CUDA support; bug fixes in cudamatrix/

msbuild/kaldi.props

@@ -26,12 +26,12 @@
     <ClCompile>
       <Optimization>Disabled</Optimization>
       <WarningLevel>Level3</WarningLevel>
-      <PreprocessorDefinitions>WIN32;NDEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+      <PreprocessorDefinitions>WIN32;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <MinimalRebuild>false</MinimalRebuild>
       <BufferSecurityCheck>false</BufferSecurityCheck>
       <EnableEnhancedInstructionSet>$(KaldiCpuInstructionSet)</EnableEnhancedInstructionSet>
       <ErrorReporting>None</ErrorReporting>
-      <WholeProgramOptimization>true</WholeProgramOptimization>
+      <WholeProgramOptimization>false</WholeProgramOptimization>
       <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
       <Cpp0xSupport>true</Cpp0xSupport>
       <!-- These properties affect the Intel compiler only -->
@@ -125,6 +125,30 @@
 
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.Default.props" />
 
+  <!-- CUDA properties may be set up only this late. -->
+  <PropertyGroup>
+    <CUDAPropsPath Condition="'$(CUDAPropsPath)'==''">$(VCTargetsPath)/BuildCustomizations</CUDAPropsPath>
+    <CUDAFound Condition="exists('$(CUDAPropsPath)/CUDA 6.5.props')" >65</CUDAFound>
+    <!--<CUDAFound Condition="exists('$(CUDAPropsPath)\CUDA 7.0.props')" >70</CUDAFound>-->
+  </PropertyGroup>
+
+  <ImportGroup Condition=" '$(KaldiCUDA)' == 'true' and '$(CUDAFound)' != '' ">
+    <Import Project="$(CUDAPropsPath)/CUDA 6.5.props" Condition=" '$(CUDAFound)' == '65' "  />
+  </ImportGroup>
+
+  <ItemDefinitionGroup  Condition=" '$(KaldiCUDA)' == 'true' and '$(CUDAFound)' != '' ">
+    <CudaCompile>
+      <CodeGeneration>compute_20,sm_20;compute_30,sm_30;compute_35,sm_35;compute_37,sm_37;compute_50,sm_50;</CodeGeneration>
+      <!--<Runtime>MT</Runtime>-->
+    </CudaCompile>
+    <ClCompile>
+      <PreprocessorDefinitions>HAVE_CUDA;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+    </ClCompile>
+    <Link>
+      <AdditionalDependencies>cublas.lib;cusparse.lib;cudart_static.lib;cuda.lib;%(AdditionalDependencies)</AdditionalDependencies>
+    </Link>
+  </ItemDefinitionGroup>
+
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.props" />
   <Import Project="$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props" Condition="exists('$(UserRootDir)\Microsoft.Cpp.$(Platform).user.props')" />
 </Project>

msbuild/kaldi.targets

@@ -6,6 +6,8 @@
            Text="Invalid BuildStage. Are you running the top-level build project?"/>
     <Error Condition=" '$(BuildStage)' == 'Binary' and '$(KaldiBinaryTarget)' == '' "
            Text="For the Binary and Test build stages, KaldiBinaryTarget= specification is required. Are you running the top-level build project?"/>
+    <Warning Condition=" '$(KaldiCUDA)' == 'true' and '$(CUDAFound)' == '' "
+             Text="CUDA has been requested in kaldi.user.props but not found. Install CUDA 6.5 (not supporting other versions at the moment)." />
   </Target>
 
   <Target Name="Build" DependsOnTargets="_PrepareBuildStages" Condition=" '$(BuildStage)' == '' ">
@@ -47,14 +49,22 @@
       --><!-- Convert all library references into project references. --><!--
       <ProjectReference Include="@(DependsOnLibs->'$(MSBuildProjectDirectory)\%(relativedir)%(filename).kwproj')" />
     </ItemGroup>-->
+    <!-- Hack CUDA 6.5 into working with Intel XE 15.0 compiler. Other combinations untested. -->
+    <ItemGroup Condition=" '$(PlatformToolset.StartsWith(`Intel C++ Compiler`))' == 'true' ">
+      <CudaCompile>
+        <VCBinDir>$(ICInstallDir)bin\ia32\icl.exe</VCBinDir>
+        <BaseCommandLineTemplate>-ccbin "%(CudaCompile.VCBinDir)" [GenerateRelocatableDeviceCode] [Include] [RequiredIncludes] [CInterleavedPTX] [GPUDebugInfo] [GenerateLineInfo] [Keep] [KeepDir] [MaxRegCount] [PtxAsOptionV] [TargetMachinePlatform] [NvccCompilation] [CudaRuntime] [AdditionalOptions]</BaseCommandLineTemplate>
+        <RuntimeApiCommandLineTemplate>%(CudaCompile.BaseCommandLineTemplate) [HostDebugInfo] [Emulation] [FastMath] [Defines] %(CudaCompile.HostCommandLineTemplate) [CompileOut] "%(CudaCompile.FullPath)"</RuntimeApiCommandLineTemplate>
+      </CudaCompile>
+    </ItemGroup>
   </Target>
 
   <Target Name ="_BeforeBuildLibrary" BeforeTargets="_PrepareForBuild" DependsOnTargets="_BeforeBuildAnyStage" Condition=" '$(BuildStage)' == 'Library' ">
     <ItemGroup>
       <!-- Fetch all *.cc files from <LibSource> into the build as C/C++ sources. -->
       <ClCompile Include="@(LibSource->'$(_KaldiSrcSubdir)%(identity)')" Condition=" '%(extension)' == '.cc' " />
-      <!-- TODO: Fetch all *.cu files as CUDA sources. -->
-
+      <!-- Fetch all *.cu files as CUDA sources. -->
+      <CudaCompile Include="@(LibSource->'$(_KaldiSrcSubdir)%(identity)')" Condition=" '%(extension)' == '.cu' and '$(KaldiCUDA)' == 'true' and '$(CUDAFound)' != '' " />
     </ItemGroup>
   </Target>
 
@@ -97,4 +107,5 @@
   </Target>
 
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" Condition=" '$(BuildStage)' != '' "/>
+  <Import Project="$(CUDAPropsPath)\CUDA 6.5.targets" Condition=" '$(BuildStage)' != '' and '$(KaldiCUDA)' == 'true' and '$(CUDAFound)' == '65' "/>
 </Project>

src/base/kaldi-utils.h

@@ -30,9 +30,12 @@
 # include <windows.h>
 #endif
 
+#if _MSC_VER < 1400
+#define __restrict__
+#endif
+
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4056 4305 4800 4267 4996 4756 4661)
-#define __restrict__
 #endif
 
 #ifdef HAVE_POSIX_MEMALIGN

src/cudamatrix/cu-device.cc

@@ -30,13 +30,15 @@
 #include <string>
 #include <vector>
 #include <algorithm>
+#ifndef _MSC_VER
 #include <dlfcn.h>
-#include <unistd.h> // for sleep
+#endif
 
 #include "cudamatrix/cu-common.h"
 #include "cudamatrix/cu-device.h"
 #include "cudamatrix/cu-matrix.h"
 #include "base/kaldi-error.h"
+#include "base/kaldi-utils.h"
 #include "util/common-utils.h"
 
 namespace kaldi {
@@ -105,7 +107,7 @@ void CuDevice::SelectGpuId(std::string use_gpu) {
       int32 sec_sleep = (use_gpu == "yes" ? 20 : 2);
       KALDI_WARN << "Will try again to get a GPU after " << sec_sleep 
         << " seconds.";
-      sleep(sec_sleep);
+      Sleep(sec_sleep);
       cudaGetLastError(); // reset the error state    
       e = cudaThreadSynchronize(); //<< 2nd trial to get CUDA context.
       if (e != cudaSuccess) {
@@ -128,7 +130,7 @@ void CuDevice::SelectGpuId(std::string use_gpu) {
                    << " seconds to get a GPU.";
       num_times++;
       wait_time += sec_sleep;
-      sleep(sec_sleep);
+      Sleep(sec_sleep);
       cudaGetLastError(); // reset the error state    
       e = cudaThreadSynchronize();
     }
@@ -377,6 +379,9 @@ std::string CuDevice::GetFreeMemory(int64* free, int64* total) const {
     //pre-fill ``safe'' values that will not cause problems
     mem_free = 1; mem_total = 1;
     //open libcuda.so
+#ifdef _MSC_VER
+    cuMemGetInfo(&mem_free, &mem_total);
+#else
     void* libcuda = dlopen("libcuda.so",RTLD_LAZY);
     if(NULL == libcuda) { 
       KALDI_WARN << "cannot open libcuda.so"; 
@@ -399,6 +404,7 @@ std::string CuDevice::GetFreeMemory(int64* free, int64* total) const {
       //close the library
       dlclose(libcuda);
     }
+#endif
   }
   // copy the output values outside
   if(NULL != free) *free = mem_free;
@@ -416,6 +422,9 @@ std::string CuDevice::GetFreeMemory(int64* free, int64* total) const {
 void CuDevice::DeviceGetName(char* name, int32 len, int32 dev) {
   //prefill with something reasonable
   strncpy(name,"Unknown GPU",len);
+#ifdef _MSC_VER
+  cuDeviceGetName(name, len, dev);
+#else
   //open libcuda.so
   void* libcuda = dlopen("libcuda.so",RTLD_LAZY);
   if(NULL == libcuda) {
@@ -434,6 +443,7 @@ void CuDevice::DeviceGetName(char* name, int32 len, int32 dev) {
     //close the library
     dlclose(libcuda);
   }
+#endif
 }
 
 

src/cudamatrix/cu-matrix-speed-test.cc

@@ -519,7 +519,7 @@ template<typename Real> void CudaMatrixSpeedTest() {
 
 
 int main() {
-  for (int32 loop = 1; loop < 2; loop++) {
+  for (int32 loop = 0; loop < 2; loop++) {
 #if HAVE_CUDA == 1
     if (loop == 0)
       CuDevice::Instantiate().SelectGpuId("no");

src/util/kaldi-io-test.cc

@@ -161,14 +161,9 @@ void UnitTestIoNew(bool binary) {
 void UnitTestIoPipe(bool binary) {
   // This is as UnitTestIoNew except with different filenames.
   {
-#ifdef _MSC_VER
-    const char *filename_out = "|more > tmpf.txt",
-        *filename_in = "type tmpf.txt |";
-#else
     const char *filename_out = "|gzip -c > tmpf.gz",
         *filename_in = "gunzip -c tmpf.gz |";
-#endif
-
+
     Output ko(filename_out, binary);
     std::ostream &outfile = ko.Stream();
     if (!binary) outfile << "\t";
@@ -214,10 +209,6 @@ void UnitTestIoPipe(bool binary) {
     if (!binary && Rand()%2 == 0) outfile << "\t";
     bool ans = ko.Close();
     KALDI_ASSERT(ans);
-#ifndef _MSC_VER
-    sleep(1);  // This test does not work without this sleep:
-    // seems to be some kind of file-system latency.
-#endif
     {
       bool binary_in;
       Input ki(filename_in, &binary_in);