_core.a")
+  if (WIN32)
+    set(INT_LIB "mkl_intel_ilp64.lib")
+    set(SEQ_LIB "mkl_sequential.lib")
+    set(THR_LIB "mkl_intel_thread.lib")
+    set(COR_LIB "mkl_core.lib")
+  else()
+    set(INT_LIB "libmkl_intel_ilp64.a")
+    set(SEQ_LIB "libmkl_sequential.a")
+    set(THR_LIB "libmkl_intel_thread.a")
+    set(COR_LIB "libmkl_core.a")
+  endif()
 else()
   set(INT_LIB "mkl_intel_ilp64")
   set(SEQ_LIB "mkl_sequential")
@@ -57,6 +64,7 @@ find_library(MKL_INTERFACE_LIBRARY
              NAMES ${INT_LIB}
              PATHS ${MKL_ROOT}/lib
                    ${MKL_ROOT}/lib/intel64
+                   ${MKL_ROOT}/lib/intel64_win
                    ${INTEL_ROOT}/mkl/lib/intel64
              NO_DEFAULT_PATH)
 
@@ -77,11 +85,17 @@ find_library(MKL_CORE_LIBRARY
 set(MKL_INCLUDE_DIRS ${MKL_INCLUDE_DIR})
 set(MKL_LIBRARIES ${MKL_INTERFACE_LIBRARY} ${MKL_SEQUENTIAL_LAYER_LIBRARY} ${MKL_CORE_LIBRARY})
 
+message("1 ${MKL_INCLUDE_DIR}")
+message("2 ${MKL_INTERFACE_LIBRARY}")
+message("3 ${MKL_SEQUENTIAL_LAYER_LIBRARY}")
+message("4 ${MKL_CORE_LIBRARY}")
+
 if (MKL_INCLUDE_DIR AND
     MKL_INTERFACE_LIBRARY AND
     MKL_SEQUENTIAL_LAYER_LIBRARY AND
     MKL_CORE_LIBRARY)
 
+
     if (NOT DEFINED ENV{CRAY_PRGENVPGI} AND
         NOT DEFINED ENV{CRAY_PRGENVGNU} AND
         NOT DEFINED ENV{CRAY_PRGENVCRAY} AND
@@ -93,7 +107,6 @@ if (MKL_INCLUDE_DIR AND
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DMKL_ILP64 ${ABI}")
 
 else()
-
   set(MKL_INCLUDE_DIRS "")
   set(MKL_LIBRARIES "")
   set(MKL_INTERFACE_LIBRARY "")
diff --git a/cmake_doze.txt b/cmake_doze.txt
new file mode 100644
index 00000000..bdacb8f6
--- /dev/null
+++ b/cmake_doze.txt
diff --git a/src/common/config.h b/src/common/config.h
index f603af61..f1dd02c8 100644
--- a/src/common/config.h
+++ b/src/common/config.h
@@ -1,7 +1,7 @@
 #pragma once
 
-#include <sys/ioctl.h>
-#include <unistd.h>
+//#include <sys/ioctl.h>
+//#include <unistd.h>
 #include <boost/program_options.hpp>
 
 #include "3rd_party/yaml-cpp/yaml.h"
diff --git a/src/common/config_parser.cpp b/src/common/config_parser.cpp
index 04f913b3..e455cc3e 100644
--- a/src/common/config_parser.cpp
+++ b/src/common/config_parser.cpp
@@ -1134,11 +1134,11 @@ std::vector<DeviceId> ConfigParser::getDevices() {
         // does this make sense?
         devices.push_back({ds.size(), DeviceType::gpu});
         for(auto d : ds)
-          devices.push_back({std::stoull(d), DeviceType::gpu});
+          devices.push_back({(size_t)std::stoull(d), DeviceType::gpu});
       }
     } else {
       for(auto d : Split(devicesStr))
-        devices.push_back({std::stoull(d), DeviceType::gpu});
+        devices.push_back({(size_t)std::stoull(d), DeviceType::gpu});
     }
 
     if(config_["cpu-threads"].as<size_t>() > 0) {
diff --git a/src/common/config_parser.h b/src/common/config_parser.h
index c51fb0ae..a9242e6a 100644
--- a/src/common/config_parser.h
+++ b/src/common/config_parser.h
@@ -2,8 +2,8 @@
 
 #include <boost/program_options.hpp>
 
-#include <sys/ioctl.h>
-#include <unistd.h>
+//#include <sys/ioctl.h>
+//#include <unistd.h>
 #include "3rd_party/yaml-cpp/yaml.h"
 #include "common/definitions.h"
 #include "common/file_stream.h"
diff --git a/src/common/file_stream.h b/src/common/file_stream.h
index 53d41faf..c5c4492f 100644
--- a/src/common/file_stream.h
+++ b/src/common/file_stream.h
@@ -1,9 +1,10 @@
 #pragma once
 
+
 #include <boost/filesystem.hpp>
 #include <boost/filesystem/fstream.hpp>
 #include <boost/iostreams/device/file_descriptor.hpp>
-#include <boost/iostreams/filter/gzip.hpp>
+//#include <boost/iostreams/filter/gzip.hpp>
 #include <boost/iostreams/filtering_stream.hpp>
 #include <iostream>
 
@@ -21,11 +22,15 @@ private:
   std::string name_;
 
   int mkstemp_and_unlink(char* tmpl) {
+    ABORT_IF(true, "NYI");
+    return 0;
+    /*
     int ret = mkstemp(tmpl);
     if(unlink_ && ret != -1) {
       ABORT_IF(unlink(tmpl), "Error while deleting '{}'", tmpl);
     }
     return ret;
+    */
   }
 
   int MakeTemp(const std::string& base) {
@@ -41,6 +46,9 @@ private:
   }
 
   void NormalizeTempPrefix(std::string& base) {
+    ABORT_IF(true, "NYI");
+    return;
+    /*
     if(base.empty())
       return;
     if(base[base.size() - 1] == '/')
@@ -51,11 +59,13 @@ private:
       return;
     if(S_ISDIR(sb.st_mode))
       base += '/';
+      */
   }
 
 public:
   TemporaryFile(const std::string base = "/tmp/", bool earlyUnlink = true)
       : unlink_(earlyUnlink) {
+      ABORT_IF(true, "NYI (see destructor)");
     std::string baseTemp(base);
     NormalizeTempPrefix(baseTemp);
     fd_ = MakeTemp(baseTemp);
@@ -65,10 +75,12 @@ public:
     if(fd_ != -1 && !unlink_) {
       ABORT_IF(unlink(name_.c_str()), "Error while deleting '{}'", name_);
     }
+    /*
     if(fd_ != -1 && close(fd_)) {
       std::cerr << "Could not close file " << fd_ << std::endl;
       std::abort();
     }
+    */
   }
 
   int getFileDescriptor() { return fd_; }
@@ -82,14 +94,15 @@ public:
     ABORT_IF(
         !boost::filesystem::exists(file_), "File '{}' does not exist", file);
 
-    if(file_.extension() == ".gz")
-      istream_.push(io::gzip_decompressor());
+    //if(file_.extension() == ".gz")
+    //  istream_.push(io::gzip_decompressor());
     istream_.push(ifstream_);
   }
 
   InputFileStream(TemporaryFile& tempfile)
       : fds_(tempfile.getFileDescriptor(), io::never_close_handle) {
-    lseek(tempfile.getFileDescriptor(), 0, SEEK_SET);
+    ABORT_IF(true, "NYI");
+    //lseek(tempfile.getFileDescriptor(), 0, SEEK_SET);
     istream_.push(fds_, 1024);
   }
 
@@ -122,14 +135,15 @@ public:
     ABORT_IF(
         !boost::filesystem::exists(file_), "File '{}' does not exist", file);
 
-    if(file_.extension() == ".gz")
-      ostream_.push(io::gzip_compressor());
+    //if(file_.extension() == ".gz")
+    //  ostream_.push(io::gzip_compressor());
     ostream_.push(ofstream_);
   }
 
   OutputFileStream(TemporaryFile& tempfile)
       : fds_(tempfile.getFileDescriptor(), io::never_close_handle) {
-    lseek(tempfile.getFileDescriptor(), 0, SEEK_SET);
+    ABORT_IF(true, "NYI");
+    //lseek(tempfile.getFileDescriptor(), 0, SEEK_SET);
     ostream_.push(fds_, 1024);
   }
 
diff --git a/src/common/logging.cpp b/src/common/logging.cpp
index 8ef5fada..e89cb5f7 100644
--- a/src/common/logging.cpp
+++ b/src/common/logging.cpp
@@ -37,7 +37,7 @@ bool setLoggingLevel(spdlog::logger& logger, std::string const level) {
     logger.set_level(spdlog::level::info);
   else if(level == "warn")
     logger.set_level(spdlog::level::warn);
-  else if(level == "err" or level == "error")
+  else if(level == "err" || level == "error")
     logger.set_level(spdlog::level::err);
   else if(level == "critical")
     logger.set_level(spdlog::level::critical);
diff --git a/src/common/utils.cpp b/src/common/utils.cpp
index c69085e6..895f9062 100644
--- a/src/common/utils.cpp
+++ b/src/common/utils.cpp
@@ -64,6 +64,9 @@ std::string Join(const std::vector<std::string>& words,
 }
 
 std::string Exec(const std::string& cmd) {
+  ABORT_IF(true, "NYI");
+  return "";
+  /*
   std::array<char, 128> buffer;
   std::string result;
   std::shared_ptr<std::FILE> pipe(popen(cmd.c_str(), "r"), pclose);
@@ -75,4 +78,5 @@ std::string Exec(const std::string& cmd) {
       result += buffer.data();
   }
   return result;
+  */
 }
diff --git a/src/data/corpus.cpp b/src/data/corpus.cpp
index e8dcfff0..7d4f6488 100644
--- a/src/data/corpus.cpp
+++ b/src/data/corpus.cpp
@@ -1,4 +1,5 @@
 #include <random>
+#include <numeric>
 
 #include "data/corpus.h"
 
diff --git a/src/functional/shape.h b/src/functional/shape.h
index 3212a3ed..62e51f51 100644
--- a/src/functional/shape.h
+++ b/src/functional/shape.h
@@ -36,7 +36,7 @@ struct ConstantShape {
         bstride_(shape.bstride_),
         elements_(shape.elements_) {}
 
-  ConstantShape(const Shape& shape) {
+  ConstantShape(const marian::Shape& shape) {
     size_t filled = shape.size();
 
     ABORT_IF(filled > N,
diff --git a/src/graph/node_operators_unary.h b/src/graph/node_operators_unary.h
index 273adf44..e14f6546 100644
--- a/src/graph/node_operators_unary.h
+++ b/src/graph/node_operators_unary.h
@@ -293,7 +293,7 @@ struct PReLUNodeOp : public UnaryNodeOp {
   }
 
 private:
-  float alpha_{0.01};
+  float alpha_{0.01f};
 };
 
 /**
diff --git a/src/tensors/cpu/device.cpp b/src/tensors/cpu/device.cpp
index 985b00f5..7d0d9feb 100644
--- a/src/tensors/cpu/device.cpp
+++ b/src/tensors/cpu/device.cpp
@@ -1,6 +1,9 @@
 #include "tensors/device.h"
 #include <iostream>
 
+//#if DOZE
+#include <malloc.h>
+
 #include <stdlib.h>
 
 namespace marian {
@@ -12,6 +15,11 @@ Device::~Device() {
   size_ = 0;
 }
 
+// #if DOZE
+void* aligned_alloc(size_t alignment, size_t size) {
+    return _aligned_malloc(size, alignment);
+}
+
 void Device::reserve(size_t size) {
   size = align(size);
   ABORT_IF(size < size_ || size == 0,
diff --git a/src/training/graph_group_async.h b/src/training/graph_group_async.h
index b9568735..d0687b80 100644
--- a/src/training/graph_group_async.h
+++ b/src/training/graph_group_async.h
@@ -42,7 +42,7 @@ protected:
   std::vector<Tensor> paramsAvg_;
   std::vector<Ptr<TensorAllocator>> paramsAllocAvg_;
   bool movingAvg_{false};
-  float mvDecay_{1e-4};
+  float mvDecay_{1e-4f};
 
   std::unique_ptr<ThreadPool> pool_;
 
diff --git a/src/training/graph_group_async_drop.cpp b/src/training/graph_group_async_drop.cpp
index 5ab2b46a..dd56ee21 100644
--- a/src/training/graph_group_async_drop.cpp
+++ b/src/training/graph_group_async_drop.cpp
@@ -70,9 +70,10 @@ void AsyncGraphGroupDrop::fetchParams(Tensor oldParams,
 
     pos += shardSize_;
   }
+  /* OMG BAD
   for(auto&& t : threads) {
     t.join();
-  }
+  }*/
   fetchStep_[device_id]++;
 }
 
diff --git a/src/training/graph_group_singleton.h b/src/training/graph_group_singleton.h
index 5562d302..73cdaaac 100644
--- a/src/training/graph_group_singleton.h
+++ b/src/training/graph_group_singleton.h
@@ -21,7 +21,7 @@ private:
 
   Ptr<ExpressionGraph> mvAvgGraph_;
   bool mvAvg_{false};
-  float mvDecay_{1e-4};
+  float mvDecay_{1e-4f};
 
   void updateMovingAverage(Tensor mvAvgParams, Tensor params, size_t batches);
 
diff --git a/src/training/graph_group_sync.h b/src/training/graph_group_sync.h
index dac7e402..ecc1a8e8 100644
--- a/src/training/graph_group_sync.h
+++ b/src/training/graph_group_sync.h
@@ -29,7 +29,7 @@ private:
   std::vector<Tensor> paramsAvg_;
   std::vector<Ptr<TensorAllocator>> paramsAllocAvg_;
   bool movingAvg_{false};
-  float mvDecay_{1e-4};
+  float mvDecay_{1e-4f};
   size_t delay_{1};
 
   void updateMovingAverage(Tensor paramsAvg, Tensor params, size_t batches);
diff --git a/src/translator/beam_search.h b/src/translator/beam_search.h
index faf21c73..e476c13a 100644
--- a/src/translator/beam_search.h
+++ b/src/translator/beam_search.h
@@ -27,7 +27,7 @@ public:
                       ? options_->get<size_t>("beam-size")
                       : 3) {}
 
-  Beams toHyps(const std::vector<uint> keys,
+  Beams toHyps(const std::vector<uint32_t> keys,
                const std::vector<float> costs,
                size_t vocabSize,
                const Beams& beams,